Datasets:

math-ai

/

AutoMathText

like 125

https://arxiv.org/abs/2105.10615

Convergence directions of the randomized Gauss--Seidel method and its extension

The randomized Gauss--Seidel method and its extension have attracted much attention recently and their convergence rates have been considered extensively. However, the convergence rates are usually determined by upper bounds, which cannot fully reflect the actual convergence. In this paper, we make a detailed analysis of their convergence behaviors. The analysis shows that the larger the singular value of $A$ is, the faster the error decays in the corresponding singular vector space, and the convergence directions are mainly driven by the large singular values at the beginning, then gradually driven by the small singular values, and finally by the smallest nonzero singular value. These results explain the phenomenon found in the extensive numerical experiments appearing in the literature that these two methods seem to converge faster at the beginning. Numerical examples are provided to confirm the above findings.

\section{Introduction} Linear least squares problem is a ubiquitous problem arising frequently in data analysis and scientific computing. Specifically, given a data matrix $A\in R^{m\times n}$ and a data vector $b\in R^{m}$, a linear least squares problem can be written as follows \begin{equation} \label{ls} \min \limits _{ x \in R^{n}}\|b-Ax\|^2_{2}. \end{equation} In the literature, several direct methods have been proposed for solving its normal equations $A^TAx=A^Tb$ through either the QR factorization or the singular value decomposition (SVD) of $A^TA$ \cite{bjorck1996numerical, Higham2002}, which can be prohibitive when the matrix is large--scale. Hence, iterative methods are considered for solving large linear least squares problem, such as the famous Gauss--Seidel method \cite{Saad2003}. In \cite{Leventhal2010}, Leventhal and Lewis proved that the randomized Gauss--Seidel (RGS) method, also known as the randomized coordinate descent method, converges to the solution at a linear rate in expectation. This method works on the columns of the matrix $A$ at random with probability proportional to their norms. Later, Ma, Needell and Ramdas \cite{Ma2015} provided a unified theory of the RGS method and the randomized Kaczmarz (RK) method \cite{Strohmer2009}, where the latter method works on the rows of $A$, and showed that the RGS method converges to the minimum Euclidean norm least squares solution $x_{\star}$ of (\ref{ls}) only when the matrix $A$ is of full column rank. To further develop the RGS method for more general matrix, inspired by the randomized extended Kaczmarz (REK) method \cite{Completion2013}, Ma et al. \cite{Ma2015} presented a variant of the RGS mehtod, i.e., randomized extended Gauss--Seidel (REGS) method, and proved that the REGS method converges to $x_{\star}$ regardless of whether the matrix $A$ has full column rank. After that, many variants of the RGS (or REGS) method were developed and studied extensively; see for example \cite{gower2015randomized, nutini2015coordinate, Hefny2017,tu2017breaking, xu2018hybrid,Dukui2019,razaviyayn2019linearly} and references therein. To the best of our knowledge, when studying the convergence properties of the RGS and REGS methods, people mainly pay attention to their convergence rates and usually give corresponding upper bounds, and no work focuses on what determines their convergence rates, what drives their convergence directions, and what their ultimate directions is. As we know, the obtained upper bound of convergence can only be used as a reference for the convergence rate, and cannot truly reflect the empirical convergence of the method. So it is interesting to consider the above three problems. In 2017, Jiao, Jin and Lu \cite{jiao2017preasymptotic} analyzed the preasymptotic convergence of the RK method. Recently, Steinerberger \cite{steinerberger2021randomized} made a more detailed analysis of the convergence property of the RK method for overdetermined full rank linear system. The author showed that the right singular vectors of the matrix $A$ describe the directions of distinguished dynamics and the RK method converges along small right singular vectors. After that, Zhang and Li \cite{zhang2021preconvergence} considered the convergence property of the REK method for all types of linear systems (consistent or inconsistent, overdetermined or underdetermined, full-rank or rank-deficient) and showed that the REK method converges to the minimum Euclidean norm least squares solution $x_{\star}$ with different decay rates in different right singular vectors spaces. In this paper, we analyze the convergence properties of the RGS and REGS methods for linear least squares problem and show that the decay rates of the sequences $\{Ax_{k}\}_{k=1}^{\infty}$ and $\{ x_{k}\}_{k=1}^{\infty}$ (resp., the sequences $\{Az_{k}\}_{k=1}^{\infty}$ and $\{ z_{k}\}_{k=1}^{\infty}$ ) generated by the RGS method (resp., the REGS method) are depend on the size of singular values of $A$. Specifically, the larger the singular value of $A$ is, the faster the error decays in the corresponding singular vector space, and the convergence directions are mainly driven by the large singular values at the beginning, then gradually driven by the small singular values, and finally by the smallest nonzero singular value. The rest of this paper is organized as follows. We first introduce some notations and preliminaries in Section \ref{sec2} and then present our main results about the RGS and REGS methods in Section \ref{sec3} and Section \ref{sec4}, respectively. Numerical experiments are given in Section \ref{sec5}. \section{Notations and preliminaries }\label{sec2} Throughout the paper, for a matrix $A$, $A^T$, $A^{(i)}$, $A_{(j)}$, $\sigma_i(A)$, $\sigma_r(A)$, $\|A\|_F$, and $\mathcal{R}(A)$ denote its transpose, $i$th row (or $i$th entry in the case of a vector), $j$th column, $i$th singular value, smallest nonzero singular value, Frobenius norm, and column space, respectively. For any integer $m\geq1$, let $[m]:=\{1, 2, 3, ..., m\}$. If the matrix $G\in R^{n\times n}$ is positive definite, we define the energy norm of any vector $x\in R^{n}$ as $\| x\|_G:=\sqrt{x^TGx}$. In addition, we denote the identity matrix by $I$, its $j$th column by $e_{(j)}$ and the expectation of any random variable $\xi$ by $\mathbb{E} [\xi]$. In the following, we use $x_{\star}=A^{\dag}b$ to denote the minimum Euclidean norm least squares solution of (\ref{ls}), where $A^{\dag}$ denotes the Moore--Penrose pseudoinverse of the matrix $A$. Because the SVD is the basic tool for the convergence analysis in next two sections, we denote the SVD \cite{golub2013matrix} of $A\in R^{m\times n}$ by \begin{align} A=U\Sigma V^{T}, \notag \end{align} where $U=[u_1, u_2, \ldots u_m]\in R^{m\times m}$ and $V=[v_1, v_2, \ldots v_n]\in R^{n\times n}$ are column orthonormal matrices and their column vectors known as the left and right singular vectors, respectively, and $\Sigma\in R^{m\times n}$ is diagonal with the diagonal elements ordered nonincreasingly, i.e., $\sigma_1(A)\geq \sigma_2(A)\geq \ldots \sigma_r(A)>0$ with $r\leq \min\{m, n\}$. \section{Convergence directions of the RGS method}\label{sec3} We first list the RGS method \cite{Leventhal2010, Ma2015} in Algorithm \ref{alg1} and restate its convergence bound in Theorem \ref{theorem0}. \begin{alg} \label{alg1} The RGS method \begin{enumerate}[ \item \mbox{INPUT:} ~$A$, $b$, $\ell$, $x_{0}\in R^{n }$ \item For $k=1, 2, \ldots, \ell-1$ do \item ~~~~Select $j\in [n]$ with probability $\frac{\|A_{(j)}\|^2_2}{\|A\|^2_F}$ \item ~~~~Set $x_{k}=x_{k-1}-\frac{A_{(j)}^T ( Ax_{k-1}-b)}{ \| A_{(j)} \|_{2}^{2}}e_{(j)}$ \item End for \end{enumerate} \end{alg} \begin{thm} \label{theorem0}\cite{Leventhal2010, Ma2015} Let $A\in R^{m\times n}$, $b\in R^{m}$, $x_{\star}=A^{\dag}b$ be the minimum Euclidean norm least squares solution, and $x_k$ be the $k$th approximation of the RGS method generated by Algorithm \ref{alg1} with initial guess $x_{0}\in R^{n }$. Then \begin{align} \mathbb{E}[\| Ax_{k}- Ax_{\star}\|_2^2]\leq(1-\frac{\sigma_r^2(A)}{\|A\|^2_F})^k \|Ax_{0}- Ax_{\star}\|_2^2.\label{th0} \end{align} \end{thm} \begin{rmk} \label{rmk1} Theorem \ref{theorem0} shows that $Ax_{k}$ converges linearly in expectation to $Ax_{\star}$ regardless of whether the matrix $A$ has full rank. Since $\| Ax_{k}- Ax_{\star}\|_2^2=\| x_{k}- x_{\star}\|_{A^TA}^2$, it follows from (\ref{th0}) that \begin{align} \mathbb{E}[\| x_{k}- x_{\star}\|_{A^TA}^2]\leq(1-\frac{\sigma_r^2(A)}{\|A\|^2_F})^k \| x_{0}- x_{\star}\|_{A^TA}^2,\notag \end{align} which implies that $ x_{k}$ converges linearly in expectation to the minimum Euclidean norm least squares solution $ x_{\star}$ when the matrix $A$ is overdetermined and of full column rank, but can not converge to $ x_{\star}$ when $A$ is not full column rank. So, we assume that the matrix $A$ is of full column rank in this section. \end{rmk} Now, we give our three main results of the RGS method. \begin{thm} \label{theorem1} Let $A\in R^{m\times n}$, $b\in R^{m}$, $x_{\star}=A^{\dag}b$ be the minimum Euclidean norm least squares solution, and $x_k$ be the $k$th approximation of the RGS method generated by Algorithm \ref{alg1} with initial guess $x_{0}\in R^{n }$. Then \begin{align} \mathbb{E}[\langle Ax_{k}- Ax_{\star}, u_{\ell} \rangle]= (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle Ax_{0}- Ax_{\star}, u_{\ell} \rangle.\label{th1} \end{align} \end{thm} \begin{pf} Let $\mathbb{E}_{k-1}[\cdot]$ be the conditional expectation conditioned on the first $k-1$ iterations of the RGS method. Then, from Algorithm \ref{alg1}, we have \begin{align} & \mathbb{E}_{k-1}[\langle Ax_{k}- Ax_{\star}, u_{\ell} \rangle]\notag \\ &= \sum\limits_{j=1}^{n}\frac{ \|A_{ (j )} \|_{2}^{2}}{\|A\|_{F}^{2}} \langle Ax_{k-1}-\frac{A_{ (j )}^T(Ax_{k-1}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )}- Ax_{\star}, u_{\ell} \rangle \notag \\ &= \langle Ax_{k-1}- Ax_{\star}, u_{\ell} \rangle - \frac{ 1}{\|A\|_{F}^{2}} \sum\limits_{j=1}^{n} \langle A_{ (j )}^T(Ax_{k-1}-b) A_{ (j )} , u_{\ell} \rangle \notag \\ &= \langle Ax_{k-1}- Ax_{\star}, u_{\ell} \rangle - \frac{ 1}{\|A\|_{F}^{2}} \sum\limits_{j=1}^{n} \langle A_{ (j )},Ax_{k-1}-b \rangle \langle A_{ (j )}, u_{\ell} \rangle \notag \\ &= \langle Ax_{k-1}- Ax_{\star}, u_{\ell} \rangle - \frac{ 1}{\|A\|_{F}^{2}} \langle A^T(Ax_{k-1}-b) , A^Tu_{\ell} \rangle, \notag \end{align} which together with the facts $A^T(b-Ax_{\star})=0$ and $A^Tu_{\ell}=\sigma_{\ell}(A) v_{\ell} $ yields \begin{align} & \mathbb{E}_{k-1}[\langle Ax_{k}- Ax_{\star}, u_{\ell} \rangle]\notag \\ &= \langle Ax_{k-1}- Ax_{\star}, u_{\ell} \rangle - \frac{ 1}{\|A\|_{F}^{2}} \langle A^T(Ax_{k-1}-Ax_{\star}) , A^Tu_{\ell} \rangle \notag \\ &= \langle Ax_{k-1}- Ax_{\star}, u_{\ell} \rangle - \frac{ 1}{\|A\|_{F}^{2}} \langle A^T(\sum\limits_{i=1}^{m} \langle Ax_{k-1}-Ax_{\star}, u_i\rangle u_i) , A^Tu_{\ell} \rangle \notag \\ &= \langle Ax_{k-1}- Ax_{\star}, u_{\ell} \rangle - \frac{ 1}{\|A\|_{F}^{2}} \langle (\sum\limits_{i=1}^{m} \langle Ax_{k-1}-Ax_{\star}, u_i\rangle \sigma_{i}(A) v_{i}) , \sigma_{\ell}(A) v_{\ell} \rangle \notag \\ &= \langle Ax_{k-1}- Ax_{\star}, u_{\ell} \rangle - \frac{\sigma_{\ell}^2(A)}{\|A\|_{F}^{2}} \langle Ax_{k-1}-Ax_{\star}, u_{\ell}\rangle \notag \\ &= (1- \frac{\sigma_{\ell}^2(A)}{\|A\|_{F}^{2}} ) \langle Ax_{k-1}-Ax_{\star}, u_{\ell}\rangle. \notag \end{align} Thus, by taking the full expectation on both sides, we have \begin{align} \mathbb{E}[\langle Ax_{k}- Ax_{\star}, u_{\ell} \rangle] = (1- \frac{\sigma_{\ell}^2(A)}{\|A\|_{F}^{2}} ) \mathbb{E}[ \langle Ax_{k-1}-Ax_{\star}, u_{\ell}\rangle ] = \ldots= (1- \frac{\sigma_{\ell}^2(A)}{\|A\|_{F}^{2}} )^k \langle Ax_{0}-Ax_{\star}, u_{\ell}\rangle, \notag \end{align} which is the estimate (\ref{th1}). \end{pf} \begin{rmk} \label{rmk2} Theorem \ref{theorem1} shows that the decay rates of $\|Ax_k-Ax_{\star}\|_2$ are different in different left singular vectors spaces. Specifically, the decay rates are dependent on the singular values: the larger the singular value of $A$ is, the faster the error decays in the corresponding left singular vector space. This implies that the smallest singular value will lead to the slowest rate of convergence, which is the one in (\ref{th0}). So, the convergence bound presented in \cite{Leventhal2010, Ma2015} is optimal. \end{rmk} \begin{rmk} \label{rmk3} Let $r_k=b-Ax_k$ be the residual vector with respect to the $k$-th approximation $x_k$, and $r_{\star}=b-Ax_{\star}$ be the true residual vector with respect to the minimum Euclidean norm least squares solution $x_{\star}$. It follows from (\ref{th1}) and $Ax_k-Ax_{\star}=-( r_{k}- r_{\star})$ that \begin{align} \mathbb{E}[\langle r_{k}- r_{\star}, u_{\ell} \rangle]= (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle r_{0}- r_{\star}, u_{\ell} \rangle.\notag \end{align} Hence, Theorem \ref{theorem1} also implies that the decay rates of $ \| r_{k}- r_{\star}\|_2 $ of the RGS method depend on the singular values. \end{rmk} \begin{rmk} \label{rmk4} Using the facts $\langle Ax_{k}- Ax_{\star}, u_{\ell} \rangle=\langle x_{k}- x_{\star},A^Tu_{\ell} \rangle$ and $A^Tu_{\ell}=\sigma_{\ell}(A) v_{\ell} $, from (\ref{th1}), we have \begin{align} \mathbb{E}[\langle x_{k}- x_{\star}, v_{\ell} \rangle]= (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle x_{0}- x_{\star}, v_{\ell} \rangle,\notag \end{align} which recovers the decay rates of the RK method in different right singular vectors spaces \cite{steinerberger2021randomized}. In this view, both RGS and RK methods are essentially equivalent. \end{rmk} \begin{thm} \label{theorem2} Let $A\in R^{m\times n}$, $b\in R^{m}$, $x_{\star}=A^{\dag}b$ be the minimum Euclidean norm least squares solution, and $x_k$ be the $k$th approximation of the RGS method generated by Algorithm \ref{alg1} with initial guess $x_{0}\in R^{n }$. Then \begin{align} \mathbb{E}[\| Ax_{k}- Ax_{\star}\|_2^2]= \mathbb{E}[(1-\frac{1}{\|A\|^2_F}\|A^T\frac{Ax_{k-1}- Ax_{\star}}{\|Ax_{k-1}- Ax_{\star}\|_2}\|_2^2)\| Ax_{k-1}- Ax_{\star}\|_2^2]. \notag \end{align} \end{thm} \begin{pf} Similar to the proof of \cite{Ma2015}, we can derive the desired result. \end{pf} \begin{rmk} \label{rmk5} Since $\|A^T\frac{Ax_{k-1}- Ax_{\star}}{\|Ax_{k-1}- Ax_{\star}\|_2}\|_2^2\geq\sigma_r^2(A)$, Theorem \ref{theorem2} implies that the RGS method actually converges faster if $Ax_{k-1}-Ax_{\star}$ is not close to left singular vectors corresponding to the small singular values of $A$ . \end{rmk} \begin{thm} \label{theorem3} Let $A\in R^{m\times n}$, $b\in R^{m}$, $x_{\star}=A^{\dag}b$ be the minimum Euclidean norm least squares solution, and $x_k$ be the $k$th approximation of the RGS method generated by Algorithm \ref{alg1} with initial guess $x_{0}\in R^{n }$. Then \begin{align} \mathbb{E}[\langle \frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}, \frac{Ax_{k+1}- Ax_{\star}}{\|Ax_{k+1}- Ax_{\star}\|_2} \rangle^2]= 1 -\frac{1}{\|A\|^2_F} \mathbb{E}[ \|A^T\frac{Ax_{k }- Ax_{\star}}{\|Ax_{k }- Ax_{\star}\|_2}\|_2^2 ]. \label{th3} \end{align} \end{thm} \begin{pf} From Algorithm \ref{alg1}, we have \begin{align} & \mathbb{E}_{k}[\langle \frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}, \frac{Ax_{k+1}- Ax_{\star}}{\|Ax_{k+1}- Ax_{\star}\|_2} \rangle^2]\notag \\ &= \mathbb{E}_{k}[\langle \frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}, \frac{Ax_{k}-\frac{A_{ (j )}^T(Ax_{k}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )}- Ax_{\star}}{\|Ax_{k}-\frac{A_{ (j )}^T(Ax_{k}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )}- Ax_{\star}\|_2} \rangle^2]\notag \\ &= \mathbb{E}_{k}[\frac{1}{\|Ax_{k}- Ax_{\star}\|_2^2\cdot \|Ax_{k}-\frac{A_{ (j )}^T(Ax_{k}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )}- Ax_{\star}\|_2^2}\langle Ax_{k}- Ax_{\star} , Ax_{k}-\frac{A_{ (j )}^T(Ax_{k}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )}- Ax_{\star} \rangle^2]. \notag \end{align} Since $\langle Ax_{k}- Ax_{\star} , Ax_{k}-\frac{A_{ (j )}^T(Ax_{k}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )}- Ax_{\star} \rangle=\|Ax_{k}-\frac{A_{ (j )}^T(Ax_{k}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )}- Ax_{\star}\|_2^2$, we have \begin{align} & \mathbb{E}_{k}[\langle \frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}, \frac{Ax_{k+1}- Ax_{\star}}{\|Ax_{k+1}- Ax_{\star}\|_2} \rangle^2]\notag \\ &= \mathbb{E}_{k}[\frac{ 1}{\|Ax_{k}- Ax_{\star}\|_2^2 }\|Ax_{k}-\frac{A_{ (j )}^T(Ax_{k}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )}- Ax_{\star}\|_2^2 ] \notag \\ &= \mathbb{E}_{k}[\frac{ 1}{\|Ax_{k}- Ax_{\star}\|_2^2 }(\|Ax_{k} - Ax_{\star}\|_2^2-2 \langle Ax_{k}- Ax_{\star}, \frac{A_{ (j )}^T(Ax_{k}-b)}{\|A_{ (j )} \|_{2}^{2}}A_{ (j )} \rangle +\frac{(A_{ (j )}^T(Ax_{k}-b))^2}{\|A_{ (j )} \|_{2}^{2}}) ] \notag \\ &= \mathbb{E}_{k}[\frac{ 1}{\|Ax_{k}- Ax_{\star}\|_2^2 }(\|Ax_{k} - Ax_{\star}\|_2^2-\frac{(A_{ (j )}^T(Ax_{k}-Ax_{\star}))^2}{\|A_{ (j )} \|_{2}^{2}}) ] \notag \\ &= \mathbb{E}_{k}[1-\frac{(A_{ (j )}^T \frac{Ax_{k}-Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2 } )^2}{\|A_{ (j )} \|_{2}^{2}} ] \notag \\ &= \sum\limits_{j=1}^{n}\frac{ \|A_{ (j )} \|_{2}^{2}}{\|A\|_{F}^{2}} (1-\frac{(A_{ (j )}^T \frac{Ax_{k}-Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2 } )^2}{\|A_{ (j )} \|_{2}^{2}}) \notag \\ &= 1-\frac{ 1}{\|A\|_{F}^{2}} \| A^T \frac{Ax_{k}-Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2 } \|_2^2. \notag \end{align} Thus, by taking the full expectation on both sides, we obtain the desired result (\ref{th3}). \end{pf} \begin{rmk} \label{rmk6} Let $u$ and $v$ are two unit vectors, i.e., $\|u\|_2=1$ and $\|v\|_2=1$. We use inner quantity $\langle u, v\rangle^2$ to represent the angle between $u$ and $v$, and the bigger the angle is, the bigger the fluctuation becomes from $u$ to $v$. Theorem \ref{theorem3} shows the fluctuation of two adjacent iterations. Specifically, when $\|A^T\frac{Ax_{k }- Ax_{\star}}{\|Ax_{k }- Ax_{\star}\|_2}\|_2^2$ is large, the angle between $\frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}$ and $\frac{Ax_{k+1}- Ax_{\star}}{\|Ax_{k+1}- Ax_{\star}\|_2}$ is large, which implies that $\frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}$ has a large fluctuation; when $\|A^T\frac{Ax_{k }- Ax_{\star}}{\|Ax_{k }- Ax_{\star}\|_2}\|_2^2$ is small, the angle between $\frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}$ and $\frac{Ax_{k+1}- Ax_{\star}}{\|Ax_{k+1}- Ax_{\star}\|_2}$ is small, which implies that $\frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}$ has very little fluctuation. Since $\|A^T\frac{Ax_{k}- Ax_{\star}}{\|Ax_{k}- Ax_{\star}\|_2}\|_2^2\geq\sigma_r^2(A)$, Theorem \ref{theorem3} implies that if $Ax_{k}-Ax_{\star}$ is mainly composed of left singular vectors corresponding to the small singular values of $A$, its direction hardly changes, which means that the RGS method finally converges along left singular vector corresponding to the small singular value of $A$. \end{rmk} \section{Convergence directions of the REGS method}\label{sec4} Recalling Remark \ref{rmk1}, when the matrix $A$ is not full column rank, the sequence $\{x_{k}\}_{k=1}^{\infty}$ generated by the RGS method does not converge to the minimum Euclidean norm least squares solution $x_{\star}$, even though $Ax_{k}$ does converge to $Ax_{\star}$. In \cite{Ma2015}, Ma et al. proposed an extended variant of the RGS method, i.e., the REGS method, to allow for convergence to $x_{\star}$ regardless of whether $A$ has full column rank or not. Now, we list the REGS method presented in \cite{ Dukui2019} in Algorithm \ref{alg2}, which is a equivalent variant of the original REGS method \cite{Ma2015}, and restate its convergence bound presented in \cite{Dukui2019} in Theorem \ref{theorem5}. From the algorithm we find that, in each iteration, $x_k$ is the $k$th approximation of the RGS method and $z_k$ is a one-step RK update for the linear system $Az=Ax_{k}$ from $z_{k-1}$. \begin{alg} \label{alg2} The REGS method \begin{enumerate}[ \item \mbox{INPUT:} ~$A$, $b$, $\ell$, $x_{0}\in R^{n }$, $z_{0}\in \mathcal{R}(A^T)$ \item For $k=1, 2, \ldots, \ell-1$ do \item ~~~~Select $j\in [n]$ with probability $\frac{\|A_{(j)}\|^2_2}{\|A\|^2_F}$ \item ~~~~Set $x_{k}=x_{k-1}-\frac{A_{(j)}^T ( Ax_{k-1}-b)}{ \| A_{(j)} \|_{2}^{2}}e_{(j)}$ \item ~~~~Select $i\in [m]$ with probability $\frac{\|A^{(i)}\|^2_2}{\|A\|^2_F}$ \item ~~~~Set $z_{k}=z_{k-1}-\frac{A^{(i)} ( z_{k-1}-x_{k})}{ \| A^{(i)} \|_{2}^{2}}(A^{(i)})^T$ \item End for \end{enumerate} \end{alg} \begin{thm} \label{theorem5}\cite{Dukui2019} Let $A\in R^{m\times n}$, $b\in R^{m}$, $x_{\star}=A^{\dag}b$ be the minimum Euclidean norm least squares solution, and $z_k$ be the $k$th approximation of the REGS method generated by Algorithm \ref{alg2} with initial $x_{0}\in R^{n }$ and $z_o\in \mathcal{R}(A^T)$. Then \begin{align} \mathbb{E}[\| z_{k}- x_{\star}\|_2^2]\leq(1-\frac{\sigma_r^2(A)}{\|A\|^2_F})^k \|z_{0}- x_{\star}\|_2^2+\frac{k}{\|A\|_F^2}(1-\frac{\sigma_r^2(A)}{\|A\|^2_F})^k \|Ax_0-Ax_{\star}\|_2^2.\label{th5} \end{align} \end{thm} For the REGS method, we first discuss the convergence behavior of $z_{k}- x_{\star}$ in Theorem \ref{theorem6} and Theorem \ref{theorem7}, and then consider its convergence behavior of $Az_{k}- Ax_{\star}$ in Theorem \ref{theorem8}. \begin{thm} \label{theorem6} Let $A\in R^{m\times n}$, $b\in R^{m}$, $x_{\star}=A^{\dag}b$ be the minimum Euclidean norm least squares solution, and $z_k$ be the $k$th approximation of the REGS method generated by Algorithm \ref{alg2} with initial $x_{0}\in R^{n }$ and $z_o\in \mathcal{R}(A^T)$. Then \begin{align} \mathbb{E}[\langle z_{k}- x_{\star}, v_{\ell} \rangle]= (1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} )^k \langle z_{0}-x_{\star}, v_{\ell}\rangle +\frac{k}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle A^T(Ax_{0}- Ax_{\star}), v_{\ell} \rangle .\label{th6} \end{align} \end{thm} \begin{pf} From Algorithm \ref{alg2}, we have \begin{align} &\mathbb{E}[\langle z_{k}- x_{\star}, v_{\ell} \rangle] \notag \\ &= \mathbb{E}[\langle z_{k-1}-\frac{A^{(i)} ( z_{k-1}-x_{k})}{ \| A^{(i)} \|_{2}^{2}}(A^{(i)})^T- x_{\star}, v_{\ell} \rangle] \notag \\ &= \mathbb{E}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle] +\mathbb{E}[\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle], \label{th65} \end{align} so we next consider $\mathbb{E}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle]$ and $\mathbb{E}[\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle]$ separately. We first consider$\mathbb{E}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle]$. Let $\mathbb{E}_{k-1}[\cdot]$ be the conditional expectation conditioned on the first $k-1$ iterations of the REGS method. That is, \begin{align} \mathbb{E}_{k-1}[\cdot]= \mathbb{E}[\cdot|j_1, i_1, j_2, i_2, \ldots, j_{k-1}, i_{k-1}], \notag \end{align} where $j_{t^*}$ is the ${t^*}$th column chosen and $i_{t^*}$ is the ${t^*}$th row chosen. We denote the conditional expectation conditioned on the first $k-1$ iterations and the $k$th column chosen as \begin{align} \mathbb{E}_{k-1}^{i}[\cdot]= \mathbb{E}[\cdot|j_1, i_1, j_2, i_2, \ldots, j_{k-1}, i_{k-1}, j_k]. \notag \end{align} Similarly, we denote the conditional expectation conditioned on the first $k-1$ iterations and the $k$th row chosen as \begin{align} \mathbb{E}_{k-1}^{j}[\cdot]= \mathbb{E}[\cdot|j_1, i_1, j_2, i_2, \ldots, j_{k-1}, i_{k-1}, i_k]. \notag \end{align} Then, by the law of total expectation, we have \begin{align} \mathbb{E}_{k-1}[\cdot]= \mathbb{E}_{k-1}^{j}[ \mathbb{E}_{k-1}^{i}[\cdot] ]. \notag \end{align} Thus, we obtain \begin{align} &\mathbb{E}_{k-1}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle] \notag \\ &=\mathbb{E}_{k-1}[\langle z_{k-1}-x_{\star}, v_{\ell} \rangle -\langle \frac{A^{(i)} ( z_{k-1}-x_{\star})}{\| A^{(i)} \|_{2}^{2}}(A^{(i)})^T, v_{\ell} \rangle] \notag \\ &=\langle z_{k-1}-x_{\star}, v_{\ell} \rangle -\frac{1}{\|A\|^2_F} \sum\limits_{i=1}^{m}\langle A^{(i)} ( z_{k-1}-x_{\star})(A^{(i)})^T, v_{\ell} \rangle \notag \\ &=\langle z_{k-1}-x_{\star}, v_{\ell} \rangle -\frac{1}{\|A\|^2_F} \sum\limits_{i=1}^{m}\langle(A^{(i)})^T, z_{k-1}-x_{\star} \rangle\langle (A^{(i)})^T, v_{\ell} \rangle \notag \\ &=\langle z_{k-1}-x_{\star}, v_{\ell} \rangle -\frac{1}{\|A\|^2_F} \langle A ( z_{k-1}-x_{\star}),A v_{\ell} \rangle. \notag \end{align} Further, by making use of $z_{k-1}-x_{\star}=\sum\limits_{i=1}^{n}\langle z_{k-1}-x_{\star}, v_i\rangle v_i$ and $Av_i=\sigma_i(A) u_i $, we get \begin{align} &\mathbb{E}_{k-1}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle] \notag \\ &=\langle z_{k-1}-x_{\star}, v_{\ell} \rangle -\frac{1}{\|A\|^2_F} \langle A \sum\limits_{i=1}^{n}\langle z_{k-1}-x_{\star}, v_i\rangle v_i,\sigma_{\ell}(A) u_{\ell} \rangle \notag \\ &=\langle z_{k-1}-x_{\star}, v_{\ell} \rangle -\frac{1}{\|A\|^2_F} \langle \sum\limits_{i=1}^{n}\langle z_{k-1}-x_{\star}, v_i\rangle \sigma_i(A) u_i,\sigma_{\ell}(A) u_{\ell} \rangle, \notag \end{align} which together with the orthogonality of the left singular vectors $u_i$ yields \begin{align} &\mathbb{E}_{k-1}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle] \notag \\ &=\langle z_{k-1}-x_{\star}, v_{\ell} \rangle -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} \langle z_{k-1}-x_{\star}, v_{\ell}\rangle \notag \\ &=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} ) \langle z_{k-1}-x_{\star}, v_{\ell}\rangle . \notag \end{align} As a result, by taking the full expectation on both sides, we have \begin{align} \mathbb{E}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle]=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} ) \mathbb{E}[ \langle z_{k-1}-x_{\star}, v_{\ell}\rangle ]. \label{th62} \end{align} We now consider $\mathbb{E}[\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle]$. It follows from \begin{align} &\mathbb{E}_{k-1}[\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle] \notag \\ &= \mathbb{E}_{k-1}^{j}[ \mathbb{E}_{k-1}^{i} [ \langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle]]\notag \\ &= \mathbb{E}_{k-1}^{j}[\frac{1}{\|A\|^2_F} \sum\limits_{i=1}^{m} \langle (A^{(i)})^TA^{(i)}( x_{k}-x_{\star}), v_{\ell} \rangle] \notag \\ &= \mathbb{E}_{k-1}^{j}[\frac{1}{\|A\|^2_F} \sum\limits_{i=1}^{m} \langle A^{(i)}( x_{k}-x_{\star}), A^{(i)} v_{\ell} \rangle] \notag \\ &= \mathbb{E}_{k-1} [\frac{1}{\|A\|^2_F} \langle A ( x_{k}-x_{\star}), A v_{\ell} \rangle], \notag \end{align} that \begin{align} \mathbb{E} [\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle]=\mathbb{E} [\frac{1}{\|A\|^2_F} \langle A ( x_{k}-x_{\star}), A v_{\ell} \rangle].\label{th63} \end{align} Since \begin{align} \mathbb{E} [\frac{1}{\|A\|^2_F} \langle A ( x_{k}-x_{\star}), A v_{\ell} \rangle]=\frac{\sigma_{\ell} (A)}{\|A\|^2_F} \mathbb{E} [ \langle A ( x_{k}-x_{\star}), u_{\ell} \rangle], \notag \end{align} by exploiting (\ref{th1}) in Theorem \ref{theorem1}, we get \begin{align} &\mathbb{E} [\frac{1}{\|A\|^2_F} \langle A ( x_{k}-x_{\star}), A v_{\ell} \rangle] \notag \\ &=\frac{\sigma_{\ell} (A)}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle Ax_{0}- Ax_{\star}, u_{\ell} \rangle \notag \\ &=\frac{1}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle Ax_{0}- Ax_{\star}, A v_{\ell} \rangle .\notag \end{align} Thus, substituting the above equality into (\ref{th63}), we have \begin{align} \mathbb{E} [\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle] &=\frac{1}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle Ax_{0}- Ax_{\star}, A v_{\ell} \rangle \notag \\ &=\frac{1}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle A^T(Ax_{0}- Ax_{\star}), v_{\ell} \rangle . \label{th64} \end{align} Combining (\ref{th65}), (\ref{th62}) and (\ref{th64}) yields \begin{align} &\mathbb{E}[\langle z_{k}- x_{\star}, v_{\ell} \rangle] \notag \\ &= \mathbb{E}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle] +\mathbb{E}[\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle] \notag \\ &=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} ) \mathbb{E}[ \langle z_{k-1}-x_{\star}, v_{\ell}\rangle ]+\frac{1}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle A^T(Ax_{0}- Ax_{\star}), v_{\ell} \rangle \notag \\ &=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} )^2 \mathbb{E}[ \langle z_{k-2}-x_{\star}, v_{\ell}\rangle ]+\frac{2}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle A^T(Ax_{0}- Ax_{\star}), v_{\ell} \rangle \notag \\ &=\ldots=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} )^k \langle z_{0}-x_{\star}, v_{\ell}\rangle +\frac{k}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle A^T(Ax_{0}- Ax_{\star}), v_{\ell} \rangle, \notag \end{align} which is the desired result (\ref{th6}). \end{pf} \begin{rmk} \label{rmk61} Theorem \ref{theorem6} shows that the decay rates of $\|z_k-x_{\star}\|_2$ are different in different right singular vectors spaces and the smallest singular value will lead to the slowest rate of convergence, which is the one in (\ref{th5}). So, the convergence bound presented by Du \cite{ Dukui2019} is optimal. \end{rmk} \begin{thm} \label{theorem7} Let $A\in R^{m\times n}$, $b\in R^{m}$, $x_{\star}=A^{\dag}b$ be the minimum Euclidean norm least squares solution, and $z_k$ be the $k$th approximation of the REGS method generated by Algorithm \ref{alg2} with initial $x_{0}\in R^{n }$ and $z_o\in \mathcal{R}(A^T)$. Then \begin{align} \mathbb{E}[ \|z_{k}- x_{\star}\|^2_2]\leq \mathbb{E}[(1-\frac{1}{\|A\|^2_F}\|A\frac{z_{k-1}-x_{\star}}{\|z_{k-1}-x_{\star}\|_2}\|_2^2)\|z_{k-1}-x_{\star}\|_2^2]+ \frac{1}{\|A\|^2_F}(1-\frac{\sigma_r^2(A)}{\|A\|^2_F})^k \| Ax_{0}-Ax_{\star} \|^2_2. \label{th7} \end{align} \end{thm} \begin{pf} Following an analogous argument to Theorem 4 of \cite{Dukui2019}, we get \begin{align} \mathbb{E} [\|z_{k}- x_{\star}\|_2^2 ]&= \mathbb{E}[ \|(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star})\|_2^2]+\mathbb{E} [\| \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star})\|_2^2], \notag \end{align} \begin{align} \mathbb{E}[ \|(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star})\|_2^2] &=\mathbb{E} [(z_{k-1}-x_{\star})^T(I-\frac{A^TA}{\|A\|_F^2})(z_{k-1}-x_{\star})]\notag \\ &=\mathbb{E}[(\|z_{k-1}-x_{\star}\|_2^2- \frac{1}{\|A\|^2_F}\|A(z_{k-1}-x_{\star})\|_2^2)]\notag \\ &=\mathbb{E}[(1-\frac{1}{\|A\|^2_F}\|A\frac{z_{k-1}-x_{\star}}{\|z_{k-1}-x_{\star}\|_2}\|_2^2)\|z_{k-1}-x_{\star}\|_2^2],\notag \end{align} and \begin{align} \mathbb{E} [\| \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star})\|_2^2]\leq\frac{1}{\|A\|^2_F}(1-\frac{\sigma_r^2(A)}{\|A\|^2_F})^k \| Ax_{0}-Ax_{\star} \|^2_2. \notag \end{align} Combining the above three equations, we have \begin{align} \mathbb{E} [ \|z_{k}- x_{\star}\|_2^2] \leq\mathbb{E}[(1-\frac{1}{\|A\|^2_F}\|A\frac{z_{k-1}-x_{\star}}{\|z_{k-1}-x_{\star}\|_2}\|_2^2)\|z_{k-1}-x_{\star}\|_2^2]+ \frac{1}{\|A\|^2_F}(1-\frac{\sigma_r^2(A)}{\|A\|^2_F})^k \| Ax_{0}-Ax_{\star} \|^2_2, \notag \end{align} which implies the desired result (\ref{th7}). \end{pf} \begin{rmk} \label{rmk71} Since $\|A\frac{z_{k-1}-x_{\star}}{\|z_{k-1}-x_{\star}\|_2}\|_2^2\geq\sigma_r^2(A)$, Theorem \ref{theorem7} implies that $z_{k}$ of the REGS method actually converges faster if $z_{k-1}-x_{\star}$ is not close to right singular vectors corresponding to the small singular values of $A$ . \end{rmk} \begin{thm} \label{theorem8} Let $A\in R^{m\times n}$, $b\in R^{m}$, $x_{\star}=A^{\dag}b$ be the minimum Euclidean norm least squares solution, and $z_k$ be the $k$th approximation of the REGS method generated by Algorithm \ref{alg2} with initial $x_{0}\in R^{n }$ and $z_o\in \mathcal{R}(A^T)$. Then \begin{align} \mathbb{E}[\langle A z_{k}- Ax_{\star}, u_{\ell} \rangle]=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} )^k \langle Az_{0}-Ax_{\star}, v_{\ell}\rangle +\frac{k}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle AA^T(Ax_{0}- Ax_{\star}), u_{\ell} \rangle .\label{th8} \end{align} \end{thm} \begin{pf} Similar to the proof of (\ref{th65}) in Theorem \ref{theorem6}, we obtain \begin{align} \mathbb{E}[\langle Az_{k}- Ax_{\star}, u_{\ell} \rangle] = \mathbb{E}[\langle A(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), u_{\ell} \rangle] +\mathbb{E}[\langle A \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), u_{\ell} \rangle]. \label{th80} \end{align} Then, we consider $\mathbb{E}[\langle A(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), u_{\ell} \rangle]$ and $\mathbb{E}[\langle A \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), u_{\ell} \rangle]$ separately. We first consider $\mathbb{E}[\langle A(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), u_{\ell} \rangle]$. It follows from $$ \langle A(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), u_{\ell} \rangle = \langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), A^Tu_{\ell} \rangle $$ and $A^Tu_{\ell}=\sigma_{\ell}(A)v_{\ell}$, that \begin{align} \mathbb{E}[\langle A(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), u_{\ell} \rangle] = \sigma_{\ell}(A)\mathbb{E}[\langle (I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), v_{\ell} \rangle], \notag \end{align} which together with (\ref{th62}), yields \begin{align} \mathbb{E}[\langle A(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), u_{\ell} \rangle] &= \sigma_{\ell}(A)(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} ) \mathbb{E}[ \langle z_{k-1}-x_{\star}, v_{\ell}\rangle ] \notag \\ &=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} ) \mathbb{E}[ \langle z_{k-1}-x_{\star},\sigma_{\ell}(A) v_{\ell}\rangle ]\notag \\ &=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} ) \mathbb{E}[ \langle Az_{k-1}-Ax_{\star},u_{\ell}\rangle ]. \label{th81} \end{align} We now consider $\mathbb{E}[\langle A \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), u_{\ell} \rangle]$. Exploiting (\ref{th64}), we have \begin{align} \mathbb{E}[\langle A \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), u_{\ell} \rangle] &=\mathbb{E}[\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), A^T u_{\ell} \rangle]\notag \\ &=\sigma_\ell (A)\mathbb{E}[\langle \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), v_{\ell} \rangle]\notag \\ &=\frac{\sigma_\ell (A)}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle A^T(Ax_{0}- Ax_{\star}), v_{\ell} \rangle \notag \\ &=\frac{1}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle AA^T(Ax_{0}- Ax_{\star}), u_{\ell} \rangle. \label{th82} \end{align} Thus, combining (\ref{th80}), (\ref{th81}) and (\ref{th82}) yields \begin{align} &\mathbb{E}[\langle Az_{k}- Ax_{\star}, u_{\ell} \rangle] \notag \\ &= \mathbb{E}[\langle A(I-\frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}})( z_{k-1}-x_{\star}), u_{\ell} \rangle] +\mathbb{E}[\langle A \frac{(A^{(i)})^TA^{(i)}}{\| A^{(i)} \|_{2}^{2}}( x_{k}-x_{\star}), u_{\ell} \rangle] \notag \\ &=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} ) \mathbb{E}[ \langle Az_{k-1}-Ax_{\star},u_{\ell}\rangle ]+\frac{1}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle AA^T(Ax_{0}- Ax_{\star}), u_{\ell} \rangle \notag \\ &=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} )^2 \mathbb{E}[ \langle Az_{k-2}-Ax_{\star},u_{\ell}\rangle ]+\frac{2}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle AA^T(Ax_{0}- Ax_{\star}), u_{\ell} \rangle \notag \\ &=\ldots=(1 -\frac{\sigma_{\ell}^2(A)}{\|A\|^2_F} )^k \langle Az_{0}-Ax_{\star}, v_{\ell}\rangle +\frac{k}{\|A\|^2_F} (1-\frac{\sigma_\ell^2(A)}{\|A\|^2_F})^k \langle AA^T(Ax_{0}- Ax_{\star}), u_{\ell} \rangle, \notag \end{align} which is the desired result (\ref{th8}). \end{pf} \begin{rmk} \label{rmk81} Theorem \ref{theorem8} shows the decay rates of $\|Az_{k}- Ax_{\star}\|$ of the REGS method and suggests that small singular values lead to poor convergence rates and vice versa. We note similar issues arise for the RK, REK, and RGS methods discussed in \cite{steinerberger2021randomized}, \cite{zhang2021preconvergence}, and Theorem \ref{theorem1}, respectively. \end{rmk} \section{Numerical experiments }\label{sec5} Now we present two simple examples to illustrate that the convergence directions of the RGS and REGS methods. To this end, let $G_0\in R^{500\times 500}$ be a Gaussian matrix with i.i.d. $N(0, 1)$ entries and $D\in R^{500\times 500}$ be a diagonal matrix whose diagonal elements are all 100. Further, we set $G_1=G_0+D$ and replace its last row $G_1^{(500)}$ by a tiny perturbation of $G_1^{(499)}$, i.e., adding 0.01 to each entry of $G_1^{(499)}$. Then, we normalize all rows of $G_1$, i.e., set $\|G_1^{(i)}\|_2=1$, $i=1, 2, \ldots, 500$. After that, we set $A_1=\begin{bmatrix} G_1\\ G_2 \end{bmatrix} \in R^{600\times 500}$ and $A_2=\begin{bmatrix} G_1, G_3 \end{bmatrix} \in R^{500\times 600}$, where $G_2\in R^{100\times 500}$ and $G_3\in R^{500\times 100}$ are zero matrices. So, the first 499 singular values of the matrices $A_1$ and $A_2$ are between $\sim 0.6$ and $\sim 1.5$, and the smallest nonzero singular value is $\sim 10^{-4}$. We first consider convergence directions of $Ax_k-Ax_{\star}$ and $x_k-x_{\star}$ of the RGS method. We generate a vector $x \in R^{500}$ using the MATLAB function \texttt{randn}, set the full column rank coefficient matrix $A=A_1$ and set the right-hand side $b=A x +z$, where $z$ is a nonzero vector belonging to the null space of $A^{T}$, which is generated by the MATLAB function \texttt{null}. With $x_0=0$, we plot $|\langle (Ax_k-Ax_{\star})/\|Ax_k-Ax_{\star}\|_2, u_{500} \rangle|$ and $\frac{\|A(x_k-x_{\star})\|_2 }{\|x_k-x_{\star}\|_2}$ in Figure \ref{fig1} and Figure \ref{fig2}, respectively. \begin{figure}[ht] \begin{center} \includegraphics [height=5.5cm,width=8.5cm ]{RGS-Ax-u-600-500.eps} \end{center} \caption{A sample evolution of $ |\langle (Ax_k-Ax_{\star})/\|Ax_k-Ax_{\star}\|_2, u_{500} \rangle|$ of the RGS method. }\label{fig1} \end{figure} From Figure \ref{fig1}, we find that $|\langle (Ax_k-Ax_{\star})/\|Ax_k-Ax_{\star}\|_2, u_{500} \rangle|$ initially is very small and almost is 0, which indicates that $Ax_k-Ax_{\star} $ is not close to the left singular vector $u_{500}$. Considering the analysis of Remark \ref{rmk5}, the phenomenon implies the `preconvergence' behavior of the RGS method, that is, the RGS method seems to converge quickly at the beginning. In addition, as $k\rightarrow\infty$, $|\langle (Ax_k-Ax_{\star})/\|Ax_k-Ax_{\star}\|_2, u_{500} \rangle|\rightarrow 1$. This phenomenon implies that $Ax_{k}-Ax_{\star}$ tends to the left singular vector corresponding to the smallest singular value of $A$. \begin{figure}[ht] \begin{center} \includegraphics [height=5.5cm,width=8.5cm ]{RGS-x-600-500.eps} \end{center} \caption{A sample evolution of $\frac{\|A(x_k-x_{\star})\|_2 }{\|x_k-x_{\star}\|_2}$ of the RGS method. }\label{fig2} \end{figure} From Figure \ref{fig2}, we observe that the values of $\frac{\|A(x_k-x_{\star})\|_2 }{\|x_k-x_{\star}\|_2}$ decreases with $k$ and finally approaches the small singular value. This phenomenon implies that the forward direction of $x_k-x_{\star}$ is mainly determined by the right singular vectors corresponding to the large singular values of $A$ at the beginning. With the increase of $k$, the direction is mainly determined by the right singular vectors corresponding to the small singular values. Finally, $x_k-x_{\star}$ tends to the right singular vector space corresponding to the smallest singular value. Furthermore, this phenomenon also allows for an interesting application, i.e., finding nonzero vectors $x$ such that $\frac{\|Ax\|_2}{\|x\|_2}$ is small. We now consider convergence directions of $Az_k-Ax_{\star}$ and $z_k-x_{\star}$ of the REGS method. We generate a vector $x \in R^{600}$ using the MATLAB function \texttt{randn}, set the coefficient matrix $A=A_2$ which does not have full column rank, and set the right-hand side $b=Ax $. With $x_0=0$ and $z_0=0$, we plot $|\langle (Az_k-Ax_{\star})/\|Az_k-Ax_{\star}\|_2, u_{500} \rangle|$ and $\frac{\|A(z_k-x_{\star})\|_2 }{\|z_k-x_{\star}\|_2}$ in Figure \ref{fig3} and Figure \ref{fig4}, respectively. \begin{figure}[ht] \begin{center} \includegraphics [height=5.5cm,width=8.5cm ]{REGS-Az-u-500-600.eps} \end{center} \caption{A sample evolution of $ |\langle (Az_k-Ax_{\star})/\|Az_k-Ax_{\star}\|_2, u_{500} \rangle|$ of the REGS method. }\label{fig3} \end{figure} \begin{figure}[ht] \begin{center} \includegraphics [height=5.5cm,width=8.5cm ]{REGS-z-500-600.eps} \end{center} \caption{A sample evolution of $\frac{\|A(z_k-x_{\star})\|_2 }{\|z_k-x_{\star}\|_2}$ of the REGS method. }\label{fig4} \end{figure} Figure \ref{fig3} and Figure \ref{fig4} show the similar results obtained in the RGS method. That is, the convergence directions of $A z_k-Ax_{\star} $ and $ z_k- x_{\star} $ of the REGS method initially are depending on the large singular values and then mainly depending on the small singular values, and finally depending on the smallest singular value of $A$.

https://arxiv.org/abs/1912.01763

A note on semi-infinite program bounding methods

Semi-infinite programs are a class of mathematical optimization problems with a finite number of decision variables and infinite constraints. As shown by Blankenship and Falk (Blankenship and Falk. "Infinitely constrained optimization problems." Journal of Optimization Theory and Applications 19.2 (1976): 261-281.), a sequence of lower bounds which converges to the optimal objective value may be obtained with specially constructed finite approximations of the constraint set. In (Mitsos. "Global optimization of semi-infinite programs via restriction of the right-hand side." Optimization 60.10-11 (2011): 1291-1308.), it is claimed that a modification of this lower bounding method involving approximate solution of the lower-level program yields convergent lower bounds. We show with a counterexample that this claim is false, and discuss what kind of approximate solution of the lower-level program is sufficient for correct behavior.

\section{Introduction} This note discusses methods for the global solution of semi-infinite programs (SIP). Specifically, the method from \cite{mitsos11} is considered, and it is shown with a counterexample that the lower bounds do not always converge. Throughout we use notation as close as possible to that used in \cite{mitsos11}, embellishing it only as necessary with, for instance, iteration counters. Consider a SIP in the general form \begin{alignat}{2} \tag{SIP} \label{eq:sip} f^* = \inf_{x}\; & f(x) \\ \st \notag & x \in X, \\ \notag & g(x,y) \le 0,\; \forall y \in Y, \end{alignat} for subsets $X$, $Y$ of finite dimensional real vector spaces and $f : X \to \mbb{R}$, $g : X \times Y \to \mbb{R}$. We may view $Y$ as an index set, with potentially uncountably infinite cardinality. Important to validating the feasibility of a point $x$ is the lower-level program: \begin{equation} \label{eq:llp} \tag{LLP} \sup_y \set{ g(x,y) : y \in Y}. \end{equation} Global solution of \eqref{eq:sip} often involves the construction of convergent upper and lower bounds. The approach in \cite{mitsos11} to obtain a lower bound is a modification of the constraint-generation/discretization method of \cite{blankenshipEA76}. The claim is that the lower-level program may be solved approximately; the exact nature of the approximation is important to the convergence of the lower bounds and this is the subject of the present note. \section{Sketch of the lower bounding procedure and claim} The setting of the method is the following. The method is iterative and at iteration $k$, for a given finite subset $Y^{LBD,k} \subset Y$, a lower bound of $f^*$ is obtained from the finite program \begin{alignat}{2} \label{eq:sip_lower} f^{LBD,k} = \inf_{x}\; & f(x) \\ \st \notag & x \in X, \\ \notag & g(x,y) \le 0, \;\forall y \in Y^{LBD,k}. \end{alignat} This is indeed a lower bound since fewer constraints are enforced, and thus \eqref{eq:sip_lower} is a relaxation of \eqref{eq:sip}. Assume that the lower bounding problem \eqref{eq:sip_lower} is feasible (otherwise we can conclude that \eqref{eq:sip} is infeasible). Let $\bar{x}^k$ be a (global) minimizer of the lower bounding problem \eqref{eq:sip_lower}. In \cite{mitsos11}, Lemma~2.2 states that we either verify $\sup_y \set{g(\bar{x}^k,y) : y \in Y} \le 0$, \textbf{or else} find $\bar{y}^k \in Y$ such that $g(\bar{x}^k,\bar{y}^k) > 0$. If $\sup_y \set{g(\bar{x}^k,y) : y \in Y} \le 0$, then $\bar{x}^k$ is feasible in \eqref{eq:sip} and thus optimal (since it also solves a relaxation). Otherwise, set $Y^{LBD,k+1} = Y^{LBD,k} \cup \set{\bar{y}^k}$ and we iterate. The precise statement of the claim is repeated here (again, with only minor embellishments to the notation to help keep track of iterations). \begin{lemma}[Lemma~2.2 in \cite{mitsos11}] \label{lem:claim} Take any $Y^{LBD,0} \subset Y$. Assume that $X$ and $Y$ are compact and that $g$ is continuous on $X \times Y$. Suppose that at each iteration of the lower bounding procedure the lower-level program is solved approximately for the solution of the lower bounding problem $\bar{x}^k$ either establishing $\max_{y \in Y} g(\bar{x}^k,y) \le 0$, or furnishing a point $\bar{y}^k$ such that $g(\bar{x}^k,\bar{y}^k) > 0$. Then, the lower bounding procedure converges to the optimal objective value, i.e. $f^{LBD,k} \to f^*$. \end{lemma} \section{Correction} \subsection{Counterexample} We now present a counterexample to the claim in Lemma~\ref{lem:claim}. Consider \begin{alignat}{2} \tag{CEx} \label{eq:counter_ex} \inf_{x}\; & -x \\ \st \notag & x \in [-1,1], \\ \notag & 2x - y \le 0,\; \forall y \in [-1,1], \end{alignat} thus we define $X = Y = [-1,1]$, $f : x \mapsto -x$, $g : (x,y) \mapsto 2x - y$. The behavior to note is this: We are trying to maximize $x$; The feasible set is \[ \set{x \in [-1,1] : x \le (\sfrac{1}{2})y, \forall y \in[-1,1] } = [-1,-\sfrac{1}{2}]; \] The infimum, consequently, is $\sfrac{1}{2}$. See Figure~\ref{fig:cex1}. \begin{figure} \begin{center} \begin{tikzpicture}[xscale=1.5,yscale=1.5] \draw[fill=gray] (-0.5,-1) -- (0.5,1) -- (1,1) -- (1,-1) -- cycle; \draw[latex-latex] (0,-1.3) -- (0,1.3) node[above]{$y$}; \draw[latex-latex] (-1.3,0) -- (1.3,0) node[right]{$x$}; \draw (-1,-1) rectangle (1,1); \draw[dashed,domain=0:1] plot (\x, \x); \end{tikzpicture} \end{center} \caption{Visualization of counterexample~\eqref{eq:counter_ex}. The box represents $[-1,1] \times [-1,1]$. The shaded grey area is the subset of $(x,y)$ such that $2x - y > 0$. The dashed line represents the approximate minimizers used in the counterexample.} \label{fig:cex1} \end{figure} Beginning with $Y^{LBD,1} = \emptyset$, the minimizer of the lower bounding problem is $\bar{x}^1 = 1$. Now, assume that solving the resulting \eqref{eq:llp} approximately, we get $\bar{y}^1 = 1$ which we note satisfies \[ 2\bar{x}^1 - \bar{y}^1 = 1 > 0 \] as required by Lemma~\ref{lem:claim}. The next iteration, with $Y^{LBD,2} = \set{1}$, adds the constraint $2x - 1 \le 0$ to the lower bounding problem; the feasible set is $[-1,\sfrac{1}{2}]$ so the minimizer is $\bar{x}^2 = \sfrac{1}{2}$. Again, assume that solving the lower-level program approximately yields $\bar{y}^2 = \sfrac{1}{2}$; again we get \[ 2\bar{x}^2 - \bar{y}^2 = \sfrac{1}{2} > 0 \] as required by Lemma~\ref{lem:claim}. The third iteration, with $Y^{LBD,3} = \set{1, \sfrac{1}{2}}$, adds the constraint $2x - \sfrac{1}{2} \le 0$ to the lower bounding problem; the feasible set is $[-1,\sfrac{1}{4}]$ so the minimizer is $\bar{x}^3 = \sfrac{1}{4}$. Again, assume that solving the lower-level program approximately yields $\bar{y}^3 = \sfrac{1}{4}$; again we get \[ 2\bar{x}^3 - \bar{y}^3 = \sfrac{1}{4} > 0 \] as required by Lemma~\ref{lem:claim}. Proceeding in this way, we construct $\bar{x}^k$ and $\bar{y}^k$ so that $g(\bar{x}^k,\bar{y}^k) > 0$ and the lower bounds satisfy $f^{LBD,k} = -\bar{x}^k = -\frac{1}{2^{k-1}}$, for all $k$. Consequently, they converge to $0$, which we note is strictly less than the infimum of $\sfrac{1}{2}$. \subsection{Modified claim} We now present a modification of the claim in order to demonstrate what kind of approximate solution of the lower-level program suffices to establish convergence of the lower bounds. To state the result, let the optimal objective value of \eqref{eq:llp} as a function of $x$ be \[ g^*(x) = \sup_y\set{g(x,y) : y \in Y}. \] The proof of the following result has a similar structure to the original proof of \cite[Lemma~2.2]{mitsos11}. \begin{lemma} \label{lem:claim_mod} Choose any finite $Y^{LBD,0} \subset Y$, and $\alpha \in (0,1)$. Assume that $X$ and $Y$ are compact and that $f$ and $g$ are continuous. Suppose that at each iteration $k$ of the lower bounding procedure \eqref{eq:llp} is solved approximately for the solution $\bar{x}^k$ of the lower bounding problem~\eqref{eq:sip_lower}, either establishing that $g^*(\bar{x}^k) \le 0$ or furnishing a point $\bar{y}^k$ such that \[ g(\bar{x}^k,\bar{y}^k) \ge \alpha g^*(\bar{x}^k) > 0. \] Then, the lower bounding procedure converges to the optimal objective value, i.e. $f^{LBD,k} \to f^*$. \end{lemma} \begin{proof} First, if the lower bounding problem~\eqref{eq:sip_lower} is ever infeasible for some iteration $k$, then \eqref{eq:sip} is infeasible and we can set $f^{LBD,k} = +\infty = f^*$. Otherwise, since $X$ is compact, $Y^{LBD,k}$ is finite, and $f$ and $g$ are continuous, for every iteration the lower bounding problem has a solution by Weierstrass' (extreme value) theorem. If at some iteration $k$ the lower bounding problem furnishes a point $\bar{x}^k$ for which $g^*(\bar{x}^k) \le 0$, then $\bar{x}^k$ is feasible for \eqref{eq:sip}, and thus optimal. The corresponding lower bound $f^{LBD,k}$, and all subsequent lower bounds, equal $f^*$. Otherwise, we have an infinite sequence of solutions to the lower bounding problems. Since $X$ is compact we can move to a subsequence $\seq[k \in \mbb{N}]{\bar{x}^{k}} \subset X$ which converges to $x^* \in X$. By construction of the lower bounding problem we have \[ g(\bar{x}^{\ell},\bar{y}^k) \le 0, \quad \forall \ell,k : \ell > k. \] By continuity and compactness of $X \times Y$ we have uniform continuity of $g$, and so for any $\epsilon > 0$, there exists a $\delta > 0$ such that \begin{equation} \label{eq:gee} g(x,\bar{y}^k) < \epsilon, \quad\forall x : \norm{x - \bar{x}^{\ell}} < \delta, \quad\forall \ell,k : \ell > k. \end{equation} Since the (sub)sequence $\seq[k \in \mbb{N}]{\bar{x}^k}$ converges, there is an index $K$ sufficiently large that \begin{equation} \label{eq:tails} \norm{\bar{x}^{\ell} - \bar{x}^k} < \delta, \quad \forall \ell,k : \ell > k \ge K. \end{equation} Using \eqref{eq:tails}, we can substitute $x = \bar{x}^k$ in \eqref{eq:gee} to get that for any $\epsilon > 0$, there exists $K$ such that \[ g(\bar{x}^k,\bar{y}^k) < \epsilon, \quad \forall k \ge K. \] By assumption $g(\bar{x}^k,\bar{y}^k) > 0$ for all $k$, and so combined with the above we have that $g(\bar{x}^k,\bar{y}^k) \to 0$. Combining $g(\bar{x}^k,\bar{y}^k) \to 0$ with $ g(\bar{x}^k,\bar{y}^k) \ge \alpha g^*(\bar{x}^k) > 0, $ for all $k$, we see $ g^*(\bar{x}^k) \to 0. $ Meanwhile $g^* : X \to \mbb{R}$ is a continuous function, by classic parametric optimization results like \cite[Theorem~1.4.16]{aubin_frankowska} (using continuity of $g$ and compactness of $Y$). Thus \[ g^*(x^*) = \lim_{k \to \infty} g^*(\bar{x}^k) = 0. \] Thus $x^*$ is feasible in \eqref{eq:sip} and so $f^* \le f(x^*)$. But since the lower bounding problem is a relaxation, $f^{LBD,k} = f(\bar{x}^k) \le f^*$ for all $k$, and so by continuity of $f$, $f(x^*) \le f^*$. Combining these inequalities we see $f^{LBD,k} \to f(x^*) = f^*$. Since the entire sequence of lower bounds is an increasing sequence, we see that the entire sequence converges to $f^*$ (without moving to a subsequence). \end{proof} \section{Remarks} The main contribution of \cite{mitsos11} is a novel \emph{upper} bounding procedure, which still stands, and combined with the modified lower bounding procedure from Lemma~\ref{lem:claim_mod} or the original procedure from \cite{blankenshipEA76}, the overall global solution method for \eqref{eq:sip} is still effective. The counterexample that has been presented may seem contrived. However, as the lower bounding method for SIP from \cite{mitsos11} is adapted to give a lower bounding method for \emph{generalized} semi-infinite programs (GSIP) in \cite{mitsosEA15}, a modification of the counterexample reveals that similar behavior may occur (and in a more natural way) when constructing the lower bounds for a GSIP. Consequently, the lower bounds fail to converge to the infimum. See \cite{Harwood19_GSIP}.

https://arxiv.org/abs/math/0610707

A fixed point theorem for the infinite-dimensional simplex

We define the infinite dimensional simplex to be the closure of the convex hull of the standard basis vectors in R^infinity, and prove that this space has the 'fixed point property': any continuous function from the space into itself has a fixed point. Our proof is constructive, in the sense that it can be used to find an approximate fixed point; the proof relies on elementary analysis and Sperner's lemma. The fixed point theorem is shown to imply Schauder's fixed point theorem on infinite-dimensional compact convex subsets of normed spaces.

\section{Introduction} In finite dimensions, one of the simplest methods for proving the Brouwer fixed point theorem is via a combinatorial result known as Sperner's lemma \cite{Sper28}, which is a statement about labelled triangulations of a simplex in $\ensuremath{\mathbb{R}} ^n$. In this paper, we use Sperner's lemma to prove a fixed point theorem on an infinite-dimensional simplex in $\ensuremath{\mathbb{R}} ^\infty$. We also show that this theorem implies the infinite-dimensional case of Schauder's fixed point theorem on normed spaces. Since $\ensuremath{\mathbb{R}} ^\infty$ is locally convex, our theorem is a consequence of Tychonoff's fixed point theorem \cite{Smar74}. However, some notable advantages of our approach are: (1) the constructive nature of Sperner's lemma provides a method for producing approximate fixed points for functions on the infinite-dimensional simplex, (2) the proof is based on elementary methods in topology and analysis, and (3) our proof provides another route to Schauder's theorem. Fixed point theorems and their constructive proofs have found many important applications, ranging from proofs of the Inverse Function Theorem \cite{Lang97}, to proofs of the existence of equilibria in economics \cite{Todd76, Yang99}, to the existence of solutions of differential equations \cite{Brow93, Smar74}. \section{Working in $\ensuremath{\mathbb{R}} ^\infty$} Let $\ensuremath{\mathbb{R}} ^\infty$ and $I^\infty = \prod [0,1]$ be the product of countably many copies of $\ensuremath{\mathbb{R}} $, and $I=[0,1]$, respectively. We equip $\ensuremath{\mathbb{R}} ^\infty$ with the standard product topology, which is metrizable \cite{BePe75} by the complete metric \[ \bar d(x,y) = \sum_{i=1}^\infty \frac{|x_i - y_i|}{2^i(1+|x_i-y_i|)}. \] In $\ensuremath{\mathbb{R}} ^n$, a $k$-dimensional simplex, or \textit{$k$-simplex}, $\sigma^k$ is the convex hull of $k+1$ affinely independent points. The \textit{standard $n$-simplex} in $\ensuremath{\mathbb{R}} ^{n+1}$, denoted $\Delta^n$, is the convex hull of the $n+1$ standard basis vectors of $\ensuremath{\mathbb{R}} ^n$. The natural extension of this definition to $\ensuremath{\mathbb{R}} ^\infty$ is to consider $\Delta^\infty$, the convex hull of the standard basis vectors $\{e_i\}$ in $\ensuremath{\mathbb{R}} ^\infty$, where $(e_i)_j= \delta_{ij}$, the Kronecker delta function. As convex combinations are finite sums, this convex hull is: \[ \ensuremath{\Delta^\infty} = \{x\in \ensuremath{\mathbb{R}} ^\infty | \sum_{i=1}^\infty x_i = 1,\ 0\leq x_i \leq 1 \mbox{, and only finitely many $x_i$ are non-zero}\}. \] Unfortunately, $\ensuremath{\Delta^\infty} $ is not closed; under the metric $\bar d$ the sequence $\{e_i\}$ converges to $\mathbf{0}$, which is not in $\ensuremath{\Delta^\infty} $. So consider, instead $\ensuremath{\Delta^\infty_0} $, the closure of $\ensuremath{\Delta^\infty} $, which can be shown to be: \[ \ensuremath{\Delta^\infty_0} = \{x\in \ensuremath{\mathbb{R}} ^\infty | \sum_{i=1}^\infty x_i \leq 1 \mbox{ and } 0\leq x_i\leq 1\}. \] It is easy to see that $\ensuremath{\Delta^\infty_0} $ is convex. It is also the closure of the convex hull of the standard basis vectors $\{e_i\}$ and $\mathbf{0}$. It is also compact because it is a closed subset of $I^\infty$, which is compact by Tychonoff's Theorem. \footnote{If one would like to avoid the Axiom of Choice, which is equivalent to Tychonoff's Theorem, it is not difficult to show that $I^\infty$ is a closed and totally bounded subset of the complete space $\ensuremath{\mathbb{R}} ^\infty$, which implies compactness.} We call $\ensuremath{\Delta^\infty_0} $ the {\em standard infinite-dimensional simplex}. It will be important for our purposes later to consider $F^n$, the $n$-dimensional face of $\ensuremath{\Delta^\infty_0} $ given by $F^n = conv\{e_1,e_2,\dots, e_{n+1}\}$. Notice that each $F^n$ is closed and thus compact. \section{Some Preliminary Machinery} \begin{comment Before stating Sperner's Lemma, which will be indispensable in our proof, some definitions are necessary. Let $\sigma^k = conv(x_0,...,x_k)$ be a $k$-simplex in $\ensuremath{\mathbb{R}} ^n$. Further, let $A$ be the simplexes of a triangulation of $\sigma^k$ and $B$ be the set of vertices of $A$. An A-Sperner map for this triangulation of $\sigma^k$ is a map $h:B\rightarrow \{0,\dots,k\}$ such that, if \[ J \subseteq \{0,...,k\} \mbox{ and } v \in conv\{x_j | j \in J\} \mbox{ then } h(v) \in J.\] A simplex in the triangulation of $\sigma^k$ is called full if the image of its vertices under $h$ maps onto $\{0,\dots,k\}$. (Definitions from van Mill p. 103) \begin{sperner} (van Mill p.103) Let $\sigma^k$ be a $k$-simplex in $\ensuremath{\mathbb{R}} ^n$ and let $A$ be the set of simplexes in a triangulation of $\sigma^k$. If $h$ is an A-Sperner map for $\sigma^k$ then the number of full simplexes in $A$ is odd and hence non-zero. \end{sperner} \end{comment Let $\sigma^k = conv(x_0,...,x_k)$ be a $k$-simplex in $\ensuremath{\mathbb{R}} ^n$. Let $T$ be a triangulation of $\sigma^k$ and $V$ be the set of vertices of $T$ (i.e., the vertices of simplices in $T$). A \emph{Sperner labelling} of the triangulation $T$ is a labelling function $\ell:V\rightarrow \{0,\dots,k\}$ such that \[ \mbox{ if } J \subseteq \{0,\dots,k\} \mbox{ and } v \in conv\{x_j | j \in J \} \mbox{, then } h(v) \in J. \] A $k$-simplex $\tau$ of $T$ is called a {\em fully-labelled simplex} (or {\em full}) if the image of the vertices of $\tau$ under $\ell$ maps onto $\{0,\dots,k\}$. Note that $\tau$ has exactly $k+1$ vertices, so all the vertices have distinct labels. \begin{sperner} Let $\sigma^k$ be a $k$-simplex in $\ensuremath{\mathbb{R}} ^n$ with triangulation $T$ and let $\ell$ be a Sperner-labelling of $T$. Then the number of full simplices of $T$ is odd (and hence, non-zero). \end{sperner} Though we will not prove this theorem here, an exposition of such proofs can be found in \cite{Su99}. In particular, there are constructive ``path-following'' proofs that locate the full simplex by tracing a path of simplices through the triangulation. Such path-following proofs have formed the basis of algorithms for locating fixed points of functions in finite-dimensional spaces, e.g., see \cite{Todd76} for a nice survey. In Section \ref{sec:fixed-dio}, we show how to use Sperner's lemma for a fixed point theorem in the infinite-dimensional space $\ensuremath{\Delta^\infty_0} $. Another crucial theorem for our purposes states that, under appropriate hypotheses, the existence of approximate fixed points implies the existence of fixed points. On the metric space $(X,d)$, we can quantify the notion of an approximate fixed point by defining an \textit{$\epsilon$-fixed point}, which for a given function $f$ is a point $x\in X$ such that $d(x,f(x))<\epsilon$. Versions of the following lemma may be found in, e.g., \cite{DuGr82, Smar74}. \begin{lemma}[Epsilon Fixed Point Theorem] \label{le epsilonfixed} Suppose that $A$ is a compact subset of the metric space $(X,d)$ and that $f:A\rightarrow A$ is continuous. If $f$ has an $\epsilon$-fixed point for every $\epsilon > 0$ then $f$ has a fixed point. \end{lemma} \begin{proof} Let $\{a_n\}$ be a sequence of $1/n$-fixed points. That is, $d(a_n,f(a_n)) < 1/n$ for all $n$. Since $A$ is compact it is sequentially compact and thus $\{a_n\}$ has a convergent subsequence, which we denote $\{a'_n\}$ with $a_n'\rightarrow x \in A$. Let $\epsilon >0$. Since $a_n'\rightarrow x$ there exists $N_1$ such that $n\geq N_1$ implies that $d(a'_n,x) < \epsilon /2$. Let $N = \max ( N_1, 2/\epsilon)$. Then $n\geq N$ implies that \[ d(x, f(a'_n)) \leq d(x,a'_n) + d(a'_n,f(a'_n)) < \epsilon, \] so that $f(a'_n) \rightarrow x$. However, since $f$ is continuous, we also know that $f(a'_n)\rightarrow f(x)$. Since limits are unique, we conclude that $f(x)=x$, which completes the proof. \end{proof} Later it will be desirable to have an isometry between $\Delta^{n-1}$, the standard $(n-1)$-simplex in $\ensuremath{\mathbb{R}} ^{n}$, and $F^{n-1}$. The easiest way to do this is to consider $\ensuremath{\mathbb{R}} ^n$ as a subspace of $\ensuremath{\mathbb{R}} ^\infty$ by projection onto the first $n$ factors, and restricting the metric on $\ensuremath{\mathbb{R}} ^\infty$ to $\ensuremath{\mathbb{R}} ^{n}$. Call this metric $\bar d_{n}$ and consider $\Delta^{n-1}$ in the metric space $(\ensuremath{\mathbb{R}} ^{n}, \bar d_{n})$. It is worthwhile to ensure that $(\ensuremath{\mathbb{R}} ^{n}, \bar d_n)$ has a rich supply of continuous functions. Before proceeding, recall that all norms on $\ensuremath{\mathbb{R}} ^n$ are equivalent and thus essentially interchangeable; we now prove that $\bar d_n$ is interchangeable with norm-induced metrics on bounded sets. \begin{lemma} \label{le metricequivalence} Let $A$ be a bounded subset of the normed space $(\ensuremath{\mathbb{R}} ^{n} , \| \cdot \|_\infty)$. On $A$, the metric $\bar d_n$ is equivalent to the metric induced by the norm $\| \cdot \|_\infty$. \end{lemma} \begin{proof} Suppose that $x,y \in \ensuremath{\mathbb{R}} ^n$. We see that \[ \bar d_n(x,y) = \sum_{i=1}^n \frac{|x_i - y_i|}{2^i(1+|x_i-y_i|)} \leq n\|x-y\|_\infty. \] Now, since $A$ is bounded, there is some $M$ such $\|x-y\|_\infty \leq M$ for $x,y \in A$. Thus we see that \[\frac{\|x-y\|_\infty}{2^n(1+M)} \leq \frac{\|x-y\|_\infty}{2^n(1+\|x-y\|_\infty)} \leq \bar d_n(x,y), \] which implies that \begin{equation} \label{eq:norm-bound} \|x-y\|_\infty \leq 2^n(1+M)\bar d_n(x,y). \end{equation} Thus $\bar d_n$ is equivalent to the metric induced by the norm on $A$.\end{proof} Lemma \ref{le metricequivalence} tells us that bounded subsets of $\ensuremath{\mathbb{R}} ^n$ have the same continuous functions regardless of whether they are considered as subsets of a normed space or as subsets of $(\ensuremath{\mathbb{R}} ^n , \bar d_n)$. Importantly, notice that $\Delta^{n-1}$ is bounded. Furthermore, the isometry $f:\Delta^{n-1} \rightarrow F^{n-1}$ between $\Delta^{n-1}$ in $(\ensuremath{\mathbb{R}} ^n,\bar d_n)$ and $F^{n-1}$ in $\ensuremath{\mathbb{R}} ^\infty$ is clearly given by $f(x) = f(x_1,x_2,\dots,x_n) = (x_1,x_2,\dots,x_n,0,0,\dots)$. This is important because it implies that $F^{n-1}$ has an arbitrarily small barycentric subdivision. Recall that the diameter of a set $X$ is $d(X) = \sup_{x,y\in X} d(x,y)$ and if $\mathscr T$ is a family of sets, then $size(\mathscr T) = \sup_{\sigma \in \mathscr T} d(\sigma)$. Thus, given $\epsilon >0$, $F^{n-1}$ has a barycentric subdivision $\mathscr T$ with $size(\mathscr T) < \epsilon$. \begin{comment}**************** \begin{proof} As proved in (include reference number) the standard $(n-1)$-simplex in the normed space $\ensuremath{\mathbb{R}} ^n$ as an arbitrarily small barycentric subdivision, and thus Lemma 2 tells us that the standard $(n-1)$-simplex in the metric space $(\ensuremath{\mathbb{R}} ^n,\bar d_n)$ has an arbitrarily fine subdivision. It is clear that this subdivision is preserved by the isometry between $F^{n-1}$ and $\Delta^{n-1}$. \end{proof} \end{comment}************************ Now we are ready to prove a fixed point theorem for $\ensuremath{\Delta^\infty_0} $. \section{A Fixed Point Theorem for $\ensuremath{\Delta^\infty_0} $} \label{sec:fixed-dio} \begin{theorem} \label{fixed-dio} Suppose that $f:\ensuremath{\Delta^\infty_0} \rightarrow \ensuremath{\Delta^\infty_0} $ is continuous. Then $f$ has a fixed point. \end{theorem} \begin{proof} Since $\ensuremath{\Delta^\infty_0} $ is compact, by Lemma \ref{le epsilonfixed}, it is sufficient to show that $f$ has an $\epsilon$-fixed point for each $\epsilon >0$. Let $\epsilon >0$ be given. Choose $N \geq \log_2(2/\epsilon)+1$. Notice that for $x,y \in \ensuremath{\Delta^\infty_0} $, this implies that \begin{equation} \label{eq:eps-over-2-N+1on} \sum_{i=N+1}^\infty \frac{|x_i - y_i|}{2^i(1+|x_i-y_i|)} \leq \sum_{i=N+1}^\infty \frac{1}{2^i} < \frac{\epsilon}{2}. \end{equation} Since $f$ maps between countably infinite-dimensional spaces, we can write $f$ in terms of its components: $f(x) = (f_1(x),f_2(x),\dots)$. Since $f$ is continuous, $f_i$ is continuous for each $i$. Consider the function \[ g(x) =(g_1(x),g_2(x),\dots) = (f_1(x),f_2(x),\dots, f_N(x) , 1-\sum_{i=1}^N f_i(x), 0,0,0,\dots ). \] Since each $f_i$ is continuous and finite sums of continuous function are continuous, $g_i$ is continuous for each $i$. Furthermore, we see that $g:F^N \rightarrow F^N$. Consequently, $g$ is continuous. Let $\epsilon_0 = \frac{\epsilon}{8(N+1)}$ and $\epsilon_1 = \frac{\epsilon}{2^{N+5}(N+1)}$. Since $g$ is continuous on a compact set, it is uniformly continuous. Thus there exists $\delta_1 >0$ such that $\bar d(x,y) < \delta_1$ implies that $\bar d(g(x),g(y)) < \epsilon_1$. Let $\delta = \min ( \delta_1 , \epsilon_1)$. Since $F^N$ can be triangulated with an arbitrarily small triangulation, let $\mathscr T$ be a triangulation with $size(\mathscr T) < \delta$. Label the vertices of $\mathscr T$ with the map \[ \ell(x) = \mbox{argmax}_i (x_i - g_i(x)). \] Recall that the \emph{argmax} function returns the index of the largest element of the argument, and if there are multiple indices that give the maximum value, the argmax function returns the least of these indices. Observe that $\ell(x)$ produces a Sperner labeling on the vertices of $\mathscr T$. Thus by Sperner's Lemma, there exists a fully-labeled simplex in $\mathscr T$. This simplex can be found using the path-following method described in \cite{Su99}. Let $\{x^1,x^2, \dots x^{N+1}\}$ be the vertices of this simplex where the index of each vertex is its Sperner label. From this, we see that for all $j$, \[ x^i_i - g_i(x^i) \geq x^i_j - g_j(x^i). \] Furthermore, since for each $x$ in $F^N$, we have \[ \sum_{j=1}^{N+1} x_j = \sum_{j=1}^{N+1} g_j(x) =1, \] there is at least one $j$ such that $g_j(x) \leq x_j$. In particular, since $\ell(x^i)=i$, this implies that for each $x^i$, \[ x^i_i - g_i(x^i) = \max_j ( x^i_j - g_j(x^i)) \geq 0. \] Since $size(\mathscr T) < \delta$ we have that, for all $i$, $\bar d(x^1, x^i) < \delta$. From the bound (\ref{eq:norm-bound}) in Lemma \ref{le metricequivalence} (note in this case $M=1$ and $n=N+1$), we find that for all $i, j$, \begin{equation} \label{eq:delta-bound} |x^1_j - x^i_j| < 2^{N+2}\delta \leq 2^{N+2} \epsilon_1 \leq \epsilon_0. \end{equation} By the same logic, we have that for all $i, j$, \begin{equation} \label{eq:delta-eps-bound} |g_j(x^1) - g_j(x^i)| < 2^{N+2}\epsilon_1 \leq \epsilon_0. \end{equation} Consequently, we have that \[ x^1_j + \epsilon_0 > x^i_j \quad \mbox{ and } \quad -g_j(x^i) < \epsilon_0 - g_j(x^1) \] which, in turn, implies that \[ 2\epsilon_0 + x^1_j - g_j(x^1) > x^i_j - g_j(x^i) \] for all $i$ and $j$. In particular, this implies that the following list of inequalities hold (simply let $i=j$ and run through all $i$): \[ \begin{array}{cccc} 2\epsilon_0 + x^1_1 - g_1(x^1) & > &x^1_1 - g_1(x^1)& \geq 0 , \\ 2\epsilon_0 + x^1_2 - g_2(x^1) & > &x^2_2 - g_2(x^2)& \geq 0 , \\ \vdots & & \vdots \\ 2\epsilon_0 + x^1_{N+1} - g_{N+1}(x^1) & > & x^{N+1}_{N+1} - g_{N+1}(x^{N+1})& \geq 0 . \end{array} \] Summing down each column yields the following inequality. \[2\epsilon_0(N+1) + \sum_{i=1}^{N+1} x^1_i - \sum_{i=1}^{N+1} g_i(x^1) > \sum_{i=1}^{N+1}\left( x_i^i-g_i(x^i)\right) \geq 0. \] Now we recall that for all $i$, $x_i^i-g_i(x^i) \geq 0$ and \[ \sum_{i=1}^{N+1} x^1_i - \sum_{i=1}^{N+1} g_i(x^1)=1-1 =0. \] Consequently, \[ \begin{split} 2\epsilon_0(N+1)& = 2\epsilon_0(N+1)+ \sum_{i=1}^{N+1} x^1_i - \sum_{i=1}^{N+1} g_i(x^1) \\ & >\sum_{i=1}^{N+1}\left( x_i^i-g_i(x^i)\right)\\ &=\sum_{i=1}^{N+1}\left| x_i^i-g_i(x^i)\right|. \end{split} \] Using (\ref{eq:delta-bound}) and (\ref{eq:delta-eps-bound}) and the continuity of $g$, for all $i$, we have that: $ |x^1_i - g_i(x^1)| \leq |x^1_i - x^i_i| + |x^i_i - g_i(x^i)| + |g_i(x^i)-g_i(x^1)| < 2 \epsilon_0 + |x^i_i - g_i(x^i)| $. Hence, \[ \begin{split} \bar d(x^1 , g(x^1)) = \sum_{i=1}^{N+1} \frac{|x^1_i - g_i(x^1)|}{2^i(1+|x^1_i - g_i(x^1)|)} & \leq \sum_{i=1}^{N+1} |x^1_i - g_i(x^1)| \\ & < \sum_{i=1}^{N+1} \left(2 \epsilon_0 + |x^i_i - g_i(x^i)|\right) \\ & < 4(N+1)\epsilon_0 \\ & = \frac{\epsilon}{2}. \end{split} \] Let $y = (x^1_1,x^1_2,\dots,x^1_N,0,0,0,\dots)$. We see that \begin{equation} \label{eq:eps-over-2-1toN} \begin{split} \sum_{i=1}^{N} \frac{|y_i - f_i(y)|}{2^i(1+|y_i - f_i(y)|)} &=\sum_{i=1}^{N} \frac{|y_i - g_i(y)|}{2^i(1+|y_i - g_i(y)|)}\\ & = \sum_{i=1}^{N} \frac{|x^1_i - g_i(x^1)|}{2^i(1+|x^1_i - g_i(x^1)|)} \\ & \leq \sum_{i=1}^{N+1} \frac{|x^1_i - g_i(x^1)|}{2^i(1+|x^1_i - g_i(x^1)|)} \\ & < \frac{\epsilon}{2}. \end{split} \end{equation} From (\ref{eq:eps-over-2-N+1on}) and (\ref{eq:eps-over-2-1toN}), we have \[ \begin{split} \bar d(y,f(y)) & = \sum_{i=1}^\infty \frac{|y_i - f_i(y)|}{2^i(1+|y_i - f_i(y)|)} \\ & =\sum_{i=1}^{N} \frac{|y_i - f_i(y)|}{2^i(1+|y_i - f_i(y)|)} +\sum_{i=N+1}^\infty \frac{|y_i - f_i(y)|}{2^i(1+|y_i - f_i(y)|)} \\ & < \frac{\epsilon}{2}+\frac{\epsilon}{2}\\ &=\epsilon . \end{split} \] Therefore, $y$ is the desired $\epsilon$-fixed point.\end{proof} Notice that the construction of the $\epsilon/2$ fixed point in $F^N$ in the proof above is identical to the construction of an $\epsilon/2$ fixed point for an arbitrary continuous function on $\Delta^N$, because of the isometry between the two sets. This construction, in conjunction with Lemmas \ref{le epsilonfixed} and \ref{le metricequivalence}, provides a proof of the Brouwer Fixed Point Theorem on the finite-dimensional simplex, which is similar to constructions found in, e.g., \cite{Todd76}. \section{Schauder's Theorem} A well-known infinite-dimensional fixed point theorem that holds for normed spaces is Schauder's theorem \cite{DuGr82, Smar74}: \begin{sch} Suppose that $X$ is a compact convex subset of the normed space $G$. If $f:X\rightarrow X$ is continuous, then $f$ has a fixed point. \end{sch} In this section we show how our proof of Theorem \ref{fixed-dio} can be used to prove Schauder's Theorem for the case where $X$ is infinite-dimensional. (The finite-dimensional version of Schauder's Theorem reduces to the Brouwer Fixed Point Theorem.) Recall that a space $X$ has the \textit{fixed point property}. if every continuous function $f:X\rightarrow X$ has a fixed point. Note that this is a topological property, so if $X$ is homeomorphic to $Y$ then $Y$ also has the fixed point property. We will establish Schauder's theorem by noting that $\ensuremath{\Delta^\infty_0} $ is homeomorphic to any infinite-dimensional compact convex subset of a normed space. Define the vector space $H$ to be \[H = \{x \in \ensuremath{\mathbb{R}} ^\infty | \sum_{i=1}^\infty \frac{|x_i|}{2^i} < \infty \}.\] It is not difficult to see that $H$ is indeed a vector space. Furthermore, we see that $\|x\| = \sum_{i=1}^\infty \frac{|x_i|}{2^i}$ defines a norm on this space and the closure of the standard simplex in $H$ is \[ \ensuremath{\Delta^H_0} = \{x\in H | \sum_{i=1}^\infty x_i \leq 1 \mbox{ and } 0\leq x_i\leq 1\} . \] \begin{prop} $\ensuremath{\Delta^\infty_0} $ is homeomorphic to $\ensuremath{\Delta^H_0} $. \end{prop} The proof of this lemma is trivial using the homeomorphism $g:\ensuremath{\Delta^\infty_0} \rightarrow \ensuremath{\Delta^H_0} $ being $g(x)=x$. Note that $\ensuremath{\Delta^H_0} $ is an infinite-dimensional compact convex subset of a normed space $H$. Now consider the following proposition \cite{Klee55}: \begin{prop} Every infinite-dimensional compact convex subset of a normed space is homeomorphic to the Hilbert Cube. \end{prop} The significance of these propositions is that \emph{every} infinite-dimensional compact convex subset of a normed space is homeomorphic to $\ensuremath{\Delta^\infty_0} $. Thus Theorem \ref{fixed-dio} implies the infinite-dimensional case of Schauder's Theorem. \begin{comment This result relies on Keller's Theorem, which states that every infinite-dimensional compact convex subset of the Hilbert Cube is homeomorphic to the Hilbert Cube, and the following theorem: \begin{theorem} Every compact convex subset of a normed space is linearly homeomorphic to a compact convex subset of the Hilbert Cube. \end{theorem} Though \cite{Smart74} states this for Banach spaces, he proves it for normed spaces.\footnote{It is worth noting that, to achieve full generality, the proof in (reference Smart) relies on a consequence of the Hahn-Banach Theorem, and thus on the Axiom of Choice. However, in specific cases, such as inner product spaces, the Hahn-Banach Theorem can be avoided. Also, to fully understand the proof in \cite{Smart74} one must recall that compact subsets of normed spaces are separable and that a continuous bijective map from a compact space to a Hausdorff space is a homeomorphism.} Furthermore, notice that linear homeomorphisms preserve the dimension of a set. Thus Theorem 3 gives us that every infinite-dimensional compact convex subset of a normed space is homeomorphic to an infinite dimensional compact convex subset of the Hilbert Cube. Theorem 3, in conjunction with Keller's Theorem, which can be found in \cite{VanM89}, gives us that all infinite-dimensional compact convex subsets of normed spaces are homeomorphic to the Hilbert Cube. \end{comment \bibliographystyle{plain}

https://arxiv.org/abs/1911.12009

Involution pipe dreams

Involution Schubert polynomials represent cohomology classes of $K$-orbit closures in the complete flag variety, where $K$ is the orthogonal or symplectic group. We show they also represent $T$-equivariant cohomology classes of subvarieties defined by upper-left rank conditions in the spaces of symmetric or skew-symmetric matrices. This geometry implies that these polynomials are positive combinations of monomials in the variables $x_i + x_j$, and we give explicit formulas of this kind as sums over new objects called involution pipe dreams. Our formulas are analogues of the Billey-Jockusch-Stanley formula for Schubert polynomials. In Knutson and Miller's approach to matrix Schubert varieties, pipe dream formulas reflect Gröbner degenerations of the ideals of those varieties, and we conjecturally identify analogous degenerations in our setting.

\section{Introduction} One can identify the equivariant cohomology rings for the spaces of symmetric and skew-symmetric complex matrices with multivariate polynomial rings. Under this identification, we show that the classes of certain natural subvarieties of (skew-)symmetric matrices are given by the \emph{involution Schubert polynomials} introduced by Wyser and Yong in \cite{wyser-yong-orthogonal-symplectic}. These classes of varieties generalize various others studied in the settings of degeneracy loci and combinatorial commutative algebra, for instance the (skew-)symmetric determinantal varieties studied by Harris and Tu~\cite{harris-tu}. Involution Schubert polynomials have a combinatorial formula for their monomial expansion~\cite{HMP1}. As a consequence of our geometric results, they must also expand as sums of products of binomials $x_i + x_j$. We give a combinatorial description of these expansions, which is a new analogue of the classic Billey-Jockusch-Stanley expansion for ordinary Schubert polynomials~\cite{BJK}. This description is far more compact than the monomial expansion. Our formulas involve novel objects that we call \emph{involution pipe dreams}. Involution pipe dreams appear to be the fundamental objects necessary to replicate Knutson and Miller's program~\cite{knutson-miller} to understand our varieties from a commutative algebra perspective. \subsection{Three flavors of matrix Schubert varieties}\label{intro1-sect} Fix a positive integer $n$. Let $\mathsf{GL}_n$ denote the general linear group of complex $n\times n$ invertible matrices, and write $B$ and $B^+$ for the Borel subgroups of lower- and upper-triangular matrices in $\mathsf{GL}_n$. Our work aims to extend what is known about the geometry of the $B$-orbits on matrix space to symmetric and skew-symmetric matrix spaces. We begin with some classical background. Consider the \emph{type A flag variety} $\Fl{n} = \mathsf{GL}_n / B$. The subgroup $B^+$ acts on $\Fl{n} $ with finitely many orbits, which are naturally indexed by permutations $w$ in the symmetric group $S_n$ of permutations of $\{1,2,\dots,n\}$. These orbits afford a CW decomposition of $\Fl{n}$, so the cohomology classes of their closures $X_w$, the \emph{Schubert varieties}, form a basis for the integral singular cohomology ring $H^*(\Fl{n})$. Borel's isomorphism explicitly identifies $H^*(\Fl{n})$ with a quotient of the polynomial ring $\mathbb{Z}[x_1,\dots,x_n]$, and the \emph{Schubert polynomials} $\mathfrak{S}_w \in \mathbb{Z}[x_1,\dots,x_n]$ are (non-unique) representatives for the \emph{Schubert classes} $[X_w] \in H^*(\Fl{n})$. The maximal torus $T$ of diagonal matrices in $\mathsf{GL}_n$ also acts on $\Fl{n}$, so we can instead consider the equivariant cohomology ring $H^*_T(\Fl{n})$. Via an extension of Borel's isomorphism, this ring is isomorphic to a quotient of $\mathbb{Z}[x_1, \ldots, x_n, y_1, \ldots, y_n]$. Lascoux and Sch\"utzenberger \cite{lascoux1982structure} introduced the \emph{double Schubert polynomials} $\mathfrak{S}_w(x,y)$ to represent the equivariant classes $[X_w]_T\in H^*_T(\Fl{n})$. These representatives are distinguished in the following sense. Let $\textsf{Mat}_n$ be the set of $n\times n$ complex matrices and write $\iota:\mathsf{GL}_n \hookrightarrow \textsf{Mat}_n$ for the obvious inclusion. The product group $ T \times T$ acts on $A\in \textsf{Mat}_n$ by $(t_1, t_2) \cdot A = t_1 A t_2^{-1}$. The \emph{matrix Schubert variety} of a permutation $w \in S_n$ is $M\hspace{-0.5mm}X_w = \overline{\iota(X_w)}$. Since $M_n$ is $T \times T$-equivariantly contractible, $H^*_{T \times T}(\textsf{Mat}_n) \cong H^*_{T \times T}(\text{point}) \cong \mathbb{Z}[x_1,\dots,x_n,y_1,\dots,y_n]$. The following theorem is roughly equivalent to work of Fulton~\cite{fulton1992flags}, and stated explicitly in~\cite{knutson-miller}; see \cite[Chpt. 15]{miller2004combinatorial}. \begin{thm}[\cite{fulton1992flags,knutson-miller}] \label{t:fulton} For all $w \in S_n$, we have $\mathfrak{S}_w(x,y) = [M\hspace{-0.5mm}X_w] \in H^*_{T \times T}(\textsf{Mat}_n)$. \end{thm} Our results are related to the geometry of certain spherical varieties studied by Richardson and Springer in \cite{richardson-springer}. Specifically, define the \emph{orthogonal group} $\O_n$ as the subgroup of $\mathsf{GL}_n$ preserving a fixed nondegenerate symmetric bilinear form on $\mathbb{C}^n$, and when $n$ is even define the \emph{symplectic group} $\mathsf{Sp}_n$ as the subgroup of $\mathsf{GL}_n$ preserving a fixed nondegenerate skew-symmetric bilinear form. We consider the actions of $\O_n$ and $\mathsf{Sp}_n$ (when $n$ is even) on $\Fl{n}$. The associated orbit closures $\hat X_y$ and $\hat X^{\fpf}_z$ are indexed by arbitrary involutions $y$ and fixed-point-free involutions $z$ in $S_n$. Let $\kappa(y)$ denote the number of 2-cycles in an involution $y=y^{-1} \in S_n$. Wyser and Yong \cite{wyser-yong-orthogonal-symplectic} constructed certain polynomials $\hat{\mathfrak{S}}_y, \hat{\mathfrak{S}}^{\fpf}_z \in \mathbb{Z}[x_1,\dots,x_n]$ and showed that the classes $[\hat X_y]$ and $[\hat X^{\fpf}_z]$ are represented in $H^*(\Fl{n})$ by $2^{\kappa(y)}\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$. We refer to $\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$ as \emph{involution Schubert polynomials}; for their precise definitions, see Section~\ref{schub-sect}. Write $\textsf{SMat}_n$ and $\textsf{SSMat}_n$ for the sets of symmetric and skew-symmetric $n\times n$ complex matrices. Let $t \in T$ act on these spaces by $t\cdot A = tAt$. One can identify the $T$-equivariant cohomology rings of both spaces with $\mathbb{Z}[x_1, \ldots, x_n]$; see the discussion in Section~\ref{subsec:eq-cohom}. For each involution $y \in S_n$, let $M\hspace{-0.5mm}\hat X_y = M\hspace{-0.5mm}X_y \cap \textsf{SMat}_n$. Similarly, for each fixed-point-free involution $z \in S_n$, let $M\hspace{-0.5mm}\hat X^{\fpf}_z = M\hspace{-0.5mm}X_z \cap \textsf{SSMat}_n$. Our first main result is a (skew-)symmetric analogue of Theorem~\ref{t:fulton}: \begin{thm} \label{thm:inv-schubert-classes} For all involutions $y$ and fixed-point-free involution $z$ in $S_n$, we have \[ 2^{\kappa(y)}\hat{\mathfrak{S}}_y = [M\hspace{-0.5mm}\hat X_y] \in H_T^*(\textsf{SMat}_n) \quad\text{and}\quad \hat{\mathfrak{S}}^{\fpf}_z = [M\hspace{-0.5mm}\hat X^{\fpf}_z] \in H_T^*(\textsf{SSMat}_n). \] Thus, involution Schubert polynomials are also equivariant cohomology representatives for symmetric and skew-symmetric matrix varieties. \end{thm} Our proof of this theorem appears in Section~\ref{sec:inv-matrix-schubert}. \begin{rmk} Another family of varieties indexed by involutions in $S_n$ has been studied by Fink, Rajchgot and Sullivant~\cite{fink2016matrix}. However, the varieties in \cite{fink2016matrix} are cut out by northeast rank conditions, while $M\hspace{-0.5mm}\hat X_y$ and $M\hspace{-0.5mm}\hat X^{\fpf}_z$ are cut out by northwest rank conditions (see \eqref{imxy-eq} and \eqref{imxz-eq} in Section~\ref{ss:classes}). \end{rmk} \subsection{Three flavors of pipe dreams} If $Z$ is a closed subvariety of $\textsf{SMat}_n$ or $\textsf{SSMat}_n$, then its $T$-equivariant cohomology class is a positive integer combination of products of binomials $x_i + x_j$ (see Corollary~\ref{cor:positivity}). Our second main result gives a combinatorial description of such an expansion for $\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$. Let $[n] = \{1, 2, \ldots, n\}$ and $\raisebox{-0.5pt}{\tikz{\draw (0,.25) -- (.25,.25) -- (0,0) -- (0,.25);}}\hspace{-0.5mm}_n = \{(i,j) \in [n] \times [n] : i+j \leq n\}$. Consider a subset $D\subseteq \raisebox{-0.5pt}{\tikz{\draw (0,.25) -- (.25,.25) -- (0,0) -- (0,.25);}}\hspace{-0.5mm}_n$. One associates to $D$ a \emph{wiring diagram} by replacing the cells $(i,j) \in \raisebox{-0.5pt}{\tikz{\draw (0,.25) -- (.25,.25) -- (0,0) -- (0,.25);}}\hspace{-0.5mm}_n$ by tiles of two types, given either by a crossing of two paths (drawn as a $\raisebox{-1pt}{\tikz[scale=0.7]{ \draw (-0.2,0) -- (0.2,0); \draw (0,-0.2) -- (0,0.2);}}$ tile) if $(i,j) \in D$ or by two paths bending away from each other (drawn as a $\raisebox{-1pt}{\tikz[scale=0.7]{ \draw (-0.2,0) to[bend right] (0,0.2); \draw (0,-0.2) to[bend left] (0.2,0); }}$ tile) if $(i,j) \notin D$. Connecting the endpoints of adjacent tiles yields a union of $n$ continuously differentiable paths, which we refer to as ``pipes.'' For example: \begin{equation}\label{first-wiring-eq} D=\{(1,3),(2,1)\} \qquad\text{corresponds to} \qquad \begin{tikzpicture}[scale=0.3,baseline=(c.base)] \node at (-0.5,3.5) {$\scriptstyle 1$}; \node (c) at (-0.5,2.5) {$\scriptstyle 2$}; \node at (-0.5,1.5) {$\scriptstyle 3$}; \node at (-0.5,0.5) {$\scriptstyle 4$}; % \node at (0.5,4.5) {$\scriptstyle 1$}; \node at (1.5,4.5) {$\scriptstyle 2$}; \node at (2.5,4.5) {$\scriptstyle 3$}; \node at (3.5,4.5) {$\scriptstyle 4$}; \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to (3,3.5) (2.5,3) to (2.5,4); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \end{equation} \begin{defn} A subset $D\subseteq \raisebox{-0.5pt}{\tikz{\draw (0,.25) -- (.25,.25) -- (0,0) -- (0,.25);}}\hspace{-0.5mm}_n$ is a \emph{reduced pipe dream} if no two pipes in the associated wiring diagram cross more than once. % \end{defn} This condition holds in the example \eqref{first-wiring-eq}. Pipe dreams as described here were introduced by Bergeron and Billey \cite{bergeron-billey}, inspired by related diagrams of Fomin and Kirillov \cite{fomin-kirillov-yang-baxter}. Bergeron and Billey originally referred to pipe dreams as \emph{reduced-word compatible sequence graphs} or \emph{rc-graphs} for short. A reduced pipe dream $D$ determines a permutation $w \in S_n$ in the following way. Label the left endpoints of the pipes in $D$'s wiring diagram by $1, 2, \ldots, n$ from top to bottom, and the top endpoints by $1, 2, \ldots, n$ from left to right. Then the associated permutation $w \in S_n$ is the element such that the pipe with left endpoint $i$ has top endpoint $w(i)$. For instance, the permutation of $D=\{(1,3),(2,1)\}$ is $w = 1423 \in S_4$. Let $\mathcal{PD}(w)$ denote the set of all reduced pipe dreams associated to $w \in S_n$. Pipe dreams are of interest for their role in formulas for $\mathfrak{S}_w$ and $\mathfrak{S}_w(x,y)$. Lascoux and Sch\"utzenberger's original definition of these Schubert polynomials in \cite{lascoux1983symmetry} is recursive in terms of \emph{divided difference operators}. However, by results of Fomin and Stanley \cite[\S4]{fomin1994schubert} we also have \begin{equation} \label{eq:double-schubert-def} \mathfrak{S}_w = \sum_{D \in \mathcal{PD}(w)} \prod_{(i,j) \in D} x_i \qquad\text{and}\qquad \mathfrak{S}_w(x,y) = \sum_{D \in \mathcal{PD}(w)} \prod_{(i,j) \in D} (x_i - y_j). \end{equation} The first identity is the Billey-Jockusch-Stanley formula for Schubert polynomials from \cite{BJK}. There are analogues of this formula for the involution Schubert polynomials $\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$, which involve the following new classes of pipe dreams. A reduced pipe dream $D\subseteq \raisebox{-0.5pt}{\tikz{\draw (0,.25) -- (.25,.25) -- (0,0) -- (0,.25);}}\hspace{-0.5mm}_n$ is \emph{symmetric} if $(i,j) \in D$ implies $(j,i) \in D$, and \emph{almost-symmetric} if both of the following properties hold: \begin{itemize} \item If $(i, j) \in D$ where $i<j$ then $(j,i) \in D$. \item If $(j, i) \in D$ where $i<j$ but $(i, j) \notin D$, then the pipes crossing at $(j, i)$ in the wiring diagram of $D$ are also the pipes that avoid each other at $(i, j)$. \end{itemize} Equivalently, $D$ is almost-symmetric if it is as symmetric as possible while respecting the condition that no two pipes cross twice, and any violation of symmetry forced by this condition takes the form of a crossing $(j,i)$ below the diagonal rather than at the transposed position $(i,j)$. Let $\mathcal{I}_n=\{ w \in S_n : w=w^{-1}\}$ and write $\I^{\fpf}_n$ for the subset of fixed-point-free elements of $\mathcal{I}_n$. Note that $n$ must be even for $\I^{\fpf}_n$ to be non-empty. Also let \[ \ltriang_n = \{(j,i) \in [n] \times [n] : i \leq j\} \qquad\text{and}\qquad \ltriang^{\!\!\neq}_n = \{(j,i) \in [n] \times [n] : i < j\}. \] \begin{defn} \label{defn:inv-pipe-dreams-1} The set of \emph{involution pipe dreams} for $y \in \mathcal{I}_n$ is \begin{equation*} \mathcal{ID}(y) = \{D \cap \ltriang_n : D \in \mathcal{PD}(y) \text{ is almost-symmetric}\}. \end{equation*} The set of \emph{fpf-involution pipe dreams} for $z \in \I^{\fpf}_{n}$ is \begin{equation*} \mathcal{FD}(z) = \{D \cap \ltriang^{\!\!\neq}_n : D \in \mathcal{PD}(z) \text{ is symmetric}\}. \end{equation*} By convention, (fpf-)involution pipe dreams are always instances of reduced pipe dreams. It would be more precise to call our objects ``reduced involution pipe dreams,'' but since we will never consider any pipe dreams that are unreduced, we opt for more concise terminology. \end{defn} We can now state our second main result, which will reappear as Theorems~\ref{inv-pipe-schubert-thm} and \ref{fpf-pipe-schubert-thm}. \begin{thm} \label{thm:inv-pipe-dream-formula} If $y \in \mathcal{I}_n$ and $z \in \I^{\fpf}_n$ then \[ \hat{\mathfrak{S}}_y = \sum_{D \in \mathcal{ID}(y)} \prod_{(i,j) \in D} 2^{-\delta_{ij}}(x_i + x_j) \quad\text{and}\quad \hat{\mathfrak{S}}^{\fpf}_z = \sum_{D \in \mathcal{FD}(z)} \prod_{(i,j) \in D} (x_i + x_j) \] where $\delta_{ij}$ denotes the usual Kronecker delta function. \end{thm} \begin{ex} The involution $y = 1432 = (2,4) \in \mathcal{I}_4$ has five reduced pipe dreams: \begin{center} \begin{tikzpicture}[scale=0.3] \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to (2,2.5) (1.5,2) to (1.5,3); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to (1,1.5) (0.5,1) to (0.5,2); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \qquad \begin{tikzpicture}[scale=0.3] \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to (3,3.5) (2.5,3) to (2.5,4); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to (1,1.5) (0.5,1) to (0.5,2); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \qquad \begin{tikzpicture}[scale=0.3] \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to (3,3.5) (2.5,3) to (2.5,4); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to (2,2.5) (1.5,2) to (1.5,3); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \qquad \begin{tikzpicture}[scale=0.3] \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to (3,3.5) (2.5,3) to (2.5,4); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to (1,1.5) (0.5,1) to (0.5,2); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \qquad \begin{tikzpicture}[scale=0.3] \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to (2,2.5) (1.5,2) to (1.5,3); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \end{center} Only the last two of these are almost-symmetric, so $|\mathcal{ID}(y)| = 2$ and Theorem~\ref{thm:inv-pipe-dream-formula} reduces to the formula $\hat{\mathfrak{S}}_{y} = (x_2+x_1)(x_3+x_1) + (x_2+x_1)(x_2 + x_2)/2 = (x_2 + x_1)(x_3 + x_1 +x_2)$. The monomial expansion has six terms, as opposed to two. In general, the expansion in Theorem~\ref{thm:inv-pipe-dream-formula} uses roughly a factor of $2^{\deg \hat{\mathfrak{S}}_y}$ fewer terms. \end{ex} Theorem~\ref{thm:inv-pipe-dream-formula} has equivalent formulations in terms of compatible sequences. These descriptions more closely mirror the Billey-Jockusch-Stanley formula for $\mathfrak{S}_w$ from \cite{BJK}. \begin{rmk} There is an alternate path towards establishing the fact that the class of a matrix Schubert variety is represented by the weighted sum of reduced pipe dreams. The defining ideal of $M\hspace{-0.5mm}X_w$ has a simple set of generators due to Fulton~\cite{fulton1992flags}. Knutson and Miller showed that Fulton's generators form a Gr\"obner basis with respect to any anti-diagonal term order~\cite{knutson-miller}. The Gr\"obner degeneration of this ideal decomposes into a union of coordinate subspaces indexed by reduced pipe dreams. Our hope is that a similar program can be implemented in the (skew-)symmetric setting, which would give a geometric proof of Theorem~\ref{thm:inv-pipe-dream-formula}. We discuss this in greater detail in Section~\ref{sec:ideals}. \end{rmk} In addition to Theorem~\ref{thm:inv-pipe-dream-formula}, we also prove a number of results about the properties of involution pipe dreams. An outline of the rest of this article is as follows. Section~\ref{prelim-sect} contains some preliminaries on involution Schubert polynomials along with a proof of Theorem~\ref{thm:inv-schubert-classes}. In Section~\ref{sec:reduced-words}, we give several equivalent characterizations of $\mathcal{ID}(y)$ and $\mathcal{FD}(z)$ in terms of {reduced words} for permutations. Section~\ref{sec:inv-polynom} contains our proof of Theorem~\ref{thm:inv-pipe-dream-formula}, which uses ideas from recent work of Knutson~\cite{Knutson} along with certain transition equations for $\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$ given in \cite{HMP3}. In Section~\ref{sec:generating-pipe-dreams} we show that both families of involution pipe dreams are obtained from distinguished ``bottom'' elements by repeatedly applying certain simple transformations. These transformations are extensions of the \emph{ladder moves} for pipe dreams described by Bergeron and Billey in \cite{bergeron-billey}. In Section~\ref{sec:future}, finally, we describe several related open problems and conjectures. \subsection*{Acknowledgements} The second author was partially supported by Hong Kong RGC Grant ECS 26305218. We thank Allen Knutson for explaining to us his proof of Equation~(\ref{eq:double-schubert-def}) prior to the appearance of~\cite{Knutson}. \section{Schubert polynomials and matrix varieties}\label{prelim-sect} Everywhere in this paper, $n$ denotes a fixed positive integer. For convenience, we realize the symmetric group $S_n$ as the group of permutations of $\ZZ_{>0} = \{1,2,3,\dots\}$ fixing all $i>n$, so that there is an automatic inclusion $S_n \subset S_{n+1}$. In this section, we present some relevant background on involution Schubert polynomials and equivariant cohomology, and then prove Theorem~\ref{thm:inv-schubert-classes}. \subsection{Involution Schubert polynomials}\label{schub-sect} To start, we provide a succinct definition of $\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$ in terms of the ordinary Schubert polynomials $\mathfrak{S}_w$ given by \eqref{eq:double-schubert-def}. Let $s_i = (i,i+1) \in S_n$ for each $i \in [n-1]$. A \emph{reduced word} for $w \in S_n$ is a minimal-length sequence $a_1a_2 \cdots a_l$ such that $w=s_{a_1}s_{a_2} \cdots s_{a_l}$. Let $\mathcal{R}(w)$ denote the set of reduced words for $w$. The \emph{length} $\ell(w)$ of $w \in S_n$ is the length of any word in $\mathcal{R}(w)$. One has $\ell(ws_i)=\ell(w)+1>\ell(w)$ if and only if $w(i)<w(i+1)$. \begin{prop}[{\cite[Thm. 7.1]{Humphreys}}] \label{demazure-prop} There is a unique associative operation $\circ : S_n \times S _n \to S_n$, called the \emph{Demazure product}, with $s_i \circ s_i = s_i$ for all $i \in [n-1]$ and $v\circ w = vw$ for all $v,w \in S_n$ with $\ell(vw) = \ell(v)+\ell(w)$. \end{prop} An \emph{involution word} for $y \in \mathcal{I}_n = \{ w \in S_n : w=w^{-1}\}$ is a minimal-length word $a_1a_2 \cdots a_l$ with \begin{equation}\label{iii-eq} y = s_{a_l} \circ \cdots \circ s_{a_2}\circ s_{a_1} \circ 1 \circ s_{a_1} \circ s_{a_2} \circ \cdots \circ s_{a_l}. \end{equation} An \emph{atom} for $y \in \mathcal{I}_n$ is a minimal-length permutation $w \in S_n$ with $ y = w^{-1} \circ w. $ Let $\hat{\mathcal{R}}(y)$ be the set of involution words for $y \in \mathcal{I}_n$ and let $\mathcal{A}(y)$ be the set of atoms for $y$. The associativity of the Demazure product implies that \label{bigsqcup-eq} \hat{\mathcal{R}}(y) = \bigsqcup_{w \in \mathcal{A}(y)} \mathcal{R}(w) $. \begin{ex} If $y = 1432 $ then $\hat{\mathcal{R}}(y) = \{ 23, 32\}$ and $\mathcal{A}(y) = \{1342, 1423\}$. \end{ex} One can show that $ \mathcal{I}_n = \{ w^{-1} \circ w : w \in S_n \} $, so $\hat{\mathcal{R}}(y)$ and $\mathcal{A}(y)$ are nonempty for all $y \in \mathcal{I}_n$. Involution words are a special case of a more general construction of Richardson and Springer \cite{richardson-springer}, and have been studied by various authors \cite{can-joyce-wyser, hansson-hultman, hu-zhang, hultman-twisted-involutions}. Our notation follows \cite{HMP2,HMP1}. \begin{defn}\label{inv-sch-def} The \emph{involution Schubert polynomial} of $y \in \mathcal{I}_n$ is $ \hat{\mathfrak{S}}_y = \sum_{w \in \mathcal{A}(y)} \mathfrak{S}_w. $ \end{defn} Wyser and Yong \cite{wyser-yong-orthogonal-symplectic} originally defined these polynomials recursively using divided difference operators; work of Brion \cite{brion} implies that our definition agrees with theirs. For a detailed explanation of the equivalence among these definitions, see \cite{HMP1}. \begin{ex} If $z = 1432 \in \mathcal{I}_4$ then $\mathcal{A}(z) = \{1342, 1423\}$ and \begin{equation*} \hat{\mathfrak{S}}_{z} = \mathfrak{S}_{1342} + \mathfrak{S}_{1423} = (x_2 x_3 + x_1 x_3 + x_1 x_2) + (x_2^2 + x_1 x_2 + x_1^2). \end{equation*} \end{ex} Assume $n$ is even, so that $\I^{\fpf}_n = \{ z \in \mathcal{I}_n: i \neq z(i)\text{ for all $i \in [n]$}\}$ is nonempty, and let \[1^{\fpf}_n = (1,2)(3,4)\cdots (n-1,n) = s_1s_3\cdots s_{n-1} \in \I^{\fpf}_n.\] An \emph{fpf-involution word} for $z \in \I^{\fpf}_{n}$ is a minimal-length word $a_1a_2 \cdots a_l$ with \[ z = s_{a_l} \cdots s_{a_2} s_{a_1} 1^{\fpf}_n s_{a_1} s_{a_2} \cdots s_{a_l} .\] This formulation avoids the Demazure product, but there is an equivalent definition that more closely parallels \eqref{iii-eq}. Namely, by \cite[Cor. 2.6]{HMP2}, an {fpf-involution word} for $z \in \I^{\fpf}_{n}$ is also a minimal-length word $a_1a_2 \cdots a_l$ with $ z = s_{a_l} \circ \cdots \circ s_{a_2} \circ s_{a_1} \circ 1^{\fpf}_n \circ s_{a_1} \circ s_{a_2} \circ \cdots \circ s_{a_l} $. An \emph{fpf-atom} for $z \in \I^{\fpf}_n$ is a minimal length permutation $w \in S_n$ with $z = w^{-1} 1^{\fpf}_n w.$ Let $\mathcal{A}^{\fpf}(z)$ be the set of fpf-atoms for $z$, and let $ \hat{\mathcal{R}}^{\fpf}(z) $ be the set of fpf-involution words for $z$. The basic properties of reduced words imply that \hat{\mathcal{R}}^{\fpf}(z) = \bigsqcup_{w \in \mathcal{A}^{\fpf}(z)} \mathcal{R}(w). $ \begin{ex} If $z = 4321 $ then $\hat{\mathcal{R}}^{\fpf}(z) = \{ 23, 21\}$ and $\mathcal{A}^{\fpf}(z) = \{1342, 3124\}$. \end{ex} Note that $a_1a_2\cdots a_l $ belongs to $\hat{\mathcal{R}}^{\fpf}(z)$ if and only if $135\cdots(n-1)a_1a_2\cdots a_l $ belongs to $\hat{\mathcal{R}}(z)$. Moreover, if $z \in \I^{\fpf}_n$ then $\hat{\mathcal{R}}^{\fpf}(z) = \hat{\mathcal{R}}^{\fpf}(zs_{n+1}) $ and $ \mathcal{A}^{\fpf}(z) = \mathcal{A}^{\fpf}(zs_{n+1}).$ Fpf-involution words are special cases of reduced words for \emph{quasiparabolic sets} \cite{RainsVazirani}. Since $\I^{\fpf}_n$ is a single $S_n$-conjugacy class, each $z \in \I^{\fpf}_n$ has at least one fpf-involution word and fpf-atom. \begin{defn}\label{fpf-inv-sch-def} The \emph{fpf-involution Schubert polynomial} of $z \in \I^{\fpf}_n$ is $ \hat{\mathfrak{S}}^{\fpf}_z = \sum_{w \in \mathcal{A}^{\fpf}(z)} \mathfrak{S}_w. $ \end{defn} These polynomials were also introduced in \cite{wyser-yong-orthogonal-symplectic}. Note that $\hat{\mathfrak{S}}^{\fpf}_{z} =\hat{\mathfrak{S}}^{\fpf}_{zs_{n+1}}$ for all $z \in \I^{\fpf}_n$. \begin{ex} If $z = 532614 \in \I^{\fpf}_6$ then $\mathcal{A}^{\fpf}(z) = \{13452, 31254\}$ and \begin{equation*} \begin{aligned} \hat{\mathfrak{S}}^{\fpf}_{z} &= \mathfrak{S}_{13452} + \mathfrak{S}_{31254} = (x_2 x_3 x_4 + x_1 x_3 x_4 + x_1 x_2 x_4 + x_1 x_2 x_3) + (x_1^2 x_4 + x_1^2 x_3 + x_1^2 x_2 + x_1^3). \end{aligned} \end{equation*} \end{ex} \subsection{Torus-equivariant cohomology} \label{subsec:eq-cohom} Suppose $V$ is a finite-dimensional rational representation of a torus $T \simeq (\mathbb{C}^\times)^n$. A character $\lambda \in \operatorname{Hom}(T, \mathbb{C}^\times)$ is a \emph{weight} of $V$ if the \emph{weight space} $V_\lambda = \{v \in V : \text{$tv = \lambda(t)v$ for all $t \in T$}\}$ is nonzero. Any nonzero $v \in V_{\lambda}$ is a \emph{weight vector}, and $V$ has a basis of weight vectors. Let $\wt(V)$ denote the set of weights of $V$. After fixing an isomorphism $T \simeq (\mathbb{C}^\times)^n$, we identify the character $(t_1, \ldots, t_n) \mapsto t_1^{a_1} \cdots t_n^{a_n}$ with the linear polynomial $a_1 x_1 + \cdots + a_n x_n \in \mathbb{Z}[x_1, \ldots, x_n]$. The \emph{equivariant cohomology ring} $H_T(V)$ is isomorphic to $\mathbb{Z}[x_1, \ldots, x_n]$, an identification we make without comment from now on. Each $T$-invariant subscheme $X \subseteq V$ has an associated class $[X] \in H_T(V)$, which we describe following \cite[Chpt. 8]{miller2004combinatorial}. First, if $X$ is a linear subspace then we define $[X] = \prod_{\lambda \in \wt(X)} \lambda$, where we identify $\lambda$ with a linear polynomial as above. More generally, fix a basis of weight vectors of $V$, and let $z_1, \ldots, z_n \in V^*$ be the dual basis; this determines an isomorphism $\mathbb{C}[V] = \operatorname{Sym}(V^*) \simeq \mathbb{C}[z_1, \ldots, z_n]$. Choose a term order on monomials in $z_1, \ldots, z_n$, and let $\init(I)$ denote the ideal generated by the leading terms of all members of a given set $I \subseteq \mathbb{C}[V]$. Given that $\init(I)$ is a monomial ideal, one can show that each of its associated primes $\mathfrak{p}$ is also a monomial ideal, and hence of the form $\langle z_{i_1}, \ldots, z_{i_r} \rangle$. The corresponding subscheme $Z(\mathfrak{p})$ is a $T$-invariant linear subspace of $V$. Now define \begin{equation} \label{eq:equivariant-class} [X] = \sum_{\mathfrak{p}} \operatorname{mult}_{\mathfrak{p}}(\init I(X)) [Z(\mathfrak{p})] \end{equation} where $I(X)$ is the ideal of $X$ and $\mathfrak{p}$ runs over the associated primes of $\init I(X)$. \subsection{Classes of involution matrix Schubert varieties} \label{sec:inv-matrix-schubert} \label{ss:classes} The matrix Schubert varieties in Theorem~\ref{t:fulton} can be described in terms of rank conditions, namely: \[ M\hspace{-0.5mm}X_w = \{A \in \textsf{Mat}_n: \text{$\rank A_{[i][j]} \leq \rank w_{[i][j]}$ for $i,j \in [n]$}\}, \] where $\textsf{Mat}_n$ is the variety of $n\times n$ matrices, $A_{[i][j]}$ denotes the upper-left $i \times j$ corner of $A \in \textsf{Mat}_n$, and we identify $w \in S_n$ with the $n\times n$ permutation matrix having $1$'s in positions $(i,w(i))$. The varieties $M\hspace{-0.5mm}\hat X_y$ and $M\hspace{-0.5mm}\hat X^{\fpf}_z$ from Theorem~\ref{thm:inv-schubert-classes} can be reformulated in a similar way. Specifically, we define the \emph{involution matrix Schubert variety} of $y \in \mathcal{I}_n$ by \begin{equation}\label{imxy-eq} M\hspace{-0.5mm}\hat X_y = M\hspace{-0.5mm}X_y \cap \textsf{SMat}_n = \{A \in \textsf{SMat}_n : \text{$\rank A_{[i][j]} \leq \rank y_{[i][j]}$ for $i,j \in [n]$}\}, \end{equation} where $\textsf{SMat}_n$ is the subvariety of symmetric matrices in $\textsf{Mat}_n$. When $n$ is even, we define the \emph{fpf-involution matrix Schubert variety} of $z \in \I^{\fpf}_n$ by \begin{equation}\label{imxz-eq} M\hspace{-0.5mm}\hat X^{\fpf}_z = M\hspace{-0.5mm}X_z \cap \textsf{SSMat}_n = \{A \in \textsf{SSMat}_n : \text{$\rank A_{[i][j]} \leq \rank z_{[i][j]}$ for $i,j \in [n]$}\}, \end{equation} where $\textsf{SSMat}_n$ is the subvariety of skew-symmetric matrices in $\textsf{Mat}_n$. \begin{ex} Suppose $y = (2,3) =\left[\begin{smallmatrix} 1 & 0 & 0 \\ 0 & 0 & 1 \\ 0 & 1 & 0 \end{smallmatrix}\right]\in \mathcal{I}_3$. Setting $R_{ij} = \rank y_{[i][j]}$, we have $R = \left[\begin{smallmatrix} 1 & 1 & 1 \\ 1 & 1 & 2\\ 1 & 2 & 3 \end{smallmatrix}\right]$. The conditions $\rank A_{[i][j]} \leq R_{ij}$ for $i,j \in [3]$ defining $M\hspace{-0.5mm}\hat X_y$ are all implied by the single condition $\rank A_{[2][2]} \leq R_{22} = 1$. Thus, $M\hspace{-0.5mm}\hat X_y = \left\{\left[\begin{smallmatrix} z_{11} & z_{21} & z_{31} \\ z_{21} & z_{22} & z_{32} \\ z_{31} & z_{32} & z_{33} \end{smallmatrix} \right] : z_{11}z_{22} - z_{21}^2 = 0\right\}$. \end{ex} Let $T \subseteq \mathsf{GL}_n$ be the usual torus of invertible diagonal matrices. Recall that $\kappa(y) = |\{ i : y(i) < i\}|$ for $y \in \mathcal{I}_n$, and that $ T$ acts on matrices in $ \textsf{Mat}_n$ by $t\cdot A = tA$ and on symmetric matrices in $ \textsf{SMat}_n$ by $t \cdot A = tAt$. We can now prove Theorem~\ref{thm:inv-schubert-classes}, which states that if $y \in \mathcal{I}_n$ and $z \in \I^{\fpf}_n$ then $2^{\kappa(y)}\hat{\mathfrak{S}}_y = [M\hspace{-0.5mm}\hat X_y] \in H_T^*(\textsf{SMat}_n)$ while $\hat{\mathfrak{S}}^{\fpf}_z = [M\hspace{-0.5mm}\hat X^{\fpf}_z] \in H_T^*(\textsf{SSMat}_n)$. \begin{rmk} It is possible, though a little cumbersome, to derive Theorem~\ref{thm:inv-schubert-classes} from \cite[Thm. 2.17 and Lem. 3.1]{MP2019}, which provide a similar statement in complex $K$-theory. We originally announced Theorem~\ref{thm:inv-schubert-classes} in an extended abstract for this paper which preceded the appearance of \cite{MP2019}. However, as the argument below is similar to the proofs of the results in \cite{MP2019}, we will be somewhat curt here in our presentation of the details. \end{rmk} \begin{proof}[Proof of Theorem~\ref{thm:inv-schubert-classes}] If $X$ and $Y$ are complex varieties with $T$-actions, and $f : X \to Y$ is a $T$-equivariant morphism, then there is a pullback homomorphism $f^* : H_T^*(Y) \to H_T^*(X)$. If $f$ is a flat morphism (e.g., an inclusion of an open subset, a projection of a fiber bundle, or a composition of flat morphisms), then $f^*([Z]) = [f^{-1}(Z)]$ for any subscheme $Z \subseteq Y$. Because $T$ acts freely on $\mathsf{GL}_n$ and since $\mathsf{GL}_n/T \twoheadrightarrow \mathsf{GL}_n/B \simeq \Fl{n}$ is a homotopy equivalence (see, e.g., \cite[\S8.1]{McGovern}), one has $H_T^*(\mathsf{GL}_n) \simeq H^*(\mathsf{GL}_n/T) \simeq H^*(\Fl{n})$. If $Z \subseteq \mathsf{GL}_n$ is a $B$-invariant subvariety, then $[Z] \in H_T^*(\mathsf{GL}_n)$ corresponds to the class of $Z/B = \{gB : g \in Z\}$ in $H^*(\Fl{n})$. Fix $y \in \mathcal{I}_n$ and define $\sigma : \mathsf{GL}_n \to \textsf{SMat}_n$ by $\sigma(g) = gg^T$. Let $\iota : \mathsf{GL}_n \hookrightarrow M_n$ be the obvious inclusion and consider the diagram \begin{equation}\label{cd-eq} \begin{CD} H_T^*(\textsf{SMat}_n) @>>\sigma^*> H_T^*(\mathsf{GL}_n) @>>\sim> H^*(\Fl{n})\\ @AA\rotatebox{90}{$\sim$}A @AA\iota^*A\\ \mathbb{Z}[x_1, \ldots, x_n] @>>\sim> H_T^*(M_n) \end{CD} \end{equation} Realize $\O_n$ as the group $\{g \in \mathsf{GL}_n : gg^T = 1\}$. The map $\sigma$ is flat because it is the composition $\mathsf{GL}_n \twoheadrightarrow \mathsf{GL}_n/\O_n \hookrightarrow \textsf{SMat}_n$, where the second map sends $g\O_n \mapsto gg^T$ and may be identified with the open inclusion $\mathsf{GL}_n \cap \textsf{SMat}_n \hookrightarrow \textsf{SMat}_n$. For fixed $i \in [n]$, one checks using the prescription of $\S \ref{subsec:eq-cohom}$ that $2x_i$ represents both the class of $Z = \{A \in \textsf{SMat}_n : A_{ii} = 0\}$ in $H_T^*(\textsf{SMat}_n)$ and the class of $Z' = \{A \in M_n : (AA^T)_{ii} = 0\}$ in $H_T^*(M_n)$. Since $\sigma^*[Z] = [\sigma^{-1}(Z)] = [\iota^{-1}(Z')] = \iota^*[Z']$, this calculation implies that \eqref{cd-eq} commutes. Now set $ \hat X_y = \sigma^{-1}(M\hspace{-0.5mm}\hat X_y)/B = \{gB \in \Fl{n} : \rank (gg^T)_{[i][j]} \leq \rank y_{[i][j]} \text{ for $i,j \in [n]$}\},$ so that the path through the upper-left corner of \eqref{cd-eq} sends the polynomial $[M\hspace{-0.5mm}\hat X_y]$ to $[\hat X_y]$. The variety $\hat X_y$ is the closure of an $\O_n$-orbit on $\Fl{n}$ \cite[\S2.1.2]{wyser-degeneracy-loci}. The path through the lower-right corner of \eqref{cd-eq} is simply the classical Borel map $\mathbb{Z}[x_1, \ldots, x_n] \to H^*(\Fl{n})$, and it is known that $2^{\kappa(y)}\hat{\mathfrak{S}}_y$ is the \emph{unique} polynomial corresponding under the Borel map to every class $[\hat X_{y \times 1^m}]$, where $y \times 1^m$ is the permutation $y(1)\cdots y(n)(n+1)\cdots (n+m)$ \cite[Thm. 2]{wyser-yong-orthogonal-symplectic}. To conclude that $[M\hspace{-0.5mm}\hat X_y] = 2^{\kappa(y)}\hat{\mathfrak{S}}_y$, it therefore suffices to show that the polynomial $[M\hspace{-0.5mm}\hat X_{y \times 1^m}]$ is constant for fixed $y$ and varying $m$. For $y \neq 1 \in S_n$, define $\operatorname{maxdes}(y) = \max \{i \in \ZZ_{\geq 0} : y(i) > y(i+1)\}$. Replacing $[n]$ in the definition \eqref{imxy-eq} by $[\operatorname{maxdes}(y)]$ yields exactly the same variety $M\hspace{-0.5mm}\hat X_y$. Since $\operatorname{maxdes}(y \times 1^m)$ is independent of $m$, as is $\rank(y \times 1^m)_{[i][j]}$ for $i, j \in [\operatorname{maxdes}(y)]$, it follows that the ideals of $M\hspace{-0.5mm}\hat X_{y \times 1^m}$ for fixed $y$ and varying $m$ have a common generating set. It is clear from \S \ref{subsec:eq-cohom} that this means that the polynomial $[M\hspace{-0.5mm}\hat X_{y \times 1^m}]$ is independent of $m$. The proof for the skew-symmetric case is the same, replacing $\O_n$ by $\mathsf{Sp}_n$ and the map $\sigma : g \mapsto gg^T$ by $g \mapsto g\Omega g^T$, where $\Omega \in \mathsf{GL}_n$ is the nondegenerate skew-symmetric form preserved by $\mathsf{Sp}_n$. \end{proof} \begin{cor} \label{cor:positivity} The polynomial $2^{\kappa(y)}\hat{\mathfrak{S}}_y$ (respectively, $\hat{\mathfrak{S}}^{\fpf}_z$) is a positive integer linear combination of products of terms $x_i+x_j$ for $1 \leq i \leq j \leq n$ (respectively, $1 \leq i < j \leq n$). \end{cor} \begin{proof} The weights of $T$ acting on $\textsf{SMat}_n$ are $x_i+x_j$ for $1 \leq i \leq j \leq n$, while the weights of $\textsf{SSMat}_n$ are the same with the added restriction $i < j$. The expression \eqref{eq:equivariant-class} makes clear that the classes $[M\hspace{-0.5mm}\hat X_y]$ and $[M\hspace{-0.5mm}\hat X^{\fpf}_z]$ are positive integer linear combinations of products of these weights. \end{proof} \begin{rmk} Let $S$ be a maximal torus in $\O_n$. Let $T \times S$ act on $\mathsf{GL}_n$ by $(t,s) \cdot g = tgs^{-1}$ and on $\textsf{SMat}_n$ by $(t,s) \cdot A = tAt$. The map $\sigma : \mathsf{GL}_n \to \textsf{SMat}_n$, $g \mapsto gg^T$ considered above is then $T \times S$-equivariant. Since the second factor of $T \times S$ acts trivially on $\textsf{SMat}_n$, the polynomial $2^{\kappa(y)}\hat{\mathfrak{S}}_y$ still represents the class $[M\hspace{-0.5mm}\hat X_y] \in H_{T \times S}(\textsf{SMat}_n)$. It follows as in the proof of Theorem~\ref{thm:inv-schubert-classes} that $2^{\kappa(y)}\hat{\mathfrak{S}}_y$ also represents the class $[\hat X_y]_S \in H_{S}(\Fl{n})$. The latter fact was proven by Wyser and Yong \cite{wyser-yong-orthogonal-symplectic}, but our approach gives an explanation for the surprising existence of a representative for $[\hat X_y]_S$ not involving the $S$-weights. Similar remarks apply in the skew-symmetric case. \end{rmk} \section{Characterizing pipe dreams} \label{sec:reduced-words} The rest of this article is focused on the combinatorial properties of involution pipe dreams and their role in the formulas in Theorem~\ref{thm:inv-pipe-dream-formula} that manifest Corollary~\ref{cor:positivity}. In the introduction, we defined (fpf-)involution pipe dreams via simple symmetry conditions. In this section, we give an equivalent characterization in terms of ``compatible sequences'' related to involution words. \subsection{Reading words} For $p \in \mathbb{Z}$, the \emph{$p\textsuperscript{th}$ antidiagonal} (respectively, \emph{$p\textsuperscript{th}$ diagonal}) in $\ZZ_{>0} \times \ZZ_{>0}$ is the set \[\{(i,j) \in \ZZ_{>0}\times \ZZ_{>0} : i+j-1 = p\} \qquad(\text{respectively, }\{(i,j) \in \ZZ_{>0}\times \ZZ_{>0}: j-i = p\}).\] Labeling the elements of $\{1,2,3\}\times\{1,2,3\}$ by their respective antidiagonal and diagonal gives \begin{equation*} \begin{array}{ccc} 1 & 2 & 3\\ 2 & 3 & 4\\ 3 & 4 & 5\\ \end{array} \qquad \text{and} \qquad \begin{array}{rrr} 0 & -1 & -2\\ 1 & 0 & -1\\ 2 & 1 & 0 \end{array} \end{equation*} Let $\textsf{adiag}: \ZZ_{>0}\times \ZZ_{>0} \to \ZZ_{>0}$ be the map sending $(i,j) \mapsto i+j-1$. \begin{defn} The \emph{standard reading word} of $D\subseteq [n]\times [n]$ is the sequence \[ \textsf{word}(D) = \textsf{adiag}(\alpha_1) \textsf{adiag}(\alpha_2) \cdots \textsf{adiag}(\alpha_{|D|})\] where $\alpha_1,\alpha_2,\dots,\alpha_{|D|}$ are the positions of $D$ read row-by-row from right to left, starting with the top row. \end{defn} If one also records the row indices of the positions $\alpha_i$ as a second word, then the resulting words uniquely determine $D$ and are the same data as a \emph{compatible sequence} for $\textsf{word}(D)$ (see \cite[(1)]{BJK}). \begin{ex} The subset $D= \{ (1,3),(1,2),(2,3),(2,2),(3,2)\}$ has $\textsf{word}(D) = 32434$. \end{ex} We introduce a more general class of reading words. Suppose $\omega : [n] \times [n] \to [n^2]$ is a bijection. For a subset $D \subseteq [n] \times [n]$ with $\omega(D) = \{ i_1 < i_2< \dots < i_{m}\}$, let \[\textsf{word}(D,\omega) =\textsf{adiag}(\omega^{-1}(i_1))\textsf{adiag}(\omega^{-1}(i_2))\cdots \textsf{adiag}(\omega^{-1}(i_m)).\] The standard reading word of $D\subseteq [n]\times [n]$ corresponds to the bijection $\omega : (i,j) \mapsto ni - j + 1$. \begin{ex} If $n=2$ and $\omega$ is such that $ \left[ \begin{array}{rr} \omega(1,1) & \omega(1,2) \\ \omega(2,1) & \omega(2,2) \end{array}\right] = \left[ \begin{array}{cc} 3 & 1 \\ 4 & 2 \end{array} \right] $ then we would have $\textsf{word}([n]\times [n],\omega) = 2312$, while if $D = \{(1,1),(2,2)\}$ then $\textsf{word}(D,\omega) = 31$. \end{ex} For us, a \emph{linear extension} of a finite poset $(P,\preceq)$ with size $m=|P|$ is a bijection $\omega : P \to [m]$ such that $\omega(s) < \omega(t)$ whenever $s \prec t$ in $P$. \begin{defn} A \emph{reading order} on $[n]\times[n]$ is a linear extension of the partial order $\leq_{\textsf{NE}}$ on $[n] \times [n]$ that has $(i,j) \leq_{\textsf{NE}} (i',j')$ if and only if both $i \leq i'$ and $j \geq j'$. If $\omega$ is a reading order, then we refer to $\textsf{word}(D,\omega)$ as a \emph{reading word} of $D\subseteq [n]\times [n]$. \end{defn} The \emph{Coxeter commutation class} of a finite sequence of integers is its equivalence class under the relation that lets adjacent letters commute if their positive difference is at least two. For example, $\{1324, 3124,1342,3142, 3412\}$ is a single Coxeter commutation class. Fix a set $D \subseteq [n] \times [n]$. \begin{lem}\label{commutation-lem} All reading words of $D$ are in the same Coxeter commutation class. \end{lem} This result can be derived using Viennot's theory of heaps of pieces; see \cite[Lem. 3.3]{Viennot}. \begin{proof} Let $s_p \in S_{n^2}$ be the simple transposition interchanging $p$ and $p+1$, and choose a reading order $\omega$ on $[n]\times [n]$. The sequence $\textsf{word}(D, s_p \omega)$ is equal to $\textsf{word}(D, \omega)$ when $\{p,p+1\}\not\subset \omega(D)$, and otherwise is obtained by interchanging two adjacent letters in $\textsf{word}(D,\omega)$. In the latter case, if $\omega^{-1}(p) = (i,j)$ and $\omega^{-1}(p+1) = (i',j')$ are not in adjacent antidiagonals, then $\textsf{word}(D,\omega)$ and $\textsf{word}(D, s_p \omega)$ are in the same Coxeter commutation class. Now suppose $\upsilon$ is a second reading order on $[n]\times [n]$. We claim that one can pass from $\omega$ to $\upsilon$ by composing $\omega$ with a sequence of simple transpositions obeying the condition just described. To check this, we induct on the number of inversions in the permutation $\upsilon \omega^{-1} \in S_{n^2}$. If $\upsilon \omega^{-1}$ is not the identity, then there exists $p$ with $\upsilon(\omega^{-1}(p)) > \upsilon(\omega^{-1}(p+1))$. Since $\upsilon$ and $\omega$ are both linear extensions of $\leq_{\textsf{NE}}$, we can have neither $\omega^{-1}(p) \leq_{\textsf{NE}} \omega^{-1}(p+1)$ nor $\omega^{-1}(p+1) \leq_{\textsf{NE}} \omega^{-1}(p)$, so the cells $\omega^{-1}(p)$ and $\omega^{-1}(p+1)$ are not in adjacent antidiagonals. Therefore $\textsf{word}(D,\omega)$ and $\textsf{word}(D, s_p \omega)$ are in the same Coxeter commutation class, which by induction also includes $\textsf{word}(D,\upsilon)$. \end{proof} Each diagonal is an antichain for $\leq_{\textsf{NE}}$, so if $\omega$ first lists the elements on diagonal $-(n-1)$ in any order, then lists the elements on diagonal $-(n-2)$, and so on, then $\omega$ is a reading order. \begin{defn}\label{unimodal-def} The \emph{unimodal-diagonal reading order} on $[n]\times [n]$ is the reading order that lists the elements of the $p\textsuperscript{th}$ diagonal from bottom to top if $p < 0$, and from top to bottom if $p \geq 0$. The \emph{unimodal-diagonal reading word} of $D\subseteq [n]\times[n]$, denoted $\textsf{udiag}(D)$, is the associated reading word. \end{defn} The unimodal-diagonal reading order on $\{1,2,3,4\} \times \{1,2,3,4\}$ has values \begin{equation*} \begin{array}{ccccc} 7 & 6 & 3 & 1 \\ 11 & 8 & 5 & 2 \\ 14 & 12& 9 & 4 \\ 16& 15& 13& 10 \\ \end{array} \end{equation*} and if $D = \{1,2,3,4\} \times \{1,2,3,4\}$ then $\textsf{udiag}(D) = 4536421357246354$. \subsection{Pipe dreams} Recall the definitions of the sets of reduced words $\mathcal{R}(w)$, involution words $\hat{\mathcal{R}}(y)$, and fpf-involution words $\hat{\mathcal{R}}^{\fpf}(z)$ for $w \in S_n$, $y \in \mathcal{I}_n$, and $z \in \I^{\fpf}_n$ from Section~\ref{schub-sect}. The main results of this section are versions of the following theorem for involution pipe dreams and fpf-involution pipe dreams. \begin{thm}\label{pipe-thm} A subset $D \subseteq [n]\times [n]$ is a reduced pipe dream for $w \in S_n$ if and only if some (equivalently, every) reading word of $D$ is a reduced word for $w$. \end{thm} \begin{proof} Fix $D \subseteq [n]\times [n]$ and $w \in S_n$. The set $\mathcal{R}(w)$ is a union of Coxeter commutation classes, so $\textsf{word}(D)\in\mathcal{R}(w)$ if and only every reading word of $D$ belongs to $\mathcal{R}(w)$ by Lemma~\ref{commutation-lem}. Saying that $D$ is a reduced pipe dream for $w$ if and only if $\textsf{word}(D) \in \mathcal{R}(w)$ is Bergeron and Billey's original definition of an \emph{rc-graph} in \cite[\S3]{bergeron-billey}, and it is clear from the basic properties of permutation wiring diagrams that this is equivalent to the definition of a reduced pipe dream in the introduction. \end{proof} \begin{cor}[{\cite[Lem. 3.2]{bergeron-billey}}] \label{transpose-cor} If $D$ is a reduced pipe dream for $w \in S_n$ then $D^T$ is a reduced pipe dream for $w^{-1}$. \end{cor} Recall that the set $\mathcal{ID}(z)$ of \emph{involution pipe dreams} for $z \in \mathcal{I}_n$ consists of all intersections $D\cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$ where $D$ is a reduced pipe dream for $z$ that is almost-symmetric and $\raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n = \{ (j,i) \in [n]\times [n] : i \leq j\}$. \begin{thm}\label{thm:almost-symmetric} Suppose $z \in \mathcal{I}_n$ and $D \subseteq [n]\times [n]$. The following are equivalent: \begin{enumerate} \item[(a)] Some reading word of $D$ is an involution word for $z$. \item[(b)] Every reading word of $D$ is an involution word for $z$. \item[(c)] The set $D$ is a reduced pipe dream for some atom of $z$. \end{enumerate} Moreover, if $D \subseteq \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$ then $D\in \mathcal{ID}(z)$ if and only if these equivalent conditions hold. \end{thm} \begin{rmk} Although this theorem implies that $\mathcal{ID}(z) \subseteq \bigsqcup_{w \in \mathcal{A}(z)} \mathcal{PD}(w)$, it is possible for an atom $w \in \mathcal{A}(z)$ to have no reduced pipe dreams contained in $\raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$, in which case $\mathcal{ID}(z)$ and $\mathcal{PD}(w)$ are disjoint. See Example~\ref{3.10ex} for an illustration of this. \end{rmk} \begin{proof} Recall that $\hat{\mathcal{R}}(z)$ is the disjoint union of the sets $\mathcal{R}(w)$, running over all atoms $w \in \mathcal{A}(z)$. The equivalences (a) $\Leftrightarrow$ (b) $\Leftrightarrow$ (c) are clear from Lemma~\ref{commutation-lem} and Theorem~\ref{pipe-thm}. Assume $D \subseteq \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$. To prove the final assertion, it suffices to show that $D\in \mathcal{ID}(z)$ if and only if the unimodal-diagonal reading word of $D$ from Definition~\ref{unimodal-def} is an involution word of $z$. Suppose $|D| = m$ and $\textsf{udiag}(D) = a_1a_2\cdots a_m$. We construct a sequence $w_0,w_1,w_2,\dots,w_m$ of involutions as follows: start by setting $w_0 = 1$, and for each $i \in [m]$ define $ w_i = s_{a_i} w_{i-1} s_{a_i}$ if we have $w_{i-1} s_{a_i} \neq s_{a_i} w_{i-1}$, or else set $w_i = w_{i-1} s_{a_i} = s_{a_i} w_{i-1}$. For example, if $m=5$ and $a_1a_2a_3a_4a_5 =13235$ then this sequence has \[w_1 =s_1, \quad w_2= s_1s_3, \quad w_3 = s_2s_1s_3s_2, \quad w_4 = s_3 s_2s_1s_3s_2 s_3, \quad\text{and}\quad w_5 =s_3 s_2s_1s_3s_2 s_3 s_5. \] Let $b_l \cdots b_2 b_1$ be the subword of $a_m \cdots a_2 a_1$ which contains $a_i$ if and only if $w_i = s_{a_i} w_{i-1} s_{a_i}$. In our example with $m=5$ and $a_1a_2a_3a_4a_5= 13235$, we have $l=2$ and $ b_2 b_1 = a_4a_3= 32$. Let $(p_1,q_1)$, $(p_2,q_2)$, \dots, $(p_m,q_m)$ be the cells in $D$ listed in the unimodal-diagonal reading order and define $E = D\sqcup \{ (q_i,p_i) : w_i = s_{a_i} w_{i-1} s_{a_i}\}.$ If $\textsf{udiag}(D) = 13235$ then we could have \[D = \left\{\begin{smallmatrix} + & \cdot & \cdot & \cdot \\ + & + & \cdot & \cdot \\ + & \cdot & \cdot & \cdot\\ \cdot & + & \cdot & \cdot \end{smallmatrix}\right\} \qquad\text{then}\qquad E = \left\{\begin{smallmatrix} + & + & + & \cdot \\ + & + & \cdot & \cdot \\ + & \cdot & \cdot & \cdot \\ \cdot & + & \cdot & \cdot \end{smallmatrix}\right\}. \] By construction $\textsf{udiag}(E) = b_l \cdots b_2 b_1 a_1a_2\cdots a_m$ is a reduced word for $z$. It follows that $E$ is almost-symmetric since each $b_i$ has a corresponding $a_j$ and the associated cells are transposes of each other. The exchange principle (see, e.g., \cite[Lem. 3.4]{hultman-twisted-involutions}) implies that if $w \in \mathcal{I}_n$, $i \in [n-1]$, and $w(i)<w(i+1)$, then either $s_i ws_i =w\neq ws_i =s_i w = s_i \circ w \circ s_i$ or $ s_iw s_i = s_i \circ w \circ s_i \neq w$. From this, it is straightforward to show that $\textsf{udiag}(D) \in \hat{\mathcal{R}}(z) $ if and only if $\textsf{udiag}(E) \in \mathcal{R}(z)$; this also follows from the results in \cite[\S2]{HMP2}. Given the previous paragraph, we conclude that $\textsf{udiag}(D)\in \mathcal{R}(z)$ if and only if $D = E \cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$ is an involution pipe dream for $z$. \end{proof} \begin{ex}\label{3.10ex} Let $z = 1432 = (2,4) \in \mathcal{I}_4$. Since $ z = s_3 \circ s_2 \circ 1 \circ s_2 \circ s_3 = s_2 \circ s_3 \circ 1 \circ s_3 \circ s_2,$ we have $ 23 \in \hat{\mathcal{R}}(z)$ and $ 32\in \hat{\mathcal{R}}(z)$. These are the standard reading words of the involution pipe dreams $\{(2,1),(3,1)\}$ and $\{(2,1),(2,2)\}$, which may be drawn as \[ \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,2) {}; \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to (1,1.5) (0.5,1) to (0.5,2); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \qquad\text{and}\qquad \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,2) {}; \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to (2,2.5) (1.5,2) to (1.5,3); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \] The only involution pipe dream for $y = 321 \in \mathcal{I}_3$ is $\{(1,1),(2,1)\}$ which has standard reading word $12$. Although $\hat{\mathcal{R}}(y) =\{12,21\}$, there is no involution pipe dream with standard reading word $21$. \end{ex} We turn to the fixed-point-free case. \begin{lem}\label{diag-sym-lem} Assume $n$ is even. Suppose $z \in \mathcal{I}_n$ is an involution with a symmetric reduced pipe dream $D=D^T$. Then $z \in \I^{\fpf}_n$ if and only if $\{(i,i) : i \in [n/2]\}\subseteq D$. \end{lem} \begin{proof} Since $D$ is symmetric, it is straightforward to check that the pipes in cell $(i,i)$ of the wiring diagram of $D$ are labeled by fixed points of $z$ if and only if $(i,i) \notin D$. \end{proof} Recall that the set $\mathcal{FD}(z)$ of \emph{fpf-involution pipe dreams} for $z \in \I^{\fpf}_n$ consists of all intersections $D\cap \ltriang^{\!\!\neq}_n$ where $D$ is a reduced pipe dream for $z$ that is symmetric and $\ltriang^{\!\!\neq}_n = \{ (j,i) \in [n]\times [n] : i < j\}$. \begin{thm}\label{thm:fpf-almost-symmetric} Suppose $n$ is even, $z \in \I^{\fpf}_n$, and $D \subseteq [n]\times [n]$. The following are equivalent: \begin{enumerate} \item[(a)] Some reading word of $D$ is an fpf-involution word for $z$. \item[(b)] Every reading word of $D$ is an fpf-involution word for $z$. \item[(c)] The set $D$ is a reduced pipe dream for some fpf-atom of $z$. \end{enumerate} Moreover, if $D \subseteq \ltriang^{\!\!\neq}_n$ then $D\in \mathcal{FD}(z)$ if and only if these equivalent conditions hold. \end{thm} \begin{proof} Recall that $\hat{\mathcal{R}}^{\fpf}(z)$ is the disjoint union of the sets $\mathcal{R}(w)$, running over all fpf-atoms $w \in \mathcal{A}^{\fpf}(z)$. Properties (a), (b), and (c) are again equivalent by Lemma~\ref{commutation-lem} and Theorem~\ref{pipe-thm}. Assume $D \subseteq \ltriang^{\!\!\neq}_n$. To prove the final assertion, it suffices to check that $D$ is an fpf-involution pipe dream for $z$ if and only if $\textsf{udiag}(D) \in \hat{\mathcal{R}}^{\fpf}(z)$. To this end, first suppose $D = E \cap \ltriang^{\!\!\neq}_n$ where $E=E^T \in \mathcal{PD}(z)$. Then $E$ is also almost-symmetric, so Theorem~\ref{thm:almost-symmetric} implies that $E\cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n\in \mathcal{ID}(z)$. This combined with Lemma~\ref{diag-sym-lem} implies that $\textsf{udiag}(E\cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n) = 1 3 5 \cdots (n-1)\textsf{udiag}(D) \in \hat{\mathcal{R}}(z)$, so $\textsf{udiag}(D) \in \hat{\mathcal{R}}^{\fpf}(z)$. Conversely, suppose every reading word of $D$ is an fpf-involution word for $z$, so that $\textsf{udiag}(D) \in \hat{\mathcal{R}}^{\fpf}(z)$. The set $D' = D \sqcup \{(i,i) : i \in [n-1]\}$ then has $\textsf{udiag}(D') \in \hat{\mathcal{R}}(z)$, so there exists an almost-symmetric $D'' \in \mathcal{PD}(z)$ with $D'' \cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n = D'$ by Theorem~\ref{thm:almost-symmetric}. By construction $D=D'' \cap \ltriang^{\!\!\neq}_n$, and since $|D''| = \ell(z) = 2|D| + n/2$ it follows that $D''$ is actually symmetric. Therefore $D\in \mathcal{FD}(z)$. \end{proof} \begin{ex} Let $z =216543= (1,2)(3,6)(4,5) \in \I^{\fpf}_6$. Then $\ell(z) = 7$ and \[ z = s_3 \cdot s_4 \cdot (s _1 \cdot s_3 \cdot s_5) \cdot s_4\cdot s_3 =s_5 \cdot s_4 \cdot (s_1 \cdot s_3 \cdot s_5) \cdot s_4\cdot s_5, \] so $3413543$ and $5413545$ are reduced words for $z$. These words are the unimodal-diagonal reading words of the symmetric reduced pipe dreams \[ \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,3) {}; \filldraw (0.5, 5.5) circle[radius=.6mm]; \draw (0,5.5) to (1,5.5) (0.5,5) to (0.5,6); \filldraw (1.5, 5.5) circle[radius=.6mm]; \draw (1,5.5) to[bend right] (1.5,6) (1.5,5) to[bend left] (2,5.5); \filldraw (2.5, 5.5) circle[radius=.6mm]; \draw (2,5.5) to (3,5.5) (2.5,5) to (2.5,6); \filldraw (3.5, 5.5) circle[radius=.6mm]; \draw (3,5.5) to[bend right] (3.5,6) (3.5,5) to[bend left] (4,5.5); \filldraw (4.5, 5.5) circle[radius=.6mm]; \draw (4,5.5) to[bend right] (4.5,6) (4.5,5) to[bend left] (5,5.5); \filldraw (5.5, 5.5) circle[radius=.6mm]; \draw (5,5.5) to[bend right] (5.5,6); \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to[bend right] (0.5,5) (0.5,4) to[bend left] (1,4.5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to (2,4.5) (1.5,4) to (1.5,5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to (3,4.5) (2.5,4) to (2.5,5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to (3,3.5) (2.5,3) to (2.5,4); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \qquad\text{and}\qquad \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,3) {}; \filldraw (0.5, 5.5) circle[radius=.6mm]; \draw (0,5.5) to (1,5.5) (0.5,5) to (0.5,6); \filldraw (1.5, 5.5) circle[radius=.6mm]; \draw (1,5.5) to[bend right] (1.5,6) (1.5,5) to[bend left] (2,5.5); \filldraw (2.5, 5.5) circle[radius=.6mm]; \draw (2,5.5) to[bend right] (2.5,6) (2.5,5) to[bend left] (3,5.5); \filldraw (3.5, 5.5) circle[radius=.6mm]; \draw (3,5.5) to (4,5.5) (3.5,5) to (3.5,6); \filldraw (4.5, 5.5) circle[radius=.6mm]; \draw (4,5.5) to (5,5.5) (4.5,5) to (4.5,6); \filldraw (5.5, 5.5) circle[radius=.6mm]; \draw (5,5.5) to[bend right] (5.5,6); \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to[bend right] (0.5,5) (0.5,4) to[bend left] (1,4.5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to (2,4.5) (1.5,4) to (1.5,5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to (3,3.5) (2.5,3) to (2.5,4); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to (1,1.5) (0.5,1) to (0.5,2); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} \] so $\{(3,1),(3,2)\}$ and $\{(4,1),(5,1)\}$ are fpf-involution pipe dreams for $z$, and their standard reading words $43$ and $45$ are fpf-involution words for $z$. \end{ex} \section{Pipe dreams and Schubert polynomials}\label{sec:inv-polynom} In this section, we derive the pipe dream formulas for involution Schubert polynomials given in Theorem~\ref{thm:inv-pipe-dream-formula}. Our arguments are inspired by a new proof due to Knutson \cite{Knutson} of the classical pipe dream formula \eqref{eq:double-schubert-def}. Knutson's approach is inductive. The key step in his argument is to show that the right side of \eqref{eq:double-schubert-def} satisfies certain recurrences which also apply to double Schubert polynomials \cite[\S4]{kohnert1997using}. Similar recurrences for $\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$ appear in~\cite{HMP3}. Adapting Knutson's strategy to our setting requires us to show that the right hand expressions in Theorem~\ref{thm:inv-pipe-dream-formula} satisfy the same family of identities. This is accomplished in Theorems~\ref{thm:inv-dominant-transition} and \ref{thm:fpf-dominant-transition}. Proving these results involves a detailed analysis of the maximal (shifted) Ferrers diagram contained in a reduced pipe dream, which we refer to as the \emph{(shifted) dominant component}. We gradually develop the technical properties of these components over the course of this section. \subsection{Dominant components of permutations} Let $\leq_{\textsf{NW}}$ be the partial order on $\ZZ_{>0} \times \ZZ_{>0}$ with $(i,j) \leq_{\textsf{NW}} (i',j')$ if $i \leq i'$ and $j \leq j'$, i.e., if $(i,j)$ is northwest of $(i',j')$ in matrix coordinates. \begin{defn} The \emph{dominant component} $\dom{D}$ of a set $D \subseteq \ZZ_{>0}\times \ZZ_{>0}$ is the union of all lower sets in $(\ZZ_{>0}\times\ZZ_{>0},\leq_{\textsf{NW}})$ that are contained in $D$. \end{defn} Equivalently, the set $\dom{D}$ consists of all pairs $(i,j) \in D$ such that whenever $(i',j') \in \ZZ_{>0}\times \ZZ_{>0}$ and $(i',j') \leq_{\textsf{NW}} (i,j)$ it holds that $(i',j') \in D$. If $D$ is finite, then its dominant component $\dom{D}$ is the Ferrers diagram $ \textsf{D}_\lambda = \{(i,j) : 1 \leq i \leq \ell(\lambda),\ 1 \leq j \leq \lambda_i\} $ of some partition $\lambda$. An \emph{outer corner} of $D$ is a pair $(i,j) \in (\ZZ_{>0} \times \ZZ_{>0}) \setminus D$ such that $\dom{D} \sqcup \{(i,j)\}$ is again a Ferrers diagram of some partition. For example, $(1,2)$ and $(2,1)$ are the outer corners of $D = \{(1,1),(1,3)\}$, since $\dom{D} = \{(1,1)\}$. \begin{lem}\label{outer-corner-lem} Suppose $w \in S_n$ and $(i,j)$ is an outer corner of some $D \in \mathcal{PD}(w)$. Then $w(i) = j$ and $D\sqcup\{(i,j)\}$ is a reduced pipe dream (for a longer permutation). \end{lem} \begin{proof} By hypothesis, $D$ contains every cell above $(i,j)$ in the $j$th column and every cell to the left of $(i,j)$ in the $i$th row. This means that in the wiring diagram associated to $D$, the pipe leaving the top of position $(i,j)$ must continue straight up and terminate in column $j$ on the top side of $D$, and after leaving the left of position $(i,j)$, the same pipe must continue straight left and terminate in row $i$ on the left side of $D$. Thus $w(i) = j$ as claimed. Suppose the other pipe at position $(i,j)$ starts at $p$ on the left and ends at $q=w(p)$ on the top. As this pipe leaves $(i,j)$ rightwards and downwards, we have $p > i$ and $q>j$, and the pipe only intersects $[i] \times [j]$ at $(i,j)$, where it avoids the other pipe. Hence $D\sqcup\{(i,j)\} \in \mathcal{PD}(w')$ for $w' =(i,p)w = w(j,q)\in S_n$ since $\ell(w') > \ell(w)$. \end{proof} \begin{ex} Suppose $w=426135 \in S_6$. If $(i,j) = (2,2)$ and \begin{equation*}\label{wiring-ex} D = \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,3) {}; \filldraw (0.5, 5.5) circle[radius=.6mm]; \draw (0,5.5) to (1,5.5) (0.5,5) to (0.5,6); \filldraw (1.5, 5.5) circle[radius=.6mm]; \draw (1,5.5) to (2,5.5) (1.5,5) to (1.5,6); \filldraw (2.5, 5.5) circle[radius=.6mm]; \draw (2,5.5) to (3,5.5) (2.5,5) to (2.5,6); \filldraw (3.5, 5.5) circle[radius=.6mm]; \draw (3,5.5) to[bend right] (3.5,6) (3.5,5) to[bend left] (4,5.5); \filldraw (4.5, 5.5) circle[radius=.6mm]; \draw (4,5.5) to (5,5.5) (4.5,5) to (4.5,6); \filldraw (5.5, 5.5) circle[radius=.6mm]; \draw (5,5.5) to[bend right] (5.5,6); \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to (3,4.5) (2.5,4) to (2.5,5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \draw (-0.3,5.5) node {$\scriptstyle 1$}; \draw (-0.3,4.5) node {$\scriptstyle 2$}; \draw (-0.3,3.5) node {$\scriptstyle 3$}; \draw (-0.3,2.5) node {$\scriptstyle 4$}; \draw (-0.3,1.5) node {$\scriptstyle 5$}; \draw (-0.3,0.5) node {$\scriptstyle 6$}; \draw (0.5,6.3) node {$\scriptstyle 1$}; \draw (1.5,6.3) node {$\scriptstyle 2$}; \draw (2.5,6.3) node {$\scriptstyle 3$}; \draw (3.5,6.3) node {$\scriptstyle 4$}; \draw (4.5,6.3) node {$\scriptstyle 5$}; \draw (5.5,6.3) node {$\scriptstyle 6$}; \end{tikzpicture} \qquad \text{so that} \qquad D\sqcup\{(i,j)\} = \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,3) {}; \filldraw (0.5, 5.5) circle[radius=.6mm]; \draw (0,5.5) to (1,5.5) (0.5,5) to (0.5,6); \filldraw (1.5, 5.5) circle[radius=.6mm]; \draw (1,5.5) to (2,5.5) (1.5,5) to (1.5,6); \filldraw (2.5, 5.5) circle[radius=.6mm]; \draw (2,5.5) to (3,5.5) (2.5,5) to (2.5,6); \filldraw (3.5, 5.5) circle[radius=.6mm]; \draw (3,5.5) to[bend right] (3.5,6) (3.5,5) to[bend left] (4,5.5); \filldraw (4.5, 5.5) circle[radius=.6mm]; \draw (4,5.5) to (5,5.5) (4.5,5) to (4.5,6); \filldraw (5.5, 5.5) circle[radius=.6mm]; \draw (5,5.5) to[bend right] (5.5,6); \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to (2,4.5) (1.5,4) to (1.5,5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to (3,4.5) (2.5,4) to (2.5,5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \draw (-0.3,5.5) node {$\scriptstyle 1$}; \draw (-0.3,4.5) node {$\scriptstyle 2$}; \draw (-0.3,3.5) node {$\scriptstyle 3$}; \draw (-0.3,2.5) node {$\scriptstyle 4$}; \draw (-0.3,1.5) node {$\scriptstyle 5$}; \draw (-0.3,0.5) node {$\scriptstyle 6$}; \draw (0.5,6.3) node {$\scriptstyle 1$}; \draw (1.5,6.3) node {$\scriptstyle 2$}; \draw (2.5,6.3) node {$\scriptstyle 3$}; \draw (3.5,6.3) node {$\scriptstyle 4$}; \draw (4.5,6.3) node {$\scriptstyle 5$}; \draw (5.5,6.3) node {$\scriptstyle 6$}; \end{tikzpicture} \end{equation*} then in the notation of the proof, we have $p = 3$, $q=6$, and $w' = 462135$. \end{ex} Recall that $D(w) = \{(i,j) \in [n] \times [n] : w(i) > j\text{, } w^{-1}(j) > i\}$ is the Rothe diagram of $w \in S_n$. \begin{defn} The \emph{dominant component} of a permutation $w \in S_n$ is $\dom{w} = \dom{D(w)}.$ We say that permutation $w \in S_n$ is \emph{dominant} if $\dom{w} \in \mathcal{PD}(w)$. \end{defn} It is more common to define $w$ to be dominant if $D(w)$ is the Ferrers diagram of a partition, or equivalently if $w$ is $132$-avoiding. The following lemma shows that our definition is equivalent. \begin{lem}\label{dom1-lem} A permutation $w \in S_n$ is dominant if and only if it holds that $\mathcal{PD}(w) = \{ \dom{w}\}$, in which case $\dom{w} = D(w)$. \end{lem} \begin{proof} If $w \in S_n$ is dominant then $\mathcal{PD}(w) = \{ \dom{w}\}= \{ D(w) \}$ since all reduced pipe dreams for $w$ have size $\ell(w) = |D(w)|$ and contain $\dom{w} \subseteq D(w)$. \end{proof} \begin{cor}\label{sym-dom-cor} Let $w \in S_n$. Then $\dom{w^{-1}} = \dom{w}^T$. If $w $ is dominant, then $\dom{w} = \dom{w}^T$ if and only if $w=w^{-1}$. \end{cor} \begin{proof} The first claim holds since $D(w^{-1}) = D(w)^T$. If $w$ is dominant and $\dom{w} = \dom{w}^T$, then $\dom{w} = D(w)$ by Lemma~\ref{dom1-lem} so $D(w) = D(w)^T = D(w^{-1})$ and therefore $w=w^{-1}$. \end{proof} We write $\mu \subseteq \lambda$ for partitions $\mu$ and $\lambda$ to indicate that $\textsf{D}_\mu \subseteq \textsf{D}_\lambda$. \begin{prop} If $\lambda$ is a partition with $\lambda \subseteq (n-1,\dots,3, 2,1)$ then there exists a unique dominant permutation $w \in S_n$ with $\dom{w} = \textsf{D}_\lambda$. \end{prop} \begin{proof} This holds by induction as adding an outer corner to the reduced pipe dream of a dominant permutation yields a reduced pipe dream of a new dominant permutation. \end{proof} Write $\leq$ for the Bruhat order on $S_n$. Since $v \leq w$ if and only if some (equivalently, every) reduced word for $w$ has a subword that is a reduced word for $v$ \cite[\S5.10]{Humphreys}, Theorem~\ref{pipe-thm} implies: \begin{lem} If $v,w \in S_n$ then $v \leq w$ if and only if some (equivalently, every) reduced pipe dream for $w$ has a subset that is a reduced pipe dream for $v$. \end{lem} \begin{cor}\label{dom-subset-cor} Let $v,w \in S_n$ with $v$ dominant. Then $v\leq w$ if and only if $\dom{v} \subseteq D$ for some (equivalently, every) $D \in \mathcal{PD}(w)$. \end{cor} \begin{proof} This holds since a dominant permutation has only one reduced pipe dream. \end{proof} \begin{prop}\label{prop:dominant-independence} If $w \in S_n$ and $D \in \mathcal{PD}(w)$ then $\dom{D} = \dom{w}$. \end{prop} \begin{proof} For each $D \in \mathcal{PD}(w)$ there exists a dominant permutation $v \in S_n$ with $\dom{v} = \dom{D}$ and $v \leq w$, in which case $\dom{D} \subseteq \dom{E}$ for all $E \in \mathcal{PD}(w)$ by Corollary~\ref{dom-subset-cor}. This can only hold if $\dom{D} = \dom{E}$ for all $E \in \mathcal{PD}(w)$. To finish the proof, it suffices to show that $\dom{w} = \dom{D_{\text{bot}}(w)}$. It is clear by definition that $\dom{w} \subseteq \dom{D_{\text{bot}}(w)}$. Conversely, each outer corner of $\dom{w}$ has the form $(i,w(i))$ for some $i \in [n]$ but no such cell is in $\dom{D_{\text{bot}}(w)}$, so we cannot have $\dom{w} \subsetneq \dom{D_{\text{bot}}(w)}$. \end{proof} In the next sections, we define an \emph{outer corner} of $w \in S_n$ to be an outer corner of $\dom{w}$. \subsection{Involution pipe dream formulas} Recall that $\mathcal{I}_n = \{ w \in S_n : w=w^{-1}\}$ and $\raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n = \{ (j,i) \in [n]\times [n] : i\leq j\}$. \begin{defn} The \emph{shifted dominant component} of $z \in \mathcal{I}_n$ is the set $ \shdom{z} = \dom{z} \cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n.$ \end{defn} Fix $z \in \mathcal{I}_n$. By Proposition~\ref{prop:dominant-independence}, we have $\shdom{z} = \dom{D} \cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$ for all $D \in \mathcal{PD}(z)$. Recall that the \emph{shifted Ferrers diagram} of a strict partition $\lambda = (\lambda_1>\lambda_2 > \dots > \lambda_k >0)$ is the set \[ \textsf{SD}_\lambda = \{ (i, i+j-1) : 1 \leq i \leq k,\ 1\leq j \leq \lambda_i\},\] which is formed from $\textsf{D}_\lambda$ by moving the boxes in row $i$ to the right by $i-1$ columns. Since $\dom{z}$ is a Ferrers diagram, the set $\shdom{z}$ is the transpose of the shifted Ferrers diagram of some strict partition. A pair $(j,i) \in \ZZ_{>0}\times \ZZ_{>0}$ with $i\leq j$ is an outer corner of $z$ if and only if the transpose of $\shdom{z} \cup \{(j,i)\}$ is a shifted Ferrers diagram, in which case $z(j) = i$. \begin{lem}\label{double-lem} If $z \in \mathcal{I}_n$ then $\dom{z} = \shdom{z} \cup \shdom{z}^T$. \end{lem} \begin{proof} This holds since $z=z^{-1}$ implies that $\dom{z} = \dom{z}^T$. \end{proof} \begin{cor}\label{maximal-lower-cor} If $z \in \mathcal{I}_n$ then $\shdom{z}$ is the union of all lower sets of $(\raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n,\leq_{\textsf{NW}})$ that are contained in some (equivalently, every) $D \in \mathcal{ID}(z)$. \end{cor} \begin{proof} This is clear from Proposition~\ref{prop:dominant-independence} and Lemma~\ref{double-lem}. \end{proof} The natural definition of ``involution'' dominance turns out to be equivalent to the usual notion: \begin{prop}\label{inv-dom-prop} Let $z \in \mathcal{I}_n$. The following are equivalent: \begin{itemize} \item[(a)] The permutation $z$ is dominant. \item[(b)] It holds that $\mathcal{PD}(z) = \{ \dom{z}\}$. \item[(c)] It holds that $\mathcal{ID}(z) = \{\shdom{z}\}$. \end{itemize} \end{prop} \begin{proof} We have (a) $\Leftrightarrow$ (b) by Lemma~\ref{dom1-lem} and (b) $\Leftrightarrow$ (c) by Lemma~\ref{double-lem}. \end{proof} \begin{prop} If $\lambda$ is a strict partition with $\lambda \subseteq (n-1,n-3,n-5,\dots)$ then there exists a unique dominant involution $z \in \mathcal{I}_n$ with $\shdom{z}^T = \textsf{SD}_\lambda$. \end{prop} \begin{proof} We have $\textsf{D}_\mu = \textsf{SD}_\lambda \cup ( \textsf{SD}_\lambda)^T$ for some $\mu$. Let $z \in S_n$ be dominant with $\dom{z} = \textsf{D}_\mu$. Then $z \in \mathcal{I}_n$ by Corollary~\ref{sym-dom-cor} and $\shdom{z}^T = (\textsf{D}_\mu \cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n)^T = \textsf{SD}_\lambda$. Uniqueness holds by Lemma~\ref{double-lem}. \end{proof} If $y,z \in \mathcal{I}_n$, then $y \leq z$ in Bruhat order if and only if some (equivalently, every) involution word for $z$ contains a subword that is an involution word for $y$ (see either \cite[Cor. 8.10]{richardson-springer} with \cite{richardson-springer2}, or \cite[Thm. 2.8]{hultman-twisted-involutions2}). The following is an immediate corollary of this property and Theorem~\ref{thm:almost-symmetric}. \begin{lem}\label{inv-bruhat-lem} Let $y,z \in \mathcal{I}_n$. Then $y \leq z$ if and only if some (equivalently, every) involution pipe dream for $z$ has a subset that is an involution pipe dream for $y$. \end{lem} \begin{cor}\label{inv-dom-subset-cor} Let $y,z \in \mathcal{I}_n$ with $y$ dominant. Then $y\leq z$ if and only if $\shdom{y} \subseteq D$ for some (equivalently, every) $D \in \mathcal{ID}(z)$. \end{cor} \begin{proof} This is clear since if $y \in \mathcal{I}_n$ is dominant then $|\mathcal{ID}(y)|=1$. \end{proof} We need to mention the following technical property of the Demazure product from \cite{KM03}. \begin{lem}[{\cite[Lem. 3.4(1)]{KM03}}] \label{0-hecke-lem} If $b_1b_2\cdots b_q$ is a subword of $a_1a_2\dots a_p$ where each $a_i \in [n-1]$, then $ s_{b_1} \circ s_{b_2}\circ \dots \circ s_{b_q} \leq s_{a_1} \circ s_{a_2}\circ \dots \circ s_{a_p} \in S_n. $ \end{lem} \begin{cor}\label{0-hecke-cor-0} If $v',w',v,w \in S_n$ and $v \leq w$ and $v'\leq w'$ then $v' \circ v \leq w' \circ w$. \end{cor} \begin{proof} This is clear from Lemma~\ref{0-hecke-lem} given the subword property of $\leq$. \end{proof} \begin{cor}\label{0-hecke-cor} If $v,w \in S_n$ and $v \leq w$ then $v^{-1}\circ v \leq w^{-1} \circ w$. \end{cor} \begin{proof} Apply Corollary~\ref{0-hecke-cor-0} with $v' = v^{-1}$ and $w'=w^{-1}$. \end{proof} Let $t_{ij} = (i,j) \in S_n$ for distinct $i,j \in [n]$. It is well-known and not hard to check that if $w \in S_n$ then $\ell(w t_{ij}) = \ell(w) +1$ if and only if $w(i)<w(j)$ and no $i<e<j$ has $w(i) < w(e) < w(j)$. Given $y \in \mathcal{I}_n$ and $1\leq i <j \leq n$, let $\mathcal{A}_{ij}(y)= \{ w t_{ij} : w \in \mathcal{A}(y)\text{, }\ell(wt_{ij})=\ell(w)+1\}.$ Each covering relation in $(S_n,\leq)$ arises as the image of right multiplication by some transposition $t_{ij}$. The following theorem characterizes certain operators $\tau_{ij}$ which play an analogous role for $(\mathcal{I}_n,\leq)$. \begin{thm}[{See \cite[\S3]{HMP3}}] \label{tau-thm} For each pair of integers $1\leq i < j \leq n$, there are unique maps $\tau_{ij} : \mathcal{I}_n \to \mathcal{I}_n$ with the following properties: \begin{enumerate} \item[(a)] If $y \in \mathcal{I}_n$ and $\mathcal{A}_{ij}(y)\cap \mathcal{A}(z)\neq\varnothing $ for some $z \in \mathcal{I}_n$ then $\tau_{ij}(y)= z$. \item[(b)] If $y \in \mathcal{I}_n$ and $\mathcal{A}_{ij}(y)\cap \mathcal{A}(z)=\varnothing$ for all $z \in \mathcal{I}_n$ then $\tau_{ij}(y)= y$. \end{enumerate} Moreover, if $y \in \mathcal{I}_n$ and $y \neq \tau_{ij}(y)=z$, then $y(i) \neq z(i)$ and $y(j)\neq z(j)$. \end{thm} This result has an extension for affine symmetric groups; see \cite{M2018,MZ2018}. \begin{rmk} The operators $\tau_{ij}$, which first appeared in \cite{incitti}, can be given a more explicit definition; see \cite[Table 1]{HMP3}. However, our present applications only require the properties in the theorem. \end{rmk} For $y \in \mathcal{I}_n$, let $\iell(y)$ denote the common value of $\ell(w)$ for any $w \in \mathcal{A}(y)$. This is also the size of any $D \in \mathcal{ID}(y)$, so $\iell(y) = |\hat c(y)| $. By Lemma~\ref{inv-bruhat-lem}, if $y,z \in \mathcal{I}_n$ and $y<z$ then $\iell(y) < \iell(z)$. Let \begin{equation* \Psi(y,j) = \left\{ z \in \mathcal{I}_{n+1} : z =\tau_{js}(y)\text{ and }\iell(z)=\iell(y)+1\text{ for some }s>j \right\} \end{equation*} for $y \in \mathcal{I}_n$ and $j \in [n]$. Since $S_n \subset S_{n+1}$ and $\mathcal{I}_n \subset \mathcal{I}_{n+1}$, this set is well-defined. \begin{thm}\label{thm:inv-dominant-transition} Let $(j,i) $ be an outer corner of $y \in \mathcal{I}_n$ with $i\leq j$. \begin{enumerate} \item[(a)] The map $D \mapsto D \sqcup \{(j,i)\}$ is a bijection $ \mathcal{ID}(y) \to \bigsqcup_{z \in \Psi(y,j)} \mathcal{ID}(z) $. \item[(b)] If $i+j \leq n$ then $\Psi(y,j) \subset \mathcal{I}_n$. \end{enumerate} \end{thm} \begin{proof} We have $y(j) = i$ and $y(i) = j$ by Lemma~\ref{outer-corner-lem}. Suppose $v \in \mathcal{A}(y)$ and $D \in \mathcal{PD}(v) \cap \mathcal{ID}(y)$. By considering the pipes crossing at position $(j,i)$ in the wiring diagram of $D$, as in the proof of Lemma~\ref{outer-corner-lem}, it follows that $D\sqcup \{(j,i)\}$ is a reduced pipe dream for a permutation $w$ that belongs to $\mathcal{A}_{js}(y)$ for some $j<s \leq n$. Set $z = w^{-1}\circ w \in \mathcal{I}_n$. We wish to show that $w \in \mathcal{A}(z)$, since if this holds then $D\sqcup \{(j,i)\} \in \mathcal{ID}(z)$ and Theorem~\ref{tau-thm} implies that $z \in \Psi(y,j)$. To this end, let $\tilde y \in \mathcal{I}_n$ be the dominant involution whose unique involution pipe dream is $\shdom{y} \sqcup\{(j,i)\}$ and let $\tilde v \in \mathcal{A}(\tilde y)$ be the (unique) atom with $\mathcal{ID}(\tilde y) \subseteq \mathcal{PD}(\tilde v) $. Corollaries~\ref{dom-subset-cor} and \ref{inv-dom-subset-cor} imply that $\tilde y \not<y$, $\tilde v \leq w $, and $v<w$ since $\shdom{\tilde y} \not\subset \shdom{y}$ and $\shdom{\tilde y} \subseteq D\sqcup \{(j,i)\}$. Hence, we have $ \tilde y = {\tilde v}^{-1} \circ \tilde v \leq w^{-1} \circ w=z $ and $ y = v^{-1} \circ v \leq w^{-1} \circ w=z $ by Corollary~\ref{0-hecke-cor}. Putting these relations together gives $\tilde y \not < y \leq z$ and $\tilde y \leq z$, so we must have $y < z$ and $\ell(w) = \iell(y)+1 \leq \iell(z)$, and therefore $w \in \mathcal{A}(z)$. Thus, the map in part (a) at least has the desired codomain and is clearly injective. To show that it is also surjective, suppose $E \in \mathcal{ID}(z)$ for some $z \in \Psi(y,j)$. Lemma~\ref{inv-bruhat-lem} implies some $(l,k) \in E$ has $E \setminus \{(l,k)\} \in \mathcal{ID}(y)$. Let $E' \in \mathcal{PD}(z)$ be the almost-symmetric reduced pipe dream with $E = E'\cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$. If $(j,i) \neq (l,k)$ then, since $\dom{y} = \shdom{y} \cup \shdom{y}^T \subset E'$, it would follow by considering the wiring diagram of $E'$ that $z(j) = i = y(j)$, contradicting the last assertion in Theorem~\ref{tau-thm}. Thus $(j,i) = (l,k)$ so the map in part (a) is surjective. Part (b) holds because an involution belongs to $\mathcal{I}_n$ if any of its involution pipe dreams is contained in $\{ (j,i) : i\leq j\text{ and }i +j \leq n\}$. \end{proof} \begin{ex}\label{inv-dom-ex} If $y = 35142 = (1,3)(2,5) \in \mathcal{I}_5$ then \[ \mathcal{ID}(y) = \left\{ \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,2.5) {}; \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to (1,1.5) (0.5,1) to (0.5,2); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} ,\quad \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,2.5) {}; \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to (2,2.5) (1.5,2) to (1.5,3); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture}\ \right\} \] so the transpose of $\shdom{y}$ is the shifted Ferrers diagram of $(2,1)$, and $(3,1)$ is an outer corner. One can show that $\Psi(y,3) = \{53241,45312\}$, and as predicted by Theorem~\ref{thm:inv-dominant-transition} with $(j,i) = (3,1)$, both elements of $\Psi(y,3)$ are dominant with \[ \mathcal{ID}(53241) = \left\{ \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,2.5) {}; \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to (1,1.5) (0.5,1) to (0.5,2); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture}\ \right\} \quad\text{and}\quad \mathcal{ID}(45312) = \left\{ \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,2.5) {}; \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to (2,2.5) (1.5,2) to (1.5,3); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture}\ \right\}. \] \end{ex} We may finally prove the pipe dream formula in Theorem~\ref{thm:inv-pipe-dream-formula} for the polynomials $\hat{\mathfrak{S}}_y$. \begin{thm}\label{inv-pipe-schubert-thm} If $z \in \mathcal{I}_n$ then $\hat{\mathfrak{S}}_z = \sum_{D \in \mathcal{ID}(z)} \prod_{(i,j) \in D} 2^{-\delta_{ij}}(x_i+x_j)$. \end{thm} \begin{proof} We abbreviate by setting $x_{(i,j)} = 2^{-\delta_{ij}}(x_i+x_j)$, so that $x_{(i,j)} = x_i$ if $i=j$ and otherwise $x_{(i,j)} = x_i+x_j$. It follows from \cite[Thm. 3.30]{HMP3} that if $(j,i) \in \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$ is an outer corner of $z \in \mathcal{I}_n$ then \begin{equation}\label{itransition-eq} 2^{-\delta_{ij}}(x_i+x_j) \hat{\mathfrak{S}}_z = \sum_{u \in \Psi(z,j)} \hat{\mathfrak{S}}_u.\end{equation} On the other hand, results of Wyser and Yong \cite{wyser-yong-orthogonal-symplectic} (see \cite[Thm. 1.3]{HMP1}) show that \begin{equation}\label{inv-rev-eq} \hat{\mathfrak{S}}_{n\cdots 321} = \prod_{1\leq i \leq j \leq n-i} x_{(i,j)}.\end{equation} Let $\mathfrak{A}_z = \sum_{D \in \mathcal{ID}(z)} \prod_{(i,j) \in D} x_{(i,j)}$. We show that $\hat{\mathfrak{S}}_z = \mathfrak{A}_z$ by downward induction on $\iell(z)$. If $\iell(z) = \max\{\iell(y): y \in \mathcal{I}_n\}$ then $z=n\cdots 321$ and the desired identity is equivalent to \eqref{inv-rev-eq} since $n\cdots 321$ is dominant. Otherwise, the transpose of $\shdom{z}$ is a proper subset of $\shdom{n\cdots 321}^T=\textsf{SD}_{(n-1,n-3,n-5,\dots)}$ by Corollary~\ref{inv-dom-subset-cor}, so $z$ must have an outer corner $(j,i)$ with $i\leq j$ and $i+j \leq n$. In this case we have $x_{(i,j)} \hat{\mathfrak{S}}_z = \sum_{u \in \Psi(z,j)} \hat{\mathfrak{S}}_u= \sum_{u \in \Psi(z,j)} \mathfrak{A}_u = x_{(i,j)} \mathfrak{A}_z$ by \eqref{itransition-eq}, induction, and Theorem~\ref{thm:inv-dominant-transition}. Dividing by $x_{(i,j)}$ completes the proof. \end{proof} \begin{ex} Continuing Example~\ref{inv-dom-ex}, we have \[ \hat{\mathfrak{S}}_{53241} = x_1x_2(x_2 + x_1)(x_3 + x_1)(x_4 + x_1) \quad\text{and}\quad \hat{\mathfrak{S}}_{45312} = x_1x_2(x_2 + x_1)(x_3 + x_1)(x_3 + x_2), \] so $ \hat{\mathfrak{S}}_{35142} = \tfrac{1}{x_3 + x_1}\left(\hat{\mathfrak{S}}_{53241} + \hat{\mathfrak{S}}_{45312}\right) =x_1x_2(x_2 + x_1)(x_1 + x_2+x_3 + x_4) . $ \end{ex} \subsection{Fixed-point-free involution pipe dream formulas} In this section, we assume $n$ is even. Recall that $\ltriang^{\!\!\neq}_n = \{(j,i) \in [n] \times [n] : i<j\}$. \begin{defn} The \emph{strictly shifted dominant component} of $z \in \I^{\fpf}_n$ is the set $ \shdomneq{z} = \dom{z} \cap \ltriang^{\!\!\neq}_n, $ which is also equal to $ \dom{D} \cap \ltriang^{\!\!\neq}_n$ for all $D \in \mathcal{PD}(z)$ by Proposition~\ref{prop:dominant-independence}. \end{defn} \begin{cor}\label{fpf-maximal-lower-cor} If $z \in \I^{\fpf}_n$ then $\shdomneq{z}$ is the maximal lower set of the poset $(\ltriang^{\!\!\neq}_n,\leq_{\textsf{NW}})$ contained in some (equivalently, every) $D \in \mathcal{FD}(z)$. \end{cor} \begin{proof} This is clear from Proposition~\ref{prop:dominant-independence}. \end{proof} For subsets $D \subseteq \mathbb{Z}\times \mathbb{Z}$, define $D^\uparrow = \{ (i-1,j): (i,j) \in D\}$ and $D^{\uparrow T} = (D^{\uparrow })^T.$ For example, if we have $z = (1,4)(2,6)(3,5) =465132 \in \I^{\fpf}_6$ then the Rothe diagram is \[ D(z)= \left\{\begin{smallmatrix} + & + & + & 1 & \cdot & \cdot \\ + & + & + & \cdot & + & 1 \\ + & + & + & \cdot & 1 & \cdot \\ 1 & \cdot & \cdot & \cdot & \cdot & \cdot \\ \cdot & + & 1 & \cdot & \cdot & \cdot \\ \cdot & 1 & \cdot & \cdot & \cdot & \cdot \end{smallmatrix}\right\}, \] with each $1$ indicating a position $(i,j) \in [n]\times [n]$ with $z(i) =j$ and each $+$ indicating a position in $ D(z)$. The relevant dominant components are \[ \dom{z} = \left\{\begin{smallmatrix} + & + & + \\ + & + & + \\ + & + & + \end{smallmatrix}\right\}, \quad \shdomneq{z} = \left\{\begin{smallmatrix} \cdot & \cdot & \cdot \\ + & \cdot & \cdot \\ + & + & \cdot \\ \end{smallmatrix}\right\}, \quad\text{and}\quad \shdomneq{z}^{\uparrow T} = \left\{\begin{smallmatrix} + & + & \cdot\\ \cdot & + & \cdot \\ \cdot & \cdot & \cdot \end{smallmatrix}\right\}. \] As we see in this example, if $z \in \I^{\fpf}_n$ is any fixed-point-free involution, then $(\shdomneq{z})^{\uparrow T} $ is the shifted Ferrers diagram of some strict partition. Moreover, a pair $(j,i) \in \ltriang^{\!\!\neq}_n$ is an outer corner of $z$ if and only if $(\shdomneq{D} \sqcup \{(j,i)\})^{\uparrow T} $ is again a shifted Ferrers diagram, in which case $z(j) = i$ by Lemma~\ref{outer-corner-lem}. The unique outer corner of $z = (1,4)(2,6)(3,5)$ in $\ltriang^{\!\!\neq}_6$ is $(4,1)$. \begin{defn} A fixed-point-free involution $z \in \I^{\fpf}_n$ is \emph{fpf-dominant} if $\shdomneq{z} \in \mathcal{FD}(z)$. \end{defn} This condition does not imply that $z$ is dominant in the sense of being $132$-avoiding. For example, $z=s_1s_3\cdots s_{n-1} = 2143\cdots n(n-1) \in \I^{\fpf}_n$ is always fpf-dominant as $\shdomneq{z} = \varnothing$. \begin{lem}\label{dom-fpf-lem} If $z \in \I^{\fpf}_n$ is fpf-dominant then $\mathcal{FD}(z) = \{ \shdomneq{z}\}.$ \end{lem} \begin{proof} This holds since $\shdomneq{z} \subseteq D$ for all $D \in \mathcal{FD}(z)$ by Proposition~\ref{prop:dominant-independence}. \end{proof} \begin{prop} If $\lambda$ is a strict partition with $\lambda \subseteq (n-2,n-4,\dots,4,2)$ then there exists a unique fpf-dominant $z \in \I^{\fpf}_n$ with $(\shdomneq{z})^{\uparrow T} = \textsf{SD}_\lambda$. \end{prop} \begin{proof} Uniqueness is clear from Lemma~\ref{dom-fpf-lem}. If $\lambda \subseteq (n-2,n-4,\dots,2)$ is empty then take $z=1^{\fpf}_n$. Otherwise, let $\mu\subset \lambda $ be a strict partition such that $ \textsf{SD}_\lambda = \textsf{SD}_\mu\sqcup \{(i,j-1)\}$ where $i<j$. By induction, there exists an fpf-dominant $y \in \I^{\fpf}_n$ with $(\shdomneq{y})^{\uparrow T} = \textsf{SD}_\mu$. Let $D \in \mathcal{PD}(y)$ be symmetric with $D \cap \ltriang^{\!\!\neq}_n = \shdomneq{y}$. Lemmas~\ref{diag-sym-lem} and \ref{outer-corner-lem} imply that $ D \sqcup \{(j,i),(i,j)\}$ is a symmetric reduced pipe dream for some $z \in \I^{\fpf}_n$, which is the desired element. \end{proof} If $y,z \in \I^{\fpf}_n$, then $y \leq z$ in Bruhat order if and only if some (equivalently, every) fpf-involution word for $z$ contains a subword that is an fpf-involution word for $y$ \cite[Thm. 4.6]{HMP3}. From this and Theorem~\ref{thm:fpf-almost-symmetric} we deduce the following: \begin{lem}\label{fpf-bruhat-lem} Suppose $y,z \in \I^{\fpf}_n$. Then $y \leq z$ if and only if some (equivalently, every) fpf-involution pipe dream for $z$ has a subset that is an fpf-involution pipe dream for $y$. \end{lem} \begin{cor}\label{fpf-dom-subset-cor} Let $y,z \in \I^{\fpf}_n$ where $y$ is fpf-dominant. Then $y\leq z$ if and only if $\shdomneq{y} \subseteq D$ for some (equivalently, every) $D \in \mathcal{FD}(z)$. \end{cor} \begin{proof} This is clear since if $y \in \I^{\fpf}_n$ is fpf-dominant then $|\mathcal{FD}(y)|=1$. \end{proof} For $y \in \I^{\fpf}_n$ and $j \in [n]$, define $ \Psi^{\fpf}(y,j)$ to be the set of fixed-point-free involutions $z \in \I^{\fpf}_{n+2}$ with length $\ell(z) = \ell(y)+2$ that can be written as $z = t_{js}\cdot ys_{n+1} \cdot t_{js}$ for an integer $s$ with $j < s \leq n+2$. We have an analogue of Theorem~\ref{thm:inv-dominant-transition}: \begin{thm}\label{thm:fpf-dominant-transition} Let $(j,i)$ be an outer corner of $y \in \I^{\fpf}_n$ with $i<j$. \begin{enumerate} \item[(a)] The map $D \mapsto D \sqcup \{(j,i)\}$ is a bijection $ \mathcal{FD}(y) \to \bigsqcup_{z \in \Psi^{\fpf}(y,j)} \mathcal{FD}(z). $ \item[(b)] If $i+j \leq n$ then $ \Psi^{\fpf}(y,j)\subset \left\{ z s_{n+1} : z \in \I^{\fpf}_{n}\right\}$. \end{enumerate} \end{thm} \begin{proof} Our argument is similar to the proof of Theorem~\ref{thm:inv-dominant-transition}. Choose $D \in \mathcal{FD}(y) = \mathcal{FD}(ys_{n+1})$. Suppose $D' \in \mathcal{PD}(ys_{n+1})$ is the symmetric reduced pipe dream with $D = D' \cap \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_{n+2}$ and set $E = D\sqcup \{(j,i)\}$ and $E' = D' \sqcup \{(i,j),(j,i)\}$. It follows Lemmas~\ref{diag-sym-lem} and \ref{outer-corner-lem} that $E'$ is a reduced pipe dream for some element $z \in \I^{\fpf}_{n+2}$. Since $E'$ is symmetric, one has $E = E' \cap \ltriang^{\!\!\neq}_n \in \mathcal{FD}(z)$. Finally, by considering the pipes crossing at position $(j,i)$ in $E$ we deduce that $z \in \Psi^{\fpf}(y,j)$. Thus $D \mapsto D \sqcup \{(j,i)\}$ is a well-defined map $\mathcal{FD}(y) \to \bigsqcup_{z\in \Psi^{\fpf}(y,j)} \mathcal{FD}(z)$. This map is clearly injective. To show that it is also surjective, suppose $E \in \mathcal{FD}(z)$ for some $z \in \Psi^{\fpf}(y,j)$. Lemma~\ref{fpf-bruhat-lem} implies that there exists a position $(l,k) \in E$ such that $E \setminus \{(l,k)\} \in \mathcal{FD}(y)$. If $(j,i) \neq (l,k)$ then it would follow as in the proof Theorem~\ref{thm:inv-dominant-transition} that $z(j) = y(j)=ys_{n+1}(j)=i$, which is impossible if $z = t_{js}\cdot ys_{n+1}\cdot t_{js}$ where $i =y(j) < j < s \leq n + 2$. Thus $(j,i) = (l,k)$ so the map in part (a) is also surjective. Part (b) holds because $\left\{ zs_{n+1} : z \in \I^{\fpf}_{n}\right\}$ contains all involutions in $\I^{\fpf}_{n+2}$ with fpf-involution pipe dreams that are subsets of $\{ (j,i) : i\leq j, i +j \leq n\}$. \end{proof} \begin{ex}\label{fpf-dom-ex} If $y = 351624 = (1,3)(2,5) \in \I^{\fpf}_6$ then \[ \mathcal{FD}(y) = \left\{ \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,3) {}; \filldraw (0.5, 5.5) circle[radius=.6mm]; \draw (0,5.5) to[bend right] (0.5,6) (0.5,5) to[bend left] (1,5.5); \filldraw (1.5, 5.5) circle[radius=.6mm]; \draw (1,5.5) to[bend right] (1.5,6) (1.5,5) to[bend left] (2,5.5); \filldraw (2.5, 5.5) circle[radius=.6mm]; \draw (2,5.5) to[bend right] (2.5,6) (2.5,5) to[bend left] (3,5.5); \filldraw (3.5, 5.5) circle[radius=.6mm]; \draw (3,5.5) to[bend right] (3.5,6) (3.5,5) to[bend left] (4,5.5); \filldraw (4.5, 5.5) circle[radius=.6mm]; \draw (4,5.5) to[bend right] (4.5,6) (4.5,5) to[bend left] (5,5.5); \filldraw (5.5, 5.5) circle[radius=.6mm]; \draw (5,5.5) to[bend right] (5.5,6); \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture} ,\quad \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,3) {}; \filldraw (0.5, 5.5) circle[radius=.6mm]; \draw (0,5.5) to[bend right] (0.5,6) (0.5,5) to[bend left] (1,5.5); \filldraw (1.5, 5.5) circle[radius=.6mm]; \draw (1,5.5) to[bend right] (1.5,6) (1.5,5) to[bend left] (2,5.5); \filldraw (2.5, 5.5) circle[radius=.6mm]; \draw (2,5.5) to[bend right] (2.5,6) (2.5,5) to[bend left] (3,5.5); \filldraw (3.5, 5.5) circle[radius=.6mm]; \draw (3,5.5) to[bend right] (3.5,6) (3.5,5) to[bend left] (4,5.5); \filldraw (4.5, 5.5) circle[radius=.6mm]; \draw (4,5.5) to[bend right] (4.5,6) (4.5,5) to[bend left] (5,5.5); \filldraw (5.5, 5.5) circle[radius=.6mm]; \draw (5,5.5) to[bend right] (5.5,6); \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to[bend right] (0.5,4) (0.5,3) to[bend left] (1,3.5); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture}\ \right\} \] so $\shdom{y}=\{(2,1)\}$, and $(3,1)$ is an outer corner. In this case, $ \Psi^{\fpf}(y,3) = \{532614,456123\}$. As predicted by the theorem with $(j,i) = (3,1)$, both elements of $ \Psi^{\fpf}(y,3)$ are fpf-dominant since \[ \mathcal{FD}(532614) = \left\{ \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,3) {}; \filldraw (0.5, 5.5) circle[radius=.6mm]; \draw (0,5.5) to[bend right] (0.5,6) (0.5,5) to[bend left] (1,5.5); \filldraw (1.5, 5.5) circle[radius=.6mm]; \draw (1,5.5) to[bend right] (1.5,6) (1.5,5) to[bend left] (2,5.5); \filldraw (2.5, 5.5) circle[radius=.6mm]; \draw (2,5.5) to[bend right] (2.5,6) (2.5,5) to[bend left] (3,5.5); \filldraw (3.5, 5.5) circle[radius=.6mm]; \draw (3,5.5) to[bend right] (3.5,6) (3.5,5) to[bend left] (4,5.5); \filldraw (4.5, 5.5) circle[radius=.6mm]; \draw (4,5.5) to[bend right] (4.5,6) (4.5,5) to[bend left] (5,5.5); \filldraw (5.5, 5.5) circle[radius=.6mm]; \draw (5,5.5) to[bend right] (5.5,6); \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to[bend right] (1.5,4) (1.5,3) to[bend left] (2,3.5); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to (1,2.5) (0.5,2) to (0.5,3); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture}\ \right\} \quad\text{and}\quad \mathcal{FD}(456123) = \left\{ \begin{tikzpicture}[scale=0.3,baseline=(b.base)] \node (b) at (0,3) {}; \filldraw (0.5, 5.5) circle[radius=.6mm]; \draw (0,5.5) to[bend right] (0.5,6) (0.5,5) to[bend left] (1,5.5); \filldraw (1.5, 5.5) circle[radius=.6mm]; \draw (1,5.5) to[bend right] (1.5,6) (1.5,5) to[bend left] (2,5.5); \filldraw (2.5, 5.5) circle[radius=.6mm]; \draw (2,5.5) to[bend right] (2.5,6) (2.5,5) to[bend left] (3,5.5); \filldraw (3.5, 5.5) circle[radius=.6mm]; \draw (3,5.5) to[bend right] (3.5,6) (3.5,5) to[bend left] (4,5.5); \filldraw (4.5, 5.5) circle[radius=.6mm]; \draw (4,5.5) to[bend right] (4.5,6) (4.5,5) to[bend left] (5,5.5); \filldraw (5.5, 5.5) circle[radius=.6mm]; \draw (5,5.5) to[bend right] (5.5,6); \filldraw (0.5, 4.5) circle[radius=.6mm]; \draw (0,4.5) to (1,4.5) (0.5,4) to (0.5,5); \filldraw (1.5, 4.5) circle[radius=.6mm]; \draw (1,4.5) to[bend right] (1.5,5) (1.5,4) to[bend left] (2,4.5); \filldraw (2.5, 4.5) circle[radius=.6mm]; \draw (2,4.5) to[bend right] (2.5,5) (2.5,4) to[bend left] (3,4.5); \filldraw (3.5, 4.5) circle[radius=.6mm]; \draw (3,4.5) to[bend right] (3.5,5) (3.5,4) to[bend left] (4,4.5); \filldraw (4.5, 4.5) circle[radius=.6mm]; \draw (4,4.5) to[bend right] (4.5,5); \filldraw (0.5, 3.5) circle[radius=.6mm]; \draw (0,3.5) to (1,3.5) (0.5,3) to (0.5,4); \filldraw (1.5, 3.5) circle[radius=.6mm]; \draw (1,3.5) to (2,3.5) (1.5,3) to (1.5,4); \filldraw (2.5, 3.5) circle[radius=.6mm]; \draw (2,3.5) to[bend right] (2.5,4) (2.5,3) to[bend left] (3,3.5); \filldraw (3.5, 3.5) circle[radius=.6mm]; \draw (3,3.5) to[bend right] (3.5,4); \filldraw (0.5, 2.5) circle[radius=.6mm]; \draw (0,2.5) to[bend right] (0.5,3) (0.5,2) to[bend left] (1,2.5); \filldraw (1.5, 2.5) circle[radius=.6mm]; \draw (1,2.5) to[bend right] (1.5,3) (1.5,2) to[bend left] (2,2.5); \filldraw (2.5, 2.5) circle[radius=.6mm]; \draw (2,2.5) to[bend right] (2.5,3); \filldraw (0.5, 1.5) circle[radius=.6mm]; \draw (0,1.5) to[bend right] (0.5,2) (0.5,1) to[bend left] (1,1.5); \filldraw (1.5, 1.5) circle[radius=.6mm]; \draw (1,1.5) to[bend right] (1.5,2); \filldraw (0.5, 0.5) circle[radius=.6mm]; \draw (0,0.5) to[bend right] (0.5,1); \end{tikzpicture}\ \right\}. \] \end{ex} We may now prove the second half of Theorem~\ref{thm:inv-pipe-dream-formula}, concerning the polynomials $\hat{\mathfrak{S}}^{\fpf}_z$. \begin{thm}\label{fpf-pipe-schubert-thm} If $z \in \I^{\fpf}_n$ then $\hat{\mathfrak{S}}^{\fpf}_z = \sum_{D \in \mathcal{FD}(z)} \prod_{(i,j) \in D} (x_i + x_j)$. \end{thm} \begin{proof} It follows from \cite[Thm. 4.17]{HMP3} that if $(j,i) \in \ltriang^{\!\!\neq}_n$ is an outer corner of $z \in \I^{\fpf}_n$ then \begin{equation}\label{ftransition-eq} (x_i+x_j) \hat{\mathfrak{S}}^{\fpf}_z = \sum_{u \in \Psi^{\fpf}(z,j)} \hat{\mathfrak{S}}^{\fpf}_u.\end{equation} Moreover, as $n$ is even, \cite[Thm. 1.3]{HMP1} implies that we have \begin{equation}\label{fpf-rev-eq} \hat{\mathfrak{S}}^{\fpf}_{n\cdots 321} = \prod_{1\leq i < j \leq n-i} (x_i + x_j).\end{equation} Let $\mathfrak{B}_z = \sum_{D \in \mathcal{FD}(z)} \prod_{(i,j) \in D} (x_i+x_j)$. If $\ell(z) = \max\left\{\ell(y): y \in \I^{\fpf}_n\right\}$ then $z=n\cdots 321$ and the identity $\hat{\mathfrak{S}}^{\fpf}_z = \mathfrak{B}_z$ is equivalent to \eqref{fpf-rev-eq}. Otherwise, the transpose of $\shdomneq{z}$ shifted up one row is a proper subset of $\shdomneq{n\cdots 321}^{\uparrow T}=\textsf{SD}_{(n-2,n-4,\dots,4,2)}$ by Corollary~\ref{fpf-dom-subset-cor}, so $z$ has an outer corner $(j,i) \in \ltriang^{\!\!\neq}_n$ with $i+j \leq n$. In this case, $(x_i+x_j) \hat{\mathfrak{S}}^{\fpf}_z = \sum_{u \in \Psi^{\fpf}(z,j)} \hat{\mathfrak{S}}^{\fpf}_u= \sum_{u \in \Psi^{\fpf}(z,j)} \mathfrak{B}_u = (x_i+x_j) \mathfrak{B}_z$ by \eqref{ftransition-eq}, induction, and Theorem~\ref{thm:fpf-dominant-transition}. Dividing by $x_i+x_j$ completes the proof. \end{proof} \begin{ex} Continuing Example~\ref{fpf-dom-ex}, we have \[ \hat{\mathfrak{S}}^{\fpf}_{532614} =(x_2 + x_1)(x_3+ x_1)(x_4+x_1) \quad\text{and}\quad \hat{\mathfrak{S}}^{\fpf}_{456123} =(x_2 + x_1)(x_3 + x_1)(x_3 + x_2), \] so $\hat{\mathfrak{S}}^{\fpf}_{351624} = \frac{1}{x_3 + x_1}\left( \hat{\mathfrak{S}}^{\fpf}_{532614}+\hat{\mathfrak{S}}^{\fpf}_{456123}\right) = (x_2 + x_1)(x_1 + x_2 + x_3 + x_4)$. \end{ex} \section{Generating pipe dreams} \label{sec:generating-pipe-dreams} Bergeron and Billey \cite{bergeron-billey} proved that the set $\mathcal{PD}(w)$ is generated by applying simple transformations to a unique ``bottom'' pipe dream. Here, we derive versions of this result for the sets of involution pipe dreams $\mathcal{ID}(y)$ and $\mathcal{FD}(z)$. This leads to algorithms for computing the sets $\mathcal{ID}(y)$ and $\mathcal{FD}(z)$ that are much more efficient than the naive methods suggested by our original definitions. \subsection{Ladder moves} Let $D$ and $E$ be subsets of $\ZZ_{>0}\times \ZZ_{>0}$, depicted as positions marked by ``$+$'' in a matrix. If $E$ is obtained from $D$ by replacing a subset of the form \[ \arraycolsep=1.5pt \begin{array}{cc} \cdot & \cdot \\ + & + \\ \vdots & \vdots \\ + & + \\ + & \cdot \end{array} \qquad\text{by}\qquad \arraycolsep=1.5pt \begin{array}{cc} \cdot & + \\ + & + \\ \vdots & \vdots \\ + & + \\ \cdot & \cdot \end{array} \] then we say that $E$ is obtained from $D$ by a \emph{ladder move} and write $D \lessdot_\mathcal{PD} E$. More formally: \begin{defn} We write $D \lessdot_\mathcal{PD} E$ if for some integers $i< j$ and $k$ the following holds: \begin{itemize} \item One has $\{i+1,i+2,\dots,j-1\}\times\{k,k+1\} \subset D$. \item It holds that $(j,k) \in D$ but $(i,k),(i,k+1),(j,k+1)\notin D$. \item One has $E = D\setminus \{(j,k)\} \cup \{(i,k+1)\}$. \end{itemize} One can have $i+1=j$ in this definition, in which case the first condition holds vacuously. Let $<_\mathcal{PD}$ be the transitive closure of $\lessdot_\mathcal{PD}$. This relation is a strict partial order. Let $\sim_\mathcal{PD}$ denote the symmetric closure of the partial order $\leq_\mathcal{PD}$. \end{defn} The \emph{Rothe diagram} of $w \in S_n$ is $ D(w) = \{(i,j) \in [n] \times [n] : w(i) > j\text{ and } w^{-1}(j) > i\}. $ It is often useful to observe that the set $D(w)$ is the complement in $[n]\times [n]$ of the union of the hooks $\{ (x,w(i)) : i < x \leq n\} \sqcup \{ (i,w(i))\} \sqcup \{ (i,y) : w(i) < y \leq n\}$ for $i \in [n]$. It is not hard to show that one always has $|D(w)| = \ell(w)$. For each $i \in [n]$ let $c_i(w) = |\{ j : (i,j) \in D(w)\}|.$ The \emph{code} of $w$ is the integer sequence $ c(w) = (c_1(w), \ldots, c_n(w))$. The \emph{bottom pipe dream} of $w$ is the set \beD_{\text{bot}}(w)=\{(i,j) \in [n]\times [n] : j \leq c_i(w)\} \end{equation} obtained by left-justifying $D(w)$. It is not obvious that $D_{\text{bot}}(w) \in \mathcal{PD}(w)$. \begin{ex} \label{ex:pipedream} If $w = 35142 \in S_5$, then $D(w)$ is the set of $+$'s below: \begin{equation*} \arraycolsep=1.5pt \def0.8{0.8} \begin{array}{ccccc} + & + & 1 & \cdot & \cdot\\ + & + & \cdot & + & 1\\ 1 & \cdot & \cdot & \cdot & \cdot\\ \cdot & + & \cdot & 1 & \cdot\\ \cdot & 1 & \cdot & \cdot & \cdot \end{array} \qquad \text{so we have} \qquad c(w) = (2,3,0,1,0) \qquad\text{and}\qquad D_{\text{bot}}(w) = \arraycolsep=1.5pt \begin{array}{ccccc} + & + & \cdot & \,\cdot\\ + & + & + & \,\cdot\\ \cdot & \cdot & \cdot & \,\cdot \\ + & \cdot & \cdot & \,\cdot \\ \cdot & \cdot & \cdot & \,\cdot \end{array} \end{equation*} \end{ex} \begin{thm}[{\cite[Thm. 3.7]{bergeron-billey}}]\label{bb-thm} Let $w \in S_n$. Then \[\mathcal{PD}(w)= \left\{ E : D_{\text{bot}}(w) \leq_\mathcal{PD} E \right\} = \left\{ E : D_{\text{bot}}(w) \sim_\mathcal{PD} E \right\} .\] Thus $\mathcal{PD}(w)$ is an upper and lower set of $\leq_\mathcal{PD}$, with unique minimum $D_{\text{bot}}(w)$. \end{thm} Define $\leq^\textsf{chute}_\mathcal{PD}$ to be the partial order with $D \leq^\textsf{chute}_\mathcal{PD} E$ if and only if $E^T \leq_\mathcal{PD} D^T$, and let $D_{\text{top}}(w) = D_{\text{bot}}(w^{-1})^T$ for $w \in S_n$. Then $\mathcal{PD}(w)= \left\{ E : E \leq^\textsf{chute}_\mathcal{PD} D_{\text{top}}(w) \right\} $ by Corollary~\ref{transpose-cor} and Theorem~\ref{bb-thm}. Bergeron and Billey \cite{bergeron-billey} refer to the covering relation in $\leq^\textsf{chute}_\mathcal{PD}$ as a \emph{chute move}. In the next sections, we will see that there are natural versions of $\leq_\mathcal{PD}$ and $D_{\text{bot}}(w)$ for (fpf-)involution pipe dreams. There do not seem to be good involution analogues of $\leq^{\textsf{chute}}_\mathcal{PD}$ or $D_{\text{top}}(w)$, however. \subsection{Involution ladder moves} \label{ss:inv-ladder} To prove an analogue of Theorem~\ref{bb-thm} for involution pipe dreams, we need to introduce a more general partial order $<_\mathcal{ID}$ on subsets of $\ZZ_{>0}\times \ZZ_{>0}$. Again let $D$ and $E$ be subsets of $\ZZ_{>0}\times \ZZ_{>0}$. Informally, we define $<_\mathcal{ID}$ to be the transitive closure of $\lessdot_\mathcal{PD}$ and the relation that has $D \lessdot_\mathcal{ID} E$ whenever $E$ is obtained from $D$ by replacing a subset of the form \begin{equation} \def0.8{0.8} \label{eq:inv-ladder} \arraycolsep=1.5pt \begin{array}{cccccc} & \cdot & \cdot & \cdot & \cdot\\ \cdot & \cdot & \cdot & \cdot\\ + & \cdot & \cdot\\ + & + \\ \vdots & \vdots \\ + & + \\ + & \cdot & {\phantom{+}} & {\phantom +} & {\phantom+} \end{array} \qquad\text{by}\qquad \arraycolsep=1.5pt \begin{array}{cccccc} & \cdot & \cdot & \cdot & \cdot\\ \cdot & \cdot & \cdot & \cdot\\ + & + & \cdot\\ + & + \\ \vdots & \vdots \\ + & + \\ \cdot & \cdot & {\phantom+} & {\phantom+} & {\phantom+} \end{array} \end{equation} where the upper parts of the antidiagonals at the top are required to be empty. For example, \[ \arraycolsep=1.5pt \begin{array}{cccccc} \cdot & \cdot & \cdot & \cdot\\ + & \cdot & \cdot & +\\ + & \cdot & \cdot& \cdot\\ \cdot& + & \cdot & \cdot \end{array} \ \lessdot_\mathcal{ID}\ \arraycolsep=1.5pt \begin{array}{cccccc} \cdot & \cdot & \cdot & \cdot\\ + & + & \cdot & +\\ \cdot & \cdot & \cdot& \cdot\\ \cdot & + & \cdot & \cdot \end{array} \qquad\text{and}\qquad \arraycolsep=1.5pt \begin{array}{cccccc} \cdot & \cdot & \cdot & \cdot\\ + & \cdot & \cdot & +\\ + & + & \cdot& \cdot\\ + & \cdot & \cdot & \cdot \end{array} \ \lessdot_\mathcal{ID}\ \arraycolsep=1.5pt \begin{array}{cccccc} \cdot & \cdot & \cdot & \cdot\\ + & + & \cdot & +\\ + & + & \cdot& \cdot\\ \cdot & \cdot & \cdot & \cdot \end{array} \] but \[ \arraycolsep=1.5pt \begin{array}{cccccc} \cdot & \cdot & + & \cdot\\ + & \cdot & \cdot & \cdot \\ + & + & \cdot& \cdot\\ + & \cdot & \cdot & \cdot \end{array} \ \not<_\mathcal{ID}\ \arraycolsep=1.5pt \begin{array}{cccccc} \cdot & \cdot & + & \cdot\\ + & + & \cdot & \cdot\\ + & + & \cdot& \cdot\\ \cdot & \cdot & \cdot & \cdot \end{array} \qquad\text{and}\qquad \arraycolsep=1.5pt \begin{array}{cccccc} \cdot & \cdot & \cdot & +\\ + & \cdot & \cdot & \cdot \\ + & + & \cdot& \cdot\\ + & \cdot & \cdot & \cdot \end{array} \ \not<_\mathcal{ID}\ \arraycolsep=1.5pt \begin{array}{cccccc} \cdot & \cdot & \cdot & +\\ + & + & \cdot & \cdot\\ + & + & \cdot& \cdot\\ \cdot & \cdot & \cdot & \cdot \end{array} \] since the relevant antidiagonals in \eqref{eq:inv-ladder} are not empty. The precise definition of $\lessdot_{\mathcal{ID}} $ is below: \begin{defn}\label{i-ladder-def} We write $ D \lessdot_{\mathcal{ID}} E$ if for some integers $i< j$ and $k$ the following holds: \begin{itemize} \item One has $\{i+1,i+2,\dots,j-1\}\times\{k,k+1\} \subset D$. \item It holds that $(i,k),(j,k) \in D$ but $(i,k+1),(i,k+2),(j,k+1)\notin D$. \item One has $E = D\setminus \{(j,k)\} \cup \{(i,k+1)\}$. \item The set $D$ contains no positions strictly northeast of and in the same antidiagonal as $(i,k-1)$, $(i,k)$, $(i,k+1)$, or $(i,k+2)$. \end{itemize} One may again have $i+1=j$, in which case the first condition holds vacuously. We define $<_\mathcal{ID}$ to be the transitive closure of $\lessdot_\mathcal{PD}$ and $\lessdot_{\mathcal{ID}}$, and write $\sim_\mathcal{ID}$ for the symmetric closure of $\leq_\mathcal{ID}$. \end{defn} Our goal is to show that $<_\mathcal{ID}$ defines a partial order on $\mathcal{ID}(z)$; for an example of this poset, see Figure~\ref{ip-fig}. To proceed, we must recall a few nontrivial properties of the set $\mathcal{A}(z)$ from Section~\ref{schub-sect}. \begin{lem}[{\cite[Lem. 6.3]{HMP2}}] \label{l:321-atoms} Let $z \in \mathcal{I}_n$ and $w \in \mathcal{A}(z)$. Then no subword $w(a)w(b)w(c)$ of $w(1)w(2)\cdots w(n)$ for $1\leq a<b<c\leq n$ has the form $(i-1)i(i+1)$ for any integer $1<i<n$. \end{lem} Fix $z \in \mathcal{I}_n$. The \emph{involution code} of $z $ is $\hat c(z) = (\hat c_1(z),\hat c_2(z),\dots,\hat c_n(z))$ with $\hat c_i(z)$ the number of integers $j>i$ with $z(i) > z(j)$ and $i\geq z(j)$. Note that we always have $\hat c_i(z) \leq i$. Suppose $a_1<a_2<\dots<a_l$ are the integers $a\in [n]$ with $a \leq z(a)$ and set $b_i = z(a_i)$. Define $\alpha_{\min}(z) \in S_n$ to be the permutation whose inverse is given in one-line notation by removing all repeated letters from $b_1 a_1 b_2 a_2\cdots b_l a_l$. For example, if $z = 4231 \in \mathcal{I}_4$ then the latter word is $412233$ and $\alpha_{\min}(z) = (4123)^{-1} = 2341 \in S_4$. Additionally, $\hat c(z) = c(\alpha_{\min}(z))$~\cite[Lem. 3.8]{HMP1}. Finally, let $\prec_\mathcal{A}$ be the transitive closure of the relation on $S_n$ that has $v \prec_\mathcal{A} w$ whenever the inverses of $v,w \in S_n$ have the same one-line representations outside of three consecutive positions where $v^{-1} = \cdots cab\cdots $ and $ w^{-1}=\cdots bca\cdots$ for some integers $a<b<c$. The relation $\prec_\mathcal{A}$ is a strict partial order. Let $\sim_\mathcal{A}$ denote the symmetric closure of the partial order $\preceq_\mathcal{A}$. \begin{thm}[{\cite[\S6.1]{HMP2}}] \label{t:atoms} Let $z \in \mathcal{I}_n$. Then \[ \mathcal{A}(z) = \{w \in S_n: \alpha_{\min}(z) \preceq_\mathcal{A} w\} = \{w \in S_n: \alpha_{\min}(z) \sim_\mathcal{A} w\}. \] Thus $\mathcal{A}(z)$ is an upper and lower set of $\preceq_\mathcal{A}$, with unique minimum $\alpha_{\min}(z)$. \end{thm} For $z \in \mathcal{I}_n$, let $\mathcal{ID}^+(z) = \bigsqcup_{w \in \mathcal{A}(z)} \mathcal{PD}(w) $, so that $\mathcal{ID}(z) = \{D \in \mathcal{ID}^+(z) : D \subseteq \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n\}$. \begin{lem}\label{inv-ladder-lem} Let $z \in \mathcal{I}_n$. Suppose $D$ and $E$ are subsets of $\ZZ_{>0}\times \ZZ_{>0}$ with $D <_{\mathcal{ID}} E $. Then $D \in \mathcal{ID}^+(z)$ if and only if $E \in \mathcal{ID}^+(z)$. \end{lem} \begin{proof} If $D \lessdot_\mathcal{PD} E$ then we have $D \in \mathcal{ID}^+(z)$ if and only if $E \in \mathcal{ID}^+(z)$ by Theorem~\ref{bb-thm}. Assume $D \lessdot_{\mathcal{ID}}E$ and let $i<j$ and $k$ be as in Definition~\ref{i-ladder-def}. Consider the reading order $\omega$ that lists the positions $(i,j)\in [n]\times [n]$ such that $(-j,i)$ increases lexicographically, i.e., the order that goes down column $n$, then down column $n-1$, and so on. In view of Theorem~\ref{thm:almost-symmetric}, we may assume without loss of generality that columns $1,2,\dots,k-1$ of $D$ and $E$ are both empty, since omitting these positions has the effect of truncating the same final sequence of letters from $\textsf{word}(D,\omega)$ and $\textsf{word}(E,\omega)$. Suppose $E \in \mathcal{PD}(w)\subseteq \mathcal{ID}^+(z)$ for some permutation $w \in \mathcal{A}(z)$. To show that $D \in \mathcal{ID}^+(z)$, it suffices by Theorem~\ref{t:atoms} to check that $D \in \mathcal{PD}(v)$ for a permutation $v \prec_\mathcal{A} w$. Consider the wiring diagram of $E$ and let $m,m+1$ and $m+2$ be the top indices of the wires in the antidiagonals containing the cells $ (i,k),$ $(i,k+1)$, and $(i+1,k+1) $, respectively. Since the northeast parts of these antidiagonals are empty, it follows that as one goes from northeast to southwest, wire $m$ of $E$ enters the top of the $+$ in cell $(i,k)$, wire $m+1$ enters the top of the $+$ in cell $(i,k+1)$, and wire $m+2$ enters the right of the $+$ in cell $(i,k+1)$. Tracing these wires through the wiring diagram of $E$, we see that they exit column $k$ on the left in relative order $m+2,m,m+1$. Since we assume columns $1,2,\dots,k-1$ are empty, the wires must arrive at the far left in the same relative order. This means that there are numbers $a<b<c$ such that $w^{-1}(m)w^{-1}(m+1)w^{-1}(m+2)=bca$. Moving the $+$ in cell $(i,k+1)$ of $E$ to $(j,k)$ gives $D$ by assumption. This transformation only alters the trajectories of wires $m$, $m+1$ and $m+2$ and causes no pair of wires to cross more than once, so $D$ is a reduced pipe dream for some $v \in S_n$. By examining the wiring diagram of $D$, we see that $v^{-1}(m)v^{-1}(m+1)v^{-1}(m+2)=cab$, so $v \prec_\mathcal{A} w$ and $D \in \mathcal{ID}^+(z)$ as needed. The same considerations show that if $D \in \mathcal{PD}(v)$ for some $v \in \mathcal{A}(z)$ then $E \in \mathcal{PD}(w)$ for a permutation $w$ with $v \prec_\mathcal{A} w$. In this case, it follows that $w \in \mathcal{A}(z)$ by Theorem~\ref{t:atoms} so $E \in \mathcal{ID}^+(z)$. \end{proof} We define the \emph{bottom involution pipe dream} of $z \in \mathcal{I}_n$ to be the set \begin{equation} \hat D_{\text{bot}}(z) = \{(i,j) \in [n] \times [n]: j \leq \hat c_i(z)\} \subseteq \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n. \end{equation} Since $\hat c(z) = c(\alpha_{\min}(z))$, it follows by Theorem~\ref{thm:almost-symmetric} that $\hat D_{\text{bot}}(z) = D_{\text{bot}}(\alpha_{\min}(z))\in \mathcal{ID}(z).$ \begin{thm}\label{pre-t:inv-ladder} Let $z \in \mathcal{I}_n$. Then $ \mathcal{ID}^+(z) = \left\{E : \hat D_{\text{bot}}(z) \leq_\mathcal{ID} E\right\} = \left\{E : \hat D_{\text{bot}}(z) \sim_\mathcal{ID} E\right\}. $ Thus $\mathcal{ID}^+(z)$ is an upper and lower set of $\leq_\mathcal{ID}$, with unique minimum $\hat D_{\text{bot}}(z)$. \end{thm} \begin{proof} Both sets are contained in $\mathcal{ID}^+(z)$ by Lemma~\ref{inv-ladder-lem}. Note that $\mathcal{ID}^+(z)$ is finite since $\mathcal{A}(z)$ is finite and each set $\mathcal{PD}(w)$ is finite. Suppose $ \hat D_{\text{bot}}(z) \neq E = D_{\text{bot}}(w)$ for some $w \in \mathcal{A}(z)$. In view of Theorem~\ref{bb-thm}, we need only show that there exists a subset $D\subset \ZZ_{>0}\times \ZZ_{>0}$ with $D\lessdot_\mathcal{ID} E$. As we assume $w \neq \alpha_{\min}(z)$, it follows from Theorem~\ref{t:atoms} that there exists some $p \in [n-2]$ with $w^{-1}(p{+}2) < w^{-1}(p) < w^{-1}(p{+}1)$. Set $i = w^{-1}(p{+}2)$, and choose $p$ to minimize $i$. We claim that if $h < i$ then $w(h) < p$. To show this, we argue by contradiction. Suppose there exists $1\leq h < i$ with $w(h) \geq p$. Choose $h$ with this property so that $w(h)$ is as small as possible. Then $w(h) > p+2 \geq 3$, and by the minimality of $w(h)$, the values $w(h){-}1$ and $w(h){-}2$ appear after position $h$ in the word $w(1)w(2)\cdots w(n)$. Therefore, by Lemma~\ref{l:321-atoms}, the one-line representation of $w$ must have the form $\cdots w(h) \cdots w(h){-}2\cdots w(h){-}1\cdots$. This contradicts the minimality of $i$, so no such $h$ can exist. Let $j > i$ be minimal with $w(j) < w(i)$ and define $k = c_i(w) - 1$. It is evident from the definition of $i$ that such an index $j$ exists and that $k$ is positive. Now consider Definition~\ref{i-ladder-def} applied to these values of $i<j$ and $k$. It follows from the claim in the previous paragraph if $h<i$ then $c_h(w) - c_i(w) \leq i-h -3$. Therefore, we see that the required antidiagonals are empty. The minimality of $j$ implies that $c_m(w) \geq c_i(w)$ for all $i < m < j$, and since we must have $j \leq w^{-1}(p)$, it follows that $c_j(w) < c_i(w) - 1$. We conclude that replacing position $(i,k+1)$ in $E$ by $(j,k)$ produces a subset $D$ with $D \lessdot_\mathcal{ID} E$, as we needed to show. \end{proof} \begin{thm} \label{t:inv-ladder} If $z \in \mathcal{I}_n$ then $ \mathcal{ID}(z) = \left\{ E \subseteq \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n : \hat D_{\text{bot}}(z) \leq_\mathcal{ID} E \right\} = \left\{ E \subseteq \raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n : \hat D_{\text{bot}}(z) \sim_\mathcal{ID} E \right\} .$ \end{thm} \begin{proof} This is clear from Theorem~\ref{pre-t:inv-ladder} since $\raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$ is a lower set under $\leq_\mathcal{ID}$. \end{proof} \begin{figure}[h] \begin{center} {\small \begin{tikzpicture}[xscale=1.2,yscale=1.2] \node at (0,0) (a) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & \cdot & & \\ + & \cdot &\cdot \\ + & + &\cdot &\cdot\\ + &\cdot &\cdot &\cdot&\cdot \end{array}$ }; \node at (-4,3) (b1) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & + & & \\ \cdot & \cdot &\cdot \\ + & + &\cdot & \cdot \\ + &\cdot &\cdot & \cdot & \cdot \end{array}$ }; \node at (0,3) (b2) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & \cdot & & \\ + & \cdot & + \\ + & \cdot & \cdot & \cdot \\ + &\cdot & \cdot& \cdot & \cdot \end{array}$ }; \node at (4,3) (b3) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & \cdot & & & \\ + & + &\cdot \\ + & + &\cdot& \cdot \\ \cdot &\cdot &\cdot & \cdot & \cdot \end{array}$ }; \node at (-4,6) (c1) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & + & & \\ \cdot & + &\cdot \\ + & + &\cdot & \cdot \\ \cdot &\cdot &\cdot & \cdot & \cdot \end{array}$ }; \node at (0,6) (c2) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & + & & \\ \cdot & \cdot& + \\ + & \cdot & \cdot & \cdot \\ + &\cdot & \cdot& \cdot & \cdot \end{array}$ }; \node at (4,6) (c3) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & \cdot & & \\ + & + & + \\ + & \cdot &\cdot& \cdot \\ \cdot &\cdot & \cdot& \cdot & \cdot \end{array}$ }; \node at (0,9) (d2) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & + & & \\ \cdot & + & + \\ \cdot & \cdot & \cdot & \cdot \\ + &\cdot & \cdot & \cdot & \cdot \end{array}$ }; \node at (4,9) (d3) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & + & & \\ + & + & + \\ \cdot & \cdot &\cdot & \cdot \\ \cdot &\cdot & \cdot& \cdot & \cdot \end{array}$ }; \node at (0,12) (e) { $\arraycolsep=1.5pt \begin{array}{ccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} \\ \cdot & + & & \\ \cdot & + & + \\ \cdot & + & \cdot & \cdot \\ \cdot &\cdot & \cdot & \cdot & \cdot \end{array}$ }; \draw[->] (a) -- (b1); \draw[->] (a) -- (b2); \draw[red,->,dashed] (a) -- (b3); \draw[->] (b2) -- (c2); \draw[->] (b1) -- (c2); \draw[->] (b1) -- (c1); \draw[red,->,dashed] (b3) -- (c3); \draw[->] (c2) -- (d2); \draw[->] (c3) -- (d3); \draw[->] (d2) -- (e); \end{tikzpicture}} \end{center} \caption{Hasse diagram of $(\mathcal{ID}(z), <_\mathcal{ID})$ for $z=(3,6)(4,5) \in \mathcal{I}_6$. The dashed red arrows indicate the covering relations of the form $D \lessdot_\mathcal{ID} E$.} \label{ip-fig} \end{figure} \subsection{Fixed-point-free involution ladder moves} In this subsection, we assume $n$ is a positive even integer. Our goal is to replicate the results in Section~\ref{ss:inv-ladder} for fixed-point-free involutions. To this end, we introduce a third partial order $<_\mathcal{FD}$. Again let $D$ and $E$ be subsets of $\ZZ_{>0}\times \ZZ_{>0}$. We define $<_\mathcal{FD}$ as the transitive closure of $\lessdot_\mathcal{PD}$ and the relation that has $D \lessdot_\mathcal{FD} E$ whenever $E$ is obtained from $D $ by replacing a subset of the form \begin{equation} \label{eq:fpf-ladder} \arraycolsep=1.5pt \def0.8{0.8} \begin{array}{ccccccc} {\phantom+} & \cdot & \cdot & \cdot &\cdot & \cdot \\ \cdot & \cdot & \cdot & \cdot & \cdot\\ \cdot & + & \cdot & \cdot\\ & + & + \\ & \vdots & \vdots \\ & + & + \\ & + & \cdot & {\phantom+} & {\phantom+} & {\phantom+}& {\phantom+} \end{array} \qquad\text{by}\qquad \arraycolsep=1.5pt \begin{array}{ccccccc} {\phantom+} & \cdot & \cdot & \cdot &\cdot & \cdot \\ \cdot & \cdot & \cdot & \cdot & \cdot\\ + & + & \cdot & \cdot\\ & + & + \\ & \vdots & \vdots \\ & + & + \\ & \cdot & \cdot & {\phantom+} & {\phantom+} & {\phantom+}& {\phantom+} \end{array} \end{equation} Here, all positions containing ``$\ \cdot\ $'' should be empty, including the five antidiagonals extending upwards. The precise definition of $\lessdot_\mathcal{FD}$ is as follows: \begin{defn} \label{fpf-ladder-def} We write $ D \lessdot_{\mathcal{FD}} E$ if for some integers $0<i<j$ and $k\geq 2$ the following holds: \begin{itemize} \item One has $\{i+1,i+2,\dots,j-1\}\times\{k,k+1\} \subset D$. \item It holds that $(i,k),(j,k) \in D$ but $(i,k-1),(i,k+1),(i,k+2),(j,k+1)\notin D$. \item One has $E = D\setminus \{(j,k)\} \cup \{(i,k-1)\}$. \item The set $D$ contains no positions strictly northeast of and in the same antidiagonal as $(i,k-2)$, $(i,k-1)$, $(i,k)$, $(i,k+1)$, or $(i,k+2)$. \end{itemize} When $i+1=j$, the first condition holds vacuously; see the lower dashed arrow in Figure~\ref{fp-fig}. Define $<_\mathcal{FD}$ to be the transitive closure of $\lessdot_\mathcal{PD}$ and $\lessdot_{\mathcal{FD}}$. Write $\sim_\mathcal{FD}$ for the symmetric closure of $\leq_\mathcal{FD}$. \end{defn} We will soon show that $<_\mathcal{FD}$ defines a partial order on $\mathcal{FD}(z)$, as one can see in the example shown in Figure~\ref{fp-fig}. For this, we will need a lemma from \cite{can-joyce-wyser} concerning the set $\mathcal{A}^{\fpf}(z)$. \begin{lem}[{\cite[Cor. 2.16]{can-joyce-wyser}}] \label{l:fpf-atoms} Let $w \in S_{n}$ and $z \in \I^{\fpf}_{n}$. Then $w \in \mathcal{A}^{\fpf}(z)$ if and only if for all $a,b,c,d \in [n]$ with $a<b=z(a)$ and $c<d=z(c)$, the following holds: \begin{enumerate} \item[(1)] One has $w(a) = 2i-1$ and $w(b) = 2i$ for some $i \in [n/2]$. \item[(2)] If $a<c$ and $b < d$, then $w(b) < w(c)$. \end{enumerate} \end{lem} The involution code and partial order $\prec_\mathcal{A}$ both have fixed-point-free versions. Fix $z \in \I^{\fpf}_n$. The \emph{fpf-involution code} of $z$ is the integer sequence \[ \hat c^{\fpf}(z) = (\hat c^{\fpf}_1(z), \hat c^{\fpf}_2(z), \dots, \hat c^{\fpf}_n(z)) \] where $\hat c^{\fpf}_i(z)$ is the number of integers $j > i$ with $z(i) > z(j)$ and $i > z(j)$. It always holds that $\hat c^{\fpf}_i(z) < i$. If $a_1<a_2<\dots <a_{n/2}$ are the numbers $a\in [n]$ with $a< z(a)$ and $b_i = z(a_i)$, then let \[ \alpha^\fpf_{\min}(z) = (a_1b_1a_2b_2\dots a_{n/2} b_{n/2})^{-1} = s_1 s_3 s_5\cdots s_{n-1} \alpha_{\min}(z) \in S_n.\] For example, if $z = 632541 \in \I^{\fpf}_6$ then we have $\alpha^\fpf_{\min}(z) = (162345)^{-1} = 134562 \in S_6$. One can check that $\hat c^{\fpf}(z) = c(\alpha^\fpf_{\min}(z))$ \cite[Lem. 3.8]{HMP1}. Define $\prec_{\mathcal{A}^{\fpf}}$ to be the transitive closure of the relation in $S_n$ that has $v \prec_{\mathcal{A}^{\fpf}} w$ whenever the inverses of $v,w \in S_n$ have the same one-line representations outside of four consecutive positions where $v^{-1} = \cdots adbc \cdots$ and $w^{-1} = \cdots bcad \cdots$ for some integers $a<b<c<d$. This is a strict partial order on $S_n$. Let $\sim_{\mathcal{A}^{\fpf}}$ denote the symmetric closure of the partial order $\preceq_{\Afpf}$. \begin{thm}[{\cite[\S6.2]{HMP2}}] \label{t:fpf-order} Let $z \in \I^{\fpf}_n$. Then \[ \mathcal{A}^{\fpf}(z) = \left\{w \in S_n: \alpha^\fpf_{\min}(z) \preceq_{\mathcal{A}^{\fpf}} w \right\} = \left\{w \in S_n: \alpha^\fpf_{\min}(z) \sim_{\mathcal{A}^{\fpf}} w \right\} . \] Thus $\mathcal{A}^{\fpf}(z)$ is an upper and lower set of $\preceq_{\mathcal{A}^{\fpf}}$, with unique minimum $\alpha^\fpf_{\min}(z)$. \end{thm} For $z \in \I^{\fpf}_n$, let $\mathcal{FD}^+(z) = \bigsqcup_{w\in\mathcal{A}^{\fpf}(z)} \mathcal{PD}(w)$, so $\mathcal{FD}(z) = \mathcal{FD}^+(z) \cap \ltriang^{\!\!\neq}_n$. \begin{lem}\label{fpf-ladder-lem} Let $z \in \I^{\fpf}_n$. Suppose $D$ and $E$ are subsets of $\ZZ_{>0}\times \ZZ_{>0}$ with $D <_{\mathcal{FD}} E $. Then $D \in \mathcal{FD}^+(z)$ if and only if $E \in \mathcal{FD}^+(z)$. \end{lem} \begin{proof} If $D \lessdot_\mathcal{PD} E$ then the result follows by Theorem~\ref{bb-thm}. Assume $D \lessdot_{\mathcal{FD}}E$ and let $i<j$ and $k$ be as in Definition~\ref{fpf-ladder-def}. As in the proof of Theorem~\ref{inv-ladder-lem}, consider the reading order $\omega$ that lists the positions $(i,j)\in [n]\times [n]$ such that $(-j,i)$ increases lexicographically. In view of Theorem~\ref{thm:fpf-almost-symmetric}, we may assume without loss of generality that columns $1,2,\dots,k-2$, as well as all positions below row $i$ in column $k-1$, are empty in both of $D$ and $E$. This follows since omitting these positions has the effect of truncating the same final sequence of letters from $\textsf{word}(D,\omega)$ and $\textsf{word}(E,\omega)$. Assume $E \in \mathcal{PD}(v) \subseteq \mathcal{FD}^+(z)$ for some $w \in \mathcal{A}^{\fpf}(z)$. To show that $D \in \mathcal{FD}^+(z)$, we will check that $D \in \mathcal{PD}(v)$ for some $v\in S_n$ with $v \prec_{\mathcal{A}^{\fpf}} w$. Consider the wiring diagram of $E$ and let $m$, $m+1$, $m+2$, and $m+3$ be the top indices of the wires in the antidiagonals containing the cells $(i,k-1)$, $(i,k)$, $(i,k+1)$, and $(i,k+2)$, respectively. Since the northeast parts of these antidiagonals are empty, it follows that as one goes from northeast to southwest, wire $m$ of $E$ enters the top of the $+$ in cell $(i,k-1)$, wire $m+1$ enters the top of the $+$ is cell $(i,k)$, wire $m+2$ enters the right of the $+$ in cell $(i,k)$, and wire $m+3$ enters the top of cell $(i+1,k+1)$, which contains a $+$ if $i+1<j$. Tracing these wires through the wiring diagram of $E$, we see that they exit column $k-1$ on the left in relative order $m+2$, $m$, $m+1$, $m+3$. Since we assume that $D$ and $E$ contain no positions in the rectangle weakly southwest of $(i+1,k-1)$, the wires must arrive at the far left in the same relative order. This means that $w^{-1}(m)w^{-1}(m+1) w^{-1}(m+2) w^{-1}(m+3) = bcad$ for some numbers $a<b<c<d$. Moving the $+$ in cell $(i,k-1)$ of $E$ to $(j,k)$ gives $D$ by assumption. This transformation only alters the trajectories of wires $m$, $m+1$, $m+2$, and $m+3$ and causes no pair of wires to cross more than once, so $D$ is a reduced pipe dream for some $v \in S_n$. By examining the wiring diagram of $D$, it is easy to check that $v^{-1}(m)v^{-1}(m+1) v^{-1}(m+2) v^{-1}(m+3) = adbc$ so $v \prec_{\mathcal{A}^{\fpf}} w$ as needed. If instead $D \in \mathcal{PD}(v)\subset \mathcal{FD}^+(z)$ for some $v \in \mathcal{A}^{\fpf}(z)$, then a similar argument shows that $E \in \mathcal{PD}(w)$ for some $w \in S_n$ with $v \prec_{\mathcal{A}^{\fpf}} w$, which implies that $E \in \mathcal{FD}^+(z)$ by Theorem~\ref{t:fpf-order}. \end{proof} We define the \emph{bottom fpf-involution pipe dream} of $z \in \I^{\fpf}_n$ to be the set \begin{equation} \hat D_{\text{bot}}^{\fpf}(z) = \left\{(i,j) \in [n]\times [n]: j \leq \hat c^{\fpf}_i(z)\right\} \subseteq \ltriang^{\!\!\neq}_n. \end{equation} Since $\hat c^{\fpf}(z) = c(\alpha^\fpf_{\min}(z))$, Theorem~\ref{thm:fpf-almost-symmetric} implies that $ \hat D_{\text{bot}}^{\fpf}(z) = D_{\text{bot}}(\alpha^\fpf_{\min}(z)) \in \mathcal{FD}(z). $ \begin{thm}\label{pre-t:fpf-ladder} Let $z \in \I^{\fpf}_n$. Then $ \mathcal{FD}^+(z) = \left\{E : \hat D_{\text{bot}}^{\fpf}(z) \leq_\mathcal{FD} E\right\} = \left\{E : \hat D_{\text{bot}}^{\fpf}(z) \sim_\mathcal{FD} E\right\}. $ Thus $\mathcal{FD}^+(z)$ is an upper and lower set of $\leq_\mathcal{FD}$, with unique minimum $\hat D_{\text{bot}}^{\fpf}(z)$. \end{thm} \begin{proof} Both sets are contained in $\mathcal{FD}^+(z)$ by Lemma~\ref{fpf-ladder-lem}, and the set $\mathcal{FD}^+(z)$ is clearly finite. Suppose $ \hat D_{\text{bot}}^{\fpf}(z) \neq E = D_{\text{bot}}(w)$ for some $w \in \mathcal{A}^{\fpf}(z)$. As in the proof of Theorem~\ref{pre-t:inv-ladder}, it suffices to show that there exists a subset $D\subset \ZZ_{>0}\times \ZZ_{>0}$ with $D\lessdot_\mathcal{FD} E$. Since $w \neq \alpha^\fpf_{\min}(z)$, Lemma~\ref{l:fpf-atoms} and Theorem~\ref{t:fpf-order} imply that there exists an odd integer $p \in [n - 3]$ such that $ w^{-1}(p)w^{-1}(p+1)w^{-1}(p+2)w^{-1}(p+3) = bcad$ for some numbers $a<b<c<d$. Choose $p$ such that $a$ is as small as possible. We claim that $a<w^{-1}(q)$ for all $q$ with $p+3 <q \leq n$. To show this, let $a_0=a$ and $b_0 =d$ and suppose $a_i $ and $b_i$ are the integers such that \[w^{-1}(p+2)w^{-1}(p+3)\cdots w^{-1}(n) = a_0b_0 a_1 b_1 \cdots a_k b_k.\] Part (1) of Lemma~\ref{l:fpf-atoms} implies that $a_i < b_i =z(a_i)$ for all $i$, so it suffices to show that $a_0<a_i$ for $i \in [k]$. This holds since if $i \in [k]$ were minimal with $a_i < a_0$, then it would follow from part (2) of Lemma~\ref{l:fpf-atoms} that $a_i < a_{i-1} < b_{i-1}< b_i$, contradicting the minimality of $a$. Now, to match Definition~\ref{fpf-ladder-def}, let $i = a=w^{-1}(p+2)$, define $j > i$ to be minimal with $w(j) < w(i)$, and set $k = c_i(w)$. It is clear from the definition of $i$ that such an index $j$ exists and that $k\geq 2$. The claim in the previous paragraph shows that if $1\leq h < i$ then $h$ must appear before position $p$ in the one-line representation of $w^{-1}$, which means that $w(h)<p$ and therefore $c_h(w) - c_i(w) \leq i-h- 4$. The antidiagonals described in Definition~\ref{fpf-ladder-def} are thus empty as needed. Since $j \leq b= w^{-1}(p)$, it follows that $c_j(w) < c_i(w)$; moreover, if $i < m < j$ then $w(m) > w(d) = p+3$ so $c_m(w) \geq c_i(w) + 1$. Collecting these observations, we conclude that replacing $(i,k-1)$ in $E$ with $(j,k)$ gives a subset $D$ with $D \lessdot_\mathcal{ID} E$, as we needed to show. \end{proof} \begin{thm} \label{t:fpf-ladder} If $z \in \I^{\fpf}_n$ then \[ \mathcal{FD}(z) = \left\{ E \subseteq \ltriang^{\!\!\neq}_n : \hat D_{\text{bot}}^{\fpf}(z) \leq_\mathcal{FD} E \right\} = \left\{ E \subseteq \ltriang^{\!\!\neq}_n : \hat D_{\text{bot}}^{\fpf}(z) \sim_\mathcal{FD} E \right\} .\] \end{thm} \begin{proof} This is clear from Theorem~\ref{pre-t:fpf-ladder} since $\ltriang^{\!\!\neq}_n$ is a lower set under $\leq_\mathcal{FD}$. \end{proof} \begin{figure}[h] \begin{center} {\scriptsize \begin{tikzpicture}[xscale=1.1,yscale=0.725] \node at (0,0) (a) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & \cdot &\cdot \\ + & \cdot &\cdot &\cdot\\ + &+ &\cdot &\cdot&\cdot \\ + &+ &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (-4,4) (b1) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ \cdot & \cdot &\cdot &\cdot\\ + &+ &\cdot &\cdot&\cdot \\ + &+ &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (0,4) (b2) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & \cdot &\cdot \\ + & \cdot &+ &\cdot\\ + &\cdot &\cdot &\cdot&\cdot \\ + &+ &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (-4,8) (c1) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ \cdot & \cdot &+ &\cdot\\ + &\cdot &\cdot &\cdot&\cdot \\ + &+ &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (0,8) (c2) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ \cdot & \cdot &+ &\cdot\\ + &\cdot &\cdot &\cdot&\cdot \\ + &+ &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (4,8) (c3) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & \cdot &\cdot \\ + & \cdot &+ &\cdot\\ + &\cdot &+ &\cdot&\cdot \\ + &\cdot &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (-4,12) (d1) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ \cdot & + &+ &\cdot\\ \cdot &\cdot &\cdot &\cdot&\cdot \\ + &+ &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (0,12) (d2) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ \cdot & \cdot &+ &\cdot\\ + &\cdot &+ &\cdot&\cdot \\ + &\cdot &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (4,12) (d3) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & \cdot &\cdot \\ + & + &+ &\cdot\\ + &\cdot &\cdot &\cdot&\cdot \\ + &\cdot &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (0,16) (e1) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ \cdot & + &+ &\cdot\\ \cdot &\cdot &+ &\cdot&\cdot \\ + &\cdot &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (4,16) (e2) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ + & + &+ &\cdot\\ \cdot &\cdot &\cdot &\cdot&\cdot \\ + &\cdot &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (0,20) (f1) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ \cdot & + &+ &\cdot\\ \cdot &+ &+ &\cdot&\cdot \\ \cdot &\cdot &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (4,20) (f2) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ \cdot & + &\cdot \\ + & + &+ &\cdot\\ \cdot &+ &\cdot &\cdot&\cdot \\ \cdot &\cdot &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \node at (4,24) (g1) { $\arraycolsep=1.5pt \begin{array}{cccccc} \cdot & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+} & {\phantom+}\\ \cdot & \cdot & & \\ + & + &\cdot \\ + & + &+ &\cdot\\ \cdot &\cdot &\cdot &\cdot&\cdot \\ \cdot &\cdot &\cdot &\cdot&\cdot & \cdot \end{array}$ }; \draw[->] (a) -- (b1); \draw[->] (a) -- (b2); \draw[->] (b1) -- (c1); \draw[->] (b2) -- (c2); \draw[->] (b2) -- (c3); \draw[->] (c1) -- (d1); \draw[->] (c1) -- (d2); \draw[->] (c2) -- (d1); \draw[->] (c2) -- (d2); \draw[->] (c3) -- (d2); \draw[red,->,dashed] (c3) -- (d3); \draw[->] (d1) -- (e1); \draw[->] (d2) -- (e1); \draw[->] (d3) -- (e2); \draw[->] (e1) -- (f1); \draw[->] (e2) -- (f2); \draw[red,->,dashed] (f2) -- (g1); \end{tikzpicture}} \end{center} \caption{Hasse diagram of $(\mathcal{FD}(z), <_\mathcal{FD})$ for $z=(1,2)(3,7)(4,8)(5,6) \in \I^{\fpf}_6$. The dashed red arrows indicate the covering relations of the form $D \lessdot_\mathcal{FD} E$. }\label{fp-fig} \end{figure} \section{Future directions} \label{sec:future} In this final section we discuss some related identities and open problems. \subsection{Enumerating involution pipe dreams} \label{sec:enumeration} Choose $w \in S_n$ and let $p =\ell(w)$. Macdonald~\cite[(6.11)]{macdonald1991notes} proved that the following specialization of a Schubert polynomial gives an exact formula for the number of reduced pipe dreams for $w$: \begin{equation}\label{bhy-eq} |\mathcal{PD}(w)| = \mathfrak{S}_w(1,1,\dots,1) = \frac{1}{p!}\sum_{(a_1,a_2,\dots,a_{p}) \in \mathcal{R}(w)} a_1 a_2\cdots a_{p}. \end{equation} Recall that $\kappa(y)$ is the number of 2-cycles in $y \in\mathcal{I}_n$. For $D \in \mathcal{ID}(y)$, define $\wt(D) = 2^{\kappa(y) -d_D}$ where $d_D $ is the number of diagonal positions in $D$. For $\mathcal{X} \subseteq \mathcal{ID}(y)$, define $\|\mathcal{X} \| = \sum_{D \in \mathcal{X}} \wt(D)$. \begin{cor}\label{bhy-cor} Suppose $y \in \mathcal{I}_n$, $z \in \I^{\fpf}_{2n}$, and $p = \iell(y) =\iellfpf(z)$. \begin{enumerate} \item[(a)] $\|\mathcal{ID}(y)\| = 2^{\kappa(y)} \hat{\mathfrak{S}}_y(\tfrac{1}{2},\tfrac{1}{2},\dots,\tfrac{1}{2}) = \frac{1}{2^pp!}\sum_{(a_1,a_2,\dots,a_{p}) \in \hat{\mathcal{R}}(y)} 2^{\kappa(y)}a_1 a_2\cdots a_{p}$. \item[(b)] $ |\mathcal{FD}(z)| = \hat{\mathfrak{S}}^{\fpf}_z(\tfrac{1}{2},\tfrac{1}{2},\dots,\tfrac{1}{2}) = \frac{1}{2^pp!}\sum_{(a_1,a_2,\dots,a_{p}) \in \hat{\mathcal{R}}^{\fpf}(z)} a_1 a_2\cdots a_{p}$. \end{enumerate} \end{cor} \begin{proof} In both parts, the first equality is immediate from Theorem~\ref{thm:inv-pipe-dream-formula} and the second equality is a consequence of \eqref{bhy-eq}, via Definitions~\ref{inv-sch-def} and \ref{fpf-inv-sch-def}. \end{proof} Billey, Holroyd, and Young gave the first bijective proof of \eqref{bhy-eq} (and of a more general $q$-analogue) in the recent paper \cite{billey2019bijective}. This follow-up problem is natural: \begin{problem} Find bijective proofs of the identities in Corollary~\ref{bhy-cor}. \end{problem} For some permutations, better formulas than Equation~\ref{bhy-eq} are available. A \emph{reverse plane partition} of shape $D \subset \ZZ_{>0}\times \ZZ_{>0}$ is a map $T: D \to \ZZ_{\geq 0}$ such that $ T(i,j) \geq T(i+1,j)$ and $T(i,j) \geq T(i,j+1)$ for all relevant $(i,j) \in D$. If $\lambda$ is a partition, then let $\mathsf{RPP}_\lambda(k)$ be the set of reverse plane partitions of Ferrers shape $\textsf{D}_\lambda = \{ (i,j) \in \ZZ_{>0}\times \ZZ_{>0} : j \leq \lambda_i\}$ with entries in $\{0,1,\dots,k\}$. Given $w \in S_n$ write $1^k \times w$ for the permutation in $S_{n+k}$ that fixes $1,2,\dots,k$ while mapping $i+k \mapsto w(i) + k$ for $i \in [n]$. Fomin and Kirillov \cite[Thm. 2.1]{fomin1997reduced} showed that if $w \in S_n$ is dominant then $ |\mathcal{PD}(1^k \times w)| = |\mathsf{RPP}_{\lambda}(k)| $ for the partition $\lambda$ with $\dom{w} = \textsf{D}_\lambda$. In particular: \begin{equation} |\mathcal{PD}(1^k \times n \cdots 321)| = |\mathsf{RPP}_{(n-1,\dots,3,2,1)}(k)| = \prod_{1 \leq i < j \leq n} \frac{ i + j+2k -1}{i + j -1}. \end{equation} Serrano and Stump gave a bijective proof of this identity in \cite{serrano2012maximal}. There are similar formulas counting (weighted) involution pipe dreams. For example: \begin{prop} Let $g_n = (1,n+1)(2,n+2)\cdots(n,2n) \in \mathcal{I}_{2n}$. Then \[|\mathcal{ID}(1^k \times g_n)| = |\mathsf{RPP}_{(n,\dots,3,2,1)}(\lfloor k/2 \rfloor)| \quad\text{for all $k \in \ZZ_{\geq 0}$.}\] \end{prop} \begin{proof} It follows from Theorem~\ref{t:atoms} that $\mathcal{A}(g_n) = \{w_n \}$ for $w_n = 246 \cdots (2n) 1 35 \cdots (2n-1) \in S_{2n}$, and that $\mathcal{A}(1^k\times g_n) = \{1^k \times w_n\}$. Moreover, we have $\hat D_{\text{bot}}(1^k\times g_n) = \{ (j+k,i) : 1 \leq i \leq j \leq n\}$. From these facts, we see that $\mathcal{ID}(1^k \times g_n)$ is connected by ordinary ladder moves that are \emph{simple} in that they replace a single cell $(i,j)$ by $(i-1,j+1)$. Now consider all ways of filling the cells $(i,j)\in \hat D_{\text{bot}}(1^k\times g_n)$ by numbers $a \in \{0,1,\dots,\lfloor k/2\rfloor\}$ such that rows are weakly increasing and columns are weakly decreasing. The set of such fillings is obviously in bijection with $\mathsf{RPP}_{(n,\dots,3,2,1)}(\lfloor k/2 \rfloor)$. On the other hand, we can transform such a filling into a subset of $\raisebox{-0.5pt}{\tikz{\draw (0,0) -- (.25,0) -- (0,.25) -- (0,0);}}_n$ by replacing each cell $(i,j)$ filled with $a$ by $(i-a,j+a)$. It is easy to see that this operation is a bijection from our set of fillings to $\mathcal{ID}(1^k\times g_n)$. \end{proof} Computations indicate that if $k,n \in \ZZ_{\geq 0}$ and $\{p,q\}= \{\lfloor n/2 \rfloor, \lceil n/2 \rceil\}$ then \begin{equation} \label{eq:pp} \begin{aligned} |\mathcal{ID}(1^k \times n \cdots 321)| &= \prod_{i=1}^{p} \prod_{j=1}^{q} \frac{i+j+k-1}{i+j-1} \end{aligned}\end{equation} and \begin{equation}\label{eq:pp2} \begin{aligned} |\mathcal{FD}(1^{\fpf}_{2k} \times 2n \cdots 321)| &= \prod_{\substack{i,j \in [n], i\neq j }} \frac{i + j +2k-1}{i + j -1}. \end{aligned} \end{equation} The right-hand side of \eqref{eq:pp} is the number of reverse plane partitions with entries at most $k$ of shifted shape $\textsf{SD}_\lambda= \{ (i,i+j-1) : (i,j) \in \textsf{D}_\lambda\}$ for $\lambda= (p+q-1,p+q-3,p+q-5,\dots)$ \cite{proctor1983shifted}. Similar formulas should hold for $\|\mathcal{ID}(1^k \times y)\|$ and $|\mathcal{FD}(1^{\fpf}_{2k} \times z)|$ when $y \in\mathcal{I}_n$ and $z \in \I^{\fpf}_{2n}$ are any (fpf-)dominant involutions. We expect that one can prove such identities algebraically using the Pfaffian formulas for $\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$ in \cite[\S5]{Pawlowski}. A more interesting open problem is the following: \begin{problem} Find bijective proofs of \eqref{eq:pp} and \eqref{eq:pp2} and their dominant generalizations. \end{problem} \subsection{Ideals of matrix Schubert varieties} \label{sec:ideals} Another open problem is to find a geometric explanation for the formulas in Theorem~\ref{thm:inv-pipe-dream-formula}. Such an explanation exists in the double Schubert case, as we briefly explain. Recall that $A_{[i][j]}$ denotes the upper-left $i\times j$ submatrix of a matrix $A$. Let $\mathcal{Z}$ be the matrix of indeterminates $(z_{ij})_{i,j \in [n]}$. For $w \in S_n$, let $I_w \subseteq \mathbb{C}[z_{ij} : i,j \in [n]] = \mathbb{C}[\textsf{Mat}_n]$ be the ideal generated by all $(\rank w_{[i][j]}+1) \times (\rank w_{[i][j]}+1)$ minors of $\mathcal{Z}_{[i][j]}$ for $i,j \in [n]$. The vanishing locus of $I_w$ in the space $\textsf{Mat}_n$ of $n\times n$ complex matrices is exactly the matrix Schubert variety $M\hspace{-0.5mm}X_w$. Let $\init(I_w)$ be the \emph{initial ideal} of leading terms in $I_w$ with respect to any term order on $\mathbb{C}[z_{ij}]$ with the property that the leading term of $\det(A)$ for any submatrix $A$ of $\mathcal{Z}$ is the product of the antidiagonal entries of $A$. For instance, lexicographic order with the variable ordering $z_{1n} < \cdots < z_{11} < z_{2n} < \cdots < z_{21} < \cdots$ has this property. \begin{thm}[\cite{knutson-miller}] For each permutation $w \in S_n$, the ideal $I_w$ is prime, and there is a prime decomposition $ \init(I_w) = \bigcap_{D \in \mathcal{PD}(w)} (z_{ij} : (i,j) \in D) $. \end{thm} Given the description of the class $[M\hspace{-0.5mm}X_w]$ in \S \ref{subsec:eq-cohom}, this result implies the pipe dream formula \eqref{eq:double-schubert-def} for $[M\hspace{-0.5mm}X_w] = \mathfrak{S}_w(x,y)$. Now let $\hat{\mathcal{Z}}$ be the symmetric matrix of indeterminates $[z_{\max(i,j), \min(i,j)}]_{i,j \in [n]}$. Define $\hat I_y \subseteq \mathbb{C}[z_{ij} : 1 \leq j < i \leq n] = \mathbb{C}[\textsf{SMat}_n]$ for $y \in \mathcal{I}_n$ to be the ideal generated by all $(1+\rank y_{[i][j]}) \times (1+\rank y_{[i][j]})$ minors of $\hat{\mathcal{Z}}_{[i][j]}$ for $i,j \in [n]$. The vanishing locus of $\hat I_y$ is $M\hspace{-0.5mm}\hat X_y$. \begin{conj} \label{conj:ideals} For $y \in \mathcal{I}_n$, the ideal $\hat I_y$ is prime, and there is a primary decomposition of $\init(\hat I_y)$ whose top-dimensional components are $\left(z_{ij}^{\cwt{i}{j}{D}} : (i,j) \in D\right)$ for $D \in \mathcal{ID}(y)$, where $\cwt{i}{j}{D}=2$ if the pipes crossing at $(i,j)$ are labeled $p$ and $z(p)$ for some $p\in[n]$, and otherwise $\cwt{i}{j}{D}=1$. \end{conj} As per \S\ref{subsec:eq-cohom}, the conjecture would give a direct geometric proof of Theorem~\ref{thm:inv-pipe-dream-formula}. \begin{ex} Let $y = 1243 = (3,4) \in \mathcal{I}_4$. Then $A \in M\hspace{-0.5mm}\hat X_y$ if and only if $\rank A_{[i][j]} \leq m_{ij}$ for \begin{equation*} (m_{ij})_{1\leq i,j\leq 4} = \left(\begin{array}{cccc} 1 & 1 & 1 & 1\\ 1 & 2 & 2 & 2\\ 1 & 2 & 2 & 3\\ 1 & 2 & 3 & 4 \end{array}\right). \end{equation*} These rank conditions all follow from $\rank A_{[3][3]} \leq 2$, so $\hat I_y$ is generated by $\det \hat{\mathcal{Z}}_{[3][3]}$. The ideals in the primary decomposition $\init(\hat I_y) = (z_{31}^2 z_{22}) = (z_{31}^2) \cap (z_{22})$ correspond to the two involution pipe dreams in the set $ \mathcal{ID}(y) = \{ \{ (3,1)\}, \{(2,2)\}\} $ for $y=(3,4)$. \end{ex} \begin{ex} Let $y = 14523 = (2,4)(3,5) \in \mathcal{I}_5$. One computes that \begin{equation*} \init(\hat I_y) = (z_{21}^2, z_{31} z_{21}, z_{22}z_{31}, z_{31}^2, z_{32}z_{31}, z_{32}^2) = (z_{21}^2, z_{31}, z_{32}^2) \cap (z_{21}, z_{22}, z_{31}^2, z_{32}). \end{equation*} There is a single involution pipe dream for $y$ given by $\{(2,1),(3,1),(3,2)\}.$ This pipe dream corresponds to the codimension $3$ component $(z_{21}^2, z_{31}, z_{32}^2)$ of $\init(\hat I_y)$, while the codimension $4$ component $(z_{21}, z_{22}, z_{31}^2, z_{32})$ does not correspond to a pipe dream of $y$. \end{ex} For $z \in \I^{\fpf}_n$, the ideal generated by the $(\rank z_{[i][j]}+1) \times (\rank z_{[i][j]}+1)$ of a skew-symmetric matrix of indeterminates need not be prime, and we do not have an analogue of Conjecture~\ref{conj:ideals}. \subsection{Pipe dream formulas for $K$-theory} \label{sec:k-theory} A third source of open problems concerns pipe dreams for $K$-theory. A subset $D\subseteq [n]\times [n]$ is a \emph{$K$-theoretic pipe dream} for $w \in S_n$ if $\textsf{word}(D) = (a_1,a_2,\dots,a_p)$ satisfies $ s_{a_1} \circ s_{a_2} \circ \dots \circ s_{a_p} = w $ where $\circ$ is the Demazure product. Let $\mathcal{KPD}(w)$ denote the set of $K$-theoretic pipe dreams of $w$. Fomin and Kirillov \cite{FominKirillov94} introduce these objects in order to state this formula for the \emph{(generalized) Grothendieck polynomial} $\mathfrak{G}_w$ of a permutation $w\in S_n$: \begin{equation} \label{eq:groth} \mathfrak{G}_w = \sum_{D \in \mathcal{KPD}(w)} \beta^{|D| - \ell(w)} \prod_{(i,j) \in D} x_i \in \mathbb{Z}[\beta][x_1,x_2,\dots,x_n]. \end{equation} This identity is nontrivial to derive from Lascoux and Sch\"utzenberger's original definition of Grothendieck polynomials in terms of isobaric divided difference operators~\cite{lascoux1990anneau,lascoux1983symmetry}. Grothendieck polynomials becomes Schubert polynomials on setting $\beta=0$. In \cite{MP2019}, continuing work of Wyser and Yong \cite{wyser-yong-orthogonal-symplectic}, the second two authors studied \emph{orthogonal} and \emph{symplectic Grothendieck polynomials} $\mathfrak{G}^{\O}_y$ and $\mathfrak{G}^{\mathsf{Sp}}_z$ indexed by $y \in \mathcal{I}_n$ and $z \in \I^{\fpf}_n$. These polynomials likewise recover the involution Schubert polynomials $\hat{\mathfrak{S}}_y$ and $\hat{\mathfrak{S}}^{\fpf}_z$ on setting $\beta=0$, and it would be interesting to know if they have analogous pipe dream formulas. The symplectic case of this question is more tractable. The polynomials $\mathfrak{G}^{\mathsf{Sp}}_z$ have a formulation in terms of isobaric divided difference operators due to Wyser and Yong \cite{wyser-yong-orthogonal-symplectic}, which suggests a natural $K$-theoretic variant of the set $\mathcal{FD}(z)$. A formula for $\mathfrak{G}^{\mathsf{Sp}}_z$ involving these objects appears in \cite[\S4]{MP2020}. By contrast, no simple algebraic formula is known for the polynomials $\mathfrak{G}^{\O}_y$. It is a nontrivial problem even to identify the correct $K$-theoretic generalization of $\mathcal{ID}(y)$. \begin{problem} Find a pipe dream formula for the polynomials $\mathfrak{G}^{\O}_y$ involving an appropriate ``$K$-theoretic'' generalization of the sets of involution pipe dreams $\mathcal{ID}(y)$. \end{problem}

https://arxiv.org/abs/1410.6535

A New Fractional Derivative with Classical Properties

We introduce a new fractional derivative which obeys classical properties including: linearity, product rule, quotient rule, power rule, chain rule, vanishing derivatives for constant functions, the Rolle's Theorem and the Mean Value Theorem. The definition, \[ D^\alpha (f)(t) = \lim_{\epsilon \rightarrow 0} \frac{f(te^{\epsilon t^{-\alpha}}) - f(t)}{\epsilon}, \] is the most natural generalization that uses the limit approach. For $0\leq \alpha < 1$, it generalizes the classical calculus properties of polynomials. Furthermore, if $\alpha = 1$, the definition is equivalent to the classical definition of the first order derivative of the function $f$. Furthermore, it is noted that there are $\alpha-$differentiable functions which are not differentiable.

\section{Introduction} The derivative of non-integer order has been an interesting research topic for several centuries. The idea was motivated by the question, ``What does it mean by $\frac{d^n f}{dx^n}$, if $n=\frac{1}{2}$ ?'', asked by L'Hospital in 1695 in his letters to Leibniz \cite{letter1,letter2, letter3}. Since then, the mathematicians tried to answer this question for centuries in several points of view. The outcomes are many folds. Various types of fractional derivatives were introduced: Riemann-Liouville, Caputo, Hadamard, Erd\'{e}lyi-Kober, Gr\"{u}nwald-Letnikov, Marchaud and Riesz are just a few to name \cite{udita1,udita2,key-9,key-8,key-2,key-12}. Most of the fractional derivatives are defined via fractional integrals \cite{key-9}. Due to the same reason, those fractional derivatives inherit some non-local behaviors, which lead them to many interesting applications including memory effects and future dependence \cite{what}. Among the inconsistencies of the existing fractional derivatives are: \begin{enumerate} \item[(1)] Most of the fractional derivatives except Caputo-type derivatives, do not satisfy $D_a^\alpha(1) = 0$ , if $\alpha$ is not a natural number. \item[(2)] All fractional derivatives do not obey the familiar Product Rule for two functions: \[ D_a^\alpha (fg) = fD_a^\alpha (g) + gD_a^\alpha(f). \] \item[(3)] All fractional derivatives do not obey the familiar Quotient Rule for two functions: \[ D_a^\alpha \Big(\frac{f}{g}\Big) = \frac{gD_a^\alpha (f) - fD_a^\alpha(g)}{g^2}. \] \item[(4)] All fractional derivatives do not obey the Chain Rule: \[ D_a^\alpha (f \circ g)(t) = f^{(\alpha)}\big(g(t)\big) \, g^{(\alpha)}(t). \] \item[(5)] Fractional derivatives do not have a corresponding Rolle's Theorem. \item[(6)] Fractional derivatives do not have a corresponding Mean Value Theorem. \item[(7)] All fractional derivatives do not obey: $D_a^\alpha D_a^\beta f = D_a^{\alpha + \beta} f$, in general. \item[(8)] The Caputo definition assumes that the function $f$ is differentiable. \end{enumerate} To overcome some of these and other difficulties, Khalil et al. \cite{Khalil}, came up with an interesting idea that extends the familiar limit definition of the derivative of a function given by the following. \begin{definition}\cite{Khalil} Let $f:[0, \infty) \rightarrow \mathbb{R}$ and $t > 0$. Then the ``conformable fractional derivative'' of $f$ of order $\alpha$ is defined by, \begin{equation} f^{(\alpha)}(t) = \lim_{\epsilon \rightarrow 0} \frac{f(t + \epsilon t^{1-\alpha}) - f(t)}{\epsilon}, \label{def1} \end{equation} for $t >0, \; \alpha \in (0, 1)$. If $f$ is $\alpha-$differentiable in some $(0, a), \; a >0$, and $\lim_{t \rightarrow 0^+} f^{(\alpha)}(t)$ exists, then define \[ f^{(\alpha)}(0) = \lim_{t \rightarrow 0^+} f^{(\alpha)}(t). \] \end{definition} As a consequence of the above definition, the authors in \cite{Khalil}, showed that the $\alpha-$derivative in (\ref{def1}), obeys the Product rule, Quotient rule and has results similar to the Rolle's Theorem and the Mean Value Theorem in classical calculus. The purpose of this work is to further generalize the results obtained in \cite{Khalil} and introduce a new fractional derivative as the most natural extension of the familiar limit definition of the derivative of a function $f$ at a point. \section{New Fractional Derivative} In this section, we give the main definition of the paper and obtain several results that are close resemblance of the results found in classical calculus. We prove that the new fractional derivative satisfies the Product rule, Quotient rule, Chain rule and have results which are natural extensions of the Rolle's theorem and the Mean Value Theorem. To this end, we start with the following definition, which is a generalization of the limit definition of the derivative. \begin{definition} Let $f:[0, \infty) \rightarrow \mathbb{R}$ and $t > 0$. Then the \emph{fractional derivative} of $f$ of order $\alpha$ is defined by, \begin{equation} \mathcal{D}^\alpha (f)(t) = \lim_{\epsilon \rightarrow 0} \frac{f(te^{\epsilon t^{-\alpha}}) - f(t)}{\epsilon}, \label{df} \end{equation} for $t >0, \; \alpha \in (0, 1)$. If $f$ is $\alpha-$differentiable in some $(0, a), \; a >0$, and $\lim_{t \rightarrow 0^+} \mathcal{D}^\alpha (f)(t)$ exists, then define \[ \mathcal{D}^\alpha (f)(0) = \lim_{t \rightarrow 0^+} \mathcal{D}^\alpha (f)(t). \] \end{definition} The first result is the generalization of Theorem 2.1 of \cite{Khalil}. \begin{theorem} If a function $f:[0, \infty) \rightarrow \mathbb{R}$ is $\alpha-$differentiable at $a >0, \; \alpha \in (0, 1]$, then $f$ is continuous at $a$. \end{theorem} \begin{proof} Since $f(ae^{\epsilon a^{-\alpha}})-f(a) = \frac{f(ae^{\epsilon a^{-\alpha}})-f(a)}{\epsilon} \epsilon$, we have \[ \lim_{\epsilon \rightarrow 0} \big[f(ae^{\epsilon a^{-\alpha}})-f(a)\big] = \lim_{\epsilon \rightarrow 0} \frac{f(ae^{\epsilon a^{-\alpha}})-f(a)}{\epsilon} .\lim_{\epsilon \rightarrow 0}\epsilon. \] Let $h= \epsilon a^{1-\alpha} + O(\epsilon^2)$. Then, $\lim_{h \rightarrow 0} \big[f(a+h)-f(a)\big] = \mathcal{D}^\alpha (f)(a) \cdot 0 = 0$, which, in turn, implies that $\lim_{h \rightarrow 0} f(a+h) = f(a)$. This completes the proof. \end{proof} The following is the main result of this paper. \vspace{.2cm} \begin{theorem} \label{thm3} Let $\alpha \in (0, 1]$ and $f, g$ be $\alpha-$differentiable at a point $t>0$. Then, \begin{enumerate} \item[(1)] $\mathcal{D}^\alpha \big[af +bg\big] = a\mathcal{D}^\alpha (f) + b\mathcal{D}^\alpha (g)$, for all $a, b \in \mathbb{R}$. \item[(2)] $\mathcal{D}^\alpha \big(t^n \big) = nt^{n-\alpha}$ for all $n \in \mathbb{R}$. \item[(3)] $\mathcal{D}^\alpha \big(C \big) = 0,$ for all constant functions, $f(t) = C.$ \item[(4)] $\mathcal{D}^\alpha \big(fg) = f\mathcal{D}^\alpha (g) + g\mathcal{D}^\alpha (f)$. \item[(5)] $\mathcal{D}^\alpha \big(\frac{f}{g}) = \frac{g\mathcal{D}^\alpha (f) - f\mathcal{D}^\alpha (g)}{g^2}$. \item[(6)] $\mathcal{D}^\alpha (f \circ g)(t) = f^\prime\big(g(t)\big) \, \mathcal{D}^\alpha g(t)$, \;\mbox{for} $f$ \mbox{differentiable at }$g(t)$. \item[(7)] If, in addition, $f$ is differentiable, then $\mathcal{D}^\alpha \big(f)(t) = t^{1-\alpha}\frac{df}{dt}(t).$ \end{enumerate} \end{theorem} \begin{proof} Part (1) and (3) follow directly from the definition. We shall prove (2), (4), (6) and (7), since the proofs are different from what appear in \cite{Khalil}. Now, for fixed $\alpha \in (0, 1]$, $n \in \mathbb{R}$ and $t>0$, we have \begin{align} \mathcal{D}^\alpha \big(t^n\big) &= \lim_{\epsilon \rightarrow 0} \frac{(te^{\epsilon t^{-\alpha}})^n- t^n}{\epsilon} \nonumber \\ &= t^n\lim_{\epsilon \rightarrow 0} \frac{e^{n\epsilon t^{-\alpha}}-1}{\epsilon} \nonumber \\ &= t^n\lim_{\epsilon \rightarrow 0} \frac{\epsilon n t^{-\alpha}+\frac{\epsilon^2 n^2 t^{-2\alpha}}{2!} +\cdots}{\epsilon} \nonumber \\ &= t^n\lim_{\epsilon \rightarrow 0} \frac{\epsilon n t^{-\alpha}+ O(\epsilon^2)}{\epsilon} \label{eq1} \\ &=nt^{n-\alpha}. \nonumber \end{align} This completes the proof of (2). Then, we shall prove (4). To this end, since $f, g$ are $\alpha-$differentiable at $t >0,$ note that, \begin{align} \mathcal{D}^\alpha \big(fg\big)(t) &= \lim_{\epsilon \rightarrow 0} \frac{f(te^{\epsilon t^{-\alpha}})g(te^{\epsilon t^{-\alpha}})-f(t)g(t)}{\epsilon} \nonumber \\ &= \lim_{\epsilon \rightarrow 0} \frac{f(te^{\epsilon t^{-\alpha}})g(te^{\epsilon t^{-\alpha}})-f(t)g(te^{\epsilon t^{-\alpha}})+f(t)g(te^{\epsilon t^{-\alpha}})-f(t)g(t)}{\epsilon} \nonumber \\ &= \lim_{\epsilon \rightarrow 0} \Big[\frac{f(te^{\epsilon t^{-\alpha}})-f(t)}{\epsilon}\cdot g(te^{\epsilon t^{-\alpha}})\Big] + f(t)\lim_{\epsilon \rightarrow 0} \frac{g(te^{\epsilon t^{-\alpha}})-g(t)}{\epsilon} \nonumber \\ &= \mathcal{D}^\alpha \big(f\big)(t)\lim_{\epsilon \rightarrow 0} g(te^{\epsilon t^{-\alpha}}) +f(t)\mathcal{D}^\alpha \big(g\big)(t). \nonumber \end{align} Since $g$ is continuous at $t$, $\lim_{\epsilon \rightarrow 0} g(te^{\epsilon t^{-\alpha}}) = g(t).$ This completes the proof of (4). The proof of (5) is similar and is left for the reader to verify. Next, we prove (6) in two different approaches. First, suppose $u=g(t)$ is $\alpha-$differentiable at a point $a>0$ and $y=f(u)$ is \textsl{differentiable} at a point $b=g(a)>0$. Let $\epsilon>0$, and $\Delta y = f(be^{\epsilon b^{-\alpha}})-f(b)$. Since $be^{\epsilon b^{-\alpha}} = b + \epsilon b^{1-\alpha}+O(\epsilon^2) = b+\Delta u$, where $\Delta u = \epsilon b^{1-\alpha}+O(\epsilon^2)$, we have \begin{equation} \Delta y = D^1 f(b) \Delta u + \epsilon_1 \Delta u \label{eq2} \end{equation} where $\epsilon_1 \rightarrow 0$ as $\Delta u \rightarrow 0$. Thus, $\epsilon_1$ is a continuous function of $\Delta u$ if we define $\epsilon_1$ to be $0$ when $\Delta u =0$. Now, if $\Delta t$ is an increment in $t$ and $\Delta u $ and $\Delta y$ (with the possibility of both being equal to $0$) are corresponding increments in $u$ and $y$, respectively. Then we may write, using Equation (\ref{eq2}), \begin{equation} \Delta u = \mathcal{D}^\alpha g(a) \Delta t + \epsilon_2 \Delta t \label{eq3} \end{equation} where $\epsilon_2 \rightarrow 0$ as $\Delta t \rightarrow 0$, and \begin{align} \Delta y &= \Big[D^1 f(b)+ \epsilon_1 \Big] \Delta u \nonumber \\ &= \Big[D^1 f(b) + \epsilon_1 \Big] \cdot \Big[\mathcal{D}^\alpha g(a) + \epsilon_2 \Big] \Delta t \nonumber \end{align} where both $\epsilon_1 \rightarrow 0$ and $\epsilon_2 \rightarrow 0$ as $\Delta t \rightarrow 0$. Taking $\Delta t = \epsilon$, we now have, \begin{align*} \mathcal{D}^\alpha (f \circ g)(t) &= \lim_{\epsilon \rightarrow 0} \frac{\Delta y}{\epsilon} \\ &= \lim_{\Delta t \rightarrow 0} \Big[D^1 f(b) + \epsilon_1 \Big] \cdot \Big[\mathcal{D}^\alpha g(a) + \epsilon_2 \Big] \\ &=D^1 f(b)\mathcal{D}^\alpha g(a) = f^\prime(g(a))\mathcal{D}^\alpha g(a). \end{align*} It can be seen that Equation~\ref{eq3}, can not be written in the form of Equation~\ref{eq2}. This makes it necessitates to use the $\alpha-$derivative in Equation~\ref{eq3}. Now, we prove the result following a standard limit-approach. To this end, in one hand, if the function $g$ is constant in a neighborhood containing $a$, then $\mathcal{D}^\alpha \big(f\circ g\big)(a) = 0$. On the other hand, assume that the function $g$ is non-constant in the neighborhood of $a$. In this case, we can find an $\epsilon_0 >0$ such that, $g(x_1) \neq g(x_2)$ for any $x_1, x_2 \in (a-\epsilon_0, a+\epsilon_0)$. Now, since $g$ is continuous at $a$, for $\epsilon$ sufficiently small, we have \begin{align*} \mathcal{D}^\alpha \big(f\circ g\big)(a) &= \lim_{\epsilon \rightarrow 0} \frac{f\big(g(ae^{\epsilon a^{-\alpha}})\big)-f\big(g(t)\big)}{\epsilon} \\ &=\lim_{\epsilon \rightarrow 0} \frac{f\big(g(ae^{\epsilon a^{-\alpha}})\big)-f\big(g(a)\big)}{g(ae^{\epsilon a^{-\alpha}})-g(a)}\cdot \frac{g(ae^{\epsilon a^{-\alpha}})-g(a)}{\epsilon} \\ &=\lim_{\epsilon \rightarrow 0} \frac{f\big(g(a)+\epsilon_1\big)-f\big(g(a)\big)}{\epsilon_1}\cdot \frac{g(ae^{\epsilon a^{-\alpha}})-g(a)}{\epsilon}, \; \mbox{ where } \; \epsilon_1 \rightarrow 0 \mbox{ as } \epsilon \rightarrow 0, \\ &=\lim_{\epsilon_1 \rightarrow 0} \frac{f\big(g(a)+\epsilon_1\big)-f\big(g(a)\big)}{\epsilon_1} \cdot \lim_{\epsilon \rightarrow 0} \frac{g(ae^{\epsilon a^{-\alpha}})-g(a)}{\epsilon} \\ &=f^\prime\big(g(a)\big) \mathcal{D}^\alpha g(a), \end{align*} for $a >0$. This establishes the Chain Rule. To prove part (7), we use a similar line of argument as in Equation (\ref{eq1}). Thus, taking $h=\epsilon t^{1-\alpha} \big(1+ O(\epsilon)\big)$, we have \begin{align*} \mathcal{D}^\alpha (f)(t) &= \lim_{\epsilon \rightarrow 0} \frac{f(te^{\epsilon t^{-\alpha}})-f(t)}{\epsilon} \\ &= \lim_{\epsilon \rightarrow 0} \frac{f(t+\epsilon t^{1-\alpha}+O(\epsilon^2))-f(t)}{\epsilon} \\ &= \lim_{\epsilon \rightarrow 0} \frac{f(t+h)-f(t)}{\frac{ht^{\alpha -1}}{1+O(\epsilon)}} \\ &=t^{1-\alpha}\frac{df}{dt}(t), \end{align*} \noindent since, by the assumption, $f$ is differentiable at $t>0$. This completes the proof of the theorem. \end{proof} \begin{remark} The result similar to (6) does not appear in \cite{Khalil}. The argument used in Equation (\ref{eq1}) was/will be utilized in several occasions in the paper. \end{remark} The following theorem lists $\alpha-$fractional derivative of several familiar functions. The results are identical to that of conformable fractional derivative discussed in \cite{Khalil}, which will be shown to be a special case of $\alpha-$fractional derivative defined in (\ref{df}). \begin{theorem} \label{thm4} Let $a, n \in \mathbb{R}$ and $\alpha \in (0, 1]$. Then, we have the following results. \begin{enumerate} \item[(a)] $\mathcal{D}^\alpha (t^n) = n t^{n-\alpha}$. \item[(b)] $\mathcal{D}^\alpha (1) = 0.$ \item[(c)] $\mathcal{D}^\alpha (e^{ax}) = ax^{1-\alpha}e^{ax}.$ \item[(d)] $\mathcal{D}^\alpha (\sin ax) = ax^{1-\alpha}\cos ax.$ \item[(e)] $\mathcal{D}^\alpha (\cos ax) = -ax^{1-\alpha}\sin ax.$ \item[(f)] $\mathcal{D}^\alpha (\frac{1}{\alpha}t^\alpha) = 1.$ \end{enumerate} \end{theorem} It is easy to see from part (6) of Theorem~\ref{thm3} that we have rather unusual results given in the next theorem. \begin{theorem} \label{thm5} Let $\alpha \in (0, 1]$ and $t >0$. Then, \begin{enumerate} \item[(i)] $\mathcal{D}^\alpha (\sin \frac{1}{\alpha}t^\alpha ) = \cos \frac{1}{\alpha}t^\alpha.$ \item[(ii)] $\mathcal{D}^\alpha (\cos \frac{1}{\alpha}t^\alpha) = -\sin \frac{1}{\alpha}t^\alpha.$ \item[(iii)] $\mathcal{D}^\alpha (e^{\frac{1}{\alpha}t^\alpha}) = e^{\frac{1}{\alpha}t^\alpha}.$ \end{enumerate} \end{theorem} Theorem \ref{thm5} suggests that there is a pseudo-invariant space corresponding to the $\alpha-$fractional derivative on which the $sine$, $cosine$ and the exponential function, $e^x$ behave as they are having familiar classical derivatives. Due to the Fourier theory, this in turn, implies that any function which possesses a Fourier expansion behaves as having classical derivatives. This would be an interesting topic for further research. \begin{figure}[h] \centering \subfloat[$\nu$= 2.0]{\includegraphics[height = 6.0cm,width=6.3cm, trim = 0 20 0 0]{v-1-2.eps}} \subfloat[$\nu$= 1.0]{\includegraphics[height=5.7cm,width=6cm, trim=0 0 0 25]{v-2-2.eps}} \caption{$\frac{1}{2}-$fractional derivatives of power function $f(x)=x^\nu$.} \label{fig:deri} \end{figure} In \cite{udita1,udita2}, the author introduced fractional integrals and derivatives, which generalize the Riemann-Liouville and the Hadamard fractional integrals and derivatives to a single form. In the next subsection, we show that the results obtained by Khalil et at. \cite{Khalil} are special cases of the more general extension discussed in this paper. \subsection{Further Generalizations} Let us first define a truncated exponential function given by, \[ e_k^x = \sum_{i=0}^k \frac{x^i}{i!} \label{e-k} \] With the help of this definition, we will define another fractional derivative given below. \begin{definition} Let $f:[0, \infty) \rightarrow \mathbb{R}$ and $t > 0$. Then the \emph{fractional derivative} of $f$ of order $\alpha$ is defined by, \begin{equation} \mathcal{D}_k^\alpha (f)(t) = \lim_{\epsilon \rightarrow 0} \frac{f(te_k^{\epsilon t^{-\alpha}}) - f(t)}{\epsilon}, \label{def2} \end{equation} for $t >0, \; \alpha \in (0, 1)$. If $f$ is $\alpha-$differentiable in some $(0, a), \; a >0$, and $\lim_{t \rightarrow 0^+} \mathcal{D}_k^\alpha (f)(t)$ exists, then define \[ \mathcal{D}_k^\alpha (f)(0) = \lim_{\epsilon \rightarrow 0} \mathcal{D}_k^\alpha (f)(t). \] \end{definition} Now, It is easy to see that \begin{equation} \mathcal{D}_1^\alpha (f)(t) = \lim_{\epsilon \rightarrow 0} \frac{f(t + \epsilon t^{1-\alpha}) - f(t)}{\epsilon}, \label{ndef1} \end{equation} and \begin{equation} \mathcal{D}_\infty^\alpha (f)(t) = \lim_{\epsilon \rightarrow 0} \frac{f(te^{\epsilon t^{-\alpha}}) - f(t)}{\epsilon}, \label{ndef2} \end{equation} where the expression on the right-hand-side of (\ref{ndef1}) is the conformable fractional derivative defined in (\ref{def1}) in \cite{Khalil}, while the one on (\ref{ndef2}) is the $\alpha-$fractional derivative defined in this paper. Furthermore, if $\alpha = 1$ the Equation (\ref{ndef1}) becomes the classical definition of the first derivative of a function $f$ at a point $t$. While, noting that $te^{\epsilon t^{-\alpha}} = t+ \epsilon +O(\epsilon^2)$, we may also show that the Equation (\ref{ndef2}) is equivalent to the classical definition of the first derivative of a function $f$. These observations further suggest that there are corresponding results similar to the Rolle's theorem and the Mean Value theorem. Next, we give those two results. \begin{theorem}[Rolle's theorem for $\alpha-$Fractional Differentiable Functions] Let $a>0$ and $f:[a, b] \rightarrow \mathbb{R}$ be a function with the properties that \begin{enumerate} \item[(1)] $f$ is continuous on $[a, b]$, \item[(2)] $f$ is $\alpha-$differentiable on $(a, b)$ for some $\alpha \in (0, 1)$, \item[(3)] $f(a)=f(b)$. \end{enumerate} Then, there exists $c\in(a,b)$, such that $\mathcal{D}^\alpha (f)(c) = 0.$ \end{theorem} \begin{proof} We prove this using contradiction. Since $f$ is continuous on $[a, b]$ and $f(a)=f(b)$, there is $c\in(a, b)$, at which the function has a local extrema. Then, \[ \mathcal{D}^\alpha f(c)=\lim_{\epsilon \rightarrow 0^-} \frac{f(ce^{\epsilon c^{-\alpha}}) - f(c)}{\epsilon}=\lim_{\epsilon \rightarrow 0^+} \frac{f(ce^{\epsilon c^{-\alpha}}) - f(c)}{\epsilon}. \] But, the two limits have opposite signs. Hence, $\mathcal{D}^\alpha f(c) =0.$ \end{proof} \begin{theorem}[Mean Value Theorem for $\alpha-$Fractional Differentiable Functions] Let $a>0$ and $f:[a, b] \rightarrow \mathbb{R}$ be a function with the properties that \begin{enumerate} \item[(1)] $f$ is continuous on $[a, b]$, \item[(2)] $f$ is $\alpha-$differentiable on $(a, b)$ for some $\alpha \in (0, 1)$, \end{enumerate} Then, there exists $c\in(a,b)$, such that $\mathcal{D}^\alpha (f)(c) = \frac{f(b)-f(a)}{\frac{1}{\alpha}b^\alpha-\frac{1}{\alpha}a^\alpha}.$ \end{theorem} \begin{proof} Consider the function, \[ g(x) = f(x) -f(a) - \frac{f(b)-f(a)}{\frac{1}{\alpha}b^\alpha-\frac{1}{\alpha}a^\alpha}\Big(\frac{1}{\alpha}x^\alpha-\frac{1}{\alpha}a^\alpha\Big). \] Then, the function $g$ satisfies the conditions of the fractional Rolle's theorem. Hence, there exists $c\in(a,b)$, such that $\mathcal{D}^\alpha (g)(c) =0$. Using the fact that $\mathcal{D}^\alpha (\frac{1}{\alpha}x^\alpha) =1$, the result follows. \end{proof} As a result of the fractional Mean Value Theorem, we also have the following proposition. \begin{proposition} Let $f:[a, b] \rightarrow \mathbb{R}$ be $\alpha-$differentiable for some $\alpha \in (0, 1)$. Suppose that $\mathcal{D}^\alpha f $ is bounded on $[a, b]$. Then, if \begin{enumerate} \item[(a)] $a>0$, or \item[(b)] $\mathcal{D}^\alpha f $ is continuous at either $a$ or $b$, \end{enumerate} the function $f$ is uniformly continuous on $[a, b]$, and hence $f$ is bounded. \end{proposition} Figure 1 depicts the graphs of the $\alpha-$fractional derivatives of the power function $f(x)=x^\nu$ for $\nu = 1$ and $2$. It is also interesting to note that the function $f(x) = \sqrt{x}$ has a constant derivative equal to $\frac{1}{2}$, and not equal to $\sqrt{\pi}/2$ as in the case of the Riemann-Liouville derivative. This also suggests that the new derivative should have a rather different interpretation in the geometrical sense, which has yet to be discovered. Further, it can be seen that a function could be $\alpha-$differential at a point but not differentiable in the classical sense. (cf. page 67 of \cite{Khalil}). For example, the function $f(t) = 3t^{\frac{1}{3}}$ is everywhere $\frac{1}{3}-$differentiable and has the derivative of $1$ while it is not differentiable at $t=0$. The similar results are there for the Riemann-Liouville and the Caputo derivatives. Next, we consider the possibility of $\alpha \in (n, n+1]$, for some $n\in \mathbb{N}$. We have the following definition. \begin{definition} \label{def5} Let $\alpha \in (n, n+1]$, for some $n\in \mathbb{N}$ and $f$ be an $n-$differentiable at $t >0$. Then the $\alpha-$fractional derivative of $f$ is defined by \[ \mathcal{D}^\alpha f (t) = \lim_{\epsilon \rightarrow 0} \frac{f^{(n)}(te^{\epsilon t^{n-\alpha}})-f^{(n)}(t)}{\epsilon}, \] if the limit exists. \end{definition} As a direct consequence of Definition \ref{def5} and part (7) of theorem \ref{thm3}, we can show that $\mathcal{D}^\alpha f (t) = t^{n+1-\alpha}f^{(n+1)}(t)$, where $\alpha \in (n, n+1]$ and $f$ is $(n+1)-$differentiable at $t>0$. Now, the question whether the fractional derivative $\mathcal{D}^\alpha f$ has a corresponding $\alpha-$fractional integral will be answered next. For simplicity, we shall only consider the class of continuous functions here. The results can easily be extended to more general settings and will be discussed in forthcoming papers. \section{Fractional Integral} As in the works of \cite{Khalil}, it is interesting to note that, in spite of the variation of the definitions of the fractional derivatives, we can still adopt the same definition of the fractional integral here due to the fact that we obtained similar results in Theorem \ref{thm4} as of the results (1)--(6) and (i)--(iii) in \cite{Khalil}. So, we have the following definition. \begin{definition}[Fractional Integral] Let $a\geq 0$ and $t \geq a$. Also, let $f$ be a function defined on $(a, t]$ and $\alpha \in \mathbb{R}$. Then, the \emph{$\alpha-$fractional integral} of $f$ is defined by, \begin{equation} \mathcal{I}^\alpha_a (f)(t) = \int_a^t \frac{f(x)}{x^{1-\alpha}} \, dx \label{int1} \end{equation} if the Riemann improper integral exists. \end{definition} It is interesting to observe that the $\alpha-$fractional derivative and the $\alpha-$fractional integral are inverse of each other as given in the next result. \begin{theorem}[Inverse property] Let $a \geq 0$, and $\alpha \in (0, 1)$. Also, let $f$ be a continuous function such that $\mathcal{I}^\alpha_a f$ exists. Then \[ \mathcal{D}^\alpha\Big(\mathcal{I}^\alpha_a (f)\Big) (t) = f(t), \quad \quad \mbox{for} \;\; t \geq a. \] \end{theorem} \begin{proof} The proof is a direct consequence of the fundamental theorem of calculus. Since $f$ is continuous, $\mathcal{I}^\alpha_a f$ is clearly differentiable. Therefore, using part (7) of theorem \ref{thm3}, we have \begin{align*} \mathcal{D}^\alpha\Big(\mathcal{I}^\alpha_a (f)\Big) (t) &= t^{\alpha -1}\frac{d}{dt} \mathcal{I}^\alpha_a (f) (t), \\ &= t^{\alpha -1}\frac{d}{dt}\int_a^t \frac{f(x)}{x^{1-\alpha}} \, dx, \\ &= t^{\alpha -1}\frac{f(t)}{t^{1-\alpha}}, \\ &= f(t). \end{align*} This completes the proof. \vspace{-.15cm} \end{proof} The applications of this $\alpha-$integral and derivative will be discussed in a forthcoming paper. Several interesting applications of a similar fractional derivative defined in (\ref{def1}) and the corresponding fractional integral appear in \cite{Khalil}, the major reference of this paper. We conclude the paper with the following questions, which have yet to be answered. \begin{enumerate} \item[(i)] What is the physical meaning or geometric interpretation of the $\alpha-$derivative? \item[(ii)]Is there a similarity between the $\alpha-$derivative and the \emph{G\^{a}teaux derivative}? \item[(iii)]One of the limitations of this version of the fractional derivative is that it assumes that the variable $t >0$. So the question is whether we can relax this condition on a special class of functions, if so, what is it? \end{enumerate} \bibliographystyle{abbrv}

https://arxiv.org/abs/1810.11932

Computing discrete equivariant harmonic maps

We present effective methods to compute equivariant harmonic maps from the universal cover of a surface into a nonpositively curved space. By discretizing the theory appropriately, we show that the energy functional is strongly convex and derive convergence of the discrete heat flow to the energy minimizer, with explicit convergence rate. We also examine center of mass methods, after showing a generalized mean value property for harmonic maps. We feature a concrete illustration of these methods with Harmony, a computer software that we developed in C++, whose main functionality is to numerically compute and display equivariant harmonic maps.

\subsection*{Acknowledgments} \addcontentsline{toc}{section}{Acknowledgments} The authors wish to thank Tarik Aougab, Benjamin Beeker, Alexander Bobenko, David Dumas, John Loftin, Jean-Marc Schlenker, Nicolas Tholozan, and Richard Wentworth for valuable conversations and correspondences pertaining to this work. We especially thank David Dumas for his extensive advice and support with the mathematical content and the development of \texttt{Harmony}{}, and Nicolas Tholozan for sharing key ideas that contributed to \autoref{thm:StrongConvexityEnergyGraphGeneral} and \autoref{thm:GeneralizedMeanValueProperty}. The first two authors gratefully acknowledge research support from the NSF Grant DMS1107367 \emph{RNMS: GEometric structures And Representation varieties} (the \emph{GEAR Network}). The third author was partially supported by the Portuguese Science Foundation FCT trough grant PTDC/MAT-STA/0975/2014 \emph{From Stochastic Geometric Mechanics to Mass Transportation Problems}. \section{Center of mass methods} \label{sec:CenterOfMass} In this section we investigate a center of mass algorithm towards the minimization of the discrete energy. We shall see that it is in some sense a variant of the fixed stepsize discrete heat flow. In the current state of our software \texttt{Harmony}{}, it is the most effective method (see \autoref{subsec:ExperimentalComparison}). First we recall facts about centers of mass in Riemannian manifolds and investigate how they relate to harmonic maps. We shall prove in particular a generalized mean value property for harmonic maps between Riemannian manifolds (\autoref{thm:GeneralizedMeanValueProperty}). \subsection{Centers of mass in metric spaces and Riemannian manifolds} \label{subsec:CenterOfMass} Let $(\Omega, \cF, \mu)$ be a probability space, $(X,d)$ be a metric space, and $h \colon \Omega \to X$ a measurable map. \begin{definition} \label{def:CenterOfMass} A \emph{center of mass} (or \emph{barycenter}) of $h$ is a minimizer of the function \begin{equation} \label{eq:DefinitionCenterOfMass} \begin{split} P_h \colon X & \to \bR\\ x & \mapsto \frac{1}{2} \int_\Omega d(x,h(y))^2 \, \upd \mu(y)~. \end{split} \end{equation} \end{definition} In general, neither existence nor uniqueness of centers of mass hold. If $X$ is a Hadamard space and $h \in \upL^2(\Omega, X)$ then existence and uniqueness do hold \cite[Lemma 2.5.1]{MR1266480}. For Riemannian manifolds we have: \begin{theorem}[Karcher \cite{MR0442975}] \label{thm:Karcher} Assume that $X$ is a complete Riemannian manifold and $h$ takes values in a ball $B = B(x_0, r) \subset X$ such that: \begin{itemize} \item $B$ is \emph{strongly convex}: any two points of $B$ are joined by a unique minimal geodesic $\gamma \colon [0,1] \to X$, and each such geodesic maps entirely into $B$. \item $B$ has nonpositive sectional curvature, or $r < \frac{\pi}{4 \sqrt{K}}$ where $K > 0$ is an upper bound for the sectional curvature in $B$. \end{itemize} Under these conditions, the function $P_h$ of $\eqref{eq:DefinitionCenterOfMass}$ only has interior minimimizers on $\bar{B}$ and is strongly convex inside $B$. Consequently, existence and uniqueness of the center of mass hold. \end{theorem} Note that if $X$ is a complete Riemannian manifold, any sufficiently small $r>0$ meets the requirements of \autoref{thm:Karcher}. The center of mass of $h$ is the unique point $G \in X$ such that \begin{equation} \label{eq:ImplicitEquationForCenterOfMass} \int_\Omega \exp_G^{-1}(h(y)) \, \upd \mu(y) = 0~. \end{equation} This simply expresses the vanishing at $G$ of the gradient of the function $P_h \colon X \to \bR$ of \eqref{eq:DefinitionCenterOfMass}. Of course, \autoref{def:CenterOfMass} generalizes the usual notion of center of mass in $\bR^n$: when $h\in\upL^2(\Omega,\bR^n)$, the center of mass given by $G = \int_\Omega h(y) \, \upd \mu(y)$. When $X$ is not Euclidean, the center of mass is only defined implicitly, but one can estimate proximity to the center of mass: \begin{lemma} \label{lem:ProximityToCenterOfMass} Assume that the conditions of \autoref{thm:Karcher} are satisfied. In particular the center of mass $G$ of $h$ is well-defined. If $G'$ is a point in $X$ such that: \begin{equation} \left\Vert \int_\Omega \exp_{G'}^{-1}(h(y)) \, \upd \mu(y) \right\Vert \leqslant \delta \end{equation} then \begin{equation} d(G, G') \leqslant C \delta~, \end{equation} where $C = 1$ when $X$ has nonpositive curvature, or $C = C(K, r) > 0$ when $X$ has sectional curvature bounded above by $K > 0$ in a strongly convex ball of radius $r$ containing the image of $h$. \end{lemma} \begin{proof} This is an immediate consequence of the fact that the function $P_h$ of \eqref{eq:DefinitionCenterOfMass} is $C$-strongly convex under the assumptions of the lemma. Also see \cite[Thm 1.5]{MR0442975}. \end{proof} \subsection{Generalized mean value property} \label{subsec:GeneralizedMeanValueProperty} In this section we show a generalized mean value property for smooth harmonic maps between Riemannian manifolds. Let $f \colon (M,g) \to (N,h)$ be a smooth map between Riemannian manifolds. Fix $x \in M$. Denote by $S_r$ (resp.~$B_r$) the sphere (resp.~the closed ball) centered at the origin of radius $r$ in the Euclidean vector space $(\upT_x M, g)$. Also denote $\hat{S}_r$ (resp.~$\hat{B}_r$) the sphere (resp.~the closed ball) centered at $x$ of radius $r$ in $(M,g)$. The topological space $S_r$ (resp. $B_r$) can be equipped with a natural Borel probability measure by taking the measure induced from the Euclidean metric $g$ in $T_x M$, renormalized so that it has total mass $1$. Similarly, $\hat{S}_r$ (resp. $\hat{B}_r$) can be equipped with a natural Borel probability measure by taking the measure induced from the Riemannian metric $g$. \begin{definition} \label{def:AverageMap} We define four functions $S_r f, B_r f, \hat{S}_r f, \hat{B}_r f \colon M \to N$ as follows. Given $x \in M$: \begin{itemize} \item $S_r f(x)$ (resp.~$B_r f(x)$) is the center of mass of $ f \circ \exp_x \colon S_r \to N$ (resp.~$f \circ \exp_x\colon B_r \to N$). \item $\hat{S}_r f(x)$ (resp.~$\hat{B}_r f(x)$) is the center of mass of $f\evalat{\hat{S}_r} \colon \hat{S}_r \to N$ (resp.~ $f\evalat{\hat{B}_r} \colon \hat{B}_r \to N$). \end{itemize} \end{definition} \begin{remark} The four functions of \autoref{def:AverageMap} are well-defined as long as $(N,h)$ is a Hadamard manifold, or as long as $r$ is small enough and $(N,h)$ has sectional curvature bounded above and injectivity radius bounded below by a positive number (\eg{} $N$ is compact). \end{remark} Note that $S_r f(x)$ and $\hat{S}_r f(x)$ are different in general, as are $B_r f(x)$ and $\hat{B}_r f(x)$. However they are very close when $r$ is small: \begin{proposition} \label{prop:ComparisonAverages} Let $f \colon M \to N$ be a smooth map. Then for all $x \in M$: \begin{equation} \begin{aligned} d(S_r f(x), \hat{S}_r f(x)) &= \bigO(r^4)\\ d(B_r f(x), \hat{B}_r f(x)) &= \bigO(r^4)~.\\ \end{aligned} \end{equation} \end{proposition} \begin{proof} The proof is technical but not very difficult. It is basically derived from a Taylor expansion of the metric in normal coordinates at $x$ and one use of \autoref{lem:ProximityToCenterOfMass}. We will do several similar proofs in what follows, so we skip the details for brevity. \end{proof} Of course, in the case where $M = \bR^m$ and $N = \bR$ (or $N = \bR^n$), $S_r f$ and $\hat{S}_r f$ (resp. $B_r f$ and $\hat{B}_r f$) coincide. We recall the celebrated mean property for harmonic functions in this setting: \begin{theorem} \label{thm:ClassicalMeanProperty} $f \colon \bR^m \to \bR$ is harmonic if and only if $S_r f = B_r f = f$ for all $r > 0$. \end{theorem} More generally, if $M$ is any Riemannian manifold and $N = \bR$, Willmore \cite{MR0033408} proved that $\hat{S}_r f = f$ characterizes harmonic maps if and only if $M$ is a \emph{harmonic manifold}. The central theorem of this subsection is the following: \begin{theorem} \label{thm:AverageMap} Let $f \colon M \to N$ be a smooth map. For all $x \in M$, as $r \to 0$: \begin{align} d\left(S_r f(x)~,~ \exp_{f(x)}\left(\frac{r^2}{2 m} \tau(f)_{x}\right)\right) &= \bigO(r^4) \label{eq:AverageMap1}\\ d\left(B_r f(x)~,~ \exp_{f(x)}\left(\frac{r^2}{2(m+2)} \tau(f)_{x}\right)\right) &= \bigO(r^4) \label{eq:AverageMap2}~. \end{align} \end{theorem} The following ``generalized mean property for harmonic functions between Riemannian manifolds'' is an immediate corollary of \autoref{thm:AverageMap}: \begin{theorem} \label{thm:GeneralizedMeanValueProperty} Let $f \colon M \to N$ be a smooth map. The following are equivalent: \begin{enumerate}[(i)] \item $f$ is harmonic. \item $d(f(x), S_r f(x)) = \bigO(r^4)$ for all $x\in M$. \item $d(f(x), B_r f(x)) = \bigO(r^4)$ for all $x\in M$. \end{enumerate} \end{theorem} \begin{remark} It is an immediate consequence of \autoref{prop:ComparisonAverages} that \autoref{thm:AverageMap} and \autoref{thm:GeneralizedMeanValueProperty} also hold for $\hat{S}_r f(x)$ instead of $S_r f(x)$, and $\hat{B}_r f(x)$ instead of $B_r f(x)$. \end{remark} In the remainder of this subsection we show \autoref{thm:AverageMap}. We only prove \eqref{eq:AverageMap1}; the proof of \eqref{eq:AverageMap2} follows exactly the same lines. Consider a smooth map $f \colon (M,g) \to (N,h)$ and fix $x \in M$. \begin{lemma} \label{lem:AverageG} Let $r > 0$. Denote by $S_r$ the Euclidean sphere of radius $r>0$ in $T_x M$ and $\sigma_r$ its area density. Then, as $r \to 0$, the following estimate holds: \begin{equation} \label{eq:AverageG} \frac{1}{\Area(S_r)} \int_{S_r} \exp_{f(x)}^{-1} \circ f \circ \exp_x(u) \, \upd \sigma_r(u) = \frac{r^2}{2 m} \tau(f)_x + \bigO(r^4) \end{equation} where $m = \dim M$. \end{lemma} \begin{proof} We write the Taylor expansion of the function $\hat{f} = \exp_{f(x)}^{-1} \circ f \circ \exp_x \colon T_x M \to T_{f(x)} N$: \begin{equation} \hat{f}(u) = \hat{f}(0) + (\upD \hat{f})_0(u) + \frac{1}{2}(\upD^2 \hat{f})_0(u,u) + \frac{1}{6}(\upD^3 \hat{f})_0(u,u,u) + \bigO(\Vert u \Vert^4)~. \end{equation} Let us integrate this identity over $S_r$. We have: \begin{itemize} \item $\hat{f}(0) = 0$, so $\frac{1}{\Area(S_r)} \int_{S_r} \hat{f}(0) \, \upd \sigma_r(u) = 0$. \item $(\upD \hat{f})_0(u) = (\upd f)_x(u)$ is an odd function of $u$, so $\frac{1}{\Area(S_r)} \int_{S_r} (\upD \hat{f})_0(u) \, \upd \sigma_r(u) = 0$. \item It is straightforward to check that $(\upD^2 \hat{f})_0(u, u) = (\Hess f)_x(u, u)$ by definition of $\Hess f$. Moreover, since this is a quadratic function of $u$, we can apply \autoref{lem:BilinearFormAverage}: \begin{equation} \frac{1}{\Area(S_r)} \int_{S_r} (\upD^2 \hat{f})_0(u, u) \, \upd \sigma_r(u) = \frac{r^2}{m}\tr((\Hess f)_x) = \frac{r^2}{m} \tau(f)_x~. \end{equation} \item $(\upD^3 \hat{f})_0(u,u,u)$ is an odd function of $u$, so $\frac{1}{\Area(S_r)} \int_{S_r} (\upD^3 \hat{f})_0(u, u, u) \, \upd \sigma_r(u) = 0$. \end{itemize} Putting all this together, we get \eqref{eq:AverageG}. \end{proof} The following lemma is required to complete the proof of \autoref{lem:AverageG}: \begin{lemma} \label{lem:BilinearFormAverage} Let $(V, g = \langle \cdot, \cdot \rangle)$ be a Euclidean vector space and let $B \colon V \times V \to \bR$ be a symmetric bilinear form. Denote by $S_r = S(0,r)$ the sphere centered at the origin in $V$ with radius $r > 0$, $\upd \sigma_r$ the area density on $S_r$ induced from the metric $g$ and $\Area(S_r) = \int_{S_r} \upd \sigma_r$ its area. Then: \begin{equation} \frac{1}{\Area(S_r)}\int_{S_r} B(x, x) \, \upd \sigma_r = \frac{r^2}{\dim V}\tr_g(B)~. \end{equation} \end{lemma} Here we have denoted by $\tr_g(B)$ the $g$-trace of $B$, \ie{} the trace of the $g$-self adjoint endomorphism of $V$ associated to $B$, or, equivalently, the trace of a matrix representing $B$ in a $g$-orthonormal basis. \begin{proof} Let $(e_1, \dots, e_n)$ be basis of $V$ which is $g$-orthonormal and $B$-orthogonal (the existence of such a basis is precisely the spectral theorem). Let $\lambda_k = B(e_k, e_k)$ for $k \in \{1, \dots n\}$. For any vector $x = \sum_{k=1}^n x_k e_k$, the quadratic form is given by $B(x,x) = \sum_{k=1}^n \lambda_k {x_k}^2$, hence: \begin{equation} \int_{S_r} B(x, x) \,\upd \sigma_r = \sum_{k=1}^n \lambda_k \int_{S_r} {x_k}^2 \,\upd \sigma_r~. \end{equation} The integrals $I_k = \int_{S_r} {x_k}^2 \,\upd \sigma_r$ can be swiftly computed starting with the observation that any two of them are equal. Indeed, for $k \neq l$, one can easily find a linear isometry $\varphi$ such that ${x_k}^2 \circ \varphi = {x_l}^2$; the change of variables theorem ensures that $I_k = I_l$. One can then write $I_k = \frac{1}{n} \sum_{l=1}^n I_l$ for any $k$. That is $I_k = \frac{1}{n}\int_{S_r} \left(\sum_{k=1}^n {x_k}^2\right) \,\upd \sigma_r$. However $\sum_{k=1}^n {x_k}^2 = g(x,x) = r^2$ for any $x \in S_r$. This yields $I_k = \frac{1}{n}\int_{S_r} r^2 \,\upd \sigma_r = \frac{r^2}{n} \Area(S_r)$. The desired result follows. \end{proof} It is easy to see that \autoref{thm:AverageMap} follows immediately from \autoref{lem:AverageG} when $N = \bR^n$. When $N$ is not Euclidean, centers of mass in $N$ are only defined implicitly (by equation \eqref{eq:ImplicitEquationForCenterOfMass}), so we have to work harder to prove \autoref{thm:AverageMap}. The trick is to use \autoref{lem:ProximityToCenterOfMass}. First we need a Riemannian geometry estimate in the following general setting. Let $A$, $B$, $C$ be three points in a Riemannian manifold $(M,g)$. We assume that $B$ and $C$ are contained in a sufficiently small ball centered at $A$ for what follows to make sense. Denote by $\vec{u_A} = (\exp_A)^{-1}(B)$, $\vec{u_B} = (\exp_B)^{-1}(C)$, and $\vec{u_C} = (\exp_C)^{-1}(A)$. If we were in a Euclidean vector space, we could write: \begin{equation} \vec{u_A} + \vec{u_B} + \vec{u_C} = 0~. \end{equation} We would like to find an approximate version of this identity in general. Note that the sum $\vec{u_A} + \vec{u_B} + \vec{u_C}$ does not make sense, because these vectors are based at different points. Denote $\vec{v}$ the parallel transport of $-\vec{u_C}$ along the geodesic segment $[C, A]$ and $\vec{w}$ the parallel transport of $\vec{u_B}$ along the geodesic segment $[B, A]$. Let us also write $\vec{u} = \vec{u_A}$ for aesthetics. Now the vectors $\vec{u}$, $\vec{v}$, $\vec{w}$ are all based at $A$, and one expects that $\vec{w} = \vec{v} - \vec{u}$ up to some error term. \begin{lemma} \label{lem:RiemannianEstimateTriangle1} Using the setting and notations above, we have as $\Vert \vec{u} \Vert \to 0$ and $\Vert \vec{v} \Vert \to 0$: \begin{equation} \vec{w} = \vec{v} - \vec{u} + \bigO(\Vert \vec{u} \Vert^2 \, \Vert \vec{v} \Vert + \Vert \vec{u} \Vert \, \Vert \vec{v} \Vert^2)~. \end{equation} \end{lemma} We prove in fact the following more precise lemma: \begin{lemma} \label{lem:RiemannianEstimateTriangle2} Let $(M,g)$ be a Riemannian manifold, fix $A \in M$. Let $\vec{U}$ and $\vec{V}$ be two tangent vectors at $A$, denote $B(t) = \exp_A(t \vec{U})$ and $C(s) = \exp_A(s \vec{V})$. Let $\vec{w}(t, s)$ be the parallel transport of $\vec{u_B} \coloneqq \exp_{B(t)}^{-1}(C(s))$ along the geodesic segment from $B(t)$ to $A$. Then: \begin{equation} \label{eq:RiemannianEstimateTriangle2} \vec{w}(t,s) = s\vec{V} - t\vec{U} - \frac{{t}^2 s}{2} R(\vec{V}, \vec{U}) \vec{U} - \frac{t {s}^2}{3} R(\vec{U}, \vec{V}) \vec{V} + \bigO(t^4 + t^3 s + t^2 s^3)~. \end{equation} where $R$ is the Riemann curvature tensor of $(M,g)$. \end{lemma} \begin{proof} First let us quickly discuss some general Riemannian geometry estimates in normal coordinates. We refer to \cite{LeoBrewinRiemannNormalCoordinates, MR2534336} for more details on the computations that follow. In normal coordinates at a point $A$, the Riemannian metric $g$ has the Taylor expansion \begin{equation} g_{ij} = \delta_{ij} - \frac{1}{3} R_{ikjl}x^k x^l + \bigO(|x|^3) \end{equation} where $R_{ijkl}$ is the Riemann curvature tensor at $A$, or rather its purely covariant version (this well-known fact of Riemannian geometry goes back to Riemann's 1854 habilitation \cite{MR3525305}). One can derive from the expression for the Christoffel symbols $\Gamma^k{}_{ij} = \frac{1}{2} g^{kl} (g_{li,j} + g_{lj,i} - g_{ij,l})$ that \begin{equation} \Gamma^k{}_{ij} = -\frac{1}{3} (R^k{}_{ijl} - R^k{}_{jil}) x^l + \bigO(|x|^3)~. \end{equation} One can then find the Taylor expansion of any geodesic $x(s)$, say with initial endpoint $x = x(0)$ and initial velocity $v$, by solving the geodesic equation $\frac{\upd^2 x^k }{\upd s^2} + \Gamma^{k}_{i j }\frac{\upd x^i }{\upd s}\frac{\upd x^j}{\upd s} = 0$. One finds: \begin{equation} \label{eq:GeodesicNormalCoordinates1} x^k(s) = x^k + s v^k - \frac{s^2}{3} R^k{}_{ilj} \, v^i v^j x^l + \bigO(s |x|^3)~. \end{equation} We can rewrite \eqref{eq:GeodesicNormalCoordinates1} as a coordinate-free expression (but still in the chart given by $\exp_A$) as \begin{equation} \label{eq:GeodesicNormalCoordinates2} x(s) = x + s v - \frac{s^2}{3} R(x, v) v + \bigO(s |x|^3)~. \end{equation} One can also compute the parallel transport of a vector $v$ along a radial geodesic $x(t) = tx$ by solving the parallel transport equation $\frac{\upd v^k}{\upd t} + \Gamma^k{}_{ij}(x(t)) v^i \frac{\upd x^j}{\upd t}$. One finds that \begin{equation} \label{eq:ParallelTransportEstimate1} v^k(t) = v^k + \frac{1}{6} R^k{}_{jil} v^i x^j(t) x^l(t) + \bigO(t |x|^3)~, \end{equation} which we can rewrite as \begin{equation} \label{eq:ParallelTransportEstimate2} v(t) = v + \frac{1}{6} R(v, x) x + \bigO(t |x|^3)~. \end{equation} Let us now come back to the setting of \autoref{lem:RiemannianEstimateTriangle2}. We shall work (implicitly) in the chart given by $\exp_A$. Note that we can write $B = t\vec{U}$ and $C = s\vec{V}$ in this chart. Let us denote by $x(\cdot)$ the unit speed geodesic from $B$ to $C$, so that $x(0) = B$, $x(r) = C$ where $r = d(B,C)$, and $x'(0) = \vec{U_B}$ is the unit vector such that $\exp_B(r \vec{U_B}) = C$. By \eqref{eq:GeodesicNormalCoordinates2} we can write \begin{equation} x(r) = x(0) + r \vec{U_B} - \frac{r^2}{3} R(x(0), \vec{U_B}) \vec{U_B} + \bigO(r |x(0)|^3)~. \end{equation} In other words, recalling that $x(0) = B = t \vec{U}$ and $x(r) = C = s\vec{V}$, we have \begin{equation} \label{eq:VminusUEstimate} s\vec{V} - t\vec{U} = r \vec{U_B} - \frac{t {r}^2}{3} R(\vec{U}, \vec{U_B}) \vec{U_B} + \bigO(r {t}^3)~. \end{equation} On the other hand, the parallel transport of $\vec{u_B} = r \vec{U_B}$ back to the origin along the radial geodesic $[A, B]$ is given by, according to \eqref{eq:ParallelTransportEstimate2}: \begin{equation} \label{eq:WEstimate1} \vec{w} = r \vec{U_B} - \frac{{t}^2 r}{6} R(\vec{U_B}, \vec{U}) \vec{U} + \bigO({t}^3 r)~. \end{equation} Comparing \eqref{eq:VminusUEstimate} and \eqref{eq:WEstimate1}, we see that \begin{equation} \label{eq:WEstimate2} \vec{w} = s\vec{V} - t\vec{U} - \frac{{t}^2 r}{6} R(\vec{U_B}, \vec{U}) \vec{U} - \frac{t {r}^2}{3} R(\vec{U}, \vec{U_B}) \vec{U_B} + \bigO({t}^3 r)~. \end{equation} Finally, let's work to have $s$'s and $\vec{V}$'s appear in this equation instead of $r$'s and $\vec{U_B}$'s. First note that $\vec{r U_B} = s \vec{V} - t \vec{U} + \bigO(t {r}^2 + r {t}^3)$ according to \eqref{eq:VminusUEstimate}, so using the fact that $R(\vec{U}, \vec{U}) = 0$, one can write: \begin{equation} \begin{split} r R(\vec{U_B}, \vec{U}) \vec{U} &= s R(\vec{V}, \vec{U}) \vec{U} + \bigO(t {r}^2 + r {t}^3)\\ {r}^2 R(\vec{U}, \vec{U_B}) \vec{U_B} &= {s}^2 R(\vec{U}, \vec{V}) \vec{V} -t s R(\vec{U}, \vec{V}) \vec{U} + \bigO(tsr^2 + t^2sr + t^2 r^3 + t^3sr + t^4 r)~. \end{split} \end{equation} We thus get in lieu of \eqref{eq:WEstimate2}: \begin{equation} \label{eq:WEstimate3} \vec{w} = s\vec{V} - t\vec{U} - \frac{{t}^2 s}{2} R(\vec{V}, \vec{U}) \vec{U} - \frac{t {s}^2}{3} R(\vec{U}, \vec{V}) \vec{V} + \bigO({t}^3 r + t^2 s r^2)~. \end{equation} The conclusion follows, noting that $r = \bigO(t + s)$ by the triangle inequality. \end{proof} \begin{remark} A direct consequence of \autoref{lem:RiemannianEstimateTriangle2} is the expansion of the distance squared: \begin{equation} d^2(\exp_A(t\vec{U}), \exp_A(s \vec{V})) = \Vert s\vec{V} - t\vec{U} \Vert^2 - \frac{1}{3}R(U, V, V, U)s^2 t^2 + \bigO((t^2 + s^2)^{\frac{5}{2}})~. \end{equation} The same formula has been observed by other authors, see \eg{} \cite{RazielMO}. \end{remark} We are now ready to wrap up the proof of \autoref{thm:AverageMap}: \begin{proof}[Proof of \autoref{thm:AverageMap}] Let $u_0 \in S_r \subset \upT_x M$, and consider the triangle in $N$ with vertices $A \coloneqq f(x)$, $B \coloneqq T_r f(x)$, and $C(u_0) \coloneqq f(\exp_x(u_0))$ in $N$. With the notations introduced above \autoref{lem:RiemannianEstimateTriangle1}, note that $\vec{u} = \vec{u_A} = \frac{r^2}{2m} \tau(f)_x$, $\vec{v}(u_0) = \exp_{f(x)}^{-1} (f(\exp_x(u_0)))$, and $\vec{w}(u_0) = P (\vec{u_B}(u_0))$ where $\vec{u_B}(u_0) = \exp_{T_r f(x)}^{-1} (f(\exp_x(u_0)))$ and $P \colon \upT_A N \to \upT_B N$ is the parallel transport along the geodesic segment $[A,B]$. By \autoref{lem:RiemannianEstimateTriangle1}, we have \begin{equation} \label{eq:AM1} P (\vec{u_B}(u_0)) = \vec{w}(u_0) = \vec{v}(u_0) - \vec{u} + \bigO(\Vert \vec{u} \Vert^2 \, \Vert \vec{v}(u_0) \Vert + \Vert \vec{u} \Vert \, \Vert \vec{v}(u_0) \Vert^2)~. \end{equation} Because we have $\Vert \vec{u} \Vert = \bigO(r^2)$ and $\Vert \vec{v}(u_0) \Vert = \bigO(r)$, \eqref{eq:AM1} may be rewritten: \begin{equation} \label{eq:AM2} P (\vec{u_B}(u_0)) = \vec{v}(u_0) - \vec{u} + \bigO(r^4)~. \end{equation} We now integrate \eqref{eq:AM2} over $u_0 \in S_r$: \begin{equation} \label{eq:AM3} \frac{1}{\Area(S_r)} \int_{S_r} P (\vec{u_B}(u_0)) \, \upd \sigma_r(u_0) = \frac{1}{\Area(S_r)} \int_{S_r} \left( \vec{v}(u_0) \, \upd \sigma_r(u_0) ~-~ \vec{u} + \bigO(r^4)\right)~, \end{equation} which we can rewrite as \begin{equation} P\left(\frac{1}{\Area(S_r)} \int_{S_r} \vec{u_B}(u_0) \, \upd \sigma_r(u_0) \right) = \left(\frac{1}{\Area(S_r)} \int_{S_r} \vec{v}(u_0) \, \upd \sigma_r(u_0)\right) ~-~ \vec{u} + \bigO(r^4)~. \end{equation} Now, \autoref{lem:AverageG} says precisely that $\left(\frac{1}{\Area(S_r)} \int_{S_r} \vec{v}(u_0) \, \upd \sigma_r(u)\right) = \vec{u} + \bigO(r^4)$. We thus get $\frac{1}{\Area(S_r)} \int_{S_r} \vec{u_B}(u_0) \, \upd \sigma_r(u_0) = P^{-1}(\bigO(r^4)) = \bigO(r^4)$. That is, \begin{equation} \int_{S_r} \exp_{T_r f(x)}^{-1} (f(\exp_x(u_0))) \upd \sigma_r(u_0) = \bigO(r^4)~. \end{equation} Recalling that $S_r f(x)$ is by definition the center of mass of the function $u_0 \in S_r \mapsto f(\exp_x(u_0))$, we can apply \autoref{lem:ProximityToCenterOfMass} to conclude that $ d(S_rf(x), T_rf(x)) = \bigO(r^4)$. \end{proof} \subsection{Center of mass methods} \label{subsec:CenterOfMassMethods} We now discuss center of mass methods as an alternative to the heat flow in order to minimize the energy functional. The basic idea is to iterate the process of replacing a function $f \colon M \to N$ by its average on balls (or spheres) of radius $r>0$, hopefully converging to a map $f_r^*$ that is almost harmonic when $r$ is small. Observe that \autoref{thm:AverageMap} shows that this method is very close to a constant step gradient flow for the energy functional, \ie{} an Euler method with fixed stepsize. The next proposition is claimed in \cite[Lemma 4.1.1]{MR1451625}. \begin{proposition} \label{prop:AveragingDecreasesApproximateEnergy} Let $(M,\mu)$ be a measure space, let $(N,d)$ be a Hadamard metric space and let $\eta \colon M \times M \to [0,+\infty)$ be a measurable symmetric function. Define the Jost energy functional by \begin{equation} \label{eq:JostEnergyFunctional} E(f) = \frac{1}{2} \int_M \int_M \eta(x,y) \, {d(f(x), f(y))}^2 \, \upd \mu(y) \, \upd \mu (x)~. \end{equation} For a measurable map $f \colon M \to N$, let $\varphi(f) \colon M \to N$ be the map such that for all $x\in M$, $\varphi(f)(x)$ is the center of mass of $f$ for the measure $\eta(x, \cdot) \mu$. Then for every $f$ with finite energy we have: \begin{equation} \label{eq:AveragingDecreasesApproximateEnergy} E(\varphi(f)) \leqslant E(f)~. \end{equation} Moreover, the following are equivalent: \begin{enumerate}[(i)] \item Equality holds in \eqref{eq:AveragingDecreasesApproximateEnergy}. \item $\varphi(f) = f$ almost everywhere in $(M, \mu)$. \item $f$ is a minimizer of $E$. \end{enumerate} \end{proposition} \subsubsection*{Center of mass method in the smooth setting} Now assume that $M$ and $N$ are both Riemannian manifolds. For $r > 0$ we take the kernel $\eta_r(x,y)$ described in \autoref{subsec:JostEnergy}, so that $E_r$ is the $r$-approximate energy. The map $\varphi(f)$ of \autoref{prop:AveragingDecreasesApproximateEnergy} is then the same as the map $\hat{B}_r f$ introduced in \autoref{def:AverageMap}. It is tempting to iterate the process of averaging $f$ in order to try and minimize $E_r$. The next theorem guarantees the success of this method under suitable conditions. Moreover, recall that the energy functional $E$ on $\upL^2(M,N)$ is the $\Gamma$-limit of $E_r$ as $r\to 0$ (\cite[Lemma 8.3.4]{MR3726907}), so that minimizers of $E_r$ converge to minimizers of $E$ (possibly up to subsequence). \begin{theorem} \label{thm:CenterOfMassMethodSmooth} Let $M$ and $N$ be Riemannian manifolds, assume $N$ is compact and with nonpositive sectional curvature. In any homotopy class of continuous maps $M \to N$ where the $r$-approximate energy $E_r$ admits a unique minimizer $f^*$, the sequence $(f_k)_{k \in \bN}$ defined by $f_{k+1} = \hat{B}_r f_k$ converges locally uniformly to $f^*$ for any choice of a locally Lipschitz continuous $f_0$. \end{theorem} \begin{proof} We reduce the proof to a combination of \autoref{lem:AveragePreservesLipschitz} and \autoref{lem:ConvergingSequence} below. Denote by $X$ the connected component of $f_0$ in $\cC(M,N)$, let $E \colon X \to \bR$ denote the restriction of $E_r$, and let $\varphi \colon X \to X$ be the map $f \mapsto \hat{B}_r f$ (it is easy to see that $\varphi$ preserves $X$). \autoref{lem:AveragePreservesLipschitz} guarantees immediately that the sequence $(f_k)_{k\in \bN}$ is equicontinuous. Since $N$ is compact, it follows from the Arzelà-Ascoli theorem that the sequence $(f_k)_{k\in \bN}$ is relatively compact in $X$ for the compact-open topology. By \autoref{prop:AveragingDecreasesApproximateEnergy} and the assumption that $f^*$ is unique, we have $E(\varphi(f)) \leqslant E(f)$ for all $f \in X$, with equality only if $f = f^*$. Conclude by application of \autoref{lem:ConvergingSequence}. \end{proof} \begin{lemma} \label{lem:AveragePreservesLipschitz} Let $f \colon M \to N$ where $M$ and $N$ are Riemannian manifolds, with $N$ complete and nonpositively curved. If $f$ is locally Lipschitz continuous, then so is $\hat{B}_r f$. Moreover, the Lipschitz constant of $\hat{B}_r f$ is bounded above by the Lipschitz constant of $f$ on any compact $K\subseteq M$. \end{lemma} \begin{proof} For simplicity, we assume that $f$ is globally $L$-Lipschitz, and argue that $\hat{B}_r f$ is also $L$-Lipschitz; the proof can easily be extended to the general case by restricting to compact sets. First we assume that $M$ is Euclidean, in fact let us put $M = \bR^m$. Let $x, y \in M$, write $y = x + h$ so that $d(x,y) = \Vert h \Vert$. By definition, $\hat{B}_r f(y)$ is the point of $N$ such that \begin{equation} \label{eq:AveragePreservesLipschitz1} \frac{1}{\vol(B(y,r))} \int_{B(y,r)} \exp_{\hat{B}_r f(y)}^{-1}(f(v)) \, \upd v_g(v) = 0~. \end{equation} Note that the map $u \mapsto u + h$ defines an isometry from $B(x,r)$ to $B(y,r)$. Making the change of variables $v = u+h$, we derive from \eqref{eq:AveragePreservesLipschitz1}: \begin{equation} \label{eq:AveragePreservesLipschitz2} \int_{B(x,r)} \exp_{\hat{B}_r f(y)}^{-1}(f(u + h)) \, \upd v_g(u) = 0~. \end{equation} It follows that \begin{equation} \label{eq:AveragePreservesLipschitz3} \int_{B(x,r)} \exp_{\hat{B}_r f(y)}^{-1}(f(u)) \, \upd v_g(u) = \int_{B(x,r)} \left[\exp_{\hat{B}_r f(y)}^{-1}(f(u)) - \exp_{\hat{B}_r f(y)}^{-1}f(u + h)) \right] \, \upd v_g(u)~. \end{equation} Assume without loss of generality that $N$ is simply connected (one can lift to the universal cover). Then $N$ is a Hadamard manifold and in particular a $\CAT(0)$ metric space, which implies that for any $p \in N$, the map $\exp_p^{-1} \colon N \to T_p N$ is distance nonincreasing (in fact, the converse is also true). We can therefore derive from \eqref{eq:AveragePreservesLipschitz3}: \begin{equation} \label{eq:AveragePreservesLipschitz4} \left\Vert \int_{B(x,r)} \exp_{\hat{B}_r f(y)}^{-1}(f(u)) \, \upd v_g(u) \right\Vert \leqslant \int_{B(x,r)} d(f(u), f(u + h)) \, \upd v_g(u)~. \end{equation} It follows from \eqref{eq:AveragePreservesLipschitz4} and the fact that $f$ is $L$-Lipschitz that \begin{equation} \label{eq:AveragePreservesLipschitz5} \left\Vert \frac{1}{\vol(B(x,r))} \int_{B(x,r)} \exp_{\hat{B}_r f(y)}^{-1}(f(u)) \, \upd v_g(u) \right\Vert \leqslant L \Vert h \Vert~. \end{equation} \autoref{lem:ProximityToCenterOfMass} now applies directly to \eqref{eq:AveragePreservesLipschitz5} to conclude that $d(\hat{B}_r f(x), \hat{B}_r f(y)) \leqslant L \Vert h \Vert$. Since $\Vert h \Vert = d(x,y)$, we have shown that $\hat{B}_r f$ is $L$-Lipschitz, as desired. Now we argue that the argument extends to the case where $M$ is an arbitrary Riemannian manifold using a \emph{local to global} trick. First note that a function is globally $L$-Lipschitz if and only if it is locally $L$-Lipschitz. Here we mean by \emph{locally $L$-Lipschitz} the property that for any $x \in M$, there exists $\delta > 0$ such that $d(y, x) < \delta$ implies $d(f(y), f(x)) \leqslant L d(y,x)$. We leave it to the reader to show that in any path metric space, locally $L$-Lipschitz in this sense implies globally $L$-Lipschitz. With this observation in mind, let us finish the proof. The key argument that worked above when $M$ is Euclidean is that there exists an isometry from $B(x,r)$ to $B(y,r)$ that displaces every point of at most $d(x,y)$. This is no longer true when $M$ is an arbitrary Riemannian manifold, however note that it is almost true when $x$ and $y$ are very close. Quantifying this properly, clearly one can show that for every $x\in M$ and for every $L' > L$, there exists $\delta > 0$ such that $d(y, x) < \delta$ implies $d(\hat{B}_r f(y), \hat{B}_r f(x)) \leqslant L' d(y,x)$. Thus we have shown that $\hat{B}_r f$ is locally $L'$-Lipschitz, and therefore globally $L'$-Lipschitz. Since this is true for all $L' > L$, $\hat{B}_r f$ is actually $L$-Lipschitz. \end{proof} \begin{lemma} \label{lem:ConvergingSequence} Let $X$ be a first-countable topological space and let $E \colon X \to \bR$ a continuous function that admits a unique minimizer $x^*$. Assume that $\varphi \colon X \to X$ is a continuous map such that: \begin{enumerate}[(i)] \item \label{lem:ConvergingSequenceHyp1} $E(\varphi(x)) \leqslant E(x)$ for all $x \in X$, with equality only if $x = x^*$. \item \label{lem:ConvergingSequenceHyp2} For all $x_0 \in X$, the set $\{\varphi^k(x_0), k\in \bN\}$ is relatively compact in $X$. \end{enumerate} Then for any $x_0 \in X$, the sequence $(\varphi^k(x_0))_{k \in \bN}$ converges to $x^*$. \end{lemma} \begin{proof} In any topological space, in order to show that a sequence $(x_k)_{k \in \bN}$ converges to a point $x^*$, it is enough to show that: \begin{enumerate}[(a)] \item \label{lem:ConvergingSequenceProofItem1} The sequence $(x_k)$ has no cluster points except possibly $x^*$. \item \label{lem:ConvergingSequenceProofItem2} Any subsequence of $(x_k)$ admits a cluster point. \end{enumerate} Indeed, assume that $(x_k)$ does not converge to $x^*$, then there exists a subsequence of $(x_k)$ that avoids a neighborhood of $x^*$. This subsequence must have a cluster point by \ref{lem:ConvergingSequenceProofItem2}, which cannot be $x^*$. However this point is also a cluster point of the sequence $(x_k)$, contradicting \ref{lem:ConvergingSequenceProofItem1}. Coming back to \autoref{lem:ConvergingSequence}, let $x_0 \in X$ and denote $x_k = \varphi^k(x_0)$. The sequence $(x_k)$ satisfies \ref{lem:ConvergingSequenceProofItem2} because of the assumption \ref{lem:ConvergingSequenceHyp2}. So we need to show that $(x_k)$ satisfies \ref{lem:ConvergingSequenceProofItem1} and we are done. Let $y$ be a cluster point of $(x_k)$, we need to show that $y = x^*$. Since $X$ is first-countable, there exists a subsequence $(x_{k_n})_{n \in \bN}$ converging to $y$. Observe that by assumption \ref{lem:ConvergingSequenceHyp1}, since $k_n \leqslant k_n +1 \leqslant k_{n+1}$, \begin{equation} \label{eq:ConvergingSequenceProof1} E(x_{k_{n+1}}) \leqslant E(x_{k_n + 1}) \leqslant E(x_{k_n}) ~. \end{equation} By continuity of $E$, we have $\lim E(x_{k_n}) = \lim E(x_{k_{n+1}}) = E(y)$, so \eqref{eq:ConvergingSequenceProof1} implies that $\lim E(x_{k_n + 1}) = E(y)$. On the other hand, since $x_{k_n + 1} = \varphi(x_{k_n})$ and $\varphi$ is continuous, we have $\lim x_{k_n + 1} = \varphi(y)$, so $\lim E(x_{k_n + 1}) = E(\varphi(y))$. Thus $E(\varphi(y)) = E(y)$, and we conclude that $y = x^*$ by \ref{lem:ConvergingSequenceHyp1}. \end{proof} \subsubsection*{Discrete center of mass method} We now prove that \autoref{thm:CenterOfMassMethodSmooth} also holds in the discrete setting developed in \autoref{sec:Discretization}, providing an alternative method to the discrete heat flow discussed in \autoref{subsec:ConvergenceOfDiscreteHeatFlow}. Let $\cG$ be a biweighted $\tilde{S}$-triangulated graph (see \autoref{subsec:Graphs}), let $N$ be a Hadamard manifold and let $\rho \colon \pi_1S\to \Isom(N)$ be a group homomorphism. We recall that the discrete energy functional $E_\cG \colon \Map_{\text{eq}}(\cG, N) \to \bR$ coincides with Jost's energy functional \eqref{eq:JostEnergyFunctional} for the appropriate choice of kernel $\eta$ (\autoref{prop:JostEnergyGraph}). Motivated by \autoref{prop:AveragingDecreasesApproximateEnergy}, we note that in this discrete setting the measure $\eta(x,\cdot) \mu$ is given by the weighted atomic measure \begin{equation} \label{eq:AtomicMeasure} \sum_{y\sim x} \frac {\omega_{xy}}{\mu(x)} \; \delta_y~, \end{equation} where $\delta_y$ is the Dirac measure at $y$. Now the averaging map $f\mapsto \hat{B}_rf$ takes the following form: \begin{definition}\label{def:discreteCenterGraph} The \emph{discrete center of mass method} on $\Map_{\text{eq}}(\cG, N) $ is given by $f \mapsto \varphi(f)$, where $\varphi(f)(x)$ is the center of mass of $f$ for the atomic measure \eqref{eq:AtomicMeasure}. \end{definition} Note that \autoref{prop:AveragingDecreasesApproximateEnergy} applies in this setting (cf.~\autoref{prop:HarmonicCenterOfMassGraph2}). Under certain assumptions on $N$ and $\rho$, we obtained strong convexity of $E_\cG$ in \autoref{thm:StrongConvexityEnergyGraphGeneral}, so that, in particular, $E_\cG$ has a unique minimum. The same assumptions have similarly useful consequences here: \begin{theorem} \label{thm:CenterOfMassMethodDiscrete} Let $N$ be a manifold of negative curvature bounded away from $0$ and let $\rho$ be a faithful representation whose image is contained in a discrete subgroup of $\Isom(N)$ acting freely, properly, and cocompactly on $N$. Given any initial discrete equivariant map $f_0 \in \Map_{\text{eq}}(\cG, N) $, the discrete center of mass method converges to the unique discrete harmonic map. \end{theorem} \begin{proof} The proof is a similar but easier version of the proof of \autoref{thm:CenterOfMassMethodSmooth}. Let $f_0 \in \Map_{\text{eq}}(\cG, N) $, and define the sequence $(f_k)_{k \in \bN}$ by $f_{k+1} = \varphi(f_k)$. The assumption on $\rho$ means that we can work in a compact quotient of $N$, making the sequence $(f_k)$ pointwise relatively compact. Since the action of $\pi_1S$ on $\cG$ is cofinite, it is easy to see that the condition that the family $\{f_k\}$ is equicontinuous is vacuous. Hence the family $\{f_k\}$ is relatively compact in $\Map_{\text{eq}}(\cG, N)$. Since \autoref{prop:AveragingDecreasesApproximateEnergy} holds in this setting, all the requirements are met to conclude with \autoref{lem:ConvergingSequence}. \end{proof} \begin{remark} In \cite{MR2346504}, Jost-Todjihounde describe an iterative process to obtain a discrete harmonic map from an edge-weighted triangulated graph $\cG$ to a target space that admits centers of mass (\eg{} Hadamard spaces). \autoref{thm:CenterOfMassMethodDiscrete} can be viewed as a strengthened version of their result in two respects: For one, we avoid Jost-Todjihounde's passage to a subsequence of $(f_k)$. Moreover, Jost-Todjihounde start by subdividing $\cG$ and pursuing centers of mass in two phases, separately for vertices and midpoints of edges. Our discrete center of mass method requires no such subdivision. \end{remark} \subsection{\texorpdfstring{$\cosh$-center of mass}{cosh-center of mass}} \label{subsec:CoshCenterOfMass} \autoref{thm:CenterOfMassMethodDiscrete} provides an effective method to compute discrete equivariant harmonic maps, alternative to the discrete heat flow (see \autoref{subsec:ConvergenceOfDiscreteHeatFlow}), as long as one is able to compute centers of mass. Unfortunately, non-Euclidean centers of mass are not easily accessible. Even finding the barycenter of three points in the hyperbolic plane is a nontrivial task. While it is possible to use gradient descent method (see \cite{MR3057324}), it is computationally expensive and, in any case, not possible to do precisely in finite time. We present a clever variant to barycenters, well-suited to hyperbolic space $\bH^n$, that avoids this issue. We thank Nicolas Tholozan for bringing this idea to our attention. \begin{definition} \label{def:CoshCenterOfMass} Let $(\Omega, \cF, \mu)$ be a probability space, $(X,d)$ be a metric space, and $h \colon \Omega \to X$ a measurable map. A \emph{$\cosh$-center of mass} of $h$ is a minimizer of the function \begin{equation} \label{eq:DefinitionCoshCenterOfMass} \begin{split} P_h \colon X & \to \bR\\ x & \mapsto \int_\Omega \left(\cosh d(x,h(y))-1 \right) \, \upd \mu(y)~. \end{split} \end{equation} \end{definition} When $X$ is a Riemannian manifold, a $\cosh$-center of mass $G$ is characterized by \begin{equation} \label{eq:ImplicitEquationForCoshCenterOfMass} \int_\Omega \sinhc d(G,h(y)) \exp_G^{-1}(h(y)) \, \upd \mu(y) = 0~, \end{equation} where $\sinhc(x) = \sinh(x)/x$ is the cardinal hyperbolic sine function. Equation \eqref{eq:ImplicitEquationForCoshCenterOfMass} implies that if $\supp(h_*\mu)$ is contained in a strongly convex region $U$ (\eg{} a ball of small enough radius), then any $\cosh$-center of mass is contained in $U$ as well: if $x$ is outside $U$, then each vector $\exp_x^{-1}(h(y))$, for $y\in \supp(h_*\mu)$, is contained in an open half-space in $\upT_x X$ containing $\exp_x^{-1}(U)$, and \eqref{eq:ImplicitEquationForCoshCenterOfMass} cannot be satisfied. Let us now specialize to the case where $X = \bH^n$ is the hyperbolic $n$-space. In this setting the function $F(x) = \cosh(d(x_0,x)) -1$ is especially amenable to computations: \begin{equation} \begin{aligned} &\grad F(x) = \sinhc(d(x_0, x)) \exp_x^{-1}(x_0)\\ &\Hess(F)_x(v,v) = F(x) \Vert v \Vert^2~. \end{aligned} \end{equation} In particular, $F$ is a strongly convex function on $\bH^n$ with modulus of strong convexity $\alpha = 1$. Existence and uniqueness of the center of mass of any function $h \in \upL^2(\Omega, \bH^n)$ quickly follows. The main advantage of the $\cosh$-center of mass is that it admits an explicit description, much like the Euclidean barycenter. For this we work in the hyperboloid model for $\bH^n$, \ie{} \begin{equation} \cH = \{x\in\bR^{n,1}: \langle x,x \rangle=-1,x_{n+1}>0\}~, \end{equation} where Minkowski space $\bR^{n,1}$ is defined as $\bR^{n+1}$ equipped with the indefinite inner product \begin{equation} \langle x, y \rangle = x_1 y_1 + \dots + x_n y_n - x_{n+1} y_{n+1}~. \end{equation} This inner product induces a Riemannian metric on $\cH$ of constant curvature $-1$. \begin{proposition} \label{prop:CoshBarycenter} The $\cosh$-center of mass in $\bH^n \approx \cH$ is equal the orthogonal projection of the Euclidean barycenter in Minkowski space $\bR^{n,1}$ to the hyperboloid $\cH \subset \bR^{n,1}$. \end{proposition} \begin{proof} We prove \autoref{prop:CoshBarycenter} for a finite collection of points for comfort; the generalization to any probability measure is immediate. Consider points $p_1,\dots, p_n\in\cH$ with weights $w_1,\dots, w_n$ satisfying $\sum_i w_i=1$, let $p$ be their Euclidean barycenter in $\bR^{n,1}$, and let $q$ indicate the orthogonal (\ie{} radial) projection of $p$ to $\cH$. By \eqref{eq:ImplicitEquationForCoshCenterOfMass}, it suffices to check that \begin{equation} \label{eq:CoshBarycenterProof} \sum_i w_i \sinhc d(q,p_i) \exp_q^{-1}(p_i) = 0~. \end{equation} Let $P$ be the tangent plane to $\cH$ at $q$, \ie{} the affine plane in $\bR^{n,1}$ which is orthogonal to the line $\bR q$ through $q$. The orthogonal projection $\pi \colon \bR^{n,1} \to P$ is an affine map, so the identity $\sum_i w_i (p_i - p) = 0$ projects to $\sum_i w_i (q_i - q) = 0$ on $P$, where $q_i = \pi(p_i)$. It is straightforward to compute $q = \frac{p}{\sqrt{-\langle p, p\rangle}}$ and $q_i = p_i + q + \frac{\langle p_i, p \rangle}{-\langle p, p \rangle}p$. Geodesics in the hyperboloid are intersections of 2-dimensional subspaces of $\bR^{n,1}$ with $\cH$, so $q_i - q$ is a vector in $\upT_q\cH$ pointing towards $p_i$, and we can compute its length: \begin{equation} \begin{aligned} \Vert q_i - q \Vert^2 &= \langle p_i, q\rangle^2 - 1\\ &=\sinh^2 d_{\cH}(q,p_i) \end{aligned} \end{equation} Thus we proved that \begin{equation} q_i - q = \frac{\sinh d_{\cH}(q,p_i)}{d_{\cH}(q,p_i)}{\exp_q}^{-1}(q_i) \end{equation} and we get \eqref{eq:CoshBarycenterProof} as desired. \end{proof} Another useful feature of the $\cosh$-center of mass is that it is a good approximation of the center of mass for small distances. This will be important in \cite{Gaster-Loustau-Monsaingeon2}. \begin{proposition} \label{prop:CoshCMvsCM} If a probability measure is supported in a ball of radius $r$, then its center of mass $p$ and its $\cosh$-center of mass $q$ and are within $O(r^3)$ of each other. \end{proposition} \begin{proof} Assume $\mu$ has finite support $\{p_1, \dots, p_n\}$ for comfort and denote $w_i = \mu(\{p_i\})$ the weights. By \eqref{eq:ImplicitEquationForCoshCenterOfMass} we can write: \begin{equation} \label{eq:CoshCMvsCM1} \sum_i w_i \exp_q^{-1}(p_i) = \sum_i w_i\left( 1- \sinhc d(p_i, q) \right) \exp_q^{-1}(p_i)~. \end{equation} Because $q$ must be contained in the same ball of radius $r$ as $\{p_i\}$, we find that $d(p_i, q) < 2 r$ for each $i$. Given that $\sinhc$ is a nondecreasing function we derive from \eqref{eq:CoshCMvsCM1} \begin{align} \left\Vert \sum_i w_i \exp_q^{-1}(p_i) \right\Vert & \leqslant \sum_i w_i \left(\sinhc d(p_i, q) -1\right) \left\Vert \exp_q^{-1}(p_i) \right\Vert \\ & \leqslant \sum_i w_i \left(\sinhc (2 r) - 1\right) \cdot (2 r) = 2 r \left(\sinhc (2 r) - 1\right)~. \end{align} \autoref{lem:ProximityToCenterOfMass} now implies that \begin{equation} d(p,q) \leqslant 2 r \left(\sinhc (2 r) - 1\right)~. \end{equation} The conclusion follows, since $2 r \left(\sinhc (2 r) - 1\right) = \frac{4 r^3}{3} + O( r^5)$. \end{proof} Note that we did not use in the proof of \autoref{prop:CoshCMvsCM} that we are working in $\bH^n$: this proposition holds in any Riemannian manifold. \section{Computer implementation: \texttt{Harmony}{}} \label{sec:Harmony} \subsection{Computer description and availability} \texttt{Harmony}{} is a computer program developed by the first two authors (Jonah Gaster and Brice Loustau). It is a cross-platform software with a graphical user interface written in \Cpp{} code using the Qt framework. In its current state, it totals about $14,000$ lines of code. \texttt{Harmony}{} is a free and open source software under the GNU General Public License. It is available on GitHub at \url{https://github.com/seub/Harmony}. For more information, including install instructions and a quick guide to get started using the software, please visit the dedicated web site: \url{https://www.brice.loustau.eu/harmony/}. \subsection{Algorithms} \label{subsec:Algorithms} We provide a brief overview of \texttt{Harmony}{}'s algorithms allowing effective computation of discrete equivariant harmonic maps $\bH^2 \to \bH^2$ with respect to a pair of Fuchsian representations. A flowchart showing how the main algorithms fit into the program is pictured in \autoref{fig:Flowchart}. We fix an identification of the closed oriented topological surface $S$ of genus $g$ as $P_0/\sim$, where $P_0$ is a topological $4g$-gon with oriented sides labelled $a_1$, $b_1$, $a_1^{-1}$, $b_1^{-1}$, $a_2$, \etc{}. We parametrize hyperbolic structures on $S$ using the famous Fenchel-Nielsen coordinates. This requires choosing a pants decomposition of $S$. \texttt{Harmony}{} is equipped to make such choices for arbitrary $g$ in a way that minimizes future error propagation. The input is a pair of Fenchel-Nielsen coordinates for hyperbolic structures $X$ and $Y$ on $S$, the domain and target hyperbolic surfaces respectively. These can be entered by the user in a `Fenchel-Nielsen selector' window: see \autoref{fig:FNselector}. \subsubsection*{Step 1: Construct the fundamental group and pants decomposition} After getting the genus $g$ as input, \texttt{Harmony}{} constructs the fundamental group of the surface as an abstract structure. It then chooses a pants decomposition of the surface, yielding a decomposition of the fundamental group in terms of amalgamated products and HNN extensions of fundamental groups of pairs of pants. This is done recursively on the genus using a binary tree structure. \subsubsection*{Step 2: Construct representations $\rho_X$ and $\rho_Y$} This step performs the translation of Fenchel-Nielsen coordinates to Fuchsian representations. \texttt{Harmony}{} starts by computing the representation of the fundamental group of each pair of pants using formulas that can be found in \eg{} \cite[Prop. 2.3]{MR1288062} or \cite{MR1703565,MR1833242}. It then computes the representation of the whole fundamental group using its decomposition discussed in Step 1. \subsubsection*{Step 3: Construct fundamental domains $P_X$ and $P_Y$} This step computes polygonal fundamental domains $P_X$ and $P_Y$ in $\bH^2$ for the Fuchsian groups in the images of $\rho_X$ and $\rho_Y$. These polygons should be `as convex as possible' in order to ensure good behavior of the discrete heat flow. Both $P_X$ and $P_Y$ come with $\pi_1S$-equivariant identifications to the topological $4n$-gon $P_0$ that record side pairings. Because the vertices of $P_0$ are all in the same $\pi_1S$-orbit, $P_X$ is determined by a single point in $\bH^2$. With this combinatorial setup, a best choice of polygon is obtained minimizing an adequate cost function $F:\bH^2\to \bR^+$. This is done with a straightforward Newton method. \subsubsection*{Step 4: Construct a triangulation of $P_X$} This step computes a triangulation of the fundamental domain $P_X$. Finer and finer meshes can then be obtained by subdivision (see \autoref{def:Refinement}). As explained in \cite{Gaster-Loustau-Monsaingeon2}, it is crucial to keep the smallest angle of the triangulation as large as possible. Unfortunately, $P_X$ already typically has very small angles. In order to avoid subdividing these angles further, we first introduce new \emph{Steiner vertices} evenly spaced along the sides of $P_X$. The resulting polygon is triangulated with a greedy recursive algorithm maximizing the smallest angle. See \autoref{fig:FNselector} for a sample output. \begin{remark} The algorithm seems to always produce acute triangulations, a necessary condition for the definition of the edge weights (see \autoref{subsec:Meshes}), but we do not know whether this always holds. Acute triangulations of surfaces are part of a fascinating area of current research \cite{MR1064872,MR301697 }. \end{remark} \subsubsection*{Step 5: Construct the $\rho_X$-invariant mesh $\cM$} The mesh $\cM$ consists of a list of meshpoints $\cM^{(0)} \subset \bH^2$, each of which is equipped with a list of references to its neighboring meshpoints, and possibly side-pairing information. This data is initially recorded from the triangulated polygon $P_X$. Then, given a user chosen \emph{mesh depth} $k\geqslant 1$, $\cM$ is replaced with the $k$th iterated midpoint refinement of $\cM$ (see \autoref{def:Refinement}). \begin{remark} Constructing the adequate data structure to efficiently store the mesh data is a difficult challenge: the corresponding \Cpp{} classes are the most sophisticated in the code of \texttt{Harmony}{}. \end{remark} \subsubsection*{Step 6: Initialize an equivariant map} \label{triangulate Y} The triangulation of $P_X$ may be transported to one for $P_Y$ via the $\pi_1S$-equivariant maps \begin{equation} P_X \approx P_0 \approx P_Y~. \end{equation} Because the mesh is built from midpoint refinements, this identification provides an initial discrete equivariant map $f_0 \colon \bH^2 \to \bH^2$. A sample initial map is showed in \autoref{fig:InitialFunctionPic}. \subsubsection*{Step 7: Run the discrete flow} \texttt{Harmony}{} is ready to run: either the discrete heat flow with fixed or optimal stepsize (see \autoref{subsec:ConvergenceOfDiscreteHeatFlow}) or the $\cosh$-center of mass method (\autoref{subsec:CenterOfMassMethods}, \autoref{subsec:CoshCenterOfMass}). \texttt{Harmony}{} uses several threads so that the flow is displayed ``live'' as it is being computed. See \autoref{fig:MidFlow} for a screenshot of \texttt{Harmony}{} mid-flow. The flow is iterated until the error reaches a preset tolerance, or when it is stopped by the user. Both the discrete heat flow and the center of mass method typically converge very well. Refer to \autoref{subsec:ExperimentalComparison} for a comparison of the methods. \autoref{fig:Output} shows a sample output equivariant harmonic map. \subsection{High energy harmonic maps between hyperbolic surfaces} \label{subsec:HighEnergy} In this final subsection we briefly explain how \texttt{Harmony}{} provides visual confirmation of a qualitative phenomenon that is well-known in Teichmüller theory. We hope that future development of the program will allow many more experimental investigation of theoretical aspects. We refer to \cite{MR2349668} for more details on what follows. As mentioned in \autoref{subsec:harmonicMapsSurfaces}, taking the Hopf differential of harmonic maps allows one to parametrize Teichmüller space by holomorphic quadratic differentials. More precisely, given a closed oriented $S$ of genus $\geqslant 2$, let $\FS$ denote the Fricke-Klein space of $S$, \ie{} the deformation space of hyperbolic structures on $S$ up to isotopically trivial diffeomorphisms. Fix a complex structure $X$ on $S$, and denote $Q(X)$ the vector space of holomorphic quadratic differentials on $X$. For any $\sigma \in \FS$, there is a unique harmonic map $f_\sigma \colon X \to (S, \sigma)$. Taking its Hopf differential yields a map \begin{equation} \label{eq:WolfParam} \begin{split} H \colon \FS &\to Q(X)\\ \sigma & \mapsto \varphi_{f_\sigma}~. \end{split} \end{equation} Wolf \cite{MR982185} proved that the map $H$ is a global diffeomorphism from $\FS$ to $Q(X)$. Furthermore, $H$ continuously extends to the ``boundaries at infinity'': \begin{equation} \label{eq:WolfParamCompactified} \partial H \colon \partial\FS \to \partial Q(X)~. \end{equation} Here the boundary $\partial\FS$ compactifying Fricke-Klein space is the \emph{Thurston boundary} $\partial \FS \coloneqq \mathcal{PMF}(S)$, the projective space of measured foliations on $S$. The boundary of the vector space $Q(X)$ is simply its projectivization: $\partial Q(X) \coloneqq \bP(Q(X))$. It can be identified to the projective space of measured laminations $\mathcal{PML}(S)$ by assigning to any holomorphic quadratic differential its horizontal foliation. Thus the boundary map $\partial H$ of \eqref{eq:WolfParamCompactified} may be described as a map \begin{equation} \partial H \colon \mathcal{PMF}(S) \to \mathcal{PML}(S)~. \end{equation} This map has a nice geometric interpretation as the well-known ``pull tight'' map which assigns to each non-singular leaf of a measured foliation the unique geodesic in its homotopy class. Concretely, this means that a high energy harmonic map $f \colon X \to (S, \sigma)$ typically has a specific behavior dictated by its Hopf differential $\varphi$, namely: the zeros of $\varphi$ are blown up to ideal hyperbolic triangles, while the rest of the surface is compressed onto the measured lamination dual to $\varphi$. This behavior is well verified by \texttt{Harmony}{}. As an example, \autoref{fig:HighEnergy} shows the image of a high energy harmonic map: observe how the most contracted (darker) regions approach a geodesic lamination, while the most dilated (lighter) regions approach a union of ideal triangles. \newpage \newgeometry{top=0.09\paperheight, bottom=0.11\paperheight, left=0.06\paperheight, right=0.06\paperheight} \subsection{Illustrations of \texttt{Harmony}{}} \begin{figure}[!ht] \centering \includegraphics[width=\textwidth]{Pic3a-reduced.png} \caption{\texttt{Harmony}{}'s main user interface.} \label{fig:HarmonyUserInterface} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=0.79\textwidth]{Pic5raw-reduced.png} \caption{Fenchel-Nielsen coordinates selector.} \label{fig:FNselector} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=1.0\textwidth]{Pic4a-reduced.png} \caption{An initial discrete equivariant map $f_0$. The highlighted blue triangles are matched.} \label{fig:InitialFunctionPic} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=\textwidth]{Pic2a-reduced.png} \caption{\texttt{Harmony}{} mid-flow.} \label{fig:MidFlow} \end{figure} \begin{figure}[!htp] \centering \includegraphics[width=\textwidth]{Pic6raw-reduced.png} \caption{Sample output harmonic map. The brighter central regions are fundamental domains.} \label{fig:Output} \end{figure} \begin{figure}[!htp] \centering \includegraphics[width=0.73\textwidth]{Pic12b1-reduced.png} \caption{Sample output high energy equivariant harmonic map.} \label{fig:HighEnergy} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=0.75\textwidth]{flowChartNoBorder.pdf} \caption{Flowchart representing \texttt{Harmony}{}'s main algorithms.} \label{fig:Flowchart} \end{figure} \restoregeometry \section{Discrete heat flow} \label{sec:DiscreteHeatFlow} \subsection{Gradient descent in Riemannian manifolds} \label{subsec:Gradient descent} The area of mathematics concerned with methods for finding the minima of a convex function $F \colon \Omega \to \bR$, called \emph{convex optimization}, has been intensely developed in the last few decades and finds countless applications. The majority of the existing literature deals with the classical case where $\Omega$ is a subset of a Euclidean space; the Riemannian setting has been far less explored although it is a natural and useful extension. Udri\c{s}te's book \cite{MR1326607} is a good standard reference for Riemannian convex optimization (see \eg{} \cite{MR2364186}, \cite{ZhangSra2016} for more recent developments). Our goal is to present a simple but effective method that can be implemented to find the minimum of the discrete energy functional, with a rigorous proof of convergence and explicit control of the convergence rate. Of course, there are more advanced and faster algorithms in practice. For example, the C++ library \emph{ROPTLIB} \cite{ROPTLIB} was developed for this purpose. \smallskip A \emph{gradient descent method} is an iterative algorithm for minimizing a function $F \colon \Omega \subseteq \bR^N \to \bR$ which produces a sequence $(x_k)_{k \geqslant 0}$ of points in $\Omega$, defined inductively by: \begin{equation} \label{eq:GradientDescentMethodEuclidean} x_{k+1} = x_k - t_k \grad F(x_k)~. \end{equation} In this relation, $t_k \geqslant 0$ is a chosen \emph{stepsize}. If $F$ has good convexity properties such as being strongly convex, then a small enough stepsize $t_k = t > 0$ guarantees convergence of the sequence $(x_k)$ to a minimum of $F$, with explicit control of the convergence rate. This method naturally extends to the Riemannian setting, \ie{} when $F \colon \Omega \subseteq M \to \bR$ is defined on a subset $\Omega$ of a Riemannian manifold $M$, in which case \eqref{eq:GradientDescentMethodEuclidean} should be understood as \begin{equation} \label{eq:GradientDescentMethodRiemannian} x_{k+1} = \exp_{x_k}( - t_k \grad F(x_k))~. \end{equation} As in the Euclidean setting, $-\grad F(x_k)$ is the direction of steepest descent for $F$ at $x_k$, so it is natural to look for $x_{k+1}$ in the geodesic ray based at $x_k$ in this direction. Note that the gradient descent method can simply be described as Euler's method for the gradient flow ODE: \begin{equation} x'(t) = -\grad F(x(t))~. \end{equation} \subsubsection*{Gradient descent method with fixed stepsize for strongly convex functions} The gradient descent method with fixed stepsize remains valid for $\cC^2$ strongly convex functions on Riemannian manifolds: \begin{theorem}[{\cite[Chap. 7, Theorem 4.2]{MR1326607}}] \label{thm:GradientDescentFixedStepRiemannian} Let $(M,g)$ be a complete Riemannian manifold and let $F \colon M \to \bR$ be a function of class $\cC^2$. Assume that there exists $\alpha, \beta >0$ such that: \begin{equation} \forall v \in \upT M \quad \alpha \, \Vert v\Vert^2 \leqslant (\Hess F)(v,v) \leqslant \beta \, \Vert v\Vert^2 \end{equation} Then $F$ has a unique minimum $x^*$. Furthermore, for $t \in (0, \frac{1}{\beta}]$, the gradient descent method with fixed stepsize $t_k = t$ converges to $x^*$ with a linear convergence rate: \begin{equation} \label{eq:ConvergenceRateFixedStepsize} d(x_k, x^*) \leqslant c \, q^k \end{equation} The constants $c \geqslant 0$ and $q \in [0,1)$ are given by: \begin{equation}\label{eq:ConstantsConvergenceRate} c = \sqrt{\frac{2}{\alpha}(F(x_0)-F(x^*))} \qquad q = \sqrt{1 - \frac{t}{2} \alpha\left(1 + \frac{\alpha}{\beta}\right)}~. \end{equation} \end{theorem} \begin{remark} A key step in the proof of \autoref{thm:GradientDescentFixedStepRiemannian} is that $(F(x_k) - F(x^*))$ is nonincreasing and limits to $0$ with a linear convergence rate. In particular, $(F(x_k))_{k \geqslant 0}$ is nonincreasing, therefore any sublevel set of $F$ is stable under the gradient descent. Moreover, such a set is convex and compact by strong convexity of $F$. Thus, the gradient descent method is valid even if the Hessian of is not bounded above: one can restrict to a sublevel set, where the Hessian of $F$ is bounded. \end{remark} \subsubsection*{Gradient descent method with optimal stepsize for strongly convex functions} There are many variants of the gradient descent method that can be more or less useful depending on the context (see \eg{} \cite{ZhangSra2016}, \cite{FerreiraLouzeiroPrudente}). One of them is the \emph{optimal stepsize gradient descent}, an instance of the gradient descent method \eqref{eq:GradientDescentMethodRiemannian} where one performs a \emph{line search} in order to determine a stepsize $t_k$ that minimizes $F(x_{k+1})$. Clearly, when $F$ is strongly convex, such a $t_k$ exists, is unique, and is $> 0$ unless $x_k = x^*$. When the Hessian of $F$ is known analytically, Newton's method offers a very fast line search. \begin{theorem} \label{thm:OptimalStepRiemannian} Let $F \colon M \to \bR$ as in \autoref{thm:GradientDescentFixedStepRiemannian}. The optimal stepsize gradient descent has a linear convergence rate at least as fast as that of \autoref{thm:GradientDescentFixedStepRiemannian}, for any choice of the fixed stepsize. \end{theorem} \begin{proof} \autoref{thm:OptimalStepRiemannian} is derived from a careful analysis of the proof of \autoref{thm:GradientDescentFixedStepRiemannian} which can be found in \cite[Chapter 7, Theorem 4.2]{MR1326607}. This proof is a combination of three observations: \begin{enumerate}[(i)] \item For any $x\in M$: \begin{equation} \label{eq:Observation1} \frac{\alpha}{2} d(x, x^*)^2 \leqslant F(x) - F(x^*) \leqslant \frac{\beta}{2} d(x, x^*)^2~. \end{equation} This follows from a Taylor expansion of $F$ at $x^*$ along the geodesic $[x^*, x]$. \item For any $x\in M$: \begin{equation} \label{eq:Observation2} \Vert \grad F(x) \Vert^2 \geqslant \alpha\left(1 + \frac{\alpha}{\beta}\right)(F(x) - F(x^*))~. \end{equation} This follows from a Taylor expansion of $F$ at $x$ along the geodesic $[x, x^*]$ and from \eqref{eq:Observation1}. \item For any $x\in M$ and for any $t \in [0, \frac{1}{\beta}]$: \begin{equation} \label{eq:Observation3} F(x) - F(x^+(t)) \geqslant \frac{t}{2} \Vert \grad F(x) \Vert^2 \end{equation} where $x^+(t) \coloneqq \exp_x(-t \grad F(x))$. This follows from a Taylor formula for $F$ along $[x, x^+(t)]$. \end{enumerate} It follows immediately from these three observations that for any $x\in M$ and for any $t \in [0, \frac{1}{\beta}]$: \begin{equation} \label{eq:Obs4} F(x^+(t)) - F(x^*) \leqslant Q(t) \left(F(x) - F(x^*)\right) \end{equation} where $Q(t) = 1 - \frac{t}{2} \alpha(1 + \frac{\alpha}{\beta}) = q^2$. When one performs the gradient descent method with fixed stepsize $t$, by assumption $x_{k+1} = x_k^+(t)$. \autoref{thm:OptimalStepRiemannian} is then easily concluded by finding $F(x_k) - F(x^*) \leqslant Q^k \left(F(x_0) - F(x^*)\right)$ from \eqref{eq:Obs4} (with an obvious induction) and making one last use of \eqref{eq:Observation1}. If instead one performs an optimal stepsize gradient descent, then $x_{k+1} = x_k^+(t_k)$, where $t_k$ is the optimal step. Fix $t \in [0, \frac{1}{\beta}]$. By definition of the optimal step, $F(x_k^+(t_k)) \leqslant F(x_k^+(t))$, so $F(x_{k+1}) - F(x_k) \leqslant F(x_k^+(t)) - F(x_k)$. Therefore we can derive from \eqref{eq:Obs4} that \begin{equation} F(x_{k+1}) - F(x^*) \leqslant Q(t) \left(F(x_k) - F(x^*)\right) \end{equation} and the conclusion follows like before. \end{proof} \subsection{Convergence of the discrete heat flow}\label{subsec:ConvergenceOfDiscreteHeatFlow} The discrete heat flow can be described as a discretization both in time and space of the heat flow on $\cC^\infty(M,N)$ . Recall that the smooth heat flow is the gradient flow of the smooth energy functional: \begin{equation} \label{eq:HeatFlowODE2} \ddt f_t = \tau(f_t) \end{equation} where $\tau(f_t) = - \grad E(f_t)$ is the tension field of $f_t$ (cf \autoref{subsec:EnergyFunctionalHarmonicMaps}, \autoref{subsec:SmoothHeatFlow}). We recall the setup of our discretization: Let $\cG$ be a biweighted $\tilde{S}$-triangulated graph (\autoref{def:BiweightedGraph}), let $N$ be a Riemannian manifold, and let $\rho \colon \pi_1S \to \Isom(N)$ be a group homomomorphism. Recall that the discrete energy is a function $E_\cG \colon \Map_{\text{eq}}(\cG, N) \to \bR$ (\autoref{def:EnergyFunctionalGraph}), where $\Map_{\text{eq}}(\cG, N)$ is the space of $\rho$-equivariant maps $\cG \to N$ . The latter space has a natural Riemannian structure with respect to which the gradient of the energy is minus the discrete tension field $\tau_\cG$ (\autoref{def:L2RiemannianMetricGraph} and \autoref{prop:FirstVariationalFormulaGraph}). Thus we define the discrete heat flow: \begin{definition} \label{def:DiscreteHeatFlow} The \emph{discrete heat flow} is the iterative algorithm which, given $f_0 \in \Map_{\text{eq}}(\cG, N)$, produces the sequence $(f_k)_{k \in \bN}$ in $\Map_{\text{eq}}(\cG, N)$ defined inductively by the relation \begin{equation} \label{eq:DiscreteHeatFlow1} f_{k+1}(x) = \exp_{f_k(x)}\left(t_k (\tau_\cG f_k)_x\right)~, \end{equation} where $t_k \in \bR$ is a chosen stepsize. \end{definition} The main theorem of this section is an immediate application of \autoref{thm:StrongConvexityEnergyGraphGeneral} and \autoref{thm:GradientDescentFixedStepRiemannian}: \begin{theorem} \label{thm:ConvergenceDiscreteHeatFlowGeneral} Let $\cG$ be a biweighted $\tilde{S}$-triangulated graph. Let $N$ be two-dimensional Hadamard manifold of nonpositive curvature and $\rho \colon \pi_1S\to \Isom(N)$ is a faithful representation whose image is contained in a discrete subgroup of $\Isom(N)$ acting freely and properly. Then there exists a unique $\rho$-equivariant harmonic map $f^* \colon \cG \to N$. Moreover, for any $f_0 \in \Map_{\text{eq}}(\cG, N)$ and for any sufficiently small $t>0$, the discrete heat flow with initial value $f_0$ and fixed stepsize $t$ converges to $f^*$ with a linear convergence rate: \begin{equation} \label{eq:ConvergenceRateDHFGeneral} d(f_k, f^*) \leqslant c q^k \end{equation} where $c>0$ and $q \in [0, 1)$ are constants, and $d(f_k, f^*)$ is the $\upL^2$ distance in $\Map_{\text{eq}}(\cG, N)$. \end{theorem} Of course, it also follows from \autoref{thm:OptimalStepRiemannian} that the discrete heat flow with optimal stepsize converges to $f^*$ as well, with a linear convergence rate as least as fast as \eqref{eq:ConvergenceRateDHFGeneral}. We emphasize that in our favorite setting where $N = \bH^2$ and $\rho \colon \pi_1S \to \Isom^+(\bH^2)$ is Fuchsian, \autoref{thm:StrongConvexityEnergyGraphH2} enables explicit estimates on the constants $c$ and $q$ in \eqref{eq:ConvergenceRateDHFGeneral}: the expressions of $c$ and $q$ are given by \eqref{eq:ConstantsConvergenceRate}, in which $\alpha$ is given by \eqref{eq:SecondDerivativeLowerBound1} and $\beta$ is given by \eqref{eq:SecondDerivativeUpperBound} with $E_0 = E(f_0)$. \subsection{Experimental comparison of convergence rates} \label{subsec:ExperimentalComparison} In \autoref{fig:graphConstants} and \autoref{fig:graphComparison} we present numerical experiments performed with the software $\texttt{Harmony}$. \subsubsection*{Comparison of different fixed stepsizes} In the first experiment (\autoref{fig:graphConstants}) we observe the number of iterations required for the discrete heat flow with fixed stepsize to converge as a function of the stepsize. Let $S$ be a closed oriented surface of genus $2$. We fix a domain Fuchsian representation $\rho_\tL \colon \pi_1S \to \Isom^+(\bH^2)$: the representation pictured on the left in \autoref{fig:InitialFunctionPic}, and let $\texttt{Harmony}$ construct an invariant mesh (depth $4$, $1921$ vertices). We let the target Fuchsian representation $\rho_\tR$ vary, taking Fenchel-Nielsen lengths $(2,2,\ell)$ and twists $(-1.5,2,0.5)$, where $\ell \in \{2.5, 1.5, 0.5, 0.2\}$. We observe that the plotted points resemble in profile functions of the form $-C_1\left( \log(1-C_2t) \right)^{-1}$, which is precisely the type of function predicted by \autoref{thm:ConvergenceDiscreteHeatFlowGeneral}. \begin{figure} \centering \begin{lpic}{graphConstantStepPerformance2(10cm)} \lbl[]{29,-7;$.01$} \lbl[]{58,-7;$.02$} \lbl[]{87,-7;$.03$} \lbl[]{116,-7;$.04$} \lbl[]{145,-7;$.05$} \lbl[]{-6.5,18;$10^4$} \lbl[]{-12,36;$2\times 10^4$} \lbl[]{-12,54;$3\times 10^4$} \lbl[]{-12,72;$4\times 10^4$} \lbl[]{-12,90;$5\times 10^4$} \lbl[]{-12,108;$6\times 10^4$} \lbl[]{184,97;$\ell=2.5$} \lbl[]{184,85.5;$\ell=1.5$} \lbl[]{184,74;$\ell=0.5$} \lbl[]{184,62.5;$\ell=0.2$} \lbl[]{-10,125;number of iterations} \lbl[]{200,0;stepsize} \end{lpic} \vspace{.8cm} \caption{Number of iterations against stepsize in the discrete heat flow with fixed stepsize performed by \texttt{Harmony}{}.} \label{fig:graphConstants} \end{figure} \subsubsection*{Comparison of our three methods} \begin{figure} \centering \begin{lpic}{graphComparison2(10cm)} \lbl[]{-6,115;number of iterations} \lbl[]{-12,26;$5\times 10^3$} \lbl[]{-12,50;$1\times 10^4$} \lbl[]{-12,74;$5\times 10^4$} \lbl[]{-12,96;$1\times 10^5$} \lbl[]{31,-5;$1$} \lbl[]{67,-5;$2$} \lbl[]{103,-5;$3$} \lbl[]{138,-5;$4$} \lbl[]{170,0;$\ell$} \lbl[l]{168,84;fixed stepsize} \lbl[l]{168,69;optimal stepsize} \lbl[l]{168,54;cosh-center of mass} \end{lpic} \vspace{.8cm} \caption{Comparison of the three methods performed by \texttt{Harmony}{}.} \label{fig:graphComparison} \end{figure} For the second experiment (\autoref{fig:graphComparison}) we compare the convergence rate, in terms of number of iterations, of our three methods: \begin{itemize} \item Discrete heat flow with fixed stepsize (see \autoref{subsec:ConvergenceOfDiscreteHeatFlow}), \item Discrete heat flow with optimal stepsize (see \autoref{subsec:ConvergenceOfDiscreteHeatFlow}), \item Cosh-center of mass method (see \autoref{subsec:CenterOfMassMethods}). \end{itemize} We keep the same setting as before, letting $\ell$ this time vary between $0.2$ and $4.4$. As the figure shows, the cosh-center of mass method is more effective than either gradient descent methods. \section{Discretization} \label{sec:Discretization} We fix some notation: for the remainder of the paper, $S$ is a smooth, closed, oriented surface of negative Euler characteristic (genus $\geqslant 2$). We denote $\pi_1S$ the fundamental group of $S$ with respect to some basepoint that can be safely ignored. \begin{remark} While this paper specializes the discretization to $2$-dimensional hyperbolic surfaces, most of the definitions can be generalized to higher-dimensional Riemannian manifolds. The sequel paper \cite{Gaster-Loustau-Monsaingeon2} treats the general Riemannian setting (while dropping the equivariant formulation, mostly for comfort). \end{remark} In \autoref{subsec:Meshes} we explain how to approach the energy minimization problem for smooth equivariant maps $\bH^2 \to N$ in order to allow effective computation by introducing meshes and subdivisions, discrete equivariant maps, and discrete energy. Several of these notions are further discussed in \cite{Gaster-Loustau-Monsaingeon2}. In the present paper they serve as a preamble to the more formal setting we develop in \autoref{subsec:Graphs} and they justify the choices made in the software \texttt{Harmony}{}. \subsection{Meshes and discrete harmonic maps} \label{subsec:Meshes} Let us fix a hyperbolic structure on $S$ given by a Fuchsian representation $\rho_\tL$, \ie{} an injective group homomorphism $\pi_1S \to \Isom^+(\bH^2)$ with discrete image. This setup can be easily generalized to any Riemannian metric on $S$, but the hyperbolic metric is best suited for computations. \subsubsection*{Meshes and subdivisions} Given a group homomorphism $\rho_\tR \colon \pi_1S \to \Isom(N)$ where $N$ is a Riemannian manifold (or a metric space), we would like to discretize $(\rho_\tL, \rho_\tR)$-equivariant maps $\bH^2 \to N$. To this end, we start by discretizing the domain hyperbolic surface with the notion of invariant mesh: \begin{definition} \label{def:Mesh} A \emph{$\rho_\tL$-invariant mesh} of $\bH^2$ is an embedded graph $\cM$ in $\bH^2$ such that: \begin{enumerate}[(i)] \item The vertex set $\cM^{(0)} \subset \bH^2$ (set of \emph{mesh points}) is invariant under a cofinite action of $\rho_\tL(\pi_1S)$. \item Every edge $e \in \cM^{(1)}$ is an embedded geodesic segment in $\bH^2$. \item The complementary components are triangles. \end{enumerate} \end{definition} For the purpose of approximating smooth maps, we will need to take finer and finer meshes. This will be discussed in detail in \cite{Gaster-Loustau-Monsaingeon2}, but let us describe the strategy that we have implemented in the software \texttt{Harmony}. A natural way to obtain a finer mesh from a given one is via geodesic subdivision. We indicate below an edge of $\cM$ with endpoints $x,y\in \cM^{(0)}$ by $e_{xy}$, and let $m(x,y) \in \bH^2$ be the midpoint of $x$ and $y$. It is easy to see that the following is well-defined: \begin{definition} \label{def:Refinement} The \emph{refinement} of a $\rho_\tL$-invariant mesh $\cM$ is the $\rho_\tL$-invariant mesh $\cM'$ such that: \begin{enumerate}[(i)] \item The vertices of $\cM'$ are the vertices of $\cM$ plus all midpoints of edges of $\cM$. \item The edges of $\cM'$ are given by $x\sim m(x,y)$ and $y\sim m(x,y)$ for each edge $e_{xy}$, and $m(x,y) \sim m(x,z)$ for each triple of vertices $x,y,z$ that span a triangle in $\cM$. \end{enumerate} \end{definition} Evidently, this refinement may be iterated. See \autoref{fig:meshes} for an illustration of a $\rho_\tL$-invariant mesh and its refinement generated by the software \texttt{Harmony}{}. \begin{figure} \centering \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=.98\textwidth]{mesh0-eps-converted-to.pdf} \caption{An invariant mesh of $\bH^2$} \label{fig:mesh0} \end{subfigure}% \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=.98\textwidth]{mesh1-eps-converted-to.pdf} \caption{Refinement of order $1$} \label{fig:mesh1} \end{subfigure} \caption{An invariant mesh of the Poincaré disk model of $\bH^2$ on the left, its refinement of order 1 on the right. The brighter central region is a fundamental domain. The blue circle arcs are the axes of the generators of $\rho_\tL(\pi_1S)$. Pictures generated by \texttt{Harmony}{}.} \label{fig:meshes} \end{figure} \subsubsection*{Discrete equivariant maps} Given a $\rho_\tL$-invariant mesh $\cM$ and a group homomorphism $\rho_\tR \colon \pi_1S \to \Isom(N)$, we call \emph{discrete equivariant map $\bH^2 \to N$ along $\cM$} a $(\rho_\tL, \rho_\tR)$-equivariant map from the vertex set $\cM^{(0)}$ to $N$. We denote $\Map_{\cM}(\bH^2, N)$ the set of discrete equivariant maps $\bH^2 \to N$ along $\cM$. Note that $\Map_{\cM}(\bH^2, N) \approx N^{V}$ where $V = \cM^{(0)}/\rho_\tL(\pi_1S)$ is the set of equivalence classes of meshpoints, which is finite. Therefore: \begin{proposition} \label{prop:EquivMapsSmoothManifold1} If $N$ is a finite-dimensional smooth manifold, then so is $\Map_{\cM}(\bH^2, N)$. \end{proposition} Denoting $\cC_\text{eq}^0(\bH^2,N)$ the space of continuous equivariant maps $\bH^2 \to N$, the forgetful map \begin{equation} \label{eq:ForgetfulMap} \cC_\text{eq}^0(\bH^2,N) \to \Map_{\cM}(\bH^2, N) \end{equation} consists in restricting a continuous function to the set of meshpoints $\cM^{(0)}$. \begin{definition} \label{def:InterpolationScheme} We shall call a right inverse of the forgetful map \eqref{eq:ForgetfulMap} an \emph{interpolation scheme}. \end{definition} While there is one most natural way to interpolate discrete maps between Euclidean spaces (affine interpolation), there is no preferred way for arbitrary Riemannian manifolds. Even in the case where both domain and target manifolds are the hyperbolic plane $\bH^2$, there are several reasonable interpolations to consider such as the barycentric interpolation and the harmonic interpolation. However, these are not explicit, and with \texttt{Harmony}{} we work with a neat variant, the $\cosh$-center of mass interpolation (see \autoref{subsec:CoshCenterOfMass}). \subsubsection*{Discretization of energy} For a smooth $(\rho_\tL,\rho_\tR)$-equivariant map $f:\bH^2\to N$ where $N$ is a Riemannian manifold, one defines the total energy of $f$ as: \begin{equation} \label{eq:EnergyFunctionalSmoothEquivariant} E(f) = \int_D \Vert \upd f \Vert^2 \, \upd v_g \end{equation} where $D \subset \bH^2$ is any fundamental domain for the action of $\pi_1S$. If $D$ is picked so that it coincides with a union of triangles defined by $\cM$, then the energy can be written as a finite sum of energy integrals over each triangle. When $f$ is discretized along $\cM$, only the values of $f$ on the meshpoints are recorded. Thus, a natural discretization of $E$ is obtained if one knows how to define the energy of a map from a triangle whose values are only known at the vertices. Given an interpolation scheme (cf \autoref{def:InterpolationScheme}), one can simply take the energy of the interpolated map. Another approach consists in defining the energy \emph{à la} Jost / Korevaar-Schoen as in \autoref{subsec:JostEnergy}. One can take the graph defined by $\cM$ with metric induced from $\bH^2$ as a domain metric space, introduce a measure that approximates the area density of $\bH^2$, and choose an appropriate kernel $\eta(x,y)$. A third approach consists in choosing a discrete energy that is a weighted sum of distances squared as in \autoref{def:DiscreteEnergyMesh} and \autoref{def:EnergyFunctionalGraph}. This approach provides a natural extension of the classical notion of real-valued harmonic functions defined on graphs \cite{MR2882891,MR1421568,MR1829620}, that is, functions whose value at any vertex is the average of the values on the neighbors. It turns out that all three approaches can be made to coincide (\autoref{prop:JostEnergyGraph}), or almost coincide for fine meshes, for the appropriate choices involved in the different definitions. This is thoroughly discussed in the sequel paper \cite{Gaster-Loustau-Monsaingeon2}. \begin{definition} \label{def:DiscreteEnergyMesh} Let $\cM$ be a $\rho_\tL$-invariant mesh in $\bH^2$ such that all the complementary triangles are acute. The \emph{discrete energy} of a discrete equivariant map $f\in \Map_\cM(\bH^2, N)$ is defined by \begin{equation} \label{eq:DiscreteEnergyMesh} E_{\cM}(f) = \frac{1}{2} \sum_{e = e_{xy} \in \cE} \omega_{xy} \, d(f(x),f(y))^2 \end{equation} where: \begin{itemize} \item $\cE \subset \cM^{(1)}$ is any fundamental domain for the action of $\pi_1S$ on the set of edges. \item Inside the sum, $x$ and $y$ are the vertices connected by the edge $e$. \item $\omega_{xy}$ is the half-sum of the cotangents of two angles: one for each of the two triangles sharing the edge $e=e_{xy}$, in which we take the angle of the vertex facing the edge $e$. \end{itemize} \end{definition} This definition is a generalization of the energy considered by Pinkall-Polthier \cite{MR1246481}, for whom the domain is a triangulated surface with a piecewise Euclidean metric and $N = \bR^n$. In their setting, the discrete energy coincides with the energy relative to the linear interpolation. In \cite{Gaster-Loustau-Monsaingeon2} we show that the discrete energy $E_\cM$ converges to the smooth energy $E$ under iterated refinement of the mesh $\cM$ in a suitable function space. Of course, we can now define a \emph{discrete equivariant harmonic map} $f\in \Map_\cM(\bH^2, N)$ as a critical point of the discrete energy functional $E_\cM$. While several authors have shown the existence and uniqueness of minimizers of the discrete energy in various contexts (\eg{} \cite{MR1775138,MR1848068,MR1938491}), in this paper we analyze its strong convexity, which makes it better suited for effective minimization. Our approach requires a Riemannian metric on $\Map_\cM(\bH^2, N)$ (cf.~\autoref{subsec:ConvexityRiemannian}), which should approach the $\upL^2$ Riemannian metric of $\mathcal{C}^\infty(M,N)$ (cf.~\autoref{subsec:EnergyFunctionalHarmonicMaps}). In the next subsection, we develop a more general framework where these ideas apply. \subsection{Equivariant harmonic maps from graphs} \label{subsec:Graphs} The definition of the discrete energy functional $E_\cM$ (\autoref{def:DiscreteEnergyMesh}) is easily generalized to any system of positive weights indexed by the edges of $\cM$. On the other hand, the Riemannian structure of $\Map_\cM(\bH^2, N)$ requires a measure on the domain: while in the smooth case one has the volume density of the Riemannian metric, in the discrete case it can be recorded by a system of weights on the vertices. All of this information can be captured using only the graph structure of $\cM$. \subsubsection*{\texorpdfstring{$\tilde{S}$-triangulated graphs}{S-triangulated graphs}} Recall that a triangulation $\cT$ of a surface is the data of a simplicial complex $K$ and a homeomorphism $h$ from $K$ to the surface. Lifting $\cT$ to the universal cover, we find a triangulation whose underlying graph $\cG$ (\ie{} 1-skeleton) is \emph{locally cyclic}, meaning that the open neighborhood of any vertex (subgraph induced on the neighbors) is a cycle\footnote{ Note that $\cG$ has the property that every face ($2$-simplex) of the triangulation is a triangle ($3$-vertex complete subgraph) in $\cG$; when the converse is also true one says that $\cT$ is a \emph{Whitney triangulation}. In other words, a Whitney triangulation can be recovered as the flag complex spanned by $\cG$. Let us cite \cite{MR2002076} here: \emph{Whitney triangulations are quite amenable for graph-theoretical considerations because they are determined by their underlying graph: the two-dimensional faces are just the triangles of the graph. In other words, we can think of a Whitney triangulation as an object wearing two hats: on one hand it is just a graph, and on the other hand it is a 2-dimensional simplicial complex which in turn can be considered either as a purely combinatorial object or as a topological surface with a fixed simplicial decomposition.} It can be shown (\cite[Prop. 14]{MR2002076}) that a simple graph $\cG$ is the underlying graph of some Whitney triangulation of a surface if and only if $\cG$ is locally cyclic.}. This motivates the following definition: \begin{definition} \label{def:TriangulatedGraph} Given a topological surface $S$, an \emph{$\tilde{S}$-triangulated graph} is a locally cyclic graph $\cG$ with a free, cofinite action of $\pi_1 S$ by graph automorphisms. \end{definition} $\tilde{S}$-triangulated graphs are precisely the graphs that arise as $1$-skeleta of triangulations. When $\cG$ is an $\tilde{S}$-triangulated graph, we denote the associated group action on the set of vertices by $\rho_\tL \colon \pi_1S \to \Aut(\cG^{(0)})$. Let $N$ be a metric space and $\rho_\tR \colon \pi_1S \to \Isom(N)$ a group homomorphism. \begin{definition} Given $\rho_\tL$ and $\rho_\tR$ as above, we call a $(\rho_\tL, \rho_\tR)$-equivariant map $\cG^{(0)} \to N$ an \emph{equivariant map from $\cG$ to $N$} . The space of such equivariant maps will be denoted $\Map_{\text{eq}}(\cG, N)$. \end{definition} As in \autoref{prop:EquivMapsSmoothManifold1} we have: \begin{proposition} \label{prop:MapGammaManifold} If $N$ is a finite-dimensional smooth manifold, so is $\Map_{\text{eq}}(\cG, N)$. \end{proposition} \subsubsection*{Edge-weighted graphs and the energy functional} \begin{definition} \label{def:EdgeWeightedGraph} Let $\cG$ be an $\tilde{S}$-triangulated graph. We say that $\cG$ is \emph{edge-weighted} if it is given a system of \emph{edge weights}, \ie{} a family of positive real numbers $(\omega_e)_{e \in \cG^{(1)}}$ indexed by the set of edges $\cG^{(1)}$, that is invariant under the action of $\pi_1S$. \end{definition} Clearly, the data of a system of edge weights is equivalent to the data of a function \begin{equation} \eta_0 \colon \cG^{(0)} \times \cG^{(0)} \to [0, +\infty) \end{equation} that is symmetric, invariant under the diagonal action of $\pi_1 S$, and such that $\eta_0(x,y) > 0$ if and only if $x$ and $y$ are adjacent. \begin{definition} \label{def:Prekernel} A function $\eta_0$ as above is called a \emph{pre-kernel} on the $\tilde{S}$-triangulated graph $\cG$. \end{definition} The motivation for introducing this notion will become clear in \autoref{def:Kernel} and \autoref{prop:JostEnergyGraph}. We are now ready to define the energy functional: \begin{definition} \label{def:EnergyFunctionalGraph} Let $\cG$ be an $\tilde{S}$-triangulated graph with a system of edge weights $(\omega_e)_{e \in \cG^{(1)}}$, and let $\rho_\tR \colon \pi_1S \to \Isom(N)$ be a group homomorphism where $N$ is a metric space. The \emph{energy functional} $E_\cG \colon \Map_{\text{eq}}(\cG, N) \to \bR$ is defined by \begin{equation} \label{eq:EnergyFunctionalGraph} E_{\cG}(f) = \frac{1}{2} \sum_{e = e_{xy} \in \cE} \omega_{xy} \, d(f(x),f(y))^2 \end{equation} where $\cE \subset \cG^{(1)}$ is any fundamental domain for the action of $\pi_1S$. \end{definition} When $N$ is a Hadamard manifold\footnote{A \emph{Hadamard manifold} is a complete, simply connected Riemannian manifold of nonpositive curvature. On a Hadamard manifold the distance squared function to a fixed point is smooth, while in general it may not be differentiable on the cut locus.}, $E_{\cG}$ is a smooth function on the manifold $\Map_{\text{eq}}(\cG, N)$. Of course we now call a map $f \in \Map_{\text{eq}}(\cG, N)$ \emph{harmonic} when it is a critical point of the energy functional $E_\cG$. When $N$ is not a Hadamard manifold but merely a metric space, one can still define (locally) energy-minimizing harmonic maps. Note that, taking $\omega_e = 1$ for all $e \in \cG^{(1)}$ and $N = \bR$, a harmonic map from $\cG$ to $\bR$ in the sense above coincides with the classical notion of harmonicity for real-valued functions on graphs. The well-known mean value property of harmonic functions is generalized: \begin{proposition} \label{prop:HarmonicCenterOfMassGraph1} Let $\cG$ be an edge-weighted triangulated graph and let $N$ be a metric space. If $f \in \Map_{\text{eq}}(\cG, N)$ is an energy-minimizing harmonic map then $f(x)$ is a center of mass of the weighted system of points $\{(f(y), \omega_{xy})\}_{y \sim x}$ in $N$ for every $x \in \cG^{(0)}$. \end{proposition} Refer to \autoref{sec:CenterOfMass} for the definition and elementary properties of centers of mass. \begin{proof} If $f(x)$ was not the center of mass of its neighbors, then the part of \eqref{eq:EnergyFunctionalGraph} that involves $x$ could be decreased by replacing $f(x)$ by the center of mass while leaving the other values unchanged. \end{proof} \subsubsection*{Vertex weighted-graphs and the Riemannian structure} \begin{definition} \label{def:VertexWeightedGraph} Let $\cG$ be an $\tilde{S}$-triangulated graph. We say that $\cG$ is \emph{vertex-weighted} if it is given a system of \emph{vertex weights}, \ie{} a family of positive real numbers $(\mu_v)_{v \in \cG^{(0)}}$ indexed by the set of vertices $\cG^{(0)}$, that is invariant under the action of $\pi_1S$. \end{definition} We think of a system of vertex weights as a $\pi_1S$-invariant Radon measure $\mu$ on $\cG^{(0)}$. Of course, this is simply a $\pi_1S$-invariant function $\mu \colon \cG^{(0)} \to (0, +\infty)$, but our viewpoint for discretization is to approximate the smooth theory where $\mu$ is the volume density of a Riemannian manifold. Assume now that $N$ is a finite-dimensional Riemannian manifold and let $\rho_\tR \colon \pi_1 S \to \Isom(N)$ be a group homomorphism. We saw (\autoref{prop:MapGammaManifold}) that $\Map_{\text{eq}}(\cG, N)$ is a smooth manifold. Moreover, it is easy to describe its tangent space. \begin{proposition} \label{prop:MapGammaTangentSpace} The tangent space at $f \in \Map_{\text{eq}}(\cG, N)$ is: \begin{equation} \label{eq:MapGammaTangentSpace1} \upT_f \Map_{\text{eq}}(\cG, N) = \Gamma_{\text{eq}}(f^* \upT N) \end{equation} where $f^* \upT N$ is the pullback of the tangent bundle $\upT N$ to $\cG^{(0)}$ and $\Gamma_{\text{eq}}(f^* \upT N)$ is its space of $\pi_1S$- equivariant smooth sections . Equivalently, if $\cV \subseteq \cG^{(0)}$ is any fundamental domain for the action of $\pi_1S$, \begin{equation} \label{eq:MapGammaTangentSpace2} \upT_f \Map_{\text{eq}}(\cG, N) = \bigoplus_{x \in \cV} \upT_{f(x)} N~. \end{equation} \end{proposition} Notice of course the similarity of \eqref{eq:MapGammaTangentSpace1} with \eqref{eq:TangentSpaceMapSmooth}. Using the measure $\mu$, one can define a natural $\upL^2$ Riemannian metric on $\Map_{\text{eq}}(\cG, N)$ analogous to \eqref{eq:InnerProductSmooth}: \begin{definition} \label{def:L2RiemannianMetricGraph} Let $(\cG, \mu)$ be an $\tilde{S}$-triangulated vertex-weighted graph, and let $\rho_\tR \colon \pi_1S \to \Isom(N)$ where $N$ is a Riemannian manifold. The $\upL^2$ Riemannian metric on $\Map_{\text{eq}}(\cG, N)$ is given by: \begin{equation} \label{eq:L2RiemannianMetricGraph1} \langle V, W \rangle = \int_\cV \langle V_x, W_x\rangle \, \upd \mu(x) \end{equation} where $V, W \in \Gamma_{\text{eq}}(f^* \upT N)$ and $\cV \subseteq \cG^{(0)}$ is any fundamental domain for the action of $\pi_1S$. \end{definition} Of course, one can write more concretely: \begin{equation} \label{eq:L2RiemannianMetricGraph2} \langle V, W \rangle = \sum_{x \in \cV} \mu(x) \langle V_x, W_x\rangle~. \end{equation} One can easily derive that the unit speed geodesics in $\Map_{\text{eq}}(\cG, N)$ are the one-parameter families of functions $(f_t(x))_{x \in \cG^{(0)}}$ given by $f_t(x) = \exp(t V_x)$, where $V \in \Gamma_{\text{eq}}(f^* \upT N)$ is a unit vector, and, provided $N$ is connected, the Riemannian distance in $\Map_{\text{eq}}(\cG, N)$ is simply given by \begin{equation} \label{eq:L2DistanceDiscrete} d(f,g)^2 = \sum_{x \in \cV} \mu(x) \, d(f(x), g(x))^2~, \end{equation} where on the right hand-side $d$ is the Riemannian distance in $N$. Of course notice that \eqref{eq:L2DistanceDiscrete} is just the discretization of \eqref{eq:L2DistanceSmooth}. \subsubsection*{Biweighted graphs} \begin{definition} \label{def:BiweightedGraph} Let $\cG$ be an $\tilde{S}$-triangulated graph (\autoref{def:TriangulatedGraph}). We say that $\cG$ is \emph{biweighted} if it is both edge-weighted (\autoref{def:EdgeWeightedGraph}) and vertex-weighted (\autoref{def:VertexWeightedGraph}). \end{definition} From the discussion of the previous paragraph, when $\cG$ is an $\tilde{S}$-triangulated biweighted graph and $N$ is a Riemannian manifold with a group homomorphism $\rho_\tR \colon \pi_1S \to N$, the space of equivariant maps $\Map_{\text{eq}}(\cG, N)$ is a Riemannian manifold and the energy is a continuous function $E_\cG \colon \Map_{\text{eq}}(\cG, N) \to \bR$. Moreover $E_\cG$ is smooth when $N$ is Hadamard. In \autoref{sec:StrongConvexity} we show that $E_\cG$ is strongly convex under suitable restrictions on $\rho_\tR$, with an explicit bound on the modulus of strong convexity (\autoref{thm:StrongConvexityEnergyGraphGeneral}). This implies that there exists a unique equivariant harmonic map $\cG \to N$ that can be computed effectively through gradient descent (\autoref{sec:DiscreteHeatFlow}). We pause to point out that our definition of the energy functional and harmonic maps in this setting coincides with Jost's theory briefly described in \autoref{subsec:JostEnergy} (we refer to \cite{MR1449406, MR1451625} for details). First we introduce the kernel function associated to a biweighted graph: \begin{definition} \label{def:Kernel} The kernel function associated to a biweighted graph $\cG$ is the function \begin{equation} \begin{split} \eta \colon \cG^{(0)} \times \cG^{(0)} \to \bR\\ (x,y) \mapsto \frac{\eta_0(x,y)}{2 \mu(x) \mu(y)} \end{split} \end{equation} where $\eta_0$ is the pre-kernel associated to the underlying edge-weighted graph (cf.~\autoref{def:Prekernel}) and $\mu$ is the measure on $\cG^{(0)}$ giving the vertex weights. \end{definition} The next proposition is trivial but conceptually significant: \begin{proposition} \label{prop:JostEnergyGraph} The energy functional on $\Map_{\text{eq}}(\cG, N)$ is given by \begin{equation} E_\cG(f) = \frac{1}{2} \iint_\cV \eta(x,y)\, {d(f(x), f(y))}^2 \, \upd \mu(y) \, \upd \mu (x) \end{equation} where $\cV \subseteq \cG^{(0)} \times \cG^{(0)}$ is a fundamental domain for the diagonal action of $\pi_1S$. \end{proposition} \autoref{prop:JostEnergyGraph} implies that, choosing $\eta_r = \eta$ for all $r > 0$, the Jost energy functional $E = \lim_{r\to 0} E_r$ (compare with \eqref{eq:EnergyFunctionalJost}) coincides with the energy functional $E_\cG$. In particular, our notion of harmonic maps from graphs is a specialization of Jost's \emph{generalized harmonic maps}. Next we observe that the Riemannian structure of $\Map_{\text{eq}}(\cG, N)$ allows us to define the discrete tension field as: \begin{definition} The \emph{tension field} of $f \in \Map_{\text{eq}}(\cG, N)$ is the vector field along $f$ denoted $\tau_\cG(f) \in \Gamma_{\text{eq}}(f^* \upT N)$ given by: \begin{equation} \tau_\cG(f)\evalat{x} = \frac{1}{\mu(x)} \sum_{y \sim x} \omega_{xy} \exp_{f(x)}^{-1}(f(y)) \end{equation} where we have denoted $\omega_{xy}$ the weight of the edge connecting $x$ and $y$. \end{definition} We have the discrete version of the first variational formula for the energy (\autoref{prop:FirstVariationalFormulaSmooth}): \begin{proposition} \label{prop:FirstVariationalFormulaGraph} The tension field is minus the gradient of the energy functional: \begin{equation} \tau_\cG(f) = -\grad E_\cG(f) \end{equation} for any $f \in \Map_{\text{eq}}(\cG, N)$. \end{proposition} \begin{proof} In a Riemannian manifold $N$, when $x_0 \in N$ is chosen such that $\exp_{x_0}$ is a diffeomorphism (any $x_0$ works when $N$ is Hadamard), the function $g \colon x \mapsto \frac{1}{2} d(x_0,x)^2$ is smooth and its gradient is given by $\grad g(x) = -\exp^{-1}_x(x_0)$. \end{proof} It follows, of course, that an equivariant map $\cG \to N$ is harmonic if and only if its tension field is zero, and we obtain a characterization of discrete harmonic maps: \begin{proposition} \label{prop:HarmonicCenterOfMassGraph2} Let $\cG$ be an edge-weighted triangulated graph and let $N$ be a Hadamard manifold. Then $f \in \Map_{\text{eq}}(\cG, N)$ is a harmonic map if and only if $f(x)$ is a center of mass of the weighted system of points $\{(f(y), \omega_{xy})\}_{y \sim x}$ in $N$ for every $x \in \cG^{(0)}$. \end{proposition} \bigskip We conclude this section by looping back to \autoref{subsec:Meshes} and the approximation problem. The point is that when $S$ is equipped with a hyperbolic structure (or more generally any nonpositively curved metric), a mesh in the sense of \autoref{def:Mesh} induces a biweighted graph structure: \begin{definition} \label{def:MeshToGraph} Let $\rho_\tL \colon \pi_1S \to \Isom^+(\bH^2)$ be a Fuchsian representation and let $\cM$ be an invariant mesh (cf.~\autoref{def:Mesh}). The biweighted graph underlying $\cM$ is the biweighted graph $\cG$ such that: \begin{itemize} \item $\cG$ is the abstract graph underlying $\cM$ (which is evidently $\tilde{S}$-triangulated). \item The edge weights are the $\omega_e$ as in \autoref{def:DiscreteEnergyMesh}. \item The vertex weights are given by, for every vertex $x$: \begin{equation} \mu(x) = \frac{1}{3} \sum_T \Area(T) \end{equation} where the sum is taken over all triangles incident to the vertex $x$. \end{itemize} \end{definition} Clearly, any discrete equivariant map along $\cM$ from $\bH^2$ to a Riemannian manifold $N$ induces an equivariant map $\cG \to N$, and the energy $E_\cM$ agrees with the energy $E_\cG$. Of course, the systems of weights are chosen so that the discrete energy functional $E_\cG$ approximates the smooth energy functional \eqref{eq:EnergyFunctionalSmoothEquivariant}, and the Riemannian structure of $\Map_\text{eq}(\cG, N)$ approximates the $\upL^2$ Riemannian metric on $\cC^\infty(M,N)$ (or $\upL^2(M,N)$), with finer approximation when one takes finer meshes. The analysis of this phenomenon is treated in \cite{Gaster-Loustau-Monsaingeon2}. \begin{remark} With this construction in mind, biweighted triangulated graphs can be roughly thought of as follows: the edge weights are a discrete record of the conformal structure of $S$, and the vertex weights, the area form. Note that the metric structure can be recovered from both, a phenomenon specific to dimension 2. This is reminiscent of \cite{MR3375525}-- though distinct-- in which two graphs with edge weights are considered \emph{conformally equivalent} if there is a function of the vertices that scales one set of weights to another. \end{remark} \section{Harmonic maps} \label{sec:HarmonicMaps} \subsection{Energy functional and harmonic maps} \label{subsec:EnergyFunctionalHarmonicMaps} Let $(M,g)$ and $(N,h)$ be two smooth Riemannian manifolds. Assuming $M$ is compact, the \emph{energy} of a smooth map $f \colon M \to N$ is: \begin{equation} \label{eq:EnergyFunctionalSmooth} E(f) = \frac{1}{2} \int_M \Vert \upd f \Vert^2 \, \upd v_g \end{equation} where $v_g$ is the volume density of the metric $g$. Note that $\upd f$ is a smooth section of the bundle $\upT^*M \otimes f^* \upT N$ over $M$, which admits a natural metric induced by $g$ and $h$, giving sense to $\Vert \upd f \Vert$. This is the so-called \emph{Hilbert-Schmidt norm} of $\upd f$, which is also described as $\Vert \upd f \Vert^2 = \tr_g(f^* h)$. \begin{definition} \label{def:HarmonicMap} A map $f \colon M \to N$ is \emph{harmonic} if it is a critical point of the energy functional \eqref{eq:EnergyFunctionalSmooth}. \end{definition} \noindent This means concretely that: \begin{equation} \ddt\evalat{t=0} E(f_t) = 0 \end{equation} for any smooth deformation $(f_t) \colon (-\delta, \delta) \times M \to N$ of $f = f_0$. Note that one should work with compactly supported deformations when $M$ is not compact, as the energy could be infinite. A more tangible characterization of harmonicity is given by the Euler-Lagrange equation for $E$, which takes the form $\tau(f) = 0$ where $\tau(f)$ is the \emph{tension field} of $f$: this is an immediate consequence of the first variational formula below (\autoref{prop:FirstVariationalFormulaSmooth}). First we define the tension field. Note that the bundle $\upT^*M \otimes f^* \upT N$ admits a natural connection $\nabla$ induced by the Levi-Civita connections of $g$ and $h$. Hence one can take the covariant derivative $\nabla (\upd f) \in \Gamma(\upT^*M \otimes \upT^*M \otimes f^* \upT N)$ (we use the notation $\Gamma$ for the space of smooth sections), also denoted $\nabla^2 f$. It is easily shown to be symmetric in the first two factors. \begin{definition} The \emph{vector-valued Hessian} of $f$ is \begin{equation} \nabla^2 f \coloneqq \nabla(\upd f) \in \Gamma(\upT^*M \otimes \upT^*M \otimes f^* \upT N)~. \end{equation} The contraction (trace) of $\nabla^2 f$ on its first two indices using the metric $g$ is the \emph{tension field} of $f$: \begin{equation} \tau(f) \coloneqq \tr_g(\nabla^2 f) \in \Gamma(f^* \upT N)~. \end{equation} \end{definition} Note that the vector-valued Hessian generalizes both the usual Hessian (when $N = \bR$) and the (vector-valued) second fundamental form (when $f$ is an isometric immersion). Accordingly, the tension field generalizes both the Laplace-Beltrami operator and the (vector-valued) mean curvature. \begin{proposition}[First variational formula for the energy] \label{prop:FirstVariationalFormulaSmooth} Let $f \colon (M,g) \to (N, h)$ be a smooth map and let $(f_t)$ be a smooth deformation of $f$. Denote by $V \in \Gamma(f^* \upT N)$ the associated infinitesimal deformation defined as $V_x = \ddt\evalat{t=0} f_t(x)$. Then \begin{equation} \label{eq:FirstVariationalFormula} \ddt\evalat{t=0} E(f_t) = -\int_M \langle \tau(f)_x, V_x\rangle \, \upd v_g(x) \end{equation} where $h = \langle \cdot , \cdot \rangle$ is the Riemannian metric in $N$. \end{proposition} One can introduce a natural $\upL^2$ inner product of two infinitesimal deformations $V, W \in \Gamma(f^* \upT N)$ (also called \emph{vector fields along $f$}): \begin{equation} \label{eq:InnerProductSmooth} \langle V, W \rangle = \int_M \langle V_x, W_x\rangle \, \upd v_g(x)~. \end{equation} There is in fact a natural smooth structure on $\cC^\infty(M,N)$, making it an infinite-dimensional manifold, which identifies the tangent space at $f$ as \begin{equation} \label{eq:TangentSpaceMapSmooth} \upT_f \cC^\infty(M,N) = \Gamma(f^* \upT N)~, \end{equation} we refer to \cite[Chapter IX]{MR1471480} for details. With respect to this smooth structure, \eqref{eq:InnerProductSmooth} defines a Riemannian metric on $\cC^\infty(M,N)$, and \eqref{eq:FirstVariationalFormula} can simply be put: \begin{equation} \label{eq:GradientEnergy} \grad E(f) = -\tau(f)~. \end{equation} Next we compute the second variation of the energy (like the first variation, this is already in \cite{MR0164306}): \begin{proposition}[Second variational formula for the energy] \label{prop:SecondVariationalFormulaSmooth} Let $(f_{st}) \colon (-\delta, \delta)^2 \times M \to N$ be a smooth deformation of $f = f_{00}$ . Denote $V=\frac{\partial f}{\partial s}\evalat{s=0}$ and $W=\frac{\partial f}{\partial t}\evalat{t=0}$. Then \begin{equation} \frac{\partial^2 E(f_{st})}{\partial s \partial t}\evalat{s=t=0} = \int_M \left(\left \langle \nabla V, \nabla W \right \rangle - \tr_g \left \langle R^N(\upd f,V) W,\upd f \right\rangle + \left \langle \nabla_{\frac \partial {\partial t}} \frac{\partial f}{\partial s} , \tau(f) \right\rangle \right) \, \upd v_g \end{equation} where $R^N$ is the Riemann curvature tensor\footnote{For us the curvature tensor is $R(X,Y)Z = \nabla^2_{X, Y} Z - \nabla^2_{Y, X} Z$. Some authors' convention differs in sign, \eg \cite{MR2088027}.} on $N$. \end{proposition} \noindent When $(f_{st})$ is a geodesic variation, \ie{} $f_{st}(x) = \exp_{f(x)}(s V_x + tW_x)$, the third term in the integral vanishes. This yields the formula for the Hessian of the energy functional: \begin{equation} \label{eq:HessianEnergySmooth} \Hess(E)\evalat{f}(V, W) = \int_M \left(\left \langle \nabla V, \nabla W \right \rangle - \tr_g \left \langle R^N(V, \upd f) \upd f, W \right\rangle \right) \, \upd v_g~. \end{equation} When $M$ is closed, this can also be written $\Hess(E)\evalat{f}(V, W) = \langle J(V), W \rangle$ using the $\upL^2$ Riemannian metric \eqref{eq:InnerProductSmooth}, where $J(V) = -\tr_g(\nabla^2 V + R^N(V, \upd f) \upd f)$ is the \emph{Jacobi operator}. \subsection{\texorpdfstring{Energy functional on $\upL^2(M,N)$ and more general spaces}{Energy functional on L2(M,N) and more general spaces}} \label{subsec:JostEnergy} The energy functional can be extended to maps that are merely in $\upL^2(M,N)$. First let us define this function space. Assume $M$ is compact. The \emph{$L^2$-distance} between two measurable maps $f_1,f_2 \colon M \to N$ is \begin{equation} \label{eq:L2DistanceSmooth} d(f_1, f_2) = \left(\int_M d( f_1(x), f_2(x))^2 \, \upd v_g(x)\right)^{\frac 12}~. \end{equation} If $f_1, f_2$ are both smooth, this is the distance induced by the $\upL^2$ Riemannian metric \eqref{eq:InnerProductSmooth}, provided there exists a geodesic between $f_1$ and $f_2$. A measurable map $f \colon M \to N$ is declared in $\upL^2(M,N)$ when it is within finite distance of a constant map. For $r> 0$, one can then define an \emph{approximate $r$-energy} of $f \in \upL^2(M,N)$: \begin{equation} \label{eq:ApproximateEnergyJost} E_r(f) = \frac{1}{2} \int_M \int_M \eta_r(x,y) \, {d(f(x), f(y))}^2 \, \upd v_g(y) \, \upd v_g (x) \end{equation} where $\eta_r(x,y)$ is a \emph{kernel} that may be chosen $\eta_r(x,y) = \frac{\mathbf{1}_r(x,y)}{r^2 V_m(r)}$, where $V_m(r)$ is the volume of a ball of radius $r$ in a Euclidean space of dimension $m = \dim M$ and $\mathbf{1}_r(x,y)$ is the characteristic function of $\{(x,y) \in M^2 : d(x,y) < r\}$ in $M \times M$ (see \cite[\S 4.1]{MR1451625} for a discussion of the choice of kernel). One can show that the functional $E_r$ is continuous on $\upL^2(M,N)$. Moreover, the limit: \begin{equation} \label{eq:EnergyFunctionalJost} E(f) \coloneqq \lim_{r \to 0} E_r(f) \end{equation} exists in $[0, \infty]$ for every $f\in \upL^2(M,N)$. The resulting energy functional $E$ is lower semi-continuous on $\upL^2(M,N)$ and coincides with \eqref{eq:EnergyFunctionalSmooth} on $\cC^\infty(M,N)$. A measurable map $f \colon M \to N$ is declared in the Sobolev space $\upH^1(M,N)$ if it is in $\upL^2(M,N)$ and has finite energy. The spaces $\upL_{\text{loc}}^2(M,N)$ and $\upH_{\text{loc}}^1(M,N)$ are similarly defined by restricting to compact sets. One can then define a (weakly) harmonic map as a critical point of the energy functional in $\upH_{\text{loc}}^1(M,N)$. Any continuous weakly harmonic map is smooth \cite[Theorem 9.4.1]{MR3726907} (the continuity assumption can be dropped when $M$ and $N$ are compact and $N$ has nonpositive curvature: \cite[Corollary 9.6.1]{MR3726907}). In addition to opening the way for tools from functional analysis, this approach can be generalized to much more general spaces than Riemannian manifolds. Indeed, assume $M = (M,\mu)$ is a measure space and $N = (N,d)$ is a metric space. The space $\upL^2(M,N)$ may be defined as before, and given a choice of kernel $\eta_r$ for $r>0$, one can define energy functionals $E_r \colon \upL^2(M,N) \to \bR$ using \eqref{eq:ApproximateEnergyJost}. For a suitable choice of $\eta_r$ and of a sequence $r_n \to 0$, the energy functional is $E = \lim_{n \to +\infty} E_{r_n}$. More precisely, one must ensure that $E$ is the $\Gamma$-limit of the functionals $E_{r_n}$. We refer to \cite[Chap. 4]{MR1451625} for details and \cite{MR1201152} for the theory of $\Gamma$-convergence. $\Gamma$-convergence is adequate here because it ensures that minimizers of $E_r$ converge to minimizers of $E$. This point of view on the theory of harmonic maps was developed by Jost \cite{MR1385525, MR1360608, MR1449406, MR1451625}. A similar approach was developed by Korevaar-Schoen \cite{MR1266480, MR1483983}. \subsection{The heat flow} \label{subsec:SmoothHeatFlow} Going back to the smooth setting, assume that $M$ is compact and $N$ is complete and has nonpositive curvature. The formula for the Hessian of the energy \eqref{eq:HessianEnergySmooth} shows that it is nonnegative, in other words $E$ is a convex function on $\mathcal{C}^\infty(M,N)$ with respect to the $\upL^2$ Riemannian metric. This makes it reasonable to expect existence and in certain cases uniqueness of harmonic maps, which are necessarily energy-minimizing in this setting (we discuss this further in \autoref{subsec:ConvexityEnergyExistenceHarmonic}). A natural approach to minimize a convex function is the gradient flow, called \emph{heat flow} in this setting: given $f_0 \in \mathcal{C}^\infty(M,N)$, consider the initial value problem $\ddt f_t = -\grad E(f_t)$, that is in light of \eqref{eq:GradientEnergy}: \begin{equation} \ddt f_t = \tau(f_t)~. \end{equation} This flow exists for all $t \geqslant 0$. Moreover, if the range of $f_t$ remains in some fixed compact subset of $N$, then $f_t$ converges to a harmonic map as $t \to \infty$, uniformly and in $\upL^2(M,N)$ (in fact, in $\mathcal{C}^\infty(M,N)$). Otherwise, there exists no harmonic map homotopic to $f$. In particular, when $N$ is compact, any $f_0 \in \mathcal{C}^\infty(M,N)$ is homotopic to a smooth energy-minimizing harmonic map. Moreover, such a harmonic map is unique, unless it is constant or maps into a totally geodesic flat submanifold of $N$ in which case non-uniqueness is realized by translating $f$ in the flat. These foundational results are due to Eells-Sampson \cite{MR0164306} and Hartman \cite{MR0214004}. \subsection{Equivariant harmonic maps} \label{subsec:EquivariantHarmonicMaps} Instead of working with maps between compact manifolds, it can be useful to study their equivariant lifts to the universal covers. Indeed, up to being careful with basepoints, any continuous map $f \colon M \to N$ lifts to a unique $\rho$-equivariant map $\tilde{f} \colon \tilde{M}\to \tilde{N}$, where $\rho \colon \pi_1 M \to \pi_1 N$ is the group homomorphism induced by $f$. Note that $\rho$ only depends on the homotopy class of $f$, and if $N$ is aspherical (\ie{} $\tilde{N}$ is contractible), then conversely any $\rho$-equivariant continuous map $M \to N$ is the lift of some continuous map $M \to N$ homotopic to $f$. This approach enables the following generalization: let $X$ and $Y$ be two Riemannian manifolds, denote $\Isom(X)$ and $\Isom(Y)$ their groups of isometries. Let $\Gamma$ be a discrete group. Given group homomorphisms $\rho_\tL \colon \Gamma \to \Isom(X)$ and $\rho_\tR \colon \Gamma \to \Isom(Y)$, a map $f \colon X \to Y$ is called $(\rho_\tL, \rho_\tR)$-equivariant if: \begin{equation} f \circ \rho_\tL(\gamma) = \rho_\tR(\gamma) \circ f \end{equation} for all $\gamma \in \Gamma$. Note that the quotients $X/\rho_\tL(\Gamma)$ and $Y/\rho_\tR(\Gamma)$ can be pathological, but the space of equivariant maps $X \to Y$ remains ripe for study. The heat flow approach of Eells-Sampson to show existence of harmonic maps between compact Riemannian manifolds when the target is nonpositively curved has been successfully adapted to the equivariant setting by various authors. The adequate condition for guaranteeing existence of equivariant harmonic maps is the \emph{reductivity} of the target representation. More precisely: \begin{theorem}[\cite{MR1049845}] \label{thm:Labourie} Let $M$ and $N$ be Riemannian manifolds, assume $N$ is Hadamard. Denote by $\rho_\tL \colon \pi_1M \to \Isom(\tilde{M})$ the action by deck transformations and let $\rho_\tR \colon \pi_1M \to \Isom(N)$ be any group homomorphism. If $\rho_\tR$ is reductive, then there exists a $(\rho_\tL, \rho_\tR)$-equivariant harmonic map $\tilde{M} \to N$. The converse also holds provided $N$ is without flat half-strips. \end{theorem} Less general versions of this theorem had previously been established by Donaldson \cite{MR887285} (for $N = \bH^3$) and Corlette \cite{MR965220} (for $N$ a Riemannian symmetric space of noncompact type). The notion of being \emph{reductive} for a group homomorphism $\rho \colon \pi_1M \to G$ can be described algebraically when $N = G/K$ is a Riemannian symmetric space of noncompact type\footnote{When $G$ is an algebraic group, a subgroup $H \subseteq G$ is \emph{completely reducible} if, for every parabolic subgroup $P \subseteq G$ containing $H$, there is a Levi subgroup of $P$ containing $H$. Equivalently, the identity component of the algebraic closure of $H$ is a reductive subgroup (with trivial unipotent radical). A $G$-valued group homomorphism $\rho$ is called reductive (or completely reducible) when its image is a completely reducible subgroup. Refer to \cite{MR2931326} for details.}. Labourie \cite{MR1049845} generalized it to Hadamard manifolds. When $N$ has negative curvature, $\rho$ is reductive if and only if it fixes no point on the Gromov boundary $\partial_\infty N$ or it preserves a geodesic in $N$. Wang \cite{MR1775138} and Izeki-Nayatani \cite{MR2174098} generalized \autoref{thm:Labourie} to Hadamard metric spaces using Jost's extended notion of reductivity \cite[Def. 4.2.1]{MR1451625}. Less general or different versions were previously established by \cite{MR1215595}, \cite[Thm 4.2.1]{MR1451625}, \cite{MR1483983}. \subsection{Harmonic maps from surfaces} \label{subsec:harmonicMapsSurfaces} When $M = S$ is a surface, \ie{} $\dim M = 2$, it is easy to check that the energy density element $e(f) \, \upd v_g \coloneqq \frac{1}{2} \Vert \upd f \Vert^2 \, \upd v_g$ is invariant under conformal changes of the metric $g$. Thus the energy functional only depends on the conformal class of $g$, as does the harmonicity of a map $S \to N$. A conformal structure on an oriented surface is equivalent to a complex structure (this follows from a result going back to Gauss \cite{Gauss1825} on the existence of conformal coordinates). Hence one may talk about the energy and harmonicity of maps $X \to (N,h)$ where $X$ is a Riemann surface. Note however that the $L^2$ metric \eqref{eq:InnerProductSmooth} does change under conformal changes of $g$, therefore the tension field $\tau(f)$ does too, as does the modulus of strong convexity of the energy (see \autoref{subsec:ConvexityRiemannian}). One can see directly that the energy density element only depends on the complex structure $X$ on $S$ by writing the pullback metric $f^* h$ on $X$. Splitting it into types, one finds that \begin{equation} f^* h = \varphi_f + g_{f} + \bar{\varphi_f} \end{equation} where $\varphi_f = (f^*h)^{(2,0)}$ is a complex quadratic differential on $X$, called the \emph{Hopf differential} of $f$, and $g_f = (f^*h)^{(1,1)}$ is $e(f)$ (more precisely, $g_f$ is the conformal metric with volume density $e(f) \, \upd v_g$). The Hopf differential $\varphi_f$ plays an important role in Teichmüller theory. First note that $f$ is conformal if and only if $\varphi_f = 0$. A key fact is that if $f$ is harmonic, then $\varphi_f$ is a holomorphic quadratic differential on $X$. Wolf \cite{MR982185} proved that the Teichmüller space of $X$ is diffeomorphic to the vector space of holomorphic quadratic differentials on $X$ by taking Hopf differentials of harmonic maps $X \to (S,h)$, where $h$ is a hyperbolic metric on $S$ (see \autoref{subsec:HighEnergy}) . We refer to \cite{MR2349668} for a beautiful review of the connections between harmonic maps and Teichmüller theory. On a closed surface $S$ of negative Euler characteristic, it is convenient to choose the \emph{Poincaré metric} within a conformal class of metrics: it is the unique metric of constant curvature $-1$ (its existence is precisely the celebrated \emph{uniformization theorem}). This provides an identification of $\tilde{S}$ with the hyperbolic plane $\bH^2$ and an action of $\pi_1 S$ on $\bH^2$ by isometries. Turning this identification around, whenever a Fuchsian (\ie{} faithful and discrete) representation $\rho_\tL \colon \pi_1 S \to \Isom^+(\bH^2)$ is chosen, we obtain a hyperbolic surface $ \bH^2 / \rho_\tL(\pi_1 S) \approx S$. \section*{Introduction} \addcontentsline{toc}{section}{Introduction} The theory of harmonic maps has its roots in the foundations of Riemannian geometry and the essential work of Euler, Gauss, Lagrange, and Jacobi. It includes the study of real-valued harmonic functions, geodesics, minimal surfaces, and holomorphic maps between Kähler manifolds. A harmonic map $f \colon M \to N$ is a critical point of the energy functional \begin{equation} E(f) = \frac{1}{2} \int_M \Vert \upd f \Vert^2 \, \upd v_M~. \end{equation} The theory of harmonic maps was brought into a modern context for Riemannian manifolds with the seminal work of Eells-Sampson \cite{MR0164306} (also Hartman \cite{MR0214004} and Al'ber \cite{MR0230254}). Eells-Sampson studied the \emph{heat flow} associated to the energy, \ie{} the nonlinear parabolic PDE \begin{equation} \ddt f_t = \tau(f_t)~, \end{equation} where $\tau(f)$ is the \emph{tension field} of $f$ (see \autoref{sec:HarmonicMaps}). The tension field can be described as minus the gradient of the energy functional on the infinite-dimensional Riemannian space $\cC^\infty(M,N)$, so that the heat flow is just the gradient flow for the energy. When $N$ is compact and nonpositively curved, the heat flow is shown to converge to an energy-minimizing map as $t\to \infty$. The theory has since been developed and generalized to various settings where the domain or the target are not smooth manifolds \cite{MR1215595,MR1266480,MR1340295,MR1451625,MR1848068, MR1938491, MR2394023}. In particular, Korevaar-Schoen developed an extensive Sobolev theory when the domain is Riemannian but the target is a nonpositively curved metric space \cite{MR1266480,MR1483983}. Jost generalized further to a domain that is merely a measure space \cite{MR756629, MR1385525,MR1360608,MR1449406}. Both took a similar approach, constructing the energy functional $E$ as a limit of approximate energy functionals. These tools have become powerful and widely used, with celebrated rigidity results \cite{MR584075,MR1215595,MR1775138,MR2827014,MR2174098} and dramatic implications for the study of deformation spaces and Teichmüller theory when the domain $M$ is a surface (see \eg \cite{MR2349668}), especially via the nonabelian Hodge correspondence \cite{MR887285,MR887284,MR965220,MR982185,MR1049845,MR1159261}. This paper and its sequel \cite{Gaster-Loustau-Monsaingeon2} are concerned with effectiveness of methods for finding harmonic maps. In addition to the mathematical content, we feature \texttt{Harmony}{}, a computer program that we developed in \Cpp{} whose main functionality is to numerically compute equivariant harmonic maps. Our project was motivated by the question: is it possible to study the nonabelian Hodge correspondence experimentally? Though the heat flow is constructive to some extent, it does not provide qualitative information about convergence in general. We show that one can design entirely effective methods to compute harmonic maps by discretizing appropriately. Some of the existing literature treats similar questions, though seldom in the Riemannian setting. Notably, Bartels applies finite element methods on submanifolds of $\bR^n$ to nonlinear PDEs such as the Euler-Lagrange equations for minimizing the energy \cite{MR2177142,MR2629993,MR3309171}. There is also an extensive literature on discrete energy functionals of the form we consider in this paper \cite{MR1151746,MR1246481,MR1721305,MR1783793,MR2346504,MR1775138,MR2174098,MR1848068,MR2471368,MR3423736}. However, several features set our work apart. Firstly, we study sequences of arbitrarily fine discretizations as opposed to one fixed approximation of the manifold, particularly in the sequel paper \cite{Gaster-Loustau-Monsaingeon2}. Moreover, we follow the maxim of Bobenko-Suris \cite[p.~xiv]{MR2467378}: \begin{center} \emph{Discretize the whole theory, not just the equations}. \end{center} Our discrete structures record two independent systems of weights on a given triangulation of the domain manifold $M$, on the set of vertices and on the set of edges respectively. The vertex weights are a discrete record of the volume form, while the edge weights generalize the well-known ``cotangent weights'' popularized by Pinkall-Polthier in the Euclidean setting \cite{MR1246481}. In the $2$-dimensional case, the edge weights can be thought of as a discrete record of the conformal structure. This discretization endows the space of discrete maps with a finite-dimensional Riemannian structure, approximating the $\upL^2$ metric on $\cC^\infty(M,N)$. We obtain the right setting for a study of the convexity of the discrete energy, and for the definitions of the discrete energy density, discrete tension field, and discrete heat flow. Among the practical benefits, we find that the discrete energy satisfies stronger convexity properties than those known to hold in the smooth setting. The theory of harmonic maps is also intimately related to centers of mass. In the Euclidean setting, harmonic functions satisfy the well-known mean value property. While the latter no longer holds \emph{stricto sensu} in the more general Riemannian setting, we prove a generalization in terms of centers of mass of harmonic maps on very small balls. More generally, averaging a function on small balls offers a viable method in order to decrease its energy, a viewpoint well adapted to Jost's theory of generalized harmonic maps \cite{MR1385525, MR1360608, MR1449406, MR1451625}. As an alternative to the discrete heat flow, we pursue a discretization of the theory of Jost by analyzing discrete center of mass methods. For both heat flow and center of mass methods, the present paper focuses on a fixed discretization of the domain, while the sequel paper \cite{Gaster-Loustau-Monsaingeon2} analyzes convergence of the discrete theory back to the smooth one as we take finer and finer discretizations approximating a smooth domain. While this paper focuses on two dimensional hyperbolic manifolds, and chooses the equivariant setting, the second paper embraces the general Riemannian setting for the most part, and drops the equivariant formulation. \bigskip Now let us describe more precisely some of the main theorems of the paper. After discussing harmonic maps in \autoref{sec:HarmonicMaps} and developing a discretized theory in \autoref{sec:Discretization}, we study the convexity of the energy functionals in \autoref{sec:StrongConvexity}. We show: \begin{theorem*}[\autoref{thm:StrongConvexityEnergyGraphGeneral}] Let $\cS$ be a discretized surface of negative Euler characteristic and let $N$ be a compact surface of nonzero Euler characteristic with nonpositive sectional curvature. The discrete energy functional is strongly convex in any homotopy class of nonzero degree. \end{theorem*} \noindent We stress that while the convexity of the energy comes for free, the content of the theorem above is a positive modulus of convexity, \ie{} a uniform lower bound for the Hessian. We actually show a more general version of this theorem involving equivariant maps and the notion of \emph{biweighted triangulated graph} which we introduce in \autoref{sec:Discretization}. See \autoref{thm:StrongConvexityEnergyGraphGeneral} for the precise statement. When the target $N$ is specialized to a hyperbolic surface, we find explicit bounds for the Hessian of the energy functional (see \autoref{thm:StrongConvexityEnergyGraphH2}). We achieve this through detailed calculations in the hyperbolic plane, which we then generalize to negatively curved target surfaces using $\CAT(k)$-type comparisons. Roughly speaking, the key idea is that if the energy of some function has a very small second variation, then one can construct an almost parallel vector field on the image of the surface; however this is not possible for topological reasons. \autoref{sec:StrongConvexity} is concerned with the significant work of making this argument precise and quantitative. In \autoref{sec:DiscreteHeatFlow} we study gradient descent methods in Riemannian manifolds and specialize to the convergence of the discrete heat flow. Fully leveraging strong convexity of the energy, we show: \begin{theorem*} [\autoref{thm:ConvergenceDiscreteHeatFlowGeneral}] Let $\cS$ be a discretized surface of negative Euler characteristic and let $N$ be a compact surface of nonzero Euler characteristic with nonpositive curvature. There exists a unique discrete harmonic map $f^* \colon \cS \to N$ in any homotopy class of nonzero degree. Moreover, for any initial value $f_0:\cS\to N$ and for any sufficiently small $t>0$, the discrete heat flow with fixed stepsize $t$ converges to $f^*$ with exponential convergence rate. \end{theorem*} \noindent This is again a simplified version of the theorem we show; see \autoref{thm:ConvergenceDiscreteHeatFlowGeneral} for the precise statement. Next we discuss center of mass methods in \autoref{sec:CenterOfMass}. An interesting alternative to the heat flow consists in averaging $f$ on balls of small radius $r>0$, producing a new map $B_r f \colon M \to N$. Repeating this process potentially produces energy-minimizing sequences for an approximate version of the energy $E_r$, a phenomenon that has been explored by Jost \cite{MR1385525}. The central theorem we prove in \autoref{subsec:GeneralizedMeanValueProperty} is that in the Riemannian setting, this iterative process is almost the same as a fixed stepsize time-discretization of the heat flow. See \autoref{thm:AverageMap} for a precise statement. We prove in particular the following generalized mean value property for harmonic maps: \begin{theorem*}[\autoref{thm:GeneralizedMeanValueProperty}] Let $f \colon M \to N$ be a smooth map between Riemannian manifolds. Then $f$ is harmonic if and only if for all $x\in M$, we have as $r\to 0$: \begin{equation} d(f(x), B_r f(x)) = \bigO(r^4)\,. \end{equation} \end{theorem*} Under suitable conditions, we show that the center of mass method converges to a minimizer of the approximate energy $E_r$ (see \autoref{thm:CenterOfMassMethodSmooth}), recovering a theorem of Jost \cite[\S3]{MR1385525}. Jost's result is more general, but our conclusion is slightly stronger. In the space-discretized setting, the approximate energy coincides with the discrete energy, making the discrete center of mass method an appropriate alternative to the discrete heat flow. \begin{theorem*}[\autoref{thm:CenterOfMassMethodDiscrete}] Let $\cS$ be a discretized surfaceand $N$ be a compact manifold of negative sectional curvature. In any $\pi_1$-injective homotopy class of maps $\cS\to N$, the center of mass method from any initial map converges to the unique discrete harmonic map. \end{theorem*} \bigskip As a concrete demonstration of the effectiveness of our algorithms, we present in \autoref{sec:Harmony} our own computer implementation of a harmonic map solver: \texttt{Harmony}{} is a freely available computer software with a graphical user interface written in \Cpp{} code, using the Qt framework. This program takes as input the Fenchel-Nielsen coordinates for a pair of Fuchsian representations $\rho_\tL, \rho_\tR \colon \pi_1 S \to \Isom(\bH^2)$ and computes and visualizes the unique equivariant harmonic map. \texttt{Harmony}{}'s main user interface is illustrated in \autoref{fig:HarmonyUserInterface}. In future development of \texttt{Harmony}{}, we plan to compute and visualize harmonic maps for more general target representations that are not necessarily discrete, and for more general target spaces, such as $\bH^3$ and other nonpositively curved symmetric spaces. We have implemented both the discrete heat flow method, with fixed and optimal stepsizes separately, and the $\cosh$-center of mass method, a clever variant of the center of mass method suggested to us by Nicolas Tholozan that is better suited for computations in hyperbolic space (discussed in \autoref{subsec:CoshCenterOfMass}). In practice, the $\cosh$-center of mass method is the most effective, both in number of iterations and execution time (see \autoref{subsec:ExperimentalComparison}). \section{Strong convexity of the energy} \label{sec:StrongConvexity} In this section we study the convexity of the discrete energy functional $E_\cG : \Map_{\text{eq}}(\cG,\bH^2) \to \bR$ introduced in the previous section (\autoref{def:EnergyFunctionalGraph}). In \autoref{subsec:ConvexityRiemannian} we recall basics about convexity and strong convexity in Riemannian manifolds. In \autoref{subsec:ConvexityEnergyExistenceHarmonic} we review the convexity of the energy functional for nonpositively curved target spaces. Next we turn to proving the strong convexity of the discrete energy when the target space is $\bH^2$ with a Fuchsian representation: we first perform some preliminary computations in the hyperbolic plane in \autoref{subsec:H2computations}, and then prove the main theorem in \autoref{subsec:StrongConvexityH2}. In \autoref{subsec:StrongConvexityGeneral} we extend this result to Hadamard spaces with negative curvature. \subsection{Convexity in Riemannian manifolds} \label{subsec:ConvexityRiemannian} The classical notion of convexity in Euclidean vector spaces naturally extends to the Riemannian setting---as Udri\c{s}te puts it \cite[Chapter 1]{MR1326607}, \emph{Riemannian geometry is the natural frame for convexity}. We first give a definition for metric spaces. Recall that geodesics in a metric space $(M,d)$ are harmonic maps from intervals of the real line; more concretely, a curve $\gamma \colon I \subseteq \bR \to M$ in $(M,d)$ is a geodesic if and only if $d(\gamma(t_1), \gamma(t_2)) = v |t_2 - t_1|$ for any sufficiently close $t_1, t_2 \in I$, where $v$ is a positive constant. A real-valued function on $M$ is then called (geodesically) convex when it is convex along geodesics. More precisely: \begin{definition} \label{def:ConvexFunction} Let $(M,d)$ be a metric space. A function $f \colon M \to \bR$ is \emph{convex} if, for every geodesic $\gamma \colon [a, b] \to M$ and for all $t \in [0,1]$: \begin{equation} f(\gamma((1-t)a + tb)) \leqslant (1-t)f(\gamma(a)) + t f(\gamma(b))~. \end{equation} When the inequality is strict for all $t \in (a,b)$, $f$ is called \emph{strictly convex}. Furthermore $f$ is called \emph{$\alpha$-strongly convex}, where $\alpha >0$, if: \begin{equation} f(\gamma((1-t)a + tb)) \leqslant (1-t)f(\gamma(a)) + t f(\gamma(b)) - \alpha \frac{t(1-t)}{2} l(\gamma)^2 \end{equation} where $l(\gamma)$ is the length of $\gamma$. The largest such $\alpha$ is called the \emph{modulus of strong convexity} of $f$. \end{definition} When $M = (M,g)$ is a Riemannian manifold and $f$ is $\cC^2$, one can quickly characterize convex functions in terms of the positivity of their Hessian as a quadratic form. Recall that the Hessian of a $\cC^2$ function $f \colon M \to \bR$ is the symmetric $2$-covariant tensor field on $M$ defined by $\Hess(f) = \nabla (\upd f)$. \begin{proposition} \label{prop:CharacConvexFunction} Let $f \colon M \to \bR$ be a $\cC^2$ function on a Riemannian manifold $(M,g)$. Then: \begin{itemize} \item $f$ is convex if and only if it has positive semidefinite Hessian everywhere. \item $f$ is strictly convex if it has positive definite Hessian everywhere. \item $f$ is $\alpha$-strongly convex if and only if it has $\alpha$-coercive Hessian everywhere: \begin{equation} \forall v \in\upT M \quad \Hess(f)(v,v) \geqslant \alpha \Vert v \Vert^2 \end{equation} \end{itemize} \end{proposition} Convex functions enjoy several attractive properties. Among them, we highlight the straightforward fact that any sublevel set of a convex function is totally convex (\ie{} it contains any geodesic whose endpoints belong to it). \autoref{def:ConvexFunction} and \autoref{prop:CharacConvexFunction} work when $M$ is an infinite-dimensional Riemannian manifold (\eg{} $\cC^\infty(M,N)$ as in \autoref{subsec:EnergyFunctionalHarmonicMaps}), however note that a convex function is not necessarily continuous in that case, whereas it is always locally Lipschitz in finite dimension. We refer to \cite[Chap. 3]{MR1326607} for convex functions on finite-dimensional Riemannian manifolds. \subsection{Convexity of the energy functional} \label{subsec:ConvexityEnergyExistenceHarmonic} We review the convexity of the energy functional when the target is nonpositively curved, whether in the Riemannian sense or in the sense of Alexandrov, and we also address the possibility of strict or strong convexity in these settings. \begin{remark} While strict convexity of the energy is a clear-cut way to prove uniqueness of harmonic maps and strong convexity their existence, neither are necessary. The existence and uniqueness of harmonic maps has been properly characterized both in the smooth case and in more general spaces: see \autoref{subsec:SmoothHeatFlow} and \autoref{subsec:EquivariantHarmonicMaps}. \end{remark} \subsubsection*{Convexity of the energy in the smooth setting} The second variation of the energy functional in the smooth context was first calculated by Eells-Sampson \cite{MR0164306} (cf.~\autoref{prop:SecondVariationalFormulaSmooth}). The next proposition follows immediately from \eqref{eq:HessianEnergySmooth}: \begin{proposition} \label{prop:SmoothEnergyHessianInequality} Let $M$ be and $N$ be smooth Riemannian manifolds. If $N$ has nonpositive sectional curvature, then the Hessian of the energy functional satisfies: \begin{equation} \label{eq:HessianEnergySmoothNegativeCurvature} \forall V\in \Gamma(f^*\upT N) \quad \Hess(E)\evalat{f}(V, V) \geqslant \int_M \Vert \nabla V \Vert^2 \, \upd v_g~. \end{equation} \end{proposition} Recall that the Hessian of the energy is taken with respect to the $L^2$ Riemannian structure on the infinite dimensional manifold $\cC^\infty(M,N)$. In particular, \eqref{eq:HessianEnergySmoothNegativeCurvature} makes it clear that the energy functional is convex. It is tempting to try and get more out of \eqref{eq:HessianEnergySmoothNegativeCurvature}: is $E$ strictly convex? Is it strongly convex? Neither can be true without some obvious restrictions: if $f$ maps into a flat (a totally geodesic submanifold of zero sectional curvature), then the energy is constant along the path that consists in translating $f$ along some constant vector field on the flat. Even when $N$ has negative sectional curvature, whence it has no flats of dimension $>1$, this issue remains for constant maps and maps into a curve. However, one can restrict to a connected component of $\cC^\infty(M,N)$ that does not contain such maps, and there the question becomes interesting. For example when $M$ is compact and $\dim N = \dim M > 1$, the degree of maps is an invariant on the components of $\cC^\infty(M,N)$, and any component of nonzero degree contains only surjective map. When the target is negatively curved, \eqref{eq:HessianEnergySmooth} does guarantee strict convexity: \begin{proposition} \label{prop:StrictConvexitySmoothEnergy} Let $M$ be a Riemannian manifold, let $N$ be a Riemannian manifold of negative sectional curvature. Then the energy functional is strictly convex on any connected component of $\cC^\infty(M,N)$ that does not contain any map of rank everywhere $\leqslant 1$. \end{proposition} \begin{proof} Let $(E_i)$ be a local orthonormal frame in $M$. The integrand for the Hessian of the energy functional $\eqref{eq:HessianEnergySmooth}$ is: \begin{equation} \Vert \nabla V \Vert^2 - \sum_{i=1}^n \left\langle R^N(V, \upd f(E_i)) \upd f(E_i), V \right\rangle \end{equation} Each term $\left\langle R^N(V, \upd f(E_i)) \upd f(E_i), V \right\rangle$ is nonpositive, and is nonzero unless $V$ and $\upd f(E_i)$ are collinear. Indeed, when $V$ and $\upd f(E_i)$ are not collinear: \begin{equation} \left\langle R^N(V, \upd f(E_i)) \upd f(E_i), V \right\rangle = K^N(V, \upd f(E_i)) \left(\Vert V \Vert^2 \, \Vert \upd f(E_i) \Vert^2 - \langle V, \upd f(E_i)\rangle^2\right) < 0 \end{equation} where $K^N(V, \upd f(E_i))$ denotes the sectional curvature of the plane spanned by $V$ and $\upd f(E_i)$. If $\Hess(E)\evalat{f}(V, V)$ vanishes, then the integrand must vanish everywhere, so that (1) $\nabla V = 0$ everywhere, and (2) $\upd f(E_i)$ and $V$ must be collinear for every $i$. From (1) it follows that $V$ has constant length, and from (2) and the fact that $V_x \neq 0$ it follows that $\upd_x f$ maps into $\operatorname{span}(V_x)$ for every $x \in M$. In particular, $f$ has rank $\leqslant 1$ everywhere. \end{proof} As far as the authors are aware, no sufficient conditions for strong convexity of the energy functional are known in the smooth setting. We believe that a quantitative refinement of the previous proof combined with a Poincar\'e-type inequality should guarantee: \begin{conjecture} \label{conj:StrongConvexitySmooth} Strong convexity holds in the setting of \autoref{prop:StrictConvexitySmoothEnergy}. \end{conjecture} \subsubsection*{Convexity of the energy for more general spaces} Defining the energy \emph{à la } Jost as in \autoref{subsec:JostEnergy}, it is straightforward that the energy functional is convex when the target space is negatively curved in a suitable sense. Indeed, let $(M, \mu)$ be a measure space and let $(N,d)$ be a \emph{Hadamard metric space}, \ie{} a complete $\CAT(0)$ metric space. Recall that a $\CAT(0)$ space is a geodesic metric space where any geodesic triangle $T$ is `thinner' than the triangle $T'$ with same side lengths in the Euclidean plane---more precisely, the comparison map $T \to T'$ is distance nonincreasing. In a Hadamard space the distance squared function \begin{equation} d^2 \colon N \times N \to \bR \end{equation} is convex (see \cite{MR1744486} for details). It follows easily that for any choice of nonnegative symmetric kernel $\eta_r$ (cf \autoref{subsec:JostEnergy}), the energy functional $E_r$ is convex on $\upL^2(M,N)$. Furthermore if the energy functional $E$ on $\upL^2(M,N)$ is obtained as a $\Gamma$-limit of $E_r$, then it must also be convex \cite[Thm 11.1]{MR1201152}. In particular, this applies to our energy functional $E_\cG$ by way of \autoref{prop:JostEnergyGraph}: \begin{proposition} \label{prop:EnergyGraphConvex} Let $\cG$ be any $\tilde{S}$-triangulated biweighted graph (\autoref{def:BiweightedGraph}) and let $N$ be a Hadamard metric space. The energy functional $E_\cG \colon \Map_{\text{eq}}(\cG, N) \to \bR$ (\autoref{def:EnergyFunctionalGraph}) is convex. \end{proposition} We stress that the convexity is relative to a metric structure on $\Map_{\text{eq}}(\cG, N)$ which depends on a system of vertex weights (see \autoref{def:VertexWeightedGraph}), but the fact that the energy is convex (respectively strictly or strongly convex) does not depend on the choice of such vertex weights. We will examine conditions that ensure $E_\cG$ is strongly convex, first for $N = \bH^2$ (\autoref{thm:StrongConvexityEnergyGraphH2}), then in Hadamard manifolds with negative curvature (\autoref{thm:StrongConvexityEnergyGraphGeneral}). We highlight some important context: Korevaar-Schoen obtained yet another form of convexity of the energy when the domain $M$ is Riemannian. Their energy functional $E$, which coincides with Jost's for suitable choices \cite{MR2348841}, satisfies the convexity inequality \begin{equation} \label{eq:KorevaarSchoenConvexity} E(f_t) \leqslant (1-t) E(f_0) + tE(f_1) - t(1-t) \int_M \Vert \nabla d(f_0,f_1) \Vert^2~, \end{equation} where $(f_t) \in \upL^2(M,N)$ is a geodesic, \ie{} $f_t(x)$ is a geodesic in $N$ for all $x\in M$. This is a weaker analog of \autoref{prop:SmoothEnergyHessianInequality}. It is again tempting to investigate strong convexity when $N$ has negative curvature bounded away from $0$ and $f$ does not have rank everywhere $\leqslant 1$, but work remains to be done. Mese \cite{MR1938491} claims without proof both an improvement of the convexity statement \eqref{eq:KorevaarSchoenConvexity} and the strict convexity of the energy functional at maps of rank $\leqslant 1$ as in \autoref{prop:StrictConvexitySmoothEnergy}, however neither of these claims are explained as far as we can tell. We also note that there is a mistake in the curvature term of \cite[eq.~(1)]{MR1938491}. (In fairness, this lack of explanation is probably due to Mese's focus on the task of extending the uniqueness of Korevaar-Schoen to the setting where $\partial M=\emptyset$.) Although we prove strong convexity for biweighted graph domains under appropriate restrictions, we suspect that a more general version of this theorem is true, namely an analog of \autoref{conj:StrongConvexitySmooth} for singular spaces. In fact, one can further explore extensions to the equivariant setting, with a suitable condition on the target representation strengthening reductivity. \subsection{Convexity estimates in the hyperbolic plane} \label{subsec:H2computations} In order to study the second variation of the discrete energy for $\bH^2$-valued equivariant maps, we first need some convexity estimates in the hyperbolic plane. The strategy in \autoref{subsec:StrongConvexityH2} will be to reach a contradiction under the assumption that the second variation of the energy is too small; here we derive necessary consequences of a small second variation of the energy in the the elementary cases consisting of two and three vertices. We start with a formula for quadrilaterals in $\bH^2$. \begin{center} \begin{figure}[!ht] \begin{minipage}{.47\textwidth} \begin{center} \begin{tikzpicture}[scale=1.0] \draw (-3,0) node[left] {$A$} node {$\bullet$}; \draw (3,0) node[right] {$B$} node {$\bullet$}; \draw (1,3) node[above] {$C$} node {$\bullet$}; \draw (-2,4) node[above] {$D$} node {$\bullet$}; \draw [thick] (-3, 0) -- (3, 0) ; \draw [thick] (3, 0) -- (1, 3) ; \draw [thick] (1, 3) -- (-2, 4) ; \draw [thick] (-2, 4) -- (-3, 0) ; \draw [thick, ->, >=latex, blue] (-2,0) arc (0:atan(4):1) ; \draw [blue] ({-3 + cos(atan(4)/2)} , {0 + sin(atan(4)/2)} ) node[above right] {$\alpha$} ; \draw [thick, ->, >=latex, blue] ({3 + cos(pi r - atan(3/2))}, {0 + sin(pi r - atan(3/2))}) arc ({pi r - atan(3/2)}:{pi r}:1) ; \draw [blue] ({3 + cos(pi r - atan(3/2)/2))} , {0 + sin(pi r - atan(3/2)/2)} ) node[above left] {$\beta$} ; \end{tikzpicture} \caption{} \label{fig:quadrilateral1} \end{center} \end{minipage} \hspace{5mm} \begin{minipage}{.47\textwidth} \begin{center} \begin{tikzpicture}[scale=1.0] \draw (-3,0) node[left] {$A$} node {$\bullet$}; \draw (3,0) node[right] {$B$} node {$\bullet$}; \draw (1,3) node[above] {$C$} node {$\bullet$}; \draw (-2,4) node[above] {$D$} node {$\bullet$}; \draw [thick] (-3, 0) -- (3, 0) ; \draw [thick] (3, 0) -- (1, 3) ; \draw [thick] (1, 3) -- (-2, 4) ; \draw [thick] (-2, 4) -- (-3, 0) ; \draw (-3, 0) -- (1, 3) ; \draw [thick, ->, >=latex, blue] (-2,0) arc (0:atan(3/4):1) ; \draw [blue] ({-3 + cos(atan(3/4)/2)} , {0 + sin(atan(3/4)/2)} ) node[above right] {$\alpha_1$} ; \draw [thick, ->, >=latex, blue] ({-3 + cos(atan(3/4)},{0 + sin(atan(3/4)}) arc (atan(3/4):atan(4):1) ; \draw [blue] ({-3 + cos(atan(3/4)/2 + atan(4)/2)} , {0 + sin(atan(3/4)/2 + atan(4)/2)} ) node[above right] {$\alpha_2$} ; \draw [thick, ->, >=latex, blue] ({3 + cos(pi r - atan(3/2))}, {0 + sin(pi r - atan(3/2))}) arc ({pi r - atan(3/2)}:{pi r}:1) ; \draw [blue] ({3 + cos(pi r - atan(3/2)/2))} , {0 + sin(pi r - atan(3/2)/2)} ) node[above left] {$\beta$} ; \end{tikzpicture} \caption{} \label{fig:quadrilateral2} \end{center} \end{minipage} \end{figure} \end{center} \begin{proposition} \label{prop:HyperbolicQuadrilateral} Let $A$, $B$, $C$, $D$ be four points in the hyperbolic plane. Let $\alpha$ and $\beta$ denote the oriented angles as shown in \autoref{fig:quadrilateral1}. Then: \begin{equation} \begin{aligned} \cosh (DC) &= \cosh (AB) \big[\cosh (DA) \cosh (BC) + \sinh (DA) \sinh (BC) \cos \alpha \cos \beta\big] \\ &\quad - \sinh (AB) \big[\cosh (DA) \sinh (BC) \cos \beta + \sinh (DA) \cosh (BC) \cos \alpha\big] \\ &\quad - \sinh (DA) \sinh (BC) \sin \alpha \sin \beta~. \end{aligned} \end{equation} \end{proposition} \begin{remark} This equation holds without restriction on $\alpha$ and $\beta$; they may be negative or obtuse. \end{remark} \begin{proof} Referring to \autoref{fig:quadrilateral2}, the hyperbolic law of cosines implies: \begin{equation} \label{eq:coshACD} \cosh(DC) = \cosh(DA)\cosh(AC) - \sinh(DA) \sinh(AC) \cos(\alpha_2)~. \end{equation} The hyperbolic laws of sines and cosines in the triangle $ABC$ give \begin{align} \cosh(AC) & = \cosh(AB)\cosh(BC) - \sinh(AB) \sinh(BC) \cos(\beta)~, \ \text{ and } \label{eq:coshABC} \\ \sinh(AC) \cos(\alpha_2) &= \sinh(AC) \cos(\alpha_1 - \alpha)\\ &= \sinh(AC) \cos (\alpha_1 ) \cos(\alpha) + \sinh(AC) \sin(\alpha_1) \sin(\alpha) \\ &= \sinh(AC) \cos (\alpha_1 ) \cos(\alpha) + \sinh(BC) \sin(\beta) \sin (\alpha)~. \label{eq:alphasumh} \end{align} Moreover, it is a consequence of the two forms of the hyperbolic law of cosines (see \eg \cite[p.~82]{MR2249478}) in the triangle $ABC$ that we have \begin{equation} \label{eq:sinhACcosalpha1} \sinh(AC) \cos(\alpha_1) = \sinh(AB) \cosh(BC) - \sinh(BC) \cosh(AB) \cos(\beta)~. \end{equation} Equation \eqref{eq:sinhACcosalpha1} allows us to rewrite equation \eqref{eq:alphasumh} as: \begin{equation} \label{eq:ACcosalphatwoh} \begin{split} \sinh(AC) \cos(\alpha_2) &= \big(\sinh(AB) \cosh(BC) - \sinh(BC) \cosh(AB) \cos(\beta)\big) \cos(\alpha) \\ &\quad + \sinh(BC) \sin(\beta) \sin(\alpha)~. \end{split} \end{equation} Together \eqref{eq:coshABC} and \eqref{eq:ACcosalphatwoh} and \eqref{eq:coshACD} imply the desired equation. \end{proof} Next we study the convexity of the energy for two points, which amounts to analyzing the second variation of the half-distance squared function $ \frac{d^2}{2} \colon \bH^2 \times \bH^2 \to \bR$. We perform this computation in two stages: first we study instead the function $(\cosh d) -1 \colon \bH^2 \times \bH^2 \to \bR$, as it is better suited to computations, and then we relate the second variation of the two functions. \begin{proposition} \label{prop:DerivativesF} Let $A$ and $B$ be two points in the hyperbolic plane at distance $D$. Let $\vec{u}$ and $\vec{v}$ be tangent vectors at $A$ and $B$ respectively. Let $A_t = \exp_A (t \vec{u})$ and $B_t = \exp_B (t \vec{v})$ for $t \in \bR$, and consider the function $F_{AB}(t) = \cosh\left(d(A_t, B_t)\right) - 1$. Then: \begin{align} \frac{\upd}{\upd t} \evalat{t=0} F_{AB}(t) & = -\sinh(D) \big( \Vert \vec{u} \Vert \cos \alpha - \Vert \vec{v} \Vert \cos \beta \big) \\ \frac{\upd^2}{\upd t^2}\evalat{t=0} F_{AB}(t) & = \cosh (D) \big(\Vert \vec{u} \Vert^2 + \Vert \vec{v} \Vert ^2 - 2 \Vert \vec{u} \Vert \Vert \vec{v} \Vert \cos \alpha \cos \beta\big) -2 \Vert \vec{u} \Vert \Vert \vec{v} \Vert \sin \alpha \sin \beta~. \end{align} where $\alpha$ (resp. $\beta$) is the oriented angle between the geodesic $AB$ and the vector $\vec{u}$ (resp. $\vec{v}$). \end{proposition} \begin{proof} Consider the quadrilateral given by the four points $A$, $B$, $C = B_t$, $D = A_t$. Note that the angle $\beta$ here corresponds to the angle $\pi - \beta$ of \autoref{prop:HyperbolicQuadrilateral}. By direct application of \autoref{prop:HyperbolicQuadrilateral}, \begin{equation} \begin{aligned} 1 + F_{AB}(t) &= \cosh (D) \big[\cosh (t \Vert \vec{u} \Vert) \cosh (t \Vert \vec{v} \Vert) - \sinh (t \Vert \vec{u} \Vert) \sinh (t \Vert \vec{v} \Vert) \cos \alpha \cos \beta\big] \\ & \quad - \sinh (D) \big[- \cosh (t \Vert \vec{u} \Vert) \sinh (t \Vert \vec{v} \Vert) \cos \beta + \sinh (t \Vert \vec{u} \Vert) \cosh (t \Vert \vec{v} \Vert) \cos \alpha\big] \\ &\quad - \sinh (t \Vert \vec{u} \Vert) \sinh (t \Vert \vec{v} \Vert) \sin \alpha \sin \beta~. \end{aligned} \end{equation} The result follows immediately by taking the first and second derivatives at $t=0$. \end{proof} \begin{proposition} \label{prop:DerivativesE} We keep the same setup as \autoref{prop:DerivativesF}, and let $E_{AB}(t) = \frac{1}{2} d(A_t, B_t)^2$. Then: \begin{align} \frac{\upd^2}{\upd t^2}\evalat{t=0} E_{AB}(t) & = a + b \; D \tanh (D/2) + c \left(D \coth D -D\tanh(D/2) \right) ~, \end{align} where $a$, $b$, and $c \geqslant 0$ are given by \begin{align} a & = \left( \Vert \vec{u} \Vert \cos \alpha - \Vert\vec{v} \Vert\cos \beta \right)^2~, \\ b & = \Vert \vec{u} \Vert^2 \sin^2 \alpha + \Vert\vec{v} \Vert^2 \sin^2 \beta ~, \text{ and } \\ c &= \left( \Vert \vec{u} \Vert \sin \alpha - \Vert\vec{v} \Vert\sin \beta \right)^2~. \end{align} \end{proposition} \begin{proof} For ease in notation, we leave the subscripts $AB$ from $E_{AB}$ and $F_{AB}$ in what follows. We have $E(t) = \phi \circ F(t)$, where $\phi(x) = \frac 12 \left( \arcosh (1+x) \right)^2$. It is straightforward to check that \begin{align} \phi'(\cosh x -1) = \frac{x}{\sinh x}~, \ \ \ \text{ and } \ \ \ \phi''(\cosh x-1) = \frac{\sinh x- x\cosh x}{\sinh^3 x}~. \end{align} Since we have $E''(0) = \phi''(F(0)) \left( F'(0) \right)^2 + \phi'(F(0)) F''(0)$, using \autoref{prop:DerivativesF} we find that \begin{align} E''(0) & = \frac{\sinh D - D\cosh D}{\sinh^3 D} \cdot \sinh^2D \left(\Vert\vec{u}\Vert\cos \alpha - \Vert\vec{v}\Vert \cos \beta\right)^2 \\ & \quad + \frac{D}{\sinh D} \left( \cosh D \left( \Vert\vec{u}\Vert^2+ \Vert\vec{v}\Vert^2 -2\Vert\vec{u}\Vert\Vert\vec{v}\Vert\cos \alpha\cos \beta\right) - 2\Vert\vec{u}\Vert\Vert\vec{v}\Vert\sin\alpha\sin\beta\right) \\ & = \left(\Vert\vec{u}\Vert\cos \alpha - \Vert\vec{v}\Vert \cos \beta\right)^2 + D\coth D \left( -\left(\Vert\vec{u}\Vert\cos \alpha - \Vert\vec{v}\Vert \cos \beta\right)^2 + \right. \\ & \quad \left. \left( \Vert\vec{u}\Vert^2+ \Vert\vec{v}\Vert^2 -2\Vert\vec{u}\Vert\Vert\vec{v}\Vert\cos \alpha\cos \beta\right) \right) - D \csch D \left( 2\Vert\vec{u}\Vert\Vert\vec{v}\Vert\sin\alpha\sin\beta\right) \\ & = a + b \; D\coth D +(c-b)\; D \csch D ~. \end{align} To finish, note that $\coth D - \csch D = \tanh(D/2)$. \end{proof} The following quantitative control is at the core of strong convexity for $E_\cG$: \begin{proposition} \label{prop:SecondDerivativeHyperbolicEstimates} Let $E_{AB}(t)$ be the function as in \autoref{prop:DerivativesE}. We have \begin{equation} \frac{\upd^2}{\upd t^2}\evalat{t=0} E_{AB}(t) \geqslant \Vert \vec{u} - P_{[BA]}\vec{v} \Vert^2~, \end{equation} where $P_{[BA]}\vec{v}$ denotes the parallel transport of $\vec{v}$ along the geodesic segment $BA$. \end{proposition} \begin{proof} By rewriting $E_{AB}''(0)$ using \autoref{prop:DerivativesE}, and noting that $2b\geqslant c$ by the Cauchy-Schwarz inequality, we find \begin{align} \label{eq:abcEAB} E_{AB}''(0) & = a + b \; D \tanh \frac D2 + c \left(D\coth D - D\tanh \frac D2\right) \\ & \geqslant a + c\left( D\coth D - \frac D2\tanh \frac D2 \right) = a + c \cdot \frac D2 \coth \frac D2~. \end{align} Because $x\coth x\geqslant1 $, we find that $E_{AB}''(0) \geqslant a+c$. The proof is completed by checking that the quantity $a+c$ is precisely $\Vert \vec{u} - P_{[BA]}\vec{v} \Vert^2$: \begin{align*} a +c & = \left( \Vert \vec{u} \Vert \cos \alpha - \Vert\vec{v} \Vert\cos \beta \right)^2 + \left( \Vert \vec{u} \Vert \sin \alpha - \Vert\vec{v} \Vert\sin \beta \right)^2~ \\ &= \Vert \vec{u} \Vert^2 - 2\Vert \vec{u} \Vert \Vert \vec{v} \Vert \cos(\alpha-\beta) + \Vert \vec{v} \Vert^2 \\ &= \langle \vec{u} ,\vec{u} \rangle - 2\langle \vec{u} , P_{[BA]}\vec{v} \rangle + \langle \vec{v},\vec{v} \rangle \\ & = \Vert \vec{u} - P_{[BA]}\vec{v} \Vert^2~. \qedhere \end{align*} \end{proof} \subsection{\texorpdfstring{Strong convexity of the discrete energy in $\bH^2$}{Strong convexity of the discrete energy in H2}} \label{subsec:StrongConvexityH2} Let $\cG$ be any $\tilde{S}$-triangulated biweighted graph (\autoref{def:BiweightedGraph}) and let $\rho_\tR \colon \pi_1S \to \Isom(\bH^2)$ be a Fuchsian representation. We are ready to prove strong convexity of the discrete energy functional $E_\cG \colon \Map_{\text{eq}}(\cG, \bH^2) \to \bR$ introduced in \autoref{def:EnergyFunctionalGraph}. Choose once and for all a fundamental domain for the action of $\pi_1S$ on $\cG^{(0)}$, consisting of vertices $\{p_1, \dots, p_n\} \subseteq \cG^{(0)}$. Recall that $\cG$ is equipped with vertex and edge weights; these are completely determined by the weights $\mu_i = \mu(p_i)$ and $\omega_{ij}=\omega(e_{p_i p_j})$ for $i,j \in \{1, \dots, n\}$. Fix an equivariant map $f\in\Map_{\text{eq}}(\cG,\bH^2)$, recorded by the tuple $(x_1,\dots, x_n) \in (\bH^2)^n$ where $x_i \coloneqq f(p_i)$. Also fix a tangent vector $ \vec{v} \in \upT_f\Map_{\text{eq}}(\cG, \bH^2)$, given by $\vec{v} = (\vec{v}_1,\dots,\vec{v}_n)$ where $v_i \in \upT_{x_i}\bH^2$ as in \eqref{eq:MapGammaTangentSpace2}. We assume that $\vec{v}$ is a unit tangent vector: by \autoref{def:L2RiemannianMetricGraph} this means $\sum_i \mu_i \Vert \vec{v_i} \Vert^2 = 1$. We want to compute the second derivative at $t=0$ of $E_\cG(t) \coloneqq E_\cG \circ \exp_{f}(t \vec{v})$. Let us denote \begin{equation} E_{ij}(t) \coloneqq \frac 12 d\;(\exp_{x_i}(t\vec{v_i}) , \exp_{x_j}(t\vec{v_j}))^2 \end{equation} for each edge $e_{ij}$ between points $p_i$ and $p_j$. First we observe that if the second variation of $E_\cG$ is small, then \autoref{prop:SecondDerivativeHyperbolicEstimates} implies that the tangent vectors $\vec{v_i}$ all have approximately the same length. Precisely, \begin{lemma} \label{lem:meshVectorsEqualLength} For all $i$ we have \begin{equation} \left| \; \Vert \vec{v_i} \Vert - \frac{1}{\sqrt A} \; \right| < 3 \sqrt{E_\cG''(0) \ \frac{ D}{\Omega}}~, \end{equation} where $D$ is the diameter of the quotient graph $\cG/\pi_1S$, $A=\sum_i \mu_i$, and $\Omega=\min \omega_{ij}$. \end{lemma} \begin{proof} Let $\varepsilon_{ij} = E_{ij}''(0)$ for each edge $e_{ij}\in\cE$, so that $ E_\cG''(0) = \sum_{e_{ij}\in\cE} \omega_{ij} \varepsilon_{ij} $. Observe that for each edge $e_{ij}\in\cE$, one finds that $|\Vert\vec{v_i}\Vert - \Vert\vec{v_j}\Vert|< \sqrt{\varepsilon_{ij}}$ by \autoref{prop:SecondDerivativeHyperbolicEstimates}. For any pair of points $p_i$ and $p_j$, choose a path $p_i= p_{i_0}, p_{i_1}, \dots, p_{i_r} = p_j$ and observe that by the triangle inequality and the Cauchy-Schwarz inequality we have: \begin{align} \left| \Vert\vec{v_i}\Vert - \Vert \vec{v_j} \Vert \right| & \leqslant \sum_{k=1}^r \sqrt{\varepsilon_{i_{k-1}i_k} } \leqslant \left( \sum_{k=1}^r \varepsilon_{i_{k-1}i_k}\right)^{1/2} \cdot \left( \sum_{k=1}^r 1\right)^{1/2} ~. \end{align} Now the path may be taken with length $r\leqslant D$, and $ \Omega \sum \varepsilon_{ij} \leqslant E_\cG''(0)$, so the above inequality implies \begin{equation} \label{eq:ViVj} \left| \; \Vert\vec{v_i}\Vert - \Vert \vec{v_j} \Vert \; \right| < \sqrt{\frac{E_\cG''(0) \ D}{\Omega}}~. \end{equation} Let $\delta=\sqrt{ E_{\cG}''(0) \ D /\Omega}$. As $| \|\vec{v_i}\| - \| \vec{v_j}\| | < \delta$ for all $i,j$, \begin{align*} | \|\vec{v_i}\|^2 - \|\vec{v_j}\|^2 | &= | \|\vec{v_i}\| - \|\vec{v_j}\| | \cdot | \|\vec{v_i}\| + \|\vec{v_j}\| | \\ & \le | \|\vec{v_i}\| - \|\vec{v_j}\| | \cdot ( | \|\vec{v_i}\| - \|\vec{v_j}\| | + 2\|\vec{v_i}\|) \\ & < \delta^2 + 2 \delta \|\vec{v_i}\|~. \end{align*} As $\vec{v}$ is a unit tangent vector and $A = \sum_j \mu_j$, we have \begin{align*} 1 &= \sum_j \mu_j \|\vec {v_j}\|^2 = A \|\vec{v_i}\|^2 + \sum_j \mu_j(\|\vec{v_j}\|^2 - \|\vec{v_i}\|^2) ~. \end{align*} Rearranging and taking absolute values we find \[ | A \|\vec{v_i}\|^2 - 1 | \leqslant \sum_j \mu_j | \|\vec{v_i}\|^2 - \|\vec{v_j}\|^2 | < A (\delta^2 + 2 \delta \|\vec{v_i}\|)~. \] Dividing by $A$ we obtain \[ \left| \|\vec{v_i}\|^2 - \frac 1A \right| < \delta^2 + 2 \delta \|\vec{v_i}\|~. \] This implies that we have \begin{align*} \|\vec{v_i}\|^2 - 2 \delta \|\vec{v_i}\| - \delta^2 &< \frac 1 A~, \text{ and} \\ \|\vec{v_i}\|^2 +2 \delta \|\vec{v_i}\| + \delta^2 &> \frac 1 A~. \end{align*} In other words, \begin{align*} (\|\vec{v_i}\| - \delta )^2 &< \frac 1 A + 2\delta^2~, \text{ and} \\ (\|\vec{v_i}\| + \delta )^2 &< \frac 1 A ~. \end{align*} We conclude that $\frac1{\sqrt{A}} -\delta < \|\vec{v_i}\| < \delta + \sqrt{ \frac 1A + 2\delta^2 } <\frac1 {\sqrt{A}} + (1+\sqrt2)\delta < \frac1 {\sqrt{A}} + 3\delta$, as desired. \end{proof} Together with \autoref{prop:SecondDerivativeHyperbolicEstimates}, this is enough to produce a lower bound for the second variation: \begin{proposition} \label{prop:SecondDerivativeEstimateBelow} We have \begin{equation} \label{eq:SecondDerivativeLowerBound1} \frac{\upd^2}{\upd t^2}\evalat{t=0} E_\cG (t) \geqslant \frac1{9A\left(1+\sqrt{\frac{D}{\Omega}}\right)^2}~, \end{equation} where $D$ is the diameter of the quotient graph $\cG/\pi_1S$, $A=\sum_i \mu_i$, and $\Omega = \min \omega_{ij}$ is the minimum edge weight. \end{proposition} \begin{remark} \label{rem:collectingConstants} Observe that this demonstrates that there is a constant $C>0$ independent of $\cG$ so that the modulus of convexity of $E_\cG$ is at least $ \frac{C \Omega}{AD}$. When $\cG$ is the biweighted graph underlying a Fuchsian representation as in \autoref{def:MeshToGraph}, the modulus of convexity is at least $\frac{C \; \Omega}{|\chi(S)|D}$. \end{remark} \begin{proof} Suppose towards contradiction that $E_\cG''(0) < \frac19 A^{-1} \left(1+\sqrt{\frac{D}{\Omega}}\right)^{-2}$, i.e.~ \begin{equation}\label{eq:contradiction} \sqrt{E_\cG''(0)} < \frac1{3\sqrt A} - \sqrt{E_\cG''(0)\cdot \frac{D}{\Omega}}~. \end{equation} Rearranging, we find that \begin{equation} \frac{\sqrt{E_\cG''(0)}}{\frac1{\sqrt{A}} - 3\sqrt{E_\cG''(0)\cdot \frac{D}{\Omega}}} < \frac13~. \end{equation} Now consider an edge $x_i\sim x_j$ in $\cG$, let $\delta_{ij} = \| \vec{v_i} - P_{[x_jx_i]} \vec{v_j} \|$, and let $\theta_{ij}$ indicate the principal value of the angle formed between $\vec{v_i}$ and $P_{[x_jx_i]}\vec{v_j}$. Evidently, $\| \vec{v_i} \| |\theta_{ij}|$ is bounded by $\frac \pi 2$ times the distance between $ \vec{v_i} $ and $\frac{\|\vec{v_i}\|}{\|\vec{v_j}\|} P_{[x_jx_i]}\vec{v_j}$, and the latter distance is bounded by $\delta_{ij} + \left| \|\vec{v_i}\| - \|\vec{v_j}\|\right| \leqslant 2\delta_{ij}$. Therefore we have \begin{equation} \| \vec{v_i} \|| \theta_{ij}| \leqslant \pi \cdot \delta_{ij} ~. \end{equation} It is not hard to see that \eqref{eq:contradiction} implies $3(E_\cG''(0)\cdot D/\Omega)^{1/2} < \frac 1{\sqrt{A}}$, so by \autoref{lem:meshVectorsEqualLength} we conclude that $ \|\vec{v_i}\| \geqslant \frac1{\sqrt{A}} - 3(E_\cG''(0)\cdot D/\Omega)^{1/2}>0$. In that case we may rearrange the above to find that \begin{equation}\label{eq:bound angle} |\theta_{ij}| \leqslant \pi \frac{\delta_{ij}}{\|\vec{v_i}\|} \leqslant \pi \cdot \frac{\sqrt{E_\cG''(0)}}{\frac1{\sqrt{A}} - 3\sqrt{E_\cG''(0)\cdot D/\Omega}} < \frac\pi 3~, \end{equation} by our assumption \eqref{eq:contradiction}. Now extend the vectors $\vec{v_i}$ at the vertices of $\cG$ over each of the edges of the graph $\cG \subset Y$ by interpolating linearly. Consider a triangle $T$ in $\cG$ with vertices $x_i,x_j,x_k$ and boundary contour $\gamma$, oriented counterclockwise. The winding number of $\vec{V}$ with respect to $\gamma$, denoted $\omega_\gamma(\vec{V})$, is given by \begin{equation} \omega_\gamma(\vec{V}) = \frac{1}{2\pi}\left( \theta_{ij}+\theta_{jk}+\theta_{ki} + \theta_i + \theta_j + \theta_k \right)~, \end{equation} where $\theta_i,\theta_j,\theta_k$ are the exterior angles of $T$ at $x_i,x_j,x_k$, respectively. Evidently, the sum of the exterior angles is at least $\pi$, and at most $3\pi$. Together with \eqref{eq:bound angle}, we find that $0<\omega_\gamma(\vec{V}) <2$, so it follows that $\omega_\gamma(\vec{V})=1$ for every triangle $T$. Now $\vec{V}$ can be extended over the triangles to a nonvanishing section of $f^*\upT Y$, contradicting \autoref{lem:no sections}. \end{proof} \begin{lemma} \label{lem:no sections} Suppose that $f:S \to Y$ is a map between surfaces with $\chi(Y)\ne0$ and $\deg f \ne 0$. Then $f^*\upT Y$ has no nonvanishing sections. \end{lemma} \begin{proof} Let $e$ indicate the \emph{Euler class}. We have \begin{equation} \label{eq:euler cl} \langle e(f^*\upT Y),[S] \rangle = \langle f^*e(\upT Y) , [S] \rangle = \langle e(\upT Y) , f_*[S] \rangle = \deg f \cdot \chi(Y)~, \end{equation} which is nonzero by assumption. Therefore $e(f^*\upT Y) \ne 0$, and $f^*\upT Y$ has no nonvanishing sections. \end{proof} The discrete heat flow will also require an upper bound for the Hessian of $E_\cG$ (see \autoref{sec:DiscreteHeatFlow}). \begin{proposition} \label{prop: second derivative estimate above} Suppose that $E_\cG(0) \leqslant E_0$. Then \begin{equation} \label{eq:SecondDerivativeUpperBound} \frac{\upd^2}{\upd t^2}\evalat{t=0} E_\cG (t) \leqslant \frac{2 V W}{ U}\left( 1+ \sqrt{\frac{E_0}{\Omega}} \coth \sqrt{\frac{E_0}{\Omega}}\right) \eqqcolon \beta \end{equation} where $V$ is the maximum valence of vertices of $\cG$, $U= \min\{\mu_i\}$ is the minimum vertex weight, $\Omega=\min\{\omega_{ij}\}$ is the minimum edge weight, and $W=\max\{\omega_{ij}\}$ is the maximum edge weight. \end{proposition} \begin{proof} It is not hard to see from \autoref{prop:DerivativesE} that we have \begin{equation} \label{eq:HessianUpperBound} \frac{\upd^2}{\upd t^2}\evalat{t=0} E_{ij}(t) \leqslant 2 \left( \Vert\vec{v_i}\Vert^2 + \Vert\vec{v_j}\Vert^2\right) \left(1+ d(x_i,x_j) \coth d(x_i,x_j) \right)~. \end{equation} Letting $L=\max\{ d(x_i,x_j)\}$ we find \begin{align} \frac{\upd^2}{\upd t^2}\evalat{t=0} E_\cG(t) &= \sum_{e_{ij}\in\cE} \omega_{ij} \frac{\upd^2}{\upd t^2}\evalat{t=0} E_{ij}(t) \leqslant 2W \left(1+ L \coth L\right) \sum_{e_{ij}\in\cE} \left(\Vert \vec{v_i}\Vert^2 + \Vert \vec{v_j}\Vert^2 \right) \\ & \leqslant \frac {2W}{U} \left(1+ L \coth L\right) \sum_{e_{ij}\in\cE} \left( \mu_i \Vert \vec{v_i}\Vert^2 + \mu_j \Vert \vec{v_j}\Vert^2 \right) \\ & \leqslant \frac{2 VW}{U} \left(1 + L \coth L\right) \sum_i \mu_i \Vert\vec{v_i}\Vert^2 = \frac{2 VW}{U} \left(1 + L \coth L\right)~. \end{align} The assumption $E_\cG(0) \leqslant E_0$ implies that $\Omega L^2 \leqslant E_0$, so we are done. \end{proof} Together, \autoref{prop: second derivative estimate above} and \autoref{prop:SecondDerivativeEstimateBelow} imply: \begin{theorem}\label{thm:StrongConvexityEnergyGraphH2} Suppose that $\cG$ is a biweighted triangulated graph, $N = \bH^2$ is the hyperbolic plane and $\rho = \rho_\tR \colon \pi_1S\to \Isom^+(\bH^2)$ is Fuchsian. Then the energy functional $E_\cG : \Map_{\text{eq}}(\cG, N) \to \bR$ is strongly convex. More precisely, \begin{equation} \forall \vec{v} \quad \alpha \Vert \vec{v} \Vert^2 \leqslant \Hess(E_\cG)(\vec{v},\vec{v}) \end{equation} where $\alpha$ is given explicitly by \eqref{eq:SecondDerivativeLowerBound1}. Moreover, on the compact convex set $\{E \leqslant E_0\} \subseteq \Map_{\text{eq}}(\cG, N)$, \begin{equation} \forall \vec{v} \quad \Hess(E_\cG)(\vec{v},\vec{v}) \leqslant \beta \Vert \vec{v} \Vert^2 \end{equation} where $\beta$ is given explicitly by \eqref{eq:SecondDerivativeUpperBound}. \end{theorem} \subsection{More general target spaces} \label{subsec:StrongConvexityGeneral} Some of the steps in the proof of \autoref{thm:StrongConvexityEnergyGraphH2} actually hold in much greater generality. For instance, \autoref{prop:SecondDerivativeHyperbolicEstimates} holds verbatim if we replace $\bH^2$ with any Hadamard manifold. Keeping the same setup as in \autoref{subsec:H2computations}, let $A,B\in N$, let $\vec{u}$ and $\vec{v}$ be vectors in $\upT_AN$ and $\upT_BN$, respectively, let $A_t= \exp_A(t\vec{u})$ and $B_t=\exp_B(t\vec{v})$, and let $E_{AB}(t) = \frac 12 d_N(A_t,B_t)^2$. \begin{proposition} \label{prop:SecondDerivativeGeneralEstimates} Suppose that $N$ is a Hadamard manifold. Then we have \begin{equation} \frac{\upd^2}{\upd t^2}\evalat{t=0} E_{AB}(t) \geqslant \Vert \vec{u} - P_{[BA]}\vec{v} \Vert^2~. \end{equation} \end{proposition} The proof below builds on \autoref{subsec:H2computations}. \begin{proof} We start by noting that the computations for $E''_{AB}(0)$ (\autoref{prop:DerivativesF} and \autoref{prop:DerivativesE}) simplify dramatically when $N=\bR^2$. In this case we have \begin{equation} \label{eq:secondDerivEuclidean} \frac{\upd^2 }{\upd t^2} E_{AB}(t) = \frac 12\Vert \vec{u} - \vec{v} \Vert^2 ~. \end{equation} In particular, if $\vec{u}\ne \vec{v}$ then $E_{AB}(t)$ is $\frac 12 \Vert \vec{u} - \vec{v} \Vert^2$-strongly convex. Now we turn to the general case. Let $r,s\in \bR$, and consider the quadrilateral through $A_{r}$, $B_{r}$, $B_{s}$, and $A_{s}$. Because Hadamard manifolds are $\CAT(0)$, this quadrilateral has a comparison quadrilateral in $\bR^2$ with vertices $ A_{r}'$, $ B_{r}'$, $ B_{s}'$, $ A_{s}'$. For any $t\in\bR$, as a consequence of \eqref{eq:secondDerivEuclidean} we have \begin{align} \label{eq:comparisonQuad} d(A_{(1-t)r+ts},B_{(1-t)r+ts})^2 & \leqslant d(A_{(1-t)r+ts}',B_{(1-t)r+ts}')^2 \\ & \leqslant (1-t) d(A_{r}',B_{r}')^2 + t d(A_{s}',B_{s}')^2 - \Vert\vec{u}' - \vec{v}'\Vert^2 \frac{ t(1-t)}{4} |r-s|^2 \\ & \leqslant (1-t) d(A_{r},B_{r})^2 + t d(A_{s},B_{s})^2 - \Vert\vec{u}' - \vec{v}'\Vert^2 \frac{ t(1-t)}{4} |r-s|^2~, \end{align} where $\vec{u}'=A_{s}'-A_{r}'$ and $\vec{v}'=B_{s}'-B_{r}'$ in $\bR^2$, respectively. In terms of $E_{AB}$ we have \begin{equation} \label{eq:EABinequality1} E_{AB}((1-t)r+ts) \leqslant (1-t) E_{AB}(r) + t E_{AB}(s) - \Vert\vec{u}' - \vec{v}'\Vert^2 \frac{ t(1-t)}{4}|r-s|^2 ~. \end{equation} This is almost an equivalent formulation of the strong convexity of $E_{AB}$ (see \autoref{def:ConvexFunction}), however the term $ \Vert\vec{u}' - \vec{v}'\Vert^2 $ depends on both $r$ and $s$. In fact, this detail is essential, as $E_{AB}$ may fail to be strongly convex. For $\delta>0$, taking $r=s=\delta$ and $t=\frac 12$, one finds \begin{equation} E_{AB}(0) = E_{AB}\left(\frac 12 \delta + \frac 12(-\delta) \right) \leqslant \frac 12 E_{AB}(\delta) + \frac 12 E_{AB}(-\delta) - \Vert\vec{u}'-\vec{v}'\Vert^2 \frac{\delta^2}{2}~. \end{equation} (We stress the dependence of $ \Vert\vec{u}'-\vec{v}'\Vert^2$ on $\delta$.) Rearranging we find that \begin{equation} \label{eq:EABinequality2} \Vert\vec{u}'-\vec{v}'\Vert^2 \leqslant \frac{ E_{AB}(\delta) - 2E_{AB}(0) + E_{AB}(-\delta)}{\delta^2}~. \end{equation} As $\delta\to 0$, the right-hand side approaches $\frac{\upd^2}{\upd t^2}\evalat{t=0} E_{AB}(t)$. As for the left-hand side, we may rewrite \begin{align} \Vert\vec{u}' - \vec{v}'\Vert^2 & = \Vert\vec{u}'\Vert^2 + \Vert\vec{v}'\Vert^2 - 2\Vert\vec{u}'\Vert \Vert\vec{v}'\Vert \cos (\alpha' -\beta') \\ & =\Vert\vec{u}_{\delta}\Vert^2 + \Vert\vec{v}_{\delta}\Vert^2 - 2\Vert\vec{u}_{\delta}\Vert \Vert\vec{v}_{\delta}\Vert \cos (\alpha' -\beta')~, \end{align} where $\vec{u}_{\delta} = \exp_{A_{-\delta}}^{-1}(A_\delta)$ and $\vec{v}_{\delta}=\exp_{B_{-\delta}}^{-1}(B_\delta)$ (so that $\Vert\vec{u}_{\delta}\Vert=\Vert\vec{u}'\Vert$ and $\Vert\vec{v}_{\delta}\Vert=\Vert\vec{v}'\Vert$), and $\alpha'$ (resp.~$\beta'$) are the oriented angles (measured in $\bR^2$) between the geodesic $A_{-\delta}B_{-\delta}$ and $\vec{u}'$ (resp.~$\vec{v}'$). By a well-known theorem of Alexandrov (see \cite{MR0049584} or \cite{MR1744486}), the interior angles of a quadrilateral of $N$ are smaller than those of the model, and we conclude that $\alpha_\delta \leqslant \alpha'$ and $\pi-\beta_\delta \leqslant \pi- \beta'$, where $\alpha_\delta$ and $\beta_\delta$ are the interior angles at $A_{-\delta}$ and $B_{-\delta}$, respectively, of the quadrilateral with vertices $A_{-\delta}$, $B_{-\delta}$, $B_{\delta}$, and $A_\delta$. Thus $\alpha_\delta - \beta_\delta \leqslant \alpha'-\beta'$, and we may conclude that \begin{equation} \Vert\vec{u}' - \vec{v}'\Vert^2 \geqslant \Vert\vec{u}_{\delta}\Vert^2 + \Vert\vec{v}_{\delta}\Vert^2 - 2\Vert\vec{u}_{\delta}\Vert \Vert\vec{v}_{\delta}\Vert \cos (\alpha_\delta -\beta_\delta)~. \end{equation} Using the latter for the lefthand side of \eqref{eq:EABinequality2} and taking the limit as $\delta\to 0$, we find that \begin{equation} \label{eq:EABinequality3} \Vert\vec{u}\Vert^2 + \Vert\vec{v}\Vert^2 - 2\Vert\vec{u}\Vert \Vert\vec{v}\Vert \cos (\alpha -\beta) \leqslant \frac{\upd^2}{\upd t^2}\evalat{t=0} E_{AB}(t)~. \end{equation} The lefthand side here is precisely $ \Vert \vec{u} - P_{[BA]}\vec{v} \Vert^2$, as in the proof of \autoref{prop:SecondDerivativeHyperbolicEstimates}. \end{proof} This in turn makes a generalization of \autoref{thm:StrongConvexityEnergyGraphH2} possible: \begin{theorem} \label{thm:StrongConvexityEnergyGraphGeneral} Let $\cG$ be a biweighted $\tilde{S}$-triangulated graph, let $N$ be a two-dimensional closed Riemannian manifold of nonzero Euler characteristic and nonpositive sectional curvature and suppose that $\rho \colon \pi_1S\to \pi_1 N$ is a group homomorphism induced by a homotopy class of maps $S \to N$ of nonzero degree. Then the energy functional $E_\cG : \Map_{\text{eq}}(\cG, \tilde{N}) \to \bR$ is strongly convex with modulus of convexity given by \eqref{eq:SecondDerivativeLowerBound1}. \end{theorem} \begin{proof} \autoref{prop:SecondDerivativeGeneralEstimates} may take the place of \autoref{prop:SecondDerivativeHyperbolicEstimates}, and we see that \autoref{lem:meshVectorsEqualLength} holds for any nonpositively curved target. Now the proof of \autoref{prop:SecondDerivativeEstimateBelow} holds verbatim. (The reader can observe that the fourth sentence of the last paragraph still holds in the setting of nonpositive curvature: the sum of exterior angles of a geodesic triangle is between $\pi$ and $3\pi$.) \end{proof}

https://arxiv.org/abs/2204.00383

A visualisation for conveying the dynamics of iterative eigenvalue algorithms over PSD matrices

We propose a new way of visualising the dynamics of iterative eigenvalue algorithms such as the QR algorithm, over the important special case of PSD (positive semi-definite) matrices. Many subtle and important properties of such algorithms are easily found this way. We believe that this may have pedagogical value to both students and researchers of numerical linear algebra. The fixed points of iterative algorithms are obtained visually, and their stability is analysed intuitively. It becomes clear that what it means for an iterative eigenvalue algorithm to "converge quickly" is an ambiguous question, depending on whether eigenvalues or eigenvectors are being sought. The presentation is likely a novel one, and using it, a theorem about the dynamics of general iterative eigenvalue algorithms is proved. There is an accompanying video series, currently hosted on Youtube, that has certain advantages in terms of fully exploiting the interactivity of the visualisation.

\section{Simple iterative eigenvalue algorithms} The (naive) QR algorithm {\cite{francis1961qr,wilkinson1968global,GoluVanl96}} evaluated on matrix $M$ begins with setting $M_0 = M$ and then repeating the following two steps until convergence: \begin{enumeratenumeric} \item Find the QR decomposition $M_n = Q_n R_n$. \item Set $\text{} M_{n + 1} = R_n Q_n$. \end{enumeratenumeric} The thing to note is that $M_{n + 1} = Q_n^T M_n Q_n$. We only concern ourselves with PSD matrices\footnote{ A PSD (positive semi-definite) matrix is a symmetric $\mathbb{R}$-matrix whose eigenvalues are all non-negative. Equivalently, it is a matrix $M$ for which $v^T M v \geq 0$ for all vectors $v$.}, so the fixed points of the iteration above are all diagonal matrices with non-negative real entries. We've also investigated a variant of the LR algorithm evaluated on a PSD matrix $M$ that begins with setting $M_0 = M$ and then repeating: \begin{enumeratenumeric} \item Find the Cholesky decomposition $M_n = L_n L_n^T$. \item Set $M_{n + 1} = L_n^T L_n$. \end{enumeratenumeric} Observe that all $M_n$ are similar to each other because $M_{n + 1} = L_n^{- 1} M_n L_n$. Each $M_n$ is also PSD. Therefore by the spectral theorem, the stronger fact follows that all $M_n$ are orthogonally similar to each other. Note that while there are more modern and sophisticated versions of the QR algorithm {\cite{GoluVanl96}}, the original one is still widely taught, and remains simpler than its more recent improvements. We think there are conceptual advantages to better understanding the original one. \section{Account of visualisation} In this paper, we present a visualisation of the QR algorithm. Our visualisation makes many properties of the naive QR algorithm, and its various improvements, intuitive. Our visualisation may be beneficial to students and researchers. Our visualisation is also qualitative, and using it we can show that many features of the QR algorithm are in fact \tmtextit{universal} to iterative eigenvalue algorithms, and not just specific to the QR algorithm. The QR algorithm is an iterative eigenvalue algorithm. By iterative, we mean that it employs a function $f$ from a set to itself, and evaluates it repeatedly on some starting value $x$ to produce the sequence: $x, f (x), f (f (x)), f (f (f (x))) \ldots$ until it converges close enough to a fixed point of $f$. It's easy to verify that for the QR algorithm the fixed points are matrices whose eigenvalues are easily found. The issue is that in general, function iterations have very complicated dynamics. For some functions $f$ and starting values $x$, the iteration can diverge, or even be chaotic. When working over $\mathbb{R}$, the usual visualisation that's employed is sometimes called the cobweb plot, and using it one can develop a good intution for the dynamics of some instances of function iteration. We cannot use cobweb plots to understand the dynamics of iterative \tmtextit{eigenvalue} algorithms however, because even in the smallest non-trivial case, the $2 \times 2$ case, the function $f$ maps $\mathbb{R}^4 \rightarrow \mathbb{R}^4$. We focus only on PSD matrices (positive semi-definite). It can be argued that this special case is sufficient for computing SVDs (Singular Value Decompositions), eigendecompositions of \tmtextit{symmetric matrices}, and eigendecompositions of orthogonal matrices.\footnote{Briefly: For symmetric matrices, one can perform a shift $M' = M + \mu I$ for some easily chosen $\mu \in \mathbb{R}$ which makes $M'$ PSD. For an orthogonal matrix $M$, we can use the matrix logarithm (or Cayley Transform) as a first step towards making a matrix $M'$ which is PSD. SVD is reducible to eigendecomposition of symmetric matrices in various ways.} The PSD case is especially amenable to visualisation. We won't try to justify further our focus on the PSD case. The visualisation is based on the one-to-one correspondence between $n \times n$ PSD matrices, and ellipsoids centred at the origin of $\mathbb{R^n}$. The correspondence is given by $M \mapsto \{ M v \mid v \in S^n \}$ where $S^n$ is the origin-centred hypersphere in $\mathbb{R^n}$. Every PSD matrix corresponds to a unique ellipsoid, and every ellipsoid (that is origin-centred) corresponds to a unique PSD matrices. We therefore sometimes talk about ellipsoids and their semi-axes, instead of about matrices and their eigen-values/vectors. We go as far as to say ``the ellipsoid'' instead of ``the PSD matrix'' sometimes.\footnote{ This all follows from the spectral theorem.} \begin{table}[h] \begin{tabular}{lll} \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=2.94721566312475cm]{paper2-1.pdf}} & \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=2.94721566312475cm]{paper2-2.pdf}} & \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=2.94721566312475cm]{paper2-3.pdf}}\\ \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=2.94721566312475cm]{paper2-4.pdf}} & \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=2.94721566312475cm]{paper2-5.pdf}} & \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=2.94721566312475cm]{paper2-6.pdf}} \end{tabular}\tmcolor{blue}{\tmcolor{black}{{\begin{tabular}{l} \tmcolor{blue}{Input PSD matrix}\\ \tmcolor{red}{1 iteration QR}\\ {\color[HTML]{008000}1 iteration LR} \end{tabular}{}}}} \ \caption{} \end{table} Our \tmcolor{blue}{input ellipse is in blue}. Our \tmcolor{red}{output ellipse for 1 iteration is in red}. We also visualise \tmcolor{green}{{\color[HTML]{008000}the LR algorithm in green} }for comparison. We visualise the naive QR algorithm. We make some immediate observations: \begin{enumerate} \item The algorithm always converges for PSD matrices. \item In every iteration, the angle between the large semi-axes and the $x$-axis diminishes. \item The large semi-axis is in the same pair of quadrants for the output ellipse as for the input ellipse. \item The fixed points only occur when the semi-axes are aligned with the coordinate axes. For a given ellipse, this can happen two ways. \item If the large semi-axis is aligned with the $x$-axis, then this fixed point is stable. \item If the large semi-axis is aligned with the $y$-axis, then this fixed point is unstable. Being close to this fixed point is undesirable, because 1 iteration moves you \tmtextit{away} from it, into the direction of the other fixed point. The closer you are to the unstable fixed point, the more slowly you move away from it. \item The rotation of the ellipse gets slower the closer it is to being a circle. This presents unsolvable difficulties for finding \tmtextit{eigenvectors} in general as opposed to \tmtextit{eigenvalues}. \item Opposite to point 7, the rotation of the ellipse gets faster the closer it is to being a degenerate line segment. This can be harnessed to speed up the algorithm. \item If the ellipse $M$ gets scaled to $M' = \lambda M$, then the angle of rotation remains the same. \item The input ellipse is congruent (in the sense of Euclidean geometry and not of linear algebra) to the output ellipse. \end{enumerate} Observation 7 has profound origins. It is a consequence of eigenvector instability. When the geometric multiplicity of some eigenvalue $\lambda$ of a matrix $M$ is 2 or more, then an infinitesimal perturbation of $M$ can violently change the eigenspaces. Therefore, the slowness of rotation in the near-circular case is a manifestation of an \tmtextit{unsolvable difficulty} which is intrinsic to the eigendecomposition problem. Changing to a different algorithm won't make it go away. Further discussion of observation 7 requires us to know what a \tmtextit{nearly diagonal matrix} looks like in the ellipse model: \begin{enumerateroman} \item A \tmtextit{near-circle}. \item An ellipse whose semi-axes are nearly coordinate-axis aligned. \end{enumerateroman} These two scenarios (i and ii) have little overlap. We can find \tmtextit{eigenvalues} in both scenarios (because of the Gershgorin circle theorem), but \tmtextit{not eigenvectors}, for which we require scenario ii. The eigenvector problem is uncomputable in general. This uncomputability is perhaps not a ``big deal''. It often suffices to find the eigendecomposition of a matrix close by to the input matrix, even though this may \tmtextup{violently} change the eigenspaces. A discussion of why this can ever be acceptable is left to a footnote.\footnote{ To understand why, we recall the distinction between \tmtextit{forwards numerical stability} and \tmtextit{backwards numerical stability}, and the difference in applications of each. An algorithm $\widehat{f}$ for computing a function $f$ is forwards-stable if $\widehat{f} (x) \approx f (x)$ for all $x$. An algorithm $\widehat{f}$ for computing a function $f$ is backwards stable if given $x$ there is an $\widehat{x} \approx x$ such that $\widehat{f} (x) \approx f (\widehat{x})$. Clearly, forwards stability is desirable, but we've shown that for the eigendecomposition problem it is unattainable. Backwards stability remains attainable, but what's the point of it? We explain. Let $f: R {\rightarrow} R$ be some continuous function. Eigendecomposition enables us to extend $f$ to the set of PSD matrices by assuming that $f (P D P^{- 1}) = P f (D) P^{- 1}$ is true. We can only compute eigendecompositions in a backwards stable way, which means that we can make $M \approx P D P^{- 1}$ true but not $M = P D P^{- 1}$. This is fine though, because by the continuity of $f$ we have that $f (M) \approx P f (D) P^{- 1}$, so the approximation error from computing $f$ via the eigendecomposition of $M$ can be made negligible. Be careful though, because while $D$ may be close to the eigenvalue matrix of $M$ (because eigenspectra vary continuously) the same is \tmtextit{not} true for $P$, which may be far from any true eigenvector matrix of $M$.} The fix for observation 6 is to make the function being iterated discontinuous. This is obvious in 2 dimensions thanks to the visualisation, but it is also true in all dimensions. A more general version of this is the statement of theorem \ref{must-be-discontinuous}. A continuous choice of function $f$ might be nicer, but continuity can't hold everywhere.\footnote{For the sake of conceptual understanding, the function $f$ may be engineered to be continuous over a subspace of its domain over which it has a globally attractive fixed point, while the discontinuities outside of this subspace may serve to ``kick'' the input into the well-behaved subspace. The Wilkinson shift happens to be discontinuous, so it provides the needed discontinuities.} Observation 8 appears to be a (kind of) converse to observation 7, but it's difficult to verify under what conditions an iterative eigenvalue algorithm behaves this way. In the case of the QR and LR algorithms, this behaviour may be directly verified, but it would be interesting to demonstrate that it is true for a large class of other algorithms. \begin{table}[h] \begin{tabular}{lll} \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=3.1920831693559cm]{paper2-7.pdf}} & \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=3.1920831693559cm]{paper2-8.pdf}} & \raisebox{0.0\height}{\includegraphics[width=3.14836678473042cm,height=3.1920831693559cm]{paper2-9.pdf}} \end{tabular}{\begin{tabular}{l} \tmcolor{blue}{Input PSD matrix}\\ \tmcolor{red}{1 iteration QR} \end{tabular}{}} \caption{$3 \times 3$ PSD matrices. Observe that when pancaked, the ellipsoid aligns 1 semi-axis in 1 iteration.} \end{table} Observation 8 can be combined with observation 9 to motivate \tmtextit{shifting}, and to demonstrate the limits of shifting and how they may be bypassed through the use of \tmtextit{deflation}. Shifting is the map $M' = M + \mu I$ for some arbitrary $\mu \in \mathbb{R}$. The effect this has is to increase the lengths of all the semi-axes by $\mu$. To better understand this effect, we exploit observation 9 to scale down the ellipse so that its largest semi-axis has length exactly equal to $1$. We then observe that the effect of shifting is that it either fattens or thins the ellipse. The fattest case is a perfect circle, which by observation 7 is the worst-case scenario. The thinnest case is a line segment, for which convergence is instant. In higher dimensions, the extreme cases are \tmtextit{flat pancakes} or perfect hyperspheres. Therefore, shifting is pancaking! Pancakes are, of course, the best; and blobs are, of course, the worst. But once you've got a perfect pancake, shifting runs out of steam as a speeding-up technique, so we may employ \tmtextit{deflation} to destroy the dimension of the ellipsoid that has been reduced to $0$, so that we may pancake the remaining dimensions, and rekindle the speed-up. The above analysis of observation 8 shows that iterative eigenvalue algorithms in which observation 8 holds can be used as improvers of cruder but faster eigenvalue estimation technique. If an eigenvalue estimation technique can crudely estimate the \tmtextit{smallest} eigenvalue, then this can be used as the value of the shift $\mu$, which can be used then to pancake the ellipsoid effectively, and which in turn will accelerate the iterative eigenvalue algorithm, and which may \tmtextit{in turn} make the crude estimation technique more accurate in subsequent iterations, and so on. \section{Account of generalisation to $n$ dimensions} It may not be obvious that the conclusions generalise straightforwardly to PSD matrices in $n$ dimensions, but they do: \begin{itemizedot} \item The algorithm converges for all PSD matrices. \item The stable fixed points are precisely the diagonal matrices of the form $\tmop{diag} (\lambda_n, \lambda_{n - 1}, \ldots, \lambda_1)$ where all the $\lambda_i$'s are non-negative reals, and $\lambda_n \geq \lambda_{n - 1} \geq \ldots \geq \lambda_1$. The other diagonal matrices are unstable fixed points, with convergence slowing down the smaller the angle an ellipsoid makes with them. \item The output of each iteration is orthogonally similar (as a matrix) to its input. In terms of Euclidean geometry, it is a congruent ellipsoid. (Be aware that ``congruence'' in Euclidean geometry means something different from in linear algebra). \item At eigenvalue clashes, the eigenvectors are unstable, but the algorithm may still converge quickly to a nearly diagonal matrix. In which case, only the eigenvalues and \tmtextit{not the eigenvectors} may be obtained under such circumstances. \item Shifting can be understood as ``pancaking'' the ellipsoid (once the semi-axes are scaled so that the largest semi-axis has length $1$). Like in the 2D case, this speeds up convergence. \item Deflation can be thought of as continuing the ``pancaking'' (shifting) once one of the semi-axes has been fully pancaked. \end{itemizedot} These conclusions can be checked by doing the ellipsoid visualisation in 3 dimensions. But this isn't necessary, as the 2D case is suggestive enough. \section{Remark about the dual purpose of Wilkinson shifts, and the search for better shifting strategies} Wilkinson shifts are curiously enough not continuous. Wilkinson shifts appear to both ``kick'' the ellipsoid away from the unstable fixed point (if it's too close to it) and to ``pancake'' the ellipsoid. Both of these features speed up convergence. One might wonder though if these two properties: The kicking away from the unstable fixed point, and the pancaking of the ellipsoid, can be achieved separately. In particular, the shift $\mu$ can perhaps be made into a continuous function, with it serving only to ``pancake'' the ellipsoid. This might simplify the search for superior shifting strategies if we can narrow the search down to only continuous functions, with the discontinuity needed to do the ``kicking'' from the fixed point coming from elsewhere. \section{Axiomatic development of the theory of iterative eigenvalue algorithms} One advantage of the visualisation is that it is mostly qualitative. In particular, we see that the LR algorithm exhibits the same qualitative behaviour as the QR algorithm. We exploit this to show that much of the behaviour and theory of the QR algorithm is not caused by the QR decomposition as such, but is intrinsic to solving the eigenvalue estimation problem by the use of fixed-point iteration. Consider the following axioms for a fixed-point iteration algorithm: \begin{enumeratenumeric} \item It consists of iterating some function $f$ over PSD matrices. \item $f (M)$ is orthogonally similar to $M$. \item The sequence $(f^n (M))_{n \in \mathbb{N}}$ converges for every $M$. It furthermore converges to a fixed point of $f$. \item A fixed point can only be a diagonal matrix. \item All fixed points are attractive. (This makes convergence fast). \end{enumeratenumeric} We consider these 5 axioms to be highly desirable. Intriguingly, these rule out the possibility that $f$ can be continuous everywhere. This is the statement of theorem \ref{must-be-discontinuous}. Morally, the theorem holds because given a non-scalar matrix $M$, the topological space of matrices orthogonally similar to $M$ is connected, but has some ``holes'' in it. The influence these ``holes'' have is that they prevent axioms 1 to 5 being realised when $f$ is continuous, even though these 5 conditions are highly desirable. A work-around might be to make $f$ continuous over some neighbourhood of each fixed point, and use the discontinuity of $f$ outside these neighbours to make $f (x)$ only map into these neighbourhoods. \begin{theorem} \label{must-be-discontinuous}Assuming axioms 1 to 5, the function $f$ being iterated must have discontinuities in the set of PSD matrices. \end{theorem} \begin{proof} We restrict the domain and codomain of $f$ to the set of matrices orthogonally similar to some PSD matrix $M$, where $M$ is not a multiple of the identity matrix. The fact that we can do this follows from condition 2. {\tmstrong{We show that $f$ has at least two fixed points under continuity}}: Assume it only has one fixed point $x$. It must be attractive by condition 5. By condition 3, all points attract to $x$. This results in the space of matrices orthogonally similar to $M$ being a contractible space, which it isn't. We get a contradiction. Therefore $f$ has at least two fixed points orthogonally similar to $M$. {\tmstrong{Now we show that under continuity, one of these two fixed points is not attractive, which contradicts conditions 3 and 5}}. The space of matrices orthogonally similar to some arbitrary matrix $M$ is connected. The set of points which attract to some attractive fixed point is always open. The basins of attraction of the attractive fixed points are disjoint, so by topological connectivity there are points not in their union, which are the points which don't attract to an attractive fixed point. This contradicts conditions 3 and 5. \end{proof} \section{Future work regarding visualisations} We ask the following questions. Forgive us if some happen to be easy: \begin{itemize} \item Can a visualisation technique for iterative eigenvalue algorithms be developed for non-symmetric matrices? The dynamics of the QR algorithm in the case of non-symmetric matrices (especially featuring shift policies) is not as well understood as in the symmetric case {\cite{banks2021global}}. \item Can a visualisation technique help to better understand Hessenberg matrices and their uses for finding eigenvalues and eigenvectors? We haven't succeeded yet in producing anything useful for this. Even symmetric Hessenberg matrices would be interesting. \end{itemize} It's important that visualisations be legible. Whether a visualisation is legible or not is perhaps subjective. \section{Novelty} We first put a visualisation of the QR algorithm onto the Wikipedia page on the QR algorithm on August 2021. We did so under a pseudonym. Afterwards, we left it for some time. We then later presented some of the material in this paper in a Youtube video series {\cite{video}}, also under a pseudonym. Based on a private correspondence with Professor David Watkins {\cite{david-email}} (who has written numerous pedagogical papers on iterative eigenvalue algorithms like the QR algorithm {\cite{watkins-tutorial2,watkins-tutorial1}}) the presentation is likely a novel one. The emphasis is likely novel. Videos and lectures are an important accompaniment when trying to present material like the one here.

https://arxiv.org/abs/2010.15204

Shortest closed curve to inspect a sphere

We show that in Euclidean 3-space any closed curve which lies outside the unit sphere and contains the sphere within its convex hull has length at least $4\pi$. Equality holds only when the curve is composed of $4$ semicircles of length $\pi$, arranged in the shape of a baseball seam, as conjectured by V. A. Zalgaller in 1996.

\section{Introduction} What is the shortest closed orbit a satellite may take to inspect the entire surface of a round asteroid? This is a well-known optimization problem \cite{zalgaller:1996,orourkeMO,hiriart:2008,ghomi:lwr,cfg,finch&wetzel} in classical differential geometry and convexity theory, which may be precisely formulated as follows. A curve $\gamma$ in Euclidean space $\mathbf{R}^3$ \emph{inspects} a sphere $S$ provided that it lies outside $S$ and each point $p$ of $S$ can be ``seen" by some point $q$ of $\gamma$, i.e., the line segment $pq$ intersects $S$ only at $p$. It is easily shown that the latter condition holds if and only if $S$ lies in the convex hull of $\gamma$. The supremum of the radii of the spheres which are contained in the convex hull of $\gamma$ and are disjoint from $\gamma$ is called the \emph{inradius} of $\gamma$. Thus we seek the shortest closed curve with a given inradius. The answer is as follows: \begin{thm}\label{thm:main} Let $\gamma\colon[a,b]\to\mathbf{R}^3$ be a closed rectifiable curve of length $L$ and inradius $r$. Then \begin{equation}\label{eq:main} L\geq 4\pi r. \end{equation} Equality holds only if, up to a reparameterization, $\gamma$ is simple, $\mathcal{C}^{1,1}$, lies on a sphere of radius $\sqrt2 \,r$, and traces consecutively $4$ semicircles of length $\pi r$. \end{thm} \begin{figure}[h] \begin{overpic}[height=1.25in]{baseball.jpg} \end{overpic} \caption{}\label{fig:baseball} \end{figure} It follows that the image of the minimal curve is unique up to a rigid motion, and resembles the shape of a baseball seam as shown in Figure \ref{fig:baseball}, which settles a conjecture of Viktor Zalgaller made in 1996 \cite{zalgaller:1996}. The previous best estimate was $L\geq 6\sqrt{3}\,r$ obtained in 2018 \cite{ghomi:lwr}. Here we use some notions from \cite{ghomi:lwr} together with other techniques from integral geometry (Crofton type formulas), geometric knot theory (unfoldings of space curves), and geometric measure theory (tangent cones, sets of positive reach) to establish the above theorem. We also derive several formulas (Sections \ref{sec:integral}, \ref{sec:horizon}, and \ref{sec:geq-sqrt2}) for efficiency of curves, which may be verified with a software package that we have provided \cite{ghomi-wenk:Mathematica2}. Our main approach for proving Theorem \ref{thm:main} is as follows. Since \eqref{eq:main} is invariant under rescaling and rigid motions, we may assume that $r=1$ and $\gamma$ inspects the unit sphere $\S^2$, in which case we say simply that $\gamma$ is an \emph{inspection curve}. Then we define the \emph{horizon} of $\gamma$ (Section \ref{sec:integral}) as the measure in $\S^2$ counted with multiplicity of the set of points $p\in\S^2$ where the tangent plane $T_p\S^2$ intersects $\gamma$: $$ H(\gamma):=\int_{p\in\S^2} \#\gamma^{-1}(T_p\S^2)\, dp. $$ Since $\gamma$ is closed, $T_p\S^2$ intersects $\gamma$ at least twice for almost every $p\in\S^2$. Thus $H(\gamma)\geq 8\pi$. Next we define the \emph{(inspection) efficiency} of $\gamma$ as \begin{equation}\label{eq:E} E(\gamma):=\frac{H(\gamma)}{L(\gamma)}. \end{equation} So to establish \eqref{eq:main} it suffices to show that $E(\gamma)\leq 2$. Now note that, since $H$ is additive, for any partition of $\gamma$ into subsets $\gamma_i$, $i\in I$, \begin{equation}\label{eq:EHL} E(\gamma)= \sum_i \frac{H(\gamma_i)}{L(\gamma)}=\sum_i \frac{L(\gamma_i)}{L(\gamma)} E(\gamma_i)\leq \sup_i E(\gamma_i) . \end{equation} So the desired upper bound for $E(\gamma)$ may be established through a partitioning of $\gamma$ into subsets $\gamma_i$ with $E(\gamma_i)\leq 2$. To find the desired partition, we may start by assuming that $\gamma$ in Theorem \ref{thm:main} has minimal length among all (closed) inspection curves, and is parameterized with constant speed (Section \ref{sec:prelim}). Then we apply an ``unfolding" procedure \cite{cks2002} to transform $\gamma$ into a planar curve $\tilde\gamma$ with the same arclength and \emph{height}, i.e., radial distance function from the origin $o$ of $\mathbf{R}^3$ (Section \ref{sec:unfolding}). It follows that $E(\gamma)=E(\tilde\gamma)$. Furthermore, the minimality of $\gamma$ will ensure that $\tilde\gamma$ is ``locally convex with respect to o" \cite{ghomi2019}. Consequently $\tilde\gamma$ may be partitioned into a collection of curves $\tilde\gamma_i$ we call spirals (Section \ref{sec:decomposition}). A \emph{spiral} is a planar curve which lies outside the unit circle $\S^1$, is locally convex with respect to $o$, has monotone height, and is orthogonal to the position vector of its closest boundary point to $o$. We will show that the efficiency of any spiral is at most $2$ by polygonal approximations and a variational argument (Section \ref{sec:inequality}), which establish \eqref{eq:main}. The rest of the paper will be devoted to characterizing minimal inspection curves. First we note that equality holds in \eqref{eq:main} only when $E(\gamma)=2$, which forces all the spirals $\tilde\gamma_i$ to have efficiency $2$ as well. Then we show that a spiral has efficiency $2$ only when it has constant height $\sqrt2$ (Section \ref{sec:leq-sqrt2}) by refining the variational procedure used earlier to establish \eqref{eq:main}. Once we know that any minimal inspection curve $\gamma$ has constant height $\sqrt{2}$, we proceed to the final stages of the characterization (Section \ref{sec:proof}). The simplicity of $\gamma$ follows from a Crofton type formula of Blaschke-Santalo. Then a characterization of $\mathcal{C}^{1,1}$ submanifolds in terms of their tangent cones \cite{ghomi-howard2014} and reach ensures the regularity of $\gamma$. Finally we show that $\gamma$ is composed of $4$ semicircles by constructing a ``nested partition" of $\gamma$ \cite{ghomi:rosenberg}, which completes the proof of Theorem \ref{thm:main}. The last technique goes back to the proofs of the classical $4$-vertex theorem due to Kneser and Bose \cite{hkneser,bose}, which has been developed further by Umehara and Thorbergsson \cite{umehara2, thorbergsson&umehara}. The question we study in this work belongs to a circle of long standing optimization problems for the length of a curve in Euclidean space subject to various constraints on its convex hull, including bounds on volume, surface area, width, and inradius \cite{ghomi:lwr,zalgaller:1994,zalgaller:1996,zalgaller:2003,finch2019translation,Schoenberg:convexhull} \cite[A28, A30]{cfg}. Most of these problems remain open; see \cite{ghomi:lwr,tilli2010} for more background and references. We should also note that these problems may be posed both for closed and open curves. In the latter case, there are connections to the ``lost in a forest problem" of Bellman \cite{bellman:1956}, or its dual version, Moser's ``worm problem" \cite{brass-moser-pach2005, fassler-orponen2018, norwood-poole2003}, which are well-known in computational geometry. \begin{comment} \begin{note} This paper is a shortened version of the original draft of this work posted on the arXiv \cite{ghomi-wenk-arXiv2020}. Some proofs have been omitted and other calculations curtailed. We refer the reader to \cite{ghomi-wenk-arXiv2020} for a more exhaustive treatment. \end{note} \end{comment} \section{Preliminaries: Minimal Inspection Curves}\label{sec:prelim} The central objects of study in this work are rectifiable curves, which become Lipschitz mappings after reparameterization with constant speed. We begin by recording basic facts we need in this regard; see \cite{cesari1958}, \cite[Chap. 2]{bbi:book}, or \cite[chap. 4]{ambrosio-tilli2004} for more background. Here $\mathbf{R}^n$ is the $n$-dimensional Euclidean space with origin $o$, inner product $\langle\cdot,\cdot\rangle$, and norm $|\cdot|:=\langle\cdot,\cdot\rangle^\frac{1}{2}$; $\S^{n-1}$, $B^n$ denote respectively the unit sphere and closed unit ball in $\mathbf{R}^n$. The \emph{interior}, \emph{closure}, and \emph{boundary} of any set $X\subset\mathbf{R}^n$ is denoted by $\inte(X)$, $\overline X$, and $\partial X$ respectively. A \emph{curve} is a continuous map $\gamma\colon[a,b]\to\mathbf{R}^n$, where $[a,b]\subset\mathbf{R}$ is an interval with $a<b$. We will also use $\gamma$ to refer to its image, $\gamma([a,b])$. We say that $\gamma$ is \emph{closed} if $\gamma(a)=\gamma(b)$. A closed curve $\gamma$ is \emph{simple} if it is one-to-one on $[a,b)$, and is $\mathcal{C}^1$ provided that it is continuously differentiable with $\gamma'_+(a)=\gamma'_-(b)$; $\gamma$ is $\mathcal{C}^{1,1}$ if $\gamma'$ is Lipschitz. The \emph{length} of $\gamma$ is $ L(\gamma):=\sup\sum_{i=1}^n \big|\gamma(t_i)-\gamma(t_{i-1})\big|, $ where the supremum is taken over all partitions $a:=t_0\leq t\leq\dots\leq t_n:=b$ of $[a,b]$; $\gamma$ is \emph{rectifiable} if $L(\gamma)<\infty$, and has \emph{constant speed} $C$ if $ L(\gamma|_{[t,s]})=C\,|t-s| $ for all $t<s\in [a,b]$. A curve $\widehat\gamma\colon [a,b]\to\mathbf{R}^n$ is a \emph{reparameterization} of $\gamma$ if there is a nondecreasing continuous map $\phi\colon [a,b]\to[a,b]$ with $\gamma=\widehat\gamma\circ\phi$. It is well-known \cite[Prop. 2.5.9]{bbi:book} that any rectifiable curve admits a reparameterization with constant speed. If $\gamma\colon[a,b]\to\mathbf{R}^n$ has constant speed $C$, then for all $t<s\in[a,b]$, \begin{equation}\label{eq:const-speed} |\gamma(t)-\gamma(s)|\leq L\big(\gamma\big|_{[t,s]}\big)= C|t-s|. \end{equation} So $\gamma$ is \emph{$C$-Lipschitz}, and therefore differentiable almost everywhere by Rademacher's theorem. Then $ L(\gamma)=\int_a^b |\gamma'(t)|\,dt $ \cite[Thm. 2.7.6]{bbi:book}. Furthermore, \eqref{eq:const-speed} implies that $|\gamma'|\leq C$ at all differentiable points of $\gamma$. On the other hand, $\int_a^b |\gamma'(t)|\,dt/(b-a)=L(\gamma)/(b-a)=C$. Thus $|\gamma'|=C$ almost everywhere. So we record: \begin{lem}\label{lem:basic} Let $\gamma\colon[a,b]\to\mathbf{R}^n$ be a rectifiable curve. Then $\gamma$ has constant speed if and only if $|\gamma'|=L(\gamma)/(b-a)$ almost everywhere. \end{lem} \noindent In particular, when $L(\gamma)>0$, we may assume after reparameterization with constant speed that $|\gamma'|\neq 0$ almost everywhere. Let $\mathcal{C}^0([a,b],\mathbf{R}^n)$ denote the space of curves $\gamma\colon[a,b]\to\mathbf{R}^n$ with the supremum norm or \emph{uniform metric} \cite[p. 47]{bbi:book} given by \begin{equation}\label{eq:sup-norm} \dist(\gamma_1,\gamma_2):=\sup_{t\in[a,b]}|\gamma_1(t)-\gamma_2(t)|. \end{equation} The functional $L\colon \mathcal{C}^0([a,b],\mathbf{R}^n)\to \mathbf{R}$ is lower semi-continuous \cite[Prop. 2.3.4(iv)]{bbi:book}. The \emph{convex hull} of a set $X\subset\mathbf{R}^n$, $\conv(X)$, is the intersection of all closed half-spaces containing $X$. We say $\gamma\colon[a,b]\to\mathbf{R}^3$ is an \emph{inspection curve} if it is closed and $\S^2\subset\conv(\gamma)$. Lemmas \ref{lem:basic} together with Arzela-Ascoli theorem \cite[Thm. 2.5.14]{bbi:book} and semicontinuity of $L$ yield \cite[Prop. 2.3]{ghomi-wenk-arXiv2020}: \begin{prop}\label{prop:minimizer} There exists an inspection curve of minimum length. \end{prop} Any curve given by the above proposition will be called a \emph{minimal} inspection curve. Let $\angle(v,w):=\cos^{-1}(\langle v, w\rangle/(|v||w|))$ denote the \emph{angle} between $v$, $w\in\mathbf{R}^n\setminus\{o\}$. For any rectifiable curve $\gamma\colon[a,b]\to\mathbf{R}^n\setminus\{o\}$, with $L(\gamma)>0$, we set $$ \alpha(t):=\angle \big(\gamma(t),\gamma'(t)\big). $$ By Lemma \ref{lem:basic}, if $\gamma$ has constant speed, then $|\gamma'|\neq 0$ almost everywhere. So $\alpha$ is well-defined for almost every $t\in[a,b]$. The \emph{tangent cone} $T_t\gamma$ of $\gamma$ at $t\in[a,b]$ is the collection of all rays emanating from $\gamma(t)$ which are limits of a sequence of secant lines emanating from $\gamma(t)$ and passing through points $\gamma(s_i)$ as $s_i$ converge to $t$. If $\gamma$ is closed and $t=a$ or $b$, then we set $T_t\gamma:=T_a\gamma\cup T_b\gamma$. See \cite[Sec. 2]{ghomi-howard2014} for basic facts on tangent cones. If $T_t\gamma$ is a line, then we call it the \emph{tangent line} of $\gamma$ at $t$. When $\gamma$ is differentiable at $t$ and $|\gamma'(t)|\neq 0$, $T_{t}\gamma$ is the line through $\gamma(t)$ spanned by $\gamma'(t)$. The following lemma generalizes an earlier observation \cite[Lem. 7.4]{ghomi:lwr} for polygonal curves. \begin{lem}\label{lem:alpha} Let $\gamma\colon [a,b]\to \mathbf{R}^3$ be a constant speed minimal inspection curve. Then tangent lines of $\gamma$ avoid $\inte(B^3)$. In particular, for almost every $t\in[a,b]$, \begin{equation}\label{eq:sine-alpha} \alpha(t)\geq\sin^{-1}\big( 1/|\gamma(t)|\big). \end{equation} \end{lem} \begin{proof} Let $T$ be a tangent line of $\gamma$ at $t\in[a,b]$. Suppose that $T$ intersects $\inte(B^3)$. Set $X:=\conv(\{\gamma(t)\}\cup B^3)$. Then $T$ intersects $\inte(X)$. So there is a open interval $U\subset [a,b]$ of the form $(t, s)$ or $(s,t)$ such that $\gamma(U)\subset \inte(X)$, and $\gamma(s)$ lies on $\partial X$. Let $\overline U$ be the closure of $U$. Replacing $\gamma(\overline U)$ with a line segment connecting $\gamma(t)$ and $\gamma(s)$ yields a closed curve $\beta$ with $L(\beta)<L(\gamma)$. But $\conv(\beta)=\conv(\gamma)$, since $\gamma(U)\subset\inte(\conv(X))\subset\inte(\conv(\gamma))$. In particular $\S^2\subset\conv(\beta)$. So $\beta$ is an inspection curve shorter than $\gamma$, which is a contradiction. Now \eqref{eq:sine-alpha} follows from basic trigonometry. \end{proof} \section{The Integral Formula for Efficiency}\label{sec:integral} As mentioned in the introduction, the \emph{efficiency} of any rectifiable curve $\gamma\colon[a,b]\to\mathbf{R}^3$ with $|\gamma|\geq 1$ is defined as $$ E(\gamma):=\frac{H(\gamma)}{L(\gamma)},\quad\quad\text{where}\quad\quad H(\gamma):=\int_{p\in\S^2} \#\gamma^{-1}(T_p\S^2)\, dp. $$ Recall that $H(\gamma)$ is called the \emph{horizon} of $\gamma$. When $\gamma$ is an inspection curve it follows from Caratheodory's convex hull theorem that $ \#\gamma^{-1}(T_p\S^2)\geq 2 $ for almost every $p\in \S^2$ \cite[Lemma 7.1]{ghomi:lwr}. Thus $H(\gamma)\geq 8\pi$. So to prove \eqref{eq:main} it suffices to show that $E(\gamma)\leq 2$. To this end we will use the area formula to compute $E(\gamma)$. This generalizes previous work \cite[Sec. 7.2]{ghomi:lwr} where the following proposition had been established for $\mathcal{C}^{1,1}$ curves. Recall that $\alpha:=\angle(\gamma,\gamma')$. \begin{prop}\label{prop:H} Let $\gamma\colon[a,b]\to\mathbf{R}^3$ be a constant speed curve with $|\gamma|\geq 1$. Then \begin{equation}\label{eq:H2} E(\gamma)= \frac{1}{b-a}\int_a^b\int_0^{2\pi} \frac{1}{|\gamma|^2}\left| \sqrt{|\gamma|^2-1}\sin\big(\alpha\big)\cos(\theta)+\cos\big(\alpha\big) \right|\,d\theta dt. \end{equation} \end{prop} \begin{proof} Let $\overline\gamma:={\gamma}/{|\gamma|}$. Since $\gamma$ is Lipschitz, $\overline\gamma$ is Lipschitz as well. So there exists a point $x\in \S^2\setminus \overline\gamma$. Let $e_1$ be a $\mathcal{C}^1$ unit tangent vector field on $\S^2\setminus\{x\}$, and $e_2(p):=p\times u(p)$. Then $(e_1, e_2)$ is a Lipschitz frame on any compact subset of $\S^2\setminus\{x\}$. So if we set $ e_1(t):=e_1(\overline\gamma(t)), \; e_2(t):=e_2(\overline\gamma(t)), $ then $t\mapsto (\overline\gamma(t), e_1(t), e_2(t))$ is a Lipschitz frame along $\gamma$. Set $ \lambda:=1/|\gamma|,\; \rho:=\sqrt{1-\lambda^2}, $ and define $F\colon[a,b]\times[0,2\pi]\to\S^2$ by $$ F(t,\theta)=\lambda(t)\overline\gamma(t)+\rho(t)\big( \cos(\theta) e_1(t)+\sin(\theta)e_2(t)\big). $$ Then $\theta\mapsto F(t,\theta)$ parameterizes the \emph{horizon circle} $H(\gamma(t))$, i.e., the set of points in $\S^2$ generated by the rays which emanate from $\gamma(t)$ and are tangent to $\S^2$. So, for all $p\in\S^2$, $ F^{-1}(p)=\gamma^{-1}(T_p \S^2). $ Thus, since $F$ is Lipschitz, the area formula \cite[Thm 3.2.3]{federer:book} yields $$ H(\gamma)=\int_{p\in\S^2}\#F^{-1}(p)\,dp=\int_a^b\int_0^{2\pi}JF(t, \theta)\,d\theta dt, $$ where $JF:=|\partial F/\partial t\times \partial F/\partial\theta|$ is the Jacobian of $F$. Next, for every differentiable point $t\in[a,b]$ of $\gamma$ let $ E_\gamma(t):= \int_0^{2\pi}JF(t,\theta)\,d\theta. $ By the Lebesgue differentiation theorem, for almost every $t\in[a,b]$, $ E_\gamma(t)=\lim_{\epsilon\to 0}\frac{1}{2\epsilon}\int_{t-\epsilon}^{t+\epsilon}E_\gamma(s)\,ds= \lim_{\epsilon\to 0}\frac{1}{2\epsilon}H(\gamma|_{[t-\epsilon,t+\epsilon]}). $ So $E_\gamma(t)$ does not depend on the choice of the frame $(e_1,e_2)$. We claim that for almost every point $t_0\in[a,b]$ of $\gamma$ we may choose a frame so that \begin{equation}\label{eq:JF} JF(t_0,\theta)= \frac{1}{|\gamma(t_0)|^2}\left| \sqrt{|\gamma(t_0)|^2-1}\sin\big(\alpha(t_0)\big)\cos(\theta)+\cos\big(\alpha(t_0)\big) \right|\,|\gamma'(t_0)|. \end{equation} This would complete the proof because $E(\gamma)=(\int_a^b E_\gamma(t)\,dt)/L(\gamma)$, and since the speed is constant $|\gamma'|=L(\gamma)/(b-a)$. To establish \eqref{eq:JF} note that if $t_0$ is a differentiable point of $\gamma$, then it is a differentiable point of $\overline\gamma$ as well. There are two cases to consider: either $\overline\gamma'(t_0)\neq 0$ or $\overline\gamma'(t_0)= 0$. First suppose that $\overline\gamma'(t_0)\neq 0$. Let $C$ be the great circle in $\S^2$ which is tangent to $\overline\gamma$ at $\overline\gamma(t_0)$. Set $e_1(\overline\gamma(t_0)):=\overline\gamma'(t_0)/|\overline\gamma'(t_0)|$. We may extend $e_1$ smoothly to a unit tangent vector field in a neighborhood of $\overline\gamma(t_0)$ on $\S^2$ so that $e_1(p)$ is tangent to $C$ when $p\in C$. Let $v:=|\overline\gamma'(t_0)|$. Then $\overline\gamma'(t_0)=ve_1(t_0)$, $e_1'(t_0)=-v\overline\gamma(t_0)$, and $e_2'(t_0)=0$. Using these rules one may compute \cite{ghomi-wenk-arXiv2020,ghomi-wenk:Mathematica2} that at $t=t_0$, $ JF=|\rho v \cos(\theta)-\lambda'| $ which is equivalent to \eqref{eq:JF}. If $\overline\gamma'(t_0)=0$, then for any choice of frame, $e_1'(t_0)=e_2'(t_0)=0$ by the chain rule. Then a computation \cite{ghomi-wenk-arXiv2020,ghomi-wenk:Mathematica2} shows that $JF=|\gamma'|/|\gamma|^2$, which establishes \eqref{eq:JF} since $\alpha=0$ or $\pi$. \end{proof} If $\gamma\colon[a,b]\to\mathbf{R}^3$ has constant speed $C$, then $ \big|\, |\gamma(s)|-|\gamma(t)|\,\big|\leq |\gamma(t)-\gamma(s)|\leq C|t-s|. $ So the function $|\gamma|\colon[a,b]\to\mathbf{R}$, which we call the \emph{height} of $\gamma$, is Lipschitz. In particular, $|\gamma|$ is differentiable almost everywhere. Furthermore note that if $t$ is a differentiable point of both $\gamma$ and $|\gamma|$, then $|\gamma|'=\langle \gamma,\gamma'\rangle/|\gamma|$ at $t$. Thus for almost every $t\in[a,b]$ \begin{equation}\label{eq:alpha-gamma-prime0} \alpha(t)=\cos^{-1}(|\gamma|'(t)/C). \end{equation} This shows, via Proposition \ref{prop:H}, that $E(\gamma)$ depends only on $|\gamma|$. Hence we conclude \begin{cor}\label{cor:lengthandheight} Let $\gamma_1$, $\gamma_2\colon[a,b]\to\mathbf{R}^3$ be constant speed curves with $L(\gamma_1)=L(\gamma_2)$. Furthermore suppose that $|\gamma_1(t)|=|\gamma_2(t)|\geq 1$ for all $t\in[a,b]$. Then $E(\gamma_1)=E(\gamma_2)$. \end{cor} \section{Unfolding of Minimal Inspection Curves}\label{sec:unfolding} Here we describe a natural ``unfolding" procedure \cite{cks2002} which transforms a space curve into a planar one. This operation preserves the arclength and height of the curve, and thus preserves its efficiency due to the results of the last section. Furthermore we will show that the unfolding of any minimal inspection curve satisfies a certain convexity condition. Let $\gamma\colon[a,b]\to\mathbf{R}^3\setminus\{o\}$ be a rectifiable curve. We set $\overline\gamma:=\gamma/|\gamma|$, and let $ \theta_\gamma(t):=L\big(\overline\gamma\big|_{[a,t]}\big)=\int_a^t|\overline\gamma'(t)|dt $ denote the arclength function of $\overline\gamma$ ($\theta_\gamma$ measures the ``cone angle" \cite{cks2002} or ``vision angle" \cite{choe-gulliver1992, choe-gulliver2017} of $\gamma$ from the point of view of $o$). The \emph{(cone) unfolding} of $\gamma$ is the planar curve $\tilde\gamma\colon[a,b]\to\mathbf{R}^2$ given by $ \tilde\gamma(t):=|\gamma(t)|e^{i\theta_\gamma(t)}, $ where $e^{i\theta_\gamma}=(\cos(\theta_\gamma), \sin(\theta_\gamma))$. In other words, $\tilde\gamma$ is generated by the isometric immersion (or unrolling) into $\mathbf{R}^2$ of the conical surface generated by the line segments $o\gamma(t)$. Note that $|\tilde\gamma(t)|=|\gamma(t)|$. Assuming $\gamma$ is reparameterized with constant speed, \begin{equation}\label{eq:theta-prime} \tilde\gamma'=(|\gamma|'+i|\gamma|\theta_\gamma')e^{i\theta_\gamma},\quad\quad\text{and}\quad\quad \theta_\gamma'=|\overline\gamma'|=\frac{1}{|\gamma|^2}\sqrt{|\gamma|^2|\gamma'|^2-\langle\gamma,\gamma'\rangle^2}, \end{equation} almost everywhere. Thus it follows that, for almost all $t\in[a,b]$, \begin{eqnarray*} |\tilde\gamma'|^2= (|\gamma|' )^2+|\gamma|^2(\theta_\gamma')^2 =\frac{\langle\gamma,\gamma'\rangle^2}{|\gamma|^2}+\frac{1}{|\gamma|^2}\Big(|\gamma|^2|\gamma'|^2-\langle\gamma,\gamma'\rangle^2\Big)=|\gamma'|^2. \end{eqnarray*} So $\gamma$ and $\tilde\gamma$ have equal height and length. Hence, by Corollary \ref{cor:lengthandheight}, \begin{prop}\label{prop:unfolding} Let $\gamma\colon[a,b]\to\mathbf{R}^3$ be a rectifiable curve with $|\gamma|\neq 0$, and $\tilde\gamma$ be the unfolding of $\gamma$. Then $E(\gamma)=E(\tilde\gamma)$. \end{prop} Next we develop some geometric properties of $\tilde\gamma$. \begin{lem}\label{lem:1-1} Let $\gamma\colon[a,b]\to\mathbf{R}^3$ be a minimal inspection curve with constant speed. Then $\tilde\gamma$ is locally one to one. \end{lem} \begin{proof} It suffices to show that $\theta'_\gamma>0$ almost everywhere. The formula for $\theta'_\gamma$ in \eqref{eq:theta-prime}, via the Cauchy-Schwartz inequality, shows that $\theta'_\gamma\geq 0$, and $\theta'_\gamma=0$ only when $\gamma'$ vanishes, or else $\gamma$ and $\gamma'$ are parallel. But $\gamma'$ can vanish only on a set of measure zero, since $\gamma$ has constant speed. Furthermore if $\gamma$ and $\gamma'$ are parallel, then $\alpha=0$. But by Lemma \ref{lem:alpha}, $\alpha\neq 0$ almost everywhere, which completes the proof. \end{proof} A \emph{convex body} $K\subset\mathbf{R}^2$ is a compact convex set with interior points. A planar curve $\gamma\colon[a,b]\to\mathbf{R}^2$ is \emph{locally convex} if it is locally one-to-one and each $t\in [a,b]$ has an open neighborhood $U\subset[a,b]$ such that $\gamma(U)$ lies on the boundary of a convex body. A \emph{local supporting line} $\ell$ for $\gamma$ at $t$ is a line passing through $\gamma(t)$ with respect to which $\gamma(U)$ lies on one side. If $\ell$ does not pass through $o$ and $\gamma(U)$ lies on the side of $\ell$ which contains $o$, then $\ell$ lies \emph{above} $\gamma$. If $\gamma$ is locally convex and through each point of it there passes a local support line which lies above $\gamma$, then $\gamma$ is locally convex \emph{with respect to $o$}. \begin{prop}\label{prop:loc-convex} Let $\gamma\colon[a,b]\to\mathbf{R}^3$ be a minimal inspection curve with constant speed. Then $\tilde\gamma$ is locally convex with respect to $o$. \end{prop} \begin{proof} By Lemma \ref{lem:1-1}, every $t\in[a,b]$ has a neighborhood $U\subset[a,b]$ such that $\tilde\gamma$ is one-to-one on $U$. Assuming that $U$ is small, $\tilde\gamma(U)$ will be star-shaped with respect to $o$, i.e., for every $s\in U$ the line passing through $o$ and $\tilde\gamma(s)$ intersects $\tilde\gamma(U)$ only at $\tilde\gamma(s)$. Thus connecting the end points of $\tilde\gamma(\overline U)$ to $o$ yields a simple closed curve, say $\Gamma$. We call the segments which run between $o$ and end points of $\tilde\gamma(\overline U)$ the sides of $\Gamma$, and let $\theta$ denote the interior angle of $\Gamma$ at $o$. We may assume that $U$ is so small that $\theta\leq\pi$. Then we claim that the region $K$ bounded by $\Gamma$ is convex, which will complete the proof. To this end let $p_0$, $p_1\in \inte (K)$. There exists a curve $p\colon [0,1]\to \inte(K)$ with $p(0)=p_0$, and $p_1=p(1)$, since $\inte(K)$ is path connected by Jordan curve theorem. Let $\overline t\in [0,1]$ be the supremum of all points $t\in[0,1]$ such that the line segment $p(0)p(t)\subset\inte(K)$. If $\overline t=1$, for all pairs of points $p_0$, $p_1\in \inte(K)$, then the line segment $p(0)p(1)\subset\inte(K)$. So $\inte(K)$ is convex, which implies that $K$ is convex, and we are done. Suppose that $\inte(K)$ is not convex. Then $\overline t<1$ for some pair of points $p_0$, $p_1\in\inte(K)$. Note also that $\overline t>0$ since $p_0\in\inte(K)$. So an interior point $x$ of $p(0)p(\overline t)$ intersects $\partial K=\Gamma$, while $p(0)p(\overline t)\subset K$. Since $\theta\leq\pi$, $x$ cannot lie on a side of $\Gamma$, for then either $p(0)$ or $p(1)$ will be forced to lie on a side of $\Gamma$ as well, which is not possible as they are interior points of $K$. So $x$ must lie on $\tilde\gamma(U)$. Now we may slightly perturb the segment $p(0)p(\overline t)$ so that a point of it leaves $K$ while its end points remain in $\inte(K)$. Then we obtain a line segment $\sigma$ whose end points lie on $\tilde\gamma(U)$ while its interior lies outside $K$. Thus if we replace the segment of $\tilde \gamma(U)$ which lies between the end points of $\sigma$ with the line segment $\sigma$, we obtain a star-shaped curve $\tilde\beta$ with $L(\tilde\beta)<L(\tilde\gamma)$. Parameterize $\tilde\beta$ by letting $\tilde\beta(t)$ be the point where the ray generated by $\tilde\gamma(t)$ intersects $\tilde\beta$. Then $|\tilde\beta(t)|\geq|\tilde\gamma(t)|$. Now set $ \beta(t):=\frac{|\tilde \beta(t)|}{|\tilde\gamma(t)|}\gamma(t). $ Then $\tilde\beta$ is the unfolding of $\beta$. So $ L(\beta)=L(\tilde\beta)<L(\tilde\gamma)=L(\gamma). $ On the other hand, $o\in\conv(\beta)$; otherwise there exists $u\in\S^2$ such that $\langle\beta(t),u\rangle>0$ for all $t\in[a,b]$, which in turn yields that $ \langle\gamma(t),u\rangle>0, $ which is not possible since $o\in\conv(\gamma)$. So $\lambda\beta(t)\in \conv(\beta)$ for all $0\leq\lambda\leq 1$. In particular $\gamma\subset\conv(\beta)$. It follows that $ \conv(\beta)\supset \conv(\gamma)\supset \S^2. $ Thus $\beta$ is an inspection curve shorter than $\gamma$, which is the desired contradiction. \end{proof} \section{Spiral Decomposition of the Unfolding}\label{sec:decomposition} Using the local convexity property established in the last section, we will show here that the unfolding of a minimal inspection curve admits a partition into certain segments we call spirals. Note that a locally convex curve is rectifiable, and therefore may be reparameterized with constant speed $C$. Then $|\gamma'|=C$ at almost all differentiable points of $\gamma$ by Lemma \ref{lem:basic}. Furthermore, local convexity also ensures \cite[Lem. 5.1]{ghomi-wenk-arXiv2020}: \begin{lem}\label{lem:gamma-speed} Let $\gamma\colon[a,b]\to\mathbf{R}^2$ be a locally convex curve with constant speed $C$. Then one sided derivatives of $\gamma$ exist at all points. Furthermore, $|\gamma'_+(a)|=|\gamma'_-(b)|=|\gamma'_\pm(t)|=C$ for all $t\in(a,b)$. \end{lem} Let $\gamma\colon[a,b]\to\mathbf{R}^2\setminus\{o\}$ be a locally convex curve with constant speed. If $t\in(a,b)$ is a differentiable point of $\gamma$ then $\gamma'_+(t)=\gamma_-'(t)=\gamma'(t)$. Thus the above lemma shows that $|\gamma'(t)|\neq 0$ at differentiable points of $\gamma$, since $C=L(\gamma)/(b-a)>0$. In particular the angle $\alpha$ is well defined at differentiable points of $\gamma$. We count $a$, $b$ among differentiable points of $\gamma$, and set $\gamma'(a):=\gamma'_+(a)$, $\gamma'(b):=\gamma'_-(b)$. We say that $\gamma$ is a \emph{spiral} if (i) $\gamma$ is locally convex with respect to $o$, (ii) $|\gamma|$ is nondecreasing, (iii) $\alpha(a)=\pi/2$, and (iv) $|\gamma(a)|\geq 1$. Note that condition (ii), via \eqref{eq:alpha-gamma-prime0}, implies that \begin{equation}\label{eq:alpha-leq-pi/2} \alpha(t)\leq \pi/2 \end{equation} at differentiable points $t\in[a,b]$ of $\gamma$. We say that $\gamma$ is a \emph{strict} spiral if $|\gamma|$ is increasing. By a \emph{spiral decomposition} of a constant speed curve $\gamma\colon[a,b]\to\mathbf{R}^2$ we mean a collection $U_i$ of mutually disjoint open subsets of $[a,b]$ such that (i) $\gamma|_{\overline U_i}$ is a strict spiral, after switching the direction of $\gamma$ if necessary, and (ii) $|\gamma|'=0$ almost everywhere on $[a,b]\setminus \cup_i \overline U_i$. By a \emph{parameter shift} we mean replacing $t$ with $(t+x)\,\text{mod}\, (b-a)$ for some $x\in[a,b]$. The main result of this section is: \begin{prop}\label{prop:decomposition} Let $\gamma\colon[a,b]\to\mathbf{R}^3$ be a minimal inspection curve. Then the unfolding of $\gamma$ admits a spiral decomposition, after a parameter shift. \end{prop} We may assume that $\gamma$ has constant speed. Let $\tilde\gamma$ be the unfolding of $\gamma$ and $x\in (a,b)$ be a local minimum point of the height function $|\gamma|=|\tilde\gamma|$. Then $\tilde\gamma$ is locally supported from below by a circle of radius $|\tilde\gamma(x)|$ centered at $o$. Thus, since $\tilde\gamma$ is locally convex with respect to $o$, there can pass only one local support line of $\tilde\gamma$ through $\tilde\gamma(x)$. Consequently $\tilde\gamma$ is differentiable at $x$ \cite[Thm. 1.5.15]{schneider2014}. Furthermore, the local support line at $\tilde\gamma(x)$ must be orthogonal to $\tilde\gamma(x)$, since $x$ is a local minimum of $|\tilde\gamma|$. So $\langle \tilde\gamma'(x),\tilde\gamma(x)\rangle =0$. Now if we shift the parameter of $\tilde\gamma$ by $x$, it follows that $\tilde\gamma$ is orthogonal to $\tilde\gamma(a)$ and $\tilde\gamma(b)$. Thus to prove Proposition \ref{prop:decomposition} it suffices to show: \begin{lem}\label{lem:decomposition} Let $\gamma\colon[a,b]\to\mathbf{R}^2$ be a constant speed curve which is locally convex with respect to $o$. Suppose that $\alpha(a)=\pi/2=\alpha(b)$, and $|\gamma|\geq 1$. Then $\gamma$ admits a spiral decomposition. \end{lem} \noindent To prove this lemma first recall that, as we had mentioned at the end of Section \ref{sec:integral}, the height function $|\gamma|$ of a constant speed curve is Lipschitz. In particular $|\gamma|$ is absolutely continuous and so it satisfies the fundamental theorem of calculus: \begin{equation}\label{eq:FTM} |\gamma(s)|-|\gamma(t)|=\int_t^s|\gamma|'(t)dt \end{equation} for every pair of points $s<t\in[a,b]$. Furthermore, let us reiterate that by \eqref{eq:alpha-gamma-prime0} if the speed of $\gamma$ is equal to $C$ then for almost every $t\in[a,b]$, $\alpha(t)=\cos^{-1}(|\gamma|'(t)/C)$. \begin{proof}[Proof of Lemma \ref{lem:decomposition}] Let $X$ be the set of points $t\in[a,b]$ such that $\gamma$ has a local support line at $\gamma(t)$ which is orthogonal to $\gamma(t)$. Then it follows from \eqref{eq:alpha-gamma-prime0} that $|\gamma|'=0$ almost everywhere on $X$. Also note that $X$ is closed, since the limit of any sequence of support lines of a convex body is a support line. Consequently each (connected) component $U$ of $[a,b]\setminus X$ is an open subinterval of $[a,b]$. We claim that $\gamma|_{\overline U}$ is a strict spiral, possibly after switching the direction of $\gamma|_{\overline U}$, which will complete the proof. To establish the above claim first note that by \eqref{eq:alpha-gamma-prime0}, $|\gamma|'$ cannot vanish at any differentiable point of $|\gamma|$ on $U$, for any such point would belong to $X$. We will show that either $|\gamma|'>0$ almost everywhere on $U$ or else $|\gamma|'<0$ almost everywhere on $U$. To this end we start by orienting each local support line $\ell$ of $\gamma$ consistent with the orientation of $\gamma$ at the point of contact with $\ell$, so that the \emph{angle} between any support line $\ell$ and the position vector of its point of contact will be consistently defined along $\gamma$. Now suppose, towards a contradiction, that there are subsets $X$ and $Y$ of $U$ with nonzero measure such that $|\gamma|'>0$ on $X$ and $|\gamma|'<0$ on $Y$. Since $X\cup Y$ is dense in $U$, there exists a point $r\in U$ which is a limit both of $X$ and $Y$. More specifically, there are sequences of differentiable points $t_i$, $s_i$ converging to $r$ such that $|\gamma|'(t_i)>0$ and $|\gamma|'(s_i)<0$, which in turn implies that $\alpha(t_i)<\pi/2$ and $\alpha(s_i)>\pi/2$ by \eqref{eq:alpha-gamma-prime0}. Hence, since the limit of support lines to a convex body is a support line, there exists a local support line through $r$ which makes an angle $\leq\pi/2$ with $r$, and there also exists a support line through $r$ which makes an angle $\geq\pi/2$ with $r$. So we conclude that there exists a local support line orthogonal to $r$, which is not possible by definition of $U$. Hence $|\gamma|'$ is always positive or always negative at differentiable points of $|\gamma|$ in $U$ as claimed. Now it follows from \eqref{eq:FTM} that $|\gamma|$ is strictly monotone on $\overline U$. Next, let $t_0$ be the boundary point of $\overline U$ which forms the minimum point of $|\gamma|$ on $\overline U$. We have to show that $\gamma|_{\overline U}$ is orthogonal to $\gamma(t_0)$. If $t_0=a$, $b$ this already holds by assumption. So suppose that $t_0\in(a,b)$. Then $t_0\in X$, and so $\gamma$ has a local support line $\ell$ at $\gamma(t_0)$ which is orthogonal to $\gamma(t_0)$. Since $\gamma$ is locally convex with respect to $o$, locally $\gamma$ lies below $\ell$. On the other hand, since $t_0$ is the minimum point of $|\gamma|$ on ${\overline U}$, then, near $\gamma(t_0)$, $\gamma|_{\overline U}$ lies above the circle $S$ with radius $|\gamma(t_0)|$ centered at $o$. Thus $\gamma|_{\overline U}$ must be orthogonal to $\gamma(t_0)$ as desired. We conclude then that $\gamma|_{\overline U}$ is a spiral, after switching the direction of $\gamma|_{\overline U}$ if necessary, so that $\gamma(t_0)$ becomes its initial point. \end{proof} \section{Efficiency of Line Segments}\label{sec:horizon} Here we derive some formulas for the horizon and therefore efficiency of line segments, which may be checked using \cite{ghomi-wenk:Mathematica2}. Suppose that we have a line segment $p_0p_1$ (with $p_0\neq p_1$) such that the line generated by $p_0p_1$ avoids $\inte(B^3)$, see Figure \ref{fig:circles}. \begin{figure}[h] \begin{overpic}[height=1.5in]{circles.pdf} \put(6,35.5){\small$p_0$} \put(36,43){\small$p_1$} \put(53,7){\small$H(p_0)$} \put(91,7){\small$H(p_1)$} \put(26,15){\small$q$} \put(21,24){\small$q'$} \put(70,6){\small$q$} \put(68,30){\small$q'$} \end{overpic} \caption{}\label{fig:circles} \end{figure} For each point $p$ on $p_0p_1$, let $C_p$ be the (inspection) cone generated by all rays which emanate from $p$ and pass through a point of $B^3$. Let $H(p)$ be the set of points where $\partial C_p$ touches $\S^2$, i.e., the \emph{horizon circle} from the point of view of $p$. Then $H(p_0p_1)$ is the area of the union of all horizon circles $H(p)$. Let $C_i:=C_{p_i}$, $H_i:=H(p_i)$, and $\{q,q'\}:=H_0\cap H_1$; it is possible that $q=q'$ which happens precisely when the line through $p_0$ and $p_1$ is tangent to $\S^2$. Note that all horizon circles $H(p)$ pass through $q$ and $q'$, because the triangles $p_0p_1q$ and $p_0p_1q'$ lie on planes which are tangent to $\S^2$. Thus $H(p_0p_1)$ consists of the two lunar regions determined by $H_0$ and $H_1$, if $q\neq q'$; otherwise, $H(p_0p_1)$ is the region lying inside one of the circles and outside the other. More precisely, if $D_i$ denote the (inspection) disks in $\S^2$ bounded by $H_i$, which lie inside $C_i$, then \begin{equation}\label{eq:Hp0p1} H(p_0p_1)=A(D_0)+ A(D_1)- 2 A(D_0\cap D_1), \end{equation} where $A$ stands for area. We may use basic spherical trigonometry to compute $H(p_0p_1)$ as follows. To start, note that if we set $ h_i:=|p_i|, \;\text{and}\; \ell:=|p_0p_1| $ then the radii of $D_i$ and the distance in $\S^2$ between the centers of $D_i$ are given respectively by $ \rho_i:=\cos^{-1}(1/h_i), $ and $ d:=\cos^{-1}((h_0^2+h_1^2-\ell^2)/(2h_0h_1)). $ It is a basic fact that $ A(D_i)=4\pi\sin^2(\rho_i/2). $ The formula for $A(D_0\cap D_1)$ is also known in terms of $\rho_i$ and $d$ \cite{tovchigrechko-vakser2001,ghomi-wenk-arXiv2020}. Substituting these formulas in \eqref{eq:Hp0p1} yields the following formula for $H(p_0p_1)$: \begin{small} \begin{multline*} H(h_0,h_1,\ell) = 4 \Big(\frac{1}{h_0}\sin ^{-1}\left(\frac{h_1^2-h_0^2-\ell^2}{\sqrt{\left(h_0^2-1\right) \left((h_0+h_1)^2-\ell^2\right) \left(\ell^2-(h_1-h_0)^2\right)}}\right) +\\ \frac{1}{h_1}\sin ^{-1}\left(\frac{h_0^2-h_1^2-\ell^2}{\sqrt{\left(h_1^2-1\right) \left((h_0+h_1)^2-\ell^2\right) \left(\ell^2-(h_1-h_0)^2\right)}}\right) +\cos ^{-1}\left(\frac{h_0^2+h_1^2-\ell^2-2}{2 \sqrt{\left(h_0^2-1\right) \left(h_1^2-1\right)}}\right)\Big). \end{multline*} \end{small} Note that $ h_1=\sqrt{h_0^2+\ell^2+2 h_0\ell\cos(\alpha)},\;\text{where}\; \alpha:=\angle(p_0,p_0p_1). $ In particular, if $\alpha=\pi/2$ then we obtain a formula for the horizon of one-edge spirals: \begin{multline}\label{eq:h0-ell} H(h_0,\ell)= 4 \left(\cos ^{-1}\left(\frac{\sqrt{h_0^2-1}}{\sqrt{h_0^2+\ell^2-1}}\right) -\frac{1}{\sqrt{h_0^2+\ell^2}}\sin ^{-1}\left(\frac{\ell}{h_0 \sqrt{h_0^2+\ell^2-1}}\right)\right). \end{multline} Finally, if we set $\ell=\sqrt{h_1^2-h_0^2}$ in the last expression we obtain another formula for the horizon of one-edge spirals in terms of $h_0$ and $h_1$: \begin{equation*}\label{eq:h0-h1} H(h_0,h_1):=4 \left(\cos ^{-1}\left(\frac{\sqrt{h_0^2-1}}{\sqrt{h_1^2-1}}\right)-\frac{1}{h_1}\sin ^{-1}\left(\frac{\sqrt{h_1^2-h_0^2}}{h_0 \sqrt{h_1^2-1}}\right)\right). \end{equation*} The graph of the corresponding efficiency function $E(h_0,h_1):=H(h_0,h_1)/\sqrt{h_1^2-h_0^2}$, for $h_1\geq h_0\geq 1$ is shown in Figure \ref{fig:line-eff}. \begin{figure}[h] \begin{overpic}[height=1.5in]{line-eff.jpg} \end{overpic} \caption{}\label{fig:line-eff} \end{figure} Note that $E(h_0,h_1)\leq 2$, and equality holds only when $h_0=h_1=\sqrt 2$, or the spiral has constant height $\sqrt2$. Below we will prove that all spirals satisfy these properties. \section{Upper Bound for Efficiency of Spirals}\label{sec:inequality} Here we apply formulas of the last section to bound the efficiency of spirals via a variational argument applied to polygonal curves. First we need to show that spirals form a locally compact space on which the efficiency functional is continuous. To this end we start by extending the definition of a spiral as follows. We say that $\gamma\colon[a,b]\to\mathbf{R}^2$ is a \emph{(generalized) spiral} provided that either $\gamma$ is a spiral as defined in Section \ref{sec:decomposition}, or else $\gamma$ is a constant map with $|\gamma|\geq 1$. We also extend the definition of efficiency by setting \begin{equation}\label{eq:LP0} E(\gamma):=\frac{4\sqrt{|\gamma|^2-1}}{|\gamma|^2}, \quad\quad \text{when} \quad\quad L(\gamma)=0. \end{equation} So by Proposition \ref{prop:H}, when $L(\gamma)=0$, $E(\gamma)$ is the efficiency of a curve of constant height $|\gamma|$. Note that then $E(\gamma)\leq 2$, and $E(\gamma)=2$ only when $|\gamma|=\sqrt2$. The space of spirals $\gamma\colon[a,b]\to\mathbf{R}^2$, with the topology induced on it by the uniform metric \eqref{eq:sup-norm}, will be denoted by $\mathcal{S}([a,b])$. To show that $E$ is continuous on $\mathcal{S}([a,b])$ first we observe that: \begin{lem}\label{lem:alpha3} Let $\gamma\colon[a,b]\to\mathbf{R}^2$ be a constant speed spiral with $L(\gamma)\neq 0$. Then \begin{equation}\label{eq:lem-alpha} \alpha(t)\geq\sin^{-1}\big(|\gamma(a)|/|\gamma(t)| \big), \end{equation} at all differentiable points $t\in[a,b]$ of $\gamma$. \end{lem} \begin{proof} We may assume, for convenience, that the speed of $\gamma$ is $1$. Then taking the cosine of both sides of \eqref{eq:lem-alpha} and squaring yields \begin{equation}\label{eq:lem-alpha2} \langle\gamma,\gamma'\rangle^2\leq |\gamma|^2-r^2. \end{equation} Recall that by \eqref{eq:alpha-leq-pi/2}, $\alpha(t)\leq\pi/2$ which in turn yields that $\langle\gamma,\gamma'\rangle\geq 0$. Thus \eqref{eq:lem-alpha2} is equivalent to \eqref{eq:lem-alpha}. To establish \eqref{eq:lem-alpha2}, first assume that $\gamma$ is $\mathcal{C}^{1,1}$. Then the left hand side of \eqref{eq:lem-alpha2} is Lipschitz; therefore, it is differentiable almost everywhere and satisfies the fundamental theorem of calculus. Furthermore, since $\gamma$ is locally convex with respect to $o$, $\langle \gamma,\gamma''\rangle\leq 0$ almost everywhere. So, since $\langle\gamma'(a),\gamma(a)\rangle=0$, and $\langle\gamma,\gamma'\rangle\geq 0$, $$ \langle\gamma(t),\gamma'(t)\rangle^2 = 2\int_a^t\langle \gamma(s),\gamma'(s)\rangle \big(1+\langle \gamma(s),\gamma''(s)\rangle\big)ds \leq 2\int_a^t\langle \gamma(s),\gamma'(s)\rangle ds = |\gamma(t)|^2-r^2, $$ as desired. To establish the general case we consider the outer parallel curves $\gamma_\epsilon$ of $\gamma$ at distance $\epsilon>0$. These curves are given by setting $\gamma_\epsilon(a):=\gamma(a)+\epsilon \gamma(a)/|\gamma(a)|$, and requiring that $\gamma_\epsilon$ maintain constant distance $\epsilon$ from $\gamma$. Since $\gamma$ is locally convex, $\gamma_\epsilon$ is $\mathcal{C}^{1,1}$ \cite[Prop. 2.4.3]{hormander}. Furthermore, it is not difficult to see that $\gamma_\epsilon$ is a spiral. So $\gamma_\epsilon$ satisfies \eqref{eq:lem-alpha2}. Next note that for each differentiable point $\gamma(t)$ of $\gamma$ there exists a unique point $\gamma_\epsilon (t_\epsilon)$ of $\gamma_\epsilon$ which is closest to $\gamma(t)$. Then $\alpha(t)=\alpha_\epsilon(t_\epsilon)$ where $\alpha_\epsilon:=\angle(\gamma_\epsilon,\gamma'_\epsilon)$. Thus $ \alpha(t)=\alpha_\epsilon(t_\epsilon)\geq\sin^{-1}\left(\frac{r+\epsilon}{|\gamma_\epsilon(t_\epsilon)|}\right). $ Letting $\epsilon\to 0$ completes the proof. \end{proof} Since for a spiral $\gamma(t)$ with $L(\gamma)\neq 0$, $\alpha(t)\leq\pi/2$, the last lemma shows that $\alpha(t)\to\pi/2$ as $\gamma(t)\to \gamma(a)$. This observation, together with some basic convex analysis, yields: \begin{lem}\label{lem:continuous} The efficiency functional $E$ is continuous on the space of spirals $\mathcal{S}([a,b])$. \end{lem} \begin{proof} For convenience we may assume that $[a,b]=[0,1]$. Let $\gamma_k\colon[0,1]\to\mathbf{R}^2$ be a sequence of spirals converging to a spiral $\gamma\colon[0,1]\to\mathbf{R}^2$. We have to show that $E(\gamma_k)\to E(\gamma)$. To this end, we may assume that all spirals have constant speed. First suppose that $L(\gamma)=0$. If $L(\gamma_k)=0$ as well, then we are done by \eqref{eq:LP0}. So we may assume that $L(\gamma_k)>0$, by passing to a subsequence. Then, by Proposition \ref{prop:H} \begin{multline}\label{eq:HP^k} E(\gamma_k) =\int_0^{1}\int_0^{2\pi} \frac{1}{|\gamma_k(t)|^2}\left| \sqrt{|\gamma_k(t)|^2-1}\sin\big(\alpha_k(t)\big)\cos(\theta)+\cos\big(\alpha_k(t)\big) \right|d\theta dt. \end{multline} Note that $\gamma_k(a)\to\gamma(a)$ and $\gamma_k(t)\to\gamma(t)=\gamma(a)$. So $\gamma_k(t)\to\gamma_k(a)$. Consequently, by Lemma \ref{lem:alpha3}, $\alpha_k(t)\to \pi/2$. So, since the integrand in \eqref{eq:HP^k} is bounded, the dominated convergence theorem yields that $$ E(\gamma_k) \to\int_0^1\int_0^{2\pi}\frac{\sqrt{|\gamma(a)|^2-1}}{|\gamma(a)|^2}|\cos(\theta)|\,d\theta dt=4\frac{\sqrt{|\gamma(a)|^2-1}}{|\gamma(a)|^2}=E(\gamma), $$ as desired. Next suppose that $L(\gamma)>0$, then we may assume that $L(\gamma_k)>0$ as well. So, again \eqref{eq:HP^k} holds. By assumption $\gamma_k\to\gamma$ uniformly. Furthermore, since $\gamma$ and $\gamma_k$ are locally convex, it follows that $\gamma'_k\to\gamma'$ almost everywhere on $[0,1]$. This can be shown by representing $\gamma_k$, $\gamma$ locally as graphs of convex functions and applying well-known results on convergence of derivatives from classical convexity theory; e.g., see \cite[C(9), p. 20]{roberts-varberg}, \cite[Lem. 2]{tsuji1952}, or \cite{lackovic1982}. So $\alpha_k\to\alpha$ almost everywhere on $[0,1]$. Thus by the dominated convergence theorem \begin{multline*} E(\gamma_k) \to \int_0^{1}\int_0^{2\pi} \frac{1}{|\gamma(t)|^2}\left| \sqrt{|\gamma(t)|^2-1}\sin\big(\alpha(t)\big)\cos(\theta)+\cos\big(\alpha(t)\big) \right|d\theta dt=E(\gamma), \end{multline*} which completes the proof. \end{proof} A \emph{polygonal curve} $P$ is a collection of line segments determined by a sequence of points $p_0,\dots, p_m\in\mathbf{R}^2$ with $p_{i+1}\neq p_{i}$. We also allow $P$ to be a single point, and use the formal notation $P=(p_0,\dots, p_m)$ to specify a polygonal curve. The line segments $p_ip_{i+1}$ are called the \emph{edges} of $P$. Each polygonal curve $P$ admits a unique constant speed parameterization $\gamma_P\colon[0,1]\to P$, with $\gamma_p(0)=p_0$ which traces the edges of $P$. The \emph{distance} between a pair of polygonal curves $P^1$, $P^2$ is defined as $ \dist\big(\gamma_{P^1},\gamma_{P^2}\big), $ the metric given by \eqref{eq:sup-norm}. Let $\P^m$ denote the space of polygonal curves with at most $m$ edges in $\mathbf{R}^2$, endowed with the topology induced by $\dist$. Then $\P^m$ is locally compact. We say that $P\in\P^m$ is a \emph{polygonal spiral} provided that $\gamma_P$ is a spiral. Let $\mc{S}^m$ be the collection of polygonal spirals with at most $m$ edges. Lemma \ref{lem:alpha3} together with Blaschke's selection principle \cite[Thm. 7.3.8]{bbi:book} quickly yields \cite[Lem. 7.3]{ghomi-wenk-arXiv2020}: \begin{lem}\label{lem:spiral-compact} The space of polygonal spirals $\mc{S}^m$ is locally compact, for every $m\geq 0$. \end{lem} Next we observe that \begin{lem}\label{lem:inequality} The efficiency of any polygonal spiral is at most $2$. \end{lem} \begin{proof} Fix an integer $m\geq 0$, and number $R>1$. By Lemma \ref{lem:spiral-compact} there exists a polygonal spiral $P=(p_0,\dots, p_k)$ which maximizes $E$ among elements of $\mathcal{S}^m$ which lie in the ball of radius $R$ centered at $o$. We need to show that $E(P)\leq 2$. If $P$ is a singleton, this is guaranteed by \eqref{eq:LP0}. So we may assume that $k\geq1$. Then $r:=|p_0|<R$. Note that for $-\epsilon< t<\epsilon$ there exists a point $p_0^t$ such that $p_0^t$ is orthogonal to $p_0^tp_1$, and $|p_0^t|=(1+t)r$, assuming that $\epsilon$ sufficiently small. Indeed $p_0^t$ lies on an arc of the circle of radius $|p_1|/2$ which is centered at the midpoint of $op_1$; see Figure \ref{fig:perturb}. \begin{figure}[h] \begin{overpic}[height=1.1in]{perturb.pdf} \put(5,46){\small$p_0$} \put(92,46){\small$p_1$} \put(9,-1){\small$o$} \end{overpic} \caption{}\label{fig:perturb} \end{figure} Furthermore, choosing $\epsilon$ sufficiently small, we can ensure that $P^t:=(p_0^t,p_1,\dots, p_k)$ is locally convex. Thus $P^t$ will be a spiral provided that $|p_0^t|\geq 1$, which will be the case for small $\epsilon$ provided that $r>1$ or else $t\geq 0$. Let us assume first that $r>1$. Then $P^t$ will be a spiral in the ball $B_R(o)$ for $-\epsilon<t<\epsilon$. Let $L(t)$, $H(t)$, and $E(t)$ denote respectively the length, horizon, and efficiency of $P^t$. Then $ 0= E'(0)=(H'(0)L(0)-H(0)L'(0))/L(0)^2, $ which in turn yields $ H'(0)/L'(0)=H(0)/L(0)=E(P). $ To compute $L'(0)$ note that $$ L(t)=L(p_0^t p_1)+L\big((p_1,\dots, p_k)\big)=\sqrt{|p_1|^2-(r+t)^2}+L\big((p_1,\dots, p_k)\big). $$ So it follows that $ L'(0)=-r/\sqrt{|p_1|^2-r^2}. $ Next, to compute $H'(0)$, note that $ H(t)=H(p_0^t p_1)+H\big((p_1,\dots, p_k)\big). $ Furthermore, by \eqref{eq:h0-ell} we have $ H(p_0^t p_1)=H\big(r+t,L(p_0^t p_1)\big). $ Now a computation \cite{ghomi-wenk:Mathematica2} yields that \begin{equation}\label{eq:H'0} H'(0)=\frac{d}{dt} H\big(r+t,L(p_0^t p_1)\big)\Big|_{t=0}=-\frac{4}{r}\frac{\sqrt{r^2-1}}{\sqrt{|p_1|^2-r^2}}. \end{equation} So we conclude that $ E(P)=\frac{H'(0)}{L'(0)}=4\frac{\sqrt{r^2-1}}{r^2}\leq 2, $ as desired. It remains to show that our earlier assumption that $r>1$ was justified. Suppose then, towards a contradiction, that $r=1$. Then $E(t)$ and $H(t)$ will still be well defined for $t\geq 0$, and so will their right-hand derivatives at $0$. By \eqref{eq:H'0}, $H'_+(0)=0$. Thus $$ E'_+(0)=\frac{-H(0)L'(0)}{L(0)^2}=\frac{H(0)}{L(0)^2}\frac{1}{\sqrt{|p_1|^2-1^2}}>0. $$ So $E(t)> E(0)$, or $E(P^t)> E(P)$, for small $t>0$ which is the desired contradiction. \end{proof} Lemma \ref{lem:inequality} together with Lemma \ref{lem:continuous} yields the main result of this section via a polygonal approximation: \begin{prop}\label{prop:inequality} The efficiency of any spiral is at most $2$. \end{prop} Proposition \ref{prop:inequality} together with Proposition \ref{prop:decomposition} establishes \eqref{eq:main}. The rest of this work will be concerned with characterizing the case of equality in \eqref{eq:main}. \section{Instantaneous Efficiency}\label{sec:geq-sqrt2} Here we investigate another method for bounding the efficiency of spirals via a notion first used in the proof of Proposition \ref{prop:H}. Let $\gamma\colon[a,b]\to\mathbf{R}^3$ be a constant speed curve with $|\gamma|\geq 1$, and $t\in[a,b]$ be a differentiable point of $\gamma$ with $|\gamma'(t)|\neq 0$. We define the \emph{instantaneous efficiency} of $\gamma$ at $t$ as $$ E_\gamma(t):=\int_0^{2\pi} \big|F(|\gamma(t)|,\alpha(t),\theta)\big|d\theta, $$ where $$ F(h,\alpha,\theta):=\frac{1}{h^2}\left(\sqrt{h^2-1}\sin(\alpha)\cos(\theta)+\cos(\alpha)\right). $$ If $t$ is a differentiable point of $t\mapsto H(\gamma |_{[a,t]})$, then by \eqref{eq:H2} $ E_\gamma(t)=\frac d {dt} H\left(\gamma |_{[a,t]}\right). $ So $E_\gamma(t)$ is the rate of change of horizon along $\gamma$. Furthermore, by Proposition \ref{prop:H}, $ E(\gamma)=\frac{1}{b-a}\int_a^b E_\gamma(t)dt\leq \sup_{[a,b]} E_\gamma(t). $ Thus to find an upper bound for $E(\gamma)$ it suffices to bound $E_\gamma$. To this end we compute the above integral as follows. Let $$ \Omega:=\{(h,\alpha)\,|\, h\geq 1,\; \sin^{-1}\left(1/h\right)\leq\alpha\leq \pi/2\} $$ be the phase space of possible values for $(|\gamma(t)|,\alpha(t))$ at differentiable points of curves $\gamma$ which lie outside $\S^2$ and whose tangent lines avoid $\inte(B^3)$. For every $(h,\alpha)\in\Omega$ we set $ \theta_0=\theta_0(h,\alpha):=\cos ^{-1}(-\cot (\alpha )/\sqrt{h^2-1}). $ Since $\sin(\alpha)\geq 1/h$, $|\cot(\alpha)/\sqrt{h^2-1}|\leq 1$. So $\theta_0$ is well defined. Also note that $F(h,\alpha, \pm\theta_0)=0$. Now we may compute that \cite{ghomi-wenk:Mathematica2} \begin{gather*} \mathcal{E}(h,\alpha):=\int_0^{2\pi}|F(h,\alpha,\theta)|\,d\theta =\int_{-\theta_0}^{\theta_0} F(h,\alpha,\theta)\,d\theta-\int_{\theta_0}^{2\pi-\theta_0} F(h,\alpha,\theta) d\theta\\ =\frac{4}{h^2} \left(\sqrt{h^2 \sin ^2(\alpha )-1}+\cos (\alpha ) \sin ^{-1}\left(\frac{\cot (\alpha )}{\sqrt{h^2-1}}\right)\right). \end{gather*} Then $ E_\gamma(t)=\mathcal{E}\big(|\gamma(t)|,\alpha(t)\big). $ For any set $X\subset[a,b]$ with measure $\mu(X)\neq 0$ we define $ E(\gamma\big|_X):=\frac{1}{\mu(X)}\int_X E_\gamma(t)dt. $ So we may record that \begin{prop}\label{prop:IE} Let $\gamma\colon[a,b]\to\mathbf{R}^3$ be a constant speed curve with $|\gamma|\geq 1$. If tangent lines of $\gamma$ avoid $\inte(B^3)$, then for any set $X\subset[a,b]$ with nonzero measure $$ E(\gamma\big|_X)=\frac{1}{\mu(X)}\int_X\mathcal{E}\big(|\gamma(t)|,\alpha(t)\big) dt. $$ \end{prop} The values of $\mathcal{E}$ on the phase space $\Omega$, which range from $0$ to about $2.6$, are shown in Figure \ref{fig:inst-eff}. \begin{figure}[h] \begin{overpic}[height=1.5in]{inst-eff.jpg} \end{overpic} \caption{}\label{fig:inst-eff} \end{figure} Since $\mathcal{E}$ may exceed $2$, it is not possible to obtain the bound $E\leq 2$ for all spirals by bounding $\mathcal{E}$; however, spirals with initial height $\geq\sqrt2$ are special. Set $$ \mathcal{E}(h):=\mathcal{E}\left(h,\frac{\pi}{2}\right)=4\frac{\sqrt{h^2 -1}}{h^2}\leq 2. $$ Note that $\mathcal{E}(h)=2$ only if $h=\sqrt2$. \begin{lem} Let $\gamma\colon[a,b]\to\mathbf{R}^2$ be a spiral with initial height $r\geq \sqrt{2}$. Then the instantaneous efficiency $E_\gamma(t)\leq \mathcal{E}(r)\leq 2$ for all $t\in[a,b]$. \end{lem} \begin{proof} Recall that $E_\gamma(t)=\mathcal{E}(|\gamma(t)|,\alpha(t))$. By Lemma \ref{lem:alpha3}, $\sin^{-1}(\sqrt2/r)\leq\alpha\leq \pi/2$. Since $\mathcal{E}(h,\pi/2)=\mathcal{E}(h)$, it suffices to check that $\alpha\mapsto \mathcal{E}(h,\alpha)$ is nondecreasing for $h\geq\sqrt{2}/\sin(\alpha)$. So we compute that \cite{ghomi-wenk:Mathematica2} \begin{eqnarray*} \frac{\partial \mathcal{E}}{\partial \alpha}(h,\alpha) &=&\frac{4}{h^2}\left(\cot (\alpha ) \sqrt{h^2 \sin ^2(\alpha )-1}-\sin (\alpha ) \sin ^{-1}\left(\frac{\cot (\alpha )}{\sqrt{h^2-1}}\right)\right)\\ &\geq& \frac{4}{h^2}\left( \frac{\cos (\alpha )}{\sqrt{1-\cos ^2(\alpha )}}-\sin ^{-1}\left(\frac{\cos (\alpha )}{\sqrt{1+\cos ^2(\alpha )}}\right)\right). \end{eqnarray*} It remains to check that $ \frac{x}{\sqrt{1-x^2}}-\sin ^{-1}\left(\frac{x}{\sqrt{1+x^2}}\right)\geq 0, $ for $0\leq x\leq 1$. Indeed this expression vanishes for $x=0$, and its derivative $ 1/(1-x^2)^{3/2}-1/(x^2+1) $ is nonnegative on $0\leq x\leq 1$. \end{proof} \begin{cor} Let $\gamma\colon[a,b]\to\mathbf{R}^2$ be a constant speed spiral with initial height $r \geq\sqrt{2}$. Then $$ E(\gamma)\leq \frac{1}{b-a}\int_a^b\mathcal{E}(|\gamma(t)|)dt\leq \mathcal{E}(r)\leq 2. $$ Equality in the second inequality holds only when $|\gamma|\equiv r$, and equality in the third inequality holds only when $r=\sqrt2$. \end{cor} \section{Spirals with Maximum Efficiency}\label{sec:leq-sqrt2} Here we refine the variational method employed in Section \ref{sec:inequality} to show that the efficiency of any spiral assumes its maximum value only when it has constant height $\sqrt{2}$. We start by considering one edge spirals $P=(p_0,p_1)$. By a \emph{lifting} of $P$ we mean any polygonal curve $\tilde P=(\tilde p_0 , p_1)$ where $\tilde p_0=\lambda p_0$ for $\lambda> 1$. \begin{lem}\label{lem:lifting} Let $P=(p_0 , p_1)$ be a spiral. For any lifting $\tilde P$ of $P$, $ H(P)<H(\tilde P). $ \end{lem} \begin{proof} Suppose $\tilde P:=(\tilde p_0, p_1)$, see Figure \ref{fig:cones}. \begin{figure}[h] \centering \begin{overpic}[height=1.75in]{cones7.pdf} \put(11.5,44){\small$p_0$} \put(30.5,43.5){\small$p_1$} \put(13,51.5){\small$\tilde p_0$} \put(96,17){\small$H(p_1)$} \put(45,17){\small$H(\tilde p_0)$} \put(73,25){\small$H(p_0)$} \end{overpic} \caption{}\label{fig:cones} \end{figure} Since $p_0p_1$ is orthogonal to $p_0$, the horizon circle $H(p_1)$ (depicted in orange) bisects the horizon circle $H(p_0)$ (depicted in dotted blue line), because the two planes which contain $p_0p_1$ and are tangent to $\S^2$ intersect $H(p_0)$ at a pair of its antipodal points. Thus the area that is gained by the horizon, as $p_0$ rises to $\tilde p_0$ exceeds the area which is lost. \end{proof} Let $P=(p_0,p_1)$ be a spiral and set $r:=|p_0|$, $R:=|p_1|$. For every $h\in[r,R]$, let $ \tilde P^h:=(\tilde p_0^{\,h},p_1) $ be the lifting of $P$ such that the distance of $\tilde p_0^{\,h}p_1$ to $o$ is equal to $h$. Let $q^h$ be the closest point of $\tilde p_0^{\,h}p_1$ to $o$; see Figure \ref{fig:lift-split}. \begin{figure}[h] \begin{overpic}[height=1.2in]{lift-split.pdf} \put(6,45){\small$p_0$} \put(92,45){\small$p_1$} \put(5.5,68){\small$\tilde p_0^{\,h}$} \put(31,67){\small$q^h$} \put(9,-1){\small$o$} \end{overpic} \caption{}\label{fig:lift-split} \end{figure} Set $ \tilde P^h_+:=(q^h,p_1)\;\text{and}\; \tilde P^h_-:=(\tilde p_0^{\,h},q^h). $ \begin{lem}\label{lem:wt} Let $P=(p_0,p_1)$ be a spiral with initial height $r$, and final height $R$. Then for every $r\leq\rho\leq R$, $$ H(P) \le \int_{r}^{\rho} w(h) \mathcal{E}(h) dh + H(\tilde P^\rho_+), $$ where $\int_r^\rho w(h)dh=L(P)-L(\tilde P^\rho_+)$, and $w\geq r/\sqrt{R^2-r^2}$. \end{lem} \begin{proof} If the desired inequality holds for all $r>1$, then it also holds for $r=1$ by continuity. So we may assume that $r>1$. We claim that $$ w(h):=-\frac{d}{ds} L(\tilde P^s_+)\Big|_{s=h} =-\frac{d}{ds} \sqrt{R^2-s^2}\Big|_{s=h}=\frac{h}{\sqrt{R^2-h^2}} $$ is the desired weight function. Clearly $\int_r^\rho w(h)dh=L(P)-L(\tilde P^\rho_+)$ and $w\geq r/\sqrt{R^2-r^2}$. Set $ \Delta h:=(\rho-r)/n, \; h_i:=r+i\Delta h,\; \text{and} \;q^i:=q^{h_i}. $ We define a sequence of liftings as follows. Set $P^0:=P$. Once $P^i$ is defined, let $\tilde p_0^i$ be its initial point, $q^i$ be its closest point to $o$, and set $P^{i}_-:=(\tilde p_0^{\,i},q^i),$ $P^{i}_+:=(q^i,p_1)$. Then we define $ P^{i+1}:=\tilde{(P^i_+)}^{h_{i+1}}; $ see Figure \ref{fig:steps}. \begin{figure}[h] \begin{overpic}[height=1.5in]{steps.pdf} \put(5,1){\small$o$} \put(2,50){\small$p_0$} \put(89,50){\small$p_1$} \end{overpic} \caption{}\label{fig:steps} \end{figure} By Lemma \ref{lem:lifting} $H(P^i_+)<H(P^{i+1})=H( P^{i+1}_-)+H(P^{i+1}_+). $ Applying this inequality iteratively yields \begin{eqnarray*} H(P) \leq \sum_{i=1}^{n} H(P^i_-)+H(\tilde P^\rho_+).\\ \end{eqnarray*} Now, for $0\leq s\leq \Delta h$, let $q^i(s):= q^{h_{i-1}+s}$; see Figure \ref{fig:triangle}. \begin{figure}[h] \begin{overpic}[height=0.6in]{triangle.pdf} \put(9,-3.5){\small$q^{i-1}$} \put(35,17){\small$q^{i}$} \put(7,24){\small$\tilde p_0^{\,i}$} \put(0,12.25){\small$x^i(s)$} \put(25,9.5){\small$q^i(s)$} \put(91.5,0){\small$p_1$} \end{overpic} \caption{}\label{fig:triangle} \end{figure} Furthermore let $x^i(s)$ be the point where the line passing through $q^i(s)$ and $p_1$ intersects $q^{i-1}\tilde p_0^{\,i}$. Set $\sigma^i_s:=x^i(s)q^i(s)$. Let $f_i(s):=H(\sigma^i_s)$. By \eqref{eq:h0-ell}, $ f_i(s)=H\big(h_{i-1}+s,L(\sigma_i(s))\big). $ So $f_i$ is $\mathcal{C}^\infty$ on $[0,\Delta h]$ provided that $h_{i-1}+s>1$ or $h_{i-1}>1$, which is the case since $r>1$. We have $ H( P^i_-)=f_i(\Delta h)-f_i(0)\leq f_i'(0)\Delta h +C_i (\Delta h)^2, $ where $C_i:=\sup_{[0,\Delta h]} f_i''(s)/2<\infty$. Note that $C_i$ depends continuously on $q^i$. So $C_i$ are bounded above by some constant $C$, independent of $i$, which yields $$ H( P^i_-)\leq f_i'(0)\Delta h +C (\Delta h)^2= f_i'(0)\Delta h +C \frac{(\rho-r)^2}{n^2}. $$ Next we compute that $$ f_i'(0)=\frac{d}{ds}L( \sigma^i_s) \Big|_{s=0} E( \sigma^i_0)+L( \sigma^i_0) \frac{d}{ds}E( \sigma^i_s)\Big|_{s=0} =\frac{d}{ds}L( \sigma^i_s) \Big|_{s=0} \mathcal{E}(h_{i-1}). $$ If we let $\tau^i_s:=q^i(s)p_1$ then at $s=0$, $ \frac{d}{ds}L( \sigma^i_s)+\frac{d}{ds}L(\tau^i_s) =\frac{d}{ds} |x^i(s)p_1|=0, $ because $x^i(0)p_1=q^{i-1}p_1$ is orthogonal to $q^{i-1}\tilde p_0^{\,i}$. Now recall that $q^h p_1=\tilde P^h_+$. Thus $$ \frac{d}{ds}L( \sigma^i_s) \Big|_{s=0}=-\frac{d}{ds}L(\tau^i_s) \Big|_{s=0} =-\frac{d}{ds}L(\tilde P^{h_{i-1}+s}_+) \Big|_{s=0} =w(h_{i-1}). $$ The last four displayed expressions yield $$ \sum_{i=1}^{n} H(P^i_-)\leq\sum_{i=0}^{n-1}\left( w(h_i)\mathcal{E}(h_i)\Delta h +C\frac{(\rho-r)^2}{n^2}\right) =\sum_{i=0}^{n-1}w(h_i)\mathcal{E}(h_i)\Delta h +C\frac{(\rho-r)^2}{n}. $$ Letting $n\to\infty$ completes the proof. \end{proof} The last lemma via an induction yields: \begin{lem}\label{lem:leqSqrt2} Let $P$ be a polygonal spiral with initial height $r$, and final height $R$. Then \begin{equation}\label{eq:leqSqrt2} H(P) \le \int_{r}^{R} w(h) \mathcal{E}(h) dh, \end{equation} where $\int_r^R w(h)dh=L(P)$, and $w\geq r/\sqrt{R^2-r^2}$. \end{lem} \begin{proof} If $P$ has only one edge \eqref{eq:leqSqrt2} holds by Lemma \ref{lem:wt}. Suppose that \eqref{eq:leqSqrt2} holds for spirals with $n$ edges and let $P=(p_0,\dots, p_{n+1})$. Let $\rho$ be the distance of the line spanned by $p_1p_2$ from $o$ and $q$ be the closest point of that line to the origin. Then $P':=(q, p_2,\dots, p_m)$ is a spiral with $n$ edges. Note that $ H(P)=H\big((p_0,p_1)\big)+H\big((p_1,\dots, p_{n+1})\big), $ and by Lemma \ref{lem:wt}, $ H\big((p_0,p_1)\big)\leq \int_r^\rho w_0(h)\mathcal{E}(h)dh +H\big((q,p_1)). $ Thus $$ H(P)\leq \int_r^{\rho} w_0(h)\mathcal{E}(h)dh +H(P'), $$ where $\int_r^\rho w_0(h)dh=L(p_0p_1)-L(q p_1)=L(P)-L(P')$, and $w_0\geq r/\sqrt{|p_1|^2-r^2}\geq r/\sqrt{R^2-r^2}$. By the inductive hypothesis $$ H(P')\leq \int_{\rho}^R w_1(h)\mathcal{E}(h)dh, $$ where $\int_r^\rho w_1(h)dh=L(P')$, and $w_1\geq \rho/\sqrt{R^2-\rho^2}\geq r/\sqrt{R^2-r^2}$. Set $w:=w_0$ for $h<\rho$ and $w:=w_1$ for $h\geq\rho$. Then the last two displayed inequalities yield \eqref{eq:leqSqrt2} . \end{proof} Now we prove the main result of this section, which extends Proposition \ref{prop:inequality}: \begin{prop}\label{prop:leqSqrt2} For any spiral $\gamma$, $E(\gamma)\leq 2$ with equality only if $|\gamma|\equiv\sqrt{2}$. \end{prop} \begin{proof} Lemma \ref{lem:leqSqrt2} together with Lemma \ref{lem:continuous} yields $E(\gamma)\leq 2$ via a polygonal approximation. Next suppose that $E(\gamma)=2$. Let $r$, $R$ be the initial and final heights of $\gamma$. If $r=R$, then $2=E(\gamma)=\mathcal E(|\gamma|)$ which yields $|\gamma|\equiv\sqrt2$. Suppose towards a contradiction that $r<R$. Let $P_i$, $i=1$, $2$, $\dots$ be a sequence of polygonal spirals converging to $\gamma$, with initial and final heights $r_i$, $R_i$. We may assume for convenience that $r\leq r_i<R_i\leq R$. Let $w_i$ be the weight functions for $P_i$ given by Lemma \ref{lem:leqSqrt2}. Set $\overline w_i:=w_i/L(P_i)$ on $[r_i,R_i]$ and $\overline w_i:=0$ elsewhere. Then $\int_{r}^{R} \overline w_i(h)dh=1$. By Lemma \ref{lem:continuous}, for any given $\epsilon>0$, we may choose $i$ so large that $ E(P_i)\geq 2-\epsilon. $ Then by Lemma \ref{lem:leqSqrt2}, $$ 2-\epsilon\leq E(P_i)\leq \int_{r}^{R} \overline w_i(h)\mathcal{E}(h) dh\leq \sup_{[r,R]}\mathcal{E}\leq 2. $$ So $\sup_{[r,R]}\mathcal{E}=2$. Since $\mathcal{E}=2$ only at $\sqrt{2}$, $[r,R]\ni\sqrt2$. So the set of heights $h\in[r,R]$ with $\mathcal{E}(h)\geq 2-\sqrt{\epsilon}$ forms a subinterval $[x_\epsilon^-, x_\epsilon^+]$. It follows that $$ 2-\epsilon\leq\int_{r}^{R} \overline w_i(h)\mathcal{E}(h)dh \le -\sqrt{\epsilon}\left(\int_{r}^{x_\epsilon^-} \overline w_i(h)dh+\int_{x_\epsilon^+}^{R} \overline w_i(h)dh \right)+ 2. $$ So $ \int_{r}^{x_\epsilon^-} \overline w_i(h)dh+\int_{x_\epsilon^+}^{R} \overline w_i(h)dh \le \sqrt\epsilon. $ But $\overline w_i\geq 1/\big(L(P_i)\sqrt{(R/r)^2-1}\,\big)$. Thus $$ R-r\leq\sqrt\epsilon L(P_i)\sqrt{(R/r)^2-1}+x_\epsilon^+-x_\epsilon^-. $$ Letting $\epsilon\to 0$, we obtain $r=R$, since $ x_\epsilon^\pm \to\sqrt{2}, $ and $L(P_i)$ is bounded above. Hence we arrive at the desired contradiction. \end{proof} \section{Proof of Theorem \ref{thm:main}}\label{sec:proof} By Proposition \ref{prop:minimizer} there exists a minimal inspection curve $\gamma\colon[a,b]\to\mathbf{R}^3$, which we may assume to have constant speed. As we described in Section \ref{sec:integral}, to establish \eqref{eq:main} it suffices to show that $E(\gamma)\leq 2$. By Proposition \ref{prop:unfolding}, $E(\gamma)=E(\tilde\gamma)$ where $\tilde\gamma$ is the unfolding of $\gamma$. By Proposition \ref{prop:decomposition}, $\tilde\gamma$ admits a spiral decomposition, generated by a collection of mutually disjoint open sets $U_i\subset[a,b]$, $i\in I$. Set $U_0:=[a,b]\setminus \cup_i\overline U_i$, and let $\tilde\gamma_i:=\tilde\gamma|_{\overline U_i}$, $\tilde\gamma_0:=\tilde\gamma|_{U_0}$. Then \begin{equation}\label{eq:E-gamma} E(\tilde\gamma)=\frac{H(\tilde\gamma)}{L(\tilde\gamma)}=\frac{1}{L(\tilde\gamma)}\sum_iH(\tilde\gamma_i)= \frac{1}{L(\tilde\gamma)}\left(L(\tilde\gamma_0)E(\tilde\gamma_0)+ \sum_iL(\tilde\gamma_i)E(\tilde\gamma_i)\right), \end{equation} where we define $L(\tilde\gamma_0):=\int_{U_0}|\tilde\gamma_0'(t)|dt$. So $L(\tilde\gamma_0)+ \sum_iL(\tilde\gamma_i)=L(\tilde\gamma)$. If $L(\tilde\gamma_0)=0$, then we may disregard the first term in the summation above. Otherwise, by definition of spiral decomposition, $\tilde\alpha(t)=\pi/2$ for almost all $t\in U_0$. Thus, by Proposition \ref{prop:IE}, \begin{equation}\label{eq:E-gamma-0} E(\tilde\gamma_0)=\frac{1}{\mu(U_0)}\int_{U_0} \mathcal{E}\left(|\tilde\gamma(t)|,\frac{\pi}{2}\right)dt=\frac{1}{\mu(U_0)}\int_{U_0}\mathcal{E}\big(|\tilde\gamma(t)|\big)dt\leq 2. \end{equation} Furthermore, by Proposition \ref{prop:inequality} or \ref{prop:leqSqrt2}, \begin{equation}\label{eq:E-gamma-1} E(\tilde\gamma_i)\leq 2, \end{equation} assuming $U_i\neq\emptyset$. So it follows that $E(\tilde\gamma)\leq 2$, as desired. It remains to characterize the case of equality in \eqref{eq:main}, which corresponds to $E(\gamma)= 2$. Then $E(\tilde\gamma)=2$, which yields that the terms $E(\tilde\gamma_i)$ and $E(\tilde\gamma_0)$ in \eqref{eq:E-gamma} must all be equal to $2$. But the inequality in \eqref{eq:E-gamma-1} must be strict by Proposition \ref{prop:leqSqrt2}, since $\tilde\gamma_i$ are strict spirals by definition of spiral decomposition. So $\tilde\gamma$ cannot contain any strict spirals or $U_i=\emptyset$, which means that $U_0=[a,b]$ or $\tilde\gamma$ has constant height. Furthermore, equality in \eqref{eq:E-gamma-0} implies that $\mathcal{E}(|\tilde\gamma(t)|)\equiv2$ which can happen only when $|\tilde\gamma(t)|\equiv\sqrt2$. So we conclude that $\gamma$ has constant height $\sqrt{2}$, since unfoldings preserve height. Now let $\overline\gamma:=\gamma/\sqrt2$ be the projection of $\gamma$ into $\S^2$. Then $L(\overline\gamma)=L(\gamma)/\sqrt2=4\pi/\sqrt2$. Recall that, since $\gamma$ is an inspection curve, the horizon circles generated by points of $\gamma$ cover $\S^2$. Since $|\gamma|\equiv\sqrt{2}$, these circles have (spherical) radius $\pi/4$ and are centered at points of $\overline\gamma$. Thus $\overline\gamma$ satisfies the hypothesis of the following proposition, which will complete the proof of Theorem \ref{thm:main}. \begin{prop}\label{prop:baseball} Let $\gamma\colon[a,b]\to\S^2$ be a closed constant speed curve with $L(\gamma)=4\pi/\sqrt2$. Suppose that the distance between any point of $\S^2$ and $\gamma$ is at most $\pi/4$. Then $\gamma$ is a simple $\mathcal{C}^{1,1}$ curve which traces consecutively $4$ semicircles of length $\pi/\sqrt2$. \end{prop} It remains then to establish the above proposition. To this end we need: \begin{lem}[Crofton-Blaschke-Santalo \cite{santalo:1942}]\label{lem:crofton} Let $\gamma\colon[a,b]\to\S^2$ be a rectifiable curve, and for every point $p\in\S^2$, and $0\leq\rho\leq\pi/2$, let $C_\rho(p)\subset\S^2$ denote the circle of radius $\rho$ centered at $p$. Then $ L(\gamma)=\frac{1}{4\sin(\rho)}\int_{p\in\S^2} \#\gamma^{-1}\big(C_\rho(p)\big)\,dp. $ \end{lem} \noindent For the rest of this section we assume that $\gamma$ satisfies the hypothesis of the last proposition. Then, since $L(\gamma)=4\pi/\sqrt2$, applying the last lemma with $\rho=\pi/4$ to $\gamma$ yields $$ \underset{p\in\S^2}{\textup{Ave}}\;\#\gamma^{-1}\big(C_{\frac\pi4}(p)\big) =\frac{1}{4\pi}\int_{p\in\S^2} \#\gamma^{-1}\big(C_{\frac\pi4}(p)\big)dp =\frac{1}{4\pi}L(\gamma)\,4\,\sin\left(\frac\pi4\right) =2. $$ Furthermore, note that $C_{\frac\pi4}(p)$ must intersect $\gamma$ for all $p\in\S^2$, since the distance of $p$ from $\gamma$ cannot be bigger than $\pi/4$ by assumption. So, since $\gamma$ is closed, $\#\gamma^{-1}(C_{\frac\pi4}(p))\geq 2$ for almost all $p\in\S^2$. Now since the average of $\#\gamma^{-1}(C_{\frac\pi4}(p))$ is $2$, it follows that \begin{lem}\label{lem:2-points} For almost every $p\in\S^2$, $ \#\gamma^{-1}\big(C_{\frac\pi4}(p)\big)=2. $ \end{lem} By a \emph{side} of a circle $C$ in $\S^2$ we mean either of the two closed disks in $\S^2$ bounded by $C$. If the radius of $C$ is less than $\pi/2$, then the disk with radius less than $\pi/2$ is called the \emph{inside} of $C$ and the other disk is called the \emph{outside} of $C$. By \emph{strictly inside} or \emph{strictly outside} we mean the interior of inside and interior of outside respectively. \begin{lem}\label{lem:outside-length} For any point $p\in\S^2$, the portion of $\gamma$ which lies outside $C_{\frac\pi4}(p)$ has length at least $\pi$. \end{lem} \begin{proof} By assumption, $\gamma$ intersects $C_{\frac\pi4}(-p)$, which has distance $\pi/2$ from $C_{\frac\pi4}(p)$. Furthermore, since $\gamma$ is closed, there must exist at least two segments of $\gamma$ which connect $C_{\frac\pi4}(-p)$ and $C_{\frac\pi4}(p)$. \end{proof} For the rest of this section, we will assume that $\gamma$ is reparameterized so that $[a,b]=[0,2\pi]$, and identify $[0,2\pi]$ with the unit circle $\S^1\simeq \mathbf{R}/(2\pi \mathbf{Z})$. Furthermore we fix an orientation on $\S^1$. Then for every pair of distinct points $t$, $s\in\S^1$, we let $[t,s]$ denote the segment in $\S^1$ whose orientation from $t$ to $s$ agrees with the orientation of $\S^1$. \begin{lem}\label{lem:tangent-cone} For every $t\in\S^1$, the tangent cone $T_t\gamma$ is a line. \end{lem} \begin{proof} Let $s_i\in\S^1$ be a sequence of points converging to $t$ from the left hand side (with respect to the orientation of $\S^1$). Since $\gamma$ has non-vanishing speed, it cannot be locally constant. Thus we may assume, after passing to a subsequence, that $\gamma(s_i)\neq\gamma(t)$. Then the secant rays $\ell_i$ in $\mathbf{R}^3$ which emanate from $\gamma(t)$ and pass through $\gamma(s_i)$ are well-defined. Let $\ell$ be a limit of $\ell_i$. Similarly, we can consider the secant rays $\ell_i'$ generated by points $s_i'\in\S^1$ converging to $t$ from the right hand side, and let $\ell'$ be a limit of $\ell_i'$. We claim that the angle between $\ell$ and $\ell'$ is $\pi$. Suppose not. Then there exist points $s$, $s'\in\S^1$ arbitrary close to $t$ and with $(s,s')\ni t$ such that the angle between the geodesic segments $\gamma(t)\gamma(s)$ and $\gamma(t)\gamma(s')$ in $\S^2$ is less than $\pi$. Consequently, there exists an open set $S$ of circles of radius $\pi/4$ in $\S^2$ such that for every $C\in S$ we have $\gamma(s)$, $\gamma(s')$ lie strictly inside $C$ while $\gamma(t)$ lies strictly outside $C$. Thus, by Lemma \ref{lem:2-points}, there exists a circle $C\in S$ which intersects $\gamma$ in only two points. So the portion of $\gamma$ which lies outside $C$ is a subset of $\gamma([s,s'])$. But since $s$ and $s'$ may be chosen arbitrarily close to $t$, the length of $\gamma([s,s'])$ may be arbitrarily small. Hence we obtain the desired contradiction via Lemma \ref{lem:outside-length}. So the angle between $\ell$ and $\ell'$ is $\pi$ as claimed. Now since $\ell$ and $\ell'$ where arbitrary limits of the right and left secant rays of $\gamma$ at $t$, and all these limits are tangent to $\S^2$, it follows that $\ell$ and $\ell'$ are unique. Hence $T_t\gamma=\ell\cup\ell'$ as desired. \end{proof} Now for each $t\in\S^1$, the left and right unit tangent vectors of $\gamma$, $ u^\pm_\gamma(t) $ are well-defined with $u^+_\gamma(t)=-u^-_\gamma(t)$. \begin{lem}\label{lem:gamma-tangent} Let $C\subset \S^2$ be a circle of radius $\pi/4$. Suppose that there exists an interval $[t,s]\subset\S^1$ such that $\gamma(t)$ lies on $C$ while $\gamma((t,s])$ lies strictly inside $C$. Then $\gamma$ is transversal to $C$ at $\gamma(t)$. \end{lem} \begin{proof} Suppose towards a contradiction that $\gamma$ is tangent to $C$ at $\gamma(t)$. Let $C'$ be a circle of radius $\pi/4$ in $\S^2$ which passes through $\gamma(t)$ and is transversal to $C$ at $\gamma(t)$ with $u^+_\gamma(t)$ pointing outside $C'$. Then there exist $r\in (t,s)$ such that $\gamma(r)$ lies strictly outside $C'$. Furthermore, choosing $C'$ sufficiently close to $C$, we can ensure that $\gamma(s)$ lies strictly inside $C'$. Next, by perturbing the center of $C'$, we may find another circle $C''$ of radius $\pi/4$ such that $\gamma(t)$ and $\gamma(s)$ lie strictly inside $C''$ while $\gamma(r)$ lies strictly outside $C''$. Since $C''$ may be chosen freely from an open set of circles in $\S^2$, we may assume by Lemma \ref{lem:2-points} that $C''$ intersects $\gamma$ at only two points. Thus the portion of $\gamma$ lying outside $C''$ is a subset of $\gamma([t,s])$. But $\gamma([t,s])$ can have arbitrarily small length, since we may choose $s$ as close to $t$ as desired. Thus we obtain a contradiction by Lemma \ref{lem:outside-length}. \end{proof} We say that a circle $C\subset\S^2$ \emph{supports} $\gamma$ at a point $p$ of $\gamma$ provided that $C$ passes through $p$ and $\gamma$ lies on one side of $C$. Furthermore, if the radius of $C$ is less than $\pi/2$, then we assume that $\gamma$ lies outside $C$. \begin{lem}\label{lem:support-circles} Through each point of $\gamma$ there pass a pair of support circles of radius $\pi/4$ which lie outside each other. \end{lem} \begin{proof} Let $C$ be one of the two circles of radius $\pi/4$ tangent to $\gamma$ at $\gamma(t)$. Suppose there is a point $t'\in\S^1$ such that $\gamma(t')$ lies strictly inside $C$. Let $D$ be the disk of radius $\pi/4$ bounded by $C$, and $I$ be the closure of the component of $\gamma^{-1}(\inte(D))$ which contains $t'$. By Lemma \ref{lem:2-points}, $\gamma$ cannot lie entirely in $C$. Thus $I$ is a proper interval in $\S^1$. By Lemma \ref{lem:gamma-tangent}, $\gamma$ is transversal to $C$ at the end points of $I$. In particular there are points $s_1$, $s_2\in\S^1$ close to each of the end points of $I$ such that $\gamma(s_i)$ lie strictly outside $C$, and $t$, $s_1$, $t'$, $s_2$ are arranged cyclically in $\S^1$. Perturbing the center of $C$, we may find a circle $C'$ of radius $\pi/4$ such that $\gamma(t)$ and $\gamma(t')$ lie strictly inside $C'$, while $\gamma(s_i)$ lie strictly outside $C'$. It follows that $C'$ intersects $\gamma$ at least $4$ times, which contradicts Lemma \ref{lem:2-points}, since $C'$ may be chosen freely from an open set of circles in $\S^2$. \end{proof} In the terminology of \cite{ghomi-howard2014}, the conclusion of Lemma \ref{lem:support-circles} means that $\gamma$ has \emph{double positive support}. Using this lemma we next show: \begin{lem}\label{lem:simple} $\gamma$ is simple. \end{lem} \begin{proof} Suppose that there are distinct points $t$, $s\in \S^1$ with $\gamma(t)=\gamma(s)$. Let $r_1$, $r_2$ be points of $\S^1$ which lie in the interior of different segments of $\S^1$ determined by $s$ and $t$, so that $s$, $r_1$, $t$, $r_2$ are cyclically arranged in $\S^1$. By Lemma \ref{lem:support-circles}, there exists a circle $C$ of radius $\pi/4$ in $\S^2$ which supports $\gamma$ at $\gamma(t)=\gamma(s)$. Furthermore, by Lemma \ref{lem:2-points}, $\gamma$ cannot lie completely on $C$. So we may choose $r_i$ so that at least one of the points $\gamma(r_1)$, $\gamma(r_2)$ lies strictly outside $C$. Then we may translate $C$ to obtain a circle $C'$ of the same radius such that $\gamma(t)=\gamma(s)$ lies strictly inside $C'$ while $\gamma(r_i)$ lie strictly outside $C'$. Hence $C'\cap\gamma$ consist of at least $4$ points. Furthermore $C'$ may be chosen from an open set of circles of radius $\pi/4$ in $\S^2$. Thus we obtain a violation of Lemma \ref{lem:2-points}. \end{proof} For a planar curve, double positive support is equivalent to \emph{positive reach} introduced by Federer \cite{federer1959}. Thus the last two lemmas imply that $\gamma$ has positive reach. \begin{lem}\label{lem:C11} $\gamma$ is $\mathcal{C}^{1,1}$. \end{lem} \begin{proof} Since $\gamma$ has finite length, there exists a point in $\S^2\setminus\gamma$, which we may assume to be $(0,0,1)$ after a rotation. Let $\pi\colon\S^2\setminus\{(0,0,1)\}\to\mathbf{R}^2$ be the stereographic projection, and set $\tilde\gamma:=\pi\circ\gamma$. Since $\pi$ preserve circles, and by Lemma \ref{lem:support-circles} $\gamma$ has double positive support, $\tilde\gamma$ has double positive support as well. Furthermore, by Lemma \ref{lem:simple} $\gamma$ has two sides in $\S^2$. The support circles of $\gamma$ must lie in opposite sides of $\gamma$ at each point; otherwise the tangent cone would be a ray (or $\gamma$ would have a cusp) which is not possible by Lemma \ref{lem:tangent-cone}. Thus the support circles of $\tilde\gamma$ must lie on the opposite sides of $\tilde\gamma$ as well. Consequently $\tilde\gamma$ is $\mathcal{C}^{1,1}$ by \cite[Thm. 1.2]{ghomi-howard2014}; see also \cite[prop. 1.4]{lytchak}. \end{proof} Next we observe that: \begin{lem}\label{lem:antipodal} Let $C$ be a support circle of $\gamma$ of radius $\pi/4$. Then $C\cap \gamma$ is either a pair of antipodal points of $C$ or else is a semicircle of $C$. \end{lem} \begin{proof} We claim that (i) every closed semicircle of $C$ intersects $\gamma$, and (ii) every open semicircle of $C$ intersects $\gamma$ in a connected set. These properties easily imply that $\gamma\cap C$ is either a pair of antipodal points of $C$, a closed semicircle of $C$, or the entire $C$. The last possibility is not allowed, because by Lemma \ref{lem:simple} $\gamma$ is simple; therefore, if $\gamma$ covers $C$, it must coincide with $C$, which would violate Lemma \ref{lem:2-points}. So it remains to establish the claims. To see (i) suppose that there exists a closed semicircle of $C$ which does not intersect $\gamma$. Then moving the center of $C$ by a small distance towards the center of that semicircle yields a circle $C'$ of radius $\pi/4$ disjoint from $\gamma$. Obviously all circles of radius $\pi/4$ which are close to $C'$ will be disjoint from $\gamma$ as well, which would violate Lemma \ref{lem:2-points}. To see (ii) suppose that there exists an open semicircle $S$ of $C$ which intersects $\gamma$ in a disconnected set. Then there exist points $t_1$, $t_2$, $s\in\S^1$, with $s\in(t_1,t_2)$ such that $\gamma(t_i)\in S$ while $\gamma(s)$ lies strictly outside $C$. Either $\gamma((t_2,t_1))$ lies entirely on $C$ or not. In the former case there exist a point $s'\in(t_2, t_1)$ such that $\gamma(s')$ lies in the open semicircle of $C$ which is disjoint from $S$; in the latter case there exists a point $s'\in(t_2, t_1)$ such that $\gamma(s')$ lies strictly outside $C$. In either case, moving the center of $C$ by a small distance towards the midpoint of $S$ will yield a circle $C'$ of radius $\pi/4$ such that $\gamma(t_i)$ lie strictly inside $C'$ while $\gamma(s)$, $\gamma(s')$ lie strictly outside $C'$. But $t_1$, $s$, $t_2$, $s'$ are cyclically arranged in $\S^1$. So perturbing the center of $C'$ yields an open set of circles of radius $\pi/4$ each of which intersects $\gamma$ at least $4$ times, which again contradicts Lemma \ref{lem:2-points}. \end{proof} The last lemma leads to the proof of Proposition \ref{prop:baseball} via the notion of \emph{nested partitions of a circle} employed in \cite{ghomi:rosenberg}, see also \cite{umehara2,thorbergsson&umehara}. By Lemma \ref{lem:simple}, $\gamma$ bounds a topological disc $D\subset\S^2$. By Lemmas \ref{lem:support-circles} and \ref{lem:C11} through each point $p\in\gamma$ there passes a circle $C_p$ of radius $\pi/4$ which lies in $D$. It follows from Lemma \ref{lem:C11} that $C_p$ is unique. Thus if we set $ [p]:=\gamma\cap C_p, \;\text{and}\; P:=\{[p]\mid p\in\gamma\}, $ then $P$ will be a partition of $\gamma$. A partition $P$ of a topological circle is \emph{nested} provided that no element of $P$ separates the components of any other element, i.e., for every $[p]\in P$ and $q\in\gamma\setminus [p]$, $[q]$ lies in a connected component of $\gamma\setminus [p]$. \begin{lem} The partition $P$ of $\gamma$ is nested. \end{lem} \begin{proof} If not there are distinct support circles $C$, $C'$ of $\gamma$ of radius $\pi/4$ contained in $D$ such that $C$ has points in different components of $\gamma\setminus C'$. In particular neither $C\cap\gamma$ nor $C'\cap\gamma$ is connected. So by Lemma \ref{lem:antipodal}, $C$ and $C'$ intersect $\gamma$ in two points each, say $C\cap\gamma=\{p,q\}$ and $C'\cap\gamma=\{p',q'\}$. Each of the segments $pq$ and $qp$ of $C$ separate $D$ into two components. Thus each of the segments $p'q'$ and $q'p'$ of $C'$ must intersect each of the segments $pq$ and $qp$. Furthermore, each of these intersections must occur in the interior of the segments, because the interior of each segment is disjoint from $\gamma$. Thus $C$ and $C'$ intersect at least $4$ times. So $C=C'$, which is the desired contradiction. \end{proof} Finally we invoke the following fact established in \cite[Lem. 2.2]{ghomi:rosenberg}. A partition is \emph{nontrivial} provided that it contains more than one element. \begin{lem} Any nontrivial nested partition of a topological circle contains at least two elements which are connected subsets of the circle. \end{lem} So there are two distinct elements $[p_1]$, $[p_2]\in P$ such that $C_{p_i}\cap\gamma$ is a connected set. Consequently, by Lemma \ref{lem:antipodal}, $C_{p_i}\cap\gamma$ are semicircles. Thus $\gamma$ contains a pair of disjoint semicircles which curve toward $D$. Similarly, by repeating the above argument for the other domain $D'$ in $\S^2$ bounded by $\gamma$, we obtain two disjoint semicircles in $\gamma$ which curve toward $D'$. Since each semicircle has length $\pi/\sqrt{2}=L(\gamma)/4$, the semicircles cover $\gamma$, which completes the proof of Proposition \ref{prop:baseball}. \section*{Acknowledgments} We thank Joseph O'Rourke for posting the sphere inspection problem on MathOverflow \cite{orourkeMO} which prompted this work, and commenters on MathOverflow, including Pietro Majer, Anton Petrunin, Jean-Marc Schlenker, and Gjergji Zaimi for their observations. Thanks also to Steven Finch for sending us a translation of Zalgaller's article \cite{finch2019translation}. \addtocontents{toc}{\protect\setcounter{tocdepth}{0} \addtocontents{toc}{\protect\setcounter{tocdepth}{1}

https://arxiv.org/abs/2107.02428

Browder's Theorem through Brouwer's Fixed Point Theorem

One of the conclusions of Browder (1960) is a parametric version of Brouwer's Fixed Point Theorem, stating that for every continuous function $f : ([0,1] \times X) \to X$, where $X$ is a simplex in a Euclidean space, the set of fixed points of $f$, namely, the set $\{(t,x) \in [0,1] \times X \colon f(t,x) = x\}$, has a connected component whose projection on the first coordinate is $[0,1]$. Browder's (1960) proof relies on the theory of the fixed point index. We provide an alternative proof to Browder's result using Brouwer's Fixed Point Theorem.

\section{Introduction} Brouwer's Fixed Point Theorem (Hadamard, 1910, Brouwer, 1911) states that every continuous function from a finite dimensional simplex into itself has a fixed point. This result was later generalized to nonempty, convex, and compact subsets of more general topological vector spaces, see, e.g., Schauder (1930), Tychonoff (1935), and Dyer (1956). The following parametric version of Brouwer's Fixed Point Theorem is a special case of a more general result of Browder (1960). To state the theorem we need the concept of connected component. A set $A \subseteq \dR^n$ is \emph{connected} if there are no two disjoint open sets $O_1,O_2$ that satisfy (a) $A \subseteq O_1 \cup O_2$, (b) $A \not\subseteq O_1$, and (c) $A \not\subseteq O_2$. A subset $B$ of $A$ is a \emph{connected component} of $A$ if every connected subset of $A$ is either contained in $B$ or disjoint of $B$. \begin{theorem}[Browder, 1960] \label{theorem:browder} Let $f : ([0,1] \times X) \to X$ be a continuous function, where $X = [0,1]^n$. Define the set of fixed points of $f$ by \begin{equation} \label{equ:98} C_f := \{ (t,x) \in [0,1] \times X \colon f(t,x)=x\}. \end{equation} Then $C_f$ has a connected component whose projection to the first coordinate is $[0,1]$. \end{theorem} \begin{example} \label{example:1} Let $X = [-1,1]$, and $f : [0,1] \times X \to X$ be given by \[ f(t,x) := \left\{ \begin{array}{lll} x & \ \ \ \ \ & t=0,\\ (1-t)x + t\sin(\tfrac{1}{t}) & & t \neq 0. \end{array} \right. \] The set $C_f$ is the union of $\{ (0,x) \colon x \in [-1,1]\}$ and $\{ (t, \sin(\tfrac{1}{t})) \colon t \in (0,1]\}$, which is connected (but not path connected). \end{example} Theorem~\ref{theorem:browder} was used in a variety of topics, like nonlinear complementarity theory (see, e.g., Eaves, 1971, or Allgower and Georg, 2012), nonlinear elliptic boundary value problems (Shaw, 1977), the study of global continua of solutions of nonlinear partial differential equations (see, e.g., Costa and Gon\c{c}alves, 1981, or Massabo and Pejsachowitz, 1984), theoretical economics (Citanna et al., 2001), and game theory (see, e.g., Herings and Peeters, 2010, or Solan and Solan, 2021). Browder's (1960) proof of Theorem~\ref{theorem:browder} uses the fixed point index \footnote{In fact, the statement of Browder's (1960) more general version of Theorem~\ref{theorem:browder} is phrased using the fixed point index.} and hence is not accessible to many researchers, and cannot be taught in an undergraduate course in topology. In this paper we prove Theorem~\ref{theorem:browder} using Brouwer's Fixed Point Theorem. In particular, our proof is accessible to all mathematicians, and can be taught in any course in which Brouwer's Fixed Point Theorem is proved. In Section~\ref{section:discussion} we discuss extensions of Theorem~\ref{theorem:browder} to more general parameter sets and more general sets $X$. Theorem~\ref{theorem:browder} easily follows from Brouwer's Fixed Point Theorem when the number of connected components of $C_f$ is finite \footnote{In this case, our Proposition~\ref{prop:1} is trivial, hence the proof reduces to our Proposition~\ref{prop:2}.} As the next example shows, the number of connected components of $C_f$ may not be finite or countable. \begin{example} Recall that the \emph{Cantor set} $K$ is the set of all real numbers in $[0,1]$ such that, in their representation in base 3, appear only the digits 0 and 2. The cardinality of the Cantor set is the continuum, and its complement is a union of countably many open intervals. Let $g : [0,1] \to [0,1]$ be the function that is the identity on $K$, and, on each maximal open subinterval $(a,b)$ of $[0,1]$ in the complement of $K$ it is given by $g(x) = x + (x-a)(b-x)$, see Figure~\arabic{figurecounter}. The function $g$ is continuous, its range is $[0,1]$, and its set of fixed points is $K$. Define now a function $f : [0,1] \times [0,1] \to [0,1]$ by $f(t,x) = g(x)$ for every $(t,x) \in [0,1] \times [0,1]$. The connected components of $C_f$ are then all sets of the form $[0,1] \times \{x\}$ for $x \in K$. \label{example:2} \end{example} \bigskip \centerline{ \includegraphics{browder_fig.1} } \vspace{0.2truecm} \centerline{{Figure \arabic{figurecounter}: The function $g$ in Example~\ref{example:2}.}} \addtocounter{figurecounter}{1} \section{Proof of Theorem~\ref{theorem:browder}} In this section we prove Theorem~\ref{theorem:browder}. The theorem will follow from Brouwer's Fixed Point Theorem once we prove the following two results. \begin{proposition} \label{prop:1} If $C_f$ has no connected component whose projection on the first coordinate is $[0,1]$, then there are two disjoint open sets $O_0$ and $O_1$ that satisfy the following properties: \begin{enumerate} \item[(C1)] The sets $O_0$ and $O_1$ cover $C_f$, that is, $C_f \subseteq O_0 \cap O_1$. \item[(C2)] Every connected component $B$ of $C_f$ that satisfies $B \cap (\{0\} \times X) \neq \emptyset$ is a subset of $O_0$. \item[(C3)] Every connected component $B$ of $C_f$ that satisfies $B \cap (\{1\} \times X) \neq \emptyset$ is a subset of $O_1$. \end{enumerate} \end{proposition} \begin{proposition} \label{prop:2} If there are two disjoint open sets $O_0$ and $O_1$ that satisfy (C1)--(C3), then there is a continuous function $F : ([0,1] \times X) \to ([0,1] \times X)$ that has no fixed point. \end{proposition} To see that the two propositions imply Theorem~\ref{theorem:browder}, note that the conclusion of Proposition~\ref{prop:2} contradicts Brouwer's Fixed Point Theorem, hence Proposition~\ref{prop:2} implies that there are no two disjoint open sets $O_0$ and $O_1$ that satisfy (C1)--(C3). Hence Proposition~\ref{prop:1} implies that $C_f$ has a connected component whose projection on the first coordinate is $[0,1]$. As Example~\ref{example:1} shows, connected components of $C_f$ may be complicated sets, and as Example~\ref{example:2} shows, the number of connected components of $C_f$ may be of the order of the continuum. In particular, the condition that $C_f$ has no connected component whose projection on the first coordinate is $[0,1]$ is difficult to use. Proposition~\ref{prop:1} turns the contrapositive assumption of Theorem~\ref{theorem:browder} into a seemingly stronger condition that is easier to use. The proof of Proposotion~\ref{prop:1} is the more challenging part of the proof of Theorem~\ref{theorem:browder}, and it goes through the following steps. \begin{itemize} \item For every $k \in \dN$ we will approximate $C_f$ by a ``simple'' set $S_k$. We will do that by covering $[0,1] \times X$ with finitely many boxes of diameter $\frac{1}{2^k}$, and defining $S_k$ to be the union of all boxes that intersect $C_f$. \item We will then prove that if $C_f$ has no connected component whose projection on the first coordinate is $[0,1]$, then there is $k$ such that $S_k$ has no connected component whose projection on the first coordinate is $[0,1]$. \item Since $S_k$ is the union of finitely many boxes, if it has no connected component whose projection on the first coordinate is $[0,1]$, then the existence of two disjoint open sets $O_0$ and $O_1$ that satisfy (C1)--(C3) w.r.t.~$S_k$ (rather than $C_f$) is clear. Since $S_k \supseteq C_f$, Proposition~\ref{prop:1} follows. \end{itemize} \subsection{Proof of Proposition~\ref{prop:1}} The maximum norm in $\dR^n$ is given by \[ d_\infty(y,y') := \max_{i=1,\dots,n}|y_i-y'_i|, \ \ \ \forall y,y' \in \dR^n. \] In the proof we will use the distance between a point and a set and the distance between two sets: for every two sets $A,A' \subseteq \dR^n$ and every point $y \in \dR^n$, \[ d_\infty(y,A) := \inf_{y' \in A} d_\infty(y,y'), \ \ \ \ \ d_\infty(A,A') := \inf_{y \in A, y' \in A'} d_\infty(y,y'). \] We will also use the Hausdorff distance between sets: \[ d_H(A,A') := \max\left\{ \sup_{y \in A} d_\infty(y,A'), \sup_{y' \in A'} d_\infty(y',A)\right\}. \] For every $k \in \dN$, let $\calT_k$ be the collection of all boxes $\prod_{i=1}^{n+1} [a_i,b_i] \subseteq [0,1]^{n+1}$ where $a_i$ and $b_i$ are rational numbers that are integer multiples of $\frac{1}{2^k}$. We note that $\calT_{k+1}$ refines $\calT_k$: every set $T \in \calT_{k+1}$ is a subset of some set $T' \in \calT_k$. Let \[ S_k := \bigcup\{ T \in \calT_k \colon T \cap C_f \neq \emptyset\}. \] This is the union of all boxes in $\calT_k$ that contain points in $C_f$, see Figure~\arabic{figurecounter}, where the set $C_f$ has three connected components. In particular, $S_k \supseteq C_f$. Since $\calT_{k+1}$ refines $\calT_k$, we have $S_{k+1} \subseteq S_k$. \bigskip \centerline{ \includegraphics{browder_fig.6} } \vspace{0.2truecm} \centerline{{Figure \arabic{figurecounter}: The boxes in $\calT_k$ and the sets $C_f$ (dark) and $S_k$ (grey).}} \addtocounter{figurecounter}{1} \begin{lemma} If there is $k \in \dN$ such that $S_k$ has no connected component whose projection on the first coordinate is $[0,1]$, then there are two disjoint open sets $O_0$ and $O_1$ that satisfy (C1)--(C3). \end{lemma} \begin{proof} Let $A_0$ be the union of all connected component of $S_k$ that intersect $\{0\} \times X$. Let $A_1 := S_k \setminus A_0$. Each of the sets $A_0$ and $A_1$ is a union of finitely many boxes, and the two sets are disjoint. It follows that $d_\infty(O_0,O_1) \geq \frac{1}{2^k}$. This implies that the open sets \[ O_0 := \left\{ (t,x) \in Y \colon d_\infty((t,x),A_0) < \frac{1}{2^{k+1}}\right\}, \] and \[ O_1 := \left\{ (t,x) \in Y \colon d_\infty((t,x),A_1) < \frac{1}{2^{k+1}}\right\}, \] satisfy (C1)--(C3). \end{proof} \bigskip From now on we assume that for every $k \in \dN$, the set $S_k$ has a connected component whose projection on the first coordinate is $[0,1]$. We will prove that in this case, $C_f$ has a connected component whose projection on the first coordinate is $[0,1]$. \begin{lemma} Suppose that for every $k \in \dN$, the set $S_k$ has a connected component whose projection on the first coordinate is $[0,1]$. There is a decreasing sequence of closed sets $(D_k)_{k \in \dN}$ that satisfies the following properties for every $k \in \dN$: \begin{enumerate} \item[(D1)] $D_k$ is a union of boxes in $\calT_k$, and in particular it is closed. \item[(D2)] $D_k \subseteq S_k$. \item[(D3)] If $k > 1$ then $D_k \subseteq D_{k-1}$. \item[(D4)] For every $l \geq k$, The set $S_l \cap D_k$ has a connected component whose projection on the first coordinate is $[0,1]$. \item[(D5)] $D_k$ is a minimal subset of $S_k$ that satisfies (D1)--(D4). \end{enumerate} \end{lemma} \begin{proof} The proof is by induction over $k$. Define $D_{-1} := [0,1] \times X$, let $k \in \dN$ be given, and assume that we already defined $(D_j)_{j=1}^{k-1}$ in a way that satisfies (D1)--(D5). We argue that the set $D_k := S_k \cap D_{k-1}$ satisfies (D1)--(D4). For $k=1$ this holds by the properties of $(S_k)_{k\in \dN}$. Assume now that $k>1$. By definition, (D2) and (D3) hold. Since $S_k$ is a union of boxes in $\calT_k$, since $D_{k-1}$ is a union of boxes in $\calT_{k-1}$, and since $\calT_k$ refines $\calT_{k-1}$, (D1) holds. (D4) holds since $D_{k-1}$ satisfies (D4). Since the set $S_k \cap D_{k-1}$ is composed of finitely many boxes in $\calT_k$, it has finitely many subsets that satisfy (D1)--(D3), and at least one of them, $S_k \cap D_{k-1}$, satisfies (D4). Let $D_k$ be a minimal (w.r.t.~set inclusion) subset of $S_k \cap D_{k-1}$ that satisfies (D1)--(D4). Then $D_k$ also satisfies (D5). \end{proof} \bigskip Since the sequence $(D_k)_{k \in \dN}$ is a decreasong sequence of closed sets, the intersection $D_* := \bigcap_{k \in \dN} D_k$ is closed and nonempty. Since the sets $(D_k)_{k \in \dN}$ are contained in the compact set $[0,1]\times X$, we have $\lim_{k \to \infty} d_H(D_k,D_*) = 0$. As we now show, the minimality of $D_k$ (Property (D5)) implies that $D_k$ is connected. In fact, this implication is the reason for requiring $D_k$ to satisfy Property~(D5). \begin{lemma} \label{lemma:2a} The set $D_k$ is connected, for every $k \in \dN$. \end{lemma} \begin{proof} Assume by way of contradiction that $D_{k_*}$ is not connected for some $k_* \in \dN$. Let $O'$ and $O''$ be two disjoint open sets that satisfy (a) $D_{k_*} \subseteq O' \cup O''$, (b) $D_{k_*} \not\subseteq O'$, and (c) $D_{k_*} \not\subseteq O''$. For every $k \in \dN$, every connected component of $D_k$ lies either in $D_k \cap O'$ or in $D_k \cap O''$. Hence, and since for every $k \in \dN$, the set $D_k$ has a connected component whose projection on the first coordinate is $[0,1]$, at least one of the sets $D_k \cap O'$ and $D_k \cap O''$ has a connected component whose projection on the first coordinate is $[0,1]$. Assume w.l.o.g.~that for infinitely many $k$'s, the set $D_k \cap O'$ has a connected component whose projection on the first coordinate is $[0,1]$. Since the sequence $(D_k)_{k \in \dN}$ is decreasing, if $D_{k+1} \cap O'$ has a connected component whose projection on the first coordinate is $[0,1]$, then $D_k \cap O'$ has a connected component whose projection on the first coordinate is $[0,1]$. It follows that for every $k \in \dN$, and in particular for $k=k_*$, the set $D_k \cap O'$ has a connected component whose projection on the first coordinate is $[0,1]$. But then the set $D_{k_*} \cap O'$ satisfies Properties (D1)--(D4) for $k_* = k$, contradicting the minimality of $D_{k_*}$ (Property (D5)). \end{proof} \begin{lemma} \label{lemma:2} The set $D_*$ is connected. \end{lemma} \begin{proof} Assume by way of contradiction that $D_*$ is not connected, and let $O'$ and $O''$ be two disjoint open sets that satisfy (a) $D_* \subseteq O' \cup O''$, (b) $D_* \not\subseteq O'$, and (c) $D_* \not\subseteq O''$. We have $D_* \cap O' = D_* \cap (O'')^c$, hence $D_* \cap O'$ is closed, and it is disjoint of the closed set $(O')^c$. It follows that $d_\infty(D_* \cap O',(O')^c) > 0$. Similarly, $d_\infty(D_* \cap O'',(O'')^c) > 0$. Since $\lim_{k \to \infty} d_H(D_k,D_*) = 0$, It follows that the sets $O'$ and $O''$ satisfy (a)--(c) w.r.t.~$D_k$ (instead of w.r.t.~$D_*$), for every $k$ sufficiently large. This implies that for every such $k$, the set $D_k$ is not connected, contradicting Lemma~\ref{lemma:2a}. \end{proof} \bigskip For every $k \in \dN$, the set $D_k$ has a connected component whose projection on the first coordinate is $[0,1]$. In particular, for every $t \in [0,1]$ there is $x_{k,t} \in X$ such that $(t,x_{k,t})\in D_k$. Since $X$ is compact, the sequence $(x_{k,t})_{k \in \dN}$ has a converging subsequence. Since $\lim_{k \to \infty} d_H(D_K,D_*) = 0$, it follows that there is $x_t \in X$ such that $(t,x_t) \in D_*$. Therefore, the projection of $D_*$ on the first coordinate is $[0,1]$. It follows that the connected component of $C_f$ that contains $D_*$ satisfies the property that its projection on the first coordinate is $[0,1]$, contradicting the assumption in the proposition. \subsection{Proof of Proposition~\ref{prop:2}} We start by defining a continuous function $g : ([0,1] \times X) \to [-1,1]$ that satisfies the following properties (see Figure~\arabic{figurecounter}, where $C_f$ has six connected components): \begin{itemize} \item $g \equiv 1$ on $B_1 := (\{0\} \times X) \cup (C_f \cap O_0)$. \item $g \equiv -1$ on $B_{-1} := (\{1\} \times X) \cup (C_f \cap O_1)$. \end{itemize} \bigskip \centerline{ \includegraphics{browder_fig.2} \ \ \ \ \ \ \ \ \ \ \includegraphics{browder_fig.3}} \vspace{0.2truecm} \centerline{{Figure \arabic{figurecounter}: The sets $C_f$ (dark), $O_0$ and $O_1$ (grey) in Part A;}} \centerline{{the sets $B_1$, $B_{-1}$ (dark) in Part B.}} \addtocounter{figurecounter}{1} \bigskip The sets $B_1$ and $B_{-1}$ are disjoint. As in the proof of Lemma~\ref{lemma:2}, $C_f\cap O_0 = C_f \cap (O_1)^c$ is closed, and similarly, $C_f \cap O_1$ is closed. It follows that $B_1$ and $B_{-1}$ and closed, and hence by Titze's Extension Theorem such a function $g$ exists, for example, \[ g(t,x) := \left\{ \begin{array}{lll} 1, & \ \ \ \ \ & (t,x) \in B_1,\\ -1, & \ \ \ \ \ & (t,x) \in B_{-1},\\ \frac{d_\infty\bigl((t,x),B_{-1}\bigr) - d_\infty\bigl((t,x),B_1\bigr)}{d_\infty\bigl((t,x),B_{-1}\bigr) + d_\infty\bigl((t,x),B_1\bigr)}, & & \hbox{Otherwise}. \end{array} \right. \] We argue that for every $\ep > 0$ sufficiently small, $t + \ep g(t,x) \geq 0$ for every $(t,x) \in [0,1] \times X$. Indeed, since $g$ is continuous over the compact set $[0,1] \times X$, it is absolutely continuous. Hence, and since $g \equiv 1$ on $\{0\} \times X$, it follows that there is $\ep > 0$ such that $g(t,x) \geq 0$ whenever $t \leq \ep$. This implies that if $t \leq \ep$ then \[ t + \ep g(t,x) \geq t \geq 0, \] and if $t \geq \ep$, then since $g(t,x) \geq -1$ we have \[ t + \ep g(t,x) \geq t - \ep \geq 0. \] Analogously, for every $\ep > 0$ sufficiently small $t + \ep g(t,x) \leq 1$ for every $(t,x) \in [0,1] \times X$. Let $\ep > 0$ be sufficiently small so that $t + \ep g(t,x) \in [0,1]$ for every $(t,x) \in [0,1] \times X$. Consider the function $F : ([0,1] \times X) \to ([0,1] \times X)$ defined by \[ F(t,x) := \bigl(t + \ep g(t,x), f(t,x)\bigr), \ \ \ \forall (t,x) \in [0,1] \times X. \] The function $F$ is continuous, and by the choice of $\ep$ its range is indeed $[0,1] \times X$. We argue that $F$ has no fixed point. Indeed, if $(t^*,x^*)$ is a fixed point of $F$, then \[ t^* = t^* + \ep g(t^*,x^*), \ \ \ x^* = f(t^*,x^*). \] This implies that $g(t^*,x^*) = 0$. Since $g$ attains the values 1 and $-1$ on $C_f$, it follows that $(t^*,x^*) \not\in C_f$. On the other hand, since $x^* = f(t^*,x^*)$, we have $(t^*,x^*) \in C_f$, a contradiction. \section{Extensions} \label{section:discussion} We proved Theorem~\ref{theorem:browder} when $X = [0,1]^n$. Our proof holds whenever $X$ is a nonempty, convex, and compact subset of a locally convex metrizable topological vector space. The only part of the proof that needs to be adapted for this extension is the definition of $\calT_k$. Since $[0,1] \times X$ is compact and metrizable, for every $k \in \dN$ there is a finite collection $(T_{k,l})_{l=1}^{L_k}$ of open sets with diamater smaller than $\frac{1}{k}$ that covers $X$. We can assume furthermore that each open set $T_{k+1,l}$ is a subset of one of the sets $(T_{k,l})_{l=1}^{L_k}$. We then define $\calT_k$ to be the collection of closures of $(T_{k,l})_{l=1}^{L_k}$, for each $k \in \dN$. We note that Browder's (1960) proof using the fixed point index implies Theorem~\ref{theorem:browder} when $X$ is a nonempty, convex, and compact subset of a locally convex topological vector space (but not necessarily metrizable). \bigskip In Theorem~\ref{theorem:browder}, the parameter set is $[0,1]$. One may wonder whether the theorem remains valid for more general parameter sets. The answer is positive. We here illustrate this extension for the parameter set $[0,1]^2$. Let $f : ([0,1]^2 \times X) \to X$ be a continuous function, where $X = [0,1]^n$, and let $\varphi : [0,1] \to [0,1]^2$ be a continuous and surjective function (e.g., the Peano's curve (Peano, 1890)). The function $h := f \circ (\varphi,{\mathrm Id}_X) : [0,1] \times X \to X$ is a composition of two continuous functions, hence continuous, and by Theorem~\ref{theorem:browder} the set $C_h$ has a connected component, denoted $B$, whose projection on the first coordinate is $[0,1]$. But then the set $\{(\varphi(t),x) \colon (t,x) \in B\}$ is a connected component of $C_f$ whose projection on the first coordinate is $[0,1]^2$. Note that this construction is valid whenever the parameter set $Y$ possesses a space-filling curve, namely, there is a continuous and surjective function $\varphi : [0,1] \to Y$. Recall that the Hahn-Mazurkiewicz Theorem (e.g., Willard, 2012, Theorem~31.5) states that a space possesses a space-filling curve if and only if it is compact, connected, locally connected, and second-countable. One example of a set that does not possess a space-filling curve is the set $C_f$ in Example~\ref{example:1}.

https://arxiv.org/abs/2107.14079

Density of binary disc packings: lower and upper bounds

We provide, for any $r\in (0,1)$, lower and upper bounds on the maximal density of a packing in the Euclidean plane of discs of radius $1$ and $r$. The lower bounds are mostly folk, but the upper bounds improve the best previously known ones for any $r\in[0.11,0.74]$. For many values of $r$, this gives a fairly good idea of the exact maximum density. In particular, we get new intervals for $r$ which does not allow any packing more dense that the hexagonal packing of equal discs.

\section{Introduction} A {\em disc packing} (or {\em circle packing}) is a set of interior-disjoint discs in the Euclidean plane. Its {\em density} $\delta$ is the proportion of the plane covered by the discs: $$ \delta:=\limsup_{k\to \infty}\frac{\textrm{area of the square $[-k,k]^2$ covered by discs}}{\textrm{area of the square $[-k,k]^2$}}. $$ If all the discs have the same radius, it has been proven by T\'oth \cite{FT43} (see also \cite{CW10}) that the density is at most $$ \delta_1:=\frac{\pi}{2\sqrt{3}}\approx 0.9069, $$ reached for the so-called {\em hexagonal compact packing}, where discs are centered on a triangular grid (of size twice the disc radius). What about more sizes of discs? In particular, what about the maximal density of {\em binary disc packings}, that is, packings with discs of two different sizes? Up to scaling, one can always assume that the largest disc has radius $1$ and the smallest one has radius $r\in (0,1)$. We then denote by $\delta(r)$ the maximal density of packings by discs of radius $1$ and $r$. Can we find the exact value of $\delta(r)$ for each $r\in(0,1)$? The first decisive step was done by Blind \cite{Bli69}, who proved that $\delta(r)=\delta_1$ for $r\geq r_B$ where $$ r_B:=\sqrt{\frac{7\tan\tfrac{\pi}{7}-6\tan\tfrac{\pi}{6}}{6\tan\tfrac{\pi}{6}-5\tan\tfrac{\pi}{5}}}\approx 0.74299. $$ In other words, binary disc packings cannot achieve higher densities than packings of equal discs when the disc sizes are too close. This yields the exact maximal density over the whole interval $[r_B,1)$. Besides this, to the best of our knowledge, there are only $9$ specific ratios for which the exact value of $\delta(r)$ is known. They are the ratios which allow {\em triangulated binary packings}, that is, packings with two sizes of discs whose contact graphs are triangulated (the vertices of the contact graph of a packing are the disc center, and the edges connect the centers of tangent discs). These ratios are algebraic numbers which have been characterized in \cite{Ken06}. The maximal density for each of them is proven in \cite{BF22} (see also \cite{Hep00,Hep03,Ken04} for the first cases). If we cannot obtain an exact value for $\delta(r)$, can we find lower and upper bounds? In particular, for which $r$ do we have $\delta(r)>\delta_1$, that is, which ratios allow binary disc packings which achieve higher densities than packings of equal discs? This is a question that is of particular interest in materials science because the maximization of density seems to be a criterion for the formation of materials (due to attractive forces of the van der Waals type). A ratio $r$ such that $\delta(r)>\delta_1$ thus suggests the possibility to obtain new materials by combining atoms or nanoparticles within this ratio. A striking example is given by the $4$ binary assemblies of nanoparticles whose synthesis is explained in \cite{PDKM15}: they all correspond faithfully to one of the $9$ triangulated binary packings that have been proven to maximize the density. To obtain a lower bound, it suffices to find a ``good'' packing. For example, if $r$ is small enough, namely $r\leq \tfrac{2}{\sqrt{3}}-1\approx 0.1547$, we can simply insert a small disc in each hole of a hexagonal compact packing of large discs to get $\delta(r)>\delta_1$. More interesting, we can get lower bounds over a whole interval of ratios by continuously modifying a ``good'' packing so that the density decreases as little as possible. A particularly efficient way to proceed is the so-called ``flipping and flowing method'', first used by T\'oth \cite{FT64}, see also \cite{CP19,CG20}. In Section~\ref{sec:lower} we detail lower bounds obtained in this way. These lower bounds seem to be mostly ``folk'' (see, {\em e.g.}, \cite{FJFS20,KenWeb}), but we provide here a SageMath worksheet \cite{sage} (the file \verb+lower_bounds.sage+ in supplementary materials) which give explicitly the transformations as well as the corresponding densities. They are depicted in Fig.~\ref{fig:density} (green curve). As a corollary, we get: \begin{corollary} \label{cor:lower_bound} One has $\delta(r)>\delta_1$ for any $r$ in $(0,a_1)\cup (a_2,a_3)\cup (a_4,a_5)$ where $a_1\approx 0.4378$, $a_2\approx 0.5165$, $a_3\approx 0.5510$, $a_4\approx 0.6276$ and $a_5\approx 0.6456$ are algebraic number whose minimal polynomials are in the file \verb+lower_bounds.sage+. \end{corollary} \begin{figure}[hbt] \centering \includegraphics[width=\textwidth]{density_plot.pdf} \caption{ Lower bound (green curve) and upper bound (red curve) on $\delta(r)$. The blue dots and lines indicate where lower and upper bounds coincide (the dots correspond to the $9$ triangulated packings, the lines to the intervals $I_i$'s in Cor.~\ref{cor:upper_bound}). The best previous upper bound is indicated by the black dotted curve. Labels $r_1$ through $r_9$, $r_b$ and $r_c$ denote disc ratios which appear in the text and are given in Table~\ref{tab:ratios}. Zooms can be found in Fig.~\ref{fig:J34}, \ref{fig:J23}, \ref{fig:J12} and \ref{fig:J0}. } \label{fig:density} \end{figure} \begin{table} \centering \begin{tabular}{|c|c|r|} \hline ratio & approx. value & minimal polynomial\\ \hline $r_1$ & $0.6375559772$ & $x^4 \textrm{-} 10x^2 \textrm{-} 8x \textrm{+} 9$\\ $r_a$ & $0.6199144044$ & $x^4 \textrm{-} 12x^3 \textrm{-} 2x^2 \textrm{+} 4x \textrm{+} 1$\\ $r_2$ & $0.5451510421$ & $x^8 \textrm{-} 8x^7 \textrm{-} 44x^6 \textrm{-} 232x^5 \textrm{-} 482x^4 \textrm{-} 24x^3 \textrm{+} 388x^2 \textrm{-} 120x \textrm{+} 9$\\ $r_3$ & $0.5332964167$ & $8x^3 \textrm{+} 3x^2 \textrm{-} 2x \textrm{-} 1$\\ $r_4$ & $0.4142135624$ & $x^2 \textrm{+} 2x \textrm{-} 1$\\ $r_5$ & $0.3861061049$ & $9x^4 \textrm{-} 12x^3 \textrm{-} 26x^2 \textrm{-} 12x \textrm{+} 9$\\ $r_b$ & $0.3691023862$ & $x^3 \textrm{-} 5x^2 \textrm{-} x \textrm{+} 1$\\ $r_6$ & $0.3491981862$ & $x^4 \textrm{-} 28x^3 \textrm{-} 10x^2 \textrm{+} 4x \textrm{+} 1$\\ $r_7$ & $0.2807764064$ & $2x^2 \textrm{+} 3x \textrm{-} 1$\\ $r_c$ & $0.2168453354$ & $x^4 \textrm{-} 4x^3 \textrm{-} 2x^2 \textrm{-} 4x \textrm{+} 1$\\ $r_8$ & $0.1547005384$ & $3x^2 \textrm{+} 6x \textrm{-} 1$\\ $r_9$ & $0.1010205144$ & $x^2 \textrm{-} 10x \textrm{+} 1$\\ \hline \end{tabular} \label{tab:ratios} \caption{Values of disc ratios appearing in Fig.~\ref{fig:density}, \ref{fig:flow841}, \ref{fig:flow873} and \ref{fig:flow9cba}.} \end{table} Finding upper bounds is more challenging. For a given $r$, it indeed amouts to proving that among the uncountably many packings of discs of size $1$ and $r$, none has density larger than the claimed upper bound. A first milestone was the proof in \cite{Flo60} that $\delta(r)$ is less than the density inside a triangle with mutually tangent discs of size $1$, $r$ and $r$ centered on its vertices,see Fig.~\ref{fig:density}, bottom center. This upper bound was later on enhanced for $r\geq 0.6735$ in \cite{Bli69}, by proving that $\delta(r)$ is less than the density inside the union of a regular heptagon and a regular pentagon, with the heptagon (resp. pentagon) being circumscribed to a large disc (resp. small disc), see Fig.~\ref{fig:density}, bottom right (the above mentioned constant $r_B$ is the value of $r$ for which this bound reaches $\delta_1$). In Section~\ref{sec:upper}, we explain how we obtain upper bounds by enhancing the computer-aided method used in \cite{BF22} to prove the maximal density of the $9$ triangulated packings. In a nutshell, the interval $(0,1)$ of the possible disc size ratio is divided into sufficiently many small intervals on which a program (the \verb-C++- code is provided in the supplementary materials) searches by dichotomy an upper bound on the maximum density. We have tried to make this paper readable as independently of \cite{BF22} as possible by recalling the main lines of the method used in \cite{BF22} (Subsec.~\ref{sec:rappel}). It nevertheless relies on many technical details of \cite{BF22} that are omitted here and is therefore not self-contained. The obtained upper bounds improve the previous ones for any $r\in[0.11,r_B)$ but the $8$ values in this interval which allow a triangulated binary packing. They are depicted in Fig.~\ref{fig:density} (red curve). As a corollary, we get: \begin{corollary} \label{cor:upper_bound} One has $\delta(r)=\delta_1$ for any $r$ in $$ \underbrace{[0.4445, 0.4532]}_{I_1} \cup \underbrace{[0.4917, 0.5145]}_{I_2} \cup \underbrace{[0.5666, 0.6270]}_{I_3} \cup \underbrace{[0.6468,1)}_{I_4}, $$ where the endpoints of the $I_i$'s are exact numerical values obtained by computer. \end{corollary} Combining the intervals given in Cor.~\ref{cor:lower_bound} and \ref{cor:upper_bound} answers whether a given ratio $r$ allows a binary packing more dense than the hexagonal compact packing of equal discs in more than $93\%$ of the cases. We have deliberately called both these results ``corollaries'' to emphasize the fact that the main result of this paper are the lower and upper bounds depicted in Fig.~\ref{fig:density}, which unfortunately cannot be stated in the form of a classical theorem. The feasibility of closing the gap between the lower and upper bounds, is discussed in Section~\ref{sec:discussion}. \section{Lower bounds by flipping and flowing} \label{sec:lower} \subsection{Principle} To date, all disc packings that have been proven to maximize density (the $9$ binary ones and the hexagonal compact packing of equal discs) are found to have a triangulated contact graph, i.e. maximum number of contacts between discs \cite{FT43,BF22,Fer19,Bli69}. This backs up the rule of thumb that the more contact between discs in a packing, the denser it is. However, a triangulated contact graph is only possible for some very particular sizes of discs. This is where {\em flipping and flowing} comes into play. The principle is, starting from a particularly dense packing, to continuously modify the ratio of disc sizes while trying to keep as much contacts between discs as possible, in the hope of decreasing the density as little as possible. The term ``flowing'' refers to the continuous modification of disc sizes. The term ``flipping'' refers to the particular (but frequent) case where the transformation connects two packings whose graphs are triangulated and differ by one or more {\em flips}, i.e. a diagonal of a quadrilateral is replaced by the other diagonal. \subsection{The flows} Figures \ref{fig:flow841}--\ref{fig:flow9cba} describe flows between (or from) especially dense packings. The ratios $r_i$'s and $a_i$'s are algebraic numbers whose exact values can be found in the file \verb+lower_bounds.sage+. These packings are all triangulated, except those for $r\in\{r_a,r_b,r_c\}$ in Fig.~\ref{fig:flow9cba}. They are all periodic and described by their fundamental domains. The density is depicted (red curve), with the horizontal axe corresponding to the density $\delta_1$ of the hexagonal compact packing. This yields the lower bound depicted in Fig.~\ref{fig:density}. \begin{figure}[hbt] \centering \includegraphics[scale=0.047]{flow841.pdf} \caption{ The flow $0\leftrightarrow r_8\leftrightarrow r_4 \leftrightarrow r_1 \leftrightarrow 1$. } \label{fig:flow841} \end{figure} \begin{figure}[hbt] \centering \includegraphics[scale=0.047]{flow873.pdf} \caption{The flow $0\leftrightarrow r_8\leftrightarrow r_7 \leftrightarrow r_3 \leftrightarrow 1$. The white discs, not in the fundamental domain, are depicted to increase clarity. } \label{fig:flow873} \end{figure} \begin{figure}[hbt] \centering \includegraphics[scale=0.047]{flow96.pdf} \caption{The flow $0\leftrightarrow r_9\leftrightarrow r_6 \leftrightarrow 1$.} \label{fig:flow96} \end{figure} \begin{figure}[hbt] \centering \includegraphics[scale=0.047]{flow5.pdf} \caption{ Flowing around $r_5$. Among the numerous way to flow for $r<r_5$, this is the one which seems to decrease the least the density. } \label{fig:flow5} \end{figure} \begin{figure}[hbt] \centering \includegraphics[scale=0.047]{flow2.pdf} \caption{ Flowing around $r_2$ (the fundamental domain tiles like scales). The flow can be extended to $r=1$ but we get only $6$ out of $7$ discs of the hexagonal compact packing (the missing one is marked by an X). } \label{fig:flow2} \end{figure} \begin{figure}[hbt] \centering \includegraphics[scale=0.047]{flow9cba.pdf} \caption{The flow $r_9\leftrightarrow r_c \leftrightarrow r_b \leftrightarrow r_a \leftrightarrow 1$.} \label{fig:flow9cba} \end{figure} \begin{remark} One checks that when the flow on the right of $r_7$ and the one on the left of $r_6$ cross, for $r\approx 0.3154$, the density is higher about $0.0076\%$ than $\tfrac{\pi}{2\sqrt{3}}$ (it is important for Cor.~\ref{cor:lower_bound}). The bottom of the green valley between $r_7$ and $r_6$ in Fig.~\ref{fig:density} is thus (very slightly) above the horizontal dotted line. For this $r$, the two flows yield packings with the same density but different structures (Fig.~\ref{fig:between_r6_r7}). \end{remark} \begin{figure}[hbt] \centering \includegraphics[width=0.8\textwidth]{between_r6_r7.pdf} \caption{ Two packings with the same density for $r\approx 0.3154$. } \label{fig:between_r6_r7} \end{figure} \subsection{Interstitials packings} For $r\leq r_8$, a small disc can fit in the holes of a hexagonal compact packing of large discs. To obtain the lower bound depicted in Fig.~\ref{fig:density} for $r\leq r_8$, we simply put as much as possible discs of a hexagonal compact packing of small discs inside the holes of a hexagonal compact packing of large discs (the center of each hole is the center of a small disc). This is not optimal (some improvements can be found in \cite{UST04}). At least, this yields $\delta(r)>\delta_1$ over $(0,r_8)$. \subsection{Computation} To compute the density along a flow, we use the fact that the contacts between discs in the packings yield quadratic equations in the coordinates of disc centers and the ratio $r$. This form a polynomial system which has to be of dimension at least $1$ to allow $r$ to vary and at most $1$ to have as much contacts as possible. Solving this system allows to compute the positions of the discs and the density as functions of the ratio $r$. The complete calculations are provided in the file \verb+lower_bound.sage+. The main tool is the simple function \verb+stick((x1,y1,r1),(x2,y2,r2),r3)+, which takes the coordinates \verb+x1,y1+ of a disc $D_1$ of radius \verb+r1+, the coordinates \verb+x2,y2+ of a disc $D_2$ of radius \verb+r2+, a real number \verb+r3+ and returns the coordinates \verb+x3,y3+ (if they exist) of the disc $D_3$ of radius \verb+r3+ which is exteriorly tangent to $D_1$ and $D_2$ such that $D_1$ sees $D_3$ on the left of $D_2$. \begin{figure}[hbt] \centering \includegraphics[width=0.7\textwidth]{flow_ex.pdf} \caption{ Disc numbering to analyze flows. } \label{fig:flow_ex} \end{figure} Consider, for example, the periodic binary packing whose fundamental domain is depicted on Fig.~\ref{fig:flow_ex}, left. Discs are numbered and we arbitrarily set the positions of the two first discs such that they are tangent: \begin{verbatim} d0=(0,0,1) d1=(2,0,1) \end{verbatim} We then stick one by one the three following discs: \begin{verbatim} d2=stick(d0,d1,r) d3=stick(d2,d1,1) d4=stick(d2,d3,1) \end{verbatim} This allows to compute the density in the fundamental domain: \begin{verbatim} d=(2+2*r^2)/((d4[1]+(d3[1]-d1[1]))*d1[0]) \end{verbatim} This yields for the density along the flow between $r_4$ and $r_1$ the expression $$ \frac{\pi(r^2 + 1)(r + 1)^4}{16(r + 2)\sqrt{r+2}r\sqrt{r}}. $$ All the considered cases are similar, except two which are slightly more complicated. The first case is the periodic binary packing whose fundamental domain is depicted on Fig.~\ref{fig:flow_ex}, right. It corresponds, in the previous subsection, to the flow between $r_6$ and $1$. In this case, it is not possible to describe the packing by sticking discs one by one to two previous discs. To get around this problem, we define a multivariate polynomial ring over $\mathbb{Q}$ whose variables are the coordinates of the disc centers, the disc ratio \verb+r+ and the density \verb+d+: \begin{verbatim} K.<x1,y1,x2,y2,x3,y3,x4,y4,x5,y5,r,d>=QQ[] \end{verbatim} We then add the equations which describe the disc contacts or symmetries of the packing (each polynomial must be equal to zero): \begin{verbatim} eqs=[y1, x1-2*x2, x1+x2-(x3+x4+x5), y1+y2-(y3+y4+y5), x3^2+y3^2-(1+r)^2, (x4-x3)^2+(y4-y3)^2-(r+r)^2, (x4-x5)^2+(y4-y5)^2-(r+r)^2, (x3-x5)^2+(y3-y5)^2-(r+r)^2, (x4-x1)^2+(y4-y1)^2-(1+r)^2, (x2-x5)^2+(y2-y5)^2-(1+r)^2, (x1-2*x3)^2+(y3+y3)^2-(r+r)^2, d*x1*y2-(1+6*r^2)] \end{verbatim} We then compute the ideal defined by these equations and eliminate all the variables but \verb+r+ and \verb+d+: \begin{verbatim} I=ideal(K,eqs) J=I.elimination_ideal([x1,y1,x2,y2,x3,y3,x4,y4,x5,y5]) P=QQ[r,d](J.gens()[0]) \end{verbatim} This yields a polynomial of degree $9$ in \verb+r+ and $6$ in \verb+d+. Actually, we can remove a factor which would correspond to a density larger than $1$. We get an irreducible polynomial of degree $2$ in \verb+d^2+ whose coefficients are polynomials in \verb+r+. This allows to get a closed-form expression for the density along the flow between $r_6$ and $1$: $$ \frac{\pi(6 r^2 + 1)\sqrt{47 r^4 + 84 r^3 + 54 r^2 + 12 r + 3 \textrm{$-$} (7 r^3 + 13 r^2 + 9 r + 3 ) \sqrt{45r^2 \textrm{$-$} 6r \textrm{$-$} 3}}}{\sqrt{6}(r^4 + 12 r^2 + 12 r + 3)}. $$ The second case corresponds, in the previous subsection, to the flow on the left of $r_5$. It is similar, except that we eventually get a polynomial in degree $6$ in \verb+d^2+: we cannot derive a closed-form expression, but an implicit plot is possible. \begin{remark} The ``algebraic'' method by ideal elimination can actually be used for all the flows. The more ``pedestrian'' method of adding the discs one by one leads however to much faster calculations in practice. \end{remark} \section{Upper bounds via localizing potentials} \label{sec:upper} \subsection{Checking an upper bound for a fixed ratio} \label{sec:rappel} Given a ratio $r$ and a candidate upper bound $\delta$ on the maximum density, we want to check whether $\delta(r)\leq\delta$ or not. The strategy used is the one in \cite{BF22}, which resembles the one used by Hales to prove the Kepler conjecture, nicely exposed in \cite{Lag02}. Here we will only sketch this strategy and omit all technical details that are not strictly necessary to understand what follows. We refer to \cite{BF22} for the complete description (which is unfortunately necessary to fully understand the \verb-C++- code provided in the supplementary materials). Given a packing, consider a triangulation of its disc center (we used the Fejes-Moln\'ar triangulation \cite{FM58} for its suitable properties). The density inside the triangles, that is, the proportion of the triangle area covered by the discs of the packing, can greatly vary from one triangle to the other. In particular, it may be larger than the candidate upper bound $\delta$ on the maximum density for any packing. The idea is to prove that whenever a triangle is too dense, there are necessarily neighbor triangles whose densities are low enough so that the density drops below $\delta$ once averaged over all these triangles. More precisely, the density of each triangle is distributed among its three vertices, and it is proven that no matter how a vertex of the triangulation of a packing is surrounded, the average of the densities distributed to this vertex by the triangles to which it belongs is at most $\delta$. Key parameters appear to be the so-called ``base vertex potentials'' $V_{111}$, $V_{11r}$, $V_{r1r}$, $V_{1r1}$, $V_{1rr}$ and $V_{rrr}$. Each $V_{ijk}$ specifies, in a triangle whose vertices are center of mutually tangent discs of radii $i$, $j$ and $k$, the amount of density distributed to the center of the disc of radius $j$. Since the density of such a triangle is determined and since there are $4$ different such triangles (depending on the radii $111$, $11r$, $1rr$ or $rrr$ of the discs centered on vertices), two of the $6$ base vertex potentials have to be chosen and the others follow. This choice has a strong influence on whether the maximum density can be proven to be less than $\delta$. This choice turned out to be relatively simple to make in each of the 9 ratios considered in \cite{BF22}. But here the ratio can be any, so it is necessary to automate this choice in a satisfactory way. To explain how this choice is made, we need to mention two other parameters that appear in \cite{BF22}, namely $m_1$ and $m_r$. The base vertex potential $V_{ijk}$ is indeed defined only in triangles with mutually tangent discs. For other triangles, $V_{ijk}$ grows linearly with the angle in $j$ with a proportionality factor $m_j$. When the angle in $j$ increases, the triangle tends to be less dense and can ``absorb'' more density from its neighbors, which encourages taking large values of $m_j$ and $V_{ijk}$. However, the total density that each triangle can absorb cannot is limited, which in turns limits the value that $m_j$ and $V_{ijk}$ can take. This can be formalized by a list of inequalities on $m_1$, $m_r$ and two of the $V_{ijk}$'s (since the $4$ other ones follow as already mentioned). In the program \verb+upper_bound.cpp+, we consider $V_{11r}$ and $V_{1rr}$. The inequalies are defined in the function \verb+add_vertex_positivity_constraints+ of the program \verb+parameters_xy.cpp+. This yields a $4$-dimensional {\em parameter polytope} (polyhedron \verb+P+ in the function \verb+find_delta+ in \verb+upper_bound.cpp+) in which a point has to be chosen. The following automatic choice was developed through trial and error, with the goal being to prove an upper bound $\delta$ as low as possible. It is implemented in the function \verb+set_xy_generic+ in the program \verb+upper_bound.cpp+. First, it sets the ratio $m_1/m_r$ to be equal to the maximal possible value of $m_1$ divided by the maximal possible value of $m_r$. This amounts to intersecting the parameter polytope with a hyperplane to get a new polytope. We then take either the barycenter of the vertices of this new polytope if $r>0.55$, or a vertex of it which minimizes $m_1$ otherwise. This defines the parameters $V_{11r}$ and $V_{1rr}$. We do not consider the values $m_1$ and $m_r$ it defines because our program deals with rational polytopes whereas $r$ can be not rational (actually, it can even be an interval as we shall later see). Instead, we keep only the numerical values of $V_{11r}$ and $V_{1rr}$ and then proceed as in \cite{BF22} to compute $m_1$ and $m_r$ (as well as the parameter $\varepsilon$, not mentioned here) which will decide exactly whether $\delta$ is a suitable upper bound or not. \subsection{Finding an upper bound for an interval of ratios} The previous subsection explained how to check a given upper bound for a given ratio. But the goal of this paper is to find an upper bound as good as possible for any ratio. This yields two problems: \begin{enumerate} \item we have no candidate upper bound on the maximal density; \item we have a continuum of ratios to consider. \end{enumerate} Assuming $r$ is fixed, the first problem can be fixed by dichotomy on the candidate density. Namely, we maintain two variables $\delta_\textrm{inf}$ and $\delta_\textrm{sup}$, such that the proof that we have an upper bound on the exact maximal density succeeds for $\delta_\textrm{sup}$ but fails for $\delta_\textrm{inf}$. We start with $\delta_\textrm{inf}$ slightly less than $\delta_1$ and $\delta_\textrm{sup}$ equal to the Florian upper bound. At each step, we check whether the proof works for $\tfrac{1}{2}(\delta_\textrm{inf}+\delta_\textrm{sup})$ and update $\delta_\textrm{inf}$ and $\delta_\textrm{sup}$ accordingly. We stop when $\delta_\textrm{sup}-\delta_\textrm{inf}$ is smaller than a fixed precision, namely $0.0001$. We finally output $\delta_\textrm{sup}$: it is an upper bound on the maximal density (and the best that we get by our method, up to the fixed precision). This is done in the function \verb+find_delta+ in \verb+upper_bound.cpp+. To fix the second problem, we can subdivide $(0,1)$ is many small intervals and rely on interval arithmetic to consider each small interval as a value for $r$. The smaller the intervals, the better the precision, thus the better the upper bound on the maximal density. We thus want to subdivide $(0,1)$ into intervals as small as possible, with the limiting factor being the computation time. Actually, we found it more relevant to compute upper bound only for regularly spaced discrete values of $r$ with maximal precision. This indeed gives a fair idea of which bounds could be obtained because the maximal density is quite regular: \begin{proposition} \label{prop:regularity} For $x<y$ in $[0,1]$, the maximal density satisfies $$ \frac{|\delta(y)-\delta(x)|}{|y-x|} \leq \frac{\pi}{y^2\sqrt{3}}. $$ \end{proposition} The proof of this proposition is given in Appendix~\ref{app:regularity}. In particular, the maximal density is $\tfrac{\pi}{a^2\sqrt{3}}$-Lipshitz over any interval $[a,b]\subset(0,1]$. This makes the computation much lighter. For example, on our computer, subdividing the interval $[0.6269, 0.6469]$ takes $27 $min. for subintervals of length $0.0001$ and $1930$ min. ($32$h.) for subintervals of length $0.00001$. In comparison, discrete values of $r$ with a step of $0.0001$ takes $42$ min. and yields for these discrete points an upper bound on the maximal density which is comparable to the one obtained with intervals of length $0.00001$, in around $50$ times less computation time. The regularity allows to extend this upper bound everywhere. Fig.~\ref{fig:precision} illustrates this point. \begin{figure}[hbt] \centering \includegraphics[width=1\textwidth]{precision.pdf} \caption{ The upper bound obtained with subintervals of length $0.0001$ (resp. $0.00001$) in light gray (resp. dark gray). The local minima of the red broken line are the upper bounds obtained with the corresponding value of $r$, and the line itself is deduced from Prop.~\ref{prop:regularity}. Since connecting the minima by segments is expected to give a fair idea of the best upper bound that can be obtained by our method, we only represented these minima in Fig.~\ref{fig:density}, \ref{fig:J34}, \ref{fig:J23}, \ref{fig:J12} and \ref{fig:J0}. } \label{fig:precision} \end{figure} Table~\ref{tab:discrete_bounds} gives the step and execution time on each connected component of the complement of $I_1\cup I_2\cup I_3\cup I_4$ (the intervals on which the maximal density has been proven to be $\delta_1$). A list of values can be found in the supplementary materials (file \verb+trace_upper_bound.txt+), see also Fig.~\ref{fig:J34}, \ref{fig:J23} and \ref{fig:J12} for a zoom of Fig.~\ref{fig:density}. \begin{table}[hbt] \centering \begin{tabular}{|c|c|c|} \hline Interval & step & execution time\\ \hline $[0.6269, 0.6469]$ & $0.0001$ & 42 min.\\ $[0.5145, 0.5666]$ & $0.0001$ & 98 min.\\ $[0.4532, 0.4917]$ & $0.0001$ & 97 min.\\ $[0.106, 0.445]$ & $0.001$ & 16h.\\ \hline \end{tabular} \caption{ Interval, step between consecutive values of $r$ and execution time. } \label{tab:discrete_bounds} \end{table} \subsection{Checking $\delta(r)=\delta_1$ over $I_1$, $I_2$, $I_3$ and $I_4$} The upper bound obtained in the previous subsection suggest $\delta(r)=\delta_1$ over the intervals $I_1$, $I_2$, $I_3$ and $I_4$ defined in Corollary~\ref{cor:upper_bound}. To prove this, it is no longer enough to consider discrete values of r: we really have to use interval arithmetic. Instead of partitionning each $I_i$ in many equally small intervals, we use again the dichotomy to subdivide as little as possible. Namely, we bisect each $I_i$ while the precision does not suffice to conclude. But we no longer need to make a dichotomy on the candidate density since it is $\delta_1$. Table~\ref{tab:subintervals} gives the total number of subintervals into which the program \verb+upper_bound_HCP.cpp+ eventually divided each interval $I_i$ as well as the execution time on our Laptop (i5-7300U CPU 2.60GHz). \begin{table}[hbt] \centering \begin{tabular}{|c|c|c|} \hline Interval & subintervals & execution time\\ \hline $I_1$ & $259$ & 30 min.\\ $I_2$ & $318$ & 20 min.\\ $I_3$ & $379$ & 32 min.\\ $I_4$ & $845$ & 20 min.\\ \hline \end{tabular} \caption{ Interval, number of subintervals and execution time. } \label{tab:subintervals} \end{table} \section{Discussion} \label{sec:discussion} Figures~\ref{fig:J34}, \ref{fig:J23}, \ref{fig:J12} and \ref{fig:J0} depicts lower and upper bound on the maximal density over each interval of interest. The upper bounds are depicted by red points and the lower bound by a green curve. The black points indicate the smallest $\delta$ that we get if we only check that the parameter polytope is not empty (this corresponds to the first pass in the function \verb+find_delta+ in \verb+upper_bound.cpp+). These points thus give a lower bound on the best upper bound that we can hope to obtain with our method, assuming that we know how to choose the parameters in an optimal way. This lower bound is however not necessarily achievable (this is in particular the case when the black points are under the green curve, i.e., under the lower bound proved on the maximum density): the parameter polytope could be not empty without containing any parameter that allows to prove the candidate bound. Indeed, the polytope is not calculated in an exact way, whereas the verification once the parameters are chosen is exact and it is used to establish the upper bound. Nevertheless, for the ratios with a large difference between red and black points, the upper bound can probably be improved by choosing the parameters more finely. The upper bound turns out to be very sensitive to the choice of parameters in the polytope, which is therefore a delicate exercise (especially since it is done automatically here because of the very large number of different ratios considered). The black points also show that the largest intervals over which one can expect our method to prove that the maximal density is $\delta_1$ are only slightly larger than the $I_i$'s in Corollary~\ref{cor:upper_bound}, namely: $$ [0.4398, 0.4644],\qquad [0.4862,0.5182],\qquad [0.5624,0.6285],\qquad [0.6455,1]. $$ \begin{figure}[hbt] \centering \includegraphics[width=\textwidth]{J34.pdf} \caption{ Between $I_3$ and $I_4$, that is, for $r\in[0.6269, 0.6469]$. The upper bound (red points) is quite close to the lower bound (green curve). We conjecture that the lower bound is tight on this interval. For $r_1\approx 0.6375$, the lower bound reaches its maximum and has been proven to be tight \cite{BF22}. } \label{fig:J34} \end{figure} \begin{figure}[hbt] \centering \includegraphics[width=\textwidth]{J23.pdf} \caption{ Between $I_2$ and $I_3$, that is, for $r\in [0.5145, 0.5666]$. For $r_3\approx 0.5332$ and $r_2\approx 0.5451$, the lower bound reaches two local maxima and has been proven to be tight \cite{BF22}. For $r\leq r_2$, the upper bound (red points) are quite close to the lower bound (green curve). We conjecture that the lower bound is tight on this interval. For $r\geq r_2$ the lower and upper bounds diverge. The difference between black and red dots is moreover relatively large between $0.552$ and $0.558$, suggesting (but not proving) that the choice of parameters could be optimized. We do not exclude the possibility that the lower bound is not optimal, i.e. that a better flow could be defined to the right of $r_2$ (remind Fig.~\ref{fig:flow2}). } \label{fig:J23} \end{figure} \begin{figure}[hbt] \centering \includegraphics[width=\textwidth]{J12.pdf} \caption{ Between $I_1$ and $I_2$, that is, for $r\in [0.4532, 0.4917]$, there is a sort of ``mysterious island''. The lower bound is indeed $\delta_1$ over this whole interval: no packing more dense than the hexagonal compact packing is known. The difference between red and black points suggest that the choice of parameters could be optimized, but does not leave any hope of reducing the upper bound to $\delta_1$ over the whole interval. Our conjecture is that the lower bound is tight, that is, this mysterious island is an artefact of our method. This possible ``artefact'' is discussed in more detail in the text. } \label{fig:J12} \end{figure} \begin{figure}[hbt] \centering \includegraphics[width=\textwidth]{J0.pdf} \caption{ On the left of $I_1$, that is, $r\in [0.106, 0.445]$. The lower bound has been proved to be tight on its local maxima at $r_8\approx 1547$, $r_7\approx 0.2808$, $r_6\approx 0.3292$, $r_5\approx 0.3861$ and $r_4\approx 0.4142$ \cite{BF22}. We conjecture that it still holds on a neighborhood of these ratios. We also conjecture that the lower bound is tight at its local maximum at $r_b\approx 0.3691$. This is less clear around $r_c\approx 0.2168$. Upper and lower bound are quite different in the valleys betweens peaks of the lower bound. This can be due to artefacts (this is our hypothesis - see text) or to unknown dense packings. } \label{fig:J0} \end{figure} In Fig.~\ref{fig:J12} and \ref{fig:J0}, we used the term ``artefact''. By this, we mean a peak in the upper bound which is thought to be well above the exact maximal density. Such peaks appear for ratios such that the discs fit together particularly well around one disc. Indeed, the method developed in \cite{BF22} is rather ``local'': it distributes the densities between each disc and its close neighbors and bounds from above the resulting average density. However, these locally dense arrangements may not combine well on a more global scale, leading to packings in the whole plane that are actually much less dense than the obtained upper bound. Figure \ref{fig:48} illustrates the case $r=0.48$. The smaller $r$, the more often discs fit together particularly well around one disc. This explain why there are more and more peaks in the red or black curves in Fig.~\ref{fig:density} or \ref{fig:J0} when $r$ becomes small. To get around this problem, it will probably be necessary to modify the method to make it less local, i.e., to distribute the densities on a larger scale. This unfortunately makes the method even more complex. \begin{figure}[hbt] \centering \includegraphics[width=\textwidth]{rm.pdf} \caption{ Discs of radius $1$ and $0.48$. The large disc can be surrounded by discs to form a pattern which is locally quite optimal in terms of density (three leftmost patterns). This is also the case, to a lesser extent, of the small discs (three rightmost patterns). However, it seems impossible to combine these patterns to form a dense packing in the plane. We can indeed start from one of these patterns, but the more we add discs, the less the local patterns can resemble those represented here. } \label{fig:48} \end{figure}

https://arxiv.org/abs/math/0411020

Resolutions of small sets of fat points

We investigate the minimal graded free resolutions of ideals of at most n+1 fat points in general position in P^n. Our main theorem is that these ideals are componentwise linear. This result yields a number of corollaries, including the multiplicity conjecture of Herzog, Huneke, and Srinivasan in this case. On the computational side, using an iterated mapping cone process, we compute formulas for the graded Betti numbers of ideals associated to two fat points in P^n, verifying a conjecture of Fatabbi, and at most n+1 general double points in P^n.

\section[1]{Introduction}\label{s:intro} The study of sets of fat points in projective space is a classical topic that has continued to receive significant attention recently. Many researchers have investigated the Hilbert function, minimal graded free resolution, and other invariants of ideals of fat points, usually with some restriction on the sets of points they consider to make the problems more tractable. In her 2001 paper \cite{Fatabbi}, Fatabbi examines the ideals of at most $n+1$ general fat points in $\mathbb P^n$. The restriction on the number of fat points allows her to work with particularly nice monomial ideals, making it possible to do a thorough analysis of the Hilbert functions, generating sets, and, in some cases, the resolutions, of the ideals of fat points. (See also \cite{tonyi} for methods for finding the Hilbert functions of at most $n+2$ general fat points in $\mathbb{P}^n$.) We extend Fatabbi's work on resolutions in this paper. We work in projective space $\mathbb P^n$ and let the corresponding polynomial ring be $R=k[x_0,\dots,x_n]$, where $k$ is a field. Let $(P_0,a_0),\dots,(P_r,a_r)$ be fat points in $\mathbb P^n$, with $a_i$ the multiplicity of the fat point $P_i$. We assume that the $a_i$ are weakly decreasing; that is, $a_0 \ge a_1 \ge \cdots \ge a_r$. Moreover, we assume that $r+1 \le n+1$ so that we have at most $n+1$ fat points in $\mathbb P^n$. Consequently, we may make a change of coordinates, allowing us to suppose that $P_i=[0:\cdots:0:1:0:\cdots:0]$, with the 1 in the $i$-th position, for each $i$. Then the ideal $\mathfrak p_i^{a_i}$ corresponding to the fat point $(P_i,a_i)$ is \[ \mathfrak p_i^{a_i} = (x_0, \dots, x_{i-1},x_{i+1}, \dots, x_n)^{a_i},\] with $x_i$ left out. Thus the ideal of a single fat point is simply a power of an ideal generated by $n$ of the $n+1$ variables. The ideal corresponding to the set of fat points $\{(P_0,a_0),\dots,(P_r,a_r)\}$ is \[ I = \mathfrak p_0^{a_0} \cap \cdots \cap \mathfrak p_r^{a_r}.\] Throughout this paper, $I$ will denote an ideal of this form. Our paper focuses on the minimal graded free resolutions of the modules $R/I$. In Section 2, we compute the graded Betti numbers of two fat points in $\mathbb P^n$, employing a splitting Fatabbi discovered and used to find the total Betti numbers. The computation proves Fatabbi's conjecture in \cite{Fatabbi} on the graded Betti numbers of two fat points in $\mathbb P^3$. In Section 3, we investigate the free resolution of ideals of at most $n+1$ general fat points with all the same multiplicity (that is, the $a_i$ are all equal). The main result of the paper is Theorem~\ref{mainthm} in Section 4, where we show that if $I$ is the ideal of at most $n+1$ general fat points in $\mathbb P^n$, then $I$ is componentwise linear, a property Herzog and Hibi first introduced in \cite{HerzogHibi}. This has a number of interesting consequences for the graded Betti numbers of $R/I$, which we survey in Section 5. Finally, in the last section, we show that the ideal of $n+2$ fat points in $\mathbb P^n$ may not be componentwise linear, and thus Theorem~\ref{mainthm} is the best we can do. I dedicate this paper to Graham and Kay Evans on the occasion of Graham's retirement and wish them many more years of happiness. \section[2]{Two fat points}\label{s:two} We begin by analyzing the graded Betti numbers of two fat points in $\mathbb P^n$, motivated by Fatabbi's conjecture about the graded Betti numbers of two fat points in $\mathbb P^3$ in \cite{Fatabbi}. We assume throughout that $n \ge 2$ since the $n=1$ case is trivial. The main result in this section, Theorem~\ref{twobetti}, was recently proven independently by Valla \cite{VallaTwo} before we discovered it. We include our proof because our approach is different, and it illustrates the iterated mapping cone technique we shall use throughout the paper. In addition, Fatabbi and Lorenzini have another method for determining the graded Betti numbers of two fat points with different multiplicities (as well as a number of other results on small sets of general fat points in $\mathbb P^n$), and we thank them for sharing their forthcoming paper \cite{FatabbiLorenzini} with us. By a change of coordinates, we may assume that the ideal corresponding to the two fat points is \[ I = (x_1,\dots,x_n)^{a_0} \cap (x_0,x_2,\dots,x_n)^{a_1},\] where $a_0 \ge a_1$. In \cite{Fatabbi}, Fatabbi finds the minimal generating set of $I$ and then uses a splitting procedure and an inductive argument to compute the total Betti numbers of $I$. She also conjectures formulas for the graded Betti numbers of two fat points in $\mathbb P^3$ and gives some examples as evidence for her conjecture. The splitting technique Fatabbi uses first appeared in the paper of Eliahou and Kervaire \cite{EK} in which they describe the minimal free resolution of stable ideals, a class of ideals we shall discuss later. Here we introduce Eliahou and Kervaire's definition of a splittable ideal \cite{EK}. \newtheorem{definition}{Definition}[section] \begin{definition} \label{splitdef} Let $M$ be a monomial ideal with minimal generating set $G(M)$. $M$ is \textbf{splittable} if there are two nonzero monomial ideals $U$ and $V$ such that: \begin{enumerate} \item $G(I)$ is the disjoint union of $G(U)$ and $G(V)$. \item There is a splitting function from $G(U \cap V)$ to $G(U) \times G(V)$ sending $w$ to $(\phi(w),\psi(w))$ such that: \begin{enumerate} \item For all $w \in G(U \cap V)$, $w=\hbox{lcm}(\phi(w),\psi(w))$. \item If $H$ is a subset of $G(U \cap V)$, then the lcm of $\phi(H)$ and the lcm of $\psi(H)$ both strictly divide the lcm of $H$. \end{enumerate} \end{enumerate} If $U$ and $V$ satisfy the conditions above, we say that $U$ and $V$ are a \textbf{splitting} of $I$. \end{definition} The reason this is a useful notion is that we can split an ideal into simpler parts in order to find the minimal free resolution. Fatabbi proves a graded version of a result of Eliahou and Kervaire on the Betti numbers of a splittable ideal \cite{Fatabbi}: \newtheorem{proposition}[definition]{Proposition} \begin{proposition} \label{splitbetti} \emph{(Eliahou-Kervaire, Fatabbi)} Let $U$ and $V$ be a splitting of a monomial ideal $M$, and let $\beta_{i,j}(R/M)$ denote the $(i,j)$-th Betti number of $R/M$. Then for all $i$ and $j$, \[\beta_{i,j}(R/M) = \beta_{i,j}(R/U) + \beta_{i,j}(R/V) + \beta_{i-1,j}(R/U \cap V).\] \end{proposition} Fatabbi shows that the following choices of $U$ and $V$ give a splitting of the ideal $I$ of two fat points: \[ U =(x_2, \dots, x_n)^{a_0}+x_1(x_2,\dots,x_n)^{a_0-1}+\cdots + x_1^{a_0-a_1}(x_2,\dots,x_n)^{a_1} \] \[ V=x_0x_1^{a_0-a_1+1}(x_2,\dots,x_n)^{a_1-1}+x_0^2x_1^{a_0-a_1+2}(x_2,\dots,x_n)^{a_1-2} + \cdots+(x_0^{a_1}x_1^{a_0}) \] We resolve $R/U$, $R/V$, and $R/U \cap V$ separately and then use Proposition~\ref{splitbetti} to get the graded Betti numbers of $R/I$. Our method uses iterated mapping cone resolutions, and we need a lemma to show that the mapping cone resolutions are minimal at each iteration. \newtheorem{example}[definition]{Example} \newtheorem{lemma}[definition]{Lemma} \begin{lemma} \label{minmap} Let $J \subset R=k[x_0,\dots,x_n]$ be a homogeneous ideal with the regularity of $R/J$ at most $d-1$. Let $m$ be a monomial of degree $d$ not in $J$ such that $J:m=(x_{i_1},\dots,x_{i_s})$. Then the mapping cone resolution of $R/(J,m)$ is minimal. \end{lemma} \newenvironment{proof}{\noindent {\it Proof}:}{} \begin{proof} The lemma follows immediately from Lemma 1.5 in \cite{HerzogTakayama}, using Herzog and Takayama's theory of linear quotients. Alternatively, the result is easy to prove by using the long exact sequence in Tor that the short exact sequence \[ 0 \longrightarrow R/(J:m)(-d) \longrightarrow R/J \longrightarrow R/(J,m) \longrightarrow 0\] induces. \hfill $\square$ \end{proof} Frequently in the next two sections, we shall need the graded Betti numbers of an ideal of the form $(x_{i_1},\dots,x_{i_s})^d$. We can compute these using Eliahou and Kervaire's resolution of stable ideals \cite{EK} or from the Eagon-Northcott complex. \begin{lemma} \label{powerbetti} Let $M=(x_{i_1},\dots,x_{i_s})^d \subset R=k[x_0,\dots,x_n]$. Then \[ \beta_{q,q+d-1}(R/M) = {d+s-1 \choose d+q-1}{d+q-2 \choose q-1} \] for $q \ge 1$, $\beta_{00}=1$, and all other graded Betti numbers are zero. \end{lemma} We begin the process of finding the Betti numbers of $R/I$ by computing the graded Betti numbers of $R/U$. \begin{lemma} \label{ubetti} For $q \ge 1$, \[\beta_{q,q+a_0-1}(R/U) = {a_0+n-2 \choose a_0+q-1}{a_0+q-2 \choose q-1} + \sum_{i=1}^{a_0-a_1} {a_0-i+n-2 \choose a_0-i}{n-1 \choose q-1}, \] $\beta_{00}(R/U)=1$, and all other graded Betti numbers are zero. \end{lemma} \begin{proof} We begin with the minimal generators of $(x_2,\dots,x_n)^{a_0}$. The graded Betti numbers of $R/(x_2,\dots,x_n)^{a_0}$ are given by Lemma~\ref{powerbetti}, and it contributes \[ {a_0+n-2 \choose a_0+q-1}{a_0+q-2 \choose q-1} \] to the Betti numbers $\beta_{q,q+a_0}(R/U)$ for $q \ge 1$. Our goal is to add in one generator of $U$ at a time, computing an iterated mapping cone resolution. The first set of generators we add in with this process are the generators of $x_1(x_2,\dots,x_n)^{a_0-1}$. The order in which we add in generators from $x_1(x_2,\dots,x_n)^{a_0-1}$ will not matter in this case, so we pick descending lex order for specificity. For example, the first ideal we resolve is $(x_2,\dots,x_n)^{a_0}+(x_1x_2^{a_0-1})$. The ideal quotient $(x_2,\dots,x_n)^{a_0}:(x_1x_2^{a_0-1})$ is $(x_2,\dots,x_n)$, which contributes ${n-1 \choose q-1}$ to the graded Betti numbers of $R/U$ since Lemma~\ref{minmap} implies that the mapping cone resolution is minimal. We then add in the generator $x_1x_2^{a_0-2}x_3$ and continue in this way until we have exhausted the generators of $x_1(x_2,\dots,x_n)^{a_0-1}$. Next, we add in the generators of $x_1^2(x_2,\dots,x_n)^{a_0-2}$ in descending lex order and continue this process. We claim that the ideal quotient at each step is $(x_2,\dots,x_n)$. That is, suppose $U'$ is the ideal \[ U'=(x_2,\dots,x_n)^{a_0} + x_1(x_2,\dots,x_n)^{a_0-1} +\cdots + x_1^{t-1}(x_2,\dots,x_n)^{a_0-t+1} \] \[ + \hbox{ an initial lex segment of generators of } x_1^t(x_2,\dots,x_n)^{a_0-t}.\] Suppose the next monomial to be added into the ideal is $m=x_1^tx_2^{b_2}\cdots x_n^{b_n}$, where the sum of the $b_i$ is $a_0-t$. We need to compute $U':m$. Multiplying $m$ by any of $x_2,\dots,x_n$ lands one in $x_1^{t}(x_2,\dots,x_n)^{a_0-t+1}$, so $(x_2,\dots,x_n) \subseteq U':m$. On the other hand, no multiplication of $m$ by a power of $x_0x_1$ will land in $U'$, and thus $U':m=(x_2,\dots,x_n)$. Again, Lemma~\ref{minmap} implies that the mapping cone resolution is minimal. Therefore, each minimal generator of $U$ other than the minimal generators of $(x_2,\dots,x_n)^{a_0}$ contributes ${n-1 \choose q-1}$ to the graded Betti numbers of $R/U$. Each $x_1^{i}(x_2,\dots,x_n)^{a_0-i}$ has ${a_0-i+n-2 \choose a_0-i}$ minimal generators; hence the generators other than those in $(x_2,\dots,x_n)^{a_0}$ combine to contribute \[ \sum_{i=1}^{a_0-a_1} {a_0-i+n-2 \choose a_0-i}{n-1 \choose q-1} \] to the graded Betti numbers. \hfill $\square$ \end{proof} Next, we compute the graded Betti numbers of $V$. \begin{lemma} \label{vbetti} For $q \ge 1$ and $2 \le i \le a_1$, \[ \beta_{q,q+a_0}(R/V)={a_1+n-3 \choose a_1+q-2}{a_1+q-3 \choose q-1}, \] \[ \beta_{q,q+a_0+i-1}(R/V)={a_1-i+n-2 \choose a_1-i}{n-1 \choose q-1}, \] $\beta_{00}(R/V)=1$, and all other graded Betti numbers are zero. \end{lemma} \begin{proof} We resolve $R/V$ the same way we resolved $R/U$. The generators of $V$ in degree $a_0+1$ are the minimal generators of $x_0x_1^{a_0-a_1+1}(x_2,\dots,x_n)^{a_1-1}$. The Betti numbers from this portion of the ideal follow from Lemma~\ref{powerbetti}, and they are the $\beta_{q,q+a_0}(R/V)$ shown above. To resolve the rest of the ideal, we begin with the generators of $x_0^2x_1^{a_0-a_1+2}(x_2,\dots,x_n)^{a_1-2}$. We can take the generators of this ideal in any order, so we pick descending lex order again. Note that \[ x_0x_1^{a_0-a_1+1}(x_2,\dots,x_n)^{a_1-1}:x_0^2x_1^{a_0-a_1+2}x_2^{a_1-2} = (x_2,\dots,x_n); \] clearly, multiplying by any of $x_2,\dots,x_n$ lands one in $x_0x_1^{a_0-a_1+1}(x_2,\dots,x_n)^{a_1-1}$, but increasing the power of $x_0$ or $x_1$ does not help. By Lemma~\ref{minmap}, the mapping cone resolution of \[ R/(x_0x_1^{a_0-a_1+1}(x_2,\dots,x_n)^{a_1-1}+(x_0^2x_1^{a_0-a_1+2}x_2^{a_1-2}))\] is minimal. Now let \[ V' = x_0x_1^{a_0-a_1+1}(x_2,\dots,x_n)^{a_1-1}+ \cdots + x_0^{t-1}x_1^{a_0+a_1+t-1}(x_2,\dots,x_n)^{a_1-(t-1)} \] \[ + \hbox{ an initial lex segment of generators of } x_0^tx_1^{a_0-a_1+t}(x_2,\dots,x_n)^{a_1-t}.\] Suppose the next monomial to add into the ideal in the mapping cone process is \[m=x_0^tx_1^{a_0-a_1+t}x_2^{b_2}\cdots x_n^{b_n},\] where the $b_i$ sum to $a_1-t$. We compute $V':m$. It is easy to see that for $i=2,\dots,n$, $x_im \in V'$ because $x_im \in x_0^{t-1}x_1^{a_0+a_1+t-1}(x_2,\dots,x_n)^{a_1-(t-1)}$. Also, $x_0^{b_0}x_1^{b_1}m \not \in V'$ for any choices of $b_0$ and $b_1$, and thus $V':m$ $=(x_2,\dots,x_n)$. Therefore the mapping cone resolution is minimal at each iteration by Lemma~\ref{minmap}, and each generator of $V$ in degrees $a_0+2,\dots,a_0+a_1$ contributes ${n-1 \choose q-1}$ to the Betti numbers of $R/V$. Each $x_0^tx_1^{a_0-a_1+i}(x_2,\dots,x_n)^{a_1-i}$ has ${a_1-i+n-2 \choose a_1-i}$ generators, and thus the graded Betti numbers in the statement of the lemma follow. \hfill $\square$ \end{proof} The last ingredient in computing the graded Betti numbers of $R/I$ is finding the graded Betti numbers of $R/U \cap V$. It is easy to show that \[ U \cap V = x_0x_1^{a_0-a_1+1}(x_2,\dots,x_n)^{a_1}. \] From Lemma~\ref{powerbetti}, we obtain the graded Betti numbers of $R/U \cap V$, which are, for $q \ge 1$, \[ \beta_{q-1,q+a_0}(R/U \cap V)={a_1+n-2 \choose a_1+q-2}{a_1+q-3 \choose q-2},\] $\beta_{00}(R/U \cap V)=1$, and all others zero. We state the graded Betti numbers in this form, using the index $q-1$ for the syzygy, to stay consistent with the formula in Proposition~\ref{splitbetti}. \newtheorem{theorem}[definition]{Theorem} \begin{theorem} \label{twobetti} Let $I \subset R$ be the ideal of two fat points of multiplicities $a_0 \ge a_1$ in $\mathbb P^n$, where $n \ge 2$. Then for $q \ge 1$, \[ \beta_{q,q+a_0-1}(R/I) = {a_0+n-2 \choose a_0+q-1}{a_0+q-2 \choose q-1} + \sum_{i=1}^{a_0-a_1} {a_0-i+n-2 \choose a_0-i}{n-1 \choose q-1}, \] \[ \beta_{q,q+a_0+i}(R/I)={a_1-i+n-3 \choose a_1-i-1}{n-1 \choose q-1} \hbox{ for } i=0,\dots,a_1-1 ,\] $\beta_{00}(R/I)=1$, and all other graded Betti numbers are zero. \end{theorem} \begin{proof} All the formulas except the one for $\beta_{q,q+a_0}$ follow from Lemmas~\ref{ubetti} and~\ref{vbetti} and Proposition~\ref{splitbetti}. (Note the shift in the range of $i$ from Lemma~\ref{vbetti}.) The formula for $\beta_{q,q+a_0}$ is a consequence of the identity \[ {a_1+n-3 \choose a_1+q-2}{a_1+q-3 \choose q-1} + {a_1+n-2 \choose a_1+q-2}{a_1+q-3 \choose q-2} = {a_1+n-3 \choose a_1-1}{n-1 \choose q-1}.\] \hfill $\square$ \end{proof} Setting $n=3$ gives the formulas for the graded Betti numbers that Fatabbi conjectured in \cite{Fatabbi}. \begin{example}\label{twopointex}\emph{ We give the graded Betti numbers of an ideal $I$ of two fat points with multiplicities four and five in $\mathbb P^5$. To display the Betti numbers, we use the notation from the computer algebra system Macaulay 2 \cite{M2}, in which we made all of our computations for this paper. The rows are indexed such that row $d$ contains the Betti numbers $\beta_{i,i+d}$, and $\beta_{i,j}$ is in column $i$ and row $j-i$, where the rows and columns are numbered starting with zero. The graded Betti numbers of $R/I$ are:} \emph{\begin{tabular}{ccccccccc} total: & 1 & 126 & 420 & 540 & 315 & 70 &\\ 0: & 1 & . & . & . & . & . &\\ 1: & . & . & . & . & . & . &\\ 2: & . & . & . & . & . & . &\\ 3: & . & . & . & . & . & . &\\ 4: & . & 91 & 280 & 330 & 175 & 35 &\\ 5: & . & 20 & 80 & 120 & 80 & 20 & \\ 6: & . & 10 & 40 & 60 & 40 & 10 &\\ 7: & . & 4 & 16 & 24 & 16 & 4 &\\ 8: & . & 1 & 4 & 6 & 4 & 1 & \end{tabular}} \end{example} Those familiar with the resolutions of stable ideals may notice that this graded Betti diagram looks like it gives the graded Betti numbers of a stable ideal. Recall the definition of a stable ideal: If $m$ is a monomial, let $\max(m)$ be the maximum index of a variable that divides $m$; for example, $\max(x_1^3x_5^3x_7)=7$. We say that a monomial ideal $M$ is \emph{stable} if, whenever $m \in M$ and $i\le \max(m)$, then $\frac{x_i}{x_{\max(m)}}m \in M$. Stable ideals are a generalization of strongly stable ideals, which, when the characteristic of $k$ is zero, are exactly the Borel-fixed ideals. They have received significant attention because generic initial ideals are strongly stable in characteristic zero, stable ideals have convenient combinatorial properties, and Eliahou and Kervaire have computed formulas for the Betti numbers for the graded Betti numbers of stable ideals \cite{EK}. The ideal in Example~\ref{twopointex} is not stable, but its Betti numbers look suspiciously like those of a stable ideal. We shall return to this observation later in the paper. \section[3]{Fat points with the same multiplicity}\label{s:same} There are a number of directions one can proceed after analyzing the case of two fat points. One possibility is to investigate sets of more than two but at most $n+1$ general fat points, also requiring that all the fat points have the same multiplicity. Fatabbi discusses an iterative splitting procedure for computing Betti numbers of ideals of at most $n$ general fat points with the same multiplicity in $\mathbb{P}^n$ and examines the first two steps in a splitting in \cite{Fatabbi}. For some recent interesting work on double point schemes using liaison techniques, see \cite{GMS}. We approach the question differently, continuing with the mapping cone idea from Section 2. We show how to use an iterated mapping cone procedure to compute the graded Betti numbers of at most $n+1$ general double points in $\mathbb P^n$. The technique works just as well for general triple points, and we give formulas for those Betti numbers as well. Unfortunately, the bookkeeping gets complicated quickly, and there seems to be no reason to compute explicit formulas for the Betti numbers of ideals of sets of higher order fat points, particularly in light of our main result, Theorem~\ref{mainthm}. We specialize a result of Fatabbi \cite{Fatabbi} to get the minimal generating set of small sets of general fat points with the same multiplicity. \begin{proposition} \label{mingens} \emph{(Fatabbi)} Let $I = \mathfrak p_0^{a} \cap \cdots \cap \mathfrak p_r^{a}$ be an ideal of $r+1 \le n+1$ general fat points in $\mathbb P^n$, all with the same multiplicity $a$. Then $I$ is minimally generated by the union of the sets of monomials $G_0,\dots,G_a$, where \[ G_0 = \{ m=x_{r+1}^{b_{r+1}}\cdots x_n^{b_n} \, | \, m \in R_{a}\},\] and for $t=1,\dots,a$, \[ G_t = \{ m=x_0^{b_0} \cdots x_n^{b_n} \in R_{a+t} \, | \, b_i \le t \, \forall \, i=0,\dots,r \hbox{ and } \, \exists \, 0 \le u < v \le r \hbox{ with } b_u=b_v=t\}. \] \end{proposition} Thus the degree $a$ generators are all the monomials of degree $a$ involving only the variables $x_{r+1},\dots,x_n$. The higher degree generators in degrees $a+t$ have power of at least two of $x_0,\dots,x_r$ equal to $t$, and no power of $x_0,\dots,x_r$ may exceed $t$. In the case $a=2$, where we have $r+1$ double points, this means that we have minimal generators in degrees 2, 3, and 4, and they have the following form: \[ G_0 = \hbox{degree two monomials in } (x_{r+1},\dots,x_n)^2 \] \[ G_1 = \{x_ix_jx_l \, | \, 0 \le i < j \le r, \, i \not = l \not = j \} \] \[ G_2 = \{ x_0^2x_1^2,\dots,x_0^2x_r^2,x_1^2x_2^2,\dots,x_1^2x_r^2,\dots,x_{r-1}^2x_r^2 \} \] We use this characterization of the minimal generators to compute the graded Betti numbers of at most $n+1$ general double points in $\mathbb P^n$. \begin{proposition} \label{doublepoints} Let $I \subset R$ be the ideal of $P_0,\dots,P_r$, a set of at most $n+1$ general double points in $\mathbb P^n$. Then, for $q \ge 1$, the graded Betti numbers of $R/I$ are \[ \beta_{q,q+1}(R/I)={n-r+1 \choose q+1}{q \choose q-1},\] \[ \beta_{q,q+2}(R/I)= \left (n-r \right ) \sum_{i=0}^{r-1} \left (r-i \right ) {n-i-1 \choose q-1}+ \sum_{i=0}^{r-2} {r-i \choose 2}{n-i-2 \choose q-1},\] \[ \beta_{q,q+3}(R/I)={r+1 \choose 2}{n-1 \choose q-1},\] $\beta_{00}(R/I)=1$, and all other graded Betti numbers are zero. \end{proposition} \begin{proof} The Betti numbers that the monomials in $G_0$ contribute are ${n-r+1 \choose q+1}{q \choose q-1}$ by Lemma~\ref{powerbetti}. To find the contribution of the monomials in $G_1$, we do an iterated mapping cone. This time, we start with the smallest element of $G_1$ in lex order and continue in ascending lex order. (If we pick the largest first, we can get some nonminimal quadratic syzygies that cancel later in the process, and, in particular, we are not in the situation of Lemma~\ref{minmap}.) We begin by computing $(G_0):(x_{r-1}x_rx_n)$. Multiplication by any of $x_{r+1}, \dots,x_n$ gives a monomial in $(x_{r+1},\dots,x_n)^2$, and we cannot land inside $(G_0)$ by multiplying by any power of $x_0,\dots,x_r$. Thus the ideal quotient is $(x_{r+1},\dots,x_n)$, and because Lemma~\ref{minmap} shows that the mapping cone resolution is minimal, $x_{r-1}x_rx_n$ contributes ${n-r \choose q-1}$ to the graded Betti numbers of $R/I$. Let $M$ be the ideal generated by $G_0$ and the first $t$ monomials in $G_1$ in ascending lex order. Suppose $x_ix_jx_l$ is the next monomial in $G_1$ to add into the ideal. There are two cases: Without loss of generality, we may assume that either $i$ and $j$ are in $\{0,\dots,r\}$ and $l \in \{r+1,\dots,n\}$, or $i$, $j$, and $l$ are all in $\{0,\dots,r\}$. We wish to compute $M:(x_ix_jx_l)$. In the first case, where $l \in \{r+1,\dots,n\}$, we may take $0 \le i < j \le r$. Then for all $p \in \{r+1,\dots,n\}$, $x_ix_jx_lx_p \in M$ since $x_lx_p \in G_0$. Thus $(x_{r+1},\dots,x_n) \subset M:(x_ix_jx_l)$. Other variables are also in the ideal quotient, however, because every monomial in $G_1$ less than $x_ix_jx_l$ in lex order is in $M$. Therefore we also have \[(x_{i+1},\dots,x_{j-1},x_{j+1},\dots,x_r) \subset M:(x_ix_jx_l),\] since $x_{i+1}x_jx_l,\dots,x_{j-1}x_jx_l \in M$ and $x_ix_{j+1}x_l,\dots,x_ix_rx_l \in M$. It is easy to see that multiplication by no product of the other variables lands inside $M$. Hence \[ (x_{i+1},\dots,x_{j-1},x_{j+1},\dots,x_n) = M:(x_ix_jx_l).\] By Lemma~\ref{minmap}, the mapping cone resolution of $R/(M,x_ix_jx_l)$ is minimal. There are $(n-j)+$ $(j-i-1)=$ $n-i-1$ variables in the ideal quotient, and by Lemma~\ref{powerbetti}, the ideal quotient contributes ${n-i-1 \choose q-1}$ to the graded Betti numbers of $R/I$. To compute the Betti numbers arising from the monomials in this case, note that there are $n-r$ choices of $l$, and after $i$ is fixed, there are $r-i$ choices for $j > i$. Since $0 \le i \le r-1$, we have a contribution of \[ \left ( n-r \right ) \sum_{i=0}^{r-1} \left (r-i \right ) {n-i-1 \choose q-1} \] to the graded Betti numbers of $R/I$. For the other case, we may assume that $0 \le i < j < l \le r$. Again, for all $r+1 \le p \le n$, $x_ix_jx_lx_p \in M$ because $x_ix_jx_p <_{lex} x_ix_jx_l$, and thus $x_ix_jx_p \in M$. Hence $(x_{r+1},\dots,x_n) \subset$ $M:(x_ix_jx_l)$. Next, we consider the variables $x_0,\dots,x_r$, asking which we could multiply by $x_ix_jx_l$ to get a monomial divisible by the elements of $G_1$ in $M$. The elements of $G_1$ in $M$ are all the monomials of $G_1$ less than $x_ix_jx_l$ in lex order, and hence any of $x_{i+1},\dots,x_{j-1},$ $x_{j+1},\dots,x_{l-1},$ $x_{l+1},\dots,x_r$ will multiply $x_ix_jx_l$ into $M$. There is no way to multiply by a product of any of the other variables and land in $G_1 \cap M$, and thus \[(x_{i+1},\dots,x_{j-1},x_{j+1},\dots,x_{l-1},x_{l+1},\dots,x_n) = M:(x_ix_jx_l).\] By Lemma~\ref{minmap} again, the mapping cone resolution of $R/(M,x_ix_jx_l)$ is minimal, and there are $(n-l)+(l-j-1)+(j-i-1)=n-i-2$ variables in the ideal quotient. Lemma~\ref{powerbetti} implies that the monomials in this case add ${n-i-2 \choose q-1}$ to the graded Betti numbers of $R/I$. Once we fix $i$ such that $0 \le i \le r-2$, there are ${r-i \choose 2}$ ways to choose $j$ and $l$ such that $i < j < l \le r$. Hence this case contributes \[ \sum_{i=0}^{r-2} {r-i \choose 2} {n-i-2 \choose q-1} \] to the graded Betti numbers of $R/I$. Combining the two cases gives the graded Betti numbers of the form $\beta_{q,q+2}$. Finally, we add in the elements of $G_2$. Let $J$ be the ideal generated by $G_0$, $G_1$, and some subset of $G_2$. Pick $0 \le i < j \le r$ such that $x_i^2x_j^2$ is an element of $G_2$ not in $J$. We claim that $J:(x_i^2x_j^2)$ is the ideal generated by all the variables except for $x_i$ and $x_j$. Multiplying $x_i^2x_j^2$ by any $x_l$, where $i \not = l \not = j$, yields a monomial divisible by $x_ix_jx_l$, which is in $G_1$. Additionally, increasing the powers of $x_i$ and $x_j$ on $x_i^2x_j^2$ is no help for getting into $J$. Therefore each element of $G_2$ contributes ${n-1 \choose q-1}$ to the Betti numbers of $R/I$ since the mapping cone resolution is minimal by Lemma~\ref{minmap}. There are ${r+1 \choose 2}$ elements in $G_2$, which gives the formula for $\beta_{q,q+3}(R/I)$. \hfill $\square$ \end{proof} The same mapping cone technique works for at most $n+1$ general triple points in $\mathbb P^n$. In this case, the ideal has generators in degrees three through six, giving four linear strands in the resolution. We record formulas for the graded Betti numbers to give an idea of what happens going from double to triple points (in particular, the formulas are messier), but we omit the proof, which is much the same as in Proposition~\ref{doublepoints}. It is possible to simplify these formulas somewhat by factoring; we chose to leave them in this form since these expressions are the ones that arise in keeping track of the ideal quotients. \begin{proposition} \label{triplepoints} Let $I \subset R$ be the ideal of $r+1 \le n+1$ general triple points in $\mathbb P^n$. Then, for $q \ge 1$, \[ \beta_{q,q+2}(R/I) = {n-r+2 \choose q+2}{q+1 \choose q-1}, \] \[ \beta_{q,q+3}(R/I) = \sum_{i=0}^{r-1} \left (r-i \right ){n-r+1 \choose 2}{n-i-1 \choose q-1} + \sum_{i=0}^{r-2} {r-i \choose 2} \left (n-r \right ){n-i-2 \choose q-1} \] \[ + \sum_{i=0}^{r-3} {r-i \choose 3}{n-i-3 \choose q-1}, \] \[ \beta_{q,q+4}(R/I)=2{r+1 \choose 3}{n-1 \choose q-1} + {r+1 \choose 3}{n-2 \choose q-1}+(n-r){r+1 \choose 2}{n-1 \choose q-1}, \] \[ \beta_{q,q+5}(R/I)={r+1 \choose 2}{n-1 \choose q-1}. \] All other graded Betti numbers are zero (except $\beta_{00}(R/I)=1$). \end{proposition} \begin{example}\label{tripleex}\emph{Consider the ideal $I$ of four triple points in general position in $\mathbb P^5$. The Betti diagram of $R/I$ is:} \emph{\begin{tabular}{ccccccccc} total: & 1 & 61 & 203 & 264 & 156 & 35 &\\ 0: & 1 & . & . & . & . & . &\\ 1: & . & . & . & . & . & . &\\ 2: & . & 4 & 3 & . & . & . &\\ 3: & . & 27 & 84 & 96 & 48 & 9 &\\ 4: & . & 24 & 92 & 132 & 84 & 20 &\\ 5: & . & 6 & 24 & 36 & 24 & 6 & \end{tabular}} \emph{Again, the Betti diagram looks like that of a stable ideal, but the ideal is not stable. We devote the next two sections to explaining this phenomenon and its consequences. } \end{example} \section[4]{Fat point ideals and componentwise linearity}\label{s:cwl} In this section, we prove that ideals of at most $n+1$ general fat points in $\mathbb P^n$ are componentwise linear. This property has many consequences for their resolutions, which we explore in Section 5. For a homogeneous ideal $J$, let $J_{<d>}$ denote the ideal generated by all the homogeneous elements of degree $d$ in $J$. Herzog and Hibi give the following definition in \cite{HerzogHibi}: \begin{definition} \label{cwdef} Let $J$ be a homogeneous ideal. We call $J$ \textbf{componentwise linear} if $J_{<d>}$ has a $d$-linear resolution for all $d$. That is, for each $d$, $J_{<d>}$ has generators only in degree $d$, first syzygies only in degree $d+1$, etc. \end{definition} There are a number of interesting examples of componentwise linear ideals, including stable ideals and Gotzmann ideals (see \cite{HerzogHibi}). Componentwise linear ideals are a natural generalization of ideals with linear resolutions, and their importance first became apparent in a combinatorial application. In \cite{EagonReiner}, Eagon and Reiner prove that a Stanley-Reisner ideal $I_{\Delta}$ associated to a simplicial complex $\Delta$ has a linear resolution if and only if the Alexander dual $\Delta^*$ is Cohen-Macaulay. Herzog and Hibi and Herzog, Reiner, and Welker generalize this result by showing that $I_{\Delta}$ is componentwise linear if and only if $\Delta^*$ is sequentially Cohen-Macaulay, a less restrictive condition than Cohen-Macaulayness that requires a nice filtration of the module $R/I_{\Delta^*}$ in which the quotients are Cohen-Macaulay \cite{HerzogHibi, HRW}. We begin the process of showing that ideals of at most $n+1$ general fat points in $\mathbb P^n$ are componentwise linear by discussing a notion of Charalambous and Evans \cite{CE:map}. \begin{definition} \label{lexwithholesdef} Let $L$ be a lex ideal in $R=k[x_0,\dots,x_n]$, and let $d_0,\dots,d_n$ be positive integers or infinity. Let $L'$ be the ideal generated by all the minimal generators of $L$ whose degree in $x_i$ is $\le$ $d_i-1$ for all $i$. Then we call $L'$ a \textbf{lex ideal with holes}. \end{definition} \begin{example}\emph{ Let $L=(a^3,a^2b,a^2c,ab^3,ab^2c,abc^2) \subset R=k[a,b,c]$. Then $L$ is a lex ideal. Suppose $(d_0,d_1,d_2)=(\infty,3,2)$. We remove all minimal generators of $L$ whose degree in $b$ is 3 or more and whose degree in $c$ is 2 or more. That leaves us with $L'=(a^3,a^2b,a^2c,ab^2c)$, which is a lex ideal with holes.} \end{example} We cannot expect the minimal resolution of an arbitrary subideal of a lex ideal to have particularly good properties. However, Charalambous and Evans show that the resolutions of lex ideals with holes do have an especially convenient description \cite{CE:map}. \begin{theorem} \label{grahamharathm} \emph{(Charalambous-Evans)} Let $L$ be a lex ideal, and let $d_0, \dots, d_n$ be positive integers or infinity. Suppose $L'$ is the lex ideal with holes obtained by removing all minimal generators of $L$ whose power of $x_i$ is $\ge d_i$ for each $i$. Then the minimal graded free resolution of $L'$ is a subcomplex of the minimal graded free resolution of $L$. Moreover, one obtains the minimal graded free resolution of $L'$ by deleting all the syzygies in the minimal resolution of $L$ whose degree in $x_i$ exceeds $d_i-1$. \end{theorem} We refer the reader to the papers of Eliahou and Kervaire \cite{EK} and Charalambous and Evans \cite{CE:map} for discussions of the basis elements of the syzygy modules (and the degrees of the syzygies) in the minimal free resolution of a lex ideal. We remark only that if a lex ideal with holes does not have too many generators, the process of determining which syzygies survive the deletion process is easy to do by hand. Let $I = \mathfrak p_0^{a_0} \cap \cdots \cap \mathfrak p_r^{a_r}$ be the ideal of at most $n+1$ general fat points in $\mathbb P^n$ with $a_0 \ge \cdots \ge a_r$ as before. Fatabbi proves in \cite{Fatabbi} that if $t \ge 0$, the set of monomials in $I_{a_0+t}$ is \[ I_{a_0+t} = \{x_0^{b_0}\cdots x_n^{b_n} \in R_{a_0+t} \, | \, b_i \le a_0-a_i+t, i=0,\dots,r\}.\] Thus the powers of $x_0,\dots,x_r$ are restricted, and the powers of $x_{r+1},\dots,x_n$ are not. Using this characterization of the monomials in $I$ in each degree, we show that for each $d$, $I_{<d>}$ is a lex ideal with holes. \begin{proposition} \label{arelexholes} Let $I$ be the ideal of at most $n+1$ general fat points in $\mathbb P^n$. Then for all $t \ge 0$, $I_{<a_0+t>}$ is a lex ideal with holes. \end{proposition} \begin{proof} Let $\mathfrak m=(x_0,\dots,x_n)$. Then $I_{<a_0+t>}$ is generated by all the monomials in $\mathfrak m^{a_0+t}$ except those with power of $x_i$ exceeding $a_0-a_i+t$ for $i=0,\dots,r$. Since $\mathfrak m^{a_0+t}$ is a lex ideal, it follows immediately that $I_{<a_0+t>}$ is a lex ideal with holes. \hfill $\square$ \end{proof} We can now state our main result. \begin{theorem} \label{mainthm} Let $I$ be the ideal of at most $n+1$ general fat points in $\mathbb P^n$. Then $I$ is componentwise linear. \end{theorem} \begin{proof} We need to show that $I_{<a_0+t>}$ has an $(a_0+t)$-linear resolution for all $t \ge 0$. By Proposition~\ref{arelexholes}, these ideals are lex ideals with holes, so Theorem~\ref{grahamharathm} implies that the minimal resolution of $I_{<a_0+t>}$ is a subcomplex of the minimal resolution of $\mathfrak m^{a_0+t}$, which is $(a_0+t)$-linear. \hfill $\square$ \end{proof} \section[5]{Consequences of componentwise linearity}\label{s:conseq} In this section, we discuss the implications of the fat point ideals $I$ being componentwise linear. We begin with a result of Aramova, Herzog, and Hibi \cite{AAH:stable}. \begin{theorem} \label{ginbetti} \emph{(Aramova-Herzog-Hibi)} Let $J$ be a homogeneous ideal in $R=k[x_0,\dots,x_n]$. Let gin$(J)$ be the reverse-lex generic initial ideal of $J$. Then $J$ is componentwise linear if and only if \[ \beta_{i,j}(R/J) = \beta_{i,j}(R/ \hbox{gin}(J)) \] for all $i$ and $j$. \end{theorem} Because generic initial ideals are strongly stable in characteristic zero, we get an immediate corollary that explains why the resolutions we examined in Sections 2 and 3 look like those of stable ideals. \newtheorem{corollary}[definition]{Corollary} \begin{corollary} \label{cwgin} Let $I$ be an ideal of at most $n+1$ general fat points in $\mathbb P^n$, where the underlying field has characteristic zero. Then $I$ has the same graded Betti numbers as a strongly stable ideal, namely its reverse-lex generic initial ideal. \end{corollary} We turn now to the question of finding the graded Betti numbers of ideals $I$ of at most $n+1$ general fat points in $\mathbb P^n$. Our goal is to express the graded Betti numbers of $R/I$ in terms of the Betti numbers of the ideals $R/I_{<d>}$; these ideals are lex ideals with holes, and we shall discuss formulas for their Betti numbers later in the section. Initially, we note that the graded Betti numbers of componentwise linear ideals satisfy a useful formula of Herzog and Hibi \cite{HerzogHibi}. \begin{proposition} \label{cwbetti} \emph{(Herzog-Hibi)} Let $J$ be a componentwise linear ideal in $R=k[x_0,\dots,x_n]$, and let $\mathfrak m = (x_0,\dots,x_n)$. Then for all $i$ and $d$, \[\beta_{i,i+d}(R/J) = \beta_i(R/J_{<d+1>}) - \beta_i(R/\mathfrak mJ_{<d>}).\] \end{proposition} Since the ideals $R/J_{<d>}$ and $R/\mathfrak mJ_{<d>}$ have only linear syzygies, we are writing total Betti numbers for simplicity. The next step is to remove the presence of $\beta_i(R/\mathfrak mJ_{<j>})$ in the formula in Proposition~\ref{cwbetti}, so we determine a formula for its Betti numbers. \begin{proposition} \label{mbetti} Let $J$ be a componentwise linear ideal in $R=k[x_0,\dots,x_n]$. Then for all $i$, \[ \beta_i(R/\mathfrak mJ_{<d>}) = \beta_{i,i+d}(J_{<d>}/\mathfrak m J_{<d>})-\beta_{i+1}(R/J_{<d>}).\] \end{proposition} \begin{proof} We have a short exact sequence \[ 0 \longrightarrow J_{<d>}/\mathfrak m J_{<d>} \longrightarrow R/\mathfrak mJ_{<d>} \longrightarrow R/J_{<d>} \rightarrow 0, \] which induces a long exact sequence of vector spaces in Tor in degree $i+d$: \[ \cdots \rightarrow \hbox{Tor}_{i+1}(k,R/\mathfrak mJ_{<d>})_{i+d} \rightarrow \hbox{Tor}_{i+1}(k,R/J_{<d>})_{i+d} \rightarrow \hbox{Tor}_{i}(k,J_{<d>}/\mathfrak m J_{<d>})_{i+d}\] \[ \rightarrow \hbox{Tor}_{i}(k,R/\mathfrak mJ_{<d>})_{i+d} \rightarrow \hbox{Tor}_{i}(k,R/J_{<d>})_{i+d} \rightarrow \cdots \] The leftmost term is zero since $\mathfrak mJ_{<d>}$ is generated in degree $d+1$. Moreover, the rightmost term is zero because the only nonzero graded Betti numbers of $R/J_{<d>}$, other than $\beta_{00}$, are those of the form $\beta_{i,i+d-1}(R/J_{<d>})$. Thus we have a short exact sequence of vector spaces, and the formula follows. \hfill $\square$ \end{proof} Finally, we compute the Betti numbers of the modules $J_{<d>}/\mathfrak m J_{<d>}$. \begin{proposition} \label{kernelbetti} Let $J \subset R=k[x_0,\dots,x_n]$ be any homogeneous ideal. Then for $i \ge 0$, \[ \beta_{i,i+d}(J_{<d>}/\frak m J_{<d>})=\beta_{i}(J_{<d>}/\frak m J_{<d>}) = \beta_{1}(R/J_{<d>}) {n+1 \choose i}.\] \end{proposition} \begin{proof} Any degree $d$ element of $J_{<d>}$ multiplied by any $x_i$ lands in $\mathfrak m J_{<d>}$, so we have $n+1$ minimal first syzygies of the form $(0,\dots,x_i,\dots,0)$, $i=0,\dots,n$, for each generator. Any other first syzygy can be written as a combination of these syzygies. The formula for the number of minimal syzygies at each step in the resolution follows immediately. \hfill $\square$ \end{proof} Combining Propositions~\ref{cwbetti}, \ref{mbetti}, and ~\ref{kernelbetti}, we have formulas for the graded Betti numbers of the fat point ideals in terms of the Betti numbers of the lex ideals with holes $I_{<d>}$. \begin{theorem} \label{bettiformulas} Let $I \subset R$ be the ideal of at most $n+1$ general fat points in $\mathbb P^n$. Then the graded Betti numbers of $R/I$ are given by \[ \beta_{i,i+d}(R/I) = \beta_i(R/I_{<d+1>})+\beta_{i+1}(R/I_{<d>}) - \beta_{1}(R/I_{<d>}) {n+1 \choose i}.\] \end{theorem} In \cite{GHP}, Gasharov, Hibi, and Peeva compute formulas for the graded Betti numbers of lex ideals with holes and, more generally, {\boldmath $a$}-stable ideals. Let $d_i$ be the bounds on the powers of $x_i$ in the lex ideal with holes (so the power of $x_i$ in any minimal generator is less than $d_i$), and for a monomial $m$, define $b(m)=\#\{i \, | \, x_i^{d_i-1}$ divides $m, 0 \le i \le \max(m)-1 \}$. Gasharov, Hibi, and Peeva prove the following theorem. \begin{theorem} \label{astable} \emph{(Gasharov-Hibi-Peeva)} Let $L$ be a lex ideal with holes in $R=k[x_0,\dots,x_n]$ with all its generators in degree $d$. Then the graded Betti numbers of $R/L$ are \[ \beta_{i,i+d-1}(R/L) = \sum_{m \in G(L)} {\max(m)-b(m) \choose i-1} \] for $i \ge 1$, and $\beta_{00}(R/L)=1$, with all other graded Betti numbers zero. \end{theorem} We have adjusted the formula from \cite{GHP} to reflect that we are working with variables $x_0,\dots,x_n$ instead of $x_1,\dots,x_n$, and we have restricted the theorem to the case we need. Corollary 2.3 in \cite{GHP} is much more general, and the Eliahou-Kervaire formulas for the Betti numbers of stable ideals actually follow from that result. As a consequence of Theorem~\ref{astable}, we can, in principle, get formulas for the graded Betti numbers of any ideal of at most $n+1$ general fat points in $\mathbb P^n$. Given the multiplicities of the points, we can use Fatabbi's characterization to list the monomials in $I$ in each degree. Theorem~\ref{astable} allows us to write down the graded Betti numbers of the $I_{<a_o+t>}$ without any difficult computation, and then we can apply Theorem~\ref{bettiformulas} to compute the graded Betti numbers of $R/I$. Our final application is to a conjecture of Herzog, Huneke, and Srinivasan on the multiplicity of a polynomial ring modulo a homogeneous ideal. \newtheorem{conjecture}[definition]{Conjecture} \begin{conjecture} \label{multconj} \emph{(Huneke-Srinivasan, Herzog-Srinivasan)} Let $J$ be a homogeneous ideal of codimension $c$ in $R=k[x_0,\dots,x_n]$ such that $R/J$ is Cohen-Macaulay, and let $e(R/J)$ be the multiplicity of $R/J$. Let $m_i$ be the minimal degree of a syzygy at step $i$ in the minimal graded free resolution of $R/J$, and let $M_i$ be the corresponding maximum. Then \[ \frac{1}{c!} \prod_{i=1}^c m_i \le e(R/J) \le \frac{1}{c!} \prod_{i=1}^c M_i.\] \end{conjecture} Conjecture~\ref{multconj} is known in a number of special cases but is open in general, even for monomial ideals in codimension three and above. Recently, there has been interest in proving it for configurations of points in $\mathbb P^n$; see, for example, the paper of Gold, Schenck, and Srinivasan \cite{GSS}. The fact that ideals of at most $n+1$ general fat points in $\mathbb P^n$ are componentwise linear gives the result for free in that case. \begin{proposition} \label{multpoints} Let $I \subset R$ be the ideal of at most $n+1$ general fat points in $\mathbb P^n$ over a field of characteristic zero. Then $R/I$ satisfies Conjecture~\ref{multconj}. \end{proposition} \begin{proof} R{\"o}mer proves Conjecture~\ref{multconj} in \cite{Roemer} for all componentwise linear ideals over a field of characteristic zero, noting that the result follows directly from Theorem~\ref{ginbetti} since Conjecture~\ref{multconj} is true for stable ideals. \hfill $\square$ \end{proof} We note that the upper bound is not hard to show regardless of characteristic. The upper bound for the ideal $I$ of a set of at most $n+1$ general fat points in $\mathbb P^n$ is \[ \frac{1}{n!} \prod_{i=0}^{n-1} (a_0+a_1+i) \] because there are generators in degree $a_0+a_1$, and $R/I$ is Cohen-Macaulay. Since $\dim R/I=1$, the multiplicity cannot exceed that of $k[x_1,\dots,x_n]/(x_1,\dots,x_n)^{a_0+a_1}$, which is equal to the upper bound. \section[6]{Larger sets of fat points}\label{s:larger} A natural question is whether we can extend the results of Section 4 to sets of more than $n+1$ general fat points in $\mathbb P^n$. The following example shows that Theorem~\ref{mainthm} does not hold for $n+2$ fat points in $\mathbb P^n$. \begin{example}\label{n+2}\emph{Consider four double points in general position in $\mathbb P^2$. We can take the ideal defining these fat points to be \[ I=(b,c)^2 \cap (a,c)^2 \cap (a,b)^2 \cap (a-b,a-c)^2 \] in $R=k[a,b,c]$. The minimal graded free resolution of $R/I$ is \[ 0 \rightarrow R^2(-6) \rightarrow R^3(-4) \rightarrow R \rightarrow R/I \rightarrow 0. \] Clearly, the resolution of $I_{<4>}$ is not 4-linear, and thus $I$ is not componentwise linear.} \end{example} There are a number of directions in which one can proceed from here. First, while Example~\ref{n+2} shows that the ideal of $n+2$ general fat points in $\mathbb P^n$ will not necessarily be componentwise linear, special arrangements of points or fat points may yield componentwise linear ideals. It would be interesting to investigate ideals corresponding to various geometric objects; we are confident there are more ideals arising from geometry that are componentwise linear. Additionally, it would be particularly useful to have more tests for componentwise linearity available to aid in checking for the condition. Finally, the natural long-term goal in this area is the question of Herzog, Reiner, and Welker: Suppose $M$ is a sequentially Cohen-Macaulay module. Does there exist a natural dual module $M^*$ that has a componentwise linear resolution? If so, this would generalize the theorem of Herzog and Hibi on sequentially Cohen-Macaulay simplicial complexes. \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}

https://arxiv.org/abs/1902.05034

Generalized ergodic problems: existence and uniqueness structures of solutions

We study a generalized ergodic problem (E), which is a Hamilton-Jacobi equation of contact type, in the flat $n$-dimensional torus. We first obtain existence of solutions to this problem under quite general assumptions. Various examples are presented and analyzed to show that (E) does not have unique solutions in general. We then study uniqueness structures of solutions to (E) in the convex setting by using the nonlinear adjoint method.

\section{Introduction} In this paper, we focus on the following equation \[ {\rm (E)} \qquad H(x,u,Du) = c \qquad \text{ in } \mathbb{T}^n. \] Here, $\mathbb{T}^n =\mathbb{R}^n/\mathbb{Z}^n$ is the flat $n$-dimensional torus, and the Hamiltonian $H=H(x,r,p):\mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n \to \mathbb{R}$ is a given continuous function. We seek for a pair of unknowns $(u,c) \in C(\mathbb{T}^n) \times \mathbb{R}$ that solves (E) in the viscosity sense. We use $Du$ to denote the spatial gradient of $u$. We are always concerned with viscosity solutions, and the adjective ``viscosity" is often omitted in the paper. Our main goals in this paper are twofold. First of all, we obtain existence results of solutions to (E) under quite general assumptions. Second, it is well-known in the theory of viscosity solutions that if $r \mapsto H(x,r,p)$ is not strictly monotone, then (E) might not have unique solutions (see Examples \ref{ex-4}--\ref{ex-7} in Section \ref{sec:ex}). It is therefore of our main interests to understand why this phenomenon appears, and to describe uniqueness structures of solutions to (E). We call (E) a generalized ergodic problem. In various other contexts, (E) is also called a Hamilton-Jacobi equation of contact type. \subsection{Assumptions} We list here the main assumptions on Hamiltonian $H$ that are used in the paper. \begin{itemize} \item[(H1)] $H$ is uniformly Lipschitz in $r$, that is, there exists a constant $C_1>0$ such that \[ |H(x,r,p) - H(x,s,p)| \leq C_1 | r-s| \quad \text{for all } (x,p) \in \mathbb{T}^n \times \mathbb{R}^n, \, r,s \in \mathbb{R}. \] \item[(H2a)] $H$ is coercive in $p$, that is, \[ \lim_{|p| \to \infty} H(x,0,p)=+\infty \quad \text{ uniformly for } x \in \mathbb{T}^n. \] \item[(H2b)] $H$ is superlinear in $p$, that is, \[ \lim_{|p| \to \infty} \frac{H(x,0,p)}{|p|}=+\infty \quad \text{ uniformly for } x \in \mathbb{T}^n. \] \end{itemize} It is clear that (H2b) is stronger than (H2a). We will assume either (H2a) or (H2b) in each of our results on existence of solutions to (E). To address the uniqueness structure, we need to assume the following assumptions. \begin{itemize} \item[(H3)] $H \in C^2(\mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n)$, and \[ \lim_{|p| \to \infty} \left(\frac{1}{2}H(x,r,p)^2 + D_xH(x,r,p)\cdot p\right)=+\infty \quad \text{ uniformly for } (x,r) \in \mathbb{T}^n \times \mathbb{R}. \] \item[(H4)] The map $r \mapsto H(x,r,p)$ is nondecreasing for all $(x,p) \in \mathbb{T}^n \times \mathbb{R}^n$. \item[(H5)] The map $(r,p) \mapsto H(x,r,p)$ is convex for all $x \in \mathbb{T}^n$. \end{itemize} It is worth noting that (H3) and (H4) are quite standard assumptions. We only require that $H$ is nondecreasing in $r$ in (H4), so it may fail to be strictly increasing. Condition (H5) however is rather strong since convexity is imposed both in $r$ and $p$. In any case, nowhere in this paper do we require $H$ to be uniformly convex in $p$. \subsection{Main results} We first state two existence results for solutions to (E). The first one is quite a standard result in light of the classical Perron method. \begin{thm}\label{thm:exist1} Assume {\rm (H1), (H2a)}. Assume further that there exist $c\in \mathbb{R}$, and $\psi, \varphi \in {\rm Lip\,}(\mathbb{T}^n)$ such that $\psi \leq \varphi$, $\psi$ and $\varphi$ are a viscosity subsolution and a viscosity supersolution to {\rm (E)}, respectively. Then, {\rm (E)} has a viscosity solution $u\in {\rm Lip\,}(\mathbb{T}^n)$ with $c\in \mathbb{R}$ given by the assumptions. \end{thm} This result is not new in the literature, and is just a variant of the classical results in \cite{Is}. What is different here is that under assumptions (H1) and (H2a), we obtain directly a Lipschitz viscosity solution $u$ with known Lipschitz constant, which is not written down explicitly in \cite{Is}. It is therefore of our interests to record it here. Next is our second existence result for solutions to (E) without prior information about the constant $c$. \begin{thm}\label{thm:exist2} Assume {\rm (H1), (H2b)}. Then, {\rm (E)} has a solution $(v,c) \in {\rm Lip\,}(\mathbb{T}^n) \times \mathbb{R}$. \end{thm} As we do not assume the existence of a subsolution $\psi$ and a supersolution $\phi$ with $\psi\le\phi$ for some given $c\in \mathbb{R}$ as in Theorem \ref{thm:exist1}, the existence of solutions to (E) cannot be obtained by the standard Perron method. Existence result for (E) was obtained in \cite[Theorem 1.5]{WWY-0} under an additional assumption that $H$ is uniformly convex. See also \cite{SWY}. Unlike \cite{SWY, WWY-0}, we do not need any convexity of $H$ here, and we believe that Theorem \ref{thm:exist2} is new in the literature. We emphasize here that, although the existence of $(v,c) \in C(\mathbb{T}^n) \times \mathbb{R}$, solution to (E), is guaranteed by Theorems \ref{thm:exist1}--\ref{thm:exist2}, we do not have uniqueness of constant $c$ in general. See Examples \ref{ex:2}, \ref{ex-4}, and \ref{ex-5} below. Furthermore, for a fixed $c\in \mathbb{R}$ such that (E) has a solution, (E) might have multiple solutions as described explicitly in Section \ref{sec:ex} (Examples \ref{ex-4}--\ref{ex-7}). It is therefore extremely important to proceed further to understand this phenomenon and investigate how such nonuniqueness appears. In particular, we aim to find a uniqueness set of (E), that is a set (hopefully the smallest) such that, if two solutions agree on it then they agree everywhere. Towards this goal, a prototype class of Hamiltonian of the following form is studied carefully in Section \ref{sec:prototype} . \begin{itemize} \item[(H6)] Assume that \[ H(x,r,p)=|p|^m - V(x) + f(r) \quad \text{ for } (x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n. \] Here, $m \geq 1$ is a given number, and $V \in C(\mathbb{T}^n)$ is the potential energy with $\min_{\mathbb{T}^n} V=0$. The function $f:\mathbb{R} \to \mathbb{R}$ is convex, and \[ \begin{cases} f(r)=0 \quad &\text{ for } r \leq 0,\\ f(r)>0 \quad &\text{ for } r>0. \end{cases} \] \end{itemize} \noindent Of course, we see that $f$ is not strictly increasing here, which makes the situation more interesting. Here is our first result on the uniqueness property of (E) for the Hamiltonians in the prototype class (H6) when $c>0$. \begin{prop}\label{prop:unique-positive} Assume {\rm (H6)}. For $c>0$ fixed, {\rm (E)} admits a unique solution $(u_c,c) \in C(\mathbb{T}^n) \times (0,\infty)$. \end{prop} Next, we consider the case that $c=0$. By using Proposition \ref{prop:unique-positive}, a priori estimates, and Arzel\`a-Ascoli's theorem, we can easily show that under (H6), (E) has a solution $(u,0) \in C(\mathbb{T}^n) \times \mathbb{R}$ (see the last part of the proof of Proposition \ref{prop:prot} for a proof of this fact). As $\min_{\mathbb{T}^n} V =0$, denote by \[ M_V = \left\{ x\in \mathbb{T}^n\,:\, V(x) = \min_{\mathbb{T}^n} V = 0\right\}. \] Here is our second result along this line. \begin{prop}\label{prop:unique-0} Assume {\rm (H6)}. Let $c=0$. Then, $M_V$ is a uniqueness set for {\rm (E)}, that is, if $(u_1,0), (u_2,0)$ are two solutions to {\rm (E)}, and $u_1=u_2$ on $M_V$, then $u_1=u_2$. \end{prop} \noindent It is worth noting that Proposition \ref{prop:unique-0} was first obtained in \cite{NR} when $f \equiv 0$. \medskip The uniqueness structure of solutions to (E), with Hamiltonians beyond the class of (H6), is more involved. To study it in a systematic way, we apply the nonlinear adjoint method and develop further the ideas in \cite{MT-P}. This is done in Section \ref{sec:structure}. We refer the readers to \cite{Ev1, T1, CGMT, MT-A, MT-P, LMT} and the references therein for the developments of the nonlinear adjoint method. Under assumptions (H1), (H2b), (E) admits a solution $(v,c) \in C(\mathbb{T}^n) \times \mathbb{R}$. As noted above, the constant $c$, that is the right hand side of (E), is not unique in general. Therefore, to discuss the uniqueness structure of (E), we fix a $c\in \mathbb{R}$ such that (E) has a solution $v\in C(\mathbb{T}^n)$. By a further normalization (setting $\tilde H(x,r,p) = H(x,r,p)-c$ for $(x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n$), we may assume that $c=0$. We hence study the uniqueness structure for the following problem \begin{equation}\label{E-0} H(x,u,Du)=0 \quad \text{ in } \mathbb{T}^n. \end{equation} Our main result on the uniqueness structure of \eqref{E-0} is as follows. \begin{thm}\label{thm:unique} Assume {\rm (H1), (H2b), (H3), (H4), (H5)}. Let $\mathcal{M}$ be the set of measures in Definition \ref{def:M} of Section \ref{sec:structure}. Then, for any two solutions $u_1, u_2$ to \eqref{E-0}, the condition \[ \int_{\mathbb{T}^n} u_1(x) \, d\nu(x) \leq \int_{\mathbb{T}^n} u_2(x)\, d\nu(x) \quad \text{ for all } \nu \in \mathcal{M} \] implies $u_1 \leq u_2$. In particular, $M := \overline{ \bigcup_{\nu \in \mathcal{M}} {\rm supp} (\nu)}$ is a uniqueness set for \eqref{E-0}. \end{thm} As described in Definition \ref{def:M}, $\mathcal{M}$ contains adjoint measures associated to solutions of \eqref{E-0}. The whole construction of these measures is done in Section \ref{sec:structure}. Theorem \ref{thm:unique} is a generalized version of \cite[Theorem 1.1]{MT-P}. See also a related work \cite{T}. To the best of our knowledge, Theorem \ref{thm:unique} is new in the literature. In two specific situations (see Subsections \ref{subsec:structure1}--\ref{subsec:structure2}), we have clear understanding about this $\mathcal{M}$. In particular, we find a natural link between Theorem \ref{thm:unique} and Proposition \ref{prop:unique-0} above (see Proposition \ref{prop:prot}). \subsection*{Organization of the paper} The paper is organized as following. In Section \ref{sec:exist}, we give the proofs of Theorems \ref{thm:exist1}--\ref{thm:exist2}. Besides, we give three examples of Hamiltonians satisfying requirements of Theorem \ref{thm:exist1} in Subsection \ref{subsec:exist}. In Section \ref{sec:ex}, we give various new examples to discuss nonuniqueness issues of solutions to (E). Then, Section \ref{sec:prototype} is devoted to further analysis on a uniqueness set for a prototype case that was discussed in Section \ref{sec:ex}. Finally, in Section \ref{sec:structure}, we use the nonlinear adjoint method to study systematically the uniqueness structure of solutions to (E). Various connections with classical results and with the prototype case in Section \ref{sec:prototype} are discussed in deep too. \section{Existence results for solutions to (E)} \label{sec:exist} \subsection{Proof of Theorem \ref{thm:exist1}} \begin{proof}[Proof of Theorem \ref{thm:exist1}] The main idea is to use the Perron method to get the existence result. Let $M= \|\psi\|_{L^\infty(\mathbb{T}^n)} + \|\varphi\|_{L^\infty(\mathbb{T}^n)} +1$. By assumptions (H1) and (H2a), there exists $C_2>0$ such that \[ H(x,r,p) \leq c \quad \text{ for some } (x,r) \in \mathbb{T}^n \times [-M, M] \quad \Longrightarrow \quad |p| \leq C_2. \] Define, for $x\in \mathbb{T}^n$, \begin{multline*} u(x) = \sup \Big\{ v(x)\,:\, \psi \leq v \leq \varphi, \, \|Dv\|_{L^\infty(\mathbb{T}^n)} \leq C_2,\\ \text{and $v \in {\rm Lip\,}(\mathbb{T}^n)$ is a viscosity subsolution to (E)} \Big\}. \end{multline*} Of course, $u$ is well-defined as $\psi$ itself is an admissible subsolution in the above formula. Furthermore, it is clear that $u$ is Lipschitz in $\mathbb{T}^n$, and $\|Du\|_{L^\infty(\mathbb{T}^n)} \leq C_2$. By the stability of viscosity subsolutions, we have that $u$ is a viscosity subsolution to (E). Hence, we only need to show that $u$ is a viscosity supersolution to (E). Assume by contradiction that this is not the case. Then, there exist a smooth test function $\phi \in C^\infty(\mathbb{T}^n)$ and a point $x_0 \in \mathbb{T}^n$ such that \[ \begin{cases} u(x_0) = \phi(x_0),\ u(x) >\phi(x) \quad \text{ for all } x \in \mathbb{T}^n \setminus \{ x_0\},\\ H(x_0,u(x_0),D\phi(x_0)) = H(x_0,\phi(x_0),D\phi(x_0)) <c. \end{cases} \] There are two cases to be considered here. The first case is when $u(x_0)=\varphi(x_0)$. This means that $\phi$ touches $\varphi$ from below at $x_0$. By the definition of viscosity supersolutions, \[ H(x_0,\phi(x_0),D\phi(x_0)) \geq c, \] which implies a contradiction immediately. The second case is when $u(x_0) <\varphi(x_0)$. There exist $r,\varepsilon>0$ sufficiently small such that \[ \begin{cases} u(x) < \varphi(x) -\varepsilon \quad &\text{ for all } x \in B(x_0,r),\\ \phi(x) < u(x) -\varepsilon \quad &\text{ for all } x\in \partial B(x_0,r),\\ H(x,\phi(x), D\phi(x)) < c - C_1 \varepsilon \quad &\text{ for all } x \in B(x_0,r),\\ |D\phi(x)| \leq C_2 \quad &\text{ for all } x \in B(x_0,r). \end{cases} \] Now, set \[ \overline{u}(x) = \begin{cases} \max\{u(x),\phi(x)+\varepsilon\} \quad &\text{ for all } x\in B(x_0,r),\\ u(x) \quad &\text{ for all } x \in \mathbb{T}^n \setminus B(x_0,r). \end{cases} \] It is quite clear that $\overline{u}$ is a viscosity subsolution to (E) thanks to (H1), and $\|D\overline{u}\|_{L^\infty(\mathbb{T}^n)} \leq C_2$. This again leads to a contradiction. The proof is complete. \end{proof} \subsection{Examples of Hamiltonians satisfying Theorem \ref{thm:exist1}} \label{subsec:exist} \begin{ex}[Classical setting with no $r$-dependence] \label{ex:1} {\rm If $H(x,r,p)= \tilde H(x,p)$ with $\tilde H$ satisfies {\rm (H2a)}, that is, \[ \lim_{|p| \to \infty} \tilde H(x,p) = +\infty \quad \text{ uniformly for } x\in \mathbb{T}^n. \] Then we use classical results (see \cite{LPV} for example) to have the existence of a unique constant $c\in \mathbb{R}$ such that \[ \tilde H(x,Dv) = c \quad \text{ in } \mathbb{T}^n \] has a viscosity solution $v\in C(\mathbb{T}^n)$. Then, we can simply choose $\psi=\varphi=v$. In this example, $c$ is unique. } \end{ex} \begin{ex}[Strictly monotone setting] \label{ex:2} {\rm Assume {\rm (H1), (H2a)}. If there exists $\alpha>0$ such that \[ H_r(x,r,p) \geq \alpha \quad \text{ for all } (x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n, \] then, for each fixed $c \in \mathbb{R}$, we can choose \[ \psi \equiv -\frac{1}{\alpha}\left(\|H(\cdot,0,0)\|_{L^\infty(\mathbb{T}^n)} +|c|\right), \quad \text{and} \quad \varphi \equiv \frac{1}{\alpha}\left(\|H(\cdot,0,0)\|_{L^\infty(\mathbb{T}^n)} +|c|\right). \] Therefore, (E) has solutions for each $c \in \mathbb{R}$. This is consistent with classical results (see \cite{CEL, CL} for example). } \end{ex} \begin{ex}[Non-monotone setting] \label{ex:3} {\rm Assume {\rm (H1), (H2b)}. Let us assume further that \begin{equation}\label{nm-condition1} \begin{cases} \max_{x\in \mathbb{T}^n} H(x,0,0) = H(x_0,0,0) \quad &\text{ for some } x_0 \in \mathbb{T}^n,\\ H(x_0,0,0) \leq H(x_0,0,p) \quad &\text{ for all } p \in \mathbb{R}^n. \end{cases} \end{equation} Note that the requirements of this example are stronger than those in Theorem \ref{thm:exist2} (as we only assume (H1) and (H2b) there). Nevertheless, this is a direct application of Theorem \ref{thm:exist1}, and hence, it is worth pointing it out here. Examples of Hamiltonians satisfying (H1), (H2b), and \eqref{nm-condition1} are many. A typical one is $H(x,r,p) = |p|^m + f(r) + V(x)$ for $(x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n$. Here, $m>1$ is fixed, $V\in C(\mathbb{T}^n)$, and $ f \in {\rm Lip\,}(\mathbb{R})$ with Lipschitz constant at most $C_1$. Of course, there is no requirement on convexity of $H$ in $p$ here. \medskip In this setting, we choose first \[ c=\max_{x\in \mathbb{T}^n} H(x,0,0) = H(x_0,0,0), \quad \text{and} \quad \psi \equiv 0. \] We now construct $\varphi$. Let $Q(x_0)= x_0 + [-1/2,1/2]^n$ be the unit cube centered at $x_0$. For $s>0$ sufficiently large, set \[ \varphi(x) = s |x-x_0| \quad \text{ for all } x\in Q(x_0), \] and extend $\varphi$ to $\mathbb{R}^n$ periodically. We claim that $\varphi$ is a supersolution to (E). It is clear that $\varphi$ is not differentiable at $x_0$ and $\partial Q(x_0)$. We do not have to worry about $\partial Q(x_0)$ as for any $x \in \partial Q(x_0)$, $D^- \varphi(x) = \emptyset$. At $x=x_0$, we have $D^-\varphi(x_0) = \overline{B}(0,s)$. By the second line in assumption \eqref{nm-condition1}, we have that \[ H(x_0,0,p) \geq H(x_0,0,0) = c \quad \text{ for all } p \in \overline{B}(0,s). \] For other values of $x$, $\varphi$ is smooth and $|D\varphi(x)|=s$. We use (H1) and (H2b), the superlinearity of $H$, to get \[ H(x,\varphi(x),D\varphi(x)) \geq H(x,0,D\varphi(x)) - C_1s \geq c, \] for $s$ sufficiently large.} \end{ex} \subsection{Proof of Theorem \ref{thm:exist2}} We always assume (H1), (H2b) in this subsection. We first formulate Theorem \ref{thm:exist2} as a fixed point problem by adding a monotone term to (E). Fix $\lambda > C_1+1$. For each $u \in C(\mathbb{T}^n)$, let $v \in {\rm Lip\,}(\mathbb{T}^n)$ be the unique viscosity solution to \begin{equation}\label{eq:u-v} \lambda v + H(x,u,Dv) - \lambda u = 0 \quad \text{ in } \mathbb{T}^n. \end{equation} Note that we use the Perron method to get directly a solution $v \in {\rm Lip\,}(\mathbb{T}^n)$. Then, uniqueness of \eqref{eq:u-v} follows immediately. Denote by $G(u) = w := v - \min_{\mathbb{T}^n} v$. It is clear that $w=G(u)$ solves \begin{equation}\label{eq:G} \lambda (w-u) + H(x,u,Dw) = -\lambda \min_{\mathbb{T}^n} v \quad \text{ in } \mathbb{T}^n. \end{equation} Our aim now is to show that the map $G: C(\mathbb{T}^n) \to C(\mathbb{T}^n)$ has a fixed point by using the Schauder fixed point theorem. We first show that $G$ is continuous. \begin{lem}\label{lem:G1} For every $u_1, u_2 \in C(\mathbb{T}^n)$, \[ \|G(u_1) - G(u_2)\|_{L^\infty(\mathbb{T}^n)} \leq 4 \|u_1-u_2\|_{L^\infty(\mathbb{T}^n)}. \] \end{lem} \begin{proof} For $i=1,2$, let $v_i \in C(\mathbb{T}^n)$ be the unique viscosity solution to \begin{equation*} \lambda v_i + H(x,u_i,Dv_i) - \lambda u_i = 0 \quad \text{ in } \mathbb{T}^n. \end{equation*} We use (H1) to deduce that $v_1$ is a subsolution to \[ \lambda v_1 + H(x,u_2, Dv_1) - \lambda u_2 \leq ( \lambda + C_1) \|u_1-u_2\|_{L^\infty(\mathbb{T}^n)} \quad \text{ in } \mathbb{T}^n. \] By the comparison principle, we yield \[ v_1 - \left(1 + \frac{C_1}{\lambda} \right) \|u_1-u_2\|_{L^\infty(\mathbb{T}^n)} \leq v_2. \] By the same argument, we obtain \[ \|v_1-v_2\|_{L^\infty(\mathbb{T}^n)} \leq \left(1 + \frac{C_1}{\lambda} \right) \|u_1-u_2\|_{L^\infty(\mathbb{T}^n)} \leq 2 \|u_1-u_2\|_{L^\infty(\mathbb{T}^n)}. \] Next, for $i=1,2$, denote by \[ w_i = G(u_i) = v_i - \min_{\mathbb{T}^n} v_i = v_i - v_i(x_i) \quad \text{ for some } x_i \in \mathbb{T}^n. \] Then, for any $x\in \mathbb{T}^n$, \begin{align*} w_1(x) - w_2(x) &= (v_1(x) - v_1(x_1)) - (v_2(x) - v_2(x_2))\\ &= (v_1(x) - v_2(x)) + (v_2(x_2) - v_1(x_1)) \\ &= (v_1(x) - v_2(x)) + (\min_{\mathbb{T}^n} v_2 - v_1(x_1))\\ &\leq (v_1(x) - v_2(x)) + (v_2(x_1) - v_1(x_1)) \leq 2 \|v_1-v_2\|_{L^\infty(\mathbb{T}^n)} \\ &\leq 4\|u_1-u_2\|_{L^\infty(\mathbb{T}^n)}. \end{align*} By a symmetric argument, the proof is complete. \end{proof} Set $C_0 = \max_{x\in \mathbb{T}^n} |H(x,0,0)|$. By (H2b), we pick $\alpha>0$ such that, if $|p| \geq \alpha$, then \[ H(x,0,p) \geq 3\lambda (C_0 + \alpha (1+\sqrt{n})). \] Denote by \[ K := \{u \in {\rm Lip\,}(\mathbb{T}^n) \,:\, u \geq 0, \|u\|_{L^\infty(\mathbb{T}^n)} +\|Du\|_{L^\infty(\mathbb{T}^n)} \le \alpha (1+\sqrt{n})\}. \] Clearly, $K$ is a non-empty convex and compact subset of $C(\mathbb{T}^n)$. \begin{lem}\label{lem:G2} We have that $G(K) \subset K$, where $G(K):=\{G(v)\,:\, v\in K\}$. \end{lem} \begin{proof} Fix $u \in K$, and let $v \in {\rm Lip\,}(\mathbb{T}^n)$ be the viscosity solution to \eqref{eq:u-v}. First of all, it is clear that $C_0 + 2\alpha (1+\sqrt{n})$ and $-C_0 - \alpha(1+\sqrt{n})$ are, respectively, a supersolution and a subsolution to \eqref{eq:u-v}. The comparison principle then gives \[ -C_0 - \alpha(1+\sqrt{n}) \leq v \leq C_0 + 2\alpha (1+\sqrt{n}) \quad \text{ in } \mathbb{T}^n. \] Thus, for a.e. $x\in \mathbb{T}^n$, \[ H(x,0,Dv(x)) \leq \lambda (u(x) - v(x)) + C_1 \|u\|_{L^\infty(\mathbb{T}^n)} < 3\lambda (C_0 + \alpha (1+\sqrt{n})), \] which, together with the choice of $\alpha$, yields $\|Dv\|_{L^\infty(\mathbb{T}^n)} \leq \alpha$. Hence, for $w=v- \min_{\mathbb{T}^n} v$, we have $w \in K$. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:exist2}] By Lemmas \ref{lem:G1}--\ref{lem:G2}, we are able to apply Schauder's fixed point theorem to imply the existence of $u \in K$ such that \begin{equation*} G(u) = u. \end{equation*} This means that, for $v \in C(\mathbb{T}^n)$ solves \eqref{eq:u-v}, $u= v - \min_{\mathbb{T}^n}v $ satisfies \begin{equation*} H(x,u,Du) = c:= -\lambda \min_{\mathbb{T}^n} v \quad \text{ in } \mathbb{T}^n. \end{equation*} \end{proof} \section{Some examples on nonuniqueness of solutions to (E)} \label{sec:ex} In this section, we give several examples to illustrate the nonuniqueness of solutions to (E). Our main guiding principle here is that, if $r \mapsto H(x,r,p)$ is not strictly monotone for each $(x,p) \in \mathbb{T}^n \times \mathbb{R}^n$, then it is highly unlikely the case that (E) has a unique solution. \begin{ex}\label{ex-4} {\rm Assume that $n=1$, and \[ H(x,r,p) = |p|^2 + V(x) - \lambda r \quad \text{ for } (x,r,p) \in \mathbb{T} \times \mathbb{R} \times \mathbb{R}, \] where $\lambda > 2$ is given. Clearly, $H_r(x,r,p) = -\lambda <0$. Here, the potential energy $V$ is defined as \begin{equation*} V(x) = \begin{cases} \frac{1}{4} (x-\frac{1}{2})^2 \quad& 0 \leq x \leq \frac{1}{2},\\ \frac{1}{4} (x+\frac{1}{2})^2 \quad& -\frac{1}{2} \leq x \leq 0. \end{cases} \end{equation*} Extend $V$ to $\mathbb{R}$ in a periodic way. It is worth noting that $V$ is $C^1$ on the torus except at $0$, and $V$ is a viscosity solution to \[ |V'|^2 - V = 0 \quad \text{ in } \mathbb{T}. \] We use this fact to imply that \begin{equation*} u_1 = \frac{\lambda + \sqrt{\lambda^2 - 4}}{2} V \quad \text{and} \quad u_2 = \frac{\lambda - \sqrt{\lambda^2 - 4}}{2} V \end{equation*} are two different viscosity solutions of the equation \begin{equation*} -\lambda u + |u'|^2 + V = 0 \quad \text{ in } \mathbb{T}. \end{equation*} In other words, $(u_1,0)$ and $(u_2,0)$ are two pairs of solutions to (E) with $c=0$ here. It is also clear that (E) has at least two solutions for every $c\in \mathbb{R}$. Indeed, for each $c\in \mathbb{R}$, and $i=1,2$, define \[ u_{i,c} = u_i -\frac{c}{\lambda}. \] Then $(u_{i,c},c)$ is a solution to (E) for $i=1,2$. } \end{ex} See also \cite[Section 1.4]{G2} for similar comments on the nonuniqueness of both $c$ and $u$. Surely, one objection that one may have for the above example is that $H_r(x,r,p) = -\lambda <0$, which is too restrictive. Nevertheless, in the following example, we will show that nonuniqueness appears even when $H_r(x,r,p) \geq 0$. \begin{ex}\label{ex-5} {\rm Assume that \[ H(x,r,p) = |p| - V(x) + f(r) \quad \text{ for } (x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n. \] Here, $f:\mathbb{R} \to \mathbb{R}$ is defined as \[ f(r)= \begin{cases} 0 \quad &\text{ for } r \leq 0,\\ r \quad &\text{ for } r>0. \end{cases} \] And $V \in C(\mathbb{T}^n)$ is the potential energy with $\min_{\mathbb{T}^n} V=0$. Let $w \in C(\mathbb{T}^n)$ be the viscosity solution to \begin{equation}\label{eq:ex5} w + |Dw| - V = 0 \quad \text{ in } \mathbb{T}^n. \end{equation} As $0$ is a subsolution to the above, $w \geq 0$. Besides, it is clear that $w \leq V$, which gives us that $\{V=0\} \subset \{w=0\}$. In particular, $f(w)=w$ always, and hence, $(w,0)$ is a solution to (E). From this, it is also clear that (E) has a solution $(w+c,c)$ for every $c\geq 0$. Let us now proceed to describe more solutions to (E) with $c=0$. Consider the usual ergodic (cell problem) \begin{equation}\label{eq:ex5-E} |Dv|-V = 0 \quad \text{ in } \mathbb{T}^n, \end{equation} which is of eikonal type. For each solution $v \in C(\mathbb{T}^n)$ of \eqref{eq:ex5-E}, take $C> \|v\|_{L^\infty(\mathbb{T}^n)}$, then $v-C$ is still a solution to \eqref{eq:ex5-E}, and $v-C \leq 0$. Thus, $f(v-C)=0$, and $(v-C,0)$ is a solution to (E). } \end{ex} \begin{ex}\label{ex-6} {\rm Let us analyze further Example \ref{ex-5}. Basically, if we put more structural condition on $V$, we are able to find more nontrivial solutions to (E) with $c=0$. Below, we use the setting in Example \ref{ex-5} and present an example in which the solution $u$ has range in both branches of the $f$ function. We assume further that $V \in C^1(\mathbb{T}^n)$, and that for some $r \in (0,\frac12)$ we have \begin{equation}\label{V-zero} \begin{cases} V \ge 0 \text{ \ in \ } \mathbb{T}^n \;\; \text{ and } \;\; \{V=0\} = \{0\} \cup \partial B(0,r), \\%\text{ for some small } r>0,\\ V(x) = \tilde V(|x|) \;\; \text{ for all } |x| \leq r. \end{cases} \end{equation} Here, $\tilde V: [0,r] \to \mathbb{R}$ is $C^1$, $\tilde V \geq 0$, and $\{\tilde V=0\} = \{0,r\}$. Let $w$ be the solution to \eqref{eq:ex5}. Then clearly $0 \leq w \leq V$, and \[ w(x)=0, \quad \text{and} \quad Dw(x)=0 \quad \text{ for each } x \in \partial B(0,r). \] Moreover, $w$ is not constantly zero in $\mathbb{T}^n\setminus \overline{B}(0,r)$. Next, we construct $\phi:[0,r] \to \mathbb{R}$ such that \[ \begin{cases} \phi'(s) = \tilde V(s) \quad \text{ for all } 0 \leq s \leq r,\\ \phi(r)=0. \end{cases} \] Then define $u:\mathbb{T}^n \to \mathbb{R}$ by \[ u(x)= \begin{cases} \phi(|x|) \quad &\text{ for } x \in B(0,r),\\ w(x) \quad &\text{ for } x \in \mathbb{T}^n \setminus B(0,r). \end{cases} \] Clearly, $u <0$ in $B(0,r)$, and $Du(x)=0$ on $\partial B(0,r)$. Besides, $Du(0)=0$, and therefore, $u$ solves \[ |Du(x)| = \phi'(|x|) = \tilde V(|x|) = V(x) \quad \text{ for } x\in B(0,r). \] We conclude that $(u,0)$ is a solution to (E). } \end{ex} Finally, let us consider the following example, where the Hamiltonian is of magnetic type. \begin{ex}\label{ex-7} {\rm Assume that \[ H(x,r,p) = |p|^2 - p\cdot D\varphi(x) + f(r) \quad \text{ for } (x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n. \] Here, $f:\mathbb{R} \to \mathbb{R}$ is defined as \[ f(r)= \begin{cases} 0 \quad &\text{ for } r \leq 0,\\ r \quad &\text{ for } r>0. \end{cases} \] And $\varphi \in C^1(\mathbb{T}^n)$ is given. Let us now proceed to describe various solutions to (E) with $c=0$. The corresponding equation reads \begin{equation}\label{eq:mag1} |Du|^2 - Du\cdot D\varphi + f(u) =0 \quad \text{ in } \mathbb{T}^n. \end{equation} It is clear that $u \equiv 0$ is a trivial solution to the above. Now, take any solution $u \in C(\mathbb{T}^n)$ of \eqref{eq:mag1}. We show that $u \leq 0$. Indeed, take $x_1 \in \mathbb{T}^n$ so that $u(x_1) = \max_{\mathbb{T}^n} u$. By the viscosity subsolution test, we deduce that \[ f(u(x_1)) \leq 0 \quad \Rightarrow \quad u(x_1) \leq 0. \] Thus, $u\leq 0$, and $u$ solves a usual ergodic (cell problem) without $f$ as following \begin{equation}\label{eq:mag2} |Du|^2 - Du\cdot D\varphi = 0 \quad \text{ in } \mathbb{T}^n, \end{equation} which is quite an interesting phenomenon. It is clear that $u_1 \equiv C_1$ for any constant $C_1 \leq 0$, and $u_2 \equiv \varphi+C_2$ for any constant $C_2 \leq -\|\varphi\|_{L^\infty(\mathbb{T}^n)}$ are solutions to \eqref{eq:mag1} and \eqref{eq:mag2}. Besides, by stability results for convex Hamiltonian, we have further that \[ u_3 = \min\{u_1,u_2\} = \min\{C_1, \varphi+C_2\} \] is also a solution to \eqref{eq:mag1} and \eqref{eq:mag2}. See also \cite[Example 6.2]{LMT}. Note that we do not claim here that we have described all solutions to \eqref{eq:mag1}. } \end{ex} \section{Further analysis on a uniqueness set for a prototype case} \label{sec:prototype} Let us now come back to Hamiltonians of type in Example \ref{ex-5} to do further analysis. In this section, we always assume (H6). That is, we consider a general class of Hamiltonian of the form \[ H(x,r,p)=|p|^m - V(x) + f(r) \quad \text{ for } (x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n. \] Here, $m \geq 1$ is a given number, and $V \in C(\mathbb{T}^n)$ is the potential energy with $\min_{\mathbb{T}^n} V=0$. The function $f:\mathbb{R} \to \mathbb{R}$ is convex, and \[ \begin{cases} f(r)=0 \quad &\text{ for } r \leq 0,\\ f(r)>0 \quad &\text{ for } r>0. \end{cases} \] It is clear that the Hamiltonian in Example \ref{ex-5} is a specific case of this class. Our goal here is to analyze more about solutions of (E) for fixed $c\geq 0$. We first show that $f$ is nondecreasing. \begin{lem} Let $f \in C(\mathbb{R})$ be given as above. Then $f$ is nondecreasing. \end{lem} \begin{proof} Take $0<r<s$. By the convexity of $f$, we have \[ 0<f(r) \leq \frac{r}{s} f(s) + \left(1-\frac{r}{s}\right) f(0) = \frac{r}{s} f(s) \leq f(s). \] \end{proof} We give a proof of our first uniqueness result when $c>0$. \begin{proof}[Proof of Proposition \ref{prop:unique-positive}] As we explain in Section \ref{sec:exist}, Example \ref{ex-5}, for every $c\ge0$, (E) has viscosity solutions. Let $u \in C(\mathbb{T}^n)$ be a solution to (E) with the given $c>0$ on the right hand side, that is, \begin{equation}\label{E-c} |Du|^m - V(x) + f(u) = c \quad \text{ in } \mathbb{T}^n. \end{equation} Then, $f(u) \leq V+c$, which means that $u \leq C$. Next, pick $x_1 \in \mathbb{T}^n$ so that $u(x_1) = \min_{\mathbb{T}^n} u$. By the viscosity supersolution test, \[ -V(x_1)+ f(u(x_1)) \geq c \quad \Rightarrow \quad f(u(x_1)) \geq c >0 \quad \Rightarrow \quad u(x_1) \geq \bar c = f^{-1}(c)>0. \] Therefore, $ \bar c \leq u \leq C$. Since $f$ is convex and increasing, we can find $0<\lambda \leq \Lambda$ such that \[ \lambda \leq f'(r) \leq \Lambda \quad \text{ for a.e. } r \in [\bar c, C]. \] We now can apply classical theory of viscosity solution to imply the uniqueness of solutions to \eqref{E-c}. For convenience later on, denote by $u_c$ this unique solution. \end{proof} One key feature we used in the above proof is that $\phi \equiv 0$ is a subsolution to (E) for all $c\geq 0$. In particular, for $c>0$, $\phi \equiv 0$ is a strict subsolution, and therefore, we were able to get that $u>0$. On the other hand, for $c=0$, we have seen in Examples \ref{ex-5} and \ref{ex-6} that we do not have uniqueness for (E). It turns out that $M_V$ is a uniqueness set for (E) in this case, which is exactly the content of the following proof. \begin{proof}[Proof of Proposition \ref{prop:unique-0}] Assume that $(u_1,0), (u_2,0)$ are two solutions to {\rm (E)}, and $u_1=u_2$ on $M_V$. If $M_V = \mathbb{T}^n$, there is nothing to prove. Hence, we assume below that $\mathbb{T}^n\setminus M_V$ is nonempty. Assume by contradiction that there exists $x_0\in\mathbb{T}^n\setminus M_V$ such that \[ \max_{\mathbb{T}^n} (u_1-u_2) = u_1(x_0) - u_2(x_0) >0. \] Take $\lambda \in (0,1)$, which is very close to $1$, such that $\lambda u_1(x_0) > u_2(x_0)$, and \[ \lambda u_1(x_0) - u_2(x_0) > \lambda u_1(x) - u_2(x) = -(1-\lambda)u_1(x) \quad \text{ for all } x\in M_V. \] Then, \[ \max_{\mathbb{T}^n}(\lambda u_1-u_2)=(\lambda u_1-u_2)(x_\lambda)>0 \] for some $x_\lambda\in\mathbb{T}^n\setminus M_V$. Due to the convexity of $r \mapsto f(r)$ and $p \mapsto |p|^m$, denote by $v=\lambda u_1= (1-\lambda)0+\lambda u_1$. Then, $v$ satisfies \[ |Dv|^m -\lambda V+ f(v) \leq 0 \quad \text{ in } \mathbb{T}^n. \] We now perform the usual doubling variables technique. For $\varepsilon>0$, consider the auxiliary function \[ \Phi^\varepsilon(x,y) = v(x) - u_2(y) - \frac{|x-y|^2}{2\varepsilon}. \] Then, $\Phi^\varepsilon$ admits a maximum at $(x_\varepsilon, y_\varepsilon)$, and by passing to a subsequence if needed, $(x_\varepsilon, y_\varepsilon) \to (x_\lambda, x_\lambda)$ as $\varepsilon \to 0$. By the viscosity solution tests, we have \[ \left|\frac{x_\varepsilon-y_\varepsilon}{\varepsilon} \right|^m - \lambda V(x_\varepsilon) + f(v(x_\varepsilon)) \leq 0, \] and \[ \left|\frac{x_\varepsilon-y_\varepsilon}{\varepsilon} \right|^m - V(y_\varepsilon) + f(u_2(y_\varepsilon)) \geq 0. \] Combine the two inequalities above to yield \[ -\lambda V(x_\varepsilon) + f(v(x_\varepsilon)) \leq - V(y_\varepsilon) + f(u_2(y_\varepsilon)). \] Then, let $\varepsilon \to 0$ to get further that \[ -\lambda V(x_\lambda) + f(v(x_\lambda)) \leq - V(x_\lambda) + f(u_2(x_\lambda)). \] Since, $v(x_\lambda) > u_2(x_\lambda)$, $f(v(x_\lambda)) \geq f(u_2(x_\lambda))$. Thus, we end up with a contradiction as $V(x_\lambda)>0$. The proof is complete. \end{proof} \section{Uniqueness structure of solutions to (E)} \label{sec:structure} In this section, we always assume (H1), (H2b), (H3), (H4), and (H5). Recall that after normalization as explained in Introduction, we assume further that $c=0$, and the ergodic problem becomes \begin{equation*} H(x,u,Du)=0 \quad \text{ in } \mathbb{T}^n. \end{equation*} In \cite{WWY-1}, the authors put an admissible condition (see \cite[Assumption (A), Theorem 1.1]{WWY-1}) to guarantee that (E) has a viscosity solution with $c=0$. Note further that in \cite{WWY-1}, they need to require a stronger condition, that is, $H_r(x,r,p)>0$, than our (H4), which basically guarantees the uniqueness of viscosity solutions to (E). See \cite{GMT, CCIZ} for related works. We now use the nonlinear adjoint method to study \eqref{E-0}. \subsection{Preliminaries} Here is a first preparatory lemma. Since this is elementary, we omit the proof. \begin{lem}\label{lem:conv} Let $u \in {\rm Lip\,}(\mathbb{T}^n)$ be a solution to \eqref{E-0}. Let $\rho \in C_c^\infty(\mathbb{R}^n,[0,\infty))$ be a standard mollifier. For $\delta>0$, let $\rho^\delta(x) = \delta^{-n} \rho(\delta^{-1}x)$ for all $x\in \mathbb{R}^n$. Denote by $u^\delta = \rho^\delta * u$. Then, \[ \|u^\delta - u\|_{L^\infty(\mathbb{T}^n)} \leq C\delta, \] and \[ \|Du^\delta\|_{L^\infty(\mathbb{T}^n)} + \delta \|D^2 u^\delta\|_{L^\infty(\mathbb{T}^n)} \leq C. \] \end{lem} Let us consider the following Cauchy problems \begin{equation}\label{eq:C1} \begin{cases} \varepsilon w^\varepsilon_t + H(x,w^\varepsilon,Dw^\varepsilon) = \varepsilon^4 \Delta w^\varepsilon \quad &\text{ in } \mathbb{T}^n \times (0,1),\\ w^\varepsilon(x,0) = u^{\varepsilon^4}(x) \quad &\text{ on } \mathbb{T}^n, \end{cases} \end{equation} and \begin{equation}\label{eq:C2} \begin{cases} \varepsilon \phi^\varepsilon_t + H(x,\phi^\varepsilon,D\phi^\varepsilon) = \varepsilon^4 \Delta \phi^\varepsilon \quad &\text{ in } \mathbb{T}^n \times (0,1),\\ \phi^\varepsilon(x,0) = u(x) \quad &\text{ on } \mathbb{T}^n. \end{cases} \end{equation} Here, $u^{\varepsilon^4}$ is $u^\delta$ with $\delta=\varepsilon^4$. \begin{lem}\label{lem:close} We have \[ \|w^\varepsilon - \phi^\varepsilon\|_{L^\infty(\mathbb{T}^n \times [0,1])} \leq C\varepsilon^4. \] \end{lem} \begin{proof} Recall that $\|u^{\varepsilon^4} - u\|_{L^\infty(\mathbb{T}^n)} \leq C\varepsilon^4$. Let $\varphi(x,t) = w^\varepsilon(x,t) + C\varepsilon^4$ for $(x,t) \in \mathbb{T}^n \times [0,1]$, then $\varphi(x,0) \geq \phi^\varepsilon(x,0)$ for $x\in \mathbb{T}^n$. Besides, thanks to (H4), \begin{align*} &\varepsilon \varphi_t + H(x,\varphi, D\varphi) - \varepsilon^4 \Delta \varphi\\ =\,& \varepsilon w^\varepsilon_t + H(x,w^\varepsilon+ C\varepsilon^4,Dw^\varepsilon) - \varepsilon^4 \Delta w^\varepsilon \geq \varepsilon w^\varepsilon_t + H(x,w^\varepsilon ,Dw^\varepsilon) - \varepsilon^4 \Delta w^\varepsilon =0, \end{align*} which means that $\varphi$ is a supersolution of \eqref{eq:C2}. By the comparison principle, $\phi^\varepsilon \leq \varphi$, and thus, \[ \phi^\varepsilon(x,t) \leq w^\varepsilon(x,t) + C\varepsilon^4 \quad \text{ for all } (x,t) \in \mathbb{T}^n \times [0,1]. \] By a symmetric argument, the proof is complete. \end{proof} The next result concerns gradient bound of $w^\varepsilon$. \begin{lem}\label{lem:grad-bound} There is a constant $C>0$ independent of $\varepsilon>0$ such that \[ \varepsilon \|w^\varepsilon_t\|_{L^\infty(\mathbb{T}^n \times [0,1])} + \|Dw^\varepsilon\|_{L^\infty(\mathbb{T}^n \times [0,1])} \leq C. \] \end{lem} \begin{proof} Denote by \[ \varphi^{\pm}(x,t) = w^\varepsilon(x,0) \pm \frac{C}{\varepsilon}t \quad \text{ for all } (x,t) \in \mathbb{T}^n \times [0,1]. \] Then, $\varphi^-, \varphi^+$ are, respectively, a subsolution, and a supersolution to \eqref{eq:C1}, thanks to (H4). Hence, by the comparison principle, \[ \varphi^- \leq w^\varepsilon \leq \varphi^+ \quad \Rightarrow \quad\|w^\varepsilon(\cdot,s)-w^\varepsilon(\cdot,0)\|_{L^\infty} \leq \frac{Cs}{\varepsilon}. \] Note next that both $w^\varepsilon$ and $w^\varepsilon(\cdot, \cdot+s)$ solve \eqref{eq:C1} with initial data $w^\varepsilon(\cdot,0)$ and $w^\varepsilon(\cdot,s)$, respectively. By the comparison principle, \[ \|w^\varepsilon(\cdot, \cdot+s)-w^\varepsilon\|_{L^\infty} \leq \|w^\varepsilon(\cdot,s)-w^\varepsilon(\cdot,0)\|_{L^\infty} \leq \frac{Cs}{\varepsilon} \quad \Rightarrow \quad \varepsilon \|w^\varepsilon_t\|_{L^\infty(\mathbb{T}^n)} \leq C. \] To prove the spatial gradient bound, we use the usual Bernstein method. Let $\psi(x,t) = \frac{|Dw^\varepsilon|^2}{2}$. Then $\psi$ satisfies \[ \varepsilon \psi_t + D_pH\cdot D\psi + 2 H_r \psi + D_xH\cdot Dw^\varepsilon = \varepsilon^4 \Delta \psi - \varepsilon^4 |D^2 w^\varepsilon|^2. \] Assume that $\max_{\mathbb{T}^n \times [0,1]} \psi = \psi(x_0,t_0)$. If $t_0=0$, then we are done. If $t_0 >0$, then by the maximum principle, noting that $H_r \geq 0$, \[ D_xH \cdot Dw^\varepsilon + \varepsilon^4 |D^2 w^\varepsilon|^2 \leq 0 \quad \text{ at } (x_0,t_0). \] For $\varepsilon<n^{-1}$, we have \[ \varepsilon^4 |D^2 w^\varepsilon|^2 \geq (\varepsilon^4 \Delta w^\varepsilon)^2 = (\varepsilon w^\varepsilon_t + H(x,w^\varepsilon, Dw^\varepsilon))^2 \geq \frac{1}{2} H(x,w^\varepsilon, Dw^\varepsilon)^2 - C. \] Therefore, \[ \frac{1}{2} H(x,w^\varepsilon, Dw^\varepsilon)^2 +D_xH \cdot Dw^\varepsilon \leq C \quad \text{ at } (x_0,t_0), \] which, together with (H3), yields the desired result. \end{proof} \begin{lem}\label{lem:ep} We have \[ \|w^\varepsilon - u\|_{L^\infty(\mathbb{T}^n \times [0,1])}+\|\phi^\varepsilon - u\|_{L^\infty(\mathbb{T}^n \times [0,1])} \leq C\varepsilon. \] \end{lem} The proof of this is similar to that of \cite[Proposition 5.5]{LMT}. Nevertheless, let us present a simple proof here for completeness. \begin{proof} We only need to show that $\|w^\varepsilon - u\|_{L^\infty(\mathbb{T}^n \times [0,1])} \leq C\varepsilon$. Let us first get an upper bound for $w^\varepsilon - u$. Define an auxiliary function \[ \Phi(x,y,t) = w^\varepsilon(x,t) - u(y) - \frac{|x-y|^2}{2 \varepsilon^2} - K\varepsilon t \quad \text{ for } (x,y,t) \in \mathbb{T}^n \times \mathbb{T}^n \times [0,1], \] where $K>0$ is to be chosen. Pick $(x_\varepsilon, y_\varepsilon, t_\varepsilon ) \in \mathbb{T}^n \times \mathbb{T}^n \times [0,1]$ so that \[ \Phi(x_\varepsilon, y_\varepsilon, t_\varepsilon ) = \max_{ \mathbb{T}^n \times \mathbb{T}^n \times [0,1]} \Phi. \] If $\Phi(x_\varepsilon, y_\varepsilon, t_\varepsilon ) \leq 0$, then we are done as \[ w^\varepsilon(x,t) - u(x) = \Phi(x,x,t) +K \varepsilon t \leq K \varepsilon. \] Therefore, we can assume $\Phi(x_\varepsilon, y_\varepsilon, t_\varepsilon ) > 0$. This gives that $w^\varepsilon(x_\varepsilon,t_\varepsilon)>u(y_\varepsilon)$. Let us consider first the case that $t_\varepsilon>0$. Since $w^\varepsilon$ and $u$ are Lipschitz, by comparing $\Phi(x_\varepsilon,y_\varepsilon,t_\varepsilon)$ with $\Phi(y_\varepsilon,y_\varepsilon,t_\varepsilon)$, we deduce first that \[ |x_\varepsilon - y_\varepsilon | \leq C\varepsilon^2. \] By the viscosity subsolution and supersolution tests, we have \[ K \varepsilon^2 + H\left(x_\varepsilon, w^\varepsilon(x_\varepsilon, t_\varepsilon), \frac{x_\varepsilon - y_\varepsilon}{\varepsilon^2} \right) \leq \varepsilon^4 \frac{n}{\varepsilon^2} = n\varepsilon^2, \] and \[ H\left(y_\varepsilon, u(y_\varepsilon), \frac{x_\varepsilon - y_\varepsilon}{\varepsilon^2} \right) \geq 0. \] Combine these two inequalities, and use (H3), (H4) to imply \begin{align*} K \varepsilon^2 &\leq n\varepsilon^2+ H\left(y_\varepsilon, u(y_\varepsilon), \frac{x_\varepsilon - y_\varepsilon}{\varepsilon^2} \right) - H\left(x_\varepsilon, w^\varepsilon(x_\varepsilon, t_\varepsilon), \frac{x_\varepsilon - y_\varepsilon}{\varepsilon^2} \right)\\ &\leq n \varepsilon^2 + C|y_\varepsilon - x_\varepsilon| +H\left(x_\varepsilon, u(y_\varepsilon), \frac{x_\varepsilon - y_\varepsilon}{\varepsilon^2} \right) - H\left(x_\varepsilon, w^\varepsilon(x_\varepsilon, t_\varepsilon), \frac{x_\varepsilon - y_\varepsilon}{\varepsilon^2} \right)\\ & \leq n \varepsilon^2 + C|y_\varepsilon - x_\varepsilon| \leq (C+n) \varepsilon^2. \end{align*} By picking $K = C+n+1$, we conclude that $t_\varepsilon$ cannot be positive. Thus, $t_\varepsilon = 0$, and \[ \Phi(x_\varepsilon, y_\varepsilon, t_\varepsilon ) \leq u^{\varepsilon^4}(x_\varepsilon) -u(y_\varepsilon) \leq C\varepsilon^4 + C|x_\varepsilon - y_\varepsilon| \leq C\varepsilon^2. \] Then, for $(x,t) \in \mathbb{T}^n \times [0,1]$, \[ w^\varepsilon(x,t) - u(x) = \Phi(x,x,t) +K \varepsilon t \leq C\varepsilon^2+ K \varepsilon \leq C\varepsilon. \] To get the other bound, we need to get an upper bound of $u-w^\varepsilon$. This can be done analogously to the above by carefully considering another auxiliary function \[ \Psi(x,y,t) = u(x) - w^\varepsilon(y,t) - \frac{|x-y|^2}{2 \varepsilon^2} - K\varepsilon t \quad \text{ for } (x,y,t) \in \mathbb{T}^n \times \mathbb{T}^n \times [0,1], \] where $K>0$ is to be chosen. We omit the proof of this part here. \end{proof} \subsection{Nonlinear adjoint method and adjoint measures} Let $u \in {\rm Lip\,}(\mathbb{T}^n)$ be a solution to \eqref{E-0}. Let $w^\varepsilon$ be the solution to \eqref{eq:C1} with this fixed $u$, that is, $w^\varepsilon(x,0) = u^{\varepsilon^4}(x)$ in $\mathbb{T}^n$. The linearized operator of \eqref{eq:C1} about the solution $w^\varepsilon$ is \[ \mathcal{L}^\varepsilon[\phi]= \varepsilon \phi_t + H_r(x,w^\varepsilon,Dw^\varepsilon) \phi + D_pH(x,w^\varepsilon,Dw^\varepsilon)\cdot D\phi - \varepsilon^4 \Delta \phi. \] The corresponding adjoint equation is \begin{equation}\label{eq:sig} \begin{cases} -\varepsilon \sigma^\varepsilon_t +H_r(x,w^\varepsilon,Dw^\varepsilon) \sigma^\varepsilon - {\rm div}(D_pH(x,w^\varepsilon,Dw^\varepsilon) \sigma^\varepsilon) = \varepsilon^4 \Delta \sigma^\varepsilon \quad &\text{in } \mathbb{T}^n \times (0,1),\\ \sigma^\varepsilon(x,1) = \delta_{x_0}. \end{cases} \end{equation} Here, $\delta_{x_0}$ is the Dirac delta measure at $x_0 \in \mathbb{T}^n$. It is clear that $\sigma^\varepsilon>0$ in $\mathbb{T}^n \times (0,1)$. \begin{prop}\label{prop:sig} The following holds \begin{align*} & \frac{d}{dt}\int_{\mathbb{T}^n}\sigma^\varepsilon(x,t)\,dx = \frac{1}{\varepsilon}\int_{\mathbb{T}^n}H_r(x,w^\varepsilon,Dw^\varepsilon)\sigma^\varepsilon\,dx\ge0, \\ & 0\le \int_{\mathbb{T}^n}\sigma^\varepsilon(x,t)\,dx\le 1 \quad\text{ for all } \quad 0\le t < 1. \end{align*} \end{prop} \begin{proof} For $t\in (0,1)$, integrate \eqref{eq:sig} on $\mathbb{T}^n$ to yield \begin{align*} \varepsilon\dfrac{d}{dt}\int_{\mathbb{T}^n}\sigma^\varepsilon\,dx &\, =\int_{\mathbb{T}^n} H_r(x,w^\varepsilon,Dw^\varepsilon)\sigma^\varepsilon-{\rm div}\,(D_pH(x,w^\varepsilon,Dw^\varepsilon)\sigma^\varepsilon) -\varepsilon^4\Delta \sigma^\varepsilon\,dx\\ &\, =\int_{\mathbb{T}^n}H_r(x,w^\varepsilon,Dw^\varepsilon)\sigma^\varepsilon\,dx\ge0, \end{align*} which gives the first claim. The second claim follows immediately. \end{proof} For each $\sigma^\varepsilon$, there exist a nonnegative Radon measure $\nu^\varepsilon \in\mathcal{R}(\mathbb{T}^n )$ satisfying \[ \int_{0}^{1}\int_{\mathbb{T}^n}\psi(x)\sigma^\varepsilon(x,t)\,dxdt =\int_{\mathbb{T}^n}\psi(x)\,d\nu^\varepsilon(x) \quad \text{ for all } \psi \in C(\mathbb{T}^n). \] In fact, for a Borel measurable set $A\subset \mathbb{T}^n$, \[ \nu^\varepsilon(A) = \int_0^1 \int_A \sigma^\varepsilon(x,t)\,dxdt. \] By Proposition \ref{prop:sig}, $\nu^\varepsilon(\mathbb{T}^n) \leq 1$. We are able to pick a subsequence $\{\varepsilon_j\} \to 0$ such that \[ \nu^{\varepsilon_j}\rightharpoonup \nu \] as $j\to\infty$ weakly in the sense of measures. \begin{defn}\label{def:M} We define the set $\mathcal{M}\subset\mathcal{R}(\mathbb{T}^n)$ as \[ \mathcal{M}:=\bigcup_{\substack{u \in \mathcal{S}, \, x_0\in\mathbb{T}^n\\ \{\varepsilon_j\} \to 0}}\{\nu\}, \] where $\mathcal{S}$ denote the family of all viscosity solutions to \eqref{E-0}. Here, we collect all possible subsequential weak limits (in the sense of measure) of $\{\nu^\varepsilon\}$ for all $x_0 \in \mathbb{T}^n$. We call each measure $\nu \in \mathcal{M}$ an adjoint measure of \eqref{E-0}. We say that $\mathcal{M}$ is the set of adjoint measures corresponding to \eqref{E-0}. \end{defn} \begin{rem} It is important noting that the set $\mathcal{M}$ is defined implicitly as it depends on all solutions to \eqref{E-0}, which are not known a priori. \end{rem} It turns out that the adjoint measures give us the uniqueness property of solutions to \eqref{E-0} as stated in Theorem \ref{thm:unique}. Here is the proof of our main theorem on uniqueness property. \begin{proof}[Proof of Theorem \ref{thm:unique}] For $i=1,2$, let $w^\varepsilon_i$ be the solution to \eqref{eq:C1} with initial data $u_i^{\varepsilon^4}$. By the convexity assumption (H5), we subtract the equations for $w^\varepsilon_1$ and $w^\varepsilon_2$ to get \begin{multline}\label{eq:diff} \varepsilon (w^\varepsilon_1 - w^\varepsilon_2)_t + H_r(x,w^\varepsilon_2, Dw^\varepsilon_2)(w^\varepsilon_1 - w^\varepsilon_2)+ D_pH(x,w^\varepsilon_2, Dw^\varepsilon_2)\cdot D(w^\varepsilon_1 - w^\varepsilon_2)\\ \leq \varepsilon^4 \Delta (w_1^\varepsilon - w_2^\varepsilon). \end{multline} Let $\sigma^\varepsilon$ be the solution to \begin{equation*} \begin{cases} -\varepsilon \sigma^\varepsilon_t +H_r(x,w^\varepsilon_2,Dw^\varepsilon_2) \sigma^\varepsilon - {\rm div}(D_pH(x,w^\varepsilon_2,Dw^\varepsilon_2) \sigma^\varepsilon) = \varepsilon^4 \Delta \sigma^\varepsilon \quad &\text{in } \mathbb{T}^n \times (0,1),\\ \sigma^\varepsilon(x,1) = \delta_{x_0}, \end{cases} \end{equation*} for $x_0 \in \mathbb{T}^n$ fixed. Multiply \eqref{eq:diff} by $\sigma^\varepsilon$, integrate on $\mathbb{T}^n$ to imply \[ \frac{d}{dt} \int_{\mathbb{T}^n} (w^\varepsilon_1 - w^\varepsilon_2)\sigma^\varepsilon\,dx \leq 0. \] In particular, \[ (w_1^\varepsilon - w_2^\varepsilon)(x_0,1) \leq \int_0^1 \int_{\mathbb{T}^n} (w^\varepsilon_1 - w^\varepsilon_2)\sigma^\varepsilon\,dxdt \] By letting $\varepsilon \to 0$ (and passing to a subsequence if needed) and using Lemma \ref{lem:ep}, we deduce that \[ (u_1-u_2)(x_0) \leq \int_{\mathbb{T}^n} (u_1 - u_2) \,d\nu(x) \leq 0, \] for some $\nu \in \mathcal{M}$. \smallskip Thus, $u_1(x_0) \leq u_2(x_0)$ for every $x_0 \in \mathbb{T}^n$. The proof is complete. \end{proof} Set \[ M := \overline{ \bigcup_{\nu \in \mathcal{M}} {\rm supp} (\nu)}. \] Then we have the following corollary, which is an immediate consequence of Theorem \ref{thm:unique}. \begin{cor} Let $u_1, u_2$ be two solutions to \eqref{E-0}. Assume that $u_1 \leq u_2$ on $M$. Then, $u_1 \leq u_2$. \end{cor} \begin{rem} Of course, this uniqueness result tells us that it is extremely important to have further understanding of the adjoint measures in $\mathcal{M}$. In case that $H(x,r,p) = \tilde H(x,p)$ for all $(x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n$ with $\tilde H$ satisfies appropriate conditions, then these adjoint measures $\nu \in \mathcal{M}$ turn out to be projected Mather measures (see \cite{FaB, MT-P}). In the general setting, one objection one might have is that $\mathcal{M}$ is defined in an abstract way, which depends on the set of solutions of \eqref{E-0} itself, and it is not clear how to analyze it. This is a fair point, and $\mathcal{M}$ should be studied much more in the near future. In particular, one question of interests is whether $\mathcal{M}$ can be defined without using $\mathcal{S}$. Nevertheless, in the following two interesting situations, we are able to provide full characterization of $\mathcal{M}$. These results are consistent with the classical literature of viscosity solutions, and also with Proposition \ref{prop:unique-0}. \end{rem} \subsection{Adjoint measures in the strictly monotone case} \label{subsec:structure1} We have the following result, which is consistent with the classical literature of viscosity solutions \cite{CEL, CL}. \begin{prop}\label{prop:st-inc} Assume that \[ H_r(x,r,p) >0 \quad \text{ for all } (x,r,p)\in\mathbb{T}^n\times \mathbb{R}\times\mathbb{R}^n. \] Then, \[ \mathcal{M}=\{0\}, \quad\text{and} \quad M=\emptyset. \] \end{prop} \begin{proof} Let $u$ be a solution of \eqref{E-0}. Let $w^\varepsilon$ be the solution to \eqref{eq:C1} with initial data $u^{\varepsilon^4}$. Let $C_2=\|u\|_{L^\infty(\mathbb{T}^n)}+\|Du\|_{L^\infty(\mathbb{T}^n)}$. Then, for $\varepsilon<1$, we can find $C_3>0$ such that $\|w^\varepsilon\|_{L^\infty(\mathbb{T}^n)}+\|Dw^\varepsilon\|_{L^\infty(\mathbb{T}^n)} \leq C_3$. By our hypothesis, we are able to find $\alpha>0$ such that \[ H_r(x,w^\varepsilon,Dw^\varepsilon) \geq \alpha \quad \text{ for all } x\in \mathbb{T}^n. \] Use this in the adjoint equation to deduce \[ -\varepsilon\sigma^\varepsilon_t+\alpha\sigma^\varepsilon-{\rm div}\,(D_pH(x,w^\varepsilon,Dw^\varepsilon)\sigma^\varepsilon) \le \varepsilon^4\Delta \sigma^\varepsilon, \] which implies \[ -\varepsilon(e^{-\frac{\alpha t}{\varepsilon}}\sigma^\varepsilon)_t-{\rm div}\,(D_pH(x,w^\varepsilon,Dw^\varepsilon)e^{-\frac{\alpha t}{\varepsilon}}\sigma^\varepsilon) \le \varepsilon^4\Delta (e^{-\frac{\alpha t}{\varepsilon}} \sigma^\varepsilon). \] Integrate the above on $\mathbb{T}^n$ to obtain \[ \dfrac{d}{dt}\left(\int_{\mathbb{T}^n}e^{-\frac{\alpha t}{\varepsilon}}\sigma^\varepsilon\,dx\right)\ge0 \quad\text{ for all } \ t\in (0,1). \] Thus, \[ \int_{\mathbb{T}^n}\sigma^\varepsilon(x,t) \,dx\le e^{-\frac{\alpha(1-t)}{\varepsilon}} \quad\text{ for all } \ t\in (0,1), \] which gives \[ \int_{\mathbb{T}^n}\,d\nu^\varepsilon(x) =\int_0^1\int_{\mathbb{T}^n}\sigma^\varepsilon(x,t) \,dxdt \leq \frac{\varepsilon}{\alpha}(1-e^{-\frac{\alpha}{\varepsilon}}). \] Sending $\varepsilon\to0$ yields the conclusion. \end{proof} Basically, Proposition \ref{prop:st-inc} says that in the strictly monotone setting, if $u_1, u_2$ are solutions to \eqref{E-0}, there is no need to compare $u_1$ and $u_2$ anywhere, and we have immediately $u_1=u_2$, which of course means that we have the unique viscosity solution to \eqref{E-0}. \subsection{Prototype example -- Revisit}\label{subsec:structure2} Let us now revisit our prototype example in Section \ref{sec:prototype}. Since we need smoothness of $H$, we consider \[ H(x,r,p)=|p|^m - V(x) + f(r) \quad \text{ for } (x,r,p) \in \mathbb{T}^n \times \mathbb{R} \times \mathbb{R}^n. \] Here, $m \geq 2$ is a given number, and $V \in C^2(\mathbb{T}^n)$ is the potential energy with $\min_{\mathbb{T}^n} V=0$. The function $f \in C^2(\mathbb{R})$ is convex, and \[ \begin{cases} f(r)=0 \quad &\text{ for } r \leq 0,\\ f(r)>0 \quad &\text{ for } r>0. \end{cases} \] It is clear that $f'(r)>0$ for $r>0$, and $f'$ is nondecreasing. In particular, \begin{equation}\label{eq:f-p} f'(s) \geq f'(r)>0 \quad \text{ for all } s\geq r>0. \end{equation} This observation will be used later on. An example of $f \in C^2(\mathbb{R})$ satisfying the above is $f(r) = (\max\{r,0\})^3$. \begin{prop}\label{prop:prot} Assume the setting in this subsection. Then, for each $\nu \in \mathcal{M}$, \[ {\rm supp}(\nu) \subset M_V=\left\{x\in \mathbb{T}^n\,:\, V(x) = \min_{\mathbb{T}^n} V =0\right\}. \] \end{prop} \begin{proof} Let $u$ be a solution of \eqref{E-0}. Let $w^\varepsilon$ be the solution to \eqref{eq:C1} with initial data $u^{\varepsilon^4}$. The corresponding adjoint equation is \[ -\varepsilon \sigma^\varepsilon_t + f'(w^\varepsilon) \sigma^\varepsilon -{\rm div}\,(m |Dw^\varepsilon|^{m-2} Dw^\varepsilon \sigma^\varepsilon) = \varepsilon^4 \Delta \sigma^\varepsilon. \] Integrate this on $\mathbb{T}^n$ to get that \[ \varepsilon \frac{d}{dt} \int_{\mathbb{T}^n} \sigma^\varepsilon(x,t)\,dx = \int_{\mathbb{T}^n} f'(w^\varepsilon) \sigma^\varepsilon\,dx. \] Next, integrate the above in $t$ on $[0,1]$ to deduce further that \[ \int_0^1 \int_{\mathbb{T}^n} f'(w^\varepsilon) \sigma^\varepsilon \,dxdt = \varepsilon \left(1 - \int_{\mathbb{T}^n} \sigma^\varepsilon (x,0)\,dx\right) \leq \varepsilon. \] Letting $\varepsilon = \varepsilon_j \to 0$ to yield, thanks to Lemma \ref{lem:ep}, \[ \int_{\mathbb{T}^n} f'(u)\, d\nu = 0. \] Thus, by using \eqref{eq:f-p}, we arrive at the fact that ${\rm supp} (\nu) \subset \{ u \leq 0\}$ for each $u \in \mathcal{S}$. We plan to pick an appropriate solution $u$ in $\mathcal{S}$ to conclude. Now, for each $c>0$, let $(u_c,c)$ be the unique solution to (E). Note that $u_c>0$. Of course, for $c\in (0,1]$, there exists $C>0$ independent of $c$ such that \[ \|u_c\|_{L^\infty(\mathbb{T}^n)} + \|Du_c\|_{L^\infty(\mathbb{T}^n)} \leq C. \] By using the Arzel\`a-Ascoli theorem, and passing to a subsequence if needed, $u_c \to u_0$ uniformly in $\mathbb{T}^n$ as $c \to 0$. By stability of viscosity solutions, $u_0$ is a solution of \eqref{E-0}. It is clear that $u_0 \geq 0$, and \[ f(u_0) \leq V \quad \text{ in } \mathbb{T}^n, \] which gives that $\{V=0\} \subset \{u_0=0\}$. Besides, for any $x_1 \in \mathbb{T}^n$ such that $u_0(x_1) = 0 = \min_{\mathbb{T}^n} u_0$, by the viscosity supersolution test, \[ 0=f(0)=f(u_0(x_1)) \geq V(x_1) \quad \Rightarrow \quad V(x_1)=0. \] Thus, $\{V=0\} = \{u_0=0\}$, and hence, ${\rm supp} (\nu) \subset M_V=\{V=0\}$. \end{proof} This last proposition is consistent with the result of Proposition \ref{prop:unique-0}. It is clear that we get $M \subset M_V$. Nevertheless, we do not get that $M=M_V$ here, and it is not clear if this holds in general. It would be very interesting if there is an example where $M \subsetneq M_V$. \section*{Acknowledgments} The work of WJ has been supported by the Recruitment Program of Global Experts of China and by the National Natural Science Foundation of China under Grant No.\,11701314. The work of HM was partially supported by the JSPS grant KAKENHI \#16H03948. The work of HT is partially supported by NSF grant DMS-1664424. \bibliographystyle{amsplain} \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace} \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR } \providecommand{\MRhref}[2]{% \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2} } \providecommand{\href}[2]{#2}

https://arxiv.org/abs/1406.0774

Set Theory or Higher Order Logic to Represent Auction Concepts in Isabelle?

When faced with the question of how to represent properties in a formal proof system any user has to make design decisions. We have proved three of the theorems from Maskin's 2004 survey article on Auction Theory using the Isabelle/HOL system, and we have produced verified code for combinatorial Vickrey auctions. A fundamental question in this was how to represent some basic concepts: since set theory is available inside Isabelle/HOL, when introducing new definitions there is often the issue of balancing the amount of set-theoretical objects and of objects expressed using entities which are more typical of higher order logic such as functions or lists. Likewise, a user has often to answer the question whether to use a constructive or a non-constructive definition. Such decisions have consequences for the proof development and the usability of the formalization. For instance, sets are usually closer to the representation that economists would use and recognize, while the other objects are closer to the extraction of computational content. In this paper we give examples of the advantages and disadvantages for these approaches and their relationships. In addition, we present the corresponding Isabelle library of definitions and theorems, most prominently those dealing with relations and quotients.

\section{Introduction} \label{sec:introduction} \label{RefIntro} \lnote{MC: Rephrased after R3 noticed there is also type theory as a third strong competitor.} When representing mathematics in formal proof systems, alternative foundations can be used, with two important examples being set theory (e.g., Mizar{} takes this approach) and higher order logic (e.g., as in Isabelle/HOL). Another dimension in the representation is the difference between classical and constructive approaches. Again, there are systems which are predominantly classical (as most first order automated theorem provers) and constructive (e.g., Coq). Isabelle{}/HOL is flexible enough to enable the user to take these different approaches in the same system (e.g., although it is built on higher order logic, it contains a library for set theory, \verb|Set.thy|). For instance, participants in an auction, i.e.\ \emph{bidders}, can be represented by a predicate \verb|Bidder x| or alternatively as a list of bidders \verb|[b1, b2, b3]|. The difference between a classical and a constructive definition can be demonstrated by the example of \lnote{MC: Changed `maximal argument' to `argument of the maximum' for Reviewer2.} the argument of the maximum for a function (which we need for determining the winner of an auction). Classically we can define it, e.g., as \Verb[commandchars=\\\{\}]|arg_max f A = \{x \(\in\) A. f x = Max(f`A)\}|. This definition is easy to understand but, unlike a constructive one, it does not tell how to compute \verb|arg_max|. The constructive definition is more complicated (and has to consider different cases). It corresponds to a recursive function recurring on the elements of the domain. From a programming perspective, the two kinds of definitions just illustrated (classical versus constructive) can be seen respectively as \emph{specification} versus \emph{implementation}. The approaches coexist in Isabelle{}/HOL. For example, for a set \verb|X| one can apply the higher order function \verb|f| by the construction \verb|f ` X| to yield the image of the set. As a result, an author does not have to make a global decision of whether to use sets or higher order functions, but has fine grained control on what to use for a new mathematical object (e.g., sets vs.\ lists or lambda functions). Such a choice will typically depend on many factors. One factor is the task at hand (e.g., whether one needs to prove a theorem about an object or needs to compute its value). Another factor is naturalness of the constructions. This will typically depend on the authors and their expected audiences. Pragmatically, users also have to consider which of the possible approaches is more viable given the support provided by existing libraries for the system being used. The ForMaRE{} project~\cite{LRK:FormareProject13:short} applies formal methods to economics. \lnote{MC: Added this sentence for reviewer 1.} One of the branches of economics the project focuses on is \emph{auction theory}, which deals with the problem of allocating a set of resources among a set of participants while maximizing one or more parameters (e.g., revenue, or social welfare) in the process. ForMaRE{} has produced the Auction Theory Toolbox (ATT{}), containing Isabelle{} code for a range of auctions and theorems about them.% \footnote{See \url{https://github.com/formare/auctions/}; the state at the time of this writing is archived at \url{https://github.com/formare/auctions/tree/1f1e7035da2543a0645b9c44a5276229a0aeb478}. } Therefore, there is a good opportunity to practically test the feasibility of the different approaches, as introduced above, in this concrete setting. We adopted a pragmatic attitude: we typically took a set theoretic approach, since firstly we felt most familiar with it and secondly we knew from our ongoing interaction with economists that it would look more natural to them. However, when we needed to generate code, this generally excluded set-theoretical constructions such as the set comprehension notation. As a consequence most of our work was done in the set-theoretical realm, but some done constructively which allowed us to produce the code we wanted. The disadvantage of this approach is that we had to provide supplementary `compatibility{}' proofs to show the equivalence of the set theoretical and the computable definitions when needed. This means that, as a byproduct of our efforts, we also generated a good amount of generic set-theoretical material which was neither provided by the Isabelle{} library nor Isabelle{}'s Archive of Formal Proofs (AFP{})\cite{afp}. In this paper, rather than discussing our progress in the application domain (auction theory), we illustrate this material and its relationship to the mathematical objects we needed. We will discuss this by the following three concrete examples that occurred while we were working on the ATT: \begin{enumerate} \item We model a function indirectly through the set-theoretical notion of the graph of a relation, rather than directly using the HOL primitive notion of a lambda-abstracted function. This allowed us to concisely define auction-related notions through two natural set-theoretical constructions (extension and restriction of a function). Moreover, this in turn allowed us to generalize some theorems from functions to relations. We also discuss how this choice does not necessarily mean giving up the advantage of computability (section \ref{RefFuncAsGraph}). Incidentally, this will also give insight into the main differences between set theory as implemented in Isabelle{}/HOL and standard set theory. \item We developed a stand-alone formalization of the definition of functions over equivalence classes (section \ref{RefQuotients}). This is a common construction when, e.g., defining the canonically induced operation on the quotient of a group by one of its normal subgroups. To define such an operation, we proved the invariance (or well-definedness) under an equivalence relation. \item \label{RefDualDefsItem} We applied a set-theoretical, non-constructive definition and a constructive one to the same object (the set of all possible partitions of a set). The two approaches were used for different purposes (theorem proving and computation, respectively) and proved formally their equivalence by a `compatibility' theorem (section \ref{RefDualDefs}). The same approach is adopted for the set of all possible injective functions between sets. \end{enumerate} In existing proof assistants based on set theory (e.g., Mizar{}, Metamath), the desirable quality of being able to compute values of functions (rather than only the truth value of predicates involving them), is typically lost% \lnote{MC: Added footnote for reviewer3.}% \footnote{We do not know if this is due to how the existing proof assistants are implemented, or to some fundamental limitation of set theory. Reading the recent ``Computational set theory{}'' thread (\url{http://www.cs.nyu.edu/pipermail/fom/2014-February/017841.html}), and its related threads, on the ``Foundations of Mathematics'' mailing list, we feel the question is currently open.}% . We will present examples which show that we can retain this Isabelle{}/HOL-induced advantage in our developments. \subsection*{Overview of the paper} Section \ref{RefFuncAsGraph} introduces and motivates the set-theoretical encoding of functions, as an alternative to a lambda representation; this is done by commenting on the relevant Isabelle{} definitions, and by illustrating a particular theorem regarding auctions, which employs those definitions. Section \ref{RefQuotients} brings this approach to a more abstract level, by showing how it can handle, given an initial function, the definition of a second function on the corresponding equivalence classes. Since this was already done in Isabelle{}'s \verb|Equiv_Relations.thy|, we also illustrate the differences with it. In some cases, defining a mathematical object in purely set-theoretical terms does not preserve computability of that object: section \ref{RefDualDefs} introduces a technique for such cases% . As mentioned in point \ref{RefDualDefsItem} above, this technique introduces a parallel, computable definition, and then formally proves the equivalence of the two definitions. In the same section, this technique is presented through two examples from our ATT{}; we then discuss how we took advantage from having preserved both the definitions. Finally, section \ref{RefApplication} explains how we applied the general machinery from section \ref{RefQuotients} to the ATT{}. \section{Set-theoretical Definition of Functions in Isabelle{}/HOL} \label{RefFuncAsGraph} In higher order logic{}, functions, function abstraction, and function application are primitives~\cite{bowen1994z}; in set theory, the primitive notion is that of a set, and a function $f$ is represented by its graph, i.e., the \emph{set} of all the ordered pairs $\left( x, f\left( x \right) \right)$. We chose to work mostly with the set-theoretical representation of the functions even though we are using a tool based on higher order logic{}. In the following we give reasons why. \begin{enumerate} \item \label{RefMotivationEnumeration} A first reason is that a set theoretical representation more easily allows to enumerate all the (injective) functions from a finite set to another finite set (\cite{CKLR:SoundCombVickCode13}). In contrast, this seems to be more complicated to do directly in higher order logic, where all the functions are assumed be total. \item \label{RefMotivationRelations} A second reason is that the set-theoretical, graph-based representation works even for generic relations, thus often allowing us to extend the results we proved to the more general situations involving relations, rather than functions (see the next subsection). \item \label{RefMotivationRestriction} A third reason is that the operation of function restriction is naturally expressed in terms of two elementary set operations: cartesian product and intersection. Indeed, this allows to extend this operation to relations, immediately giving an instance of what we argued in point~\eqref{RefMotivationRelations} above. Restriction is a fundamental operation for representing the concept of weakly dominant strategy, a key concept in auction theory, see e.g., \cite[proposition~2]{mas-04}. The definition of function restriction is arguably more complicated in standard higher order logic{} since functions are always total. That is, restricting them to a set requires carrying the restricted domain set with the function, e.g.\ by forming a pair $(R, f|_R)$, whereas the set-theoretical representation naturally includes this set. \item Finally, specific partial, finite set theoretical functions can be very concisely and quickly defined; this made it possible to prompt\-ly test Isabelle{} code while we were working on it. For example, functions can be written in the form of a set as \verb|{(0,10), (1,11), (2,12)}| and fed to the Isabelle{} definitions very easily in order to test the correctness of related computations empirically, whereas a lambda expression would be more complex to define.\rnote{MK: Added ``to define''} \end{enumerate} \subsection{Two basic operators on relations: “outside” and pasting} \label{RefOutside} In this section, we discuss two general mathematical operations we encountered specifically during formalization of auctions. Assume a set of bidders $N $ and a function $b:N\to\mathbb{R}$ that determines the corresponding bids. The first operation removes one bidder $i$ from the domain $N$ of the bid function. The second operation alters the bid function in one point $i$ of its domain, $b\,+\!*\,(i,b_i)$, which is equal to $b$ except for argument $i$, where the value is changed to $b_i$. In set theory, a function (or a relation) inherently specifies its domain and range. A generalization of the first operation is thus obtained by writing (in Isabelle): \begin{mytable} \begin{alltt} definition Outside :: "('a \( \times \) 'b) set ⇒ 'a set ⇒ ('a × 'b) set" (infix "outside" 75) where "R outside X = R - (X × Range R)", \end{alltt} \end{mytable} where 75 denotes the binding strength of the infix operator. The following specialization to singletons … \begin{mytable} \begin{alltt} abbreviation singleoutside (infix "--" 75) where "b -- i ≡ b outside \{i\}", \end{alltt} \end{mytable} … turned out handy for our purposes. Another circumstance making the set theoretical approach convenient is that now the second operation can be obtained in simple terms of the first as follows: \begin{mytable} \begin{alltt} definition paste (infix "+*" 75) where "P +* Q = (P outside Domain Q) ∪ Q", \end{alltt} \end{mytable} which can be specialized to the important case of \verb|Q| being a singleton function: \begin{mytable} \begin{alltt} abbreviation singlepaste where "singlepaste F f ≡ F +* \{(fst f, snd f)\}" notation singlepaste (infix "+<" 75). \end{alltt} \end{mytable} While we often applied \verb|outside| and \verb|+*| to functions rather than relations (i.e., assuming right uniqueness, see section~\ref{RefRuniq}), many of its properties were proved for relations in general. For example, the associativity theorem \begin{mytable} \begin{alltt} lemma ll53: "(P +* Q) +* R = P +* (Q +* R)" \end{alltt} \end{mytable} has no hypotheses in its statement and holds for general relations. \subsection{Specializing relations to functions: right-uniqueness and evaluation} \label{RefRuniq} The operators \verb|outside| and \verb|+*| are building blocks on which the statements of many theorems we proved for auctions are based. In turn, a number of preparatory lemmas have been proven about those objects in the generic case of relations; however, others only hold when considering relations which actually are functions. Hence, we need the following predicate for \emph{right-uniqueness} of a relation, which we define in terms of the \emph{triviality} of a set (i.e.\ being empty or singleton): \begin{mytable} \begin{alltt} definition trivial where "trivial x = (x ⊆ \{the_elem x\})", \end{alltt} \end{mytable} where \verb|the_elem| extracts an element from a set, being undefined when this cannot be done in a unique way.\rnote{MK: changed `one unique way' to `a unique way'} The predicate for right-uniqueness is called \verb|runiq|, and it uses the operator \verb|R``X|, which yields the image of the set \verb|X| through the relation \verb|R|: \begin{mytable} \begin{alltt} definition runiq :: "('a × 'b) set ⇒ bool" where "runiq R = (∀ X . trivial X → trivial (R `` X))". \end{alltt \end{mytable} We note that, contrary to other proof assistants based on set-theory (e.g., Mizar{}), which in general cannot directly compute values of functions (at most the truth or falsity of predicates involving those values), we are able to preserve from Isabelle{}/HOL the ability of actually computing the evaluation of these set-theoretical flavoured functions, when they are right-unique, through the following operator: \begin{mytable} \begin{alltt} fun eval_rel :: "('a × 'b) set ⇒ 'a ⇒ 'b" (infix ",," 75) where "R ,, a = the_elem (R `` \{a\})" \end{alltt \end{mytable} Now, indeed, set theoretical functions can be evaluated via \verb|,,|.% This can be tested through the Isabelle{} command \verb|value|: for example we can write \begin{mytable} \begin{alltt} value "\{(0::nat,10),(1,11),(1,12::nat)\} ,, 0" \end{alltt} \end{mytable} and obtain \verb!10! as an answer. This holds also when combining \verb|eval_rel| with the operators as from the beginning of this section; for example \begin{mytable} \begin{alltt} value "(\{(0::nat,10),(1,11),(1,12)\} +< (1,13::nat)) ,, 1" \end{alltt} \end{mytable} yields the answer \verb|13|.\rnote{MK@MC: should $\tt +<$ not be $\tt +*$? And on page 9 as well?} A right-unique relation and a standard higher order logic{}, lambda abstracted function represent the same mathematical object, hence it should be possible to pass from one representation to another. \verb|graph| from \verb|Function_Order.thy|, defined as \begin{mytable} \begin{alltt} definition graph where "graph X f = \{(x, f x) | x. x ∈ X\}" \end{alltt} \end{mytable} does exactly that. The opposite conversion can be achieved easily as follows: \begin{mytable} \begin{alltt} definition toFunction (* inverts graph *) where "toFunction R = (λ x . (R ,, x))" \end{alltt} \end{mytable} However, the degree of computability of set-theoretical functions is less than with original HOL functions; for example we cannot evaluate \begin{small} \begin{equation*} \label{RefFunctionNotEvaluable} \verb|value "(graph {x::nat. x<3} (λx. (10::nat))),,(1::nat)"|, \end{equation*} \end{small} \noindent while the following works as expected \begin{mytable} \begin{alltt} value "(graph \{0,1,2\} (λx. (10::nat))),,(1::nat)". \end{alltt} \end{mytable} We also note that, since Isabelle formalizes set theory inside higher order logic{}, types still impose some rigidity, compared to stand-alone set theory: see section~\ref{RefSetTheory}. For example, the following alternative Isabelle{} definition would be exactly equivalent to \verb|eval_rel| in a standard (untyped) set theory: \label{RefEvalRel2} \begin{mytable} \begin{alltt} abbreviation "eval_rel2 (R::('a×('b set)) set) (x::'a) ≡ ⋃ (R``\{x\})" notation eval_rel2 (infix ",,," 75), \end{alltt \end{mytable} It is, however, actually defined (and equivalent to it) \emph{only} for set-yielding relations:\rnote{MK: split long sentence into two.} \begin{mytable} \begin{alltt} lemma lll82: assumes "runiq (f::(('a × ('b set)) set))" "(x::'a) ∈ Domain f" shows "f,,x = f,,,x" \end{alltt} \end{mytable} However, when it is applicable, \verb|eval_rel2| has the desirable qualities of evaluating to the empty set outside the domain of \verb|f|, and in general to something defined when right-uniqueness does not hold (in which case \verb|eval_rel| is undefined). This allowed us to give more concise proofs in such cases. \verb|runiq| is a central definition in our formalization; its many possible equivalent formulations have turned out to be useful in different steps when proving various lemmas. Here we present the possible alternative definitions we have proven to be equivalent in the ATT{}: \begin{mytable}\renewcommand{\baselinestretch}{0.65} \begin{alltt} lemma lll33: "runiq P=inj_on fst P" lemma runiq_alt: "runiq R ↔ (∀ x . trivial (R `` \{x\}))" lemma runiq_basic: "runiq R ⟷ (∀ x y y' . (x, y) ∈ R ∧ (x, y') ∈ R ⟶ y = y')" lemma runiq_wrt_eval_rel: "runiq R ⟷ (∀x . R `` \{x\} ⊆ \{R ,, x\})" lemma runiq_wrt_eval_rel': "runiq R ⟷ (∀x ∈ Domain R . R `` \{x\} = \{R ,, x\})" lemma runiq_wrt_ex1: "runiq R ⟷ (∀ a ∈ Domain R . ∃! b . (a, b) ∈ R)" lemma runiq_wrt_THE: "runiq R ⟷ (∀ a b . (a, b) ∈ R ⟶ b = (THE b . (a, b) ∈ R))" \end{alltt \end{mytable} In general, we found that, especially for basic and ubiquitous concepts such as \verb|runiq|\rnote{MK: put runiq into verbose and `as' to `such as'}, the more equivalent definitions we have\lnote{MK: `one has' to `we have'}, the better. One reason is that this improves the understandability of the formalization: different readers will find different definitions easier to grasp. Another reason is that automated theorem proving tools, such as Sledgehammer% \footnote{Sledgehammer is an Isabelle{} tool that applies automatic theorem provers (ATPs) and satisfiability-modulo-theories (SMT) solvers to automatically produce proofs~\cite{isabelle-sledgehammer}.}% , will be more likely to find automated justification in single steps of subsequent proofs: by picking the appropriate equivalent definition, sledgehammer can find a justification, while, upon removing that definition, it is no longer able to do that. We actually experienced this phenomenon with proofs involving \verb|runiq|: the form of \verb|runiq| given in \verb|lll33| allowed Sledgehammer to find the proof of this technical lemma: \begin{mytable} \begin{alltt} lemma lll34: assumes "runiq P" shows "card (Domain P) = card P". \end{alltt} \end{mytable} \noindent \verb|lemma lll34| above was in turn used to formalize proposition 3 from \cite{mas-04}. \subsection{Application to auctions} \label{RefFuncAsGraphApplication} Next, we give one example of the roles of the operations introduced in this section in our practical setting of auctions, for the simple case of a single-good auction. In this case, the input data for the auction are given through a function \verb|b| associating to each bidder the amount she{} bids for the good. Given a fixed bidder \verb|i|, the outcome of the auction is determined by two functions, \verb|a| and \verb|p|. Both take \verb|b| as an argument: the first yields whether that bidder won the item (\verb|a,,b = 1|) or not (\verb|a,,b = 0|); the second, \verb|p,,b|, yields how much she{} has to pay. We take the Isabelle{} formalization of the second proposition in \cite{mas-04}, which is theorem \verb|th10| in file \verb|Maskin2.thy|. This result proves, given some general requirements, the logical equivalence of two properties, each binding \verb|a| and \verb|p|: \begin{itemize} \item The first property we called \verb|genvick| (for generalized Vickrey auction, see below); it states that the payment imposed to \verb|i| is the sum of a `fee{}' term \verb|t(b--i)|, to be paid irrespectively of the outcome, and of a proper price term \verb|(a,,b - a1)*w(b--i)|, which is to be added only in case she{} obtains the good. Moreover, the first term does not depend on how much \verb|i| herself{} bids, but only on others' bids. The proper price is determined by the auxiliary function \verb|w|, which also does not depend on \verb|i|'s bid (since Vickrey auctions are second price auctions). Hence, \verb|i|'s bid can influence only whether \verb|i| pays or not the proper price, but not its amount. \item The second property, called \verb|dom4|, states that \verb|i| can never be worse off if she{} changes her{} bid to her{} real valuation \verb|v| of the good. \end{itemize} Hence, \verb|genvick| assumes the following form: \begin{mytable} \begin{alltt} abbreviation genvick where "genvick a p i w ≡ (∃ (a1::allocation) t. (∀ b ∈ Domain a ∩ (Domain p). p,,b = (a,,b - a1)*w(b--i) + t(b--i) ))", \end{alltt} \end{mytable} \noindent while \verb|dom4| is the following inequality: \begin{mytable} \begin{alltt} definition dom4 where "dom4 i a p = (∀ b::bid. ∀ v.( \{b,b+<(i,v)\}⊆(Domain a ∩ (Domain p)) ∧ i∈Domain b)⟶ v*(a,,b)-(p,,b) ≤ v*(a,,(b<(i,v)))-(p,,(b+<(i,v))))". \end{alltt} \end{mytable} The operators \verb|--| and \verb|+<| are central in expressing those two conditions; their respective general properties (collected in \verb|RelationProperties.thy|) permitted to streamline the proof of the theorem, whose thesis reads: \begin{equation} \label{RefMaskin2Thesis} \verb|genvick a p i w = dom4 i a p|. \end{equation} \rnote{MK: Can we formulate the theorem also in plain English?} \section{Quotients between relations} \label{RefQuotients} We built a library of basic facts centred around our new constructs of right-uniqueness (\verb|runiq|), evaluation (\verb|,,|), pasting (\verb|+*|), and considering a function \verb|outside| some subset of its domain. The library also contains more advanced results. In particular, we describe here our approach to building quotients; then, we derive functions on equivalence classes of points from functions defined on single points. These methods are common in many areas of mathematics, especially algebra and topology; they are used when a given property holds on classes of objects, and one wants to abstract away from the specific representative of a class, and rather define the given property on the whole class. Such classes are typically the equivalence classes induced by an equivalence relation over its domain, which form the \emph{quotient} of the original set. For example, in group theory, the operation of a group $G$ is canonically transported to the set of the cosets yielded by a normal subgroup $N$. This is what makes the quotient group $G/N$ a group, and is only possible if the group operation is \emph{class-invariant}% \footnote{In this case we can also say that the group operation is \emph{well-defined}, or that it \emph{respects} the corresponding equivalence relation, or even that it is \emph{compatible} with it.}% : the product of two representative elements of two cosets must be in the same coset, irrespective of how those representative elements are selected. This is ensured by the definition of normal subgroup. In our case, to construct \verb|t| appearing in the definition of \verb|genvick| towards the end of the previous section, we need to define some function taking as an argument a bid vector with the $i$-th component removed (where $i$ is a bidder). The formal way to do that was to aggregate all possible bid vectors differing only on their $i$-th component and to define that function on the classes obtained this way; hence, using quotients naturally emerged as one elegant approach. More details on this particular application of quotients are in section \ref{RefApplication}. As illustrated in \cite{paulson2006defining}, existing Isabelle{}'s theory \verb|Equiv_Relations.thy| already introduces tools for these general mathematical techniques. There are, however, the following problems: \begin{enumerate} \item The operation of passing from a pointwise function to the `abstracted{}' version defined on equivalence classes is done using type-theoretical \verb|Abs_| and \verb|Rep_|. In paper-based mathematics, on the other hand, everything is done using set theory; hence, a mathematician would probably not know how to use this implementation without first getting some knowledge of the underlying type-theoretical foundations. \item This operation must be performed `manually{}' in each separate case: there is no generic definition to do that given a pair $\left( f, R \right)$, where $f$ is the function to be abstracted and $R$ is an equivalence relation on its domain. In contrast, in our treatment we introduce a function called \verb|quotient| doing exactly this. \item $f$ and $R$ are typed as a lambda function and a set-theoretical relation, respectively,\rnote{MK: put respectively at the end of the construct} while there is no reason to preemptively preventing $f$ from being a generic relation (i.e., not necessarily being right unique). \end{enumerate} For these reasons, we coded a purely set-theoretical Isabelle{} implementation of this machinery, via three simple definitions and one theorem establishing the right-uniqueness of the abstracted function given basic requirements on the pointwise function. The first definition gives a map to pass from a point of the domain of a relation \verb|R| to the corresponding equivalence class: \begin{mytable} \begin{alltt} definition projector where "projector R = \{(x,R``\{x\}) | x. x ∈ Domain R \}" \end{alltt \end{mytable} The second definition builds, given a pointwise relation \verb|R| and two equivalence relations \verb|P|, \verb|Q| (working on its domain and codomain, respectively) the corresponding, abstracted relation on the resulting equivalence classes: \begin{mytable} \begin{alltt} definition quotient where "quotient R P Q = \{(p,q)| p q. q ∈ (Range (projector Q)) ∧ p ∈ Range (projector P) ∧ p × q ∩ R ≠ \{\} \}". \end{alltt} \end{mytable} While this definition is typically given for a function \verb|R| and equivalence relations \verb|P| and \verb|Q|, it still makes sense if these additional conditions are not satisfied. This allows us to lift these requirements in some preparatory lemmas before assuming them in the following main result: \label{RefWellDefinednessTheorem} \begin{mytable} \begin{alltt} lemma l23: assumes "compatible f P Q" "runiq f" "trans P" "sym P" "equiv (Domain Q) Q" shows "runiq (quotient f P Q)", \end{alltt} \end{mytable} where the predicates \verb|equiv X P|, \verb|trans P|, \verb|sym P| exist in the Isabelle{} library, and state, respectively, that \verb|P| is an equivalence relation over the set \verb|X|; that \verb|P| is a transitive relation; and that \verb|P| is a symmetric relation. Note that in Isabelle{} there is a definition for a quotient of a relation \verb|R| written as \verb|quotient R| as the set of all equivalence classes associated with \verb|R|. Here, however, we assume a function (or relation) \verb|f| with respect to relations \verb|P| and \verb|Q| and define the quotient of \verb|f| with respect to \verb|R| as a function with the domain \verb|quotient R|. The notion of compatibility above asks that \verb|f| respects \verb|P| and \verb|Q|:\smallskip \noindent\begin{minipage}{0.6\textwidth} \begin{mytable} \begin{alltt} definition compatible where "compatible R P Q = (∀ x . (R``(P``\{x\}) ⊆ Q``(R``\{x\})))" \end{alltt \end{mytable} Note that the definition of \verb|compatible| is a generalization of the usual commutativity of the application of functions as displayed to the right. \end{minipage}\qquad \begin{minipage}{0.35\textwidth} \includegraphics[width=1\textwidth]{compatible.pdf} \end{minipage}\bigskip\rnote{MK:Changed wrapfigure to minipage} It should be noted that this condition is not required at the time of defining a quotient, which can be defined for any triple of relations; rather, it is only required when asking that the quotient behaves in the expected way, as stated by \verb|lemma l23| above. This allows us to freely use the construction \verb|quotient| in advance, to show a number of intermediate results not requiring compatibility themselves. For example, the following lemma holds for general relations: \begin{small} \begin{alltt} lemma quotientFactors: assumes "equiv (Domain p) p" "equiv (Domain q) q" shows "quotient r p q = (projector p)¯ O r O (projector q)". \end{alltt} \end{small} where \verb|R¯| is the converse of the relation \verb|R| and \verb|O| stands for relation composition. We also note that, for similar reasons, we devised a definition of compatibility taking any triple of relations as arguments: as in the definition of quotient, one can use \verb|compatible R P Q| even before showing that \verb|R| is right-unique and that \verb|P|, \verb|Q| are equivalence relations. In the latter case, however, the definition of compatibility reduces to asking that any two \verb|Q|-equivalent points have \verb|P|-equivalent images through \verb|R|. \section{Injective functions and partitions} \label{RefDualDefs} In section \ref{RefFuncAsGraphApplication}, we introduced the mathematical description of a single-good auction. An important family of more complex schemes is given by \emph{combinatorial} auctions. In these, there are several objects at stake (a set of goods $G$), and each bidder (of a set $N$) can bid for each possible combination of them. The outcome of the auction is still described by a pair of maps $\left( a, p \right)$, yielding respectively what a bidder gets and how much she has to pay. However, now the ``what a bidder gets{}'' part must be represented by a mathematical object more articulate than a $\left\{ 0,1 \right\}$-valued function. It is represented by a partition of $G$, and by an injective function (or injection) from that partition to $N$. Treating the latter function inside set-theory gives an advantage: the pair (partition, injection) is conveniently represented by the injection alone, because the partition of $G$ will be simply the domain of the former. This allowed us to type the relevant objects plainly as follows: \begin{mytable} \begin{alltt} type_synonym bidder = "nat" type_synonym goods = "nat set" type_synonym allocation_rel = "(goods × bidder) set". \end{alltt} \end{mytable} Since we wanted to extract code from our Isabelle{} formalization, we had to implement recursive definitions for both the set of all possible partitions of a finite set and for the set of all possible injections from a finite set to another finite set. In both cases, those recursive definitions turned out to be inconvenient when it came to prove mathematical facts involving them. Hence a separate, more natural definition was needed that is equivalent but not computable. We illustrate this dual approach in the case of injections. \begin{mytable} \begin{alltt} fun injections_alg :: "'a list ⇒ 'b::linorder set ⇒ ('a × 'b) set list" where "injections_alg [] Y = [\{\}]" | "injections_alg (x # xs) Y = concat [[R+*\{(x,y)\}. y←sorted_list_of_set (Y-Range R)]. R ← injections_alg xs Y]" \end{alltt} \mycaption{Constructive definition of injection}\label{table:injective_alg} \end{mytable} In Table~\ref{table:injective_alg}, we find a definition of all injective functions between two finite sets \verb|X| and \verb|Y|, which recurs on \verb|X|, while in Table~\ref{table:injections} the set of all injective functions is defined axiomatically. \begin{mytable} \begin{alltt} definition injections :: "'a set ⇒ 'b set ⇒ ('a × 'b) set set" where "injections X Y = \{R. Domain R=X ∧ Range R⊆Y ∧ runiq R ∧ runiq(R¯)\}" \end{alltt} \mycaption{Axiomatic definition of injection}\label{table:injections} \end{mytable} We use the constructive definition in computations and the axiomatic for proofs and mathematical manipulations. In order to do that we have to prove their equivalence in the theorem stated in Table~\ref{table:theoremInjections_equiv}. \begin{mytable} \begin{alltt} theorem injections_equiv: fixes xs::"'a list" and Y::"'b::linorder set" assumes non_empty: "card Y > 0" shows "distinct xs ⟹ (set(injections_alg xs Y)::('a×'b)set set)=injections (set xs) Y" \end{alltt} \mycaption{\xmakefirstuc{\junction}{} of constructive and axiomatic definitions of injection}\label{table:theoremInjections_equiv} \end{mytable} Similarly, we have a constructive and an axiomatic definition for partitions (see Tables~\ref{table:partition_alg} and \ref{table:partition_ax}, respectively). \begin{mytable} \begin{alltt} definition insert_into_member_list :: "'a ⇒ 'a set list ⇒ 'a set ⇒ 'a set list" where "insert_into_member_list new_el Sets S = (S ∪ \{ new_el \}) # (remove1 S Sets)" definition coarser_partitions_with_list ::"'a ⇒ 'a set list ⇒ 'a set list list" where "coarser_partitions_with_list new_el P = (\{ new_el \} # P) # (map ((insert_into_member_list new_el P)) P)" definition all_coarser_partitions_with_list ::"'a ⇒ 'a set list list ⇒ 'a set list list" where "all_coarser_partitions_with_list elem Ps = concat (map (coarser_partitions_with_list elem) Ps)" fun all_partitions_list :: "'a list ⇒ 'a set list list" where "all_partitions_list [] = [[]]" | "all_partitions_list (e # X) = all_coarser_partitions_with_list e (all_partitions_list X)" \end{alltt} \mycaption{Constructive definition of partition}\label{table:partition_alg} \end{mytable} The constructive definition above represents a partition of a finite set as a list of (disjoint) subsets of it; it works by induction on the cardinality of the set to be partitioned, as follows. Given a set of cardinality $n+1$, we write it as a disjoint union $X \cup \left\{ x \right\}$, and, given a partition $P$ of $X$, we insert $x$ into each set belonging to $P$ (this is done by the operator \verb|insert_into_member_list| above). We thus obtain $|P|$ many partitions of $X \cup \left\{ x \right\}$, to which we add the distinct partition $\left\{ \left\{ x \right\} \right\} \cup P$. Making $P$ range over all possible partitions of $X$, we obtain in this way all possible partitions of $X \cup \left\{ x \right\}$. When implementing this idea into the code above, however, we chose to represent the finite set $X$ as a list (rather than a set) of elements of it, and a partition of it as a list (rather than a set) of subsets of it. This is because the algorithm just described is iterated in two ways: first, $x$ is inserted into each set of a partition; secondly, the construction is iterated for each possible partition of $X$. Since iterations are easier with lists than with finite sets \cite{nipkow2005proof}, we adopted that choice. \begin{mytable} \begin{alltt} definition is_partition where "is_partition P = (∀ X∈P . ∀ Y∈ P . (X ∩ Y ≠ \{\} ↔ X = Y))" definition is_partition_of (infix "partitions" 75) where "is_partition_of P A = (\(\bigcup\) P = A ∧ is_partition P)" definition all_partitions where "all_partitions A = \{P . P partitions A\}." \end{alltt} \mycaption{Axiomatic definition of partition}\label{table:partition_ax} \end{mytable} It should be noted that passing from a list to a finite set (via the operator \verb|set|) is easier than the converse, hence the equivalence theorem for partitions in Table~\ref{table:theoremPartition_equiv} is stated taking as input data a list \verb|xs|: \begin{mytable} \begin{alltt} theorem all_partitions_paper_equiv_alg: fixes xs::"'a list" shows "distinct xs ⟹ set(map set (all_partitions_list xs)) = all_partitions(set xs)" \end{alltt} \mycaption{\xmakefirstuc{\junction}{} of constructive and axiomatic definitions of partition}\label{table:theoremPartition_equiv} \end{mytable} Thanks to the constructive definitions above, we were able to extract, from Isabelle{} code \cite{haftmann2010code}, executable code running Vickrey combinatorial auctions \cite{CKLR:SoundCombVickCode13}. Furthermore, we were able to prove fundamental theoretical properties about those auctions by using the non-constructive versions of those definitions; for example, that the price for any bidder is non-negative. The compatibility{} theorems in Tables~\ref{table:theoremInjections_equiv} and \ref{table:theoremPartition_equiv} allow to certify that such theoretical properties hold also for the extracted code. \lnote{MC: Added for clarification following reviewer2.} We note that, while non-constructive versions hold in the general case, constructive versions are obviously limited to the finite case (e.g., calculating all the partitions of a finite set). Therefore, the two compatibility{} theorems must restrict to the finite case: this is reflected by the fact that the argument of \verb|injections| and \verb|all_partitions|, appearing in Tables~\ref{table:theoremInjections_equiv} and \ref{table:theoremPartition_equiv}, respectively, is \verb|set xs|.\rnote{MK: respectively after} This means that such an argument is automatically a finite set, being the result of converting a list (\verb|xs|) to a set using the Isabelle{} function \verb|set|. We stress the fact that our approach allows to prove most theorems (e.g., thesis~\eqref{RefMaskin2Thesis} at the end of section~\ref{RefFuncAsGraph}) without restricting to the finite case, and to add the additional hypothesis of finiteness only and exactly for the theorems needing it. \section{Application of quotients to auctions} \label{RefApplication} We now give enough formalization details to illustrate the exact point in our proofs of auction theory in which we needed to employ \verb|projector| and \verb|quotient|, introduced in section \ref{RefQuotients}. To prove the thesis \eqref{RefMaskin2Thesis} at the end of section~\ref{RefFuncAsGraph}, we needed to explicitly build the \verb|t| appearing in the definition of \verb|genvick|. This function is uniquely determined by \verb|a| and \verb|p|; however, the latter takes \verb|b| as an argument, while \verb|t| takes \verb|b--i| (which represents a bid vector with bidder \verb|i|'s bid removed, and is called a \emph{reduced bid}). The algebraic way to pass from \verb|b| to \verb|b--i| is to consider an equivalence relation associating any two \verb|b| \verb|b'| differing at most in the point \verb|i|. Correspondingly, to obtain from \verb|p| a function on arguments of the form \verb|b--i|, we form the quotient of \verb|p| according to that equivalence relation. The equivalence relation we need is exactly the kernel% \footnote{We recall that the kernel of a function $f$ is the equivalence relation $\circ_f$ given by $x_1 \circ_f x_2 \iff{} f \left( x_1 \right) = f \left( x_2 \right)$. See \cite[Definition~1.18]{bergman2011universal}. The kernel notion was missing in the Isabelle{} library, and we also provided it in ours.} of the following function: \begin{mytable} \begin{alltt} definition reducedbid:: "bidder ⇒ (bid × allocation) set ⇒ (bid × bidder set × bid × allocation) set" where "reducedbid i a = \{(b, (Domain b, b outside \{i\}, a ,, b))| b. b ∈ Domain a\}". \end{alltt} \end{mytable} So that the explicit construction of \verb|t|, used inside the proof for \eqref{RefMaskin2Thesis}, ends up as \begin{mytable} \begin{alltt} "λx. reducedprice p i a ,, (\{i\} ∪ Domain x, x, Min (Range a))", \end{alltt}% \end{mytable} where \begin{mytable} \begin{alltt} definition reducedprice:: "(bid × price) set ⇒ bidder ⇒ (bid × allocation ) set ⇒ ((bidder set × bid × allocation) × price) set" where "reducedprice p i a = (projector ((reducedbid i a)¯)) O (quotient p (Kernel (reducedbid i a)) Id) O ((projector Id)¯)", \end{alltt} \end{mytable} \verb|Id| being the identity function. With this definition in place, an important part of the theorem consists in showing that \verb|reducedprice| is right-unique. This reduces to showing that \verb|quotient p | \verb|(Kernel(reducedbid i a))| \verb| Id| appearing in the definition above is right-unique. Thanks to \verb|lm23| (see section \ref{RefWellDefinednessTheorem}), this in turn means proving the compatibility between \verb|p| and the equivalence relation just introduced, i.e., \verb|Kernel (reducedbid i a)|, which is provided by \begin{mytable} \begin{alltt} lemma l24b: assumes "functional (Domain a)" "Domain a ⊆ Domain p" "dom4 i a p" "runiq p" shows "compatible p (Kernel (reducedbid i a)) Id". \end{alltt} \end{mytable} \section{Discussion and related work} \label{RefSetTheory} \label{RefRelatedWork} The work presented here is based on the specific set theory implemented in Isabelle{}/HOL, sometimes called simply-typed set theory \cite[Section~6.1]{paulson2013proof}, \cite[Section~1]{Paulson:stfv93}% . It differs from standard set theories, as Zermelo-Fraenkel (ZF), in that the primitives of the latter are encoded using primitives of higher order logic{}, as follows: a given set $p$ is actually a term $p$ of type $\tau \Rightarrow \text{bool}$, and the writing $x \in p$ is actually the application $p\ x$. This means that in Isabelle/HOL{} one cannot write a set like $ \left\{ x, \left\{ x \right\} \right\}$ (which is not well-typed): the standard hierarchy of ZF is no longer constructable. As long as the objective is to formalize ordinary (as defined in \cite[Section~I.1]{simpson2009subsystems}) mathematics, this usually causes no problem: while this difference prevents the encoding of relevant mathematical objects (e.g., natural numbers, integers, reals, cartesian products) as usually done in ZF, those objects can be directly represented in higher order logic{}. This means that, for many mathematical branches, what is lost with respect to ZF is limited to its more technical side-effects, as $\pi \cap \mathbb{Q} \neq \emptyset$, or $1 \cap 2 \subseteq 3$. Since the latter are often regarded as strange or meaningless writings \cite{leinster2012rethinking}, \cite[Section~1]{Paulson:stfv93}, this could even be a desirable consequence. On the other hand, when trying to exploit technical ZF `hacks{}', as we tried to do with \verb|eval_rel2|, problems can arise: we discussed that in section \ref{RefEvalRel2}. Moreover, ZF is a more appropriate tool for studying the remaining branches of mathematics, starting with set theory itself. The Z pattern catalogue \cite[Section~IV]{valentine2004az} allows to represent functions as sets of pairs, as done here. However, Z is a specification language, while we are interested in both specification and implementation. Mizar{} is based on untyped set theory, thus modelling functions and relations as done here, and also provides many relevant existing theorems in its library; however, it is not possible to extract code or to do computations in general. The Ssreflect extension library for the Coq proof assistant is extensive and provides a lot of operations; on the other hand, most of this material seems to apply to functions as represented by lists, thereby inherently limiting to the finite case. More details on how Isabelle{}/HOL generally compares with other systems are in the comparative study \cite{la-ca-ke-mo-ro-we-ww-13}. \section{Conclusions}\label{RefConclusions} We have built an extensive library of results about functions, represented as right-unique relations as an alternative to the lambda abstractions that have so far been typical of Isabelle{}/HOL. In this paper we explained how we employed this alternative technique, and the concepts in our library, in the application domain of auction theory. Our library ranges from simpler constructions such as pasting a relation onto another, to more sophisticated ones such as quotients.\rnote{MK: I think, we should mention again the one advantage mentioned previously, namely, that set theory allows for the generalization of theorems.} We described how set-theoretical constructs can concisely express constructions that frequently occur in formalization, giving concrete examples from the application domain of auction theory. In the particular case of representing functions, we discussed the advantages and disadvantages of the set-theoretical representations with respect to the alternative, more natural (given the foundations of Isabelle/HOL) lambda representation. We also showed that computability, a typical feature of lambda functions, can be achieved for set-theoretical functions in Isabelle/HOL. Moreover, even in those cases in which it cannot be achieved, we showed how we were still able to employ a dual approach, giving non-constructive, more expressive definitions along with constructive (and thus computable) ones, finally proving their equivalence through compatibility{} theorems. This allows us to obtain the best of both worlds, that is, expressiveness and proof-friendly definitions together with computability, at the cost of having to prove the additional compatibility{} theorems. Finally, we took our application of this approach to a more advanced, abstract level by describing in a simple, close-to-paper set-theoretical style notions such as \emph{quotients}, \emph{compatibility} and \emph{kernel}. We presented a concrete application of this material in a hands-on case encountered in formalizing auction theory within the ForMaRE{} project. We think that our library is rich and generic enough to be of possible use to other Isabelle{} users, and we hope that this paper can serve as a first guiding example to show how it can be employed. \begin{spacing}{0.95} \printbibliography \end{spacing} \end{document} Things to add in response to reviewers: Reviewer1: -* make it a bit clearer to people outside the HOL world what sets in Isabelle/HOL are -* It would be helpful to have sentence or two in the introduction to explain that auction theory is a branch of economics and to outline the kind of problems it deals with. -* What about related work? Z comes to mind as a typed notation that uses sets of pairs to represent functions. The Z pattern catalogue of Valentine, Stepney and Toyn may be of interest. -* >[...]"ell,ell63" or what?And why?That's just a label.We'll give it a more meaningful name. Reviewer2 (the only one not writing any follow-up to our response): -* Consequently,most theorems hold irrespective of those restrictions,making them apply to Isabelle generated code as a particular case:imposing those only when needed can be regarded as an upside of our approach.However,this remark prompted us to better explain where we impose them. -* >[...]SSREFLECT library seems to me to be better. We'll mention this related work.Still,we stress that we don't restrict to finite sets(see above). -* In general, we found that, especially for basic and ubiquitous concepts as runiq, the more equivalent definitions one has, the better. One reason is that automated theorem proving tools, as sledgehammer, will be more likely to find automated justification in single steps of subsequent proofs: by picking the appropriate equivalent definition, sledgehammer can find a justification, while, upon removing that definition, it is no longer able to do that. We actually experienced that with proofs involving runiq. From the human point of view, we find that the preference of one definition with respect to another one can be quite arbitrary: indeed, we are not sure of which one the reviewer would find most natural, and we would be interested in knowing it and possibly add the corresponding equivalence theorem to our Isabelle library. Indeed, by providing as many equivalent definitions as possible, we hope to maximize the understandability of the formalization, too. - >this part of the paper sounds a bit like black magic[...] While we find important to keep the paper readable for a reader not necessarily interested in the more technical implementation details, we will try to give references and some context for those. -* Their equivalence is thus actually the proof of correctness of the implementation. Is that useful to present both terminologies? Do you have something to discuss here about this? Reviewer3: - Perhaps it could be made a bit > clearer how much of this formalization has already been published > elsewhere. -* The response honestly says that the authors are rather pragmatically using the Isabelle techniques that work best for them, instead of explaining the underlying Isabelle mechanisms. This is a bit disappointing for the reader of the paper in its current form. For example this sentence in the conclusion: "We also showed how computability, a typical feature of lambda functions, can be achieved for set-theoretical functions in Isabelle/HOL", suggests that some explanation of _how_ things work is given in the paper, but it is not the case. I really recommend that the paper gets clarified in this point, so that it does not give the reader any false hope of some deep foundational (or even implementational) discussions. *- set theory (e.g., Mizar takes this approach) and higher order logic (e.g., as in Isabelle/HOL)." ==> Type theory is another strong competitor (Coq, Agda, Nuprl)

https://arxiv.org/abs/1811.06389

Semi-perfect 1-Factorizations of the Hypercube

A 1-factorization $\mathcal{M} = \{M_1,M_2,\ldots,M_n\}$ of a graph $G$ is called perfect if the union of any pair of 1-factors $M_i, M_j$ with $i \ne j$ is a Hamilton cycle. It is called $k$-semi-perfect if the union of any pair of 1-factors $M_i, M_j$ with $1 \le i \le k$ and $k+1 \le j \le n$ is a Hamilton cycle.We consider 1-factorizations of the discrete cube $Q_d$. There is no perfect 1-factorization of $Q_d$, but it was previously shown that there is a 1-semi-perfect 1-factorization of $Q_d$ for all $d$. Our main result is to prove that there is a $k$-semi-perfect 1-factorization of $Q_d$ for all $k$ and all $d$, except for one possible exception when $k=3$ and $d=6$. This is, in some sense, best possible.We conclude with some questions concerning other generalisations of perfect 1-factorizations.

\section{Introduction} A 1-factorization of a graph $H$ is a partition of the edges of $H$ into disjoint perfect matchings $\{M_1,M_2,\ldots,M_n\}$, also known as 1-factors. Let $\mathcal{M} = \{M_1,M_2,\ldots,M_n\}$ be such a 1-factorization. We say that $\mathcal{M}$ is a perfect factorization if every pair $M_i \cup M_j$ with $i,j$ distinct forms a Hamilton cycle. A 1-factorization $\mathcal{M}$ is called \emph{semi-perfect} if $M_1 \cup M_i$ forms a Hamilton cycle for all $i \ne 1$. Kotzig \cite{Kotzig} conjectured that the complete graph $K_{2n}$ has a perfect 1-factorization for all $n \ge 2$. This has long been outstanding and has so far only been shown to hold for $n$ prime and $2n - 1$ prime (independently by Anderson and Nakamura \cite{Anderson, Nakamura}), as well as certain other small values of $n$ (see \cite{Wallis} for references). The existence or non-existence of perfect or semi-perfect 1-factorizations has been studied for various other families of graphs, in particular for the hypercube $Q_d$, for $d \ge 2$. The hypercube graph $Q_d$ has vertices the subsets of $\{1,2,\ldots,d\}$ and two vertices joined by an edge if they differ in a single element. We say a vertex of $Q_d$ is even if the set contains an even number of elements, and odd if not. Note that every edge of $Q_d$ goes from an odd vertex to an even vertex and so $Q_d$ is bipartite with one vertex class of odd vertices and one vertex class of even vertices, each of size $2^{d-1}$. We say an edge is in direction $i$ if its two endpoints differ in element $i$. This allows us to define some natural 1-factors of $Q_d$, called the \emph{directional matchings}: for each direction $i = 1,\ldots,d$ let $D_i$ be all edges in direction $i$. The collection of all directional matchings is a 1-factorization of $Q_d$, and note that the union of any pair $D_i \cup D_j$, with $i,j$ distinct, is a disjoint union of 4-cycles. Thus any perfect or semi-perfect 1-factorization of $Q_d$ must be in some sense far from this. Craft \cite{Craft} conjectured that for every integer $d \ge 2$ there is a semi-perfect 1-factorization of $Q_d$. This was proved independently by Gochev and Gotchev \cite{Gochev} and by Kr\'{a}lovi\v{c} and Kr\'{a}lovi\v{c} \cite{Kralovic} in the case where $d$ is odd, and settled for $d$ even by Chitra and Muthusamy \cite{Chitra}. Gochev and Gotchev in fact went further and defined $\mathcal{M}$ to be $k$-semi-perfect if $M_i \cup M_j$ forms a Hamilton cycle for every $1\le i \le k$ and $k+1 \le j \le d$. They proved that there is a $k$-semi-perfect factorization of $Q_d$ whenever $k$ and $d$ are both even with $k<d$. This leads us to wonder how close to a perfect factorization we can get. Is there a k-semi-perfect factorization of $Q_d$ for all $k<d$? Is there a perfect factorization of $Q_d$? If not, what is the maximal number of pairs of 1-factors whose union is a Hamilton cycle? Let us introduce some definitions. For a 1-factorization $\mathcal{M}= \{M_1,M_2,\ldots,M_d\}$ of $Q_d$, we define a graph $G[\mathcal{M}]$ with vertices labelled $M_1, \ldots, M_d$ and an edge between $M_i$ and $M_j$ if $M_i \cup M_j$ is a Hamilton cycle on $H$. Note that the definitions above can be easily restated using $G[\mathcal{M}]$: $\mathcal{M}$ is perfect if $G[\mathcal{M}]$ is complete, $\mathcal{M}$ is semi-perfect if $G[\mathcal{M}]$ contains $K_{1,d-1}$ as a subgraph, and $\mathcal{M}$ is $k$ semi-perfect if $G[\mathcal{M}]$ contains $K_{k,d-k}$ as a subgraph. With this new notation, we can rephrase our questions and ask what is the maximal number of edges that $G[\mathcal{M}]$ can contain if $\mathcal{M}$ is a 1-factorization of $Q_d$? Which graphs can $G[\mathcal{M}]$ be isomorphic to? It is in fact not possible for $G[\mathcal{M}]$ to be complete when $d > 2$ (i.e. $\mathcal{M}$ cannot be perfect). More than this, we can show that $G[\mathcal{M}]$ must be bipartite. \begin{theorem}[\cite{Laufer}] \label{thm:Bipartite} Let $H$ be a bipartite graph on two vertex classes each of size $n$, where $n$ is even. Let $\mathcal{M}$ be a partition of $H$ into perfect matchings. Then $G[\mathcal{M}]$ must be bipartite. \end{theorem} A version of Theorem \ref{thm:Bipartite} with the weaker conclusion that $G[\mathcal{M}]$ is not complete has been, according to Bryant, Maenhaut and Wanless \cite{Bryant} proved many times, including by Laufer in 1980 \cite{Laufer}. We re-prove it here for a few reasons, the main one being that we extend the argument slightly to show that $G[\mathcal{M}]$ is bipartite. The proof also introduces ideas that we will be using later (in Theorem \ref{thm:why3hard}). In addition, it is hard to find the theorem and its proof in the literature -- in particular, when making the conjecture that there is a semi-perfect 1-factorization of $Q_d$, Craft also asked whether a \emph{perfect} 1-factorization of $Q_d$ could be found. Theorem \ref{thm:Bipartite} is not mentioned in any of the papers that proved Craft's semi-perfect conjecture. \begin{proof} Let $X$ and $Y$ be the vertex classes of $H$. A perfect matching $M$ naturally induces a function ${M: X \rightarrow Y}$, where $(x,M(x))$ is an edge of $M$. For two perfect matchings $M_i$ and $M_j$, let $\pi_{j,i}$ be the permutation $M_j^{-1}M_i$ on $X$. Note that $\pi_{i,i} = id$, $\pi_{i,j} = \pi_{j,i}^{-1}$ and $\pi_{k,j}\pi_{j,i} = \pi_{k,i}$. Note further that if $M_iM_j$ is an edge of $G[\mathcal{M}]$ then $M_i \cup M_j$ is a Hamilton cycle and so $\pi_{j,i}$ is a cycle of length $n$ on $X$. Suppose for a contradiction that $G[\mathcal{M}]$ contains a odd cycle and let $M_{i_1}$, $M_{i_2}$, $\ldots$, $M_{i_k}$, $M_{i_1}$ be such a cycle. The permutations $\pi_{i_2,i_1}, \pi_{i_3,i_2}, \ldots, \pi_{i_k,i_{k-1}}, \pi_{i_1,i_k}$ are all cycles of length $n$. Since $n$ is even, all of these are odd permutations. Now, \begin{align*} 1 = \sign(\pi_{i_1,i_1}) &= \sign(\pi_{i_1,i_k}\pi_{i_k,i_{k-1}} \pi_{i_{k-1},i_{k-2}} \ldots \pi_{i_3,i_2} \pi_{i_2,i_1}) \\ &= \sign(\pi_{i_1,i_k})\sign(\pi_{i_k,i_{k-1}})\ldots \sign(\pi_{i_3,i_2})\sign(\pi_{i_2,i_1})\\ &= (-1)^k = -1 \end{align*} We have a contradiction, hence $G[\mathcal{M}]$ contains no odd cycles. \end{proof} In the light of Theorem \ref{thm:Bipartite}, the only remaining question is whether for any $k,d$ there is a 1-factorization $\mathcal{M}$ of $Q_d$ such that $G[\mathcal{M}]$ is isomorphic to the complete bipartite graph $K_{k,d-k}$. (Equivalently, whether there is a $k$-semi-perfect 1-factorization of $Q_d$ for every $k$ and $d$, in the language of Gochev and Gotchev.) Section \ref{sec:main} of this paper fully resolves this problem, except for whether $G[\mathcal{M}]$ can be isomorphic to $K_{3,3}$. We also explain, in section \ref{sec:direction}, why the $K_{3,3}$ case cannot be resolved with our methods. In particular, the 1-factorizations we construct in the proof of the main theorem have a direction respecting property. We show that any 1-factorization $\mathcal{M}$ of $Q_{6}$ satisfying this direction respecting property cannot have $G[\mathcal{M}]$ is isomorphic to $K_{3,3}$. We finish with some open questions. \section{Main Theorem} \label{sec:main} \begin{theorem} For $k,l \in \mathbb{N}$ not both equal to 3, there is a 1-factorization $\mathcal{M}$ of the hypercube $Q_{k+l}$ such that $G[\mathcal{M}]$ is isomorphic to the complete bipartite graph $K_{k,l}$. \label{thm:main} \end{theorem} To prove the theorem, we will use the following result due to Stong, which concerns the symmetric directed hypercube $\overleftrightarrow{Q_d}$, obtained from $Q_d$ by replacing each edge with two directed edges, one in each direction. \begin{theorem}[\cite{Stong}] \label{thm:Stong} For $d \ne 3$, the symmetric directed hypercube $\overleftrightarrow{Q_d}$ can be partitioned into $d$ directed Hamilton cycles. \end{theorem} Stong's result applies to directed cubes, but the following corollary allows us to use it for undirected cubes. \begin{corollary}\label{cor:2Stong} For $d \ne 3$, the cube $Q_d$ can be partitioned into 1-factors $A_1, A_2, \ldots, A_d$ and also partitioned into 1-factors $B_1, B_2, \ldots, B_d$ such that $A_i \cup B_i$ is a Hamilton cycle for all $i = 1,2,\ldots,d$. \end{corollary} \begin{proof} Using Theorem \ref{thm:Stong}, partition $\overleftrightarrow{Q_d}$ into directed Hamilton cycles $H_1,H_2,\ldots,H_d$. Let $E$ be the even vertices of $\overleftrightarrow{Q_d}$ and $O$ the odd vertices, so that $\overleftrightarrow{Q_d}$ is bipartite with respect to the vertex classes $E$ and $O$. For each $H_i$, we define $A_i$ to be the edges of $H_i$ that go from $E$ to $O$, and $B_i$ to be the edges that go from $O$ to $E$. Since $H_1,H_2,\ldots,H_d$ partition $\overleftrightarrow{Q_d}$ , every edge from $E$ to $O$ is in a unique $A_i$ and every edge from $O$ to $E$ is in a unique $B_j$. If we now ignore the directions on the edges, every edge of $Q_d$ is in a unique $A_i$ and a unique $B_j$. It is clear that $A_i$ and $B_i$ are perfect matchings and $A_i \cup B_i$ is a Hamilton cycle by construction. \end{proof} Note that we have slightly abused notation in the case $d=1$, since $A_1 = B_1 = Q_1$ and so $A_1 \cup B_1$ is a single edge rather than a cycle. This will not matter in the cases $k \ne 3, l=1$, and we will consider the case $k=3,l=1$ separately. Corollary \ref{cor:2Stong} together with a theorem of Gochev and Gotchev \cite[Theorem~3.1]{Gochev} is enough to show that it is possible to have $G[\mathcal{M}]$ isomorphic to $K_{k,n-k}$ for all $k\ne 3$ and all even $n-k$. We will improve on their arguments to deal with all but one of the remaining cases. We will split the theorem for three different cases and prove each separately. Before we do so, let us outline the ideas involved. We can view the hypercube $Q_{k+l}$ as a $k$-dimensional hypercube whose `vertices' are copies of $Q_l$ (i.e. as the Cartesian product of $Q_k$ and $Q_l$). Let us formalise this idea: Label the vertices of $Q_k$ as subsets of $\{1,2,\ldots,k\}$ in the usual way. For each vertex $u$ of $Q_k$, we define a different copy of $Q_l$ within $Q_{l+k}$: let $Q_l^u$ be the induced subgraph of $Q_{k+l}$ on all vertices $w$ where $w \cap \{1,2,\ldots,k\} = u$. Conversely, we can view $Q_{k+l}$ as a $l$-dimensional hypercube whose `vertices' are copies of $Q_k$. This time, label the vertices of $Q_l$ as subsets of $\{k+1,k+2,\ldots,k+l\}$ in the natural way. For each vertex $v$ of $Q_l$, we define a different copy of $Q_k$ within $Q_{l+k}$: let $Q_k^v$ be the induced subgraph of $Q_{k+l}$ on all vertices $x$ with $x \cap \{k+1,k+2,\ldots k+l\} = v$. The most straightforward case of the theorem is when neither $k$ nor $l$ is equal to $3$, proved in Proposition \ref{case:neither3}. To prove this we use a generalisation of Gochev and Gotchev's construction \cite{Gochev}. The idea of the proof is as follows: first, we construct $k$ disjoint matchings that use only edges in directions $1,\ldots k$. The matchings used within the $Q_k^v$s are those obtained from applying Corollary \ref{cor:2Stong} to $Q_k$. Next we construct $l$ disjoint matchings that use only edges in directions $k+1,\ldots, k+l$. Similarly, the matchings used within the $Q_l^u$s are those obtained from applying Corollary \ref{cor:2Stong} to $Q_l$. We then prove that taking the union of a matching of the first kind and a matching of the second kind gives a Hamilton cycle. The second case of the theorem is when $k=3$ and $l$ is not equal to $1$ or $3$, proved in Proposition \ref{case:k=3}. We use a similar construction to the first case, the only difference being that while we can use Corollary \ref{cor:2Stong} on $Q_l$, we cannot apply it to $Q_3$. We will instead take directional matchings on the copies of $Q_3$; it turns out this can be made to work here. Finally, we are left with two cases: $(k,l) = (3,1)$ and $(k,l) =(3,3)$. The first of these is proved in Proposition \ref{case:k=3,l=1} by means of an explicit example. The case $(k,l) =(3,3)$ is left unsolved. The difficulty of these final two cases is discussed in the section \ref{sec:direction}. The following useful notation is common to the proofs of propositions \ref{case:neither3} and \ref{case:k=3}. For a perfect matching $M$ and a vertex $v$, we define $M(v)$ to be the other endpoint of the edge containing $v$ in $M$. (Note that this clashes slightly with our notation in Theorem \ref{thm:Bipartite}: by that notation we are here conflating $M$ and $M^{-1}$.) \begin{prop} When neither $k$ nor $l$ is equal to 3, there is a 1-factorization $\mathcal{M}$ of the hypercube $Q_{k+l}$ such that $G[\mathcal{M}]$ is isomorphic to the complete bipartite graph $K_{k,l}$. \label{case:neither3} \end{prop} \begin{proof} Using Corollary \ref{cor:2Stong} partition $Q_k$ into matchings $A_1,A_2,\ldots,A_k$ and matchings $B_1,B_2,\ldots,B_k$ such that $A_i \cup B_i$ is a Hamilton cycle for all $i$. For $i = 1,2,\ldots,k$ define $M_i$ to be the matching on $Q_{k+l}$ defined by taking the following edges: $$\begin{cases} A_i & \text{ on } Q_k^{\emptyset} \\ B_i & \text{ on } Q_k^v \text{ for } v\ne \emptyset \end{cases} $$ Note that the $M_i$ are all disjoint, and they only use edges in directions $1,2,\ldots,k$. Also partition $Q_l$ into matchings $X_1,X_2,\ldots,X_l$ and matchings $Y_1,Y_2,\ldots,Y_l$ such that $X_j \cup Y_j$ is a Hamilton cycle for all $j$. For $j=1,2,\ldots,l$ define $N_j$ to be the matching on $Q_{k+l}$ defined by taking the following edges: $$\begin{cases} X_j & \text{ on } Q_l^u \text{ for } u \text{ even} \\ Y_j & \text{ on } Q_l^u \text{ for } u \text{ odd} \end{cases} $$ Another way to think of $N_j$ is as containing edges between copies of $Q_k^v$. From an even vertex in $Q_k^v$ we add an edge to the corresponding vertex in $Q_k^{X_j(v)}$, and from an odd vertex in $Q_k^v$ we add an edge to the corresponding vertex in $Q_k^{Y_j(v)}$. Note that the $N_j$ are all disjoint, and they only use edges in directions $k+1,k+2,\ldots,k+l$. Thus the matchings $\{M_i\}_{i=1}^k \cup \{N_j\}_{j=1}^l$ are all disjoint and form a 1-factorization of $Q_{k+l}$. \begin{figure}[h] \centering \begin{subfigure}[b]{0.45\textwidth} \includegraphics[width=\textwidth]{Matching1.eps} \caption{$M_i$} \end{subfigure} \qquad \begin{subfigure}[b]{0.45\textwidth} \includegraphics[width=\textwidth]{Matching2.eps} \caption{$N_j$} \end{subfigure} \caption{An example when $k=2$ and $l=4$}\label{fig:general_k,l} \end{figure} All that is left is to show that $M_i \cup N_j$ is a Hamilton cycle for all $i,j$. Consider following the cycle starting at a vertex $u$ that lies in $Q_k^{\emptyset}$ and alternating between edges first in $N_j$ and then in $M_i$. Every time we travel along an edge in $M_i$ the parity of the vertex in $Q_k^v$ switches, and so we will alternate using edges from $X_j$ and edges from $Y_j$ in $N_j$. As $X_j \cup Y_j$ is a Hamilton cycle, the first time the cycle returns to $Q_k^{\emptyset}$ we will have travelled through each other $Q_k^v$ exactly once. Each time we travel through a different $Q_k^v$ we use an edge from $B_i$ within it. After passing through $2^l - 1$ copies of $Q_k^v$ we will have bounced between $u$ and $B_i(u)$ an odd number of times, so the first vertex we encounter in our return to $Q_k^{\emptyset}$ is $B_i(u)$. The next vertex would then be $A_i(B_i(u))$. After passing through $2(2^l)$ distinct vertices (two in each $Q_k^v$) we have moved from $u$ to $A_i(B_i(u))$, i.e. made two steps of the Hamilton cycle $A_i \cup B_i$ within $Q_k^{\emptyset}$. Thus the first time we will return to $u \cup \emptyset$ is after passing through $2^k2^l$ vertices, which is the total number of vertices in the graph. Hence we have a Hamilton cycle. \end{proof} \begin{prop} For $l$ not equal to $1$ or $3$, there is a 1-factorization $\mathcal{M}$ of the hypercube $Q_{3+l}$ such that $G[\mathcal{M}]$ is isomorphic to the complete bipartite graph $K_{3,l}$. \label{case:k=3} \end{prop} \begin{proof} Using Corollary \ref{cor:2Stong}, partition $Q_l$ into matchings $A_1,A_2,\ldots,A_l$ and $B_1,B_2,\ldots, B_l$ such that $A_i\cup B_i$ is a Hamilton cycle for all $j$. Let $X_1, X_2$ and $X_3$ be the three directional matchings of $Q_3$ -- that is, $X_j$ contains all edges in direction $j$. For $i=1,2,\ldots,l$ define $M_i$ to be the matching on $Q_{3+l}$ defined by taking the following edges: $$\begin{cases} A_i & \text{ on } Q_l^{\emptyset}, Q_l^{\{1,2\}}, Q_l^{\{1,3\}}, Q_l^{\{2,3\}} \text{ and } Q_l^{\{1,2,3\}} \\ B_i & \text{ on } Q_l^{\{1\}}, Q_l^{\{2\}} \text{ and } Q_l^{\{3\}} \end{cases} $$ For $j = 1,2,3$ define $N_j$ to be the matching on $Q_{3+l}$ defined by taking the following edges, where the subscripts for the $X$s are taken modulo $3$: $$\begin{cases} X_{j} & \text{ on } Q_3^v \text{ for $v$ odd} \\ X_{j+1} & \text{ on } Q_3^v \text{ for $v$ even and } v\ne \emptyset \\ X_{j+2} & \text{ on } Q_3^{\emptyset} \end{cases} $$ \begin{figure}[h] \centering \begin{subfigure}[b]{0.28\textwidth} \includegraphics[width=\textwidth]{3Matching1.eps} \caption{$M_i$} \end{subfigure} \quad \begin{subfigure}[b]{0.28\textwidth} \includegraphics[width=\textwidth]{3Matching2.eps} \caption{$N_1$} \end{subfigure} \qquad \begin{subfigure}[b]{0.28\textwidth} \includegraphics[width=\textwidth]{3AlmostMatching.eps} \caption{Sketch of cycle ${M_i\cup N_1}$} \label{subfig:sketch} \end{subfigure} \caption{An example when $l=2$}\label{fig:l=3} \end{figure} Now $\{M_i\}_{i=1}^l \cup \{N_j\}_{j=1}^3$ is a set of $3+l$ disjoint perfect matchings. It remains to show that $M_i \cup N_j$ is a Hamilton cycle for any $i$ and $j$. Note that $\{M_i\}$ is invariant under the permutation that cycles directions 1,2 and 3. Since $N_2$ and $N_3$ are obtained from $N_1$ by such cyclic permutations, we can without loss of generality assume that $j=1$. Consider $M_i \cup N_1$ with the edges in $Q_3^{\emptyset}$ removed; that is, the edges $\emptyset \{3\}$, $\{1\}\{1,3\}$, $\{2\}\{2,3\}$ and $\{1,2\}\{1,2,3\}$. We will show that the resulting graph comprises four paths, from $\emptyset$ to $\{2\}$, from $\{2,3\}$ to $\{1,2,3\}$, from $\{1,2\}$ to $\{1\}$ and from $\{1,3\}$ to $\{3\}$. Thus when we add back the four edges in direction 3, we get a Hamilton cycle. See figure \ref{subfig:sketch} for an example. View $Q_{3+l}$ as an $l$-dimensional hypercube whose `vertices' are copies of $Q_3$. Starting at a vertex in $Q_3^\emptyset$ and following the path from it, we will not return to $Q_3^\emptyset$ until we have made $2^l$ steps around $A_i \cup B_i$. A path starting at $\emptyset$ will move in directions according to $A_i$ then $X_1$ then $B_i$ then $X_2$ and then repeat this pattern. It will return to $Q_3^\emptyset$ after $2^l$ moves from $A_i \cup B_i$ and $2^l- 1$ moves from $X_1 \cup X_2$. Since $l \ge 2$, this means we end at the vertex $\{2\}$, and the path contains $2(2^l)$ vertices. The same argument works to show that there is a path from $\{1,2\}$ to $\{1\}$ containing $2(2^l)$ vertices. A path starting at $\{2,3\}$ will move in directions according to $A_i$ then $X_1$ then $A_i$, ending at the vertex $\{1,2,3\}$ and containing $4$ vertices. A path starting at $\{1,3\}$ will move in directions according to $A_i, X_1, B_i, X_2, A_i, X_1, A_i, X_2$, and then repeat this pattern. It will return to $Q_3^\emptyset$ after $2(2^l)-2$ moves from $A_i \cup B_i$ and $2(2^l)- 3$ moves from $X_1 \cup X_2$. Thus we end at the vertex $\{3\}$, and the path contains $4(2^l)-4$ vertices. The sum of the lengths of these paths is $8(2^l)$, and so every vertex is contained in one of these paths. \end{proof} \begin{prop} There is a 1-factorization $\mathcal{M}$ of the hypercube $Q_{4}$ such that $G[\mathcal{M}]$ is isomorphic to the complete bipartite graph $K_{3,1}$. \label{case:k=3,l=1} \end{prop} \begin{proof} The four matchings are shown in figure \ref{fig:3-1}. It is easy to check that the top matching forms a Hamilton cycle with any of the three bottom matchings (in fact, by symmetry you need only check one pair). \begin{figure}[h!] \centering \includegraphics[width=.8\textwidth]{3-1Matching.eps} \caption{The matchings for $k=1$ and $l=3$}\label{fig:3-1} \end{figure} \end{proof} \begin{proof}[Proof of Theorem \ref{thm:main}] Combine the results of Propositions \ref{case:neither3}, \ref{case:k=3} and \ref{case:k=3,l=1}. \end{proof} \section{Direction Respecting 1-Factorizations} \label{sec:direction} The only case not covered by Theorem \ref{thm:main} is whether $G[\mathcal{M}]$ can be isomorphic to $K_{3,3}$. This case cannot be resolved with our methods alone. To explain why, we will introduce a notion of direction respecting 1-factorizations. Fix $k$ and $l$ and let $\mathcal{M} = M_1,M_2,\ldots M_{k+l}$ be a 1-factorization of $Q_{k+l}$. We call the 1-factorization $\mathcal{M}$ \emph{direction respecting} if $M_1,M_2,\ldots, M_k$ only use edges in directions $1,\ldots,k$ and $M_{k+1},M_{k+2},\ldots M_{k+l}$ only use edges in directions $k+1,\ldots k+l$. Note that the matchings constructed in Propositions \ref{case:neither3} and \ref{case:k=3} were direction respecting for the appropriate $k$ and $l$. However, the 1-factorisation given in the proof of proposition \ref{case:k=3,l=1} was not direction respecting. We shall prove that there is no direction respecting 1-factorization $\mathcal{M}$ with $G[\mathcal{M}]$ isomorphic to $K_{3,3}$ or $K_{3,1}$. \begin{theorem} \label{thm:why3hard} There is a direction respecting 1-factorization $\mathcal{M}$ of $Q_{k+l}$ with $G[\mathcal{M}] = K_{k,l}$ if and only if $k,l$ are not $3,1$ or $3,3$. \end{theorem} \begin{proof} First note that the proof of Theorem \ref{thm:main} shows that such a 1-factorization exists when $k,l$ are not $1,3$ or $3,3$. Let $d = 3+l$ where $l$ is 3 or 1. For $M$ a perfect matching on $Q_d$, think of $M$ as a bijection from the odd vertices of $Q_d$ to the even vertices (as in Theorem \ref{thm:Bipartite}). If $M$ and $N$ are perfect matchings then $M N^{-1}$ is a permutation on the even vertices of $Q_d$. We define the \emph{sign} of a 1-factorization $\{M_i\}_{i=1}^{d}$ of $Q_{d}$ to be the product of the signs of the permutations $M_i M_j^{-1}$ for all $i < j$. That is, $$\sign(\mathcal{M}) = \prod_{i<j} \sign\left(M_i M_j^{-1}\right).$$ Let $\mathcal{D}^{(d)} = \{D^{(d)}_i\}_{i=1}^{n}$ be the directional matchings of $Q_d$, where $D^{(d)}_i$ contains all edges in direction $i$. For $i \ne j$ the permutation $D^{(d)}_i \left(D^{(d)}_j\right)^{-1}$ consists of $2^{d-2}$ disjoint 4-cycles. Thus $\sign\left(D^{(d)}_i \left(D^{(d)}_j \right)^{-1}\right) = (-1)^{2^{d-2}} = 1$ for all $i,j$, and so $\sign(\mathcal{D}^{(d)})=1$. Suppose $\mathcal{M} = \{M_i\}_{i=1}^3\cup\{N_j\}_{j=1}^l$ is a 1-factorization of $Q_d$ where $M_i \cup N_j$ is a Hamilton cycle for all $i,j$. The permutation $M_i N_j^{-1}$ is a cycle of length $2^{d-1}$ and so has sign $-1$. Note that $\sign\left(M_i M_s^{-1}\right) = \sign\left(M_i N_j^{-1}\right)\left((M_s N_j^{-1})^{-1}\right) = (-1)(-1) = 1$, and similarly $\sign(N_j N_t^{-1}) = 1$. Thus $\sign(\mathcal{M}) = (-1)^{3l} = -1$ for any $\mathcal{M}$ with $G[\mathcal{M}] = K_{3,l}$. We will define a switching operation on 1-factorizations that preserves their sign. We will further show that any direction respecting 1-factorization $\mathcal{M}$ can be obtained from $\mathcal{D}^{(d)}$ using a series of switches. Since the sign of $\mathcal{D}^{(d)}$ is $1$, this is enough to show that $G[\mathcal{M}] \ne K_{3,l}$. Let $\mathcal{M} = \{M_i\}_{i=1}^d$ be a 1-factorization of $Q_d$. Take a 4-cycle $x,y,v,w$ in $Q_{d}$ and suppose that the edges $xy$ and $vw$ are in matching $M_s$ and $vy$ and $xw$ are in matching $M_t$. A switch on $w,v,y,w$ replaces $\mathcal{M}$ by the 1-factorization $\mathcal{M'} = \{M'_i\}_{i=1}^n$ where $M'_s = M_s \cup \{ vy, xw\} \setminus \{ xy, vw\}$, $M_t' = M_t \cup \{ xy, vw\} \setminus \{vy, xw\}$, and $M'_i = M_i$ for $i \ne s,t$. Viewing the 1-factors $M_i$ as bijections from the even vertices to the odd vertices, we have composed $M_s$ and $M_t$ with the function swapping $x$ and $v$, where $x$ and $v$ are the even vertices of $x,y,v,w$. Therefore the permutations $M_s M_i^{-1}$ and $M'_s M_i^{-1}$, where $i \ne s,t$, differ from each other by the transposition $(a,c)$ and so have opposite sign. Similarly $M_t M_i^{-1}$ and $M'_t M_i^{-1}$ have opposite sign. From this second interpretation of the switch it is clear that: \begin{align*} \sign(\mathcal{M'}) &= \prod_{i<j} \sign\left(M'_i (M'_j)^{-1}\right)) \\ &= \prod_{\substack{\text{exactly one of} \\ \text{$i,j$ is $s$ or $t$}}} - \sign\left(M_i M_j^{-1}\right) \prod_{\substack{\text{neither or both of} \\ \text{$i,j$ is $s$ or $t$}}} \sign\left(M_i M_j^{-1}\right) \\ &= (-1)^{2(d-2)} \sign(\mathcal{M}) = \sign(\mathcal{M}) \end{align*} All that is left to show is that a 1-factorization satisfying the conditions of the theorem can be obtained from $\mathcal{D}^{(d)}$ by a series of switches. We will use the following claim. \begin{claim} Let $D^{(3)}_1,D^{(3)}_2,D^{(3)}_3$ be the directional matchings on $Q_3$ and let $A_1,A_2,A_3$ be another 1-factorization of $Q_3$. Then there are a series of switches that transform $D^{(3)}_1$, $D^{(3)}_2$, $D^{(3)}_3$ into $A_1,A_2,A_3$, respecting the ordering. \end{claim} \begin{proof}[Proof of Claim] It is easy to check that there are only 4 ways to partition $Q_3$ into perfect matchings, up to ordering -- one way uses three directional matchings and the other three ways each use one directional matching. Without loss of generality say that $A_1$ is a directional matching. Note that we can use switches to re-order $D^{(3)}_1,D^{(3)}_2,D^{(3)}_3$. To swap $D^{(3)}_i$ and $D^{(3)}_j$ switch on $\emptyset, \{i\}, \{j\}, \{i,j\}$ and on $\{k\}, \{i,k\}, \{j,k\}, \{i,j,k\}$, where $i,j,k$ is $1,2,3$ in some order. Thus we can assume without loss of generality that $A_1 = D^{(3)}_1$. If $A_2$ and $A_3$ are also directional matchings then we are done. If not, then we can switch on $\emptyset, \{2\},\{2,3\},\{3\}$ to make them both directional matchings. \end{proof} Let $\mathcal{M} = \{M_i\}_{i=1}^3\cup\{N_j\}_{j=1}^l$ be a 1-factorization of $Q_d$ satisfying the conditions of the theorem. As in Theorem \ref{thm:main}, we can view $Q_{3+l}$ as an $l$ dimensional hypercube whose `vertices' are copies of $Q_3$. For $v \subset \{3+1,\ldots, 3+l\}$ let $Q_3^v$ be the induced subgraph of $Q_{3+l}$ on vertices of the form $u \cup v$ for all $u \subset \{1,2,3\}$. For each $v$ in turn, apply the claim to $Q_3^v$ and $M_1,M_2,M_3$ restricted to $Q_3^v$. In this way we obtain a series of switches that turns $D^{(d)}_1,D^{(d)}_2,D^{(d)}_3$ into $M_1,M_2,M_3$. If $l= 1$, $N_1 = D^{(n)}_4$ and we are done. If $l = 3$, apply an analagous process to above to find switches that turn $D^{(d)}_4,D^{(d)}_5,D^{(d)}_6$ into $N_1,N_2,N_3$. Note that these switches will be only on edges in directions 4,5,6 and so will not interfere with $M_1,M_2,M_3$ in any way. \end{proof} \section{Open Questions} \label{sec:questions} The most obvious question is the missing case from Theorem \ref{thm:main}. \begin{question} Is it possible to find a 1-factorization $\mathcal{M}$ of $Q_6$ such that $G[\mathcal{M}] = K_{3,3}$? \end{question} Theorem \ref{thm:why3hard} and its proof show that any such matching $\mathcal{M}$ cannot be obtained from applying a series of switches to the directional matchings. However, there is an example where $G[\mathcal{M}] = K_{3,1}$, and computer checking suggests that in 4 or 5 dimensions there are many other ways to 1-factorize $Q_d$ and get complete bipartite graphs than the way shown in this paper. We know from Theorem \ref{thm:Bipartite} that we cannot have a perfect 1-factorization of $Q_d$ for $d>2$. In fact, the maximum possible number of pairs of 1-factors whose union forms a Hamilton cycle is $\left\lfloor \frac{d^2}{4} \right\rfloor$, obtained when $G[\mathcal{M}] = K_{\lfloor d/2 \rfloor,\lceil d/2 \rceil}$. What can be said about the other pairs -- can their union be close to a Hamilton cycle in some way? \begin{question} Let $\mathcal{M} = \{M_i\}_{i=1}^d$ be a 1-factorization of $Q_d$. Is it possible for $M_i \cup M_j$ to contain a cycle of length $(1-o(1))2^d$ for every $i\ne j$? \end{question} \begin{question} Let $\mathcal{M} = \{M_i\}_{i=1}^d$ be a 1-factorization of $Q_d$. Is it possible for $M_i \cup M_j$ to consist of at most 2 cycles for every $i\ne j$? \end{question} Computer checking shows that for $n \le 5$ the answer to the latter question is `yes' and in dimensions 4 and 5 there are actually several different 1-factorizations that work. One could also phrase more general versions of these questions in terms of finding bounds on an appropriate minimax or maximin function. For example, \begin{question} For a 1-factorization $\mathcal{M} = \{M_i\}_{i=1}^d$ of $Q_d$, let $c_{i,j}$ be the length of the longest cycle in $M_i \cup M_j$ and let $f(\mathcal{M}) = min_{i\ne j}(c_{i,j})$. Can one find bounds on $max_{\mathcal{M}}(f(\mathcal{M}))$ in terms of $d$? \end{question} We can prove that $max_{\mathcal{M}}(f(\mathcal{M}))$ is non-decreasing with $d$. We suspect that it grows exponentially in $d$, but we cannot yet prove it is even better than constant. A different way of thinking of Hamilton cycles is as connected 2-factors. Thus a different generalisation of the problem would be to ask about the connectivity of other $r$-factors. For example, \begin{question} For each $d$, let $r=r(d)$ be minimal subject to there existing a 1-factorization $\mathcal{M}$ of $Q_d$ where the union of any $r$ distinct 1-factors is connected. What is the value of $r(d)$? \end{question} Theorem \ref{thm:Bipartite} shows that $r(d)$ is greater than $2$ for $d>2$. The 1-factorization given by Theorem \ref{thm:main} in the case $k = \left\lfloor \frac{d}{2} \right\rfloor$ and $l = \left\lceil \frac{d}{2} \right\rceil$ has the property that the union of any $\left( \left\lceil \frac{d}{2} \right\rceil + 1 \right)$ 1-factors is connected, hence $r(d) \le \left\lceil \frac{d}{2} \right\rceil + 1$ for $d \ne 6$. It seems possible that $r$ is constant and it could be even as small as 3. \nocite{*} \bibliographystyle{alpha}

https://arxiv.org/abs/1407.4339

Extension from Precoloured Sets of Edges

We consider precolouring extension problems for proper edge-colourings of graphs and multigraphs, in an attempt to prove stronger versions of Vizing's and Shannon's bounds on the chromatic index of (multi)graphs in terms of their maximum degree $\Delta$. We are especially interested in the following question: when is it possible to extend a precoloured matching to a colouring of all edges of a (multi)graph? This question turns out to be related to the notorious List Colouring Conjecture and other classic notions of choosability.

\section{Introduction} Let $G = (V,E)$ be a (multi)graph and let $\mathcal{K}=[K]=\{1,\dots,K\}$ be a palette of available colours. (In this paper, a \emph{multigraph} can have multiple edges, but no loops; while a \emph{graph} is always simple.) We consider the following question: given a subset $S\subseteq E$ of edges and a proper colouring of elements of~$S$ (i.e., adjacent edges must receive distinct colours) using only colours from~$\mathcal{K}$, is there a proper colouring of all edges of~$G$ (again using only colours from~$\mathcal{K}$) in concordance with the given colouring on~$S$? We may consider the set~$S$ as a set of \emph{precoloured} edges, while the full colouring, if it exists, may be considered as \emph{extending} the precolouring. Clearly, if the set~$S$ forms a matching in~$G$, then the precolouring of~$S$ can be arbitrary from~$\mathcal{K}$. An early appearance of a problem regarding precolouring extension of edge-colourings can be found in Marcotte and Seymour~\cite{MS90}. Nevertheless, since then its counterpart for vertex-colourings has been more comprehensively studied. We hope to provoke interest in edge-precolouring extension and in particular introduce the following conjecture. \begin{conjecture}\label{conj:main}\mbox{}\\* Let~$G$ be a multigraph with maximum degree~$\Delta(G)$ and maximum multiplicity~$\mu(G)$. Using the palette $\mathcal{K}=[\Delta(G)+\mu(G)]$, any precoloured matching can be extended to a proper edge-colouring of all of~$G$. \end{conjecture} \noindent If true, Conjecture~\ref{conj:main} would extend Vizing's theorem~\cite{Viz64}, which is independently due to Gupta, cf.~\cite{Gup74}. It strengthens a conjecture of Albertson and Moore \cite[Conj.~1]{AlMo01}, which deals with graphs and imposes a stronger minimum distance condition between precoloured edges. A weaker form of Conjecture~\ref{conj:main} was proved by Berge and Fournier~\cite[Cor.~2]{BeFo91} --- they showed that extension is guaranteed, provided that all edges of the matching have been precoloured with the same colour. Note that even for trees, Conjecture~\ref{conj:main} is false if we replace the palette~$\mathcal{K}$ by $[\Delta(G)]$ or by $[\chi'(G)]$ (where $\chi'(G)$ is, as usual, the chromatic index of $G$): consider a star with all leaves subdivided exactly once; see Figure~\ref{fig:extree}. \begin{figure} \centering \input{extree} \caption{A representative of a class of bipartite planar graphs~$G$, with a non-extendable precoloured (distance-$2$) matching using palette $[\Delta(G)]=[\chi'(G)]$. Dashed lines indicate edges precoloured with colour~$1$.\label{fig:extree}} \end{figure} The main contribution of this paper is to provide evidence in support of Conjecture~\ref{conj:main} by confirming it in some important special cases, in particular, for bipartite multigraphs, subcubic multigraphs, and planar graphs of large enough maximum degree. We discuss this further in Subsection~\ref{sub:main}; however, first allow us to place the conjecture in context by giving some preliminary observations. The obvious relationship between edge-precolouring and its vertex counterpart --- in which we can see edge-precolouring extension of~$G$ as vertex-precolouring extension in its line graph~$L(G)$ --- yields immediate implications. For us, the distance between two edges in~$G$ is their corresponding distance in~$L(G)$, i.e., the number of \emph{vertices} contained in a shortest path in~$G$ between any of their end-vertices. A \emph{distance-$t$ matching} is a set of edges having pairwise distance greater than $t$. (This means that a matching is a distance-$1$ matching, while an induced matching is a distance-$2$ matching. Any set of edges is a distance-$0$ matching.) We point out the following consequence of a result of Albertson \cite[Thm.~4]{Alb98} (see Subsection~\ref{sub:background}) and Vizing's theorem. This consequence is in two senses a weaker statement than that of Conjecture~\ref{conj:main}. \begin{proposition}\label{prop:albertson1}\mbox{}\\* Let~$G$ be a multigraph with maximum degree $\Delta(G)$ and maximum multiplicity $\mu(G)$. Using the palette $\mathcal{K}=[\Delta(G)+\mu(G)+1]$, any precoloured distance-$3$ matching can be extended to a proper edge-colouring of all of~$G$. \end{proposition} \noindent Evidently, we believe that in edge-precolouring the distance requirement ought to be not as strong as it is for vertex-precolouring extension. Naturally, one might wonder if a strong enough distance requirement on the precoloured set yields an improved palette bound, i.e., with palette $[\Delta(G)]$ or $[\chi'(G)]$. This fails however even for bipartite graphs; see Figure~\ref{fig:exdistant}. \begin{figure} \centering \input{exdistant} \caption{A representative of a class of bipartite graphs $G$, with a non-extendable matching consisting of two edges that can be made arbitrarily distant, using the palette $[\Delta(G)]=[\chi'(G)]$. Dashed lines indicate edges precoloured with colour $1$. (There is a general construction for $\Delta(G)$ even: within each diamond-shaped block replace each of the two independent sets of size three with one of size $\tfrac12\Delta(G)$ and replace the central independent set of size five with one of size $\Delta(G)-1$; next increase the distance between the precoloured edges by replacing the two blocks by a chain of arbitrarily many blocks.) \label{fig:exdistant}} \end{figure} By the following easy observation, Conjecture~\ref{conj:main} is also related to list edge-colouring, and therefore to the \emph{List Colouring Conjecture} (LCC) which states that $\ch'(G)=\chi'(G)$ for any multigraph~$G$ (where $\ch'(G)$ is, as usual, the list chromatic index of $G$). For a non-precoloured edge, we define its \emph{precoloured degree} as the number of adjacent precoloured edges. \begin{proposition}\label{prop:choose}\mbox{}\\* Let~$G$ be a multigraph with list chromatic index $\ch'(G)$. For a positive integer~$k$, take the palette as $\mathcal{K}=[\ch'(G)+k]$. If~$G$ is properly precoloured so that the precoloured degree of any non-precoloured edge is at most~$k$, then the precolouring can be extended to a proper edge-colouring of all of~$G$. \end{proposition} \noindent So, if we assume that the LCC holds, then the following weak form of Conjecture~\ref{conj:main} holds as well: any precoloured matching can be extended to a proper edge-colouring of all of~$G$ with the palette $\mathcal{K}=[\Delta(G)+\mu(G)+2]$; moreover, if the precoloured matching is induced, then we may use $\mathcal{K}=[\Delta(G)+\mu(G)+1]$. Due to the remarkable work of Kahn~\cite{Kah96b,Kah96a,Kah00} on edge-colourings and list edge-colourings of (multi)graphs, not only does an asymptotic form of Conjecture~\ref{conj:main} hold, but so does a precolouring extension of an asymptotic form of the Goldberg--Seymour Conjecture (which we review in Subsection~\ref{sub:background}). Kahn's theorem and Proposition~\ref{prop:choose} together imply the following. \begin{proposition}\label{prop:kahn}\mbox{}\\* For any $\varepsilon>0$, there exists a constant~$C_\varepsilon$ such that the following holds. For any multigraph~$G$ with $\chi'(G)\ge C_\varepsilon$, any precoloured matching using the palette $\mathcal{K}=[(1+\varepsilon)\chi'(G)]$ can be extended to a proper edge-colouring of all of~$G$. \end{proposition} \noindent If we replace $\chi'(G)$ in the statement above by $\Delta(G)+\mu(G)$ or by the Goldberg--Seymour bound, then the statement remains valid, either due to Vizing's theorem or due to another theorem of Kahn. One of our motivations for the formulation and study of Conjecture~\ref{conj:main} comes from the close connections with vertex-precolouring and with the LCC. \subsection{Main Results}\label{sub:main} Although it appears that the LCC and our conjecture are independent statements, we have obtained several results corresponding to specific areas of success in list edge-colouring. For clarity, in this subsection we mostly present our main results restricted to precoloured matchings, but more general statements hold. In summary, we have confirmed Conjecture~\ref{conj:main} for bipartite multigraphs, subcubic multigraphs, and planar graphs of large enough maximum degree; we have also obtained a precolouring extension variant of Shannon's theorem. Our first result is an edge-precolouring extension of K\H{o}nig's theorem that any bipartite multigraph~$G$ is $\Delta(G)$-edge-colourable, whereas the subsequent result is an edge-precolouring analogue of Shannon's theorem that any multigraph~$G$ is $\bigl\lfloor\tfrac32\Delta(G)\bigr\rfloor$-edge-colourable. \begin{theorem}\label{thm:bipartite}\mbox{}\\* Let~$G$ be a bipartite multigraph with maximum degree $\Delta(G)$. With the palette $\mathcal{K}=[\Delta(G)+1]$, any precoloured matching can be extended to a proper edge-colouring of all of~$G$. \end{theorem} \noindent As we saw in Figures~\ref{fig:extree} and~\ref{fig:exdistant}, the palette size in Theorem~\ref{thm:bipartite} is sharp. \begin{theorem}\label{thm:shannonmain}\mbox{}\\* Let~$G$ be a multigraph with maximum degree $\Delta(G)$. With the palette $\mathcal{K}=\bigl[\,\bigl\lfloor\tfrac32\Delta(G)+\tfrac12\bigr\rfloor\,\bigr]$, any precoloured matching can be extended to a proper edge-colouring of all of~$G$. \end{theorem} \noindent Due to the Shannon multigraphs, this last statement is sharp if $\Delta(G)$ is even, and within~$1$ of being sharp if $\Delta(G)$ is odd. Theorems~\ref{thm:bipartite} and~\ref{thm:shannonmain} are proved in Section~\ref{sec:bipartite} using powerful list colouring tools developed by Borodin, Kostochka and Woodall~\cite{BKW97}. \medskip The following theorem concerns multigraphs that are subcubic, i.e, of maximum degree at most~$3$. Note that Theorem~\ref{thm:subcubic} improves upon Theorem~\ref{thm:shannonmain} for $\Delta(G)=3$. \begin{theorem}\label{thm:subcubic}\mbox{}\\* Let~$G$ be a subcubic multigraph. With the palette $\mathcal{K}=[4]$, any precoloured matching can be extended to a proper edge-colouring of all of~$G$. \end{theorem} \noindent A form of Theorem~\ref{thm:subcubic}, for subcubic graphs and with a distance condition on the precoloured matching, was observed by Albertson and Moore~\cite{AlMo01}. Although the LCC remains open for subcubic graphs, Juvan, Mohar and \v{S}krekovski~\cite{JMS98} have made a significant attempt. (They showed that for any subcubic graph~$G$, if lists of~$3$ colours are given to the edges of a subgraph~$H$ with $\Delta(H)\le2$ and lists of~$4$ colours to the other edges, then~$G$ has a proper edge-colouring using colours from those lists.) Theorem~\ref{thm:subcubic} is a direct corollary of the following theorem, which may be of interest in its own right. Its proof uses a degree-choosability condition and can be found in Section~\ref{sec:gallai}. For a vertex, we define its \emph{precoloured degree} as the number of incident precoloured edges. \begin{theorem}\label{thm:gallai}\mbox{}\\* Let~$G$ be a connected multigraph. Choose a non-negative integer~$k$ so that $\Delta(L(G))\le\Delta(G)+k$, and take the palette as $\mathcal{K}=[\Delta(G)+k]$. If~$G$ is properly precoloured so that the precoloured degree of any vertex is at most~$k$, then the precolouring can be extended to a proper edge-colouring of all of~$G$, except in the following cases: \smallskip \qitem{(a)}$k=0$ and~$G$ is a simple odd cycle; \smallskip \qitem{(b)}$G$ is a triangle with edges of multiplicity $m_1,m_2,m_3$ and $k=\min\{m_1,m_2,m_3\}-1$. \end{theorem} \noindent Note that, when restricted to precoloured matchings, this theorem produces weak or limited bounds for larger maximum degree. On the other hand, if we replace every precoloured edge in the example of Figure~\ref{fig:extree} by a precoloured multi-edge of multiplicity $k$ (or $k+1$) and a precolouring from $[k]$ (or $[k+1]$), we see that the palette bound (or precoloured degree condition) is best possible. In the case where $k=\Delta(L(G))-\Delta(G)$ in Theorem~\ref{thm:gallai}, the number of colours used is equal to the maximum degree of the line graph. In that sense the theorem can be considered as a precolouring extension of Brooks's theorem restricted to line graphs. It is relevant to mention that vertex-precolouring extension versions of Brooks's theorem~\cite{AKW05,Axe03,Rac09} require, among other conditions, a large minimum distance between the precoloured vertices. \medskip The class of planar graphs could be of particular interest. There is a prominent line of work on (list) edge-colouring for this class, which we discuss further in Subsection~\ref{sub:background} and Section~\ref{sec:planar}. Our main contributions to this area are the following results, the second one of which can be viewed as a strengthening of an old result of Vizing~\cite{Viz65}, provided the graph's maximum degree is large enough. \begin{theorem}\label{thm:planar}\mbox{}\\* Let~$G$ be a planar graph with maximum degree $\Delta(G)\ge17$. Using the palette $\mathcal{K}=[\Delta(G)+1]$, any precoloured matching can be extended to a proper edge-colouring of all of~$G$. \end{theorem} \begin{theorem}\label{thm:planar2}\mbox{}\\* Let~$G$ be a planar graph with maximum degree $\Delta(G)\ge20$. Using the palette $\mathcal{K}=[\Delta(G)]$, any precoloured distance-$3$ matching can be extended to a proper edge-colouring of all of~$G$. \end{theorem} \noindent Due to the trees exhibited in Figure~\ref{fig:extree}, the palette size in Theorem~\ref{thm:planar} cannot be reduced, while the minimum distance condition in Theorem~\ref{thm:planar2} cannot be weakened. In Section~\ref{sec:planar}, we give some more results on when it is possible for a precoloured matching in a planar graph to be extended. A summary of the results is given in Table~\ref{tab:planar}. \begin{table} \centering \begin{tabular}{clclp{42mm}} \toprule & Palette $\mathcal{K}$ & Distance $t$ & Max.\ degree $\Delta$ \qquad\qquad\quad & Reference\\ \midrule 1. & $[\Delta+4]$ & 1 & all $\Delta$ & Thm.~\ref{thm:shannonmain} ($\Delta\le7$) and\hspace*{\fill}\linebreak Prop.~\ref{prop:choose} with~\cite{Bon13+} ($\Delta\ge8$)\\ 2. & $[\Delta+3]$ & 1 & $\Delta\le5$; $\Delta\ge8$ & Thm.~\ref{thm:shannonmain}; Prop.~\ref{prop:choose} with~\cite{Bon13+}\\ 3. & $[\Delta+2]$ & 1 & $\Delta\ge12$ & Prop.~\ref{prop:choose} with~\cite{BKW97}\\ 4. & $[\Delta+2]$ & 2 & $\Delta\le4$; $\Delta\ge8$ & \cite{JMS99}; Prop.~\ref{prop:choose} with~\cite{Bon13+}\\ 5. & $[\Delta+2]$ & 3 & all $\Delta$ & Prop.~\ref{prop:albertson1}\\ 6. & $[\Delta+1]$ & 1 & $\Delta\le3$; $\Delta\ge17$ & Thm.~\ref{thm:subcubic}; Thm.~\ref{thm:planar}\\ 7. & \mbox{}$[\Delta+1]$ & \mbox{}2 & $\Delta\ge12$ & Prop.~\ref{prop:choose} with~\cite{BKW97}\\ 8. & $[\Delta]$ & 3 & $\Delta\ge20$ & Thm.~\ref{thm:planar2}\\ \bottomrule \end{tabular} \caption{Summary of edge-precolouring extension results for planar graphs with maximum degree~$\Delta$, when a distance-$t$ matching~$M$ is precoloured using the palette~$\mathcal{K}$. See Section~\ref{sec:planar} for further details how these results can be obtained.}\label{tab:planar} \end{table} \medskip Some basic knowledge of edge-colouring is a prerequisite to the consideration of edge-precolouring extension problems --- we provide some related background in the next subsection. To our frustration, many of the major methods for colouring edges (such as Kempe chains, Vizing fans, Kierstead paths, Tashkinov trees) seem to be rendered useless by precoloured edges. Though Conjecture~\ref{conj:main} may at first seem as if it should be an ``easy extension'' of Vizing's theorem, it might well be very difficult to confirm (if true at all). We are keen to learn of related edge-precolouring results independent of current list colouring methodology. \subsection{Further Background}\label{sub:background} Edge-colouring is a classic area of graph theory. We give a quick overview of some of the most relevant history for our study. The reader is referred to the recent book by Stiebitz, Scheide, Toft and Favrholdt~\cite{SSTF12} for detailed references and fuller insights. The lower bound $\chi'(G)\ge\Delta(G)$ is obviously true for any multigraph $G$. Close to a century ago, K\H{o}nig proved that all bipartite multigraphs meet this lower bound with equality. Shannon~\cite{Sha49} in 1949 proved that $\chi'(G)\le\bigl\lfloor\tfrac32\Delta(G)\bigr\rfloor$ for any multigraph $G$. Somewhat later, Gupta (as mentioned in~\cite{Gup74}) and, independently, Vizing~\cite{Viz64} proved that $\chi'(G)\le\Delta(G)+\mu(G)$ for any multigraph~$G$, so $\chi'(G)\in\{\Delta(G),\Delta(G)+1\}$ if~$G$ is simple. Both the Shannon bound and the Gupta--Vizing bound are tight in general due to the Shannon multigraphs, which are triangles whose multi-edges have balanced multiplicities. (Note however that the latter bound can be improved for specific choices of $\Delta(G)$ and $\mu(G)$ as described in the work of Scheide and Stiebitz~\cite{ScSt09}.) A notable conjecture on edge-colouring arose in the 1970s, on both sides of the iron curtain. The Goldberg--Seymour Conjecture, due independently to Goldberg~\cite{Gol73} and Seymour~\cite{Sey79}, asserts that $\chi'(G)\in\{\Delta(G),\Delta(G)+1,\lceil\rho(G)\rceil\}$ for any multigraph~$G$, where \[\rho(G)=\max\biggl\{\,\frac{2|E(G[T])|}{|T|-1}\;:\; T\subseteq V,\;|T|\ge3,\;\text{$|T|$ odd}\,\biggr\}.\] The parameter $\rho(G)$ is a lower bound on $\chi'(G)$ based on the maximum ratio between the number of edges in~$H$ and the number of edges in a maximum matching of~$H$, taken over induced subgraphs~$H$ of~$G$. This conjecture remains open and is regarded as one of the most important problems in chromatic graph theory. Perhaps the most outstanding progress on this problem is due to Kahn, who established an asymptotic form~\cite{Kah96a}. The list variant of edge-colouring can be traced as far back as list colouring itself. The concept of list colouring was devised independently by Vizing~\cite{Viz76} and Erd\H{o}s, Rubin and Taylor~\cite{ERT80}, with the iron curtain playing its customary role here too. The List Colouring Conjecture (LCC) was already formulated by Vizing as early as 1975 and was independently reformulated several times, a brief historical account of which is given in, e.g., H{\"a}ggkvist and Janssen~\cite{HaJa97}. For more on the LCC, particularly with respect to the probabilistic method, consult the monograph of Molloy and Reed~\cite{MoRe02}. The results on the LCC most relevant to our investigations also happen to be two of the most striking, both from the mid-1990s. First, Galvin~\cite{Gal95} used a beautiful short argument to prove Dinitz's Conjecture (which concerned the extension of partially completed Latin squares), thereby confirming the LCC for bipartite multigraphs. Not long after Galvin's work, Kahn applied powerful probabilistic methods, with inspiration from extremal combinatorics and statistical physics, to asymptotically affirm the LCC~\cite{Kah96a,Kah00}. For more background on Kahn's proof, related methods, and improvements, consult~\cite{HaJa97,MoRe00,MoRe02}. Inspiration for this class of problems may also be taken from list vertex-colouring. For instance, we utilise a degree-choosability criterion due independently to Borodin~\cite{Bor77} and Erd\H{o}s, Rubin and Taylor~\cite{ERT80}. See for example a survey of Alon~\cite{Alo93} for an excellent (if older) survey on list colouring in somewhat more generality. We should mention that part of the motivation for studying list colouring was to use it to attack other, less constrained colouring problems. The connection has gone back in the other direction as well, as precolouring extension demonstrates. Activity in the area of precolouring extension increased dramatically as a result of the startling proof by Thomassen of planar $5$-choosability~\cite{Tho94}; a key ingredient in that proof was a particular type of precolouring extension from some pair of adjacent vertices, according to a specific planar embedding. A little bit later, Thomassen asked about precolouring extension for planar graphs under a more general setup~\cite{Tho97}. Eliding the planarity condition, Albertson~\cite{Alb98} quickly answered Thomassen's question and proved more: in any $k$-colourable graph, for any set of vertices with pairwise minimum distance at least~$4$, any precolouring of that set from the palette $[k+1]$ can be extended to a proper colouring of the entire graph. (This implies Proposition~\ref{prop:albertson1} above.) Since Albertson's seminal work, a large body of research has developed around precolouring extension. But this research has focused almost exclusively on extension of vertex-colourings. One of the few papers we are aware of that deals with edge-precolouring extension is by Marcotte and Seymour~\cite{MS90}, in which a different type of necessary condition for extension is studied --- curiously, this paper predates the above mentioned activity in vertex-precolouring. For planar graphs, there has been significant interest in both edge-colouring and list edge-colouring. It is known that planar graphs $G$ with $\Delta(G)\ge7$ satisfy $\chi'(G)=\Delta(G)$. This was proved in 1965 by Vizing~\cite{Viz65} in the case $\Delta(G)\ge8$, and much later by Sanders and Zhao \cite{SaZh01} for $\Delta(G)=7$. We remark that Theorem~\ref{thm:planar2} strengthens this for $\Delta(G)$ somewhat larger. Vizing conjectured that the same can be said for planar graphs $G$ with $\Delta(G)=6$, but this long-standing question remains open. Vizing also noted that not every planar graph $G$ with $\Delta(G) \in \{4,5\}$ is $\Delta(G)$-edge-colourable. Regarding list edge-colouring, Borodin, Kostochka and Woodall~\cite{BKW97} proved the LCC for planar graphs with maximum degree at least $12$, i.e., they proved that such graphs have list chromatic index equal to their maximum degree. The LCC remains open for planar graphs with smaller maximum degree, though it is known that if $\Delta(G)\le4$, or $\Delta(G)\ge8$, then $\ch'(G)\le\Delta(G)+1$. This is due to Juvan, Mohar and \v{S}krekovski~\cite{JMS99}, and to Bonamy~\cite{Bon13+} and Borodin~\cite{Bor90}, respectively. As noted above, it is not true that planar graphs $G$ with $\Delta(G)\in\{4,5\}$ are always $\Delta(G)$-edge-choosable. \section{Extensions of K\H{o}nig's and Shannon's Theorems} \label{sec:bipartite} Theorem~\ref{thm:predegree} below implies Theorem~\ref{thm:bipartite}, and hence verifies Conjecture~\ref{conj:main} for bipartite multigraphs. Theorem~\ref{thm:shannon} implies Theorem~\ref{thm:shannonmain}. Recall that, for a vertex, we define its precoloured degree as the number of incident precoloured edges. \begin{theorem}\label{thm:predegree}\mbox{}\\* Let~$G$ be a bipartite multigraph and $k\ge1$. Take the palette as $\mathcal{K}=[\Delta(G)+k]$. If~$G$ is properly precoloured so that the precoloured degree of any vertex is at most~$k$, then this precolouring can be extended to a proper edge-colouring of all of $G$. \end{theorem} \begin{theorem}\label{thm:shannon}\mbox{}\\* Let~$G$ be a multigraph and $k\ge1$. Take the palette as $\mathcal{K}=\bigl[\,\bigl\lfloor\tfrac32\Delta(G)+\tfrac12k\bigr\rfloor\,\bigr]$. If~$G$ is properly precoloured so that the precoloured degree of any vertex is at most~$k$, then this precolouring can be extended to a proper edge-colouring of all of~$G$. \end{theorem} \noindent The two results are corollary to two theorems of Borodin, Kostochka and Woodall~\cite{BKW97}. Let~$G$ be a multigraph and $f\colon E(G)\rightarrow\mathbb{Z^+}$. We say~$G$ is \emph{$f$-edge-choosable} if, for any assignment of lists where every edge~$e$ receives a list of size at least~$f(e)$, there is a proper edge-colouring of~$G$ using colours from the lists. \begin{theorem}[Borodin, Kostochka \& Woodall \cite{BKW97}]\label{thm:bkw}\mbox{}\\* Let~$G$ be a bipartite multigraph, and set $f(uv)=\max\{d(u),d(v)\}$ for each edge $uv\in E(G)$. Then~$G$ is $f$-edge-choosable. \end{theorem} \begin{theorem}[Borodin, Kostochka \& Woodall~\cite{BKW97}]\label{thm:bkwshannon}\mbox{}\\* Let~$G$ be a multigraph, and set $f(uv)=\max\{d(u),d(v)\}+ \bigl\lfloor\tfrac12\min\{d(u),d(v)\}\bigr\rfloor$ for each edge $uv\in E(G)$. Then~$G$ is $f$-edge-choosable. \end{theorem} \noindent Note that Theorem~\ref{thm:bkw} is a strengthening of Galvin's theorem; while Theorem~\ref{thm:bkwshannon} is a list colouring version of Shannon's theorem (and in fact follows from Theorem~\ref{thm:bkw}). \begin{proof}[Proof of Theorems~\ref{thm:predegree} and~\ref{thm:shannon}] For a vertex~$v$ in a properly precoloured multigraph~$G$, let $k(v)$ be the number of precoloured edges incident with~$v$. Let~$G'$ be obtained from~$G$ by deleting all precoloured edges. To each edge $e\in E(G')$, assign a list $\ell(e)$ consisting of those colours in~$\mathcal{K}$ that do not appear on precoloured edges adjacent to~$e$ in~$G$. Next consider any uncoloured edge $e=uv$, and assume that $d_G(u)-k(u)=d_{G'}(u)\le d_{G'}(v)=d_G(v)-k(v)$. In the bipartite case, since $\Delta(G)\ge d_G(v)$ and $k\ge k(u)$, we infer that \begin{align*} |\ell(e)|&\ge(\Delta(G)+k)-k(u)-k(v)=(\Delta(G)-k(v))+(k-k(u))\\ &\ge d_{G'}(v)=\max\{d_{G'}(u),d_{G'}(v)\}. \end{align*} As this holds for every edge $e=uv$ in~$G'$, Theorem~\ref{thm:bkw} guarantees a colouring of the edges of~$G'$ from their lists. This colouring is an extension of the precolouring of~$G$, completing the proof of Theorem~\ref{thm:predegree}. In the general case, again because $\Delta(G)\ge\max\{\,d_G(u),d_G(v)\,\}$ and $k\ge k(u)$, we infer that \begin{align*} |\ell(e)|&\ge\bigl\lfloor\tfrac32\Delta(G)+ \tfrac12k\bigr\rfloor-k(u)-k(v)=(\Delta(G)-k(v))+ \bigl\lfloor\tfrac12\Delta(G)+\tfrac12k-k(u)\bigr\rfloor\\ &\ge d_{G'}(v)+\bigl\lfloor\tfrac12(\Delta(G)-k(u))\bigr\rfloor\ge d_{G'}(v)+\bigl\lfloor\tfrac12d_{G'}(u)\bigr\rfloor\\ &=\max\{d_{G'}(u),d_{G'}(v)\}+ \bigl\lfloor\tfrac12\min\{d_{G'}(u),d_{G'}(v)\}\bigr\rfloor. \end{align*} Analogously to the previous paragraph, we use Theorem~\ref{thm:bkwshannon} to complete the proof of Theorem~\ref{thm:shannon}. \end{proof} \section{An Approach using Gallai Trees} \label{sec:gallai} In this section, we use a slight refinement of a result due independently to Borodin~\cite{Bor77} and Erd\H{o}s, Rubin and Taylor~\cite{ERT80}. This is a list version of an older result of Gallai~\cite{Gal63} on colour-critical graphs, and it follows for instance from Theorem~4.2 in~\cite{Tho97}. A connected graph all of whose blocks are either complete graphs or odd cycles is called a \emph{Gallai tree}. \begin{theorem}[Borodin~\cite{Bor77}, Erd\H{o}s, Rubin \& Taylor~\cite{ERT80}]\label{thm:gallaitree}\mbox{}\\* Given a connected graph $G=(V,E)$, let $\ell(v)$, for $v\in V$, be an assignment of lists where each vertex~$v$ receives at least~$d(v)$ colours. Then there is a proper colouring of~$G$ using colours from the lists, unless $G$ is a Gallai tree and $|\ell(v)|=d(v)$ for all~$v$. \end{theorem} \noindent With this we prove Theorem~\ref{thm:gallai}, which implies Theorem~\ref{thm:subcubic}. \begin{proof}[Proof of Theorem~\ref{thm:gallai}] Assume to the contrary that the connected multigraph~$G$ and the non-negative integer~$k$ satisfy $\Delta(L(G))\le\Delta(G)+k$, but that, using the palette $\mathcal{K}=[\Delta(G)+k]$, there is a proper edge-precolouring of~$G$ of the required type that does not extend to a proper edge-colouring of~$G$. For a vertex~$v$, let $K(v)\subseteq\mathcal{K}$ be the set of colours appearing on the precoloured edges incident with~$v$, and set $k(v)=|K(v)|$. Let~$G'$ be obtained from~$G$ by deleting all precoloured edges. To each edge $e=uv$ in~$G'$, we assign a list $\ell(e)$ containing those colours in~$\mathcal{K}$ not appearing on precoloured edges adjacent to~$e$ in~$G$. For any edge $e=uv$ in~$G'$ we obtain, using that $\Delta(G)+k\ge\Delta(L(G))\ge d_{L(G)}(e)$, \begin{equation}\label{eq1} \begin{array}{@{}r@{}l@{}} |\ell(e)|&{}=|\mathcal{K}|-|K(u)\cup K(v)|= (\Delta(G)+k)-|K(u)\cup K(v)|\\[1mm] &{}\ge d_{L(G)}(e)-|K(u)\cup K(v)|=d_{L(G')}(e). \end{array} \end{equation} Since there is no extension of the precolouring of $L(G)$ to a full colouring of $L(G)$, it follows that $L(G')$ is not edge-choosable with the lists $\ell(e)$, for $e\in E(G')$. In particular, there is a component~$C'$ of~$G'$ such that $L(C')$ is not edge-choosable with the lists $\ell(e)$, for $e\in E(C')$. By Theorem~\ref{thm:gallaitree}, $L(C')$ must be a Gallai tree such that $|\ell(e)|=d_{L(C')}(e)$ for every~$e$. This also means that we must have equality in all inequalities used to derive~\eqref{eq1}; in particular: \begin{subequations} \begin{align} &\forall e\in E(C'),\qquad d_{L(G)}(e)=\Delta(L(G))=\Delta(G)+k;\label{eq2a}\\ &\forall e\in E(C'),\qquad |\ell(e)|=d_{L(C')}(e)=d_{L(G)}(e)-|K(u)\cup K(v)|.\label{eq2b} \end{align} \end{subequations} Now note that $d_{C'}(v)+k(v)=d_G(v)\le\Delta(G)$ for every vertex~$v$. So, analogously to~\eqref{eq1} above, we infer that for each edge $e=uv$ in~$C'$ the order of $\ell(e)$ is at least the degree in~$C'$ of each of its end-vertices: \begin{equation}\label{eq3} \begin{array}{@{}r@{}l@{}} |\ell(e)|&{}=(\Delta(G)+k)-|K(u)\cup K(v)|= (\Delta(G)+k)-k(u)-k(v)+|K(u)\cap K(v)|\\[1mm] &{}\ge d_{C'}(v)+(k-k(u))+|K(u)\cap K(v)|\ge d_{C'}(v). \end{array} \end{equation} We require the following statements. \begin{claim}\label{cl1} \ Every vertex in~$C'$ has at least two neighbours. \end{claim} \begin{proof} Suppose to the contrary that the vertex~$u$ has the vertex~$v$ as its unique neighbour. Then for the edge $e=uv$ we have $d_{L(C')}(e)=d_{C'}(v)-1$ (this holds even if~$uv$ is a multi-edge). But since $|\ell(e)|=d_{L(C')}(e)$, this gives $|\ell(e)|<d_{C'}(v)$, contradicting~\eqref{eq3}. \end{proof} \begin{claim}\label{cl2} \ If~$C'$ is a simple odd cycle, then $k=0$ and $G=C'$. \end{claim} \begin{proof} Suppose that~$C'$ is a simple odd cycle. If $e=uv$ is an edge in~$C'$, then $|\ell(e)|=d_{L(C')}(e)=d_{C'}(v)=2$. From this we can assume, by permuting the colours, that $\ell(e)=\{1,2\}$ for every $e\in L(C')$. (Indeed, the only way to assign lists of length~$2$ to the edges of an odd cycle in such a way that there is no proper colouring of the cycle using colours from the lists is by making all lists identical.) There must also be equality everywhere in~\eqref{eq3}. Combining that with~\eqref{eq2b} means in particular that for every edge $e=uv$ we have $K(u)\cup K(v)=\{3,4,\ldots,\Delta(G)+k\}$ and $K(u)\cap K(v)=\varnothing$. By an easy parity argument, we can see that this is only possible if all the sets~$K(u)$, for $u\in V(G)$, are empty. This means that $k=0$. Since~$G$ is connected, if there are no precoloured edges, then~$G$ can have only one component, which must be~$C'$. \end{proof} \noindent We continue by considering the case that~$C'$ is not an odd cycle. Since line graphs are claw-free, it follows that odd cycle blocks of length at least five are impossible in $L(C')$. We deduce that all blocks of $L(C')$ are cliques. The only way that a leaf block~$B$ of $L(C')$ could be part of a nontrivial block structure is if it corresponds to a set of edges in~$C'$ that are all incident with a unique vertex, with one of the edges corresponding to the cut-vertex of~$B$. This is ruled out by Claim~\ref{cl1}. We conclude that $L(C')$ must itself be a clique. In turn, the only way that a line graph $L(C')$ of a multigraph is a clique is if~$C'$ is a star or a triangle, with possibly multiple edges. The first option is ruled out by Claim~\ref{cl1}, so~$C'$ must be a triangle, possibly with multi-edges. Let $u,v,w$ be the vertices in~$C'$ and set $m=|E(C')|$. Then for all $e\in E(C')$ we have $|\ell(e)|=d_{L(C')}=m-1$. It is easy to check that with lists of this size, the only way that~$C'$ is not edge-choosable is if all the lists are the same. This also means that the sets $K(u)\cup K(v)$, $K(u)\cup K(w)$ and $K(v)\cup K(w)$ are the same. Let $A(u)$ be the set of colours that appear on precoloured edges which are incident with~$u$, but not with~$v$ or~$w$; define $A(v)$ and $A(w)$ analogously. (In other words, these are colours on the edges that connect~$C'$ to the rest of the graph~$G$.) Let~$D$ be the set of colours that appear on precoloured edges with end-vertices contained in $\{u,v,w\}$. From~\eqref{eq2a} and~\eqref{eq2b} we deduce that $|\ell(e)|=d_{L(C')}(e)$ for every edge $e$ in $C'$, which, applied to an edge between~$u$ and~$v$, implies that $A(u)\cap A(v)=\varnothing$, $A(u)\cap D=\varnothing$ and $A(v)\cap D=\varnothing$. By symmetry, $A(v)\cap A(w)=\varnothing$, $A(u)\cap A(w)=\varnothing$ and $A(w)\cap D=\varnothing$. Now recall that all edges in~$C'$ must have the same list. Consequently, the disjointness of the sets $A(u)$, $A(v)$ and $A(w)$ implies that these three sets are empty. Thus we find that there are no precoloured edges between any of $u,v,w$ and the rest of the graph. Since~$G$ is connected, it follows that $V(G)=\{u,v,w\}$. Let $m(uv),m(uw),m(vw)$ be the multiplicities of the edges of~$G$. Then $\Delta(L(G))=m(uv)+m(uw)+m(vw)-1$, while $\Delta(G)=m(uv)+m(uw)+m(vw)-\min\{m(uv),m(uw),m(vw)\}$. Since $\Delta(L(G))=\Delta(G)+k$, we have shown that part~(b) of the statement of the theorem holds, completing the proof. \end{proof} \section{Planar Graphs}\label{sec:planar} In this section, for brevity we usually write~$\Delta$ for $\Delta(G)$. In the next subsection we prove Conjecture~\ref{conj:main} for planar graphs of large enough maximum degree (at least~$17$), which is the assertion of Theorem~\ref{thm:planar}. As mentioned earlier, the LCC is known to hold for planar graphs with maximum degree at least~$12$. This is yet another result of Borodin, Kostochka and Woodall~\cite{BKW97}: they indeed show that $\ch'(G)\le\Delta$ for such graphs~$G$. Combining this with Proposition~\ref{prop:choose} gives the bounds in lines~3 and~7 of Table~\ref{tab:planar}. Borodin~\cite{Bor90} showed $\ch'(G)\le\Delta+1$ for planar graphs~$G$ of maximum degree $\Delta\ge9$. Recently, Bonamy~\cite{Bon13+} extended this last statement to the case $\Delta=8$. Combining this result with Proposition~\ref{prop:choose} implies that for planar graphs with maximum degree $\Delta\ge8$ a precoloured matching can be extended to a proper colouring of the entire graph with the palette $[\Delta+3]$, while a precoloured distance-$2$ matching can be extended with the palette $[\Delta+2]$. For smaller values of~$\Delta$, we can use Theorems~\ref{thm:shannonmain} and~\ref{thm:subcubic}, and the result of Juvan, Mohar and \v{S}krekovski~\cite{JMS99} that $\ch'(G)\le\Delta(G)+1$ for a planar graph~$G$ with $\Delta(G)\le4$, to achieve several of the bounds in Table~\ref{tab:planar}. In particular, it follows that $\Delta+4$ colours suffice for any planar graph with maximum degree~$\Delta$. The final proof we present is of Theorem~\ref{thm:planar2}. As discussed in Subsection~\ref{sub:background}, Vizing conjectured~\cite{Viz65} that any planar graph with maximum degree $\Delta\ge6$ has a $\Delta$-edge-colouring. The examples in Figure~\ref{fig:extree} show that this statement is false if we allow an adversarial precolouring of a distance-$2$ matching. But does it remains true with the adversarial precolouring of any distance-$3$ matching? We prove that this is indeed the case if $\Delta\ge20$. We expect that this lower bound on~$\Delta$ can be reduced, though, as noted before, certainly not below~$6$. \medskip The proofs of Theorems~\ref{thm:planar} and~\ref{thm:planar2} can be found in the next two subsections. They use a common framework, terminology and notation, which we outline now. Whenever considering a planar graph~$G$, we fix a drawing of~$G$ in the plane. (So we really should talk about a \emph{plane graph}.) Because of this fixed embedding we can talk about the \emph{faces} of the graph. If~$G$ is connected, then the boundary of any face~$f$ forms a closed walk~$W_f$. We adopt the following notation to classify vertices of a graph~$G$ according to degree and incidence with vertices of degree~$1$. Let~$V_i$ be the set of vertices of degree~$i$. Also, identify by $T_i\subseteq V_i$ those vertices of degree~$i$ that are adjacent to a vertex of degree~$1$, and set $U_i=V_i\setminus T_i$. Write $T=\cup_{i\ge1}T_i$ and $U=V(G)\setminus T$. We also adopt the shorthand notation $V_{[i,j]}$, $U_{[i,j]}$ and $T_{[i,j]}$ to mean, respectively, the sets of vertices in $V$, $U$ and $T$ with degrees between~$i$ and~$j$ inclusively. \subsection{Proof of Theorem~\ref{thm:planar}} The statement of Theorem~\ref{thm:planar} is true for graphs with maximum degree~$17$ and exactly~$17$ edges. We use induction on $E(G)$, and proceed with the induction step. We may easily assume that~$G$ is connected with at least~$18$ vertices, since $\Delta\ge17$. Let~$M$ be a precoloured matching. We first observe that \begin{equation} \label{eq:sum1} \text{if $uv\in E(G)\setminus M$, then $d(u)+d(v)\ge\Delta+3$.} \end{equation} Indeed, suppose that the inequality does not hold for some edge $uv\notin M$. Then, by induction, $M$ can be extended to a colouring of $G-uv$ so that at most~$\Delta$ colours are used on the edges adjacent to~$uv$, and so we can easily extend the colouring further to~$uv$. It follows from this observation that~$G$ has no vertices of degree~$2$, that every vertex with degree~$1$ is incident with an edge of~$M$ and that any vertex has at most one neighbour of degree~$1$. We will use these facts often without reference in the remainder of the proof. For a face~$f$, let $V^-(f)=V(f)\setminus V_1$, and denote by $W^-_f$ the sequence of vertices on the boundary walk~$W_f$ after removing vertices from~$V_1$. For a vertex~$v$, let $v_1,v_2,\ldots,v_{d(v)}$ be the neighbours of~$v$, listed in clockwise order according to the drawing of~$G$. Write~$f_i$ for the face incident with~$v$ lying between the edges $vv_i$ and $vv_{i+1}$ (taking addition modulo~$d(v)$). If $v\in T$ has a (unique) neighbour in~$V_1$, then we always choose~$v_1$ to be this neighbour. In that case we have $f_{d(v)}=f_1$; we denote that face by~$f_1$ again. Note that it is possible for other faces to be the same as well (if~$v$ is a cut-vertex), but we will not identify those multiple names of the same face. So, if $v\in U$, then the faces around~$v$ in consecutive order are $f_1,f_2,\ldots,f_{d(v)}$; while, if $v\in T$, then the faces around~$v$ are $f_1,f_2,\ldots,f_{d(v)-1}$. \begin{claim}\label{claim:vdeltav3} \ $|V_\Delta|>|V_3|$. \end{claim} \begin{proof} Consider the set~$F$ of edges in $E(G)\setminus M$ with one end-vertex in~$V_3$ and the other in~$V_\Delta$. The subgraph with vertex set $V_3\cup V_\Delta$ and edge set~$F$ is bipartite; we claim it is acyclic. For suppose there exists an (even) cycle $C\in F$. By induction, we can extend the precolouring of $E(G)\setminus M$ to $G-C$. But then we can further extend this colouring to the edges of~$C$, since each edge of~$C$ is adjacent to only $\Delta-1$ coloured edges, and even cycles are $2$-edge-choosable. Since each vertex in~$V_3$ is incident with at least two edges in~$F$, we have $|V_\Delta|+|V_3|>|F|\ge2|V_3|$. The claim follows. \end{proof} \noindent We use a discharging argument to continue the proof of the theorem. First, let us assign to each vertex~$v$ a charge \smallskip \qitem{$\alpha1$:} $\alpha(v)=3d(v)-6$, \smallskip\noindent and to each face~$f$ a charge \smallskip \qitem{$\alpha2$:} $\alpha(f)=-6$. \smallskip For each vertex~$v$ we define $\beta(v)$ as follows. \smallskip \qitem{$\beta1$:} If $v\in V_\Delta$, then $\beta(v)=-2$. \smallskip \qitem{$\beta2$:} If $v\in V_3$, then $\beta(v)=2$. \smallskip \qitem{$\beta3$:} In all other cases, $\beta(v)=0$. \smallskip For each vertex~$v$ and edge $e=vu$, we define $\gamma_e(v)$ and $\gamma_e(u)$ as follows. \smallskip \qitem{$\gamma1$:} If $v\in V_1$, then $\gamma_e(v)=-\gamma_e(u)=3$. \smallskip \qitem{$\gamma2$:} If $v,u\notin V_1$, then $\gamma_e(v)=\gamma_e(u)=0$. \smallskip Finally, for each face~$f$ and vertex $v\in W^-_f$ we define $\delta_{f}(v)$ and $\delta_v(f)$ as follows. \smallskip \qitem{$\delta1$:} If $v\in T_3$, then $\delta_v(f)=-\delta_{f}(v)=1$. \smallskip \qitem{$\delta2$:} If $v\in U_3$, then $\delta_v(f)=-\delta_{f}(v)=\tfrac53$. \smallskip \qitem{$\delta3$:} If $v\in T$ and $4\le d(v)\le\Delta-2$, then $\delta_v(f)=-\delta_{f}(v)=3-\dfrac{6}{d(v)-1}$. \smallskip \qitem{$\delta4$:} If $v\in U$ and $4\le d(v)\le\Delta-2$, then $\delta_v(f)=-\delta_{f}(v)=3-\dfrac{6}{d(v)}$. \smallskip \qitem{$\delta5$:} If $d(v)\ge\Delta-1$, $|V^-(f)|=3$, and both neighbours of~$v$ in $V^-(f)$ are vertices in $U_{[3,8]}$ that are joined by an edge in~$M$, then $\delta_v(f)=-\delta_{f}(v)=3$. \smallskip \qitem{$\delta6$:} If $d(v)\ge\Delta-1$, $|V^-(f)|=3$, and~$v$ has a neighbour in $V^-(f)\cap(T_{[3,6]}\cup U_{[3,5]})$, then $\delta_v(f)=-\delta_{f}(v)=\tfrac52$. \smallskip \qitem{$\delta7$:} If $d(v)\ge\Delta-1$, $|V^-(f)|=3$, and none of $\delta6$ and $\delta7$ applies, then $\delta_v(f)=-\delta_{f}(v)=2$. \smallskip \qitem{$\delta8$:} If $d(v)\ge\Delta-1$, $|V^-(f)|\ge4$, and~$v$ has a neighbour in $V^-(f)\cap{T_{[3,6]}}$, then $\delta_v(f)=-\delta_{f}(v)=2$. \smallskip \qitem{$\delta9$:} If $d(v)\ge\Delta-1$, $|V^-(f)|\ge4$, and~$\delta9$ does not apply, then $\delta_v(f)=-\delta_{f}(v)=\tfrac32$. \smallskip For a vertex~$v$, write $\gamma(v)$ for the sum of $\gamma_e(v)$ over all edges~$e$ that have~$v$ as an end-vertex. For a vertex~$v$ of degree~$1$ we set $\delta(v)=0$. For any other vertices, write~$\delta(v)$ for the sum over the faces~$f$ around~$v$ of $\delta_f(v)$. Similarly, for a face~$f$, write $\delta(f)$ for the sum over the vertices~$v$ on the reduced walk~$W^-_f$ around~$f$ of the values of $\delta_v(f)$. By the definitions of~$\gamma$ and~$\delta$, \[\textstyle\sum_v\gamma(v)+\sum_v\delta(v)+\sum_{f}\delta(f)=0.\] It follows from Claim~\ref{claim:vdeltav3} that \[\textstyle\sum_v\beta(v)<0.\] Finally, from Euler's Formula for simple plane graphs, we obtain \[\textstyle\sum_{v}\alpha(v)+\sum_{f}\alpha(f)<0.\] Thus, in order to reach a contradiction, it is enough to show that for every vertex~$v$: \begin{equation}\label{eq:discvertex} \alpha(v)+\beta(v)+\delta(v)+\gamma(v)\ge0, \end{equation} and that for every face~$f$: \begin{equation}\label{eq:discregion} \alpha(f)+\delta(f)\ge0. \end{equation} Let~$f$ be a face. As~$G$ is simple, $|V^-(f)|\ge3$. Since $\alpha(f)=-6$, to establish~\eqref{eq:discregion} it is enough to show that $\delta(f)\ge6$. Let~$v$ be the vertex in $V^-(f)$ for which $\delta_v(f)$ is minimum. If $\delta_v(f)\cdot|V^-(f)|\ge6$, then~\eqref{eq:discregion} clearly holds, and so we only need to deal with cases $\delta1$\,--\,$\delta4$. Also, if $v\in T_{[7,\Delta-2]}\cup U_{[6,\Delta-2]}$, then~$\delta3$ and~$\delta4$ give $\delta_v(f)\ge2$, and hence again~\eqref{eq:discregion} is verified. If $v\in T_{[3,4]}$, then $\delta_{v}(f)\ge1$ by~$\delta1$ or~$\delta3$ and, by~\eqref{eq:sum1}, the neighbours~$u$ and~$w$ of~$v$ in $V^-(f)$ have degree at least $\Delta-1$. If $|V^-(f)|=3$, then~$\delta6$ applies to both~$u$ and~$w$, so $\delta_{u}(f)=\delta_{w}(f)=\tfrac52$. If $|V^-(f)|\ge4$, then~$\delta8$ applies, so $\delta_{u}(f)=\delta_{w}(f)=2$, while a fourth vertex~$z$ in~$V^-(f)$ satisfies $\delta_z(f)\ge1$. So~\eqref{eq:discregion} always follows. If $v\in T_{[5,6]}$, then $\delta_v(f)\ge\tfrac32$, so we may assume that $|V^-(f)|=3$ (as $\delta_v(f)\le\delta_u(f)$ whenever~$u\in V^-(f)$). Moreover, $v$ has neighbours $u,w$ in $V^-(f)$ with degree at least $\Delta-3\ge9$. If $d(u)\ge\Delta-1$, then~$\delta6$ gives $\delta_u(f)=\tfrac52$. If $d(u)\in\{\Delta-3,\Delta-2\}$, then~$\delta3$ and~$\delta4$ give $\delta_u(f)\ge\tfrac94$, as $\Delta\ge5$. Since similar bounds hold for $\delta_w(f)$, we deduce that~\eqref{eq:discregion} holds. We are left with the case where $v\in U_{[3,5]}$. By~$\delta2$ and~$\delta4$ we find that $\delta_v(f)\ge\tfrac32$, and hence we again only have to consider the case where $|V(f)|=3$. Rules $\delta3$\,--\,$\delta6$ ensure that any other vertex~$u$ in~$V^-(f)$ with $d(u)\ge9$ satisfies $\delta_u(f)\ge\tfrac94$. So we may suppose that there is a vertex $u\in V^-(f)$ with $d(u)\le8$. Since $|V^-(f)|=3$, we must in fact have $uv\in E(G)$. Moreover, as $\Delta\ge12$, we know by~\eqref{eq:sum1} that the edge~$uv$ belongs to the matching~$M$. This means that $u\in U_{[3,8]}$. Let~$w$ be the third vertex in~$V^-(f)$. If $v\in U_{[3,4]}$, then $d(w)\ge\Delta-1$ by~\eqref{eq:sum1}, and so $\delta_w(f)=3$ by~$\delta5$, confirming~\eqref{eq:discregion}. As the final case, assume that $v\in U_5$ and $\delta_u(f)\ge\delta_v(f)=\tfrac95$. Since also $d(w)\ge\Delta-2\ge11$, one of $\delta3$\,--\,$\delta6$ applies, yielding that $\delta_w(f)\ge\tfrac{12}{5}$. So again $\delta(f)\ge6$, confirming~\eqref{eq:discregion} for all faces. \medskip Now let~$v$ be a vertex. Recall the convention that if $v\in T$, then the two consecutive faces incident with~$v$ neighbouring the neighbour of degree~$1$ are counted as one face, while all other faces are counted separately. If $d(v)=1$, then $\alpha(v)=-3$ and $\gamma(v)=3$. Since $\beta(v)=\delta(v)=0$, we immediately obtain~\eqref{eq:discvertex}. Recall that~$G$ has no vertices of degree~$2$. If $d(v)=3$, then $\alpha(v)=3$, while $\beta(v)=2$ by~$\beta2$. If $v\in T_3$, then $\gamma(v)=-3$ and~$\delta1$ implies that $\delta(v)=-2$. If $v\in U_3$, then $\gamma(v)=0$ and~$\delta2$ implies that $\delta(v)=-5$. This confirms~\eqref{eq:discvertex} if $d(v)=3$. Next suppose that $4\le d(v)\le\Delta-2$. Recall that $\alpha(v)=3d(v)-6$, and observe that $\beta(v)=0$. If $v\in T$, then $\gamma(v)=-3$ by~$\gamma1$. By~$\delta3$ we have $\delta(v)=(d(v)-1)\cdot\Bigl(-3-\dfrac{6}{d(v)-1}\Bigr)=9-3d(v)$. Similarly, if $v\in U$, then $\gamma(v)=0$, and~$\delta4$ implies that $\delta(v)=6-3d(v)$. This proves~\eqref{eq:discvertex} for those vertices~$v$. Now suppose that $d(v)\ge\Delta-1$. As a next step towards proving~\eqref{eq:discvertex}, we consider the average value of $\delta_{f}(v)$ over the faces incident with~$v$. \begin{claim}\label{claim:avg} \ For any two consecutive faces~$f_i$ and $f_{i+1}$ incident with~$v$ it holds that $\delta_{f_i}(v)+\delta_{f_{i+1}}(v)\ge-5$ (where the addition is modulo $d(v)$ if $v\in U$, and modulo $d(v)-1$ if $v\in T$). \end{claim} \begin{proof} We may assume that one of $\delta_{f_i}(v)$ or $\delta_{f_{i+1}}(v)$ is smaller than~$-\tfrac52$, otherwise the stated inequality holds. Suppose first that $\delta_{f_i}(v)<-\tfrac52$. By the definitions of $\delta5$\,--\,$\delta9$, it follows that~$\delta5$ was applied to $f_i$, and consequently $\delta_{f_i}(v)=-3$. Hence~$v_i$ and~$v_{i+1}$ both belong to~$U$, have degree at most~$8$, and there is an edge from~$M$ between them. Since~$M$ is a matching, it cannot also be the case that $\delta_{f_{i+1}}(v)=-3$. If $\delta_{f_{i+1}}(v)=-\tfrac52$, then~$\delta6$ was applied to the face~$f_{i+1}$. But then the vertex~$v_{i+2}$ has degree at most~$6$ and there is an edge $v_{i+1}v_{i+2}$, which contradicts~\eqref{eq:sum1}. So $\delta_{f_{i+1}}(v)\ge-2$ and the statement follows. By symmetry, the same arguments apply if $\delta_{f_{i+1}}(v)<-\tfrac52$. \end{proof} \noindent If we add up the values of $\delta_{f_i}(v)+\delta_{f_{i+1}}(v)$ for $i\in\{1,2,\ldots,d(v)\}$ if $v\in U$, or for $i\in\{1,2,\ldots,\linebreak[1]d(v)-1\}$ if $v\in T$, then Claim~\ref{claim:avg} yields that \begin{equation}\label{eq:sum2} \textstyle\delta(v)= \tfrac12\sum_i\bigl(\delta_{f_i}(v)+\delta_{f_{i+1}}(v)\bigr)\ge \left\{\begin{array}{l@{}l@{}} -\tfrac52d(v)+\tfrac52,&\quad\text{if $v\in T$};\\[1mm] -\tfrac52d(v),&\quad\text{if $v\in U$}.\end{array}\right. \end{equation} First suppose that $d(v)=\Delta-1$. Then $\alpha(v)=3\Delta-9$ and $\beta(v)=0$. If $v\in T$, then $\gamma(v)=-3$ and~\eqref{eq:sum2} gives $\delta(v)\ge-\tfrac52(\Delta-1)+\tfrac52$. Since $\Delta\ge14$, inequality~\eqref{eq:discvertex} follows. If $v\in U$, then $\gamma(v)=0$ and $\delta(v)\ge-\tfrac52(\Delta-1)$. The hypothesis that $\Delta\ge13$ guarantees that~\eqref{eq:discvertex} is valid again. Finally, suppose that $d(v)=\Delta$. Now $\alpha(v)=3\Delta-6$ and $\beta(v)=-2$. If $v\in T$, then $\gamma(v)=-3$ and $\delta(v)\ge-\tfrac52\Delta+\tfrac52$. We see that~\eqref{eq:discvertex} holds, as $\Delta\ge17$. If $v\in U$, then~$\gamma(v)=0$ and $\delta(v)\ge-\tfrac52\Delta$. So~\eqref{eq:discvertex} is verified, provided that $\Delta\ge16$. This confirms~\eqref{eq:discvertex} for all vertices and completes the proof of the theorem.\qquad\hspace*{\fill}$\Box$ \subsection{Proof of Theorem~\ref{thm:planar2}} Recall the notation and terminology introduced in the introduction of this section. The statement of Theorem~\ref{thm:planar2} is true for graphs with maximum degree~$20$ and exactly~$20$ edges. We use induction on $E(G)$, and proceed with the induction step. We may easily assume that~$G$ is connected with at least~$21$ vertices, since $\Delta\ge20$. Let~$M$ be a precoloured \mbox{distance-$3$} matching. We first observe that \begin{equation} \label{eq:sum3} \text{if $uv\in E(G)\setminus M$, then $d(u)+d(v)\ge\Delta+2$.} \end{equation} Indeed, suppose that the inequality does not hold for some $uv\notin M$. Then~$M$ can be extended to a colouring of $G-uv$ so that at most $\Delta-1$ colours are used on the edges adjacent to~$uv$, and so we can easily extend the colouring further to~$uv$. From~\eqref{eq:sum3} it follows that every vertex with degree~$1$ is incident with an edge in~$M$ and that if~$v$ has degree~$2$ and $uv\notin M$, then $d(u)=\Delta$. In particular, if a vertex~$v$ with degree greater than~$1$ has a neighbour in~$T_2$, then $d(v)=\Delta$. Moreover, since edges in~$M$ are at distance at least~$4$ in $G$, a vertex can have at most one neighbour in $V_1\cup T_2$. Let~$V_2'$ be the set of vertices of degree~$2$ that are not incident with an edge of~$M$. For a face~$f$, let $V^-(f)=V(f)\setminus(V_1\cup T_2)$, and let $W^-_f$ be the sequence of vertices on the boundary walk~$W_f$ after removing vertices from $V_1\cup T_2$. For a vertex~$v$, let $v_1,v_2,\ldots,v_{d(v)}$ be the neighbours of~$v$, listed in clockwise order according to the drawing of~$G$. Write~$f_i$ for the face incident with~$v$ lying between the edges $vv_i$ and $vv_{i+1}$ (taking addition modulo~$d(v)$). If a vertex~$v$ has a (unique) neighbour in $V_1\cup T_2$, then we always choose~$v_1$ to be this neighbour. In that case $f_{d(v)}=f_1$, and that face is called~$f_1$ again. Note that it is possible for other faces to be the same as well (if~$v$ is a cut-vertex), but we will not identify those multiple names of the same face. \begin{claim}\label{claim:vdeltav2} \ $|V_{\Delta}|>|V_2'|$. \end{claim} \begin{proof} Consider the set~$F$ of edges in $E(G)\setminus M$ with one end-vertex in~$V_2'$ and the other in~$V_{\Delta}$. The subgraph with vertex set $V_2'\cup V_{\Delta}$ and edge set~$F$ is bipartite; we claim it is acyclic. For suppose there exists an (even) cycle $C\in F$. By induction, we can extend the precolouring of $E(G)\setminus M$ to $G-C$. But then we can further extend this colouring to the edges of~$C$, since each one sees only $\Delta-2$ coloured edges, and even cycles are $2$-edge-choosable. Since each vertex in~$V_2'$ is incident with at least two edges in~$F$, we have $|V_\Delta|+|V_2'|>|F|\ge2|V_2'|$. The claim follows. \end{proof} \noindent We use a discharging argument to complete the proof. First, let us assign to each vertex~$v$ a charge \smallskip \qitem{$\alpha1$:} $\alpha(v)=3d(v)-6$, \smallskip\noindent and to each face~$f$ a charge \smallskip \qitem{$\alpha2$:} $\alpha(f)=-6$. \smallskip For each vertex~$v$ we define $\beta(v)$ as follows. \smallskip \qitem{$\beta1$:} If $v\in V_\Delta$, then $\beta(v)=-2$. \smallskip \qitem{$\beta2$:} If $v\in V_2'$, then $\beta(v)=2$. \smallskip \qitem{$\beta3$:} In all other cases, $\beta(v)=0$. \smallskip For each vertex~$v$ and edge $e=vu$, we define $\gamma_e(v)$ and $\gamma_e(u)$ as follows. \smallskip \qitem{$\gamma1$:} If $v\in V_1$, then $\gamma_e(v)=-\gamma_e(u)=3$. \smallskip \qitem{$\gamma2$:} If $v\in T_2$ and $u\in V_\Delta$, then $\gamma_e(v)=-\gamma_e(u)=3$. \smallskip \qitem{$\gamma3$:} If $v\in U_2\setminus V_2'$ and $u\in V_\Delta$, then $\gamma_e(v)=-\gamma_e(u)=2$. \smallskip \qitem{$\gamma4$:} In all other cases, $\gamma_e(v)=0$. \smallskip Finally, for each face~$f$ and vertex $v\in W^-_f$ we define $\delta_{f}(v)$ and $\delta_v(f)$ as follows. \smallskip \qitem{$\delta1$:} If $v\in U_2$, then $\delta_v(f)=-\delta_f(v)=1$. \smallskip \qitem{$\delta2$:} If $v\in T$ and $3\le d(v)\le\Delta-4$, then $\delta_v(f)=-\delta_f(v)=3-\dfrac{6}{d(v)-1}$. \smallskip \qitem{$\delta3$:} If $v\in U$ and $3\le d(v)\le\Delta-4$, then $\delta_v(f)=-\delta_f(v)=3-\dfrac{6}{d(v)}$. \smallskip \qitem{$\delta4$:} If $d(v)\ge\Delta-3$, $|V^-(f)|=3$, and both neighbours of~$v$ in $V^-(f)$ are joined by an edge in~$M$, then $\delta_v(f)=-\delta_f(v)=4$. \smallskip \qitem{$\delta5$:} If $d(v)\ge\Delta-3$ and~$v$ has a neighbour in $V^-(f)\cap T_3$, then $\delta_v(f)=-\delta_f(v)=3$. \smallskip \qitem{$\delta6$:} If $d(v)\ge\Delta-3$ and none of~$\delta4$ and~$\delta5$ applies, then $\delta_v(f)=-\delta_f(v)=\tfrac52$. \smallskip For a vertex~$v$ and face~$f$, let $\gamma(v)$, $\delta(v)$ and~$\delta(f)$ be defined as in the proof of Theorem~\ref{thm:planar}. By definition we have $\sum_v\gamma(v)+\sum_v\delta(v)+\sum_{f}\delta(f)=0$. It follows from Claim~\ref{claim:vdeltav2} that $\sum_v\beta(v)<0$. From Euler's Formula we obtain $\sum_{v}\alpha(v)+\sum_{f}\alpha(f)<0$. Thus, in order to reach a contradiction, it is enough to show that for every vertex~$v$: \begin{equation}\label{eq:discvertex3} \alpha(v)+\beta(v)+\delta(v)+\gamma(v)\ge0, \end{equation} and that for every face~$f$: \begin{equation}\label{eq:discregion3} \alpha(f)+\delta(f)\ge0. \end{equation} Let~$f$ be a face. As~$G$ is simple, $|V^-(f)|\ge3$. Since $\alpha(f)=-6$, it follows that~\eqref{eq:discregion3} is verified if we can show that $\delta(f)\ge6$. Let~$v$ be the vertex in $V^-(f)$ for which $\delta_v(f)$ is minimum. If $\delta_v(f)\cdot|V^-(f)|\ge6$, then~\eqref{eq:discregion3} clearly holds. So, by checking $\delta1$\,--\,$\delta6$, we see we only have to consider the case where $v\in T_{[3,6]}\cup U_{[2,5]}$. (Recall that vertices from $V_1\cup T_2$ do not appear in~$W^-_f$.) If $v\in U_2$, then let~$u$ and~$w$ be the neighbours of~$v$. Consider first the case where both~$u$ and~$w$ have degree~$\Delta$. Then they both belong to $V^-(f)$, so~\eqref{eq:discregion3} follows, since $\delta_v(f)=1$ and $\delta_u(f)\ge\tfrac52$, $\delta_w(f)\ge\tfrac52$ by $\delta4$\,--\,$\delta6$. Suppose now that~$u$ has degree less than~$\Delta$, which implies by~\eqref{eq:sum3} that $uv\in M$ and, consequently, $vw\notin M$. In particular, $w\in V^-(f)$ and~$w$ has degree~$\Delta$. Note also that necessarily $u\in V^-(f)$. If $|V^-(f)|=3$, then $\delta_w(f)=4$ by~$\delta4$. As $\delta_u(f)\ge\delta_v(f)=1$, it follows that~\eqref{eq:discregion3} holds. If $|V^-(f)|\ge4$, then~$u$ has a neighbour~$u'$ in $V^-(f)\setminus\{v,w\}$. We assert that $\delta_u(f)+\delta_{u'}(f)\ge\tfrac{5}{2}$. Indeed, this holds if either of~$u$ and~$u'$ has degree at least $\Delta-3$, by $\delta5$ and $\delta6$. Otherwise, $\delta2$ or~$\delta3$ applies to both~$u$ and~$u'$. So $\delta_u(f)+\delta_{u'}(f)\ge6-\dfrac{6}{d(u)-1}-\dfrac{6}{d(u')-1}$. Since $d(u)\ge3$, $d(u')\ge3$ and $d(u')\ge\Delta+2-d(u)$, it is easy to see that the expression $6-\dfrac{6}{d(u)-1}-\dfrac{6}{d(u')-1}$ is minimised if $d(u)=d(u')=\tfrac12(\Delta+2)$. We infer that $\delta_u(f)+\delta_{u'}(f)\ge6-2\cdot\dfrac{6}{\tfrac12\Delta} \ge\tfrac52$, because $\Delta\ge7$. The assertion follows. As a result, we deduce that~\eqref{eq:discregion3} holds, since $\delta_w(f)\ge\tfrac52$ by~$\delta5$\ and~$\delta6$. If $v\in T_3$, then $\delta_v(f)=0$, but~$v$ has two neighbours in $V^-(f)$ that have degree at least $\Delta-1$ each. Equation~\eqref{eq:discregion3} then follows from~$\delta5$. For the remaining cases we always have $\delta_v(f)\ge1$. Rules $\delta2$\,--\,$\delta6$ ensure that any vertex $u\in V^-(f)$ with $d(u)\ge13$ satisfies $\delta_u(f)\ge\tfrac52$; hence there can be at most one such vertex and, in particular, a neighbour~$u$ of~$v$ in $V^-(f)$ must have degree at most~$12$. As~$v$ itself has degree at most~$6$, by~\eqref{eq:sum3} we have $uv\in M$, which also implies that $u,v\in U$. Hence in particular $v\in U_{[3,5]}$. Let~$w$ be the neighbour of~$v$ in $V^-(f)\setminus\{u\}$. Since $v\in U_{[3,5]}$ and $vw\notin M$, it necessarily holds that $d(w)\ge\Delta-3$. If $|V^-(f)|=3$, then~\eqref{eq:discregion3} holds by~$\delta4$ since $\delta_u(f)\ge\delta_v(f)\ge1$. If $|V^-(f)|\ge4$, then~$u$ has a neighbour~$u'$ in $V^-(f)\setminus\{v,w\}$, which has degree at least $\Delta+2-d(u)\ge10$. Consequently, $\delta_{u'}(f)\ge\tfrac52$ by $\delta3$, $\delta5$ or~$\delta6$. We deduce that~\eqref{eq:discregion3} holds, as $\delta_w(f)\ge\tfrac{5}{2}$ by~$\delta5$ or~$\delta6$. This confirms~\eqref{eq:discregion3} for all faces. \medskip Now let~$v$ be a vertex. Recall that $\alpha(v)=3d(v)-6$. Furthermore, if~$v$ has a neighbour in $V_1\cup T_2$, then the two consecutive faces incident with that neighbour are counted as one face; all other faces are counted separately. Finally, as noted earlier, a vertex can have at most one neighbour in $V_1\cup T_2$ If $d(v)=1$, then $\alpha(v)=-3$ and $\gamma(v)=3$. Since $\beta(v)=\delta(v)=0$, we immediately obtain~\eqref{eq:discvertex3}. If $d(v)=2$, then $\alpha(v)=0$. If $v\in T_2$, then both~$\gamma1$ and~$\gamma2$ apply; hence $\gamma(v)=0$. Again one can check that $\beta(v)=\delta(v)=0$, confirming~\eqref{eq:discvertex3}. Otherwise $v\in U_2$, and~$\delta1$ implies that $\delta(v)\ge-2$, as~$v$ is incident with at most two faces. If $v\in V_2'$ as well, then~$\beta2$ yields that $\beta(v)=2$ and $\gamma(v)=0$. If $v\notin V_2'$, then $\gamma(v)=2$ while $\beta(v)=0$. In either case~\eqref{eq:discvertex3} follows. Next suppose that $3\le d(v)\le\Delta-4$. Observe that $\beta(v)=0$. If $v\in T$, then $\gamma(v)=-3$ by~$\gamma1$. Since~$v$ has a neighbour with degree one, we know that~$v$ is incident with $d(v)-1$ regions, and so~$\delta2$ yields that $\delta(v)=(d(v)-1)\cdot\bigl({-3}-\dfrac{6}{d(v)-1}\bigr)= 9-3d(v)$. Similarly, if $v\in U$, then $\gamma(v)=0$, and~$\delta4$ yields that $\delta(v)=6-3d(v)$. This proves~\eqref{eq:discvertex3} for those vertices~$v$. Suppose now that $d(v)\in\{\Delta-3,\Delta-2,\Delta-1\}$. Then $\beta(v)=0$. If $v\in T$, then $\gamma(v)=-3$ by~$\gamma1$. Since $M$ is distance-$3$, none of~$\delta4$ and~$\delta5$ applies to~$v$, and~$v$ is incident with $d(v)-1$ faces. From~$\delta6$ we deduce that $\delta(v)=-\tfrac52(d(v)-1)$. Since $d(v)\ge\Delta-3\ge13$, it follows that $3d(v)-6-3-\tfrac52(d(v)-1)=\tfrac12d(v)-\tfrac{13}2\ge0$, and hence~\eqref{eq:discvertex3} is satisfied again. Next assume that $v\in U$, and so $\gamma(v)=0$. The fact that~$M$ is distance-$3$ ensures that~$\delta6$ applies to at least $d(v)-1$ faces incident with~$v$. This implies that $\delta(v)\ge-(4+\tfrac52(d(v)-1))$. Combined with the assumption that $\Delta\ge15$, this is always enough to satisfy~\eqref{eq:discvertex3}. Finally, suppose that $d(v)=\Delta$. In this case $\beta(v)=-2$. If $v\in T$, then the distance condition on~$M$ ensures that $\gamma(v)=-3$ and $\delta(v)=-\tfrac52(\Delta-1)$. Since $\Delta\ge17$ this confirms~\eqref{eq:discvertex3}. So we are left with the case where $v\in U$. Since~$M$ is a distance-$3$ matching, at most one of $\gamma2$, $\gamma3$, $\delta4$ or $\delta5$ applies. Moreover, if~$\gamma2$ does apply, then $\gamma(v)=-3$, the vertex~$v$ is incident with $\Delta-1$ faces and for all those faces~$f$ we have $\delta_f(v)=-\frac52$. If~$\gamma3$ does apply, then $\gamma(v)=-2$, the vertex~$v$ is incident with~$\Delta$ faces and for all those faces~$f$ we have $\delta_f(v)=-\frac52$. If~$\delta4$ or~$\delta5$ does apply, then $\gamma(v)=0$, the vertex~$v$ is incident with~$\Delta$ faces, for all those faces~$f$ but one we have $\delta_f(v)=-\tfrac52$, while for the final face~$f$ we have $\delta_f(v)\in\{-4,-3\}$. Finally, if none of $\gamma2$, $\gamma3$, $\delta4$ or $\delta5$ applies, then $\gamma(v)=0$, the vertex~$v$ is incident with~$\Delta$ faces and for all those faces~$f$ we have $\delta_f(v)=-\tfrac52$. Using that $\Delta\ge20$, we can check that~\eqref{eq:discvertex3} is satisfied in all cases. This confirms~\eqref{eq:discvertex3} for all vertices and completes the proof of the theorem.\qquad\hspace*{\fill}$\Box$ \section{A Final Conjecture} Suppose that we would go to any means to get an extension form of Vizing's theorem, say, by weakening the precolouring condition. We still let $\mathcal{K}=[K]$ be a palette of available colours. Given a subset $S\subseteq E$ of edges and an arbitrary (i.e., not necessarily proper) colouring of elements of~$S$ using only colours from~$\mathcal{K}$, is there a proper colouring of all edges of~$G$ (using colours from~$\mathcal{K}$) that \emph{disagrees} with the given colouring on every edge of~$S$? We may consider the coloured set~$S$ as a set of \emph{forbidden} (coloured) edges, while the full colouring, if it can be produced, is called an \emph{avoidance} of the forbidden edges. The following is a significant weakening of Conjecture~\ref{conj:main}, and is not implied by the LCC, nevertheless its proof eludes us. \begin{conjecture}\label{conj:weak}\mbox{}\\* Let~$G$ be a multigraph with maximum degree $\Delta(G)$ and maximum multiplicity $\mu(G)$. Using the palette $\mathcal{K}=[\Delta(G)+\mu(G)]$, any forbidden matching can be avoided by a proper edge-colouring of all of~$G$. \end{conjecture}

https://arxiv.org/abs/2207.09323

Thin polytopes: Lattice polytopes with vanishing local $h^*$-polynomial

In this paper we study the novel notion of thin polytopes: lattice polytopes whose local $h^*$-polynomials vanish. The local $h^*$-polynomial is an important invariant in modern Ehrhart theory. Its definition goes back to Stanley with fundamental results achieved by Karu, Borisov & Mavlyutov, Schepers, and Katz & Stapledon. The study of thin simplices was originally proposed by Gelfand, Kapranov and Zelevinsky, where in this case the local $h^*$-polynomial simply equals its so-called box polynomial. Our main results are the complete classification of thin polytopes up to dimension 3 and the characterization of thinness for Gorenstein polytopes. The paper also includes an introduction to the local $h^*$-polynomial with a survey of previous results.

\section{Introduction} In this paper we propose to investigate {\em thin polytopes}: lattice polytopes with vanishing local $h^*$-polynomials. Local $h^*$-polynomials are also called $l^*$-polynomials or $\tilde{S}$-polynomials. In the case of lattice simplices, they equal the so-called box polynomial, see Example~\ref{ex:simplex}. Thin simplices were first defined in the context of regular $A$-determinants and $A$-discri\-minants by Gelfand, Kapranov and Zelevinsky \cite[11.4.B]{GKZ94} as those lattice simplices whose Newton numbers are zero, see Remark~\ref{rem:gkz}. As has been noted in \cite{GKZ94}, ``a classification of thin lattice simplices seems to be an interesting problem in the geometry of numbers.'' In this paper, we extend this endeavor to thin lattice polytopes, which we throughout refer to for simplicity as thin polytopes. Our main results are a complete classification of thin polytopes up to dimension $3$ (Theorem~\ref{thm:3d}) and a characterization of thin Gorenstein polytopes in any dimension (Theorem~\ref{thm:main}). The latter relies crucially on a recent non-negativity result by Katz and Stapledon \cite[Theorem~6.1]{Katz2016Local}. As a consequence, we answer some questions posed by Borisov, Schepers and the last named author. \smallskip We also hope that this paper is a step in bringing renewed interest to the study of the local $h^*$-polynomial as a fundamental invariant of a lattice polytope with manifold fruitful connections as pioneered in the work of Stanley \cite{Stanley1992Subdivisions}, Karu \cite{Karu08}, Batyrev, Borisov, Mavlyutov \cite{batyrev2008combinatorial, BorisovMavlyutov}, Schepers \cite{Schepers2012Stringy, NS13}, and Katz, Stapledon \cite{Katz2016Local}. \smallskip Let us give an overview of this paper. In Section~\ref{sec:local} we give a comprehensive survey on the local $h^*$-polynomial of a lattice polytope. In Section~\ref{sec:thin} we define thin polytopes, present the main examples and discuss several open questions (e.g., Question~\ref{question}). Section~\ref{sec:3d} contains the complete classification of three-dimensional thin polytopes. In particular, we prove that three-dimensional lattice simplices are thin if and only if they are lattice pyramids (Corollary~\ref{cor:tetra}). Section~\ref{sec:gorst} presents the characterization of thin Gorenstein polytopes (Theorem~\ref{thm:main}). In particular, we deduce that thin Gorenstein polytopes have lattice width $1$ (Corollary~\ref{cor:flat-gorst}), being thin is invariant under the duality of Gorenstein polytopes (Corollary~\ref{cor:dual}) and show that Gorenstein simplices are thin if and only if they are lattice pyramids (Corollary~\ref{cor:gorsimplex}), thereby answering the original question of \cite{GKZ94} in the Gorenstein case. For these results, we study in Section~\ref{sec:join} the behavior of local $h^*$-polynomials under joins, particularly for Gorenstein polytopes. \section*{Acknowledgment} Our deepest thanks go to Jan Schepers whose notes and insights during and after the collaboration with the third author on \cite{NS13} contained in particular Lemma~\ref{lem:scheperscons} and were the basis of the proof of Theorem~\ref{thm:main}. We thank Lev Borisov for his comments, questions and his kindness to share the proof of Proposition~\ref{prop:borisov}. We are also grateful for Sam Payne and Liam Solus for their interest. All the authors have been supported by the Research Training Group Mathematical Complexity Reduction funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 314838170, GRK 2297 MathCoRe. \section{A primer on the local $h^*$-polynomial of a lattice polytope} \label{sec:local} As the local $h^*$-polynomial is still not as well known in Ehrhart theory as the $h^*$-polynomial and also has been studied with different names and notations, we will give a slightly more thorough account on previous research than strictly necessary for the mere purpose of the results of this paper. \subsection{Toric $g$- and $h$-polynomials of lower Eulerian posets} In \cite{Stanley1987Generalized}, Stanley generalized the notion of $h$-vectors of simplicial complexes and simplicial polytopes significantly. For this, let us recall some basic terminology. \begin{defn} The dual of a finite poset $\mathcal{P}$ is denoted $\mathcal{P}^\ast$. A finite poset $\mathcal{P}$ is \emph{locally graded} if every inclusion-maximal chain in every interval $[x,y]$ has the same length $r(x,y)$. The \emph{rank} $\rk(\mathcal{P})$ is the length of the longest chain in $\mathcal{P}$. If in addition there exists a rank function $\rho: \mathcal{P} \rightarrow \mathbb{Z}$, i.e., $r(x,y) = \rho(y) - \rho(x)$ for every interval $[x,y]$, then $\mathcal{P}$ is called \emph{ranked}. If $\mathcal{P}$ is ranked and every interval $[x,y]$ with $x \neq y$ has the same number of even rank and odd rank elements, then $\mathcal{P}$ is \emph{locally Eulerian}. If $\mathcal{P}$ is locally Eulerian and contains a minimal element $\hat{0}$, then it is called \emph{lower Eulerian}. If it also contains a maximal element $\hat{1}$, then $\mathcal{P}$ is called \emph{Eulerian}. In presence of a minimal element $\hat{0}$ of a ranked poset $\mathcal{P}$, we will always assume that the rank function satisfies $\rho(\hat{0}) = 0$. \end{defn} Here is the definition of the $g$-polynomial and the $h$-polynomial for lower Eulerian posets according to Stanley \cite{Stanley1987Generalized}. \begin{defn} Let $\mathcal{P}$ be a lower Eulerian poset with rank function $\rho$ and rank $d$. We define the \emph{$g$-polynomial $g_\mathcal{P}(t)$} and the \emph{$h$-polynomial $h_\mathcal{P}(t)$} recursively by introducing a third polynomial $f_\mathcal{P}(t)$ as an intermediate step. Let \begin{gather*} f_\emptyset(t) = g_\emptyset(t) = h_\emptyset(t) = 1 \end{gather*} and if $\mathcal{P}\not=\emptyset$, we set \begin{gather*} f_\mathcal{P}(t) = \sum_{x \in \mathcal{P}} (t-1)^{d - \rho(x)} g_{[\hat{0},x)}(t) \end{gather*} and define for $f_{\mathcal{P}}(t)=\sum_{i=0}^df_i t^i$, \begin{gather*} g_\mathcal{P}(t) = \sum_{i=0}^{\lfloor d/2 \rfloor} (f_i-f_{i-1}) t^i, \text{ and }\\ h_\mathcal{P}(t) = \sum_{i=0}^d f_{d-i} t^i. \end{gather*} \label{def:fgh} \end{defn} Hence, $h_\mathcal{P}(t)$ is a polynomial with constant term $1$ of degree $\leq d$, and $g_\mathcal{P}(t)$ is a polynomial with degree $\le d/2$.\\ \begin{rmk} If for $x \in \mathcal{P}$, the interval $[\hat{0},x]$ is boolean, then $g_{[\hat{0},x)}(t) = 1$, see \cite[Prop.~2.1]{Stanley1987Generalized}.\label{simplicial} \end{rmk} \begin{rmk} Let us recall the situation of simplicial complexes $\Delta$ (see \cite{Stanley1987Generalized}). We identify $\Delta$ with its face poset which is a lower Eulerian poset with minimal element $\emptyset \in \Delta$. Throughout, we use the convention that $\dim(\emptyset) = -1$. If $\Delta$ has dimension $d-1$, then $f_\Delta(t) = \sum_{i=0}^d f_{i-1} (t-1)^{d-i}$, where $f_j$ denotes the number of faces of $\Delta$ of dimension $j$. It follows that \[h_\Delta(t) = \sum_{\sigma \in \Delta} t^{\dim(\sigma)+1} (1-t)^{d-1-\dim(\sigma)}\] equals the usual $h$-polynomial of $\Delta$, and its coefficients are the usual $h$-vector of $\Delta$ \cite[p.~199]{Stanley1987Generalized}. \end{rmk} Stanley proved in \cite[Theorem~2.4]{Stanley1987Generalized} the following combinatorial palindromicity result generalizing the Dehn-Sommerville equations for face numbers of simplicial polytopes. \begin{thm} Let $\hat{\mathcal{P}}$ be an Eulerian poset and $\mathcal{P} \coloneqq \hat{\mathcal{P}} \setminus \hat{1}$ with $\rk(\mathcal{P}) = d$. Then the $h$-polynomial $h_\mathcal{P}(t) = \sum_{i=0}^d h_i t^i$ is symmetric of degree $d$, i.e. $h_i = h_{d-i}$ for all $i=0, \ldots, d$.\label{thm:pali} \end{thm} In particular, we have $f_{\mathcal{P}}(t) = h_{\mathcal{P}}(t)$ in this case. \begin{rmk} We emphasize that in the situation of Theorem~\ref{thm:pali} it is important to distinguish between the $g$- and $h$-polynomials of $\mathcal{P}$ and $\hat{\mathcal{P}}$. Indeed, $g_{\hat{\mathcal{P}}}(t) = 0$ and $h_{\hat{\mathcal{P}}}(t) = g_\mathcal{P}(t)$. Unfortunately, in this regard the different notations employed in the literature can be confusing. Our notation follows that of Stanley while Katz and Stapledon in \cite{Katz2016Local} write $g(\hat{\mathcal{P}};t)$ for our $g_\mathcal{P}(t)$ but also use $h(\mathcal{P};t)$ for our $h_\mathcal{P}(t)$. In Borisov and Mavylutov \cite{BorisovMavlyutov}, as well as in \cite{batyrev2008combinatorial, NS13}, our $g_\mathcal{P}(t)$ and $h_\mathcal{P}(t)$ would be $g_{\hat{\mathcal{P}}^\ast}(t)$ and $h_{\hat{\mathcal{P}}^\ast}(t)$. \label{rmk:ambiguous} \end{rmk} Let us give the definition of $h$- and $g$-polynomials of polytopes. \begin{defn} For $P$ a polytope we define its {\em (toric) $h$-polynomial} $h_P(t)$, and its {\em (toric) $g$-polynomial} $g_P(t)$ as the $h$-, resp., $g$-polynomial of the face lattice $[\emptyset, P)$ of proper faces of $P$. Note that $g_P(t) = h_{[\emptyset,P]}(t)$, see Remark~\ref{rmk:ambiguous}. \end{defn} Note that by Remark~\ref{simplicial}, we have $g_P(t)=1$ if $P$ is a simplex. \begin{thm}\label{thm:h-pol} Let $P$ be a polytope of dimension $d$. \label{item:h_pol_unimodal} Then $h_P(t)=\sum_{i=0}^d h_i t^i$ is a palindromic polynomial with positive integer coefficients that form a unimodal sequence, i.e., \begin{equation*} 1 = h_0 \leq h_1 \leq \cdots \leq h_{\lfloor \frac{d}{2} \rfloor}. \end{equation*} Equivalently, $g_P(t)$ has non-negative coefficients. \end{thm} \begin{proof} Palindromicity follows from Theorem~\ref{thm:pali}. For rational polytopes $P$, nonnegativity follows from the interpretation of the coefficients of $h_P(t)$ as the dimensions of the even intersection cohomology groups of the toric variety associated with $P$ and the unimodality property follows from the hard Lefschetz theorem \cite[Theorem~3.1, Corollary~3.2]{Stanley1987Generalized}. The non-rational case has been treated by Karu in \cite{Karu04Hard}. \end{proof} Let us mention the following less well-known duality property of $g$-polynomials that will be of importance in Section~\ref{sec:gorst}. This is a result by Kalai, published in \cite[Theorem~4.5]{BradenKalai} as a consequence of the main result in that paper by Braden. Here, $\mathcal{P}^*$ denotes the {\em dual} poset of a poset $\mathcal{P}$. \begin{thm} \label{thm:Kalai} Let $P$ be a polytope. Then \begin{equation*} \deg(g_{[\emptyset, P)}) = \deg(g_{(\emptyset,P]^\ast}). \end{equation*} In other words, if $Q$ is any polytope which is combinatorially dual to $P$, then $\deg(g_P) = \deg(g_Q)$. \end{thm} \subsection{(Relative) local $h$-polynomials of triangulations} We give the definition of the local $h$-polynomial (and its generalized relative version) of a triangulation $\Delta$ of a polytope $P$, following \cite{Stanley1992Subdivisions} and \cite{Katz2016Local} (the relative version was introduced independently in \cite{Athanasiadis} and \cite{Report}). Here, we define the {\em link} of a face $\sigma \in \Delta$ as $\link(\Delta,\sigma) := \{\rho \in \Delta \,:\, \rho \cap \sigma = \emptyset, \,\rho \cup \sigma \in \Delta\}$. \begin{defn} Let $P$ be a polytope, $\Delta$ a triangulation of $P$, and $\sigma \in \Delta$. The \emph{relative local $h$-polynomial} of $\Delta$ with respect to $\sigma$ is defined as \begin{equation*} l_{\Delta,\sigma}(t) \coloneqq \sum_{\sigma \subseteq F \leq P} (-1)^{\dim(P) - \dim(F)} h_{\link(\Delta_F,\sigma)}(t) \, g_{(F,P]^\ast}(t), \end{equation*} where $F \le P$ means that $F$ is a face of $P$ (including $\emptyset$ and $P$) and $\Delta_F := \{\rho \in \Delta \colon \rho \subseteq F\}$. We call $l_\Delta (t) := l_{\Delta,\emptyset}$ the \emph{local $h$-polynomial} of $\Delta$. \label{def:local-h} \end{defn} We suppress $P$ in this notation as it equals $|\Delta| = \bigcup_{\sigma \in \Delta} \sigma$, the support of $\Delta$. We remark that the same definition of the local $h$-polynomial can be given for an arbitrary \emph{polyhedral} subdivision of $P$, and this can be extended to so called strong formal subdivisions of Eulerian posets, see \cite[Definition~4.1]{Katz2016Local}. \begin{thm}\label{thm:local_h} Let $P$ be a polytope of dimension $d$, $\Delta$ a triangulation of $P$, and $\sigma \in \Delta$. Write $l_{\Delta,\sigma}(t) = \sum_{i=0}^d l_i t^i$. \begin{enumerate} \item $l_{\Delta,\sigma}(t)$ has nonnegative integer coefficients. \item \label{item:local_h_symmetric} $l_{\Delta,\sigma}(t)$ is palindromic, i.e. $l_i = l_{d-\dim(\sigma)-i}$ for $i=0, \ldots, d-\dim(\sigma)$. \item \label{item:local_h_unimodal} If $\Delta$ is regular, then the coefficients of $l_{\Delta,\sigma}(t)$ form a unimodal sequence. \end{enumerate} \end{thm} \begin{proof} (2): For the local $h$-polynomial this is a special case of \cite[Corollary~7.7]{Stanley1992Subdivisions}, for the relative local $h$-polynomial see \cite[Corollary~4.5]{Katz2016Local}. (1) and (3): For $\Delta$ a rational triangulation this has been proven in \cite[Theorem~7.9]{Stanley1992Subdivisions}, respectively, \cite[Theorem~6.1]{Katz2016Local} using the decomposition theorem (cf. \cite{Beilinson, Decomposition, Propermaps}). As pointed out in \cite[Remark~6.6]{Katz2016Local}, the only missing ingredient to drop the rationality hypothesis was the relative hard Lefschetz theorem for the intersection cohomology of fans which was subsequently proven in \cite{Karu2019Relative}. \end{proof} The following decomposition theorem was one of the main motivations of Stanley for the notion of {\em local} $h$-vectors. This is proven in \cite[Theorem~7.8]{Stanley1992Subdivisions}, and the general version in \cite{Katz2016Local} (see, e.g., second equation in proof of Lemma~6.4). To stress the analogy to Theorem~\ref{thm:bm}, we state the equality also using the $h$-polynomial. \begin{prop} Let $P$ be a polytope of dimension $d$, $\Delta$ a triangulation of $P$, and $\sigma \in \Delta$. Then \[h_{\link(\Delta,\sigma)}(t) = \sum_{\sigma \subseteq F \le P} l_{\Delta_F,\sigma}(t) g_{[F,P)}(t) = \sum_{\sigma \subseteq F \le P} l_{\Delta_F,\sigma}(t) h_{[F,P]}(t).\] In particular for $\sigma=\emptyset$, we get $l_\Delta(t) \le h_\Delta(t)$ and $g_P(t)= h_{[\emptyset,P]} \le h_\Delta(t)$ coefficientwise. \label{prop:h-decomp} \end{prop} As a consequence of above nonnegativity results, Stanley and later Katz and Stapledon show that $h$-polynomials as well as relative local $h$-polynomials are coefficientwise monotone under subdivision refinement \cite[Corollary~6.10]{Katz2016Local}. \subsection{The $h^*$-polynomial of a lattice polytope} We quickly recall the basic notions of Ehrhart theory. Let $P \subseteq \mathbb{R}^d$ be a lattice polytope with respect to some lattice $M \subseteq \mathbb{R}^d$ of maximal rank $d$, usually $M=\mathbb{Z}^d$. The \emph{Ehrhart series of $P$} (with respect to $M$) is the formal power series \begin{equation*} \mathrm{Ehr}_P(t) \coloneqq 1 + \sum_{n \geq 1} |(nP) \cap M| t^n \in \mathbb{Z}[[t]]. \end{equation*} By a theorem of Ehrhart \cite{Ehr62}, the map $\mathrm{ehr}_P \colon \mathbb{Z}_{\geq 1} \rightarrow \mathbb{Z}, \ n \mapsto |(nP) \cap M|$ is a polynomial in $n$, called the \emph{Ehrhart polynomial of $P$}. It has degree $\dim(P)$, constant term $1$ and leading coefficient equal to the volume of $P$ normalized with respect to $M$. It follows that \begin{equation*} \mathrm{Ehr}_P(t) = \frac{h^\ast_P(t)}{(1-t)^{\dim(P) + 1}}, \end{equation*} for a unique polynomial $h^\ast_P(t) \in \mathbb{Z}[t]$ of degree at most $\dim(P)$, called the \emph{$h^\ast$-polynomial of $P$}. Here, $h^\ast_P(t)$ has non-negative integer coefficients by \cite{Sta80}. Moreover, $h^\ast_P(0) = 1$ and $h^\ast_P(1)$ equals the {\em lattice volume} ${\rm vol}_\mathbb{Z}(P) \in \mathbb{Z}_{\ge 1}$, which is defined as $d!$ times the volume of $P$ normalized with respect to $M$. Note that the lattice volume of a lattice point equals~$1$. The \emph{degree of $P$}, denoted $\deg(P)$, is the degree of its $h^\ast$-polynomial $h^\ast_P(t)$. The \emph{codegree of $P$}, denoted $\codeg(P)$, is the smallest integer $k \geq 1$ such that the $k$-th dilate $kP$ of $P$ contains a lattice point of $M$ in its relative interior. By convention, a point has codegree~$1$. It follows from Ehrhart-MacDonald reciprocity \cite{Macdonald} that \begin{equation*} \deg(P) + \codeg(P) = \dim(P) + 1. \end{equation*} Let us recall that two lattice polytopes (with respect to the lattice $M$) are called {\em isomorphic} (or {\em unimodularly equivalent}) if there is an affine-linear automorphism of $\mathbb{R}^d$ that maps $M$ bijectively onto itself. We say $P$ is a {\em unimodular simplex} if $P$ is isomorphic to the convex hull of an affine lattice basis of $M$. Now, $P$ is a unimodular simplex if and only if $h^*_P(t)=1$, or equivalently, $\deg(P)=0$. For $M=\mathbb{Z}^d$ let us also define the {\em standard unimodular simplex} $\Delta_d := \conv(0,e_1, \ldots, e_d)$ for the standard lattice basis $e_1, \ldots, e_d$. \subsection{The local $h^*$-polynomial of a lattice polytope} Let us introduce the main player of this paper (see \cite[Example~7.13]{Stanley1992Subdivisions} and \cite[Def.~7.2]{Katz2016Local}). \begin{defn} Let $P$ be a lattice polytope. The \emph{local $h^\ast$-polynomial} or \emph{$l^\ast$-polynomial} of $P$ is defined as \begin{equation*} l^\ast_P(t) \coloneqq \sum_{\emptyset \leq F \leq P} (-1)^{\dim(P)-\dim(F)} h^\ast_F(t) g_{(F,P]^\ast}(t). \end{equation*} \label{def:lstar} \end{defn} Let us note that the local $h^*$-polynomial of the empty face equals $1$, while for a point it equals $0$. \begin{rmk} The local $h^*$-polynomial has been studied by Borisov, Batyrev, Mavlyutov, Schepers, and the third author under the name {\em $\tilde{S}$-polynomial}, see \cite[Definition~5.3]{BorisovMavlyutov}. It was used by Borisov and Mavlyutov to simplify the formulas for the stringy $E$-polynomial of Calabi-Yau complete intersections in Gorenstein toric Fano varieties originally described via so-called $B$-polynomials \cite{BB96a}. We remark that the reader should be aware that in these papers in the definition of $h$- and $g$-polynomials the dual poset was used compared to the one given here. \end{rmk} \begin{ex} For lattice simplices $P$ (of dimension $d>0$) the $h^*$- and $l^*$-polynomial can be easily understood, as in this case the face posets are all Boolean. Let $\Pi$ denote the half-open parallelepiped spanned by the vertices of $P \times \{1\}$. Then $h^*_P(t)$ (resp., $l^*_P(t)$) enumerates the number of lattice points in $\Pi$ (resp., in the interior of $\Pi$). More precisely, we have $h^*_P(t)=\sum_{i=0}^{d+1} h^*_i t^i$ and $l^*_P(t)=\sum_{i=0}^{d+1} l^*_i t^i$, where for $i=0, \ldots, d+1$ the coefficient $h^*_i$ (resp. $l^*_i$) equals the number of lattice points in $\Pi$ (resp., in the interior of $\Pi$) with last coordinate $i$. We refer to \cite[Prop. 4.6]{batyrev2008combinatorial}. This polynomial $l^*_P(t)$ of a lattice simplex $P$ has sometimes also been called {\em box polynomial}, cf. \cite{Liam}. For instance, we have $h^*_P(t)=1$ if and only if $P$ is a unimodular simplex; in this case, $l^*_P(t)=0$. Let us note that for $h^*$-polynomials this combinatorial interpretation of its coefficients can also be extended to arbitrary lattice polytopes, e.g., by half-open decompositions \cite{Halfopen}. On the other hand, there is not yet a combinatorial counting interpretation for the coefficients of the local $h^*$-polynomial of lattice polytopes known. \label{ex:simplex} \end{ex} Let us summarize some of the basic properties of the local $h^*$-polynomial. Throughout, we use the convention that the degree of the zero-polynomial is zero. \begin{thm} Let $P$ be a lattice polytope of dimension $d > 0$ and local $h^*$-polynomial $l^*_P(t)$. Then the following holds: \begin{enumerate} \item $l^*_P(t)$ has nonnegative integer coefficients. \item $l^*_P(t)= \sum_{i=0}^{d+1} l^*_i t^i$ is palindromic: we have $l^*_i = l^*_{d+1-i}$ for $i=0, \ldots, d+1$. \item The degree of $l^*_P(t)$ equals at most the degree of $h^*_P(t)$, and its subdegree (i.e., the smallest $i$ such that the $i$th coefficient of $l^*(t)$ is nonzero) is at least the codegree of $P$. \item We have $l^*_0 = l^*_{d+1}=0$, and $l^*_1 = l^*_d$ equals the number of lattice points in the interior of $P$. \end{enumerate} \label{thm:lstar} \end{thm} \begin{proof} Let us give the corresponding references: (1) This was conjectured by Stanley \cite[Conj.7.14]{Stanley1992Subdivisions} and proven by Karu \cite{Karu08}. Using the $\tilde{S}$-notation for $l^*$ it also follows from its interpretation as the Hilbert function of a graded vector space by Borisov, Mavlyutov \cite[Prop.~5.5]{BorisovMavlyutov}. (2) This was observed in \cite[Remark 5.4]{BorisovMavlyutov}. (3) This follows directly, see also \cite[Cor.2.16(2)]{NS13}. (4) For this observation, see \cite[Example~4.7]{batyrev2008combinatorial}. \end{proof} In particular, the number $\intr_\mathbb{Z}(P)$ of interior lattice points completely determines the local $h^*$-polynomial up to dimension $2$. If $d=0$, then $l^*_P(t)=0$; if $d=1$, then $l^*_P(t) = \intr_\mathbb{Z}(P) t$; and if $d=2$, then $l^*_P(t) = \intr_\mathbb{Z}(P) t + \intr_\mathbb{Z}(P) t^2$. \subsection{Relating the $h,l,h^*,l^*$-polynomials in a nonnegative way} The following classical result by Betke and McMullen explains the relation of $h^*$-polynomials to $h$-polynomials of a {\em lattice} triangulation (i.e., a triangulation whose vertices are lattice points). Recall that a lattice triangulation is called {\em unimodular} if all its simplices are unimodular simplices. \begin{thm}[\cite{BetkeMcMullen85}] Let $P$ be a lattice polytope with a lattice triangulation $\Delta$. Then the following holds: \[h^*_P(t) = \sum_{\sigma \in \Delta}\, l^*_\sigma(t)\, h_{\link(\Delta,\sigma)}(t).\] In particular, we have $h_\Delta(t) \le h^*_P(t)$ coefficientwise, with equality if and only if $\Delta$ is a unimodular triangulation. \label{thm:bm} \end{thm} We recall that the consequence follows from two facts: First, the combinatorial description of the $h^*$- and $l^*$-polynomial of a lattice simplex, Example~\ref{ex:simplex}, implies that a lattice simplex $S$ is a unimodular simplex if and only if $l^*_\sigma(t) = 0$ for all non-empty faces $\sigma$ of $S$. Second, the links in the Betke-McMullen formula are all triangulations of balls or spheres, for which the $h$-polynomial is at least $1$ coefficientwise \cite{BetkeMcMullen85, Stanley75}. Let us note that this result was one motivation for Stanley to define local $h$-polynomials, as these allowed him to prove an analogous result in the combinatorial setting, namely, Proposition~\ref{prop:h-decomp} above. And similar to that formula positively expressing the $h$-polynomial of a triangulation into local $h$-polynomials and toric $h$-polynomials, one can also decompose the $h^*$-polynomial of a lattice polytope positively into local $h^*$-polynomials and toric $h$-polynomials. \begin{prop} Let $P$ be a lattice polytope. Then \[h^*_P(t) = \sum_{\emptyset \leq F \leq P} l^\ast_F(t) g_{[F,P)}(t)=\sum_{\emptyset \leq F \leq P} l^\ast_F(t) h_{[F,P]}(t).\] In particular, $l^*_P(t) + g_P(t) = l^*_P(t) + h_{[\emptyset,P]}(t) \le h^*_P(t)$ coefficientwise. \label{prop:schepers} \end{prop} This follows by inverting the alternating expression of Definition~\ref{def:lstar} in the sense of \cite{Stanley1992Subdivisions}. A proof in more generality is given in \cite[Prop.~2.9]{Schepers2012Stringy}, see also \cite[Cor.~1.1]{Karu08} and \cite[Prop.~2.5]{NS13}. We remark that Proposition~\ref{prop:schepers} gives significance to thinking of the local $h^*$-polynomial as the "Ehrhart core" of the $h^*$-polynomial. This is most prominently clear in the case of lattice simplices, see Example~\ref{ex:simplex}. Now, just as the Betke-McMullen formula transparently separates the lattice data (the $l^*$-polynomials of the cells) and combinatorial data (the $h$-polynomials of the links of the cells) of the $h^*$-polynomial of the support of a lattice triangulation, the same can be done for the local $h^*$-polynomial. This was observed in \cite{Report}, see also \cite[Lemma~7.12(4)]{Katz2016Local}. \begin{prop} Let $P$ be a lattice polytope with a lattice triangulation $\Delta$. Then the following holds: \[l^*_P(t) = \sum_{\sigma \in \Delta} l^*_\sigma(t)\; l_{\Delta,\sigma}(t).\] In particular, $l_\Delta(t) \le l^*_ P(t)$ coefficientwise, with equality if $\Delta$ is a unimodular triangulation. \label{prop:decomp} \end{prop} Here, we used the nonnegativity of the relative local $h$-polynomial, Theorem~\ref{thm:local_h}(1), for the consequence. In particular, we get another proof of the nonnegativity of the local $h^*$-polynomial. Moreover, as already observed in \cite{Report}, we have the following {\em intrinsic} obstruction for a lattice polytope to have a unimodular triangulation (apply Theorem~\ref{thm:local_h}(3) with $\sigma=\emptyset$). \begin{cor} If $P$ admits a regular, unimodular triangulation, then its local $h^*$-polynomial has unimodal coefficients. \end{cor} For the purpose of this paper, the following innocent looking consequence of the nonnegativity of relative local $h$-polynomials is crucial. \begin{cor} Let $P$ and $P'$ be lattice polytopes such that $P'$ is obtained from $P$ by refining the lattice. Then $h^*_P(t)\le h^*_{P'}(t)$ and $l^*_P(t) \le l^*_{P'}(t)$ coefficientwise. \label{cor:lattice} \end{cor} \begin{proof} This follows from Theorem~\ref{thm:bm}, respectively, Proposition~\ref{prop:decomp}, the explicit combinatorial description of the $l^*$-polynomial of a lattice simplex, see Example~\ref{ex:simplex}, and the nonnegativity of the $h$-polynomial, respectively, of the relative local $h$-polynomial. \end{proof} We remark that for $h^*$-polynomials this lattice-monotonicity can also easily be seen combinatorially, e.g., using half-open decompositions. However, for local $h^*$-polynomials there seems to be no such combinatorial argument known. Finally, let us mention that in the Katz-Stapledon paper \cite{Katz2016Local}, the $h^*$- and $l^*$-polynomials are further refined leading to the notion of $h^*$- and $l^*$-diamonds. One beautiful consequence is the following lower bound theorem on the coefficients of the $l^*$-polynomial \cite[p.184]{Katz2016Local}. \begin{thm} Let $P$ be a lattice polytope of dimension $d$ with $l^*_P(t)=\sum_{i=1}^d l^*_i t^*$. Then \[l^*_1 \le l^*_i \quad \text{ for } i=2, \ldots, d.\] \label{thm:lower} \end{thm} \section{Definition, basic properties and examples of thin polytopes} \label{sec:thin} \subsection{Main definition and known results from \cite{GKZ94}} The following notion is the main focus of this paper. \begin{defn} A lattice polytope $P$ is called {\em thin} if its local $h^*$-polynomial $l^*_P$ vanishes. By the nonnegativity of the coefficients, Theorem~\ref{thm:lstar}(1), this is equivalent to $l^*_P(1)=0$. \end{defn} Let us note that lattice polytopes of dimension $0$ as well as unimodular simplices are thin, see Example~\ref{ex:simplex}. \begin{rmk} Thin simplices appeared first in \cite[11-4-B]{GKZ94} in the context of regular $A$-determinants and $A$-discriminants, more precisely, in the characterization of so-called $D$-equivalence classes of regular triangulations of $A$. There a lattice simplex $S$ was defined to be {\em thin} if its {\em Newton number} $\nu(S)$ equals zero. Here, the Newton number is defined as follows: \begin{equation} \nu(S) := \sum_{\emptyset \le F \le S} (-1)^{\dim(S)-\dim(F)} {\rm vol}_\mathbb{Z}(F) =0, \label{newton} \end{equation} where, ${\rm vol}_\mathbb{Z}(\emptyset):=1$ (also ${\rm vol}_\mathbb{Z}(F)=1$ if $\dim(F)=0$). Recall from Example~\ref{ex:simplex}, that ${\rm vol}_\mathbb{Z}(F)=h^*_F(1)$ counts the number of lattice points in the half-open parallelepiped over $F$. Hence, by inclusion-exclusion, it is straightforward to deduce $\nu(S)=l^*_S(1)$, the number of interior lattice points in the half-open parallelepiped over $S$. Thus, for lattice simplices the definitions agree. Let us note that in \cite{GKZ94} the nonnegativity of $\nu(S)$ follows from quite deep algebro-geometric arguments, while it is combinatorially obvious from the interpretation of $l^*_S$ as the box polynomial of the lattice simplex $S$.\label{rem:gkz} The reader should also be warned that the expression in equation~\eqref{newton} may be negative for lattice {\em polytopes}. For instance, it equals $-1$ for the 0-1-cube $[0,1]^3$. \end{rmk} Thin simplices were classified in \cite{GKZ94} up to dimension $2$. Here, we can deduce the following statement directly from Theorem~\ref{thm:lstar}(4). Let us define a lattice polytope to be {\em hollow} if it has no lattice points in its interior. Here, a $0$-dimensional lattice polytope is not hollow (but thin). \begin{prop} Thin polytopes of dimension $> 0$ are hollow. The converse also holds in dimensions $1$ and $2$. \label{prop:dim2} \end{prop} In particular, $\Delta_1$ is the only thin polytope of dimension $1$. Hollow polytopes in dimension $2$ are well-known. They are either isomorphic to $2 \Delta_2$ or have lattice width $1$ (i.e., all vertices lie on two parallel hyperplanes of lattice distance one). Note that hollow three-dimensional lattice polytopes do not have to be thin., e.g., $2 \Delta_3$ and $[0,1]^3$ are not thin. One important construction for thin polytopes is to take lattice pyramids. \begin{defn} Let $P \subset \mathbb{R}^d$ be a lattice polytope. Then \[\conv(P \times \{0\}, \{0\} \times \{1\}) \subset \mathbb{R}^d \times \mathbb{R}\] is called the {\em lattice pyramid} over $P$. By convention, a lattice point is also considered a lattice pyramid. \end{defn} It is well-known that the $h^*$-polynomial, and particularly the degree, does not change under taking lattice pyramids. The following result has already been observed in \cite{GKZ94} for lattice simplices and in general in \cite{batyrev2008combinatorial} for lattice polytopes. \begin{prop}Lattice pyramids are thin.\label{prop:pyr} \end{prop} Using this notation we can state the classification of thin simplices up to dimension two as follows. \begin{cor} A lattice simplex of dimension at most $\le 2$ is thin if and only if it is isomorphic to $2 \Delta_2$ or it is a lattice pyramid. \end{cor} \begin{rmk} In \cite{GKZ94} {\em thin triangulations} were intensively studied. Recently, this notion has also been investigated by \cite{Thintriangulations} where it was completely characterized up to dimension $3$. As Stanley observed at the end of Section~7 in \cite{Stanley1992Subdivisions}, a thin triangulation may be defined by the vanishing of its local $h$-polynomial. Now, it follows from Proposition~\ref{prop:decomp} that {\em all} lattice triangulations of thin polytopes are thin. This seems to be a quite strong combinatorial obstruction worth of further study. \end{rmk} \begin{rmk} By Corollary~\ref{cor:lattice}, a thin polytope stays thin if the lattice is coarsened. We do not know of a purely combinatorial proof of this fact.\label{rmk:refine} \end{rmk} \begin{rmk} Let us note that thin simplices turn up in \cite{unimodal} when studying conditions for IDP polytopes to have an unimodal $h^*$-polynomial. In particular, it is asked whether every IDP lattice polytope has a regular triangulation into lattice simplices that have vanishing or monomial $l^*$-polynomial. \end{rmk} \subsection{Two classes of examples of thin polytopes} Let us describe two ways to get thin polytopes in higher dimensions. The first observation is that lattice polytopes of small degree (in other words, "very hollow" lattice polytopes) are always thin. \begin{defn} We say, $P$ is {\em trivially thin} if $\dim(P) \ge 2 \deg(P)$. \end{defn} \begin{prop} Trivially thin polytopes are thin. \end{prop} \begin{proof} A lattice polytope $P$ is trivially thin if and only if $\deg(P) < \codeg(P)$. Now, the statement follows from Theorem~\ref{thm:local_h}(3).\end{proof} Typical examples of trivially thin polytopes are Cayley polytopes with low-dimensional factors (see Remark~\ref{rem:cayley}). \medskip A second way to get high-dimensional thin polytopes is to use free joins. \begin{defn} Let $P \subset \mathbb{R}^n$ and $Q \subset \mathbb{R}^m$ be lattice polytopes. We call \[P \circ_\mathbb{Z} Q := \conv(P \times \{0\} \times \{0\}, \{0\} \times P \times \{1\}) \subset \mathbb{R}^n \times \mathbb{R}^m \times \mathbb{R},\] the {\em free join} of $P$ and $Q$. \end{defn} For instance, the free join of $[0,1]$ with itself is a unimodular $3$-simplex. Note that isomorphic factors lead to isomorphic free joins. From the Ehrhart-theoretic viewpoint the free join construction is important because of the following multiplicativity property, see \cite[Lemma~1.3]{Join} and \cite[Remark~4.6(5)]{NS13}. \begin{prop} Let $P \subset \mathbb{R}^n$ and $Q \subset \mathbb{R}^m$ be lattice polytopes. Then \[h^*_{P \circ_\mathbb{Z} Q}(t) = h^*_P(t) \, h^*_Q(t),\; \text{ and }\; l^*_{P \circ_\mathbb{Z} Q}(t) = l^*_P(t) \, l^*_Q(t).\] \label{prop:mult} \end{prop} \begin{cor} The free sum of two lattice polytopes is thin if and only if at least one of the two factors is thin. \end{cor} As a lattice pyramid is the free join of a point (which is thin) and a lattice polytope, this generalizes Proposition~\ref{prop:pyr}. \subsection{Are there other examples of thin polytopes?} It is not obvious to give examples of thin polytopes (such as Example~\ref{ex:nonsp} below) that do not fall in above described two classes. In order to formulate a natural question in this respect, let us recall two notions. First, a lattice polytope $P$ is called {\em spanning} if every lattice point in its affine hull is an integer affine combination of the lattice points in $P$. Note that every lattice polytope becomes spanning after a possible coarsening of the ambient lattice (we refer to \cite{spanning} for more background and results on spanning lattice polytopes). Second, let us call a lattice polytope $P$ a {\em join} if there are two non-empty faces $F$ and $G$ of $P$ such that the free join of $F$ and $G$ is affinely-isomorphic to $P$. Let us remark that if $P$ is spanning and a join of $F$ and $G$ where every lattice point in $P$ is contained in $F$ or $G$, then it is automatically a free join. \begin{question}\label{question}\ \begin{enumerate} \item Is every thin polytope trivially thin or a join? \item Is every spanning thin polytope trivially thin or a free join? \end{enumerate} \end{question} Both questions are closely related but not directly. The reason is that the degree of the polytope can drop under coarsenings of the lattice, so a non-spanning thin but not trivially thin polytope could be trivially thin with respect to its spanning lattice. We also note that trivially thin polytopes are often not joins. For example the unit square $[0,1]^2$ has degree $1$ and is hence trivially thin, while triangles are the only polygons which are joins. As the following example shows, the spanning hypothesis in the second part of Question~\ref{question} is indeed important. It is one of the apparently rare thin polytopes that are not trivially thin and not a free join. \begin{ex} Consider the $4$-simplex $P = \conv(0, e_1, e_2, (1, 2, 4, 0), (2, 1, 0, 4)) \subseteq \mathbb{R}^4$. The sublattice $N$ of $\mathbb{Z}^4$ spanned by all lattice points of $P$ has index $2$ and the quotient $\mathbb{Z}^4/N \cong \mathbb{Z}/2\mathbb{Z}$ is generated by $\overline{e_3} = \overline{e_4}$. A computation in \texttt{SageMath} with backend \texttt{Normaliz} shows that $P$ is thin and $h_P^\ast(t) = t^3 + 11t^2 + 3t + 1$, in particular $\deg(P) = 3$, so $P$ is not trivially thin. A computation in \texttt{Polymake} shows that the lattice width of $P$ is $2$, so that $P$ is not a Cayley polytope, in particular not a free join. It can be checked that with respect to $N$, $P$ is the lattice pyramid over a reflexive $3$-simplex of lattice volume $8$. \label{ex:nonsp} \end{ex} Question~\ref{question} should be understood as a guiding question for finding interesting high-dimensional thin polytopes. Let us discuss this problem in more detail in the following. As being hollow is equivalent to $\deg(P) < \dim(P)$, it is evident that every hollow lattice polytope in dimension $\le 2$ is trivially thin. Hence, by Proposition~\ref{prop:dim2} every thin polytope in dimension $\le 2$ is trivially thin. It will be proven in our first main result Theorem~\ref{thm:3d} that in dimension $3$ all non-trivially thin polytopes are lattice pyramids. In particular, Question~\ref{question} has an affirmative answer in dimensions $\le 3$. Note that $\conv(e_1,e_2,-e_1-e_2) \circ_\mathbb{Z} 2 \Delta_2$ is an example of a thin simplex in dimension $5$ that is not trivially thin (it has degree $3$), but is not a lattice pyramid, while being a free join with a (trivially) thin factor. In higher dimensions our second main result shows that non-trivially thin Gorenstein polytopes are Gorenstein joins with a trivially thin factor, so that Question~\ref{question} has an affirmative answer also in the Gorenstein case (cf. Corollary~\ref{cor:spanning}). Computationally, we have verified that Question~\ref{question} has an affirmative answer for all $4$-dimensional lattice polytopes of lattice volume $\leq 21$, for all $5$-dimensional lattice simplices of lattice volume $\leq 20$ and for all $6$-dimensional lattice simplices of lattice volume $\leq 16$. We provide some of the relevant data at \cite{github}. \subsection{Interesting thin empty simplices?} A lattice simplex is called {\em empty} if its vertices are its only lattice points. Among the hollow polytopes this is the class of lattice simplices that has been studied most intensively, see e.g. \cite{Santos4d} and the references therein. However, it turns out that there are no interesting thin empty simplices in dimension at most $4$. Let us give the easy reasoning. For this, we recall that the {\em quotient group} of a $d$-dimensional lattice simplex $P \subset \mathbb{R}^d$ is defined as the quotient of $\mathbb{Z}^{d+1}$ by the subgroup generated by the vertices of $P \times \{1\}$. \begin{prop} Let $P$ be a lattice simplex with cyclic quotient group. Then $P$ is thin if and only if $P$ is a lattice pyramid. \end{prop} \begin{proof} Let $P \subset \mathbb{R}^d$ be $d$-dimensional. We denote by $\Pi$ the half-open parallelepiped from Example~\ref{ex:simplex}. Clearly, every element in the quotient group of $P$ has a unique representative in $\Pi \cap \mathbb{Z}^{d+1}$. Let $g \in \Pi \cap \mathbb{Z}^{d+1}$ be the representative of a generator of the quotient group of $P$. We assume that $P$ is thin. Hence, there is a proper, non-empty subset $V'$ of the vertex set of $S \times \{1\}$ such that $g$ is a linear combination of vertices of $V'$. In particular, this also holds for the representatives of all the elements in the quotient group of $P$. Now, it follows from \cite[Lemma~12]{NillSimplices} that $P$ is a lattice pyramid. \end{proof} It is well-known that all empty lattices simplices in dimension at most $4$ have cyclic quotient group \cite{barile}. As also in higher dimensions most empty simplices constructed (but not all of them) have this property (see e.g. \cite{WideSimplices}), it seems to be a challenge to find examples of empty simplices that are thin but not simply lattice pyramids. \subsection{Are thin polytopes `flat'?} We observed above that all thin polytopes in dimension at most two have lattice width $1$ except for $2 \Delta_2$. We leave it as an exercise to the reader to show that $2 \Delta_d$ for even $d$ is the only thin simplex among all lattice simplices of the form $\conv(0,k_1 e_1, \ldots, k_d e_d) \subset \mathbb{R}^d$ with $k_1, \ldots, k_d \in \mathbb{Z}_{\ge 1}$ that are not lattice pyramids (i.e., $k_i > 1$ for all $i$). It will follow from our main results that thin polytopes in dimension three (Corollary~\ref{cor:flat-3d}) as well as thin Gorenstein polytopes in arbitrary dimension (Corollary~\ref{cor:flat-gorst}) have lattice width $1$. In dimension four Example~\ref{ex:nonsp} has lattice width $2$. As thin polytopes (of dimension $> 0$) are hollow, in fixed dimension their lattice width is bounded. Now, our lack of 'non-flat' examples motivates the following question. \begin{question}\label{question2} Can one find (spanning) thin polytopes of arbitrarily large lattice width? \end{question} We expect that such examples with increasing lattice width should exist with increasing dimension. Note that if one assumes that Question~\ref{question}(2) has an affirmative answer, then for Question~\ref{question2} it would be important to find the maximum width of trivially thin spanning polytopes $P$. However, it is a folklore open question, often called Cayley conjecture (see \cite{Alicia,HNP09,Higashitani}), that any lattice polytope with $\dim(P) > 2 \deg(P)$ has lattice width $1$. Thus, assuming also that the Cayley conjecture holds essentially reduces the previous question to the study of spanning lattice polytopes with $\dim(P) = 2 \deg(P)$. \section{Classification of thin polytopes in dimension $3$} \label{sec:3d} As observed above, three-dimensional lattice polytopes $P$ that are lattice pyramids over polygons or have degree at most one are automatically thin. Our first main result, Theorem~\ref{thm:3d}, shows that in dimension three indeed all the thin polytopes are of this type. Lattice polytopes of degree at most one are completely known in any dimension. For this, let us recall the following definition. \begin{defn} A {\em Lawrence prism} is a $d$-dimensional lattice polytope in $\mathbb{R}^d$ isomorphic to $\conv(0,e_1,\ldots, e_{d-1}, k_0 e_d, e_1+k_1 e_d, \ldots, e_{d-1}+k_{d-1} e_d)$ with $k_0, k_1, \ldots, k_{d-1} \in \mathbb{Z}_{\ge 1}$. \end{defn} The following result was proven in \cite{batyrev2007multiples}. \begin{thm} Any lattice polytopes of degree $1$ is either a lattice pyramid, a Lawrence prism or isomorphic to $2 \Delta_2$. \label{thm:BN} \end{thm} Here is the main result of this section. \begin{thm} Let $P$ be a $3$-dimensional lattice polytope. Then $P$ is thin if and only if $P$ is a lattice pyramid over a lattice polygon or $\deg(P) \le 1$. Equivalently, $P$ is thin if and only if \begin{itemize} \item $P$ is a lattice pyramid over a lattice polygon, or \item $P$ is a Lawrence prism. \end{itemize} \label{thm:3d} \end{thm} \begin{cor} Every three-dimensional thin polytope has lattice width $1$.\label{cor:flat-3d} \end{cor} The proof of Theorem~\ref{thm:3d} relies on two instances that seem to be exceptional to small dimensions. First, in dimension three all the coefficients of the local $h^*$-polynomial can be explicitly determined. \begin{prop} Let $P \subseteq \mathbb{R}^3$ be a $3$-dimensional lattice polytope. Then \begin{equation*} l^\ast_P(t) = |\interior_{\mathbb{Z}}(P)| (t + t^3) + \left(|\interior_{\mathbb{Z}}(2P)| - 4 |\interior_{\mathbb{Z}}(P)| - \sum_{F \leq P \text{ facet}} |\interior_{\mathbb{Z}}(F)|\right) t^2. \end{equation*} \label{prop:lstar-3d} \end{prop} \begin{proof} Recall from Theorem~\ref{thm:local_h} that $l^\ast_1 = l^\ast_{3} = h^\ast_{3} = |\interior_{\mathbb{Z}}(P)|$. Hence, we need only determine $l^\ast_2$. From Stanley reciprocity, we deduce $h^\ast_2=|\interior_{\mathbb{Z}}(2P)| - 4 |\interior_{\mathbb{Z}}(P)|$. Now, in the notation of the $h^\ast$-diamond introduced in \cite{Katz2016Local}, we have $h^\ast_2 = h^\ast_{1,0}+h^\ast_{1,1}$, where $h^\ast_{1,1}=l^\ast_2$ and $h^\ast_{1,0} = \sum_{F \leq P \text{ facet}} |\interior_{\mathbb{Z}}(F)|$ by \cite[Example~8.9]{Katz2016Local}. This implies the statement. \end{proof} Let us note that we get from the lower bound theorem of Katz-Stapledon, Theorem~\ref{thm:lower}, $l^*_1 \le l^*_2$. This leads to the following non-obvious corollary. It would be very interesting to find a purely combinatorial proof. \begin{cor} Let $P \subseteq \mathbb{R}^3$ be a $3$-dimensional lattice polytope. Then \[|\interior_{\mathbb{Z}}(2P)| \ge 5 \,|\interior_{\mathbb{Z}}(P)| + \sum_{F \leq P \text{ facet}} |\interior_{\mathbb{Z}}(F)|.\] \label{cor:comb3d} \end{cor} For our purposes, let us note the following numerical characterization of thinness in dimension three. \begin{cor} \label{cor:thinCondition} Let $P \subseteq \mathbb{R}^3$ be a $3$-dimensional lattice polytope. Then $P$ is thin if and only if $P$ is hollow and \begin{equation*} |\interior_{\mathbb{Z}}(2P)| = \sum_{F \leq P \text{ facet}} |\interior_{\mathbb{Z}}(F)|. \end{equation*} \end{cor} The second result that is not yet available in higher dimensions is a complete classification of hollow three-dimensional lattice polytopes. For this, we denote by a {\em lattice projection} $\mathbb{R}^d \to \mathbb{R}^m$ an affine-linear map mapping $\mathbb{Z}^d$ surjectively onto $\mathbb{Z}^m$. \begin{thm}[\cite{averkov2011maximal}]\label{thm:3Dhollow} Let $P \subseteq \mathbb{R}^3$ be a $3$-dimensional hollow lattice polytope. Then one of the following holds: \begin{enumerate} \item \label{it:exceptions} $P$ is contained in one of the $12$ maximal hollow lattice polytopes classified in \cite{averkov2011maximal}. \item \label{it:width1} There is a lattice projection $\mathbb{R}^3 \rightarrow \mathbb{R}^1$ mapping $P$ onto $\Delta_1$. \item \label{it:2delta2} There is a lattice projection $\mathbb{R}^3 \rightarrow \mathbb{R}^2$ mapping $P$ onto $2 \Delta_2$. \end{enumerate} \end{thm} Before giving the proof of Theorem~\ref{thm:3d} let us also recall the following well-known formula for the mixed volume (e.g. \cite[Corollary~3.2]{nill2020mixed}): \begin{lem}\label{lem:MVinterior} Let $P_1, P_2 \subseteq \mathbb{R}^2$ be lattice polytopes. Then \begin{align*} \MV(P_1, P_2) = 1 &+ (-1)^{\dim(P_1 + P_2)}|\interior_{\mathbb{Z}}(P_1 + P_2)| \\ &+(-1)^{\dim(P_1) - 1}|\interior_{\mathbb{Z}}(P_1)| + (-1)^{\dim(P_2) - 1}|\interior_{\mathbb{Z}}(P_2)|. \end{align*} \end{lem} \begin{proof}[Proof of Theorem~\ref{thm:3d}] By Corollary~\ref{cor:thinCondition}, $P$ is hollow. We treat the three cases of Theorem~\ref{thm:3Dhollow} separately. A direct computation in \texttt{Magma} deals with case~\ref{it:exceptions}, see \cite{github}. For case~\ref{it:width1}, denote by $P_1, P_2 \subseteq \mathbb{R}^2$ the preimages in $P$ of the vertices of $\Delta_1$. Note that $P_1$ and $P_2$ are faces of $P$ such that every lattice point of $P$ is either contained in $P_1$ or $P_2$. (We remark that $P$ is a Cayley polytope of $P_1$ and $P_2$ in the notation of Definition~\ref{def:cayley}.) We denote by $\interior_{\mathbb{Z}}^2(Q)$ the set of lattice points in the \emph{absolute} interior of a lattice polytope $Q \subseteq \mathbb{R}^2$. Then \begin{align*} \sum_{F \leq P \text{ facets}} |\interior_{\mathbb{Z}}(F)| &= |\interior_{\mathbb{Z}}^2(P_1)| + |\interior_{\mathbb{Z}}^2(P_2)|, \\ |\interior_{\mathbb{Z}}(2P)| &= |\interior_{\mathbb{Z}}(P_1 + P_2)|, \end{align*} where the second equation follows from the so-called Cayley trick. Therefore, Corollary~\ref{cor:thinCondition} translates into \begin{equation*} |\interior_{\mathbb{Z}}^2(P_1)| + |\interior_{\mathbb{Z}}^2(P_2)| = |\interior_{\mathbb{Z}}(P_1 + P_2)|. \end{equation*} In case $\dim(P_1) = \dim(P_2) = 2$, plugging this into Lemma~\ref{lem:MVinterior} yields $\MV(P_1, P_2) = 1$, thus $(P_1, P_2) \cong (\Delta_2, \Delta_2)$ by \cite[Proposition~2.7]{cattani2013mixed}, hence $\deg(P) = 1$. If $\dim(P_1)=2$ and $\dim(P_2)=1$, then Lemma~\ref{lem:MVinterior} yields $\MV(P_1,P_2) = 1 + |\interior_{\mathbb{Z}}(P_2)|$. On the other hand, $\MV(P_1,P_2) = V(\pi_{P_2}(P_1))(|\interior_{\mathbb{Z}}(P_2)|+1)$ by \cite[Theorem~5.3.1]{schneider_2013}, where $\pi_{P_2}$ is a lattice projection along the line segment $P_2$ and $V(\pi_{P_2}(P_1))$ denotes the lattice volume. Hence, $V(\pi_{P_2}(P_1)) = 1$ and therefore $\pi_{P_2}(P_1) \cong \Delta_1$. The lattice projection of $P$ along $P_2$ is then a lattice projection onto $\Delta_2$. Thus, $\codeg(P) \ge 3$ and hence $\deg(P) \leq 1$, so either $\deg(P) = 1$ or $P \cong \Delta_3$ is a lattice pyramid. The case $\dim(P_1) = \dim(P_2) = 1$ is similar. Lemma~\ref{lem:MVinterior} yields $\MV(P_1,P_2)= 1 + |\interior_{\mathbb{Z}}(P_1)| + |\interior_{\mathbb{Z}}(P_2)|$. On the other hand, again $\MV(P_1,P_2)=V(\pi_{P_2}(P_1)) (|\interior_{\mathbb{Z}}(P_2)|+1)$. We may assume $|\interior_{\mathbb{Z}}(P_1)| \leq |\interior_{\mathbb{Z}}(P_2)|$. If $V(\pi_{P_2}(P_1)) \geq 2$ or $V(\pi_{P_2}(P_1)) = 0$, we obtain a contradiction, so $V(\pi_{P_2}(P_1)) = 1$. The same argument as above shows $\deg(P) \leq 1$. Finally, if one of the $P_i$ is zero-dimensional, then $P$ is a lattice pyramid.\vspace{1em} It is left to study case~\ref{it:2delta2}, and we may assume $P$ to be of lattice width at least $2$ because width $1$ is equivalent to $P$ being a Cayley polytope which is precisely case \ref{it:width1}. We distinguish several cases and always start by showing how, in each case, we can associate to each lattice point in the interior of a facet of $P$, in an injective way, a lattice point in the interior of $2P$. We then prove that there always exists an additional lattice point in the interior of $2P$, therefore showing that case~\ref{it:2delta2} does not yield any new thin polytopes by Corollary~\ref{cor:thinCondition}. We may assume that $P$ projects onto $2 \Delta_2$ along the $z$-axis. As lattice projections map interior lattice points to interior lattice points, all interior lattice points of a facet of $P$ are of the form $x_1^a=(1,0,a)$, $x_2^a=(0,1,a)$, or $x_3^a=(1,1,a)$ for suitable $a \in \mathbb{Z}$. By fixing vertices $v_1=(0,2,\alpha)$, $v_2=(2,0,\beta)$, $v_3=(0,0,\gamma)$ of $P$ we hence obtain points $\frac{1}{2}(x_i^a + v_i) \in \interior(P)$, and therefore $(x_i^a + v_i) \in \interior_{\mathbb{Z}}(2P)$ for all $a \in \mathbb{Z}$ such that $x_i^a$ is an interior point of a facet of $P$. Then $(x_i^a+v_i) \neq (x_j^b+v_j)$ if~$i \neq j$ or~$a \neq b$. Now we show the existence of an additional interior lattice point. Indeed, $P$ can have at most three facets containing interior lattice points, namely at most those facets, if there are such, that project to one of the three edges of $2 \Delta_2$. We proceed by distinguishing these different cases. If there is no such facet at all, then Corollary~\ref{cor:thinCondition} implies that $2P$ is hollow, so $\deg(P) \leq 1$, contradicting the fact that there is no lattice polytope of degree $\leq 1$ with width $>1$ by~Theorem~\ref{thm:BN}. Next, assume that $P$ has exactly two facets containing interior lattice points, and say these are the facets opposite to $v_1$ and $v_2$. We pick two such points $(0,1,q)$, $(1,0,r)$. Then we obtain as many interior lattice points in $2P$ of the form $x_1^a+v_1$ or $x_2^a+v_2$ as there are points in the interiors of facets of $P$, and $(1,1,q+r)$ is an additional interior lattice point of $2P$. Next, assume $P$ has three facets containing interior lattice points. If there exists $i \in \{1,2,3\}$ such that the fiber of $2P$ containing $v_i$ contains more than one lattice point, then we can similarly construct an additional interior lattice point of $2P$ by considering the two points of minimal and maximal height in this fiber. Therefore, we may assume $v_1, v_2, v_3$ to be the unique lattice points of $P$ over $(0,2)$, $(2,0)$ and $(0,0)$, respectively. This implies that all three facets $F_1$, $F_2$, $F_3$ of $P$ lying over the three edges of $2 \Delta_2$ have a special form. E.g., the facet projecting to the edge $[(0,0), (2,0)]$ is a quadrangle or triangle which, after a unimodular equivalence, looks similar to the following: \begin{center} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[->, line width=0.8pt] (-1,-2) -- (-1,2) node[midway, left]{$z$};\draw[->, line width=0.8pt] (0,-3) -- (2,-3) node[midway, below]{$x$};\draw[line width=1pt] (1, -2) -- (2, -1);\draw[line width=1pt] (1, 2) -- (2, -1);\draw[line width=1pt] (0, 0) -- (1, 2);\draw[line width=1pt] (0, 0) -- (1, -2);\draw[line width=1pt, dotted] (0, 0) -- (2, -1);\node at (1, -1/4) {}; \node at (1, 0) {}; \node[above] at (1,2) {$p_{\mathrm{max}}$}; \fill (0,0) circle (0.08);\fill (1,-2) circle (0.08);\fill (1,-1) circle (0.08);\fill (1,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (2,-1) circle (0.08);\fill (0,1) circle (0.08);\fill (0,2) circle (0.08);\fill (0,-1) circle (0.08);\fill (0,-2) circle (0.08);\fill (2,0) circle (0.08);\fill (2,-2) circle (0.08);\fill (2,1) circle (0.08);\fill (2,2) circle (0.08); \end{tikzpicture} \end{center} Now, we fix one interior lattice point $u_i \coloneqq x_i^{a_i}$ in each of the facets for some suitable $a_i \in \mathbb{Z}$. The three maps $x_i^a \mapsto x_i^a + u_j$ for $(i,j) \in \{(1,2),(2,3),(3,1)\}$ map each interior lattice point of the three facets $F_i$ injectively to an interior lattice point of $2P$. Moreover, the images of these three maps are disjoint by construction. Now, we may assume that in the facet projecting to the edge $[(0,0), (2,0)]$, the lattice point $p_{\mathrm{max}}$ lying over $(1,0)$ with maximal third coordinate is a vertex as in the picture. But then we obtain the additional interior lattice point $p_{\mathrm{max}}+u_2$ of $2P$ not covered by the images of the three maps above.\vspace{1em} The only remaining case is the one where only one facet $F$ of $P$ contains interior lattice points. We may assume $F$ is the facet which projects onto the edge $[(0,0), (2,0)]$. Then all interior lattice points of $F$ project to $(1,0)$ and are of the form $x_1^a$ for $a \in \mathbb{Z}$ ranging in a suitable interval. From this we obtain $|\interior_{\mathbb{Z}}(F)|$ interior lattice points $v_1 + x_1^a$ of $2P$. Observe that all of them have second coordinate $2$. Therefore, it is enough to show that $2P$ contains an interior lattice point with second coordinate $1$. Again we proceed by distinguishing cases. If $F$ contains at least three lattice points projecting to $(1,0)$, then it contains the convex hull of $(1,0,a)$, $(1,0, a+1)$, $(1,0,a+2)$, $(2,0,b)$ for some $a,b \in \mathbb{Z}$. Recall that $v_1=(0,2,\alpha)$. Then the point \begin{equation*} 2 \cdot \left(\frac{1}{4} v_1 + \frac{1}{4} (1,0,a) + \frac{1}{4} (2,0,b) + \frac{1}{4} (1,0,a+n)\right) = (2,1,a + \frac{\alpha + b + n}{2}) \end{equation*}is an interior lattice point of $2P$ with second coordinate $1$ for precisely one choice of $n \in \{1, 2\}$. Hence, by Corollary~\ref{cor:thinCondition}, $P$ is not thin if $F$ contains at least three lattice points over $(1,0)$, which in particular includes the case $|\interior_{\mathbb{Z}}(F)| \geq 3$. Before we proceed further let us observe that the fibers of $P$ over the points $(0,0)$ and $(2,0)$ both consist of at most three lattice points because otherwise another facet than $F$ would contain interior lattice points. This is because, up to unimodular equivalence, there are exactly two lattice pyramids of height $2$ over a lattice segment of length $\geq 3$, and both contain an interior lattice point: \begin{center} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (0, 2) -- (3, 0);\draw[line width=1pt] (0, 0) -- (3, 0);\draw[line width=1pt] (0, 0) -- (0, 2);\draw[line width=1pt] (0, 0) -- (1, 2);\draw[line width=1pt] (1, 2) -- (3, 0);\node at (1, 2/3) {}; \fill (0,0) circle (0.08);\fill (1,0) circle (0.08);\fill (2,0) circle (0.08);\fill (3,0) circle (0.08);\fill (0,1) circle (0.08);\fill (1,1) circle (0.11);\fill (0,2) circle (0.08);\fill (1,2) circle (0.08);\fill (2,1) circle (0.08);\fill (2,2) circle (0.08);\fill (3,1) circle (0.08);\fill (3,2) circle (0.08); \end{tikzpicture} \end{center} Now, we can deal with the remaining case $|\interior_{\mathbb{Z}}(F)| \in \{1,2\}$. Let $F' \subseteq F$ run through the inclusion-minimal subpolygons of $F$ which contain the same interior lattice points as $F$. We will show that there are only two possibilities for $F'$. Let us first consider the case $|\interior_{\mathbb{Z}}(F)| = 2$. By the previous argument, after a suitable shear fixing the $x = 1$ line (within the $x$-$z$-plane, i.e. $y=0$), $F'$ fits inside the following box: \begin{center} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[->, line width=0.8pt] (-1,0) -- (-1,3) node[midway, left]{$z$};\draw[->, line width=0.8pt] (0,-1) -- (2,-1) node[midway, below]{$x$};\draw[line width=1pt] (0, 3) -- (2, 3);\draw[line width=1pt] (2, 0) -- (2, 3);\draw[line width=1pt] (0, 0) -- (2, 0);\draw[line width=1pt] (0, 0) -- (0, 3);\node at (1, 3/2) {}; \fill (0,0) circle (0.08);\fill (0,1) circle (0.08);\fill (0,2) circle (0.08);\fill (0,3) circle (0.08);\fill (1,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (1,3) circle (0.08);\fill (2,0) circle (0.08);\fill (2,1) circle (0.08);\fill (2,2) circle (0.08);\fill (2,3) circle (0.08); \end{tikzpicture} \end{center} As $F'$ is inclusion-minimal, it is isomorphic to one of the following polygons: \begin{center} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (1, 3) -- (2, 0);\draw[line width=1pt] (0, 0) -- (2, 0);\draw[line width=1pt] (0, 0) -- (1, 3);\node at (1, 1) {}; \fill (0,0) circle (0.08);\fill (1,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (1,3) circle (0.08);\fill (2,0) circle (0.08); \end{tikzpicture} \hspace{10pt} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (1, 3) -- (2, 1);\draw[line width=1pt] (0, 0) -- (2, 1);\draw[line width=1pt] (0, 0) -- (1, 3);\node at (1, 4/3) {}; \fill (0,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (1,3) circle (0.08);\fill (2,1) circle (0.08); \end{tikzpicture} \hspace{10pt} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (1, 0) -- (2, 2);\draw[line width=1pt] (1, 3) -- (2, 2);\draw[line width=1pt] (0, 2) -- (1, 3);\draw[line width=1pt] (0, 2) -- (1, 0);\node at (1, 7/4) {}; \fill (0,2) circle (0.08);\fill (1,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (1,3) circle (0.08);\fill (2,2) circle (0.08); \end{tikzpicture} \hspace{10pt} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (1, 0) -- (2, 2);\draw[line width=1pt] (1, 3) -- (2, 2);\draw[line width=1pt] (0, 1) -- (1, 3);\draw[line width=1pt] (0, 1) -- (1, 0);\node at (1, 3/2) {}; \fill (0,1) circle (0.08);\fill (1,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (1,3) circle (0.08);\fill (2,2) circle (0.08); \end{tikzpicture} \hspace{10pt} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (0, 2) -- (2, 3);\draw[line width=1pt] (2, 1) -- (2, 3);\draw[line width=1pt] (0, 0) -- (2, 1);\draw[line width=1pt] (0, 0) -- (0, 2);\node at (1, 3/2) {}; \fill (0,0) circle (0.08);\fill (0,1) circle (0.08);\fill (0,2) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (2,1) circle (0.08);\fill (2,2) circle (0.08);\fill (2,3) circle (0.08); \end{tikzpicture} \end{center} But only the last one does not contain three lattice points over $(1,0)$. Let us also consider the case of $|\interior_{\mathbb{Z}}(F)| = 1$. Here, we can fit $F'$ inside the standard square $[0,2]^2$ (inside the $x$-$z$-plane) after a suitable shear fixing the $x = 1$ line, so $F'$ can be taken to be one of the following polygons: \begin{center} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (1, 2) -- (2, 0);\draw[line width=1pt] (0, 0) -- (2, 0);\draw[line width=1pt] (0, 0) -- (1, 2);\node at (1, 2/3) {}; \fill (0,0) circle (0.08);\fill (1,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (2,0) circle (0.08); \end{tikzpicture} \hspace{10pt} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (1, 0) -- (2, 1);\draw[line width=1pt] (1, 2) -- (2, 1);\draw[line width=1pt] (0, 1) -- (1, 2);\draw[line width=1pt] (0, 1) -- (1, 0);\node at (1, 1) {}; \fill (0,1) circle (0.08);\fill (1,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (2,1) circle (0.08); \end{tikzpicture} \hspace{10pt} \begin{tikzpicture}[scale=0.5,baseline=-5pt] \draw[line width=1pt] (1, 2) -- (2, 1);\draw[line width=1pt] (0, 0) -- (2, 1);\draw[line width=1pt] (0, 0) -- (1, 2);\node at (1, 1) {}; \fill (0,0) circle (0.08);\fill (1,1) circle (0.08);\fill (1,2) circle (0.08);\fill (2,1) circle (0.08); \end{tikzpicture} \end{center} Again, only the last one does not contain three lattice points over $(1,0)$. Lastly, for each of these two remaining polygons $F'$ we may choose a lattice subpolytope $P'$ of $P$ that is a pyramid of height $2$ over $F'$. We observe that for given $F'$ there are at most four non-isomorphic possibilities for $P'$ to consider as the first two coordinates of an apex in $\mathbb{R}^2 \times \{2\}$ over a base polytope in $\mathbb{R}^2 \times \{0\}$ may be chosen by a unimodular shearing to be in $\{(0,0),(1,0),(0,1),(1,1)\}$. Now, an explicit computation in \texttt{SageMath} shows that for all these at most eight cases we have $|\interior_{\mathbb{Z}}(2P')| > |\interior_{\mathbb{Z}}(F')| = |\interior_{\mathbb{Z}}(F)|$, concluding the proof. \end{proof} We can now answer the original question in \cite{GKZ94} in dimension $3$. \begin{cor} A three-dimensional lattice simplex is thin if and only if it is a lattice pyramid.\label{cor:tetra} \end{cor} This follows directly from Theorem~\ref{thm:3d}. The reader is cautioned not to jump to the conclusion that the same result may be true in higher dimensions. In dimension $4$, $[-1,1] \circ_\mathbb{Z} 2 \Delta_2$ is an example of a (trivially) thin simplex that is not a lattice pyramid. \section{Thin Gorenstein polytopes and Gorenstein joins} \label{sec:join} \subsection{Gorenstein polytopes and their duals} \begin{defn} A lattice polytope $P$ is called {\em Gorenstein} if $h^*_P$ is palindromic. \end{defn} Let us recall that {\em reflexive polytopes} are precisely the Gorenstein polytopes of codegree one. For more background on reflexive and Gorenstein polytopes, its relevance in toric geometry and mirror symmetry, as well as alternative characterizations we refer to \cite{BB96b, batyrev2008combinatorial, NS13}. Here, let us summarize definition and properties of the dual Gorenstein polytope. We remark that the codegree of a Gorenstein polytope is often called its index. \begin{prop} Let $P \subset \mathbb{R}^d$ be a $d$-dimensional Gorenstein polytope. Then $\codeg(P) P$ is a reflexive polytope (up to lattice translation), we denote its unique interior lattice point by $w$. Then \[P^\times := \{y \in (\mathbb{R}^{d+1})^* \;:\; \pro{y}{w} = 1 \text{ and } \pro{y}{x} \ge 0 \;\forall\, x \in P \times \{1\}\}\] is a Gorenstein polytope, called the {\em dual Gorenstein polytope} of $P$. It has the same dimension and degree as $P$ and is combinatorially dual to $P$. If $F$ is a face of $P$, we denote by $F^*$ the {\em dual face}, i.e., the corresponding face of $P^\times$. \label{prop:dual-def} \end{prop} If the Gorenstein polytope is lower-dimensional, we consider, as usual, its ambient lattice in order to get its dual Gorenstein polytope. Local $h^*$-polynomials of Gorenstein polytopes (often called $\tilde{S}$-polynomials) allow to give an elegant formula for computing stringy $E$-polynomials of Calabi-Yau complete intersections in toric Gorenstein Fano varieties (we refer to \cite{BorisovMavlyutov, batyrev2008combinatorial}). In this context, several questions about stringy $E$-polynomials are still open, see \cite{batyrev2008combinatorial,NS13}. Here, we make some progress in this direction by addressing the question when the local $h^*$-polynomial of a Gorenstein polytope vanishes. As one consequence of our main result, Theorem~\ref{thm:main}, we will see that not only the degree of the $h^*$-polynomials of Gorenstein polytopes and their duals are the same but also of their $l^*$-polynomials (Corollary~\ref{cor:dualdeg}). \subsection{Joins, Cayley polytopes, and Cayley joins} In the following let us discuss some important notions of decomposing lattice polytopes that turn up naturally when studying Gorenstein polytopes (we refer to \cite{batyrev2008combinatorial, NS13}). \begin{defn} Let $P \subseteq \mathbb{R}^d$ be a polytope and $F, G$ non-empty subsets of $P$. Then $P$ is the {\em join} of $F$ and $G$, written $P = F \circ G$, if $P = \conv(F,G)$ and $\dim(P) = \dim(F) + \dim(G) + 1$. \end{defn} Note that in this case, $P$ is affinely-isomorphic to the free join $F \circ_\mathbb{Z} G$. In particular, $F$ and $G$ are automatically faces of $P$. \begin{rmk}{\rm Note that the join property is associative. Namely, given faces $F,G,H$ of $P$, then $P=F \circ (G \circ H)$, respectively, $P=(F \circ G) \circ H$, are both equivalent to $P=\conv(F,G,H)$ and $\dim(P) = \dim(F) + \dim(G) + \dim(H) + 2$.\label{remark:assoc} } \end{rmk} \begin{defn} Let $P \subseteq \mathbb{R}^d$ be a lattice polytope and $F, G$ non-empty subsets of $P$. Then $P$ is the {\em Cayley polytope} of $F$ and $G$, written $P = F * G$, if $P = \conv(F,G)$ and there exists an affine-linear map $\mathbb{R}^d \to \mathbb{R}$ mapping $\mathbb{Z}^d \to \mathbb{Z}$, such that $F \mapsto 0$ and $G \mapsto 1$. in other words, $P$ is a Cayley polytope if and only if there is a lattice projection mapping $P$ onto $\Delta_1$. \label{def:cayley} \end{defn} If $P=F*G$, then $F$ and $G$ are necessarily faces of $P$. Cayley polytopes can also be characterized as lattice polytopes with lattice width one. \begin{rmk} Given lattice polytopes $F$ and $G$ in $\mathbb{R}^d$, the convex hull of $F \times \{0\}$ and $G \times \{1\}$ is called the {\em Cayley sum} of $F$ and $G$. Its dimension is one larger than the dimension of the Minkowski sum of $F$ and $G$. If $P=F*G$, then $P$ is isomorphic to the Cayley sum of $F$ and $G$.\label{rem:cayley} \end{rmk} Cayley sums are important in the construction of high-dimensional Gorenstein polytopes, see e.g. \cite[Theorem~2.6]{batyrev2008combinatorial}. Note that the degree of a Cayley polytope is at most the degree of the dimension of the Minkowski sums of its factors, see Proposition~\cite[Proposition~1.12]{batyrev2007multiples}. Let us remark that Cayley sums and polytopes are usually defined for more than just two factors, which is sufficient here for our purposes. \begin{defn} Let $P \subseteq \mathbb{R}^d$ be a full-dimensional lattice polytope and $F,G \subseteq P$ faces. Then $P$ is the {\em Cayley join} of $F$ and $G$, written $P = F \circ_{\Cay} G$, if $P=F \circ G$ and $P=F * G$.\label{def:cayley-join} \end{defn} Clearly, the notion of a Cayley join is more restrictive than that of a Cayley polytope (e.g., $[0,1]^2$ is a Cayley polytope of two edges but not a Cayley join). The reader should be aware that Cayley polytopes and Cayley joins are in general not associative in the sense of Remark~\ref{remark:assoc}, see Example~\ref{ex:tetra} below. Let us recall some properties of a Gorenstein polytope that is a join or Cayley join, see \cite[Lemma~4.8, Proposition~4.9]{NS13}. \begin{prop} \label{prop:NS-prop} Let $P$ be a Gorenstein polytope. \begin{itemize} \item If $P=F \circ G$, then $P^\times = F^* \circ G^*$ with $F$ and $G^*$ (respectively, $G$ and $F^*$) being combinatorially dual to each other. \item If $P=F \circ_{\Cay} G$, then $F^\ast$ is a Gorenstein polytope with dual Gorenstein polytope $(F^\ast)^\times$, which can be identified with the lattice polytope $G$ with respect to a refined lattice. \end{itemize} In the last statement, the lattice does not have to be refined if the Cayley join is even a free join. \end{prop} \subsection{Gorenstein joins} In \cite{NS13} the following notion was defined. \begin{defn} Let $F$ and $G$ be faces of a Gorenstein polytope $P$. We say $P$ is a {\em Gorenstein join} of $F$ and $G$, denoted by $P = F \circ_{\Gor} G$, if $P = F \circ_{\Cay} G$ and $P^\times = F^* \circ_{\Cay} G^*$. We call $F$ and $G$ the {\em factors} of the Gorenstein join. \end{defn} We remark that Gorenstein joins do not have to be free joins, see \cite[Example~4.14]{NS13}. The following result, a strengthening of Stanleys monotonicity theorem in the case of faces of Gorenstein polytopes, motivated the previous definition of a Gorenstein join and gives a direct enumerative criterion for its existence (\cite[Theorems~3.6+4.12]{NS13}). \begin{thm} Let $P$ be a Gorenstein polytope and $F$ a non-empty proper face of $P$. Then $\codeg(P) \le \codeg(F) + \codeg(F^*)$ (equivalently, $\deg(F) + \deg(F^*) \le \deg(P)$), with equality if and only if $P$ is a Gorenstein join with factor $F$. In this case, $F$ is a Gorenstein polytope. \label{thm:NS} \end{thm} Gorenstein polytopes that are not Gorenstein joins have been previously also called {\em irreducible} in \cite{NS13}. As we see from the following result it is not necessary to compute the dual Gorenstein polytope to check whether a Cayley join is a Gorenstein join. \begin{lem}\label{lemma:GorensteinCodegree} Let $P = F \circ_{\Cay} G$ be a Gorenstein polytope which is the Cayley join of two faces $F, G \leq P$. Then $P = F \circ_{\Gor} G$ if and only if $\codeg(P) = \codeg(F) + \codeg(G)$ (or equivalently, $\deg(P) = \deg(F)+\deg(G)$). \end{lem} \begin{proof} By Theorem~\ref{thm:NS}, $P = F \ast_{\Gor} G$ if and only if $\codeg(F) + \codeg(F^\ast) = \codeg(P) =: r$, and in this case by \cite[Theorem~4.12]{NS13} $\codeg(G) = \codeg(P) - \codeg(F) = \codeg(F^\ast)$. Conversely, assume $\codeg(G) = \codeg(P) - \codeg(F)$. By Theorem~\ref{thm:NS}, the inequality $\codeg(F) + \codeg(F^\ast) \geq r$ always holds, so that by Theorem~\ref{thm:NS} again we only need to prove $\codeg(F^\ast) \leq \codeg(G)$. As $P$ is the Cayley join of $F$ and $G$, Proposition~\ref{prop:NS-prop} yields that $\codeg(F^\ast) = \codeg((F^\ast)^\times) \leq \codeg(G)$ as the codegree can only decrease under refinements of the lattice. \end{proof} \begin{rmk}As we will need it for the upcoming proofs, let us recall how to characterize Gorenstein polytope via cones. For more details, we refer to \cite{batyrev2008combinatorial}. The {\em cone over $P$} is denoted by $C_P \subseteq \mathbb{R}^{d+1}$ spanned by $P \times \{1\} \subset \mathbb{R}^{d+1}$. Any polyhedral cone in $\mathbb{R}^{d+1}$ that is unimodularly equivalent to some $C_P$ is called a {\em Gorenstein cone}. Now, $P$ is a Gorenstein polytope if and only if the dual cone $C_P^\vee = \{y \in (\mathbb{R}^{d+1})^*: \langle y, x \rangle \ge 0 \;\forall\, x \in C_P\}$ is a Gorenstein cone. In this case, $C_P^\vee$ is unimodularly equivalent to the cone over $P^\times$. \end{rmk} The following proposition contains a positive result regarding associativity of Gorenstein joins. In general, however, we do not expect associativity to hold. \begin{prop} Let $P \subseteq \mathbb{R}^d$ be a $d$-dimensional Gorenstein polytope with faces $F, G, H \leq P$ such that $P = (F \ast_{\Gor} G) \ast_{\Gor} H$. If $F$ is a vertex or $H$ is a vertex, then $P = F \ast_{\Gor} (G \ast_{\Gor} H)$. In particular, Gorenstein joins are associative for $\dim(P) \leq 3$. \end{prop} \begin{proof} By Lemma~\ref{lemma:GorensteinCodegree}, the faces $\conv(F,G)$ and $H$ of $P$ are themselves Gorenstein polytopes with $r := \codeg(P) = \codeg(\conv(F,G)) + \codeg(H)$. Applying the same result to the Gorenstein polytope $\conv(F,G) = F \ast_{\Gor} G$, we obtain that $F$ and $G$ are Gorenstein polytopes with $\codeg(\conv(F,G)) = \codeg(F) + \codeg(G)$. Hence, $r = \codeg(F) + \codeg(G) + \codeg(H)$. By Lemma~\ref{lemma:GorensteinCodegree}, it hence suffices to show $P = F \ast_{\Cay} (G \ast_{\Cay} H)$. That the join of $G$ and $H$ is a Cayley join is immediate from the assumption, so it is enough to show that the join of $F$ and $\conv(G,H)$ is a Cayley join.\medskip Let first $F$ be a vertex. Hence, $P$ is a pyramid with vertex $F$ and base $\conv(G,H)$. Consider the cone $C_P \subset \mathbb{R}^{d+1}$. As $P$ is Gorenstein of codegree $r$, there is a unique interior lattice point $p \in C_P$ on height $p_{d+1} = r$. Similarly, let $f$, $g$ and $h$ be the unique interior lattice points of $C_F, C_G, C_H \subseteq C_P$ on heights $\codeg(F), \codeg(G), \codeg(H)$, respectively. Then necessarily $f+g \in C_{\conv(F,G)} \subseteq C_P$ is the unique interior lattice point on height $\codeg(\conv(F,G))$ of the Gorenstein polytope $\conv(F,G)$. Hence, $p = f + g + h$. Let $u \in (\mathbb{Z}^{d+1})^*$ be the primitive inner facet normal of the hyperplane containing $\conv(G,H)$. As $u$ is a vertex of $P^\times$, we have $\langle u, p \rangle = 1$ (cf. Proposition~\ref{prop:dual-def}). Therefore, \begin{equation*} 1 = \langle u, p \rangle = \langle u, f \rangle + \langle u, g+h \rangle = \langle u, f \rangle. \end{equation*} But this means that $F$ and $\conv(G,H)$ have lattice distance equal to $1$, i.e., the combinatorial join $F \circ \conv(G,H)$ is a Cayley join.\medskip Let now $H$ be a vertex, so $P$ is a lattice pyramid with vertex $H$ and base $\conv(F,G)$. As $F \circ_{\Cay} G$, we may assume that ${\rm lin}(F,G)=\mathbb{R}^{d-1} \times\{0\}$, $H = \{e_d\}$, and there exists some $u \in (\mathbb{Z}^{d-1} \times \{0\})^*$ such that $\langle u, F \rangle = 0$ and $\langle u, G \rangle = 1$. Now, $\langle u + e^*_d, F \rangle = 0$ and $\langle u + e^*_d, \conv(G,H) \rangle = 1$. In particular, the join of $F$ and $\conv(G,H)$ is a Cayley join. \smallskip Finally, if $d = \dim(P) \leq 3$ and $P$ is the join of $F$, $G$ and $H$, then necessarily at least one of $F$ and $H$ is a vertex for dimension reasons, concluding the proof. \end{proof} We observe that a Gorenstein polytope $P$ is a lattice pyramid over a face $F$ with apex a vertex $v$ of $P$ if and only if $P$ is a Gorenstein join of $F$ and $v$. Hence, the previous result has the following consequence. \begin{cor} If $P$ is a Gorenstein join of two faces with one face a lattice pyramid, then $P$ is also a lattice pyramid. \label{cor:pyr} \end{cor} This result is already contained in the master's thesis \cite{masterthesis}. Let us give an example that shows that it fails for Cayley joins. \begin{ex} Consider $F := \conv(e_1,e_2) \times \{0\}$ and $G := \conv(0,-e_1-e_2) \times \{1\}$ in $\mathbb{R}^3$. Then its convex hull $P$ is a tetrahedron that is a Gorenstein polytope of lattice volume~$2$ (with $h^*_P(t)=1+t^2$ and $l^*_P(t)=t^2$). It is a Cayley join $P = F \circ_{\Cay} G$ but not a Gorenstein join as $\deg(P)=2\not=0=\deg(F)+\deg(G)$. Note that $F$ and $G$ are lattice pyramids, but $P$ is not. In particular, this example shows that the Cayley join property is not associative, and moreover, a Cayley join does not have to be thin if a factor of the Cayley join is thin.\label{ex:tetra} \end{ex} \subsection{Local $h^*$-polynomials of joins} Recall that $h^\ast$- and $l^\ast$-polynomials are multiplicative with respect to free joins (Proposition~\ref{prop:mult}). For general joins, one still gets inequalities. \begin{lem}\label{lemma:inequalities} Let $P$ be a lattice polytope which is the join of two faces $F$ and $G$. Then \begin{equation*} l^\ast(F;t) \cdot l^\ast(G;t) \leq l^\ast(P;t) \ \text{ and } \ h^\ast(F;t) \cdot h^\ast(G;t) \leq h^\ast(P;t). \end{equation*} If moreover $P$ is a Gorenstein polytope which is the Gorenstein join of $F$ and $G$, then also \begin{equation*} l^\ast(P^\times;t) \leq l^\ast(F^\times;t) \cdot l^\ast(G^\times;t) \ \text{ and } \ h^\ast(P^\times;t) \leq h^\ast(F^\times;t) \cdot h^\ast(G^\times;t). \end{equation*} \end{lem} \begin{proof} We will use the notation of \cite{NS13}.Let $\overline{M} = \mathbb{Z}^{d+1}$, and $M(F)$ denote the sublattice of $\overline{M}$ spanned by the lattice points in the linear hull of $F \times \{1\}$. Relative to the sublattice $M(F) \oplus_\mathbb{Z} M(G)$ of $\overline{M}$, $P$ becomes the free join of $F$ and $G$. Recall that by Corollary~\ref{cor:lattice} both the $h^\ast$-polynomial and the $l^\ast$-polynomial are (weakly) monotonically increasing under refinements of the lattice. It hence follows from Proposition~\ref{prop:mult} that $l^\ast(F;t) \cdot l^\ast(G;t) \leq l^\ast(P;t)$ and $h^\ast(F;t) \cdot h^\ast(G;t) \leq h^\ast(P;t)$.\medskip For the second claim, assume that $P \subseteq \mathbb{R}^d$ is a full-dimensional Gorenstein polytope of codegree $r$ with respect to the lattice $M = \mathbb{Z}^d \subseteq \mathbb{R}^d$. By assumption, $P$ is the Gorenstein join of two faces $F$ and $G$. By Theorem~\ref{thm:NS} and Lemma~\ref{lemma:GorensteinCodegree}, $F$ and $G$ are Gorenstein polytopes and $\codeg(F) + \codeg(G) = r$. For $\overline{M} = \mathbb{Z}^{d+1}$ we define the dual lattice $\overline{N} \coloneqq \mathrm{Hom}_\mathbb{Z}(\overline{M},\mathbb{Z}) \subseteq (\mathbb{R}^{d+1})^*$. By definition, as $P$ is a Gorenstein polytope, $C_P^\vee$ is a Gorenstein cone with respect to $\overline{N}$. Let $n = e_{d+1}^\ast \in \overline{N}$ be the unique interior lattice point of $C_P^\vee$ with $P \times \{1\} = C_P \cap \{x \in \mathbb{R}^{d+1}: \langle n, x \rangle = 1\}$. In the same way, we denote by $m \in \overline{M}$ the unique interior lattice point of $C_P$ such that $P^\times = C_P^\vee \cap \{y \in (\mathbb{R}^{d+1})^*: \langle y, m \rangle = 1\}$. Recall that $\langle n, m \rangle = r$ and hence $m = (p,r) \in M \oplus \mathbb{Z} = \overline{M}$ with $p \in M$ the unique interior lattice point of the $r$-th dilate $rP$ of $P$. Now, let us consider the sublattice $M(F) \oplus M(G) \subseteq \overline{M}$. With respect to this coarser lattice, the polytope $P$ is the \emph{free} join of $F$ and $G$, and this is clearly a Gorenstein polytope of codegree $\codeg(F) + \codeg(G) = r$. Hence, the $r$-th dilate $rP$ contains a unique interior lattice point in the original as well as in the coarser lattice. These two points must therefore agree, so the unique interior lattice point $m = (p,r) \in (rP) \times \{r\} \subseteq C_P$ with respect to the original lattice actually lies in $M(F) \oplus M(G)$. Let now $\Tilde{N} \subset (\mathbb{R}^{d+1})^*$ be the dual lattice of $M(F) \oplus M(G)$. Hence, $P^\times$ is with respect to the finer lattice $\Tilde{N}$ the dual Gorenstein polytope of the Gorenstein polytope $P$ considered with respect to the coarser lattice $M(F) \oplus M(G)$. By Proposition~\ref{prop:NS-prop}, the Gorenstein dual of the free join of the Gorenstein polytopes $F$ and $G$, is the free join of the Gorenstein duals $F^\times$ and $G^\times$. Again, monotonicity and multiplicativity proves the second claim: $l^\ast(P^\times;t) \leq l^\ast(F^\times;t) \cdot l^\ast(G^\times;t)$ and $h^\ast(P^\times;t) \leq h^\ast(F^\times;t) \cdot h^\ast(G^\times;t)$, where $P^\times$ is considered with respect to the original lattice $\overline{N}$ again. \end{proof} \begin{rmk} It follows in the situation of the second part of Lemma~\ref{lemma:inequalities} from Proposition~\ref{prop:NS-prop} and Remark~\ref{rmk:refine} that $l^\ast(F;t) \leq l^\ast((G^\ast)^\times;t)$ and $l^\ast(G;t) \leq l^\ast((F^\ast)^\times;t)$, because $(G^\ast)^\times$ is just the polytope $F$ with a possibly finer lattice, and analogously for $(F^\ast)^\times$ and $G$. The same holds for the $h^\ast$-polynomial. \end{rmk} \section{Characterization of thin Gorenstein polytopes} \subsection{The main result} \label{sec:gorst} The following notion will occur naturally in the proof of Theorem~\ref{thm:main}. \begin{defn} A lattice polytope $P$ is called {\em $g$-thin} if $\deg(g_P) = \deg(P)$. \end{defn} For instance, any unimodular simplex is $g$-thin. Note that by Proposition~\ref{prop:schepers}, we always have $\deg(g_P) \le \deg(P)$. As by construction (Definition~\ref{def:fgh}) $\deg(g_P) \le \dim(P)/2$, we observe that \[g\text{-thin} \quad\Longrightarrow\quad \text{trivially thin}\] \begin{ex} Let $P$ denote the lattice pyramid over $[-1,1]$. Then $P$ has dimension $2$, degree $1$ and $\deg(g_P)=0$ as it is a simplex. This is an example of a trivially thin Gorenstein polytope that is not $g$-thin. For another example, consider $2 \Delta_2$ which is a trivially thin (non-Gorenstein) simplex that is not $g$-thin. This example shows that a spanning thin polytope that is not a free join does not have to be $g$-thin (i.e., in Question~\ref{question} 'trivially thin' cannot be strengthened by 'g-thin'). \end{ex} Here is our main result. \begin{thm}\label{thm:main} Let $P$ be a Gorenstein polytope. Then the following are equivalent: \begin{enumerate}[label=(\roman*)] \item $P$ is thin,\label{item:thin} \item $P$ is trivially thin or $P = F \ast_{\Gor} G$ with at least one factor trivially thin,\label{item:trivially_thin} \item $P$ is $g$-thin or $P = F \ast_{\Gor} G$ with $\deg(l^*_F) = \deg(F)$ and $G$ $g$-thin.\label{item:g-thin} \end{enumerate} Moreover, if $P$ is not thin, then $\deg(l_P^\ast) = \deg(P)$. \end{thm} Let us remark that if $P$ is not thin, the last statement implies that $l^*_P(t)$ and $h^*_P(t)$ have the same leading coefficient $1$, as $h^*_P(t)$ is palindromic with constant coefficient $1$. We have to leave it as an open question whether it is possible to strengthen in the previous result `Gorenstein join' to `free join'. Let us note the following situation in which being a Gorenstein join (or even just a spanning join of faces) automatically implies being a free join. \begin{cor} Let $P$ be a spanning Gorenstein polytope. Then $P$ is thin if and only if it is trivially thin or a free join with a trivially thin factor (necessarily also a spanning Gorenstein polytope).\label{cor:spanning} \end{cor} The proof of Theorem~\ref{thm:main} relies critically on the decomposition of the $h^*$-polynomial into $l^*$-polynomials and $g$-polynomials (Proposition~\ref{prop:schepers}), valid also for general lattice polytopes. \begin{lem} Let $P$ be a lattice polytope with $\deg(l^*_P) < \deg(P)$. Then $P$ is $g$-thin or there exists a non-empty, proper face $F$ of $P$ with $\deg(P)=\deg(l^*_F)+\deg(g_{[F,P)})$. \label{lem:scheperscons} \end{lem} Here, we recall $\deg(l^*_{\emptyset})=\deg(1)=0$ and $\deg(g^*_{\emptyset})=\deg(1)=0$. \begin{proof} By the nonnegativity of $l^*$- and $g$-polynomials, Proposition~\ref{prop:schepers} implies that there exists a face $F$ of $P$ with $\deg(l^*_F)+\deg(g_{[F,P)}) = \deg(h^*_P)$. By our assumption, $F\not=P$. If $F=\emptyset$, then $\deg(h^*_P) = \deg(g_{[\emptyset,P)})$, so $P$ is $g$-thin. \end{proof} Let us note the following observation: \begin{lem} Let $P = F \ast_{\Gor} G$. Then the Gorenstein polytopes $F$, $G^*$, $F^\times$, $(G^*)^\times$ have the same degree, dimension, and degree of their $g$-polynomials. In particular, if any of these Gorenstein polytopes are trivially thin (respectively, $g$-thin), then all of them are. \label{lem:deg} \end{lem} \begin{proof} The Gorenstein property follows from Proposition~\ref{prop:NS-prop}. It is well-known, cf. \cite{batyrev2008combinatorial}, that duality of Gorenstein polytopes keeps dimension and degree invariant. It follows from Theorem~\ref{thm:Kalai} that this is also true for the degree of the $g$-polynomial. By Proposition~\ref{prop:NS-prop}, $F$ and $G^*$ are combinatorially dual to each other, hence, have the same dimension and by Theorem~\ref{thm:Kalai} the same degree of the $g$-polynomial. Finally, by Lemma~\ref{lemma:GorensteinCodegree} and Theorem~\ref{thm:NS}, \[\deg(P) - \deg(F) = \deg(G) = \deg(P) - \deg(G^*),\] hence, $\deg(F)=\deg(G^*)$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:main}] The implication \ref{item:g-thin} $\Rightarrow$ \ref{item:trivially_thin} is immediate. \ref{item:trivially_thin} $\Rightarrow$ \ref{item:thin}: Let $P = F \ast_{\Gor} G$ with $F$ trivially thin. By Lemma~\ref{lem:deg}, it follows that $(G^\ast)^\times$ is trivially thin as well. Now, applying Lemma~\ref{lemma:inequalities} to the factorization $P^\times = F^\ast \ast_{\Gor} G^\ast$ yields $l^\ast(P;t) \leq l^\ast((F^\ast)^\times; t) \cdot l^\ast((G^\ast)^\times; t) = 0$, so $P$ is thin. \ref{item:thin} $\Rightarrow$ \ref{item:g-thin}: Let $P$ be a Gorenstein polytope. We assume only that $\deg(l_P^\ast) < \deg(P)$ and will deduce \ref{item:g-thin}, so that $P$ is in particular thin by the implications we already proved (and thus, if $P$ is not thin, then $\deg(l^\ast_P) = \deg(P)$). Let us assume that $P$ is not $g$-thin. Now, by Lemma~\ref{lem:scheperscons} there exists a non-empty, proper face $F$ of $P$ with $\deg(P)=\deg(l^*_F)+\deg(g_{[F,P)})$. Theorem~\ref{thm:Kalai} shows that $\deg(g_{[F,P)})=\deg(g_{(F, P]^*}) = \deg(g_{F^*})$. Thus, Proposition~\ref{prop:schepers} and Theorem~\ref{thm:NS} imply that \[\deg(F^*) \ge \deg(g_{F^*}) = \deg(P) - \deg(l^*_F) \ge \deg(P) - \deg(F) \ge \deg(F^*).\] Therefore, $F^*$ is $g$-thin, $\deg(l^*_F)=\deg(F)$, and $\deg(F)+\deg(F^*)=\deg(P)$, which implies \ref{item:g-thin} by Theorem~\ref{thm:NS} and Proposition~\ref{prop:NS-prop} (with the roles of $F$ and $G$ exchanged). \end{proof} \begin{cor} Every thin Gorenstein polytope (of dimension $>0$) has lattice width $1$.\label{cor:flat-gorst} \end{cor} \begin{proof} As Gorenstein joins are Cayley joins, it remains by Theorem~\ref{thm:main} to show that a trivially thin Gorenstein polytope of dimension $>0$ is a Cayley polytope. This is precisely the statement of Theorem~3.1 in \cite{HNP09}. \end{proof} \begin{ex} Let us illustrate Theorem~\ref{thm:main} by showing that all thin Gorenstein polytopes $P$ of dimension $d=3$ are lattice pyramids over Gorenstein polygons (without using Theorem~\ref{thm:3d} directly). Let us assume otherwise. If $P$ is trivially thin, then $\deg(P) \le 1$, so by Theorem~\ref{thm:BN} $P$ is a Lawrence prism. Palindromicity implies $h^*_P(t)=1 + t$, so lattice volume $2$, which is a contradiction because any three-dimensional Lawrence prism has at least lattice volume $3$. Hence, by Theorem~\ref{thm:main} $P$ must be a lattice pyramid or a Gorenstein join of two Gorenstein intervals one of them being thin. As a thin interval is a unimodular simplex, thus a lattice pyramid, also $P$ is a lattice pyramid by Corollary~\ref{cor:pyr}. \end{ex} \subsection{Borisov's proof of the degree of $l^*$-polynomials of non-thin Gorenstein polytopes} Theorem~\ref{thm:main} answers affirmatively Question~6.3(b) in \cite{NS13} asking whether for Gorenstein polytopes having a non-vanishing $l^*$-polynomial forces its degree to be maximal (i.e., equal to the degree of the $h^*$-polynomial). Lev Borisov has provided us with an alternative algebraic proof of this fact that we reproduce here. It uses the description of the local $h^*$-polynomial of a lattice polytope as a Hilbert series of a graded ideal given in \cite{BorisovMavlyutov}. \begin{prop}Let $P \subseteq \mathbb{R}^d$ be a Gorenstein polytope of codegree $r$. Then either $P$ is thin or $l^\ast(P;t)$ starts with $t^r$. In this case, $l^\ast(P;t)$ has degree $\deg(P)$ and leading coefficient~$1$.\label{prop:borisov} \end{prop} \begin{proof} Let $K \subseteq \mathbb{Z}^{d+1}$ be the lattice points in the Gorenstein cone over $P \times \{1\}$. Denote by $\mathbb{C}[K]$ the associated affine semi-group algebra with $\mathbb{N}_0$-grading given by the exponent of $x_{d+1}$, viewing $\mathbb{C}[K] \subseteq \mathbb{C}[x_1^{\pm 1}, \ldots, x_d^{\pm 1}, x_{d+1}]$. As in \cite[Section~4]{BorisovMavlyutov}, we let $f \in \mathbb{C}[K]_1$ be non-degenerate and $I \subseteq \mathbb{C}[K]$ the homogeneous ideal generated by the so called logarithmic derivatives of $f$. Let moreover $J \subseteq \mathbb{C}[K]$ be the homogeneous ideal generated by all lattice points in the relative interior $K^\circ$ of $K$. Then Borisov and Mavlyutov define $R_1(f,K)$ to be the image of $J$ in the quotient ring $\mathbb{C}[K]/I$, i.e., $R_1(f,K)$ is the homogeneous ideal $(I+J)/I$ of $\mathbb{C}[K]/I$.\\ Now, by \cite[Proposition~5.5]{BorisovMavlyutov}, $l^\ast(P;t)$ is the Hilbert series of $R_1(f,K)$. Moreover, as $P$ is Gorenstein of codegree $r$, we have $K^\circ = (p,r) + K$, where $p \in (rP) \cap \mathbb{Z}^d$ is the unique interior lattice point of $rP$. Therefore, $R_1(f,K)$ is just the image of the principal ideal $(x^p x_{d+1}^r)$ in the quotient $\mathbb{C}[K]/I$. Hence, $R_1(f,K)$ is $0$ if and only if $x^p x_{d+1}^r \in I$, and otherwise the lowest degree of its non-zero homogeneous components is $r$ with $R_1(f,K)_r = \langle x^p x_{d+1}^r \rangle$ of vector space dimension $1$. This proves the first claim, and the second follows from reciprocity, Theorem~\ref{thm:lstar}(2). \end{proof} In dimensions $\leq 4$ it is a consequence of the reciprocity of $l^\ast_P(t)$ that for any lattice polytope $P$ either $P$ is thin or $\deg(l^\ast_P) = \deg(P)$. In higher dimensions, this property fails for non-Gorenstein lattice polytopes. \begin{ex} Consider the full-dimensional lattice simplex $P \subseteq \mathbb{R}^5$ given as the convex hull $P = \conv(0, e_1, e_2, e_3, (0,1,1,2,0), (5,3,3,2,6))$. Then $l_P^\ast(t) = 4t^3$ while $h^\ast_P(t) = t^4 + 5t^3 + 4t^2 + t + 1$. In particular, $P$ is not thin but $\deg(l^\ast_P) < \deg(h^\ast_P) = \deg(P)$. This is the only such example among lattice simplices of dimension $5$ with lattice volume $\leq 15$. It was found using the database \cite{gabriele}. The computations were performed in \texttt{SageMath} with backend \texttt{Normaliz}. \end{ex} \begin{ex} Consider the full-dimensional lattice simplex $P \subseteq \mathbb{R}^5$ given as the convex hull $P = \conv(0, e_1, e_2, (1, 1, 2, 0, 0), (3, 5, 6, 8, 0), (1, 1, 0, 0, 2))$. Then $l_P^\ast(t) = t^3$ while $h^\ast_P(t) = 7t^3 + 19t^2 + 5t + 1$. So $\deg(l_P^\ast) = \deg(h_P^\ast)$ but the leading coefficient of $l_P^\ast$ is strictly smaller. \end{ex} \subsection{Thinness is invariant under duality} It was noted in Lemma~\ref{lem:deg} that being trivially thin as well as being $g$-thin is invariant under duality of Gorenstein polytopes. Let us explain how this allows us to deduce that also thinness has this beautiful duality property: \begin{cor} Let $P$ be a Gorenstein polytope. Then $P$ is thin if and only if $P^\times$ is thin.\label{cor:dual} \end{cor} \begin{proof} By Theorem~\ref{thm:main}(ii) we may assume that $P$ is a thin Gorenstein polytope such that $P$ is a Gorenstein join of faces $F$ and $G$ with $F$ being trivally thin. Hence, by Lemma~\ref{lem:deg} we also have $P^\times = F^* \ast_{\Gor} G^*$ with $G^*$ being trivially thin. Again, by Theorem~\ref{thm:main} this implies that $P^\times$ is thin. \end{proof} According to Lev Borisov, this statement might also be proven using vertex algebra techniques. In particular, as Theorem~\ref{thm:main} implies that there are only two choices for the degree of the $l^*$-polynomial of a Gorenstein polytope we see that its degree is also invariant under duality (as it holds for the degrees of the $h^*$-polynomial and the $g$-polynomial). \begin{cor} Let $P$ be a Gorenstein polytope. Then $\deg(l^*_P) = \deg(l^*_{P^\times})$. \label{cor:dualdeg} \end{cor} \begin{ex} The reader should be aware that the local $h^*$-polynomials of a Gorenstein polytope $P$ and its dual $P^\times$ may differ. For instance, for $P=[-1,1]^3$ we have $l^*_P(t) = t + 17 t^2 + t^3$ and $l^*_{P^\times}(t)=t + 3 t^2 + t^3$. \end{ex} \subsection{Thin Gorenstein simplices} For the special case of Gorenstein simplices, we can answer the original question in \cite{GKZ94} about classifying thin simplices. \begin{cor} Let $P$ be a Gorenstein simplex. Then $P$ is thin if and only if $P$ is a lattice pyramid.\label{cor:gorsimplex} \end{cor} \begin{proof} Let $P$ be thin. If $P$ is $g$-thin, then $\deg(P)=\deg(g_P)=0$ as $P$ is a simplex. Hence, $P$ is a unimodular simplex, in particular, a lattice pyramid. Otherwise, Theorem~\ref{thm:main}(iii) implies that there are faces $F$ and $G$ of $P$ such that $P=F \ast_{\Gor} G$ with $G$ $g$-thin. As $G$ is also a simplex, the previous consideration shows that $G$ is a unimodular simplex, thus, a lattice pyramid. Hence, Corollary~\ref{cor:pyr} implies that $P$ is also a lattice pyramid. \end{proof} In particular, if a Gorenstein simplex $P$ satisfies $\deg(P) \ge 2 \deg(P)$, then $P$ is a lattice pyramid. This statement can also be deduced from \cite[Cor.~3.10(2)]{Adjunction}. \medskip \bibliographystyle{alpha}

https://arxiv.org/abs/math/0602191

On the maximum number of cliques in a graph

A \emph{clique} is a set of pairwise adjacent vertices in a graph. We determine the maximum number of cliques in a graph for the following graph classes: (1) graphs with $n$ vertices and $m$ edges; (2) graphs with $n$ vertices, $m$ edges, and maximum degree $\Delta$; (3) $d$-degenerate graphs with $n$ vertices and $m$ edges; (4) planar graphs with $n$ vertices and $m$ edges; and (5) graphs with $n$ vertices and no $K_5$-minor or no $K_{3,3}$-minor. For example, the maximum number of cliques in a planar graph with $n$ vertices is $8(n-2)$.

\section{Introduction} \seclabel{Intro} The typical question of extremal graph theory asks for the maximum number of edges in a graph in a certain family; see the surveys \citep{SimonSos-DM01, Simonovits97, Bollobas95, Simonovits83}. For example, a celebrated theorem of \citet{Turan41} states that the maximum number of edges in a graph with $n$ vertices and no $(k+1)$-clique is $\ensuremath{\protect\tfrac{1}{2}}(1-\frac{1}{k})n^2$. Here a \emph{clique} is a (possibly empty) set of pairwise adjacent vertices in a graph. For $k\geq0$, a \emph{$k$-clique} is a clique of cardinality $k$. Since an edge is nothing but a $2$-clique, it is natural to consider the maximum number of $\ell$-cliques in a graph. The following generalisation of Tur\'{a}n's Theorem, first proved by \citet{Zykov49}, has been rediscovered and itself generalised by several authors \citep{FisherRyan-DM92, HadNen77, PR00, NenKhad75, KN75, Had77, Roman-DM76, Erdos-CPM69, Erdos62, Sauer-JCTB71}. \begin{theorem}[\citep{Zykov49}] \thmlabel{Zykov} For all integers $k\geq\ell\geq 0$, the maximum number of $\ell$-cliques in a graph with $n$ vertices and no $(k+1)$-clique is $\binom{k}{\ell}\bracket{\frac{n}{k}}^{\ell}$. \end{theorem} A simple inductive proof of \thmref{Zykov} is included in \appref{BoundedClique}. In this paper we determine the maximum number of cliques in a graph in each of the following classes: \begin{itemize} \item graphs with $n$ vertices and $m$ edges (\secref{General}), \item graphs with $n$ vertices, $m$ edges, and maximum degree $\Delta$ (\secref{Degree}), \item $d$-degenerate graphs with $n$ vertices and $m$ edges (\secref{Degenerate}), \item planar graphs with $n$ vertices and $m$ edges (\secref{Planar}), and \item graphs with $n$ vertices and no $K_5$-minor or no $K_{3,3}$-minor (\secref{K5Minor}). \end{itemize} We now review some related work from the literature. \citet{Eckhoff-DM04, Eckhoff-DM99} determined the maximum number of cliques in a graph with $m$ edges and no $(k+1)$-clique. Lower bounds on the number of cliques in a graph have also been obtained \citep{FisherNonis90, Moon-CMB65, Larman69, LovSim76, LovSim83, Fisher-JGT89, BolErdSze-DM75}. The number of cliques in a random graph has been studied \citep{OT-MS97, Schurger-PMH79, BolErd76}. Bounds on the number of cliques in a graph have recently been applied in the analysis of an algorithm for finding small separators \citep{ReedWood-EuroComb05} and in the enumeration of minor-closed families \citep{NSTW-JCTB06}. \section{Preliminaries} \seclabel{Prelim} Every graph $G$ that we consider is undirected, finite, and simple. Let $V(G)$ and $E(G)$ be the vertex and edge sets of $G$. Let $\Delta(G)$ be the maximum degree of $G$. We say $G$ is a $(|V(G)|,|E(G)|)$-graph or a $(|V(G)|,|E(G)|,\Delta(G))$-graph. Let $C(G)$ be the set of cliques in $G$. Let $c(G):=|C(G)|$. Let $C_k(G)$ be the set of $k$-cliques in $G$. Let $c_k(G):=|C_k(G)|$. Our aim is to prove bounds on $c(G)$ and $c_k(G)$. A clique is not necessarily maximal\footnote{\citet{MoonMoser} proved that the maximum number of \emph{maximal} cliques in a graph with $n$ vertices is approximately $3^{n/3}$; see \citep{GKSV05, SV05, Wilf-SJDM86, HL74, Spencer71, Erdos-IJM66, JC00, FHT-Networks93, FL91, Sato-DM86} for related results.}. In particular, $\emptyset$ is a clique of every graph, $\{v\}$ is a clique for each vertex $v$, and each edge is a clique. Thus every graph $G$ satisfies \begin{equation} \eqnlabel{LowerBound} c(G)\geq c_0(G)+c_1(G)+c_2(G)=1+|V(G)|+|E(G)|\enspace. \end{equation} A \emph{triangle} is a $3$-clique. \Eqnref{LowerBound} implies that \begin{equation} \eqnlabel{TriangleFree} c(G)=1+|V(G)|+|E(G)|\text{ if and only if $G$ is triangle-free.} \end{equation} Triangle-free graphs have the fewest cliques. Obviously the complete graph $K_n$ has the most cliques for a graph on $n$ vertices. In particular, $c(K_n)=2^n$ since every set of vertices in $K_n$ is a clique. Say $v$ is a vertex of a graph $G$. Let $G_v$ be the subgraph of $G$ induced by the neighbours of $v$. Observe that $X$ is a clique of $G$ containing $v$ if and only if $X=Y\cup\{v\}$ for some clique $Y$ of $G_v$. Thus the number of cliques of $G$ that contain $v$ is exactly $c(G_v)$. Every clique of $G$ either contains $v$ or is a clique of $G\setminus v$. Thus $C(G)=C(G\setminus v)\cup\{Y\cup\{v\}:Y\in C(G_v)\}$ and \begin{equation} \eqnlabel{DeleteVertex} c(G)=c(G\setminus v)+c(G_v)\leq c(G\setminus v)+2^{\deg(v)}\enspace. \end{equation} Let $G$ be a graph with induced subgraphs $G_1$, $G_2$ and $S$ such that $G=G_1\cup G_2$ and $G_1\cap G_2=S$. Then $G$ is obtained by \emph{pasting} $G_1$ and $G_2$ on $S$. Observe that $C(G)=C(G_1)\cup C(G_2)$ and $C(G_1)\cap C(G_2)=C(S)$. Thus \begin{equation} \eqnlabel{Pasting} c(G)=c(G_1)+c(G_2)-c(S)\enspace. \end{equation} \begin{lemma} \lemlabel{PastingPasting} Let $G$ be an $(n,m)$-graph that is obtained by pasting $G_1$ and $G_2$ on $S$. Say $G_i$ has $n_i$ vertices and $m_i$ edges. Say $S$ has $n_S$ vertices and $m_S$ edges. If $c(G_i)\leq xn_i+ym_i+z$ and $c(S)\geq xn_s+ym_S+z$, then $c(G)\leq xn+ym+z$. \end{lemma} \begin{proof} By \Eqnref{Pasting}, \begin{align*} c(G) & = c(G_1)+c(G_2)-c(S)\\ & \leq (xn_1+ym_1+z)+(xn_2+ym_2+z)-(xn_s+ym_s+z)\\ & = x(n_1+n_2-n_S)+y(m_1+m_2-m_S)+z\\ & = xn+ym+z\enspace. \end{align*} \qed\end{proof} The following special case of \lemref{PastingPasting} will be useful. \begin{corollary} \corlabel{PastingClique} Let $G$ be an $(n,m)$-graph that is obtained by pasting $G_1$ and $G_2$ on a $k$-clique. Say $G_i$ has $n_i$ vertices and $m_i$ edges. Assume that $c(G_i)\leq xn_i+ym_i+z$ and that $xk+y\binom{k}{2}+z\leq 2^k$. Then $c(G)\leq xn+ym+z$.\qed \end{corollary} \section{General Graphs} \seclabel{General} We now determine the maximum number of cliques in an $(n,m)$-graph. \begin{theorem} \thmlabel{GeneralGraphs} Let $n$ and $m$ be non-negative integers such that $m\leq\binom{n}{2}$. Let $d$ and $\ell$ be the unique integers such that $m=\binom{d}{2}+\ell$ where $d\geq1$ and $0\leq\ell\leq d-1$. Then the maximum number of cliques in an $(n,m)$-graph equals $2^d+2^\ell+n-d-1$. \end{theorem} \begin{proof} First we prove the lower bound. Let $V(G):=\{v_1,v_2,\dots,v_n\}$ and $E(G):=\{v_iv_j:1\leq i<j\leq d\}\cup\{v_iv_{d+1}:1\leq i\leq\ell\}$, as illustrated in \figref{Example}. Then $G$ has $\binom{d}{2}+\ell$ edges. Now $\{v_1,v_2,\dots,v_d\}$ is a clique, which contains $2^d$ cliques (including $\emptyset$). The neighbourhood of $v_{d+1}$ is an $\ell$-clique with $2^{\ell}$ cliques. Thus there are $2^\ell$ cliques that contain $v_{d+1}$. Finally $v_{d+2},v_{d+3},\dots,v_n$ are isolated vertices, which contribute $n-d-1$ cliques to $G$. In total, $G$ has $2^d+2^\ell+n-d-1$ cliques. \Figure{Example}{\includegraphics{Example}}{A $(14,31)$-graph with $269$ cliques ($d=8$ and $\ell=3$).} Now we prove the upper bound. That is, every $(n,m)$-graph $G$ has at most $2^d+2^\ell+n-d-1$ cliques. We proceed by induction on $n+m$. For the base case, suppose that $m=0$. Then $d=1$, $\ell=0$, and $c(G)=n+1=2^d+2^\ell+n-d-1$. Now assume that $m\geq1$. Let $v$ be a vertex of minimum degree in $G$. Then $\deg(v)\leq d-1$, as otherwise every vertex has degree at least $d$, implying $m\geq\frac{dn}{2}\geq \frac{d(d+1)}{2}=\binom{d+1}{2}$, which contradicts the definition of $d$. By \Eqnref{DeleteVertex}, $c(G)\leq c(G\setminus v)+2^{\deg(v)}$. To apply induction to $G\setminus v$ (which has $n-1$ vertices and $m-\deg(v)$ edges) we distinguish two cases. First suppose that $\deg(v)\leq\ell$. Thus $m-\deg(v)=\binom{d}{2}+\ell-\deg(v)$. By induction, $c(G)\leq 2^d+2^{\ell-\deg(v)}+n-1-d-1+2^{\deg(v)}$. Hence the result follows if $2^d+2^{\ell-\deg(v)}+n-1-d-1+2^{\deg(v)}\leq2^d+2^\ell+n-d-1$. That is, $2^{\ell-\deg(v)}-1\leq(2^{\ell-\deg(v)}-1)2^{\deg(v)}$, which is true since $0\leq\deg(v)\leq\ell$. Otherwise $\ell+1\leq\deg(v)\leq d-1$. Thus $m-\deg(v)=\binom{d-1}{2}+d-1+\ell-\deg(v)$. By induction, $c(G)\leq2^{d-1}+2^{d-1+\ell-\deg(v)}+n-1-d+2^{\deg(v)}$. Hence the result follows if $2^{d-1}+2^{d-1+\ell-\deg(v)}+n-1-d+2^{\deg(v)}\leq2^d+2^\ell+n-d-1$. That is, $2^\ell(2^{\deg(v)-\ell}-1) \leq 2^{d-1-\deg(v)+\ell}(2^{\deg(v)-\ell}-1)$. Since $\deg(v)\geq\ell+1$, we need $2^\ell \leq 2^{d-1-\deg(v)+\ell}$, which is true since $\deg(v)\leq d-1$. \qed\end{proof} \section{Bounded Degree Graphs} \seclabel{Degree} We now determine the maximum number of cliques in an $(n,m,\Delta)$-graph. \citet{West84} proved a related result. \begin{theorem} \thmlabel{Degree} The number of cliques in an $(n,m,\Delta)$-graph $G$ is at most \begin{equation*} 1+n+\bracket{\frac{2^{\Delta+1}-\Delta-2}{\binom{\Delta+1}{2}}}m\leq 1+\bracket{\frac{2^{\Delta+1}-1}{\Delta+1}}n \enspace. \end{equation*} \end{theorem} \begin{proof} $G$ has one $0$-clique and $n$ $1$-cliques. For $k\geq2$, each edge is in at most $\binom{\Delta-1}{k-2}$ $k$-cliques, and each $k$-clique contains $\binom{k}{2}$ edges. Thus $G$ has at most $m\binom{\Delta-1}{k-2}/\binom{k}{2}$ $k$-cliques. Thus the number of cliques (not counting $0$- and $1$-cliques) is at most \begin{align*} \sum_{k=2}^{\Delta+1}\frac{m\binom{\Delta-1}{k-2}}{\binom{k}{2}} &= m\sum_{k=2}^{\Delta+1} \frac{2}{k(k-1)}\cdot\frac{(\Delta-1)!}{(k-2)!(\Delta-1-k+2)!}\\ &= \frac{m}{\binom{\Delta+1}{2}} \sum_{k\geq2}^{\Delta+1}\frac{2(\Delta-1)!\binom{\Delta+1}{2}}{k!(\Delta+1-k)!}\\ &= \frac{m}{\binom{\Delta+1}{2}} \sum_{k=2}^{\Delta+1}\frac{(\Delta+1)!}{k!(\Delta+1-k)!}\\ &= \frac{m}{\binom{\Delta+1}{2}} \bracket{ \bracket{\sum_{k=0}^{\Delta+1}\binom{\Delta+1}{k}} -\frac{(\Delta+1)!}{1!(\Delta+1-1)!} -\frac{(\Delta+1)!}{0!(\Delta+1-0)!} }\\ &= \frac{m}{\binom{\Delta+1}{2}} \bracket{2^{\Delta+1}-\Delta-2}\enspace. \end{align*} The result follows since $m\leq\frac{\Delta n}{2}$. \qed\end{proof} The bound in \thmref{Degree} is tight for many values of $m$. \begin{proposition} \proplabel{DegreeLowerBound} For all $n$ and $m$ such that $m\leq\frac{\Delta n}{2}$ and $m\equiv0\pmod{\binom{\Delta+1}{2}}$, there is an $(n,m,\Delta)$-graph $G$ with \begin{equation*} c(G)=1+n+\bracket{\frac{2^{\Delta+1}-\Delta-2}{\binom{\Delta+1}{2}}}m\enspace. \end{equation*} \end{proposition} \begin{proof} Let $p:=m/\binom{\Delta+1}{2}$. Let $G$ consist of $p$ copies of $K_{\Delta+1}$, plus $n-p(\Delta+1)$ isolated vertices. Then $G$ is an $(n,m,\Delta)$-graph. Each copy of $K_{\Delta+1}$ contributes $2^{\Delta+1}-\Delta-2$ cliques with at least two vertices. Thus $G$ has $1+n+(2^{\Delta+1}-\Delta-2)p$ cliques. \qed\end{proof} \section{Degenerate Graphs} \seclabel{Degenerate} A graph $G$ is $d$-degenerate if every subgraph of $G$ has a vertex with degree at most $d$. The following simple result is well known; see \citep{ReedWood-EuroComb05, Eppstein-JGT93} for example. \begin{proposition} \proplabel{Degen} Every $d$-degenerate graph $G$ with $n\geq d$ vertices has at most $2^d(n-d+1)$ cliques. \end{proposition} \begin{proof} We proceed by induction on $n$. If $n=d$ then $c(G)\leq2^d=2^d(n-d+1)$. Now assume that $n\geq d+1$. Let $v$ be a vertex of $G$ with $\deg(v)\leq d$. By \Eqnref{DeleteVertex}, $c(G)\leq c(G\setminus v)+2^{\deg(v)}$. Now $G\setminus v$ is $d$-degenerate since it is a subgraph of $G$. Moreover, $G\setminus v$ has at least $d$ vertices. By induction, $c(G\setminus v)\leq 2^d(n-1-d+1)$. Thus $c(G)\leq2^d(n-1-d+1)+2^d=2^d(n-d+1)$. \qed\end{proof} The bound in \propref{Degen} is tight. \begin{proposition} \proplabel{DegenLowerBound} For all $n\geq d$, there is a $d$-degenerate graph $G_n$ with $n$ vertices and exactly $2^d(n-d+1)$ cliques \textup{(}and with a $d$-clique\textup{)}. \end{proposition} \begin{proof} Let $G_d$ be the complete graph $K_d$. Then $G_d$ has the desired properties. For $n\geq d+1$, let $G_n$ be the graph obtained by adding one new vertex $v$ adjacent to every vertex in some $d$-clique in $G_{n-1}$. Then $G_n$ is $d$-degenerate and contains a $d$-clique. ($G_n$ is a chordal graph called a \emph{$d$-tree}; see \citep{Bodlaender-TCS98}.)\ By \Eqnref{DeleteVertex}, $c(G_n)=c(G_{n-1})+2^{\deg(v)}=2^d(n-1-d+1)+2^d=2^d(n-d+1)$. \qed\end{proof} \propref{Degen} can be made sensitive to the number of edges as follows. \begin{theorem} \thmlabel{SensitiveDegen} For all $d\geq1$, every $d$-degenerate graph $G$ with $n$ vertices and $m\geq\binom{d}{2}$ edges has at most \begin{equation*} n+\frac{(2^d-1)m}{d}-\frac{(d-3)2^d+d+1}{2} \end{equation*} cliques. \end{theorem} \begin{proof} We proceed by induction on $n+m$. For the base case, suppose that $m=\binom{d}{2}+\ell$ where $d\geq1$ and $0\leq\ell\leq d-1$. Thus $c(G)\leq 2^d+2^\ell+n-d-1$ by \thmref{GeneralGraphs}, and the result follows if \begin{equation*} 2^d+2^\ell+n-d-1\leq n+\frac{(2^d-1)m}{d}-\frac{(d-3)2^d+d+1}{2}\enspace. \end{equation*} That is, $d(2^\ell-1)\leq\ell(2^d-1)$, which we prove in \lemref{Ineq} below. Now assume that $m\geq\binom{d+1}{2}$. Now $G$ has a vertex $v$ with $\deg(v)\leq d$. By \Eqnref{DeleteVertex}, $c(G)\leq c(G\setminus v)+2^{\deg(v)}$. The graph $G\setminus v$ has $m-\deg(v)\geq\binom{d}{2}$ edges, and is $d$-degenerate since it is a subgraph of $G$. By induction, \begin{equation*}c(G\setminus v)\leq n-1+\frac{(2^d-1)(m-\deg(v))}{d}-\frac{(d-3)2^d+d+1}{2}\enspace. \end{equation*} Thus the result follows if \begin{equation*}-1+\frac{(2^d-1)(m-\deg(v))}{d}+2^{\deg(v)}\leq \frac{(2^d-1)m}{d}\enspace.\end{equation*} That is, $d(2^{\deg(v)}-1)\leq (2^d-1)\deg(v)$, which holds by \lemref{Ineq} below. \qed\end{proof} \begin{lemma} \lemlabel{Ineq} $d(2^\ell-1)\leq \ell(2^d-1)$ for all integers $d\geq\ell\geq0$. \end{lemma} \begin{proof} The case $\ell=0$ is trivial. Now assume that $\ell\geq1$. We proceed by induction on $d$. The base case $d=\ell$ is trivial. Assume that $d\geq\ell+1\geq2$ and by induction, \begin{equation} \eqnlabel{AAA} (d-1)(2^\ell-1)\leq\ell(2^{d-1}-1). \end{equation} Since $d\geq2$, \begin{equation} \eqnlabel{BBB} \frac{d}{d-1}\leq 2<2+\frac{1}{2^{d-1}-1}=\frac{2^d-1}{2^{d-1}-1}. \end{equation} Equations \eqnref{AAA} and \eqnref{BBB} imply that \begin{equation*} (d-1)(2^\ell-1)\cdot\frac{d}{d-1}<\ell(2^{d-1}-1)\cdot\frac{2^d-1}{2^{d-1}-1}. \end{equation*} That is, $d(2^\ell-1)<\ell(2^d-1)$, as desired. \qed\end{proof} Note that a $d$-degenerate $n$-vertex graph has at most $dn-\binom{d+1}{2}$ edges, and \thmref{SensitiveDegen} with $m=dn-\binom{d+1}{2}$ is equivalent to \propref{Degen}. The bound in \thmref{SensitiveDegen} is tight for many values of $m$. \begin{proposition} \proplabel{SensitiveDegenLowerBound} Let $d\geq1$. For all $n$ and $m$ such that $\binom{d}{2}\leq m\leq dn-\binom{d+1}{2}$ and \begin{equation*} m\bmod{d} \begin{cases} 0 & \text{if }d\text{ is odd} \\ \frac{d}{2} & \text{if }d\text{ is even ,} \end{cases} \end{equation*} there is a $d$-degenerate $(n,m)$-graph $G$ with \begin{equation*} c(G)=n+\frac{(2^d-1)m}{d}-\frac{(d-3)2^d+d+1}{2}\enspace. \end{equation*} \end{proposition} \begin{proof} Let $n':=\frac{m}{d}+\ensuremath{\protect\tfrac{1}{2}}(d+1)$. Then $n'$ is an integer and $d\leq n'\leq n$. Let $G$ consist of a $d$-degenerate $n'$-vertex graph with $2^d(n'-d+1)$ cliques (from \propref{DegenLowerBound}), plus $n-n'$ isolated vertices. Then $G$ has $m$ edges and $c(G)=2^d(n'-d+1)+n-n'=n+(2^d-1)\frac{m}{d}-\ensuremath{\protect\tfrac{1}{2}}((d-3)2^d+d+1)$. \qed\end{proof} A graph is $1$-degenerate if and only if it is a forest. Thus \thmref{SensitiveDegen} with $d=1$ implies that every forest has at most $n+m-1$ cliques, which also follows from \Eqnref{TriangleFree}. In particular, $c(T)=2n$ for every $n$-vertex tree $T$. \thmref{SensitiveDegen} with $d=2$ implies that every $2$-degenerate graph has at most $n+\ensuremath{\protect\tfrac{1}{2}}(3m+1)$ cliques. Outerplanar graphs are $2$-degenerate. The construction in \twopropref{DegenLowerBound}{SensitiveDegenLowerBound} can produce outerplanar graphs. (Add each new vertex adjacent to two consecutive vertices on the outerface.)\ Thus this bound is tight for outerplanar graphs. \section{Planar Graphs} \seclabel{Planar} \citet{PY-IPL81} and \citet{Storch-GECCO06} proved that every $n$-vertex planar graph has \Oh{n} cliques; see \citep{Eppstein-JGT93} for a more general result. The proof is based on the corollary of Euler's Formula that planar graphs are $5$-degenerate. By \thmref{SensitiveDegen}, if $G$ is a planar $(n,m)$-graph with $m\geq 10$, then $c(G)<n+\frac{31}{5}m<\frac{98}{5}n$. We now prove that the bound for $3$-degenerate graphs in \thmref{SensitiveDegen} also holds for planar graphs. \begin{theorem} \thmlabel{PlanarGraphs} Every planar $(n,m)$-graph $G$ with $m\geq3$ has at most $n+\frac{7}{3}m-2$ cliques. \end{theorem} \begin{proof} We proceed by induction on $n+m$. The result is easily verified if $m=3$. Suppose that $G$ has a separating triangle $T$. Thus $G$ is obtained by pasting two induced subgraphs $G_1$ and $G_2$ on $T$. Say $G_i$ has $n_i$ vertices and $m_i$ edges. Then $m_i\geq3$ since $T\subset G_i$. By induction, $c(G_i)\leq n_i+\frac{7}{3}m_i-2$. By \corref{PastingClique} with $k=3$, $x=1$, $y=\frac{7}{3}$ and $z=-2$, we have $c(G)\leq n+\frac{7}{3}m-2$ (since $1\cdot3+\frac{7}{3}\binom{3}{2}-2=2^3$). Now assume that $G$ has no separating triangle. Let $v$ be a vertex of $G$. We have $c(G)=c(G\setminus v)+c(G_v)$ by \Eqnref{DeleteVertex}. The graph $G\setminus v$ has $m-\deg(v)$ edges. Suppose that $m-\deg(v)\leq 2$. (Then we cannot apply induction to $G\setminus v$.)\ Then $G$ has no $4$-clique and at most two triangles. If $G$ has at most one triangle, then $c(G)\leq 1+n+m+1\leq n+\frac{7}{3}m-2$ since $m\geq3$. Otherwise $G$ has two triangles, and $c(G)\leq 1+n+m+2<n+\frac{7}{3}m-2$ since $m\geq5$. Now assume that $m-\deg(v)\geq3$. By \eqnref{DeleteVertex}, applying induction to $G\setminus v$, \begin{equation*} c(G)=c(G\setminus v)+c(G_v) \leq(n-1)+\tfrac{7}{3}(m-\deg(v))-2+c(G_v)\enspace. \end{equation*} Fix a plane embedding of $G$. If $uw$ is an edge of $G_v$, then the edges $vu$ and $vw$ are consecutive in the circular ordering of edges incident to $v$ defined by the embedding (as otherwise $G$ would contain a separating triangle). Thus $\Delta(G_v)\leq 2$ and $c(G_v)\leq1+\frac{7}{3}\deg(v)$ by \thmref{Degree}. Hence \begin{equation*} c(G)\leq(n-1)+(\tfrac{7}{3}(m-\deg(v))-2)+(1+\tfrac{7}{3}\deg(v)) =n+\tfrac{7}{3}m-2\enspace. \end{equation*} \qed\end{proof} If $n\geq3$ in \thmref{PlanarGraphs} then $m\leq3(n-2)$ by Euler's Formula. Thus we have the following corollary. \begin{corollary} \corlabel{MaximalPlanarGraphs} Every planar graph with $n\geq3$ vertices has at most $8(n-2)$ cliques.\qed \end{corollary} We now prove bounds on the number of $3$- and $4$-cliques in a planar graph. \begin{proposition} \proplabel{PlanarTrianglesSquares} For every planar graph $G$ with $n\geq3$ vertices, $c_3(G)\leq 3n-8$ and $c_4(G)\leq n-3$. \end{proposition} \begin{proof} We proceed by induction on $n$. The result is trivial if $n\leq 4$. Now assume that $n\geq5$. First suppose that $G$ has no separating triangle. Then $c_4(G)=0$, and every triangle of $G$ is a face. By Euler's Formula, $c_3(G)\leq 2n-4<3n-8$ faces. Now suppose that $G$ has a separating triangle $T$. Thus $G$ is obtained by pasting two induced subgraphs $G_1$ and $G_2$ on $T$. Say $G_i$ has $n_i$ vertices. Then $n_i\geq3$ since $T\subset G_i$. By induction, $c_3(G_i)\leq 3n_i-8$ and $c_4(G_i)\leq n_i-3$. Every clique of $G$ is a clique of $G_1$ or $G_2$. Thus $c_4(G)=c_4(G_1)+c_4(G_2)\leq n_1-3+n_2-3=n-3$. Moreover, $T$ is a triangle in both $G_1$ and $G_2$. Thus $c_3(G)\leq (3n_1-8)+(3n_2-8)-1=3(n_1+n_2)-17=3(n+3)-17=3n-8$. \qed\end{proof} Note that \propref{PlanarTrianglesSquares} and Euler's Formula (which implies $c_2(G)\leq 3n-6$) reprove \corref{MaximalPlanarGraphs}, since $1+n+3(n-2)+(3n-8)+(n-3)=8(n-2)$. We now show that all our bounds for planar graphs are tight. \begin{proposition} \proplabel{MaximalPlanarLowerBound} For all $n\geq3$ there is a maximal planar $n$-vertex graph $G_n$ with $c_2(G_n)=3(n-2)$, $c_3(G_n)=3n-8$, $c_4(G_n)=n-3$, and $c(G_n)=8(n-2)$. \end{proposition} \begin{proof} Let $G_3:=K_3$. Then $c_2(G_3)=3$, $c_3(G_3)=1$, $c_4(G_3)=0$, and $c(G_3)=8$. Say $G_{n-1}$ is a maximal planar $(n-1)$-vertex graph with $c_2(G_{n-1})=3(n-3)$, $c_3(G_{n-1})=3n-11$, $c_4(G_{n-1})=n-4$, and $c(G_n)=8(n-3)$. Let $G_n$ be the maximal planar $n$-vertex graph obtained by adding one new vertex $v$ adjacent to each vertex of some face of $G_{n-1}$, as illustrated in \figref{Stellated}. Then $c_2(G_n)=c_2(G_{n-1})+3=3(n-2)$, $c_3(G_n)=c_3(G_{n-1})+3=3n-8$, $c_4(G_n)=c_4(G_{n-1})+1=n-3$, and $c(G_n)=c(G_{n-1})+c(G_n(v))=8(n-3)+8=8(n-2)$. (Note that $G_n$ is also an example of a $3$-degenerate graph with the maximum number of cliques; see \propref{DegenLowerBound}.)\ \qed\end{proof} \Figure{Stellated}{\includegraphics[width=\textwidth]{StellatedTriangle6}}{A planar graph with $124$ vertices, $366$ edges, $364$ triangles, $121$ $4$-cliques, and $976$ cliques. It is obtained by repeatedly adding one degree-$3$ vertex inside each internal face (starting from $K_3$).} \begin{proposition} \proplabel{PlanarLowerBound} For all $n\geq3$ and $m\in\{3,6,\dots,3n-6\}$, there is a planar $(n,m)$-graph $G$ with $c(G)=n+\frac{7}{3}m-2$. \end{proposition} \begin{proof} Let $n':=\frac{m}{3}+2$. Let $G$ consist of a maximal planar graph on $n'$ vertices with $8(n'-2)$ cliques (from \propref{MaximalPlanarLowerBound}), plus $n-n'$ isolated vertices. Then $G$ has $n$ vertices and $m$ edges, and $c(G)=8(n'-2)+n-n'=n+7n'-16=n+7(\frac{m}{3}+2)-16=n+\frac{7}{3}m-2$. \qed\end{proof} \section{Graphs with no $K_5$-Minor} \seclabel{K5Minor} A graph $H$ is a \emph{minor} of a graph $G$ if $H$ can be obtained from a subgraph of $G$ by contracting edges. The graphs with no $K_3$-minor are the forests, which have at most $2n$ cliques, and this bound is tight. The graphs with no $K_4$-minor (called \emph{series-parallel}) are $2$-degenerate, and thus have at most $4(n-1)$ cliques, and this bound is tight. The Kuratowski-Wagner Theorem characterises planar graphs as those with no $K_5$-minor and no $K_{3,3}$-minor. We now extend \corref{MaximalPlanarGraphs} for graphs with no $K_5$-minor (but possibly a $K_{3,3}$-minor). \begin{theorem} \thmlabel{NoK5Minor} Every graph $G$ with $n\geq3$ vertices and no $K_5$-minor has at most $8(n-2)$ cliques. \end{theorem} \begin{proof} Let $V_8$ be the graph obtained from the $8$-cycle by adding an edge between each pair of antipodal vertices; see \figref{W}. Let $G$ be a minimum counterexample to the theorem. We can assume that $G$ is edge-maximal with no $K_5$-minor. \citet{Wagner37} proved that (a) $G$ is a maximal planar graph, (b) $G=V_8$, or (c) $G$ is obtained by pasting two smaller graphs (that are thus not counterexamples), each with no $K_5$-minor, on an edge or a triangle $T$. In case (a) the result is \corref{MaximalPlanarGraphs}. In case (b), since $V_8$ is triangle-free, $c(V_8)=1+|V(V_8)|+|E(V_8)|=21<8(|V(V_8)|-2)$ by \Eqnref{TriangleFree}. In case (c), if $T$ is an edge, we have $c(G)\leq 8(n-2)$ by \corref{PastingClique} with $k=2$, $x=8$, $y=0$ and $z=-16$ (since $8\cdot2+0-16<2^2$). In case (c), if $T$ is a triangle, we have $c(G)\leq 8(n-2)$ by \corref{PastingClique} with $k=3$, $x=8$, $y=0$ and $z=-16$ (since $8\cdot3+0-16=2^3$). \qed\end{proof} \Figure{W}{\includegraphics{W}}{The graph $V_8$.} A similar result is obtained for graphs with no $K_{3,3}$-minor. \begin{theorem} \thmlabel{NoK33Minor} Every graph $G$ with $n\geq3$ vertices and no $K_{3,3}$-minor has at most $\frac{4}{3}(7n-11)$ cliques. Conversely, for all $n\equiv 2\pmod{3}$ with $n\geq5$ there is an $n$-vertex graph with no $K_{3,3}$-minor and $c(G)=\frac{4}{3}(7n-11)$. \end{theorem} \begin{proof} Let $G$ be a minimum counterexample. We can assume that $G$ is edge-maximal with no $K_{3,3}$-minor. \citet{Wagner37} proved that (a) $G$ is a maximal planar graph, (b) $G=K_5$, or (c) $G$ is obtained by pasting two smaller graphs (that are thus not counterexamples), each with no $K_{3,3}$-minor, on an edge. In case (a) the result follows from \corref{MaximalPlanarGraphs} since $8n-16<\frac{4}{3}(7n-11)$. In case (b), $c(K_5)=32=\frac{4}{3}(7\cdot 5-11)$. In case (c), we have $c(G)\leq \frac{4}{3}(7n-11)$ by \corref{PastingClique} with $k=2$, $x=\frac{28}{3}$, $y=0$ and $z=-\frac{44}{3}$ (since $\frac{28}{3}\cdot2+0-\frac{44}{3}=2^2$). By the same analysis, the graph obtained from $K_5$ by repeatedly pasting copies of $K_5$ on an edge has no $K_{3,3}$-minor and $\frac{4}{3}(7n-11)$ cliques. \qed\end{proof} We finish with an open problem: What is the maximum number of cliques in an $n$-vertex graph $G$ with no $K_t$-minor? \citet{Kostochka82} and \citet{Thomason84} independently proved that $G$ is \Oh{t\sqrt{\log t}}-degenerate\footnote{Moreover, this bound is best possible; \citet{Thomason01} even determined the asymptotic constant.}. Thus \propref{Degen} implies that $G$ has at most $2^{\Oh{t\sqrt{\log t}}}n$ cliques; similar bounds can be found in \citep{NSTW-JCTB06, ReedWood-EuroComb05}. It is unknown whether this bound can be improved to $c^tn$ for some constant $c$ (possibly for sufficiently large $n$). We have proved that $c(G)\leq 2^{t-2}(n-t+3)$ whenever $t\leq5$. Moreover, the graph $G$ in \propref{DegenLowerBound} (with $t=d+2$) has no $K_t$-minor and $c(G)=2^{t-2}(n-t+3)$. However, for large values of $t$ this upper bound does not hold for the complete $k$-partite graph $K_{2,2,\dots,2}$. By \thmref{HadwigerCompletePartite} in \appref{Hadwiger}, the maximum order of a clique minor in $K_{2,2,\dots,2}$ is $\floor{\frac{3}{2}k}$. But by \propref{CompletePartite}, $c(K_{2,2,\dots,2})=3^k>2^{\floor{3k/2}-1}(2k-\floor{\frac32k}+2)$ for all $k\geq 42$. \begin{acknowledgement} Thanks to a referee for pointing out reference \citep{Stiebitz-DM92}. \end{acknowledgement} \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} \def$'${$'$} \def\leavevmode\lower.6ex\hbox to 0pt{\hskip-.23ex \accent"16\hss}D{\leavevmode\lower.6ex\hbox to 0pt{\hskip-.23ex \accent"16\hss}D}

https://arxiv.org/abs/1911.06619

Monotone Sobolev functions in planar domains: level sets and smooth approximation

We prove that almost every level set of a Sobolev function in a planar domain consists of points, Jordan curves, or homeomorphic copies of an interval. For monotone Sobolev functions in the plane we have the stronger conclusion that almost every level set is an embedded $1$-dimensional topological submanifold of the plane. Here monotonicity is in the sense of Lebesgue: the maximum and minimum of the function in an open set are attained at the boundary. Our result is an analog of Sard's theorem, which asserts that for a $C^2$-smooth function in a planar domain almost every value is a regular value. As an application we show that monotone Sobolev functions in planar domains can be approximated uniformly and in the Sobolev norm by smooth monotone functions.

\section{Introduction}\label{Section:Introduction} The classical theorem of Sard \cite{Sard:Theorem} asserts that for a $C^2$-smooth function $f$ in a planar domain $\Omega$ almost every value is a regular value. That is, for almost all $t\in \mathbb R$ the set $f^{-1}(t)$ does not intersect the critical set of $f$, and hence, $f^{-1}(t)$ is an embedded $1$-dimensional $C^2$-smooth submanifold of the plane. This theorem is sharp, in the sense that the $C^2$-smoothness cannot be relaxed to $C^1$-smoothness, as was shown by Whitney \cite{Whitney:SardOptimal}. In fact, Sard's theorem and some of the other theorems that we quote below have more general statements that hold for maps defined in subsets of $\mathbb R^n$, taking values in $\mathbb R^m$, and having appropriate regularity. In order to facilitate the comparison to our results, we will only give formulations in the case of real-valued functions defined in planar domains. Several generalizations and improvements of Sard's theorem have been proved since the original theorem was published. In particular, Dubovitskii \cite{Dubovitskii:Sard} proved that a $C^1$-smooth function $f$ in a planar domain has the property that for a.e.\ value $t\in \mathbb R$ the set $f^{-1}(t)$ intersects the critical set in a set of Hausdorff $1$-measure zero. De Pascale \cite{DePascale:Sard} extended the conclusion of Sard's theorem to Sobolev functions of the class $W^{2,p}$, where $p>2$. For other versions of Sard's theorem in the setting of H\"older and Sobolev spaces see \cite{Bates:Sard}, \cite{BojarskiHajlaszStrzelecki:Sard}, \cite{Figalli:Sard}, \cite{Norton:Sard}. Now, we turn our attention to the structure of the level sets of functions, instead of discussing the critical set. Theorem 1.6 in \cite{BojarskiHajlaszStrzelecki:Sard} states that if $f\in W^{1,p}_{\loc}(\mathbb R^2)$, then there exists a Borel representative of $f$ such that for a.e.\ $t\in \mathbb R$ the level set $f^{-1}(t)$ is equal to $Z\cup \bigcup_{j\in \mathbb N} K_j$, where $\mathcal H^1(Z)=0$, $K_j\subset K_{j+1}$, and $K_j$ is \textit{contained} in an $1$-dimensional $C^1$-smooth submanifold $S_j$ of $\mathbb R^2$ for $j\in \mathbb N$. Under increased regularity, Bourgain, Korobkov, and Kristensen \cite{BourgainEtAl:Sard} proved that if $f\in W^{2,1}(\mathbb R^2)$ then for a.e.\ $t\in \mathbb R$ the level set $f^{-1}(t)$ is an $1$-dimensional $C^1$-smooth manifold. We direct the reader to \cite{BourgainEtAl:Sard} and the references therein for a more detailed exposition of generalizations of Sard's theorem. Our first theorem is the following: \begin{theorem}\label{Intro:Sobolev:Theorem} Let $\Omega\subset \mathbb R^2$ be an open set and $u\colon \Omega \to \mathbb R$ be a continuous function that lies in $W^{1,p}_{\loc}(\Omega)$ for some $1\leq p\leq \infty$. Then for a.e.\ $t\in \mathbb R$ the level set $u^{-1}(t)$ has locally finite Hausdorff $1$-measure and each component of $u^{-1}(t)$ is either a point, or a Jordan curve, or it is homeomorphic to an interval. \end{theorem} This result generalizes a result of Alberti, Bianchini, and Crippa \cite[Theorem 2.5(iv)]{AlbertiEtAl:LipschitzLevelSets}, who obtained the same conclusion under the weaker assumptions that $u$ is Lipschitz and compactly supported. Under no further regularity assumptions, we do not expect in this case the level sets to be $1$-dimensional manifolds. Instead, we pose some topological restriction: \begin{definition}\label{Monotone:Definition} Let $\Omega\subset \mathbb R^n$ be an open set and $f\colon \Omega\to \mathbb R$ be a continuous function. We say that $f$ is monotone if for each open set $U\subset \subset \Omega$ the maximum and the minimum of $f$ on $\overline U$ are attained on $\partial U$. That is, \begin{align*} \max_{ \overline U}f= \max_{\partial U}f \,\,\, \textrm{and} \,\,\, \min_{\overline U}f=\min_{\partial U}f. \end{align*} \end{definition} Here the notation $U\subset \subset \Omega$ means that $\overline U$ is compact and is contained in $\Omega$. This definition can also be extended to real-valued functions defined in a locally compact metric space $X$. \begin{remark}\label{Monotone:Remark} If $f$ extends continuously to $\overline \Omega$, then we may take $U\subset\subset \overline \Omega$ in the above definition. \end{remark} Monotonicity in Definition \ref{Monotone:Definition} is in the sense of Lebesgue \cite{Lebesgue:Monotone}. There are other more general notions of monotonicity; e.g.\ there is a notion of \textit{weak monotonicity} due to Manfredi \cite{Manfredi:Monotone} that agrees with Lebesgue monotonicity for the spaces $W^{1,p}$, $p\geq 2$. We now state our main theorem: \begin{theorem}\label{Intro:MonotoneSobolev:Theorem} Let $\Omega\subset \mathbb R^2$ be an open set and $u\colon \Omega \to \mathbb R$ be a continuous monotone function that lies in $W^{1,p}_{\loc}(\Omega)$ for some $1\leq p\leq \infty$. Then for a.e.\ $t\in \mathbb R$ the level set $u^{-1}(t)$ has locally finite Hausdorff $1$-measure and it is an embedded $1$-dimensional topological submanifold of $\mathbb R^2$. \end{theorem} Recall that a set $J\subset \mathbb R^2$ is an embedded $1$-dimensional sumbanifold of $\mathbb R^2$ if for each point $x\in J$ there exists an open set $U$ in $\mathbb R^2$ and a homeomorphism $\phi\colon U\to \mathbb R^2$ such that $\phi(J)=\mathbb R$. By the classification of $1$-manifolds (see Theorem \ref{Classification:theorem}), each component of $J$ is homeomorphic to $S^1$ or to $\mathbb R$. Monotone Sobolev functions appear as infimizers of certain energy functionals, not only in the Euclidean setting, such as $\mathcal A$-harmonic functions \cite{Heinonenetal:DegenerateElliptic}, but also in metric spaces. For example, they appear as real and imaginary parts of ``uniformizing" quasiconformal mappings into the plane from metric spaces $X$ that are homeomorphic to the plane; see \cite{Rajala:uniformization}, \cite{RajalaRomney:reciprocal}. Our proof is partly topological and can be applied also in these settings, leading to the understanding of the level sets of the ``uniformizing" map, which is crucial for proving injectivity properties. The results in this paper can be used to simplify some of the arguments in \cite{Rajala:uniformization}. More specifically, our techniques yield the following result in the metric space setting. \begin{theorem}\label{Intro:metric:theorem} Let $(X,d)$ be a metric space homeomorphic to $\mathbb R^2$ and $\Omega$ be an open subset of $X$. Suppose that $u\colon \Omega \to \mathbb R$ is a continuous function with the property that for a.e.\ $t\in \mathbb R$ the level set $u^{-1}(t)$ has locally finite Hausdorff $1$-measure (in the metric of $X$). Then for a.e.\ $t\in \mathbb R$ each component of the level set $u^{-1}(t)$ is either a point, or a Jordan curve, or it is homeomorphic to an interval. If, in addition, $u$ is monotone, then for a.e.\ $t\in \mathbb R$ the level set $u^{-1}(t)$ is an embedded $1$-dimensional topological submanifold of $X$. \end{theorem} Monotone Sobolev functions enjoy some important further regularity properties. For example, if $f=(u_1,\dots,u_n)\colon \mathbb R^n\to \mathbb R^n$ is continuous and monotone, in the sense that the coordinate functions $u_i$ are, and $f\in W^{1,n}_{\loc}(\mathbb R^n)$, then $f$ is differentiable a.e.\ and it has the \textit{Luzin property}, i.e, it maps null sets to null sets; see \cite[Lemma 1.3]{HajlaszMaly:approximation}. For this reason, the approximation of Sobolev functions $u$ in the Sobolev norm by {locally weakly monotone} Sobolev functions $u_n$, $n\in \mathbb N$, has been established in \cite[Theorem 1.3]{HajlaszMaly:approximation}; a property of the approximants $u_n$ in this theorem that is important in applications is that the gradient of $u_n$ vanishes on the set $\{u\neq u_n\}$. As it is pointed out in that paper, in some cases the approximating functions cannot be taken to be smooth, not even continuous. As an application of Theorem \ref{Intro:MonotoneSobolev:Theorem} we give the following approximation theorem: \begin{theorem}\label{Intro:Approximation:Theorem} Let $\Omega\subset \mathbb R^2$ be a bounded open set and $u \colon \Omega\to \mathbb R$ be a continuous monotone function that lies in $W^{1,p}(\Omega)$ for some $1< p <\infty$. Then there exists $\alpha>0$ and a sequence $\{u_n\}_{n\in \mathbb N}$ of monotone functions in $\Omega$ such that \begin{enumerate}[\upshape(A)] \item $u_n$ is $C^{1,\alpha}$-smooth in $\Omega$, $n\in \mathbb N$, \item $u_n$ converges uniformly to $u$ in $\Omega$ as $n\to\infty$, \item $u_n$ converges to $ u$ in $W^{1,p}(\Omega)$ as $n\to\infty$, \item $\|\nabla u_n\|_{L^p(\Omega)}\leq \|\nabla u\|_{L^p(\Omega)}$, $n\in \mathbb N$, \item $u_n-u \in W^{1,p}_0(\Omega)$, $n\in \mathbb N$. \end{enumerate} In fact, $u_n$ may be taken to be $p$-harmonic on a subset of $\Omega$ having measure arbitrarily close to full and $C^\infty$-smooth, except at a discrete subset of $\Omega$. If $p=2$, then the functions $u_n$ can be taken to be $C^\infty$-smooth in $\Omega$. \end{theorem} Here, a function $f\colon \Omega\to \mathbb R$ is $C^{1,\alpha}$-smooth if it has derivatives of first order that are locally $\alpha$-H\"older continuous. We remark that standard mollification with a smooth kernel does not produce a monotone function and one has to use the structure of the level sets of the monotone function in order to construct its approximants. Therefore, Theorem \ref{Intro:MonotoneSobolev:Theorem} provides us with a powerful tool in this direction. For our proof we follow the strategy of \cite{IwaniecKovalevOnninen:W1papproximation} and \cite{IwaniecKovalevOnninen:W12approximation}, where it is proved that Sobolev homeomorphisms of planar domains can be approximated by $C^\infty$-smooth diffeomorphisms uniformly and in the Sobolev norm. In our proof we have to deal with further technicalities, not present in the mentioned results for homeomorphisms, which are related to the fact that our approximants $u_n$ might have critical points. In fact, $u_n$ will be $p$-harmonic in a large subset of $\Omega$ and it is known that at a critical point a $p$-harmonic function, $p\neq 2$, is only expected to be $C^{1,\alpha}$-smooth, rather than $C^{\infty}$-smooth. The main motivation for Theorem \ref{Intro:Approximation:Theorem} was to study regularity properties of a certain type of infimizers that appear in the setting of Sierpi\'nski carpets and are called \textit{carpet-harmonic functions}; see \cite[Chapter 2]{Ntalampekos:CarpetsThesis}. Namely, these infimizers are restrictions of monotone Sobolev functions (under some geometric assumptions) and the approximation Theorem \ref{Intro:Approximation:Theorem} would imply some absolute continuity properties for these functions. We will not discuss these applications any further in this paper. We pose some questions for further study. One of the reasons that we were not able to prove approximation by smooth functions for all $p\in (1,\infty)$ in Theorem \ref{Intro:Approximation:Theorem} was the presence of critical points of $p$-harmonic functions. \begin{question} Can a $p$-harmonic function be approximated uniformly and in the $W^{1,p}$ norm by $C^\infty$-smooth monotone functions near a critical point, when $p\neq 2$? \end{question} A positive answer to this question would imply that we may use smooth functions in Theorem \ref{Intro:Approximation:Theorem}. It would be very surprising if this fails. Moreover, what can one say for higher dimensions? \begin{question} Do analogs of Theorems \ref{Intro:MonotoneSobolev:Theorem} and \ref{Intro:Approximation:Theorem} hold in higher dimensions? \end{question} The smooth approximation of Sobolev homeomorphisms is still open in higher dimensions (\cite[Question 1.1]{IwaniecKovalevOnninen:W12approximation}, \cite{FoundCompMath:book-Ball}). Since the coordinate functions of a homeomorphism are monotone, a first step in studying this problem would be to study the smooth approximation of monotone Sobolev functions as in the above question. The paper is structured as follows. In Section \ref{Section:Level} we study the level sets of Sobolev functions. In particular, in Subsection \ref{Section:Sobolev} we prove Theorem \ref{Intro:Sobolev:Theorem} and in Subsection \ref{Section:levelmonotone} we prove Theorem \ref{Intro:MonotoneSobolev:Theorem} and discuss the generalization that gives Theorem \ref{Intro:metric:theorem}. In Subsection \ref{Section:glue} we include some gluing results for monotone functions that are used in the proof of the approximation Theorem \ref{Intro:Approximation:Theorem}, but also are of independent interest. Section \ref{Section:Preliminaries} contains preliminaries on $p$-harmonic functions. The approximation Theorem \ref{Intro:Approximation:Theorem} is proved in Section \ref{Section:approximationTheorem}. Finally, in Section \ref{Section:Smoothing} we include some quite standard smoothing results for monotone functions that are needed for the proof of the approximation theorem. \subsection*{Acknowledgments} The author would like to thank Tadeusz Iwaniec for a motivating discussion and Matthew Romney for his comments on the manuscript. \section{Level sets of Sobolev and monotone functions}\label{Section:Level} Throughout the entire section we assume that $u\colon \Omega\to \mathbb R$ is a non-constant continuous function on an open set $\Omega\subset \mathbb R^2$. We define $A_t=u^{-1}(t)$ for $t\in \mathbb R$, which can be the empty set. For $ s<t$ we define $A_{s,t}=u^{-1}((s,t))$. A \textit{Jordan arc} in a metric space $X$ is the image of the interval $[0,1]$ under a homeomorphic embedding $\phi \colon [0,1] \to X$. In this case, the set $\phi((0,1))$ is called an \textit{open Jordan arc}. Finally, a \textit{Jordan curve} is the image of $S^1$ under a homeomorphic embedding $\phi\colon S^1\to X$. \subsection{Level sets of Sobolev functions}\label{Section:Sobolev} In this subsection we study the level sets of Sobolev functions and prove Theorem \ref{Intro:Sobolev:Theorem}. \begin{lemma}\label{FiniteLength:Lemma} Suppose that $u\in W^{1,p}(\Omega)$ for some $1\leq p \leq \infty$ and $\Area(\Omega)<\infty$. Then for a.e.\ $t\in \mathbb R$ the level set $A_t$ has finite Hausdorff $1$-measure. \end{lemma} \begin{proof} This follows from the co-area formula \cite[Theorem 1.1]{MalySwansonZiemer:Coarea}, which is attributed to Federer \cite[Section 3.2]{Federer:gmt}, and the $L^p$-integrability of $\nabla u$: \begin{align*} \int \mathcal H^1(u^{-1}(t))\, dt= \int_\Omega |\nabla u| \leq \|\nabla u\|_{L^p(\Omega)} \cdot \Area(\Omega)^{1/p'}<\infty, \end{align*} where $\frac{1}{p}+\frac{1}{p'}=1$. \end{proof} Now, we restate and prove Theorem \ref{Intro:Sobolev:Theorem}. \begin{theorem}\label{Jordan:Theorem} Suppose that $u\in W^{1,p}_{\loc}(\Omega)$ for some $1\leq p\leq \infty$. Then for a.e.\ $t\in \mathbb R$ the level set $A_t$ has locally finite Hausdorff $1$-measure and each component $E$ of $A_t$ is either a point, or a Jordan curve, or it is homeomorphic to an interval. \end{theorem} Recall that $u$ is assumed to be continuous. If a level set $A_t=u^{-1}(t)$ is the empty set then it has no components so the conclusions of the theorem hold trivially; this is also true for other statements later in the paper, so we will not mention again the possibility that the level sets can be empty. \begin{proof} \textbf{Main Claim:} Consider an open set $U$ with $U\subset \subset \Omega$. We restrict $u$ to a neighborhood of $\overline U$ that is compactly contained in $\Omega$ and apply Lemma \ref{FiniteLength:Lemma}. It follows that for a.e.\ $t\in \mathbb R$ we have $\mathcal H^1(A_t\cap \overline U)<\infty$. We shall show as our Main Claim that if we further exclude countably many values $t$, then each component $E_0$ of $A_t\cap \overline U$ is either a point, or a Jordan arc, or a Jordan curve. \vspace{1em} \noindent \textbf{Compact exhaustion argument:} Assuming that the Main Claim holds for each such $U$, we consider an exhaustion of $\Omega$ by a nested sequence of open sets $\{U_n\}_{n\in \mathbb N}$, each compactly contained in $\Omega$, such that for a.e.\ $t\in \mathbb R$ the following holds for the level set $A_t$: $\mathcal H^1(A_t\cap \overline{U_n})<\infty$ and each component $E_n$ of $A_t\cap \overline {U_n}$ is either a point, or a Jordan arc, or a Jordan curve, for all $n\in \mathbb N$. We let $A_t$ be such a level set and fix $x_0\in A_t$. We will show that the component $E$ of $A_t$ containing $x_0$ is either a point, or a Jordan curve, or it is homeomorphic to an interval. We have $x_0\in U_n$ for all sufficiently large $n$. To simplify our notation, we assume that this holds for all $n\in \mathbb N$. Let $E, E_n$ be the component of $A_t, A_t\cap \overline{U_n}$, respectively, that contains $x_0$. We have $ E_n \subset E_{n+1} \subset E$ for each $n\in \mathbb N$, which follows from the definition of a connected component (as the largest connected set containing a given point). By the continuity of $u$, $A_t$ is rel.\ closed in $\Omega$, so $E_n$ is compact for all $n\in \mathbb N$. If $E\subset U_n$ for some $n\in \mathbb N$, then $E_n=E$ and therefore $E$ itself is either a point, or a Jordan curve, or a Jordan arc by the Main Claim. This completes the proof in this case. Suppose now that $E$ is not contained in $U_n$ for any $n\in \mathbb N$. Then $E_n$ necessarily intersects $\partial U_n$ for each $n\in \mathbb N$ as we see below using the next lemma, which is a direct consequence of \cite[IV.5, Corollary 1, p.~83]{Newman:Topology}. \begin{lemma}\label{Separation:Lemma} Suppose that $F$ is a connected component of a compact set $A$ in the plane. Then for each open set $U\supset F$ there exists an open set $V$ with $F\subset V \subset \subset U$ and $\partial V \cap A=\emptyset$. \end{lemma} In our case, if $E_n \subset U_n$, then by the preceding lemma there exists an open set $V\supset E_n$ with $ V\subset\subset U_n$ and $\partial V\cap (A_t\cap \overline{U_n})=\emptyset$. This implies that $\partial V\cap A_t=\emptyset$. Then $\partial V\cap E=\emptyset$ and it follows that $V\cap E$ is both open and closed in $E$, so $V\cap E=E$ by the connectedness of $E$. Then $E=V\cap E\subset U_n$, which contradicts our assumption that $E\not\subset U_n$ for all $n\in \mathbb N$. Therefore, $E_n\cap \partial U_n \neq \emptyset$ for all $n\in \mathbb N$. This implies that $E_n$ cannot be a single point since it also contains $x_0\in U_n$, so $E_n$ is either a Jordan arc, or a Jordan curve for each $n\in \mathbb N$, by the Main Claim. Note that $E_{n+1} \supsetneq E_n$ for all $n\in \mathbb N$, since $E_{n+1}\cap \partial U_{n+1}\neq \emptyset$ and $\partial U_{n+1}\cap \partial U_n=\emptyset$, as $U_n\subset \subset U_{n+1}$. If $E_n$ is a Jordan curve, then $E_{n+1}$ can be neither a Jordan curve, nor a Jordan arc, as it is strictly larger than $E_n$ (this uses the fundamental fact that $S^1$ is not homeomorphic to $[0,1]$). Therefore, $E_n$ is a Jordan arc for all $n\in \mathbb N$, and there exist homeomorphisms $\phi_n\colon [0,1]\to E_n$, $n\in \mathbb N$. Since $E_n\subsetneq E_{n+1}$, these homeomorphisms can be pasted appropriately to obtain a continuous injective map $\phi \colon I\to \bigcup_{n\in \mathbb N} E_n \eqqcolon \mathcal E$, where $I\subset \mathbb R$ is either $\mathbb R$ or $[0,\infty)$ (after a change of variables). Assume that $I=[0,\infty)$; the other case is treated in the same way. The map $\phi$ has the property that $\phi^{-1}(E_n)$ is a compact subinterval of $I$ and $\phi^{-1}(E_n)\subsetneq \phi^{-1}(E_{n+1})$ for $n\in \mathbb N$. Moreover, $\phi(I)=\mathcal E$ exits all compact subsets of $\Omega$. It now remains to show that $\phi$ is a homeomorphism and that $E= \bigcup_{n\in \mathbb N} E_n$, concluding therefore that $E$ is homeomorphic to an interval as desired. This is subtle and requires to use the assumption that $\mathcal H^1(A_t\cap \overline {U_n}) <\infty$ for all $n\in \mathbb N$. We first claim that $\phi(s)$ does not accumulate at any point of $\Omega$ as $s\to \infty$. Assume for the sake of contradiction that $\phi(s)$ converges to a point $y_0\in U_{n_0}\subset \Omega$ along a subsequence of $s\to \infty$, for some $n_0\in \mathbb N$. Since $\phi(s)$ exits all sets $U_n$, we may find disjoint intervals $[a_n,b_n]$, $n\in \mathbb N$, such that $\phi(a_n) \to y_0$ as $n\to\infty$, $\phi([a_n,b_n])\subset \overline {U_{n_0}}$, and $\phi(b_n)\in \partial U_{n_0}$ for all $n\in \mathbb N$. It follows that $\liminf_{n\to\infty}\diam (\phi([a_n,b_n])) \geq \dist(y_0, \partial U_{n_0})>0$. Since the diameter is a lower bound for the Hausdorff $1$-measure in connected spaces (see \cite[Section 18, p.~18]{Semmes:Hausdorff}), we have \begin{align*} \mathcal H^1( A_t\cap \overline{U_{n_0}}) \geq \sum_{n\in \mathbb N} \mathcal H^1(\phi([a_n,b_n])) \geq \sum_{n\in \mathbb N} \diam (\phi([a_n,b_n]))=\infty, \end{align*} a contradiction. This now implies that $\phi$ is a homeomorphism of $I$ onto $\mathcal E$. Indeed, if $\phi(s_n)=y_n\in \mathcal E$ is a sequence converging to $\phi(s_0)=y_0\in \mathcal E$, then $y_n$ is contained in a compact subset of $\Omega$ for all $n\in \mathbb N$ and therefore $s_n,s_0$ lie in a compact subinterval $I_0$ of $I$ for all $n\in \mathbb N$. By the injectivity oh $\phi$, $\phi|I_0$ is a homeomorphism, so $s_n\to s_0$, and $\phi^{-1}$ is continuous on $\mathcal E$, as desired. Another implication of the previous paragraph is that $\overline{\mathcal E} \setminus \mathcal E$ is contained in $\partial \Omega$ and $\mathcal E=\overline{\mathcal E}\cap \Omega= \overline{\mathcal E}\cap E$ is rel.\ closed in $\Omega$ and in $E$. We will show that $\mathcal E$ is also rel.\ open in $E$ and by the connectedness of $E$ we will have $E=\mathcal E$ as desired. Let $x\in \mathcal E$, so $x\in U_{n_0}$ for some $n_0\in \mathbb N$ and consider an open neighborhood $V\subset \subset U_{n_0}$ of $x$. We wish to show that $V\cap E \subset \mathcal E$, after shrinking the neighborhood $V$ if necessary. This will complete the proof that $\mathcal E$ is rel.\ open in $E$. If $y_0 \in V\cap E\setminus \mathcal E$, then arguing as we did for the construction of $\mathcal E$ and using the Main Claim, one can construct a set $\mathcal F \subset E$ containing $y_0$ and a homeomorphism $\phi' \colon I' \to \mathcal F$, where $I'=\mathbb R$ or $I'=[0,\infty)$. Moreover, the set $\mathcal F$ is, by construction, necessarily disjoint from $\mathcal E$. To see this, assume that they intersect at a point $z_0=\phi(s_0)= \phi'(s_0')$. Then there exist non-trivial closed intervals $I_0\subset I$ and $I_0'\subset I'$ such that $\phi(I_0)$ is the subarc of $\mathcal E$ from $x_0$ to $z_0$ and $\phi'(I_0')$ is the subarc of $\mathcal F$ from $y_0$ to $z_0$. We choose a large $k\in \mathbb N$ so that $\phi(I_0)\cup \phi'(I_0')\subset U_k$. Then by the definition of $E_k$ we must have $E_k \supset \phi(I_0)\cup \phi'(I_0')$, since the latter is connected and contains $x_0$. Then we have $y_0\in \phi'(I_0')\subset E_k\subset \mathcal E$, which is a contradiction. Hence, we have indeed $\mathcal F \cap \mathcal E=\emptyset$. Moreover, as in the construction of $\mathcal E$, since $E$ exits all compact subsets of $\Omega$, the set $\mathcal F$ must also have this property; see the statement right before Lemma \ref{Separation:Lemma}. Therefore, \begin{align*} \mathcal H^1(\mathcal F\cap \overline {U_{n_0}}) \geq \dist (y_0, \partial U_{n_0}) \geq \dist (\partial V, \partial U_{n_0}) \eqqcolon \delta>0. \end{align*} Another property of $\mathcal F$ is that it is rel.\ closed in $\Omega$, for the same reason as $\mathcal E$. If $V\cap E \setminus (\mathcal E\cup \mathcal F)\neq \emptyset$, by repeating the above procedure we may find sets $\mathcal F_i\subset A_t$, $i=1,2,\dots,N$, with the same properties as $\mathcal F\eqqcolon \mathcal F_1$ so that they are disjoint with each other and with $\mathcal E$, and they intersect $V$. We have \begin{align*} \infty > \mathcal H^1(A_t\cap \overline{U_{n_0}} ) \geq \sum_{i=1}^N \mathcal H^1(\mathcal F_i\cap \overline{U_{n_0}} ) \geq N\delta. \end{align*} This implies that we can find only a finite number of such sets $\mathcal F_i$. Since the compact sets $\mathcal F_i\cap \overline V$ have positive distance from $\mathcal E\cap \overline V$, we may find a smaller neighborhood $W\subset V$ of $x\in \mathcal E$ such that $W\cap E \subset \mathcal E$. This completes the proof that $\mathcal E$ is rel.\ open in $E$, as desired. \vspace{1em} \noindent \textbf{Proof of Main Claim:} We will show that for all but countably many $t\in \mathbb R$ for which $\mathcal H^1(A_t\cap \overline U)<\infty$ we have that each component $E_0$ of $A_t\cap \overline U$ is either a point, or a Jordan arc, or a Jordan curve. Suppose that $\mathcal H^1(A_t\cap \overline U)<\infty$ and let $E_0$ be a component of $A_t\cap \overline U$, so $L\coloneqq \mathcal H^1(E_0)<\infty$. Since $E_0$ is a continuum, it follows that there exists a $2L$-Lipschitz continuous parametrization $\gamma\colon [0,1]\to E_0$ of $E_0$; see \cite[Theorem 2a]{EilenbergHarrold:Curves} or \cite[Proposition 5.1]{RajalaRomney:reciprocal}. Hence, $E_0$ is a locally connected, compact set; see \cite[Theorem 9.2, p.~60 and Theorem 27.12, p.~200]{Willard:topology}. Now, we need the following topological lemma that we prove later: \begin{lemma}\label{Junction:Lemma} Let $X$ be a Peano space, i.e., a connected, locally connected, compact metric space. If $X$ contains more than one points and is not homeomorphic to $[0,1]$ or $S^1$, then it must have a \textit{junction point} $x$, i.e., there exist three Jordan arcs $A_1,A_2,A_3\subset X$ that meet at $x$ but otherwise they are disjoint. \end{lemma} By our previous discussion, if $\mathcal H^1(A_t \cap \overline U)<\infty$ then each component $E_0$ of $A_t\cap \overline U$ is a Peano space. If there is a component $E_0$ of $A_t\cap \overline U$ that is not a point or a Jordan arc or a Jordan curve, then by Lemma \ref{Junction:Lemma}, $E_0$ must have a junction point. Hence, $A_t\cap \overline U$ has a junction point. A theorem of Moore \cite[Theorem 1]{Moore:triods} (see also \cite[Proposition 2.18]{Pommerenke:conformal}) states that there can be no uncountable collection of disjoint compact sets in the plane such that each set has a junction point. Note that $A_s\cap A_t=\emptyset$ for $s\neq t$. Hence, for at most countably many $t\in \mathbb R$ the set $A_t\cap \overline U$ can have a junction point. Summarizing, for at most countably many $t\in \mathbb R$ for which $\mathcal H^1(A_t\cap \overline U)<\infty$ the set $A_t\cap \overline U$ has a component $E_0$ that has a junction point and is not a Jordan arc or a Jordan curve. \end{proof} \begin{proof}[Proof of Lemma \ref{Junction:Lemma}] Suppose that $X$ contains more than one points and is not homeomorphic to $[0,1]$. Then there exist at least three \textit{non-cut points} $x_1,x_2,x_3\in X$, i.e., points whose removal does not disconnect $X$; see \cite[Theorems (6.1) and (6.2), p.~54]{Whyburn:topology}. Since $X$ is a Peano space, it is locally path-connected \cite[Theorem 31.4, p.~221]{Willard:topology}. It follows that each of the spaces $X\setminus \{x_{i}\}$, $i=1,2,3$, is connected and locally path-connected. Moreover, a connected, locally path-connected space is path-connected \cite[Theorem 4.3, p.~162]{Munkres:topology}. Hence, we may find Jordan arcs $J_{ij} \subset X \setminus \{x_k\}$ with endpoints $x_i$ and $x_j$, where $i,j,k\in \{1,2,3\}$, are distinct and $i<j$; see also \cite[Corollary 31.6, p.~222]{Willard:topology}. If two of the arcs $J_{ij}$ intersect at an interior point, i.e., a point different from $x_1,x_2,x_3$, then it is straightforward to consider three subarcs $A_1,A_2,A_3$ of these arcs that intersect at one point, but otherwise are disjoint, as required in the statement of the lemma. If the arcs $J_{ij}$ intersect only at the endpoints, then we can concatenate the three arcs to obtain a Jordan curve $J$, i.e., homeomorphic to $S^1$. Suppose now that the space $X$ is, in addition, not homeomorphic to $S^1$. Then, there must exist a point $x\in X\setminus J$. We claim that there exists a Jordan arc $J_x$ that connects $x$ to $J$ and intersects $J$ at one point. Assuming that claim, one can now define $A_1\coloneqq J_x$ and $A_2,A_3$ to be subarcs of $J$ that meet at $x$ but otherwise are disjoint. To prove the claim note that since $X$ is a {Peano space}, any two points in $X$ can be joined with a Jordan arc \cite[Theorem 31.2, p.~219]{Willard:topology}. We connect $x$ to any point $y\in J$ with a Jordan arc $J_x$, parametrized by $\gamma\colon[0,1]\to J_x$, so that $\gamma(0)=x$ and $\gamma(1)=y$. If there exists $s\in (0,1)$ such that $\gamma(s)\in J$ then we consider the smallest such $s$ and we restrict $\gamma$ to $[0,s]$. This gives the desired Jordan arc. \end{proof} \begin{remark}\label{remark:metric spaces} In Theorem \ref{Jordan:Theorem} (Theorem \ref{Intro:MonotoneSobolev:Theorem}) the assumption that $u\in W^{1,p}_{\loc}(\Omega)$ is only used to deduce that almost every level set of $u$ has locally finite Hausdorff $1$-measure; the latter is proved in Lemma \ref{FiniteLength:Lemma} using the co-area formula. All the other arguments rely on planar topology. Thus our proof can be generalized to functions defined on metric spaces homeomorphic to $\mathbb R^2$ and the first part of Theorem \ref{Intro:metric:theorem} follows. \end{remark} \subsection{Level sets of monotone functions}\label{Section:levelmonotone} In this subsection we study the level sets of monotone functions and prove Theorem \ref{Intro:MonotoneSobolev:Theorem}. Recall that $u\colon \Omega\to \mathbb R$ is assumed to be continuous. \begin{lemma}\label{Monotonicity:Corollary} If $u$ is monotone in $\Omega$ then for each $t\in \mathbb R$ each component $E$ of the level set $A_t$ satisfies either of the following: \begin{enumerate}[\upshape(i)] \item $E$ exits all compact subsets of $\Omega$, or \item all bounded components of $\mathbb R^2 \setminus E$ intersect $\partial \Omega$ and there exists at least one such component. \end{enumerate} In particular, $$\diam (E) \geq \sup_{x\in E} \dist(x,\partial \Omega)>0.$$ \end{lemma} \begin{proof} In the proof we will use the following lemma, known as Zoretti's theorem, which is in the same spirit as Lemma \ref{Separation:Lemma}. \begin{lemma}[{\cite[Corollary 3.11, p.~35]{Whyburn:TopologicalAnalysis}}]\label{Zoretti:Lemma} Let $F$ be a connected component of a compact set $A$ in the plane. Then for each open set $U \supset F$ there exists a Jordan region $V$ with $F\subset V$, $\partial V\subset U$, and $\partial V\cap A=\emptyset$. \end{lemma} Suppose that a component $E$ of $A_t$ is compactly contained in $\Omega$. First we will show that $\mathbb R^2\setminus E$ has a bounded component. Suppose that this is not the case. Then $\widehat\mathbb C\setminus E$ is simply connected and contains $\partial \Omega$, so by using the Riemann mapping theorem we may find arbitrarily close to $E$ Jordan curves surrounding $E$ and separating $E$ from $\partial \Omega$. Hence, we may find a Jordan region $U \subset \subset \Omega$ containing $E$. Consider the compact set $A_t\cap \overline U$. Then $E$ is also a component of $A_t\cap \overline U$. By Lemma \ref{Separation:Lemma}, we can find a Jordan region $V$ such that $E\subset V$, $\partial V\subset U$, and $\partial V\cap A_t=\emptyset$. It follows that $V\subset U\subset \subset \Omega$. On $\partial V$ we must have $u>t$ or $u<t$. Without loss of generality, suppose that we have $u>t$ on $\partial V$. Then by the monotonicity of $u$ we have $u>t$ on $ V\supset E$, a contradiction. Therefore, $\mathbb R^2\setminus E$ has a bounded component. Next, we will show that all bounded components of $\mathbb R^2\setminus E$ must intersect $\partial \Omega$. Suppose that a bounded component $U$ of $\mathbb R^2\setminus E$ does not intersect $\partial \Omega$. Then $U\subset \Omega$ and $\partial U \subset E\subset \subset \Omega$ so $U\subset \subset \Omega$. Since $u=t$ on $\partial U$, by the monotonicity of $u$ we have that $u=t$ on $\overline U$. Since $E$ is a connected component of $A_t$, we must have $U\subset E$, a contradiction. Now we prove the claim involving the diameters. If $\Omega=\mathbb R^2$, then $E$ necessarily satisfies the first alternative, so it escapes to $\infty $ and $\diam(E)=\infty$. If $\Omega \subsetneq \mathbb R^2$, then for each $x\in E$ we may consider the largest ball $B(x,r)$ contained in $\Omega$, where $r=\dist(x,\partial \Omega)$. Then $E$ cannot be contained in a ball $B(x,r-\varepsilon)$ for any $\varepsilon>0$ since this would violate both alternatives. Therefore, $\diam(E) \geq r-\varepsilon$ for all $\varepsilon>0$, which implies that $\diam(E) \geq \dist(x,\partial \Omega)$, as desired. \end{proof} We record an immediate corollary: \begin{corollary} If $\Omega$ is simply connected, then $u$ is monotone in $\Omega$ if and only if each component of each level set of $u$ exits all compact subsets of $\Omega$. \end{corollary} \begin{proof} Suppose that $u$ is monotone. If $\Omega$ is simply connected, then $\partial \Omega$ is connected, so it cannot be separated by a level set $A_t$ of $u$. Thus, in Lemma \ref{Monotonicity:Corollary} only the first alternative can occur, as desired. Conversely, suppose that only the first alternative occurs and let $U\subset\subset \Omega$. Then for any $x_0\in U$ the component $E$ of the level set $A_t$, $t=u(x_0)$, that contains $x_0$ has to intersect $\partial U$. Thus, there exists $y_0\in \partial U$ with $u(x_0)=u(y_0)$. This implies the monotonicity of $u$. \end{proof} Next, we add the assumption that $u$ lies in a Sobolev space: \begin{lemma}\label{MonotoneSobolev:Lemma} Suppose that $u$ is monotone in $\Omega$ and lies in $W^{1,p}_{\loc}(\Omega)$ for some $1\leq p \leq \infty$. Then for a.e.\ $t\in \mathbb R$ the components of the level set $A_t$ are rel.\ open in $A_t$. In other words, if $E$ is a component of $A_t$ and $x\in E$ then there exists an open neighborhood $U$ of $x$ such that $E\cap U= A_t\cap U$. \end{lemma} \begin{proof} We consider an exhaustion of $\Omega$ by a nested sequence of open sets $\{U_n\}_{n\in \mathbb N}$, each compactly contained in $\Omega$, such that for a.e.\ $t\in \mathbb R$ we have $\mathcal H^1(A_t\cap \overline{U_n})<\infty$ for all $n\in \mathbb N$; the existence of such an exhaustion can be justified using Lemma \ref{FiniteLength:Lemma}. We fix $t\in \mathbb R$ such that $A_t\neq \emptyset$, and consider a component $E\subset A_t$ and $x\in E$. We claim that for each neighborhood $V\subset \subset \Omega$ of $x$ there are at most finitely many components of $A_t$ intersecting $V$. There exists $n_0\in \mathbb N$ such that $V\subset \overline V\subset U_{n_0}$. Suppose that $F$ is a component of $A_t$ intersecting $V$. We consider the restriction of $u$ to $U_{n_0}$, which is still a monotone function and we let $G\subset F$ be a component of the level set $A_t\cap U_{n_0}$ of $u\big|{U_{n_0}}$ such that $G\cap V\neq \emptyset$. For a point $y\in G\cap V$ we have $$\diam(G) \geq \dist(y,\partial U_{n_0})$$ by Lemma \ref{Monotonicity:Corollary}. Since $V\subset\subset U_{n_0}$, we have $\dist(y,\partial U_{n_0}) \geq \dist(\overline V, \partial U_{n_0}) >0$. Moreover, by the connectedness of $G$ we have $\mathcal H^1(G) \geq \diam (G)$; see \cite[Section 18, p.~18]{Semmes:Hausdorff}. Combining these inequalities, we have \begin{align*} \mathcal H^1(F\cap \overline{U_{n_0}} )\geq \mathcal H^1(G)\geq \diam(G)\geq \dist(y,\partial U_{n_0})\geq \dist(\overline V, \partial U_{n_0}) >0. \end{align*} Since $\mathcal H^1(A_t\cap\overline{U_{n_0}})<\infty$, there can be at most finitely many components $F$ of $A_t$ intersecting $V$. Since the compact sets $\overline V\cap F$ and $\overline V\cap E$ have positive distance, it follows that if we choose a smaller neighborhood $U\subset V$ of $x$, then we have $E\cap U=A_t\cap U$ as desired. \end{proof} We will also need the following general lemma: \begin{lemma}\label{Extremal:Lemma} Let $(X,d)$ be a separable metric space and $f\colon X \to \mathbb R$ be any function. Then the set of local extremal values of $f$ is at most countable. \end{lemma} \begin{proof} Let $E$ be the set of local maximum values of $f$. Then, by definition, for each $y\in E$ there exists $x\in X$ with $f(x)=y$ and a ball $B(x,r)$ such that for all $z\in B(x,r)$ we have $f(z)\leq y$. We can write $E=\bigcup_{n\in \mathbb N} E_n$, where $$E_n=\{y\in \mathbb R: y=f(x) \,\,\,\textrm{for some}\,\,\, x\in X \,\,\, \textrm{and}\,\,\, f(z)\leq y \,\,\, \textrm{for all}\,\,\, z\in B(x,2/n)\}.$$ We will show that $E_n$ is at most countable for each $n\in \mathbb N$. Let $y_1,y_2\in E_n$ be distinct, so there exist points $x_1,x_2\in X$ with $f(x_i)=y_i$, $i=1,2$, as in the definition of $E_n$. Then the balls $B(x_1, 1/n)$, $B(x_2,1/n)$ are necessarily disjoint. Indeed, otherwise, we have $f(x_2)\leq y_1$ and $f(x_1)\leq y_2$, so $y_1=y_2$, a contradiction. Therefore, the set $E_n$ is in one-to-one correspondence with a collection of disjoint balls in $X$. The separability of $X$ implies that there can be at most countably many such balls. The same proof shows that the set of local minimum values of $f$ is at most countable. \end{proof} Now we have completed the preparation for the proof of Theorem \ref{Intro:MonotoneSobolev:Theorem}, restated below: \begin{theorem}\label{Sard:Theorem} Suppose that $u$ is monotone in $\Omega$ and lies in $W^{1,p}_{\loc}(\Omega)$ for some $1\leq p \leq \infty$. Then for a.e.\ $t\in \mathbb R$ the level set $A_t$ is an embedded $1$-dimensional topological submanifold of $\mathbb R^2$ that has locally finite Hausdorff $1$-measure. \end{theorem} \begin{proof} By Theorem \ref{Jordan:Theorem} and Lemma \ref{MonotoneSobolev:Lemma} we conclude that for a.e.\ $t\in \mathbb R$ the level set $A_t$ has locally finite Hausdorff $1$-measure, each component $E$ of the level set $A_t$ is rel.\ open in $A_t$, and it is either a point, or a Jordan curve, or it is homeomorphic to an interval. Using Lemma \ref{Extremal:Lemma}, we further exclude the countably many local extremal values $t\in \mathbb R$ of $u$ in $\Omega$. We fix a level set $A_t$ satisfying all the above. In particular, $A_t$ has the property that if $x\in A_t$ then each neighborhood $U$ of $x$ contains points with $y_1,y_2$ with $u(y_1)>t$ and $u(y_2)<t$. Our goal is to prove that every component $E$ of $A_t$ is either a Jordan curve, or it is homeomorphic to $\mathbb R$. Since each component of $A_t$ is rel.\ open in $A_t$, it will then follow that each $x\in A_t$ has a neighborhood $U$ such that $U\cap A_t$ is homeomorphic to an open interval. This shows that $A_t$ is a $1$-manifold. Since there are no wild arcs in the plane (see Remark \ref{Straighten:Remark} below), each Jordan arc of $A_t$ can be mapped to $[0,1]\times \{0\}$ with a global homeomorphism of $\mathbb R^2$. This shows that $A_t$ is an embedded submanifold of $\mathbb R^2$, as desired. We now focus on proving that every component $E$ of $A_t$ is either a Jordan curve, or it is homeomorphic to $\mathbb R$ We already know that $E$ is either a point, or a Jordan curve, or it is homeomorphic to an interval. If $E$ is a point or it is homeomorphic to the closed interval $[0,1]$, then $E$ is compactly contained in $\Omega$, so by Lemma \ref{Monotonicity:Corollary} the set $\mathbb R^2\setminus E$ must have at least one bounded component. This is a contradiction. Finally, assume that $E$ is homeomorphic to $I=[0,\infty)$ under a map $\phi\colon I \to E$. Then $\phi(s)$ cannot accumulate to any point of $\Omega$ as $s\to\infty$. This is because $\phi$ is a homeomorphism and $E$ is rel.\ closed in $A_t$, and thus in $\Omega$. Let $x_0=\phi(0)$ and consider a ball $B(x_0,r)\subset \subset \Omega$. Then there exists $s_0$ such that $\phi(s_0)\in \partial B(x_0,r)$ and $\phi(s)\notin B(x_0,r)$ for all $s\geq s_0$. We may straighten $\phi([0,s_0])$ to the segment $[0,1]\times \{0\}$ with a homeomorphism of $\mathbb R^2$ (see Remark \ref{Straighten:Remark} below) and, using that, we can find a topological ball $U\subset B(x_0,r)$ containing $x_0$ such that $U\setminus E$ is connected. If $U$ is sufficiently small, then by Lemma \ref{MonotoneSobolev:Lemma} we have that $U\setminus A_t=U\setminus E$, so $U\setminus A_t$ is connected. This implies that $u>t$ or $u<t$ in $U\setminus E$. This is a contradiction, since $u(x_0)=t$ would then be a local extremal value of $u$ in this case. \end{proof} \begin{remark}\label{Straighten:Remark} In the proof we used the fact that a Jordan arc in $\mathbb R^2$ can be mapped to $[0,1]\times \{0\}$ with a homeomorphism of $\mathbb R^2$. That is, there are no \textit{wild} arcs in the plane. To see this, one can first embed the Jordan arc in a Jordan curve, and then apply the Schoenflies theorem to straighten the Jordan curve. See also the discussion in \cite{NoWildArcs}. \end{remark} \begin{remark}\label{remark:metric spaces monotone} As in Remark \ref{remark:metric spaces} the proof of Theorem \ref{Sard:Theorem} (Theorem \ref{Intro:MonotoneSobolev:Theorem}) can be generalized to monotone functions defined on metric spaces homeomorphic to $\mathbb R^2$. Indeed, monotonicity is a topological property. This observation yields the second part of Theorem \ref{Intro:metric:theorem}. \end{remark} \subsection{Gluing monotone functions}\label{Section:glue} In this section we include some results that allow us to paste monotone functions in order to obtain a new monotone function. These results will be useful towards the proof of the approximation Theorem \ref{Intro:Approximation:Theorem}. The proofs are elementary but the assumptions of the statements are finely chosen and cannot be relaxed. Recall that $u$ is continuous in $\Omega$. \begin{lemma}[Gluing Lemma]\label{Gluing:Lemma} Suppose that $u$ is monotone in $\Omega$ and consider $t_1,t_2\in \mathbb R$ with $t_1<t_2$. Let $\Upsilon=u^{-1}((t_1,t_2))$ and consider a continuous function $v$ on $\Upsilon$ such that \begin{enumerate}[\upshape(i)] \item $v$ is monotone in $\Upsilon$, \item $v$ extends continuously to $\partial \Upsilon\cap \Omega$ and agrees there with $u$, and \item $t_1\leq v \leq t_2$ on $\Upsilon$. \end{enumerate} Then the function $\widetilde u$ that is equal to $u$ in $\Omega\setminus \Upsilon$ and to $v$ in $\Upsilon$ is monotone in $\Omega$. \end{lemma} The proof we give below is elementary but delicate. Note that it is important to assume that $u$ is monotone in all of $\Omega$, since the function $u(x)=|x|$ on $\{x\in \mathbb R^2: 1/2<|x|<1\}$ does not have a monotone extension in the unit disk. Moreover, (iii) cannot be relaxed. Indeed, the function $u(x)=|x|$ is monotone in the punctured unit disk, considered to be the set $\Omega$; however if we set $\widetilde u=u$ in $\{x\in \mathbb R^2: 1/2\leq |x|<1\}$ and $\widetilde u = 1-|x|$ in $\{x\in \mathbb R^2 : 0<|x|<1/2\}=u^{-1}((0,1/2))$, then $\widetilde u$ is not monotone in $\Omega$. \begin{proof} Aiming for a contradiction, suppose that $\widetilde u$ is not monotone in $\Omega$, so, without loss of generality, there exists an open set $U\subset \subset \Omega$ and $x_0\in U$ such that $\widetilde u(x_0)=\max_{\overline U} \widetilde u >\max_{\partial U} \widetilde u$. Suppose first that $x_0\in U\cap \Upsilon$, so $\widetilde u(x_0)=v(x_0)$. Note that $\overline{U\cap \Upsilon} \subset \overline \Upsilon \cap \Omega$, so $v$ is continuous in $\overline{U\cap \Upsilon} \subset \overline \Upsilon \cap \Omega$ by assumption (ii) and monotone in $U\cap \Upsilon$ by (i). By Remark \ref{Monotone:Remark} we conclude that there exists $y_0\in \partial (U\cap \Upsilon) \subset (U\cap \partial \Upsilon)\cup \partial U$ such that $$\widetilde u(y_0)=v(y_0)=\max_{\overline{U\cap \Upsilon}}v \geq v(x_0)= \widetilde u(x_0)= \max_{\overline U}\widetilde u.$$ It follows that $\widetilde u(y_0)= \widetilde u(x_0)=\max_{\overline U}\widetilde u> \max_{\partial U}\widetilde u$. Hence, $y_0\notin \partial U$, and we have $y_0\in U\cap \partial \Upsilon$. Since $\partial \Upsilon \cap \Omega \subset A_{t_1}\cup A_{t_2}$, we have $\widetilde u(y_0)=u(y_0)=t_1$ or $\widetilde u(y_0)=u(y_0)=t_2$. By the monotonicity of $u$ in $\Omega$, it follows that there exists $z_0\in \partial U$ such that $$u(z_0)=\max_{\overline U}u\geq u(y_0).$$ If $z_0\notin \Upsilon$, then $\widetilde u(z_0)=u(z_0)$, so $\widetilde u(z_0)\geq u(y_0)=\widetilde u(y_0)$. If $z_0\in \Upsilon$, then $\max_{\overline U}u=u(z_0)<t_2$ so on $U\setminus \Upsilon \supset U\cap \partial \Upsilon$ we necessarily have $u\leq t_1$. Since $y_0\in U\cap \partial \Upsilon$, we have $\widetilde u(y_0)=u(y_0)=t_1$. Moreover, $\widetilde u(z_0)=v(z_0)$ and by assumption (iii) we have $\widetilde u(z_0)\geq t_1$. It follows that $\widetilde u(z_0)\geq \widetilde u(y_0)$ also in this case. Therefore, $$\max_{\overline U} \widetilde u= \widetilde u(y_0) \leq \widetilde u(z_0) \leq \max_{\partial U} (\widetilde u),$$ which is a contradiction. It remains to treat the case that $x_0\in U\setminus \Upsilon$, so $\widetilde u(x_0)=u(x_0) \notin (t_1,t_2)$. If $u(x_0)\leq t_1$, then $\max_{\partial U}\widetilde u < u(x_0)\leq t_1$, so $\widetilde u<t_1$ on $\partial U$. By assumption (iii) we have $\partial U\cap \Upsilon =\emptyset$, so $u=\widetilde u$ on $\partial U$, and $\max_{\partial U} u =\max_{\partial U} \widetilde u <u(x_0)$, where $x_0\in U$. This contradicts the monotonicity of $u$ in $\Omega$. If $u(x_0)\geq t_2$, by the monotonicity of $u$ in $\Omega$, there exists $z_0\in \partial U$ such that $u(z_0)=\max_{\overline U}u \geq u(x_0) \geq t_2$. This implies that $z_0\notin \Upsilon$, so $\widetilde u(z_0)=u(z_0)$. Therefore, $$\max_{\partial U} \widetilde u \geq \widetilde u(z_0) =u(z_0)\geq u(x_0)= \widetilde u(x_0) =\max_{\overline U} \widetilde u,$$ which is a contradiction. \end{proof} \begin{lemma}\label{ConvergenceMonotonicity:Lemma} Let $\{u_n\}_{n\in \mathbb N}$ be a sequence of monotone functions in $\Omega$ converging locally uniformly to a function $u$. Then $u$ is monotone in $\Omega$. \end{lemma} \begin{proof} Let $U\subset \subset \Omega$. Then $\max_{\overline U}u_n =\max_{\partial U}u_n$. Since $u_n\to u$ uniformly in $\overline U$, we have $\max_{\overline U}u_n \to \max_{\overline U} u$ and $\max_{\partial U}u_n \to \max_{\partial U}u$. The claim for the minima is proved in the same way. \end{proof} \begin{corollary}\label{Gluing:Corollary} Suppose that $u$ is monotone in $\Omega$ and consider a bi-infinite sequence of real numbers $\{t_i\}_{i\in \mathbb Z}$, such that $t_i<t_{i+1}$, $i\in \mathbb Z$, and $\lim_{i\to \pm \infty} t_i= \pm \infty$. In each region $\Upsilon_i\coloneqq u^{-1}((t_i,t_{i+1}))$ consider a function $v_i$ such that \begin{enumerate}[\upshape(i)] \item $v_i$ is monotone in $\Upsilon_i$, \item $v_i$ extends continuously to $\partial \Upsilon_i\cap \Omega$ and agrees there with $u$, and \item $t_i\leq v_i\leq t_{i+1}$ on $\Upsilon_i$. \end{enumerate} Then the function $\widetilde u$ that is equal to $u$ on $\Omega\setminus \bigcup_{i\in \mathbb Z} \Upsilon_i$ and to $v_i$ in $\Upsilon_i$, $i\in \mathbb Z$, is continuous and monotone in $\Omega$. \end{corollary} \begin{proof} By Lemma \ref{Gluing:Lemma} and induction, one can show that for each $n\in \mathbb N$ the function $$\widetilde u_n \coloneqq u \cdot\x_{\Omega\setminus \bigcup_{|i|\leq n} \Upsilon_i}+ \sum_{|i|\leq n} v_i \cdot \x_{\Upsilon_i}$$ is continuous and monotone in $\Omega$. The important observation here is that $$u^{-1}((t_j,t_{j+1}))=\Upsilon_j= \widetilde u_n^{-1}((t_j,t_{j+1}))$$ for all $n\in \mathbb N$ and $|j|>n$, by assumption (iii). If $U\subset \subset \Omega$ is an open set, then there exists $n\in \mathbb N$ such that $$t_{-n} < \min_{\overline U} u \leq \max_{\overline U} u <t_n.$$ Hence, $\Upsilon_j\cap \overline U=\emptyset$ for $|j|> n$ and $\widetilde u_m=\widetilde u_n=\widetilde u$ on $\overline U$ for all $m\geq n$. It follows that $\{\widetilde u_n\}_{n\in \mathbb N}$ converges locally uniformly in $\Omega$ to $\widetilde u$, and therefore $\widetilde u$ is continuous and monotone by Lemma \ref{ConvergenceMonotonicity:Lemma}. \end{proof} In order to establish the approximation by $C^{\infty}$-smooth functions in Theorem \ref{Intro:Approximation:Theorem}, we need to introduce the notion of strict monotonicity and prove some further, more specialized, gluing lemmas. \begin{definition} A continuous function $f\colon \Omega\to \mathbb R$ is called strictly monotone if it is monotone and for each open set $U\subset\subset \Omega$ the maximum and minimum of $f$ on $\overline U$ are not attained at any point of $U$. \end{definition} \begin{example} If a function $f$ is of class $C^1$ and has no critical points in $\Omega$, then it has no local maxima and minima in $\Omega$ so it is strictly monotone. \end{example} \begin{lemma}[Gluing Lemma]\label{GluingMonotonicity:Lemma} Let $A,V$ be open subsets of $\Omega$ with $\overline V\cap \Omega\subset A$. Suppose that the function $u$ is monotone when restricted to $\Omega \setminus \overline V$ and strictly monotone when restricted to $A$. Then $u$ is monotone in $\Omega$. \end{lemma} \begin{proof} If the statement fails, there exists an open set $U\subset \subset \Omega$ such that the maximum or minimum of $u$ in $U$ is not attained at $\partial U$, but at an interior point $x_0\in U$. Without loss of generality, assume that $\max_{\partial U} u < \max_{\overline U}u =u(x_0)$. Note that $U$ cannot be contained in $\Omega\setminus \overline V$ or in $A$, by the monotonicity of $u$ there. Hence, $U$ intersects both $\overline V$ and $\Omega\setminus \overline A$. If $x_0\in U\cap \overline V \subset U\cap A$, then $u(x_0)= \max_{\overline U}u\geq \max_{\overline{U\cap A}}u$ and this contradicts the strict monotonicity of $u$ in $A$. If $x_0\in U\setminus \overline V\subset \Omega\setminus \overline V$, then by the monotonicity of $u$ there, there exists $x_1\in \partial (U\setminus \overline V)$ such that $u(x_1)\geq u(x_0)> \max_{\partial U}u$. We necessarily have that $x_1\notin \partial U$. Since $\partial (U\setminus \overline V)\subset \partial U \cup (U\cap \partial V)$, it follows that $x_1\in U\cap \partial V \subset U\cap A$. Then we have $u(x_1)\geq u(x_0) = \max_{\overline U}u \geq \max_{\overline{U\cap A}}u$. Again, this contradicts the strict monotonicity of $u$ in $A$. \end{proof} \begin{lemma}\label{GluingConstant:Lemma} Suppose that $J$ is a connected closed subset of $\Omega$ that exits all compact subsets of $\Omega$. If $u$ is monotone in $\Omega\setminus J$ and $u$ is constant in $J$, then $u$ is monotone in $\Omega$. \end{lemma} \begin{proof} Assume that $u=t$ in $J$. Suppose that $u$ is not monotone in $\Omega$, so, without loss of generality, we can find an open set $U\subset\subset \Omega$ and a point $x_0\in U$ with $u(x_0)=\max_{\overline U}u >\max_{\partial U}u$. By the monotonicity of $u$ in $\Omega\setminus J$, we necessarily have that $U$ intersects both $\Omega\setminus J$ and $J$, and $x_0\in J \cap U$. Thus, $u(x_0)=t=\max_{\overline U}u$. Since $J$ is connected and is not contained in $\overline U$, we have $J\cap \partial U\neq \emptyset$. Hence, there exists $y_0\in \partial U$ such that $u(y_0)=t\leq \max_{\partial U}u$, a contradiction. \end{proof} \section{Preliminaries on $p$-harmonic functions}\label{Section:Preliminaries} Let $\Omega\subset \mathbb R^2$ be an open set. A function $u\colon \Omega\to \mathbb R$ is called $p$-harmonic, $1<p<\infty$, if $u\in W^{1,p}_{\loc}(\Omega)$ and \begin{align*} \Delta_pu \coloneqq \textrm{div}(|\nabla u|^{p-2} \nabla u) =0 \end{align*} in the sense of distributions. That is, \begin{align*} \int_\Omega \langle |\nabla u|^{p-2} \nabla u , \nabla \phi \rangle =0 \end{align*} for all $\phi \in C^\infty_c(\Omega)$, i.e., compactly supported smooth functions in $\Omega$. We mention some standard facts about $p$-harmonic functions. There exists an exponent $\alpha \in (0,1]$ such that every $p$-harmonic function $u$ on $\Omega$ lies in $C^{1,\alpha}_{\loc}(\Omega)$ \cite{Uralceva:C1alpha}, \cite{Evans:C1alpha}, \cite{Lewis:C1alpha}. In fact, outside the \textit{singular set}, i.e., the set where $\nabla u=0$, the function $u$ is $C^\infty$-smooth by elliptic regularity theory; see \cite[Corollary 8.11, p.~186]{GilbargTrudinger:pde}. The singular set consists of isolated points, unless $u$ is constant \cite[Corollary 1]{Manfredi:pharmonicplane}. Finally, the maximum principle \cite[p.~111]{Heinonenetal:DegenerateElliptic} implies that $p$-harmonic functions are monotone. \begin{prop}[Solution to the Dirichlet Problem]\label{DirichletProblem:Prop} Let $\Omega\subset \mathbb R^2$ be a bounded open set and let $u_0\in W^{1,p}(\Omega)$ be given Dirichlet data. There exists a unique $p$-harmonic function $u\in W^{1,p}(\Omega)$ that minimizes the $p$-harmonic energy \begin{align*} E_p(v)\coloneqq \int_\Omega |\nabla v|^p \end{align*} among all functions $v\in W^{1,p}(\Omega)$ with $v-u_0 \in W^{1,p}_0(\Omega)$. \end{prop} See e.g.\ \cite[Chapter 5]{Heinonenetal:DegenerateElliptic} for a general approach. An open Jordan arc $J$ in $\mathbb R^2$ is $C^\infty$-\textit{smooth} if there exists an open set $U\supset J$ and a $C^\infty$-smooth diffeomorphism $\phi \colon U\to \mathbb R^2$ such that $\phi(J)=\mathbb R$. In this case, the open set $U\supset J$ may be taken to be arbitrarily close to $J$; see also the classification of $1$-manifolds in Theorem \ref{Classification:theorem}. Let $\Omega\subset \mathbb R^2$ be an open set and $J\subset \partial \Omega$ be a Jordan arc. We say that $J$ is an \textit{essential} boundary arc if for each $x_0\in J$ there exists a neighborhood $U$ of $x_0$ and a homeomorphism $\phi\colon U \to \mathbb R^2$ such that $\phi(J\cap U)=\mathbb R=\{(s,0)\in \mathbb R^2:s\in \mathbb R\}$ and $\phi(\Omega\cap U)= \mathbb R^2_+= \{(s,t)\in \mathbb R^2: t>0\}$. \begin{lemma}[Continuous extension]\label{ContinuousExtension:Lemma} Suppose that $u$ and $u_0$ are as in Proposition \ref{DirichletProblem:Prop}. Assume further that $J\subset \partial \Omega$ is an essential open Jordan arc. If $u_0$ extends continuously to $\Omega\cup J$, then $u$ also extends continuously to $\Omega\cup J$ and $u|J=u_0|J$. \end{lemma} This follows from \cite[Theorem 6.27]{Heinonenetal:DegenerateElliptic}, which implies that each point $x_0\in J$ is a \textit{regular point} for the $p$-Laplace operator, since $\partial \Omega$ is \textit{$p$-thick} at $x_0$; see \cite[Lemma 2]{Lehrback:CapacityEstimate} for a relevant capacity estimate. \begin{lemma}[Comparison]\label{Comparison:Lemma} Suppose that $u$ and $u_0$ are as in Proposition \ref{DirichletProblem:Prop}. If there exists $M\in \mathbb R$ such that $u_0\leq M$ in $\Omega$, then $u\leq M$ in $\Omega$. \end{lemma} To see this, one can take an exhaustion of $\Omega$ by regular open sets $\Omega_n$ (see \cite[Corollary 6.32]{Heinonenetal:DegenerateElliptic}) and solve the $p$-Laplace equation in $\Omega_n$ with boundary data $u_{0,n}$ that smoothly approximates in $W^{1,p}(\Omega_n)$ the function $u_0$ and satisfies $u_{0,n}\leq M+1/n$. The solution $u_n$ of the $p$-Laplace equation satisfies $u_n=u_{0,n}$ on $\partial \Omega_n$, and by the maximum principle \cite[p.~111]{Heinonenetal:DegenerateElliptic} we have $u_n\leq M+1/n$ in $\Omega_n$. Now, by passing to a limit \cite[Theorem 6.12]{Heinonenetal:DegenerateElliptic}, we obtain a $p$-harmonic function $\widetilde u\leq M$ that necessarily solves the Dirichlet Problem in $\Omega$. Since the solution is unique, we have $u=\widetilde u\leq M$. \begin{prop}[Non-degeneracy up to boundary]\label{NonDegeneracy:Prop} Suppose that $u$ is a $p$-harmonic function in $\Omega$. Moreover, suppose that $J\subset \partial \Omega$ is a $C^\infty$-smooth essential open Jordan arc on which $u$ extends continuously and is equal to a constant $t$, and suppose that $u\neq t$ in a neighborhood of $J$ in $\Omega$. Then $|\nabla u| \neq 0$ in a neighborhood of $J$ in $\Omega$ and in fact each point $x_0\in J$ has a neighborhood in $\Omega$ in which $|\nabla u|$ is bounded away from $0$. \end{prop} This result follows from \cite[Theorem 2.8]{LewisNystrom:BoundaryRegularityPHarmonic} and \cite[Corollary 6.2]{AikawaKilpelainenShanmugalingamZhong:BoundaryHarnack}. The first result implies that each $x_0\in J$ has a neighborhood $B(x_0,r)$ such that $|\nabla u(x)| \geq c |u(x)-t|/\dist(x,\partial \Omega)$ for all $x\in B(x_0,r)\cap \Omega$, where $c>0$ is a constant depending on $p$ and on the geometry of $J$. The second cited result implies that $|u(x)-t|/ \dist(x,\partial \Omega) \geq c'$ for all $x$ in a possibly smaller ball $B(x_0,r')\cap \Omega$, where $c'>0$ is some constant depending on $u$. If $J\subset \partial \Omega$ is a $C^\infty$-smooth essential open Jordan arc, then a function $u\colon \Omega \to \mathbb R$ is said to be ($C^\infty$-)\textit{smooth up to $J$} if there exists an open set $U\supset J$ and $u$ extends to a $C^\infty$-smooth function on $U$. Note that this does \textit{not} require that $u$ is smooth in all of $\Omega$. If $u$ is also smooth in $\Omega$, then we write that $u$ is smooth in $\Omega\cup J$. \begin{prop}[Boundary regularity]\label{BoundaryRegularity:Prop} Suppose that $u$ is a $p$-harmonic function in $\Omega$. Moreover, suppose that $J\subset \partial \Omega$ is a $C^\infty$-smooth essential open Jordan arc and that $u$ is $C^\infty$-smooth when restricted to $J$. If each $x_0\in J$ has a neighborhood in $\Omega$ in which $|\nabla u|$ is bounded away from $0$, then $u$ is $C^{\infty}$-smooth up to the arc $J$. \end{prop} This follows from \cite[Theorem 6.19, p.~111]{GilbargTrudinger:pde}, since under the assumptions the $p$-harmonic equation is uniformly elliptic up to each compact subarc of the Jordan arc $J$. Combining Propositions \ref{NonDegeneracy:Prop} and \ref{BoundaryRegularity:Prop} we have: \begin{corollary}\label{NonDegeneracy:Corollary} Suppose that $u$ is a $p$-harmonic function in $\Omega$. Moreover, suppose that $J\subset \partial \Omega$ is a $C^\infty$-smooth essential open Jordan arc on which $u$ extends continuously and is equal to a constant $t$, and suppose that $u\neq t$ in a neighborhood of $J$ in $\Omega$. Then $u$ is $C^\infty$-smooth up to the arc $J$ and each point of $J$ has a neighborhood in $\Omega\cup J$ in which $|\nabla u|$ is bounded away from $0$. \end{corollary} \section{Proof of the approximation Theorem}\label{Section:approximationTheorem} In this section we prove Theorem \ref{Intro:Approximation:Theorem}. Let us start with an elementary lemma: \begin{lemma}\label{EssentialArcs:Lemma} Let $f\colon \Omega\to \mathbb R$ be a continuous function and suppose that $\{s_i\}_{i\in \mathbb Z}$ is a bi-infinite sequence of real numbers with $s_i<s_{i+1}$, $i\in \mathbb Z$, and $\lim_{i\to\pm\infty} s_i= \pm\infty$. If each of the level sets $f^{-1}(s_i)$, $i\in \mathbb Z$, is an embedded $1$-dimensional topological submanifold of $\mathbb R^2$, then \begin{enumerate}[\upshape(i)] \item $\bigcup_{i\in \mathbb Z} f^{-1}(s_i)$ is also an embedded $1$-dimensional topological submanifold, \vspace{.3em} \item $\partial f^{-1}((s_j,s_{j+1}))\cap \Omega \subset f^{-1}(s_j) \cup f^{-1}(s_{j+1})$ and $f^{-1}(s_j)\subset\partial f^{-1}((s_{j-1},s_{j})) \cup \partial f^{-1}((s_j,s_{j+1})) $ for all $j\in \mathbb Z$, and \vspace{.3em} \item if $x\in f^{-1}(s_j)$ for some $j\in \mathbb Z$, then there exist disjoint Jordan regions $V_+, V_-$ such that the closure of $V_+\cup V_-$ contains a neighborhood of $x$ and there exists an open Jordan arc $J\subset \partial V_+\cap \partial V_- \cap f^{-1}(s_j)$ containing $x$ such that $J$ is an essential boundary arc of $ V_+$ and $ V_-$. Moreover, each of $V_+,V_-$ is contained in a component of $ f^{-1}((s_{j-1},s_j))\cup f^{-1}((s_j,s_{j+1}))$. \end{enumerate} \end{lemma} \begin{proof} Let $x\in \bigcup_{i\in \mathbb Z} f^{-1}(s_i)$, so $x\in f^{-1}(s_j)$ for some $j\in \mathbb Z$. Then we claim that there exists a neighborhood $V$ of $x$ such that $V\cap \bigcup_{i\in \mathbb Z} f^{-1}(s_i)= V\cap f^{-1}(s_j)$. This claim trivially implies (i). Note that $x$ has positive distance from $f^{-1}(s_i)$, for all $i\neq j$ by the continuity of $f$. If the claim were not true, then one would be able to find a sequence $x_n\in f^{-1}(s_{i_n})$, where $i_n$ are distinct integers with $i_n\neq j$, $n\in \mathbb N$, such that $x_n\to x$ as $n\to \infty$. By continuity $f(x_n)\to f(x)$ so $s_{i_n}\to s_j$ as $n\to \infty$. This contradicts the assumptions on the sequence $\{s_i\}_{i\in \mathbb Z}$. Part (ii) follows by the continuity of $f$. Indeed, if $x_n \in f^{-1}((s_j,s_{j+1}))$ converges to $x \in \partial f^{-1}((s_j,s_{j+1}))\cap \Omega$, then $f(x_n)$ converges to $f(x)$. Note that $f(x)$ cannot lie in $(s_j,s_{j+1})$, otherwise $f(y)\in (s_j,s_{j+1})$ for all $y$ in a neighborhood of $x$ by continuity, so $x$ would not be a boundary point of $f^{-1}((s_j,s_{j+1}))$. Therefore $f(x)=s_j$ or $f(x)=s_{j+1}$. For the second claim in (ii) note that $f^{-1}(s_j)$ has empty interior (as a subset of the plane), since it is a $1$-manifold. Therefore for each point $x\in f^{-1}(s_j)$ we can find a sequence of points $x_n$ converging to $x$ with $f(x_n)\neq s_j$, so $f(x_n)\in (s_{j-1},s_j)\cup (s_j,s_{j+1})$ for all sufficiently large $n$. The claim follows. For (iii), let $x\in f^{-1}(s_j)$. By part (i), we conclude that there exists a neighborhood $V$ of $x$ that does not intersect $f^{-1}(s_i)$ for any $i\neq j$. Since $f^{-1}(s_j)$ is an embedded $1$-dimensional manifold, after shrinking $V$ if necessary, there exists a homeomorphism $\phi\colon V \to \mathbb R^2$ such that $\phi(f^{-1}(s_j)\cap V)=\mathbb R$; see Remark \ref{Straighten:Remark}. Consider small Jordan regions $ V_+ \subset \phi^{-1}(\mathbb R^2_+)$, $ V_- \subset \phi^{-1}(\mathbb R^2_-)$ such that the closure of $V_+\cup V_-$ contains a neighborhood of $x$. It follows that there exists an open Jordan arc $J\subset \partial V_+\cap \partial V_- \cap f^{-1}(s_j)$ containing $x$ such that $J$ is an essential boundary arc of $ V_+$ and $ V_-$. By (ii) we have that $x\in \partial f^{-1}((s_{j-1},s_{j}))\cup \partial f^{-1}((s_j,s_{j+1}))$. Hence, $V_+,V_-$ must intersect the open set $f^{-1}((s_{j-1},s_{j}))\cup f^{-1}((s_j,s_{j+1}))$. Since $V_+$ and $V_-$ are connected and they do not intersect the boundary of that set (by the choice of $V$), it follows that they are contained in connected components of $f^{-1}((s_{j-1},s_{j}))\cup f^{-1}((s_j,s_{j+1}))$. \end{proof} We let $u\colon \Omega\to \mathbb R $ be a continuous monotone function with $u\in W^{1,p}(\Omega)$, as in the statement of Theorem \ref{Intro:Approximation:Theorem}. In our proof we will follow the steps of \cite{IwaniecKovalevOnninen:W12approximation} and \cite{IwaniecKovalevOnninen:W1papproximation}. Namely, using Theorem \ref{Intro:MonotoneSobolev:Theorem}, we will first partition $\Omega$ into disjoint open subsets $\Upsilon_i$ and smooth the function $u$ in these subsets by replacing it with a $p$-harmonic function. This is the first step below. Next, we have to mollify the new function along the boundaries of $\Upsilon_i$. For this purpose, we partition even further $\Omega$, so that the boundaries of the regions $\Upsilon_i$ are contained in regions $\Psi_j$. Following \cite{IwaniecKovalevOnninen:W12approximation}, we call this new partition a \textit{lens-type partition}. In the regions $\Psi_j$ we consider another $p$-harmonic replacement. This completes the second step. In the third and final step we have to smooth our function along the boundaries of the regions $\Psi_j$, using smoothing results from Section \ref{Section:Smoothing}. Throughout all steps we have to ensure that our functions are monotone. This is guaranteed by the gluing lemmas from Subsection \ref{Section:glue} that allow us to paste together the various $p$-harmonic functions. \subsection{Step 1: Approximation by a piecewise smooth function} \subsubsection{Initial partition of $\Omega$}\label{InitialPartition:Section} For fixed $\delta>0$ we consider a bi-infinite sequence of real numbers $\{t_i\}_{i\in \mathbb Z}$ such that $t_i<t_{i+1}$, $|t_{i+1}-t_i|<\delta$ for $i\in \mathbb Z$, and $\lim_{i\to \pm\infty} t_i=\pm\infty$. Moreover, we may have that the conclusions of Theorem \ref{Sard:Theorem} are satisfied for the level sets $A_{t_i}$; that is, $A_{t_i}$ is an embedded $1$-dimensional topological submanifold of the plane. Note that $u^{-1}(t_i)$ or $u^{-1}((t_i,t_{i+1}))$ will be empty if $t_i\notin u(\Omega)$ or $(t_i,t_{i+1})\cap u(\Omega)=\emptyset$, respectively. \subsubsection{Solution of the Dirichlet Problem}\label{Dirichlet:Section} In each region $ \Upsilon_i =u^{-1}((t_i,t_{i+1}))$, using Proposition \ref{DirichletProblem:Prop}, we solve the Dirichlet Problem with boundary data $u$ and obtain a function $u_i\in W^{1,p}(\Upsilon_i)$. The function $u_i$ is monotone in $\Upsilon_i$ since it is $p$-harmonic. By Lemma \ref{Comparison:Lemma}, we have $t_i\leq u_i\leq t_{i+1}$. Moreover, if $x_0\in \partial \Upsilon_j \cap \Omega$, then $x_0 \in A_t$, where $t=t_j$ or $t=t_{j+1}$ by Lemma \ref{EssentialArcs:Lemma}(ii). Part (iii) of the same lemma implies that there exist disjoint Jordan regions $V_+, V_-$ such that the closure of $V_+\cup V_-$ contains a neighborhood of $x_0$ and there exists an open Jordan arc $J\subset \partial V_+\cap \partial V_-\cap A_t$ containing $x_0$ such that $J$ is an essential boundary arc of $ V_+$ and $ V_-$. Moreover, $V_+ \subset \Upsilon_j$ and $V_- \subset \Upsilon_i$ for some $i\in \mathbb Z$. By Lemma \ref{ContinuousExtension:Lemma} the functions $u_j,u_i$ extend continuously to $V_+\cup J$, $V_-\cup J$, respectively, and $u_j\big |J=t$, $u_i\big|J=t$. Since $\overline{V_+\cup V_-}$ contains a neighborhood of $x$, we have that $u_j$ extends continuously to $x_0$. Hence, $u_j$ extends continuously to each point of $\partial \Upsilon_j\cap \Omega$ and agrees there with $u$. Using Corollary \ref{Gluing:Corollary}, we ``glue" the functions $u_i$, $i\in \mathbb Z$, together with $u$, to obtain a continuous monotone function $u_\delta$ on $\Omega$. Note that $$u_\delta-u= \sum_{i\in \mathbb Z} [u_i-u]_0,$$ where $[u_i-u]_0 \in W^{1,p}_0(\Omega)$ denotes the extension of $u_i-u$ by $0$ outside $u^{-1}((t_i,t_{i+1}))$. The completeness of $W^{1,p}_0(\Omega)$ implies that $u_\delta-u\in W^{1,p}_0(\Omega)$. Therefore, by Proposition \ref{DirichletProblem:Prop} we conclude that $$E_p(u_\delta)\leq E_p(u)= \int_\Omega |\nabla u|^p.$$ If $u$ is not already $p$-harmonic, which we may assume, then the above inequality is strict by the uniqueness part of Proposition \ref{DirichletProblem:Prop}. Since, $t_i\leq u_\delta\leq t_{i+1}$ in $ A_{t_i,t_{i+1}}$, we have $$|u-u_\delta| \leq |u-t_i|+|t_i-u_\delta|\leq 2|t_{i+1}-t_i|<2\delta$$ in $\Omega$ and so $u_\delta$ is uniformly close to $u$. We also observe here that \begin{align}\label{Dirichlet:vu} u_\delta^{-1}((t_j,t_{j+1}))\subset A_{t_j,t_{j+1}}, \,\,\, j\in \mathbb Z, \end{align} which will be used later. This follows from the decomposition of $\Omega$ as $$\Omega= \bigcup_{i\in \mathbb Z} (A_{t_i,t_{i+1}} \cup A_{t_i})$$ and the observation that $u_\delta^{-1}((t_i,t_{i+1}))$ cannot intersect $A_{t_j}$ for any $i,j\in \mathbb Z$, since $u_\delta=u=t_j$ on $A_{t_j}$. Since $\Omega$ is bounded, it follows that $u_\delta \to u$ in $L^p(\Omega)$ as $\delta\to 0$. Moreover, $\|u_\delta\|_{W^{1,p}(\Omega)}= \|u_\delta\|_{L^p(\Omega)}+ \|\nabla u_\delta\|_{L^p(\Omega)}$ is uniformly bounded as $\delta\to 0$. By the Banach-Alaoglu theorem it follows that, along a sequence of $\delta\to 0$, $u_\delta$ converges weakly in $W^{1,p}(\Omega)$ to $u$, which is also the pointwise limit of $u_\delta$. Since the limits are unique, we do not need to pass to a subsequence. Hence, by Fatou's lemma for weak convergence \begin{align*} \|\nabla u\|_{L^p(\Omega)}\leq \liminf_{\delta\to 0} \|\nabla u_\delta\|_{L^p(\Omega)} \leq \limsup_{\delta\to 0} \|\nabla u_\delta\|_{L^p(\Omega)}. \end{align*} The latter is bounded by $E_p(u)^{1/p}=\|\nabla u\|_{L^p(\Omega)}$, hence $\lim_{\delta\to 0}\|\nabla u_\delta\|_{L^p(\Omega)} = \|\nabla u\|_{L^p(\Omega)}$. This implies that $\nabla u_\delta\to \nabla u$ in $L^p(\Omega)$, due to the uniform convexity of $L^p$ spaces, when $1<p<\infty$; see for example \cite[pp.~95--97]{Brezis:Functional} and \cite[Proposition 3.32, p.~78]{Brezis:Functional}. We remark that $u_\delta$ might be constant in some component of a level region $u^{-1}((t_i,t_{i+1}))$. Summarizing, for each $\varepsilon>0$ there exists $\delta_0>0$ such that for $0<\delta<\delta_0$ there exists a monotone function $u_\delta$ on $\Omega$ satisfying the following: \begin{enumerate}[(A)] \item $u_\delta$ is $p$-harmonic in $u^{-1}((t_i,t_{i+1}))$, $i\in \mathbb Z$, \item $\sup_{\Omega}|u_\delta-u|<\varepsilon$, \item $\|\nabla u_\delta -\nabla u\|_{L^p(\Omega)}<\varepsilon$, \item $E_p(u_\delta)< E_p(u)$, and \item $u_\delta-u \in W^{1,p}_0(\Omega)$. \end{enumerate} In the next step, our goal is to approximate in the sense of (B)--(E) the function $v\coloneqq u_\delta$, for a small fixed $\delta$, by functions that are smooth along the level sets $A_{t_i}$, $i\in \mathbb Z$; note that these are the level sets of the original function $u$! \subsection{Step 2: Smoothing along the level sets $u^{-1}(t_j)$} \subsubsection{Lens-type partition}\label{Lens:Section} Note that if a level set $v^{-1}(t)$ is a $1$-dimensional manifold, then it cannot intersect any component of $\bigcup_{i\in \mathbb Z}A_{t_i,t_{i+1}}$ where $v$ is constant. Since $v$ is $p$-harmonic in $\bigcup_{i\in \mathbb Z}A_{t_i,t_{i+1}}$, it has at most countably many critical points in each component of $A_{t_i,t_{i+1}}$, unless it is constant there; see Section \ref{Section:Preliminaries}. For $\eta>0$ and for each $j\in \mathbb Z$ we consider real numbers $t_j^-, t_j^+$ with $|t_j^+-t_j^-|<\eta$ such that $t_{j-1}^+<t_j^- <t_j<t_j^+<t_{j+1}^-$, $v^{-1}(t_j^{\pm})$ does not contain any critical points of $v$, and the conclusions of Theorem \ref{Sard:Theorem} hold for the level sets $v^{-1}(t_j^\pm)$. In particular, the level sets $v^{-1}(t_j^{\pm})$ intersect only components of $\bigcup_{i\in \mathbb Z}A_{t_i,t_{i+1}}$ in which $v$ is non-constant, and therefore the critical points of $v$ form a discrete set within these components. It follows that each point $x\in v^{-1}(t_j^\pm)$ has a neighborhood in $\bigcup_{i\in \mathbb Z}A_{t_i,t_{i+1}}$ that does not contain any critical points of $v$. In each region $\Psi_j\coloneqq v^{-1}( (t_{j}^-, t_j^+))$ we solve the Dirichlet Problem with boundary data $v$ as in Section \ref{Dirichlet:Section} and obtain $p$-harmonic functions $v_j$ with $t_{j}^-\leq v_j\leq t_j^+$ that extend continuously to $\partial \Psi_j\cap \Omega$ and agree there with $v$. We define a function $v_\eta$ to be equal to $v_j$ in $\Psi_j$, $j\in \mathbb Z$, and equal to $v$ on $\Omega\setminus \bigcup_{j\in \mathbb Z} \Psi_j$. This procedure results in a continuous monotone function $v_\eta$, by Corollary \ref{Gluing:Corollary}, that approaches $v$ in the uniform norm and also in $W^{1,p}(\Omega)$, as $\eta\to 0$. Moreover, we have $E_p(v_\eta)< E_p(v)$ (unless $v$ is $p$-harmonic in $\Omega$, which we may assume that is not the case) and $v_\eta-v\in W^{1,p}_0(\Omega)$. That is, (B)--(E) from Section \ref{Dirichlet:Section} are satisfied. \subsubsection{Regularity of the approximation}\label{Section:LipschitzRegularity} As far as the regularity of $w\coloneqq v_\eta$ is concerned, we claim the following: \begin{enumerate}[(a)] \item The function $w$ is $p$-harmonic in $\Psi_j$ and in $v^{-1}((t_j^+,t_{j+1}^-))$, for all $j\in \mathbb Z$. Therefore, $w$ is $C^{1,\alpha}$-smooth in each of these sets and if $p=2$, it is actually $C^\infty$-smooth. \item Each point $x_0\in v^{-1}(t_j^\pm)$, $j\in \mathbb Z$, has a neighborhood $V=V_+\cup V_-$ in $$\mathcal W\coloneqq \Omega \setminus \bigcup_{i\in \mathbb Z} v^{-1}(\{t_i^-,t_i^+\})= \bigcup_{i\in \mathbb Z} (\Psi_i \cup v^{-1}((t_i^+,t_{i+1}^-))),$$ where $V_+,V_-$ are disjoint Jordan regions contained in connected components of $\mathcal W$, $\overline V$ contains a neighborhood of $x_0$, and there exists a $C^\infty$-smooth open Jordan arc $J\subset \partial V_+\cap \partial V_- \cap v^{-1}(t_j^{\pm})$ containing $x_0$ such that $J$ is an essential boundary arc of $V_+$ and $V_-$. Moreover, inside each one of $V_+, V_-$, $|\nabla w|$ is either vanishing or bounded below away from $0$; in the second case we have $w\neq t_{j}^\pm$ in the corresponding region. The function $w$ is smooth up to the boundary arc $J$ in each of $V_+,V_-$. \item The function $w$ is $C^{\infty}$-smooth in $\mathcal W$, except possibly for a discrete set of points of $\mathcal W$ that accumulates at $\partial \Omega$. \end{enumerate} For (a) note that, by definition, $w$ is $p$-harmonic on $\Psi_j$, while on $v^{-1}((t_j^+ ,t_{j+1}^-)) \subset \Omega \setminus \bigcup_{i\in \mathbb Z} \Psi_i$ we have $w=v$ (by the definition of $w$) and $v$ is $p$-harmonic in $v^{-1}((t_j^+ ,t_{j+1}^-))=u_\delta^{-1}((t_j^+ ,t_{j+1}^-))\subset u^{-1}( (t_j,t_{j+1}))$; see \eqref{Dirichlet:vu}. Therefore, by the regularity of $p$-harmonic functions (see Section \ref{Section:Preliminaries}), $w$ is $C^{1,\alpha}$-smooth in the open set $\mathcal W$ and in fact $w$ is $C^\infty$-smooth in $\mathcal W$, except possibly for the (at most) countably many critical points that are contained in components $U$ of $\mathcal W$ in which $w$ is non-constant. The critical points form a discrete subset of such components $U$ of $\mathcal W$, so they could only accumulate in points of $\partial U$. Assuming that (b) is true, we will show (c). It suffices to show that the critical points contained in a component $U$ of $\mathcal W$ where $w$ is non-constant do not have any accumulation point in $\Omega$, but they can only accumulate at $\partial \Omega$. Suppose for the sake of contradiction that the critical points contained in $U$ accumulate at a point $x_0\in\partial U\cap \Omega$. Observe that the values $t_i^\pm$, $i\in \mathbb Z$, form a discrete set of real numbers with no finite accumulation points, and the level sets $v^{-1}(t_{i}^{\pm})$, $i\in \mathbb Z$, are embedded $1$-dimensional manifolds, as in the setting of Lemma \ref{EssentialArcs:Lemma}. By Lemma \ref{EssentialArcs:Lemma}(ii) there exists $j\in \mathbb Z$ such that $x_0 \in v^{-1}(\{t_j^-, t_j^+\})$. Consider the Jordan regions $V_+,V_- \subset \mathcal W$ as in (b) such that $\overline {V_+ \cup V_-}$ contains a neighborhood of $x_0$. This implies that there are infinitely many critical points of $w$ in, say, $V_+$, where $V_+\subset U$; note that at least one of $V_+,V_-$ has to be contained in $U$. Since $w$ is non-constant on $U\supset V_+$, the conclusion of (b) implies that $w$ does not have any critical points in $V_+$, a contradiction. Finally we prove (b). Let $x_0\in v^{-1}(t_j^-)$ for some $j\in \mathbb N$; the argument is the same if $x_0\in v^{-1}(t_j^+)$. By Lemma \ref{EssentialArcs:Lemma}(iii) we conclude that there exist disjoint Jordan regions $V_+,V_-$ such that the closure of $V_+\cup V_-$ contains a neighborhood of $x_0$ and there exists an open Jordan arc $J\subset \partial V_+\cup \partial V_- \cap v^{-1}(t_j^-) $ containing $x_0$ that is an essential boundary arc of $V_+$ and $V_-$. Moreover, the last statement of Lemma \ref{EssentialArcs:Lemma}(iii) implies that $V_+,V_-$ are contained in connected components of $ v^{-1}(( t_{j-1}^+,t_j^-))\cup v^{-1}((t_j^-,t_j^+))$. We will show that $|\nabla w|$ is vanishing or bounded below in each of $V_+,V_-$ (after possibly shrinking them) and that $w$ is smooth up to the arc $J$ in $V_+$ and $V_-$. We work with $V_+$. Let $U$ be the component of $v^{-1}(( t_{j-1}^+,t_j^-))\cup v^{-1}((t_j^-,t_j^+))$ containing $V_+$. If $w$ is constant in $U$ then $w$ is obviously smooth up to the arc $J$ inside $V_+$ and we have nothing to show. We suppose that $w$ is non-constant in $U$. Assume first that $V_+\subset U$ is contained in $v^{-1}((t_{j-1}^+,t_{j}^-))$. Then $w=v<t_j^-$ on $V_+$ by the definition of $w$. By the choice of $t_j^-$ in Section \ref{Lens:Section}, each point of $ v^{-1}(t_j^-)\supset J$ has a neighborhood in $\bigcup_{i\in \mathbb Z}A_{t_i,t_{i+1}}$ that does not contain any critical points of $v$. In particular, each point of the arc $J$ has a neighborhood in which $v$ is non-constant, $p$-harmonic, and $|\nabla v|$ is non-zero. It follows that $v$ is smooth in a neighborhood of $J$; see Section \ref{Section:Preliminaries}. Therefore, after shrinking the Jordan region $V_+$ if necessary, $w$ is smooth in $V_+$ up to the arc $J$, and each point of $J$ has a neighborhood in $V_+$ in which $|\nabla w|=|\nabla v|$ is bounded away from $0$. Suppose now that $V_+\subset U$ is contained in $\Psi_j=v^{-1}((t_j^-,t_j^+))$. We have $J\subset v^{-1}(t_j^-)$, so $w=v=t_j^-$. Since $w$ is $p$-harmonic on $V_+\subset U$ and non-constant, we must have $w>t_j^-$ on $V_+$ by the strong maximum principle \cite[p.~111]{Heinonenetal:DegenerateElliptic}. Note that the arc $J$ is $C^\infty$-smooth, because $J\subset v^{-1}(t_j^-)$ and the $t_j^-$ is a regular (i.e., non-critical) value of $v$ by the choice of $t_{j}^{-}$. Now we are exactly in the setting of Corollary \ref{NonDegeneracy:Corollary}, which implies that each point of $J$ has a neighborhood in $V_+$ that does not contain any critical points of $w$ and $|\nabla w|$ is bounded away from $0$. If we shrink the Jordan region $V_+$, we may have that these hold in $V_+$. By Corollary \ref{NonDegeneracy:Corollary} it also follows that $w$ is smooth up to the boundary in the region $V_+$. The proof of (b) is completed.\qed \vspace{0.5em} We have managed to obtain a function $w$ that is smooth along the level sets $u^{-1}(t_j)$, but it is not necessarily smooth along $v^{-1}(t_j^{\pm})$. It has, however, some smoothness up to $v^{-1}(t_j^\pm)$, as described in (b). In the next subsection we prove that $w$ can be $C^\infty$-smoothed at arbitrarily small neighborhoods of the level sets $v^{-1}(t_{j}^{\pm})$ so as to complete the proof of Theorem \ref{Intro:Approximation:Theorem}. For this purpose we utilize smoothing results from Section \ref{Section:Smoothing}. \subsection{Step 3: Smoothing along the level sets $v^{-1}(t_j^{\pm})$} We fix $j\in \mathbb Z$ and a component $J$ of $v^{-1}(t_j^-)$. Recall that $J$ is a smooth $1$-manifold, since $t_j^-$ was chosen to be a regular value of $v$, and $v$ is $C^\infty$-smooth in a neighborhood of $v^{-1}(t_j^-)$. There are two cases: either $J$ is homeomorphic to $\mathbb R$ or it is homeomorphic to $S^1$ (see the classification in Theorem \ref{Classification:theorem}). By claim (b) from Subsection \ref{Section:LipschitzRegularity}, each $x_0\in J$ has a neighborhood $V=V_+\cup V_-$ in $\Omega\setminus v^{-1}(t_j^-)$, where $V_+,V_-$ are disjoint Jordan regions, $\overline V$ contains a neighborhood of $x_0$ and there exists an open arc $J_0\subset J$ containing $x_0$ that is contained in the common boundary of $V_+$ and $V_-$. Moreover, $w$ is $C^\infty$-smooth in ${ V_+}\cup J_0$ and in ${V_-}\cup J_0$, and $|\nabla w|$ is either vanishing or bounded below away from $0$ in each of $V_+, V_-$; in the second case we have $w\neq t_j^-$ in the corresponding set. Assume first that $J$ is homeomorphic to $\mathbb R$. By Theorem \ref{Classification:theorem}, there exists a neighborhood $U$ of $J$ in $\mathbb R^2$ and a diffeomorphism $\phi\colon \mathbb R^2\to U$ such that $\phi(\mathbb R)=J$. The previous paragraph implies that there exist connected neighborhoods $\widetilde B^+$ and $\widetilde B^-$ of $\mathbb R$ in $\{y>0\}$ and $\{y<0\}$, respectively, such that $w\circ \phi$ is smooth in $\widetilde B^+ \cup \mathbb R$ and in $\widetilde B^-\cup \mathbb R$, and each point of $\mathbb R$ has a neighborhood in $\widetilde B^+\cup\widetilde B^-$ in which $|\nabla (w\circ \phi)|$ is either vanishing or bounded away from $0$. Moreover, in each of $\widetilde B^+$ and $\widetilde B^-$ the function $w\circ \phi$ is either constant or we have $w\circ \phi > t_j^-$ or $w\circ \phi<t_j^-$ . Therefore, there exist disjoint connected open sets $B^+=\phi^{-1}(\widetilde B^+)$ and $B^-=\phi^{-1}(\widetilde B^-)$ so that $B=B^+\cup B^-\cup J$ is an open set containing $J$ and $w$ satisfies one of the following conditions, with the roles of $B^+$,$B^-$ possibly reversed: \begin{enumerate} \item[(i)] $w=t_j^-$ on $J$, $w>t_j^-$ on $B^+$, $w<t_j^-$ on $B^-$, \item[(i$'$)] $w=t_j^-$ on $J$, $w>t_j^-$ on $B^+$, $w=t_j^-$ on $B^-$, \item[(i$''$)] $w=t_j^-$ on $J$, $w>t_j^-$ on $B^+$, $w>t_j^-$ on $B^-$, \item[(i$'''$)] $w=t_j^-$ on $J$, $w=t_j^-$ on $B^+$, $w=t_j^-$ on $B^-$. \end{enumerate} Note that (i$'''$) immediately implies that $w$ is smooth at each point of $J$. If (i) holds, then $|\nabla w|$ must be non-zero in $B^+\cup B^-$, hence each point of $J$ has a neighborhood in $B^+\cup B^-$ in which $|\nabla w|$ is bounded away from $0$. We are now exactly in the setting of the smoothing Lemma \ref{SmoothingArc:Lemma}. Thus, there exists a monotone function $\widetilde w$ in $\Omega$ that is smooth in $B$, agrees with $w$ outside an arbitrarily small neighborhood $A$ of $J$, and is arbitrarily close to $w$ in the uniform and in the Sobolev norm. In particular, since $E_p(w)<E_p(u)$ (cf.\ (D) from Subsection \ref{Dirichlet:Section}) we may have $E_p(\widetilde w) < E_p(u)$. Finally, we have $\widetilde w -w \in W^{1,p}_0(\Omega)$ from Lemma \ref{SmoothingArc:Lemma}(d) as claimed in (E) in the statement of Theorem \ref{Intro:Approximation:Theorem}. If (i$'$) or (i$''$) holds instead, then we are in the setting of Lemma \ref{SmoothingArc:Lemma2} and obtain the same conclusions. Now, if $J$ is homeomorphic to $S^1$ we can find connected open sets $B^+$, $B^-$ so that $J$ is their common boundary. In this case, we have the following alternatives: \begin{enumerate} \item[(i)] $w=t_j^-$ on $J$, $w>t_j^-$ on $B^+$, $w<t_j^-$ on $B^-$, \item[(i$'$)] $w=t_j^-$ on $J$, $w>t_j^-$ on $B^+$, $w=t_j^-$ on $B^-$, \item[(i$'''$)] $w=t_j^-$ on $J$, $w=t_j^-$ on $B^+$, $w=t_j^-$ on $B^-$. \end{enumerate} We note that the alternative (i$''$) of the previous case does not occur here, since it would violate the monotonicity of $w$. We can now apply Lemma \ref{SmoothingCurve:Lemma} to smooth the function $w$ near $J$. We apply the above smoothing process countably many times in pairwise disjoint, arbitrarily small neighborhoods of components of $v^{-1}(t_j^{\pm})$, $j\in \mathbb Z$. Since these components do not accumulate in $\Omega$ (see Lemma \ref{EssentialArcs:Lemma}(i)), the resulting limiting function $\widetilde w$ will be $C^\infty$-smooth in $\Omega$, except possibly at a discrete set of points of $\Omega$ (by (a),(c) in Subsection \ref{Section:LipschitzRegularity}), in which $\widetilde w$ is $C^{1,\alpha}$-smooth; if $p=2$ then $\widetilde w$ is $C^\infty$-smooth everywhere. If the neighborhoods of $v^{-1}(t_j^{\pm})$, $j\in \mathbb Z$, where the smoothing process takes place are arbitrarily small, then $\widetilde w$ will be $p$-harmonic on a subset of $\Omega$ having measure arbitrarily close to full. Moreover, by Lemma \ref{ConvergenceMonotonicity:Lemma}, $\widetilde w$ is monotone in $\Omega$. Finally, it satisfies the corresponding conditions (B)--(E) from Subsection \ref{Dirichlet:Section}. This completes the proof of Theorem \ref{Intro:Approximation:Theorem}. \qed \section{Smoothing results}\label{Section:Smoothing} \begin{theorem}\label{Classification:theorem} Let $J$ be an embedded $1$-dimensional smooth submanifold of $\mathbb R^2$ that is connected. Then $J$ is diffeomorphic to $X$, where $X=\mathbb R\times \{0\}\subset \mathbb R^2$ or $X=S^1\subset \mathbb R^2$. Moreover, there exists a neighborhood $U$ of $J$ in $\mathbb R^2$ and a diffeomorphism $\phi\colon \mathbb R^2\to U$ such that $\phi(X)=J$. \end{theorem} We direct the reader to \cite{Viro:1manifolds} for an account on the classification of $1$-manifolds. The last statement of the lemma can be proved in the same way as the existence of tubular neighborhoods of embedded submanifolds of $\mathbb R^n$; see for example \cite[Theorem 6.24, p.~139]{Lee:manifolds}. \begin{lemma}[Smoothing along a line] \label{SmoothingArc:Lemma} Let $\Omega\subset \mathbb R^2$ be an open set and let $J \subset \Omega$ be an embedded $1$-dimensional smooth submanifold of $\mathbb R^2$, homeomorphic to $\mathbb R$, such that $J$ has no accumulation points in $\Omega$; that is, the topological ends of $J$ lie in $\partial \Omega$. Suppose that $J$ is contained in an open set $B\subset \Omega$ that it is the common boundary of two disjoint regions $B^+$ and $B^-$ with $B=B^+\cup B^-\cup J$. Moreover, let $u\in W^{1,p}(\Omega)$ be a monotone function with the following properties: \begin{enumerate}[\upshape(i)] \item \label{SmoothingArc:Lemma:sign} $u=0$ on $J$, $u>0$ on $B^+$, $u<0$ on $B^-$, \item \label{SmoothingArc:Lemma:smooth} $u$ is smooth in $B^+\cup J$ and in $B^-\cup J$, \item \label{SmoothingArc:Lemma:nabla} each point of $J$ has a neighborhood in $B^+\cup B^-$ in which $|\nabla u|$ is bounded away from $0$. \end{enumerate} Then for any open set $U\subset B$ with $U\supset J$ and each $\varepsilon>0$ there exists an open set $A\subset U$ containing $J$ and a monotone function $\widetilde u$ in $\Omega$ such that \begin{enumerate}[\upshape(a)] \item $\widetilde u$ agrees with $u$ in $\Omega\setminus A$ and is smooth in $B$, \item $|\widetilde u- u|<\varepsilon$ in $A$, \item $\|\nabla \widetilde u-\nabla u\|_{L^p(A)}<\varepsilon$, and \item $\widetilde u-u\in W^{1,p}_{0}(A)$. \end{enumerate} \end{lemma} \begin{proof} By using a diffeomorphism $\phi$ defined in a neighborhood of $J$ in $B$, given by Theorem \ref{Classification:theorem}, we may straighten $J$ and assume that $J=\mathbb R=\{(x,y):y=0\}$, $B=\mathbb R^2$, $B^+=\{(x,y):y>0\}$, $B^-=\{(x,y):y<0\}$. For each $\varepsilon>0$ we will construct a smooth function $\widetilde u$ that agrees with $u$ outside an arbitrarily small neighborhood $A$ of $\mathbb R$, $|\widetilde u-u|<\varepsilon$ in $A$ and $\|\nabla \widetilde u -\nabla u\|_{L^q(A)}<\varepsilon$, where $q>p$ is a fixed exponent (e.g., $q=2p$). If the neighborhood $A$ is sufficiently thin, then $\widetilde u-u\in W^{1,q}_0(A)$. Pulling back $\widetilde u$ under the diffeomorphism $\phi$ and using H\"older's inequality will yield the desired conclusions in $\Omega$. We argue carefully for the monotonicity of $\widetilde u$ in the end of the proof. For the construction of $\widetilde u$, essentially, we are going to interpolate between the function $u$ away from $\mathbb R$ and the ``height function" $(x,y)\mapsto y$ near $\mathbb R$. Several technicalities arise though, in order to ensure that the new function is monotone. By assumption \ref{SmoothingArc:Lemma:nabla}, for each $x\in \mathbb R$ there exists a constant $m>0$ such that $|\nabla u |>m$ in a neighborhood of $x$ in $\mathbb R^2\setminus \mathbb R$. Since $u=0$ on $\mathbb R$, we have $u_x=0$ on $\mathbb R$. Hence, using the smoothness of $u$ in $B^+\cup \mathbb R$ and in $B^-\cup \mathbb R$ we conclude that for each point $x\in \mathbb R$ there exist constants $m,M>0$ such that $M>|u_y|>m$ in a neighborhood of $x$ in $\mathbb R^2\setminus \mathbb R$. Thus, we may consider positive smooth functions $\gamma,\delta \colon \mathbb R \to \mathbb R$ such that $\delta(x)>|u_y(x,y)|>\gamma(x)>0$ for all $(x,y)$ lying in a small neighborhood of $\mathbb R$ in $\mathbb R^2\setminus \mathbb R$. By assumptions \ref{SmoothingArc:Lemma:sign} and \ref{SmoothingArc:Lemma:smooth}, we have $u_y\geq 0$ in a neighborhood of $\mathbb R$ in $\mathbb R^2\setminus \mathbb R$. Therefore, \begin{align}\label{SmoothingArc:Lemma:gammadelta} \delta(x)>u_y(x,y)>\gamma(x)>0 \end{align} for all $(x,y)$ lying in a neighborhood of $\mathbb R$ in $\mathbb R^2\setminus \mathbb R$. By choosing a possibly larger $\delta$ we may also have \begin{align}\label{SmoothingArc:Lemma:gammadelta2} \delta(x)>|u_x(x,y)| \end{align} in that neighborhood. We let $A$ be a sufficiently small open neighborhood of $\mathbb R$ in $\mathbb R^2$ so that the preceding inequalities hold. Later we will shrink $A$ even further. For a positive smooth function $\beta\colon \mathbb R\to \mathbb R$ we define $V(\beta)=\{(x,y): |y|<\beta(x)\}$. Note that we may choose $\beta$ so that $\overline{V(\beta)}\subset A$. By scaling $\beta$, we may also have that \begin{align}\label{SmoothingArc:Lemma:beta} |\beta'|\leq 1. \end{align} Moreover, we consider a non-negative smooth function $\alpha\colon \mathbb R\to \mathbb R$ with $\alpha(t)=1$ for $t\geq 1$, $\alpha(t)=0$ for $t\leq 1/2$, and \begin{align}\label{SmoothingArc:Lemma:alpha} 0\leq \alpha'\leq 4. \end{align} We now define $s=s(x,y)= |y|/\beta(x)$ and \begin{align*} \widetilde u(x,y)= \alpha(s)u(x,y)+ (1-\alpha(s))y\gamma(x). \end{align*} This function is smooth, agrees with $u$ outside $V(\beta)$ (where $s\geq 1$), and agrees with $(x,y)\mapsto y\gamma(x)$ in $V(\beta/2)$ (where $0\leq s<1/2$). We have $|\widetilde u-u| \leq |1-\alpha(s)| |u-y\gamma|\leq |u|+ |y|\gamma$. By \eqref{SmoothingArc:Lemma:gammadelta} in $A$ we have \begin{align}\label{SmoothingArc:Lemma:u bound} |u(x,y)| = \left| \int_0^y u_y(x,t)dt \right| \leq |y|\delta(x). \end{align} Therefore, $|u|+|y|\gamma < |y|(\gamma+\delta)$. If $A$ is so small that it is contained in the open set $\{(x,y): |y|<\varepsilon/(\gamma(x)+\delta(x))\}$, then we have $|\widetilde u-u|<\varepsilon$ in $A$, which proves claim (b). We now compute the derivatives: \begin{align*} \widetilde u_x&= -\alpha'(s) \beta'(x) s^2 \frac{u-y\gamma}{|y|} +\alpha(s)u_x + (1-\alpha(s)) y\gamma' \,\,\,\,\, \textrm{and}\\ \widetilde u_y&= \alpha'(s)s \frac{u-y\gamma}{y}+ \alpha(s)u_y+(1-\alpha(s))\gamma. \end{align*} By \eqref{SmoothingArc:Lemma:beta}, \eqref{SmoothingArc:Lemma:alpha}, and \eqref{SmoothingArc:Lemma:u bound} we have \begin{align*} |\widetilde u_x| &\leq 4(\gamma+\delta)+\delta +|y||\gamma'| \,\,\,\,\, \textrm{and}\\ |\widetilde u_y| & \leq 4(\gamma+\delta)+ \delta+ \gamma \end{align*} in $V(\beta)\subset A$. Since the above bounds and \eqref{SmoothingArc:Lemma:gammadelta}, \eqref{SmoothingArc:Lemma:gammadelta2} do not depend on $A$ but only on the function $u$, if $A$ is sufficiently small, then $\|\nabla \widetilde u- \nabla u\|_{L^q(A)}\leq \|\nabla \widetilde u\|_{L^q(V(\beta))}+ \| \nabla u\|_{L^q(V(\beta))}$ can be made as small as we wish. We next prove that $\widetilde u$ is monotone in $\mathbb R^2$. Note that $\frac{u-y\gamma}{y} \geq 0$ on $A$, since $u_y>\gamma$. Therefore, \begin{align*} \widetilde u_y \geq \gamma>0. \end{align*} This implies that $\nabla \widetilde u \neq 0$ in $A$, so $\widetilde u$ does not have any local maximum or minimum in $A$, and it is strictly monotone there. On the other hand, $\widetilde u =u$ outside $\overline{V(\beta)}\subset A$, and therefore $\widetilde u$ is monotone outside $V(\beta)$. By the Gluing Lemma \ref{GluingMonotonicity:Lemma} it follows that $\widetilde u$ is monotone in $\mathbb R^2$. The argument of the previous paragraph and the use of Lemma \ref{GluingMonotonicity:Lemma} can be carried in $\Omega$, after precomposing $\widetilde u$ with the straightening diffeomorphism $\phi^{-1}$. We denote the composition, for simplicity, by $\widetilde u$. So $\widetilde u$ is a function in $\Omega$ that is strictly monotone in $\phi^{-1}(A)$ and can be extended to agree with $u$ in $\Omega\setminus \overline{\phi^{-1}(V(\beta))}$; in particular, $\widetilde u$ is monotone in $\Omega\setminus \overline{\phi^{-1}(V(\beta))}$. If we ensure that $\overline{\phi^{-1}(V(\beta))}\cap \Omega \subset \phi^{-1}(A)$, then Lemma \ref{GluingMonotonicity:Lemma} can be applied to yield the monotonicity of $\widetilde u$ in $\Omega$. The latter can be achieved by shrinking $\beta$ (and thus the neighborhood $V(\beta)$) if necessary, using the assumption that the topological ends of $J$ lie in $\partial \Omega$ and not in $\Omega$. \end{proof} \begin{lemma}\label{SmoothingArc:Lemma2} The conclusions of Lemma \ref{SmoothingArc:Lemma} hold if the assumptions \textup{(i)},\textup{(iii)} are replaced by the assumptions \begin{enumerate}\upshape \item[(i$'$)] $u=0$ on $J$, $u>0$ on $B^+$, $u=0$ on $B^-$, and \item[(iii$'$)] each point of $J$ has a neighborhood in $B^+$ in which $|\nabla u|$ is bounded away from $0$, \end{enumerate} or if \textup{(i)} is replaced by the assumption \begin{enumerate}\upshape \item[(i$''$)] $u=0$ on $J$, $u>0$ on $B^+$, $u>0$ on $B^-$. \end{enumerate} \end{lemma} \begin{proof} In the first case, one argues as in the proof of Lemma \ref{SmoothingArc:Lemma}, by straightening the line $J$ to the real line with a diffeomorphism $\phi$. This time we interpolate between $u$ and the function $e^{-1/y}$ instead of the height function $y$: \begin{align*} \widetilde u(x,y)= \begin{cases} \alpha(s)u(x,y)+ (1-\alpha(s)) e^{-1/y} \gamma(x) & y>0\\ u & y\leq 0. \end{cases} \end{align*} Here $\gamma$, $\delta$, $\beta$, $s$, $\alpha(s)$ are as in the previous proof, but working only in the upper half plane, and $A$, $V(\beta)=\{(x,y):0<y<\beta(x)\}$ are appropriate neighborhoods of $\mathbb R$ in the upper half plane $\{y>0\}$. The function $\widetilde u$ is smooth and the claims (a)--(d) from Lemma \ref{SmoothingArc:Lemma} follow as in the previous proof. We only have to argue differently for the monotonicity of $\widetilde u$. We compute for $y>0$ $$\widetilde u_y= \alpha'(s) s \frac{u-e^{-1/y}\gamma}{y}+ \alpha(s) u_y + (1-\alpha(s)) \frac{e^{-1/y}}{y^2}\gamma.$$ Since $u_y>\gamma$ on $A$, we have $\frac{u-e^{-1/y}\gamma}{y} \geq 0$ for all sufficiently small $y$. In particular this holds in $A$ if we shrink $A$. It follows that $\widetilde u_y>0$ on $A$. Thus, $\widetilde u$ is strictly monotone in $A$. Moreover, outside $\overline {V(\beta)}\cap \{y\neq 0\} \subset A$ the function $\widetilde u$ agrees with $u$. Summarizing, in the domain $\widetilde\Omega =\mathbb R^2\setminus \mathbb R=\{y\neq 0\}$ we have $\overline{V(\beta)} \cap \widetilde \Omega \subset A$ and the function $\widetilde u$ is strictly monotone in $A$, and monotone in $\widetilde \Omega\setminus \overline{V(\beta)}$. By Lemma \ref{GluingMonotonicity:Lemma} we have that $\widetilde u$ is monotone in $\widetilde \Omega$. Arguing as in the previous proof, we transfer the conclusions to $\Omega$ and deduce that $\widetilde u$ is monotone in $\Omega\setminus J$. Note that $J$ is a connected closed subset of $\Omega$. Since $\widetilde u=0$ in $J$ and $J$ exists all compact subsets of $\Omega$, by Lemma \ref{GluingConstant:Lemma} we conclude that $\widetilde u$ is monotone in $\Omega$. This completes the proof under the assumptions (i$'$) and (iii$'$). Now, we assume that (i$'')$ holds in the place of (i). After straightening $J$, we define \begin{align*} \widetilde u(x,y)= \begin{cases} \alpha(s)u(x,y)+ (1-\alpha(s)) e^{-1/|y|} \gamma(x) & y\neq 0\\ 0 & y= 0 \end{cases} \end{align*} This is a smooth function in $\mathbb R^2$. The conclusions (a)--(d) are again straightforward, so we only argue for the monotonicity. In $V(\beta)= \{(x,y): 0<|y|<\beta(x)\}$ and in a slightly larger open set $A\supset \overline{V(\beta)}$ we have $\widetilde u_y\neq 0$. This implies that $\widetilde u$ is strictly monotone in $A$. Using Lemma \ref{GluingMonotonicity:Lemma}, we have that $\widetilde u$ is monotone in $\mathbb R^2\setminus \mathbb R$. We transfer the conclusions to $\Omega$ and we obtain a function $\widetilde u$ that is monotone in $\Omega\setminus J$. Since $J$ is closed, connected and exits all compact subsets of $\Omega$, and $\widetilde u=0$ on $J$, by Lemma \ref{GluingConstant:Lemma} we conclude that $\widetilde u$ is monotone in $\Omega$. \end{proof} \begin{lemma}[Smoothing along a Jordan curve] \label{SmoothingCurve:Lemma} Let $\Omega\subset \mathbb R^2$ be an open set and let $J \subset \Omega$ be an embedded $1$-dimensional smooth submanifold of $\mathbb R^2$, homeomorphic to $S^1$. Suppose that $J$ is contained in an open set $B\subset \Omega$ that it is the common boundary of two disjoint regions $B^+$ and $B^-$ with $B=B^+\cup B^-\cup J$. Moreover, let $u\in W^{1,p}(\Omega)$ be a monotone function satisfying one of the following triples of conditions: \begin{enumerate}[\upshape(i)] \item $u=0$ on $J$, $u>0$ on $B^+$, $u<0$ on $B^-$, \item $u$ is smooth in $B^+\cup J$ and in $B^-\cup J$, \item each point of $J$ has a neighborhood in $B^+\cup B^-$ in which $|\nabla u|$ is bounded away from $0$, \end{enumerate} or \begin{enumerate}[\upshape(i$'$)] \item $u=0$ on $J$, $u>0$ on $B^+$, $u=0$ on $B^-$, \item $u$ is smooth in $B^+\cup J$ and in $B^-\cup J$, \item each point of $J$ has a neighborhood in $B^+$ in which $|\nabla u|$ is bounded away from $0$. \end{enumerate} Then for any open set $U\subset B$ with $U\supset J$ and each $\varepsilon>0$ there exists an open set $A\subset U$ containing $J$ and a monotone function $\widetilde u$ in $\Omega$ such that \begin{enumerate}[\upshape(a)] \item $\widetilde u$ agrees with $u$ in $\Omega\setminus A$ and is smooth in $B$, \item $|\widetilde u- u|<\varepsilon$ in $A$, \item $\|\nabla \widetilde u-\nabla u\|_{L^p(A)}<\varepsilon$, and \item $\widetilde u-u \in W^{1,p}_0(A)$. \end{enumerate} \end{lemma} \begin{proof} We will only sketch the differences with the proofs of Lemmas \ref{SmoothingArc:Lemma} and \ref{SmoothingArc:Lemma2}. By Theorem \ref{Classification:theorem} we may map $B$ with a diffeomorphism to a neighborhood of $S^1$. After shrinking $B$, we may assume that this neighborhood of $S^1$ is an annulus. Then, using logarithmic coordinates, we map $S^1$ to $\mathbb R$. Let $\phi$ denote the composition of the two maps described. By, precomposing $u$ with $\phi^{-1}$, we can obtain a periodic function, still denoted by $u$, in a strip $\{(x,y):|y|< c\}$. We consider a strip $V(\beta)= \{ |y|<\beta\}$ under assumption (i), or $V(\beta)=\{0<y<\beta\}$ under (i$'$), where $\beta$ is a constant rather than a function, and a strip $A \supset \overline {V(\beta)}$. We consider the function $\widetilde u$ with the same definition as in the proofs of Lemmas \ref{SmoothingArc:Lemma} and \ref{SmoothingArc:Lemma2}. The function $\widetilde u$ is also periodic and smooth, so it gives a function in the original domain $\Omega$, by extending it to be equal to $u$ outside $B$. We still denote this function by $\widetilde u$. The properties (a)--(d) are straightforward to obtain, upon shrinking the strip $A$. We only argue for the monotonicity. Under the first set of assumptions, and in particular under (i), the function $\widetilde u$ is strictly monotone in $\phi^{-1}(A)$ and agrees with $u$ outside $\phi^{-1}(\overline{V(\beta)}) \subset A$. Hence, by Lemma \ref{GluingMonotonicity:Lemma}, $\widetilde u$ is monotone in $\Omega$. Under the second set of assumptions, and in particular under (i$'$), we can only conclude that $\widetilde u$ is monotone in $\Omega\setminus J$; see also the proof of Lemma \ref{SmoothingArc:Lemma2}. However, now we cannot apply Lemma \ref{GluingConstant:Lemma}, since $J$ does not exit all compact subsets of $\Omega$. Instead, we are going to use Lemma \ref{Gluing:Lemma}. In $\phi^{-1}(V(\beta))$, which is precompact, by continuity we have $0<\widetilde u< t$ and $0<u<t$ for some $t>0$. We set $\Upsilon= u^{-1}((0,t))$, so we have $\phi^{-1}(V(\beta))\subset \Upsilon$. Observe that $\widetilde u$ is monotone in $\Upsilon \subset \Omega\setminus J$ and $0\leq \widetilde u\leq t$ in $\Upsilon$. By Lemma \ref{Gluing:Lemma} we conclude that the function $$\widetilde u= \widetilde u \x_{\Upsilon} + u \x_{\Omega\setminus \Upsilon}$$ is monotone in $\Omega$. \end{proof}

https://arxiv.org/abs/1406.0768

Toeplitz Lemma, Complete Convergence and Complete Moment Convergence

In this paper, we study the Toeplitz lemma, the Cesàro mean convergence theorem and the Kronecker lemma. At first, we study "complete convergence" versions of the Toeplitz lemma, the Cesàro mean convergence theorem and the Kronecker lemma. Two counterexamples show that they can fail in general and some sufficient conditions for "complete convergence" version of the Cesàro mean convergence theorem are given. Secondly we introduce two classes of complete moment convergence, which are stronger versions of mean convergence and consider the Toeplitz lemma, the Cesàro mean convergence theorem, and the Kronecker lemma under these two classes of complete moment convergence.

\section{Introduction} The Toeplitz lemma and its two corollaries (the Ces\`{a}ro mean convergence theorem and the Kronecker lemma) are useful tools in the study of probability limit theorems. For the reader's convenience, we spell out them in the following and their proofs may be found in Lo\`{e}ve (1977). \begin{thm} {\rm (Toeplitz lemma)} Let $\{a_{nk},1\leq k\leq k_n,n\ge1 \}$ be a double array of real numbers such that for any $k\geq 1, \lim_{n\to\infty}a_{nk}=0$ and $\sup_{n\geq 1}\sum_{k=1}^{k_n}|a_{nk}|<\infty$. Let $\{x_n,n\geq 1\}$ ba a sequence of real numbers.\\ (i) If $\lim_{n\to\infty}x_n=0$, then $\lim_{n\to\infty}\sum_{k=1}^{k_n}a_{nk}x_k=0.$\\ (ii) If $\lim_{n\to\infty}x_n=x\in \mathbf{R}$ and $\lim_{n\to\infty}\sum_{k=1}^{k_n}a_{nk}=1$, then $\lim_{n\to\infty}\sum_{k=1}^{k_n}a_{nk}x_k=x$. \end{thm} \begin{cor}{\rm (Ces\`{a}ro mean convergence theorem)} Let $\{x_n,n\geq 1\}$ be a sequence of real numbers and let $\bar{x}_n=\sum_{k=1}^nx_k/n,n\geq 1$. If $\lim_{n\to\infty}x_n=x\in \mathbf{R}$, then $\lim_{n\to\infty}\bar{x}_n=x$. \end{cor} \begin{cor}{\rm (Kronecker lemma)} Let $\{x_n,n\geq 1\}$ and $\{b_n,n\geq 1\}$ be sequences of real numbers such that $0<b_n\uparrow\infty$. If the series $\sum_{k=1}^{\infty}x_k/b_k$ converges, then $\lim_{n\to\infty}\frac{1}{b_n}\sum_{k=1}^nx_k=0$. \end{cor} By the definition of almost sure (a.s.) convergence, we know that the Toeplitz lemma and its two corollaries (the Ces\`{a}ro mean convergence theorem and the Kronecker lemma) still hold when the numerical sequence $\{x_n,n\geq 1\}$ and real number $x$ are replaced by a sequence of random variable $\{X_n,n\geq 1\}$ and a random variable $X$, respectively, and the limit is taken to be a.s. convergence. Recently, Linero and Rosalsky (2013) showed among other things that ``convergence in probability" versions of the Toeplitz lemma, the Ces\`{a}ro mean convergence theorem and the Kronecker lemma can fail, and their ``mean convergence" versions are true. We know that both a.s. convergence and mean convergence imply convergence in probability. Then there are the following two questions: {\it Question 1. Do the Toeplitz lemma, the Ces\`{a}ro mean convergence theorem and the Kronecker lemma hold under stronger convergence than a.s. convergence?} {\it Question 2. Do the Toeplitz lemma, the Ces\`{a}ro mean convergence theorem and the Kronecker lemma hold under stronger convergence than mean convergence?} In Section 2, we study Question 1 and consider ``complete convergence " versions of the Toeplitz lemma, the Ces\`{a}ro mean convergence theorem and the Kronecker lemma. At first, we will give two examples to show that they can fail in general. Then we give some sufficient conditions for the Ces\`{a}ro mean convergence theorem under complete convergence. Let $\{X, X_n,n\geq 1\}$ be a sequence of random variables on some probability space $(\Omega,\mathcal{F},P)$. If $\forall \varepsilon>0$, $$ \sum_{n=1}^{\infty}P\{|X_n-X|\geq \varepsilon\}<\infty, $$ then $\{X_n,n\geq 1\}$ is said to converge completely to $X$ (write $X_n\stackrel{c.c.}{\longrightarrow}X$, or $X_n\to X$ c.c. for short). This concept was introduced by Hsu and Robbins (1947). Let $\{X,X_n,n\geq 1\}$ be a sequence of independent and indentically distributed (i.i.d.) random variables and set $S_n=\sum_{k=1}^nX_k,n\geq 1$. Hsu and Robbins (1947) proved that if $E[X]=0$ and $E[X^2]<\infty$, then $S_n/n\stackrel{c.c.}{\longrightarrow}0$. The converse was proved by Erd\"{o}s (1949, 1950). The Hsu-Robbins-Erd\"{o}s theorem was generalized in various ways, see, Baum and Katz (1965), Gut (1978, 1980), Li et al. (1995), Lanzinger (1998), Sung and Volodin (2006), Sung (2007), Gut and Stadtm\"{u}ller (2011), and Chen and Sung (2014). In Section 3, we study Question 2. To that end, in view of the relations between convergence in probability and complete convergence, we introduce two classes of complete moment convergences, which are stronger versions of mean convergence. Let $p>0$. \begin{defi} $\{X_n,n\geq 1\}$ is said to s-$L^p$ converge to $X$ (denote $X_n\stackrel{s\mbox{-}L^p}{\longrightarrow} X$ for short), if \begin{eqnarray*} \sum_{n=1}^{\infty}E[|X_n-X|^p]<\infty. \end{eqnarray*} \end{defi} \begin{defi} $\{X_n,n\geq 1\}$ is said to s$^*$-$L^p$ converge to $X$ (denote $X_n\stackrel{s^*\mbox{-}L^p}{\longrightarrow} X$ for short), if $$ \sum_{n=1}^{\infty}\|X_n-X\|_p<\infty, $$ \end{defi} where $\|X_n-X\|_p=(E[|X_n-X|^p])^{1/p}$. \begin{rem} (i) Obviously, if $X_n\stackrel{s\mbox{-}L^p}{\longrightarrow} X$ or $X_n\stackrel{s^*\mbox{-}L^p}{\longrightarrow} X$ for some $p>0$, then $\|X_n-X\|_p\to 0$.\\ (ii) By Markov's inequality, we know that if $X_n\stackrel{s\mbox{-}L^p}{\longrightarrow} X$ for some $p>0$, then $X_n\stackrel{c.c.}{\longrightarrow}X$ and thus $X_n\stackrel{a.s.}{\longrightarrow}X$ by the Borel-Cantelli lemma.\\ (iii) If $p>1$ and $X_n\stackrel{s^*\mbox{-}L^p}{\longrightarrow} X$, then $X_n\stackrel{s\mbox{-}L^p}{\longrightarrow} X$; if $0<p<1$ and $X_n\stackrel{s\mbox{-}L^p}{\longrightarrow} X$, then $X_n\stackrel{s^*\mbox{-}L^p}{\longrightarrow} X$. \end{rem} Chow (1988) first investigated the complete moment convergence, and obtained the following result. Let $\{X,X_n,n\geq 1\}$ be a sequence of i.i.d. random variables with $E[X]=0$. Let $1\leq p<2$ and $\gamma\geq p$. If $E[|X|^{\gamma}+|X|\log(1+|X|)]<\infty$, then \begin{eqnarray}\label{Chow-a} \sum_{n\geq 1}n^{\frac{\gamma}{p}-2-\frac{1}{p}}E\left[\left(|S_n|-\varepsilon n^{\frac{1}{p}}\right)^{+}\right]<\infty\ \mbox{for all}\ \varepsilon>0, \end{eqnarray} where $x^+=\max\{0,x\}$. Chow's result has been generalized in various directions. Wang and Su (2004), Wang et al. (2005), Chen (2006), Guo and Xu (2006), Rosalsky et al. (2006), Ye and Zhu (2007), and Qiu et al. (2014) studied complete moment convergence for sums of Banach space valued random elements. Li and Zhang (20004), Chen et al. (2007), Kim et al. (2008), and Zhou (2010) considered complete moment convergence for moving average processes. Jiang and Zhang (2006), Li (2006), Liu and Lin (2006), Ye et al. (2007), Fu and Zhang (2008), Zhao and Tao (2008), and Chen and Zhang (2010) studied precise asymptotics for complete moment convergence. Wang and Zhao (2006), Liang et al. (2010), and Guo (2013) considered complete moment convergence for negatively associated random variables. Qiu and Chen (2014) studied complete moment convergence for i.i.d. random variables, and extended two results in Gut and Stadtm\"{u}ller (2011) to complete moment convergence. \begin{exa} Let $\{X_n,n\geq 1\}$ be a sequence of random variables with $\sup_{i\geq 1}E[|X_i|]\leq C$ for some positive constant $C$. Then for any $\alpha>1$, we have $\frac{S_n}{n^2(\ln n)^{\alpha}}\stackrel{s\mbox{-}L^1}{\longrightarrow} 0$. In fact, \begin{eqnarray*} \sum_{n=1}^{\infty}E\left[\left|\frac{S_n}{n^2(\ln n)^{\alpha}}\right|\right]\leq \sum_{n=1}^{\infty}\frac{1}{n^2(\ln n)^{\alpha}}\sum_{k=1}^nE[|X_i|]\leq C\sum_{n=1}^{\infty}\frac{1}{n(\ln n)^{\alpha}}<\infty. \end{eqnarray*} \end{exa} \begin{exa} Let $\{X_n,n\geq 1\}$ be a sequence of pairwise uncorrelated random variables with $\sup_{i\geq 1}Var(X_i)\leq C$ for some positive constant $C$, where $Var(X_i)$ stands for the variance of $X_i$. Then for any $\alpha>1$, we have $\frac{S_n-E[S_n]}{n^{3/2}(\ln n)^{\alpha}}\stackrel{s^*\mbox{-}L^2}{\longrightarrow} 0$. In fact, \begin{eqnarray*} \sum_{n=1}^{\infty}\left\|\frac{S_n-E[S_n]}{n^{3/2}(\ln n)^{\alpha}}\right\|_2&=&\sum_{n=1}^{\infty}\frac{1}{n^{3/2}(\ln n)^{\alpha}}(Var(S_n))^{1/2}\\ &=&\sum_{n=1}^{\infty}\frac{1}{n^{3/2}(\ln n)^{\alpha}}\left(\sum_{k=1}^nVar(X_i)\right)^{1/2}\\ &\leq& \sqrt{C}\sum_{n=1}^{\infty}\frac{1}{n(\ln n)^{\alpha}}<\infty. \end{eqnarray*} \end{exa} In Section 3, we consider ``s-$L^p$ convergence " versions and ``s$^*$-$L^p$ convergence " versions of the Toeplitz lemma, the Ces\`{a}ro mean convergence theorem and the Kronecker lemma. Four counterexamples will be given to show that they can fail in general. Some sufficient conditions for the Ces\`{a}ro mean convergence theorem under these two complete moment convergences will be presented. \section{Complete convergence} \subsection{Counterexamples} In this subsection, we will construct two counterexamples to show that ``complete convergence " versions of the Toeplitz lemma, the Ces\`{a}ro mean convergence theorem and the Kronecker lemma can fail in general. The next example shows that complete convergence version of the Ces\`{a}ro mean convergence theorem fails. \begin{exa}\label{exa2.1} Suppose that $\{X_n,n\geq 1\}$ is a sequence of independent random variables such that $P(X_n=n)=\frac{1}{n^2},P(X_n=0)=1-\frac{1}{n^2}$. For any $\varepsilon>0$, we have $$ \sum_{n=1}^{\infty}P(|X_n-0|\geq \varepsilon)= \sum_{n=1}^{\infty}P(X_n=n)=\sum_{n=1}^{\infty}\frac{1}{n^2}<\infty, $$ i.e. $X_n\to 0$ c.c. Let $\bar{X}_n=\frac{1}{n}\sum_{k=1}^nX_k,n\geq 1$. In the following, we will show that $\bar{X}_n\nrightarrow 0$ c.c. Let $n=2k,k\geq 2$ and define $k$ sets $A_1,\cdots,A_k$ as follows: \begin{eqnarray*} &&A_1:=\{X_{2k}=2k\},\\ &&A_2:=\{X_{2k}=0,X_{2k-1}=2k-1\},\\ &&\ \ \ \ \ \cdots\\ &&A_k:=\{X_{2k}=0,\cdots,X_{k+2}=0,X_{k+1}=k+1\}. \end{eqnarray*} Then we have $\bigcup_{i=1}^kA_i\subset\{\bar{X}_n\geq \frac{1}{2}\}$, and thus \begin{eqnarray*} &&P\left(\bar{X}_n\geq \frac{1}{2}\right)\geq \sum_{i=1}^kP(A_i)\\ &&=\frac{1}{(2k)^2}+\left(1-\frac{1}{(2k)^2}\right)\frac{1}{(2k-1)^2}+\cdots+ \prod_{j=2}^k\left(1-\frac{1}{(k+j)^2}\right)\frac{1}{(k+1)^2}\\ &&\geq \prod_{j=2}^k\left(1-\frac{1}{(k+j)^2}\right)\sum_{i=k+1}^{2k}\frac{1}{i^2}. \end{eqnarray*} Denote $I_k= \prod_{j=2}^k\left(1-\frac{1}{(k+j)^2}\right)$. Then \begin{eqnarray*}\label{exa2.1-a} I_k&=&\frac{(2k+1)(2k-1)}{(2k)^2}\frac{2k(2k-2)}{(2k-1)^2}\cdots\frac{(k+4)(k+2)}{(k+3)^2}\frac{(k+3)(k+1)}{(k+2)^2}\\ &=&\frac{(2k+1)(k+1)}{2k(k+2)}\to 1\ \mbox{as}\ k\to\infty. \end{eqnarray*} Thus there exists a large number $K$ such that for any $k\geq K$, we have $I_k\geq \frac{1}{2}$. So, for any $n=2k\geq 2K$, we have \begin{eqnarray*} P\left(\bar{X}_n\geq \frac{1}{2}\right)\geq I_k\sum_{i=k+1}^{2k}\frac{1}{i^2}\geq\frac{1}{2}\sum_{i=k+1}^{2k}\frac{1}{(2k)^2}=\frac{1}{8k}. \end{eqnarray*} It follows that $$ \sum_{n=1}^{\infty}P\left(\bar{X}_n\geq \frac{1}{2}\right)\geq \sum_{k=K}^{\infty}\frac{1}{8k}=\infty. $$ Hence $\bar{X}_n\nrightarrow 0$ c.c. \end{exa} \begin{rem} The above example also shows the failure of the Toeplitz lemma when the mode of convergence is ``complete convergence", taking $a_{nk}=1/n,1\leq k\leq k_n=n,n\geq 1$. \end{rem} The next example shows that complete convergence version of the Kronecker lemma also fails. The basic idea comes from Linero and Rosalsky \cite[Example 2.3]{LR13}. \begin{exa} Let $\{Y_n,n\geq 1\}$ be a sequence of independent random variables such that $P(Y_n=16^{n-1})=\frac{1}{n^2},P(Y_n=0)=1-\frac{1}{n^2}$. Denote $X_{2n-1}=Y_n,X_{2n}=-2Y_n,n\geq 1$. Then for any $n\geq 1$, we have \begin{eqnarray}\label{Exa2.2-a} \frac{X_{2n-1}}{2^{2n-1}}+\frac{X_{2n}}{2^{2n}}=0. \end{eqnarray} By the above definitions, for any $\varepsilon>0$, we have $$ \sum_{n=1}^{\infty}P(|X_n-0|\geq \varepsilon)\leq 2\sum_{n=1}^{\infty}\frac{1}{n^2}<\infty, $$ and thus $X_n\to 0$ c.c. By (\ref{Exa2.2-a}), we know that $\sum_{k=1}^n\frac{X_k}{2^k}=\frac{X_n}{2^n}I(n\ \mbox{is odd})$. Hence $\sum_{k=1}^n\frac{X_k}{2^k}\to 0$ c.c. In the following, we will show that $\frac{1}{2^n}\sum_{k=1}^nX_k\nrightarrow 0$ c.c. It's enough to show one of its subsequence \begin{eqnarray}\label{Exa2.2-b} \frac{1}{2^{4n}}\sum_{k=1}^{4n}X_k\nrightarrow 0\ c.c. \end{eqnarray} For any odd integer $k$, $$ X_k+X_{k+1}=X_k-2X_k=-X_k. $$ Thus, for any $n\geq 1$, $$ \frac{1}{2^{2n}}\sum_{k=1}^{2n}X_k=-\frac{1}{2^{2n}}\sum_{k=1}^nX_{2k-1}. $$ And so (\ref{Exa2.2-b}) can be expressed to be \begin{eqnarray}\label{Exa2.2-c} \frac{1}{16^n}\sum_{k=1}^{2n}X_{2k-1}\nrightarrow 0\ c.c. \end{eqnarray} For $k=n+1,\cdots,2n$, we have \begin{eqnarray}\label{Exa2.2-d} P(X_{2k-1}=16^{k-1})=P(Y_k=16^{k-1})=\frac{1}{k^2},\ P(X_{2k-1}=0)=P(Y_k=0)=1-\frac{1}{k^2}. \end{eqnarray} Define $n$ sets $A_1,\cdots,A_n$ as follows: \begin{eqnarray*} &&A_1:=\{X_{2(2n)-1}=16^{2n-1}\},\nonumber\\ &&A_2:=\{X_{2(2n)-1}=0,X_{2(2n-1)-1}=16^{(2n-1)-1}\},\nonumber\\ &&\ \ \ \ \cdots\nonumber\\ &&A_n:=\{X_{2(2n)-1}=0,\cdots,X_{2(n+2)-1}=0,X_{2(n+1)-1}=16^n\}. \end{eqnarray*} Then $\bigcup_{k=1}^nA_k\subset\{|\frac{1}{16^n}\sum_{k=1}^{2n}X_{2k-1}-0|\geq 1\}$, and thus \begin{eqnarray}\label{Exa2.2-d} &&P\left\{\left|\frac{1}{16^n}\sum_{k=1}^{2n}X_{2k-1}-0\right|\geq 1\right\}\geq \sum_{k=1}^nP(A_k)\nonumber\\ &&=\frac{1}{(2n)^2}+\left(1-\frac{1}{(2n)^2}\right)\frac{1}{(2n-1)^2}+\cdots+ \prod_{j=2}^n\left(1-\frac{1}{(n+j)^2}\right)\cdot\frac{1}{(n+1)^2}\nonumber\\ &&\geq \prod_{j=2}^n\left(1-\frac{1}{(n+j)^2}\right)\cdot\sum_{k=n+1}^{2n}\frac{1}{k^2}. \end{eqnarray} By (\ref{Exa2.2-d}) and following the deduction in Example 2.1, we can obtain that $$ \sum_{n=1}^{\infty}P\left\{\left|\frac{1}{16^n}\sum_{k=1}^{2n}X_{2k-1}-0\right|\geq 1\right\}=\infty, $$ i.e. (\ref{Exa2.2-c}) holds. \end{exa} \subsection{Sufficient conditions} \begin{pro}\label{pro2.1} Let $\{X_1,X_2,\cdots\}$ be pairwise uncorrelated random variables satisfying \begin{eqnarray}\label{pro2.1-a} \sum_{n=1}^{\infty}\frac{Var(X_n)}{n^{\alpha}}<\infty, \end{eqnarray} where $\alpha>0$, then for any $\varepsilon>0$, \begin{eqnarray}\label{pro2.1-b} \sum_{n=1}^{\infty}n^{1-\alpha}P\left\{\left|\frac{S_n-E(S_n)}{n}\right|\geq \varepsilon\right\}<\infty. \end{eqnarray} If $\{X_1,X_2,\cdots\}$ is a sequence of independent random variables satisfying (\ref{pro2.1-a}), then for any $\varepsilon>0$, \begin{eqnarray}\label{pro2.1-c} \sum_{n=1}^{\infty}n^{1-\alpha}P\left\{\max_{1\leq k\leq n}|S_k-E(S_k)|\geq n\varepsilon\right\}<\infty. \end{eqnarray} \end{pro} {\bf Proof.} For any $\varepsilon>0$, by Chebyshev's inequality and (\ref{pro2.1-a}), we get \begin{eqnarray}\label{pro2.1-d} &&\sum_{n=1}^{\infty}n^{1-\alpha}P\left\{\left|\frac{S_n-E(S_n)}{n}\right|\geq \varepsilon\right\}\leq\sum_{n=1}^{\infty}n^{1-\alpha}\frac{Var(S_n)}{(n\varepsilon)^2}\nonumber\\ &&=\frac{1}{\varepsilon^2}\sum_{n=1}^{\infty}\frac{1}{n^{1+\alpha}}\sum_{k=1}^{n}Var(X_k)\nonumber\\ &&=\frac{1}{\varepsilon^2}\sum_{k=1}^{\infty}Var(X_k)\sum_{n=k}^{\infty}\frac{1}{n^{1+\alpha}}\nonumber\\ &&\leq\frac{M}{\varepsilon^2}\sum_{k=1}^{\infty}\frac{Var(X_k)}{k^{\alpha}}<\infty, \end{eqnarray} where $M$ is a positive constant. Hence (\ref{pro2.1-b}) holds. By Kolmogorov's inequality and the deduction of (\ref{pro2.1-d}), we get (\ref{pro2.1-c}).\hfill\fbox \begin{rem}\label{rem2.2} (i) Letting $\alpha=1$ in the above proposition, we get that if $ \sum_{n=1}^{\infty}\frac{Var(X_n)}{n}<\infty$, then $\frac{S_n-E(S_n)}{n}\to 0$ c.c. (ii) By Kolmogorov's strong law of large numbers, we know that if $\{X_1,X_2,\cdots\}$ are independent random variables satisfying $ \sum_{n=1}^{\infty}\frac{Var(X_n)}{n^2}<\infty, $ then $\frac{S_n-E(S_n)}{n}\to 0$ a.s. By Proposition \ref{pro2.1} and Baum and Katz \cite[Proposition 1(b)]{BK65}, we know that if $\{X_1,X_2,\cdots\}$ is a sequence of pairwise uncorrelated random variables satisfying $ \sum_{n=1}^{\infty}\frac{Var(X_n)}{n^2}<\infty, $ and $|X_i|<i,\forall i\in \mathbb{N}$, then $\frac{S_n-E(S_n)}{n}\to 0$ a.s. \end{rem} \begin{cor}\label{cor2.3} Let $\{X_1,X_2,\cdots\}$ be pairwise uncorrelated random variables satisfying \linebreak $ \sum_{n=1}^{\infty}\frac{Var(X_n)}{n}<\infty, $ and $E(X_n)\to 0$. Then $\frac{S_n}{n}\to 0$ c.c. \end{cor} {\bf Proof.} In this case, $\frac{E(S_n)}{n}=\frac{\sum_{k=1}^{n}E(X_k)}{n}\to 0$. Then the result follows from Proposition \ref{pro2.1}.\hfill\fbox \begin{rem} By the above corollary, we know that if $\{X_1,X_2,\cdots\}$ is a sequence of pairwise uncorrelated random variables satisfying that $X_n\to 0$ c.c. (or $X_n$ converges to 0 in probability), $ \sum_{n=1}^{\infty}\frac{Var(X_n)}{n}<\infty, $ and there exists an integrable random variable $X$ such that $|X_n|\leq X\ a.s.,\ \forall n\geq 1$. Then by the dominated convergence theorem, we have that $E(X_n)\to 0$ and thus by the above corollary, the Ces\`{a}ro sum $\frac{S_n}{n}$ of $\{X_n,n\geq 1\}$ satisfies $\frac{S_n}{n}\to 0$ c.c. \end{rem} \section{Complete moment convergence} \subsection{Counterexamples} In this subsection, we will construct four counterexamples to show that s-$L^1$ convergence versions and s$^*$-$L^2$ convergence versions of the Toeplitz lemma, the Ces\`{a}ro mean convergence theorem and the Kronecker lemma can fail in general. The next example shows that s-$L^1$ convergence versions of the Ces\`{a}ro mean convergence theorem and the Toeplitz lemma fail. \begin{exa}\label{exa3.1} Let $\{X_n,n\geq 1\}$ be a sequence of random variables such that $P(X_n=n)=\frac{1}{n^3},P(X_n=0)=1-\frac{1}{n^3}$. Then we have $E[|X_n|]=\frac{1}{n^2}$ and thus \begin{eqnarray*} \sum_{n=1}^{\infty}E[|X_n|]=\sum_{n=1}^{\infty}\frac{1}{n^2}<\infty, \end{eqnarray*} i.e. $X_n\stackrel{s\mbox{-}L^1}{\longrightarrow}0$. Let $\bar{X}_n=\frac{1}{n}\sum_{k=1}^nX_k,n\geq 1$. Then \begin{eqnarray*} E[|\bar{X}_n|]=\frac{1}{n}\sum_{k=1}^nE[|X_k|]=\frac{1}{n}\sum_{k=1}^n\frac{1}{k^2}. \end{eqnarray*} Since $\sum_{k=1}^{\infty}\frac{1}{k^2}=\frac{\pi^2}{6}$, there exists a large $N$ such that $\forall n\geq N,\sum_{k=1}^n\frac{1}{k^2}\geq \frac{\pi^2}{12}$. Hence $$ \sum_{n=1}^{\infty}E[|\bar{X}_n|]\geq \sum_{n=N}^{\infty}\frac{1}{n}\cdot \frac{\pi^2}{12}=\infty, $$ and so it doesn't hold that $\bar{X}_n\stackrel{s\mbox{-}L^1}{\longrightarrow} 0$. \end{exa} The next example shows that s-$L^1$ convergence version of the Kronecker lemma fails. The basic idea comes from Linero and Rosalsky \cite[Example 2.3]{LR13}. \begin{exa}\label{exa3.2} Let $\{Y_n,n\geq 1\}$ be a sequence of independent random variables such that $P(Y_n=n)=\frac{1}{n^2(\ln n)^{1+\alpha}},P(Y_n=0)=1-\frac{1}{n^2(\ln n)^{1+\alpha}}$, $\alpha>0$. Denote $X_{2n-1}=(2n-1)Y_n,X_{2n}=-2nY_n,n\geq 1$. Then for any $n\geq 1$, we have \begin{eqnarray}\label{Exa3.2-a} \frac{X_{2n-1}}{2n-1}+\frac{X_{2n}}{2n}=0. \end{eqnarray} By (\ref{Exa3.2-a}), we know that $\sum_{k=1}^n\frac{X_k}{k}=\frac{X_n}{n}I(n\ \mbox{is odd})$. If $n=2k-1$, then we have $$ E[|X_n/n|]=E[|Y_k|]=\frac{1}{k(\ln k)^{1+\alpha}}, $$ which implies that $$ \sum_{n=1}^{\infty}E\left[\left|\sum_{k=1}^n\frac{X_k}{k}\right|\right]= \sum_{k=1}^{\infty}\frac{1}{k(\ln k)^{1+\alpha}}<\infty, $$ i.e. $\sum_{k=1}^n\frac{X_k}{k}\stackrel{s\mbox{-}L^1}{\longrightarrow} 0$. In the following, we will show that $\frac{1}{n}\sum_{k=1}^nX_k\stackrel{s\mbox{-}L^1}{\nrightarrow} 0$. It's enough to show one of its subsequence \begin{eqnarray}\label{Exa3.2-b} \frac{1}{2n}\sum_{k=1}^{2n}X_k\stackrel{s\mbox{-}L^1}{\nrightarrow} 0. \end{eqnarray} For any integer $k$, we have $ X_{2k-1}+X_{2k}=-Y_k.$ Thus (\ref{Exa3.2-b}) can be expressed to be \begin{eqnarray}\label{Exa3.2-c} \frac{1}{2n}\sum_{k=1}^{n}Y_k\stackrel{s\mbox{-}L^1}{\nrightarrow} 0. \end{eqnarray} By the Fubini theorem, we have \begin{eqnarray*} \sum_{n=1}^{\infty}E\left[\left|\frac{1}{2n}\sum_{k=1}^{n}Y_k\right|\right] &= &\sum_{n=1}^{\infty}\frac{1}{2n}\sum_{k=1}^{n}E\left[Y_k\right] =\sum_{n=1}^{\infty}\frac{1}{2n}\sum_{k=1}^{n}\frac{1}{k(\ln k)^{1+\alpha}}\\ &=&\sum_{k=1}^{\infty}\frac{1}{k(\ln k)^{1+\alpha}}\sum_{n=k}^{\infty}\frac{1}{2n}=\infty. \end{eqnarray*} Hence (\ref{Exa3.2-c}) holds. \end{exa} The next example shows that s$^*$-$L^2$ convergence versions of the Ces\`{a}ro mean convergence theorem and the Toeplitz lemma fail. \begin{exa}\label{exa3.3} Let $\{X_n,n\geq 1\}$ be a sequence of random variables such that $P(X_n=\sqrt{n})=\frac{1}{n^5},P(X_n=0)=1-\frac{1}{n^5}$. Then we have $E[|X_n|^2]=\frac{1}{n^4}$ and thus \begin{eqnarray*} \sum_{n=1}^n\|X_n\|_2=\sum_{n=1}^{\infty}\left(\frac{1}{n^4}\right)^{\frac{1}{2}}=\sum_{n=1}^{\infty}\frac{1}{n^2}<\infty, \end{eqnarray*} i.e. $X_n\stackrel{s^*\mbox{-}L^2}{\longrightarrow} 0$. Let $\bar{X}_n=\frac{1}{n}\sum_{k=1}^nX_k,n\geq 1$. Then \begin{eqnarray*} E[|\bar{X}_n|^2]&=&\frac{1}{n^2}\left(\sum_{k=1}^nE[|X_k|^2]+2\sum_{1\leq i<j\leq n}E[X_iX_j]\right)\\ &\geq&\frac{1}{n^2}\sum_{k=1}^nE[|X_k|^2]=\frac{1}{n^2}\sum_{k=1}^n\frac{1}{k^4}. \end{eqnarray*} Denote $c=\sum_{k=1}^{\infty}\frac{1}{k^4}$. Then $c$ is a positive constant and there exists a large $N$ such that $\forall n\geq N,\sum_{k=1}^n\frac{1}{k^4}\geq \frac{c}{2}$. It follows that \begin{eqnarray*} \sum_{n=1}^{\infty}\|\bar{X}_n\|_2\geq \sum_{n=N}^{\infty}\left(\frac{1}{n^2}\cdot \frac{c}{2}\right)^{\frac{1}{2}}=\sqrt{\frac{c}{2}}\sum_{n=N}^{\infty}\frac{1}{n}=\infty. \end{eqnarray*} Hence it doesn't hold that $\bar{X}_n\longrightarrow^{\!\!\!\!\!\!\!\!\!\!\! \!\!\!s^*\mbox{-}L^2} 0$. \end{exa} Following Examples \ref{exa3.2} and \ref{exa3.3}, we construct the following example, which shows that s$^*$-$L^2$ convergence version of the Kronecker lemma fails. \begin{exa}\label{exa3.4} Let $\{Y_n,n\geq 1\}$ be a sequence of independent random variables such that $P(Y_n=\sqrt{n})=\frac{1}{n^5},P(Y_n=0)=1-\frac{1}{n^5}$. Denote $X_{2n-1}=(2n-1)Y_n,X_{2n}=-2nY_n,n\geq 1$. Then for any $n\geq 1$, we have \begin{eqnarray}\label{Exa3.4-a} \frac{X_{2n-1}}{2n-1}+\frac{X_{2n}}{2n}=0. \end{eqnarray} By (\ref{Exa3.4-a}), we know that $\sum_{k=1}^n\frac{X_k}{k}=\frac{X_n}{n}I(n\ \mbox{is odd})$. If $n=2k-1$, then we have $$ \|X_n/n\|_2=\|Y_k\|_2=\frac{1}{k^2}. $$ Hence $$ \sum_{n=1}^{\infty}\left\|\sum_{k=1}^n\frac{X_k}{k}\right\|_2=\sum_{k=1}^{\infty}\|Y_k\|_2=\sum_{k=1}^{\infty}\frac{1}{k^2}<\infty, $$ i.e. $\sum_{k=1}^n\frac{X_k}{k}\stackrel{s^*\mbox{-}L^2}{\longrightarrow} 0$. In the following, we will show that $\frac{1}{n}\sum_{k=1}^nX_k\stackrel{s^*\mbox{-}L^2}{\nrightarrow} 0$. It's enough to show one of its subsequence \begin{eqnarray}\label{Exa3.4-b} \frac{1}{2n}\sum_{k=1}^{2n}X_k\stackrel{s^*\mbox{-}L^2}{\nrightarrow} 0. \end{eqnarray} For any integer $k$, we have $ X_{2k-1}+X_{2k}=-Y_k.$ Thus (\ref{Exa3.4-b}) can be expressed to be \begin{eqnarray}\label{Exa3.4-c} \frac{1}{2n}\sum_{k=1}^{n}Y_k\stackrel{s^*\mbox{-}L^2}{\nrightarrow} 0. \end{eqnarray} Denote $c=\sum_{k=1}^{\infty}\frac{1}{k^4}$. Then $0<c<\infty$, and there exists $N$ such that for any $n\geq N$, we have $\sum_{k=1}^{n}\frac{1}{k^4}\geq \frac{c}{2}$. Hence we have \begin{eqnarray*} \sum_{n=1}^{\infty}\left\|\frac{1}{2n}\sum_{k=1}^{n}Y_k\right\|_2&=&\sum_{n=1}^{\infty}\frac{1}{2n}\left(\sum_{k=1}^{n}E[Y_k^2]+2\sum_{1\leq i<j\leq n}E[Y_iY_j]\right)^{1/2}\\ &\geq&\sum_{n=1}^{\infty}\frac{1}{2n}\left(\sum_{k=1}^{n}E[Y_k^2]\right)^{1/2}=\sum_{n=1}^{\infty}\frac{1}{2n}\left(\sum_{k=1}^{n}\frac{1}{k^4}\right)^{1/2}\\ &\geq&\sum_{n=N}^{\infty}\frac{1}{2n}\left(\sum_{k=1}^{n}\frac{1}{k^4}\right)^{1/2}\\ &\geq&\sum_{n=N}^{\infty}\frac{1}{2n}\sqrt{\frac{c}{2}}=\infty. \end{eqnarray*} Hence (\ref{Exa3.4-c}) holds. \end{exa} \subsection{Sufficient conditions} By Example 3.1, we know that, if $\sum_{n=1}^{\infty}E[|X_n|^p]<\infty$, then we don't have $\sum_{n=1}^{\infty}E[|S_n/n|^p]<\infty$ necessarily. In general, we have the following result. \begin{pro}\label{pro3.1} Suppose that $1\leq p< \infty$ and $\sum_{n=1}^{\infty}E[|X_n|^p]<\infty$, then $\forall \varepsilon>0$, we have \begin{eqnarray*} \sum_{n=1}^{\infty}\frac{1}{(\ln n)^{1+\varepsilon}}E\left[\left|S_n/n\right|^p\right]<\infty. \end{eqnarray*} \end{pro} {\bf Proof.} By the convexity of the function $f(x)=|x|^p$, we have \begin{eqnarray*} \sum_{n=1}^{\infty}\frac{1}{(\ln n)^{1+\varepsilon}}E\left[\left|S_n/n\right|^p\right]&\leq& \sum_{n=1}^{\infty}\frac{1}{n(\ln n)^{1+\varepsilon}}\left(\sum_{k=1}^nE[|X_k|^p]\right)\\ &\leq &\left(\sum_{k=1}^{\infty}E[|X_k|^p]\right)\sum_{n=1}^{\infty}\frac{1}{n(\ln n)^{1+\varepsilon}}<\infty. \end{eqnarray*} \hfill\fbox \begin{pro}\label{pro3.2} Let $\{X_1,X_2,\cdots\}$ be pairwise uncorrelated random variables satisfying \linebreak $\sum_{n=1}^{\infty}Var(X_n)<\infty,$ then for any $1<q\leq 2$, we have \begin{eqnarray*} \sum_{n=1}^{\infty}E\left[\left|\frac{S_n-E(S_n)}{n}\right|^q\right]<\infty, \end{eqnarray*} in particular, $\frac{S_n-E(S_n)}{n}\to 0$ c.c. \end{pro} {\bf Proof.} By the assumptions, we have \begin{eqnarray*} \sum_{n=1}^{\infty}E\left[\left|\frac{S_n-E(S_n)}{n}\right|^q\right]&=&\sum_{n=1}^{\infty}\left(\left\|\frac{S_n-E(S_n)}{n}\right\|_q\right)^q\\ &\leq&\sum_{n=1}^{\infty}\left(\left\|\frac{S_n-E(S_n)}{n}\right\|_2\right)^q\\ &=&\sum_{n=1}^{\infty}\frac{1}{n^q}\left(\sum_{i=1}^{n}Var(X_i)\right)^{q/2}\\ &\leq&\sum_{n=1}^{\infty}\frac{1}{n^q}\left(\sum_{i=1}^{\infty}Var(X_i)\right)^{q/2}<\infty. \end{eqnarray*} \hfill\fbox By Example \ref{exa3.3}, we know that, if $\sum_{n=1}^{\infty}\|X_n\|_p<\infty$, then we don't have $\sum_{n=}^{\infty}\|\frac{S_n}{n}\|_p<\infty$ necessarily. In general, we have the following two propositions. \begin{pro}\label{pro3.3} Suppose that $1\leq p< \infty$ and $\sum_{n=1}^{\infty}\|X_n\|_p<\infty$, then $\forall \varepsilon>0$, we have \begin{eqnarray}\label{pro3.3-a} \sum_{n=1}^{\infty}\frac{1}{(\ln n)^{1+\varepsilon}}\left\|S_n/n\right\|_p<\infty. \end{eqnarray} \end{pro} {\bf Proof.} By Minkowski's inequality and the definition of the norm $\|\cdot\|_p$, we have that $$ \left\|S_n/n\right\|_p\leq \frac{1}{n}(\sum_{k=1}^{n}\|X_k\|_p). $$ Then we can prove (\ref{pro3.3-a}) by following the proof of Proposition \ref{pro3.1}.\hfill\fbox \begin{pro}\label{pro3.4} Suppose that $1<p<\infty$ and $ \sum_{n=1}^{\infty}\|X_n\|_p<\infty, $ then for any $1<q\leq p$, we have \begin{eqnarray*} \sum_{n=1}^{\infty}E[|S_n/n|^q]<\infty, \end{eqnarray*} in particular, $S_n/n\to 0$ c.c. \end{pro} {\bf Proof.} By the fact that $\|\cdot\|_q\leq \|\cdot\|_p$, Minkowski's inequality and the assumption, we have \begin{eqnarray*} \sum_{n=1}^{\infty}E[|S_n/n|^q]&=&\sum_{n=1}^{\infty}\left(\|S_n/n\|_q\right)^q\leq\sum_{n=1}^{\infty}\left(\|S_n/n\|_p\right)^q\\ &\leq&\sum_{n=1}^{\infty}\left(\frac{\sum_{k=1}^n\|X_k\|_p}{n}\right)^q\\ &=&\sum_{n=1}^{\infty}\frac{1}{n^q}\left(\sum_{k=1}^n\|X_k\|_p\right)^q\\ &\leq&\left(\sum_{k=1}^{\infty}\|X_k\|_p\right)^q\sum_{n=1}^{\infty}\frac{1}{n^q}<\infty. \end{eqnarray*} \hfill\fbox \begin{pro}\label{pro3.5} Suppose that $\sum_{n=1}^{\infty}\|X_n\|_{\infty}<\infty$. Then\\ (i) for any $\varepsilon>0$, we have \begin{eqnarray}\label{pro3.4-a} \sum_{n=1}^{\infty}\frac{1}{(\ln n)^{1+\varepsilon}}\|S_n/n\|_{\infty}<\infty; \end{eqnarray} (ii) for any $1<q<\infty$, we have \begin{eqnarray*}\label{pro3.4-b} \sum_{n=1}^{\infty}E[|S_n/n|^q]<\infty, \end{eqnarray*} in particular, $S_n/n\to 0$ c.c. \end{pro} {\bf Proof.} (i) By the definition of the norm $\|\cdot\|_{\infty}$, we have that $$ \|S_n/n\|_{\infty}\leq \frac{1}{n}\left(\sum_{k=1}^n\|X_k\|_{\infty}\right). $$ Then we can prove (\ref{pro3.4-a}) by following the proof of Proposition \ref{pro3.1}. (ii) It's a direct consequence of Proposition \ref{pro3.4} by noting that for any $1<p<\infty$ and any random variable $X$, $\|X\|_p\leq \|X\|_{\infty}$.\hfill\fbox \bigskip { \noindent {\bf\large Acknowledgments} \vskip 0.1cm \noindent We acknowledge the helpful suggestions and comments of three anonymous referees, which improved the presentation of this paper. We are grateful to the support of NNSFC (Grant No. 11371191) and Jiangsu Province Basic Research Program (Natural Science Foundation) (Grant No. BK2012720).}

https://arxiv.org/abs/1004.2446

Spanning and independence properties of frame partitions

We answer a number of open problems in frame theory concerning the decomposition of frames into linearly independent and/or spanning sets. We prove that in finite dimensional Hilbert spaces, Parseval frames with norms bounded away from 1 can be decomposed into a number of sets whose complements are spanning, where the number of these sets only depends on the norm bound. We also prove, assuming the Kadison-Singer conjecture is true, that this holds for infinite dimensional Hilbert spaces. Further, we prove a stronger result for Parseval frames whose norms are uniformly small, which shows that in addition to the spanning property, the sets can be chosen to be independent, and the complement of each set to contain a number of disjoint, spanning sets.

\section{} \section{Introduction} A family of vectors $\{f_i\}_{i\in I}$ is a \textit{frame} for a Hilbert space $\mathbb{H}$ if there are constants $0<A\le B<\infty$ satisfying \[ A\|x\|^2 \le \sum_{i\in I}|\langle x,f_i\rangle|^2 \le B \|x\|^2,\ \ \mbox{for all $x\in \mathbb{H}$}.\] The theory of frames in Hilbert spaces has applications covering a broad spectrum of problems in pure mathematics, applied mathematics and engineering \cite{CT}. Many fundamental questions in frame theory involve determining the extent to which frames can be decomposed into subsets which to some extent resemble bases. It is known that these problems are generally difficult to resolve. For example, work of the second author shows that decomposing frames into subsets which are Riesz basic sequences is equivalent to an important open problem in analysis -- the 1959 Kadison-Singer Problem \cite{CT}. In this note we will answer a number of open problems concerning the decomposition of frames into linearly independent and/or spanning sets. The solutions to these problems, for finite dimensional Hilbert spaces, requires some non-trivial variations of the Rado-Horn Theorem, which is itself rather delicate. We prove that some infinite dimensional analogues of these results would have a positive answer if the Kadison-Singer Problem has a positive answer. In particular, the fact that the number $R$ appearing in Theorem~3.2 can be chosen independent of the dimension of the underlying Hilbert space is implied by the assumption that the Kadison-Singer Problem has a positive answer. Indeed, one of the motivations that led to the study of these consequences of Kadison-Singer was, initially, a quest for a negative answer to the Kadison-Singer Problem. However, our results verify that these consequences of a positive answer to the Kadison-Singer Problem are, in fact, true. In Section~2, we derive a result about Parseval frames with norms bounded away from $1$ that is a consequence of the assumption that the Kadison-Singer Problem has a positive answer. We will see that these results involve questions about spanning sets. In Section~3, we then show that the finite dimensional versions of these results are uniformly true, that is, with constants that do not depend on the dimension, without the need to assume that the Kadison-Singer Problem has a positive answer. The results of Section~3 are in some sense refinements of the Rado-Horn theory and rely strongly on earlier refinements of Rado-Horn obtained by the second and fourth authors together with Kutyniok. Section~4 is a further development which investigates the consequences of having Parseval frames with uniformly small norms. \section{Kadison-Singer and Spanning Properties for Frame Partitions} In this section, we begin with a few observations about Parseval frames and prove a result about spanning sets for Parseval frames whose norms are bounded away from $1$, assuming that the Kadison-Singer Problem has a positive answer. Given a family of vectors $\mathcal F = \{ f_i \}_{i \in S}$ in a Hilbert space $\mathbb{H},$ where $S$ is some index set, and a subset $B \subseteq S,$ we write $\mathcal F_B = \{ f_i\}_{i \in B }$ and let $\mathbb{H}_B$ denote the closed linear span of $\mathcal F_B.$ Recall that if $\{ f_i \}_{i \in S}$ is a Parseval frame for $\mathbb{H}$ and $P$ is the orthogonal projection onto some closed subspace, then $\{ P f_i \}_{i \in S}$ is a Parseval frame for $P(\mathbb{H}).$ Also, recall that $\{ f_i \}_{i \in S}$ is a Parseval frame for $\mathbb{H}$ if and only if the Gram matrix $G= ( \langle f_j, f_i \rangle)_{i,j \in S}$ is the matrix of a projection operator on $\ell^2(S).$ We begin with a few useful observations. \begin{proposition}\label{2.1} Let $\{ f_i \}_{i \in S}$ be a Parseval frame for $\mathbb{H},$ let $P$ be an orthogonal projection onto a closed subspace of $\mathbb{H}$ and let $I$ denote the identity operator on $\mathbb{H}.$ Then $G= ( \langle f_j, f_i \rangle)_{i,j,\in S},$ $R = ( \langle Pf_j, Pf_i \rangle)_{i,j\in S}$ and $Q= ( \langle (I-P)f_j, (I-P)f_i \rangle )_{i,j\in S}$ are the matrices of projection operators on $\ell^2(S)$ with $G= R + Q.$ Moreover, $P=I$ if and only if $1$ is not an eigenvalue of $Q.$ \end{proposition} \begin{proof} The equality $G = R+Q$ is immediate from $R=( \langle Pf_j, f_i \rangle)$ and $Q= ( \langle (I-P)f_j, f_i \rangle )$ for each $i,j\in S$, and from the linearity of the inner product in the first entry. The fact that $G, Q$ and $R$ are matrices of projections follows from the fact that the vectors $\{f_i\}_{i \in S}$ form a Parseval frame. For the final statement, note that $P=I$ if and only if $Q=0.$ But since $Q$ is a projection, $Q=0$ if and only if $1$ is not an eigenvalue. \end{proof} Given a subset $B \subseteq S,$ we let $D_B=(d_{i,j})_{i,j \in S}$ denote the bounded operator on $\ell^2(S)$ whose matrix is the diagonal matrix with $d_{i,i} =1$ when $i \in B$ and $d_{i,j}=0$ when $i \in B^c$ or $j \in B^c$, the complement of the set $B.$ \begin{proposition}\label{2.2} Let $\{ f_i \}_{i \in S}$ be a Parseval frame for $\mathbb{H},$ let $G=(\langle f_j , f_i \rangle)_{i,j \in S}$ denote its Gram matrix, and let $B \subseteq S.$ Then $\mathbb{H}_B = \mathbb{H}$ if and only if $1$ is not an eigenvalue of $D_{B^c}GD_{B^c}.$ \end{proposition} \begin{proof} Let $P$ denote the projection onto $\mathbb{H}_B$ and apply Proposition \ref{2.1}. Since for $j \in B, f_j \in \mathbb{H}_B,$ we have that when $j \in B$ then $\langle f_j, f_i \rangle = \langle P f_j, P f_i \rangle .$ More generally, the matrices $G$ and $R$ are equal in any entry $(i,j)$ provided that $i \in B$ or $j \in B.$ Thus, the matrix $Q$ must be $0$ in any such entry. Hence we obtain the operator inequalities $0 \le Q = D_{B^c}QD_{B^c} \le D_{B^c}GD_{B^c} \le D_{B^c}.$ Now, if $1$ is not an eigenvalue of $D_{B^c}GD_{B^c},$ then these inequalities imply that $1$ is not an eigenvalue of $Q.$ Invoking the preceding proposition, we get $P=I,$ and so $\mathbb{H}_B = \mathbb{H}.$ Conversely, assume that $1$ is an eigenvalue of $D_{B^c}GD_{B^c}.$ Write $G=V V^*$ where $V: \mathbb{H} \to \ell^2(S)$ is the analysis operator of the Parseval frame. Since $D_{B^c}GD_{B^c} = (V^*D_{B^c})^*(V^*D_{B^c}),$ we have that $(V^*D_{B^c})(V^*D_{B^c})^* = V^* D_{B^c} V$ also has eigenvalue $1.$ By the Parseval property, $V$ is an isometry, $V^* V = I$, necessarily $V^* D_{B} V= I - V^* D_{B^c} V$ has eigenvalue zero. Thus the range of $V^* D_B$ is orthogonal to the corresponding eigenvectors. Since the closure of the range of $V^* D_B$ is by definition $\mathbb{H}_B$, it is not equal to $\mathbb{H}$. \end{proof} The last result yields a complementarity principle between spanning and linear independence. \begin{proposition} Let $\mathbb{H}$ be a Hilbert space with orthonormal basis $\{ e_j\}_{j \in S },$ let $P$ be the orthogonal projection onto a closed subspace of $\mathbb{H},$ and let $B \subseteq S.$ Then the linear span of $\{Pe_j\}_{ j \in B }$ is dense in $P(\mathbb{H})$ if and only if the operator $(\langle (I-P)e_j, (I-P)e_i \rangle)_{i,j \in B^c}$ on $\ell^2(B^c)$ is one-to-one. \end{proposition} \begin{proof} Note that the set $\{Pe_j: j \in S \}$ is a Parseval frame for $P(\mathbb{H}).$ Hence, the span of $\{Pe_j:j \in B \}$ is dense in $P(\mathbb{H})$ if and only if the matrix $Q= (\langle Pe_j, Pe_i \rangle)_{i,j \in B_c}$ does not have 1 as an eigenvalue. But since $I_{\ell^2(B^c)} - Q = (\langle (I-P)e_j, (I-P)e_i \rangle)_{i,j \in B^c},$ $Q$ not having eigenvalue 1 is equivalent to the latter matrix having a trivial kernel. \end{proof} \begin{corollary} If $\mathbb{H}$ is finite dimensional, then $\{Pe_j\}_{ j \in B }$ spans $P(\mathbb{H})$ if and only if the set $\{ (I-P)e_j: j \in B^c \}$ is linearly independent. \end{corollary} For our next result, we will be assuming that the Anderson Paving Problem has an affirmative answer. The Anderson Paving Problem and the Kadison-Singer Problem are known to be equivalent \cite{An}. There are several equivalent versions of Anderson's Paving Problem. The particular version that we shall use asserts the following: For each $0< s <1,$ there exists an $r$ depending only on $s,$ such that if $H=(h_{i,j}) \in B(\ell^2(\mathbb{N}))$ is any operator with $h_{i,i} =0,$ for every $i,$ then there exists a partition of $\mathbb{N}$ into $r$ disjoint sets $A_1 \cup \cdots \cup A_r = \mathbb{N},$ with $\|D_{A_k}HD_{A_k}\| \le s \|H\|,$ for $k=1,...,r.$ \begin{theorem} Let $0 < \delta < 1.$ If the Anderson Paving Problem has a positive answer, then there exists an $r$ depending on $\delta,$ such that whenever $\{ f_n \}_{n \in \mathbb{N}}$ is a Parseval frame for a Hilbert space $\mathbb{H},$ with $\|f_n\|^2 \le 1 - \delta $ for all $n \in \mathbb{N}$, then there exists a partition of $\mathbb{N}$ into $r$ disjoint sets, $A_1 \cup \cdots \cup A_r = \mathbb{N},$ such that $\mathbb{H}_{A_k^c} = \mathbb{H},$ for $k=1,..., r.$ \end{theorem} \begin{proof} Let $G= ( \langle f_j, f_i \rangle)$ and let $E(G)$ denote the diagonal part of $G,$ so that $0 \le E(G) \le (1 - \delta) I.$ Since $0 \le G \le I,$ we have that $(\delta -1)I \le -E(G) \le G - E(G) \le G \le I.$ Hence, $H = G - E(G)$ has 0 diagonal and $\|H\| \le 1.$ Set $s = \delta/2$ in the statement of Anderson's Paving and let $r$ be the corresponding integer. Then we may pick disjoint sets, $A_1 \cup \cdots A_r = \mathbb{N}$ such that $\|D_{A_k}HD_{A_k}\| \le \delta/2,$ for every $k \in \{1,2, \dots,r\}.$ Hence, we have that $0 \le D_{A_k}GD_{A_k} = D_{A_k}E(G)D_{A_k} + D_{A_k}HD_{A_k} \le (1 - \delta)I + (\delta/2)I = (1 - \delta/2)I.$ Thus, $\|D_{A_k}GD_{A_k}\| < 1$ and it follows that $1$ can not be an eigenvalue. From the preceding proposition, it follows that $\mathbb{H}_{A_k^c}= \mathbb{H},$ for each $k \in \{1,2, \dots,r\}.$ \end{proof} The study of Parseval frames with norms bounded away from $1$ is in some sense complementary to other results relating Parseval frames and the Kadison-Singer Problem, e.g.~\cite{CCLV}, since most other work relating these problems focuses on Parseval frames whose norms are bounded away from $0$ rather than $1$ and focuses on linear independence rather than spanning. However, having Parseval frames that are norm-bounded away from $1$ has the added advantage that when one projects onto a subspace, then the projections of these vectors is a Parseval frame for the subspace that is bounded away from $1$ by the same bound. In contrast, a Parseval frame with norms that are bounded away from $0$ might no longer have norms bounded away from $0$ when one projects it onto a subspace. \section{Spanning properties for partitions of Parseval frames with norms bounded away from one} The previous section illustrates that Anderson's paving would provide a partition of certain norm-bounded Parseval frames $\{f_i\}_{i\in I}$ into a number of sets with specific spanning properties. Now we show that the existence of such a number, $r$, can be obtained independently of the assumption of Anderson Paving, if the Hilbert space is finite dimensional. Moreover, our choice of $r$ depends only on the norm bound $1-\delta,$ and not the dimension of the space, and an explicit formula for $r$ as a function of $\delta$ is provided. Recall that a {\emph {matroid}} is a finite set $X$ together with a collection of subsets of $X$, ${\mathcal {I}}$, which satisfy three properties: \begin{enumerate} \item $\emptyset\in {\mathcal {I}}$ \item if $I_1\in {\mathcal {I}}$ and $I_2\subset I_1$, then $I_1\in {\mathcal {I}}$, and \item if $I_1,I_2\in {\mathcal {I}}$ and $|I_1| < |I_2|$, then there exists $x\in I_2 \setminus I_1$ such that $I_1 \cup\{x\} \in {\mathcal {I}}$. \end{enumerate} We will say that elements of ${\mathcal {I}}$ are independent. We also recall that the rank of a set $E\subset X$ is defined to be the cardinality of a maximal independent set contained in $E$. Now, given a set of vectors $\{f_j:j\in J\}$ which spans $\mathbb H_N$, we say $J\in {\mathcal {J}}$ if $\{f_j:j\not\in J\}$ spans $\mathbb H_N$. It is straightforward to verify that $(J, {\mathcal {J}})$ forms a matroid. Indeed, the first two properties are immediate and the third property reduces after taking complements to the fact that if $\{f_j : j\in E_1\}$ and $\{f_j:j\in E_2\}$ both span $\mathbb H_N$, and $|E_1| > |E_2|$, then there exists $x\in E_1\setminus E_2$ such that $\{f_j:j\in E_1, j\not= x\}$ spans. We note here that for a natural number $n$, rank$(E) \ge n$ if and only if there is a set $F\subset E$ such that $|F| = n$ and $\{f_j:j\not \in F\}$ spans $\mathbb H_N$. Finally, we recall the Rado-Horn Theorem \cite{H, R} in the context of matroids. \begin{theorem}\cite{EF} Let $(X, {\mathcal {I}})$ be a matroid, and let $R$ be a positive integer. A set $J\subset X$ can be partitioned into $R$ independent sets if and only if for every subset $E\subset J$, \begin{equation}\label{eq1} \frac {|E|}{\operatorname{rank}(E)} \le R. \end{equation} \end{theorem} \begin{theorem} \label{T1} Let $\delta > 0$. Suppose that $\{f_j:j\in J\}$ is a Parseval frame for $\mathbb H_N$ with $\|f_j\|^2 \le 1 - \delta$ for all $j\in J$. Let $R \in \mathbb{N}$, $R \ge \frac1\delta$. Then, it is possible to partition $J$ into $R$ sets $\{A_1,\ldots, A_R\}$ such that for each $1\le r\le R$, the family $\{f_j:j\not\in A_r\}$ spans $\mathbb H_N$. \end{theorem} \begin{proof} Let ${\mathcal {J}} = \{E\subset J: \operatorname{span} \{f_j:j\not\in E\} = \mathbb H_N \}$. Since a Parseval frame must span, we have that $(J, {\mathcal {J}})$ is a matroid. By the Rado-Horn Theorem, it suffices to show (\ref{eq1}) for each subset of $J$. Let $E\subset J$. Define $S = \operatorname{span}\{f_j:j\not\in E\}$, and let $P$ be the orthogonal projection onto $S^\perp$. Since the orthogonal projection of a Parseval frame is again a Parseval frame, we have that $\{Pf_j:j\in J\}$ is a Parseval frame for $S^\perp$. Moreover, we have \begin{eqnarray*} \dim S^\perp &=& \sum_{j\in J} \|Pf_j\|^2 = \sum_{j\in E} \|Pf_j\|^2 \\ &\le& |E|(1 - \delta). \end{eqnarray*} Let $M$ be the largest integer smaller than or equal to $|E|(1 - \delta)$. Since $\dim S^\perp \le M$, we have that there exists a set $E_1 \subset E$ such that $|E_1| = M$ and $\operatorname{span} \{P f_j:j\in E_1\} = S^\perp$. Let $E_2 = E\setminus E_1$. We show $E_2$ is independent. Write $h \in \mathbb H_N$ as $h = h_1 + h_2$, where $h_1\in S$ and $h_2 \in S^\perp$. We have that $h_2 = \sum_{j\in E_1} \alpha_j Pf_j$ for some choice of $\{\alpha_j:j\in E_1\}$. Write $\sum_{j\in E_1} \alpha_j f_j = g_1 + h_2$, where $g_1 \in S$. Then, there exist $\{\alpha_j:j\not\in E\}$ such that $\sum_{j\not\in E} \alpha_j f_j = h_1 - g_1$. Then, we have \[ \sum_{j\not\in E_2} \alpha_j f_j = h, \] as desired. Now, since $E$ contains an independent set of cardinality $|E| - M$, it follows that $\operatorname{rank}(E) \ge |E| - M \ge |E| - |E|(1 - \delta) = \delta |E|$. Therefore, \[ \frac {|E|}{\operatorname{rank}(E)} \le \frac 1\delta \le R, \] as desired. \end{proof} We note that it is not possible, in general, to get the partition in Theorem \ref{T1} to have the property that the $\{f_j\}_{j\in A_i}$ are linearly independent. Again, the problem is that we do not have a lower bound on the norms of the frame vectors and so there can be an arbitrarily large number of them. That is, there can be too many frame vectors to be able to partition them into $R$ linearly independent sets. However, we will see that it is possible to achieve a partition in which all sets but one are linearly independent and spanning, if the norms of the vectors are uniformly small. \section{Spanning and linear independence properties for Parseval frames with uniformly small norms} In this section we obtain a strengthening of the preceding section with the help of a generalization of the Rado-Horn Theorem due to Casazza, Kutyniok and Speegle \cite{CKS}. \begin{theorem}\label{T2} Let $\{f_i\}_{i\in I}$ be a finite collection of vectors in a vector space $X$ and let $M\in {\mathbb N}$. The following conditions are equivalent: (1) There exists a partition $\{I_j\}_{j=1}^M$ of $I$ so that for each $j$, $\{f_i\}_{i\in I_j}$ is linearly independent. (2) For all $J\subset I$, \[ \frac{|J|}{dim\ span\ \{f_i\}_{i\in J}} \le M.\] Moreover, in the case that the above conditions fail, there exists a partition $\{I_j\}_{j=1}^M$ of $I$ and a subspace $S$ of $X$ such that the following three conditions hold. \vskip12pt \ \ \ \ (a) For all $1\le j\le M$, $S= span\ \{f_i:i\in I_j,\ \mbox{and}\ f_i\in S\}$. \vskip12pt \ \ \ \ (b) For $J = \{i\in I:f_i\in S\}$, \[ \frac{|J|}{dim\ span\ \{f_i\}_{i\in J}}>M.\] \ \ \ \ \ \ \ (c) For each $1\le j\le M$, \[ \sum_{i\in I_j,f_i\notin S}\alpha_if_i =0,\ \ \mbox{implies}\ \ \alpha_i=0,\ \ \mbox{for all i}.\] In particular, for each $1\le j \le M$, $\{f_i:i\in I_j,\ f_i\notin S\}$ is linearly independent. \end{theorem} We also need a slight generalization of a result of Casazza and Tremain \cite{CT}. \begin{proposition}\label{P5} Let $r,k,N$ be natural numbers with $0< k <N$ and let $\{f_i\}_{i=1}^{rN+k}$ be an equal norm Parseval frame for an $N$-dimensional Hilbert space $\mathbb H_N$. Then $\{f_i\}_{i=1}^{rN+k}$ can be partitioned into $r+1$ linearly independent sets. If $k=0$, $\{f_i\}_{i=1}^{rN}$ can be partitioned into $r$ linearly independent spanning sets. \end{proposition} \begin{proof} Since $\{f_i\}_{i=1}^{rN+k}$ is an equal norm Parseval frame, we have \[ N = \sum_{i=1}^{rN+k}\|f_i\|^2 = (rN+k)\|f_j\|^2,\ \ \mbox{for all $j=1,2,\ldots,rN+k$}.\] That is, \[ \|f_i\|^2 = \frac{N}{rN+k},\ \ \mbox{for all $i=1,2,\ldots,N$}.\] We will verify that the assumption of the Rado-Horn Theorem holds for $r+1$. Choose $J\subset \{1,2,\ldots,rN+k\}$. Let $P$ be the orthogonal projection of $\mathbb H_N$ onto span $\{f_i\}_{i\in J}$. Since $\{Pf_i\}_{i\in J}$ is a Parseval frame for its span we have \[ \dim \span\ \{f_i\}_{i\in J} = \sum_{i=1}^{rN+k}\|Pf_i\|^2 \ge \sum_{i\in J}\|Pf_i\|^2 = \sum_{i\in J}\|f_i\|^2 = \frac{N|J|}{rN+k}.\] That is, \[ \frac{|J|}{\dim \span\ \{f_i\}_{i\in J}}\le \frac{rN+k}{N}.\] That is, \[ \frac{|J|}{\dim \span\ \{f_i\}_{i\in J}}\le \begin{cases} r & \mbox{if $k=0$}\\ r+1 & \mbox{if $0<k<N$} \end{cases} \] The result now follows by the Rado-Horn Theorem and the fact that in the case $k=0$, we have partitioned an $rN$ element set into $r$ linearly independent sets in an $N$-dimensional Hilbert space $\mathbb H_N$, and hence, each must contain exactly $N$ elements and so it must be a spanning set. \end{proof} We now want to strengthen Proposition \ref{P5} to show that we can actually partition our family of vectors into a linearly independent set and $r$ linearly independent spanning sets. \begin{lemma}\label{lem1} Let $\{f_i\}_{i\in I_j}$, $j=1,2,\ldots r$ be linearly independent families of vectors in an $N$-dimensional Hilbert space $\mathbb H_N$. Assume there is a partition of $\cup_{j=1}^r I_j$ into $\{A_j\}_{j=1}^r$ so that \[ \span \ \{f_i\}_{i\in A_j} = \mathbb H_N,\ \ \mbox{for all $j=1,2,\ldots,r$}.\] Then \[ \span \ \{f_i\}_{i\in I_j} = \mathbb H_N,\ \ \mbox{for all $j=1,2,\ldots,r$}.\] \end{lemma} \begin{proof} For all $j=1,2,\ldots,r$, the fact that $\{f_i\}_{i\in I_j}$ are linearly independent implies that the dimension of the span of $\{f_i\}_{i\in I_j} = |I_j|.$ Also, the fact that $\{f_i\}_{i\in A_j}$ span $\mathbb{H}_N$ implies $|A_j|\ge N$. Now, we have \[ Nr \ge \sum_{j=1}^r \dim \span\ \{f_i:i\in I_j\} = \sum_{j=1}^r |I_j| = |\cup_{j=1}^r I_j| = |\cup_{j=1}^r A_j| = \sum_{j=1}^r |A_j| \ge Nr.\] Hence, \[ \sum_{j=1}^r \dim \span\ \{f_i:i\in I_j\} = Nr,\] and so \[ \dim \span\ \{f_i:i\in I_j\} =N,\ \ \mbox{for every $j=1,2,\ldots,r$}.\] \end{proof} Now we can partition frames into spanning sets. \begin{proposition}\label{P6} Let $\{f_i\}_{i\in I}$ be a frame for $\mathbb H_N$ with lower frame bound $A$ and $\|f_i\|^2 \le 1$ for all $i \in I$. Let $r = \lfloor A \rfloor$. Then there exists a partition $\{I_j\}_{j=1}^r$ of $I$ so that \[ \span\ \{f_i:i\in I_j\} = \mathbb H_N,\ \ \mbox{for all $j=1,2,\ldots,r$}.\] In particular, the number of frame vectors in a unit norm frame with lower frame bound $A$ is greater than or equal to$\lfloor A\rfloor N$. \end{proposition} \begin{proof} We replace $\{f_i\}_{i\in I}$ by $\{\frac{1}{\sqrt{r}}f_i\}_{i\in I}$ so that our frame has lower frame bound greater than or equal to $1$ and \[ \|f_i\|^2 \le \frac{1}{r},\ \ \mbox{for all $i\in I$}.\] Assume the frame operator for $\{f_i\}_{i\in I}$ has eigenvectors $\{e_j\}_{j=1}^N$ with respective eigenvalues $\lambda_{1}\ge \lambda_{2}\ge \ldots \lambda_N \ge 1$. We proceed by induction on $N$. \vskip12pt \noindent $N=1$: Since \begin{equation}\label{E1} \sum_{i\in I}\|f_i\|^2 \ge1,\ \ \mbox{and}\ \ \|f_i\|^2 \le \frac{1}{r}, \end{equation} it follows that $|\{i\in I:f_i \not= 0\}|\ge r$ and so we have a partition into $r$ spanning sets. \vskip12pt \noindent Assume the inductive hypothesis holds for $\mathbb H_N$ and consider $\mathbb H_{N+1}$. \vskip12pt We check two cases: \vskip12pt \noindent {\bf Case I}: Suppose there exists a partition $\{I_j\}_{j=1}^r$ of $I$ so that $\{f_i\}_{i\in I_j}$ is linearly independent for all $j=1,2,\ldots,r$. \vskip12pt In this case, \[ N+1 \le (N+1)\lambda_N \le \sum_{j=1}^{N+1}\lambda_j= \sum_{i\in I}\|f_i\|^2 \le |I| \frac{1}{r},\] and hence, \[ |I|\ge r(N+1).\] However, by linear independence, we have \[ |I|= \sum_{j=1}^r|I_j| \le r(N+1) .\] Thus, $|I_j|=N+1$ for every $j=1,2,\ldots,r$ and so $\{f_i\}_{i\in I_j}$ are all spanning. \vskip12pt \noindent {\bf Case II}: Our family cannot be partitioned into $r$ linearly independent (spanning) sets. \vskip12pt In this case, let $\{I_j\}_{j=1}^r$ and a subspace $\emptyset \not= S \subset \mathbb H_{N+1}$ be given by Theorem \ref{T2}. If $S = \mathbb H_{N+1}$, we are done. Otherwise, let $P$ be the orthogonal projection onto the subspace $S$. Let \[ I_j' = \{i\in I_j:f_i\notin S\},\ \ I' = \cup_{j=1}^r I_j'.\] Theorem \ref{T2} (c) implies that $\{f_i\}_{i\in I_j'}$ is linearly independent for all $j=1,2,\ldots,r$. To see this, note that the non-zero elements of $\{(I-P)f_i\}_{i\in I}$ are $\{(I-P)f_i\}_{i\in I'}$. Fix $1\le j\le r$ and assume there are scalars $\{\alpha_i\}_{i\in I_j'}$ with \[ \sum_{i\in I_j'}\alpha_i(I-P) f_i = 0.\] This implies {by} Theorem \ref{T2} (c): \[ \sum_{i\in I_j'}\alpha_if_i \in S,\ \ \mbox{and so}\ \ \alpha_i =0,\ \mbox{for all $i\in I_j'$}.\] Now, $\{(I-P)f_i\}_{i\in I'}$ has lower frame bound 1 in $(I-P)(\mathbb H_{N+1})$, dim $(I-P)(\mathbb H_{N+1})\le N$ and \[ \|(I-P)f_i\|^2 \le \|f_i\|^2 \le \frac{1}{r},\ \ \mbox{for all $i\in I'$}.\] Applying the induction hypothesis, we can find a partition $\{A_j\}_{j=1}^r$ of $I'$ with \[ \span \ \{(I-P)f_i\}_{i\in A_j} = (I-P)(\mathbb H_{N+1}),\ \ \mbox{for all $j=1,2,\ldots,r$}.\] Now, we can apply Lemma \ref{lem1} together with the partition $\{A_j\}_{j=1}^r$ to conclude: \[ \span\ \{(I-P)f_i\}_{i\in I_j'}= (I-P)(\mathbb H_{N+1}),\] and hence \[ \span\ \{f_i\}_{i\in I_j} = \span\ \{S,\{(I-P)f_i\}_{i\in I_j'}\} = \mathbb H_{N+1}.\] \end{proof} Note that we cannot expect to get any linear independence in Proposition \ref{P6} because our vectors can have arbitrarily small norms and hence there can be an arbitrarily large number of them. However, we can remove appropriate vectors from the last $r-1$-sets until they are linearly independent and spanning. Putting the removed vectors into the first set, we get a partition into a spanning set and $r-1$ linearly independent spanning sets. \begin{corollary}\label{P6.5} Let $\{f_i\}_{i\in I}$ be a Parseval frame for $\mathbb H_N$ and $r$ a natural number so that $\|f_i\|^2 \le \frac{1}{r}$ for every $i\in I$. Then there is a partition $\{I_j\}_{j=1}^r$ of $I$ so that \[ \span\ \{f_i:i\in I_j\} = \mathbb H_N,\ \ \mbox{for all $j=1,2,\ldots,r$}.\] \end{corollary} In the following we answer a question concerning the partition of equal norm Parseval frames into spanning sets which continues Proposition \ref{P5}, and had been left open in \cite{CT}. We will be considering partitions which maximize dimensions in a very particular way. \begin{definition} Let $\{f_i\}_{i \in I}$ be a family of vectors. We say that a partition $\{I_1,\ldots,I_M\}$ of the index set $I$ has the maximality property (MD)\ if whenever $\{J_i\}_{i=1}^M$ is any partition of $I$ satisfying that for all $1\le i\le M$, dim span $\{f_j\}_{j\in J_i} \ge$ dim span $\{f_j\}_{j\in I_i}$, then dim span $\{f_j\}_{j\in I_i} =$ dim span $\{f_j\}_{j\in J_i}$ for all $i=1,2,\ldots,M$. \end{definition} A straightforward consequence of maximality is the following: \begin{lemma}\label{keylemma} Let $\mathcal F = \{f_i: i\in I\}$ be a finite collection of vectors in a vector space. Let $M\in \mathbb N$ and $\{I_j: j=1,\ldots,M\}$ be a partition of $I$ satisfying property (MD). If $f_k\in I_p$ and $f_k = \sum_{l \in I_p, l\not= k} \alpha_l f_l$, then $f_k\in \span(\mathcal{F}_{I_j})$ for all $1\le j \le M$. \end{lemma} \begin{proof} Assuming the hypothesis of the lemma, if $f_k = \sum_{l \in I_p, l\not= k} \alpha_l f_l$, then removing $f_k$ from $I_p$ keeps $\dim \span (\mathcal{F}_{I_p})$ constant. By property (MD), moving $f_k$ into another $I_j$, $j\not= p$ cannot increase $\dim \span(\mathcal{F}_{I_j})$, and the result follows. \end{proof} If there are linear dependent sets in a partition having property (MD)\ then we can move suitable vectors from one set to another. The following definition will be used to help us keep track of which vectors are being moved. \begin{definition} Let $\{f_i: i\in I\}$ be a collection of vectors in a vector space and let $\{I_j: j = 1,\ldots,M\}$ be a partition of $I$. We define a \emph{chain of length one} to be a set $\{(a,b)\}$ with $a \in I_b$, $b \in \{1, 2, \dots, M\}$ and $f_{a} = \sum_{j\in I_{b}, j \ne a} \alpha_j f_j$ for some choice of constants $\{\alpha_j\}_{j \in I_b, j \ne a}$. We define a \emph{chain of length $n$} to be a finite sequence $\{(a_1, b_1),\ldots, (a_n, b_n)\}$, where $a_i \in I$ and $b_i\in \{1,\ldots,M\}$, such that \begin{itemize} \item $(a_1,b_1)$ is a chain of length one, \item for $2\le i \le n$, $a_i \in I_{b_i}$ and $f_{a_i} = \alpha f_{a_{i-1}} + \sum_{j\in I_{b_i}, j\not= a_i} \alpha_j f_j$ for some $\alpha \not= 0$, and \item $a_i \not= a_k$ for $i \not= k$. \end{itemize} A chain of length $n$ starting with $a_1 \in L\subset I$ and ending at $a_n\in I$ is a \emph{chain of minimal length starting in $L$ and ending at $a_n$} if every chain starting in $L$ and ending at $a_n$ has length greater than or equal to $n$. \end{definition} We recall the following lemma. \begin{lemma}[Casazza, Kutyniok, Speegle]\label{chainlem} Let $\{f_i: i\in I\}$ be a collection of vectors in a vector space, let $\{I_j: j = 1,\ldots,M\}$ be a partition of $I$, and let $L \subset I_1$. If $\{(a_1,b_1),\ldots,(a_n, b_n)\}$ is a chain of minimal length starting in $L$ and ending at $a_n$, then for each $1\le i \le n$, $\{(a_1,b_1),\ldots,(a_i,b_i)\}$ is a chain of minimal length starting in $L$ and ending at $a_i$. \end{lemma} \begin{proof} By induction it suffices to show that $\{(a_1,b_1),\ldots,(a_{n-1}, b_{n-1})\}$ is a chain of minimal length. Suppose, for the sake of contradiction, that there did exist a chain $\{(u_1,v_1), \ldots, (u_k,v_k)\}$ such that $u_k = a_{n-1}$ and $k < n- 1$. Since $\{\chain ab1n\}$ is a chain, $$ f_{a_n} = \alpha f_{a_{n-1}} + \sum_{j\in I_{b_{n}}, j\not= a_n} \alpha_j f_j $$ for some $\alpha \not=0$. Therefore, either $\{\chain uv1k, (a_n, b_n)\}$ is a chain with length $k+1<n$ or $a_n = u_i$ for some $i\le k $, either of which contradicts the minimality of $n$. \end{proof} \begin{lemma} Let $\mathcal F = \{f_i: i\in I\}$ be a finite collection of vectors, and $M\in \mathbb{N}$. There exists among all the partitions of $I$ into $M$ non-empty subsets a partition $\{I_1, I_2, \dots I_M\}$ with the property (MD). This partition can be chosen so that $\mathcal{F}_{I_j}$ is linearly independent for all $2 \le j \le M$. \end{lemma} \begin{proof} The set of partitions of $I$ into $M$ sets has a partial ordering with respect to which two partitions $\{I_j\}_{j=1}^M$ and $\{J_j\}_{j=1}^M$ satisfy $\{I_j\}_{j=1}^M \le \{J_j\}_{j=1}^M$ if $\dim \mathcal{F}_{I_j} \ge \dim \mathcal{F}_{J_j}$ for all $j \in \{1, 2, \dots, M\}$. $I$ is a finite set, so there are maximal elements. By definition, these partitions have the property (MD). Assume that there is a partition with property (MD)\ which contains more than one set for which the associated vectors are linearly dependent, say $I_1$ and $I_2$. We can then successively remove indices from $I_2$ and place them into $I_1$ if the associated vectors are linear combinations of others remaining in the set indexed by $I_2$. After finitely many such moves, $\mathcal{F}_{I_2}$ is linearly independent. Moreover, by Lemma~\ref{keylemma}, the span of $\mathcal{F}_{I_1}$ and $\mathcal{F}_{I_2}$ retain their dimensions, which means the maximality is preserved. \end{proof} If $\mathcal{F}_{I_2},\ldots, \mathcal{F}_{I_M}$ are linearly independent, $L = \{i\in I_1: f_i = \sum_ {j\in I_1, j\not= i} \alpha_j f_j\}$, and $\{\chain ab1n\}$ is a chain of minimal length starting in $L$, it follows that for each $1\le i < n$, $b_i \not= b_{i + 1}$. In this case, we can track the changes in the partition as vectors are moved among the sets in a straightforward manner. \begin{definition} If $\mathcal{F}_{I_2},\ldots, \mathcal{F}_{I_M}$ are linearly independent, then proceeding by induction, we can define \[ U_k^1 = I_k, \,\,\,\,\, 1\le k \le M, \] and for $2\le i\le n$, \begin{eqnarray*} U_k^i &=& U_k^{i-1} \,\text{ for } k\not= b_{i-1}, k\not= b_i,\\ U_{b_i}^i &=& U_{b_i}^{i-1} \cup \{a_{i-1}\},\\ U_{b_{i-1}}^i &=& U_{b_{i-1}}^{i-1} \setminus \{a_{i-1}\}. \end{eqnarray*} \end{definition} \begin{lemma} \label{subc2} Let $\mathcal F = \{f_i: i\in I\}$ be a finite collection of vectors, and $\{I_1, I_2, \dots I_M\}$ a partition with the property (MD)\ for which $\mathcal{F}_{I_2}, \dots, \mathcal{F}_{I_M}$ are linearly independent. Let $L$ be as above and assume that $\{\chain ab1n\}$ is a minimal chain starting in $L$. For each $1\le i \le n$, $f_{a_i}$ can then be written as the sum \begin{equation}\label{4.5} f_{a_i} = \sum_{j\in I_{b_i}, j\not\in \{a_p: 1\le p \le n\}} \alpha_j f_j + \sum_{j\in U_{b_i}^i \cap \{a_p: 1 \le p < i\}} \alpha_j f_j. \end{equation} \end{lemma} \begin{proof} For the case $i = 1$, note that $a_1\in L$ implies that $f_{a_1} = \sum_{j\in L, j\not= a_1} \alpha_j f_j$ for some choice of $\alpha_j$. By Lemma~\ref{chainlem} none of these $j\in L$ can be in $\{a_p: 1\le p \le n\}$ since this would not be a chain of minimal length starting in $L$. Recalling that $b_i = 1$, the claim is proven for $i = 1$. Proceeding by induction, let $i\in \{1,\ldots,n\}$ and we assume (\ref{4.5}) is true for $1\le k < i$. We will show that it is also true for $i$. Note that \begin{eqnarray} f_{a_i} &=& \alpha f_{a_{i-1}} + \sum_{j\in I_{b_i}, j\not= a_{i}} \alpha_j f_j\\ &=& \alpha f_{a_{i-1}} + \sum_{j\in I_{b_i} \cap U_{b_i}^i, j \not= a_i} \alpha_j f_j + \sum_{j\in I_{b_i} \setminus U_{b_i}^i} \alpha_j f_j\nonumber\\ \label{eqqq} &=& \alpha f_{a_{i-1}} + \sum_{j\in I_{b_i} \cap U_{b_i}^i, j\not = a_i} \alpha_j f_j + \sum_{j\in I_{b_i} \cap \{a_p: 1\le p < i - 1\}} \alpha_j f_j, \end{eqnarray} where we have used in the last two lines that $I_{b_i} \cap \{a_p: 1\le p < i-1\} = I_{b_i} \setminus U_{b_i}^i.$ Now, suppose for the sake of contradiction that there is a $j\in I_{b_i} \cap U_{b_i}^i$ such that $\alpha_j \not= 0$ and $j = a_p$ for some $p > i$. Then $\{\chain ab1{i-1}, (a_p,b_i)\}$ is a chain starting in $L$, which contradicts the minimality of the chain $\{\chain ab1n\}$. So, using the induction hypothesis on each term in the last sum in (\ref{eqqq}) and combining terms, one obtains \[ f_{a_i} = \alpha f_{a_{i-1}} + \sum_{j\in I_{b_i}, j\not\in \{a_p: 1\le p \le n\}} \tilde\alpha_j f_j + \sum_{j\in U_{b_i}^i \cap \{a_p : 1\le p < i\} }\tilde \alpha_j f_j \] with an apporpriate choice of $\tilde\alpha_j$'s. \end{proof} \begin{lemma}\label{claim1} Let $\mathcal F = \{f_i: i\in I\}$ be a finite collection of vectors, and $\{I_j\}_{j =1}^M$ a partition of the index set $I$ into $M\in \mathbb N$ non-empty sets which has the proeprty (MD)\ and for which sets $I_2$, $I_3$, \dots $I_{M}$ index linearly independent sets. Moreover, let $L = \{i \in I_1: f_i = \sum_{j\in I_1, j\not= i} \alpha_j f_j\}$, $L_0 = \{i\in I: $ there is a chain starting in $L$ and ending at $i\}$, and $L_j = L_0 \cap I_j$ for $1\le j \le M$. If $\{(a_1, b_1),\ldots, (a_n, b_n)\}$ is a chain of minimal length starting in $L$ and ending at $a_n$, then $f_{a_n} \in \span (\mathcal{F}_{L_m})$ for all $1\le m \le M$. \end{lemma} \begin{proof} We show that, if $\{\chain ab1n\}$ is a chain of minimal length starting in $L$ and ending at $a_n$, then $f_{a_n}\in \span(\mathcal{F}_{L_m})$ for each $1\le m\le M$. For $n = 1$, fix $m\in \{1,\ldots,M\},$ and observe that $a_1\in L$. Hence, by Lemma~\ref{keylemma}, we can write $f_{a_1} = \sum_{l\in I_m} \alpha_l f_l$. For each $l$ such that $\alpha_l\not= 0$, $(a_1,1), (l,m)$ is a chain ending at $l$. Therefore, $f_{a_1} \in \span(\mathcal{F}_{L_m})$, as desired. By Lemma \ref{subc2} and the fact that $I_{b_i} \setminus \{a_p: 1\le p \le n\}\subset U^k_{b_i}$ for all $1\le k \le n$, we have that $f_{a_i} \in \span (\mathcal{F}_{U_{b_i}^i \setminus \{a_i\}})$. Therefore, $\dim\span(\mathcal{F}_{U_{b_i}^i}) = \dim\span(\mathcal{F}_{U_{b_i}^{i+ 1}})$. In particular, the partition $\{U_k^i:1\le k \le M\}$ satisfies property (MD). By property (MD), Lemma~\ref{subc2}, and Lemma~\ref{keylemma}, $f_{a_n} \in \span(\mathcal{F}_{U_m^n})$ for each $1\le m \le M$. Therefore, for $m \not = b_n$, there exist $\alpha_j^{(0)}$ such that \begin{eqnarray} f_{a_n} &=& \sum_{j\in U_m^n} \alpha_j^{(0)} f_j = \sum_{j\in U_m^n \cap I_m} \alpha_j^{(0)} f_j + \sum_{j\in U_m^n \setminus I_m} \alpha_j^{(0)} f_j \nonumber \\ &=& \label{ppp} \sum_{j\in U_m^n\cap I_m} \alpha_j^{(0)} f_j + \sum_{j\in \{a_p : b_{p + 1} = m, 1 \le p < n-1\}} \alpha_j^{(0)} f_j. \end{eqnarray} By definition of a chain, for each $a_p$ such that $b_{p + 1} = m$ and $1\le p < n - 1$, \begin{equation}\label{iii} f_{a_p} = \alpha^p f_{a_{p + 1}} + \sum_{j\in I_m, j\not= a_{p + 1}} \alpha_j^{(p)} f_j, \end{equation} for some choice of $\alpha_j^{(p)}$ and some $\alpha^{(p)}\not= 0$. Fix $j_0$ such that $\alpha_{j_0}^{(0)} \not= 0$ in (\ref{ppp}). We show that $j_0\in L_m$, which finishes the proof of the lemma. Clearly, if $j_0\in \{a_1,\ldots, a_n\}$, then we are done, so we assume that $j_0 \not\in \{a_1,\ldots, a_n\}$. Case 1: There is some $1\le p < n - 1$ such that $b_{p + 1} = m$ and $\alpha_{j_0}^{(p)} \not= 0$. Then, one can solve (\ref{iii}) for $f_{j_0}$ to obtain \[ f_{j_0} = \beta f_{a_p} + \sum_{j\in I_m, j\not= j_0, j\not= a_p} \beta_j f_j \] for some $\beta \not= 0$. Hence, $\chain ab1p, (j_0, m)$ is a chain and $j_0 \in L_m$. Case 2: For each $1\le p< n-1$ such that $b_{p + 1} = m$, we have $\alpha_{j_0}^{(p)} = 0$. We have \begin{eqnarray*} f_{a_n} &=& \sum_{j\in U_m^n \cap I_m} \alpha_j^{(0)} f_j + \sum_{j\in \{a_p : b_{p + 1} = m, 1\le p < n - 1\}} \alpha_j^{(0)} f_j \\ \label{belangrik} &=&\sum_{j\in U_m^n \cap I_m} \alpha_j^{(0)} f_j + \sum_{p\in \{p : b_{p+1} = m, 1\le p < n - 1\}} \alpha_{a_p}^{(0)} f_{a_p}\\ &=&\sum_{j\in U_m^n \cap I_m} \alpha_j^0 f_j + \sum_{p\in \{p : b_{p+1} = m, 1\le p < n - 1\}} \alpha_{a_p}^{(0)} \bigl( \alpha^{(p)} f_{a_{p + 1}} + \sum_{j\in I_m, j\not= a_{p + 1}} \alpha_j^{(p)} f_{j}\bigr) \\ &=&\alpha_{j_0}^{(0)} f_{j_0} + \sum_{j\in I_m, j\not= j_0} \tilde \alpha_j f_j, \end{eqnarray*} where the first equality is (\ref{ppp}), the second equality is a re-indexing, the third equality follows from (\ref{iii}), and the last equality holds for some choice of $\tilde \alpha_j$ by combining sums, since $\alpha_{j_0}^{(p)} = 0$ for all $1\le p < n- 1$ such that $b_{p + 1} = m$, and $j_0\not\in \{a_1,\ldots,a_n\}$. Therefore, $\{\chain ab1{n}, (j_0, m)\}$ is a chain and $j_0\in L_m$. \end{proof} The purpose of this is to prove the following: \begin{theorem}\label{cor5} Let $\{f_i\}_{i\in I}$ be a finite collection of vectors in a finite dimensional vector space $X$. Assume (1) $\{f_i\}_{i\in I}$ can be partitioned into $r+1$-linearly independent sets, and (2) $\{f_i\}_{i\in I}$ can be partitioned into a set and $r$ linearly independent spanning sets. Then there is a partition $\{I_i\}_{i=1}^{r+1}$ so that $\{f_j\}_{j\in I_i}$ is a linearly independent spanning set for all $i=2,3,\ldots,r+1$ and $\{f_i\}_{i\in I_{1}}$ is a linearly independent set. \end{theorem} \begin{proof} We choose the partition $\{I_i\}_{i=1}^{r+1}$ of $I$ that maximizes dim span $\{f_j\}_{j\in I_{1}}$ taken over all partitions so that the last $r$ sets span $X$. If $\{J_i\}_{i = 1}^{r + 1}$ is a partition of $I$ such that for all $1\le i\le r + 1$, dim span $\{f_j\}_{j\in J_i} \ge$ dim span $\{f_j\}_{j\in I_i}$, then dim span $\{f_j\}_{j\in I_i} =$ dim span $\{f_j\}_{j\in J_i}$ for all $i=2,\ldots,r+1$ since dim span $\{f_j\}_{j\in I_i} = \dim X$, and $\dim \span \{f_j\}_{j\in I_{1}} = \dim \span \{f_j\}_{j\in J_{1}}$ by construction. This means, the chosen partition has the property (MD)\ and the properties asserted by Lemma~\ref{claim1}. Suppose that this does not partition $\mathcal{F}_I$ into linearly independent sets, i.e. $\mathcal{F}_{I_1}$ is not linearly independent. As in Lemma~\ref{claim1}, let $L = \{i\in I_1: f_i = \sum_ {j\in I_1, j\not= i} \alpha_j f_j\}$ be the index set of the ``linearly dependent vectors" in $I_1$, $L_0 = \{i\in I: \text {\rm there is a chain starting in $L$ ending at $i$}\},$ and $L_j = L_0\cap I_j, 1\le j \le r+1.$ Let $S = \span (\mathcal{F}_{L_0})$. By Lemma~\ref{claim1}, $S = \span(\mathcal{F}_{L_j})$ for all $1\le j\le r+1$. Moreover, for $1\le j \le r+1$, $i\in L_j$ implies that $i\in I_j$ and $f_i \in S$. Therefore, \begin{equation*} S \subset \span \{f_i: i\in L_j\} \subset \span \{f_i: i\in I_j, f_i\in S\} = S. \end{equation*} Let $J = \{i\in I : f_i\in S\}$. By construction, $L\subset J$. Let $d = \dim(S)$ and see that, by the preceding portion of this proof, $\dim \span(\mathcal{F}_J) = d$. Moreover, \[|J| = |L_1| + \cdots + |L_M| = |L_1| + rd > d(r+1),\] because $L_1$ is linearly dependent, since it contains $L$ by virtue of chains of length one. Therefore, for $J = \{i\in I: f_i\in S\}$, $\frac {|J|} {\dim\span(\mathcal{F}_J)} > r+1$. This is in contradiction with assumption (1), which implies by the Rado Horn theorem that $|J|/d \le r+1$. \end{proof} Combining Propositions \ref{P5} and \ref{P6} with Theorem~\ref{cor5} we conclude: \begin{corollary} Let $\{f_i\}_{i\in I}$ be an equal norm Parseval frame for $\mathbb H_N$ with $|I|=rN+k$ with $0\le k <N$. Then there is a partition $\{I_i\}_{i=1}^{r+1}$ of $I$ so that for $i\in \{2,\ldots,r+1\}$, $\{f_j\}_{j\in I_i}$ is a linearly independent spanning set and $\{f_j\}_{j\in I_{1}}$ is linearly independent. \end{corollary} If $r\ge 2$ then this result implies that each set of frame vectors has a complement which is spanning, which was already obtained in Section~3. The insight of this last corollary is that with the lower norm bound implicit in the equal-norm Parseval property, the partition can be chosen to consist of linearly independent sets. Moreover, the complement of each set can then be partitioned into at least $r-1$ spanning sets. \paragraph{\bf Acknowledgments.} The authors would like to thank the American Institute of Mathematics for hospitality and the support of the Structured Quartet Research Ensemble on the Kadison Singer Problem.

https://arxiv.org/abs/math/0505080

Conway's napkin problem

The napkin problem was first posed by John H. Conway, and written up as a `toughie' in "Mathematical Puzzles: A Connoisseur's Collection," by Peter Winkler. To paraphrase Winkler's book, there is a banquet dinner to be served at a mathematics conference. At a particular table, $n$ men are to be seated around a circular table. There are $n$ napkins, exactly one between each of the place settings. Being doubly cursed as both men and mathematicians, they are all assumed to be ignorant of table etiquette. The men come to sit at the table one at a time and in random order. When a guest sits down, he will prefer the left napkin with probability $p$ and the right napkin with probability $q=1-p$. If there are napkins on both sides of the place setting, he will choose the napkin he prefers. If he finds only one napkin available, he will take that napkin (though it may not be the napkin he wants). The third possibility is that no napkin is available, and the unfortunate guest is faced with the prospect of going through dinner without any napkin!We think of the question of how many people don't get napkins as a statistic for signed permutations, where the permutation gives the order in which people sit and the sign tells us whether they initially reach left or right. We also keep track of the number of guests who get a napkin, but not the napkin they prefer. We build a generating function for the joint distribution of these statistics, and use it to answer questions like: What is the probability that every guest receives a napkin? How many guests do we expect to be without a napkin? How many guests are happy with the napkin they receive?

\section{Introduction} The problem studied in this article first appeared in the book ``Mathematical Puzzles: A Connoisseur's Collection," by Peter Winkler \cite{Winkler}, and was inspired by a true story. Rather than recounting the problem and the story ourselves, we prefer to quote directly from ``Mathematical Puzzles": \begin{quote} \textsc{The Malicious Maitre D'} {\small At a mathematics conference banquet, 48 male mathematicians, none of them knowledgeable about table etiquette, find themselves assigned to a big circular table. On the table, between each pair of settings, is a coffee cup containing a cloth napkin. As each person is seated (by the maitre d'), he takes a napkin from his left or right; if both napkins are present, he chooses randomly (but the maitre d' doesn't get to see which one he chose). In what order should the seats be filled to maximize the expected number of mathematicians who don't get napkins? \ldots This problem can be traced to a particular event. Princeton mathematician John H. Conway came to Bell Labs on March 30, 2001 to give a ``General Research Colloquium." At lunchtime, [Winkler] found himself sitting between Conway and computer scientist Rob Pike (now of Google), and the napkins and coffee cups were as described in the puzzle. Conway asked how many diners would be without napkins if they were seated in \emph{random} order, and Pike said: ``Here's an easier question---what's the \emph{worst} order?" } \end{quote} The problem of the Malicious Maitre D' is not horribly difficult; if you're having trouble finding a solution, you can see Winkler's book for a nice explanation. In this paper, it is Conway's problem that we focus on. Again, from \cite{Winkler}: \begin{quote} \textsc{Napkins in a Random Setting} {\small Remember the conference banquet, where a bunch of mathematicians find themselves assigned to a big circular table? Again, on the table, between each pair of settings, is a coffee cup containing a cloth napkin. As each person sits down, he takes a napkin from his left or right; if both napkins are present, he chooses randomly. This time there is no maitre d'; the seats are occupied in random order. If the number of mathematicians is large, what fraction of them (asymptotically) will end up without a napkin? } \end{quote} Let $p$ be the probability that a diner prefers the left napkin and $q=1-p$ be the probability that a diner prefers the right napkin. For the case $p=q=1/2$, Winkler's book gives two proofs of the answer: $(2- \sqrt{e})^2 \approx .12339675$. One is combinatorial, while the other, taken from a more general result due to Aidan Sudbury \cite{Sudbury}, is analytical. In fact Sudbury gives the expected proportion of diners without a napkin as (asymptotically) \begin{equation}\label{eq:Sud} \frac{(1 - pe^q)(1 - qe^p)}{pq}. \end{equation} (As an aside, we took an informal survey of 55 mathematicians and found about 69\% would prefer the napkin on the left. According to Sudbury's result, we would thus expect about 10.58 percent of the guests to get stuck without napkins.) In this paper, we use combinatorial methods to produce the generating function for the probability that at a table for $n$ people, $i$ of them have no napkin and $j$ of them have a napkin, but not the napkin they prefer. This generating function allows for a thorough statistical analysis of the problem, including a new proof of \eqref{eq:Sud}. >From our point of view, the number of people without a napkin is a statistic for signed permutations; just not one so well studied as inversions, descents, and such. We consider the order in which guests sit down at a place as a permutation of $[n]= \{1,2,\ldots,n\}$, while their preference for the right or left napkin is given by a plus or minus sign. We label the places $1,2,3,\ldots,n$ counter-clockwise, so that place $i$ has place $i-1$ on its left, place $i+1$ on its right, and place $n$ is to the left of place 1. With this convention, the signed permutation $(2,-3,4,-1)$ describes the following sequence of events at a table for four. The person sitting in place 4 sits down first and takes the napkin on his left. The person in place 1 sits next and takes the napkin on his right. The person at place 2 sits third and wants to take the napkin on his left, but since that napkin is already taken, he is forced to take the napkin on the right. Finally, poor person 3 sits last to find no napkin at his place. \begin{figure}\label{fig:intro_ex} $$ \begin{array}{|l|l|} \hline \xymatrix@!0@C=4ex@R=2.9ex{ \\ &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&&\\ \\ && *=0{\bullet} && *=0{\bullet} && *=0{\bullet} && *=0{\bullet}\ar@{-}[luu] \\ && \e{2} && \e{-3} && \e{4} && \e{-1} \\ \\ } & \xymatrix@!0@C=4ex@R=2.9ex{ \\ &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&&\\ \\ && *=0{\bullet}\ar@{-}[ruu] && *=0{\bullet} && *=0{\bullet} && *=0{\bullet}\ar@{-}[luu] \\ && \e{2} && \e{-3} && \e{4} && \e{-1} \\ \\ }\\ \hline \xymatrix@!0@C=4ex@R=2.9ex{ \\ &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&&\\ \\ && *=0{\bullet}\ar@{-}[ruu] && *=0{\bullet}\ef\de && *=0{\bullet} && *=0{\bullet}\ar@{-}[luu] \\ && \e{2} && \e{-3} && \e{4} && \e{-1} \\ \\ }& \xymatrix@!0@C=4ex@R=2.9ex{ \\ &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&&\\ \\ && *=0{\bullet}\ar@{-}[ruu] && *=0{\bullet}\ef\de && *=0{\bullet}\ar@{--}[ruu] && *=0{\bullet}\ar@{-}[luu] \\ && \e{2} && \e{-3} && \e{4} && \e{-1} \\ \\ }\\ \hline \end{array} $$ \caption{A table for four.} \end{figure} In Figure~\ref{fig:intro_ex} we illustrate this sequence of events. There a ``$\bullet$'' and a ``$\curlywedge$'' represent a seat and a napkin, respectively. Moreover, a solid line symbolizes an arm taking a napkin and a broken line symbolizes an imaginary arm reaching for the preferred napkin. When looking at these diagrams, keep in mind that the table is circular: the guests at places $1$ and $4$ are neighbors. Each signed permutation thus corresponds to a particular set of diners who will have no napkin. In particular, for each permutation there is some number of diners without napkins, and we seek to determine what proportion of signed permutations leave a given number of diners napkinless. We remark that this problem makes for interesting mealtime conversation, and we have heard many suggestions for variations on the theme. What if instead of a math conference banquet with all men, it is a dinner party for couples, and the couples enter one at a time, lady sitting first? What if the dinner party is a mixer for singles, where now all the ladies enter first, sitting in alternating seats?\footnote{This variation is actually not very interesting mathematically, as it is not difficult to see the women will always get their choice of napkin, and that 1/4 (or $pq$ in general) of the men are expected to be without a napkin.} What if the guests don't mind looking farther afield than simply to their immediate left and right, say, reaching as many as two places over for a napkin? Exploring answers to these questions may lead to some interesting mathematics, but in this article we stay within the original question, where the guests are all male mathematicians, they enter in a totally random order, and they are too shy to reach beyond their immediate left or right to find a napkin. In section \ref{sec:def} we define the generating functions that we use throughout the paper. Section \ref{sec:every} explores the question of when everyone gets a napkin, and makes connections with combinatorial objects called \emph{ordered bipartitions}, due to Dominique Foata and Doron Zeilberger \cite{FoataZeil}. This connection makes subsequent proofs much simpler, as in section \ref{sec:ordgenfun}, where we derive powerful identities involving our generating functions that ultimately lead to exact formulas. In section \ref{sec:expect} we answer the original question of how many guests are expected to be without a napkin, as well as provide some other statistics of interest. Section \ref{sec:alternate} gives a second derivation of the generating function for the expected number of guests without a napkin. \section{Definitions}\label{sec:def} Let $p$ be the probability that a diner will reach to the left, and $q = 1-p$ be the probability of reaching right. We denote the set of signed permutations of $[n]$ by $\mathcal{C}_n$. For any $\pi \in \mathcal{C}_n$, let $|\pi|_{-}$ and $|\pi|_{+}$ be the number of negative and positive entries in $\pi$, respectively. Also, let $|\pi|=|\pi|_{-} + |\pi|_{+} = n$ be the length of $\pi$. Let $o(\pi)$ be the number of guests without a napkin after every guest has been seated as described by $\pi$. Note that the number of guests without a napkin is equal to the number of napkins left on the table. Furthermore, let $m(\pi)$ be the number of people who get a napkin, but not their first choice. We say a guest is \emph{napkinless} if he has no napkin, and a guest is \emph{frustrated} if he gets a napkin, but not the napkin he originally wanted.\footnote{When discussing this problem, French mathematician Sylvie Corteel argued that if the mathematicians were French, they would never take the ``incorrect" napkin. If the napkin they wanted was not there, they would simply cross their arms and refuse to eat. But of course, that is a different problem (how does the number of napkinless guests change as the proportion of French diners changes?).} Otherwise, we say the guest is \emph{happy}. Define the \emph{weight} of $\pi$ as $$ w(\pi) := p^{|\pi|_{-}}q^{|\pi|_{+}}x^{o(\pi)}y^{m(\pi)}. $$ For example, if $\pi = (2, -1, 3, 4)$, then $|\pi|_{-} = 1$, $|\pi|_{+} = 3$, and $|\pi| = 4$. Also, the napkin between places 2 and 3 is unused (guest 4 is the unlucky one), and although person 1 gets a napkin, it was not the one he wanted. Thus, $o(\pi) = 1$, $m(\pi) = 1$, and two of the guests are happy: $w(\pi) = pq^3xy$. The generating function we are interested in is \begin{equation}\label{eq:Cgf} C(p; x,y,z) := \sum_{n \geq 1}\sum_{\pi \in \mathcal{C}_n}w(\pi)\frac{z^n}{n!} = \sum_{\substack{ i,j\geq0 \\ n \geq 1}} \p(i,j,n)x^i y^j z^n, \end{equation} where $\p(i,j,n)$ denotes the probability that at a table for $n$ people, $i$ of them are napkinless, and $j$ of them are frustrated. The main result of this paper, Theorem \ref{thm:exact}, is an exact formula for $C(p; x,y,z)$. Our approach is to first ``straighten" the table. Suppose that instead of a circular table with $n$ places and $n$ napkins, we look at a straight table with $n$ places and $n+1$ napkins, so that each place has a napkin on both its left and right. If we know all that can possibly happen in this situation, then in order to determine the circular case, we just consider that the last person to enter the room sits ``between" the first and last person on the linear table. Let us make this connection more precise. Let $\mathcal{N}_n$ be the set of all signed permutations of $[n]$ that result in taking neither napkins from the ends of the table. Similarly, $\mathcal{L}_n$ (resp. $\mathcal{R}_n$) denotes the set of all signed permutations of $[n]$ that result in the left end napkin being taken but not the right (resp. right but not left), and $\mathcal{B}_n$ denotes those signed permutations that result in both end napkins being taken. We note that $\mathcal{C}_n = \mathcal{N}_n \cup \mathcal{L}_n \cup \mathcal{R}_n \cup \mathcal{B}_n$. For the straight table we define the weight of $\pi$ as $$ w_S(\pi) := p^{|\pi|_{-}}q^{|\pi|_{+}}x^{o(\pi)}y^{m(\pi)}, $$ where $o(\pi)$ is still the number of people without a napkin at the (now linear) table, so that $o(\pi)+1$ is the number of napkins left on the table. Moreover, we define the following generating functions for the linear table: \begin{align*} S(p;x,y,z) &:= \sum_{n\geq 0} \sum_{\pi \in \mathcal{C}_n} w_S(\pi) \frac{z^{n}}{n!};\\ N(p;x,y,z) &:= \sum_{n\geq 0} \sum_{\pi \in \mathcal{N}_n} w_S(\pi) \frac{z^{n}}{n!};\\ L(p;x,y,z) &:= \sum_{n\geq 1} \sum_{\pi \in \mathcal{L}_n} w_S(\pi) \frac{z^{n}}{n!};\\ R(p;x,y,z) &:= \sum_{n\geq 1} \sum_{\pi \in \mathcal{R}_n} w_S(\pi) \frac{z^{n}}{n!};\\ B(p;x,y,z) &:= \sum_{n\geq 1} \sum_{\pi \in \mathcal{B}_n} w_S(\pi) \frac{z^{n}}{n!}. \end{align*} By construction, we have \begin{equation}\label{eq:SNLRB} S(p;x,y,z) = N(p;x,y,z) + L(p;x,y,z) + R(p;x,y,z) + B(p;x,y,z). \end{equation} Further, we can observe that by symmetry \begin{alignat*}{3} C(p;x,y,z) &= C(q;x,y,z), &\;\,&\;\,& S(p;x,y,z) &= S(q;x,y,z), \\ N(p;x,y,z) &= N(q;x,y,z), &\;\,&\;\,& B(p;x,y,z) &= B(q;x,y,z), \end{alignat*} and \begin{equation} R(p;x,y,z) = L(q;x,y,z).\label{eq:RL} \end{equation} >From now on we will usually suppress the $p$ and write $C(x,y,z)$ instead of $C(p;x,y,z)$, $S(x,y,z)$ instead of $S(p;x,y,z)$, etc. Now let us recast the generating function $C(x,y,z)$ in terms of the generating functions for the linear table. Everything that has happened before the last person sits down at a table for $n$ people can be considered the result of a signed permutation of $[n-1]$ playing out on a linear table. If the last person sits at place $n$, this is obvious, but in general the labeling of the seats is unimportant, i.e., we can always cyclically permute the labels on the seats without changing the weight of the permutation. In what follows, $\pi = \sigma\tau$ means that $\pi$ is the concatenation of $\sigma$ with $\tau$. For instance, $(3,-1)(-6)(2,-4,5) = (3,-1,-6,2,-4,5)$. Let us consider a circular table for $n$ guests and assume that $\pi = \sigma(\pm n)\tau$ is a member of $\mathcal{C}_n$. Let $\pi' = \tau\sigma$. If the last person walks in and has both napkins available, then $\pi'$ must have resulted in leaving both end napkins on a linear table, that is, $\pi' \in \mathcal{N}_{n-1}$. Whether the last guest prefers the left napkin or the right napkin, he will get his choice. So, the weight of $\pi$ is: $$ w(\pi) = \begin{cases} pw_S(\pi') & \mbox{if person } n \mbox{ prefers left,}\\ qw_S(\pi') & \mbox{if person } n \mbox{ prefers right.} \end{cases} $$ If person $n$ walks in to find only the left napkin available, $\pi' \in \mathcal{L}_{n-1}$, and the last person will take that napkin regardless of preference, getting the one he wants with probability $p$, and getting frustrated with probability $q$: $$w(\pi) = \begin{cases} pw_S(\pi') & \mbox{if person } n \mbox{ prefers left,}\\ qyw_S(\pi') & \mbox{if person } n \mbox{ prefers right.} \end{cases} $$ Similarly, if person $n$ walks in to find only the right napkin available then $$w(\pi) = \begin{cases} pyw_S(\pi') & \mbox{if person } n \mbox{ prefers left,}\\ qw_S(\pi') & \mbox{if person } n \mbox{ prefers right.} \end{cases} $$ Finally, guest $n$ can walk in to find both napkins already taken. In this case, we know that $\pi' \in \mathcal{B}_{n-1}$ and that guest $n$ will be one of the napkinless guests: $$w(\pi) = \begin{cases} pxw_S(\pi') & \mbox{if person } n \mbox{ prefers left,}\\ qxw_S(\pi') & \mbox{if person } n \mbox{ prefers right.} \end{cases} $$ Figure~\ref{fig:C_NLRB} illustrates these, altogether 8, possibilities in the special case when the last person to sits at place $n$. All other cases can be reduced to this case by cyclically shifting the picture. \begin{figure}\label{fig:C_NLRB} $$ \begin{array}{|c|c|} \hline &\\ \xymatrix@!0@C=3.4ex@R=2.2ex{ &{\curlywedge} && {\curlywedge} && \ldots && {\curlywedge} && {\curlywedge} &&& \\ \\ && *=0{\bullet}\ar@{-}[ruu] && *=0{\bullet} & \ldots & *=0{\bullet} && *=0{\bullet}\ar@{-}[luu] && \c\ar@{-}[luu] \\ &&&&& N &&&&& pz \\ }& \xymatrix@!0@C=3.4ex@R=2.2ex{ &{\curlywedge} && {\curlywedge} && \ldots && {\curlywedge} && {\curlywedge} &&& \\ \\ && *=0{\bullet}\ar@{-}[ruu] && *=0{\bullet} & \ldots & *=0{\bullet} && *=0{\bullet}\ar@{-}[luu] && \c\ar@{-}[ruu] \\ &&&&& N &&&&& qz \\ }\\ &\\ \hline\hline &\\ \xymatrix@!0@C=3.4ex@R=2.2ex{ &{\curlywedge} && {\curlywedge} && \ldots && {\curlywedge} && {\curlywedge} &&& \\ \\ && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet} & \ldots & *=0{\bullet} && *=0{\bullet}\ar@{-}[luu] && \c\ar@{-}[luu] \\ &&&&& L &&&&& pz \\ }& \xymatrix@!0@C=3.4ex@R=2.2ex{ &{\curlywedge} && {\curlywedge} && \ldots && {\curlywedge} && {\curlywedge} &&& \\ \\ && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet} & \ldots & *=0{\bullet} && *=0{\bullet}\ar@{-}[luu] && \c\fe\ed \\ &&&&& L &&&&& qyz \\ }\\ &\\ \hline &\\ \xymatrix@!0@C=3.4ex@R=2.2ex{ &{\curlywedge} && {\curlywedge} && \ldots && {\curlywedge} && {\curlywedge} &&& \\ \\ && *=0{\bullet}\ar@{-}[ruu] && *=0{\bullet} & \ldots & *=0{\bullet} && *=0{\bullet}\ar@{-}[ruu] && \c\ef\de \\ &&&&& R &&&&& pyz \\ }& \xymatrix@!0@C=3.4ex@R=2.2ex{ &{\curlywedge} && {\curlywedge} && \ldots && {\curlywedge} && {\curlywedge} &&& \\ \\ && *=0{\bullet}\ar@{-}[ruu] && *=0{\bullet} & \ldots & *=0{\bullet} && *=0{\bullet}\ar@{-}[ruu] && \c\ar@{-}[ruu] \\ &&&&& R &&&&& qz \\ }\\ &\\ \hline\hline &\\ \xymatrix@!0@C=3.4ex@R=2.2ex{ &{\curlywedge} && {\curlywedge} && \ldots && {\curlywedge} && {\curlywedge} &&& \\ \\ && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet} & \ldots & *=0{\bullet} && *=0{\bullet}\ar@{-}[ruu] && \c\ar@{--}[luu] \\ &&&&& B &&&&& pxz \\ }& \xymatrix@!0@C=3.4ex@R=2.2ex{ &{\curlywedge} && {\curlywedge} && \ldots && {\curlywedge} && {\curlywedge} &&& \\ \\ && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet} & \ldots & *=0{\bullet} && *=0{\bullet}\ar@{-}[ruu] && \c\ar@{--}[ruu] \\ &&&&& B &&&&& qxz \\ }\\ & \\ \hline \end{array} $$ \caption{$C = z\big( N + (p+qy)L + (py+q)R + xB \big)$.} \end{figure} In terms of generating functions we have showed that \begin{equation}\label{eq:CNLRB} C(x,y,z) = z\big( N(x,y,z) + (p+qy)L(x,y,z) + (py+q)R(x,y,z) + xB(x,y,z) \big). \end{equation} \section{A warm-up problem: when does everyone have a napkin?} \label{sec:every} A natural question to ask (perhaps easier than the general question) is: what is the probability, $\p(0,1,n) = \p(0,n)$, that at a circular table for $n$ people, every guest has a napkin? The generating function for this probability is $C(0,y,z)$. By equation \eqref{eq:CNLRB}, we have \[C(0,y,z) = z\big( N(0,y,z) + (p+qy)L(0,y,z) + (py+q)R(0,y,z) \big), \] but since (on a straight table with at least one person) there is always at least one person without a napkin if neither end napkin is taken, $N(0,y,z) = 1$, and the above equation reduces to \begin{equation}\label{eq:C0} C(0,y,z) = z\big( 1 + (p+qy)L(0,y,z) + (py+q)R(0,y,z) \big). \end{equation} This makes intuitive sense because if everyone is to have a napkin, they all need to take the left napkin, or all need to take the right napkin. Therefore we turn our attention to $L(0,y,z)$ (since $R(0,y,z)$ follows by swapping $p$ and $q$). For $n$ a positive integer, let us consider a straight table for $n$ guests and let $\pi\in\mathcal{L}_n$. Further, let $\pi = \sigma(\pm n)\tau$, where $\sigma$ and $\tau$ are the signed permutations consisting of the letters to the left and to the right of $n$, respectively. If the last person to enter sits down at the rightmost seat (i.e., $\tau$ is empty) then, since $\pi\in\mathcal{L}_n$, he necessarily prefers the left napkin. So, $$ w_S(\pi) = w_S(\sigma)p. $$ Assume that the last person to enter sits down with at least one guest to his right (i.e., $\tau$ is not empty). Then he will be happy with probability $p$ and frustrated with probability $q$: $$ w_S(\pi) = \begin{cases} w_S(\sigma)pw_S(\tau) & \mbox{if person } n \mbox{ prefers left,}\\ w_S(\sigma)qyw_S(\tau) & \mbox{if person } n \mbox{ prefers right.} \end{cases} $$ Thus the function $L(0,y,z)$ satisfies the following differential equation: \begin{equation}\label{eq:L0deq} \frac{d}{dz}L(0,y,z) = \big( L(0,y,z) + 1 \big)\big( p + (p+qy)L(0,y,z) \big). \end{equation} Se Figure~\ref{fig:L0} for an illustration. \begin{figure}\label{fig:L0} $$ \begin{array}{|c|} \hline \xymatrix@!0@C=3.4ex@R=2.2ex{ \\ &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&& &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&{\curlywedge} && {\curlywedge} &\\ \\ && *=0{\bullet}\fe\ed && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet}\ar@{-}[luu] && \c\ar@{-}[luu] &&& \leftrightarrow &&& *=0{\bullet}\fe\ed && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet}\ar@{-}[luu] &&*=0{\bullet}\ar@{-}[luu]&&,&& \\ \\ &&&&&& \frac{d}{dz}L &&&&&&&&&&&&&(L+1)&&&&& p\\ \\ }\\ \hline \hline \xymatrix@!0@C=3.4ex@R=2.2ex{ \\ &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&& &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&{\curlywedge} && {\curlywedge} &\\ \\ && *=0{\bullet}\fe\ed && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet}\ar@{-}[luu] && \c\ar@{-}[luu] && *=0{\bullet}\ar@{-}[luu] &&& \leftrightarrow &&& *=0{\bullet}\fe\ed && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet}\ar@{-}[luu] &&,&& *=0{\bullet}\ar@{-}[luu] && \\ \\ &&&&&& \frac{d}{dz}L &&&&&&&&&&&&(L+1) &&&& p && L\\ \\ }\\ \hline \xymatrix@!0@C=3.4ex@R=2.2ex{ \\ &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} & && &{\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} &&{\curlywedge} && {\curlywedge} &\\ \\ && *=0{\bullet}\fe\ed && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet}\ar@{-}[luu] && \c\fe\ed && *=0{\bullet}\ar@{-}[luu] &&& \leftrightarrow &&& *=0{\bullet}\fe\ed && *=0{\bullet}\ar@{-}[luu] && *=0{\bullet}\ar@{-}[luu] &&,&& *=0{\bullet}\ar@{-}[luu] && \\ \\ &&&&&& \frac{d}{dz}L &&&&&&&&&&&& (L+1) &&&& qy && L\\ \\ }\\ \hline \end{array} $$ \caption{ $\frac{d}{dz}L = (L+1)p + (L+1)pL + (L+1)qyL$, where $L=L(0,y,z)$.} \end{figure} Let $L_n = [z^n]L(1/2;0,1,z)$, where $[z^n]F(z)$ denotes the coefficient of $z^n$ in $F(z)$. Using \eqref{eq:L0deq} we compute the $L_n$ for some small values: $L_1 = 1$, $L_2 = 3$, $L_3 = 13$, $L_4 = 75$, $L_5 = 541$, \dots, and plug them into Sloane's Encyclopedia \cite{Sloane}. Luckily, we get a hit with sequence A000670! These numbers happen to be fairly well known as the ``ordered Bell numbers," or the number of ordered set partitions of $[n]$. We shall now give a bijection between the ordered set partitions of $[n]$ and permutations for which everyone takes the left napkin. This bijection will also give the general solution to \eqref{eq:L0deq}. \subsection{Ordered set partitions} We now give a bijection between ordered set partitions of $[n]$ and signed permutations that correspond to everyone at a linear table taking the napkin on the left. An \emph{ordered set partition}, $\alpha$, of $[n]$ is a word $$\alpha = B_1 B_2 \cdots B_k $$ where the ``blocks" $B_1$, $B_2$, \dots, $B_k$ are subsets of $[n]$ such that $\{B_1,B_2,\ldots,B_k\}$ is a set partition of $[n]$. By convention we require that the elements of each block are written in decreasing order. We describe the bijection with an example. Let \[\alpha = \{5,2\}\{6\}\{7,4,1\}\{3\}. \] First, we give a minus sign to the least element of each block, then we remove the braces, to obtain \[\pi = (5,-2,-6,7,4,-1,-3), \] a permutation corresponding to a situation where everyone takes the napkin on the left. We can see that we indeed have produced a permutation in which everyone takes the napkin on the left by thinking of $\pi$ as a diagram for the entry times and napkin preferences of everyone at the table. With $\pi$ as above, we see that the person seated second from the right entered first and took the napkin on the left. The person seated third from the right wanted to take that napkin, but entered fourth, and so was forced to take the napkin on his left. Because we required that the elements of a block are in decreasing order, anytime someone has a preference for the napkin on the right, they find that it was taken by the person to their right. Clearly this process is reversible. Given a permutation where everyone takes the napkin on the left, it must have at least one minus sign. In particular, it must have a minus sign on 1, since the first person to enter must take the napkin on his left. Anybody with a plus sign immediately to the left of a person with a minus sign must enter after that person. And if two or more people with plus signs are sitting (consecutively) to the left of a person with a minus sign, they must arrive in order of closeness to the minus sign; the closest to the minus sign first, followed by the second closest to the minus sign, and so on. This gives us the block structure of the partition on $[n]$. Reading $\pi$ from left to right, we separate the blocks by putting walls immediately to the right of any number appearing with a minus sign. For example, if \[\pi = (7, -5, -6, 4, -1, 3, -2), \] we convert this permutation into \[\alpha = \{7,5\}\{6\}\{4,1\}\{3,2\}. \] Now we can build the generating function for $L(0,y,z)$ by purely combinatorial means. Taking an approach from \cite{FoataZeil} (more on that paper in a bit), let $$ H(p; y,z) := \frac{pz}{1!} + \frac{pqyz^2}{2!} + \cdots + \frac{p(qy)^{n-1}z^n}{n!} + \cdots = \frac{p(e^{qyz}-1)}{qy} $$ Then $H(y,z) := H(p; y,z)$ is the generating function for single blocks, $\{n, \ldots, 2, 1\}$, since for every block, only the person corresponding to the least number gets the napkin he wants, leaving the other $n-1$ people frustrated. For example with $n = 5$: $$ w\Bigg( \vcenter{\xymatrix@!0@C=3.4ex@R=2.2ex{ {\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} && {\curlywedge} \\ \\ & *=0{\bullet}\fe\ed && *=0{\bullet}\fe\ed && *=0{\bullet}\fe\ed && *=0{\bullet}\fe\ed && *=0{\bullet}\ar@{-}[luu] \\ & qy && qy && qy && qy && p }}\, \Bigg) = p (qy)^4. $$ Therefore $$(1-H(y,z))^{-1} = 1 + H(y,z) + (H(y,z))^2 + \cdots $$ is the generating function for ordered sequences of blocks, and thus \[ L(0,y,z) = (1 - H(y,z))^{-1} - 1 = \frac{p(e^{qyz}-1)}{qy-p(e^{qyz}-1)}. \] One can easily check that this expression indeed satisfies the differential equation~\eqref{eq:L0deq}. Due to equations \eqref{eq:RL} and \eqref{eq:C0} we are now in a position to give the generating function for the probability that everyone gets a napkin on a circular table: \begin{align*} C(0,y,z) &= z\big( 1 + (p+qy)L(p;0,y,z) + (py+q)L(q;0,y,z) \big) \\ &= z\!\left(1 + \frac{(p+qy)p(e^{qyz}-1)}{qy-p(e^{qyz}-1)} + \frac{(py+q)q(e^{pyz}-1)}{py-q(e^{pyz}-1)} \right). \end{align*} \subsection{Ordered bipartitions} Now we will generalize the correspondence just described. The paper of Foata and Zeilberger \cite{FoataZeil} introduces objects called \emph{ordered bipartitions}, which are easiest to think of as ordered set partitions with some of the subsets underlined. A \emph{compatible} bipartition is an ordered bipartition where all the underlined subsets are on the right, e.g., \[\alpha =\{5,2\}\{6\}\underline{\{1,4,7\}}\, \underline{\{3\}}, \] where we adopt the convention that underlined subsets have their elements written in ascending order. The bijection works as follows. For every non-underlined group in $\alpha$, we perform the same operation as above, while for every underlined group we perform the ``opposite" operation. Specifically, we put minus signs on all \emph{but} the least element before removing the braces. This produces \[ \pi = ( 5, -2, -6, 1, -4, -7, 3 ), \] a permutation where everyone on a linear table receives a napkin. All we really need to observe is that the underlined groups correspond to the part of the table where people all take napkins on their right, while the non-underlined groups all take napkins on the left. Thus, \begin{align} S(0,y,z) &= (L(0,y,z)+1)(R(0,y,z)+1)\nonumber \\ &= \frac{pqy^2}{\big(q(e^{pyz}-1)-py\big)\big(p(e^{qyz}-1)-qy\big)} \label{eq:S0} \end{align} We can take this more general correspondence and see now that the permutations for which everyone takes the napkin on the left correspond to the ordered bipartitions with no underlined subsets, the ordered bipartitions with all subsets underlined correspond to permutations where everyone takes the right napkin, and the compatible bipartitions with at least one underlined subset and one non-underlined subset correspond to the permutations where everyone gets a napkin and both end napkins are taken. This last observation gives \[ B(0,y,z) = L(0,y,z)R(0,y,z). \] \section{Ordered bipartitions and generating functions}\label{sec:ordgenfun} Using ordered bipartitions (not simply the compatible ones), we can encode more than just those permutations where everyone gets a napkin. Using the algorithm below, we can encode the set of \emph{all} signed permutations. Let $\varphi_S$ be this map, where the subscript $S$ reminds us that we are dealing with the straight table. The image of the injection $\varphi_S$ will lead us to the main theorem of this section, which gives some wonderful relationships between the generating functions for the linear table. Given a signed permutation $\pi$ of $[n]$, we form its image $\varphi_S(\pi)$ as follows: \begin{itemize} \item[(1)] Find the least element, $\pi(i)$ (ignoring signs), that is not already included in some subset. \item[(2a)] If $\pi(i)$ is positive, then underline the set including $\pi(i),$ and set $j = i+1$.\\ While $|\pi(j)| > |\pi(j-1)|$ and $\pi(j)$ negative,\\ $\mbox{\quad}$ add $\pi(j)$ to the set containing $\pi(i)$, and set $j = j+1$. \item[(2b)] If $\pi(i)$ is negative, then set $j = i-1$.\\ While $|\pi(j)| > |\pi(j+1)|$ and $\pi(j)$ positive,\\ $\mbox{\quad}$ add $\pi(j)$ to the set containing $\pi(i)$, and set $j = j-1$. \item[(3)] If every element is contained in a set, then delete all minus signs and quit. Else, go to (1). \end{itemize} Clearly, no two permutations can be mapped to the same bipartition. We will demonstrate the injection with an example. Start with \[\pi = (9, 1, -3, 2, 5, 6, -4, -7, 8).\] The number 1 is the least element, and it is positive, so its set will be underlined. We start searching to the right of 1, looking for negative numbers with absolute value bigger than 1. We get \[ 9 \;\; \underline{\{\,1, -3\,\}} \;\; 2 \;\; 5 \;\; 6 \;\; {-4} \;\; {-7} \;\; 8 \] Now 2 is the least element, and it is also positive, but there are no negative numbers immediately to its right, so it forms a singleton set, \[ 9 \;\; \underline{\{\,1, -3\,\}} \; \underline{\{\,2\,\}} \;\; 5 \;\; 6 \;\; {-4} \;\; {-7} \;\; 8 \] Now 4 is the least element not already in a set, and it is negative. So we start searching to the left of 4, looking for positive numbers with bigger absolute value. We get \[ 9 \;\; \underline{\{\,1, -3\,\}} \; \underline{\{\,2\,\}} \;\; 5 \;\; \{\,6 , -4\,\} \;\; {-7} \;\; 8 \] The next steps give, \[ 9 \;\; \underline{\{\,1, -3\,\}} \; \underline{\{\,2\,\}} \; \underline{\{\,5\,\}} \; \{\,6 , -4\,\} \;\; {-7} \;\; 8 \] \[ 9 \; \underline{\{\,1, -3\,\}} \; \underline{\{\,2\,\}} \; \underline{\{\,5\,\}} \; \{\,6 , -4\,\} \; \{\,-7\,\} \;\; 8 \] \[ 9 \;\; \underline{\{\,1, -3\,\}} \; \underline{\{\,2\,\}} \; \underline{\{\,5\,\}} \; \{\,6 , -4\,\} \; \{\,-7\,\} \; \underline{\{\,8\,\}} \] \[ \underline{\{\,9\,\}} \; \underline{\{\,1, -3\,\}} \; \underline{\{\,2\,\}} \; \underline{\{\,5\,\}} \; \{\,6 , -4\,\} \; \{\,-7\,\} \;\underline{\{\,8\,\}} \] Then we drop the minus signs to get, \[ \underline{\{\,9\,\}} \; \underline{\{\,1, 3\,\}} \; \underline{\{\,2\,\}} \; \underline{\{\,5\,\}} \; \{\,6 , 4\,\} \; \{\,7\,\} \; \underline{\{\,8\,\}} \] so that we have an ordered bipartition of $[n]$ where we write underlined sets in increasing order. We will now explain how the weight $w(\pi)$ of the signed permutation $\pi$ can be read from the corresponding ordered bipartition $\alpha = \varphi_S(\pi)$. To start with, the number of napkinless diners, $o(\pi)$, is exactly the number of occurrences of an underlined set followed immediately by a non-underlined set. The frustrated diners are those who are not the least element in a block, less the people without any napkin: $$m(\pi) = |\alpha| - \ell(\alpha) - o(\pi), $$ where $|\alpha|$ is the number of elements of the underlying set of $\alpha$ (same as $|\pi|$ here), and $\ell(\alpha)$ is the number of blocks in $\alpha$. Let $\un{\alpha}$ (resp. $\nu{\alpha}$) be the the bipartition formed from the blocks of $\alpha$ that are underlined (resp. non-underlined). For underlined blocks of $\alpha$ we have that the least element is positive in $\pi$ and the other elements are negative in $\pi$. Similarly, for non-underlined blocks of $\alpha$, the least element is negative in $\pi$ and the other elements are positive in $\pi$. Thus, \begin{align*} |\pi|_{-} &= |\un{\alpha}| - \ell(\un{\alpha}) + \ell(\nu{\alpha}); \\ |\pi|_{+} &= |\nu{\alpha}| - \ell(\nu{\alpha}) + \ell(\un{\alpha}). \\ \end{align*} Let $\varphi_S(\mathcal{C}_n)$ be the image set of bipartitions corresponding to all signed permutations of $[n]$. By examining the algorithm describing $\varphi_S$, we see that the set $\varphi_S(\mathcal{C}_n)$ consists of all ordered bipartitions that never contain the patterns $$ \underline{\cdots a\}} \, \{b\} \cdots \quad\text{or}\quad \cdots \underline{\{b\}} \, \{ a \cdots, \qquad\text{where $a < b$}. $$ To simplify notation in what follows, we write $C$ for $C(x,y,z)$, $S$ for $S(x,y,z)$, etc. Further, we write $H$ for $H(p;y,z)$ and $\underline{H}$ for $H(q;y,z)$. The following theorem takes advantage of our combinatorial model. Its power is reflected in the subsequent results it implies. \begin{thm}\label{thm:BL} We have the following formulas: \begin{align} L + B &= H S \\ B &= H S \underline{H} \end{align} \end{thm} \begin{proof} If we add a non-underlined block of size $r$ to the left of any bipartition in $\varphi_S(\mathcal{C}_n)$, then clearly we get a bipartition in $\varphi_S(\mathcal{C}_{n+r})$ that corresponds to a permutation in $\mathcal{L}_{n+r} \cup \mathcal{B}_{n+r}$. Furthermore, this new block will not change the number of people without a napkin (occurrences of underlined blocks immediately to the left of non-underlined blocks), and the number of new people who get a napkin they don't want is exactly $r-1$. Therefore, \[ L + B = H S. \] Similarly, if we add a non-underlined block of size $r$ to the left of a a bipartition in $\varphi_S(\mathcal{C}_n)$, and an underlined block of size $s$ to the right, then we get a bipartition in $\varphi_S(\mathcal{C}_{n+r+s})$ that corresponds to a permutation in $\mathcal{B}_{n+r+s}$. Thus, \[B = H S \underline{H}, \] which concludes the proof. \end{proof} \begin{cor}\label{thm:KBLNJS} We have the following formulas: \begin{align*} N &= S \big(1-H\big) \big(1-\underline{H}\big)\\ L &= S H \big(1-\underline{H}\big)\\ R &= S \big(1-H\big) \underline{H}\\ B &= S H \underline{H} \\ C &= z S \big( 1 + q(y-1)H + p(y-1)\underline{H} + (x-y)H \underline{H} \big) \end{align*} \end{cor} \begin{proof} From Theorem~\ref{thm:BL} we immediately get the formulas for $B$ and $L$. Plug these formulas into \eqref{eq:RL} and \eqref{eq:SNLRB} and the formulas for $R$ and $N$ follows. Finally, \eqref{eq:CNLRB} yields the formula for $C$. \end{proof} What Corollary \ref{thm:KBLNJS} tells us is that if we can find an explicit formula for $S$, then we will have explicit formulas for all the other generating functions. We will derive such an explicit formula shortly. First we need to introduce the notion of a \emph{cyclic} bipartition. A cyclic bipartition is a bipartition for which one element is distinguished, and only the cyclic ordering of the blocks matters. As a convention, we put the block containing the distinguished element at the far right if that block is not underlined, or at the far left if it is underlined. In our notation we will enclose the distinguished element in parentheses. For example, \[ \underline{\{\,1,6\,\}}\; \{\,8,2\,\} \; \underline{\{\,3,5\,\}} \; \{\,9,(7),4\,\} \] is a cyclic bipartition, which we also write \[7,4\,\}\;\underline{\{\,1,6\,\}}\; \{\,8,2\,\} \; \underline{\{\,3,5\,\}} \;\{\,9, \] so that the distinguished element is equivalently the first element in a bipartition where a block can ``wrap around." Similarly to how we encoded signed permutations playing out on a straight table with ordered bipartitions, we can encode the case of the circular table with cyclic bipartitions. Here the distinguished element will correspond to a ``distinguished guest," who sits in place number 1 at the circular banquet table. Let $\varphi$ be the map encoding the circular table case. For any signed permutation $\pi$ of $[n]$, we form its image $\varphi(\pi)$ as follows (the only difference in our algorithm is that the searches are cyclic): \begin{enumerate} \item[(1)] Find the least element, $\pi(i)$ (ignoring signs), that is not already included in some subset. \item[(2a)] If $\pi(i)$ is positive, then underline the set including $\pi(i),$ and set $j = i+1 \mod n$.\\ While $|\pi(j)| > |\pi(j-1)|$ and $\pi(j)$ negative,\\ $\mbox{\quad}$ add $\pi(j)$ to the set containing $\pi(i)$, and set $j = j+1 \mod n$. \item[(2b)] If $\pi(i)$ is negative, then set $j = i-1 \mod n$.\\ While $|\pi(j)| > |\pi(j+1)|$ and $\pi(j)$ positive,\\ $\mbox{\quad}$ add $\pi(j)$ to the set containing $\pi(i)$, and set $j = j-1 \mod n$. \item[(3)] If every element is contained in a set, then delete all minus signs and quit. Else, go to (1). \end{enumerate} As an example, start with the permutation \[ \pi = ( -7, 1, -3, 4, -2, 5, -6). \] The steps of the encoding are \[ -7 \; \underline{\{\,1,-3\,\}} \; 4 \; -2 \; 5\; -6 \] \[ -7 \; \underline{\{\,1,-3\,\}} \; \{\,4, -2\,\} \; 5\; -6 \] and this last step, which requires us to perform a cyclic search, \[ \underline{-7\,\}} \; \underline{\{\,1,-3\,\}} \; \{\,4, -2\,\} \; \underline{\{\,5, -6, } \] Now we can drop the signs to get \[ \underline{7\,\}} \; \underline{\{\,1,3\,\}} \; \{\,4, 2\,\} \; \underline{\{\,5, 6,} \] or \[ \underline{\{\,5, 6, (7)\,\}} \; \underline{\{\,1,3\,\}} \; \{\,4, 2\,\}. \] Let $\varphi(\mathcal{C}_n)$ be the set of all cyclic bipartitions corresponding to signed permutations on a circular table. We will use this set of cyclic bipartitions along with $\varphi_S(\mathcal{C}_n)$ to obtain the following theorem regarding the derivative of $S(x,y,z)$. \begin{thm}\label{thm:dSdz} We have the following formula relating the generating function for the straight table and the generating function for the circular table: \begin{equation}\label{eq:dSdz} z \frac{d}{dz} S(x,y,z) = C(x,y,z)S(x,y,z). \end{equation} \end{thm} \begin{proof} To prove equation \eqref{eq:dSdz}, it will suffice to equate coefficients and prove \begin{equation}\label{eq:[z^n]dSdz} nS_n(x,y) = \sum_{i=1}^n \binom{n}{i}C_k(x,y)S_{n-i}(x,y), \end{equation} in which the polynomials $$ S_n(x,y):=\sum_{\pi \in \mathcal{C}_n}w_S(\pi)\quad\text{and}\quad C_n(x,y):=\sum_{\pi \in \mathcal{C}_n}w(\pi) $$ are the coefficients of $z^n/n!$ in $S(x,y,z)$ and $C(x,y,z)$, respectively. We will prove \eqref{eq:[z^n]dSdz} bijectively. The left-hand side of equation \eqref{eq:[z^n]dSdz} can be thought of as counting the weights of permutations corresponding to bipartitions in $\varphi_S(\mathcal{C}_n)$ with a distinguished element, or a straight table with a distinguished guest. We simply put parentheses around one of the $n$ elements of the bipartition, as in \[ \underline{\{\,2,8\,\} }\;\{\, 1 \,\}\;\{\, 7, (4), 3 \,\}\; \underline{\{\,9\,\}}\;\underline{\{\,5,6\,\}}. \] Given any bipartition with a distinguished element, we can form a pair $(c,s)$, where $c$ is a cyclic bipartition of a subset $A \subset [n]$ (corresponding to one in $\varphi(\mathcal{C}_n)$), and $s$ is an ordered bipartition of $[n]\setminus A$ (corresponding to one in $\varphi_S(\mathcal{C}_n)$). We simply split the table into two pieces just before or after the block containing the distinguished element. If the block containing the distinguished element is not underlined, we make the split just after that block. The above example yields the pair \[ \big(\; \underline{\{\,2,8\,\}}\;\{\, 1 \,\}\;\{\, 7, (4), 3 \,\} , \, \underline{\{\,9\,\}}\;\underline{\{\,5,6\,\}} \;\big). \] If instead the block with the distinguished element is underlined, as in \[ \{\,2,8\,\} \;\{\, 1 \,\}\;\underline{\{\, 3,(4),7 \,\}}\;\underline{\{\,9\,\}} \;\underline{\{\,5,6\,\}}, \] we make the split before the underlined block. Now the right half becomes the circular table, and the left half is the straight table: \[ \big(\; \underline{\{\, 3,(4),7 \,\}} \; \underline{\{\,9\,\}}\;\underline{\{\,5,6\,\}}, \, \{\,2,8\,\} \;\{\, 1\, \}\;\big). \] By splitting the bipartition as we do, all of the guests retain their status as happy, frustrated, or napkinless. In other words, the product of the weights of the pair of tables equals the weight of original table. Now, given any pair $(c,s)$, where $c \in \varphi(\mathcal{C}_i)$ and $s \in \varphi_S(\mathcal{C}_{n-i})$, we can form a bipartition in $\varphi_S(\mathcal{C}_n)$ with a distinguished element. First, in any of $\binom{n}{i}$ ways, we choose a subset $A = \{a_1 < a_2 < \cdots < a_i\} \subset [n]$, and replace $k$ with $a_k$ in $c$. For $s$, we replace $k$ with $b_k$, where $[n]\setminus A = \{ b_1 < b_2 < \cdots < b_{n-i} \}$. Now we concatenate the bipartitions. If the distinguished element of $c$ is in an underlined block we put the straight table on the left: $sc$. If the distinguished element is not in an underlined block, we put the straight table on the right: $cs$. \end{proof} Now, thanks to Corollary \ref{thm:KBLNJS} and Theorem \ref{thm:dSdz}, we have: \[S' = \big((x-y)H\underline{H} +q(y-1)H + p(y-1)\underline{H} +1\big)S^2.\] We can solve this differential equation to get the following exact formula: \begin{equation}\label{eq:Sfunc} S(x,y,z) = \frac{ pqy^3}{ D }, \end{equation} where the denominator $D$ is: \begin{equation}\label{eq:D} \begin{aligned} pq(y-x)e^{yz} + (qx - pqy -q^2y^2)e^{pyz} & + (px-pqy -p^2y^2)e^{qyz} \\ & + pq\big(y(y-1)^2 + x(1-yz)\big) + y^2-x. \end{aligned} \end{equation} We can now obtain exact formulas for any of the other generating functions discussed here by plugging equation \eqref{eq:Sfunc} into the formulas of Corollary \ref{thm:KBLNJS}. In particular, we have our main result. \begin{thm}\label{thm:exact} At a table for $n$ people, the probability that $i$ people are napkinless and $j$ people are frustrated is given by the coefficient of $x^i y^j z^k$ in the following function: \begin{equation}\label{eq:Cfunc} C(x,y,z) = \frac{ pqyz\big( (x-y)e^{yz} + (qy^2 + py -x)e^{pyz} + (py^2 + qy -x)e^{qyz} + x \big) }{ D }, \end{equation} where $D$ is given in \eqref{eq:D}. \end{thm} \section{The expected number of napkinless guests and other statistics}\label{sec:expect} With the generating function from \eqref{eq:Cfunc}, we can in principle extract any sort of statistics related to the number of napkinless guests or frustrated guests. We will highlight a few statistical results that are of interest to us. In particular, we find the expected number of napkinless and frustrated guests, the variance for each of these distributions, and the covariance for their joint distribution.\footnote{This task is daunting by hand, but luckily we have technology to help us. Computer software such as Maple, for example, is very helpful, both with solving the differential equation leading to \eqref{eq:Sfunc}, and in obtaining residues.} \subsection{The expected number of napkinless guests} Recall the definition of the polynomial $C_n(x,y) = \sum_{\pi \in \mathcal{C}_n} w(\pi)$ from the proof of Theorem \ref{thm:dSdz}. Suppose we know exactly what $C_n(x,y)$ is for some $n$. If we want to obtain the expected number of people without a napkin, we want to compute the weighted average \[ E_n\big(o(\pi)\big) := \sum_{\pi \in \mathcal{C}_n} p^{|\pi|_-}q^{|\pi|_+} o(\pi). \] In terms of the polynomial $C_n (x,1)$ (we set $y=1$ since we're not interested in frustrated diners at the moment), this just means that we differentiate with respect to $x$ and set $x =1$, since \begin{align*} C'_n(x,1) &= \sum_{\pi \in \mathcal{C}_n} p^{|\pi|_-}q^{|\pi|_+} o(\pi) x^{o(\pi) -1} \\ C'_n(1,1) &= \sum_{\pi \in \mathcal{C}_n} p^{|\pi|_-}q^{|\pi|_+} o(\pi). \end{align*} Therefore, we want to find the generating function for the numbers $E_n\big(o(\pi)\big) = C'_n(1,1)$, or \begin{align} E(z) &:= \sum_{n\geq 0} E_n\big(o(\pi)\big) z^n \nonumber\\ &= \frac{d}{dx}\Big[ C(x,1,z) \Big]_{x=1} \nonumber\\ &= \frac{ z\big( pq(2-z)e^z + (p^2 + pqz -1)e^{pz} + (q^2 + pqz -1)e^{qz} +1 \big) }{pq(1-z)^2}. \label{expect_gf} \end{align} \begin{thm} The expected number of napkinless guests on a circular table for $n$ people is \begin{equation}\label{expect_formula} E_n\big(o(\pi)\big) = \frac{n}{pq}\big(1-p\exp_n(q)-q\exp_n(p)+pq\exp_n(1)\big), \end{equation} where $\exp_n(x)=\sum_{k=0}^n x^k/k!$ is the truncated exponential function. \end{thm} \begin{proof} Let $f(n) = n\big(1-p\exp_n(q)-q\exp_n(p)+pq\exp_n(1)\big)/(pq)$ be the right hand side of equation \eqref{expect_formula}, and let $F(z)$ be the ordinary generating function for the numbers $f(n)$. Note that $$nf(n+1) = (n+1)f(n) + \frac{1-p^n-q^n}{(n-1)!}, $$ which implies $$z\frac{d}{dz}\biggl[\frac{1}{z}F(z)\biggr] = \frac{d}{dz}\bigl[z F(z)\bigr] + z(e^z - pe^{pz} - qe^{qz}) $$ or equivalently, $$(1-z)F'(z) = \biggl(1+\frac{1}{z}\biggr)F(z) + z(e^z - pe^{pz} - qe^{qz}). $$ It is easy to check that $E(z)$, given by formula \eqref{expect_gf}, satisfies this differential equation. \end{proof} Formula \eqref{expect_formula} implies Sudbury's result (equation \eqref{eq:Sud}). \begin{cor} The expected value of $o(\pi)$ satisfies $$E_n\big(o(\pi)\big) = \frac{n(1 - pe^q)(1 - qe^p)}{pq} + O\Bigl(\frac{1}{n!}\Bigr). $$ In particular, when $p = q = 1/2$ we have $$E_n\big(o(\pi)\big) = n(2-\sqrt{e})^2 + O\Bigl(\frac{1}{n!}\Bigr). $$ \end{cor} So, the answer to Winkler's problem of napkins in a random setting is $(2-\sqrt{e})^2 \approx 0.12339675$. It was quite a bit of work for this answer, but of course our work pays off in being able to find the following statistics as well. \subsection{Other statistics} In the rest of this section we present asymptotic estimates for further statistics regarding the napkin problem. Since the formulas are less messy, we will primarily restrict our attention to the $p = q = 1/2$ case, but our approach is general. It is straightforward to obtain asymptotic estimates for functions with a finite number of poles such as $E(z)$. See, for example, chapter 5 of Herbert Wilf's book \cite{Wilf}. We use the same technique to obtain all the estimates given here. We briefly outline the approach. Suppose we have a power series $f(z) = \sum a_n z^n$, with a singularity at $z=1$ of multiplicity $m$ (and no other singularities). Then the Laurent expansion of $f$ around $1$ is: \[ f(z) = \frac{b_{-m}}{(1-z)^m} + \cdots + \frac{b_{-1}}{(1-z)} + b_0 + b_1(1-z) + b_2(1-z)^2 + \cdots. \] If we let \[ g(z) = \sum c_n z^n = \frac{b_{-m}}{(1-z)^m} + \cdots + \frac{b_{-1}}{(1-z)},\] (called the \emph{principal part} of $f$), then the function $h(z) = f(z) - g(z)$ is entire, and its coefficients vanish very quickly. Thus for large $n$, the coefficient of $z^n$ in $g(z)$ closely approximates the coefficient of $z^n$ in $f(z)$, i.e., \[ a_n \sim c_n = \binom{m+n-1}{m-1} b_{-m} + \cdots + \binom{n+1}{1}b_{-2} + b_{-1}. \] As an illustration, we apply this technique to $E(z)$ to, again, derive Sudbury's result. We first expand $E(z)$ as a series in $u= 1-z$ to get \begin{align*} E(u) & = \sum_{n \geq -2} b_n u^n \\ & = \frac{ (1-u)\big( pq(1+u)e^{1-u} + (p^2 + pq(1-u) -1)e^{p(1-u)} + (q^2 + pq(1-u) -1)e^{q(1-u)} +1 \big) }{pqu^2}. \end{align*} We let $\overline{E}(u):= u^2 E(u)$, then set $u = 0$ to obtain \[\overline{E}(0) = b_{-2} = \frac{pqe + (p^2+pq-1)e^p + (q^2+pq-1)e^q+1}{pq}.\] Next, we differentiate $\overline{E}$ before setting $u= 0$. This gives \[ \overline{E}'(0) = b_{-1} = - \frac{pq(e + e^p + e^q) + (1+p)(p^2 + pq -1)e^p + (1+q)(q^2 + pq -1)e^q - 1}{pq}.\] Now we can obtain our estimate: $$ a_n \sim (n + 1)b_{-2} + b_{-1} = \frac{n(1 - pe^q)(1 - qe^p)}{pq}. $$ Now, if we want to get the variance of the distribution of napkins on the table, we need to compute the sums of the squares of the number of napkins left on the table, since \[ \var_n( o(\pi) ) = E_n\big( o(\pi)^2 \big) - E_n\big( o(\pi) \big)^2,\] and we already have $E_n\big( o(\pi) \big)$. We get \begin{align*} E_n\big( o(\pi)^2 \big) &= \sum_{\pi \in \mathcal{C}_n} p^{|\pi|_-}q^{|\pi|_+} o(\pi)^2 \\ &= C_n''(1,1) + C_n'(1,1). \end{align*} Therefore the generating function for the second moment is \[ \sum_{n\geq 0} E_n \big( o(\pi)^2 \big) z^n = \frac{d^2}{dx^2}\Big[ C(x,1,z) \Big]_{x=1} + E(z), \] and all we need to find is $\displaystyle\frac{d^2}{dx^2}\Big[ C(x,1,z) \Big]_{x=1}$. Computing, we find: \begin{align*} C_{x^2}(1/2; 1,1,z) &= \frac{2z(2 - e^{z/2}) \Big( e^{z/2}(1-z)^2 + (2-e^{z/2})\big( (e^{z/2}-1)^2(3-z) -e^z + 2z \big) \Big)} {(1-z)^3}. \end{align*} Using the same method, we find that the variance is asymptotically \[ \var_n( o(\pi) ) \sim \frac{ n(1-pe^q)(1-qe^p)\big( 1 - (p^2 -pq)e^q - (q^2 - pq)e^p - pq(e + 1) \big)}{p^2q^2},\] or $n(3-e)(2-\sqrt{e})^2 \approx n(.0347631)$ for the $p=q = 1/2$ case. To get the expectation and variance for the number of frustrated guests, we follow the same procedure, except now we differentiate $C(1,y,z)$ with respect to $y$. The covariance is \[E\big(o(\pi)m(\pi)\big) - E\big(o(\pi)\big)E\big(m(\pi)\big),\] so we differentiate with respect to $x$, then with respect to $y$ to get the generating function for $E\big(o(\pi)m(\pi)\big)$, the only piece we don't know. The statistics are summarized in Table \ref{table}. Notice in particular that if the expected number of napkinless guests is $n(2-\sqrt{e})^2$, and the expected number of frustrated guests is $n(6\sqrt{e} - e - 7)$, then the expected number of happy guests is $n(4-2\sqrt{e}) \approx n(.702557)$. Seventy percent of the guests are happy! \begin{table}[t] \vspace{.5cm} \begin{tabular}{|c|c|c|} \hline & & \\ $X$: & $o(\pi)$ (Napkinless) & $m(\pi)$ (Frustrated) \\ \hline & & \\ $E(X)$ & $n(2-\sqrt{e})^2\approx n(.12339675)$ & $n(6\sqrt{e} - e - 7)\approx n(.174046)$\\ & & \\ \hline & & \\ $\var(X)$ & $n(3-e)(2-\sqrt{e})^2\approx n(.0347631)$ & $n( 6 \sqrt{e^3} - e^2 - e - 38\sqrt{e} +46)$\\ & & $\approx n(.13138819)$ \\ \hline & \multicolumn{2}{c|}{} \\ $\covar(X,Y)$ & \multicolumn{2}{c|}{$n\big( -(2 - \sqrt{e})( \sqrt{e^3} - 3e - 5\sqrt{e} + 12 )\big) \approx n( -.029239461)$ } \\ & \multicolumn{2}{c|}{} \\ \hline \end{tabular} \vspace{.5cm} \caption{Statistics for napkinless and frustrated guests with $p = q = 1/2$.\label{table}} \end{table} \section{Another proof for the expected number of napkinless guests}\label{sec:alternate} Upon reviewing a draft of this paper, Ira Gessel pointed out that the generating function for the expected number of napkinless guests satisfies the following differential equation (with $p=q=1/2$): \begin{equation}\label{eq:EdN} E(z) = z\frac{d}{dz}\left[\frac{(2-e^{z/2})^2}{1-z} \right] = z\frac{d}{dz}\Big[ N(1,1,z)\Big], \end{equation} where $N(1,1,z)$ is the generating function for the proportion of signed permutations for which neither napkin on a straight table is taken. We will think of such permutations in terms of their image under $\varphi_S$: specifically, those ordered bipartitions in for which the leftmost block is underlined and the rightmost block is not underlined. Gessel suggested that there may be a simple combinatorial explanation for \eqref{eq:EdN}, and there is. First, if all we want is the expected number of napkinless guests on the circular table, then because of the symmetry of the table, we have $E_n\big(o(\pi)\big) = np'$, where $p'$ is the probability that any particular guest (say guest 1) has no napkin. Upon equating coefficients, equation \eqref{eq:EdN} claims that $|\mathcal{N}_n| = |\mathcal{C}_n^1|$, where $\mathcal{C}_n^1$ is the set of all signed permutations for which guest 1 gets no napkin on a circular table. We can give a bijection between the ordered bipartitions in $\varphi_S(\mathcal{N}_n)$ and the cyclic bipartitions in $\varphi(\mathcal{C}_n^1)$ as follows. Given an ordered bipartition for which the leftmost block is underlined and the rightmost block is not underlined, there must be a guest in the middle of the table who is napkinless. Make the leftmost such guest distinguished, and cyclically permute the blocks of the bipartition until the block with the distinguished guest is first or last, depending on whether that block is underlined. For example, \[ \underline{\{\,8\,\}} \; \{\,(9),6,1\,\} \; \{\,2\,\} \; \underline{\{\,3,5\,\}} \; \{\, 7,4\,\} \] is an ordered bipartition in $\varphi_S(\mathcal{N}_n)$ with the leftmost napkinless guest highlighted (guest 2 in this case). We cyclically permute the blocks to get \[ \{\,2\,\} \; \underline{\{\,3,5\,\}}\; \{\, 7,4\,\} \; \underline{\{\,8\,\}} \; \{\,(9),6,1\,\}, \] which is a cyclic bipartition where guest 1 gets no napkin on a circular table. The inverse of this bijection is given by simply cyclically permuting the blocks until we have an ordered bipartition for which the distinguished guest is the leftmost guest without a napkin, the leftmost block is underlined, and the rightmost block is not. It is straightforward to check that we can always achieve this state. \section{Acknowledgements} We would like to extend a sincere thank you to Don Knuth for bringing this problem to our attention. His help and encouragement were invaluable. In particular, the ideas of reducing the problem to a straight table and keeping track of frustrated guests are his. Peter Winkler deserves credit for writing his wonderful book and for reading an early draft of this paper. We thank the Institut Mittag-Leffler and the organizers of the spring 2005 session in algebraic combinatorics. If not for our time there, this work may never have been done. Lastly, we thank Ira Gessel, whose observation motivated Section \ref{sec:alternate}.

https://arxiv.org/abs/2208.07459

Nesterov smoothing for sampling without smoothness

We study the problem of sampling from a target distribution in $\mathbb{R}^d$ whose potential is not smooth. Compared with the sampling problem with smooth potentials, this problem is much less well-understood due to the lack of smoothness. In this paper, we propose a novel sampling algorithm for a class of non-smooth potentials by first approximating them by smooth potentials using a technique that is akin to Nesterov smoothing. We then utilize sampling algorithms on the smooth potentials to generate approximate samples from the original non-smooth potentials. We select an appropriate smoothing intensity to ensure that the distance between the smoothed and un-smoothed distributions is minimal, thereby guaranteeing the algorithm's accuracy. Hence we obtain non-asymptotic convergence results based on existing analysis of smooth sampling. We verify our convergence result on a synthetic example and apply our method to improve the worst-case performance of Bayesian inference on a real-world example.

\section{Introduction} \looseness=-1 Sampling from a target distribution $\pi (x ) \propto \exp(-s(x)) $ known up to a normalization constant is an important problem in many areas such as machine learning, due to its pivotal role in Bayesian statistics and inference~\citep{gelman1995bayesian,durmus2018efficient,krauth2006statistical, fan2021variational}. Over the past decades, there has been a vast amount of research carried out on sampling problems from smooth distributions~\citep{dalalyan2017theoretical,vempala2019rapid}, i.e., the gradient of the potential $s(x)$ is Lipschitz. In contrast, the understanding of non-smooth sampling is still quite limited ~\citep{chatterji2020langevin,salim2020primal,liang2021proximal,liang2022proximal,chewi2021analysis}. In this work, we consider the task of sampling from a specific class of non-smooth distributions $\pi(x) \propto \exp(-s(x))$ whose potentials $s(x)$ permit an explicit max-structure as \begin{align}\label{eq:potential} s(x) = f(x) + \max_{y \in {\mathcal Y}} \{ \<h(x), y\> -g(y) \} \end{align} where $f$ is smooth, $g$ is convex, ${\mathcal Y}$ is convex and bounded, and $h$ is Lipschitz and smooth. \fan{$s(x)$ can be non-smooth.} For instance, when $f(x)=0 $ and $h(x) =Ax$, $s(x)$ becomes a piece-wise affine function and is clearly non-smooth. \chen{A special instance of particular interest is \begin{equation}\label{eq:maxh} s(x) = f(x)+ \max_{i} \{h_1(x), h_2(x),\ldots, h_n(x)\}. \end{equation} The optimization of minimizing $s(x)$ has been widely used for science and engineering problems to optimize the worst-case performance \citep{ben2009robust,diehl2006approximation}. Its Bayesian counterpart that accounts for uncertainties can be captured by a sampling problem from the potential $s(x)$.} Compared to other recent work ~\citep{chatterji2020langevin,salim2020primal,liang2021proximal,liang2022proximal,chewi2021analysis} for sampling from black box non-smooth distribution, we assume certain structures in the target distribution and hope to leverage these structures to achieve better performance. Indeed, in practice, it is rare to have a black box model as a target distribution; we almost always know some structures of the problem. It could be advantageous to utilize such structures in the algorithm design. With this in mind, we take one step forward in the direction of sampling beyond black box models and propose a sampling method for non-smooth distribution that utilizes the specific max-structure \eqref{eq:potential}. We develop a sampling algorithm for non-smooth potentials whose complexity has the same dimensional dependency as its smooth counterparts. Our strategy is akin to Nesterov smoothing~\citep{nesterov2005smooth} in optimization that constructs a smooth function $s_\beta(x)$ whose difference with $s(x)$ can be controlled through the smoothing intensity $\beta$. In Nesterov smoothing, the idea is to use a fast smooth optimization algorithm to minimize $s_\beta(x)$ for the purpose of optimizing the non-smooth function $s(x)$. In our problem, instead, we sample from $\pi_\beta (x) \propto \exp(-s_\beta(x))$ for the purpose of sampling from $\pi(x) \propto \exp(-s(x))$. Thanks to the smoothness of $s_\beta(x)$, we can take advantage of the many existing sampling algorithms for smooth potentials to sample from $\pi_\beta (x) \propto \exp(-s_\beta(x))$. Our smoothing strategy is compatible with any such algorithms. With state-of-the-art sampling algorithms for smooth potentials and a proper smoothing intensity $\beta$, we obtain a series algorithms with competitive complexity bounds. Surprisingly, the complexity bounds we obtained with Nesterov smoothing have exactly the same dimension dependency as their smooth counterparts. This is crucial for high dimensional problems. In a high level, this work is along the direction of the recent line of research that tries to bridge optimization and sampling~\citep{wibisono2018sampling,barp2022geometric,liang2022proximal}. Our results prove that some smoothing techniques used in optimization are equally effective in sampling. \paragraph{Related Works:} Over the last few years, several algorithms for sampling from potentials that are not smooth have been developed. A majority of them is devoted to the analysis of original LMC in the non-smooth setting \citep{erdogdu2021convergence,nguyen2021unadjusted,chewi2021analysis}. These theorectical results are challenging and significantly different from those in the smooth setting. In \citet{bernton2018langevin,durmus2019analysis,salim2020primal}, the non-smooth sampling problem is studied and analyzed from an optimization point of view. Another algorithm that is effective for non-smooth sampling is the proximal sampler developed in the sequence of work \citep{lee2021structured,liang2021proximal,liang2022proximal,chen2022improved}. The methods that are most related to ours are \citet{chatterji2020langevin,durmus2018efficient,lau2022bregman}, in which the authors use Gaussian smoothing and Moreau envelope. \chen{ Gaussian smoothing can be applied to any \fan{convex} potential, however the evaluation of the smoothed potential is difficult. Normally one can only get unbiased estimation of it through Gaussian sampling, inducing additional variance in the algorithm. Moreau envelope can be applied to any \fan{weakly-convex} potential. To evaluate the smoothed potential or its gradient, one needs to solve an optimization problem (to get the proximal map), inducing additional complexity, \fan{especially when the potential is not convex. We include a more detailed comparison with these two smoothing techniques in the end of Section \ref{sec:error}.}} \paragraph{Contributions:} We summarize our contributions as follows. \\ i) ~~~Inspired by Nesterov smoothing in optimization, we develop a smoothing technique, different from existing ones like Gaussian smoothing and Moreau envolep, for sampling problems for a class of potentials that permit the max-structure \eqref{eq:potential}. \\ ii) ~~Our smoothing technique is compatible with any first-order sampling algorithm for smooth potentials and the performance is guaranteed via a proper smoothing intensity $\beta$. \\ iii) ~Combining this smoothing technique with start-of-the-art sampling algorithms for smooth potentials we obtain competitive complexity bounds for strongly-log-concave distributions, log-concave distributions, and distributions satisfying log-Sobolev inequality (see Table \ref{table}). In particular, we achieve ${d}^{1/3}$ dependence in sampling non-smooth strongly-log-concave distributions, better than all existing methods. Besides, our analysis covers certain composite potentials with non-smooth (Lipschitz continuous) components and the distribution satisfies LSI. The complexity bound in this setting has not been established before. \begin{remark} \fan{ In Table \ref{table}, our result contains one additional problem-dependent parameter $D$ defined in \eqref{eq:bound_l}. It can hide dimension dependency, like any other parameter $\lambda_h,\lambda_\rho$ or $ L_f $. In Abstract and Contribution, our claim in terms of dimension dependency assumes other parameters are dimension-free. In realistic examples, this is unlikely to be true, so one need to include the dimension dependency hidden in all parameters for a fair comparison. An existing caveat is if $D$ hides dimension dependency, then our algorithm is clearly not state-of-art. \chen{Interestingly, for the class of problems with potentials of the form \eqref{eq:maxh}, $D = \log (n)$ will be seen as in Section \ref{sec:ex}.} } \end{remark} \begin{table}[] \caption{ Convergence comparison for \fan{state-of-the-art} non-smooth sampling. We highlight the dependency on dimension $d$, accuracy ${\mathcal E}$, \fan{smoothness $L_f$ of $f(x)$,} Lipschitz $\lambda_h$ of $h(x)$, Lipschitz $\lambda_\rho$ of $\rho(x):= \max_{y \in {\mathcal Y}} \{ \<h(x), y\> -g(y) \}$, \fan{boundness $D$ defined in \eqref{eq:bound_l},} and log-Sobolev constant $C_\pi$. A trivial upper bound for $\lambda_\rho $ is $\lambda_h \max \|y\|_2 $. \fan{ Other than \citet{durmus2019analysis}, all results are the last-iterate convergence. } } \label{table} \centering {\renewcommand{\arraystretch}{2.1}% \begin{tabular}{|c|c|c|c|c|} \hline Assumption & Source & Complexity & Metric & Note \\ \hline \multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}$\alpha$-strongly \\ log-concave\end{tabular}} & \citep{liang2022proximal} & $\widetilde{\mathcal O} \left( \frac{(\lambda_\rho^{2} \vee L_f) d }{\alpha } \right)$ & R\'enyi \xspace & \\\cline{2-5} & \citep{salim2020primal} & $\widetilde{\mathcal O} \left( \frac{\lambda_\rho^{2} + L_f d }{\alpha^2 {\mathcal E}^2 } \right)$ & $W_2$ & \\\cline{2-5} & \citep{chatterji2020langevin} & $\widetilde{\mathcal O} \left( \frac{(\lambda_\rho+L_f)^{2} d^{5/2} }{\alpha^4 {\mathcal E}^4 } \right)$ & $W_2$ & Gaussian smoothing \\\cline{2-5} & Theorem \ref{thm:str_convex} (1) & $\widetilde{\mathcal O} \left( \frac{\lambda_h^{2} D d^{1/3} }{\alpha^{11/6} {\mathcal E}^{5/3}} + \frac{\lambda_h^{7/3} D^{7/6} d^{1/6} }{\alpha^{23/12} {\mathcal E}^{3/2}} \right)$ & $W_2$ & \\\cline{2-5} & Theorem \ref{thm:str_convex} (2) & $ \widetilde{\mathcal O} \left( \frac{ \lambda_h^2 D d^{1/2}}{\alpha {\mathcal E} } \right)$ & TV & Assume warm-start \\\hline \multirow{5}{*}{log-concave} & \citep{liang2022proximal} & ${\mathcal O} \left(\frac{( \lambda_\rho^2 \vee L_f ) d^2 }{ {\mathcal E}} \right)$ & R\'enyi \xspace & \\\cline{2-5} & \citep{lehec2021langevin} & ${\mathcal O} \left(\frac{\lambda_\rho^2 d^2 }{ {\mathcal E}^{4}} \right)$ & $W_2$ & Assume $M_\pi= {\mathcal O}({d})$ \\\cline{2-5} & \citep{durmus2019analysis} & ${\mathcal O} \left(\frac{ L_f d^2 + \lambda_\rho^2 d }{ {\mathcal E}^2} \right)$ & KL & \\\cline{2-5} & \citep{durmus2018efficient} & ${\mathcal O} \left(\frac{ \lambda_\rho^2 d^5 }{ {\mathcal E}^2} \right)$ & TV & Moreau envelope \\\cline{2-5} & Theorem \ref{thm:convex} & ${\mathcal O} \left(\frac{\lambda_h^2 D M_\pi d }{ {\mathcal E}^{3}} \right)$ & TV & \\\hline log-Sobolev & Theorem \ref{thm:lsi} & $\widetilde {\mathcal O} \left( \frac{\lambda_h^2 D C_\pi d }{{\mathcal E}} \right)$ & TV & \\ \hline \end{tabular} } \end{table} \section{Background}\label{sec:back} Before we proceed, we briefly introduce the notation basics. We assume the norm in the space of $x$ and $y$ can be different and denote them as $\|\cdot \|_{\mathcal X}$ and $\|\cdot \|_{\mathcal Y}$ respectively. We use $J_h(x): {\mathbb R}^d \rightarrow {\mathbb R}^{n \times d} $ to represent the Jacobian matrix of the function $h$ at $x$. The norm of a matrix $A: {\mathbb R}^d \rightarrow {\mathbb R}^{n \times d} $ is defined as $\|A\|_{{\mathcal X},{\mathcal Y}} = \max_{x,y} \{ \< A x, y\>_{\mathcal Y} : \|x\|_{\mathcal X} =1, \|y\|_{\mathcal Y} =1 \}.$ \fan{Given a primal space equipped with norm $\|\cdot\|$, we denote the norm in its dual space as $\|\cdot\|^*$.} In this article, we assume the probability measure have density with respect to the Lebesgue measure. \subsection{Nesterov smooothing} It is well-known that with subgradient methods, the optimal bound of non-smooth convex minimization complexity is ${\mathcal O}( 1/\epsilon^2)$ to achieve $\epsilon$ error in the function value~\citep{nemirovskij1983problem,nesterov2003introductory}. However, beyond the black-box oracle model, there are various structures that can be used to improve this bound. Nesterov smoothing~\citep{nesterov2005smooth} is an interesting technique for functions with explicit max structure \eqref{eq:potential}. In doing so, the complexity order can be improved to ${\mathcal O}(1/{\mathcal E})$. The underlying principle of Nesterov smoothing is that \[ \textit{The dual of a strongly-convex function is smooth and vice versa~\citep{kakade2009duality}.} \] Specifically, it adds a $\sigma$-strongly-convex penalty function $\ell(y)$ to the original maximization such that \begin{align}\label{eq:s_beta} s_\beta(x) = f(x) +\max_{y \in {\mathcal Y}} \{\<h(x), y\> -g(y) - {\beta} \ell(y) \} , \end{align} where $\beta$ is the smoothing intensity. In the original paper of Nesterov smoothing, they consider the specific task $h(x) =Ax$ and find $s_\beta$ is smooth with constant $L_f + \|A\|^2_{{\mathcal X},{\mathcal Y}}/\beta \sigma. $ Then they apply the Nesterov accelerated gradient descent algorithm to minimize $s_\beta$, an optimal scheme for smooth optimization. Our idea is very similar on the high level: use smoothing for the non-smooth potential first, and then apply the existing sampling algorithm for the smoothed distribution. \subsection{Smooth sampling}\label{sec:smooth} In this section, we summarize several state-of-art sampling algorithms for smooth distributions. In general, there are several dominant popular algorithms: Unadjusted Langevin Monte Carlo (LMC)~\citep{parisi1981correlation}, Underdamped/Kinetic Langevin Monte Carlo (KLMC)~\citep{cheng2018underdamped} Metropolis-Adjusted Langevin Algorithm (MALA)~\citep{roberts1996exponential}, and Hamiltonian Monte Carlo (HMC)~\citep{neal2011mcmc}. Suppose we are given the target distribution $\pi \propto \exp(-V(x))$ for some smooth potential $V(x)$. The LMC algorithm runs the update \begin{align} \label{eq:lmc} x_{t+1} = x_{t} - \gamma \nabla V(x) + \sqrt{2\gamma} {\epsilon}_t, \end{align} where $\epsilon_t$ is the standard Gaussian random variable in ${\mathbb R}^d$, $x_0$ is a user-defined starting point and $\gamma>0$ is the step size that is small enough. Indeed, \eqref{eq:lmc} is the Euler discretization of Langevin diffusion dynamics, \begin{align}\label{eq:langevin} {\text{d}} x_t = - \nabla V(x_t) {\text{d}} t + \sqrt{2} {\text{d}} B_t. \end{align} This dynamics has an invariant density $\pi$ no matter what $x_0$ is. The underdamped Langevin diffusion adds an Hamiltonian ingredient to the standard Langevin diffusion \eqref{eq:langevin}, so that an additional random variable $v_t := {\text{d}} x_t /{\text{d}} t$ that controls the velocity in the dynamics. Similarly, HMC maintains a velocity term, but with a different randomization manner. MALA is also a variant of LMC. As its name implies, it adds a Metropolis-Hastings~\citep{metropolis1953equation,hastings1970monte} rejection step to LMC, in order to eliminate the bias induced by discretization. We refer the readers to \citet{mackay2003information} for the an overview of these algorithms. In spite of the canonical role of LMC in sampling, it is not the fastest algorithm empirically. Theoretically, in the standard strongly-convex potential case, the complexity of LMC can achieve $\widetilde {\mathcal O}(d)$~\citep{dalalyan2017theoretical,vempala2019rapid,erdogdu2022convergence}, whereas KLMC, HMC algorithms can achieve $\widetilde {\mathcal O}(d^{1/2})$~\citep{dalalyan2020sampling,chen2022optimal}. \fan{The complexity $\widetilde {\mathcal O}(d^{1/2})$ also holds for MALA, however, under a warm start~\citep{wu2021minimax}. Regarding the accuracy dependency, MALA is notably the best with a benefit from Metropolis step.} With a considerable research accumulated for decades, several novel sampling algorithms or the variants are proposed recently, such as the proximal algorithm~\citep{lee2021structured,chen2022improved,liang2022proximal} based on restricted Gaussian oracle, and the randomized midpoint method for KLMC~\citep{shen2019randomized}. The proximal algorithm has been proved to have the state-of-art dimension dependency for log-concave distribution~\citep{chen2022improved}. There exist some accelerating algorithms that can achieve even lower dimension order in complexity~\citep{mangoubi2018dimensionally,sanz2021wasserstein}. However, they require stronger regularity assumptions, e.g. bounded third-order derivative. Although we do not consider these higher-order algorithms in this paper for the sake of generality, one can still apply them in our framework if the higher-order smoothness of $s_\beta(x)$ can be verified in practice. \section{Nesterov smoothing for sampling}\label{sec:algo} In this section, we will introduce our Nesterov smoothing sampling algorithm in Section \ref{sec:setup} and bound the error caused by smoothing in Section \ref{sec:error}. \subsection{Problem setup and algorithm}\label{sec:setup} Formally, we consider the task to sample from a non-smooth distribution $ \pi(x) \propto \exp(-s(x))$ where \begin{align} s(x) = f(x) +\max_{y \in {\mathcal Y}} \{\<h(x), y\> -g(y) \}. \end{align} We will assume $f(x):{\mathcal X} \rightarrow {\mathbb R}$ is $L_f$-smooth; $h(x): {\mathcal X} \rightarrow {\mathcal Y} $ is $\lambda_h$-Lipschitz and $L_h$-smooth; $g(y): {\mathcal Y} \rightarrow {\mathbb R}$ is continuous and convex through out the paper. A function $f(x) $ is $L$-smooth if \fan{$\|\nabla f(x_1) -\nabla f(x_2)\|^*_{\mathcal X} \leq L \|x_1-x_2\|_{\mathcal X} $} ~for~ $\forall x_1,x_2 \in {\mathbb R}^d.$ Correspondingly, we say a multivariate function $h(x) $ is $L$-smooth if $\|J_h(x_1) - J_h(x_2) \|_{{\mathcal X},{\mathcal Y}} \le L \|x_1 - x_2\|_{\mathcal X} $ ~for~ $\forall x_1,x_2 \in {\mathcal X}.$ We say a multivariate function $h(x)$ is $\lambda$-Lipschitz if \fan{$\|h(x_1) -h(x_2) \|^*_{\mathcal Y} \le \lambda \|x_1 -x_2\|_{\mathcal X}$, which is equivalent to $\| J_h(x) \|_{{\mathcal X},{\mathcal Y}} \le \lambda $ ~for~ $\forall x \in {\mathcal X}$ if $h(x)$ is differentiable. } This boils down to that the \fan{operator norm} of $A$ is bounded when $h(x) =Ax$. And this allows us to compare the smoothing effect with the case considered in \citet{nesterov2005smooth}. \fan{Without more specification, in the rest of paper, we assume $\|\cdot\|^*_{\mathcal X} =\|\cdot\|_{\mathcal X} =\|\cdot\|_2 $ is Euclidean norm to be consistent with the sampling literature.} It is very likely that $s(x)$ is not differentiable since it is a "piece-wise" function segmented by different maximizer $y$. As such, it is not feasible to directly apply the smooth sampling algorithms. Inspired by the Nesterov smoothing technique, our strategy is to \[ \textit{1) smooth $s(x)$ as $s_\beta(x)$, 2) apply smooth sampling algorithm to $\pi_\beta(x)$.} \] \begin{algorithm}[tb] \caption{ Nesterov smoothing sampling framework} \label{algo} \begin{algorithmic} \STATE{{\bfseries Input:} A first-order based smooth sampling algorithm, the target distribution $\pi \propto \exp(-s(x))$, number of iterations $K$, initial point $x_0 \sim \mu_0$ } \STATE{{\bfseries Initialization:} Choose the smoothing intensity $\beta$ } \FOR{ $k=0, \ldots, K-1$} \STATE{\it{\# Calculate $\nabla s_\beta(x) $ according to \eqref{eq:grad_s_beta}} } \STATE{Iterate $x_{k+1} $ by the smooth sampling algorithm } \ENDFOR \STATE {{\bfseries Output:} $x_K $ } \end{algorithmic} \end{algorithm} For smoothing, we consider the surrogate function $$s_\beta(x) = f(x) +\max_{y \in {\mathcal Y}} \{\<h(x), y\> -g(y) - {\beta} \ell(y) \} , $$ where $\ell(y)$ is a $\sigma$-strongly-convex prox-function $$\ell(y) \ge \frac{\sigma}{2} \| y-y_0\|_{\mathcal Y}^2 .$$ The center of ${\mathcal Y}$ is defined as $y_0 = \argmin_y \{ \ell (y): y \in {\mathcal Y}\} $ and we can assume that $\ell(y_0) =0.$ The compactness of ${\mathcal Y}$ also allows us to define \begin{align}\label{eq:bound_l} D = \max_y\{\ell(y): y \in {\mathcal Y} \} \end{align} and $R = \max_y\{\|y\|_{\mathcal Y} : y \in {\mathcal Y} \} $. From the definition, $D$ is the squared diameter of the set ${\mathcal Y}$, and the $R$ is the furthest distance of any point in the set to the origin. By the triangular inequality, it holds that $R \le \|y_0\| + \sqrt{\frac{ 2D}{\sigma}}$. In this way, $s_\beta(x)$ is a smooth function, which is justified by Lemma \ref{lem:phi_smooth}. The proof is postponed to Appendix. \begin{lemma}\label{lem:phi_smooth} Let $\phi_\beta(x,y): = \<h(x), y\> -g(y) - {\beta} \ell(y) $, and $y_\beta(x) := \argmax_{y \in {\mathcal Y}} \phi_\beta(x,y) $, then \begin{align}\label{eq:grad_s_beta} \nabla s_\beta(x) = \nabla f(x) +J_h^\top (x) y_\beta(x). \end{align} Moreoever, $s_\beta(x)$ is $L_{s_\beta}$-smooth with \begin{align}\label{eq:L_phi} L_{s_\beta}= L_f + R L_h + \frac{\lambda_h^2 }{\beta \sigma}. \end{align} \end{lemma} If $h(x) =Ax$, then $L_h=0$ and $\lambda_h = \|A\|_{{\mathcal X},{\mathcal Y}}$, we get $L_{s_\beta} = L_f + \frac{1 }{\beta \sigma} \|A\|^2_{{\mathcal X},{\mathcal Y}} $ in this special case. This recovers the smoothness constant in \citet[Theorem 1]{nesterov2005smooth}. In many cases, $\lambda_h$ is dimensional-free. For example, when $h(x) =Ax $ and $A = c I_m$, then $\lambda_h=c$ regardless of the dimensionality. We will also assume that we can solve $ \max_y \phi_\beta(x,y)$ for any $x \in {\mathcal X}$ in a fast dimension-free manner, which has been verified by the complexity analysis~\citep[Section 3]{bubeck2015convex} and especially several examples in \citet[Section 4]{nesterov2005smooth}. Thus we can query the gradient of $s_\beta(x)$ according to \eqref{eq:grad_s_beta} and resort to first-order sampling algorithm to sample from $\pi_\beta \propto \exp(- s_\beta (x))$. The pseudo-code of our scheme is presented in Algorithm \ref{algo}. \subsection{Bounded distance between $\pi$ and $\pi_\beta$} \label{sec:error} When $\beta \rightarrow 0,$ it holds that $\pi_\beta \rightarrow \pi $. In this section, we rigorously bound the distance between $\pi$ and $\pi_\beta$. A key observation is that $s_\beta(x)$ is a uniform smooth approximation of $s(x)$, i.e., $ s_\beta(x) \le s(x) \le s_\beta(x)+ \beta D$ (see \eqref{eq:s_diff})~\citep[Eq. 2.7]{nesterov2005smooth}. To bound the difference between $\pi$ and $\pi_\beta$, we select two widely-used distributional distance as the error criteria. The first one is the total variation ${\textup{TV}} (p,q) = \frac{1}{2} \| p-q\|_1 $, which can be bounded by Kullback–Leibler divergence ${\textup{KL}} (p \|q)$ through Pinsker's inequality \begin{align}\label{eq:pinsker} {\textup{TV}} (p,q) \le \sqrt{\frac{1}{2} {\textup{KL}}(p \| q)}. \end{align} The second one is the Wasserstein-2 distance~\citep{villani2021topics} $$ W_2 (p,q) := \sqrt{\inf_{\nu \in \Pi (p,q)} \int_{{\mathbb R}^d \times {\mathbb R}^d } \|a-b\|_2^2 d\nu(a,b) }, $$ where $\Pi (p,q) $ denotes the set of the all joint distributions of $p$ and $q$. The $W_2$ distance can also be bounded by ${\textup{KL}}(p \| q)$ through the Talagrand inequality~\citep{otto2000generalization} \begin{align} \label{eq:tala} \tag{Talagrand} W_2(p,q) \le \sqrt{2 C_{\textup{LSI}} {\textup{KL}}( p\|q)}. \end{align} if $q$ satisfies the following logarithmic-Sobolev inequality (LSI) with constant $C_{\textup{LSI}}$. \begin{definition}[Log-Sobolev inequality] \label{def:lsi} A probability distribution $p$ satisfies a logarithmic-Sobolev inequality with constant $C_{\textup{LSI}}$ if, for all smooth functions $u: {\mathbb R}^d \rightarrow {\mathbb R}$, it holds that \begin{align}\tag{LSI} \label{eq:lsi} {\mathbb E}_p[u^2 \log{u^2} ] -{\mathbb E}_p[u^2] \log {\mathbb E}_p[u^2] \le {2}{C_{\textup{LSI}}} {\mathbb E}_p [ \|\nabla u\|^2]. \end{align} \end{definition} LSI is a powerful tool for the sampling convergence analysis when there is no log-concavity. It has nice properties, for instance, it implies the sub-Gaussian concentration property~\citep{ledoux1999concentration}. Moreoever, if $s(x)$ is $\alpha$-strongly-convex, then $\pi(x)$ satisfies \eqref{eq:lsi} with constant $1/\alpha$~\citep{bakry1985diffusions}. With the help of the above inequalities, we are able to show that Proposition \ref{prop:tv}, \ref{prop:w} hold. \begin{proposition}\label{prop:tv} ${\textup{TV}}(\pi , \pi_\beta) \le \frac{\beta D}{2}$. \end{proposition} \begin{proposition}\label{prop:w}If $\pi$ satisfies \eqref{eq:lsi} with constant $C_\pi$, then $W_2(\pi , \pi_\beta) \le \sqrt{C_\pi} {\beta D} $. \end{proposition} We next compare the smoothing effect of Nesterov smoothing, Moreau envelope and Gaussian smoothing. Recall that we define the non-smooth part in our potential as $\rho(x)$, and Lipschitz constant of $\rho(x)$ to be $\lambda_\rho.$ All of three smoothing methods can deal with composite smooth$+$non-smooth potential, but \emph{they require the convexity or weakly-convexity of $\rho(x)$, while Nesterov smoothing does not require so.} \paragraph{Moreau envelope} The Moreau-Yosida envelope of a l.s.c. convex function $\rho(x)$ is defined as \begin{align*} \rho_\beta(x) := \min_{y \in {\mathbb R}^d} \left\{ \rho(y) + \frac{1}{2\beta} \| x-y\|_2^2 \right\}. \end{align*} $\rho_\beta$ is $L_\rho$-smooth with $ L_\rho = \frac{1}{\beta}. $ The smoothed composite potential is then written as $s_\beta(x) = f(x) + \rho_\beta(x).$ The smoothing effect of Moreau envelope applies to not only l.s.c. convex functions, but also l.s.c. weakly-convex functions with an upper bound constraint on $\beta$~\citep{bohm2021variable,hoheiselproximal}. According to Proposition 1 in \citet{durmus2018efficient}, there is $ {\textup{TV}} (\pi, \pi_\beta) \le \frac{1}{2} \beta \lambda_\rho^2, $ which is implied by $0 \le s(x) -s_\beta(x) \le {\beta \|\lambda_\rho\|^2}/{2} $. \paragraph{Gaussian smoothing} Unlike the proposed method in \citet{chatterji2020langevin}, where they apply Gaussian smoothing to both the smooth part $f(x)$ and the non-smooth part $\rho(x)$ in $s(x)$, our discussion is based on only smoothing the $\rho_\beta(x)$ here. We find smoothing both parts would introduce a worse dimension dependency in bounding the difference between $\pi(x)$ and $\pi_\beta(x)$. The Gaussian smoothing for $\rho(x)$ is defined as $\rho_\beta (x) = {\mathbb E}_\xi[ \rho ( x + \beta \xi )] ,$ where $\xi \sim \mathcal{N}(0, I_{d\times d}). $ The smoothing effect of Gaussian smoothing only applies to convex functions. Indeed, according to Lemma 2.2 in~\citet{chatterji2020langevin}, if $\rho(x)$ is convex, then $ \rho_\beta (x)$ is $L_{ \rho}$-smooth with $ L_{\rho} = {\lambda_\rho d^{1/2}}/{\beta}.$ To bound the difference between $\pi $ and $\pi_\beta $, one doesn't need a stronger assumption. By Lemma E.2 in \citet{chatterji2020langevin}, $ {\textup{TV}} (\pi, \pi_\beta) \le \frac{1}{2} \beta \lambda_\rho d^{1/2} . $ Additionally, if $f(x)$ is $\alpha$-strongly convex, by Lemma 3.3 in \citet{chatterji2020langevin}, $ W_2 (\pi, \pi_\beta) = \widetilde{\mathcal{O}}\left( \frac{ \beta \lambda_\rho d}{\alpha} \right). $ Now, if we want to achieve ${\textup{TV}}(\pi,\pi_\beta) = {\mathcal E}/2$, then Moreau envelope results in $L_{\rho} = \frac{\lambda_\rho^2}{{\mathcal E}};$ Gaussian smoothing results in $L_{\rho} = \frac{\lambda_\rho^2 d}{{\mathcal E}};$ Nesterov smoothing results in $L_{\rho} = R L_h + \frac{\lambda_h^2 D }{\mathcal{E} \sigma}.$ Given the same smooth sampling algorithm, the dimension dependency in $L_\rho$ determines the dimension dependency in the final convergence rate. Normally, $\sigma=1$. Assuming $\lambda_\rho = \lambda_h , D =\widetilde{\mathcal{O}}(1) $, then Nesterov smoothing effect is the same order as Moreau envelope, and better than Gaussian smoothing. If $D$ hides dimension dependency, then Nesterov smoothing is worse than Moreau envelope. However, Moreau envelope requires more strict assumption on $\rho(x)$, i.e. weakly-convexity; and since the potential is not smooth, solving the proximal operator of Moreau envelope requires additional complexity in each sampling step, even for structured potential in \eqref{eq:potential}. \section{Non-asymptotic convergence analysis}\label{sec:complexity} In this section, we present our convergence analysis for several fundamental scenarios: strongly log-concavity, log-concavity, and log-Sobolev inequality. Denote the underlying distribution of $x_{k}$ as $\mu_{k}$. We will give the sufficient number of gradient queries for $\mu_k$ to reach a given error tolerance. As discussed in Section \ref{sec:smooth}, there are multiple algorithm candidates to use for the smooth sampling in Algorithm \ref{algo}. Our strategy is to test every algorithm among them and select the one with the best complexity. In the main theorems of this section, each complexity is accompanied by a state-of-art smoothing algorithm, which we will specify concretely. \subsection{Strong log-concavity}\label{sec:strong_concave} The strong log-concavity condition is mostly-widely studied for sampling convergence since it can be viewed as the strongest assumption. To our best knowledge, under this condition and without higher-order smoothness,\footnote{Here, higher-order smoothness means the condition that is higher than bounded Hessian.} the randomized midpoint KLMC has the best dimension dependency $\widetilde {\mathcal O}(d^{1/3})$ among the first-order sampling algorithms. Whereas MALA doesn't have the best dependency on dimension, it has best accuracy dependency. So we will adopt these two algorithms as our smooth sampling algorithm. We also need to point out that, to have such good complexity, MALA needs a warm-start initialization. The initial distribution $\mu_0$ is called $G$-warm if the $\sup_{S \in {\mathcal B}({\mathbb R}^d) } \mu_0(S) / \pi(S) \le G $, where ${\mathcal B}$ means the Borel sigma algebra. If $G$ does not scale with dimension or condition number, then $\mu_0$ is called a warm start. This warm-start may not hold without a delicate initialization. For example, if Gaussian initialization is used, $G={\mathcal O}(\exp(d))$. \begin{theorem}\label{thm:str_convex} Assume $f$ is $\alpha$-strongly-convex, and $ \< h(\cdot), y\>$ is convex for any $ y \in {\mathcal Y}$.\\ 1) In Algorithm \ref{algo}, if we choose the smooth sampling algorithm to be randomized midpoint KLMC~\citep{shen2019randomized} with initial point $x_0 = \argmin s(x)$, $\beta = \frac{\sqrt{\alpha} {\mathcal E}}{2D} $, then for any $0<{\mathcal E}< 2\sqrt{\frac{d}{\alpha}}$, the iterate $\mu_K$ satisfies $ W_2(\mu_K , \pi) \le {\mathcal E}$ for $K = \widetilde{\mathcal O} \left( \left( \frac{L_{s_\beta}}{\alpha} \right)^{7/6} ( \frac{2}{{\mathcal E}} \sqrt{\frac{d}{\alpha} } )^{1/3} + \left( \frac{L_{s_\beta}}{\alpha} \right) ( \frac{2}{{\mathcal E}} \sqrt{\frac{d}{\alpha} } )^{2/3} \right) $. \fan{If ${\mathcal E}$ is sufficiently small such that $\frac{L_f + R L_h}{\lambda_h^2 D}= {\mathcal O}(1/{\mathcal E})$, } we obtain $$\fan{K= \widetilde{\mathcal O} \left( \frac{\lambda_h^{2} D d^{1/3} }{\alpha^{11/6} {\mathcal E}^{5/3}} + \frac{\lambda_h^{7/3} D^{7/6} d^{1/6} }{\alpha^{23/12} {\mathcal E}^{3/2}} \right) .} $$ Note that $L_f$ and $ L_h$ show up in the coefficient of lower order of $1/ {\mathcal E}$. \\ 2) In Algorithm \ref{algo}, if we choose the smooth sampling algorithm to be MALA~\citep{wu2021minimax} with the initial $\mu_0$ being a warm start, $\beta = \frac{{\mathcal E}}{D} $, then for any $0<{\mathcal E}< 2$, the iterate $\mu_K$ satisfies ${\textup{TV}}(\mu_K , \pi) \le {\mathcal E}$ for some $K = \widetilde{\mathcal O} \left( \frac{L_{s_\beta}}{\alpha} \sqrt{d} \right) $. If ${\mathcal E}$ is sufficiently small such that $\frac{L_f + R L_h}{\lambda_h^2 D}= {\mathcal O}(1/{\mathcal E})$, we get $$K = \widetilde{\mathcal O} \left( \frac{ \lambda_h^2 D \sqrt{d}}{\alpha {\mathcal E} } \right) . $$ \end{theorem} We need the assumption $ \< h(\cdot), y\>$ is convex for any $ y \in {\mathcal Y}$ since it not only promises $s(x)$ is convex, but also $s_\beta(x)$. Thus we can utilize the convergence result for strong log-concave smooth sampling. \fan{However, this assumption does not necessarily restrict $h(x)$ to be linear.} Here, we give several examples where this assumption hold: (i) $h(x) =Ax$, (ii) $h_i(x)$ is convex for $\forall i$, and ${\mathcal Y} \subset {\mathbb R}_{+}^n$, (iii) $h_i(x)$ is concave for $\forall i$, and ${\mathcal Y} \subset {\mathbb R}_{-}^n$, where ${\mathbb R}^n_+$ means the non-negative part of ${\mathbb R}^n$ and ${\mathbb R}^n_-$ means the non-positive part. \subsection{Log-concavity} Denote the second moment of $\pi$ is $M_\pi$. Under the log-concavity condition, the smooth sampling algorithm with best dimension dependency~\citep[Corollary 6]{chen2022improved} is the proximal algorithm. Similar to Section \ref{sec:strong_concave}, MALA can achieve the better accuracy order~\citep{chen2020fast} and worse dimension order compared to the proximal algorithm. For the sake of conciseness, we do not list MALA result in this section. \begin{theorem}\label{thm:convex} Assume $f$ is convex, and $\< h(\cdot), y\>$ is convex for any $ y \in {\mathcal Y}$. In Algorithm \ref{algo}, if we choose the smooth sampling algorithm to be proximal algorithm~\citep{lee2021structured} with $\beta = \frac{{\mathcal E}}{D} $, then for any ${\mathcal E}> 0 $, the iterate $\mu_K$ satisfies $ {\textup{TV}}(\mu_K , \pi) \le {\mathcal E} $ within $K = {\mathcal O} \left( \frac{L_{s_\beta} d W_2^2(\mu_0, \pi_\beta ) }{{\mathcal E}^2} \right) $ iterations in expectation. For simplification, assume that $W_2^2(\mu_0, \pi_\beta ) = {\mathcal O}(M_\pi)$, and ${\mathcal E}$ is sufficiently small such that $\frac{L_f + R L_h}{\lambda_h^2 D}= {\mathcal O}(1/{\mathcal E})$, then we have \begin{align}\label{eq:concave_K} K= {\mathcal O} \left(\frac{\lambda_h^2 D M_\pi d }{ {\mathcal E}^{3}} \right) . \end{align} \end{theorem} One may notice that an additional term $M_\pi$ shows up in Theorem \ref{thm:convex}. This is because in Theorem \ref{thm:str_convex}, this term is absorbed into logarithm. Generally, $M_\pi$ would scale with dimension $d$. For example $M_\pi = {\mathcal O}(d)$ for a product measure. One needs to take this additional dimension dependency into account when applying Theorem \ref{thm:convex}. The randomness of $K$ in \eqref{eq:concave_K} comes from the rejection sampling step in proximal algorithm~\citet{lee2021structured}. The bound is in expectation but one can also derive a high probability bound by Multiplicative Chernoff bound since each iteration in rejection sampling is independent. \subsection{Log-Sobolev} A notable condition when there is no log-concavity is the LSI condition. For example, the double-well potential $s(x) = ax^4 -bx^2$ would satisfy the LSI condition but it is not convex. For the LSI condition, the convergence analysis still concentrate on LMC~\citep{vempala2019rapid,chewi2021analysis}. The recent study on proximal algorithm also has the convergence guarantee~\citep{chen2022improved}. After the comparison on their convergence, no one is strictly better than another. So we list the complexity for both algorithms here. In this part, we will use the second order R\'enyi \xspace divergence, defined as ${\mathcal R}_2 (p \| q) : = \log \left\|\frac{{\text{d}} p}{{\text{d}} q} \right\|^2_{L^2(q)} $, to control the initialization error. It is a strong divergence since it can upper bound both ${\textup{KL}}$ divergence, and squared total variation. It has been a popular analysis tool in sampling~\citep{chewi2021analysis}. \begin{lemma}\label{lem:q_lsi} If $\pi(x)$ satisfies \eqref{eq:lsi} with constant $C_\pi$, then $\pi_\beta$ satisfies \eqref{eq:lsi} with constant $C_{\beta} =C_\pi \exp \left( 4 \beta D \right) $. \end{lemma} \begin{remark} \fan{Although $C_\beta$ has exponential dependence on $D$, we will choose $\beta = \frac{{\mathcal E}}{D}$ in the algorithm such that $\exp (4 \beta D) = \exp(4{\mathcal E}) \approx 1 $. In this manner, $C_\beta$ has the same order of $C_\pi$.} \end{remark} With Lemma \ref{lem:q_lsi}, we are able to apply the convergence of smooth sampling algorithms in our scheme and reach Theorem \ref{thm:lsi}. \begin{theorem}\label{thm:lsi} Assume $\pi(x)$ satisfies \eqref{eq:lsi} with constant $C_\pi$. In Algorithm \ref{algo}, if we choose the smooth sampling algorithm to be proximal algorithm~\citep{lee2021structured}, $\beta = \frac{ {\mathcal E}}{D} $, then for any ${\mathcal E} > 0 $, the iterate $\mu_K$ satisfies ${\textup{TV}}(\mu_K , \pi) \le {\mathcal E} $ within $K = {\mathcal O} \left( L_{s_\beta} C_\beta d \log(2/{\mathcal E}^2) \right) $ iterations in expectation. Assume \fan{${\mathcal E}$ is sufficiently small such that $\frac{L_f + R L_h}{\lambda_h^2 D}= {\mathcal O}(1/{\mathcal E})$, } we obtain \begin{align}\label{eq:lsi_prox} K= \widetilde {\mathcal O} \left( \frac{\lambda_h^2D C_\pi d }{ {\mathcal E}} \right). \end{align} \end{theorem} \subsection{Example}\label{sec:ex} In this section, we present case study to illustrate our algorithm. We consider the case where $s(x) $ is the summation of a quadratic term and the maximum of absolute values (also see Section 4.4 of \citep{nesterov2005smooth}) \begin{align}\label{eq:ex} s(x) = \|x\|_{\mathcal X}^2 + \max_{1 \le j \le m } \{ |\< a_j , x\> -b^{(j)} | \}, \quad a_j \in {\mathbb R}^d, ~b^{(j)} \in {\mathbb R}, \end{align} where we choose $\|\cdot\|_{\mathcal X}$ as Euclidean norm $\|\cdot\|_2$. It is not difficult to find that $\lambda_\rho = \max_{1\le j \le m} \|a_j\|_2$ since it is a piece-wise affine function. To write \eqref{eq:ex} in the form of \eqref{eq:potential}, we denote the matrix $A $ with rows $a_j, j=1, \ldots, m$, and $b = (b_1, \ldots, b_m)$. We also choose $n=2m,$ ${\mathcal Y} = \Delta_{n}$, where $\Delta_{n}$ is the probability simplex on ${\mathbb R}^n$. Then with $\widehat A = \begin{pmatrix} A \\ -A \end{pmatrix} $ and $\widehat b = \begin{pmatrix} b \\ -b \end{pmatrix}, $ \begin{align} s(x) = \|x\|_2^2 + \max_{y \in \Delta_{n} } \{ \< \widehat A x, y\> - \< \widehat b, y\> \}. \end{align} \chen{This potential is associated with a Bayesian inference problem with a Gaussian prior and likelihood consists of (absolute values of) linear functions. The goal is to infer a posterior distribution that would optimize the worst-case performance.} To construct the smoothed $s_\beta(x)$, we choose $\| \cdot\|_{\mathcal Y} $ as 1-norm $ \|\cdot\|_1 $, ~$y_0$ is the uniform distribution, and $\ell(y) = \log (n) + \sum_{j=1}^n y^{(j)} \log y^{(j)} $. Thus by Lemma 3 in \citet{nesterov2005smooth}, we result in $\sigma=1$ and $D = \log(n)$. Clearly, $s(x)$ is strongly-convex, and we have strongly-convex constant $\alpha=2 $, smoothness $L_f= 2, L_h=0$, and Lipschitz $$\lambda_h = \|\widehat A \|_{{\mathcal X},{\mathcal Y}} = \max_{x} \{ \max_{1\le j \le m} |\< a_j, x\>| : \|x\|_2 =1 \} = \max_{1\le j \le m} \|a_j\|_2 =\lambda_\rho $$ Putting these together in Theorem \ref{thm:str_convex} bound, with \fan{randomized midpoint KLMC,} it is enough to have $\widetilde {\mathcal O} \left( \frac{ \lambda_h^{7/3} d^{1/3} }{ {\mathcal E}^{5/3}} \right)$, however, the best bound up to now for strongly-convex non-smooth sampling is $\widetilde{\mathcal O} \left( { \lambda_h^2 d } \right)$ \citep{liang2022proximal}. \fan{If we assume $\lambda_h$ is dimensional free,} we have better convergence result in terms of dimension. \section{Conclusion} We consider sampling from a non-smooth distribution $\pi(x) \propto \exp(-s(x))$ that permits the explicit max structure~\eqref{eq:potential}. Inspired by Nesterov smoothing, we firstly transfer the non-smooth potential $s(x)$ to a smooth potential $s_\beta(x)$ by adding a strongly convex penalty in the max structure, then we apply the smooth sampling algorithms directly to $\pi_\beta(x) \propto \exp(- s_\beta(x)) $. Our smoothing technique is compatible with any first-order sampling algorithm for smooth potentials. We also show that the error introduced by smoothing can be controlled by a careful choice of smoothing intensity $\beta$. Finally, we systematically study the convergence rate under several common conditions, and compare them with the state-of-the-art results in Table \ref{table}. Our convergence rates are better than or on par with existing methods \fan{if no parameter is hiding dimension dependency.} \paragraph{Limitation and future work} In this paper, we only consider one class of non-smooth potential with max structure~\eqref{eq:potential} for smoothing. It is interesting to explore more structures that can benefit from smoothing techniques. In addition, from Table \ref{table}, one can see that our accuracy dependency does not achieve state-of-art, since we need to compensate the error caused by smoothing. \fan{If some parameter like $D$ is hiding dimension dependency, then our method may not be state-of-art anymore.} Finally, in the Nesterov smoothing work~\citep{nesterov2005smooth}, the optimal estimate for smooth convex optimization is proved. However, the lower bound of smooth sampling is still unclear and seems more challenging. Only a specific class of sampling algorithms has been shown to have lower bound complexity~\citep{wu2021minimax}. It would also be pivotal to derive the lower bound of smooth sampling. \newpage

https://arxiv.org/abs/0811.1531

Homotopy groups and twisted homology of arrangements

Recent work of M. Yoshinaga shows that in some instances certain higher homotopy groups of arrangements map onto non-resonant homology. This is in contrast to the usual Hurewicz map to untwisted homology, which is always the zero homomorphism in degree greater than one. In this work we examine this dichotomy, generalizing both results.

\section{Introduction} Let $\mathcal{A}$ be an arrangement of hyperplanes in $\mathbf{C} ^{\ell }$. Thus $\mathcal{A}$ is a finite non-empty collection $ \left\{ H_{1},\ldots ,H_{n}\right\} $ where $H_{i}=\alpha _{i}^{-1}(b_i)$ with $b_i \in \mathbf{C}$ and each $\alpha _{i}$ is a linear homogeneous form in the variables $ (z_{1},\ldots ,z_{\ell })$. (See \cite{OrlikTerao} for material on arrangements). \ We call $\mathcal{A}$ an $\ell $\emph{-arrangement}. We let $M$ be the complement of the union $H$ of the hyperplanes \begin{equation*} M=\mathbf{C}^{\ell }\setminus \cup H_{i}. \end{equation*} Now $M$ is the complement of a real codimension two subset of $\mathbf{R} ^{2\ell }$ and so has numerous topological properties of interest. Here we focus on the homotopy theory of the complement. Since the fundamental group of the complement $\pi=\pi _{1}(M)$is fairly rich in structure, it makes sense to look at covers of the complement. Equivalently one may look at homology or homotopy with coefficients in $R=\mathbb{Z}[\pi]$ modules or at local system homology. Two results motivate this study. The first is found in \cite{RandellHomotopy}; the untwisted Hurewicz homomorphism $h:\pi _{k}(M)\rightarrow~H_{k}(M;\mathbb{Z})$ is trivial for $ k>1$. \ On the other hand, in \cite{Yoshinaga} Yoshinaga has shown that the twisted Hurewicz homomorphism maps \emph{onto} certain twisted homology for non-resonant local systems, in the case that $M$ is a generic section of another hyperplane arrangement. In this work we generalize both these results while giving a unified treatment. It is known that an arrangement complement has a minimal CW structure (\cite{DimcaPapadima}, \cite{RandellMinimal}), and this fact is used in Yoshinaga's proof. Part of our goal in this work is to understand exactly where this minimality is useful in studying the topology of arrangement complements. Thus we assume minimality only when necessary, and in particular reprove Yoshinaga's result without use of this property. We begin in the next section with a discussion of the twisted homology and Hurewicz homomorphisms and general relationships between higher homotopy groups and twisted homology groups. We then specialize to the case of hyperplane arrangement complements, and derive consequences in various situations. \section{Local Systems and the Twisted Hurewicz homomorphism} Let $(X,Y)$ be an $(n-1)-$connected topological pair with $n\geq 3$. Thus $\pi _{j}(X,Y)\cong 0$ for $j\leq n-1$, and by the long exact homotopy sequence of the pair, $\pi _{j}(X)\cong \pi _{j}(Y)$ for $j\leq n-2$. In particular, the fundamental group $\pi =\pi _{1}(Y)$ of $Y$ includes isomorphically into the fundamental group of $X$. Then there is a generalized (twisted) Hurewicz isomorphism, which we now explain. We follow the discussion of Rong \cite{Rong}, with notation as in Hatcher \cite{Hatcher}. Let $N$ be a left $\mathbf{Z}[\pi ]$-module, characterized by the left action \begin{equation*} \varphi : \pi_{1}(X) \times N \rightarrow N \end{equation*} or equivalently by the homomorphism \begin{equation*} \varphi : \pi_{1}(X) \rightarrow Aut(N). \end{equation*} Then as usual there is an associated local system $N_{\varphi}$ on $X$. (See Hatcher \cite{Hatcher} for details about local systems). Then one has homology and cohomology with coefficients in $N_{\varphi }$, and by the usual left action of $\pi $ on $\pi _{n}(X,Y)$ we may form the tensor product \begin{equation*} \pi _{n}(X,Y)\otimes _{\pi }N \end{equation*} where we use the subscript``$\pi "$ to indicate the tensor product over $\mathbf{Z}[\pi ]$. Note that since we are working over a generally noncommutative ring, we need to make the action on $\pi_n(X,Y)$ into a \emph{right} action, as in \cite{Hatcher}. In what follows we will switch as appropriate among the various viewpoints and notations for local systems, but we will always have the above set-up in mind. We will generally suppress $\varphi$ from the notation. \begin{theorem} (Twisted Hurewicz Theorem) Let $(X,Y)$ be an $(n-1)-$connected topological pair with $n\geq 3$. Then there is a natural isomorphism \begin{equation} h:\pi_n(X,Y) \otimes _{\pi } N\rightarrow H_{n}(X,Y;N) \end{equation} \end{theorem} Notice that the right hand side involves local system homology, while the left hand side is an algebraic tensor product. We sketch the proof of this theorem below to highlight the property of naturality. Here ``Naturality" means that with these assumptions there exist homomorphisms $h:\pi_n(X)\otimes _{\pi } N \rightarrow H_{n}(X;N) $ and $h:\pi_n(Y) \otimes _{\pi } N\rightarrow H_{n}(Y;N)$ so that the following diagram commutes. \begin{equation} \begin{array}{ccccccc} \rightarrow & H_{n}(X;N) & \rightarrow & H_{n}(X,Y;N) & \stackrel{i_*}{\rightarrow} & H_{n-1}(Y;N) & \rightarrow \\ & \uparrow h & & \uparrow h\cong & & \uparrow h & \\ \rightarrow & \pi_n(X) \otimes _{\pi } N & \rightarrow & \pi_n(X,Y) \otimes _{\pi } N & \rightarrow & \pi_{n-1}(Y) \otimes _{\pi } N & \rightarrow \end{array} \end{equation} \begin{equation*} \begin{array}{cccc} \rightarrow H_{n-1}(X;N) & \rightarrow & H_{n-1}(X,Y;N) & \rightarrow \\ \uparrow h & & \uparrow h & \\ \rightarrow \pi _{n-1}(X) \otimes _{\pi } N & \rightarrow &\pi _{n-1}(X,Y) \otimes _{\pi }N & \rightarrow \end{array} \end{equation*} \pagebreak \begin{proof}(Sketch \cite{Rong}) Let $(\tilde{X},\tilde{Y})$ denote the universal covers. There is a sequence of isomorphisms $H_n((X,Y);N)$ $\cong H_0(\pi,H_n(\tilde{X},\tilde{Y};N))$, by the Eilenberg-Moore spectral sequence $\cong [H_n(\tilde{X},\tilde{Y};N)]_\pi$ $\cong H_n(\tilde{X},\tilde{Y})\otimes_{\pi}N$, by the universal coefficient theorem $\cong \pi_n(X,Y)\otimes_{\pi}N$, by the usual Hurewicz theorem. Naturality may be traced through these isomorphisms. \end{proof} Note that since tensor product is not exact, we have no guarantee that the lower row of this diagram is exact. It will be exact, of course, if the module $N$ is flat. Of course the upper row is exact always. Then we have the following general result: \begin{theorem} Let $(X,Y)$ be an $(n-1)-$connected topological pair with $n\geq 3$. Then $\ker (i_{\ast }:H_{n-1}(Y;N)\rightarrow H_{n-1}(X;N))\subset im(h_{Y}:N\otimes _{\pi }\pi _{n-1}(Y)\rightarrow H_{n-1}(Y;N))$. \end{theorem} \begin{proof} Since $\pi _{n-1}(X,Y)\cong 0$, and tensor product is right exact, the sequence \begin{displaymath} \begin{array}{ccccccc} \pi _{n}(X,Y) \otimes_{\pi } N & \rightarrow & \pi_{n-1}(Y)\otimes _{\pi }N & \rightarrow & \pi _{n-1}(X)\otimes _{\pi }N & \rightarrow & 0 \end{array} \end{displaymath} is exact, so that the commuting diagram above yields a commuting diagram with exact rows \begin{equation} \begin{array}{cccccccc} H_{n}(X,Y;N) & \rightarrow & H_{n-1}(Y;N) & \stackrel{i_*}{\rightarrow} & H_{n-1}(X;N) & \rightarrow & 0 & \rightarrow \\ \uparrow h\cong & & \uparrow h_{Y} & & \uparrow h & & & \\ \pi _{n}(X,Y) \otimes _{\pi }N & \rightarrow &\pi _{n-1}(Y) \otimes _{\pi } N & \rightarrow & \pi_{n-1}(X) \otimes_{\pi } N & \rightarrow & 0 & \rightarrow \end{array} \label{eq:BasicDiagram} \end{equation} A simple diagram chase then gives the result. \end{proof} Next we examine the above result in various special cases. \section{Consequences of the Twisted Hurewicz Theorem} The following result follows immediately from the exact sequence (\ref{eq:BasicDiagram}). \begin{proposition}Let $(X,Y)$ be an $(n-1)-$connected topological pair with $n\geq 3$. If $h_{Y}$ is the zero homomorphism, then $i_{\ast }:H_{n-1}(Y;N)\rightarrow H_{n-1}(X;N)$ is injective and hence an isomorphism. \end{proposition} \begin{definition} An \emph{arrangement pair} $(M,M'')$ is an $(n-1)$-connected pair of topological spaces with $M$ an arrangement complement and $M''$ a generic hyperplane section of $M$. \end{definition} The notation is usual; $M''$ is the ``restriction" of $M$ to a hyperplane. A hyperplane section is generic provided that the intersection lattice of $M''$ agrees with that of $M$ through rank $n-1$. It is well-known that a generic section $M''$ of $M$ yields an arrangement pair with the homotopy type of a CW pair, and that in fact the number of $n$-cells attached to $M''$ to yield $M$ is exactly equal to the $n$-th betti number of $M$. (See (\cite{DimcaPapadima}, \cite {RandellMinimal}). \begin{proposition} If $N$ is the trivial system, $N=\mathbb{Z}$, and $(M,M'')$ is an arrangement pair, then $i_{\ast }$ is an isomorphism, as is $j_{\ast }:H_{n}(M;N)\rightarrow H_{n}(M,M'';N)$. \end{proposition} \begin{proof}For the trivial system, it was shown in \cite{RandellHomotopy} that the Hurewicz map is always trivial on higher homotopy groups. Thus $h_{M''}=0$. It then follows from (\ref{eq:BasicDiagram}) that $i_{\ast}$ and $j_{\ast}$ are isomorphisms. \end{proof} In the case of hyperplane arrangments, the injectivity of $i_{\ast }$ follows from the Orlik-Solomon algebra (or, more basically from the Lefschetz theorem on hyperplane sections). More interesting is the next result, as mentioned due to Yoshinaga in the case of a generic pair of hyperplane complements and a non-resonant local system $N$, where $N$ as an abelian group is just $\mathbf{C}^r$ (a local system of rank $r$). Our method of proof uses nothing about the \emph{minimal} CW structure of arrangement complements, however. \begin{corollary} Suppose $(X,Y)$ is an $(n-1)-$connected pair of topological spaces, $n\geq 3$ . Suppose that $N$ is a local system on $X$ so that $ H_{n-1}(X;N)=0$. \ Then $h_{Y}:\pi_{n-1}(Y) \otimes _{\pi } N \rightarrow H_{n-1}(Y;N)$ is onto. \end{corollary} Stated for arrangement pairs we have \begin{theorem}(Yoshinaga \cite{Yoshinaga}) Suppose $(M,M'')$ is an arrangement pair and $N$ is a non-resonant local system of rank $r$ on $M$. Then $h_{M''}:\pi _{n-1}(M'')\otimes _{\pi } N \rightarrow H_{n-1}(M'';N)$ is onto. \end{theorem} \begin{proof} For a non-resonant system $ H_{n-1}(M;N) = 0$. \end{proof} Actually the Hurewicz map here differs slightly from the one considered by Yoshinaga, but the result above clearly implies that of Yoshinaga. Here is the relationship. Let $X$ be a topological space with basepoint $x$. Then, as in \cite{Yoshinaga} we have a twisted Hurewicz map \[h_j:\pi_j(X,x) \otimes_{\mathbf{Z}}\mathcal{L}_x \rightarrow H_j(X,\mathcal{L})\] defined by setting $h(f \otimes t)$ equal to the twisted cycle it determines. This Hurewicz homomorphism differs by a change of ring (from $\mathbf{Z}[\pi]$ to $\mathbf{Z}[1] \cong \mathbf{Z}$) from the one we considered earlier. The homomorphisms are related by the obvious commuting triangle. \section{The Image of Hurewicz Maps} \subsection{Basic Topological Properties of Arrangement Complements} An arrangement complement $M$ has several basic topological properties. Before listing them we need a bit of notation \cite{OrlikTerao}. To any hyperplane arrangement one has the lattice whose elements are the various intersections of the hyperplanes, ordered by reverse inclusion. To any lattice element $Z$ one has the ``localization" $\mathcal{A}_Z=\{H \in \mathcal{A} \mid Z \subset H\}$. The rank $r=rk(Z)$ of any lattice element $Z$ is its codimension. Note that it is always the case that there is the inclusion $M(\mathcal{A}) \subset M(\mathcal{A}_Z)$. Here are several basic properties of arrangement complements. \begin{description} \item[Locality] The ``Brieskorn homomorphism" $i:\oplus H^r(M(\mathcal{A}_Z)) \rightarrow H^r(M(\mathcal{A}))$ is an isomorphism, the direct sum taken over all rank $r$ lattice elements. \item[Toroidality] The cohomology of $M$ is generated by cohomology in degree one, in particular by the logarithmic one forms $d\alpha_i / \alpha_i$, where $\alpha_i$ is a linear form defining the hyperplane $H_i$. \item[Minimality] $M$ has the homotopy type of a minimal CW-complex, where ``minimal" here means that the number of $q$ cells is equal to the $q$-th betti number, for all $q$. \end{description} In the context of local systems locality is not usually relevant, since the Brieskorn homomorphisms are usually not defined: to do so one needs a local system on $M(\mathcal{A})$ which is the restriction of one on $M(\mathcal{A}_Z)$, and so is trivial around hyperplanes not containing $Z$. Deletion-restriction for this situation (when $rk(Z)=1$) is examined by D. Cohen in \cite{CohenDR}. Minimality in the local systems setting has been examined in \cite{DimcaPapadima2}. In particular they show there that the $\pi$-equivariant chain complex associated to a Morse-theoretic minimal CW structure on an arrangement complement is independent of the CW structure. In this section we will examine the role of toroidality for local systems. We will prove the analog of the result mentioned earlier, that the Hurewicz map is trivial for the trivial local system, finding a part of the kernel for any local system. Later we will speculate about injectivity--can the Hurewicz map in local systems detect homotopy groups? Now for a non-resonant (or ``generic") local system it is known that the first homology and cohomology groups vanish (\cite{CDO}). Thus for generic local systems, toroidality fails badly. Toroidality is a key ingredient in the proof that the usual Hurewicz homomorphism is zero. In place of that result we have the result below, which first requires some preparation. \subsection{The Hurewicz Image for General Arrangement Covers} Now since $M$ is an arrangement complement, it is path-connected, locally path-connected and semi-locally simply connected. Therefore $M$ has a universal cover and there is a bijective correspondence between subgroups $\pi'$ of $\pi = \pi_1(M)$ and connected covering spaces of $M$. Let $M'$ be the cover corresponding to $\pi'$. Then the free abelian group $\mathbf{Z}[\pi/\pi']$ with basis the cosets $\gamma \pi'$ is a $\mathbf{Z} [\pi]$-module and the homology groups of $M'$ with coefficients in $\mathbf{Z}$ are the same as the homology groups of $M$ with coefficients in the local system $\mathbf{Z}[\pi/\pi']$. The same holds in cohomology and there are isomorphisms \[ H_j(M,\mathbf{Z}[\pi/\pi']) \cong H_j(M',\mathbf{Z}) \] and \[ H^j(M,\mathbf{Z}[\pi/\pi']) \cong H^j(M',\mathbf{Z}) \] Let us now consider the case where we have a local system on $M$ of the form $N=\mathbf{Z}[\pi/\pi'])$ . \begin{proposition}$Im(h': \pi_j(M') \rightarrow H_j(M')) \subseteq ker (p_*: H_j(M') \rightarrow H_j(M))$. \end{proposition} Note that $H_j(M') = H_j(M,\mathbf{Z}[\pi/\pi'])$. In case $M=M'$ this is the result of \cite{RandellHomotopy}. \begin{proof} Fix $j \geq 2$. Let $p:\tilde{M} \rightarrow M$ be the universal cover of $M$. Then there is a commuting diagram \begin{equation} \begin{array}{ccc} \pi_j(\tilde{M}) & \stackrel{\tilde{h}}{\rightarrow} & H_{j}(\tilde{M}) \\ p_* \downarrow \cong & \ & \downarrow p_* \\ \pi_j(M') & \stackrel{h'}{\rightarrow} & H_j(M') \\ p_* \downarrow \cong & \ & \downarrow p_* \\ \pi_j(M) & \stackrel{h}{\rightarrow} & H_j(M) \\ \end{array} \end{equation} Now by the result of \cite{RandellHomotopy} the bottom horizontal arrow is the zero homomorphism, and the result follows \end{proof} Note that one can tensor the left-hand higher homotopy groups with $\mathbf{Z}[\pi])$-modules, maintaining the commutativity and the isomorphisms in the left-hand column, to get analogous results. We further note that finite covers of $M$, such as the Milnor fiber, fit into this framework. For instance, in the case of the Milnor fiber one may see easily that one does not have isomorphism in the above proposition. Examples below show that the inclusion in the above Proposition may be strict. \subsection{Examples} \begin{example} Consider the reflection arrangement $A_3$, as an arrangement in projective space. In that case $H_1(M)$ is free abelian of rank five, while the six-fold cyclic cover $M'$ (which is the Milnor fiber $F$ of the associated central arrangement) has $H_1(F)$ free of rank seven. One may use standard calculations and euler characteristic arguments to conclude that the second homology groups have ranks six and eighteen respectively. Also, $M$ and $M'=F$ are aspherical, so that $Im(h':~\pi_j(M') \rightarrow H_j(M'))$ is zero, while it may be seen easily that $p_*: H_2(M')~\rightarrow H_2(M)$ has non-trivial kernel (in fact, $p_*$ is surjective, so that the kernel is free abelian of rank twelve.) \end{example} The next example shows that the Hurewicz map on the universal cover may be an isomorphism, and that this is reflected in twisted homology. \begin{example} Let $\mathcal{A}$ be any arrangement, with complement $M$, and let $\tilde{M}$ be its universal cover. Then on the first non-trivial homotopy group, the Hurewicz map is an isomorphism. For example, one may take $\mathcal{A}$ to be the arrangement of the three hyperplanes $x=0$, $y=0$, $x+y-1=0$ in $\mathbf{C^2}$. Now by a well-known result of A. Hattori \cite{Hattori}, the complement $M$ has the homotopy type of the two-skeleton of the three torus $T^3$. That is, $M$ has the homotopy type of $T^3$ with a single $3$-cell removed. It is then easily seen that $F$ has the homotopy type of $T^3$ with four $3$-cells removed. Then $\pi_2(M)$ is the free $Z[\pi]$-module of rank one (where $\pi = \pi_1(M) $ is the free abelian group of rank three.) Then the Hurewicz homomorphism \begin{displaymath} \begin{array}{ccccc} h:\pi_2(M) \otimes_\pi Z[\pi] & \rightarrow & H_2(\tilde{M}) & \cong & H_2(M,\mathbf{Z}[\pi]) \end{array} \end{displaymath} is an isomorphism. \end{example} In fact, one recalls the following well-known fact (Hurewicz) on \emph{detectability} of higher homotopy groups. We will say (an element of) a homotopy group is detectable if there is some twisted Hurewicz map $h$ which does not send it to zero. \begin{theorem}The first non-trivial higher homotopy group of an space $X$ possessing a universal cover $\tilde{X}$ is detected by local system homology. \end{theorem} \begin{proof}The groups $\pi_k(X)$ and $\pi_k(\tilde{X})$ are isomorphic, and the Hurewicz theorem gives an isomorphism between $\pi_k(\tilde{X})$, and $H_k(\tilde{X})$. The latter is isomorphic to $H_k(X;\mathbf{Z}[\pi])$. \end{proof} The next example points out that the twisted Hurewicz map may have the maximal image, subject to the above results. \begin{example}\label{generic example} Take $\mathcal{A}$ to be $x=0$, $y=0$, $x+y-1=0$ in $\mathbf{C^2}$ giving $M$ as in the previous example. Consider the Milnor fiber $F$ associated to the cone of this arrangement, $\{x=0, y=0, x+y-z=0, z=0\}$. This fiber {F} is the four-fold cyclic cover of $M$ associated to the homomorphism sending each meridianal loop of $M$ to the generator of $\mathbf{Z}/4\mathbf{Z}$. Since $F$ is a cover of $M$, $\pi_2(F)$ is also the free $Z[\pi]$ module of rank $1$, where $\pi$ is free abelian of rank $3$, and of course the homomorphism induced by the covering map is an isomorphism. Now regarding $\pi_2(F)$ as a module over $\mathbf{Z}[\pi_1(F)]$ one can show that the image of $\pi_2(F) \otimes_\pi \mathbf{Z}$ in $H_2(F)$ is free of rank three. Thus this image is of finite index in $ker(p_*)$, which is easily seen to be of rank three \end{example} Finally, we may consider the same arrangement, but with various rank one local systems. \begin{example}Take the same arrangement as in Example \ref{generic example}, but take rank one local systems $\mathcal{L}_a$ corresponding to the automorphisms sending all generators of first homology to $a=1$, $a=i$, $a=-1$, $a=-i$ respectively. Then by Corollary 1.5 of \cite{cohensuciu} the homology $H_2(F;\mathbf{C})$ is the direct sum of the four corresponding local system homology groups $H_2(F;\mathcal{L}_a)$ where $a=1,i,-1,-i$. The latter three values of $a$ yield ``non-resonant"' local systems, with homology concentrated in degree two (of dimension one). The value $a=1$ yields the homology of $M$ with betti numbers $(1,3,3)$. In each case $a\neq1$ the twisted Hurewicz map \[h_2:\pi_2(M,m) \otimes_{\mathbf{Z}}\mathcal{L}_m \rightarrow H_2(M,\mathcal{L}_a)\] is onto. \end{example} We close with a general \textbf{Question:} Let $\mathcal{A}$ be any complex hyperplane arrangement in $\mathbf{C}^{\ell}$, $\rho \in \pi_k(M)$ with $k \leq \ell$. When is there a rank one local system $\mathcal{L}$ on $M$ so that for the appropriate Hurewicz map $h:\pi_k(M) \otimes_Z \mathcal{L}_m \rightarrow H_k(M;\mathcal{L})$, one has $h(\rho) \neq 0$?

https://arxiv.org/abs/1507.04571

GPU-based visualization of domain-coloured algebraic Riemann surfaces

We examine an algorithm for the visualization of domain-coloured Riemann surfaces of plane algebraic curves. The approach faithfully reproduces the topology and the holomorphic structure of the Riemann surface. We discuss how the algorithm can be implemented efficiently in OpenGL with geometry shaders, and (less efficiently) even in WebGL with multiple render targets and floating point textures. While the generation of the surface takes noticeable time in both implementations, the visualization of a cached Riemann surface mesh is possible with interactive performance. This allows us to visually explore otherwise almost unimaginable mathematical objects. As examples, we look at the complex square root and the folium of Descartes. For the folium of Descartes, the visualization reveals features of the algebraic curve which are not obvious from its equation.

\section{Introduction} \subsection{Mathematical background} \label{sec:mathematical-background} The following basic example illustrates what we would like to visualize. \begin{example} Let $y$ be the square root of $x$, \[y = \sqrt{x}.\] If $x$ is a non-negative real number, we typically define $y$ as the non-negative real number whose square equals $x$, i.e.\ we always choose the non-negative solution of the equation \begin{equation} \label{eq:parabola} y^2 - x = 0 \end{equation} as $y = \sqrt{x}$. For negative real numbers $x$, no real number $y$ solves \autoref{eq:parabola}. However, if we define the imaginary unit $\mathrm{i}$ as a number with the property that $\mathrm{i}^2 = -1$ then the square root of $x$ becomes the purely imaginary number \[y = \mathrm{i} \sqrt{|x|}.\] Together, these two conventions yield a continuous square root function \[\sqrt{\cdot}\colon \mathbb{R} \to \mathbb{C}.\] For complex numbers $x$, \autoref{eq:parabola} has exactly two complex solutions (counted with multiplicity), the square roots of $x$. \end{example} \noindent We have seen that, for real numbers $x$, we can choose one solution of~\autoref{eq:parabola} for the square root and obtain a square root function that is continuous over the real numbers. In contrast, we cannot for every complex number $x$ choose one solution of~\autoref{eq:parabola} so that we obtain a square root function that is continuous over the complex numbers: If we plot the two solutions of~\autoref{eq:parabola} as $x$ runs along a circle centred at the origin of the complex plane, we observe that $y$ moves at half the angular velocity of $x$ (see~\autoref{fig:square-root-angular-velocity}). \begin{figure} \centering \begin{tikzpicture} \draw[dashed,gray] (0,0) circle (2); \draw[dashed,gray] (0,0) circle ({sqrt(2)}); \draw[->,>=latex] (-2.5,0) -- (2.5,0) node[right] {$\operatorname{Re}$}; \draw[->,>=latex] (0,-2.5) -- (0,2.5) node[above] {$\operatorname{Im}$}; \foreach \phi in {0,10,...,90} { \filldraw (\phi:2) circle (0.05); \draw[fill=white] (0.5*\phi:{sqrt(2)} ) circle (0.05); \draw[fill=white] (0.5*\phi+180:{sqrt(2)} ) circle (0.05); } \end{tikzpicture} \hfill \begin{tikzpicture} \draw[->,>=latex] (-2.5,0) -- (2.5,0) node[right] {$\operatorname{Re}$}; \draw[->,>=latex] (0,-2.5) -- (0,2.5) node[above] {$\operatorname{Im}$}; \filldraw (0:2) circle (0.05); \draw[fill=white] (0:{sqrt(2)} ) circle (0.05); \draw[fill=white] (180:{sqrt(2)} ) circle (0.05); \draw[thick,->,>=triangle 45] (2:{sqrt(2)} ) arc (2:178:{sqrt(2)} ); \draw[thick,->,>=triangle 45] (182:{sqrt(2)} ) arc (182:358:{sqrt(2)} ); \draw[thick,->,>=triangle 45] (2:2) arc (2:358:2); \end{tikzpicture} \caption{When a complex number (black points) runs along a circle centred at the origin of the complex plane, its square roots (white points) move at half the angular velocity (left image). After a full turn of $x$, the square roots have interchanged positions. The real part of the square root that initially had positive real part has become negative, and vice versa (right image).} \label{fig:square-root-angular-velocity} \end{figure} When $x$ completes one full circle and reaches its initial position again, the square roots have interchanged signs. Therefore, a discontinuity occurs when $x$ returns to its initial position after one full turn and the square root jumps back to its initial position. Note that by choosing the values of the square root in a different manner, or, equivalently, letting $x$ start at at a different position, we can move the discontinuity to an arbitrary position on the circle. Moreover, note that there is (at least) one discontinuity on any circle of any radius centred at the origin. In order to define the principal branch of the complex square root function, we usually align the discontinuities along the negative real axis, the canonical branch cut of the complex square root function, and choose those values that on the real axis agree with the square root over the real numbers. Alternatively, we can extend the domain of the complex square root to make it a single-valued and continuous function. To that end, we take two copies of the extended complex plane and slit them along the negative real axis. On the first copy, we choose the solution of~\autoref{eq:parabola} with non-negative real part as the complex square root of $x$; on the second copy, we choose the other solution. We glue the upper side (lower side) of the slit of the first copy to the lower side (upper side) of the slit of the second copy and obtain a Riemann surface of the complex square root. (In three dimensions, this is not possible without self-intersections.) \begin{remark} On the Riemann surface, the complex square root is single-valued and continuous. It is even analytic except at the origin and at infinity, which are exactly the points where the two solutions of \autoref{eq:parabola} coincide. \end{remark} \begin{remark} The branch cut is not a special curve of the Riemann surface. When we glue the Riemann surface together, the branch cut becomes a curve like every other curve on the Riemann surface. If we had used a different curve between the origin and infinity as branch cut, we would have obtained the same result. \end{remark} \begin{remark} \autoref{eq:parabola} describes a parabola. We can proceed analogously to obtain Riemann surfaces for other plane algebraic curves. \end{remark} \subsection{Previous work} Probably the most common approach for visualization of functions is to plot a function graph. However, for a complex function \[g\colon \mathbb{C} \to \mathbb{C},\] the function graph \[\{(z, g(z)) \mid z \in \mathbb{C}\} \subset \mathbb{C} \times \mathbb{C} \simeq \mathbb{R}^2 \times \mathbb{R}^2\] is a (real) two-dimensional surface in (real) four-dimensional space. One way to visualize a four-dimensional object is to plot several two- or three-dimensional slices. This approach seems less useful for understanding the overall structure of the object. Another traditional method to visualize complex functions is domain colouring. The principle of domain colouring is to colour every point in the domain of a function with the colour of its function value in a reference image. If we choose the reference image wisely, a lot of information about the complex function can be read off from the resulting two-dimensional image (see e.g.~\cite{PoelkePolthier2012} and~\cite{Wegert2012}). The idea of lifting domain colouring to Riemann surfaces is due to \ocite{PoelkePolthier2009}. We can interpret a Riemann surface of a plane algebraic curve \begin{equation} \mathcal{C}\colon f(x,y) = 0 \end{equation} as a function graph of a multivalued complex function, which maps every $x$ to multiple values of $y$. If $f(x,y)$ is a polynomial of degree $n$ in $y$, there are exactly $n$ values of $y$ for every value of $x$ that satisfy $f(x,y) = 0$ (counted with multiplicity). Every such pair $(x,y)$ corresponds to a point on the Riemann surface. In other words, the Riemann surface is an $n$-fold cover of the complex plane. Let $\pi\colon (x,y) \mapsto x$ denote a projection function on the Riemann surface. Then the values of $y$ at $x$ correspond to the elements of the fibre $\pi^{-1}(x)$. The situation is analogous to function graphs of single-valued functions from the real numbers (or the real plane) to the real numbers, where one function value lies above every point in the domain. We can transfer the Riemann surface from (real) four-dimensional space into (real) three-dimensional space by introducing a height function $H\colon \mathbb{C} \to \mathbb{R}$. We typically use the real part as a height function. We plot the surface \[\{(\operatorname{Re} x, \operatorname{Im} x, H (y)) \mid x, y \in \mathbb{C}\colon f (x, y) = 0\}\] and use domain colouring to represent the value of $y$ at every point of the surface. In practice, we want to generate a triangle mesh that approximates the Riemann surface as the graph of a multivalued function over a triangulated domain in the complex plane. The Riemann surface mesh approximates the continuous Riemann surface in the following sense: The $y$-values at the vertices of a triangle of the Riemann surface mesh result from each other under analytic continuation along the edges of the underlying triangle in the triangulated domain. If $f(x,y) = 0$ is a polynomial of degree $n$ in $y$ there are $n$ values of $y$ above every vertex $x$ of the triangulated domain. Hence, we have to determine which of the $3 n$ values of $y$ above a triangle in the triangulated domain form triangles of the Riemann surface mesh. A wrong combination of values of $y$ to triangles might for example occur due to discontinuity if we used the principal branch of the square root function for the computation of $y$. This would produce artefacts in the visualization for which there is no mathematical justification. For the generation of such a Riemann surface mesh, previous algorithms have solved systems of differential equations~\cites{Trott2008, NieserPoelkePolthier2010} or explicitly identified and analyzed branch cuts to remove discontinuities~\citelist{\cite{Kranich2012} \cite{Wegert2012}*{Section~7.6}}. In the next section, we discuss an algorithm based on a different idea: We can exploit that $y(x)$ is continuous almost everywhere on the Riemann surface and therefore, if $x$ changes little, so does $y(x)$. \section{Algorithms} In this section, we describe algorithms for generating and visualizing domain-coloured Riemann surface meshes of plane algebraic curves. Let \begin{equation} \label{eq:plane-algebraic-curve} \mathcal{C}\colon f (x, y) = 0 \end{equation} be a complex plane algebraic curve. In particular, let $f$ be a polynomial with complex coefficients of degree $n$ in $y$. Moreover let $U \subset \mathbb{C}$ be a triangulated domain in the complex plane. (In practice, $U$ is typically rectangular.) We want to generate a Riemann surface mesh of $\mathcal{C}$. The mesh discretizes a part of a (real) two-dimensional surface in (real) four-dimensional space. We can visualize it using a height function and domain colouring, as described in the previous section. We obtain a Riemann surface mesh of $\mathcal{C}$ as a graph of the multivalued function induced by~\autoref{eq:plane-algebraic-curve}, which maps every value of $x$ in $U$ to $n$ values of $y$ such that $f (x, y) = 0$. For every triangle in $U$, we thus obtain $n$ values of $y$ at each of its three vertices. The problem is to determine whether, and if so, how, the $3 n$ values of $y$ can be combined to form triangles of the Riemann surface mesh. The resulting triangles should be consistent with the fact that $y$ as a function of $x$ is analytic almost everywhere on the Riemann surface. This is impossible if the triangle in $U$ contains a ramification point of $y(x)$. In this case, we subdivide the triangle to obtain smaller triangles mostly free of ramification points. Otherwise, the triangles of the Riemann surface mesh are uniquely determined by analytic continuation of $y(x)$ along the edges of the triangle in $U$. In order to find these triangles of the Riemann surface mesh, we use the following idea: Consider a triangle $\triangle x_1 x_2 x_3$ in $U$ that is free of ramification points of $y(x)$. Under this assumption, $y(x)$ is continuous on those parts of the Riemann surface that lie above $\triangle x_1 x_2 x_3$. Hence, for every $\varepsilon > 0$ there exists $\delta > 0$ such that $|y(x_1) - y(x_2)| < \varepsilon$ for all $x_1, x_2$ with $|x_1 - x_2| < \delta$. If $\varepsilon$ is half the minimum distance between the $n$ values of $y(x)$ at $x_1$ and $|x_1 - x_2|$ is smaller than the corresponding $\delta$, then the values of $y(x)$ at $x_2$ are closer to the corresponding values of $y(x)$ at $x_1$ than to any other value of $y(x)$ at $x_1$. In other words, if triangle $\triangle x_1 x_2 x_3$ is small enough, we can combine the values of $y$ at its vertices to triangles of the Riemann surface mesh based on proximity: Among the $3 n$ values of $y$ at the vertices of triangle $\triangle x_1 x_2 x_3$, every three values of $y$ closest to each other form a triangle of the Riemann surface mesh. We can algorithmically compute a $\delta > 0$ as above using the epsilon-delta bound for plane algebraic curves of~\autoref{thm:epsilon-delta-bound}. \autoref{thm:epsilon-delta-bound} is of essential importance for our approach. Our approach only works because \autoref{thm:epsilon-delta-bound} provides us with a reliable bound computable as a function of $x$ that depends only on a few constants derived from the coefficients of $f(x,y)$. If triangle $\triangle x_1 x_2 x_3$ is not small enough to correctly combine the values of $y$ at its vertices based on proximity, we subdivide the triangle. In summary, we obtain the following algorithm: \begin{algorithm}[Generation of a Riemann surface mesh] \label{alg:riemann-surface-mesh} Let $U \subset \mathbb{C}$ be a triangulated domain in the complex plane. Let \[\mathcal{C}\colon f (x, y) = 0\] be a complex plane algebraic curve and $f (x, y)$ a polynomial of degree $n$ in $y$. We prescribe a maximal subdivision depth (as a maximal number of iterations or as a minimal edge length). \begin{enumerate} \item Compute the global ingredients of the epsilon-delta bound of \autoref{thm:epsilon-delta-bound} for $y (x)$. \item For every triangle $\triangle x_1 x_2 x_3$ in $U$: \begin{enumerate} \item Compute the $3 n$ values of $y(x)$ at $x_1, x_2, x_3$, \[\{y_k (x_j) \mid f (x_j, y_k (x_j)) = 0,\ j = 1, 2, 3,\ k = 1, 2, \dots, n\}.\] \item Compute half the minimum distance between the values of $y(x)$ at each of the vertices of $\triangle x_1 x_2 x_3$, \[\varepsilon (x_j) = \tfrac{1}{2} \min_{k \neq l} |y_k(x_j) - y_l(x_j)|, \quad j = 1, 2, 3.\] \item Compute $\delta (x_j)$ by the epsilon-delta bound of \autoref{thm:epsilon-delta-bound} so that \[|y(x_j) - y(x)| < \varepsilon (x_j), \text{ if } |x_j - x| < \delta (x_j), \quad j = 1, 2, 3.\] \item Determine which of the edges of $\triangle x_1 x_2 x_3$ are longer than the minimum of the $\delta (x_j)$ at their endpoints and must be subdivided. \item Select the right adaptive refinement pattern (see~\autoref{fig:adaptive-refinement-patterns}) and subdivide $\triangle x_1 x_2 x_3$ accordingly. \end{enumerate} \item Repeat step 2 until the maximal subdivision depth is reached. \item Discard every triangle in $U$ with an edge longer than the minimum of the $\delta (x_j)$ at its endpoints. \item For every triangle $\triangle x_1 x_2 x_3$ in $U$, combine the values of $y(x)$ at its vertices to triangles of the Riemann surface mesh based on proximity. More formally, the triangles added to the Riemann surface mesh comprise the vertices \begin{gather*} (x_1, y_k(x_1)),\\ (x_2, \operatorname*{argmin}_{y_l (x_2)} |y_k (x_1) - y_l(x_2)|),\\ (x_3, \operatorname*{argmin}_{y_l (x_3)} |y_k(x_1) - y_l(x_3)|), \end{gather*} for $k = 1, 2, \dots, n$. \item Output the Riemann surface mesh and stop. \end{enumerate} \end{algorithm} \begin{figure} \centering \begin{tikzpicture}[scale=0.75] \coordinate (A) at (210:1); \coordinate (B) at (330:1); \coordinate (C) at (90:1); \draw (A) -- (B) -- (C) -- cycle; \end{tikzpicture} \hfill \begin{tikzpicture} \draw (A) -- (B) -- (C) -- cycle; \draw (B) -- ($(C)!0.5!(A)$); \end{tikzpicture} \hfill \begin{tikzpicture} \draw (A) -- (B) -- (C) -- cycle; \draw (A) -- ($(B)!0.5!(C)$); \end{tikzpicture} \hfill \begin{tikzpicture} \draw (A) -- (B) -- (C) -- cycle; \draw (A) -- ($(B)!0.5!(C)$); \draw ($(C)!0.5!(A)$) -- ($(B)!0.5!(C)$); \end{tikzpicture} \hfill \begin{tikzpicture} \draw (A) -- (B) -- (C) -- cycle; \draw (C) -- ($(A)!0.5!(B)$); \end{tikzpicture} \hfill \begin{tikzpicture} \draw (A) -- (B) -- (C) -- cycle; \draw ($(A)!0.5!(B)$) -- ($(C)!0.5!(A)$); \draw (B) -- ($(C)!0.5!(A)$); \end{tikzpicture} \hfill \begin{tikzpicture} \draw (A) -- (B) -- (C) -- cycle; \draw (C) -- ($(A)!0.5!(B)$); \draw ($(A)!0.5!(B)$) -- ($(B)!0.5!(C)$); \end{tikzpicture} \hfill \begin{tikzpicture} \draw (A) -- (B) -- (C) -- cycle; \draw ($(A)!0.5!(B)$) -- ($(B)!0.5!(C)$); \draw ($(B)!0.5!(C)$) -- ($(C)!0.5!(A)$); \draw ($(C)!0.5!(A)$) -- ($(A)!0.5!(B)$); \end{tikzpicture} \caption{Adaptive refinement patterns used in \autoref{alg:riemann-surface-mesh}} \label{fig:adaptive-refinement-patterns} \end{figure} \begin{remark} By construction, \autoref{alg:riemann-surface-mesh} generates a Riemann surface mesh that is consistent with the analytic structure of the Riemann surface of $\mathcal{C}$. \end{remark} \begin{remark} The adaptive refinement patterns used for the subdivision of triangles, whose edges are too long, produce a watertight subdivision. \end{remark} \begin{remark} Step 4 of~\autoref{alg:riemann-surface-mesh} produces holes around the ramification points of $y(x)$. We can make these holes very small if we choose the maximal subdivision depth appropriately. \end{remark} \noindent For the visualization of a Riemann surface mesh, we use the following algorithm: \begin{algorithm}[Visualization of a Riemann surface mesh] \label{alg:visualization} Let a Riemann surface mesh and a domain colouring reference image be given. We choose a height function $H\colon \mathbb{C} \to \mathbb{R}$ to transform a point on the Riemann surface mesh from (real) four-dimensional space to a point in (real) three-dimensional space, \[(x, y(x)) \mapsto (\operatorname{Re} x, \operatorname{Im} x, H (y(x))).\] \begin{enumerate} \item Draw the mesh that results from transforming every vertex of the Riemann surface mesh as above. \item Interpolate the value of $y(x)$ on the transformed mesh. \item Assign to every point on the transformed mesh the colour in the reference image of the value that $y(x)$ attains at that point on the transformed mesh. \end{enumerate} \end{algorithm} \begin{remark} If we choose the real (or imaginary) part of $y(x)$ as a height function, the transformation from (real) four-dimensional to (real) three-dimensional space becomes a projection. \end{remark} \begin{remark} Using the real part of $y(x)$ as a height function has the advantage that the visualization then contains the image of $\mathcal{C}$ interpreted as a real plane algebraic curve. It is the intersection of the visualization of the Riemann surface mesh in (real) three-dimensional space with the $\operatorname{Re} x$-$\operatorname{Re} y$-plane (the $xz$-plane, if we label the coordinate axes of real three-dimensional space such that the $x$-axis points to the right and the $z$-axis points upwards). \end{remark} \begin{remark} The computation of the Riemann surface mesh by \autoref{alg:riemann-surface-mesh} is independent of the choice of height function used for its visualization. \end{remark} \section{Implementation} In this section, we discuss how~\autoref{alg:riemann-surface-mesh} and~\autoref{alg:visualization} can be implemented using OpenGL and WebGL.\@ Since WebGL targets a much wider range of devices, its API is more limited than that of OpenGL.\@ Consequently, our implementation using WebGL differs substantially from our implementation using OpenGL.\@ Before we discuss each setup separately, let us talk about what they have in common. The main part of our programs is written in shading language (GLSL for OpenGL and ESSL for WebGL) and runs on the GPU.\@ We use the CPU to compute the global ingredients for the epsilon-delta bound, to generate shading language code that computes the epsilon-delta bound for $\mathcal{C}\colon f(x,y) = 0$ as a function of $x$, and to generate a coarse triangulation of the input domain. The implementations in OpenGL and WebGL share some shading language code. Since there is no native type for complex numbers, we represent them using two-dimensional floating point vectors. Common routines include complex arithmetic, numerical root-finding algorithms, the computation of the epsilon-delta bound, and domain colouring. The implementation of complex arithmetic is straightforward and we shall not go into detail about it. We need numerical root-finding algorithms to approximate roots of polynomials in order to compute values of $y(x)$ (and to compute the global ingredients of the epsilon-delta bound). For instance, Laguerre's method~\cite{PressTeukolskyVetterlingFlannery2007}*{Section~9.5.1} and deflation~\cite{PressTeukolskyVetterlingFlannery2007}*{Section~9.5.3} or Weierstraß--Durand--Kerner method~\cites{Weierstrass1891,Durand1960,Kerner1966} are well-suited. The latter may be a little easier to implement in shading language (due to the absence of variable-length arrays). For the computation of the epsilon-delta bound, we use the following theorem: \begin{theorem}[cf.~\cite{Kranich2015}*{Theorem~2.1}] \label{thm:epsilon-delta-bound} Let $\mathcal{C}\colon f(x, y) = 0$ be a complex plane algebraic curve, where \[f(x, y) = \sum\limits_{k = 0}^n a_k(x) y^{n - k}\] is a polynomial of degree $n$ in $y$ whose coefficients $a_k(x)$ are polynomials in $x$ of the form \[a_k(x) = \sum_{l = 0}^{m_k} a_{kl} x^{m_k - l}.\] Let $x_1 \in \mathbb{C}$ be a point in the complex plane at which neither the leading coefficient $a_0(x)$ nor the discriminant of $f(x,y)$ w.r.t.\ $y$ vanish. Then for every $\varepsilon > 0$, we can algorithmically compute $\delta > 0$ such that \[|y_j(x_1) - y_j(x_2)| < \varepsilon\] for all holomorphic functions $y_j(x)$, $j = 1, 2, \dots, n$, which satisfy $f(x, y_j(x)) = 0$ in a neighbourhood of $x_1$ and for all $x_2$ with $|x_1 - x_2| < \delta$. We obtain \begin{equation} \label{eq:epsilon-delta-bound} \delta = \frac{\rho \left(\sqrt{{(\rho Y - \varepsilon)}^2 + 4 \varepsilon M} - (\rho Y + \varepsilon)\right)}{2 (M - \rho Y)}, \end{equation} where \begin{gather*} \rho < \min\{|x_1 - x| \colon a_0(x) \cdot \Delta_y(f(x,y))(x) = 0\},\\ Y:= \max\limits_j \left|\dfrac{f_x(x_1, y_j(x_1))}{f_y(x_1, y_j(x_1))}\right|,\quad M := 2 \max\limits_k {\left(\frac{\tilde{a}_k}{\tilde{a}_0}\right)}^{\frac{1}{k}},\\ \tilde{a}_0 := |a_{00}| \prod_{l = 1}^{m_0} (|\bar{x}_l - x_1| - \rho) > 0,\quad \tilde{a}_k := \sum_{l = 0}^{m_k} |a_{kl}| {(|x_1| + |\rho|)}^{n-l}, \end{gather*} $\Delta_y(f(x,y))(x)$ denotes the $y$-discriminant of $f(x,y)$, and $\bar{x}_l$, $l = 1, 2, \dots, m_0$, are the zeros of $a_0 (x)$. \end{theorem} Note that the computation is parallelizable since the epsilon-delta bound can be implemented as a function of $x$ that depends on only a few constants derived from the coefficients of $f(x,y)$. Instead of computing texture coordinates, which would depend on the range of $y(x)$ on the input domain, we generate the domain colouring procedurally on-the-fly. To that end, we use a variation of the enhanced phase portrait colour scheme of~\cite{Wegert2012}*{Section~2.5}. The reference image is shown in~\autoref{fig:domain-colouring}. We discuss the colour scheme in~\autoref{sec:reference-image}. The main difference between the implementations in OpenGL and WebGL is how the common routines can be combined to realize \autoref{alg:riemann-surface-mesh} and~\autoref{alg:visualization}. \subsection{An implementation in OpenGL} \subsubsection{Implementation of~\autoref{alg:riemann-surface-mesh} in OpenGL} Our implementation of~\autoref{alg:riemann-surface-mesh} in OpenGL comprises three GLSL programs, for initialization, subdivision, and assembly of the Riemann surface mesh. We cache the output of each program using transform feedback and feed it back to the next program. The initialization program consists only of a vertex shader, which operates on the vertices of the triangulated input domain. For every vertex $x$, we compute $y_k (x)$, $k = 1, 2, \dots, n$, and $\delta (x)$. After initialization, we run the subdivision program. The program consists of a pass-through vertex shader and a geometry shader. The geometry shader operates on the triangles of the triangulated input domain or of its last subdivision, respectively. We have access to the values of $x_j$, $\delta (x_j)$, and $y_k (x_j)$, $k = 1, 2, \dots, n$, $j = 1, 2, 3$, at the vertices of each triangle $\triangle x_1 x_2 x_3$. We determine which edges of triangle $\triangle x_1 x_2 x_3$ are longer than the minimum of the $\delta (x_j)$ at their endpoints. In order to subdivide these edges, we compute their midpoints $x$, and $\delta (x)$ and $y_k (x)$, $k = 1, 2, \dots, n$, at the midpoints. We use the appropriate adaptive refine pattern of \autoref{fig:adaptive-refinement-patterns} and output between one and four triangles for every input triangle. In doing so, we reuse previously computed values rather than recomputing them. We run the subdivision program iteratively until we reach the prescribed maximal subdivision depth. The assembly program consists of a pass-through vertex shader and a geometry shader. The geometry shader operates on the triangles of the adaptively subdivided input domain. We again have access to the values of $x_j$, $\delta (x_j)$, and $y_k (x_j)$, $k = 1, 2, \dots, n$, $j = 1, 2, 3$, at the vertices of each triangle $\triangle x_1 x_2 x_3$. For every triangle $\triangle x_1 x_2 x_3$, we test whether one of its edges is longer than the minimum of the $\delta (x_j)$ at its endpoints. In this case, we discard the triangle. Otherwise, we determine the triangles of the Riemann surface mesh by proximity (see~\autoref{alg:riemann-surface-mesh}, step~5) and output these $n$ triangles. We also cache the assembled Riemann surface mesh using transform feedback so that we can pass it as input to our implementation of the visualization algorithm (\autoref{alg:visualization}). \subsubsection{Implementation of~\autoref{alg:visualization} in OpenGL} Our implementation of~\autoref{alg:visualization} in OpenGL consists of one GLSL program with a vertex and a fragment shader. The vertex shader operates on the vertices of a Riemann surface mesh generated by our implementation of~\autoref{alg:riemann-surface-mesh}. We apply height function $H\colon \mathbb{C} \to \mathbb{R}$ to map each (real) four-dimensional vertex \[(\operatorname{Re} x, \operatorname{Im} x, \operatorname{Re} y (x), \operatorname{Im} y (x))\] to a (real) three-dimensional vertex \[(\operatorname{Re} x, \operatorname{Im} x, H (y (x))).\] We homogenize the coordinates of this vertex and transform them using the model-view-projection matrix. We pass $y (x)$ as a varying variable to the fragment shader. The fragment shader operates on the interpolated value of $y (x)$ at a fragment of a pixel of the output device. We compute the colour of $y (x)$ according to our domain colouring reference image. \subsubsection{Remarks} Using our implementation, the generation of a Riemann surface mesh takes little but noticeable time. The bottlenecks of the implementation are numerical root-finding and iterative subdivision. However, if we use transform feedback to cache the Riemann surface mesh and pass it to the implementation of the visualization algorithm, we obtain interactive performance. Another advantage of using transform feedback to cache the Riemann surface mesh is that we can easily export the data. If we additionally compute texture coordinates and a high-resolution reference image, we can even print our visualization using a full colour 3D printer (see \autoref{fig:3d-printed-models}). \begin{figure}[ht!] \centering \includegraphics[width=\linewidth]{3d-printed-models.jpg} \caption{3D-printed models of domain-coloured Riemann surfaces of square root, folium of Descartes, and unit circle (clockwise). The merchandise coffee mug of DFG Collaborative Research Center TRR 109, ``Discretization in Geometry and Dynamics'', is included in the picture as an indicator for the size of the models.} \label{fig:3d-printed-models} \end{figure} \subsection{An implementation in WebGL} In order to support a wider range of devices, the WebGL API is much more limited than the OpenGL API.\@ Particularly, in WebGL, geometry shaders and transform feedback are currently unavailable. (The WebGL 2 draft includes transform feedback and compute shaders.) Therefore our implementation in WebGL differs substantially from our implementation in OpenGL.\@ \subsubsection{How to replace transform feedback} Instead of transform feedback, our implementation in WebGL uses floating point textures (specified in the \texttt{OES\_texture\_float} extension) and multiple render targets (specified in the \texttt{WEBGL\_draw\_buffers} extension). I do not claim originality of this approach. It is commonly used for running simulations on the GPU.\@ The original idea may be due to~\cite{Crane2007}. We number the vertices of every mesh consecutively and pass this number (index) to the vertex shaders along with the other attributes. In particular, vertices that are shared among several triangles must be duplicated and numbered separately. Hence, we assume that every triangle appears as three consecutive vertices in array buffer storage (triangle soup). We use floating point textures essentially as we would arrays of floats, indexed by vertex number. We store values corresponding to the $k$-th vertex in the $k$-th pixel of a texture. We can store up to four floats per pixel of a floating point texture, namely one float each in the red, green, blue, and alpha channel. If we need to store more than four floats, we use multiple render targets which allows us to colour the same pixel of several textures simultaneously. We want to store values we compute for a vertex in textures (`transform' in transform feedback). To that end, we bind the array buffers and draw the contents as points (as opposed to triangles). In the vertex shader, we compute the positions of the point with index $k$ (in normalized device coordinates) so that it is rasterized as the $k$-th pixel of the render target textures. Recall that normalized device coordinates range in ${\left[-1,1\right]}^3$. For example, if $h$ and $w$ denote the height and width of the textures in pixels, we assign to the point with index $k$ the position \[\left(\frac{2 \cdot (k \text{ mod } w) + 1}{w} - 1, \frac{2 \cdot \lfloor k / w\rfloor + 1}{h}- 1, 0\right).\] In the fragment shader, we compute the values to be stored and assign them as output colours in a specific order (as we later want to retrieve them). We want to read stored values for a vertex from textures (`feedback' in transform feedback). To that end, we bind the textures and an array buffer containing a range of vertex numbers (indices). We draw the contents of the array buffer as points or triangles (depending on whether we want to send the output to different textures or to the screen). In the vertex shader, we compute texture coordinates for the point with index $k$ which allow us to lookup the $k$-th pixel from the textures. Recall that texture coordinates range in ${\left[0,1\right]}^2$. For a texture of height $h$ and width $w$, we compute the texture coordinates \[\left(\frac{(k \text{ mod } w) + 0.5}{w},\frac{(k \text{ mod } w) + 0.5}{h}\right).\] Adding $0.5$ in the numerators accounts for the fact that we want to obtain coordinates for the centre of a pixel in order to avoid interpolation with adjacent pixels. We pass the texture coordinates to the fragment shader, where we can use them to perform a texture lookup. In order to access data of a whole triangle (as in geometry shaders), we can, in the vertex shader, determine the indices of the other vertices of the triangle. For example, the point with index $k$ is part of the triangle whose vertices have indices \[k - (k \text{ mod } 3),\ k - (k \text { mod } 3) + 1, \text{ and } k - (k \text { mod } 3) + 2.\] We compute normalized device coordinates or texture coordinates for all three indices and pass them to the fragment shader, together with the index of the triangle vertex currently under consideration. \subsubsection{How to replace geometry shaders} We replace the geometry shader of the subdivision program of our implementation in OpenGL using a variation of a method proposed by \ocite{BoubekeurSchlick2008}. The method works as follows: We precompute all adaptive refinement patterns up to a certain subdivision depth, in our case eight adaptive refinement patterns up to depth one (see~\autoref{fig:adaptive-refinement-patterns}). We use barycentric coordinates to store the positions of the triangle vertices of each refinement pattern in an array buffer. Using array buffers of different lengths allows us to achieve variable-length output, as with geometry shaders. For every triangle of a coarse input mesh, we draw the triangles in the array buffer of the appropriate adaptive refinement pattern. We use the vertex positions of the input triangle (read from a texture or from uniform variables) and the barycentric coordinates of the triangles of the adaptive refinement pattern to compute the vertex positions of the output triangle. We can combine this method with floating point texture and multiple render targets as outlined above, if we number the vertices of each adaptive refinement pattern consecutively and store those indices together with the barycentric coordinates. We pass an offset as a uniform variable to the vertex shader that needs to be added to the indices. We draw the adaptive refinement pattern and increment the offset by the number of vertices in the adaptive refinement pattern. The geometry shader of the assembly program has fixed-length output. It generates exactly $n$ triangles of the Riemann surface mesh per triangle of the (subdivided) input mesh. We can replace it with $n$ invocations of a vertex shader, one for every sheet of the Riemann surface mesh. We pass the number of the current sheet to the vertex shader as a uniform variable. \subsubsection{Remarks} We cannot expect our WebGL implementation to reach the same performance as our OpenGL implementation. In the subdivision program, since we draw a different adaptive refinement pattern for every triangle of the input mesh, we lose parallelism. Consequently, subdivision in WebGL is much slower than its OpenGL counterpart. However, if we cache the assembled Riemann surface mesh (in textures) and pass it to our implementation of the visualization algorithm, we can still achieve interactive performance. \section{Examples} In this section, we discuss domain-coloured Riemann surface meshes for the complex square root function and for the folium of Descartes. Before that, let us explain our domain colouring reference image so that we can interpret the domain-coloured Riemann surface meshes. \subsection{Domain colouring reference image} \label{sec:reference-image} Recall that the basic idea of domain colouring is the following: If we want to visualize a complex function \[f\colon K \subset \mathbb{C} \to \mathbb{C},\] we face the problem that its graph is real four-dimensional. However, we can visualize the behaviour of the function by colouring every point in its domain with the colour of the function value at that point in a reference image. The reference image is the domain colouring of the complex identity function. Depending on what reference image we choose, we can read off various properties of a function from its domain colouring. For an overview of different colour schemes, we refer to~\cites{PoelkePolthier2012,Wegert2012}. As our reference image, we use a variation of the enhanced phase portrait colour scheme of~\cite{Wegert2012}*{Section~2.5}. The reference image is best described using polar coordinates \[r \mathrm{e}^{\mathrm{i} \varphi} = r (\cos\varphi + \mathrm{i} \sin\varphi)\] of a complex number with \emph{modulus} $r > 0$ and \emph{phase} $\varphi \in \left[0,2\pi\right)$. Firstly, we encode the phase at any point in the domain as the hue of its colour (in HSI colour space). In a square with side length $10$ centred at the origin, we thus obtain the colour wheel shown in~\autoref{fig:domain-colouring-hue}. As the phase changes from $0$ to $2 \pi$, we obtain every colour of the rainbow. Positive real numbers, which have phase $0$, are coloured in pure red. Negative real numbers, which have phase $\pi$, are coloured in cyan. Purely imaginary numbers do not have such distinctive colours. (This can be fixed using the NIST continuous phase mapping, which scales the phase piecewise linearly so that purely imaginary numbers with positive imaginary part become yellow and purely imaginary numbers with negative imaginary part become blue. See~\cite{DLMF}*{\url{http://dlmf.nist.gov/help/vrml/aboutcolor\#S2.SS2}}. For simplicity, we do not follow this approach here.) \begin{figure}[ht] \centering \setcounter{subfigure}{0} \begin{subfigure}[b]{0.32\linewidth} \includegraphics[width=\linewidth]{hue} \caption{}\label{fig:domain-colouring-hue} \end{subfigure} \begin{subfigure}[b]{0.32\linewidth} \includegraphics[width=\linewidth]{blackp} \caption{}\label{fig:domain-colouring-blackp} \end{subfigure} \begin{subfigure}[b]{0.32\linewidth} \includegraphics[width=\linewidth]{reference_image} \caption{}\label{fig:domain-colouring-reference-image} \end{subfigure} \caption{Composition of domain colouring reference image: If we \subref{fig:domain-colouring-hue} represent phase by hue, \subref{fig:domain-colouring-blackp} add contour lines of phase and \subref{fig:domain-colouring-reference-image} add contour lines of modulus, we obtain our domain colouring reference image, a variation of the enhanced phase portrait colour scheme of~\cite{Wegert2012}*{Section~2.5}.} \label{fig:domain-colouring} \end{figure} Secondly, we add contour lines of complex numbers of the same phase at integer multiples of $15$ degrees (see~\autoref{fig:domain-colouring-blackp}). To that end, we change the intensity of the colour by multiplying it with a sawtooth function \[0.7 + (1.0 - 0.7) \cdot \left(\varphi / (\pi / 12) - \lfloor\varphi / (\pi / 12)\rfloor\right).\] Because phase corresponds to hue, the points of such a contour line are all of the same colour. Finally, we add contour lines of complex numbers of the same modulus on a log-scale (see~\autoref{fig:domain-colouring-reference-image}). To that end, we change the intensity of the colour by multiplying it with a sawtooth function \[0.7 + (1.0 - 0.7) \cdot \left(\log (r) / (\pi / 12) - \lfloor\log (r) / (\pi / 12)\rfloor\right).\] Note that the contour lines of phase and modulus intersect each other orthogonally. The scaling factor $1 / (\pi / 12)$ in the sawtooth function for the modulus contour lines deliberately matches the scaling factor used in the sawtooth function for the phase contour lines. Consequently, the regions enclosed by the contour lines of phase and modulus are squarish in appearance. \subsection{Complex square root} \label{sec:square-root-visualization} Recall the construction of a Riemann surface of the complex square root from \autoref{sec:mathematical-background} where we glued together its two branches at a branch cut along the negative real axis. \begin{figure}[ht] \centering \setcounter{subfigure}{0} \begin{subfigure}[b]{0.49\linewidth} \includegraphics[width=\linewidth]{squareRoot2} \caption{}\label{fig:square-root-domain-colouring-negative} \end{subfigure} \begin{subfigure}[b]{0.49\linewidth} \includegraphics[width=\linewidth]{squareRoot1} \caption{}\label{fig:square-root-domain-colouring-positive} \end{subfigure} \caption{Domain colouring of the two sheets of the complex square root} \label{fig:square-root-domain-colouring} \end{figure} The domain colouring of the two branches of the complex square root over a square of side length $10$ centred at the origin is shown in \autoref{fig:square-root-domain-colouring}. On the sheet shown in~\autoref{fig:square-root-domain-colouring-negative}, the complex square root takes values with negative real part (coloured green to blue). On the sheet shown in~\autoref{fig:square-root-domain-colouring-positive}, it takes values with positive real part (coloured purple to yellow). (The sheet shown in~\autoref{fig:square-root-domain-colouring-positive} corresponds to the principal branch of the complex square root.) On both sheets, twelve contour lines of phase are visible, half as many as in the reference image. We can see that the phase of the complex square root function changes at half the angular velocity of its argument. Moreover, the discontinuity at the branch cut along the negative real axis is clearly visible. We also see that there is a smooth transition between the second (third) quadrant of \autoref{fig:square-root-domain-colouring-negative} and the third (second) quadrant of~\autoref{fig:square-root-domain-colouring-positive}. If we cut the two sheets along the negative real axis and glue the upper side of the cut of one sheet to the lower side of the cut of the other sheet, and vice versa, we obtain a Riemann surface of the complex square root. The resulting Riemann surface, produced with \autoref{alg:riemann-surface-mesh} and~\autoref{alg:visualization} using real part as height function, is shown in \autoref{fig:square-root-perspective} (perspective) and \autoref{fig:square-root-multiview} (multiview orthogonal). \begin{figure}[ht] \centering \includegraphics[width=0.5\linewidth]{squareRoot_75_30} \caption{Domain-coloured Riemann surface of the complex square root in perspective projection} \label{fig:square-root-perspective} \end{figure} \begin{figure}[ht] \centering \setcounter{subfigure}{0} \begin{subfigure}[b]{0.24\linewidth} \includegraphics[width=\linewidth]{squareRoot_left} \caption{}\label{fig:square-root-multiview-left} \end{subfigure} \begin{subfigure}[b]{0.24\linewidth} \includegraphics[width=\linewidth]{squareRoot_front} \caption{}\label{fig:square-root-multiview-front} \end{subfigure} \begin{subfigure}[b]{0.24\linewidth} \includegraphics[width=\linewidth]{squareRoot_right} \caption{}\label{fig:square-root-multiview-right} \end{subfigure} \begin{subfigure}[b]{0.24\linewidth} \includegraphics[width=\linewidth]{squareRoot_back} \caption{}\label{fig:square-root-multiview-back} \end{subfigure} \caption{Domain-coloured Riemann surface of the complex square root in orthogonal projection from left, front, right, and back (from left to right)} \label{fig:square-root-multiview} \end{figure} Note that the self-intersection of the surface in \autoref{fig:square-root-perspective} is only an artefact of using a height function to map the Riemann surface mesh from real four-dimensional to real three-dimensional space. Evidently, the two values of the complex square root at each point of the self-intersection do not agree: they are coloured differently, in green and purple, respectively. In~\autoref{fig:square-root-multiview-front}, we see the parabola that the real parts of the $y$-values describe according to the equation $y^2 - x = 0$ when $x$ takes values on the non-negative real axis. \subsection{Folium of Descartes} \begin{figure}[ht] \centering \begin{tikzpicture}[scale=0.5] \draw[help lines] (-5,-5) grid (5,5); \draw[thick,->, >=latex] (-5,0) -- (5,0) node[below left] {$x$}; \draw[thick,->, >=latex] (0,-5) -- (0,5) node[below right] {$y$}; \clip (-5,-5) rectangle (5,5); \draw[very thick,domain=0:0.36,samples=100,lightgray] plot ({(6*\x*(-1+\x^2)^2)/(-1+3*\x^2+8*\x^3-3*\x^4+\x^6)}, {12*(\x^4-\x^2)/(3*\x^2+8*\x^3-3*\x^4+\x^6-1)}); \draw[very thick,domain=0:2.1, samples=100,lightgray] plot ({(-6*\x*(-1+\x^2)^2)/(-1+3*\x^2-8*\x^3-3*\x^4+\x^6)}, {12*(\x^4-\x^2)/(3*\x^2-8*\x^3-3*\x^4+\x^6-1)}); \end{tikzpicture} \caption{The folium of Descartes as a real plane algebraic curve} \label{fig:folium-real} \end{figure} \noindent The folium of Descartes is a classical plane algebraic curve of order three, \begin{equation} \label{eq:folium} \mathcal{C}\colon f (x, y) = x^3 + y^3 - 3 x y = 0. \end{equation} The cubic curve is nowadays called `folium' after the leaf-shaped loop that it describes in the first quadrant of the real plane (see~\autoref{fig:folium-real}). It is named in honour of the French geometer René Descartes (1596--1650), who was among the first mathematicians to introduce coordinates into geometry. Originally, the curve was called \emph{fleur de jasmin} since Descartes and some of his contemporaries, who were working out the principles of dealing with negative and infinite coordinates, initially wrongly believed that the leaf-shaped loop repeated itself in the other quadrants and therefore resembled a jasmine flower~\cite{Loria1910}*{p.~53}. \begin{figure}[ht] \centering \setcounter{subfigure}{0} \begin{subfigure}[b]{0.32\linewidth} \includegraphics[width=\linewidth]{folium1} \caption{}\label{fig:folium-domain-colouring-1} \end{subfigure} \begin{subfigure}[b]{0.32\linewidth} \includegraphics[width=\linewidth]{folium2} \caption{}\label{fig:folium-domain-colouring-2} \end{subfigure} \begin{subfigure}[b]{0.32\linewidth} \includegraphics[width=\linewidth]{folium3} \caption{}\label{fig:folium-domain-colouring-3} \end{subfigure} \caption{Domain colouring of three sheets of the folium of Descartes} \label{fig:folium-domain-colouring} \end{figure} \noindent \autoref{fig:folium-domain-colouring} shows three domain-coloured sheets of the folium of Descartes over a square of side length $10$ centred at the origin of the complex plane. We can generate these sheets by sorting the $y$-values that satisfy~\autoref{eq:folium} at every point $x$ of the domain according to their real part. The sheet shown in~\autoref{fig:folium-domain-colouring-1} uses the $y$-value with the smallest real part, the sheet shown in~\autoref{fig:folium-domain-colouring-2} the $y$-value with the second-smallest real part, and the sheet shown in~\autoref{fig:folium-domain-colouring-3} the $y$-value with the largest real part. We see that the first sheet carries $y$-values with negative real part (coloured green to blue). At the centre of the second sheet, we identify a zero of order two, which we can recognize from the fact that the colours of the colour wheel used in our reference image wind around it twice in the same order as in the reference image. It is the node of the leaf-shaped loop. The third sheet carries $y$-values with positive real part (coloured purple to yellow). There are three branch cuts (discontinuities of hue) on the first sheet, six on the second sheet and three on the third sheets. We can see how the sheets of the Riemann surface are connected to each other along the branch cuts: First and second sheet are connected at the branch cuts of the first sheet. Second and third sheet are connected at the branch cuts of the third sheet. First and third sheet are not connected directly with each other. (Imagine how much harder it would be to read this off from \autoref{eq:folium}.) Apart from the branch cuts, the map from $x$ to $y(x)$ is conformal (angle-preserving) on every sheet. We can see that the contour lines of phase and modulus intersect each other orthogonally on every sheet, as in our reference image. If we cut the sheets along the branch cuts and glue them together correctly, we obtain a Riemann surface for the folium of Descartes. The resulting Riemann surface, produced with \autoref{alg:riemann-surface-mesh} and~\autoref{alg:visualization} using real part as height function, is shown in~\autoref{fig:folium-perspective} (perspective) and~\autoref{fig:folium-multiview} (multiview orthogonal). \begin{figure}[ht] \centering \includegraphics[width=0.5\linewidth]{folium_75_30} \caption{Domain-coloured Riemann surface of the folium of Descartes in perspective projection} \label{fig:folium-perspective} \end{figure} Again, the self-intersections of the surface in \autoref{fig:folium-perspective} are only an artefact of using a height function to map the Riemann surface mesh from (real) four-dimensional to (real) three-dimensional space. \autoref{fig:folium-perspective} makes it obvious that cutting a Riemann surface into sheets by sorting $y$-values by real part may be the most straightforward but not necessarily the geometrically most appropriate method. Our Riemann surface of the folium of Descartes in large part appears to be composed of three copies of the complex plane (which looks like our reference image). Complications seem to arise only near the origin. \begin{figure}[ht] \centering \setcounter{subfigure}{0} \begin{subfigure}[b]{0.24\linewidth} \includegraphics[width=\linewidth]{folium_left} \caption{}\label{fig:folium-multiview-left} \end{subfigure} \begin{subfigure}[b]{0.24\linewidth} \includegraphics[width=\linewidth]{folium_front} \caption{}\label{fig:folium-multiview-front} \end{subfigure} \begin{subfigure}[b]{0.24\linewidth} \includegraphics[width=\linewidth]{folium_right} \caption{}\label{fig:folium-multiview-right} \end{subfigure} \begin{subfigure}[b]{0.24\linewidth} \includegraphics[width=\linewidth]{folium_back} \caption{}\label{fig:folium-multiview-back} \end{subfigure} \caption{Domain-coloured Riemann surface of the folium of Descartes in orthogonal projection from left, front, right, and back (from left to right)} \label{fig:folium-multiview} \end{figure} If we look closely at~\autoref{fig:folium-multiview-front}, we may see how we obtain the real folium of Descartes (as a real plane algebraic curve) as the intersection of our Riemann surface mesh with the $\operatorname{Re} x$-$\operatorname{Re} y$-plane. The leaf-shaped loop is clearly visible as a hole in our visualization. One of the `complex planes of which the Riemann surface is composed' is so thin that it is barely visible from this perspective. It is almost asymptotic to the `wings' of the folium of Descartes (as a real plane algebraic curve) in the second and fourth quadrant of the real $xy$-plane. Right below the centre of~\autoref{fig:folium-multiview-right}, we see two leaf-shaped loops in complex directions. Perhaps Descartes and his contemporaries were not entirely wrong after all to believe that the folium of Descartes has more than one leaf. Indeed, if we let $x' = \mathrm{e}^{\pm\mathrm{i} \pi / 3} x$, we discover that in the $\operatorname{Re} x'$-$\operatorname{Re} y$-plane the curve describes a leaf-shaped loop, which is exactly half as high as that in the $\operatorname{Re} x$-$\operatorname{Re} y$-plane (this also holds for the `wings') and rotated into a different quadrant (see~\autoref{fig:folium-more-leaves}). \begin{figure}[!ht] \centering \setcounter{subfigure}{0} \begin{subfigure}[b]{0.49\linewidth} \includegraphics[width=\linewidth]{folium_green_loop} \caption{}\label{fig:folium-green-loop} \end{subfigure} \begin{subfigure}[b]{0.49\linewidth} \includegraphics[width=\linewidth]{folium_blue_loop} \caption{}\label{fig:folium-blue-loop} \end{subfigure} \caption{Leaf-shaped loops of the folium of Descartes in complex directions.} \label{fig:folium-more-leaves} \end{figure} \section{Conclusion} We have discussed algorithms for the generation of a Riemann surface mesh of a plane algebraic curve (\autoref{alg:riemann-surface-mesh}) and its visualization as a domain-coloured surface (\autoref{alg:visualization}) and their implementation using OpenGL and WebGL. The WebGL implementation combines floating point textures, multiple render targets, and a method due to \ocite{BoubekeurSchlick2008} to replace the use of transform feedback and geometry shaders of the OpenGL implementation. While the generation of the surface takes noticeable time in both implementations, the visualization of a cached Riemann surface mesh is possible with interactive performance. This allows us to visually explore otherwise almost unimaginable mathematical objects. Sometimes the visualization makes properties of the plane algebraic curves immediately apparent that may not so easily be read off from its equation. It is possible to turn these domain-coloured Riemann surface meshes into physical models using a full colour 3D printer. \section*{Funding} This research was supported by DFG Collaborative Research Center TRR 109, ``Discretization in Geometry and Dynamics''. \begin{bibdiv} \begin{biblist} \bib{BoubekeurSchlick2008}{article}{ title={A Flexible Kernel for Adaptive Mesh Refinement on GPU}, author={Boubekeur, Tamy}, author={Schlick, Christophe}, journal={Computer Graphics Forum}, volume={27}, number={1}, pages={102--113}, date={2008}, doi={10.1111/j.1467-8659.2007.01040.x} } \bib{Crane2007}{article}{ title={Real-Time Simulation and Rendering of 3D Fluids}, author={Crane, Keenan}, author={Llamas, Ignacio}, author={Tariq, Sarah}, book={ title={GPU gems 3}, editor={Nguyen, Hubert}, publisher={Addison-Wesley} }, pages={633--675}, date={2007} } \bib{Durand1960}{book}{ title={Equations du type F(x), racines d'un polynôme}, author={Durand, Émile}, series={Solutions numériques des équations algébriques}, volume={1}, date={1960}, publisher={Masson}, address={Paris} } \bib{Kerner1966}{article}{ title={Ein Gesamtschrittverfahren zur Berechnung der Nullstellen von Polynomen}, author={Kerner, Immo O.}, date={1966}, journal={Numerische Mathematik}, volume={8}, number={3}, pages={290--294}, publisher={Springer}, doi={10.1007/BF02162564}, } \bib{Kranich2012}{thesis}{ title={Real-time Visualization of Geometric Singularities}, author={Kranich, Stefan}, institution={Technische Universität München}, type={Master's thesis}, date={2012} } \bib{Kranich2015}{article}{ title={An epsilon-delta bound for plane algebraic curves and its use for certified homotopy continuation of systems of plane algebraic curves}, author={Kranich, Stefan}, journal={arXiv:1505.03432 [math.CV]}, eprint={http://arxiv.org/abs/1505.03432}, date={2015} } \bib{NieserPoelkePolthier2010}{article}{ title={Automatic Generation of Riemann Surface Meshes}, author={Nieser, Matthias}, author={Poelke, Konstantin}, author={Polthier, Konrad}, book={ title={Advances in Geometric Modeling and Processing}, series={Lecture Notes in Computer Science}, editor={Mourrain, Bernard}, editor={Schaefer, Scott}, editor={Xu, Guoliang}, publisher={Springer}, address={Berlin}, volume={6130}, }, pages={161--178}, date={2010}, doi={10.1007/978-3-642-13411-1\_11} } \bib{DLMF}{misc}{ editor={NIST}, title={NIST Digital Library of Mathematical Functions}, date={2014}, url={http://dlmf.nist.gov/}, note={Release 1.0.9 of 2014-08-29. Online companion to~\cite{OlverLozierBoisvertClark2010}. \url{http://dlmf.nist.gov}} } \bib{Loria1910}{book}{ author={Loria, Gino}, translator={Schütte, Fritz}, language={German}, title={Spezielle algebraische und transzendente ebene Kurven}, subtitle={Theorie und Geschichte}, volume={1}, publisher={Teubner}, address={Leipzig}, edition={2}, date={1910} } \bib{OlverLozierBoisvertClark2010}{book}{ editor={Olver, F. W. J.}, editor={Lozier, D. W.}, editor={Boisvert, R. F.}, editor={Clark, C. W.}, title={NIST Handbook of Mathematical Functions}, publisher={Cambridge University Press}, address={New York, NY}, date={2010}, note={Print companion to~\cite{DLMF}} } \bib{PoelkePolthier2009}{article}{ title={Lifted Domain Coloring}, author={Poelke, Konstantin}, author={Polthier, Konrad}, journal={Computer Graphics Forum}, pages={735--742}, volume={28}, number={3}, date={2009}, doi={10.1111/j.1467-8659.2009.01479.x} } \bib{PoelkePolthier2012}{article}{ title={Domain Coloring of Complex Functions}, subtitle={An Implementation-Oriented Introduction}, author={Poelke, Konstantin}, author={Polthier, Konrad}, journal={IEEE Computer Graphics and Applications}, pages={90--97}, volume={32}, number={5}, date={2012}, doi={10.1109/MCG.2012.100} } \bib{PressTeukolskyVetterlingFlannery2007}{book}{ title={Numerical Recipes}, subtitle={The Art of Scientific Computing}, author={Press, William H.}, author={Teukolsky, Saul A.}, author={Vetterling, William T.}, author={Flannery, Brian P.}, edition={3}, publisher={Cambridge University Press}, address={New York}, date={2007} } \bib{Trott2008}{article}{ title={The Return of the Riemann Surface}, author={Trott, Michael}, journal={The Mathematica Journal}, volume={10}, number={4}, date={2008}, pages={626--656}, doi={10.3888/tmj.10.4-1} } \bib{Wegert2012}{book}{ title={Visual Complex Functions}, subtitle={An Introduction with Phase Portraits}, author={Wegert, Elias}, publisher={Birkhäuser}, address={Basel}, date={2012}, doi={10.1007/978-3-0348-0180-5} } \bib{Weierstrass1891}{article}{ title={Neuer Beweis des Satzes, dass jede ganze rationale Function einer Veränderlichen dargestellt werden kann als ein Product aus linearen Functionen derselben Veränderlichen}, author={Weierstraß, Karl}, date={1891}, journal={Sitzungsberichte der Königlich Preußischen Akademie der Wissenschaften zu Berlin}, volume={2}, pages={1085--1101} } \end{biblist} \end{bibdiv} \end{document}

https://arxiv.org/abs/1203.1462

A simple proof of the Fundamental Theorem of Calculus for the Lebesgue integral

This paper contains a new elementary proof of the Fundamental Theorem of Calculus for the Lebesgue integral. The hardest part of our proof simply concerns the convergence in ${\rm L}^1$ of a certain sequence of step functions, and we prove it using only basic elements from Lebesgue integration theory.

\section{Introduction} Let $f:[a,b]\longrightarrow {{\Bbb R}}$ be absolutely continuous on $[a,b]$, i.e., for every $\varepsilon>0$ there exists $\delta>0$ such that if $\{(a_j,b_j)\}_{j=1}^{n}$ is a family of pairwise disjoint subintervals of $[a,b]$ satisfying $$\sum_{j=1}^{n}{(b_j-a_j)}<\delta$$ then $$\sum_{j=1}^{n}{|f(b_j)-f(a_j)|}<\varepsilon.$$ Classical results ensure that $f$ has a finite derivative almost everywhere in $I=[a,b]$, and that $f'\in {\rm L}^1(I)$, see \cite{bot} or \cite[Corollary 6.83]{str}. These results, which we shall use in this paper, are the first steps in the proof of the main connection between absolute continuity and Lebesgue integration: the Fundamental Theorem of Calculus for the Lebesgue integral. \begin{theorem} \label{tfc} If $f:I=[a,b] \longrightarrow {{\Bbb R}}$ is absolutely continuous on $I$ then $$f(b)-f(a)=\int_{a}^b{f'(x) \, dx} \quad \mbox{in Lebesgue's sense.}$$ \end{theorem} In this note we present a new elementary proof to Theorem \ref{tfc} which seems more natural and easy than the existing ones. Indeed, our proof can be sketched simply as follows: \begin{enumerate} \item We consider a well--known sequence of step functions $\{h_n\}_{n \in {\scriptsize {\Bbb N}}}$ which tends to $f'$ almost everywhere in $I$ and, moreover, $$\int_a^b{h_n(x) \, dx}=f(b)-f(a) \quad \mbox{for all $n \in {{\Bbb N}}$.}$$ \item We prove, by means of elementary arguments, that $$\lim_{n \to \infty}\int_a^b{h_n(x) \, dx}=\int_a^b{f'(x) \, dx}.$$ \end{enumerate} More precise comparison with the literature on Theorem \ref{tfc} and its several proofs will be given in Section 3. In the sequel $m$ stands for the Lebesgue measure in ${{\Bbb R}}$. \section{Proof of Theorem \ref{tfc}} For each $n \in {{\Bbb N}}$ we consider the partition of the interval $I=[a,b]$ which divides it into $2^n$ subintervals of length $(b-a)2^{-n}$, namely $$x_{n,0}<x_{n,1}<x_{n,2}<\cdots<x_{n,2^n},$$ where $x_{n,i}=a+i(b-a)2^{-n}$ for $i=0,1,2,\dots, 2^n$. Now we construct a step function $h_n:[a,b) \longrightarrow {{\Bbb R}}$ as follows: for each $x \in [a,b)$ there is a unique $i \in \{0,1,2,\dots, 2^n-1\}$ such that $$x \in [x_{n,i},x_{n,i+1}),$$ and we define $$h_n(x)=\dfrac{f(x_{n,i+1})-f(x_{n,i})}{x_{n,i+1}-x_{n,i}}=\dfrac{2^n}{b-a}[f(x_{n,i+1})-f(x_{n,i})].$$ On the one hand, the construction of $\{h_n\}_{n \in {\scriptsize {\Bbb N}}}$ implies that \begin{equation} \label{derlim} \lim_{n \to \infty}{h_n(x)}=f'(x) \quad \mbox{for all $x \in [a,b) \setminus N$,} \end{equation} where $N \subset I$ is a null--measure set such that $f'(x)$ exists for all $x \in I \setminus N$. On the other hand, for each $n \in {{\Bbb N}}$ we compute \begin{equation} \nonumber \int_a^b{h_n(x) dx}=\sum_{i=0}^{2^n-1}{\int_{x_{n,i}}^{x_{n,i+1}}{h_n(x) dx}}=\sum_{i=0}^{2^n-1}[f(x_{n,i+1})-f(x_{n,i})]=f(b)-f(a), \end{equation} and therefore it only remains to prove that $$\lim_{n \to \infty}\int_a^b{h_n(x) \, dx}=\int_a^b{f'(x) \, dx}.$$ Let us prove that, in fact, we have convergence in ${\rm L}^1(I)$, i.e., \begin{equation} \label{goal} \lim_{n \to \infty}\int_a^b{|h_n(x)-f'(x)| \, dx}=0. \end{equation} Let $\varepsilon>0$ be fixed and let $\delta>0$ be one of the values corresponding to $\varepsilon/4$ in the definition of absolute continuity of $f$. Since $f' \in {\rm L}^1(I)$ we can find $\rho>0$ such that for any measurable set $E \subset I$ we have \begin{equation} \label{eq1} \int_E{|f'(x)| \, dx}<\dfrac{\varepsilon}{4} \quad \mbox{whenever $m(E)<\rho$.} \end{equation} The following lemma will give us fine estimates for the integrals when $|h_n|$ is ``small". We postpone its proof for better readability. \begin{lemma} \label{le1} For each $\varepsilon>0$ there exist $k, n_k \in {{\Bbb N}}$ such that $$k \cdot m \left( \left\{x \in I \, : \, \sup_{n \ge n_k}|h_n(x)|>k \right\} \right)<\varepsilon.$$ \end{lemma} Lemma \ref{le1} guarantees that there exist $k, n_k \in {{\Bbb N}}$ such that \begin{equation} \label{etdos} k \cdot m \left( \left\{x \in I \, : \, \sup_{n \ge n_k}|h_n(x)|>k \right\} \right)<\min\left\{\delta,\dfrac{\varepsilon}{4}, \rho \right\}. \end{equation} Let us denote $$A=\left\{x \in I \, : \, \sup_{n \ge n_k}|h_n(x)|>k \right\},$$ which, by virtue of (\ref{etdos}) and (\ref{eq1}), satisfies the following properties: \begin{eqnarray} \label{intdos} m(A) <\delta, \\ \label{et2} k\cdot m(A)<\dfrac{\varepsilon}{4},\\ \label{int1} \int_{A}{|f'(x)| \, dx}<\dfrac{\varepsilon}{4}. \end{eqnarray} We are now in a position to prove that the integrals in (\ref{goal}) are smaller than $\varepsilon$ for all sufficiently large values of $n \in {{\Bbb N}}$. We start by noticing that (\ref{int1}) guarantees that for all $n \in {{\Bbb N}}$ we have \begin{align} \nonumber \int_{I}{|h_n(x)-f'(x)| \, dx}&=\int_{I \setminus A}{|h_n(x)-f'(x)| \, dx}+\int_{ A}{|h_n(x)-f'(x)| \, dx}\\ \label{int2} &<\int_{I \setminus A}{|h_n(x)-f'(x)| \, dx}+\int_{ A}{|h_n(x)| \, dx}+\dfrac{\varepsilon}{4}. \end{align} The definition of the set $A$ implies that for all $n \in {{\Bbb N}}$, $n \ge n_k$, we have $$|h_n(x)-f'(x)| \le k+|f'(x)| \quad \mbox{for almost all $x \in I \setminus A$},$$ so the Dominated Convergence Theorem yields \begin{equation} \label{et3} \lim_{n \to \infty}\int_{I \setminus A}{|h_n(x)-f'(x)| \, dx}=0. \end{equation} From (\ref{int2}) and (\ref{et3}) we deduce that there exists $n_{\varepsilon} \in {{\Bbb N}}$, $n_{\varepsilon} \ge n_k$, such that for all $n \in {{\Bbb N}}$, $n \ge n_{\varepsilon}$, we have \begin{equation} \label{int3} \int_{I}{|h_n(x)-f'(x)| \, dx}<\dfrac{\varepsilon}{2}+\int_{ A}{|h_n(x)| \, dx}. \end{equation} Finally, we estimate $\int_A{|h_n|}$ for each fixed $n \in {{\Bbb N}}$, $n \ge n_{\varepsilon}$. First, we decompose $A=B \cup C$, where $$B=\{ x \in A \, : \, |h_n(x)| \le k\} \quad \mbox{and} \quad C=A \setminus B.$$ We immediately have \begin{equation} \label{int4} \int_B{|h_n(x)| \, dx} \le k \cdot m(B) \le k \cdot m(A) < \dfrac{\varepsilon}{4} \quad \mbox{by (\ref{et2}).} \end{equation} Obviously, $\int_C{|h_n|}<\varepsilon/4$ when $C=�\varnothing$. Let us see that this inequality holds true when $C \neq \varnothing$. For every $x \in C=\{ x \in A \, : \, |h_n(x)| >k \}$ there is a unique index $i \in \{0,1,2,\dots, 2^n-1\}$ such that $x \in [x_{n,i},x_{n,i+1})$. Since $|h_n|$ is constant on $[x_{n,i},x_{n,i+1})$ we deduce that $[x_{n,i},x_{n,i+1}) \subset C$. Thus there exist indexes $i_l \in \{0,1,2,\dots, 2^n-1\}$, with $l=1,2,\dots, p$ and $i_l \ne i_{\tilde l}$ if $l \neq \tilde l$, such that $$C=\bigcup_{l=1}^{p}[x_{n,i_l},x_{n,i_l+1}).$$ Therefore $$\sum_{l=1}^{p}(x_{n,i_l+1}-x_{n,i_l})=m(C) \le m(A) < \delta \quad \mbox{by (\ref{intdos}),}$$ and then the absolute continuity of $f$ finally comes into action: \begin{align*} \int_C{|h_n(x)| \, dx}&=\sum_{l=1}^{p}\int_{x_{n,i_l}}^{x_{n,i_l+1}}{|h_n(x)| \, dx} \\ &=\sum_{l=1}^{p}|f(x_{n,i_l+1})-f(x_{n,i_l})|< \dfrac{\varepsilon}{4}. \end{align*} This inequality, along with (\ref{int3}) and (\ref{int4}), guarantee that for all $n \in {{\Bbb N}}$, $n \ge n_{\varepsilon}$, we have $$\int_{I}{|h_n(x)-f'(x)| \, dx}<\varepsilon,$$ thus proving (\ref{goal}) because $\varepsilon$ was arbitrary. \hbox to 0pt{}\hfill$\rlap{$\sqcap$}\sqcup$\medbreak \noindent {\bf Proof of Lemma \ref{le1}.} Let $\varepsilon>0$ be fixed and let $\rho>0$ be such that for every measurable set $E \subset I$ with $m(E)< \rho$ we have $$\int_{E}{|f'(x)| \, dx}< \dfrac{\varepsilon}{2}.$$ Let $N \subset I$ be as in (\ref{derlim}) and let $k \in {{\Bbb N}}$ be sufficiently large so that $$m \left( \left\{x \in I \setminus N \, : \, |f'(x)| \ge k \right\}\right) < \rho,$$ which implies that \begin{equation} \label{l1} k \cdot m \left( \left\{x \in I \setminus N \, : \, |f'(x)| \ge k \right\}\right) \le \int_{\left\{x \in I \setminus N \, : \, |f'(x)| \ge k \right\}}{|f'(x)| \, dx}<\dfrac{\varepsilon}{2}. \end{equation} Let us define $$E_j=\left\{ x \in I \setminus N \, : \, \sup_{n \ge j}|h_n(x)| >k \right\} \quad (j \in {{\Bbb N}}).$$ Notice that $E_{j+1}\subset E_j$ for every $j \in {{\Bbb N}}$, and $m(E_1)<\infty$, hence \begin{equation} \label{pl1} \lim_{j \to \infty}m(E_j)=m \left( \bigcap_{j=1}^{\infty}E_j\right). \end{equation} Clearly, $\cap_{j=1}^{\infty}E_j \subset \left\{x \in I \setminus N \, : \, |f'(x)| \ge k \right\}$, so we deduce from (\ref{pl1}) that we can find some $n_k \in {{\Bbb N}}$ such that $$m(E_{n_k}) \le m \left( \left\{x \in I \setminus N \, : \, |f'(x)| \ge k \right\}\right) +\dfrac{\varepsilon}{2k},$$ and then (\ref{l1}) yields $k \cdot m(E_{n_k}) < \varepsilon$. \hbox to 0pt{}\hfill$\rlap{$\sqcap$}\sqcup$\medbreak \section{Final remarks} The sequence $\{h_n\}_{n \in {\scriptsize {\Bbb N}}}$ is used in other proofs of Theorem \ref{tfc}, see \cite{bar} or \cite{vol}. The novelty in this paper is our elementary and self--contained proof of (\ref{goal}). Incidentally, a revision of the proof of our Lemma \ref{le1} shows that it holds true for any sequence of measurable functions $\tilde h_n: E \subset {{\Bbb R}} \longrightarrow {{\Bbb R}}$ which converges pointwise almost everywhere to some $h \in {\rm L}^1(E)$ and $m(E) < \infty$. Our proof avoids somewhat technical results often invoked to prove Theorem \ref{tfc}. For instance, we do not use any sophisticated estimate for the measure of image sets such as \cite[Theorem 7.20]{bbt}, \cite[Lemma 6.88]{str} or \cite[Proposition 1.2]{vol}, see also \cite{kol}. We do not use the following standard lemma either: an absolutely continuous function having zero derivative almost everywhere is constant, see \cite[Theorem 7.16]{bbt} or \cite[Lemma 6.89]{str}. It is worth having a look at \cite{gor} for a proof of that lemma using tagged partitions; see also \cite{bot1} for a proof based on full covers \cite{tho}. Concise proofs of Theorem \ref{tfc} follow from the Radon--Nikodym Theorem, see \cite{bar}, \cite{bbt} or \cite{rud}, but this is far from being elementary. Finally, it is interesting to note that (\ref{goal}) easily follows from the Dominated Convergence Theorem when $f$ is Lipschitz continuous on $I$. This fact made the author think about the following project for students in an introductory course to Lebesgue integration. \begin{center} {\sc Project: Two important results for the price of one.} \end{center} \begin{enumerate} \item Let $f:I=[a,b] \longrightarrow {{\Bbb R}}$ be Lispchitz continuous on $I$. A deep result (worth to know without proof) guarantees that $f'(x)$ exists for almost all $x \in I$, see \cite[Theorem 7.8]{bbt}. Consider the sequence $\{h_n\}_{n \in {\scriptsize {\Bbb N}}}$ as defined in Section 2 and prove \begin{enumerate} \item $\{h_n (x) \}_{n \in {\scriptsize {\Bbb N}}}$ tends to $f'(x)$ for almost all $x \in I$; \item $\int_I{h_n(x) \, dx}=f(b)-f(a)$ for all $n \in {{\Bbb N}}$; \item (Use the Dominated Convergence Theorem) $f' \in L^1(I)$ and $$f(b)-f(a)=\int_a^b{f'(x) \, dx} \quad \mbox{in Lebesgue's sense.}$$ \end{enumerate} \item Let $g:I=[a,b] \longrightarrow {{\Bbb R}}$ be Riemann--integrable on $I$ and define $$f(x)=(R)\int_a^x{g(s) \, ds} \quad (x \in I),$$ where $(R) \int$ stands for the Riemann integral. Use the information in Exercise 1 to deduce that $g \in {\rm L}^1(I)$ and $$(R) \int_a^b{g(x) \, dx}=\int_a^b{g(x) \, dx} \quad \mbox{in Lebesgue's sense.}$$ \end{enumerate}

https://arxiv.org/abs/2003.03197

Bounding probability of small deviation on sum of independent random variables: Combination of moment approach and Berry-Esseen theorem

In the context of bounding probability of small deviation, there are limited general tools. However, such bounds have been widely applied in graph theory and inventory management. We introduce a common approach to substantially sharpen such inequality bounds by combining the semidefinite optimization approach of moments problem and the Berry-Esseen theorem. As an application, we improve the lower bound of Feige's conjecture from 0.14 to 0.1798.

\section{Introduction} \label{} The problem of upper bounding \begin{equation} \label{sum version} \text{Prob}\left[ \sum_{i=1}^n X_i \geq \sum_{i=1}^n \mathop{\mathbb{E}}(X_i) + \delta \right]. \end{equation} for independent random variables $X_i$ and a given constant $\delta>0$, has been studied for years. Many classic tail bounds of this type, such as Markov's inequality, Chebyshev's inequality \cite{lin2011probability}, Hoeffding's inequality \cite{hoeffding1963}, Bennett's inequality \cite{bennett1962probability} and Bernstein's inequality \cite{bernstein1924modification} and applications have been well-studied in literature and textbooks \cite{boucheron2013concentration,lin2011probability}. However, those inequalities are designed when $\delta$ is large. For example, Hoeffding's inequality indicates the following: \begin{equation*} \text{Prob}\left[ \sum_{i=1}^n X_i \ge \sum_{i=1}^n \mathop{\mathbb{E}}(X_i) + \delta \right] \le e^{-\frac{2\delta^{2}}{\sum_{i=1}^n (b_i - a_i)^{2}}}. \end{equation*} where $X_i \in [a_i, b_i]$. If $\delta$ is a relatively small constant, this inequality only provides bounds that are not so sharp. In particular, when $\delta$ is $0$, it yields a trivial bound. \begin{table*} \caption{: Feige's bound in literature and our paper.} \label{tab1} \centering \setlength{\tabcolsep}{9mm}{ \begin{tabular}{cccc} \hline Name&Bound&Method&Theorem\\ \hline Feige \cite{Feige04onsums}& $\frac{1}{13}$&\\ He \cite{he2010}&0.125& approximate MP(1,2,4) &\\ Garnett \cite{GARNETT2020105119}& 0.14 & approximate MP(1,2,3,4) & \\ \multirow{4}*{Our paper}&0.1536& approximate MP(1,2,4) and B-E& Theorem \ref{thm3}\\ &0.1541& MP(1,2,4) and B-E & Theorem \ref{thm2}\\ &0.1587& MP(1,2,3,4) and B-E& Theorem \ref{thm4}\\ &0.1798& MP(1,2,3,4) and B-E with refinement& Theorem \ref{thm5}\\ \hline \end{tabular}} \end{table*} In this context of small deviation, which is widely applied in graph theory \cite{Feige04onsums} and inventory management \cite{wang2015process}, there are limited general tools to derive such bound. One approach is to formulate this problem as a moments problem (MP). Given the moments information, we can further derive an equivalent semidefinite programing (SDP) problem to the original moments problem through duality theory \cite{bertsimas2005optimal} and sum-of-square technique \cite{Lasserre07}, based on the classical theorem established by the great mathematician Hilbert in 1888 as the univariate case of Hilbert's 17th problem, which states that an univariate polynomial is nonnegative if and only it can be represented as sum of squares of polynomials. It is also worth to mention that Berry-Esseen theorem, a uniform bound between the cdf of sum of independent random variables and the cdf of standard normal distribution, is a quite powerful inequality, no matter $\delta$ is large or small. Specifically, we can define a random variable $S=\frac{\sum_{i=1}^n X_i}{\sqrt{\sum_{i=1}^n \mathop{\mathbb{E}}[X_i^2]}}$ with $\text{Var}[S]=1$ to represent the normalized sum, and a third moment bound $\psi_0=\frac{\sum_{i=1}^n \mathop{\mathbb{E}}[|X_i|^3]}{(\sum_{i=1}^n \mathop{\mathbb{E}}[X_i^2])^{3/2}}$. Then Berry-Esseen theorem indicates the following: \begin{equation} \sup_{x \in \mathbb{R}} |F(x)-\Phi(x)| \leq c_0 \psi_0 \end{equation} where $F(x)$ and $\Phi(x)$ are the cdf of $S$ and standard normal distribution respectively with best known $c_0=0.56$ \cite{Shevtsova2010An}. Unsurprisingly, if all random variables are independent identically distributed, then central limit theorem states that their properly scaled sum tends to towards a normal distribution. Berry-Esseen theorem just provides a quantitative rate of convergence. Therefore, it is a nature idea to combine moment approach and Berry-Eseen theorem to achieve a better bound of small deviation problems. As an application to better illustrate the way of combination, Feige\cite{Feige04onsums} first established a bound $\alpha=\frac{1}{13}$ and conjectured the true bound to be $\alpha=\frac{1}{e}$ of the following small deviation problem: \begin{equation} \label{Feige} \text{Prob}\left[ \sum_{i=1}^n X_i \geq 1 \right] \leq 1-\alpha > 0 \end{equation} where $X_1,X_2,...,X_n$ are independent random variables, with $\mathop{\mathbb{E}}[X_i]=0$ and $X_i \geq -1$ for each $i$. This inequality has many applications in the field of graph theory \cite{ferber2019uniformity}, combinatorics \cite{Alon2012}, and evolutionary algorithms \cite{Corus2014}. One contribution of this paper is to improve the Feige's bound from best-known $\alpha=0.14$ to $0.1798$ step by step as it is shown in Table \ref{tab1}. The other contribution of this paper is to introduce a general approach to bound probability of small deviation by merging the moment problem approach and the Berry-Esseen Theorem. Suppose we have a sequence of independent random variables $X_i$ and a constant $\delta$, and define $\sum_{i=1}^n \mathop{\mathbb{E}}(X_i^2)=D$. Here is the guideline of our approach. \begin{itemize} \item In order to achieve the upper bound of (\ref{sum version}), without loss of generality, we can assume $X_i$ has support set of at most $k$ discrete points where $k$ is the number of given moment information (including the trivial $0$-th order moment) on $\sum X_i$ by constructing an associated linear programming. This insight is an extension of lemma 6 in Feige \cite{Feige04onsums}, and has been established by Bertismas et. al. \cite{bertsimas2005optimal}. \item For both the Berry-Esseen theorem and moment approach, it requires the distributions $X_i$ has bounded support, i.e., there exists a certain constant $K$ such that $|X_i| \leq K$ for all $i$. When the distributions are not bounded, we divide the distributions into bounded and unbounded groups, and treats the unbounded group separately as in \cite{he2010}. \item We can derive a bound of (\ref{sum version}) by Berry-Esseen theorem. Suppose we have $|X_i| \leq K$ for all $i$. Then $\psi_0 =\frac{\sum_{i=1}^n \mathop{\mathbb{E}}[|X_i^3|]}{D^{3/2}}\leq \frac{K}{D^{1/2}}$. When $D$ is large, it follows that $\psi_0$ is relatively small, and therefore Berry-Esseen theorem can provide a rather tight bound. \item We can also bound (\ref{sum version}) by the moment approach. In particular, when the distributions are bounded, we can bound the third moment or above through the second moment $D$. Theorem \ref{more} indicates the more moments we use, the better bound we can achieve. In addition, when $D$ is small, the bound of moment approach is often better as it is shown in theorem \ref{thmMono}. \item We observe that often the worse-case scenario of these two approaches do not agree with each other, which provides us a great opportunity to merge these two methods together to further improve the bound estimation. In Section 3 and 4, we discuss how to synthetically merge these two approaches together to achieve better results. \end{itemize} \section{An SDP formulation of the moment problem} In this section, we introduce the classical SDP formulation of the moment problem. Supposing $X$ is a real random variable and $P$ is a set of given moments, then we formulate moments problem as the following. \begin{equation}\label{MP} \begin{aligned} \max_X & &\text{Prob}[X \geq 0] & \\ \text{subject to} & & \mathop{\mathbb{E}}[X^i]=M_i, & \text{ where } i \in P \end{aligned} \end{equation} Note that $M_0=1$. The corresponding dual problem is \begin{equation}\label{SPD} \begin{aligned} \min_y \,\,\, & & \sum_{i \in P} y_i M_i & \\ \text{subject to} & & \sum_{i \in P} y_i x^i \geq \text{\bf 1}_{x \geq 0}, & \,\,\,\text{ for all } x \in R \end{aligned} \end{equation} where $\text{\bf 1}_{x \geq 0}$ is an indicator function. This is an well-studied optimization problem and the dual formulation was first established in \cite{isii1960extrema} and treated extensively in \cite{bertsimas2005optimal}. In fact, the property of strong duality was shown in \cite[Theorem 2.2]{bertsimas2005optimal}. Moreover, the dual constraint requires the polynomial function to be nonnegative, which is equivalent to certain matrices being positive semidefinite, as it is shown in the following theorem. \begin{theorem}\cite[Section 3.a]{reznick2000some}\label{sos} \noindent A real polynomial function $f(x) =\sum_{r=0}^{2n} y_rx^r$ is nonnegative if and only if $f(x) = g(x)^2 + h(x)^2$ for some polynomial function $g(x)$ and $h(x)$. Furthermore, the nonnegativity of $f(x)$ implies the existence of an $(n + 1) \times (n + 1)$ positive semidefinite matrix $V$ such that $f(x) = X^TVX$ with $X = (1,x,x^2,...x^n)^{T}$. \end{theorem} \noindent{\bf Proof.} (SOS Decomposition): If $f(x)=g(x)^2+h(x)^2$, then it is obviously nonnegative. If $f(x)$ is nonnegative, then all real roots of $f(x)$ are of even multipliers, because otherwise $f(x)$ will be negative locally. By the Fundamental Theorem of Algebra, $$f(x)=\prod_{j=1}^m (x - r_j)^{2n_j}\prod_{i=1}^{n-\sum_{j=1}^{m} n_j}(x - z_i)(x - \overline z_i).$$ Notice that $(x - z_i)(x - \overline z_i) = (x - a_i)^2 + b_i^2$, $f(x)$ can be decomposed as products of sum square of two functions. Since $(a^2 +b^2)(c^2+d^2) = (ac+bd)^2 + (ad-bc)^2$, then we can reformulate $f$ into $g^2 + h^2$. (SDP Representation): Suppose $f(x) = g(x)^2 + h(x)^2$, and the coefficient vector of $g$ and $h$ are $u$ and $v$, respectively. Then $g(x) = (1,x,x^2,...x^n)u$, and $h(x) = (1,x,x^2,...x^n)v$. Let $V=uu^T + vv^T$, and we have $f(x) = g(x)^2 + h(x)^2 = X^Tuu^TX + X^Tvv^TX = X^T(uu^T+vv^T)X = X^TVX$. \qed Theorem 2.1 is the univariate case for Hilbert’s 17th problem. In fact, this theorem together with the further work by Lasserre \cite{Lasserre07} can help us transform the dual problem into a semidefinite program, as it is shown in the following proposition. \begin{prop}\cite[Proposition 3.1]{bertsimas2005optimal} \label{Prop} \begin{itemize} \item The polynomial $g(x)=\sum_{r=0}^{2n}y_rx^r$ satisfies $g(x) \geq 0$ for all $x \in R$ \text{\bf if and only if} there exists a positive semidefinite matrix $V=[v]_{i,j=0,1,...,n}$ such that $$y_r= \sum_{i,j:i+j=r}v_{ij}, \,\, r=0,...,2n $$ \item The polynomial $g(x)=\sum_{r=0}^{n}y_rx^r$ satisfies $g(x) \geq 0$ for all $x \geq 0$ \text{\bf if and only if} there exists a positive semidefinite matrix $V=[v]_{i,j=0,1,...,n}$ such that $$0= \sum_{i,j:i+j=2l-1}v_{ij}, \,\, l=1,...,n $$ $$y_r= \sum_{i,j:i+j=2l}v_{ij}, \,\, l=0,...,n $$ \end{itemize} \end{prop} In all, moments problem (\ref{MP}) can be solved by its SDP formulation. One key property of the moments problem is to achieve a better bound by taking advantage of additional moment information, as extra moment information yields a more restrictive constraint set in the moment problem. \begin{theorem} \label{more} Given a real random variable $X$, consider the primal problem \begin{equation*} \begin{aligned} \text{\bf opt }(P)=\max_X & \,\,\,\,\, \text{Prob}[X \geq 0] & \\ \text{subject to} & \,\,\,\,\, \mathop{\mathbb{E}}[X^i]=M_i, & \text{ where } i \in P \end{aligned} \end{equation*} Supposing we have $P_1 \subset P_2$, then $\text{\bf opt }(P_1) \geq \text{\bf opt }(P_2)$. \end{theorem} Similarly, the following theorem explores the monotonicity of the moments problem with upper and lower bounds, as the feasible region enlarges as $D$ grows larger. \begin{theorem} \label{thmMono} Given a real random variable $X$ and mutually exclusive sets $P_1, P_2, P_3$, consider the primal problem where $B_i$ and $L_i$ are the upper and lower of moments in a function of a real number $D$. \begin{align*} \text{\bf opt }(D)= \max_X & &\text{Prob}[X \geq 0] & \\ \text{subject to} & & \mathop{\mathbb{E}}[X^i]=M_i, & \text{ where } i \in P_1 \\ & & \mathop{\mathbb{E}}[X^i] \leq B_i(D), & \text{ where } i \in P_2 \\ & & \mathop{\mathbb{E}}[X^i] \geq L_i(D), & \text{ where } i \in P_3 \end{align*} Supposing we have \begin{itemize} \item $B_i(D) \geq 0$ and $B_i(D)$ is an increasing function in $D$ for all $i \in P_2$; \item $L_i(D) \leq 0$ and $L_i(D)$ is an decreasing function in $D$ for all $i \in P_3$, \end{itemize} then $\text{\bf opt }(D)$ is an monotonically increasing function in $D$. \end{theorem} \section{A combination of moment approach and Berry-Esseen theorem} \begin{theorem} \label{MPBE} Let $X=\sum_{i=1}^n X_i$, and $D=Var(X)=\sum_{i=1}^n \mathop{\mathbb{E}}[ X_i]^2$ for independent random variables $X_i$ with bound $|X_i| \leq K$. In addition, suppose there exist mutually exclusive sets $P_1=\{0,1,2\}$, $P_2$, $P_3$ with increasing nonnegative functions $B_i(D)$ for $i \in P_2$ and decreasing nonpositive functions $L_i(D)$ for $i \in P_3$. Then $$\text{Prob}[X \geq 0] \leq \min_{D > 0}\max\{\text{\bf opt}(D), F(D)\}$$ where $$F(D)=0.5+0.56 \frac{K}{D^{1/2}}$$ and \begin{align*} \text{\bf opt }(D)= \max_X & \,\,\,\, \text{Prob}[X \geq 0] & \\ \text{subject to} \,\,\,\, & \mathop{\mathbb{E}}[X^0]=1, & \\ & \mathop{\mathbb{E}}[X]=\sum_{i=1}^n \mathop{\mathbb{E}}[X_i], & \\ & \mathop{\mathbb{E}}[X^2]=D, & \\ & \mathop{\mathbb{E}}[X^i] \leq B_i(D), & \text{ where } i \in P_2 \\ & \mathop{\mathbb{E}}[X^i] \geq L_i(D), & \text{ where } i \in P_3 \end{align*} Moreover, $\arg\min_{D > 0}\max\{\text{\bf opt}(D), F(D)\}$ is at the intersection of function $\text{\bf opt}(D)$ and function $F(D)$, if it exists. \end{theorem} \noindent{\bf Proof.} We can apply Berry-Esseen theorem on $\text{Prob}[X \geq 0]$. \begin{align*} \text{Prob}[X \geq 0] & \leq 1 - \Phi(0) + c_0 \psi_0 \\ & \leq 0.5 + 0.56 \frac{\sum_{i=1}^n |X_i|^3}{D^{3/2}} \\ & \leq 0.5+0.56 \frac{K}{D^{1/2}} = F(D) \end{align*} When $D$ is large, Berry-Esseen theorem is effective, because $F(D)$ is a decreasing function. When $D$ is small, the moments problem performs well because $\text{\bf opt}(D)$ is an increasing function by theorem \ref{thmMono}. Therefore, if $\text{\bf opt}(D)$ and $F(D)$ exists an intersection, then it is the optimal solution of $\min_{D > 0}\max\{\text{\bf opt}(D), F(D)\}$ due to the monotonicity of these two functions. \qed \section{Example: improve the bound of Feige's inequality} In this section, we will show that a combination of moment approach and Berry-Essen theorem can improve Feige’s bound. As we see, Feige's conjecture (\ref{Feige}) has no assumptions on the upper bound of each random variable, though we know the lower bound is $-1$. The following theorem \ref{thm1} allows us to transform such variables $X_i$ into a group of corresponding $Y_i$ with both upper and lower bounds, through truncating the sufficiently large negative part of $X_i$ and rescale the rest. Similar technique was used in \cite{he2010, GARNETT2020105119}. In this way, we can apply the inequalities of sum of independent random variables with both upper bound and lower bound, as theorem \ref{MPBE} indicates. \begin{theorem}\label{thm1} Supposing for $m$ random variables $Y_1$, $Y_2$,...,$Y_m$ with mean zero and $ -\xi \leq Y_i \leq 1$ for some fixed $0 < \xi \leq 1$, there exists an universal bound $\omega>0$ independent of $m$ and the choice of $Y_i$ such that $$\text{Prob}[\sum_{i=1}^m Y_i \leq \xi] \geq \omega$$. Then, consider $n$ random variables $X_1$, $X_2$,...,$X_n$ with mean zero and $X_i \geq -1$. $$\text{Prob}[\sum_{i=1}^n X_i <1] \geq e^{- {\xi} } \cdot \omega$$ \end{theorem} \noindent{\bf Proof.} As it is shown in \cite{Feige04onsums}, without loss of generality, we can assume $X_i$ follows a two-point distribution. Therefore, we can assume that there exists $0<a_i \leq 1$ and $b_i >0$ such that $$X_i=\left\{ \begin{array}{c l} -a_i \text{ with probability } \frac{b_i}{a_i+b_i}\\ \\ b_i \text{ with probability } \frac{a_i}{a_i+b_i} \end{array}\right.$$ given $E[X_i]=0$. Suppose $b_1 \geq b_2 \geq ... \geq b_n$. Then we consider to make a partition and define $A=\{1,2,...,N\}$ and $B=\{N+1,...,n\}$ by a fixed number $\tau>1$ where $$N=\max\{\,0,\,\max\{\,k \,|\, b_k \geq \tau (\sum_{i=1}^k a_i), 1\leq k \leq n\}\}$$ Define $a=\sum_{i=1}^k a_i$ and we have \[b_i \geq b_N \geq \tau a , \text{ for every } i \leq N\] \[b_i \leq b_{N+1} \leq \tau(a+a_{N+1}) \leq \tau(a+1) , \text{ for every } i > N\] If $N >0$, then \begin{align*} \text{Prob}[\sum_{i=1}^N X_i=-a ] &= \Pi_{i=1}^N \text{Prob}[X_i=-a_i]\\ &=\Pi_{i=1}^N (1-\frac{a_i}{a_i+b_i}) \geq \Pi_{i=1}^N (1-\frac{a_i}{a_i+\tau a})\\ & \geq \Pi_{i=1}^N e^{-\frac{a_i}{\tau a}}=e^{-\frac{1}{\tau}} \end{align*} Therefore, \begin{align*} \text{Prob}[\sum_{i=1}^n X_i < 1 ] & \geq \text{Prob}[\sum_{i \in A} X_i=-a ] \text{Prob}[\sum_{i \in B} X_i < a+1 ]\\ & \geq e^{-\frac{1}{\tau}} \text{Prob}[\sum_{i \in B} X_i < a+1 ] \end{align*} Let $Y_i=\frac{1}{\tau} \frac{X_i}{a+1}$ for $i \in B$. Note that $$Y_i =\frac{1}{\tau} \frac{X_i}{a+1} \leq \frac{1}{\tau} \frac{\tau (a+1)}{a+1}=1$$ and $$Y_i =\frac{1}{\tau} \frac{X_i}{a+1} \geq -\frac{1}{\tau}.$$ Then, if we set $\xi=\frac{1}{\tau}$, \begin{align*} \text{Prob}[\sum_{i=1}^n X_i < 1 ] & \geq e^{-\frac{1}{\tau}} \text{Prob}[ \sum_{i \in B} X_i < a+1 ] \\ &= e^{-\frac{1}{\tau}} \text{Prob}[ \sum_{i \in B} Y_i < \frac{1}{\tau} ] \\ & \geq e^{- \xi} \cdot \omega \end{align*} \qed For the rest of work, we will consider the following problem: Let $Y_1,Y_2,...,Y_n$ be independent random variable with mean zero and $-\xi \leq Y_i \leq 1$ for some $0< \xi \leq 1$. Without loss of any generality, we can assume \begin{equation} \label{setupY} Y_i=\left\{ \begin{array}{c l} -a_i \text{ with probability } \frac{b_i}{a_i+b_i}\\ \\ b_i \text{ with probability } \frac{a_i}{a_i+b_i} \end{array}\right. \end{equation} where $0 \leq a_i \leq \xi$ and $0 \leq b_i \leq 1$. Then for any $n$, we are interested in the lower bound of $$\text{Prob}[\sum_{i=1}^n Y_i \leq \xi] , $$ as a key to improve Feige's bound $\alpha$ in (\ref{Feige}). \subsection{Grouping the first, second and fourth moment information} \label{124} When it comes to Feige's bound (\ref{Feige}), He and et al. improved it to $1/8$ by solving the moments problem with the first, second and fourth moment information. Therefore, we only consider the same moment information in this subsection as a fair comparison. Suppose we have $Y_1,Y_2,...,Y_n$ be independent random variables with mean zero and $-\xi \leq Y_i \leq 1$ for some $0< \xi \leq 1$, as it is stated in (\ref{setupY}). Let $Y=\sum_{i=1}^n Y_i$ and $D=Var(Y)=\sum_{i=1}^n E(Y_i^2)=\sum_{i=1}^n a_ib_i$. Berry-Esseen theorem implies the following: \begin{align*} \text{Prob}[ Y \leq \xi ] &=\text{Prob}[ \frac{Y}{\sqrt{D}} \leq \frac{\xi}{\sqrt{D}} ] \\ & \geq \Phi(\frac{\xi}{\sqrt{D}}) - c_0 \psi_0 \\ & \geq \Phi(\frac{\xi}{\sqrt{D}}) - 0.56 \frac{\sum_{i=1}^n E[|Y_i^3|]}{D^{3/2}} \\ & \geq \Phi(\frac{\xi}{\sqrt{D}}) - 0.56 \frac{\max_i\{|Y_i|\} \sum_{i=1}^n E[|Y_i^2|]}{D^{3/2}} \\ & = \Phi(\frac{\xi}{\sqrt{D}}) - 0.56 \frac{1}{\sqrt{D}} \end{align*} Define \begin{equation} \label{F1} F_1(\xi,D)=\Phi(\frac{\xi}{\sqrt{D}}) - 0.56 \frac{1}{\sqrt{D}} \end{equation} as a lower bound of $\text{Prob}[ Y \leq \xi ]$. For moments problem, we can define $Z=Y-\xi$. Then: \begin{itemize} \item $M_1=\mathop{\mathbb{E}}[Z]=-\xi$ \item $M_2=\mathop{\mathbb{E}}[Z^2]=D+\xi^2$ \item $M_4 = \mathop{\mathbb{E}}[Z^4] $ \begin{align*} &=3D^2+6\xi^2D+\xi^4+\sum_{i=1}^n(\mathop{\mathbb{E}}[Y_i^4]-4\xi \mathop{\mathbb{E}}[Y_i^3]-3(\mathop{\mathbb{E}}[Y_i^2])^2)\\ &=3D^2+6\xi^2D+\xi^4+\sum_{i=1}^n a_i b_i(a_i^2+b_i^2-4a_ib_i-4\xi(b_i-a_i)) \end{align*} \end{itemize} Since $a_i^2+b_i^2-4a_ib_i-4\xi(b_i-a_i)$ is a convex function of $a_i$ when $b_i$ and $\xi$ are fixed, and is a convex function of $b_i$ when $a_i$ and $\xi$ are fixed. Then supposing we fix $\xi$, the optimal solution of $$\max_{0 \leq a_i \leq \xi, 0 \leq b_i \leq 1} a_i^2+b_i^2-4a_ib_i-4\xi(b_i-a_i)$$ is in the set $\{(0,0), (\xi,0),(0,1),(\xi,1)\}$ with the optimal value $S(\xi)$. It follows that $$M_4 \leq 3D^2+6\xi^2D+\xi^4+S(\xi) D$$ Let $\text{\bf opt}(\xi,D)$ be the optimal value of the following moments problem given $\xi$ and $D$. \begin{equation*} \begin{aligned} \text{\bf opt}(\xi,D)=\max_X & \,\,\,\, \text{Prob}[Z \geq 0] & \\ \text{subject to} \,\,\,\, & \mathop{\mathbb{E}}[Z^0]=1, & \\ & \mathop{\mathbb{E}}[Z]=-\xi, & \\ & \mathop{\mathbb{E}}[Z^2]=D+\xi^2, & \\ & \mathop{\mathbb{E}}[Z^4] \leq 3D^2+6\xi^2D+\xi^4+S(\xi) D, & \end{aligned} \end{equation*} Define \begin{equation} \label{F2} F_2(\xi,D)=1- \text{\bf opt}(\xi,D) \end{equation} to be another lower bound of $\text{Prob}[ Y \leq \xi ]$. In addition, we can calculate $F_2(\xi,D)$ numerically by solving a corresponding SDP problem introduced in proposition \ref{Prop}, when $\xi$ and $D$ are fixed. \begin{theorem}\label{thm2} Let $X_1$, $X_2$,...,$X_n$ be n independent random variables with $E[X_i]=0$ and $X_i \ge -1$ for each $i$ and let $X = \sum_{i=1}^n X_i$, then $$\text{Prob}[X \le 1] \ge 0.1541.$$ \end{theorem} \noindent{\bf Proof.} Set $\xi=0.2$. Then $S(\xi) \leq 0.2$. \begin{itemize} \item If $D \geq 2.374$, then $$\text{Prob}[\sum_{i=1}^n X_i <1] \geq e^{-0.2}F_1(0.2, 2.374) >0.1541,$$ as $F_1(0.2,D)$ is an increasing function in $D$. \item If $D \leq 2.374$, then $$\text{Prob}[\sum_{i=1}^n X_i <1] \geq e^{-0.2}F_2(0.2, 2.374) > 0.1541,$$ as $F_2(0,2,D)$ is an decreasing function in $D$ by theorem (\ref{thmMono}). \end{itemize} In figure \ref{sosBE}, we plot $F_1(0.2, D)$ and $F_2(0.2,D)$ over the value of $D$. \begin{figure}[h!] \centering \caption{: Feige's bound by $F_1(0.2, D)$ and $F_2(0.2, D)$.} \includegraphics[width=9cm]{SOSvsBE.png} \label{sosBE} \end{figure} In all, the bound is improved to 0.1541. \qed Instead of achieving bound 0.1541 numerically, we can roughly verify this result by an approximation of $F_2(\xi, D)$ in an explicit form. Specially, we can derive bound 0.1536 exactly as theorem \ref{thm3} indicates in the appendix. \subsection{Add the third moment information} Recently, Garnett improved Feige's bound to 0.14 by a finer consideration of first four moments of the corresponding moments problem \cite{GARNETT2020105119}. If adding the third moment information, then we have the following lower bound in the same set-up as the previous section. \begin{equation}\label{M3} \begin{aligned} M_3 & = \mathop{\mathbb{E}}[Z^3]\\ & = -\xi^3-3\xi D +\sum_{i=1}^n \mathop{\mathbb{E}}[Y_i^3] \\ & = -\xi^3-3\xi D - \sum_{i=1}^n a_i b_i ( a_i-b_i)\\ & \geq -\xi^3-3\xi D -\xi D = -\xi^3-4\xi D \end{aligned} \end{equation} Let $\text{\bf opt}(\xi,D)$ be the optimal value of the following moments problem given $\xi$ and $D$. \begin{equation*} \begin{aligned} \max_X & \,\,\,\, \text{\bf opt}(\xi,D)=\text{Prob}[Z \geq 0] & \\ \text{subject to} \,\,\,\, & \mathop{\mathbb{E}}[Z^0]=1, & \\ & \mathop{\mathbb{E}}[Z]=-\xi, & \\ & \mathop{\mathbb{E}}[Z^2]=D+\xi^2, & \\ & \mathop{\mathbb{E}}[Z^3] \geq -\xi^3-3\xi D -\xi D = -\xi^3-4\xi D, & \\ & \mathop{\mathbb{E}}[Z^4] \leq 3D^2+6\xi^2D+\xi^4+S(\xi) D, & \end{aligned} \end{equation*} Define \begin{equation} \label{F4} F_4(\xi,D)=1- \text{\bf opt}(\xi,D) \end{equation} to be another lower bound of $\text{Prob}[ Y \leq \xi ]$. Unsurprisingly, $F_4(\xi,D)$ should be better than $F_2(\xi,D)$. \begin{theorem}\label{thm4} Let $X_1$, $X_2$,...,$X_n$ be n independent random variables with $E[X_i]=0$ and $X_i \ge -1$ for each $i$ and let $X = \sum_{i=1}^n X_i$. Then $$\text{Prob}[X \le 1] \ge 0.1587.$$ \end{theorem} \noindent{\bf Proof.} Set $\xi=0.2$. Then $S(\xi) \leq 0.2$. \begin{itemize} \item If $D \geq 2.464$, then $$\text{Prob}[\sum_{i=1}^n X_i <1] \geq e^{-0.2}F_1(0.2, 2.464) >0.1587,$$ as $F_1(0.2,D)$ is an increasing function in $D$. \item If $D \leq 2.464$, then $$\text{Prob}[\sum_{i=1}^n X_i <1] \geq e^{-0.2}F_4(0.2, 2.464) > 0.1587,$$ as $F_4(0.2,D)$ is an decreasing function in $D$ by theorem \ref{thmMono}. \end{itemize} In figure \ref{sosThird}, we plot $F_1(0.2, D)$ and $F_4(0.2,D)$ over the value of $D$. \begin{figure}[h!] \centering \caption{: Feige's bound by $F_1(0.2, D)$ and $F_4(0.2,D)$.} \includegraphics[width=9cm]{SOSvsThird.png} \label{sosThird} \end{figure} In all, the bound is improved to 0.1587. \qed As we see, from 0.1541 to 0.1587, the Feige's bound was improved only a little. The reason is that we bound $M_3$ purely by $a_i - b_i \leq 1$ in (\ref{M3}), which is not enough. In fact, we observe that often the worse-case scenario of bound $M_3$ and the bound of Berry-Esseen term $\sum_{i=1}^n \mathop{\mathbb{E}}[|Y_i^3|]$ do not agree with each other. Therefore, we are able to better bound $M_3$ through the term $\sum_{i=1}^n \mathop{\mathbb{E}}[|Y_i^3|]$ (which is bounded as $\max{|Y_i|} \cdot D$ through this paper). In general, this technique is significant to improve our result when we hybrid the moments method and Berry-Esseen theorem. \begin{theorem}\label{thm5} Let $X_1$, $X_2$,...,$X_n$ be n independent random variables with $E[X_i]=0$ and $X_i \ge -1$ for each $i$ and let $X = \sum_{i=1}^n X_i$. Then $$\text{Prob}[X \le 1] \ge 0.1798.$$ \end{theorem} \noindent{\bf Proof.} Define $T_B=\sum_{i=1}^n \mathop{\mathbb{E}}[Y_i^3]=\sum_{i=1}^n a_i b_i \frac{a_i^2+b_i^2}{a_i+b_i}$ in the same set-up as (\ref{setupY}). When applying Berry-Esseen theorem, we can define $\hat{F}_1(\xi,D,T_B)$ to be following \begin{align*} \text{Prob}[\sum_{i=1}^n Y_i \geq \xi] & \geq \Phi(\frac{\xi}{\sqrt{D}})-0.56\frac{\sum_{i=1}^n \mathop{\mathbb{E}}[Y_i^3] }{D^{3/2}} \\ & = \Phi(\frac{\xi}{\sqrt{D}})-0.56\frac{T_B}{D^{3/2}} = \hat{F}_1(\xi,D,T_B). \end{align*} At the same time, define $T_M=\sum_{i=1}^n a_i b_i (a_i - b_i)$, and we have $$M_3=\mathop{\mathbb{E}}[Z^3]= -\xi^3-3\xi D - T_M.$$ Note that $$T_B+T_M = \sum_{i=1}^n a_i b_i \frac{2a_i^2}{a_i+b_i} \leq \sum_{i=1}^n a_i b_i \cdot 2a_i \leq 2 \xi D $$ In this way, we can better bound the third moment $M_3$. \begin{itemize} \item If $T_B=D$, then the Berry-Esseen bound remains the same i.e. $F_1(\xi,D)=\hat{F_1}(\xi,D,D)$. In this way, $T_M \leq (2\xi-1)D$ implying $M_3 \geq -\xi^3-3\xi D -(2\xi-1)D = -\xi^3-5\xi D +D$. \item If $\xi D \leq T_B \leq D$, then the Berry-Esseen bound improves i.e. $F_1(\xi,D) \leq \hat{F_1}(\xi,D,T_B)$. In this way, $T_M \leq 2\xi D- T_B$ implying $M_3 \geq -\xi^3-3\xi D -(2\xi D -T_B)=-\xi^3-5\xi D +T_B$. \item If $T_B < \xi D$, then the bound $T_B+T_M \leq 2 \xi D $ is no longer effective. We have $M_3 \geq -\xi^3-3\xi D -(2\xi D -T_B)=-\xi^3-4\xi D$, the same as the bound as (\ref{M3}) when $T_B=\xi D$. \end{itemize} For each given $T_B$, we can achieve corresponding bound of $M_3$ by the analysis above. Then we can define $\hat{F_4}(\xi,D,T_B)$ in a similar way. Fix $\xi=0.2$. Suppose $T_B=s \cdot D$ for some $0 < s \leq 1$. Define function possible Feige's bound $g(s)$ to be the following: $$g(s) = \min_{D}\max\{e^{-0.2} \cdot \hat{F_1}(0.2,D,sD), \,\, e^{-0.2} \cdot \hat{F_4}(0.2,D,sD)\}$$ Figure \ref{sosTB} plots the value of g(s) under different value of s. \begin{figure}[h!] \centering \caption{: Feige's bound under different s.} \includegraphics[width=9cm]{SOSTB.png} \label{sosTB} \end{figure} Note that $\hat{F_1}(0.2,D,sD)$ is a decreasing function in $s$ for each given $D$, and $\hat{F_4}(0.2,D,sD)$ is an increasing function in $s$ for each given $D$. Figure \ref{sosTB} indicates the influence of improving Berry-Essen bound $\hat{F}_1$ dominates the influence of improving moment bound $\hat{F}_4$. Therefore, we can set $T_B=D$. \begin{itemize} \item If $D \geq 2.938$, then $$\text{Prob}[\sum_{i=1}^n X_i <1] \geq e^{-0.2}\hat{F_1}(0.2, 2.938, 2.938) >0.1798,$$ as $\hat{F}_1(0.2,D,D)$ is an increasing function in $D$. \item If $D \leq 2.938$, then $$\text{Prob}[\sum_{i=1}^n X_i <1] \geq e^{-0.2}\hat{F}_4(0.2, 2.938,2.938) > 0.1798,$$ as $\hat{F}_4(0.2, D,D)$ is an decreasing function in $D$. \end{itemize} In figure \ref{sosThird(Interplay))}, we plot $\hat F_1(0.2, D, D)$ and $\hat F_4(0.2,D, D)$ over the value of $D$. \begin{figure}[h!] \centering \caption{: Feige's bound by $\hat F_1(0.2, D, D)$ and $\hat F_4(0.2, D, D)$.} \includegraphics[width=9cm]{SOSvsThird_Interplay.png} \label{sosThird(Interplay))} \end{figure} In all, the bound is improved to 0.1798. \qed \section{Summary} \label{} In this paper, we show that the combination of Berry-Esseen theorem and moment approach can better bound probability in small deviation. As an application, we improve Feige's bound from 0.14 to 0.1798 using first four moments. However, there is still a gap between 0.1798 to the conjectured $\frac{1}{e}$. Due to the length of this paper, we leave the readers to further improve it by including higher order moments, or better bounding fourth moment via $T_B$. More importantly, we expect this common approach to be widely applied on other interesting small deviation problems. For example, Ben-Tal and et al. \cite{ben2002robust} conjectured the following: Consider a symmetric matrix $B \in R^{n\times n}$, and let $\xi=(\xi_1,\xi_2,...,\xi_n) \in R^n$ with coordinates $\xi_i$ of $\xi$ being independently identically distributed random variables with $$Pr(\xi_i=1)=Pr(\xi_i=-1)=\frac{1}{2}.$$ Then, $$Pr(\xi^T B\xi \leq Tr(B))\geq \frac{1}{4}.$$ Define the lower bound $$y(n)=\inf_{B \in S^{n \times n}} Pr(\xi^T B\xi \leq Tr(B)).$$ The best known of result is $y(n) \geq \frac{3}{100}$ \cite{he2010}. Besides, Yuan showed the upper bound of $y(n)$ is $\frac{14}{64}$ by an example \cite{yuan2013counter}. Straightforward application of the approach in this paper leads to improved bound of $\frac{6}{100}$ at least. Due to the space limitation, and since we believe finer consideration could vastly improve that bound, we omit the detailed proof and leave it for future research. In all, it will be interesting to see our approach substantially sharpening inequality bound of small deviation problems and facilitating their applications.

https://arxiv.org/abs/2202.02092

Couplings and Matchings: Combinatorial notes on Strassen's theorem

Some mathematical theorems represent ideas that are discovered again and again in different forms. One such theorem is Hall's marriage theorem. This theorem is equivalent to several other theorems in combinatorics and optimization theory, in the sense that these results can easily be derived from each other. In this paper it is shown that this equivalence extends to a finite version of Strassen's theorem, a celebrated result on couplings of probability measures. Though this equivalence is known, probabilistic or combinatorial proofs of this fact are lacking. A novel combinatorial lemma will be introduced that can be used to deduce both Hall's and Strassen's theorems.

\section{Introduction} In the original paper from \citeyear{hall1935representatives} \citeauthor{hall1935representatives} already mentions a similarity between his \emph{marriage theorem} and a result by \citeauthor{konig1916graphen} from \citeyear{konig1916graphen}. Since then numerous other results have been found that are `equivalent' to Hall's theorem. This equivalence is an informal concept and simply means that two results can be derived from each other via simple proofs. This class of equivalent theorems includes among others \emph{Menger's theorem} \citeyearpar{menger1927kurventheorie}, \emph{K\"{o}nig's minimax theorem} \citeyearpar{konig1931graphen}, the \emph{Birkhoff-von Neumann theorem} \citeyearpar{birkhoff1946algebra}, \emph{Dilworth's theorem} \citeyearpar{dilworth1950decomposition} and the \emph{max-flow min-cut theorem} by \citeauthor{ford1956flow} \citeyearpar{ford1956flow}. An extensive discussion on these equivalences can be found in \cite{reichmeider1984equivalence}. It is often overlooked that \emph{Strassen's theorem} \citeyearpar{strassen1965existence} also belongs to this class of equivalent statements. The equivalence of Strassen's theorem and Hall's theorem is already known in the literature, as it is mentioned in e.g. \cite{feldman1996doublystochastic}. However, explicit proofs that witness this equivalence are difficult to find. As Strassen's theorem is a result from probability theory, the original proof made use of analytical tools rather than the combinatorial methods used in the proofs of the above mentioned theorems. Therefore, it is remarkable that this result is, in fact, equivalent to these combinatorial statements. In this paper we consider a finite version of Strassen's theorem, which is stated in \cref{thm:strassen_finite}. For a discussion on the general version of the theorem the reader is referred to \cite{lindvall1999strassen}. The goal of this paper is twofold: firstly to give a combinatorial proof of the finite version of Strassen's theorem directly from first principles, and secondly to give a simple proof of the equivalence between Strassen's theorem and Hall's theorem. For both of these objectives we will make use of a novel lemma, that will be introduced in \cref{sec:subforest_lemma}, and which will be referred to as the \emph{subforest lemma}. As will be discussed in \cref{rem:transport}, this lemma could be derived from a more abstract result within the theory of optimal transport. In \cref{sec:main_theorems} we will introduce the two main theorems. The original part of this paper is contained in the subsequent sections, whose content is outlined in \cref{fig:outline}. \begin{figure}[h!t] \centering \includegraphics[scale=0.9]{Graphics_Outline_of_equivalences.pdf} \caption{A graphical outline of this paper, where the arrows represent the different proofs.} \label{fig:outline} \end{figure} \subsection{The two main theorems}\label{sec:main_theorems} We will start by introducing the two theorems that are the main topic of this paper. If $\P$ and $\P'$ are probability measures on two finite sets $A$ and $B$, respectively, then a \emph{coupling} of $\P$ and $\P'$ is any probability measure $\hat{\P}$ on the product set $A \times B$ for which its marginals correspond to $\P$ and $\P'$. That is, for all $U \se A$ and $S \se B$ it holds that $\P(U)=\hat{\P}(U\times B)$ and $\P'(S)=\hat{\P}(A\times S)$. \begin{theorem}[Strassen's theorem for finite sets]\label{thm:strassen_finite} Let $A$ and $B$ be finite sets and $R\se A \times B$ a relation between them. Let $\P$ and $\P'$ be probability measures on $A$ and $B$, respectively. Then there exists a coupling $\hat{\P}$ of $\P$ and $\P'$ with $\hat{\P}(R)=1$ if and only if \begin{equation}\label{eq:coupling_condition} \P(U) \leq \P'(N_R(U)), \quad \text{ for all }U \se A, \end{equation} where $N_R(U)=\cbrac{y \in B\c \exists x \in U\text{ s.t. }(x,y) \in R}$. \end{theorem} We will refer to \eqref{eq:coupling_condition} as the \emph{coupling condition}. In \cite{feldman1996doublystochastic} it is shown how the general version of Strassen's theorem can be derived from this finite version. In this paper we will use the graph theoretic formulation of the marriage theorem. All graphs in this paper are assumed to be simple finite undirected graphs. A \emph{bipartite graph} is a graph of which the vertices can be partitioned into two sets $\cbrac{A,B}$ such that all edges have one endpoint in $A$ and the other endpoint in $B$. This partition $\cbrac{A,B}$ will be called the \emph{bipartition} of the graph. A \emph{matching} of a graph is a subset $M$ of its edges such that all vertices are incident to at most one edge in $M$. If all vertices are incident to an edge in $M$, then $M$ is called a \emph{perfect matching}. \begin{theorem}[Hall's marriage theorem]\label{thm:marriage_theorem} Let $G$ be a bipartite graph with bipartition $\cbrac{A,B}$ such that $\abs{A}=\abs{B}$. Then $G$ contains a perfect matching if and only if it holds that \begin{equation}\label{eq:marriage_condition} \abs{U} \leq \abs{N_G(U)}, \quad \text{ for all }U \se A. \end{equation} \end{theorem} Here $N_G(U)$ denotes the set of vertices that are neighbors of vertices in $U$. If the underlying graph is clear, then the subscript will be dropped. We will refer to \eqref{eq:marriage_condition} as the \emph{marriage condition}. \section{Independent proof of Strassen's theorem} \subsection{The subforest lemma}\label{sec:subforest_lemma} A graph that does not contain any cycles is called a \emph{forest}. A \emph{weighted graph} is a graph that is equipped with a vertex weight function $w:V\to[0,\infty)$. For such weight functions we write $w(U)=\sum_{x \in U}w(x)$ for $U \se V$. Unless otherwise specified, a subgraph of a weighted graph is equipped with the restriction of the weight function to the vertices of the subgraph. So, in particular a spanning subgraph has the same weight function as the underlying full graph. For brevity we will call a spanning subgraph a \emph{subforest} when it is a forest. \begin{lemma}[Subforest lemma]\label{thm:subforest} Let $G=(V,E,w)$ be a weighted bipartite graph with bipartition $\cbrac{A,B}$ such that $w(B)=w(B)$. If it holds that \begin{equation} w(U) \leq w(N_G(U)), \quad \text{for all } U \se A, \label{eq:subforest_condition} \end{equation} then $G$ contains a subforest that satisfies \eqref{eq:subforest_condition}. \end{lemma} We will refer to \eqref{eq:subforest_condition} as the \emph{subforest condition}. Note that both the marriage condition \eqref{eq:marriage_condition} and the coupling condition \eqref{eq:coupling_condition} are special cases of this subforest condition. For the marriage condition all vertices have unit weight, while for the coupling condition the weight function is normalized so that $w(A)=w(B)=1$. Note that these three conditions seem to break the symmetry between sets $A$ and $B$ that is present in the setting of the theorems. This is in fact not the case, as it can be easily verified that \eqref{eq:subforest_condition} implies that $w(U) \leq w(N_G(U))$ for all $U \se B$. Here we give an independent proof of \cref{thm:subforest} directly from first principles. The proof uses the same strategy used in the inductive proof of the marriage theorem by \citet{halmos1950marriage}, in which the induction hypothesis acts as a marriage broker. That is, we distinguish between the case where the graph contains a `critical set' of vertices and the case where no such set exists. \begin{proof}[Proof of \cref{thm:subforest}] We will apply induction on $\abs{V}$. Let $\mathcal{S}=\cbrac{U \se A \c 0<\abs{U}<\abs{A}} \cup \cbrac{U \se B \c 0<\abs{U}<\abs{B}}$ denote the collection of non-empty strict subsets of either $A$ or $B$. We will distinguish two cases. In the first case we assume that there exists a $U \in \mathcal{S}$ with $w(U)=w(N_G(U))$. Without loss of generality we assume that $U \se A$. Let $\cbrac{V_1, V_2}$ be the partition of $V$ given by $V_1=U\cup N_G(U)$ and $V_2=(A\setminus U) \cup (B \setminus N_G(U))$. Then both induced subgraphs $G[V_1]$ and $G[V_2]$ satisfy the subforest condition \eqref{eq:subforest_condition}. Thus by the induction hypothesis there exist subforests $F_1$ and $F_2$ of $G[V_1]$ and $G[V_2]$, respectively, that both satisfy the subforest condition \eqref{eq:subforest_condition}. The graph $F=(V,E(F_1)\cup E(F_2))$, that contains all edges of $F_1$ and $F_2$, is a subforest of $G$ that satisfies the subforest condition with respect to $w$. Also note that $F$ contains at least two connected components, since none of the vertices in $V_1$ is connected to any of the vertices in $V_2$. For the second case we assume that $w(U)<w(N_G(U))$ for all $U \in \mathcal{S}$. Let $\varepsilon$ denote the minimal weight of any vertex of $G$, i.e. $\varepsilon = \min_{v \in V}w(v)$. Let $x \in V$ be any vertex with $w(x)=\varepsilon$. Without loss of generality we can assume that $x \in A$. Let $y \in N_G(x)$ be any neighbor of $x$. Let $\mathcal{U}=\cbrac{U \se A \c x \notin U \text{ and }U \cap N_G(y) \neq \varnothing}$ and take \begin{equation*} \delta = \min_{U \in \mathcal{U}} w(N_G(U))-w(U). \end{equation*} Since $A-x \in \mathcal{U}$, we have that $\delta \leq \varepsilon$. Let $D \in \mathcal{U}$ be such that $w(N_G(D))-w(D)=\delta$. We add a new element $\tilde{x}$ to $A$ to obtain $\tilde{A}=A+\tilde{x}$. Let $\tilde{V}=V+\tilde{x}$, $\tilde{E}=E(G)+\cbrac{\tilde{x},y}$ and $\tilde{G}=(\tilde{V},\tilde{E})$. Define the weight function $\tilde{w}$ on $V+\tilde{x}$ by $\tilde{w}=w+\delta\Ind_{\cbrac{\tilde{x}}}-\delta\Ind_{\cbrac{x}}$. The weighted graph $(\tilde{G},\tilde{w})$ now satisfies the subforest condition. It also holds that $w(D+\tilde{x})=w(N_{\tilde{G}}(D+\tilde{x}))$. If $D \neq A - x$, then $\abs{(D+\tilde{x})\cup N_{\tilde{G}}(D+\tilde{x})} < \abs{V}$. It follows from the induction hypothesis, in the same manner as in the previous case, that there exists a subforest $\tilde{F}$ of $\tilde{G}$ with $x$ and $y$ in two distinct components such that $(\tilde{F}, \tilde{w})$ satisfies \eqref{eq:subforest_condition}. The graph $F=(V,E(\tilde{F})-\cbrac{\tilde{x},y}+\cbrac{x,y})$ is a spanning subgraph of $G$. Since $x$ and $y$ are contained in distinct components of $\tilde{T}$, we also have that $F$ is a forest. It is also clear that $(F,w)$ satisfies the subforest condition. If instead we have that $D = A - x$, then $N_{G}(D+\tilde{x})=B$. This follows since we have for all $v \in V$ that $w(x) \leq w(v)$ and $w(v)<w(N_G(v))$, so there does not exists a $v \in B$ with $N_{G}(v)\se \cbrac{x}$. This means that $\varepsilon = \delta$. Define the weight function $w'$ on $V-x$ by $w'=w-\delta\Ind_{\cbrac{y}}$. Then the weighted graph $(G[V-x],w')$ satisfies the subforest condition. Hence, by the induction hypothesis, there exists a spanning subforest $F'$ of $G[V-x]$ satisfying the subforest condition. Let $F=(V,E(F')+\cbrac{x,y})$. Then $F$ is a subforest of $G$ satisfying the subforest condition. In both cases we have shown the existence of a spanning subforest that satisfies the subforest condition, thus completing the proof. \end{proof} \begin{remark}\label{rem:transport} The problem of finding a coupling that satisfies the coupling condition \eqref{eq:coupling_condition} can also be phrased as an optimal transport problem. A solution to such a transportation problem corresponds to a bipartite graph with weights assigned to the edges. \citet{klee1968facets} showed that the polytope of feasible solutions has at its vertices exactly those solutions whose accompanying bipartite graph corresponds to a forest. Hence, the subforest lemma can also be derived from this result of \citeauthor{klee1968facets}. \end{remark} \subsection{Deriving Strassen's theorem from the subforest lemma} For the independent proof of Strassen's theorem for finite sets, we show how it can easily be derived from the subforest lemma. It is natural to translate the setting of \cref{thm:strassen_finite} to a weighted bipartite graph $G=(V,E,w)$ defined by \begin{equation}\label{eq:coupling_graph} \begin{gathered} V=A \cup B,\quad E=\cbrac{\cbrac{x,y} \c (x,y) \in R}, \text{ and } \\ w(x)=\begin{cases}\P(x) & \text{ if }x \in A \\ \P'(x) & \text{ if }x \in B\end{cases} \end{gathered} \end{equation} (Here we assume w.l.o.g. that $A\cap B=\varnothing$.) The coupling condition then translates to $w(U) \leq w(N_G(U))$ for all $U \se A$, while the sought coupling becomes an edge weight function $\hat{w}:E \to [0,\infty)$ that satisfies $w(x)=\sum_{e \sim x}\hat{w}(e)$ for all $x \in A$, where the sum is taken over all edges incident to $x$. This translation gives us the following equivalent formulation of \cref{thm:strassen_finite}, which resembles a weighted version of Hall's marriage theorem. \begin{proposition}[Combinatorial formulation of Strassen's theorem ]\label{thm:strassen_finite_combinatorial} Let $G=(V,E,w)$ be a weighted bipartite graph with bipartition $\cbrac{A,B}$ such that $w(A)=w(B)$. Then the following are equivalent: \begin{enumerate}[$(i)$] \item for all $U \se A$ it holds that $w(U) \leq w(N(U))$; \label{it:strassen_condition} \item there exists an edge weight function $\hat{w}:E \to [0,\infty)$ such that for all $x \in V$ it holds that $w(x)=\sum_{e \sim x}\hat{w}(e)$, where the sum is taken over all edges incident to $x$. \label{it:strassen_edge_weight} \end{enumerate} \end{proposition} \begin{proof}[Proof of \cref{thm:strassen_finite_combinatorial} using \cref{thm:subforest}] The implication from \eqref{it:strassen_edge_weight} to \eqref{it:strassen_condition} is easily shown. If $\hat{w}$ satisfies \eqref{it:strassen_edge_weight}, then \begin{equation*} w(U) = \sum_{x \in U}\sum_{e \sim x}\hat{w}(e) \leq \sum_{y \in N(U)}\sum_{e \sim y}\hat{w}(e) = w(N(U)). \end{equation*} The reverse implication will be proven by induction on $\abs{V}$. Since $w$ satisfies \eqref{it:strassen_condition}, by the subforest lemma there exists a subforest $F$ of $G$ satisfying \eqref{it:strassen_condition}. Since $F$ is a forest, there exists a vertex $x$ in $F$ with degree $1$. Without loss of generality we can assume that $x \in A$. Let $y \in B$ be the unique neighbor of $x$ in $F$. Note that it follows from \eqref{it:strassen_condition} that $w(x)\leq w(y)$. Set $\varepsilon = w(x)$. Consider the induced subgraph $F[V-x]$ obtained by removing vertex $x$ from $F$ and equip it with the vertex weight function $\tilde{w}:V-x\to[0,\infty)$ given by $\tilde{w}(v)=w(v)-\varepsilon\Ind_{\cbrac{v=y}}$. The weighted graph $(F[V-x],\tilde{w})$ satisfies \eqref{it:strassen_condition}, hence by the induction hypothesis there exists an edge weight function $\hat{w}$ on $F[V-x]$ satisfying \eqref{it:strassen_edge_weight}. Now define an edge weight function on the edges of $G$ by \begin{equation*} \check{w}(e) = \begin{cases} \hat{w}(e) & \text{if } e \in E(F[V-x]) \\ \varepsilon & \text{if } e = \cbrac{x,y} \\ 0 & \text{otherwise.} \end{cases} \end{equation*} Then $\check{w}$ is the sought edge weight function satisfying \eqref{it:strassen_edge_weight}. \end{proof} \Cref{thm:strassen_finite} follows directly from \cref{thm:strassen_finite_combinatorial} by normalizing the vertex and edge weights, so that these form probability measures. This independent proof of Strassen's theorem for finite sets is constructive and can in principle be used to find the required coupling. However, far more efficient methods for finding such a coupling exists. In \cite[corollary 2.1.5]{lovasz1986matching} it is mentioned that \cref{thm:strassen_finite_combinatorial} can be derived from the max-flow min-cut theorem. The method is similar to the derivation of the marriage theorem from the max-flow min cut theorem, that is given in \cite{ford1958representatives}. This derivation is not only elegant, it also shows that any method for finding maximal network flows can also be used to find such a coupling. \section{Equivalence of Hall's theorem and Strassen's theorem} In the second part of this paper we prove the equivalence of Strassen's theorem for finite sets and Hall's marriage theorem. \subsection{Deriving Hall's theorem from Strassen's theorem} The derivation of Hall's theorem from Strassen's theorem will go via the subforest lemma. \begin{proof}[Proof of \cref{thm:subforest} using Strassen's theorem] The statement will be proven by induction on $\abs{E}$. If $\abs{E}=1$, then $G$ is itself a forest, so there is nothing to prove. Now assume that $\abs{E} \geq 2$ and that the statement holds when $\abs{E}$ is smaller. Define the probability measures $\P$ and $\P'$ on $A$ and $B$, respectively, by setting $\P(x)=\frac{w(x)}{w(A)}$ and $\P(y)=\frac{w(y)}{w(B)}$ for $x \in A$ and $y$ in $B$. Since $G$ satisfies the subforest condition, these two probability measures then satisfy the coupling condition with respect to the relation $E$. Hence, by Strassen's theorem there exists a coupling $\hat{\P}$ of $\P$ and $\P'$ that is supported on $E$. If $G$ is not a forest, then there is a subset of edges $C \se E$ that constitute a cycle. Now take $\varepsilon = \min\cbrac{\hat{\P}(e) \c e \in C}$ and let $e^* \in C$ be such that $\hat{\P}(e)=\varepsilon$. Since $G$ is a bipartite graph, the cycle $C$ contains an even number of edges. Hence, we can partition $C$ into two sets $\cbrac{I_C,J_C}$ such that edges in $I_C$ are only incident to edges in $J_C$ and vice-versa. Without loss of generality we can assume that $e^* \in I_C$. Now define a new probability measure $\tilde{\P}$ on $E$ by \begin{equation*} \tilde{\P}(e) = \begin{cases} \hat{\P}(e)-\varepsilon & \text{ if }e \in I_C \\ \hat{\P}(e)+\varepsilon & \text{ if }e \in J_C \\ \hat{\P}(e) & \text{ otherwise.} \end{cases} \end{equation*} Since each vertex in $G$ is incident to the same number of edges in $I_C$ as to edges in $J_C$, we have that $\tilde{\P}$ is also a coupling of $\P$ and $\P'$. Moreover, the coupling $\tilde{\P}$ is supported on $E\setminus \cbrac{e^*}$, since by construction it holds that $\tilde{\P}(e^*)=0$. Thus, by applying Strassen's theorem in the reverse direction, we find that the relation $E \setminus \cbrac{e^*}$ satisfies the coupling condition with respect to $\P$ and $\P'$. It follows that the weighted graph $G-e^*$ satisfies the subforest condition. By the induction hypothesis $G-e^*$ contains a subforest $F$ satisfying that condition. Clearly, $F$ is also a subforest of $G$, which finishes the proof. \end{proof} \begin{proof}[Proof of \cref{thm:marriage_theorem} using the subforest lemma] We will only prove the sufficiency of the marriage condition, which will be done by induction on $\abs{E}$. So, we assume that $G$ satisfies the marriage condition. Clearly the statement holds if $\abs{E}=1$. Now assume that $\abs{E} \geq 2$ and that the statement holds if $\abs{E}$ is smaller. Note that the marriage condition is a special case of the subforest condition where each vertex has unit weight. Hence, by \cref{thm:subforest} there exists a subforest $F$ of $G$ that satisfies the marriage condition. Since $F$ is a forest, there exists a vertex $x$ in $F$ with degree $1$. Let $y$ be the unique neighbor of $x$ in $F$. Then the induced subgraph $G[V\setminus\cbrac{x,y}]$ still satisfies the marriage condition. Thus by the induction hypothesis $G[V\setminus\cbrac{x,y}]$ has perfect matching $M$. Taking $M \cup \cbrac{x,y}$ gives a perfect matching of $G$. \end{proof} \subsection{Deriving Strassen's theorem from the marriage theorem} To finish our reciprocal derivations we still have to prove Strassen's theorem from the marriage theorem. This will be done using two well-known generalizations of both theorems, \cref{thm:hall_deficiency,thm:strassen_deficiency} below, that allow for some small deficiencies in the conditions. The used generalization of Hall's marriage theorem is due to \citet{ore1955matching} and can be found in e.g. \cite[Thm.~1.3.1]{lovasz1986matching}. It can be easily derived from the marriage theorem itself, which led \citeauthor{mirsky1969selfrefining} to call the marriage theorem a \emph{self-refining result} \cite{mirsky1969selfrefining}. For completeness we also give this derivation. \begin{proposition}[Hall's theorem with deficiency]\label{thm:hall_deficiency} Let $G$ be a bipartite graph with bipartition $\cbrac{A,B}$ with $\abs{A}=\abs{B}=n$. Then $G$ contains a matching $M$ with $\abs{M}\geq n-k$ if and only if it holds that \begin{equation}\label{eq:marriage_condition_deficiency} \abs{U} \leq \abs{N_G(U)}+k, \quad \text{ for all }U \se A. \end{equation} \end{proposition} \begin{proof}[Proof of \cref{thm:hall_deficiency} using Hall's marriage theorem] We only prove the sufficiency of \eqref{eq:marriage_condition_deficiency}. Construct the bipartite graph $\tilde{G}$ by adding $k$ new vertices $a_1,\ldots a_k$ to $A$ and $k$ new vertices $b_1,\ldots b_k$ to $B$. Set $\tilde{A}=A\cup\cbrac{a_1,\ldots a_k}$ and $\tilde{B}=B\cup\cbrac{b_1,\ldots b_k}$. Also add edges between all $a_i$ and all vertices in $\tilde{B}$ and all $b_i$ and all vertices in $\tilde{A}$. That is, $\tilde{G}=(\tilde{A}\cup \tilde{B}, \tilde{E})$ with $\tilde{E}=E \cup \cbrac{\cbrac{a_i,v} \c 1 \leq i \leq k, \ v \in \tilde{B}} \cup \cbrac{\cbrac{b_i,v} \c 1 \leq i \leq k, \ v \in \tilde{A}}$. Since $G$ satisfies \eqref{eq:marriage_condition_deficiency} and all vertices have $k$ more neighbors in $\tilde{G}$ than in $G$, we have that $\tilde{G}$ satisfies the marriage condition \eqref{eq:marriage_condition}. Hence, by \cref{thm:marriage_theorem} $\tilde{G}$ contains a perfect matching $\tilde{M}$. Note that $\abs{\tilde{M}}=n+k$. Hence, $M:=\tilde{M}\cap E$ is a matching of $G$ with $\abs{M}\geq n-k$, since at most $2k$ edges in $\tilde{M}$ are incident to any of the $2k$ vertices that are not in $G$. \end{proof} As with \cref{thm:hall_deficiency} the generalized version of Strassen's theorem also follows from its original. However, for our purposes we will derive it from \cref{thm:hall_deficiency} instead. \begin{proposition}[Strassen's theorem with deficiency]\label{thm:strassen_deficiency} Let $A$ and $B$ be finite sets and $R\se A \times B$ a relation between them. Let $\P$ and $\P'$ be probability measures on $A$ and $B$, respectively. Let $\varepsilon \geq 0$ be given. Then there exists a coupling $\hat{\P}$ of $\P$ and $\P'$ with $\hat{\P}(R)\geq 1-\varepsilon$ if and only if \begin{equation}\label{eq:coupling_condition_perturbation} \P(U) \leq \P'(N_R(U)) + \varepsilon, \quad \text{ for all }U \se A. \end{equation} \end{proposition} \begin{proof}[Proof of \cref{thm:strassen_deficiency} using \cref{thm:hall_deficiency}] For the necessity of \eqref{eq:coupling_condition_perturbation} we note that if $\hat{\P}$ is a coupling of $\P$ and $\P'$ with $\hat{\P}(R)\geq 1- \varepsilon$, then it holds for all $U\se A$ that \begin{equation*} \P(U)=\hat{\P}(U\times B) \leq \hat{\P}(U\times N_R(U))+\varepsilon \leq \hat{\P}(A\times N_R(U))+\varepsilon = \P'(N_R(U))+\varepsilon. \end{equation*} It remains to prove its sufficiency. This will be done in two steps. In the first step we assume that $\P$ and $\P'$ are both rational valued and that $\varepsilon\in\Q$, and in the second step we derive the result for arbitrary $\P$, $\P'$ and $\varepsilon$. \subsubsection*{(1)} First we assume that $\P$ and $\P'$ are rational valued and we also take $\varepsilon$ rational. Define $G=(V,E,w)$ as in \eqref{eq:coupling_graph}. Then we have that $w(x)\in \Q$ for all $x \in V$. Since $V$ is finite, there exists a large enough $N \in \N$ such that the product $Nw(x)$ is an integer for all $x \in V$ and such that $k:=\varepsilon N$ is an integer as well. Let $\tilde{V}=\bigcup_{x \in A}\bigcup_{i =1}^{Nw(x)}\cbrac{x_i}$ be the set consisting of $Nw(x)$ copies of each element $x \in V$. Now consider the bipartite graph $\tilde{G}=(\tilde{V}, \tilde{E})$, where the edge set is given by $\tilde{E}=\cbrac{\cbrac{x_i, y_j} \c \cbrac{x, y} \in E}$. That is, two vertices $x_i$ and $y_j$ in $\tilde{G}$ are connected by an edge if and only if their originals $x$ and $y$ are adjacent in $G$. Denote the bipartition of $\tilde{G}$ by $\cbrac{\tilde{A}, \tilde{B}}$. Since $\P$ and $\P'$ satisfy \eqref{eq:coupling_condition_perturbation}, we then have for all $\tilde{U} \se \tilde{A}$ that \begin{equation*} \abs{\tilde{U}} \leq \abs{N_{\tilde{G}}(\tilde{U})}+k. \end{equation*} By \cref{thm:hall_deficiency} there exists a matching $\tilde{M}$ of $\tilde{G}$ with $\abs{\tilde{M}}=N-k$. Let $a_1,\ldots,a_k$ and $b_1,\ldots,b_k$ denote the $k$ vertices in $\tilde{A}$ and $\tilde{B}$, respectively, that are unmatched by $\tilde{M}$. Now consider the set of edges $M^+:=\tilde{M} \cup \cbrac{\cbrac{a_i,b_i}\c i \in [k]}$, which is obtained from $\tilde{M}$ by adding $k$ arbitrary edges, not necessarily belonging to $\tilde{E}$, between the unmatched vertices. For each pair $(x,y)\in A\times B$ let \begin{equation*} \hat{w}(x,y)=\abs*{\bigcup_{i=1}^{Nw(x)}\bigcup_{j=1}^{Nw(y)}\cbrac{x_i,y_j} \cap M^+} \end{equation*} denote the number of edges between copies of $x$ and copies of $y$ that occur in $M^+$. Since $M^+$ is a perfect matching of the complete bipartite graph on $\tilde{A}\cup \tilde{B}$, we find that $\sum_{y \in B}\hat{w}(x,y)=Nw(x)$ for all $x \in A$ and similarly that $\sum_{x \in A}\hat{w}(x,y)=Nw(y)$ for all $y \in B$. So, the probability measure $\hat{\P}$ on $A \times B$ defined by $\hat{\P}(x,y)=\frac{\hat{w}(x,y)}{N}$ is a coupling of $\P$ and $\P'$. Since only $k$ of the edges of $M^+$ do not belong to $\tilde{M}$ we also find that $\hat{\P}(R)=1-\varepsilon$, so $\hat{\P}$ is the sought coupling. \subsubsection*{(2)} Let $\P$, $\P'$ and $\varepsilon$ be arbitrary. Let $(\varepsilon_i)_{i \in \N}$ be a rational sequence converging to $\varepsilon$ from above. Since $\P$ and $\P'$ satisfy \eqref{eq:coupling_condition_perturbation}, we can find two sequences $(\P_i)_{i \in \N}$ and $(\P'_i)_{i \in \N}$ of rational valued probability measures on $A$ and $B$ that converge such that for every $i \in \N$ it holds that \begin{equation*} \P_i(U) \leq \P_i'(N_R(U)) + \varepsilon_i, \quad \text{ for all }U \se A. \end{equation*} By the first part of the proof, for each $i$ there exists a coupling $\hat{\P}_i$ of $\P_i$ and $\P'_i$ with $\hat{\P}_i(R) \geq 1-\varepsilon_i$. Note that we can interpret $(\hat{\P}_i)_{i \in \N}$ as a sequence in the compact metric space $[0,1]^{\abs{E}}$. Thus it contains a converging subsequence $(\hat{\P}_{i_j})_{j \in \N}$ with limit $\hat{\P}$. It follows that $\hat{\P}(R)\geq 1-\varepsilon$. It remains to be shown that $\hat{\P}$ is a coupling of $\P$ and $\P'$. Let $\delta > 0$ be given. Then for all $x \in A$ there exists a $k \in \N$ such that for all $j \geq k$ it holds that both $\abs{\P_{i_j}(x)-\P(x)} < \delta$ and \begin{equation*} \abs{\hat{\P}_{i_j}(\cbrac{x}\times B)-\hat{\P}(\cbrac{x}\times B)} < \delta. \end{equation*} It follows that \begin{align*} \abs{\P(x)-\hat{\P}(\cbrac{x}\times B)} & < \abs{\P(x)-\hat{\P}_{i_k}(\cbrac{x}\times B)}+ \delta \\ & = \abs{\P(x)-\P_{i_k}(x)}+ \delta \\ & < 2\delta. \end{align*} Similarly, we find that $ \abs{\P'(x)-\hat{\P}(A\times \cbrac{x})}<2\delta$ for all $x \in B$. As this holds for all $\delta > 0$, it follows that $\hat{\P}$ is a coupling of $\P$ and $\P'$. \end{proof} \section*{Acknowledgments} This paper originated as follow-up on my bachelor's thesis. I would like to thank my supervisors Luca Avena and Siamak Taati for their support, Siamak for introducing me to this topic and the many lengthy discussions, and Luca for guiding me through the subsequent process leading to this paper. I also thank Frits Spieksma and Leen Stougie for their useful comments. \printbibliography \end{document}

https://arxiv.org/abs/1807.04670

The Hausdorff-Young inequality on Lie groups

We prove several results about the best constants in the Hausdorff-Young inequality for noncommutative groups. In particular, we establish a sharp local central version for compact Lie groups, and extend known results for the Heisenberg group. In addition, we prove a universal lower bound to the best constant for general Lie groups.

\section{Introduction} For $f \in L^1(\mathbb{R}^n)$, define the Fourier transform $\hat f$ of $f$ by \[ \hat f(\xi) = \int_{\mathbb{R}^n} f(x) \, e^{2\pi i \xi \cdot x} \, dx \qquad\forall \xi \in \mathbb{R}^n. \] Then the Riemann--Lebesgue lemma states that $\hat f \in C_0(\mathbb{R}^n)$ and \[ \| \hat f \|_\infty \leq \| f\|_1. \] Further, the Plancherel theorem entails that if $f \in L^2(\mathbb{R}^n)$, then \[ \| \hat f \|_2 = \| f \|_2. \] Suppose that $1 \leq p \leq 2$ and $p'$ is the conjugate exponent to $p$, that is, $1/p' = 1 - 1/p$. Then interpolation implies the Hausdorff--Young inequality, namely, \begin{equation}\label{eq:HY} \| \hat f \|_{p'} \leq C \|f\|_{p} \end{equation} for all $f \in L^p(\mathbb{R}^n)$, where $C \leq 1$. We denote the best constant for this inequality, that is, the smallest possible value of $C$, by $H_p(\mathbb{R}^n)$. This was found many years after the original result. We define the Babenko--Beckner constant $B_p$ by \[ B_p = \frac{ p^{1/2p} }{ (p')^{ 1/2p' } }. \] Then $B_p < 1$ when $1 < p < 2$. \begin{theorem}[Babenko \cite{Bab-61}, Beckner \cite{Bec-1975}] For all $p \in [1,2]$, \[ H_p(\mathbb{R}^n) = (B_p)^n. \] \end{theorem} Babenko treated the case where $p' \in 2 \mathbb{Z}$, and Beckner proved the general case. The extremal functions are gaussians; see \cite{lieb_1990} for an alternative proof. One can extend the Babenko--Beckner theorem to more general contexts than $\mathbb{R}^n$, such as locally compact abelian groups $G$. For instance, the best constant $H_p(G)$ for the inequality \eqref{eq:HY} when $G = \mathbb{R}^a \times \T^b \times \mathbb{Z}^c$ is $(B_p)^a$. The extremal functions are of the form $\gamma \otimes \chi \otimes \delta$, where $\gamma$ is a gaussian on $\mathbb{R}^a$, $\chi$ is a character of $\T^b$, and $\delta$ is the characteristic function of a point in $\mathbb{Z}^c$. For nonabelian groups, matters are more complicated, in part because the interpretation of the $L^{q}$ norm of the Fourier transform for $q \in (2,\infty)$ is trickier. We refer the reader to Section \ref{s:FLq} below for details. General versions of the Hausdorff--Young inequality \eqref{eq:HY} were obtained by Kunze \cite{Kun-1958} and Terp \cite{terp_1980} for arbitrary locally compact groups $G$, and a number of works in the literature are devoted to the study of the corresponding best constants $H_p(G)$. It is known, at least in the unimodular case, that $H_p(G)<1$ for $p \in (1,2)$ if and only if $G$ has no compact open subgroups \cite{Russo-1974,fournier_1977}. On the other hand, when $H_p(G)$ is not $1$, its value is known only in few cases, and typically only for exponents $p$ whose conjugate exponent is an even integer; in addition, as shown by Klein and Russo, extremal functions need not exist \cite{KR-1978}. Recently, various authors considered local versions of the Hausdorff--Young inequality. Namely, for each neighbourhood $U$ of the identity $e \in G$, define $H_p(G;U)$ as the best constant in the inequality \eqref{eq:HY} with the additional support constraint $\supp f \subseteq U$, and let $H_p^\mathrm{loc}(G)$ be the infimum of the constants $H_p(G;U)$. Clearly $H_p^\mathrm{loc}(G) \leq H_p(G)$, and equality holds whenever $G$ has a contractive automorphism. For other groups, however, the inequality may be strict, which makes the study of $H_p^\mathrm{loc}(G)$ interesting also for groups where $H_p(G) = 1$, such as compact groups. Indeed, in the case of the torus $G = \T^n$, the value of $H_p^\mathrm{loc}(G)$ is known and is strictly less than $1$ for $p \in (1,2)$. \begin{theorem}[Andersson \cite{And-PhD,And-1994}, Sj\"olin \cite{Sjo-1995}, Kamaly \cite{Kam-2000}]\label{thm:local-HY-Tn} For all $p \in [1,2]$, \[ H_p^\mathrm{loc}(\T^n) = (B_p)^n. \] \end{theorem} Here we are interested in analogues of the above result for noncommutative Lie groups $G$. We also study what happens when additional symmetries are imposed by restricting to functions $f$ on $G$ which are invariant under a compact group $K$ of automorphisms of $G$. Let us denote by $H_{p,K}(G)$ and $H_{p,K}^\mathrm{loc}(G)$ the corresponding global and local best Hausdorff--Young constants. Note that the original constants $H_p(G)$ and $H_{p}^\mathrm{loc}(G)$ correspond to the case where $K$ is trivial. When $K$ is nontrivial, \emph{a priori} the new constants $H_{p,K}(G)$ and $H_{p,K}^\mathrm{loc}(G)$ might be smaller. However we can prove a universal lower bound, which is independent of the symmetry group $K$ and depends only on $p$ and the dimension of $G$. \begin{theorem}\label{thm:local-HY} Let $G$ be a Lie group and $K$ be a compact group of automorphisms of $G$. For all $p \in [1,2]$, \[ H_{p,K}^\mathrm{loc}(G) \geq (B_p)^{\dim(G)}. \] \end{theorem} Recall that a function $f$ on a group $G$ is \emph{central} if $f(xy) = f(yx)$, that is, if $f$ is invariant under the group $\Inn(G)$ of inner automorphisms of $G$. Garc{\'\i}a-Cuerva, Marco and Parcet \cite{GCMP-2003} and Garc{\'\i}a-Cuerva and Parcet \cite{GCP-2004} studied the Hausdorff--Young inequality for compact semi\-simple Lie groups $G$ restricted to central functions; in particular, they obtained the inequality $H^\mathrm{loc}_{p,\Inn(G)}(G) > 0$, which they applied to answer questions about Fourier type and cotype of operator spaces (see also \cite{Par-2006}). Theorem \ref{thm:local-HY} gives a substantially more precise lower bound to $H^\mathrm{loc}_{p,\Inn(G)}(G)$. As a matter of fact, in this case we can prove that equality holds. \begin{theorem}\label{thm:local-central-HY-compact-Lie} Suppose that $G$ is a compact connected Lie group. Then, for all $p \in [1,2]$, \[ H_{p,\Inn(G)}^\mathrm{loc}(G) = (B_p)^{\dim(G)}. \] \end{theorem} Note on the one hand that, in the abelian case $G = \T^n$, all functions are central, so Theorem \ref{thm:local-central-HY-compact-Lie} extends Theorem \ref{thm:local-HY-Tn}. On the other hand, it would be interesting to know whether the result holds also without the restriction to central functions. More generally, one may ask whether the inequality in Theorem \ref{thm:local-HY} is actually an equality for an arbitrary Lie group $G$. As a matter of fact, the equality \[ H_{p,K}^\mathrm{loc}(G) = (B_p)^{\dim(G)} \] holds for arbitrary $G$ and $K$ whenever $p' \in 2\mathbb{Z}$, as a consequence of a recent result of Bennett, Bez, Buschenhenke, Cowling and Flock \cite{BBBCF_2018} and the relation between the best constants for the Young and the Hausdorff--Young inequalities (see Proposition \ref{prop:basic} below). In particular, by interpolation, \[ H_{p,K}^\mathrm{loc}(G) < 1 \] for all $p \in (1,2)$ and arbitrary $G$ and $K$ with $\dim(G) > 0$. Moreover, the equality \begin{equation}\label{eq:hy_conjecture} H_{p}(G) = H_p^\mathrm{loc}(G) = (B_p)^{\dim(G)} \end{equation} holds when $p' \in 2\mathbb{Z}$ for all Lie groups $G$ with a contractive automorphism (which are nilpotent---see \cite{Siebert-1986}), and also for all solvable Lie groups $G$ admitting a chain of closed subgroups \[ \{e\} = G_0 < G_1 < \dots < G_{n-1} < G_n = G, \] where $G_j$ is normal in $G_{j+1}$ and $G_{j+1} / G_j$ is isomorphic to $\mathbb{R}$ (here $n=\dim(G)$). For many of those groups $G$, the upper bound $H_{p}(G) \leq (B_p)^{\dim(G)}$ for $p'\in 2\mathbb{Z}$ was proved in \cite{KR-1978}, but the question of the lower bound was left open there, except for the Heisenberg groups. Hence Theorem \ref{thm:local-HY} proves the sharpness of a number of results in \cite{KR-1978}. The Heisenberg groups $\Heis_n$ are among the simplest examples of groups in the above class. Nevertheless, determining the value of $H_{p}(\Heis_n) = H_{p}^\mathrm{loc}(\Heis_n)$ appears to be a nontrivial problem when $p' \notin 2\mathbb{Z}$, and is related to a similar problem for the so-called Weyl transform. Recall that the Weyl transform $\rho$ on $\mathbb{C}^n$ maps functions on $\mathbb{C}^n$ to integral operators on $L^2(\mathbb{R}^n)$ \cite{Folland-1989}, and an inequality of Hausdorff--Young type can be proved for $\rho$ \cite{KR-1978,Russo-1979}: for all $p \in [1,2]$, \begin{equation}\label{eq:HY_weyl} \| \rho(f) \|_{\Sch^{p'}(L^2(\mathbb{R}^n))} \leq C \|f\|_{L^p(\mathbb{C}^{n})}, \end{equation} where $\Sch^{q}(\Hilb)$ denotes the $q$th Schatten class of operators on the Hilbert space $\Hilb$, and $C \leq 1$. As above, we can define $W_p(\mathbb{C}^n)$ as the best constant in \eqref{eq:HY_weyl}, as well as corresponding local and symmetric versions $W_{p}^\mathrm{loc}(\mathbb{C}^n), W_{p,K}(\mathbb{C}^n), W_{p,K}^\mathrm{loc}(\mathbb{C}^n)$. A scaling argument (see Proposition \ref{prp:weylheisenberg} below) then shows that, for all compact subgroups $K$ of the unitary group $\group{U}(n)$, \begin{equation}\label{eq:HY_heis_weyl} H_{p,K}(\Heis_n) = B_p \, W_{p,K}(\mathbb{C}^n) \end{equation} (here $\group{U}(n)$ acts naturally on $\mathbb{C}^n$ and the first layer of $\Heis_n$). So the problem of determining the best Hausdorff--Young constants for the Heisenberg group $\Heis_n$ is equivalent to the analogous problem for the Weyl transform. In particular, \eqref{eq:HY_heis_weyl} and Theorem \ref{thm:local-HY} yield that \[ W_{p,K}(\mathbb{C}^n) \geq (B_p)^{2n} \] for all $p \in [1,2]$. As an indication that equality may well hold, here we prove the following local result. \begin{theorem}\label{thm:local-HY-weyl} Let $K$ be a compact subgroup of $\group{U}(n)$. Then, for all $p \in [1,2]$, \[ W_{p,K}^\mathrm{loc}(\mathbb{C}^n) \geq (B_p)^{2n}. \] Moreover, if $K \supseteq \group{U}(1) \times \dots \times \group{U}(1)$, then, for all $p \in [1,2]$, \[ W_{p,K}^\mathrm{loc}(\mathbb{C}^n) = (B_p)^{2n}. \] \end{theorem} Functions on $\mathbb{C}^n$ or $\Heis_n$ that are invariant under $\group{U}(1) \times \dots \times \group{U}(1)$ are called polyradial. Equality in Theorem \ref{thm:local-HY-weyl} is obtained as a consequence of the following weighted Hausdorff--Young inequality for polyradial functions $f$: \begin{equation}\label{eq:weighted_HY_weyl} \| \rho(f) \|_{\Sch^{p'}(\mathbb{R}^n)} \leq (B_p)^{2n} \| f e^{(\pi/2)|\cdot|^2} \|_{L^p(\mathbb{C}^n)}. \end{equation} Unfortunately we have not found a way to remove the weight and obtain the equality $W_{p,K}(\mathbb{C}^n) = W_{p,K}^\mathrm{loc}(\mathbb{C}^n)$ for arbitrary $p \in [1,2]$; note however that $W_{p,K}(\mathbb{C}^n) = W_{p,K}^\mathrm{loc}(\mathbb{C}^n) = (B_p)^{2n}$ when $p' \in 2\mathbb{Z}$, as proved in \cite{KR-1978}. Both cases where we can prove equalities in Theorems \ref{thm:local-central-HY-compact-Lie} and \ref{thm:local-HY-weyl} for general $p \in [1,2]$ correspond to Gelfand pairs (see, for example, \cite{carcano_1987}): indeed, central functions on a compact group $G$ and polyradial functions on the Heisenberg group $\Heis_n$ form commutative subalgebras of the respective convolution algebras $L^1(G)$ and $L^1(\Heis_n)$. It seems a reasonable intermediate question to ask for best constants in Hausdorff--Young inequalities in the context of Gelfand pairs, since here the group Fourier transform reduces to the Gelfand transform for the corresponding commutative algebra of invariant functions, which makes the $L^q$ norm of the Fourier transform in these settings more accessible. Indeed, in both the proofs of Theorems \ref{thm:local-central-HY-compact-Lie} and \ref{thm:local-HY-weyl}, this additional commutativity allows one to relate the group Fourier transform and the Weyl transform with the Euclidean Fourier transform, for which the Babenko--Beckner result is available. Regrettably, even in the case of polyradial functions on the Heisenberg group we are not able yet to fully answer the question. Indeed, as we discuss in Section \ref{s:weyl}, in this case it seems unlikely that the best Hausdorff--Young constant on the Heisenberg group can be obtained by a direct reduction to the corresponding sharp Euclidean estimate, and new ideas appear to be needed. As for the universal lower bound of Theorem \ref{thm:local-HY}, the intuitive idea behind its proof is that, at smaller and smaller scales, the group structure of a Lie group $G$ looks more and more like the abelian group structure of its Lie algebra $\Lie{g}$, whence $H_p^\mathrm{loc}(G)$ is likely to be related to $H_p(\Lie{g}) = (B_p)^{\dim(G)}$. Indeed, a scaling argument based on this idea readily yields the analogue of Theorem \ref{thm:local-HY} for Young's convolution inequality (see the discussion in Section \ref{s:FLq} below). This appears to have been overlooked in \cite{KR-1978}, where a number of upper bounds for Young constants on Lie groups are proved, which are actually equalities in view of this observation. The additional complication with the Hausdorff--Young inequality is that it involves the $L^q$ norm of the Fourier transform. While it is reasonably clear that, at small scales, the noncommutative convolution on $G$ approximates the commutative convolution on $\Lie{g}$, the same is not so evident for the Fourier transform: indeed, if the group Fourier transform is defined, as it is common, in terms of irreducible unitary representations, then it is not immediately clear how to relate the representation theories of $G$ and $\Lie{g}$ for an arbitrary Lie group $G$, let alone the corresponding Fourier transforms and $L^q$ norms thereof. Here we completely bypass the problem, by characterising the $L^q$ norm of the Fourier transform in terms of an operator norm of a fractional power of an integral operator, acting on functions on $G$: \begin{equation}\label{eq:FTnorm} \| \hat f \|_q^q = \| |L_f \Delta^{1/q}|^q \|_{1 \to \infty}. \end{equation} Here $L_f$ is the operator of convolution on the left by $f$ and $\Delta$ is the operator of multiplication by the modular function of $G$. A transplantation argument, not dissimilar from those in \cite{mitjagin_1974,KST-1982,martini_joint_2017}, allows us to relate the operator $L_f \Delta^{1/q}$ on $G$ to its counterpart on $\Lie{g}$ and obtain the desired lower bound. Although it might be evident to some experts in noncommutative integration, we are not aware of the characterisation \eqref{eq:FTnorm} being explicitly observed before. What is interesting about \eqref{eq:FTnorm} is that it allows one to access the $L^q$ norm of the Fourier transform through properties of a more ``geometric'' convolution-multiplication operator on $G$, which appears to be more tractable. As a matter of fact, when dealing with convolution, one can use induction-on-scales methods to completely determine the best local constants for the Young convolution inequality on any Lie group $G$; this remarkable result has been recently proved in \cite{BBBCF_2018}, as a corollary of a more general result for nonlinear Brascamp--Lieb inequalities. It would be interesting to know whether similar methods could be applied to the Hausdorff--Young inequality on noncommutative Lie groups as well. \subsection*{Plan of the paper} In Section \ref{s:FLq} we discuss the definition of the $L^q$ norm of the Fourier transform for an arbitrary Lie group, by comparing a number of definitions available in the literature, and prove the characterisation \eqref{eq:FTnorm}; we also present a proof of the universal lower bound of Theorem \ref{thm:local-HY}, as well as its analogue for the Young convolution inequality, and discuss relations between best constants for Young and Hausdorff--Young inequalities. The sharp local central Hausdorff--Young inequality for arbitrary compact Lie groups (Theorem \ref{thm:local-central-HY-compact-Lie}) is proved in Section \ref{s:compact}; to better explain the underlying idea without delving into technicalities, the proof of the abelian case (Theorem \ref{thm:local-HY-Tn}) is briefly revisited in Section \ref{s:torus}. Finally, in Section \ref{s:weyl} we discuss the relations between Hausdorff--Young constants for the Heisenberg group and the Weyl transform and prove Theorem \ref{thm:local-HY-weyl}, together with the weighted inequality \eqref{eq:weighted_HY_weyl} for polyradial functions. \section{\texorpdfstring{$L^q$}{Lq} norm of the Fourier transform}\label{s:FLq} Let $G$ be a Lie group (or, more generally, a separable locally compact group) with a fixed left Haar measure. In order to discuss best Hausdorff--Young constants in this generality, we first need to clarify what is meant by the ``Fourier transform'' in this setting and how Hausdorff--Young inequalities --- even the endpoint ones, such as the Plancherel formula --- can be stated in this context. A common way to generalise the Fourier transformation to this setting exploits irreducible unitary representations of $G$ (see, for example, \cite{lipsman_1974} or \cite[Chapter 7]{folland_course_1995} for a survey). Namely, let $\widehat G_\unit$ be the ``unitary dual'' of $G$, that is, the set of (equivalence classes of) irreducible unitary representations of $G$, endowed with the Fell topology and the Mackey Borel structure. The (unitary) Fourier transform $\Four_\unit f$ of a function $f \in L^1(G)$ is then defined as the operator-valued function on $\widehat G_\unit$ given by \[ \widehat G_\unit \ni \pi \mapsto \pi(f) = \int_G f(x) \pi(x) \,dx \in \Lin(\Hilb_\pi); \] here $\Lin(\Hilb_\pi)$ denotes the space of bounded linear operators on the Hilbert space $\Hilb_\pi$ on which the representation $\pi$ acts, and integration is with respect to the Haar measure. In case $G$ is unimodular and type I (this includes the cases where $G$ is abelian or compact), the Plancherel formula can be stated in the form \begin{equation}\label{eq:plancherel_rep} \|f\|^2_{L^2(G)} = \int_{\widehat G_\unit} \| \pi(f) \|_{\HS(\Hilb_\pi)}^2 \,d\pi \end{equation} for all $f \in L^1 \cap L^2(G)$. Here $\HS(\Hilb_\pi)$ denotes the space of Hilbert--Schmidt operators on $\Hilb_\pi$, and integration on $\widehat G_\unit$ is with respect to a suitable measure, called the Plancherel measure, which is uniquely determined by the above formula; in addition, the Fourier transformation $f \mapsto \Four_\unit f$ extends to an isometric isomorphism between $L^2(G)$ and the direct integral $L^2_\unit(\widehat G) := \int^\oplus_{\widehat G_\unit} \HS(\Hilb_\pi) \,d\pi$. Interpolation then leads to the Hausdorff--Young inequality \begin{equation}\label{eq:Lq_rep} \|\Four_\unit f\|_{L^{p'}_\unit(\widehat G)} := \left(\int_{\widehat G_\unit} \| \pi(f) \|_{\Sch^{p'}(\Hilb_\pi)}^{p'} \,d\pi\right)^{1/p'} \leq C \|f\|_{L^p(G)} \end{equation} when $1 < p < 2$, where $C = 1$; here, for all $q \in [1,\infty]$, $\Sch^q(\Hilb_\pi)$ denotes the $q$th Schatten class of operators on $\Hilb_\pi$, and the operator-valued $L^q$-spaces $L^q_\unit(\widehat G)$ are defined in terms of measurable fields of operators as in \cite{lipsman_1974}. The fact that the spaces $L^q_\unit(\widehat G)$ constitute a complex interpolation family, that is, \begin{equation}\label{eq:interpol_unitLp} [L^{q_0}_\unit(\widehat G),L^{q_1}_\unit(\widehat G)]_\theta = L^q_\unit(\widehat G) \end{equation} with equal norms for $q_0,q_1,q \in [1,\infty]$, $\theta \in (0,1)$, $1/q = (1-\theta)/q_0 + \theta/q_1$, readily follows from standard interpolation results for vector-valued Lebesgue spaces and Schatten classes (see, for example, \cite{triebel_1978,hytonen_2016,Pisier-Xu-2003}) and the structure of the measurable field of separable Hilbert spaces $\pi \mapsto \Hilb_\pi$ \cite[Proposition 7.19]{folland_course_1995}. In the case where $G$ is not unimodular, under suitable type I assumptions it is possible to prove a Plancherel formula similar to \eqref{eq:plancherel_rep}, where the right-hand side is adjusted by means of ``formal dimension operators'' \cite{tatsuuma_1972,kleppner_lipsman_1972,kleppner_lipsman_1973,duflo_moore_1976,Fuehr_2005}. Analogous modifications of \eqref{eq:Lq_rep} lead to a version of the Hausdorff--Young inequality that has been studied in a number of works \cite{Eymard-Terp-1979,Russo-1979,Inoue-1992,Fuehr_2006,Baklouti-Ludwig-Scuto-Smaoui-2007}. When $G$ is not type I, the above approach to the Plancherel formula based on irreducible unitary representation theory does not work as neatly. This however does not prevent one from studying the Hausdorff--Young inequality. Indeed, what is possibly the first appearance in the literature of the Hausdorff--Young inequality in a noncommutative setting, that is, the work of Kunze \cite{Kun-1958} for arbitrary unimodular locally compact groups (not necessarily of type I), does not express the Fourier transform in terms of irreducible unitary representations, but uses instead the theory of noncommutative integration (the same theory was used in earlier works of Mautner \cite{mautner_unitary_1950} and Segal \cite{segal-1950} to express the Plancherel formula). This point of view was subsequently developed by Terp \cite{terp_1980} to cover the case of non-unimodular groups and more recently has been further extended to the context of locally compact quantum groups \cite{caspers_2013,cooney_2010}. One way of thinking of noncommutative $L^q$ spaces is as complex interpolation spaces between a von Neumann algebra $M$ and its predual $M_*$ (which play the role of $L^\infty$ and $L^1$ respectively) \cite{terp_1982,Kosaki-1984,Izumi-1997,Pisier-Xu-2003}. In general this requires establishing a ``compatibility'' between $M$ and $M_*$, which may involve a number of choices, but in our case there appears to be a natural way to proceed (see also \cite{forrest_2011,daws_2011}). Namely, the von Neumann algebra $\VN(G)$ of $G$ (that is, the weak${}^*$-closed $*$-subalgebra of $\Lin(L^2(G))$ of the operators which commute with right translations) can be identified with the space $\Cv^2(G)$ of left convolutors of $L^2(G)$, that is, those distributions on $G$ which are left convolution kernels of $L^2(G)$-bounded operators. Moreover, the predual $\VN(G)_*$ can be identified with the Fourier algebra $A(G)$, an algebra of continuous functions on $G$ defined by Eymard \cite{Eymard-1964} for arbitrary locally compact groups $G$. Now $A(G)$ and $\Cv^2(G)$ are naturally compatible as spaces of distributions on $G$ (see \cite[Propositions (3.26) and (3.27)]{Eymard-1964}), so we can use complex interpolation to define Fourier--Lebesgue spaces of distributions on $G$: for $q\in[1,\infty]$, we set \[ \Four L^q(G) = \begin{cases} A(G) &\text{if } q=1,\\ \Cv^2(G) &\text{if } q=\infty,\\ [A(G),\Cv^2(G)]_{1-1/q} &\text{if } 1 < q < \infty. \end{cases} \] One can check that this definition corresponds to Izumi's left $L^p$ spaces \cite{Izumi-1997,Izumi-1998} for the von Neumann algebra $\VN(G)$ with respect to the Plancherel weight, and therefore it matches the construction given in \cite{caspers_2013,cooney_2010} for quantum groups. In particular $\Four L^2(G) = L^2(G)$ with equality of norms (see \cite[Section 5]{Izumi-1998} and \cite[Proposition 2.21(iii)]{caspers_2013}; this corresponds to the Plancherel theorem), while clearly $L^1(G) \subseteq \Cv^2(G)$ with norm-decreasing embedding. Interpolation then leads to the following formulation of the Hausdorff--Young inequality: $L^p(G) \subseteq \Four L^{p'}(G)$ and \begin{equation}\label{eq:HY_FLq} \| f \|_{\Four L^{p'}(G)} \leq C \| f\|_{L^p(G)} \end{equation} where $C= 1$ and $p \in [1,2]$. We then define the $L^p$ Hausdorff--Young constant $H_p(G)$ on the group $G$ as the minimal constant $C$ for which \eqref{eq:HY_FLq} holds for all $f \in L^p(G)$. Similarly, if $U$ is a neighbourhood of the identity in $G$, we let $H_p(G;U)$ be the minimal constant $C$ in \eqref{eq:HY_FLq} when $f$ is constrained to have support in $U$, and define the local $L^p$ Hausdorff--Young constant $H_p^\mathrm{loc}(G)$ as the infimum of the constants $H_p(G;U)$ where $U$ ranges over the neighbourhoods of the identity of $G$. The approach to Hausdorff--Young constants via $\Four L^q$ spaces is consistent with the unitary Fourier transformation approach described above, when the latter is applicable. Indeed, as discussed in \cite[Theorems 2.1 and 3.1]{lipsman_1974}, in the case where $G$ is unimodular and type I, the unitary Fourier transformation $\Four_\unit$ induces isometric isomorphisms $\Cv^2(G) \cong L^\infty_\unit(\widehat G)$ and $A(G) \cong L^1_\unit(\widehat G)$, besides the Plancherel isomorphism $L^2(G) \cong L^2_\unit(\widehat G)$ (analogous results in the nonunimodular case can be found in \cite[Theorems 3.48 and 4.12]{Fuehr_2005}); so by interpolation $\Four_u$ induces an isometric isomorphism between $\Four L^q(G)$ and $L^q_\unit(\widehat G)$ for all $q \in [1,\infty]$. Hence defining Hausdorff--Young constants in terms of the inequality \eqref{eq:Lq_rep} would lead to the same constants $H_p(G)$ and $H_p^\mathrm{loc}(G)$ as those we have defined in terms of $\Four L^q$ spaces. On the other hand, the approach via $\Four L^q$ spaces does not require type I assumptions, or even separability, and can be applied to every locally compact group $G$. There is an alternative characterisation of the noncommutative $L^q$ spaces associated to $\VN(G)$, namely as certain spaces $L^q_{\VN}(\widehat G)$ of (closed, possibly unbounded) operators on $L^2(G)$. This characterisation, which is that originally used in the works of Kunze and Terp on the Hausdorff--Young inequality, corresponds to Hilsum's approach to noncommutative $L^q$ spaces \cite{Hilsum-1981} based on Connes's ``spatial derivative'' construction \cite{Connes-1980} (the work of Kunze is actually based on an earlier version of the theory \cite{dixmier_1953,segal-1953} that only applies to semifinite von Neumann algebras). We will not enter into the details of this construction and only recall two important properties. First, if the operator $T$ belongs to $L^q_{\VN}(\widehat G)$ for some $q \in [1,\infty)$, then $|T|^q = (T^* T)^{q/2}$ belongs to $L^1_{\VN}(\widehat G)$ and \begin{equation}\label{eq:LqL1} \|T\|_{L^q_{\VN}(\widehat G)}^q = \||T|^q\|_{L^1_{\VN}(\widehat G)}. \end{equation} Moreover, for all $q \in [1,\infty]$, an isometric isomorphism from $\Four L^q(G)$ to $L^q_{\VN}(\widehat G)$ is given by \begin{equation}\label{eq:LqFT} f \mapsto L_f \Delta^{1/q}, \end{equation} where $L_f$ is the left-convolution operator by $f$, and we identify the modular function $\Delta$ of $G$ with the corresponding multiplication operator (see \cite[Proposition 2.21(ii)]{caspers_2013}). Recall that convolution on $G$ is given by \[ L_f \phi(x) = f * \phi(x) = \int_G f(xy) \, \phi(y^{-1}) \,dy, \] at least when $f$ and $\phi$ are in $C_c(G)$. Note that, when $q=p'$, \eqref{eq:LqFT} matches the definitions by Kunze and by Terp of the $L^p$ Fourier transformation $\Four_p : L^p(G) \to L^{p'}_{\VN}(\widehat G)$ for $p \in [1,2]$ \cite{Kun-1958,terp_1980}. In other words, the $L^p$ Fourier transformation $\Four_p : L^p(G) \to L^{p'}_{\VN}(\widehat G)$ factorises as the inclusion map $L^p(G) \to \Four L^{p'}(G)$ and the isometric isomorphism $\Four L^{p'}(G) \to L^{p'}_{\VN}(\widehat G)$, whence the compatibility with the Kunze--Terp approach of the above definition of the best Hausdorff--Young constants based on \eqref{eq:HY_FLq}. Another consequence of the above discussion is the following characterisation of the $\Four L^q(G)$ norm in terms of a more ``concrete'' operator norm. \begin{proposition}\label{prp:fouriernorm} For all $q \in [1,\infty)$ and $f \in \Four L^q(G)$, \begin{equation}\label{eq:nu_fouriernorm} \| f \|_{\Four L^q(G)} = \| |L_f \Delta^{1/q}|^q \|_{L^1(G) \to L^\infty(G)}^{1/q}. \end{equation} \end{proposition} \begin{proof} By \eqref{eq:LqL1} and \eqref{eq:LqFT}, \[ \| f \|_{\Four L^q(G)} = \| L_f \Delta^{1/q} \|_{L^q_{\VN}(\widehat G)} = \| |L_f \Delta^{1/q}|^q \|_{L^1_{\VN}(\widehat G)}^{1/q} = \| g \|_{A(G)}^{1/q}, \] where $g \in A(G)$ satisfies $L_g \Delta = |L_f \Delta^{1/q}|^q$. On the other hand, the operator $L_g \Delta$ is given by \[ L_g \Delta \phi(x) = \int_G g(xy) \, \Delta(y^{-1}) \, \phi(y^{-1}) \,dy = \int_G g(xy^{-1}) \, \phi(y) \,dy; \] since $L_g \Delta = |L_f \Delta^{1/q}|^q$ is a positive operator, the kernel $g$ must be a function of positive type (see, for example, \cite[Section 3.3]{folland_course_1995}), whence \[ \|g\|_{A(G)} = g(e) = \|g\|_\infty = \| L_g \Delta \|_{L^1(G) \to L^\infty(G)} \] and we are done. \end{proof} A classical way of accessing Hausdorff--Young constants is through their relations with best constants in the Young convolution inequalities. Recall that, for a possibly nonunimodular group $G$, the $k$-linear version of Young's inequality takes the following form: for all $p_1,\dots,p_k,r \in [1,\infty]$ such that $\sum_{j=1}^k 1/p_j' = 1/r'$, \begin{equation}\label{eq:nu_young} \Bigl\| \bigast_{j=1}^k (f_j \Delta^{\sum_{l=1}^{j-1} 1/p_l'}) \Bigr\|_{L^r(G)} \leq C \prod_{j=1}^k \|f_j\|_{L^{p_j}(G)} \end{equation} where $C \leq 1$ (see \cite[Lemma 1.1]{terp_1980}, or \cite[Corollary 2.3]{KR-1978} where the inequality is written for the right Haar measure). As in the case of the Hausdorff--Young inequality, we can define the Young constant $Y_{p_1,\dots,p_k}(G)$ for $G$ as the smallest constant $C$ for which \eqref{eq:nu_young} holds for all $f_1 \in L^{p_1}(G), \dots, f_k \in L^{p_k}(G)$, as well as the localised versions $Y_{p_1,\dots,p_k}(G;U)$ for neighbourhoods $U$ of the identity of $G$ (corresponding to the constraint $\supp f_1, \dots, \supp f_k \subseteq U$) and $Y^\mathrm{loc}_{p_1,\dots,p_k}(G)$. Note that the above Young inequality \eqref{eq:nu_young} is ``dual'' to the following H\"older-type inequality for $\Four L^p$-spaces: for all $p_1,\dots,p_k,r \in [1,\infty]$ such that $\sum_{j=1}^k 1/p_j = 1/r$, \begin{equation}\label{eq:nu_nchoelder} \Bigl\| \bigast_{j=1}^k (f_j \Delta^{\sum_{l=1}^{j-1} 1/p_l}) \Bigr\|_{\Four L^r(G)} \leq \prod_{j=1}^k \|f_j\|_{\Four L^{p_j}(G)}; \end{equation} this is a rephrasing of H\"older's inequality for Hilsum's noncommutative $L^p$ spaces, \[ \| T_1 \cdots T_k \|_{L^r_{\VN}(\widehat G)} \leq \prod_{j=1}^k \| T_j \|_{L^{p_j}_{\VN}(\widehat G)} \] \cite[Proposition 8]{Hilsum-1981}, via the isomorphism \eqref{eq:LqFT} from $\Four L^q(G)$ to $L^q_{\VN}(\widehat G)$ and the identities \begin{equation}\label{eq:conv_modular} \Delta^{\alpha} (f*g) = (\Delta^\alpha f) * (\Delta^\alpha g) \qquad\text{and}\qquad L_{\Delta^{\alpha} f} = \Delta^{\alpha} L_f \Delta^{-\alpha}, \end{equation} valid for all $\alpha \in \mathbb{C}$. Let us also recall that \begin{equation}\label{eq:adj_involution} L_{f^*} = L_{f}^*, \end{equation} where $f \mapsto f^*$ is the isometric conjugate-linear involution of $L^1(G)$ given by \[ f^*(x) = \Delta^{-1}(x) \, \overline{f(x^{-1})}. \] The proposition below summarises a number of relations between Young and Hausdorff--Young constants that can be found in the literature, at least in particular cases (see, for example, \cite{Bec-1975} and \cite{KR-1978}), as well as corresponding local versions. \begin{proposition}\label{prop:basic} Let $G$ be a locally compact group. \begin{enumerate}[label=(\roman*)] \item\label{en:yhineq} For all $p_1,\dots,p_k,q \in [1,2]$ such that $\sum_j 1/p_j' = 1/q$, \begin{align*} Y_{p_1,\dots,p_k}(G) &\leq H_{q}(G) \, H_{p_1}(G) \cdots H_{p_k}(G), \\ Y_{p_1,\dots,p_k}^\mathrm{loc}(G) &\leq H_{q}^\mathrm{loc}(G) \, H_{p_1}^\mathrm{loc}(G) \cdots H_{p_k}^\mathrm{loc}(G). \end{align*} \item\label{en:yheq} For all $p \in [1,2)$ such that $p'=2k$, $k \in \mathbb{Z}$, if $p_1=\dots=p_k=p$, then \begin{align*} H_{p}(G) &= Y_{p_1,\dots,p_k}(G)^{1/k}, \\ H_{p}^\mathrm{loc}(G) &= Y_{p_1,\dots,p_k}^\mathrm{loc}(G)^{1/k}. \end{align*} \item\label{en:ext} If $N$ is a closed normal subgroup of $G$, then, for all $p_1,\dots,p_k \in [1,\infty]$ such that $\sum_{j=1}^k 1/p_j' \in [0,1]$, \begin{align*} Y_{p_1,\dots,p_k}(G) &\leq Y_{p_1,\dots,p_k}(N) \, Y_{p_1,\dots,p_k}(G/N),\\ Y_{p_1,\dots,p_k}^\mathrm{loc}(G) &\leq Y_{p_1,\dots,p_k}^\mathrm{loc}(N) \, Y_{p_1,\dots,p_k}^\mathrm{loc}(G/N), \end{align*} with equality when $G \cong N \times (G/N)$. \end{enumerate} \end{proposition} \begin{proof} \ref{en:yhineq}. For all $f_1,\dots,f_k,g \in C_c(G)$, by \eqref{eq:HY_FLq} and \eqref{eq:nu_nchoelder}, \[\begin{split} \left\langle \bigast_{j=1}^k (f_j \Delta^{\sum_{l=1}^{j-1} 1/p_l'}) , g \right\rangle &\leq \left\| \bigast_{j=1}^k (f_j \Delta^{\sum_{l=1}^{j-1} 1/p_l'}) \right\|_{\Four L^q} \| g \|_{\Four L^{q'}} \\ &\leq \| f_1\|_{\Four L^{p_1'}} \cdots \|f_k\|_{\Four L^{p_k'}} \|g\|_{\Four L^{q'}} \\ &\leq H_q(G) H_{p_1}(G) \cdots H_{p_k}(G) \|f_1\|_{L^{p_1}} \cdots \|f_k\|_{L^{p_k}} \|g\|_{L^q}, \end{split}\] which proves that \[ \left\|\bigast_{j=1}^k (f_j \Delta^{\sum_{l=1}^{j-1} 1/p_l'}) \right\|_{L^{q'}} \leq H_q(G) H_{p_1}(G) \cdots H_{p_k}(G) \|f_1\|_{L^{p_1}} \cdots \|f_k\|_{L^{p_k}}, \] that is, $Y_{p_1,\dots,p_k}(G) \leq H_{q}(G) \, H_{p_1}(G) \cdots H_{p_k}(G)$. Note now that, if $f_1,\dots,f_k$ are supported in a neighbourhood $U$ of the identity, then $\bigast_{j=1}^k (f_j \Delta^{\sum_{l=1}^{j-1} 1/p_l'})$ is supported in $U^k$ and, to estimate its $L^{q'}$ norm, it is enough to test it against functions $g$ that are also supported in $U^k$; the same argument as above then also gives \[ Y_{p_1,\dots,p_k}(G;U) \leq H_{q}(G;U^k) \, H_{p_1}(G;U) \cdots H_{p_k}(G;U) \] and $Y_{p_1,\dots,p_k}^\mathrm{loc}(G) \leq H_{q}^\mathrm{loc}(G) \, H_{p_1}^\mathrm{loc}(G) \cdots H_{p_k}^\mathrm{loc}(G)$. \ref{en:yheq}. Part \ref{en:yhineq} gives us the inequality $H_{p}(G) \geq Y_{p_1,\dots,p_k}(G)^{1/k}$ and its local version. On the other hand, for all $f \in C_c(G)$, if we define $\tilde f = \Delta^{1/p'} f^*$, then \[ \|\tilde f\|_p = \|f\|_p \] and, by \eqref{eq:conv_modular} and \eqref{eq:adj_involution}, \[ L_{\tilde f} \Delta^{1/p'} = (L_f \Delta^{1/p'})^*. \] For all $j=1,\dots,k$, let $f_j$ be either $\tilde f$ or $f$, according to whether $k-j$ is odd or even, and define $g = \bigast_{j=1}^k (f_j \Delta^{(j-1)/p'})$. Then, since $p'=2k$, \[\begin{split} |L_f \Delta^{1/p'}|^{p'} &= [(L_f \Delta^{1/p'})^* (L_f \Delta^{1/p'}) ]^k \\ &= (L_{\tilde f} \Delta^{1/p'}) (L_{f} \Delta^{1/p'}) \cdots (L_{\tilde f} \Delta^{1/p'}) (L_{f} \Delta^{1/p'}) \\ &= |(L_{f_1} \Delta^{1/p'}) \cdots (L_{f_k} \Delta^{1/p'})|^2 \end{split}\] and, by \eqref{eq:conv_modular}, \[ (L_{f_1} \Delta^{1/p'}) \cdots (L_{f_k} \Delta^{1/p'}) = L_g \Delta^{1/2}. \] So $|L_f \Delta^{1/p'}|^{p'} = |L_g \Delta^{1/2}|^2$ and, by \eqref{eq:nu_fouriernorm} and \eqref{eq:nu_young}, \[ \| f \|_{\Four L^{p'}}^{p'} = \| g \|_{\Four L^2}^2 = \| g \|_{L^2}^2 \leq Y_{p_1,\dots,p_k}(G)^2 \|f_1\|_{L^p}^2 \cdots \| f_k \|_{L^p}^2 = Y_{p_1,\dots,p_k}(G)^2 \|f\|_{L^p}^{p'}, \] which gives the inequality $H_{p}(G) \leq Y_{p_1,\dots,p_k}(G)^{1/k}$. The same argument also gives $H_{p}(G;U) \leq Y_{p_1,\dots,p_k}(G;U)^{1/k}$ and $H_{p}^\mathrm{loc}(G) \leq Y_{p_1,\dots,p_k}^\mathrm{loc}(G)^{1/k}$. \ref{en:ext}. The inequalities are proved by a simple extension of Klein and Russo's argument for the case of semidirect products \cite[proof of Lemma 2.4]{KR-1978}, using the ``measure disintegration'' in \cite[Theorem (2.49)]{folland_course_1995}. In the case of direct products, equalities follow by testing on tensor product functions (see \cite[Lemma 5]{Bec-1975}). \end{proof} The next lemma contains the fundamental approximation results that allow us to relate Hausdorff--Young constants on a Lie group $G$ and on its Lie algebra $\Lie{g}$ by means of a ``transplantation'' or ``blow-up'' technique. The Lie algebra $\Lie{g}$ will be considered as an abelian group with addition, and the Lebesgue measure on $\Lie{g}$ is normalised so that the Jacobian determinant of the exponential map $\exp : \Lie{g} \to G$ is equal to $1$ at the origin. The context will make clear whether the notation for convolution, involution and convolution operators ($f*g$, $f^*$, $L_f$) refers to the group structure of $G$ or the abelian group structure of $\Lie{g}$. Denote by $C_{\pg}([0,\infty))$ the space of continuous functions $\Phi : [0,\infty) \to \mathbb{C}$ with at most polynomial growth, that is, $|\Phi(u)| \leq C(1+u)^N$ for some $C,N \in (0,\infty)$ and all $u \in [0,\infty)$. \begin{lemma}\label{lem:nu_localisation} Let $G$ be a Lie group with Lie algebra $\Lie{g}$ of dimension $n$, and let $\exp : \Lie{g} \to G$ be the exponential map. Let $\Omega$ be an open neighbourhood of the origin in $\Lie{g}$ such that $\Omega = -\Omega$ and $\exp|_\Omega : \Omega \to \exp(\Omega)$ is a diffeomorphism. For all $f \in C_c(\Lie{g})$, $\lambda \in (0,\infty)$, $\alpha \in \mathbb{R}$ and $p \in [1,\infty]$, define $f^{\lambda,p,\alpha} : G \to \mathbb{C}$ by \begin{equation}\label{eq:nu_associatedp} f^{\lambda,p,\alpha}(x) = \begin{cases} \lambda^{-n/p} \Delta(x)^{-\alpha} f(\lambda^{-1} \exp|_\Omega^{-1}(x)) &\text{if } x \in \exp(\Omega),\\ 0 &\text{otherwise.} \end{cases} \end{equation} Set also $f^{\lambda,p} = f^{\lambda,p,0}$. Then the following hold. \begin{enumerate}[label=(\roman*)] \item\label{en:nu_asspnorm} For all $f \in C_c(\Lie{g})$, $\alpha \in \mathbb{R}$ and $p \in [1,\infty]$, \begin{equation}\label{eq:nu_asspnormest} \|f^{\lambda,p,\alpha}\|_{L^p(G)} \leq C_{\alpha,p,\Omega} \, \|f\|_{L^p(\Lie{g})} \end{equation} for all $\lambda \in (0,\infty)$, and \begin{equation}\label{eq:nu_asspnormlim} \|f^{\lambda,p,\alpha}\|_{L^p(G)} \to \|f\|_{L^p(\Lie{g})} \end{equation} as $\lambda \to 0$. \item\label{en:nu_assinnerprod} For all $k \in \mathbb{N}$, $\alpha_1,\dots,\alpha_k,\beta\in \mathbb{R}$, $f_1,\dots,f_k,g \in C_c(\Lie{g})$, \begin{equation}\label{eq:nu_innerconv} \langle f_1^{\lambda,1,\alpha_1} * \cdots * f_k^{\lambda,1,\alpha_k} , g^{\lambda,\infty,\beta} \rangle_{L^2(G)} \to \langle f_1 * \cdots * f_k , g \rangle_{L^2(\Lie{g})} \end{equation} as $\lambda \to 0$. \item\label{en:nu_assfunccalc} For all $\alpha \in \mathbb{R}$, $f,g,h \in C_c(\Lie{g})$, $\Phi \in C_{\pg}([0,\infty))$, \begin{equation}\label{eq:nu_continuousfc} \langle \Phi(\Delta^\alpha L_{(f^{\lambda,1})^* * f^{\lambda,1}} \Delta^\alpha) g^{\lambda,2}, h^{\lambda,2}\rangle_{L^2(G)} \to \langle\Phi(L_{f^* * f}) g, h\rangle_{L^2(\Lie{g})} \end{equation} as $\lambda \to 0$. \item\label{en:nu_asspower} For all $\alpha \in \mathbb{R}$, $f,g,h \in C_c(\Lie{g})$ and $q \in [0,\infty)$, \[ \lambda^{-n(q-1)} \langle |L_{f^{\lambda,\infty}} \Delta^\alpha|^{q} g^{\lambda,1}, h^{\lambda,1}\rangle_{L^2(G)} \to \langle|L_{f}|^{q} g, h\rangle_{L^2(\Lie{g})} \] as $\lambda \to 0$. \end{enumerate} \end{lemma} \begin{proof} Let $J : \Lie{g} \to \mathbb{R}$ denote the modulus of the Jacobian determinant of $\exp$, and define $\Delta_e : \Lie{g} \to (0,\infty)$ to be $\Delta \circ \exp$ \ref{en:nu_asspnorm}. Note that \[ \|f^{\lambda,p,\alpha}\|_{p}^{p} = \lambda^{-n} \int_\Omega |f(\lambda^{-1} X)|^{p} (J\Delta_e^{-\alpha p})(X) \,dX = \int_{\lambda^{-1}\Omega} |f(X)|^{p} (J \Delta_e^{-\alpha p})(\lambda X) \,dX. \] From this, \eqref{eq:nu_asspnormest} follows (with $C_{\alpha,p,\Omega}^p = \sup_{\Omega} J \Delta_e^{-\alpha p}$), and \eqref{eq:nu_asspnormlim} follows as well because $f$ is compactly supported and $\lim_{X \to 0} (J \Delta_e^{-\alpha p})(X) = J(0) \Delta(e)^{-\alpha p} = 1$. \ref{en:nu_assinnerprod}. By the Baker--Campbell--Hausdorff formula, \[ \exp(X_1) \cdots \exp(X_k) = \exp(X_1 + \dots + X_k + B(X_1,\dots,X_k)), \] where $B(X_1,\dots,X_k) = \sum_{m \geq 2} B_m(X_1,\dots,X_k)$ and, for all $m \geq 2$, $B_m(X_1,\dots,X_k)$ is a homogeneous polynomial function of $X_1,\dots,X_k$ of degree $m$; indeed we can find a sufficiently small neighbourhood $\tilde\Omega \subseteq \Omega$ of the origin in $\Lie{g}$ so that, if $X_1,\dots,X_k \in \tilde\Omega$, then $X_1 + \dots + X_k + B(X_1,\dots,X_k) \in \Omega$. Note that \begin{multline*} \langle f_1^{\lambda,1,\alpha_1} * \cdots * f_k^{\lambda,1,\alpha_k} , g^{\lambda,\infty,\beta} \rangle_{L^2(G)} \\ = \int_{G^k} f_1^{\lambda,1,\alpha_1}(x_1) \, \cdots \, f_k^{\lambda,1,\alpha_k}(x_k) \, \overline{g^{\lambda,\infty,\beta}(x_1\cdots x_k)} \,dx_1 \dots \,dx_k \end{multline*} If $\lambda$ is sufficiently small that $\bigcup_{j=1}^k \lambda \supp f_j \subseteq \exp(\tilde\Omega)$, then the last integral may be rewritten as \[ \int_{\Lie{g}^k} \bar g\Biggl(\sum_{j=1}^k X_j + \lambda^{-1}B(\lambda X_1,\dots,\lambda X_k)\Biggr) \prod_{j=1}^k ( f_j(X_j) (J \Delta_e^{-\alpha_j-\beta})(\lambda X_j) ) \,dX_1\cdots \,dX_k. \] Since $\lambda^{-1} B(\lambda X_1,\dots,\lambda X_k) = \lambda \sum_{m \geq 2} \lambda^{m-2} B_m(X_1,\dots,X_k)$ tends to $0$ as $\lambda \to 0$, the last integral tends to $\langle f_1 * \cdots * f_k , g \rangle_{L^2(\Lie{g})}$. \ref{en:nu_assfunccalc}. Note first that $\Delta^\alpha L_{(f^{\lambda,1})^* * f^{\lambda,1}} \Delta^\alpha$ is a nonnegative self-adjoint operator on $L^2(G)$ (which may be unbounded when $G$ is nonunimodular) and that, for all $N \in \mathbb{N}$, the $L^2$-domain of $(\Delta^\alpha L_{(f^{\lambda,1})^* * f^{\lambda,1}} \Delta^\alpha)^N$ contains all compactly supported functions in $L^2(G)$, so the left-hand side of \eqref{eq:nu_continuousfc} is well-defined. Note moreover that \begin{equation}\label{eq:nu_star} (f^{\lambda,p,\alpha})^* = (f^*)^{\lambda,p,1-\alpha} \end{equation} whence, by \eqref{eq:conv_modular}, \[\begin{split} &\langle (\Delta^\alpha L_{(f^{\lambda,1})^* * f^{\lambda,1}} \Delta^\alpha)^N g^{\lambda,2}, h^{\lambda,2} \rangle_{L^2(G)} \\ &= \left\langle \left(\bigast_{j=1}^N ((f^*)^{\lambda,1,1-(2j-1)\alpha} * f^{\lambda,1,-(2j-1)\alpha}) \right) * g^{\lambda,1,-2N\alpha}, h^{\lambda,\infty} \right\rangle. \end{split}\] So, in the case where $\Phi(u) = u^N$ for some $N \in \mathbb{N}$, \eqref{eq:nu_continuousfc} follows from \eqref{eq:nu_innerconv}. Note that, by shrinking $\Omega$ if necessary, we may assume that $\Omega$ and $\exp(\Omega)$ have compact closures in $\Lie{g}$ and $G$, and moreover the topological boundary of $\exp(\Omega)$ has null Haar measure (indeed shrinking $\Omega$ does not change the left-hand side of \eqref{eq:nu_continuousfc} for $\lambda$ sufficiently small). As in \cite[proof of Theorem 5.2]{martini_joint_2017}, we can now extend the diffeomorphism $\phi := \exp|_\Omega^{-1} : \exp(\Omega) \to \Omega$ to a diffeomorphism $\phi_* : U \to V$, where $U$ and $V$ are open sets in $G$ and $\Lie{g}$ containing $\exp(\Omega)$ and $\Omega$, and moreover $G \setminus U$ has null Haar measure. Finally, let $J_* : V \to (0,\infty)$ be the density of the push-forward via $\phi_*$ of the Haar measure with respect to the Lebesgue measure (so $J_* = J$ on $\Omega$), and define an isometric isomorphism $\Psi : L^2(G) \to L^2(V)$ by \[ \Psi(F) = (F \circ \phi_*^{-1}) \, J_*^{1/2} . \] Since $A_\lambda := \Delta^\alpha L_{(f^{\lambda,1})^* * f^{\lambda,1}} \Delta^\alpha$ is a self-adjoint operator on $L^2(G)$, we can define a self-adjoint operator $\tilde A_\lambda$ on $L^2(\Lie{g}) = L^2(V) \oplus L^2(\Lie{g} \setminus V)$ by \[ \tilde A_\lambda = \begin{pmatrix} \Psi A_\lambda \Psi^{-1} & 0 \\ 0 & 0 \end{pmatrix} \] and another self-adjoint operator $\hat A_\lambda$ on $L^2(\Lie{g})$ by $\hat A_\lambda = T_\lambda^{-1} \tilde A_\lambda T_\lambda$, where $T_\lambda$ is the isometry on $L^2(\Lie{g})$ defined by \[ T_\lambda f(X) = \lambda^{-n/2} f(X/\lambda). \] It is now not difficult to check that, for all $\Phi \in C_{\pg}([0,\infty))$ and $g,h \in C_c(\Lie{g})$, \begin{equation}\label{eq:nu_conj_fcinner} \langle \Phi(\hat A_\lambda) g, h \rangle_{L^2(\Lie{g})} = \langle \Phi(A_\lambda) g^{\lambda,2}, h^{\lambda,2} \rangle_{L^2(G)} \end{equation} for all $\lambda$ sufficiently small that $\supp T_\lambda g \cup \supp T_\lambda h \subseteq \Omega$. For all $N \in \mathbb{N}$, from the cases $\Phi(u) = u^N$ and $\Phi(u) = u^{2N}$ of \eqref{eq:nu_continuousfc} and \eqref{eq:nu_conj_fcinner} it follows that, for all $g,h \in C_c(\Lie{g})$, \begin{equation}\label{eq:nu_conj_conv} \langle \hat A_\lambda^N g, h \rangle_{L^2(\Lie{g})} \to \langle A^N g,h \rangle_{L^2(\Lie{g})}, \qquad \| \hat A_\lambda^N g \|_{L^2(\Lie{g})} \to \| A^N g \|_{L^2(\Lie{g})} \end{equation} as $\lambda \to 0$, where $A := L_{f^* * f}$. In particular, from this and the density of $C_c(\Lie{g})$ in $L^2(\Lie{g})$ it is not difficult to conclude that, for all $g \in C_c(\Lie{g})$, \begin{equation}\label{eq:nu_strongpower} \hat A_\lambda^N g \to A^N g \end{equation} in $L^2$-norm as $\lambda \to 0$ \cite[Proposition 3.32]{brezis}. Since $A$ is a bounded self-adjoint operator on $L^2(\Lie{g})$, $C_c(\Lie{g})$ is a core for $A$ and \cite[Theorem 9.16]{weidmann} implies that \[ \hat A_\lambda \to A \] in the sense of strong resolvent convergence as $\lambda \to 0$. In turn this implies that, for all bounded continuous functions $\Phi : [0,\infty) \to \mathbb{C}$, \begin{equation}\label{eq:nu_soc} \Phi(\hat A_\lambda) \to \Phi(A) \end{equation} in the sense of strong operator convergence as $\lambda \to 0$ \cite[Theorem 9.17]{weidmann}. Suppose now that $\Phi \in C_{\pg}([0,\infty))$. Then we can write $\Phi(u) = \tilde\Phi(u) \, (1+u^N)$ for some bounded continuous function $\tilde \Phi : [0,\infty) \to \mathbb{C}$ and $N \in \mathbb{N}$. For all $g,h \in C_c(\Lie{g})$, by \eqref{eq:nu_conj_fcinner}, \[ \langle \Phi(A_\lambda) g^{\lambda,2}, h^{\lambda,2} \rangle_{L^2(G)} = \langle \Phi(\hat A_\lambda) g, h \rangle_{L^2(\Lie{g})} = \langle \tilde\Phi(\hat A_\lambda) g, h \rangle_{L^2(\Lie{g})} + \langle \tilde\Phi(\hat A_\lambda) g, \hat A_\lambda^N h \rangle_{L^2(\Lie{g})} \] for all $\lambda$ sufficiently small, and the last quantity tends to \[ \langle \tilde\Phi(A) g, h \rangle_{L^2(\Lie{g})} + \langle \tilde\Phi(A) g, A^N h \rangle_{L^2(\Lie{g})} = \langle \Phi(A) g, h \rangle_{L^2(\Lie{g})} \] as $\lambda \to 0$, by \eqref{eq:nu_strongpower} and \eqref{eq:nu_soc}. \ref{en:nu_asspower}. This is just a restatement of part \ref{en:nu_assfunccalc} in the case where $\Phi(u) = u^{q/2}$. \end{proof} We can finally prove the enunciated relation between Hausdorff--Young constants of a Lie group and its Lie algebra. We find it convenient to state the result together with its analogue for Young constants, since both follow by the approximation results of Lemma \ref{lem:nu_localisation}. Part \ref{en:hyloc} of Proposition \ref{prop:nu_loc}, together with the following Remark \ref{rmk:symmetry} and the Babenko--Beckner theorem for $\mathbb{R}^n$, prove Theorem \ref{thm:local-HY}. As in \cite{Siebert-1986}, we define a \emph{contractive automorphism} of a locally compact group $G$ as an automorphism $\tau$ such that $\lim_{k\to\infty} \tau^k(x)=e$ for all $x \in G$. \begin{proposition}\label{prop:nu_loc} Let $G$ be a locally compact group. \begin{enumerate}[label=(\roman*)] \item\label{en:yloc} For all $p_1,\dots,p_k \in [1,\infty]$ such that $\sum_{j=1}^k 1/p_j' \in [0,1]$, \begin{equation}\label{eq:ytrivialineq} Y_{p_1,\dots,p_k}(G) \geq Y_{p_1,\dots,p_k}^\mathrm{loc}(G), \end{equation} with equality when $G$ has a contractive automorphism; moreover, if $G$ is a Lie group with Lie algebra $\Lie{g}$, \begin{equation}\label{eq:ylocineq} Y_{p_1,\dots,p_k}^\mathrm{loc}(G) \geq Y_{p_1,\dots,p_k}(\Lie{g}). \end{equation} \item\label{en:hyloc} For all $p \in [1,2]$, \begin{equation}\label{eq:hytrivialineq} H_p(G) \geq H_p^\mathrm{loc}(G), \end{equation} with equality if $G$ has a contractive automorphism. Moreover, when $G$ is an $n$-dimensional Lie group with Lie algebra $\Lie{g}$, \begin{equation}\label{eq:hylocineq} H_p^\mathrm{loc}(G) \geq H_p(\Lie{g}). \end{equation} \end{enumerate} \end{proposition} \begin{proof} \ref{en:yloc}. The first inequality is obvious. Moreover, in case $G$ has a contractive automorphism, the reverse inequality follows from a scaling argument. Indeed, for all automorphisms $\gamma$ of $G$, there exists $\kappa_\gamma \in (0,\infty)$ such that the push-forward via $\gamma$ of the Haar measure on $G$ is $\kappa_\gamma$ times the Haar measure. So, if $R_\gamma f = f \circ \gamma^{-1}$, then \[ \| R_\gamma f \|_{L^p(G)} = \kappa_\gamma^{-1/p} \| f\|_{L^p(G)}, \quad R_\gamma \Delta = \Delta, \quad R_\gamma \left( \bigast_{j=1}^k f_j \right) = \kappa_\gamma^{k-1} \bigast_{j=1}^k R_\gamma f_j, \] whence it is immediate that both sides of Young's inequality \eqref{eq:nu_young} are scaled by the same factor when each $f_j$ is replaced with $R_\gamma f_j$. Now, by density, the value of the best constant $Y_{p_1,\dots,p_k}(G)$ may be determined by testing \eqref{eq:nu_young} on arbitrary $f_1,\dots,f_k \in C_c(G)$. Moreover, if $\tau$ is a contractive automorphism of $G$ and $U$ is any neighbourhood of the identity, then, for all compact subsets $K \subseteq G$, there exists $N \in \mathbb{N}$ such that $\tau^N(K) \subseteq U$ \cite[Lemma 1.4(iv)]{Siebert-1986}; in particular, for all $f_1,\dots,f_k \in C_c(G)$, by taking $\gamma = \tau^N$ for sufficiently large $N \in \mathbb{N}$, we see that $\supp R_\gamma f_j \subseteq U$. This shows that $Y_{p_1,\dots,p_k}(G) \leq Y_{p_1,\dots,p_k}(G;U)$ for all neighbourhoods $U$ of $e \in G$, and consequently $Y_{p_1,\dots,p_k}(G) \leq Y_{p_1,\dots,p_k}^\mathrm{loc}(G)$. As for the second inequality, let $U$ be an arbitrary neighbourhood of $e \in G$. To conclude, it is sufficient to show that $Y_{p_1,\dots,p_k}(\Lie{g}) \leq Y_{p_1, \dots, p_k}(G;U)$. Let $r \in [1,\infty]$ be defined by $\sum_j 1/p_j' = 1/r'$. Consider $g,f_1,\dots,f_k \in C_c(\Lie{g})$. For all $\lambda \in (0,\infty)$, $\alpha \in \mathbb{C}$ and $p\in[1,\infty]$, define $g^{\lambda,p},f_j^{\lambda,p},f_j^{\lambda,p,\alpha}$ as in Lemma \ref{lem:nu_localisation}. Then $\bigcup_{j=1}^k \supp f_j^{\lambda,1} \subseteq U$ for all sufficiently small $\lambda$, and therefore, by \eqref{eq:nu_young}, \begin{multline*} \left\langle \bigast_{j=1}^k (f_j^{\lambda,1} \Delta^{\sum_{l=1}^{j-1} 1/p_l'}) , g^{\lambda,\infty} \right\rangle_{L^2(G)} \\ \leq Y_{p_1, \dots, p_k}(G;U) \, \| f_1^{\lambda,1} \|_{L^{p_1}(G)} \cdots \|f_k^{\lambda,1} \|_{L^{p_k}(G)} \|g^{\lambda,\infty} \|_{L^{r'}(G)}. \end{multline*} Note that $\sum_{j=1}^k 1/p_j + 1/r' = k$. So the last inequality can be rewritten as \begin{multline*} \left\langle f_1^{\lambda,1,\alpha_1} * \dots * f_k^{\lambda,1,\alpha_k} , g^{\lambda,\infty} \right\rangle_{L^2(G)} \\ \leq Y_{p_1, \dots, p_k}(G;U) \, \| f_1^{\lambda,p_1} \|_{L^{p_1}(G)} \cdots \|f_k^{\lambda,p_k} \|_{L^{p_k}(G)} \|g^{\lambda,r'} \|_{L^{r'}(G)}, \end{multline*} where $\alpha_j = -\sum_{l=1}^{j-1} 1/p_l'$. Hence, by Lemma \ref{lem:nu_localisation}, by taking the limit as $\lambda \to 0$, we obtain \[ \langle f_1 * \dots * f_k , g \rangle_{L^2(\Lie{g})} \leq Y_{p_1, \dots, p_k}(G;U) \, \| f_1\|_{L^{p_1}(\Lie{g})} \cdots \|f_k \|_{L^{p_k}(\Lie{g})} \|g \|_{L^{r'}(\Lie{g})}. \] The arbitrariness of $f_1,\dots,f_k,g \in C_c(\Lie{g})$ implies that $Y_{p_1,\dots,p_k}(\Lie{g}) \leq Y_{p_1, \dots, p_k}(G;U)$. \ref{en:hyloc}. Much as in part \ref{en:yloc}, the first inequality is obvious, and equality follows from a rescaling argument when $G$ has a contractive automorphism, since \[ \| R_\gamma f \|_{\Four L^q(G)} = \kappa_\gamma^{-1/q'} \| f \|_{\Four L^q(G)} \] for all automorphisms $\gamma$ of $G$. As for the second inequality, we need to show that $H_p(\Lie{g}) \leq H_p(G;U)$ for all neighbourhoods $U$ of $e \in G$. Set $q = p'$ and note that, by \eqref{eq:nu_fouriernorm}, \[ \|f\|_{\Four F^{q}(G)}^{q} = \sup_{\|g\|_{L^1(G)}, \|h\|_{L^1(G)} \leq 1} \langle |L_f \Delta^{1/q}|^{q} g, h \rangle_{L^2(G)}. \] For $\lambda \in (0,\infty)$, $r \in [1,\infty]$ and $f,g,h \in C_c(\Lie{g})$, we define $f^{\lambda,r},g^{\lambda,r},h^{\lambda,r} : G \to \mathbb{C}$ as in Lemma \ref{lem:nu_localisation}. For all sufficiently small $\lambda$, $\supp f^{\lambda,r} \subseteq U$ and therefore \[ \langle |L_{f^{\lambda,\infty}} \Delta^{1/q}|^{q} g^{\lambda,1}, h^{\lambda,1} \rangle_{L^2(G)} \leq H_p(G;U)^{q} \|f^{\lambda,\infty} \|_{L^p(G)}^{q} \|g^{\lambda,1}\|_{L^1(G)} \|h^{\lambda,1}\|_{L^1(G)}, \] that is, \begin{multline*} \lambda^{-n(q-1)} \langle |L_{f^{\lambda,\infty}} \Delta^{1/q}|^{q} g^{\lambda,1}, h^{\lambda,1} \rangle_{L^2(G)}\\ \leq H_p(G;U)^{q} \|f^{\lambda,p} \|_{L^p(G)}^{q} \|g^{\lambda,1}\|_{L^1(G)} \|h^{\lambda,1}\|_{L^1(G)}. \end{multline*} As $\lambda \to 0$, by Lemma \ref{lem:nu_localisation} we then deduce that \[ \langle |L_{f}|^{q} g, h \rangle_{L^2(\Lie{g})} \leq H_p(G;U)^{q} \|f \|_{L^p(\Lie{g})}^{q} \|g\|_{L^1(\Lie{g})} \|h\|_{L^1(\Lie{g})}. \] By the arbitrariness of $g,h \in C_c(\Lie{g})$, \[ \|f\|_{\Four L^q(\Lie{g})} \leq H_p(G;U) \|f\|_{L^p(\Lie{g})} \] and finally, by the arbitrariness of $f \in C_c(\Lie{g})$, $H_p(\Lie{g}) \leq H_p(G;U)$. \end{proof} \begin{remark}\label{rmk:symmetry} The argument in Proposition \ref{prop:nu_loc} can be extended to the case of inequalities restricted to particular classes of functions on $G$. In particular, suppose that the class of functions is determined by invariance with respect to the action of a compact group $K$ of automorphisms of $G$. Then it is possible to choose a positive inner product on $\Lie{g}$ so that $K$ acts on $\Lie{g}$ by isometries (take any inner product on $\Lie{g}$ and average it with respect to the action of $K$), and the correspondence \eqref{eq:nu_associatedp} preserves $K$-invariance whenever $\Omega$ is a ball centred at the origin. Moreover the class of functions on $\Lie{g}$ under consideration contains all radial functions. Since the extremisers for Young and Hausdorff--Young constants on $\Lie{g}$ are centred gaussians \cite{Bec-1975,brascamp_best_1976}, which may be assumed to be radial, the resulting lower bounds do not change. This observation completes the proof of Theorem \ref{thm:local-HY}. \end{remark} \begin{remark}\label{rmk:Y_HY} While the inequalities \eqref{eq:ytrivialineq} and \eqref{eq:hytrivialineq} may be strict for certain Lie groups $G$ (note that, when $G$ is compact, the global Young and Hausdorff--Young constants are equal to $1$), it appears natural to ask whether the inequalities \eqref{eq:ylocineq} and \eqref{eq:hylocineq} are actually equalities. We are not aware of any counterexample. As a matter of fact, a particular case of a recent result of Bennett, Bez, Buschenhenke, Cowling and Flock about nonlinear Brascamp--Lieb inequalities \cite{BBBCF_2018} entails that equality \emph{always} holds in \eqref{eq:ylocineq} for all Lie groups $G$: \[ Y_{p_1,\dots,p_k}^\mathrm{loc}(G) = Y_{p_1,\dots,p_k}(\Lie{g}) \] for \emph{all} $p_1,\dots,p_k \in [1,\infty]$ such that $\sum_{j=1}^k 1/p_j' \in [0,1]$. By Proposition \ref{prop:basic}\ref{en:yheq}, this in turn implies that \[ H_p^\mathrm{loc}(G) = H_p(\Lie{g}) = (B_p)^{\dim G} \] for all $p \in [1,2]$ such that $p'$ is an even integer, and \emph{a fortiori} the same equality holds for the $K$-invariant version of the constants for any compact group of automorphisms $K$. \end{remark} As a consequence of the above results, we strengthen some results of Klein and Russo \cite[Corollaries 2.5' and 2.8]{KR-1978}, where upper bounds for Young and Hausdorff--Young constants are obtained for particular solvable Lie groups. Klein and Russo explicitly remark that they are able to obtain equalities instead of upper bounds in the particular case of the Heisenberg groups and only for special exponents (through a different argument, involving the analysis of the Weyl transform) and seem to leave the general case open. Here instead we obtain equality for all the Young constants, as well as a lower bound for the Hausdorff--Young constants (which becomes an equality in the case of Babenko's exponents). \begin{corollary} Let $G$ be a $n$-dimensional solvable Lie group admitting a chain of closed subgroups \[ \{e\} = G_0 < \dots < G_n = G, \] where $G_j$ is normal in $G_{j+1}$ and $G_{j+1}/G_j$ is isomorphic to $\mathbb{R}$. Denote by $B_p$ the Babenko--Beckner constant. Then the following hold. \begin{enumerate}[label=(\roman*)] \item\label{en:nu_ysolvable} for all $p_1,\dots,p_k,r \in [1,\infty]$ such that $\sum_{j=1}^k 1/p_j' = 1/r'$, \[ Y_{p_1,\dots,p_k}(G) = Y^\mathrm{loc}_{p_1,\dots,p_k}(G) = (B_{r'} B_{p_1} \cdots B_{p_k})^n; \] \item\label{en:nu_hysolvable} for all $p \in [1,2]$, \[ H_p(G) \geq H_p^\mathrm{loc}(G) \geq (B_p)^n, \] with equalities if $p' \in 2\mathbb{Z}$. \end{enumerate} \end{corollary} \begin{proof} \ref{en:nu_ysolvable}. The inequality $Y_{p_1,\dots,p_k}(G) \leq (B_{r'} B_{p_1} \cdots B_{p_k})^n$ can be obtained, as in \cite{KR-1978}, by iteratively applying Proposition \ref{prop:basic}\ref{en:ext} and the fact that $Y_{p_1,\dots,p_k}(\mathbb{R}) = B_{r'} B_{p_1} \cdots B_{p_k}$ \cite{Bec-1975,brascamp_best_1976}. On the other hand, by Propositions \ref{prop:nu_loc}\ref{en:yloc} and \ref{prop:basic}\ref{en:ext}, \[ Y_{p_1,\dots,p_k}(G) \geq Y^\mathrm{loc}_{p_1,\dots,p_k}(G) \geq Y_{p_1,\dots,p_k}(\Lie{g}) = Y_{p_1,\dots,p_k}(\mathbb{R})^n = (B_{r'} B_{p_1} \cdots B_{p_k})^n, \] and we are done. \ref{en:nu_hysolvable}. From part \ref{en:nu_ysolvable} and Proposition \ref{prop:basic}\ref{en:yheq}, we deduce immediately that $H_p(G) = (B_p)^n$ whenever ${{q}}$ is an even integer. On the other hand, by Proposition \ref{prop:nu_loc}\ref{en:hyloc}, \[ H_{p}(G) \geq H^\mathrm{loc}_{p}(G) \geq H_{p}(\Lie{g}) = (B_p)^n, \] by \cite{Bec-1975}, and we are done. \end{proof} \section{The \texorpdfstring{$n$}{n}-torus \texorpdfstring{$\T^n$}{Tn} revisited}\label{s:torus} The proof of the central local Hausdorff--Young theorem on a compact Lie group mimics that of the local Hausdorff--Young theorem on $\T^n$, and we present this case first to make the proof of the general case more evident. \begin{proof}[Proof of Theorem \ref{thm:local-HY-Tn}] There is no loss of generality in supposing functions smooth; this ensures that all the sums and integrals that occur in the proof below converge. Let us identify $\T^n$ with the subset $(-1/2, 1/2]^n$ of $\mathbb{R}^n$. For $f \in L^1(\T^n)$, the Fourier transform $\hat f : \mathbb{Z}^n \to \mathbb{C}$ of $f$ is given by \[ \hat f(\mu) = \int_{\T^n} f(x) \, e^{2\pi i \mu \cdot x} \, dx. \] for all $\mu \in \mathbb{Z}^n$. We denote by $V$ the open subset $(-1/2,1/2)^n$ of $\mathbb{R}^n$. For any function $f \in L^1(\T^n)$ such that $\supp f \subseteq V$, we define $F$ on $\mathbb{R}^n$ by \[ F(x) = \begin{cases} f(x) & \text{when $x \in V$,} \\ 0 & \text{otherwise;} \end{cases} \] we say that $F$ corresponds to $f$. Clearly $F \in L^1(\mathbb{R}^n)$ and $\hat F|_{\mathbb{Z}^n} = \hat f$; further, if $f$ is smooth, so is $F$. We are going to transfer the sharp Hausdorff--Young theorem for $F$ to $f$. The Plancherel formulae for Fourier series and Fourier integrals imply that \[ \| \hat f \|_{\ell^2(\mathbb{Z}^n)} = \| f \|_{L^2(\T^n)} = \| F \|_{L^2(\mathbb{R}^n)} = \| \hat F \|_{L^2(\mathbb{R}^n)} . \] In particular, since $\hat F|_{\mathbb{Z}^n} = \hat f$, \begin{equation}\label{eq:restriction} \| \hat F|_{\mathbb{Z}^n} \|_{\ell^2(\mathbb{Z}^n)} \leq \| \hat F \|_{L^2(\mathbb{R}^n)} . \end{equation} Further, trivially, \[ \|\hat F|_{\mathbb{Z}^n} \|_{\ell^\infty(\mathbb{Z}^n)} \leq \| \hat F \|_{L^\infty(\mathbb{R}^n)}. \] If we could interpolate between these inequalities, then it would follow that \begin{equation}\label{eq:key} \| \hat F|_{\mathbb{Z}^n} \|_{\ell^q(\mathbb{Z}^n)} \leq \| \hat F \|_{L^q(\mathbb{R}^n)} \end{equation} for all $q \in [2,\infty]$ and $\hat F$ in $L^{{{q}}}(\mathbb{R}^n)$, whence \[ \| \hat f \|_{\ell^{p'}(\mathbb{Z}^n)} = \| \hat F|_{\mathbb{Z}^n} \|_{\ell^{p'}(\mathbb{Z}^n)} \leq \| \hat F \|_{L^{p'}(\mathbb{R}^n)} \leq (B_p)^n \| F \|_{L^p(\mathbb{R}^n)} = (B_p)^n \| f \|_{L^p(\T^n)} , \] and we would be done. But we can \emph{not} interpolate, because \eqref{eq:restriction} does not hold for all $\hat F$ in $L^{2}(\mathbb{R}^n)$, or even for all $\hat F$ in a dense subspace of $L^{2}(\mathbb{R}^n)$, but only for those $\hat F$ where $\supp F \subseteq V$; \emph{inter alia}, this ensures that $\hat F$ is smooth so that $\hat F|_{\mathbb{Z}^n}$ is well-defined. So we prove a variant of \eqref{eq:key}. Let $U$ be a small neighbourhood $U$ of $0$ in $\T^n$ such that $\overline U \subseteq V$, and take $\phi \in A(\mathbb{R}^n)$ such that $\supp \phi \subseteq V $ and $\phi(x) = 1$ for all $x \in U$. We now define \[ T G = (\hat\phi * G)|_{\mathbb{Z}^n} \qquad\forall G \in L^1(\mathbb{R}^n) + L^\infty(\mathbb{R}^n). \] We claim that when $q \in [2, \infty]$, \begin{equation}\label{eq:key2} \| TG \|_{\ell^q(\mathbb{Z}^n)} \leq \| \hat\phi \|_{L^1(\mathbb{R}^n)} \| G \|_{L^q(\mathbb{R}^n)} \qquad\forall G \in L^{q}(\mathbb{R}^n). \end{equation} To prove the claim, observe that the inverse Fourier transform of $\hat\phi * G$ is supported in $V$, whence \begin{equation*} \| TG \|_{\ell^2(\mathbb{Z}^n)} = \| (\hat\phi * G)|_{\mathbb{Z}^n} \|_{\ell^2(\mathbb{Z}^n)} \leq \| \hat\phi * G \|_{L^2(\mathbb{R}^n)} \leq \| \hat\phi \|_{L^1(\mathbb{R}^n)} \| G \|_{L^2(\mathbb{R}^n)}, \end{equation*} for all $G \in L^{2}(\mathbb{R}^n)$, by \eqref{eq:restriction} and a standard convolution inequality. Similarly, since $\hat\phi * G$ is continuous, the same inequalities hold when $2$ is replaced by $\infty$. Thus \eqref{eq:key2} holds when $q$ is $2$ or $\infty$. The Riesz--Thorin interpolation theorem establishes \eqref{eq:key2} for all $q \in [2, \infty]$. To conclude the proof, take $f \in C^\infty(\T^n)$ such that $\supp f \subseteq U$, and let $F$ correspond to $f$. Then $\hat F \in L^1(\mathbb{R}^n) \cap L^{\infty}(\mathbb{R}^n)$ and $\hat\phi * \hat F = \hat F$. Thus \[ \| \hat f \|_{\ell^q(\mathbb{Z}^n)} = \| T\hat F \|_{\ell^q(\mathbb{Z}^n)} \leq \| \hat\phi \|_{L^1(\mathbb{R}^n)} \| \hat F \|_{L^q(\mathbb{R}^n)} \] by \eqref{eq:key2}. This now gives \[\begin{split} \| \hat f \|_{\ell^{p'}(\mathbb{Z}^n)} &\leq \| \hat\phi \|_{L^1(\mathbb{R}^n)} \| \hat F \|_{L^{p'}(\mathbb{R}^n)}\\ &\leq \| \hat\phi \|_{L^1(\mathbb{R}^n)} (B_p)^n \| F \|_{L^p(\mathbb{R}^n)} = \| \hat\phi \|_{L^1(\mathbb{R}^n)} (B_p)^n \| f \|_{L^p(\mathbb{Z}^n)}. \end{split}\] This proves that $H_p(\T^n;U) \leq \|\hat\phi\|_{L^1(\mathbb{R}^n)} (B_p)^n$. By choosing $U$ small enough, we may make $\|\hat\phi \|_{L^1(\mathbb{R}^n)}$ as close to $1$ as we like (see \cite{leptin}): indeed, we can take $\phi = |K|^{-1} \chrfn_{U+K} * \chrfn_K$, where $K=-K$ is a fixed small neighbourhood of the origin (here $\chrfn_\Omega$ denotes the characteristic function of a measurable set $\Omega \subseteq \mathbb{R}^n$ and $|\Omega|$ its Lebesgue measure), so that $\supp \phi \subseteq U +2K$ and \[ 1 = \phi(0) \leq \|\hat\phi\|_{L^1(\mathbb{R}^n)} \leq |K|^{-1} \|\chrfn_K\|_{L^2(\mathbb{R}^n)} \|\chrfn_{U+K}\|_{L^2(\mathbb{R}^n)} = (|U+K|/|K|)^{1/2}. \] So $H_p^\mathrm{loc}(\T^n) \leq (B_p)^n$, and the converse inequality is given by Theorem \ref{thm:local-HY}. \end{proof} \section{Compact Lie groups}\label{s:compact} Before entering into the proof of Theorem \ref{thm:local-central-HY-compact-Lie}, we present a summary of the theory of representations and characters of compact connected Lie groups $G$. For more details, the reader may consult, for example, \cite{BtD_1985,Knapp-2002}. We assume throughout that $G$ is not abelian, since the abelian case was treated in Theorem \ref{thm:local-HY-Tn}. A compact connected Lie group $G$ comes with a set $\Lambda^+$ of \emph{dominant weights}, which parametrise the collection of irreducible unitary representations $\pi_\lambda$ of $G$ modulo equivalence. Each such representation $\pi_\lambda$ is of finite dimension $d_\lambda$ and has a character $\chi_\lambda$ given by $\trace \pi_\lambda(\cdot)$. Assume that the Haar measure on $G$ is normalised so as to have total mass $1$. The Peter--Weyl theory gives us the Plancherel formula: if $f \in L^2(G)$, then \[ \| f\|_2^2 = \sum_{\lambda \in \Lambda^+} d_\lambda \| \pi_\lambda(f) \|_\HS^2 . \] In other words, the group Plancherel measure on the unitary dual of $G$ can be identified with the discrete measure on $\Lambda^+$ that assigns mass $d_\lambda$ to the point $\lambda$. From the discussion in Section \ref{s:FLq}, we deduce that \[ \| f \|_{\Four L^q} = \left( \sum_{\lambda \in \Lambda^+} d_\lambda \| \pi_\lambda(f) \|_{\Sch^q}^q \right)^{1/q}. \] for all $q \in [1,\infty)$. If $f$ is a central function, then $\pi_\lambda(f)$ is a multiple of the identity and \[ \tilde f(\lambda) := \int_G f(x) \, \chi_\lambda(x) \,dx = \trace \pi_\lambda(f), \] whence \[ \| f \|_{\Four L^q} = \left( \sum_{\lambda \in \Lambda^+} d_\lambda^{2-q} |\tilde f(\lambda)|^q \right)^{1/q}. \] For $q=2$, this corresponds to the fact that the characters $\chi_\lambda$ form an orthonormal basis for the space of square-integrable central functions. A more precise description of the set $\Lambda^+$ of dominant weights and the characters $\chi_\lambda$ can be given as follows. Recall that the conjugation action of the group $G$ on itself determines the adjoint representation of $G$ on $\Lie{g}$: \[ \exp( \Ad(x) Y) = x \exp(Y) x^{-1} \qquad\forall x \in G \quad\forall Y \in \Lie{g}. \] Since $G$ is compact, there exists an $\Ad(G)$-invariant inner product on $\Lie{g}$, which in turn determines a Lebesgue measure on $\Lie{g}$; we scale the inner product so that the Jacobian determinant $J : \Lie{g} \to \mathbb{R}$ of the exponential mapping is $1$ at the origin. Clearly $J$ is an $\Ad(G)$-invariant function. The group $G$ contains a maximal torus $T$, that is, a maximal closed connected abelian subgroup, which is unique up to conjugacy; its Lie algebra $\Lie{t}$ is a maximal abelian Lie subalgebra of $\Lie{g}$. The set $\Gamma$ of $X$ in $\Lie{t}$ such that $\exp X = e$ is a lattice in $\Lie{t}$, and $T$ may be identified with $\Lie{t} / \Gamma$. The \emph{weight lattice} $\Lambda$ is the dual lattice to $\Gamma$, that is, the set of elements $\lambda$ of the dual space $\Lie{t}^*$ taking integer values on $\Gamma$: equivalently, $\Lambda$ is the set of the $\lambda \in \Lie{t}^*$ such that $X \mapsto e^{2\pi i \lambda(X)}$ descends to a character $\kappa_\lambda$ of $T$. We say that a weight $\lambda \in \Lambda$ occurs in a unitary representation $\pi$ of $G$ if the character $\kappa_\lambda$ of $T$ is contained in the restriction of $\pi$ to $T$. Weights occurring in the (complexified) adjoint representation are called \emph{roots}. A choice of an ordering splits roots into into \emph{positive} and \emph{negative} roots. We denote by $\rho$ half the sum of the positive roots. The set $\Lambda^+$ of \emph{dominant weights} is the set of the $\lambda \in \Lambda$ having nonnegative inner product with all positive roots. The irreducible representation $\pi_\lambda$ of $G$ corresponding to $\lambda \in \Lambda^+$ is determined, up to equivalence, by the fact that $\lambda$ is the highest weight occurring in $\pi_\lambda$ (that is, $\lambda$ occurs in $\pi_\lambda$, while $\lambda + \alpha$ does not occur in $\pi_\lambda$ for any positive root $\alpha$). Via the orthogonal projection of $\Lie{g}$ onto $\Lie{t}$, we can identify $\Lie{t}^*$ with a subspace of $\Lie{g}^*$. Given $\lambda$ in $\Lie{g}^*$, we write $O_{\lambda}$ for the compact set $\Ad(G)^* \lambda$, usually called the \emph{orbit} of $\lambda$. Kirillov's character formula \cite[p.\ 459]{Kir} states that, for all $X \in \Lie{g}$ and all $\lambda \in \Lambda^+$, \begin{equation}\label{eq:kirillov} J(X)^{1/2} \, \chi_\lambda( \exp( X) ) = \int_{O_{\lambda+\rho}} \exp( 2 \pi i \xi \cdot X) \,d\sigma(\xi), \end{equation} where $\sigma$ is a canonical $\Ad(G)^*$-invariant measure on $O_{\lambda+\rho}$, and $\xi \cdot X$ denotes the duality pairing between $\xi \in \Lie{g}^*$ and $X \in \Lie{g}$. When $X = 0$, this formula becomes the normalisation \[ \int_{O_{\lambda+\rho}} \,d\sigma(\xi) = d_\lambda . \] \begin{proof}[Proof of Theorem \ref{thm:local-central-HY-compact-Lie}] Take a small connected conjugation-invariant neighbourhood $U$ of the identity in $G$ that is also symmetric, that is, $U^{-1} = U$. Then $U = \bigcup_{x \in G} x (U \cap T) x^{-1}$. Let $V$ be the small connected neighbourhood of $0$ in $\Lie{g}$ such that $U = \exp V$ and $\exp$ is a diffeomorphism from a neighbourhood of $\overline V$ onto a neighbourhood of $\overline U$ in $G$. To a function $f$ on $G$ supported in $U$, we associate the function $F$ on $\Lie{g}$ supported in $V$ by the formula \[ F(X) = \begin{cases} J(X)^{1/2} \, f(\exp(X)) &\text{when $X \in V$,} \\ 0 &\text{otherwise}. \end{cases} \] Then $\| J^{1/p -1/2} F \|_p = \| f \|_p$. We define the Fourier transform $\hat F$ of $F$ as follows: \[ \hat F (\xi ) = \int_{\Lie{g}} F(X) \, \exp( 2\pi i \xi \cdot X) \, dX \qquad \forall \xi \in \Lie{g}^*. \] The following conditions are equivalent: $f$ is central on $G$; $F$ is $\Ad(G)$-invariant on $\Lie{g}$; and $\hat F$ is $\Ad(G)^*$-invariant on $\Lie{g}^*$. Assume that $f$ is central and supported in $U$, and let $F$ be the associated function on $\Lie{g}$. From the character formula \eqref{eq:kirillov}, a change of variables, and a change of order of integration, \begin{align*} \tilde f(\lambda) &= \int_G f(x) \, \chi_\lambda(x) \,dx = \int_{\Lie{g}} F(X) \int_{O_{\lambda+\rho}} \exp( 2 \pi i \xi \cdot X) \,d\sigma(\xi)\,dX \\ &= \int_{O_{\lambda+\rho}} \int_{\Lie{g}} F(X) \exp(2 \pi i \xi \cdot X) \,dX\, d\sigma(\xi) = \int_{O_{\lambda+\rho}} \hat F(\xi) \, d\sigma(\xi) = d_\lambda \hat F(\lambda+\rho). \end{align*} This, combined with the Plancherel theorems for central functions on $G$ and for functions on $\Lie{g}$, implies that \[ \sum_{\lambda \in \Lambda^+} d_\lambda^2 |\hat F(\lambda+\rho)|^2 = \| f\|_2^2 = \| F \|_2^2 = \| \hat F \|_2^2. \] For such functions, moreover, $\hat F$ is continuous and so \[ \sup_{\lambda \in \Lambda^+} | \hat F (\lambda + \rho) |_\infty \leq \| \hat F \|_\infty. \] For a function $H$ on $\Lie{g}^*$, we define \[ H^G(\lambda) = \int_G H(\Ad(g)^*\lambda) \, dg. \] Much as in the case of $\T^n$, we choose an $\Ad(G)$-invariant function $\phi \in A(\Lie{g})$ which vanishes off $V$ and takes the value $1$ on the open $\Ad(G)$-invariant subset $W$ of $V$. For $H$ in $L^1(\Lie{g}^*) + L^\infty(\Lie{g}^*)$, we define the function $TH$ by \[ TH(\lambda) = \hat\phi*H^G(\lambda + \rho) \qquad\forall\lambda \in \Lambda^+. \] For such functions $H$, the inverse Fourier transform $F$ of $\hat\phi*H^G$ is supported in $V$ and is $\Ad(G)$-invariant, so the corresponding function $f$ on $G$ is central and supported in $U$. From our previous discussion, \[ \left( \sum_{\lambda \in \Lambda^+} d_\lambda^2 | TH (\lambda)|^2 \right)^{1/2} = \| \hat\phi* H^G \|_{2} \leq \|\hat\phi\|_1 \| H^G \|_2 \leq \|\hat\phi\|_1 \| H \|_2 \] and \[ \sup_{\lambda \in \Lambda^+} | TH (\lambda)| \leq \| TH \|_\infty \leq \|\hat\phi\|_1 \| H^G \|_\infty \leq \|\hat\phi\|_1 \| H \|_\infty . \] By Riesz--Thorin interpolation, when $2 \leq q < \infty$, \[ \left( \sum_{\lambda \in \Lambda^+} d_\lambda^2 | TH (\lambda)|^q \right)^{1/q} \leq \|\hat\phi\|_1 \| H \|_q . \] Much as in the proof of Theorem \ref{thm:local-HY-Tn}, if $f$ is a central function on $G$ supported in $\exp(W) \subseteq U$, and $F$ is the $\Ad(G)$-invariant function on $\Lie{g}$ corresponding to $f$, then $T\hat F(\lambda) = \hat\phi * \hat F(\lambda+\rho) = \hat F(\lambda+\rho)$ for all $\lambda \in \Lambda^+$. Hence, if $n=\dim G$, from the Hausdorff--Young inequality on $\mathbb{R}^n$ we deduce that \begin{multline*} \|f\|_{\Four L^{p'}}= \left( \sum_{\lambda \in \Lambda^+} d_\lambda^{2-p'} | \tilde f(\lambda) |^{p'} \right)^{1/{p'}} = \left( \sum_{\lambda \in \Lambda^+} d_\lambda^2 | \hat F (\lambda+\rho) |^{p'} \right)^{1/{p'}} \\ \leq \| \hat\phi \|_1 \| \hat F \|_{p'} \leq \| \hat\phi \|_1 (B_p)^n \| F \|_p \leq \| \hat\phi \|_1 (B_p)^n \sup_{X \in W} J(X)^{1/2-1/p} \| f \|_p, \end{multline*} which shows that $H_{p,\Inn(G)}(G;\exp(W)) \leq \| \hat\phi \|_1 (B_p)^n \sup_{X \in W} J(X)^{1/2-1/p}$. By taking $W$ small, we may make both $\sup_{X \in W} J(X)^{1/2-1/p}$ and $\| \hat\phi \|_1$ close to $1$. So $H_{p,\Inn(G)}^\mathrm{loc}(G) \leq (B_p)^n$, and the converse inequality is given by Theorem \ref{thm:local-HY}. \end{proof} \section{The Weyl transform}\label{s:weyl} In this section, we shall mostly adopt the notation from Folland's book \cite{Folland-1989}. The Weyl transform $\rho(f)$ of a function $f \in L^1(\mathbb{C}^n)$ can be written as the operator \[ \rho(f) =\int_{\mathbb{R}^n} \int_{\mathbb{R}^n} f(u+iv) \, e^{2\pi i(uD+vX)}\, du \,dv \] on $L^2(\mathbb{R}^n)$, where $uD=\sum_{j=1}^n u_j D_j$ and $vX=\sum_{j=1}^n v_j X_j$, and where $D_j$ and $X_j$ denote the operators \[ D_j\phi(x)=\frac 1{2\pi i}\frac \partial{\partial x_j} f(x) \qquad\text{and}\qquad X_j\phi(x)=x_j\phi(x). \] Explicitly, $\rho(f)$ is the integral operator given by \[ \rho(f) \phi(x)=\int_{\mathbb{R}^n} K_f(x,y) \, \phi(y) \, dy, \] with integral kernel given by \[ K_f(x,y)=\int_{\mathbb{R}^n} f(y-x+iv) \, e^{\pi i v(x+y)} \, dv. \] As Folland points out on page 24 of his monograph, this notion of ``Weyl transform'' is historically incorrect---the Weyl transform of $f$ should rather be $\rho(\hat f)$, the pseudodifferential operator associated to the symbol $f$ in the Weyl calculus \cite[Chapter 2]{Folland-1989}. Nevertheless, we shall use the definition of Weyl transform above. In \cite{KR-1978}, the authors consider the operator $\nu(f)$ given by \[ \nu(f) =\int_{\mathbb{R}^n}\int_{\mathbb{R}^n} f(u +iv) \, e^{2\pi iuD} \, e^{2\pi ivX}\, du \, dv , \] and call this the Weyl operator associated to $f$---this appears to be even more inappropriate, as $\nu(f)$ is actually more closely related to the Kohn--Nirenberg calculus (see, for example, \cite[(2.32)]{Folland-1989}). In any case, it is easily seen that the operators $\nu(f)$ and $\rho(f)$ are related by the identity \begin{equation}\label{eq:nurho} \nu(f)=\rho(e^{i\pi u \cdot v} f) \end{equation} (compare also \cite[Proposition 2.33]{Folland-1989}). We are interested in best constants in Hausdorff--Young inequalities of the form \begin{equation}\label{eq:hyweyl} \|\rho(f)\|_{\Sch^{p'}}\leq C \|f\|_{L^p(\mathbb{C}^n)}, \end{equation} for suitable functions $f$, for instance Schwartz functions. In light of \eqref{eq:nurho}, we may work with $\nu(f)$ in place of $\rho(f)$ equally well. As discussed in the introduction, we denote by $W_p(\mathbb{C}^n)$ the best constant $C$ in \eqref{eq:hyweyl}, and use the symbols $W_p^\mathrm{loc}(\mathbb{C}^n)$, $W_{p,K}(\mathbb{C}^n)$ and $W_{p,K}^\mathrm{loc}(\mathbb{C}^n)$ for the corresponding local and $K$-invariant variants. If $p=2$, then $\rho$ is indeed isometric from $L^2(\mathbb{C}^{n})$ onto the space of Hilbert--Schmidt operators \cite[Theorem (1.30)]{Folland-1989}, and thus the following ``Plancherel identity'' for the Weyl transform holds true: \begin{equation}\label{eq:weyl_plancherel} \|\rho(f)\|_{\HS}=\|f\|_2. \end{equation} This tells us that $W_2(\mathbb{C}^n) = 1$ and, by interpolation, $W_p(\mathbb{C}^n) \leq 1$ for all $p \in [1,2]$. However, as Klein and Russo have shown, $W_p(\mathbb{C}^n) < 1$ when $1<p<2$. Indeed, \cite[Theorem 1]{KR-1978} may be restated by saying that \begin{equation}\label{eq:weylconstant} W_p(\mathbb{C}^n) = (B_p)^{2n} \end{equation} when $p'\in 2\mathbb{Z}$. Moreover, in contrast with the Euclidean case, there are no extremal functions for the optimal estimate---the best constant can only be found as a limit, for instance along a suitable family of Gaussian functions $f$. This raises the question whether \eqref{eq:weylconstant} holds for more general $p\in[1,2]$. Besides being of interest in its own right, the determination of the best constants in the Hausdorff--Young inequality \eqref{eq:hyweyl} for the Weyl transform on $\mathbb{C}^n$ is relevant to the analysis of the analogous inequality on the Heisenberg group $\Heis_n$. Indeed, the proof of Klein and Russo \cite[Theorem 3]{KR-1978} that \begin{equation}\label{eq:heisconstant} H_p(\Heis_n) = (B_p)^{2n+1} \end{equation} when $p' \in 2\mathbb{Z}$ is based on a reduction, via a scaling argument, to the corresponding result \eqref{eq:weylconstant} for the Weyl transform. A somewhat refined version of the scaling argument, presented below, shows that the problem of determining the best Hausdorff--Young constants for the Heisenberg group is completely equivalent to the analogous problem for the Weyl transform, irrespective of the exponent $p \in [1,2]$, and also in case of restriction to functions with symmetries. \begin{proposition}\label{prp:weylheisenberg} For all compact subgroups $K$ of $\group{U}(n)$ and all $p \in [1,2]$, \[ H_{p,K}(\Heis_n) = B_p W_{p,K}(\mathbb{C}^n). \] \end{proposition} \begin{proof} Let us identify $\Heis_n$ with $\mathbb{C}^n \times \mathbb{R}$ with group law \[ (z,t) \cdot (z',t') = (z+z',t+t'+\Im(\bar z \cdot z')/2). \] The Lebesgue measure on $\mathbb{C}^n \times \mathbb{R}$ is a Haar measure on $\Heis_n$, which we fix throughout. The Schr\"odinger representation $\pi$ of $\Heis_n$ on $L^2(\mathbb{R}^n)$ is given by \[ \pi(u+iv,t) \phi(x) = e^{2\pi i t + 2\pi i v \cdot x + \pi i u \cdot v} \phi(u+x) \] \cite[(1.25)]{Folland-1989}. For all $\lambda \in \mathbb{R} \setminus \{0\}$, the map $A_\lambda : \Heis_n \to \Heis_n$, given by \[ A_\lambda(z,t) = \begin{cases} (\sqrt{|\lambda|} \, z, \lambda t) &\text{if $\lambda > 0$,}\\ (\sqrt{|\lambda|} \, \bar z, \lambda t) &\text{if $\lambda < 0$}, \end{cases} \] is an automorphism of $\Heis_n$. The representations $\pi_\lambda = \pi \circ A_\lambda$ form a family of pairwise inequivalent irreducible unitary representations of $\Heis_n$, in terms of which we can express the Plancherel formula for $\Heis_n$: \[ \| F \|^2_{L^2(\Heis_n)} = \int_{\mathbb{R} \setminus \{0\}} \| \pi_\lambda(F) \|_{\HS}^2 \, |\lambda|^n \,d\lambda. \] \cite[p.\ 39]{Folland-1989}. Hence, by the discussion in Section \ref{s:FLq}, for all $q \in [1,\infty)$, \begin{equation}\label{eq:FLq_Heis} \| F \|^q_{\Four L^q} = \int_{\mathbb{R} \setminus \{0\}} \| \pi_\lambda(F) \|_{\Sch^q}^q \, |\lambda|^n \,d\lambda. \end{equation} For all $F \in L^1(\Heis_n)$ and $\lambda \in \mathbb{R}$, let us set \[ F^\lambda(z) = \int_\mathbb{R} F(z,t) \, e^{2\pi i t \lambda} \,dt. \] Then \begin{equation}\label{eq:Heis_rep_weyl} \pi_\lambda(F) = \rho(Z_\lambda F^\lambda), \end{equation} where, for all functions $f$ on $\mathbb{C}^n$, \[ Z_\lambda f(z) = \begin{cases} |\lambda|^{-n} f(|\lambda|^{-1/2} z) & \text{if $\lambda>0$,}\\ |\lambda|^{-n} f(|\lambda|^{-1/2} \bar z) & \text{if $\lambda<0$.} \end{cases} \] From the definition of $\rho$, it is not difficult to show that \[ \rho(Z_{-1} f) = S \rho(f)^* S, \] where $S f(z) = f^*(z) = \overline{f(-z)}$. From this it readily follows that \begin{equation}\label{eq:weyl_conj} \|\rho(Z_{-\lambda} f)\|_{\Sch^q} = \|\rho(Z_\lambda f)\|_{\Sch^q} \end{equation} for all $\lambda \in \mathbb{R} \setminus \{0\}$ and $q \in [1,\infty]$. Let $F \in C^\infty_c(\Heis_n)$ be $K$-invariant. Then $Z_\lambda F^\lambda$ is also $K$-invariant for all $\lambda >0$. Hence, by \eqref{eq:FLq_Heis}, \eqref{eq:Heis_rep_weyl} and \eqref{eq:weyl_conj}, \[\begin{split} \|F\|_{\Four L^{p'}} &= \left( \int_{\mathbb{R} \setminus \{0\}} \| \rho(Z_{|\lambda|} F^\lambda) \|_{\Sch^{p'}}^{p'} \, |\lambda|^n \,d\lambda \right)^{1/p'} \\ &\leq W_{p,K}(\mathbb{C}^n) \left( \int_{\mathbb{R} \setminus \{0\}} \| Z_{|\lambda|} F^\lambda \|_{p}^{p'} \, |\lambda|^n \,d\lambda \right)^{1/p'} \\ &= W_{p,K}(\mathbb{C}^n) \left( \int_{\mathbb{R} \setminus \{0\}} \| F^\lambda \|_{p}^{p'} \,d\lambda \right)^{1/p'} \\ &\leq W_{p,K}(\mathbb{C}^n) \left( \int_{\mathbb{C}^n} \left( \int_\mathbb{R} | F^\lambda(z) |^{p'} \,d\lambda \right)^{p/p'} \,dz \right)^{1/p} \\ &\leq W_{p,K}(\mathbb{C}^n) B_p \|F\|_p, \end{split}\] where we applied, in order, the sharp Hausdorff--Young inequality for the Weyl transform and $K$-invariant functions, a scaling, the Minkowski integral inequality (note that $p'/p \geq 1$) and the sharp Hausdorff--Young inequality on $\mathbb{R}$. This shows that $H_{p,K}(\Heis_n) \leq B_p W_{p,K}(\mathbb{C}^n)$. Conversely, let $f \in C^\infty_c(\mathbb{C}^n)$ be $K$-invariant and $\phi : \mathbb{R} \to \mathbb{C}$ be in the Schwartz class, and let $F = f \otimes \phi$. Then $F$ is also $K$-invariant, and moreover $F^\lambda = \hat\phi(\lambda) f$. So, by applying the sharp Hausdorff--Young inequality on $\Heis_n$ to $F$ we obtain that \begin{equation}\label{eq:test_sharp_HY_heis_tensor} \left( \int_{\mathbb{R} \setminus \{0\}} \| \rho(Z_\lambda f) \|_{\Sch^{p'}}^{p'} \, |\hat\phi(\lambda)|^{p'} |\lambda|^n \,d\lambda \right)^{1/p'} \leq H_{p,K}(\Heis_n) \|f\|_p \|\phi\|_p. \end{equation} For $\mu \in (0,\infty)$ and $\lambda_0 \in \mathbb{R} \setminus \{0\}$, take \[ \phi(t) = e^{-\pi \mu t^2 - 2\pi i t \lambda_0}, \] so that \[ \hat\phi(\lambda) = \mu^{-1/2} e^{-(\pi/\mu) (\lambda-\lambda_0)^2} \qquad\text{and}\qquad \|\hat \phi\|_{p'} = B_p \|\phi\|_p, \] since gaussians are extremal functions for the Hausdorff--Young inequality on $\mathbb{R}$. With this choice of $\phi$, the inequality \eqref{eq:test_sharp_HY_heis_tensor} can be rewritten as \[ B_p ( R_f * \Phi_\mu (\lambda_0) )^{1/p'} \leq H_{p,K}(\Heis_n) \|f\|_p, \] where $*$ denotes convolution on $\mathbb{R}$ and \[ R_f(\lambda) = \| \rho(Z_\lambda f) \|_{\Sch^{p'}}^{p'} |\lambda|^n, \qquad \Phi_\mu(\lambda) = \frac{e^{-(\pi p'/\mu) \lambda^2}}{\int_\mathbb{R} e^{-(\pi p'/\mu) s^2} \,ds}. \] Note that $\lambda \mapsto Z_\lambda f$ is continuous $\mathbb{R} \setminus \{0\} \to L^p(\mathbb{C}^n)$ and $\rho : L^p(\mathbb{C}^n) \to \Sch^{p'}(L^2(\mathbb{R}^n))$ is continuous too, so $R_f$ is a continuous function on $\mathbb{R} \setminus \{0\}$. Moreover $\Phi_\mu$ is an approximate identity as $\mu \to 0$. Hence, by taking the limit as $\mu \to 0$, we obtain \[ B_p \sup_{\lambda \in\mathbb{R} \setminus \{0\}} \| \rho(Z_\lambda f) \|_{\Sch^{p'}} |\lambda|^{n/p'} \leq H_{p,K}(\Heis_n) \|f\|_p, \] which for $\lambda = 1$ gives \[ B_p \| \rho(f) \|_{\Sch^{p'}} \leq H_{p,K}(\Heis_n) \|f\|_p, \] that is, $B_p W_{p,K}(\mathbb{C}^n) \leq H_{p,K}(\Heis_n)$. \end{proof} Let us come back to the question whether the identity \eqref{eq:weylconstant} holds for arbitrary $p \in [1,2]$. The following result, which allows for arbitrary $p$ but restricts the class of functions $f$ and, regrettably, also requires a weight in the $p$-norm, gives another indication that this might be true. Recall that a function $f$ on $\mathbb{C}^{n}$ is \emph{polyradial} if \[ f(z) = f_0(|z_1|,\dots,|z_n|), \] or, equivalently, if $f$ is invariant under the $n$-fold product group $\group{U}(1) \times \dots \times \group{U}(1)$. \begin{proposition}\label{prop:locweyl} If $f\in C_c^\infty(\mathbb{C}^{n})$ is polyradial, then, for all $p \in [1,2]$, \begin{equation}\label{eq:hyweyl2} \|\rho(f)\|_{\Sch^{p'} }\leq (B_p)^{2n}\|f e^{(\pi/2) |\cdot|^2}\|_{L^p(\mathbb{C}^{n})}. \end{equation} \end{proposition} As observed in the introduction, this inequality implies that $W^\mathrm{loc}_{p,K}(\mathbb{C}^n) \leq (B_p)^{2n}$ for $K = \group{U}(1) \times \dots \times \group{U}(1)$, and \emph{a fortiori} also for any larger group $K$. \begin{proof} We present a proof of Proposition \ref{prop:locweyl} which follows the philosophy of the proof of Theorem \ref{thm:local-central-HY-compact-Lie}. The key is the following identity relating Laguerre polynomials to Bessel functions: \begin{equation}\label{eq:bela1} L^\alpha_k(x)=\frac {e^x x^{-\alpha/2}}{k!} \int_0^\infty t^{k+\alpha/2} \, J_\alpha(2\sqrt{xt}) \, e^{-t} \, dt \qquad\forall x>0, \end{equation} where $\alpha \in (-1,\infty)$ \cite[(4.19.3)]{lebedev}. In order to avoid technicalities, let us concentrate on the case where $n=1$; we shall later indicate the straightforward changes in the argument which are needed to deal with general $n\geq 1$. If $f(z)=f_0(|z|)$ is a radial $L^1$-function on $\mathbb{C}$, then one may use the orthonormal basis of Hermite functions $h_k$ ($k\in \mathbb{N}$) of $L^2(\mathbb{R})$ to represent the operator $\rho(f)$ as an infinite diagonal matrix, with diagonal elements given by \begin{equation}\label{eq:diag1} \tilde f(k):=\langle \rho(f) h_k,h_k\rangle = \int_{\mathbb{C}} f(z) \, \chi_k(z) \, dz \qquad\forall k\in\mathbb{N}, \end{equation} where $\chi_k$ is the Laguerre function \[ \chi_k(z)=e^{-(\pi/2)|z|^2}L^0_k(\pi |z|^2). \] (see \cite[(1.45) and (1.104)]{Folland-1989}; see also \cite[(1.4.32)]{thangavelu_1998}). In particular, \begin{equation}\label{eq:weyl_schatten_rad} \|\rho(f)\|_{\Sch^q}=\|\tilde f\|_{\ell^q} \end{equation} for all $q \in [1,\infty]$. Recall also that the Euclidean Fourier transform of any radial $L^1$-function $g$ on $\mathbb{C} \cong \mathbb{R}^2$ can be written in polar coordinates as \begin{equation}\label{eq:hatrad} \hat g(\zeta)=2\pi \int_0^\infty g_0(r) \, J_0(2\pi |\zeta| r) r\, dr, \end{equation} where $g(z) = g_0(|z|)$. We assume that $f$ has compact support, and put $F(z)=e^{(\pi/2) |z|^2} f(z)$. Since also $\hat F$ is radial, we may write $\hat F(\zeta)=\hat F_0(|\zeta|).$ Combining \eqref{eq:bela1} and \eqref{eq:hatrad}, we obtain \begin{equation}\label{eq:tk1} \tilde f(k)=\int_0^\infty \hat F_0\big(\sqrt{t/\pi}\big)\, \frac {t^k}{k!} e^{-t} \,dt, \end{equation} which can be re-written as \begin{equation}\label{eq:tk2} \tilde f(k)=\int_{\mathbb{C}} \hat F(\zeta) \frac {\pi^k|\zeta|^{2k}}{k!} e^{-\pi|\zeta|^2} \,d\zeta=\int_{\mathbb{C}}\hat F(\zeta) \, d\mu_k(\zeta), \end{equation} where the measures $d\mu_k$, $k\in \mathbb{N}$, are probability measures on $\mathbb{C}$. Combining the aforementioned Plancherel identity for the Weyl transform, which leads to \[ \sum_{k\in \mathbb{N}}\left|\int_{\mathbb{C}}\hat F(\zeta) \, d\mu_k(\zeta) \right|^2 = \|f\|_2^2 \leq \|F\|_2^2 =\|\hat F\|_2^2, \] with the trivial estimate \[ \sup_{k\in\mathbb{N}} \left|\int_{\mathbb{C}}\hat F(\zeta) \, d\mu_k(\zeta) \right| \leq \|\hat F\|_\infty, \] we see that from here on we can easily modify the argument in the proof of Theorem \ref{thm:local-central-HY-compact-Lie} in order to arrive at \eqref{eq:hyweyl2}. Indeed, an even simpler interpolation argument is possible here, which avoids any smallness assumption on the support of $f$. For suitable functions $\phi$ on the positive real line, let us write \[ \breve \phi(k)=\int_0^\infty \phi(t) \, \frac{t^k}{k!} \, e^{-t} \,dt \] for all $k \in \mathbb{Z}$. We claim that \begin{equation}\label{eq:pqest} \|\breve \phi\|_{\ell^q}\le \|\phi\|_{L^q(\mathbb{R}^+, dt)} \end{equation} for all $q \in [1,\infty]$. Indeed, this estimate is trivial for $q=\infty$, since the $\frac{t^k}{k!} \, e^{-t} \,dt$ are probability measures, and for $q=1$ we may estimate as follows: \[ \sum_{k=0}^\infty|\breve \phi(k)|\le\int_0^\infty |\phi(t)| \sum_{k=0}^\infty \frac{t^k}{k!} \, e^{-t} \,dt =\|\phi\|_1. \] Thus, \eqref{eq:pqest} follows by Riesz--Thorin interpolation. From \eqref{eq:pqest} and \eqref{eq:tk1}, \begin{align*} \|\tilde f \|_{\ell^q}\le \left(\int _0^\infty \left| \hat F_0\big(\sqrt{t/\pi}\big) \right|^q \, dt\right)^{1/q}=\|\hat F\|_q, \end{align*} and thus, by \eqref{eq:weyl_schatten_rad} and the sharp Hausdorff--Young inequality on $\mathbb{R}^2$, we obtain \[ \|\rho(f)\|_{\Sch^{p'}}=\|\tilde f \|_{\ell^{p'}}\le (B_p)^2 \|F\|_p, \] whence \eqref{eq:hyweyl2} follows. Let us finally indicate the changes needed to deal with the case of arbitrary $n$. The Laguerre functions must be replaced by the $n$-fold tensor products \[ \chi_k(z_1,\dots, z_n)=\chi_{k_1}(z_1) \dots \chi_{k_1}(z_1),\] where $k=(k_1,\dots,k_n)\in \mathbb{N}^n$, and thus, in place of \eqref{eq:diag1}, \[ \tilde f(k)=\int_{\mathbb{C}^{n}} f(z_1,\dots, z_n) \,\chi_k(z_1,\dots, z_n)\, dz_1\dots dz_n \] where $k\in \mathbb{N}^n$. Accordingly, the measures $d\mu_k$ must be replaced by the $n$-fold tensor products $d\mu_k=d\mu_{k_1}\otimes \dots \otimes d\mu_{k_n}$, which are again probability measures, and so on. It then becomes evident that the proof carries over without any difficulty to this general case. \end{proof} \begin{remark}\label{rem:newideasneeded} There are indications that it may not be possible to establish \eqref{eq:hyweyl2} without the presence of the weight $e^{(\pi/2) |\cdot|^2}$ by means of a reduction to the Euclidean Fourier transform and the Babenko--Beckner estimate, and that new techniques are required. Let us again restrict our discussion for simplicity to the case $n=1$. There is another interesting identity relating Laguerre functions and Bessel functions, namely \[ e^{-x/2} x^{\alpha/2}L^\alpha_k(x)=\frac{(-1)^k}2 \int_0^\infty J_\alpha(\sqrt{xy}) \, e^{-y/2} \, y^{\alpha/2} \, L_k^\alpha(y) \,dy \qquad \forall x>0, \] where $\alpha \in (-1,\infty)$ \cite[(4.20.3)]{lebedev}. For $\alpha=0$, this in combination with \eqref{eq:hatrad} implies the well-known identity \begin{equation}\label{eq:ch_ft} \chi_k(z)=e^{-(\pi/2)|z|^2} L^0_k(\pi |z|^2)= \frac{(-1)^k}{2} \widehat{\chi_k}(z/2) \end{equation} (see \cite[Remark after Theorem (1.105)]{Folland-1989}, which is based on a more conceptual approach based on the Wigner transform). This easily leads to the identity \begin{equation}\label{eq:tk3} \tilde f(k)=\int_{\mathbb{C}} \hat f(\zeta) \, (-1)^k \, 2 \, \chi_k(2\zeta)\, d\zeta=\int_{\mathbb{C}} \hat f(\zeta) \,d\nu_k(\zeta). \end{equation} In contrast with \eqref{eq:tk2}, the signed measure $d\nu_k$ oscillates when $k \geq 1$ and is no longer a probability measure. Indeed, by \cite[Lemma 1]{markett}, we have \[ \|\chi_k\|_1\sim k^{1/2}\quad \mbox{as}\quad k\to \infty. \] Thus we cannot use \eqref{eq:tk3} in place of \eqref{eq:tk2} as before in order to get a sharp Hausdorff--Young estimate for $\rho(f)$ without a weight. Even the case where $p'=2m$ for some $m \in \mathbb{Z}$ does not seem to allow one to reduce to the Euclidean estimate. Indeed, note that, for all $f \in L^1(\mathbb{C}^n)$, \begin{equation}\label{eq:twisted_weyl} \rho(f^*) = \rho(f)^* \qquad\text{and}\qquad \rho(f) \, \rho(g)=\rho(f\times g), \end{equation} where $f^*(z) = \overline{f(-z)}$ and $f\times g$ denotes the twisted convolution of $f$ and $g$, that is, \begin{equation}\label{eq:twisted} f \times g(z) = \int_{\mathbb{C}^n} f(z-w) \, g(w) \, e^{\pi i \Im (\bar z \cdot w)} \,dw \end{equation} \cite[(1.32)]{Folland-1989}. In particular, if $f$ is radial and real-valued, then $f = f^*$ and therefore \[ \|\rho(f)\|_{\Sch^{2m}}^{2m}=\|\rho(f)^m\|_{\HS}^2 =\| \rho(f\times \cdots \times f)\|_\HS^2 =\|f\times \cdots \times f\|_2^2, \] with $m$ factors $f$. A reduction to the sharp estimate for the Euclidean Fourier transform $\hat f$ of $f$ would therefore require the validity of an estimate of the form \begin{equation}\label{nored} \|f\times \cdots \times f\|_2\le \|\hat f\|_{2m}^m=\|f* \cdots * f\|_2, \end{equation} where $*$ denotes the Euclidean convolution. However this estimate is false, even when $m=2$. Indeed, it is sufficient to test the estimate \eqref{nored} when $f = \chi_k$. Note that, from \eqref{eq:diag1} and the orthogonality of Laguerre polynomials, \[ \tilde\chi_k(l) = \langle \rho(\chi_k) h_l, h_l \rangle = \langle \chi_k, \chi_l \rangle = \delta_{kl}. \] In particular $\rho(\chi_k \times \chi_k) = \rho(\chi_k)$, that is, \[ \chi_k \times \chi_k = \chi_k. \] Therefore \[ \|\chi_k \times \chi_k\|_2 = \|\chi_k\|_2 = 1, \] while \[ \|\hat\chi_k\|_4 = 2^{-1/2} \|\chi_k\|_4 \sim k^{-1/4} (\log k)^{1/4} \quad\text{as } k \to \infty, \] by \eqref{eq:ch_ft} and \cite[Lemma 1]{markett}. This shows that \eqref{nored} cannot hold when $m=2$ and for all radial real-valued functions $f$ (not even with some constant larger than one multiplying the right-hand side). \end{remark} In order to conclude the proof of Theorem \ref{thm:local-HY-weyl}, we need to prove the lower bound \begin{equation}\label{eq:weyl_lowerbd} W^\mathrm{loc}_K(\mathbb{C}^n) \geq (B_{p})^{2n} \end{equation} for any compact subgroup $K$ of $\group{U}(n)$. As we will see, this can be done much as in Section \ref{s:FLq}. For a function $f \in L^1(\mathbb{C}^n) + L^2(\mathbb{C}^n)$, let $T_f$ denote the operator of twisted convolution on the left by $f$, that is, \[ T_f \phi = f \times \phi. \] In analogy with Proposition \ref{prp:fouriernorm}, we can characterise the Schatten norms of Weyl transforms $\rho(f)$ as follows. \begin{proposition} For all $q \in [2,\infty]$ and $f \in C_c(\mathbb{C}^n)$, \[ \|\rho(f)\|_{\Sch^q}^q = \||T_f|^q\|_{L^1(\mathbb{C}^n) \to L^\infty(\mathbb{C}^n)}. \] \end{proposition} \begin{proof} From the Plancherel formula \eqref{eq:weyl_plancherel} for the Weyl transform, together with \eqref{eq:twisted_weyl}, it is easily seen that, for all $f \in L^1(G)$, \[ \|\rho(f)\|_{L^2(\mathbb{R}^n)\to L^2(\mathbb{R}^n)} = \|T_f\|_{L^2(\mathbb{C}^n) \to L^2(\mathbb{C}^n)}. \] This corresponds to the well-known fact that the norm of a linear operator on $L^2(\mathbb{R}^n)$ is the same as the norm of the corresponding left-multiplication operator on $\HS(L^2(\mathbb{R}^n))$. Note, moreover, that the analogue of \eqref{eq:twisted_weyl} holds: \[ T_{f^*} = T_f^* \qquad\text{and}\qquad T_{f \times g} = T_f T_g. \] Hence the correspondence $\rho(f) \mapsto T_f$ induces an isometric $*$-isomorphism between $\Lin(L^2(\mathbb{R}^n))$ and the von Neumann algebra of operators on $L^2(\mathbb{C}^n)$ generated by $\{ T_f \,:\, f \in L^1(\mathbb{C}^n) \}$. Take now $f \in C_c(\mathbb{C}^n)$. Then $\rho(f) \in \Sch^q(\mathbb{C}^n)$ and \[ \| \rho(f) \|_{\Sch^q(\mathbb{C}^n)}^q = \| |\rho(f)|^{q/2} \|_{\HS}^2. \] Since $|\rho(f)|^{q/2} \in \HS(L^2(\mathbb{R}^n))$, by the Plancherel theorem for the Weyl transform there exists $g \in L^2(\mathbb{C}^n)$ such that \[ \rho(g) = |\rho(f)|^{q/2}. \] Since isomorphisms between von Neumann algebras preserve the polar decomposition and the functional calculus, \[ T_g = |T_f|^{q/2}. \] In order to conclude, then it is enough to show that \[ \|\rho(g)\|_{\HS}^2 = \|T_g^2\|_{L^1(\mathbb{C}^n) \to L^\infty(\mathbb{C}^n)}. \] On the other hand, $T_g = |T_f|^{q/2}$ is a nonnegative self-adjoint operator, so \[ \|T_g^2\|_{L^1(\mathbb{C}^n) \to L^\infty(\mathbb{C}^n)} = \|T_g\|^2_{L^1(\mathbb{C}^n) \to L^2(\mathbb{C}^n)} \] and, according to \eqref{eq:twisted}, $T_g$ is an integral operator with kernel $\tilde K_g$ given by \[ \tilde K_g(z,w) = g(z-w) \, e^{\pi i \Im(\bar z \cdot w)}, \] whence \[ \|T_g\|_{L^1(\mathbb{C}^n) \to L^2(\mathbb{C}^n)} = \esssup_{w \in \mathbb{C}^n} \|\tilde K_g(\cdot,w)\|_2 = \|g\|_2 = \|\rho(g)\|_{\HS}, \] and we are done. \end{proof} Given the above characterisation, the proof of the inequality \eqref{eq:weyl_lowerbd} proceeds, much as in Section \ref{s:FLq}, via a ``blow-up'' argument. The main observation here is that, if $S_\lambda$ denotes the $L^1$-isometric scaling on $\mathbb{C}^n$, \[ S_\lambda f(z) = \lambda^{-2n} f(z/\lambda), \] then \[ (S_\lambda f) \times (S_\lambda g) = S_\lambda(f \times_\lambda g), \] where \[ f \times_\lambda g(z) = \int_{\mathbb{C}^n} f(z-w) \, g(w) \, e^{\pi i \lambda^2 \Im (\bar z \cdot w)} \,dw; \] moreover, from the above formula it is clear that, as $\lambda \to 0$, the scaled twisted convolution $\times_\lambda$ tends to the standard convolution on $\mathbb{C}^n \cong \mathbb{R}^{2n}$ (see also \cite{cowling_1981}). Following this idea, it is not difficult to prove the analogues of Lemma \ref{lem:nu_localisation} and Proposition \ref{prop:nu_loc}, where the twisted convolution $\times$ and the standard convolution on $\mathbb{C}^n$ take the place of the convolutions on the Lie group and the Lie algebra respectively. In addition, the action of $\group{U}(n)$ on functions on $\mathbb{C}^n$ commutes with the scaling operators $S_\lambda$ and the twisted convolution, so the analogue of Remark \ref{rmk:symmetry} applies here. We leave the details to the interested reader. \begin{remark} Given the noncommutative subject of this paper, it is natural to ask whether the best constants $H_p(G), H^\mathrm{loc}_p(G),\dots$ are the same in the category of operator spaces (that is, quantized or noncommutative Banach spaces). To be more precise, let us equip the (commutative and noncommutative) $L^q$-spaces involved in the corresponding Hausdorff--Young inequality with their natural operator space structures \cite{Pisier-2003}. Does the complete $L^p \to L^{p'}$ norm of the Fourier transform coincide with the corresponding norm $H_p(G)$ in the category of Banach spaces? In the Euclidean case of $H_p(\mathbb{R}^n)$, this problem was asked by Pisier in 2002 to the fourth-named author, but it is still open. \'Eric Ricard recently noticed that such a result for the Euclidean Fourier transform (that is, its completely bounded norm is still given by the Babenko--Beckner constant raised to the dimension of the underlying space) would give the expected constants for the Weyl transform in CCR algebras and, therefore, also for the Fourier transform in the Heisenberg group. Unfortunately, Beckner's original strategy crucially uses hypercontractivity, which has been recently proved to fail in the completely bounded setting \cite{AAA}. In conclusion, the above discussion indicates one more time (see Remark \ref{rem:newideasneeded}) that some new ideas seem to be necessary to solve these questions. \end{remark} \newcommand{\arttitle}[1]{\lq #1\rq,} \newcommand{\journal}[1]{\textit{#1} } \newcommand{\jvolume}[1]{\textbf{{#1}}} \newcommand{\booktitle}[1]{\emph{{#1}.}} \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}

https://arxiv.org/abs/1412.0840

Natural operations on differential forms

We prove that the only natural operations between differential forms are those obtained using linear combinations, the exterior product and the exterior differential. Our result generalises work by Palais and Freed-Hopkins.As an application, we also deduce a theorem, originally due to Kolar, that determines those natural differential forms that can be associated to a connection on a principal bundle.

\section*{Introduction} Let $\, \omega_1\dots,\omega_k\,$ be differential forms on a smooth manifold, of positive degree. In this paper we determine those differential forms that can be associated in a {\it natural} way to the given forms $\, \omega_1, \ldots , \omega_k$. Loosely speaking, our result says that the only natural operations between differential forms are those obtained using linear combinations, the exterior product and the exterior differential. To be more precise, let us fix positive integers $\, p_1,\dots,p_k\,$ and an integer $\, q\geq 0$. Let us suppose that, for each manifold $\,X\,$ and each collection $\,\omega_1\dots,\omega_k\,$ of differential forms on $\,X\,$ of degree $\,p_1,\dots,p_k$, we have defined a $q$-form $\,P(\omega_1,\dots,\omega_k)\,$ on $\,X$. Let us also assume that the assignment $\,(\omega_1\dots,\omega_k)\longmapsto P(\omega_1,\dots,\omega_k)\,$ is compatible with inverse images: for any smooth map $\,\tau\colon\bar X\to X\,$ it holds $$\tau^*\left[ P(\omega_1,\dots,\omega_k)\right]\,=\, P(\tau^*\omega_1,\dots,\tau^*\omega_k)\ .$$ Then, our main result (Theorem \ref{3.3}) proves that there exists a unique polynomial $\,\mathbf{P}(x_1,y_1,\dots,x_k,y_k)$, homogeneous of degree $q$ in anti-commutative variables of degree $\,\text{deg}\, x_i=p_i$, $\text{deg}\, y_i=p_i+1$, such that: $$P(\omega_1,\dots,\omega_k)\,=\, \mathbf{P}(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega) \ , $$ for any smooth manifold $\,X\,$ and any collection $\,\omega_1\dots,\omega_k\,$ of differential forms on $\,X\,$ of degrees $\,p_1,\dots,p_k$. In this formula, the product of variables on the polynomial $\,\mathbf{P}\,$ has been replaced by the exterior product of forms. The case $\,p_1=\dots=p_k=1\,$ has been recently obtained by Freed-Hopkins (\cite{FreedHopkins}, Thm. 8.1) using a different language and quite different methods to those applied here. When the given collection reduces to a single form $\,\omega\,$ of degree $\,p$, and we take $\,q=p+1$, it follows that the assignment $\,P\,$ is a constant multiple of the exterior differential: $\,P(\omega)=\lambda \, \mathrm{d}\omega$. This is a classical statement due to Palais (\cite{Palais}, Thm. 10.5), who assumed $\, P\,$ to be linear, and to Kol\'{a}\v{r}-Michor-Slov\'{a}k (\cite{KMS}, Prop. 25.4) in the general case. We have included two versions of our main result: a first one, Theorem \ref{2.3}, whose proof relies on general methods developed in \cite{KMS}, and a second one, Theorem \ref{3.3}, which corresponds to the statement announced above. The exposition of this paper will be self-contained, except for the Peetre-Slov\'{a}k theorem on differential operators and some basic facts about the invariant theory of the general linear group. As an application of our results, we prove in the Appendix a version of a beautiful theorem originally due to Kol\'{a}\v{r} (\cite{Kolar}), that determines those differential forms that one can construct in a natural way from a connection on a principal bundle (see also \cite{FreedHopkins}, Thm. 7.20). \section{Preliminaries} \subsection{Invariant theory of the general linear group} Let $\,V\,$ be an $\mathbb{R}$-vector space of finite dimension $\,n$, and let $\,Gl_n\,$ be the Lie group of its $\mathbb{R}$-linear automorphisms. As $\,Gl_n\,$ is linearly semisimple, the following holds: \begin{proposicion}\label{Semisimple} Let $\,E\,$ and $\,F\,$ be linear representations of $\,Gl_n$, and let $\,E'\subset E\,$ be a sub-representation. Any equivariant linear map $\,E' \to F\,$ is the restriction of an equivariant linear map $\,E \to F$. \end{proposicion} On the other hand, the Main Theorem of the invariant theory for the general linear group states that the only $\,Gl_n$-equivariant linear maps $\,\otimes^p V \longrightarrow \otimes^p V\,$ are the linear combinations of permutations of indices (e. g., \cite{KMS} Sect. 24). In this paper, we will only use the following consequence: \begin{proposicion}\label{CorolarioMain} The only $\,Gl_n$-equivariant linear maps $\, \otimes^p V \longrightarrow \Lambda^p V\,$ are the multiples of the skew-symmetrisation operator. \end{proposicion} Combining the previous two propositions, it follows: \begin{corolario}\label{0.3} Let $\,E\subseteq\,\otimes^pV\,$ be a $\,Gl_n$-sub-representation. The only $\,Gl_n$-equi\-va\-riant linear maps $\,\otimes^pV\supseteq E \longrightarrow \Lambda^p V\,$ are the multiples of skew-symme\-trisation operator. \end{corolario} \begin{remark}\label{0.4} We will make use of the following properties of the skew-sy\-mme\-trisa\-tion operator $\,h$:\smallskip -- $h(T\otimes T')=h(T)\wedge h(T')\,$ for any covariant tensors $\,T,T'$,\smallskip -- $h(\omega)=q!\,\omega$ for any $q$-form $\,\omega$,\smallskip -- $h(\nabla\omega)=q!\,\mathrm{d}\omega$ for any differential $q$-form $\omega$ on $\,\mathbb{R}^n$, where $\,\nabla\,$ denotes the standard flat connection. \end{remark} \subsection{Differential operators on forms} Let $\, E\to X\,$ be a bundle (or a submersion) on a smooth manifold $\,X$. Let us denote by $\,\mathcal{E}\,$ the sheaf of sections of that bundle: $\,\mathcal{E}\,$ is a contravariant functor, defined over the category of open sets of $\,X\,$ and inclusions between them, that assigns, to each open set $\,U\subseteq X$, the set $\,\mathcal{E}(U)\,$ of smooth sections defined over $\,U$, and to each inclusion between open sets $\,V\hookrightarrow U$, the restriction map $\,\mathcal{E}(U)\to\mathcal{E}(V)$. Let us denote by $\,J^rE\to X\,$ the bundle of $r$-jets of sections of $\,E\to X$.\medskip \begin{definition} The \textbf{bundle of $\infty$-jets} is the inverse limit $$J^\infty E\, :=\,\varprojlim J^rE\ ,$$ endowed with the initial topology of the projections $\, \pi_r\colon J^\infty E\to J^rE$. \end{definition} Let $\,Y\,$ be a smooth manifold. A continuous map $\,\varphi\colon J^\infty E\to Y\,$ is said \textbf{smooth} if, for any $\infty$-jet $\,j^\infty_xs\in J^\infty E\,$ there exist a natural number $\,r\,$ and a smooth map $\,\varphi_r\colon J^rE\supseteq U\to Y$, defined on a neighbourhood $\,U\,$ of $\,j^r_xs$, such that $\,\varphi=\varphi_r\circ\pi_r$ in the neighbourhood $\,\pi_r^{-1}(U)\,$ of $\,j^\infty_xs$.\medskip \begin{definition} Let $\,E\to X\,$ and $\,F\to X\,$ be two bundles. A \textbf{differential operator} $\,\tilde P\colon E\leadsto F\,$ is a smooth morphism of bundles $\,\tilde P\colon J^\infty E\to F$. A differential operator $\,\tilde P\,$ is said of \textbf{order} $\leq r$ if there exists a morphism of bundles $\,\tilde P_r\colon J^rE\to F\,$ such that $\,\tilde P=\tilde P_r\circ\pi_r$. \end{definition} Let $\,\mathcal{E}\,$ and $\,\mathcal{F}\,$ be the sheaves of sections of two bundles $\,E\to X\,$ and $\,F\to X$. A differential operator $\,\tilde P\colon E\leadsto F$ can be understood as a morphism of sheaves (that is, a morphism of functors): $$P\colon\mathcal{E}\,\longrightarrow\,\mathcal{F}\quad,\quad P(s)(x):=\,\tilde P(j^\infty_xs)\ .$$ Conversely, let us see that any morphism of sheaves, satisfying certain re\-gu\-larity condition, is a differential operator.\medskip If $\,T\,$ is a smooth manifold, let us denote $\,X_T := X\times T$. Any open set $\,U \subset X_T\,$ can be thought as a family of open sets $\,U_t \subset X$, where $\,U_t=U\cap(X\times t)$. A family of sections $\,\{ \, s_t \colon U_t \to E \, \}_{t \in T}\,$ defines a map: $$ s \colon U \to E \quad , \quad s (t,x) := s_t (x) \ $$ and the family $\,\{ s_t \}_{t \in T}\,$ is said {\it smooth} (with respect to the parameters $\,t \in T$) precisely when the map $\,s \colon U \to E\,$ is smooth. \begin{definition} With the previous notations, a morphism of sheaves $\,P \colon \mathcal{E} \to \mathcal{F}\,$ is \textbf{regular} if, for any smooth manifold $\,T\,$ and any smooth family of sections $\,\{ s_t \colon U_t \to E \}_{t \in T}$, the family $\,\{ P(s_t) \colon U_t \to F \}_{t\in T}\,$ is also smooth. \end{definition} \begin{theorem}[Peetre-Slovak]\label{PeetreSlovak} There exists a bijection $$\mathrm{Diff}(E,F)\,=\,\mathrm{Hom}_{\mathrm{reg}}(\mathcal{E},\mathcal{F})\quad ,\quad \tilde P\mapsto P\ ,$$ where $\,\mathrm{Diff}(E,F)\,$ stands for the set of differential operators and $\,\mathrm{Hom}_{\mathrm{reg}}(\mathcal{E},\mathcal{F})\,$ for that of regular morphisms of sheaves. \end{theorem} This theorem was first proved by J. Peetre, in the case of $\mathbb{R}$-linear differential operators. Later on, J. Slov\'{a}k (\cite{Peetre}) established a general result that includes the above statement as a particular case (see \cite{PeetreRevisited}).\medskip \begin{notation} If $\,X\,$ is a smooth manifold, let us write $\,\Omega^p_X\,$ for the sheaf of differential $p$-forms on $\,X$; that is, for the sheaf of sections of the bundle $\,\Lambda^pT^*X\to X$. \end{notation} \begin{lema}\label{1.2} Any differential operator $\,\tilde P\colon\Lambda^pT^*\mathbb{R}^n\leadsto\Lambda^qT^*\mathbb{R}^n$, that is invariant under translations and homotheties, has finite order. \end{lema} The condition of a differential operator $\,\tilde P\,$ being invariant by a diffeomorphism $\,\tau\colon\mathbb{R}^n\to\mathbb{R}^n\,$ means that the following square is commutative $$\begin{CD} J^\infty\Lambda^pT^*\mathbb{R}^n @>{\tau^*}>> J^\infty\Lambda^pT^*\mathbb{R}^n\\ @V{\tilde P}VV @VV{\tilde P}V \\ \Lambda^qT^*\mathbb{R}^n @>{\tau^*}>> \Lambda^qT^*\mathbb{R}^n \end{CD}$$ In terms of the corresponding morphism of sheaves $\,P\colon\Omega^p_{\mathbb{R}^n}\to\Omega^q_{\mathbb{R}^n}$, this amounts to saying $\,P(\tau^*\omega)=\tau^*P(\omega)$.\medskip \noindent{\it Proof of \ref{1.2}:} As the operator $\,\tilde P\,$ is invariant under translations, it is determined by its restriction $\,\tilde P\colon J^\infty_0(\Lambda^pT^*\mathbb{R}^n) \longrightarrow \Lambda^qT^*_0\mathbb{R}^n$ to the fibres over the origin $\,0\in\mathbb{R}^n$. The smoothness of this map at the zero $\infty$-jet $\,j^\infty_00\in J^\infty_0(\Lambda^pT^*\mathbb{R}^n)\,$ implies that there exist a natural number $\,r\,$ and a smooth map $\,\tilde P_r\colon J^r_0(\Lambda^pT^*\mathbb{R}^n)\supseteq U \longrightarrow \Lambda^qT^*_0\mathbb{R}^n$, defined on an open set $\,U\,$ of the zero $r$-jet, such that $\,\tilde P=\tilde P_r\circ\pi_r\,$ over $\,\pi_r^{-1}(U)$. Now, a simple computation in local coordinates shows that, for any fixed jet $\,j^\infty_0\omega\in J^\infty_0\Lambda^pT^*\mathbb{R}^n$, there exists an homothety $\,\tau_\lambda\colon\mathbb{R}^n\to\mathbb{R}^n$, $\,\tau_\lambda(x)=\lambda x$, with $\,\lambda\,$ sufficiently small, such that $\,\tau_\lambda^*(j^r_0\omega)\,$ belongs to $\,U$, that is, $\,\tau_\lambda^*(j^\infty_0\omega)\in \pi_r^{-1}(U)$. Therefore, \begin{align*} \tilde P(j^\infty_0\omega)\, &=\,\tilde P(\tau_{\lambda^{-1}}^*\tau_\lambda^*(j^\infty_0\omega))\,=\,\tau_{\lambda^{-1}}^*\left(\tilde P(\tau_\lambda^*(j^\infty_0\omega))\right)\,=\,\tau_{\lambda^{-1}}^*\left(\tilde P_r\left(\pi_r\tau_\lambda^*\left(j^\infty_0\omega\right)\right)\right) \\ &=\, \tau_{\lambda^{-1}}^*\left(\tilde P_r\left( \tau_\lambda^*\pi_r\left(j^\infty_0\omega\right)\right)\right)\,=\,\tau_{\lambda^{-1}}^*\left(\tilde P_r\left(\tau_\lambda^*\left(j^r_0\omega\right)\right)\right) =\,\tilde Q_r(j^r_x\omega)\ . \end{align*} where $\,\tilde Q_r :=\tau_{\lambda^{-1}}^*\circ\tilde P_r\circ\tau_\lambda^*$. The same formula $\,\tilde P=\tilde Q\circ\pi_r\,$ holds for any point in the neighbourhood of $\,j^\infty_0\omega\,$ formed by jets $\,j^\infty_0\bar\omega\,$ such that $\,\tau_\lambda^*(j^\infty_0\bar\omega)\in \pi_r^{-1}(U)$. Hence, $\,\tilde P\,$ factors through the projection $\,\pi_r$. \hfill$\square$ \medskip \begin{notation} For the rest of the section, let us fix a natural number $\, n$. Let $\,\textbf{Man}_n\,$ denote the category of smooth manifolds of dimension $\,n\,$ and local diffeomorphisms between them. \end{notation} Let $\, \Omega^p\,$ denote the contravariant functor over $\,\textbf{Man}_n\,$ that assigns, to each smooth $n$-manifold $\,X$, the set $\,\Omega^p(X)\,$ of differential $p$-forms on $\,X$, and, to each local diffeomorphism $\,\tau\colon X\to Y$, the map $\,\tau^*\colon\Omega^p(Y)\to\Omega^p(X)\,$ (inverse image of differential forms). \begin{definition} A \textbf{natural} morphism $\,P\colon\Omega^p\to\Omega^q\,$ is a morphism of functors over $\,\textbf{Man}_n$. In other words, a natural morphism $\,P\colon\Omega^p\to\Omega^q\,$ amounts to giving, for each $n$-manifold $\,X$, a map $\,P\colon\Omega^p(X)\to\Omega^q(X)$ such that for each local diffeomorphism $\,\tau\colon X\to Y$, it holds $$P(\tau^*\omega)\,=\,\tau^*(P(\omega))\qquad \forall \, \omega\in\Omega^p(Y)\ .$$ \end{definition} \medskip Let $\,X\,$ be an $n$-manifold. The sheaf of $p$-forms $\,\Omega_X^p\,$ over $\,X\,$ is the restriction of $\,\Omega^p\,$ to the subcategory of $\,\textbf{Man}_n\,$ formed by the open sets of $\,X\,$ and the inclusions between them. A natural morphism $\,P\colon\Omega^p\to\Omega^q\,$ defines, by restriction to such subcategory, a morphism $\,P\colon\Omega_X^p\to\Omega_X^q\,$ of sheaves over $X$. \begin{definition} A natural morphism $\,P\colon\Omega^p\to\Omega^q\,$ is said \textbf{regular} if its restriction $\,P\colon\Omega^p_X\to\Omega^q_X\,$ to any $n$-manifold $\,X\,$ is a regular morphism of sheaves over $\,X$. \end{definition} Let $\,P\colon\Omega^p\to\Omega^q\,$ be a regular and natural morphism. As any $n$-manifold $\,X\,$ is locally diffeomorphic to $\,\mathbb{R}^n$, the natural morphism $\,P\colon\Omega^p\to\Omega^q\,$ is determined by its restriction to the manifold $\,\mathbb{R}^n$, that is, by the regular morphism of sheaves $\,P\colon\Omega^p_{\mathbb{R}^n}\to\Omega^q_{\mathbb{R}^n}$. By the Peetre-Slov\'{a}k theorem, such a morphism is in fact defined by a differential operator $\,\tilde P\colon J^\infty\Lambda^pT^*\mathbb{R}^n\longrightarrow \Lambda^qT^*\mathbb{R}^n$. By naturalness, this differential operator is invariant under the action of any diffeomorphism $\,\mathbb{R}^n\to\mathbb{R}^n$, and hence, due to Lemma \ref{1.2}, has finite order. Summing up, \begin{proposicion}\label{1.3} Any regular and natural morphism $\,P\colon\Omega^p\to\Omega^q\,$ is a differential operator of finite order.\end{proposicion} To be more precise, the restriction of $\,P\,$ to each $n$-manifold $\,X\,$ is a morphism of sheaves $\,P\colon\Omega^p_X\to\Omega^q_X\,$ corresponding to a differential operator $\,\tilde P\colon \Lambda^pT^*X\leadsto\Lambda^qT^*X\,$ of finite order. This order, that is common for all manifolds, will be called the \textbf{order} of the morphism $\,P\colon\Omega^p\to\Omega^q$. \begin{definition} For each natural number $\,r$, let us write $\,G^r\,$ for the Lie group of $r$-jets $\,j^r_0\tau\,$ at the origin $\,0\in\mathbb{R}^n\,$ of local diffeomorphisms satisfying $\,\tau(0)=0$. \end{definition} The group $\,G^{r+1}\,$ acts on the fibre $\,J^r_0(\Lambda^pT^*\mathbb{R}^n)$, $$(j^{r+1}_0\tau)\cdot (j^r_0\omega):=\, j^r_0\left(\tau_*\omega\right)\ .$$ \begin{proposicion}\label{1.4} There exists an injective map: $$\left\{\begin{aligned}\mathrm{Regular\ and \ natural \ morphisms \ }\\ P\colon\Omega^p\longrightarrow\Omega^q\ \mathrm{\ of \ order \ }r\quad\end{aligned}\right\}\,\subseteq\, \left\{\begin{aligned}\mathrm{Smooth \ }G^{r+1}\text{--}\,\mathrm{equivariant \ maps \ }\\ \tilde P\colon J^r_0(\Lambda^pT^*\mathbb{R}^n)\longrightarrow \Lambda^qT^*_0\mathbb{R}^n\qquad\end{aligned}\right\}$$ \end{proposicion} The relation between $\,P\,$ and $\,\tilde P\,$ is determined by the equality: $$P(\omega)_{x=0}\,=\,\tilde P(j^r_0\omega)\qquad\forall\, \omega\in\Omega^p(\mathbb{R}^n)\ .$$ \noindent \textit{Proof: } Any regular and natural morphism $\,P\colon\Omega^p\to\Omega^q$ of order $\,r\,$ is determined by its restriction to $\mathbb{R}^n$, that is a regular morphism of sheaves $\,P\colon\Omega^p_{\mathbb{R}^n}\to\Omega^q_{\mathbb{R}^n}\,$ invariant under diffeomorphisms. As it is regular, it corresponds to a differential operator $\,\tilde P\colon J^r(\Lambda^pT^*\mathbb{R}^n)\longrightarrow \Lambda^qT^*\mathbb{R}^n$, that, by naturalness, is invariant under diffeomorphisms. Due to this last property, $\tilde P$ is determined by its restriction to the fibres over the origin, $\,\tilde P\colon J^r_0(\Lambda^pT^*\mathbb{R}^n)\longrightarrow \Lambda^qT^*_0\mathbb{R}^n$, which, in turn, is a smooth, $G^{r+1}$-equivariant map. \hfill $\square$ With more generality, in this paper we will consider the functors $\,\Omega^{p_1}\oplus\cdots\oplus\Omega^{p_k}\,$ over $\,\textbf{Man}_n$. The definition of naturalness, as well as Propositions \ref{1.3} and \ref{1.4}, all trivially extend to morphisms $\,\Omega^{p_1}\oplus\cdots\oplus\Omega^{p_k}\longrightarrow\Omega^q$. In particular, \begin{proposicion}\label{1.5} There exists an injective map $$\begin{CD} \left\{\begin{aligned}\mathrm{Regular \ and \ natural \ morphisms \ of \ order \ }r\\ P\colon\Omega^{p_1}\oplus\cdots\oplus\Omega^{p_k}\,\longrightarrow\,\Omega^q\qquad\qquad\end{aligned}\right\}\\ |\bigcap^{\phantom{\bigcap}} _{\phantom{\bigcap}} \\ \left\{\begin{aligned} &\mathrm{Smooth\ }G^{r+1}\text{--}\,\mathrm{equivariant \ maps\ }\ \\ \tilde P\colon J^r_0(\Lambda^{p_1}&T^*\mathbb{R}^n)\times\cdots\times J^r_0(\Lambda^{p_k}T^*\mathbb{R}^n)\,\longrightarrow\, \Lambda^qT^*_0\mathbb{R}^n\end{aligned}\right\}\end{CD}$$ \end{proposicion} The relation between $\,P\,$ and $\,\tilde P\,$ is determined by the equality $$P(\omega_1,\dots,\omega_k)_{x=0}\,=\,\tilde P(j^r_0\omega_1,\dots,j^r_0\omega_k)$$ for all $\,(\omega_1,\dots,\omega_k)\in\Omega^{p_1}(\mathbb{R}^n)\times\dots\times \Omega^{p_k}(\mathbb{R}^n)$.\medskip \begin{remark} Later on, we will compute those smooth, $G^{r+1}$-equi\-va\-riant maps, and it will result that the inclusions of \ref{1.4} and \ref{1.5} are in fact equalities (see \cite{KMS} for more general results of this kind). \end{remark} \section{Main result: first version} In this section we still work in category $\,\textbf{Man}_n\,$ of $n$-dimensional manifolds and local diffeomorphisms between them. Let us introduce the following notations for the exterior and symmetric powers: $\,\Lambda^p:=\Lambda^p(T^*_0\mathbb{R}^n)\,$ and $\, S^p:=S^p(T^*_0\mathbb{R}^n)$. We will also use that $\,S^p\,$ is canonically isomorphic to the vector space of homogeneous polynomials of degree $\,p\,$ over $\,T^*_0\mathbb{R}^n$. There exists a diffeomorphism: $$\begin{CD} J^r_0(\Lambda^pT^*\mathbb{R}^n) @= (\Lambda^p)\oplus(S^1\otimes \Lambda^p)\oplus\cdots\oplus(S^r\otimes \Lambda^p)\\ j^r_0\omega & \longmapsto & (\omega^0,\omega^1,\dots\dots ,\omega^r)\qquad\qquad \end{CD}$$ where $\,\omega^s\,$ is the homogeneous component of degree $\,s\,$ in the Taylor expansion (on cartesian coordinates) of the $p$-form $\,\omega$. If we write $\,\nabla\,$ for the standard flat connection on $\,\mathbb{R}^n$, then $\,\omega^s=(\nabla^s\omega)_{x=0}$. This diffeomorphism is not invariant under arbitrary changes of coordinates, but it is so under linear changes of coordinates. In other words, this diffeomorphism is equivariant with respect to the action of the linear group $\,Gl_n \subseteq G^{r+1}$, although it is not equivariant with respect to the whole group $\,G^{r+1}$. Therefore, Proposition \ref{1.5} implies the following: \begin{lema}\label{2.1} There exists an injective map: $$\begin{CD} \left\{\begin{aligned}\mathrm{Regular\ and \ natural \ morphisms \ of \ order \ }r\\ P\colon\Omega^{p_1}\oplus\cdots\oplus\Omega^{p_k}\,\longrightarrow\,\Omega^q\qquad\qquad\end{aligned}\right\}\\ |\bigcap^{\phantom{\bigcap}} _{\phantom{\bigcap}} \\ \left\{\begin{aligned} &\mathrm{Smooth\ }Gl_n\text{--}\,\mathrm{equivariant \ maps \ }\quad\\ \tilde P\colon\prod_{i=1,\dots,k} &\left[(S^0\otimes \Lambda^{p_i})\oplus\cdots\oplus(S^r\otimes \Lambda^{p_i})\right]\,\longrightarrow\, \Lambda^q\end{aligned}\right\}\end{CD}$$ \end{lema} The relation between $\,P\,$ and $\,\tilde P\,$ is determined by the equality \begin{align*} P(\omega_1,\dots,\omega_k)_0\, &=\,\tilde P(\dots,\left[\omega_i^0,\dots,\omega_i^r\right],\dots) \\ &=\, \tilde P(\dots,\left[(\nabla^0\omega_i)_0,\dots,(\nabla^r\omega_i)_0\right],\dots)\ . \end{align*} \medskip Again, this inclusion will be proved to be an equality, once we compute the $Gl_n$-equivariant maps. The first step in this computation is to prove that these maps satisfy certain homogeneity condition. \begin{lema}\label{2.2} Any smooth, $\,Gl_n$-equivariant map $$\begin{CD}\prod_{i=1}^k\left[(S^0\otimes \Lambda^{p_i})\oplus\cdots\oplus(S^r\otimes \Lambda^{p_i})\right] @>{\tilde P}>> \Lambda^q \end{CD}$$ satisfies the following homogeneity condition \begin{equation}\label{CondicionHomog} \tilde P\left(\dots,(\lambda^{p_i+0}\omega_i^0,\dots,\lambda^{p_i+r}\omega_i^r),\dots\right)\,=\,\lambda^q\tilde P\left(\dots,(\omega_i^0,\dots,\omega_i^r),\dots\right) \end{equation} for any real number $\, \lambda\neq 0$. \end{lema} \noindent \textit{Proof: } Let $\,\tau_\lambda\colon\mathbb{R}^n\to\mathbb{R}^n$, $\tau_\lambda(x)=\lambda x$, be the homothety of ratio $\,\lambda\neq 0$. On the one hand, for any tensor $\,\omega_i^j\in S^j\otimes\Lambda^{p_i}$, covariant of order $\,p_i+j$, it holds $\,\tau_\lambda^*\omega_i^j=\lambda^{p_i+j}\omega_i^j$. On the other, as the map $\, P\,$ is $Gl_n$-equivariant, it holds $\,\tilde P\circ\tau_\lambda^*=\tau_\lambda^*\circ\tilde P$. Consequently: \begin{align*} \tilde P & \left(\dots,(\lambda^{p_i+0}\omega_i^0,\dots,\lambda^{p_i+r}\omega_i^r),\dots\right)\, =\, \tilde P\left(\dots,(\tau_\lambda^*\omega_i^0,\dots,\tau_\lambda^*\omega_i^r),\dots\right) \\ & =\,\tau_\lambda^*\tilde P\left(\dots,(\omega_i^0,\dots,\omega_i^r),\dots\right) \,=\,\lambda^q\tilde P\left(\dots,(\omega_i^0,\dots,\omega_i^r),\dots\right)\ . \end{align*} \hfill $\square$ \medskip A smooth map between vector spaces satisfying a homogeneity condition has necessarily to be a polynomial, in virtue of the following elementary result (e. g. \cite{KMS}, Thm. 24.1): \medskip \noindent {\bf Homogeneous Function Theorem:} {\it Let $E_1,\dots, E_k$ be finite dimensional $\mathbb{R}$-vector spaces. Let $\, f \, \colon \prod E_i \to \mathbb{R}$ be a smooth function such that there exist positive real numbers $a_i > 0$, and $ w \in \mathbb{R}$ satisfying: \begin{equation}\label{CondicionHomogeneidadLemma} f ( \lambda^{a_1} e_1 , \ldots , \lambda^{a_k} e_k) = \lambda^w \, f(e_1 , \ldots , e_k) \end{equation} for any positive real number $\lambda >0$ and any $(e_1 , \ldots , e_k) \in \prod E_i$. Then $f$ is a sum of monomials of degree $(d_1,\dots,d_k)$ in the variables $e_1,\dots,e_k$ satisfying the relation \begin{equation}\label{CondicionMonomios} a_1 d_1 + \cdots + a_k d_k = w \ . \end{equation} If there are no natural numbers $d_1,\dots,d_k \in \mathbb{N} \cup \{ 0 \}$ satisfying this equation, then $f$ is the zero map. } \medskip In other words, for any finite dimensional vector space $W$, there exists an $\mathbb{R}$-linear isomorphism: $$\begin{CD} \left[ \text{Smooth maps }\, f \colon \prod E_i \to W\,\text{ satisfying } (\ref{CondicionHomogeneidadLemma})\right] \\ @| \\ \bigoplus \limits _{(d_1 , \ldots , d_k)}\text{Hom}_{\mathbb{R}}(S^{d_1} E_1 \otimes \ldots \otimes S^{d_k} E_k,\, W) \end{CD}$$ where $(d_1, \ldots , d_k)$ runs over the non-negative integers solutions of equation (\ref{CondicionMonomios}). A smooth map $\,f\colon\prod E_i\to W$, satisfying (\ref{CondicionHomogeneidadLemma}), and the corresponding linear map $\,\oplus f^{d_1\dots d_k}\in \bigoplus_{(d_1 , \ldots , d_k)}\text{Hom}_{\mathbb{R}}(S^{d_1} E_1 \otimes \ldots \otimes S^{d_k} E_k,\, W)$, are related by the equality $$f(e_1,\dots,e_k)\,=\,\sum_{(d_1 , \ldots , d_k)}f^{d_1\dots d_k}\left((e_1\otimes\overset{d_1}{\dots}\otimes e_1) \otimes\cdots\cdots\otimes (e_k\otimes\overset{d_k}{\dots}\otimes e_k) \right)\ .$$ \medskip \begin{definition} Let us fix a finite sequence of positive integers $\,(p_1,\dots,p_k)$. Let us denote $\,\mathbb{R}\{ u_1,\dots,u_k\}\,$ the {\it anti-commutative} algebra of polynomials with real coefficients in the variables $\,u_1,\dots, u_k$, where each variable $\,u_i\,$ is assigned degree $\,p_i$. The anti-commutative character of this algebra is expressed by the relations $$u_iu_j\,=\, (-1)^{p_ip_j}u_ju_i\ .$$ The degree of a monomial $\,u_1^{a_1}\dots u_k^{a_k}\,$ is defined as $\,\sum a_ip_i$. A polynomial $\,\mathbf{P}(u_1,\dots, u_k)\in \mathbb{R}\{ u_1,\dots,u_k\}\,$ is said homogeneous of degree $\,q\,$ if it is a linear combination of monomials of degree $\,q$. \end{definition} Let $\,\mathbf{P}(x_1,\dots, x_k)\in \mathbb{R}\{ u_1,\dots,u_k\}\,$ be a homogeneous polynomial of degree $\,q$, and let $\,\omega_1,\dots,\omega_k\,$ differential forms of degree $\,p_1,\dots,p_k\,$ on a smooth manifold $\,X$. Then $\,\mathbf{P}(\omega_1,\dots,\omega_k)$, where the product of variables is replaced by the exterior product of forms, is a differential form of degree $\,q\,$ on $\,X$. \begin{teorema}\label{2.3} Let $\,p_1,\dots,p_k\,$ be positive integers, and let $\,\mathbb{R}\{ u_1,v_1,\dots,u_k,v_k\}\,$ be the anti-commutative algebra of polynomials on the variables $\,u_1,v_1,\dots,u_k,v_k$, of degree $\,\mathrm{deg}\, u_i=p_i$, $\,\mathrm{deg}\, v_i=p_i+1$. Any regular and natural morphism over $\mathbf{Man}_n$ $$ P\colon \Omega^{p_1} \oplus \ldots \oplus \Omega^{p_k} \ \longrightarrow \ \Omega^q\qquad (0\leq q\leq n) $$ can be written as $$P(\omega_1,\dots,\omega_r)\,=\, \mathbf{P}(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k)$$ for a unique homogeneous polynomial $\,\mathbf{P}(u_1,v_1,\dots,u_k,v_k)\in\mathbb{R}\{ u_1,v_1,\dots,u_k,v_k\}\,$ of degree $\,q$. \end{teorema} \noindent \textit{Proof: } Via Lemma \ref{2.1}, a regular natural morphism $\,P\colon \Omega^{p_1} \oplus \ldots \oplus \Omega^{p_k}\longrightarrow\Omega^q$, of order $r$, corresponds to a $Gl_n$-equivariant smooth map $$\tilde P\colon \prod_{i=1,\dots,k}\left[(S^0\otimes \Lambda^{p_i})\oplus\cdots\oplus(S^r\otimes \Lambda^{p_i})\right]\,\longrightarrow\, \Lambda^q\ ,$$ satisfying the homogeneity condition (\ref{CondicionHomog}). Due to the Homogeneous Function Theorem, such an $\,\tilde P\,$ has to be a polynomial; that is to say, it is a sum $\,\oplus_{\{d_{i,j}\}}\, \tilde P^{\{d_{i,j}\}}\,$ of $Gl_n$-equivariant linear maps: $$ \tilde P^{\{ d_{i,j}\}}: \bigotimes_{i=1,\dots,k}\left[S^{d_{i,0}} (S^0\otimes\Lambda^{p_i}) \otimes\cdots\otimes S^{d_{i,r}} (S^r\otimes\Lambda^{p_i})\right]\, \longrightarrow \, \Lambda^{q} $$ where each sequence $\,\{ d_{i,j}\}\,$ of non-negative integers is a solution to the equation: \begin{equation}\label{Soluciones} \left[p_1d_{1,0}+\cdots+(p_1+r)d_{1,r}\right]+ \cdots + \left[p_kd_{k,0}+\cdots+(p_k+r)d_{k,r}\right] \, = \, q \ . \end{equation} Observe that this condition implies that $\,\tilde P^{\{ d_{i,j}\}}\,$ is in fact defined on a vector subspace of $\,\otimes^qT_0^*\mathbb{R}^n$. Then, by Proposition \ref{0.3}, the linear map $\,\tilde P^{\{ d_{i,j}\}}\,$ is a multiple of the skew-symmetrisation operator $\,h$, $$\tilde P^{\{ d_{i,j}\}}\,=\, \lambda^{\{ d_{i,j}\}}\cdot h\ .$$ The skew-symmetrisation of two symmetric indices vanishes, so that we may assume, from now on, that $$ d_{i,2} =d_{i,3} =\ldots =d_{i,r}= 0 \quad , \quad \mbox{ for all } i = 1 , \ldots , k.$$ That is to say, we only consider solutions $\,\{ d_{1,0},d_{1,1},\dots,d_{k,0},d_{k,1}\}\,$ to the equation: \begin{equation}\label{SolucionesII} \left[p_1d_{1,0}+(p_1+1)d_{1,1}\right]+ \cdots + \left[p_kd_{k,0}+(p_k+1)d_{k,1}\right] \, = \, q \ . \end{equation} Bringing all this together, $$P(\omega_1,\dots,\omega_k)_{x=0}\,\overset{2.1}{=}\,\tilde P\left((\omega_1)_0,(\nabla\omega_1)_0,\dots,(\omega_k)_0,(\nabla\omega_k)_0\right) $$ $$=\,\sum_{\{ d_{i,j}\}}\tilde P^{\{ d_{i,j}\}}\left(\left[(\omega_1)_0\otimes\overset{d_{1,0}}{\dots}\otimes(\omega_1)_0\right]\otimes\left[(\nabla\omega_1)_0\otimes\overset{d_{1,1}}{\dots}\otimes(\nabla\omega_1)_0\right]\otimes\dots\right.$$ $$ \left. \dots\otimes \left[(\omega_k)_0\otimes\overset{d_{k,0}}{\dots}\otimes(\omega_k)_0\right]\otimes\left[(\nabla\omega_k)_0\otimes\overset{d_{k,1}}{\dots}\otimes(\nabla\omega_1)_0\right]\right)$$ $$=\,\sum_{\{ d_{i,j}\}}\lambda^{\{ d_{i,j}\}}h\left(\left[(\omega_1)_0\otimes\overset{d_{1,0}}{\dots}\otimes(\omega_1)_0\right]\otimes\left[(\nabla\omega_1)_0\otimes\overset{d_{1,1}}{\dots}\otimes(\nabla\omega_1)_0\right]\otimes\dots\right.$$ $$ \left. \dots\otimes \left[(\omega_k)_0\otimes\overset{d_{k,0}}{\dots}\otimes(\omega_k)_0\right]\otimes\left[(\nabla\omega_k)_0\otimes\overset{d_{k,1}}{\dots}\otimes(\nabla\omega_1)_0\right]\right)$$ (using Properties \ref{0.4} of the skew-symmetrisation operator) $$=\,\sum_{\{ d_{i,j}\}}\mu^{\{ d_{i,j}\}}\left(\left[(\omega_1)_0\wedge\overset{d_{1,0}}{\dots}\wedge(\omega_1)_0\right]\wedge\left[(\mathrm{d}\omega_1)_0\wedge\overset{d_{1,1}}{\dots}\wedge(\mathrm{d}\omega_1)_0\right]\wedge\dots\right.$$ $$ \left. \dots\wedge \left[(\omega_k)_0\wedge\overset{d_{k,0}}{\dots}\wedge(\omega_k)_0\right]\wedge\left[(\mathrm{d}\omega_k)_0\wedge\overset{d_{k,1}}{\dots}\wedge(\mathrm{d}\omega_1)_0\right]\right)$$ where $\mu^{\{ d_{i,j}\}}:=\lambda^{\{ d_{i,j}\}}(p_1!)^{d_{1,0}+d_{1,1}}\cdots(p_k!)^{d_{k,0}+d_{k,1}}$. Therefore, taking $$\mathbf{P}(u_1,v_1,\dots,u_k,v_k):=\,\sum_{\{ d_{i,j}\}}\mu^{\{ d_{i,j}\}}u_1^{d_{1,0}}v_1^{d_{1,1}}\cdots u_k^{d_{k,0}}v_k^{d_{k,1}}$$ it follows $$P(\omega_1,\dots,\omega_r)_{x=0}\,=\, \mathbf{P}(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k)_{x=0}\ .$$ By naturalness, we conclude \begin{equation}\label{CondUnicidad} P(\omega_1,\dots,\omega_k)\,=\, \mathbf{P}(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k) \end{equation} for any $n$-manifold $\,X\,$ and any $\,(\omega_1,\dots,\omega_k)\in \Omega^{p_1}(X)\oplus\dots\oplus\Omega^{p_k}(X)$. Finally, the uniqueness of the polynomial $\,\mathbf{P}\,$ is proved applying Lemma \ref{2.4} below to the difference of two polynomials satisfying (\ref{CondUnicidad}). \hfill $\square$ \begin{lema}\label{2.4} Let $\,\mathbb{R}\{ u_1,v_1,\dots,u_k,v_k\}\,$ be as in Theorem \ref{2.3}. Let $\,\mathbf{P}\in\mathbb{R}\{ u_1,v_1,\dots,u_k,v_k\}\,$ be a non-zero homogeneous polynomial of degree $\,q\leq n$. Then, the morphism of sheaves over $\,X=\mathbb{R}^n$ $$ P\colon \Omega^{p_1}_X \oplus \ldots \oplus \Omega^{p_k}_X \ \longrightarrow \ \Omega^q_X\quad ,\qquad P(\omega_1,\dots,\omega_k)=\mathbf{P}(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k) $$ is not identically zero. \end{lema} \noindent \textit{Proof: } Let us first consider the case $\,k=1$, that is, $\mathbf{P}\in\mathbb{R}\{u,v\}$. In this case, $\,\mathbf{P}\,$ is (up to a scalar factor) a monomial of one of the following four types, depending on the parities of $\,p_1\,$ and $\,q$, $$\left\{\begin{aligned} &\ u^0v^s\qquad (p_1\text{ odd, }q \text{ even}) \\ &\ u^1v^s\qquad (p_1\text{ odd, }q\text{ odd}) \\ &\ u^sv^0\qquad (p_1\text{ even, }q \text{ even}) \\ &\ u^sv^1\qquad (p_1\text{ even, }q\text{ odd})\end{aligned}\right.$$ where $\,s\,$ has to be taken in each case, for the monomial to have degree $\,q$. Depending on each of the four cases, let us consider the $\,p_1$-form $$\omega_1\,=\left\{\begin{aligned} \sum_{j=1}^sy_{j_0}\mathrm{d} y_{j_1}\wedge\dots\wedge\mathrm{d} y_{j_{p_1}} \quad , \quad \text{case }u^0v^s\\ \mathrm{d} x_1\wedge\dots\wedge\mathrm{d} x_{p_1}+\sum_{j=1}^sy_{j_0}\mathrm{d} y_{j_1}\wedge\dots\wedge\mathrm{d} y_{j_{p_1}}\quad , \quad \text{case }u^1v^s\\ \sum_{j=1}^s\mathrm{d} y_{j_1}\wedge\dots\wedge\mathrm{d} y_{j_{p_1}} \quad , \quad \text{case }u^sv^0\\ x_0\mathrm{d} x_1\wedge\dots\wedge\mathrm{d} x_{p_1}+\sum_{j=1}^s\mathrm{d} y_{j_1}\wedge\dots\wedge\mathrm{d} y_{j_{p_1}} \quad , \quad \text{case }u^sv^1 \end{aligned}\right.$$ on $\,X=\mathbb{R}^q$ (each formula employs $\,q\,$ variables). In each of the four cases, it holds $$P(\omega_1)\,=\,\mathbf{P}(\omega_1,\mathrm{d}\omega_1)\,=\left\{\begin{aligned} (\omega_1)^0\wedge(\mathrm{d}\omega_1)^s\,=\, s! \,\omega_{\mathbb{R}^q} \\ (\omega_1)^1\wedge(\mathrm{d}\omega_1)^s\,=\, s! \,\omega_{\mathbb{R}^q} \\ (\omega_1)^s\wedge(\mathrm{d}\omega_1)^0\,=\, s! \,\omega_{\mathbb{R}^q} \\ (\omega_1)^s\wedge(\mathrm{d}\omega_1)^1\,=\, s! \,\omega_{\mathbb{R}^q} \end{aligned}\right.$$ where $\,\omega_{\mathbb{R}^q}\,$ is the volume form of $\,\mathbb{R}^q$. Hence, the morphism $\,P\colon \Omega^{p_1}_X\longrightarrow \Omega^q_X\,$ is not identically zero for $\,X=\mathbb{R}^q$. Let us now consider the case where $\,k\,$ is arbitrary. Let $$\mathbf{P}\,=\,\sum\lambda_{a_1b_1\dots a_kb_k}\, u_1^{a_1}v_1^{b_1}\dots u_k^{a_k}v_k^{b_k}$$ be a homogeneous polynomial of degree $\,q$. Let us fix an index $\,a_1b_1\dots a_kb_k\,$ whose corresponding coefficient is non-zero. On each pair $\,a_ib_i\,$ one of the two terms is $0$ or $1$, so that we have again four possibilities $\,0s_i$, $\,1s_i$, $\,s_i0$ and $\,s_i1$. Let us now define on $\,\mathbb{R}^q\,$ a $p_i$-form $\,\omega_i\,$ using the four formulas above, depending on the case (but taking care of writing the different forms $\,\omega_1,\dots,\omega_k\,$ on $\,\mathbb{R}^q$ with disjoint variables). It is now easy to check that: $$P(\omega_1,\dots,\omega_k)\,=\,\mathbf{P}(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k)\,=\,\lambda_{a_1b_1\dots a_kb_k}s_1!\cdots s_k!\cdot\omega_{\mathbb{R}^q}\ ,$$ so that $\,P\colon \Omega^{p_1}_X \oplus \ldots \oplus \Omega^{p_k}_X\longrightarrow \Omega^q_X\,$ is not identically zero when $\,X=\mathbb{R}^q$. Finally, the same expressions for $\,\omega_1,\dots,\omega_k\,$ prove the thesis for $\,X=\mathbb{R}^n\,$ when $\,n\geq q$. \hfill $\square$ \medskip A particular case of Theorem \ref{2.3} is the following characterisation of the exterior differential, that was first proved for $\mathbb{R}$-linear morphisms by Palais (\cite{Palais}, Thm. 10.5), and in the general case by Kol\'{a}\v{r}-Michor-Slov\'{a}k (\cite{KMS}, Prop. 25.4). \begin{corolario} Let $\,p\,$ be a positive integer. Up to a constant factor, the only regular and natural morphism $\,P\colon \Omega^p\to\Omega^{p+1}\,$ over $\,\mathbf{Man}_n\,$ is the exterior differential: $\,P(\omega)=\mathrm{d}\omega$. \end{corolario} \noindent \textit{Proof: } According to the previous Theorem, we have to consider the algebra $\,\mathbb{R}\{ u,v\}$, where $\,\text{deg}\, u=p$, $\,\text{deg}\, v=p+1$. In this algebra, the only (up to scalar factors) homogeneous polynomial of degree $\,p+1\,$ is the monomial $\,v$, that corresponds to the natural morphism $\,P\colon \Omega^p\to\Omega^{p+1}$, $\,P(\omega)=\mathrm{d}\omega$. \hfill $\square$ \section{Main result: second version} Let us now prove a variation of Theorem \ref{2.3}, that corresponds to the statement announced in the Introduction.\medskip \begin{notation} Let \textbf{Man} be the category of all smooth manifolds and arbitrary smooth maps between them. \end{notation} Firstly, let us check that any morphism of functors over $\,\textbf{Man}\,$ automatically satisfies the regularity condition. \begin{lema} Let $\,P\colon\Omega^p\to\Omega^q\,$ be a morphism of functors over the category $\,{\bf Man}$. The restriction of $\,P\,$ to a smooth manifold $\,X\,$ is a regular morphism $\,P\colon\Omega^p_X\to\Omega_X^q\,$ of sheaves over $X$. \end{lema} \noindent \textit{Proof: } Let $\,T\,$ be a smooth manifold (of parameters), and let us consider an open set $\,U\subseteq X\times T\,$ and a smooth family $\,\{\omega_t\in\Omega^p_X(U_t)\}_{t\in T}\,$ of $p$-forms on the open sets $\,U_t=U\cap(X\times t)$. Regularity is a local condition, so we can assume $\,X=\mathbb{R}^n$, $\,T=\mathbb{R}^k\,$ and $\,U=X\times T=\mathbb{R}^n\times\mathbb{R}^k$. Let us write $$\omega_t\,=\, \sum_{i_1<\dots<i_p} f_{i_1\dots i_p}(x,t)\, \text{d}x_{i_1}\wedge\dots\wedge\text{d}x_{i_p}\,\in\,\Omega_X^p(U_t=\mathbb{R}^n)\ .$$ The smoothness of the family $\,\{\omega_t\in\Omega^p_X(U_t)\}_{t\in T}\,$ means that the functions $\,f_{i_1\dots i_p}(x,t)\,$ are smooth on $\,\mathbb{R}^n\times\mathbb{R}^k$, that is, $$\omega:=\, \sum_{i_1<\dots<i_p} f_{i_1\dots i_p}(x,t)\,\text{d}x_{i_1}\wedge\dots\wedge\text{d}x_{i_p}$$ is a differential $p$-form on $\,\mathbb{R}^n\times\mathbb{R}^k$. Then, $\,P(\omega)\in\Omega^q(\mathbb{R}^n\times\mathbb{R}^k)$, let us say $$P(\omega)\,=\, \sum_{j_1<\dots<j_q} g_{j_1\dots j_q}(x,t)\,\text{d}x_{j_1}\wedge\dots\wedge\text{d}x_{j_q}+\cdots\,\text{ terms with }\text{d}t_j\,\cdots$$ for some smooth functions $\,g_{j_1\dots j_q}(x,t)\,$ on $\,\mathbb{R}^n\times\mathbb{R}^k$. Now we can write $$P(\omega_t)\,=\, P(\omega_{|\mathbb{R}^n\times t})$$ (by functoriality) $$=\, P(\omega)_{|\mathbb{R}^n\times t}\,=\, \sum_{j_1<\dots<j_q} g_{j_1\dots j_q}(x,t)\,\text{d}x_{j_1}\wedge\dots\wedge\text{d}x_{j_q}\ ,$$ so that the family $\,\{ P(\omega_t)\}_{t\in T}\,$ is smooth. \hfill $\square$ \medskip In a similar way, it can be proved that, if $\,P\colon\Omega^{p_1}\oplus\cdots\oplus\Omega^{p_k}\to\Omega^q\,$ is a morphism of functors over $\,\textbf{Man}$, then its restriction to a smooth manifold $\,X\,$ is a regular morphism $\, P\colon\Omega^{p_1}_X\oplus\cdots\oplus\Omega^{p_k}_X\to\Omega^q_X\,$ of sheaves over $X$. Consequently, \begin{corolario} Let $\,P\colon\Omega^{p_1}\oplus\cdots\oplus\Omega^{p_k}\to\Omega^q\,$ be a morphism of functors over $\,\mathbf{Man}$. Its restriction to the category $\,\mathbf{Man}_n\,$ is a regular and natural morphism. \end{corolario} \begin{teorema}\label{3.3} Let $\,p_1,\dots,p_k\,$ be positive integers, and let $\,\mathbb{R}\{ u_1,v_1,\dots,u_k,v_k\}\,$ be the anti-commutative algebra of polynomials in the variables $\,u_1,v_1,\dots,u_k,v_k$, of degree $\,\mathrm{deg}\, u_i=p_i$, $\,\mathrm{deg}\, v_i=p_i+1$. Any morphism of functors over the category $\,\mathbf{Man}$ $$ P\colon \Omega^{p_1} \oplus \ldots \oplus \Omega^{p_k} \ \longrightarrow \ \Omega^q\qquad (q\geq 0) $$ can be written as $$P(\omega_1,\dots,\omega_k)\,=\, \mathbf{P}(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k)$$ for a unique homogeneous polynomial $\,\mathbf{P}(u_1,v_1,\dots,u_k,v_k)\in\mathbb{R}\{ u_1,v_1,\dots,u_k,v_k\}$ of degree $\, q$. \end{teorema} \noindent \textit{Proof: } Let $\,P_n\,$ be the restriction of $\,P\,$ to the category $\,\mathbf{Man}_n$. By Theorem \ref{2.3}, for $\,n\geq q$ there exists a unique homogeneous polynomial $\,\mathbf{P}_n\in\mathbb{R}\{ u_1,v_1,\dots,u_k,v_k\}\,$ of degree $\,q\,$ such that $$ P(\omega_1,\dots,\omega_k)\,=\,P_n(\omega_1,\dots,\omega_k)\,=\, \mathbf{P}_n(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k)$$ for any $n$-manifold $\,X\,$ and any $\,(\omega_1,\dots,\omega_k)\in \Omega^{p_1}(X)\oplus\dots\oplus\Omega^{p_k}(X)$. Let us check that $\,\mathbf{P}_n=\mathbf{P}_m\,$ for all $\,n\,$ and $\,m\,$ greater that $\,q$. Suppose $\,m<n$. For any $m$-manifold $\,X_m$, consider the $n$-manifold $\,X_m\times\mathbb{R}^{n-m}$, the natural projection $\,\pi\colon X_m\times\mathbb{R}^{n-m}\to X_m\,$ and the inclusion $\,i\colon X_m\hookrightarrow X_m\times\mathbb{R}^{n-m}$, $\,x\mapsto(x,0)$, so that $\,\pi\circ i=\text{Id}$. For any $\,(\omega_1,\dots,\omega_k)\in \Omega^{p_1}(X_m)\oplus\dots\oplus\Omega^{p_k}(X_m)$, on the one hand $$P_m(\omega_1,\dots,\omega_k)\,=\, \mathbf{P}_m(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k) \ , $$ and, on the other, \begin{align*} P_m(&\omega_1,\dots,\omega_k)\,=\, P(\omega_1,\dots,\omega_k)\,=\,P(i^*\pi^*\omega_1,\dots,i^*\pi^*\omega_k)\\ &\,=\,i^*P(\pi^*\omega_1,\dots,\pi^*\omega_k)\,=\,i^*\mathbf{P}_n(\pi^*\omega_1,\pi^*\mathrm{d}\omega_1,\dots,\pi^*\omega_k,\pi^*\mathrm{d}\omega_k)\\ &\,=\,\mathbf{P}_n(i^*\pi^*\omega_1,i^*\pi^*\mathrm{d}\omega_1,\dots,i^*\pi^*\omega_k,i^*\pi^*\mathrm{d}\omega_k)\,=\,\mathbf{P}_n(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k) \end{align*} so that $\,\mathbf{P}_n=\mathbf{P}_m$. Let us write $\,\mathbf{P}\,$ for this identical polynomials $\,\mathbf{P}_n$. By definition of the $\,\mathbf{P}_n$, it holds $$P(\omega_1,\dots,\omega_k)\,=\, \mathbf{P}(\omega_1,\mathrm{d}\omega_1,\dots,\omega_k,\mathrm{d}\omega_k)$$ for any $\,(\omega_1,\dots,\omega_k)\in \Omega^{p_1}(X)\oplus\dots\oplus\Omega^{p_k}(X)\,$ on any manifold $\,X\,$ of dimension $\,n\geq q$. This equality also holds for manifolds $\,X\,$ of dimension $\,n<q$, because in such a case, both terms of the formula are zero, for they take values on $\,\Omega^q(X)=0$. \hfill $\square$ The case $\,p_1=\dots=p_k=1\,$ of this theorem has been proved by Freed-Hopkins (\cite{FreedHopkins}, Thm. 8.1). \section{Appendix: $\mathfrak{g}$-Valued forms associated to a connection} Let us fix a Lie group $G$ and let $\mathfrak{g}$ be its Lie algebra. Given a principal $G$-bundle $P\to X$ with a principal connection $\alpha$ on it, let us denote $\Theta=\Theta(P,\alpha)$ the curvature form, which is a $\mathfrak{g}$-valued $2$-form on $P$ with the following standard properties: \begin{enumerate} \item[{\it 1.}] {\it It is horizontal}: If $D$ is a vector field tangent to the fibres of $P \to X$, then $$i_D \Theta = 0 \ . $$ \item[{\it 2.}] {\it It is $G$-equivariant}: For any element $g \in G$, it holds: $$ (R_g^* \Theta) (D , \bar{D} ) \, = \, {\rm Ad}\, (g^{-1}) (\Theta (D , \bar{D} )) \ , $$ where $R_g \colon P \to P$, $R_g (p) = p \cdot g$, denotes right translation by $g \in G$. \item[{\it 3.}] {\it It is invariant under isomorphisms}: if $\Phi \colon (P , \alpha) \to (\bar{P} , \bar{\alpha})$ is an isomorphism of principal $G$-bundles over $X$ such that $\Phi^* \bar{\alpha} = \alpha$, then $$ \Phi^* \left[ \Theta (\bar{P} , \bar{\alpha}) \right] \, = \, \Theta (P , \alpha) \ . $$ \item[{\it 4.}] {\it It is stable under arbitrary base changes}: For any smooth map $f \colon Y \to X$, it holds: $$ f^* \left[ \Theta (P, \alpha) \right] \, = \, \Theta (f^* P , f^* \alpha )\ . $$ \end{enumerate} In this section we shall show that the properties {\it 1-4} essentially characterise the curvature $2$-form. More generally, we will determine all the differential forms of arbitrary order satisfying the properties {\it 1-4}, with values on a linear representation of $G$ (Theorem \ref{TeoremaGeneral}).\medskip Let $P\to X$ be a principal fibre bundle with group $G$. We follow the conventions of \cite{Kobayashi1}; in particular, $G$ acts on $P$ on the right. \begin{lemma}\label{LemaUno} Let $P= \mathbb{R}^n \times G \to \mathbb{R}^n$ be the trivial $G$-bundle, let $\alpha$ be a principal connection on it, let $s \colon \mathbb{R}^n \to P$ be a global section, and let $x_0 \in \mathbb{R}^n$ be a point. There exists an isomorphism $\Phi \colon P \to P$ such that: $$ (s^* \Phi^* \alpha)_{x_0} \, = \, 0 \ . $$ \end{lemma} \noindent \textit{Proof: } Let $\bar s\colon \mathbb{R}^n \to P$ be a global section such that $\bar s(x_0)=s(x_0)$ and $\bar s_{*}(T_{x_0}\mathbb{R}^n)=\text{ker}\,\alpha_{s(x_0)}$, i.e., $(\bar s^*\alpha)_{x_0}=0$. Let us write $s(x)=(x,f(x))$ and $\bar s(x)=(x,\bar f(x))$. Let $g(x)\in G$ be such that $\bar f(x)=g(x)\cdot f(x)$; then the isomorphism $\Phi\colon P\to P$, $\Phi(x,g)=(x,g(x)\cdot g)$, satisfies $\bar s=\Phi\circ s$, and hence $\,(s^* \Phi^* \alpha)_{x_0}=(\bar s^*\alpha)_{x_0}=0$. \hfill$\square$ \medskip The adjoint action on the Lie algebra of an element $g \in G$ will be written ${\rm Ad}\, (g) \colon \mathfrak{g} \to \mathfrak{g}$. This map induces a linear map ${\rm Ad}\, (g) \colon S^q\mathfrak{g} \to S^q\mathfrak{g}$, where $S^q\mathfrak{g}$ denotes the $q$-th symmetric tensor power of $\mathfrak{g}$; in other words, the adjoint action of $G$ over $\mathfrak{g}$ induces an action on $S^q\mathfrak{g}$.\medskip The following statement is a reformulation of a result due to Kol\'{a}\v{r} (\cite{Kolar}). The original proof is based on a generalisation of the classical Utiyama theorem on gauge-invariant Lagrangians. \begin{theorem}\label{TeoremaGeneral} Let $G$ be a Lie group, let $V$ be a linear representation of $G$ and let $q$ be a natural number. Associated to any pair $(P \to X , \alpha)$ of a principal $G$-bundle and a principal connection $\alpha$ on it, let $\theta (P , \alpha)$ be a $2q$-form on $P$ with values on $V$, satisfying the properties {\it 1-4 } written above. Then, there exists a $G$-equivariant linear map $T \colon S^q\mathfrak{g} \to V$ such that \begin{equation}\label{Tesis} \theta (P , \alpha) \, = \, T \circ \left[ \Theta (P, \alpha) \, \wedge \stackrel{q}{\ldots} \wedge \, \Theta (P , \alpha) \right] \ , \end{equation} for all pairs $(P \to X , \, \alpha)$. Any differential form of odd order $\theta (P, \alpha)$ satisfying properties {\it 1-4} above is identically zero. \end{theorem} \begin{proof} Let $\theta (P, \alpha)$ be a $2q$-form as in the statement, and let us prove formula (\ref{Tesis}). As any principal bundle is locally isomorphic to the trivial bundle, properties {\it 3} and {\it 4} allow to reduce the reasoning to the case of the trivial bundle $P = X \times G \to X$. Moreover, due to properties {\it 1} and {\it 2}, the form $\theta (P, \alpha)$, for any connection $\alpha$ on the trivial bundle $P$, is determined by its restriction to a fixed global section $s \colon X \to P$. Hence, it is enough to prove that there exists a $G$-equivariant linear map $T \colon S^q\mathfrak{g} \to V$ such that \begin{equation}\label{ToProve} \theta (P, \alpha)_{|s} = T \circ \left[ \Theta (P, \alpha)_{|s}\, \wedge \stackrel{q}{\ldots} \wedge \, \Theta (P , \alpha)_{|s} \right] \end{equation} for any connection $\alpha$ on the trivial bundle $P$. To this end, observe that property {\it 4} implies that the map: \begin{equation}\label{MorfismoDeHaces} \Omega^1 \otimes \mathfrak{g} \longrightarrow \Omega^{2q} \otimes V \quad , \quad \alpha_{|s} \longmapsto \theta (P, \alpha)_{|s} \ , \end{equation} is a well-defined morphism of functors over the category ${\bf Man}$. Let us fix a basis $(v_1, \ldots , v_s)$ of $V$, and write: $$ \alpha = \sum_{j=1}^s \alpha_j \otimes v_j \quad , \quad \Theta ( P , \alpha) = \sum_{j=1}^s \Theta_j (P , \alpha) \otimes v_j \quad , \quad \theta (P , \alpha) = \sum_{j=1}^s \theta_j (P , \alpha) \otimes v_j $$ for some ordinary forms $\alpha_j , \Theta_j (P, \alpha)$, $\theta_j (P, \alpha)$ on $P$. Applying Theorem \ref{3.3} to the components of the morphism (\ref{MorfismoDeHaces}), it follows the existence of a unique collection of scalars $\lambda^I_{ j} , \delta^{KL}_j \in \mathbb{R}$ such that, for any connection $\alpha$ on the trivial bundle $P$, \begin{align*} \theta_j (P , \alpha)_{|s} \ &= \ \sum_{|I|= q} \lambda^{I}_{j} \, \mathrm{d} (\alpha_{{i_1}|s} ) \, \wedge \stackrel{q}{\ldots} \wedge \, \mathrm{d} (\alpha_{{i_q}|s} ) \\ & \quad + \, \sum_{ 2r + s = 2q } \delta^{KL}_j \mathrm{d} (\alpha_{{k_1}|s} )\, \wedge \stackrel{r}{\ldots} \wedge \, \mathrm{d} (\alpha_{{k_r}|s} ) \wedge (\alpha_{l_1})_{|s} \wedge \stackrel{s}{\ldots} \wedge (\alpha_{l_{s}})_{|s} \ . \end{align*} As $\Theta = \mathrm{d} \alpha +\frac{1}{2}\, \alpha \wedge \alpha$, the formula above can be re-written, with other unique\smallskip \noindent coefficients $\lambda^I_{ j} , \mu^{KL}_j \in \mathbb{R}$, \begin{align*} \theta_j (P , \alpha)_{|s} \ &= \ \sum_{|I|=q} \lambda^I_{j} \ \Theta_{i_1} (P, \alpha)_{|s}\, \wedge \stackrel{q}{\ldots} \wedge \, \Theta_{i_q} (P , \alpha)_{|s} \\ & \quad + \, \sum_{ 2r + s = 2q } \mu^{KL}_j \mathrm{d} (\alpha_{{k_1}|s} )\, \wedge \stackrel{r}{\ldots} \wedge \, \mathrm{d} (\alpha_{{k_r}|s} ) \, \wedge \, (\alpha_{l_1})_{|s} \, \wedge \stackrel{s}{\ldots} \wedge \, (\alpha_{l_{s}})_{|s} \ . \end{align*} At any point $x \in X$, there always exists a connection $\bar{\alpha}=\Phi^*\alpha$, isomorphic to $\alpha$, such that $(\bar{\alpha} _{|s})_{x}= 0$ (Lemma \ref{LemaUno}). As $\theta_j(P, \alpha)=(\Phi^{-1})^*[\theta_j (P, \bar\alpha)]$, the coefficients $\mu_j^{KL}$ above have to be zero; that is to say, $$ \theta_j (P , \alpha)_{|s} \, =\, \sum_{|I|=q} \lambda^I_{j} \ \Theta_{i_1} (P, \alpha)_{|s} \, \wedge \stackrel{q}{\ldots} \wedge \, \Theta_{i_q} (P , \alpha)_{|s} \ , $$ which is the required formula (\ref{ToProve}), where $T\colon S^q\mathfrak{g} \to V$ is the linear map defined by the scalars $\lambda^I_{j}$. To check the $G$-equivariance of $T$, let us apply $R_g^*$ in formula (\ref{Tesis}). As $\theta$ and $\Theta$ are $G$-equivariant, it follows: $$ {\rm Ad}\, (g^{-1}) ( T( \Theta ( D_1 , \bar{D}_1) \wedge \ldots \wedge \Theta (D_q , \bar{D}_q) )$$ $$= \, T \left( {\rm Ad}\, (g^{-1}) (\Theta (D_1 , \bar{D}_1) \wedge \ldots \wedge \Theta (D_q , \bar{D}_q) ) \right) $$ for any vector fields $D_1 , \ldots , \bar{D}_q$ on any principal bundle $P \to X$. When varying the pair $(P \to X, \alpha)$, the values $\Theta (D_1 , \bar{D}_1) \wedge \ldots \wedge \Theta (D_q , \bar{D}_q) $ linearly span $S^q \mathfrak{g}$ (see Lemma \ref{LemaDos} below). Hence, the relation above allows to conclude that $T \colon S^q \mathfrak{g} \to V$ is $G$-equivariant. Finally, if $\theta (P, \alpha)$ is a $(2q-1)$-form, a similar reasoning applies. Nevertheless, in this case Theorem \ref{3.3} implies that (\ref{MorfismoDeHaces}) is the zero map, and therefore $\theta (P, \alpha)$ is the zero form. \hfill $\square$ \end{proof} \medskip Observe that the last assertion of this theorem implies that the $2q$-forms $\,\theta=\theta(P,\alpha)\,$ have vanishing covariant differential; that is, they satisfy a kind of Bianchi identity. \begin{lemma}\label{LemaDos} Let $\,v_1\cdots v_q\in S^q\mathfrak{g}\,$ be a product of elements $\, v_1,\dots,v_q\in\mathfrak{g}$. There exists a principal bundle $P \to X$, a connection $\alpha $ on it, a point $p \in P$ and vectors $D_1, \ldots , D_{2q} \in T_pP$ such that: $$ \left( \Theta \, \wedge \stackrel{q}{\cdots} \wedge \, \Theta \right)_p (D_1, \ldots , D_{2q} ) \, = \, v_1\cdots v_q \ , $$ where $\Theta$ denotes the curvature form of $\alpha$. \end{lemma} \noindent \textit{Proof: } Let $P =\mathbb{R}^{2q}\times G \to \mathbb{R}^{2q}$ be the trivial bundle and let $(x_1, \ldots ,x_q,y_1,\ldots , y_q)$ be linear coordinates on $\mathbb{R}^{2q}$. Let $s\colon\mathbb{R}^{2q}\to P$ be the unit section: $s(x)=(x,1)$. Let us consider the unique connection $\alpha$ on $P$ such that $$\alpha_{|s}\,=\, \sum_{i=1}^qx_i\mathrm{d} y_i\otimes v_i\ .$$ The restriction to $s$ of the curvature form $\Theta=\mathrm{d}\alpha+\frac{1}{2}\alpha\wedge\alpha$ is $$\Theta_{|s}\,=\, \sum_i \mathrm{d} x_i\wedge\mathrm{d} y_i\otimes v_i+\frac{1}{2}\sum_{i,j}x_ix_j\mathrm{d} y_i\wedge\mathrm{d} y_j\otimes[v_i,v_j]\ ,$$ hence $$\left( \Theta \, \wedge \stackrel{q}{\cdots} \wedge \, \Omega \right)_{|s}\,=\, q!\,(\mathrm{d} x_1\wedge\mathrm{d} y_1\wedge\cdots\wedge\mathrm{d} x_q\wedge\mathrm{d} y_q)\otimes(v_1\cdots v_q)\ .$$ Therefore, at any point $p$ of the unit section $s$, we have $$\left( \Theta \, \wedge \stackrel{q}{\cdots} \wedge \, \Theta \right)_p (\partial_{x_1},\partial_{y_1},\dots,\partial_{x_q},\partial_{y_q})\,=\, q!\, v_1\cdots v_q\ .$$ \hfill$\square$\medskip The case $q=1$, $V=\mathfrak{g}$ in Theorem \ref{TeoremaGeneral} is essentially a characterisation of the curvature form: \begin{corollary}\label{caso q=1} Associated to any pair $(P \to X , \alpha)$ of a principal $G$-bundle and a principal connection $\alpha$ on it, let $\theta (P , \alpha)$ be a $\mathfrak{g}$-valued $2$-form on $P$ satisfying properties {\it 1-4} above. Then, there exists a $G$-equivariant endomorphism $T \colon \mathfrak{g} \to \mathfrak{g}$ such that\begin{equation} \theta (P , \alpha) \, = \, T \circ \Theta (P, \alpha) \ , \end{equation} for all pairs $\,(P \to X , \, \alpha)$. \end{corollary} If the Lie group $G$ is simple, then the adjoint representation is irreducible. This is the case of the compact groups $U(1)$, $SU(n)$, $SO(n\neq 4)$, $Sp(n)$, $Spin(n\neq 4)$, and the exceptional groups $G_2$, $F_4$, $E_6$, $E_7$ and $E_8$. Then any $G$-equivariant endomorphism $T \colon \mathfrak{g} \to \mathfrak{g}$ is an homothety, and the previous corollary reads: \begin{corollary}\label{Unitario} Let $G$ be a simple Lie group. Associated to any pair $(P \to X , \alpha)$ of a principal $G$-bundle and a principal connection $\alpha$ on it, let $\theta (P , \alpha)$ be a $\mathfrak{g}$-valued $2$-form on $P$ satisfying the properties {\it 1-4} above. Then, there exists $\lambda \in \mathbb{R}$ such that, for all pairs $(P \to X , \, \alpha)$, $$ \theta (P , \alpha) \, = \, \lambda \cdot \Theta (P , \alpha) \ . $$ \end{corollary} \medskip Let us now consider the trivial representation $V=\mathbb{R}$ in Theorem \ref{TeoremaGeneral}. Now $\theta(P,\alpha)$ is an ordinary form on the principal bundle $\pi\colon P\to X$ satisfying the properties {\it 1}-{\it 4}. From properties {\it 1} and {\it 2}, it follows that $\theta(P,\alpha)=\pi^*\widetilde\theta(P,\alpha)$ for an unique ordinary form $\widetilde\theta(P,\alpha)$ on the base manifold $X$. Therefore we may reformulate Theorem \ref{TeoremaGeneral} (in the case $V=\mathbb{R}$) using the language of ordinary forms on base manifolds. Let us precise the statement: Let us fix a Lie group $G$. Let $\mathcal{C}$ denote the functor on ${\bf Man}$ that assigns, to any smooth manifold $X$, the set of isomorphy classes of pairs $(P \to X , \alpha)$, where $P \to X$ is a principal $G$-bundle and $\alpha$ is a principal connection on it. A \textbf{ $q$-form naturally associated to a connection} is a morphism of functors $\widetilde\theta \colon \mathcal{C} \to \Omega^q$. By definition, $\,\widetilde\theta\,$ assigns, to any pair $\,(P \to X, \alpha)\,$, a differential $q$-form $\widetilde\theta (P, \alpha)$ on $\,X\,$ satisfying the following two properties: \begin{itemize} \item[-] If $\,(P \to X, \alpha)\,$ and $\,(P' \to X , \alpha')\,$ are isomorphic, then $\,\widetilde\theta (P, \alpha) = \widetilde\theta(P',\alpha')$. \item[-] For any pair $\,(P \to X , \alpha)\,$ and any smooth map $\,f \colon Y \to X$, it holds: $$ \widetilde\theta ( f^* P , f^* \alpha) = f^* (\widetilde\theta (P, \alpha))\ . $$ \end{itemize} The following statement is again a reformulation of a theorem due to Kol\'{a}\v{r} (\cite{Kolar}). See also (\cite{FreedHopkins}, Th. 7.20) for another formulation. \begin{corollary} The $2q$-forms naturally associated to a connection biunivocally correspond with the $G$-invariant linear maps $T \colon S^q \mathfrak{g} \to \mathbb{R}$. Any form of odd order naturally associated to a connection is identically zero. \end{corollary} More precisely, the natural $2q$-form $\,\widetilde\theta(P,\alpha)\,$ corresponding to a $G$-invariant linear map $\,T \colon S^q \mathfrak{g} \to \mathbb{R}\,$ is the projection on $\,X\,$ of the following $2q$-form over $\,P\,$ $$\theta (P, \alpha) \, := \, T \circ \left[ \Theta (P, \alpha) \, \wedge \stackrel{q}{\ldots} \wedge \, \Theta (P, \alpha) \right] \ , $$ where $\,\Theta\,$ is the $\mathfrak{g}$-valued curvature $2$-form on $\,P$. This result shows that the Chern-Weil forms, defined by the Weil homomorphism (\cite{Kobayashi2}), are the {\it only} natural differential forms one can construct from a $G$-connection. Observe that the second part of the corollary automatically implies that any $2q$-form naturally associated to a connection is closed. \medskip \begin{example} If $G = Gl_n (\mathbb{C})$ is the general complex linear group of rank $n$, the algebra of $Gl_n$-invariant, real polynomials $T(x_{rs},y_{rs})$ on $\mathfrak{gl}_n=M_n (\mathbb{C})$ is generated by the real and imaginary parts of the coefficients $c_1(z_{rs}), \ldots , c_n(z_{rs})$ of the characteristic polynomial of each matrix $(z_{rs}=x_{rs}+iy_{rs})$. The corresponding forms associated to a $G$-connection are essentially the Chern forms. \end{example}

https://arxiv.org/abs/1607.04813

Infinite families of 2-designs and 3-designs from linear codes

The interplay between coding theory and $t$-designs started many years ago. While every $t$-design yields a linear code over every finite field, the largest $t$ for which an infinite family of $t$-designs is derived directly from a linear or nonlinear code is $t=3$. Sporadic $4$-designs and $5$-designs were derived from some linear codes of certain parameters. The major objective of this paper is to construct many infinite families of $2$-designs and $3$-designs from linear codes. The parameters of some known $t$-designs are also derived. In addition, many conjectured infinite families of $2$-designs are also presented.

\section{Introduction} Let ${\mathcal{P}}$ be a set of $v \ge 1$ elements, and let ${\mathcal{B}}$ be a set of $k$-subsets of ${\mathcal{P}}$, where $k$ is a positive integer with $1 \leq k \leq v$. Let $t$ be a positive integer with $t \leq k$. The pair ${\mathbb{D}} = ({\mathcal{P}}, {\mathcal{B}})$ is called a $t$-$(v, k, \lambda)$ {\em design\index{design}}, or simply {\em $t$-design\index{$t$-design}}, if every $t$-subset of ${\mathcal{P}}$ is contained in exactly $\lambda$ elements of ${\mathcal{B}}$. The elements of ${\mathcal{P}}$ are called points, and those of ${\mathcal{B}}$ are referred to as blocks. We usually use $b$ to denote the number of blocks in ${\mathcal{B}}$. A $t$-design is called {\em simple\index{simple}} if ${\mathcal{B}}$ does not contain repeated blocks. In this paper, we consider only simple $t$-designs. A $t$-design is called {\em symmetric\index{symmetric design}} if $v = b$. It is clear that $t$-designs with $k = t$ or $k = v$ always exist. Such $t$-designs are {\em trivial}. In this paper, we consider only $t$-designs with $v > k > t$. A $t$-$(v,k,\lambda)$ design is referred to as a {\em Steiner system\index{Steiner system}} if $t \geq 2$ and $\lambda=1$, and is denoted by $S(t,k, v)$. A necessary condition for the existence of a $t$-$(v, k, \lambda)$ design is that \begin{eqnarray}\label{eqn-tdesignnecessty} \binom{k-i}{t-i} \mbox{ divides } \lambda \binom{v-i}{t-i} \end{eqnarray} for all integer $i$ with $0 \leq i \leq t$. There has been an interplay between codes and $t$-designs for decades. The incidence matrix of any $t$-design spans a linear code over any finite field ${\mathrm{GF}}(q)$. A lot of progress in this direction has been made and documented in the literature (see, for examples, \cite{AK92}, \cite{DingBook}, \cite{Tonchev,Tonchevhb}). On the other hand, both linear and nonlinear codes may hold $t$-designs. Some linear and nonlinear codes were employed to construct $2$-designs and $3$-designs \cite{AK92,Tonchev,Tonchevhb}. Binary and ternary Golay codes of certain parameters hold $4$-designs and $5$-designs \cite{AK92}. However, the largest $t$ for which an infinite family of $t$-designs is derived directly from codes is $t=3$. It looks that not much progress on the construction of $t$-designs from codes has been made so far, while many other constructions of $t$-designs are documented in the literature (\cite{BJL,CMhb,KLhb,RR10}). The main objective of this paper is to construct infinite families of $2$-designs and $3$-designs from linear codes. In addition, we determine the parameters of some known $t$-designs, and present many conjectured infinite families of $2$-designs that are based on projective ternary cyclic codes. \section{The classical construction of $t$-designs from codes and highly nonlinear functions} Let ${\mathcal{C}}$ be a $[v, \kappa, d]$ linear code over ${\mathrm{GF}}(q)$. Let $A_i:=A_i({\mathcal{C}})$, which denotes the number of codewords with Hamming weight $i$ in ${\mathcal{C}}$, where $0 \leq i \leq v$. The sequence $(A_0, A_1, \cdots, A_{v})$ is called the \textit{weight distribution} of ${\mathcal{C}}$, and $\sum_{i=0}^v A_iz^i$ is referred to as the \textit{weight enumerator} of ${\mathcal{C}}$. For each $k$ with $A_k \neq 0$, let ${\mathcal{B}}_k$ denote the set of supports of all codewords of Hamming weight $k$ in ${\mathcal{C}}$, where the coordinates of a codeword are indexed by $(0,1,2, \cdots, v-1)$. Let ${\mathcal{P}}=\{0, 1, 2, \cdots, v-1\}$. The pair $({\mathcal{P}}, {\mathcal{B}}_k)$ may be a $t$-$(v, k, \lambda)$ design for some positive integer $\lambda$. The following theorems, developed by Assumus and Mattson, show that the pair $({\mathcal{P}}, {\mathcal{B}}_k)$ defined by a linear code is a $t$-design under certain conditions. \begin{theorem}\label{thm-AM1}[Assmus-Mattson Theorem \cite{AM74}, \cite[p. 303]{HP03}] Let ${\mathcal{C}}$ be a binary $[v, \kappa, d]$ code. Suppose ${\mathcal{C}}^\perp$ has minimum weight $d^\perp$. Suppose that $A_i=A_i({\mathcal{C}})$ and $A_i^\perp=A_i({\mathcal{C}}^\perp)$, for $0 \leq i \leq v$, are the weight distributions of ${\mathcal{C}}$ and ${\mathcal{C}}^\perp$, respectively. Fix a positive integer $t$ with $t < d$, and let $s$ be the number of $i$ with $A_i^\perp \ne 0$ for $0 < i \leq v-t$. Suppose that $s \leq d -t$. Then \begin{itemize} \item the codewords of weight $i$ in ${\mathcal{C}}$ hold a $t$-design provided that $A_i \ne 0$ and $d \leq i \leq v$, and \item the codewords of weight $i$ in ${\mathcal{C}}^\perp$ hold a $t$-design provided that $A_i^\perp \ne 0$ and $d^\perp \leq i \leq v$. \end{itemize} \end{theorem} The Assmus-Mattson Theorem for nonbinary codes is given as follows [Assmus-Mattson Theorem \cite{AM74}, \cite[p. 303]{HP03}] \begin{theorem}\label{thm-AM2} Let ${\mathcal{C}}$ be a $[v, \kappa, d]$ code over ${\mathrm{GF}}(q)$. Suppose ${\mathcal{C}}^\perp$ has minimum weight $d^\perp$. Let $w$ be the largest integer with $w \leq v$ satisfying $$ w - \left\lfloor \frac{w+q-2}{q-1} \right\rfloor < d. $$ (So $w=v$ when $q=2$.) Define $w^\perp$ analogously using $d^\perp$. Suppose that $A_i=A_i({\mathcal{C}})$ and $A_i^\perp=A_i({\mathcal{C}}^\perp)$, for $0 \leq i \leq v$, are the weight distributions of ${\mathcal{C}}$ and ${\mathcal{C}}^\perp$, respectively. Fix a positive integer $t$ with $t < d$, and let $s$ be the number of $i$ with $A_i^\perp \ne 0$ for $0 < i \leq v-t$. Suppose that $s \leq d -t$. Then \begin{itemize} \item the codewords of weight $i$ in ${\mathcal{C}}$ hold a $t$-design provided that $A_i \ne 0$ and $d \leq i \leq w$, and \item the codewords of weight $i$ in ${\mathcal{C}}^\perp$ hold a $t$-design provided that $A_i^\perp \ne 0$ and $d^\perp \leq i \leq \min\{v-t, w^\perp\}$. \end{itemize} \end{theorem} The Assmus-Mattson Theorems documented above are very powerful tools in constructing $t$-designs from linear codes. We will employ them heavily in this paper. It should be noted that the conditions in Theorems \ref{thm-AM1} and \ref{thm-AM2} are sufficient, but not necessary for obtaining $t$-designs. To construct $t$-designs via Theorems \ref{thm-AM1} and \ref{thm-AM2}, we will need the following lemma in subsequent sections, which is a variant of the MacWilliam Identity \cite[p. 41]{vanLint}. \begin{theorem} \label{thm-MI} Let ${\mathcal{C}}$ be a $[v, \kappa, d]$ code over ${\mathrm{GF}}(q)$ with weight enumerator $A(z)=\sum_{i=0}^v A_iz^i$ and let $A^\perp(z)$ be the weight enumerator of ${\mathcal{C}}^\perp$. Then $$A^\perp(z)=q^{-\kappa}\Big(1+(q-1)z\Big)^vA\Big(\frac {1-z} {1+(q-1)z}\Big).$$ \end{theorem} A function $f$ from ${\mathrm{GF}}(q^m)$ to itself is called {\em planar} or \textit{perfect nonlinear} (PN) if \[\max_{0\ne a\in{\mathrm{GF}}(q^m)}\max_{b\in{\mathrm{GF}}(q^m)}|\{x\in{\mathrm{GF}}(q^m): f(x+a)-f(x)=b\}|=1,\] and {\em almost perfect nonlinear} (APN) if \[\max_{0\ne a\in{\mathrm{GF}}(q^m)}\max_{b\in{\mathrm{GF}}(q^m)}|\{x\in{\mathrm{GF}}(q^m): f(x+a)-f(x)=b\}|=2.\] Later in this paper, we will employ such functions in the constructions of linear codes and thus our constructions of $t$-designs. \section{Infinite families of $3$-designs from the binary RM codes}\label{sec-brmdesigns} It was known that Reed-Muller codes give families of $3$-$(2^m, k, \lambda)$ designs (\cite[Chapter 15]{MS77}, \cite{Tonchevhb}). However, the parameters of $k$ and $\lambda$ may not be specifically given in the literature. The purpose of this section is to determine the parameters of some $3$-designs derived from binary Reed-Muller codes. We use ${\mathrm{RM}}(r, m)$ to denote the binary Reed-Muller code of length $2^m$ and order $r$. Note that ${\mathrm{RM}}(m-r, m)^\perp={\mathrm{RM}}(r-1, m)$, where $2 \leq r < m$. The definition and information about binary Reed-Muller codes can be found in \cite[Section 4.5]{vanLint} and \cite[Chapters 13 and 14]{MS77}. \begin{lemma}\label{lem-RMwt1} The weight distribution of ${\mathrm{RM}}(m-2, m)$ (except $A_i=0$) is given by \begin{eqnarray*} A_{4k}= \frac{1}{2^{m+1}} \left[2\binom{2^m}{4k} + (2^{m+1}-2) \binom{2^{m-1}}{2k}\right] \end{eqnarray*} for $0 \leq k \leq 2^{m-2}$, and by \begin{eqnarray*} A_{4k+2}= \frac{1}{2^{m+1}} \left[2\binom{2^m}{4k+2} - (2^{m+1}-2) \binom{2^{m-1}}{2k+1}\right] \end{eqnarray*} for $0 \leq k \leq 2^{m-2}-1$. \end{lemma} \begin{proof} It is well known that the weight enumerator of ${\mathrm{RM}}(1, m)$ is $$ 1 + (2^{m+1}-2)z^{2^{m-1}} + z^{2^m}. $$ By Theorem \ref{thm-MI}, the weight enumerator of RM$(m-2, m)$, which is the dual of RM$(1, m)$, is given by \begin{eqnarray*} B(z) &=& \frac{1}{2^{m+1}} (1+z)^{2^m}\left[ 1 + (2^{m+1}-2) \left(\frac {1-z} {1+z}\right)^{2^{m-1}} + \left(\frac {1-z} {1+z}\right)^{2^m} \right] \\ &=& \frac{1}{2^{m+1}} \left[ (1+z)^{2^m} + (2^{m+1}-2) (1-z^2)^{2^{m-1}} + (1-z)^{2^m} \right] \\ &=& \frac{1}{2^{m+1}} \left[ 2 \sum_{i=0}^{2^{m-1}} \binom{2^m}{2i} z^{2i} +(2^{m+1}-2) \sum_{i=0}^{2^{m-1}} \binom{2^{m-1}}{i} (-1)^i z^{2i}\right] \\ &=&\frac{1}{2^{m+1}} \sum_{k=0}^{2^{m-2}} \left[ 2\binom{2^m}{4k} + (2^{m+1}-2) \binom{2^{m-1}}{2k} \right] z^{4k} + \\ & & \frac{1}{2^{m+1}}\sum_{k=0}^{2^{m-2}-1} \left[ 2\binom{2^m}{4k+2} - (2^{m+1}-2) \binom{2^{m-1}}{2k+1}\right] z^{4k+2}. \end{eqnarray*} The desired conclusion then follows. \end{proof} The following theorem gives parameters of all the $3$-designs in both ${\mathrm{RM}}(m-2, m)$ and ${\mathrm{RM}}(1, m)$. \begin{theorem}\label{thm-brmdesign1} Let $m \geq 3$. Then ${\mathrm{RM}}(m-2, m)$ has dimension $2^m -m-1$ and minimum distance $4$. For even positive integer $\kappa$ with $4 \leq \kappa \leq 2^m-4$, the supports of the codewords with weight $\kappa$ in ${\mathrm{RM}}(m-2, m)$ hold a $3$-$(2^m, \kappa, \lambda)$ design, where \begin{eqnarray*} \lambda=\left\{ \begin{array}{ll} \frac{\frac{1}{2^{m+1}}\binom{\kappa}{3}\left(2\binom{2^m}{4k} + (2^{m+1}-2) \binom{2^{m-1}}{2k} \right)}{\binom{2^m}{3}} & \mbox{ if } \kappa = 4k, \\ \frac{\frac{1}{2^{m+1}}\binom{\kappa}{3}\left(2\binom{2^m}{4k+2} - (2^{m+1}-2) \binom{2^{m-1}}{2k+1} \right)}{\binom{2^m}{3}} & \mbox{ if } \kappa = 4k+2. \end{array} \right. \end{eqnarray*} The supports of all codewords of weight $2^{m-1}$ in ${\mathrm{RM}}(1, m)$ hold a $3$-$(2^m, 2^{m-1}, 2^{m-2}-1)$ design. \end{theorem} \begin{proof} Note that the weight distribution of ${\mathrm{RM}}(1,m)$ is given by $$ A_0=1, \ A_{2^m}=1, \ A_{2^{m-1}}=2^{m+1}-2, \ \mbox{ and } A_i =0 \mbox{ for all other $i$.} $$ It is known that the minimum distance $d$ of ${\mathrm{RM}}(m-2, m)$ is equal to $4$. Put $t=3$. The number of $i$ with $A_i^\perp \neq 0$ and $1 \leq i \leq 2^m -3$ is $s=1$. Hence, $s=d-t$. Notice that two binary vectors have the same support if and only if they are equal. The desired conclusions then follow from Theorem \ref{thm-AM1} and Lemma \ref{lem-RMwt1}. \end{proof} As a corollary of Theorem \ref{thm-brmdesign1}, we have the following \cite[p. 63]{MS77}, which is well known. \begin{corollary} The minimum weight codewords in ${\mathrm{RM}}(m-2, m)$ form a $3$-$(2^m, 4, 1)$ design, i.e., a Steiner system. \end{corollary} The following theorem is also well known, and tells us that Reed-Muller codes give much more $3$-designs \cite{Tonchevhb}. \begin{theorem}\label{thm-brmdesign2} Let $m \geq 4$ and $2 \leq r < m$. Then ${\mathrm{RM}}(m-r, m)$ has dimension $2^m -\sum_{i=0}^{r-1} \binom{m}{i}$ and minimum distance $2^r$. For every nonzero weight $\kappa$ in ${\mathrm{RM}}(m-r, m)$, the codewords of weight $\kappa$ in RM$(m-r, m)$ hold a $3$-$(2^m, \kappa, \lambda)$ design. \end{theorem} \begin{proof} Since $2 \leq r < m$, by Theorem 24 in \cite[p. 400]{MS77}, the automorphism group of ${\mathrm{RM}}(m-r, m)$ is triply transitive. The desired conclusion then follows from Theorem 8.4.7 in \cite[p. 308]{HP03}. \end{proof} Determining the weight distribution of ${\mathrm{RM}}(m-r, m)$ may be hard for $3 \leq r \leq m-3$ in general. Therefore, it may be difficult to find out the parameters $(\kappa, \lambda)$ of all the $3$-designs. The following problem is open in general. \begin{open} Determine the weight distribution of ${\mathrm{RM}}(m-r, m)$ for $3 \leq r \leq m-3$. \end{open} Some progress on the open problem above was made by Kasami and Tokura \cite{KT} and Kasami, Tokura and Azumi \cite{KTA}. Detailed information on this problem can be found in \cite[Chapter 15]{MS77}. \section{Designs from cyclic Hamming codes}\label{sec-hmdesigns} Let $\alpha$ be a generator of ${\mathrm{GF}}(q^m)^*$. Set $\beta=\alpha^{q-1}$. Let $g(x)$ be the minimal polynomial of $\beta$ over ${\mathrm{GF}}(q)$. Let ${\mathcal{C}}_{(q,m)}$ denote the cyclic code of length $v=(q^m-1)/(q-1)$ over ${\mathrm{GF}}(q)$ with generator polynomial $g(x)$. Then ${\mathcal{C}}_{(q,m)}$ has parameters $[(q^m-1)/(q-1), (q^m-1)/(q-1)-m, d]$, where $d \in \{2,3\}$. When $\gcd(q-1, m)=1$, ${\mathcal{C}}_{(q,m)}$ has minimum weight $3$ and is equivalent to the Hamming code. \begin{lemma}\label{lem-HCwt} The weight distribution of ${\mathcal{C}}_{(q,m)}$ is given by \begin{eqnarray*} A_{k}= \frac{1}{q^m} \sum_{\substack {0 \le i \le (q^{m-1}-1)/(q-1) \\0 \le j \le q^{m-1} \\ i+j=k}}\left[\binom{\frac{q^{m-1}-1}{q-1}}{i} \binom{q^{m-1}}{j}\Big((q-1)^k+(-1)^j(q-1)^i(q^m-1)\Big)\right] \end{eqnarray*} for $0 \leq k \leq (q^m-1)/(q-1)$. \end{lemma} \begin{proof} ${\mathcal{C}}_{(q,m)}^\perp$ is the simplex code, as $\gcd(q-1, (q^{m}-1)/(q-1))=1$. Its weight enumerator is $$ 1+(q^m-1)z^{q^{m-1}}. $$ By Theorem \ref{thm-MI}, the weight enumerator of ${\mathcal{C}}_{(q,m)}$ is given by \begin{eqnarray*} A(z) &=& \frac{1}{q^m} (1+(q-1)z)^{v}\left[ 1 + (q^{m}-1) \left(\frac {1-z} {1+(q-1)z}\right)^{q^{m-1}} \right] \\ &=& \frac{1}{q^m} \left[ (1+(q-1)z)^{v} + (q^m-1) (1-z)^{q^{m-1}}(1+(q-1)z)^{\frac {q^{m-1}-1}{q-1}} \right] \\ &=& \frac{1}{q^m} (1+(q-1)z)^{\frac {q^{m-1}-1}{q-1}} \left[ (1+(q-1)z)^{q^{m-1}} + (q^m-1) (1-z)^{q^{m-1}} \right]. \end{eqnarray*} The desired conclusion then follows. \end{proof} A code of minimum distance $d=2e+1$ is \textit{perfect}, if the spheres of radius $e$ around the codewords cover the whole space. The following theorem introduces a relation between perfect codes and $t$-designs and is due to Assmus and Mattson \cite{AM74}. \begin{theorem}\label{thm-perfectcodedesign} A linear $q$-ary code of length $v$ and minimum distance $d=2e+1$ is perfect if and only if the supports of the codewords of minimum weight form a simple $(e+1)$-$(v, 2e+1, (q-1)^e)$ design. In particular, the minimum weight codewords in a linear or nonlinear perfect code, which contains the zero vector, form a Steiner system $S(e+1, 2e+1, v)$. \end{theorem} It is known that the Hamming code over ${\mathrm{GF}}(q)$ is perfect, and the codewords of weight $3$ hold a $2$-design by Theorem \ref{thm-perfectcodedesign}. The $2$-designs documented in the following theorem may be viewed as an extension of this result. \begin{theorem}\label{thm-HMdesign171} Let $m \geq 3$ and $q = 2$ or $m \geq 2$ and $q >2$, and let $\gcd(q-1, m)=1$. Let ${\mathcal{P}}=\{0,1,2, \cdots, (q^m-q)/(q-1)\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords of Hamming weight $k$ with $A_k \neq 0$ in ${\mathcal{C}}_{(q,m)}$, where $3 \leq k \leq w$ and $w$ is the largest such that $w-\lfloor (w+q-2)/(q-1) \rfloor < 3$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$((q^m-1)/(q-1), k, \lambda)$ design. In particular, the supports of codewords of weight $3$ in ${\mathcal{C}}_{(q,m)}$ form a $2$-$((q^m-1)/(q-1), 3, q-1)$ design. The supports of all codewords of weight $q^{m-1}$ in ${\mathcal{C}}_{(q,m)}^\perp$ form a $2$-$((q^m-1)/(q-1), q^{m-1}, \lambda)$ design, where $$ \lambda=(q-1)q^{m-2}. $$ \end{theorem} \begin{proof} ${\mathcal{C}}_{(q,m)}^\perp$ is the simplex code, as $\gcd(q-1, (q^{m}-1)/(q-1))=1$. Its weight enumerator is $$ 1+(q^m-1)z^{q^{m-1}}. $$ A proof of this weight enumerator is straightforward and can be found in \cite[Theorem 15]{DY13}. Recall now Theorem \ref{thm-AM2} and the definition of $w$ for ${\mathcal{C}}_{(q,m)}$ and $w^\perp$ for ${\mathcal{C}}_{(q,m)}^\perp$. Since ${\mathcal{C}}_{(q,m)}$ has minimum weight $3$. Given that the weight enumerator of ${\mathcal{C}}_{(q,m)}^\perp$ is $1+(q^m-1)z^{q^{m-1}},$ we deduce that $w^\perp=q^{m-1}$. Put $t=2$. It then follows that $s=1=d-t$. The desired conclusion on the $2$-design property then follows from Theorem \ref{thm-AM2} and Lemma \ref{lem-HCwt}. We now prove that the supports of codewords of weight $3$ in ${\mathcal{C}}_{(q,m)}$ form a $2$-$((q^m-1)/(q-1), 3, q-1)$ design. We have already proved that these supports form a $2$-$((q^m-1)/(q-1), 3, \lambda)$ design. To determine the value $\lambda$ for this design, we need to compute the total number $b$ of blocks in this design. To this end, we first compute the total number of codewords of weight $3$ in ${\mathcal{C}}_{(q,m)}$. It follows from Lemma \ref{lem-HCwt} that $$ A_3=\frac{(q^m-1)(q^m-q)}{6}. $$ Since $3$ is the minimum nonzero weight in ${\mathcal{C}}_{(q,m)}$, it is easy to see that two codewords of weight $3$ in ${\mathcal{C}}_{(q,m)}$ have the same support if and only one is a scalar multiple of another. Thus, the total number $b$ of blocks is given by $$ b:=\frac{A_3}{q-1}=\frac{(q^m-1)(q^m-q)}{6(q-1)}. $$ It then follows that $$ \lambda=\frac{b\binom{3}{2}}{\binom{\frac{q^m-1}{q-1}}{2}}=q-1. $$ Let $\alpha$ be a generator of ${\mathrm{GF}}(q^m)^*$, and set $\beta=\alpha^{q-1}$. Then $\beta$ is a $v$-th primitive root of unity, where $v=(q^m-1)/(q-1)$. It is known that $$ {\mathcal{C}}_{(q,m)}^\perp=\{{\mathbf{c}}_u: u \in {\mathrm{GF}}(q^m)\}, $$ where ${\mathbf{c}}_u=(({\mathrm{Tr}}(u), {\mathrm{Tr}}(u\beta), \cdots, {\mathrm{Tr}}(u\beta^{v-1}))$ and ${\mathrm{Tr}}(x)$ is the trace function from ${\mathrm{GF}}(q^m)$ to ${\mathrm{GF}}(q)$. It is then easily seen that ${\mathbf{c}}_{u}$ and ${\mathbf{c}}_{v}$ have the same support if and only if $u=av$ for some $a \in {\mathrm{GF}}(q)^*$. We then deduce that the total number $b^\perp$ of blocks in the design is given by $$ b^\perp = \frac{q^m-1}{q-1}. $$ Consequently, $$ \lambda^\perp = \frac{\frac{q^m-1}{q-1} \binom{q^{m-1}}{2}}{\binom{\frac{q^{m-1}-1}{q-1}}{2}} =(q-1)q^{m-2}. $$ Thus, the supports of all codewords of weight $q^{m-1}$ in ${\mathcal{C}}_{(q,m)}^\perp$ form a $2$-design with parameters $$ \left((q^m-1)/(q-1), \ q^{m-1}, \ (q-1)q^{m-2} \right). $$ \end{proof} Theorem \ref{thm-HMdesign171} tells us that for some $k \geq 3$ with $A_k \neq 0$, the supports of the codewords with weight $k$ in ${\mathcal{C}}_{(q,m)}$ form $2$-$((q^m-1)/(q-1), k, \lambda)$ design. However, it looks complicated to determine the parameter $\lambda$ corresponding to this $k \geq 4$. We draw the reader's attention to the following open problem. \begin{open} Let $q \geq 3$ and $m \geq 2$. For $k \geq 4$ with $A_k \neq 0$, determine the value $\lambda$ in the $2$-$((q^m-1)/(q-1), k, \lambda)$ design, formed by the supports of the codewords with weight $k$ in ${\mathcal{C}}_{(q,m)}$. \end{open} Notice that two binary codewords have the same support if and only if they are equal. When $q=2$, Theorem \ref{thm-HMdesign171} becomes the following. \begin{corollary}\label{cor-HMdesign171} Let $m \geq 3$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-2\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords with Hamming weight $k$ in ${\mathcal{C}}_{(2,m)}$, where $3 \leq k \leq 2^m-3$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(2^m-1, k, \lambda)$ design, where $$ \lambda=\frac{(k-1)kA_k}{(2^m-1)(2^m-2)} $$ and $A_k$ is given in Lemma \ref{lem-HCwt}. The supports of all codewords of weight $2^{m-1}$ in ${\mathcal{C}}_{(2,m)}^\perp$ form a $2$-$(2^m-1, 2^{m-1}, 2^{m-2})$ design. \end{corollary} Corollary \ref{cor-HMdesign171} says that each binary Hamming code ${\mathcal{C}}_{(2,m)}$ and its dual code give a total number $2^m-4$ of $2$-designs with varying block sizes. The following are examples of the $2$-designs held in the binary Hamming code. \begin{example} Let $m \geq 4$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-2\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords with Hamming weight $3$ in ${\mathcal{C}}_{(2,m)}$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(2^m-1, \, 3, \, 1)$ design. \end{example} \begin{proof} By Lemma \ref{lem-HCwt}, we have $$ A_3=\frac{(2^{m-1}-1)(2^m-1)}{3}. $$ The desired value for $\lambda$ then follows from Corollary \ref{cor-HMdesign171}. \end{proof} \begin{example}\label{exam-hmdesign4} Let $m \geq 4$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-2\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords with Hamming weight $4$ in ${\mathcal{C}}_{(2,m)}$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(2^m-1, \, 4, \, 2^{m-1}-2)$ design. \end{example} \begin{proof} By Lemma \ref{lem-HCwt}, we have $$ A_4=\frac{(2^{m-1}-1)(2^{m-1}-2)(2^m-1)}{6}. $$ The desired value for $\lambda$ then follows from Corollary \ref{cor-HMdesign171}. \end{proof} \begin{example}\label{exam-hmdesign5} Let $m \geq 4$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-2\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords with Hamming weight $5$ in ${\mathcal{C}}_{(2,m)}$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(2^m-1, \, 5, \, \lambda)$ design, where $$ \lambda=\frac{2(2^{m-1}-2)(2^{m-1}-4)}{3} $$ \end{example} \begin{proof} By Lemma \ref{lem-HCwt}, we have $$ A_5=\frac{(2^{m-1}-1)(2^{m-1}-2)(2^{m-1}-4)(2^m-1)}{15}. $$ The desired value for $\lambda$ then follows from Corollary \ref{cor-HMdesign171}. \end{proof} \begin{example}\label{exam-hmdesign6} Let $m \geq 4$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-2\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords with Hamming weight $6$ in ${\mathcal{C}}_{(2,m)}$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(2^m-1, \, 6, \, \lambda)$ design, where $$ \lambda=\frac{(2^{m-1}-2)(2^{m-1}-3)(2^{m-1}-4)}{3} $$ \end{example} \begin{proof} By Lemma \ref{lem-HCwt}, we have $$ A_6=\frac{(2^{m-1}-1)(2^{m-1}-2)(2^{m-1}-3)(2^{m-1}-4)(2^m-1)}{45}. $$ The desired value for $\lambda$ then follows from Corollary \ref{cor-HMdesign171}. \end{proof} \begin{example}\label{exam-hmdesign7} Let $m \geq 4$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-2\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords with Hamming weight $7$ in ${\mathcal{C}}_{(2,m)}$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(2^m-1, \, 7, \, \lambda)$ design, where $$ \lambda=\frac{(2^{m-1}-2)(2^{m-1}-3)(4 \times 2^{2(m-1)}-30 \times 2^{m-1} +71)}{30}. $$ \end{example} \begin{proof} By Lemma \ref{lem-HCwt}, we have $$ A_7=\frac{(2^{m-1}-1)(2^{m-1}-2)(2^{m-1}-3)(2^m-1)(4 \times 2^{2(m-1)}-30 \times 2^{m-1} +71)}{630}. $$ The desired value for $\lambda$ then follows from Corollary \ref{cor-HMdesign171}. \end{proof} \section{Designs from a class of binary codes with two zeros and their duals}\label{sec-newdesigns} In this section, we construct many infinite families of $2$-designs and $3$-designs with several classes of binary cyclic codes whose duals have two zeros. These binary codes are defined by almost perfect nonlinear (APN) functions over ${\mathrm{GF}}(2^m)$. \begin{table}[ht] \caption{Weight distribution for odd $m$.}\label{tab-CG1} \centering \begin{tabular}{ll} \hline Weight $w$ & No. of codewords $A_w$ \\ \hline $0$ & $1$ \\ $2^{m-1}-2^{(m-1)/2}$ & $(2^m-1)(2^{(m-1)/2}+1)2^{(m-3)/2}$ \\ $2^{m-1}$ & $(2^m-1)(2^{m-1}+1)$ \\ $2^{m-1}+2^{(m-1)/2}$ & $(2^m-1)(2^{(m-1)/2}-1)2^{(m-3)/2}$ \\ \hline \end{tabular} \end{table} \begin{lemma}\label{lem-TZwt} Let $m \geq 5$ be odd. Let ${\mathcal{C}}_m$ be a binary linear code of length $2^m-1$ such that its dual code ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{tab-CG1}. Then the weight distribution of ${\mathcal{C}}_m$ is given by \begin{eqnarray*} 2^{2m}A_k&=& \sum_{\substack{0 \le i \le 2^{m-1}-2^{(m-1)/2} \\ 0\le j \le 2^{m-1}+2^{(m-1)/2}-1 \\i+j=k}}(-1)^ia\binom{2^{m-1}-2^{(m-1)/2}} {i} \binom{2^{m-1}+2^{(m-1)/2}-1}{j}\\ & & + \binom {2^m-1}{k}+\sum_{\substack{0 \le i \le 2^{m-1} \\ 0\le j \le 2^{m-1}-1 \\i+j=k}}(-1)^ib\binom{2^{m-1}} {i}\binom{2^{m-1}-1} {j} \\ & & + \sum_{\substack{0 \le i \le 2^{m-1}+2^{(m-1)/2} \\ 0\le j \le 2^{m-1}-2^{(m-1)/2}-1 \\i+j=k}}(-1)^ic\binom{2^{m-1}+2^{(m-1)/2}}{i}\binom{2^{m-1}-2^{(m-1)/2}-1}{j} \end{eqnarray*} for $0 \le k \le 2^m-1$, where \begin{eqnarray*} a &=& (2^m-1)(2^{(m-1)/2}+1)2^{(m-3)/2}, \\ b &=& (2^m-1)(2^{m-1}+1), \\ c &=& (2^m-1)(2^{(m-1)/2}-1)2^{(m-3)/2}. \end{eqnarray*} In addition, ${\mathcal{C}}_m$ has parameters $[2^m-1, 2^m-1-2m, 5]$. \end{lemma} \begin{proof} By assumption, the weight enumerator of ${\mathcal{C}}_m^\perp$ is given by $$ A^\perp(z)=1+az^{2^{m-1}-2^{(m-1)/2}}+bz^{2^{m-1}}+cz^{2^{m-1}+2^{(m-1)/2}}. $$ It then follows from Theorem \ref{thm-MI} that the weight enumerator of ${\mathcal{C}}_m$ is given by \begin{eqnarray*} A(z) &=& \frac{1}{2^{2m}} (1+z)^{2^m-1}\left[ 1 + a\left(\frac {1-z} {1+z}\right)^{2^{m-1}-2^{(m-1)/2}} \right] + \\ & & \frac{1}{2^{2m}} (1+z)^{2^m-1}\left[ b\left(\frac {1-z} {1+z}\right)^{2^{m-1}}+c\left(\frac {1-z} {1+z}\right)^{2^{m-1}+2^{(m-1)/2}} \right] \\ &=& \frac{1}{2^{2m}} \Bigg[ (1+z)^{2^m-1} + a(1-z)^{2^{m-1}-2^{(m-1)/2}}(1+z)^{2^{m-1}+2^{(m-1)/2}-1} \\ & & + b(1-z)^{2^{m-1}}(1+z)^{2^{m-1}-1} + c(1-z)^{2^{m-1}+2^{(m-1)/2}}(1+z)^{2^{m-1}-2^{(m-1)/2}-1} \Bigg]. \end{eqnarray*} Obviously, we have \begin{eqnarray*} (1+z)^{2^m-1} &=& \sum_{k=0}^{2^m-1} \binom{2^m-1}{k}z^k. \end{eqnarray*} It is easily seen that \begin{eqnarray*} \lefteqn{(1-z)^{2^{m-1}-2^{(m-1)/2}}(1+z)^{2^{m-1}+2^{(m-1)/2}-1} } \\ &=& \sum_{k=0}^{2^m-1} \left[ \sum_{\substack{0 \le i \le 2^{m-1}-2^{(m-1)/2} \\ 0\le j \le 2^{m-1}+2^{(m-1)/2}-1 \\i+j=k}}(-1)^i \binom{2^{m-1}-2^{(m-1)/2}} {i} \binom{2^{m-1}+2^{(m-1)/2}-1}{j} \right] z^k \end{eqnarray*} and \begin{eqnarray*} \lefteqn{(1-z)^{2^{m-1}+2^{(m-1)/2}}(1+z)^{2^{m-1}-2^{(m-1)/2}-1} } \\ &=& \sum_{k=0}^{2^m-1} \left[ \sum_{\substack{0 \le i \le 2^{m-1}+2^{(m-1)/2} \\ 0\le j \le 2^{m-1}-2^{(m-1)/2}-1 \\i+j=k}}(-1)^i \binom{2^{m-1}+2^{(m-1)/2}}{i}\binom{2^{m-1}-2^{(m-1)/2}-1}{j} \right] z^k. \end{eqnarray*} Similarly, we have \begin{eqnarray*} (1-z)^{2^{m-1}}(1+z)^{2^{m-1}-1}= \sum_{k=0}^{2^m-1} \left[ \sum_{\substack{0 \le i \le 2^{m-1} \\ 0\le j \le 2^{m-1}-1 \\i+j=k}}(-1)^i \binom{2^{m-1}} {i}\binom{2^{m-1}-1} {j} \right] z^k. \end{eqnarray*} Combining these formulas above yields the weight distribution formula for $A_k$. The weight distribution in Table \ref{tab-CG1} tells us that the dimension of ${\mathcal{C}}_m^\perp$ is $2m$. Therefore, the dimension of ${\mathcal{C}}_m$ is equal to $2^m-1-2m$. Finally, we prove that the minimum distance $d$ of ${\mathcal{C}}_m$ equals $5$. After tedious computations with the formula of $A_k$ given in Lemma \ref{lem-TZwt}, one can verify that $A_1=A_2=A_3=A_4=0$ and \begin{eqnarray}\label{eqn-minimumwt5} A_5=\frac{4\times 2^{3m-5} - 22 \times 2^{2m-4} + 26 \times 2^{m-3} -2}{15}. \end{eqnarray} When $m \geq 5$, we have $$ 4\times 2^{3m-5} = 4 \times 2^{m-1} 2^{2m-4} \geq 64 \times 2^{2m-4} > 22 \times 2^{2m-4} $$ and $$ 26 \times 2^{m-3} -2 >0. $$ Consequently, $A_5>0$ for all odd $m$. This proves that $d=5$. \end{proof} \begin{theorem}\label{thm-newdesigns1} Let $m \geq 5$ be odd. Let ${\mathcal{C}}_m$ be a binary linear code of length $2^m-1$ such that its dual code ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{tab-CG1}. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-2\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords of ${\mathcal{C}}_m$ with weight $k$, where $A_k \neq 0$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(2^m-1, k, \lambda)$ design, where \begin{eqnarray*} \lambda=\frac{k(k-1)A_k}{(2^m-1)(2^m-2)}, \end{eqnarray*} where $A_k$ is given in Lemma \ref{lem-TZwt}. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-2\}$, and let ${\mathcal{B}}^\perp$ be the set of the supports of the codewords of ${\mathcal{C}}_m^\perp$ with weight $k$ and $A_k^\perp \neq 0$. Then $({\mathcal{P}}, {\mathcal{B}}^\perp)$ is a $2$-$(2^m-1, k, \lambda)$ design, where \begin{eqnarray*} \lambda=\frac{k(k-1)A_k^\perp}{(2^m-1)(2^m-2)}, \end{eqnarray*} where $A_k^\perp$ is given in Lemma \ref{lem-TZwt}. \end{theorem} \begin{proof} The weight distribution of ${\mathcal{C}}_m$ is given in Lemma \ref{lem-TZwt} and that of ${\mathcal{C}}_m^\perp$ is given in Table \ref{tab-CG1}. By Lemma \ref{lem-TZwt}, the minimum distance $d$ of ${\mathcal{C}}_m$ is equal to $5$. Put $t=2$. The number of $i$ with $A_i^\perp \neq 0$ and $1 \leq i \leq 2^m-1 -t$ is $s=3$. Hence, $s=d-t$. The desired conclusions then follow from Theorem \ref{thm-AM1} and the fact that two binary vectors have the same support if and only if they are equal. \end{proof} \begin{example} Let $m \geq 5$ be odd. Then ${\mathcal{C}}_m^\perp$ gives three $2$-designs with the following parameters: \begin{itemize} \item $(v,\, k, \, \lambda)=\left(2^m-1,\ 2^{m-1}-2^{(m-1)/2}, \ 2^{m-3} (2^{m-1} - 2^{(m-1)/2} -1) \right).$ \item $(v, \, k, \, \lambda)=\left(2^m-1, \ 2^{m-1}+2^{(m-1)/2}, \ 2^{m-3} (2^{m-1} + 2^{(m-1)/2} -1) \right).$ \item $(v, \, k, \, \lambda)=\left(2^m-1, \ 2^{m-1}, \ (2^m-1)(2^{m-1}+1) \right).$ \end{itemize} \end{example} \begin{example} Let $m \geq 5$ be odd. Then the supports of all codewords of weight $5$ in ${\mathcal{C}}_m$ give a $2$-$(2^m-1,\, 5,\, (2^{m-1}-4)/3)$ design. \end{example} \begin{proof} By Lemma \ref{lem-TZwt}, $$ A_5 = \frac{(2^{m-1}-1) (2^{m-1}-4) (2^m-1)}{30} $$ The desired value for $\lambda$ then follows from Theorem \ref{thm-newdesigns1}. \end{proof} \begin{example} Let $m \geq 5$ be odd. Then the supports of all codewords of weight $6$ in ${\mathcal{C}}_m$ give a $2$-$(2^m-1,\, 6,\, \lambda)$ design, where $$ \lambda= \frac{(2^{m-2}-2)(2^{m-1}-3)}{3}. $$ \end{example} \begin{proof} By Lemma \ref{lem-TZwt}, $$ A_6 = \frac{(2^{m-1}-1) (2^{m-1}-4) (2^{m-1}-3) (2^m-1)}{90} $$ The desired value for $\lambda$ then follows from Theorem \ref{thm-newdesigns1}. \end{proof} \begin{example} Let $m \geq 5$ be odd. Then the supports of all codewords of weight $7$ in ${\mathcal{C}}_m$ give a $2$-$(2^m-1,\, 7,\, \lambda)$ design, where $$ \lambda= \frac{2 \times 2^{3(m-1)} - 25 \times 2^{2(m-1)} + 123 \times 2^{m-1} - 190}{30}. $$ \end{example} \begin{proof} By Lemma \ref{lem-TZwt}, $$ A_7 = \frac{(2^{m-1}-1) (2^m-1) (2 \times 2^{3(m-1)} - 25 \times 2^{2(m-1)} + 123 \times 2^{m-1} - 190)}{630}. $$ The desired value for $\lambda$ then follows from Theorem \ref{thm-newdesigns1}. \end{proof} \begin{example} Let $m \geq 5$ be odd. Then the supports of all codewords of weight $8$ in ${\mathcal{C}}_m$ give a $2$-$(2^m-1,\, 8,\, \lambda)$ design, where $$ \lambda= \frac{ (2^{m-2}-2) (2 \times 2^{3(m-1)} - 25 \times 2^{2(m-1)} + 123 \times 2^{m-1} - 190)}{45}. $$ \end{example} \begin{proof} By Lemma \ref{lem-TZwt}, $$ A_8 = \frac{(2^{m-1}-1) (2^{m-1}-4) (2^m-1) (2 \times 2^{3(m-1)} - 25 \times 2^{2(m-1)} + 123 \times 2^{m-1} - 190)}{8 \times 315}. $$ The desired value for $\lambda$ then follows from Theorem \ref{thm-newdesigns1}. \end{proof} \begin{lemma}\label{lem-TZEwt} Let $m \geq 5$ be odd. Let ${\mathcal{C}}_m$ be a linear code of length $2^m-1$ such that its dual code ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{tab-CG1}. Denote by $\overline{{\mathcal{C}}}_m$ the extended code of ${\mathcal{C}}_m$ and let $\overline{{\mathcal{C}}}_m^\perp$ denote the dual of $\overline{{\mathcal{C}}}_m$. Then the weight distribution of $\overline{{\mathcal{C}}}_m$ is given by \begin{eqnarray*} 2^{2m+1}\overline{A}_k&=& (1+(-1)^k) \binom{2^m}{k} + \frac{1+(-1)^k}{2} (-1)^{\lfloor k/2 \rfloor} \binom{2^{m-1}}{\lfloor k/2 \rfloor} v + \\ & & u \sum_{\substack{0 \le i \le 2^{m-1}-2^{(m-1)/2} \\ 0\le j \le 2^{m-1}+2^{(m-1)/2} \\i+j=k}}(-1)^i \binom{2^{m-1}-2^{(m-1)/2}} {i} \binom{2^{m-1}+2^{(m-1)/2}}{j} + \\ & & u \sum_{\substack{0 \le i \le 2^{m-1}+2^{(m-1)/2} \\ 0\le j \le 2^{m-1}-2^{(m-1)/2} \\i+j=k}}(-1)^i \binom{2^{m-1}+2^{(m-1)/2}}{i}\binom{2^{m-1}-2^{(m-1)/2}}{j} \end{eqnarray*} for $0 \le k \le 2^m$, where $$ u=2^{2m-1}-2^{m-1} \mbox{ and } v = 2^{2m}+2^m-2. $$ In addition, $\overline{{\mathcal{C}}}_m$ has parameters $[2^m, 2^m-1-2m, 6]$. The code $\overline{{\mathcal{C}}}_m^\perp$ has weight enumerator \begin{eqnarray}\label{eqn-wtenumerator} \overline{A}^\perp(z) = 1+uz^{2^{m-1}-2^{(m-1)/2}}+vz^{2^{m-1}}+uz^{2^{m-1}+2^{(m-1)/2}}+z^{2^m}, \end{eqnarray} and parameters $[2^m, \ 2m+1, \ 2^{m-1}-2^{(m-1)/2}]$. \end{lemma} \begin{proof} It was proved in Lemma \ref{lem-TZwt} that ${\mathcal{C}}_m$ has parameters $[2^m-1, 2^m-1-2m, 5]$. By definition, the extended code $\overline{{\mathcal{C}}}_m$ has parameters $[2^m, 2^m-1-2m, 6]$. By Table \ref{tab-CG1}, all weights of ${\mathcal{C}}_m^\perp$ are even. Note that ${\mathcal{C}}_m^\perp$ has length $2^m-1$ and dimension $2m$, while $\overline{{\mathcal{C}}}_m^\perp$ has length $2^m$ and dimension $2m+1$. By definition, $\overline{{\mathcal{C}}}_m$ has only even weights. Therefore, the all-one vector must be a codeword in $\overline{{\mathcal{C}}}_m^\perp$. It can be shown that the weights in $\overline{{\mathcal{C}}}_m^\perp$ are the following: $$ 0, \ w_1,\ w_2, \ w_3,\ 2^m-w_1,\ 2^m-w_2, \ 2^m-w_3, \ 2^m, $$ where $w_1, w_2$ and $w_3$ are the three nonzero weights in ${\mathcal{C}}_m^\perp$. Consequently, $\overline{{\mathcal{C}}}_m^\perp$ has the following four weights $$ 2^{m-1}-2^{(m-1)/2}, \ 2^{m-1}, \ 2^{m-1}+2^{(m-1)/2}, \ 2^m. $$ Recall that $\overline{{\mathcal{C}}}_m$ has minimum distance $6$. Employing the first few Pless Moments, one can prove that the weight enumerator of $\overline{{\mathcal{C}}}_m^\perp$ is the one given in (\ref{eqn-wtenumerator}). By Theorem \ref{thm-MI}, the weight enumerator of $\overline{{\mathcal{C}}}_m$ is given by \begin{eqnarray}\label{eqn-j18-1} 2^{2m+1}\overline{A}(z) &=& (1+z)^{2^m}\left[ 1 + u\left(\frac {1-z} {1+z}\right)^{2^{m-1}-2^{(m-1)/2}}+ v\left(\frac {1-z} {1+z}\right)^{2^{m-1}} \right] + \nonumber \\ && (1+z)^{2^m}\left[ u\left(\frac {1-z} {1+z}\right)^{2^{m-1}+2^{(m-1)/2}} + \left(\frac{1-z}{1+z}\right)^{2^m} \right] \nonumber \\ &=& (1+z)^{2^m} + (1-z)^{2^m} + v (1-z^2)^{2^{m-1}} + \nonumber \\ & & u(1-z)^{2^{m-1}-2^{(m-1)/2}}(1+z)^{2^{m-1}+2^{(m-1)/2}} + \nonumber \\ & & u(1-z)^{2^{m-1}+2^{(m-1)/2}}(1+z)^{2^{m-1}-2^{(m-1)/2}} . \end{eqnarray} We now treat the terms in (\ref{eqn-j18-1}) one by one. We first have \begin{eqnarray}\label{eqn-j18-2} (1+z)^{2^m} + (1-z)^{2^m} = \sum_{k=0}^{2^m} \left(1+(-1)^k \right) \binom{2^m}{k}. \end{eqnarray} One can easily see that \begin{eqnarray}\label{eqn-j18-3} (1-z^2)^{2^{m-1}} = \sum_{i=0}^{2^{m-1}} (-1)^i \binom{2^{m-1}}{i} z^{2i} = \sum_{k=0}^{2^{m}} \frac{1+(-1)^k}{2} (-1)^{\lfloor k/2 \rfloor} \binom{2^{m-1}}{\lfloor k/2 \rfloor} z^{k}. \end{eqnarray} Notice that \begin{eqnarray*} (1-z)^{2^{m-1}-2^{(m-1)/2}}=\sum_{i=0}^{2^{m-1}-2^{(m-1)/2}} \binom{2^{m-1}-2^{(m-1)/2}}{i} (-1)^i z^i \end{eqnarray*} and \begin{eqnarray*} (1+z)^{2^{m-1}+2^{(m-1)/2}}=\sum_{i=0}^{2^{m-1}+2^{(m-1)/2}} \binom{2^{m-1}+2^{(m-1)/2}}{i} z^i \end{eqnarray*} We have then \begin{eqnarray}\label{eqn-j18-4} \lefteqn{(1-z)^{2^{m-1}-2^{(m-1)/2}} (1+z)^{2^{m-1}+2^{(m-1)/2}} } \nonumber \\ & & = \sum_{k=0}^{2^m} \left[ \sum_{\substack{0 \le i \le 2^{m-1}-2^{(m-1)/2} \\ 0\le j \le 2^{m-1}+2^{(m-1)/2} \\i+j=k}}(-1)^i \binom{2^{m-1}-2^{(m-1)/2}} {i} \binom{2^{m-1}+2^{(m-1)/2}}{j} \right] z^k. \end{eqnarray} Similarly, we have \begin{eqnarray}\label{eqn-j18-5} \lefteqn{(1-z)^{2^{m-1}+2^{(m-1)/2}} (1+z)^{2^{m-1}-2^{(m-1)/2}} } \nonumber \\ & & = \sum_{k=0}^{2^m} \left[ \sum_{\substack{0 \le i \le 2^{m-1}+2^{(m-1)/2} \\ 0\le j \le 2^{m-1}-2^{(m-1)/2} \\i+j=k}}(-1)^i \binom{2^{m-1}+2^{(m-1)/2}} {i} \binom{2^{m-1}-2^{(m-1)/2}}{j} \right] z^k. \end{eqnarray} Plugging (\ref{eqn-j18-2}), (\ref{eqn-j18-3}), (\ref{eqn-j18-4}), and (\ref{eqn-j18-5}) into (\ref{eqn-j18-1}) proves the desired conclusion. \end{proof} \begin{theorem}\label{thm-newdesigns2} Let $m \geq 5$ be odd. Let ${\mathcal{C}}_m$ be a linear code of length $2^m-1$ such that its dual code ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{tab-CG1}. Denote by $\overline{{\mathcal{C}}}_m$ the extended code of ${\mathcal{C}}_m$ and let $\overline{{\mathcal{C}}}_m^\perp$ denote the dual of $\overline{{\mathcal{C}}}_m$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-1\}$, and let $\overline{{\mathcal{B}}}$ be the set of the supports of the codewords of $\overline{{\mathcal{C}}}_m$ with weight $k$, where $\overline{A}_k \neq 0$. Then $({\mathcal{P}}, \overline{{\mathcal{B}}})$ is a $3$-$(2^m, k, \lambda)$ design, where \begin{eqnarray*} \lambda=\frac{\overline{A}_k\binom{k}{3}}{\binom{2^m}{3}}, \end{eqnarray*} where $\overline{A}_k$ is given in Lemma \ref{lem-TZEwt}. Let ${\mathcal{P}}=\{0,1,2, \cdots, 2^m-1\}$, and let $\overline{{\mathcal{B}}}^\perp$ be the set of the supports of the codewords of $\overline{{\mathcal{C}}}_m^\perp$ with weight $k$ and $\overline{A}_k^\perp \neq 0$. Then $({\mathcal{P}}, \overline{{\mathcal{B}}}^\perp)$ is a $3$-$(2^m, k, \lambda)$ design, where \begin{eqnarray*} \lambda=\frac{\overline{A}_k^\perp\binom{k}{3}}{\binom{2^m}{3}}, \end{eqnarray*} where $\overline{A}_k^\perp$ is given in Lemma \ref{lem-TZEwt}. \end{theorem} \begin{proof} The weight distributions of $\overline{{\mathcal{C}}}_m$ and $\overline{{\mathcal{C}}}_m^\perp$ are described in Lemma \ref{lem-TZEwt}. Notice that the minimum distance $d$ of $\overline{{\mathcal{C}}}_m$ is equal to $6$. Put $t=3$. The number of $i$ with $\overline{A}_i^\perp \neq 0$ and $1 \leq i \leq 2^m -t$ is $s=3$. Hence, $s=d-t$. The desired conclusions then follow from Theorem \ref{thm-AM1} and the fact that two binary vectors have the same support if and only if they are identical. \end{proof} \begin{example} Let $m \geq 5$ be odd. Then $\overline{{\mathcal{C}}}_m^\perp$ gives three $3$-designs with the following parameters: \begin{itemize} \item $(v,\, k, \, \lambda)=\left(2^m,\ 2^{m-1}-2^{(m-1)/2}, \ (2^{m-3}-2^{(m-3)/2}) (2^{m-1}-2^{(m-1)/2}-1) \right).$ \item $(v, \, k, \, \lambda)=\left(2^m, \ 2^{m-1}+2^{(m-1)/2}, \ (2^{m-3}+2^{(m-3)/2}) (2^{m-1}-2^{(m-1)/2}-1) \right).$ \item $(v, \, k, \, \lambda)=\left(2^m, \ 2^{m-1}, \ (2^{m-1}+1)(2^{m-2}-1) \right).$ \end{itemize} \end{example} \begin{example} Let $m \geq 5$ be odd. Then the supports of all codewords of weight $6$ in $\overline{{\mathcal{C}}}_m$ give a $3$-$(2^m,\, 6,\, \lambda)$ design, where $$ \lambda= \frac{2^{m-1}-4}{3}. $$ \end{example} \begin{proof} By Lemma \ref{lem-TZEwt}, $$ \overline{A}_6 = \frac{2^{m-1} (2^{m-1}-1) (2^{m-1}-4) (2^m-1)}{90} $$ The desired value for $\lambda$ then follows from Theorem \ref{thm-newdesigns2}. \end{proof} \begin{example} Let $m \geq 5$ be odd. Then the supports of all codewords of weight $8$ in $\overline{{\mathcal{C}}}_m$ give a $3$-$(2^m,\, 8,\, \lambda)$ design, where $$ \lambda= \frac{2 \times 2^{3(m-1)} - 25 \times 2^{2(m-1)} + 123 \times 2^{m-1} - 190}{30}. $$ \end{example} \begin{proof} By Lemma \ref{lem-TZEwt}, $$ \overline{A}_8 = \frac{2^{m-1}(2^{m-1}-1) (2^m-1) (2 \times 2^{3(m-1)} - 25 \times 2^{2(m-1)} + 123 \times 2^{m-1} - 190)}{8 \times 315}. $$ The desired value for $\lambda$ then follows from Theorem \ref{thm-newdesigns2}. \end{proof} \begin{example} Let $m \geq 5$ be odd. Then the supports of all codewords of weight $10$ in $\overline{{\mathcal{C}}}_m$ give a $3$-$(2^m,\, 10,\, \lambda)$ design, where $$ \lambda= \frac{ (2^{m-1}-4) (2 \times 2^{4(m-1)} - 34 \times 2^{3(m-1)} + 235 \times 2^{2(m-1)} - 931 \times 2^{m-1} + 1358)}{315}. $$ \end{example} \begin{proof} By Lemma \ref{lem-TZEwt}, $$ \overline{A}_{10} = \frac{2^{h} (2^{h}-1) (2^{h}-4) (2^{h+1}-1) (2\times 2^{4h} - 34 \times 2^{3h} + 235 \times 2^{2h} - 931 \times 2^{h} + 1358)}{4\times 14175}, $$ where $h=m-1$. The desired value for $\lambda$ then follows from Theorem \ref{thm-newdesigns2}. \end{proof} To demonstrate the existence of the $2$-designs and $3$-designs presented in Theorems \ref{thm-newdesigns1} and \ref{thm-newdesigns2}, respectively, we describe a list of binary codes that have the weight distribution of Table \ref{tab-CG1} below. Let $\alpha$ be a generator of ${\mathrm{GF}}(2^m)^*$. Let $g_s(x)=\mathbb{M}_1(x)\mathbb{M}_s(x)$, where $\mathbb{M}_i(x)$ is the minimal polynomial of $\alpha^i$ over ${\mathrm{GF}}(2)$. Let ${\mathcal{C}}_m$ denote the cyclic code of length $v=2^m-1$ over ${\mathrm{GF}}(2)$ with generator polynomial $g_s(x)$. It is known that ${\mathcal{C}}_m^\perp$ has dimension $2m$ and the weight distribution of Table \ref{tab-CG1} when $m$ is odd and $s$ takes on the following values \cite{DLLZ}: \begin{enumerate} \item $s=2^h+1$, where $\gcd(h, m)=1$ and $h$ is a positive integer. \item $s=2^{2h}-2^h+1$, where $h$ is a positive integer. \item $s=2^{(m-1)/2}+3$. \item $s=2^{(m-1)/2}+2^{(m-1)/4}-1$, where $m \equiv 1 \pmod{4}$. \item $s=2^{(m-1)/2}+2^{(3m-1)/4}-1$, where $m \equiv 3 \pmod{4}$. \end{enumerate} In all these cases, ${\mathcal{C}}_m$ has parameters $[2^m-1, 2^m-1-2m, 5]$ and is optimal. It is also known that the binary narrow-sense primitive BCH code with designed distance $2^{m-1}-2^{(m-1)/2}$ has also the weight distribution of Table \ref{tab-CG1} \cite{DFZ}. These codes and their extended codes give $2$-designs and $3$-designs when they are plugged into Theorems \ref{thm-newdesigns1} and \ref{thm-newdesigns2}. It is known that ${\mathcal{C}}_m$ has parameters $[2^m-1, 2^m-1-2m, 5]$ if and only if $x^e$ is an APN monomial over ${\mathrm{GF}}(2^m)$. However, even if $x^e$ is APN, the dual code ${\mathcal{C}}_m^\perp$ may have many weights, and thus the code ${\mathcal{C}}_m$ and its dual ${\mathcal{C}}_m^\perp$ may not give $2$-designs. One of such examples is the inverse APN monomial. \section{Infinite families of $2$-designs from a type of ternary linear codes}\label{sec-june28} In this section, we will construct infinite families of $2$-designs with a type of primitive ternary cyclic codes. \begin{table}[ht] \caption{Weight distribution of some ternary linear codes}\label{tab-CG3} \centering \begin{tabular}{|l|l|} \hline Weight $w$ & No. of codewords $A_w$ \\ \hline $0$ & $1$ \\ $2\times 3^{m-1}-3^{(m-1)/2}$ & $(3^m-1)(3^{m-1}+3^{(m-1)/2})$ \\ $2\times 3^{m-1}$ & $(3^m-1)(3^{m-1}+1)$ \\ $2\times 3^{m-1}+3^{(m-1)/2}$ & $(3^m-1)(3^{m-1}-3^{(m-1)/2})$ \\ \hline \end{tabular} \end{table} \begin{table}[ht] \caption{Weight distribution of some ternary linear codes}\label{tab-CG328} \centering \begin{tabular}{|l|l|} \hline Weight $w$ & No. of codewords $A_w$ \\ \hline $0$ & $1$ \\ $2\times 3^{m-1}-3^{(m-1)/2}$ & $3^{2m}-3^m$ \\ $2\times 3^{m-1}$ & $(3^m+3)(3^m-1)$ \\ $2\times 3^{m-1}+3^{(m-1)/2}$ & $3^{2m}-3^m$ \\ \hline $3^m$ & $2$ \\ \hline \end{tabular} \end{table} \begin{lemma}\label{lem-TZEwt28} Let $m \geq 3$ be odd. Assume that ${\mathcal{C}}_m$ is a ternary linear code of length $3^m-1$ such that its dual code ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{tab-CG3}. Denote by $\overline{{\mathcal{C}}}_m$ the extended code of ${\mathcal{C}}_m$ and let $\overline{{\mathcal{C}}}_m^\perp$ denote the dual of $\overline{{\mathcal{C}}}_m$. Then we have the following conclusions. \begin{enumerate} \item The code ${\mathcal{C}}_m$ has parameters $[3^m-1, \, 3^m-1-2m, \, 4]$. \item The code ${\mathcal{C}}_m^\perp$ has parameters $[3^m-1, \, 2m, \, 2\times 3^{m-1}-3^{(m-1)/2}]$. \item The code $\overline{{\mathcal{C}}}_m^\perp$ has parameters $[3^m, \, 2m+1, \, 2\times 3^{m-1}-3^{(m-1)/2}]$, and its weight distribution is given in Table \ref{tab-CG328}. \item The code $\overline{{\mathcal{C}}}_m$ has parameters $[3^m, 3^m-1-2m, 5]$, and its weight distribution is given by \begin{eqnarray*} 3^{2m+1}\overline{A}_k &=& (2^k+(-1)^k 2) \binom{3^m}{k} + \\ & & v \sum_{\substack{0 \le i \le 2 \times 3^{m-1} \\ 0\le j \le 3^{m-1} \\i+j=k}}(-1)^i \binom{2 \times 3^{m-1}} {i} 2^j \binom{3^{m-1}}{j} + \\ & & u \sum_{\substack{0 \le i \le 2\times 3^{m-1}-3^{\frac{m-1}{2}} \\ 0\le j \le 3^{m-1}+3^{\frac{m-1}{2}} \\i+j=k}}(-1)^i \binom{2 \times 3^{m-1}-3^{\frac{m-1}{2}}} {i} 2^j \binom{3^{m-1}+3^{\frac{m-1}{2}}}{j} + \\ & & u \sum_{\substack{0 \le i \le 2 \times 3^{m-1}+3^{\frac{m-1}{2} } \\ 0\le j \le 3^{m-1}-3^{\frac{m-1}{2}} \\i+j=k}}(-1)^i \binom{2 \times 3^{m-1}+3^{\frac{m-1}{2}}}{i}2^j \binom{3^{m-1}-3^{\frac{m-1}{2}}}{j} \end{eqnarray*} for $0 \le k \le 3^m$, where $$ u=3^{2m}-3^{m} \mbox{ and } v = (3^m+3)(3^m-1). $$ \end{enumerate} \end{lemma} \begin{proof} The proof is similar to that of Lemma \ref{lem-TZEwt} and is omitted here. \end{proof} \begin{theorem}\label{thm-newdesigns228} Let $m \geq 3$ be odd. Let ${\mathcal{C}}_m$ be a linear code of length $3^m-1$ such that its dual code ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{tab-CG3}. Denote by $\overline{{\mathcal{C}}}_m$ the extended code of ${\mathcal{C}}_m$ and let $\overline{{\mathcal{C}}}_m^\perp$ denote the dual of $\overline{{\mathcal{C}}}_m$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 3^m-1\}$, and let $\overline{{\mathcal{B}}}$ be the set of the supports of the codewords of $\overline{{\mathcal{C}}}_m$ with weight $k$, where $5 \leq k \leq 10$ and $\overline{A}_k \neq 0$. Then $({\mathcal{P}}, \overline{{\mathcal{B}}})$ is a $2$-$(3^m,\, k,\, \lambda)$ design for some $\lambda$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 3^m-1\}$, and let $\overline{{\mathcal{B}}}^\perp$ be the set of the supports of the codewords of $\overline{{\mathcal{C}}}_m^\perp$ with weight $k$ and $\overline{A}_k^\perp \neq 0$. Then $({\mathcal{P}}, \overline{{\mathcal{B}}}^\perp)$ is a $2$-$(3^m, \, k, \, \lambda)$ design for some $\lambda$. \end{theorem} \begin{proof} The weight distributions of $\overline{{\mathcal{C}}}_m$ and $\overline{{\mathcal{C}}}_m^\perp$ are described in Lemma \ref{lem-TZEwt28}. Notice that the minimum distance $d$ of $\overline{{\mathcal{C}}}_m$ is equal to $5$. Put $t=2$. The number of $i$ with $\overline{A}_i^\perp \neq 0$ and $1 \leq i \leq 3^m -t$ is $s=3$. Hence, $s=d-t$. The desired conclusions then follow from Theorem \ref{thm-AM2}. \end{proof} \begin{corollary} Let $m \geq 3$ be odd. Let ${\mathcal{C}}_m$ be a ternary linear code of length $3^m-1$ such that its dual code ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{tab-CG3}. Denote by $\overline{{\mathcal{C}}}_m$ the extended code of ${\mathcal{C}}_m$ and let $\overline{{\mathcal{C}}}_m^\perp$ denote the dual of $\overline{{\mathcal{C}}}_m$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 3^m-1\}$, and let $\overline{{\mathcal{B}}}^\perp$ be the set of the supports of the codewords of $\overline{{\mathcal{C}}}_m^\perp$ with weight $2\times 3^{m-1}-3^{(m-1)/2}$. Then $({\mathcal{P}}, \overline{{\mathcal{B}}}^\perp)$ is a $2$-$(3^m, \, 2\times 3^{m-1}-3^{(m-1)/2}, \, \lambda)$, where $$ \lambda=\frac{(2\times 3^{m-1}- 3^{(m-1)/2})(2\times 3^{m-1}- 3^{(m-1)/2} -1)}{2}. $$ \end{corollary} \begin{proof} It follows from Theorem \ref{thm-newdesigns228} that $({\mathcal{P}}, \overline{{\mathcal{B}}}^\perp)$ is a $2$-design. We now determine the value of $\lambda$. Note that $\overline{{\mathcal{C}}}_m^\perp$ has minimum weight $2\times 3^{m-1}-3^{(m-1)/2}$. Any two codewords of minimum weight $2\times 3^{m-1}-3^{(m-1)/2}$ have the same support if and only if one is a scalar multiple of the other. Consequently, $$ \left|\overline{{\mathcal{B}}}^\perp \right|=\frac{3^{2m}-3^m}{2}. $$ It then follows that $$ \lambda=\frac{3^{2m}-3^m}{2}\frac{\binom{2\times 3^{m-1}-3^{(m-1)/2}}{2}}{\binom{3^m}{2}} = \frac{(2\times 3^{m-1}- 3^{(m-1)/2})(2\times 3^{m-1}- 3^{(m-1)/2} -1)}{2}. $$ \end{proof} \begin{corollary} Let $m \geq 3$ be odd. Let ${\mathcal{C}}_m$ be a ternary linear code of length $3^m-1$ such that its dual code ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{tab-CG3}. Denote by $\overline{{\mathcal{C}}}_m$ the extended code of ${\mathcal{C}}_m$ and let $\overline{{\mathcal{C}}}_m^\perp$ denote the dual of $\overline{{\mathcal{C}}}_m$. Let ${\mathcal{P}}=\{0,1,2, \cdots, 3^m-1\}$, and let $\overline{{\mathcal{B}}}$ be the set of the supports of the codewords of $\overline{{\mathcal{C}}}_m$ with weight $5$. Then $({\mathcal{P}}, \overline{{\mathcal{B}}})$ is a $2$-$(3^m,\, 5,\, \lambda)$ design, where $$ \lambda=\frac{5(3^{m-1}-1)}{2}. $$ \end{corollary} \begin{proof} It follows from Theorem \ref{thm-newdesigns228} that $({\mathcal{P}}, \overline{{\mathcal{B}}})$ is a $2$-design. We now determine the value of $\lambda$. Using the weight distribution formula in Lemma \ref{lem-TZEwt28}, we obtain that $$ \overline{A}_5=\frac{3^{3m-1}-4 \times 3^{2m-1}+3^m}{4}. $$ Recall that $\overline{{\mathcal{C}}}_m$ has minimum weight $5$. Any two codewords of minimum weight $5$ have the same support if and only if one is a scalar multiple of the other. Consequently, $$ \left|\overline{{\mathcal{B}}}^\perp \right|=\frac{\overline{A}_5}{2}. $$ It then follows that $$ \lambda= \frac{\overline{A}_5}{2} \frac{\binom{5}{2}}{\binom{3^m}{2}}=\frac{5(3^{m-1}-1)}{2}. $$ \end{proof} Theorem \ref{thm-newdesigns228} gives more $2$-designs. However, determining the corresponding value $\lambda$ may be hard, as the number of blocks in the design may be difficult to derive from $\overline{A}_k$ or $\overline{A}_k^\perp$. \begin{open} Determine the value of $\lambda$ of the $2$-$(3^m,\, k,\, \lambda)$ design for $6 \leq k \leq 10$, which are described in Theorem \ref{thm-newdesigns228}. \end{open} \begin{open} Determine the values of $\lambda$ of the $2$-$(3^m,\, 3^{m-1},\, \lambda)$ design and the $2$-$(3^m,\, 2\times 3^{m-1}-3^{(m-1)/2},\, \lambda)$ design, which are described in Theorem \ref{thm-newdesigns228}. \end{open} To demonstrate the existence of the $2$-designs presented in Theorem \ref{thm-newdesigns228}, we present a list of ternary cyclic codes that have the weight distribution of Table \ref{tab-CG3} below. Put $n=3^m-1$. Let $\alpha$ be a generator of ${\mathrm{GF}}(3^m)^*$. Let $g_s(x)=\mathbb{M}_{n-1}(x)\mathbb{M}_{n-s}(x)$, where $\mathbb{M}_i(x)$ is the minimal polynomial of $\alpha^i$ over ${\mathrm{GF}}(3)$. Let ${\mathcal{C}}_m$ denote the cyclic code of length $n=3^m-1$ over ${\mathrm{GF}}(3)$ with generator polynomial $g_s(x)$. It is known that ${\mathcal{C}}_m^\perp$ has dimension $2m$ and the weight distribution of Table \ref{tab-CG3} when $m$ is odd and $s$ takes on the following values \cite{CDY,YCD}: \begin{enumerate} \item $s=3^h+1$, $h \geq 0$ is an integer. \item $s=(3^h + 1)/2$, where $h$ is a positive integer and $\gcd(m, h)=1$. \end{enumerate} In these two cases, $x^s$ is a planar function on ${\mathrm{GF}}(3^m)$. Hence, these ternary codes are extremal in the sense that they are defined by planar functions whose differentiality is extremal. More classes of ternary codes such that their duals have the weight distribution of Table \ref{tab-CG3} are documented in \cite{DLLZ}. They give also $2$-designs via Theorem \ref{thm-newdesigns228}. There are also ternary cyclic codes with three weights but different weight distributions in \cite{DLLZ}. They may also hold $2$-designs. \section{Conjectured infinite families of $2$-designs from projective cyclic codes}\label{sec-conjectureddesigns} Throughout this section, let $m \geq 3$ be an odd integer, and let $v=(3^m-1)/2$. The objective of this section is to present a number of conjectured infinite families of $2$-designs derived from linear projective ternary cyclic codes. \begin{table} \begin{center} \caption{The weight distribution for odd $m \ge3$}\label{Tab-GG2} \begin{tabular}{|c|c|} \hline Weight & Frequency \\ \hline $0$ & $1$ \\ \hline $3^{m-1}-3^{(m-1)/2}$ & $\frac{(3^{m-1}+3^{(m-1)/2})(3^m-1)}{2}$ \\ \hline $3^{m-1}$ & $(3^m-3^{m-1}+1)(3^m-1)$ \\ \hline $3^{m-1}+3^{(m-1)/2}$ & $\frac{(3^{m-1}-3^{(m-1)/2})(3^m-1)}{2}$ \\ \hline \end{tabular} \label{table4} \end{center} \end{table} \begin{lemma}\label{lem-june21} Let ${\mathcal{C}}_m$ be a linear code of length $v$ over ${\mathrm{GF}}(3)$ such that its dual ${\mathcal{C}}_m^\perp$ has the weight distribution in Table \ref{Tab-GG2}. Then the weight distribution of ${\mathcal{C}}_m$ is given by \begin{eqnarray*} 3^{2m}A_k &=& \sum_{\substack{0 \le i \le 3^{m-1}-3^{(m-1)/2} \\ 0\le j \le \frac {3^{m-1}+2\cdot 3^{(m-1)/2}-1} {2} \\i+j=k}}(-1)^i2^ja\binom{3^{m-1}-3^{(m-1)/2}} {i} \binom{\frac {3^{m-1}+2 \cdot 3^{(m-1)/2}-1}{2}}{j}\\ & & + \binom {\frac {3^m-1} {2}}{k}2^k+ \sum_{\substack{0 \le i \le 3^{m-1} \\ 0\le j \le \frac {3^{m-1}-1} {2} \\i+j=k}}(-1)^i2^jb\binom{3^{m-1}} {i}\binom{\frac {3^{m-1}-1} {2}} {j} \\ & & + \sum_{\substack{0 \le i \le 3^{m-1}+3^{(m-1)/2} \\ 0\le j \le \frac {3^{m-1}-2 \cdot 3^{(m-1)/2}-1}{2} \\i+j=k}}(-1)^i2^jc\binom{3^{m-1}+3^{(m-1)/2}}{i}\binom{\frac {3^{m-1}-2 \cdot 3^{(m-1)/2}-1}{2}}{j} \end{eqnarray*} for $0 \le k \le \frac {3^m-1} {2}$, where \begin{eqnarray*} a &=& \frac{(3^{m-1}+3^{(m-1)/2})(3^m-1)}{2}, \\ b &=& (3^m-3^{m-1}+1)(3^m-1), \\ c &=& \frac{(3^{m-1}-3^{(m-1)/2})(3^m-1)}{2}. \end{eqnarray*} In addition, ${\mathcal{C}}_m$ has parameters $[(3^m-1)/2, (3^m-1)/2-2m, 4]$. \end{lemma} \begin{proof} Note that the weight enumerator of ${\mathcal{C}}_m^\perp$ is $$1+az^{3^{m-1}-3^{(m-1)/2}}+bz^{3^{m-1}}+cz^{3^{m-1}+3^{(m-1)/2}}.$$ The proof of this theorem is similar to that of Lemma \ref{lem-TZwt} and is omitted. \end{proof} Below we present two examples of ternary linear codes ${\mathcal{C}}_m$ such that their duals ${\mathcal{C}}_m^\perp$ have the weight distribution of Table \ref{Tab-GG2}. \begin{example}\label{exam-j271} Let $m \geq 3$ be odd. Let $\alpha$ be a generator of ${\mathrm{GF}}(3^m)^*$. Put $\beta=\alpha^2$. Let $\mathbb{M}_i(x)$ denote the minimal polynomial of $\beta^i$ over ${\mathrm{GF}}(3)$. Define $$ \delta=3^{m-1}-1-\frac{3^{(m+1)/2}-1}{2} $$ and $$ h(x)=(x-1){\mathrm{lcm}}(\mathbb{M}_1(x), \, \mathbb{M}_2(x), \, \cdots, \, \mathbb{M}_{\delta-1}(x)), $$ where ${\mathrm{lcm}}$ denotes the least common multiple of the polynomials. Let ${\mathcal{C}}_m$ denote the cyclic code of length $v=(3^m-1)/2$ over ${\mathrm{GF}}(3)$ with generator polynomial $g(x):=(x^v-1)/h(x)$. Then ${\mathcal{C}}_m$ has parameters $[(3^m-1)/2, (3^m-1)-2m, 4]$ and ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{Tab-GG2}. \end{example} \begin{proof} A proof of the desired conclusions was given in \cite{LDXG}. \end{proof} \begin{example}\label{exam-j272} Let $m \geq 3$ be odd. Let $\alpha$ be a generator of ${\mathrm{GF}}(3^m)^*$. Let $\beta=\alpha^2$. Let $g(x)=\mathbb{M}_{n-1}(x)\mathbb{M}_{n-2}(x)$, where $\mathbb{M}_i(x)$ is the minimal polynomial of $\beta^i$ over ${\mathrm{GF}}(3)$. Let ${\mathcal{C}}_m$ denote the cyclic code of length $v=(3^m-1)/2$ over ${\mathrm{GF}}(3)$ with generator polynomial $g(x)$. Then ${\mathcal{C}}_m$ has parameters $[(3^m-1)/2, (3^m-1)-2m, 4]$ and ${\mathcal{C}}_m^\perp$ has the weight distribution of Table \ref{Tab-GG2}. \end{example} \begin{proof} The desired conclusions can be proved similarly as Theorem 19 in \cite{LDXG}. \end{proof} \begin{conj}\label{conj-j261} Let ${\mathcal{P}}=\{0,1,2, \cdots, v-1\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords of ${\mathcal{C}}_m$ with Hamming weight $k$, where $A_k \neq 0$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(v, k, \lambda)$ design for all odd $m \ge 3$. \end{conj} \begin{conj}\label{conj-j262} Let ${\mathcal{P}}=\{0,1,2, \cdots, v-1\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords of ${\mathcal{C}}_m$ with Hamming weight $4$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a Steiner system $S(2, 4, (3^m-1)/2)$ for all odd $m \ge 3$. \end{conj} There are a survey on Steiner systems $S(2,4,v)$ \cite{RR10} and a book chapter on Steiner systems \cite{CMhb}. It is known that a Steiner system $S(2,4,v)$ exists if and only if $v \equiv 1 \mbox{ or } 4 \pmod{12}$ \cite{Hanani}. If Conjecture \ref{conj-j261} is true, so is Conjecture \ref{conj-j262}. In this case, a coding theory construction of a Steiner system $S(2, 4, (3^m-1)/2)$ for all odd $m \ge 3$ is obtained. \begin{conj}\label{conj-j263} Let ${\mathcal{P}}=\{0,1,2, \cdots, v-1\}$, and let ${\mathcal{B}}$ be the set of the supports of the codewords of ${\mathcal{C}}_m^\perp$ with Hamming weight $k$, where $A_k^\perp \neq 0$. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(v, k, \lambda)$ design for all odd $m \ge 3$. \end{conj} Even if some or all of the three conjectures are not true for ternary codes with the weight distribution of Table \ref{Tab-GG2}, these conjectures might still be valid for the two classes of ternary cyclic codes descried in Examples \ref{exam-j271} and \ref{exam-j272}. Note that Theorem \ref{thm-AM2} does not apply to the three conjectures above. We need to develop different methods for settling these conjectures. \section{Summary and concluding remarks} In the last section of this paper, we mention some applications of $t$-designs and summarize the main contributions of this paper. \subsection{Some applications of $2$-designs} Let ${\mathcal{P}}$ be an Abelian group of order $v$ under a binary operation denoted by $+$. Let ${\mathcal{B}}=\{B_1, B_2, \cdots, B_b\}$, where all $B_i$ are $k$-subsets of ${\mathcal{P}}$ and $k$ is a positive integer. We define $\Delta(B_i)$ to be the multiset $\{ x-y: x \in B_i,\ y \in B_i\}$. If every nonzero element of ${\mathcal{P}}$ appears exactly $\delta$ times in the multiset $\bigcup_{i=1}^b \Delta(B_i)$, we call ${\mathcal{B}}$ a $(v, k, \delta)$ difference family in $({\mathcal{P}}, +)$. The following theorems are straightforward and should be well known. \begin{theorem}\label{thm-june261} Let ${\mathcal{P}}$ be an Abelian group of order $v$ under a binary operation denoted by $+$. Let ${\mathcal{B}}=\{B_1, B_2, \cdots, B_b\}$, where all $B_i$ are $k$-subsets of ${\mathcal{P}}$ and $k$ is a positive integer. Then $({\mathcal{P}}, {\mathcal{B}})$ is a $2$-$(v, k, \lambda)$ design if and only if ${\mathcal{B}}$ is a $(v, k, \lambda v)$ difference family in $({\mathcal{P}}, +)$. \end{theorem} \begin{theorem}\label{thm-june262} Let $({\mathcal{P}}, {\mathcal{B}})$ be a $t$-$(v, k, \lambda)$ design, where ${\mathcal{P}}$ is an Abelian group. If $t \geq 2$, then ${\mathcal{B}}$ is a $(v, k, \delta)$ difference family in ${\mathcal{P}}$, where $$ \delta=\frac{v \lambda \binom{v-2}{t-2}}{\binom{k-2}{t-2}}. $$ \end{theorem} Difference families have applications in the design and analysis of optical orthogonal codes, frequency hopping sequences, and other engineering areas. By Theorems \ref{thm-june261} and \ref{thm-june262}, $t$-designs with $t \geq 2$ have also applications in these areas. In addition, $2$-designs give naturally linear codes \cite{AK92,DingBook}. These show the importance of $2$-designs in applications. \subsection{Summary} It is well known that binary Reed-Muller codes hold $3$-designs. Hence, the only contribution of Section \ref{sec-brmdesigns} is the determination of the specific parameters of the $3$-designs held in ${\mathrm{RM}}(m-2, m)$ and its dual code, which are documented in Theorem \ref{thm-brmdesign1}. It has also been known for a long time that the codewords of weight $3$ in the Hamming code hold a $2$-$((q^m-1)/(q-1), 3, q-1)$ design. The contribution of Section \ref{sec-hmdesigns} is Theorem \ref{thm-HMdesign171}, which may be viewed as an extension of the known $2$-$((q^m-1)/(q-1), 3, q-1)$ design held in the Hamming code, and also the parameters of the infinite families of $2$-designs derived from the binary Hamming codes, which are documented in Examples \ref{exam-hmdesign4}, \ref{exam-hmdesign5}, \ref{exam-hmdesign6}, and \ref{exam-hmdesign7}. A major contribution of this paper is presented in Section \ref{sec-newdesigns}, where Theorems \ref{thm-newdesigns1} and \ref{thm-newdesigns2} document many infinite families of $2$-design and $3$-designs. The parameters of these $2$-designs and $3$-designs are given specifically. These designs are derived from binary cyclic codes that are defined by special almost perfect nonlinear functions. Another major contribution of this paper is documented in Section \ref{sec-june28}, where Theorem \ref{thm-newdesigns228} and its two corollaries describe several infinite families of $2$-designs. These $2$-designs are related to planar functions. It is noticed that the total number of $3$-designs presented in this paper (see Theorems \ref{thm-brmdesign1} and \ref{thm-newdesigns2}) are exponential. All of them are derived from linear codes. After comparing the list of infinite families of $3$-designs in \cite{KLhb} with the $3$-designs presented in this paper, one may conclude that many, if not most, of the known infinite families of $3$-designs are from coding theory. Section \ref{sec-conjectureddesigns} presents many conjectured infinite families of $2$-designs. The reader is cordially invited to attack these conjectures and solve other open problems presented in this paper.

https://arxiv.org/abs/2006.04355

$K_{r+1}$-saturated graphs with small spectral radius

For a graph $H$, a graph $G$ is $H$-saturated if $G$ does not contain $H$ as a subgraph but for any $e \in E(\overline{G})$, $G+e$ contains $H$. In this note, we prove a sharp lower bound for the number of paths and walks on length $2$ in $n$-vertex $K_{r+1}$-saturated graphs. We then use this bound to give a lower bound on the spectral radii of such graphs which is asymptotically tight for each fixed $r$ and $n\to\infty$.

\section{Introduction} \subsection{Notation and preliminaries} In this note we deal with finite undirected graphs with no loops or multiple edges. For a graph $H$, a graph $G$ is $H$-{\em saturated} if $H$ is not a subgraph of $G$ but after adding to $G$ any edge results in a graph containing $H$. For a positive integer $n$ and a graph $H$, the {\em extremal number} $ex(n, H)$ is the maximum number of edges in an $n$-vertex graph not containing $H$. Clearly, an extremal $n$-vertex graph $G$ not containing $H$ with $|E(G)|=ex(n, H)$ is $H$-saturated. Thus, one can also say that $ex(n, H)$ is the maximum number of edges in an $n$-vertex $H$-saturated graph. On the other hand, the {\em saturation number of $H$}, $sat(n, H)$, is the least number of edges in an $H$-saturated graph with $n$ vertices. Initiating the study of extremal graph theory, Tur{\'a}n~\cite{T1} determined the extremal number $ex(n, K_{r+1})$. He also proved that there is the unique extremal graph, $T_{n,r}$, the $n$-vertex complete $r$-partite graph whose partite sets differ in size at most 1. The first result on saturation numbers is due to Erd{\H{o}}s, Hajnal and Moon~\cite{EHM}: \bigskip {\bf Theorem A~\cite{EHM}.} {\em If $2\leq r<n$, then $sat(n, K_{r+1})=(r-1)(n-r+1)+\binom{r-1}{2}$. The only $n$-vertex $K_{r+1}$-saturated graph with $sat(n, K_{r+1})$ edges is the graph $S_{n,r}$ obtained from a copy of $K_{r-1}$ with vertex set $S$ by adding $n - r + 1$ vertices, each of which has neighborhood $S$. } \medskip Graph $S_{n,r}$ has clique number $r$ and no $r$-connected subgraphs; in particular, $S_{n,2}$ is a star. For an excellent survey on saturation numbers, we refer the reader to Faudree, Faudree, and Schmitt~\cite{FFS}. Recently, there was a series of publications on eigenvalues of $H$-free graphs. For a graph $G$, let $A(G)$ be its adjacency matrix, and we index the eigenvalues of $A(G)$ in nonincreasing order, $\lambda_1(G)\ge \cdots \ge \lambda_n(G)$. The value $\lambda_1(G)$ is also called the {\em spectral radius} of $G$, and denoted by $\rho(G)$. Studying properties of quasi-random graphs, Chung, Graham, and Wilson~\cite{CGW} proved a theorem implying that, if $n$ is sufficiently large, $0 < c < \frac 12$ and $G$ is an $n$-vertex $K_r$-free graph with $\lceil cn^2 \rceil$ edges, then either $\lambda_n(G) < -c'n$ or $\lambda_2(G)>c'n$, where $c' = c'(r, c)$ is a positive constant. However, the methods in~\cite{CGW} fail to indicate which of the two inequalities actually holds. Bollob{\'a}s and Nikiforov~\cite{BN} observed that if $G$ is a dense $K_r$-free graph, then $\lambda_n(G) < - cn $ for some $c > 0$ independent of $n$. Nikiforov~\cite{N1} gave a more precise statement that if $G$ is a $K_{r+1}$-free graph with $n$ vertices and $m$ edges, then $\lambda_n(G) < -\frac{2^{r+1}m^r}{rn^{2r-1}}.$ Nikiforov~\cite{N2} also proved that if $G$ ia a $K_{r+1}$-free graph with $n$ vertices, then $\rho(G) \le \rho(T_{n,r}).$ Since each $K_{r+1}$-saturated graph is $K_{r+1}$-free, his theorem implies the following. \bigskip {\bf Theorem B~\cite{N1}.} {\em If $G$ is a $K_{r+1}$-saturated graph with $n$ vertices, then $$\rho(G) \le \rho(T_{n,r}).$$ } In this note, we give a new lower bound for the spectral radius of an $n$-vertex $K_{r+1}$-saturated graph. This bound is asymptotically tight when $r$ is fixed or grows as $o(n)$. For this, we give a tight lower bound on the sum of the squares of the vertex degrees in an $n$-vertex $K_{r+1}$-saturated graph. \subsection{Results} Our main tool will be the following. \begin{thm} \label{main1} If $n\geq r+1$ and $G$ is a $K_{r+1}$-saturated graph with $n$ vertices, then \begin{equation}\label{ne1} \sum_{v \in V(G)}d^2(v) \ge (n-1)^2(r-1)+(r-1)^2(n-r+1). \end{equation} For $r=2$, equality in the bound holds only when $G$ is $S_{n,2}$ or a Moore graph with diameter 2. For $r \ge 3$, equality in the bound holds only when $G$ is $S_{n,r}$. \end{thm} The reason why it is helpful is the following simple observation. \begin{lem}\label{lem2} For every $n$-vertex graph $G$ with adjacency matrix $A$, \begin{equation}\label{l2} \rho^2(A)\geq \frac{1}{n}\sum_{v\in V(G)}d^2(v). \end{equation} \end{lem} Theorem~\ref{main1} together with this observation immediately yield \begin{thm}\label{main2} If $2\leq r<n$ and $G$ is a $K_{r+1}$-saturated graph with $n$ vertices, then \begin{equation}\label{so} \rho(G) \ge \sqrt{\frac{(n-1)^2(r-1)+(r-1)^2(n-r+1)}{n}}. \end{equation} \end{thm} This bound asymptotically is tight because the spectral radius of $S_{n,r}$ is close to $f(n,r)$, where $f(n,r)$ is the lower bound for $\rho(G)$ in Theorem~\ref{main2}. More specifically, note that $\rho(S_{n,2})=f(2,n)$ and for $r \ge 3$, we have $\rho(S_{n,r})=f(r,n)+\frac{r-2}2+\Theta(\frac{r^{1.5}}{\sqrt{n}})$. \begin{prop}\label{jk} For integers $2\leq r<n$, $$\rho(S_{n,r})=\frac{r-2+\sqrt{(r-2)^2+4(r-1)(n-r+1)}}{2}.$$ \end{prop} In the next section we prove Theorem~\ref{main1} (in a somewhat stronger form) and in the last section we present proofs for Lemma~\ref{lem2} and Proposition~\ref{jk} For undefined terms, see Brouwer and Haemers~\cite{BH}, Godsil and Royle~\cite{GR}, or West~\cite{W}. \section{Proof of Theorem~\ref{main1}} We will derive Theorem~\ref{main1} from the following slightly stronger statement. \begin{thm} \label{main3} If $n\geq r+1$ and $G$ is a $K_{r+1}$-saturated graph with $n$ vertices, then \begin{equation}\label{ne3} \sum_{v \in V(G)}(d(v)+1)(d(v)+1-r) \ge (r-1)n(n-r). \end{equation} \end{thm} \begin{proof Let $m=|E(G)|$ and $\overline{m}=|E(\overline{G})|={n\choose 2}-m$. For $v\in V(G)$, let $f(v)$ be the number of pairs of non-adjacent vertices $x$ and $y$ in $N(v)$ such that $G[N(x)\cap N(y)\cap N(v)]$ contains $K_{r-2}$ as a subgraph. Note that if $G[N(v)]$ is a copy of $K_{r-1}$, then $f(v)=0$.\\ \noindent {\it Claim 1. $\overline{m}\leq \frac{1}{r-1}\sum_{v \in V(G)} f(v)$.}\\ We construct an auxiliary bipartite graph $H$ with parts $A$ and $B$ as follows. Let $A=E(\overline{G})$ and $B=V(G)$. The graph $H$ has an edge between $xy \in A$ and $v \in B$ iff $x,y \in N(v)$ and $G[N(x)\cap N(y)\cap N(v)]$ contains $K_{r-2}$ as a subgraph. Then for each $v \in B$, we have $|N_{H}(v)|=f(v)$. Also, since $G$ is $K_{r+1}$-saturated, for each $xy \in A$, $G+xy$ contains $K_{r+1}$ as a subgraph. Thus there exist at least $r-1$ vertices $v$ such that $x, y \in N(v)$ and $G[N(x)\cap N(y)\cap N(v)]$ contains $K_{r-2}$ as a subgraph, which implies \begin{equation}\label{eq2} |N_{H}(xy)| \ge r-1. \end{equation} By~\eqref{eq2}, \begin{equation}\label{eq21} (r-1)\overline{m} \le \sum_{xy\in A} d_H(xy)= |E(H)|=\sum_{v \in V(G)}f(v). \end{equation} This proves Claim 1. \\ \noindent {\it Claim 2. For each $v \in V(G)$, we have $\displaystyle f(v)\le {d(v)-r+2 \choose 2}$.} Let $H_v=G[N(v)]$, and let $d(v)=p$. Since $G$ contains no $K_{r+1}$, the graph $H_v$ has no $K_{r}$. Partition the pairs of vertices in $N(v)$ into the sets $E_1,E_2$ and $E_3$ as follows:\\ (i) $E_1=E(H_v)$, \\ (ii) $E_2$ is the set of the edges $xy\in E(\overline{H_v})$ such that $H_v+xy$ does not contain $K_r$,\\ (iii) $E_3$ is the set of the edges $xy\in E(\overline{H_v})$ such that $H_v+xy$ contains $K_r$. \\ Let $m_i=|E_i|$ for $1\leq i\leq 3$. By definition, $m_3=f(v)$ and $m_1+m_2+m_3={p \choose 2}.$ As any $K_r$-free graph is a subgraph of $K_r$-saturated graph on the same vertex set, there exists a $K_r$-saturated graph $H'$ with vertex set $N(v)$ containing $H_v$. Then $E(H')\supseteq E_1$. Furthermore, since $H'$ is $K_{r}$-free and contains $E_1$, $E(H')\cap E_3=\emptyset$. By Theorem A, $|E(H')|\geq (r-2)(p-r+2)+\binom{r-2}{2}$. Hence $$m_3\leq {p\choose 2}-|E(H')|\leq {p\choose 2}-(r-2)(p-r+2)-\binom{r-2}{2}={p-r+2 \choose 2}.$$ This proves Claim 2. \\ Now we are ready to prove the theorem. By Claims 1 and 2, $$ {n \choose 2}=m+\overline{m} \le m + \frac 1{r-1}\sum_{v\in V(G)} f(v) \le \sum_{v\in V(G)}\left[ \frac {d(v)}2+\frac 1{r-1}\frac{(d(v)-r+2)(d(v)-r+1)}2\right]. $$ Multiplying both sides by $2(r-1)$, we get $$(r-1)n(n-1) \le \sum_{v \in V(G)}\left[(r-1)d(v)+(d(v)+1)(d(v)-r+1)-(r-1)(d(v)-r+1)\right].$$ This yields $$\sum_{v \in V(G)}(d(v)+1)(d(v)+1-r)\ge (r-1)n(n-1)-(r-1)^2n=(r-1)n(n-r),$$ and Theorem~\ref{main3} is proved. \end{proof} To obtain Theorem~\ref{main1}, observe that~\eqref{ne3} implies $$\sum_{v \in V(G)}d^2(v) \ge (r-1)n(n-r)+(r-1)n+(r-2)2m.$$ So, by Theorem A, \begin{equation} \label{eq3} \sum_{v \in V(G)}d^2(v) \ge (r-1)n(n-r)+(r-1)n+2(r-2)\left[{n \choose 2}-{n-r+1 \choose 2}\right] \end{equation} $$ = (n-1)^2(r-1)+(r-1)^2(n-r+1). $$ This proves the first part of Theorem~\ref{main1}. Furthermore, for $r \ge 3$, equality in the bound requires equality in Theorem A. Thus equality holds only for $S_{n,r}$. Suppose now $r=2$ and $G$ is an $n$-vertex $K_3$-saturated graph for which~\eqref{ne1} holds with equality. As $G$ is $K_3$-saturated, $G$ has diameter 2. Equality in the bound requires equality in~\eqref{eq21}, and hence equality in~(\ref{eq2}) for every $xy\in E(\overline{G})$. This means $G$ has no $C_4$, which implies that $G$ has girth at least 5. If $G$ has no cycles, then $G$ is a copy of $S_{n,2}$. Otherwise, $G$ is a Moore graph with diameter 2. \medskip Recall that there are at most four Moore graphs with diameter 2: $C_5$, the Petersen graph, the Hoffman-Singleton graph, and possibly one 57-regular graph of girth $5$ with 3250 vertices. \section{Spectral radius} We will use the following standard tool. \begin{thm}[Rayleigh Quotient Theorem]\label{rqt} For a real matrix $A$ \begin{equation} \label{rqt1} \rho(A)=\max_{x \in \mathrm{R}^n \setminus \{0\}}\frac{x^TAx}{x^Tx}. \end{equation} \end{thm} First, we present a proof of Lemma~\ref{lem2}. By~\eqref{rqt1}, $$\rho^2(A)=\rho(A^2)=\max_{x \in \mathrm{R}^n \setminus \{0\}}\frac{x^TA^2x}{x^Tx}= \max_{x \in \mathrm{R}^n \setminus \{0\}}\frac{(x^TA^T)(Ax)}{x^Tx}\geq \frac{({\bf 1}^TA^T)(A{\bf 1})}{{\bf 1}^T{\bf 1}}=\frac{1}{n}\sum_{v\in V(G)}d^2(v). $$ Thus,~\eqref{l2} holds. Together with Theorem~\ref{main1}, this implies Theorem~\ref{main2}. \bigskip To show that Theorem~\ref{main2} is asymptotically tight, we will determine the spectral radius of $S_{n,r}$, i.e. prove Proposition~\ref{jk}. We will need a new notion. Consider a partition $V(G) = V_1 \cup \cdots \cup V_s$ of the vertex set of a graph $G$ into $s$ non-empty subsets. For $1 \le i, j \le s$, let $q_{i,j}$ denote the average number of neighbors in $V_j$ of the vertices in $V_i$. The quotient matrix $Q$ of this partition is the $s \times s$ matrix whose $(i, j)$-th entry equals $q_{i,j}$. The eigenvalues of the quotient matrix interlace the eigenvalues of $G$. This partition is {\it equitable} if for each $1 \le i, j \le s$, each vertex $v \in V_i$ has exactly $q_{i,j}$ neighbors in $V_j$. In this case, the eigenvalues of the quotient matrix are eigenvalues of $G$ and the spectral radius of the quotient matrix equals the spectral radius of $G$ (see \cite{BH}, \cite{GR} for more details). \bigskip \noindent [Proof of Proposition \ref{jk}] \\ Partition $V(S_{n,r})$ into sets $A$ and $B$ such that $S_{n,r}[A]$ is a copy of $K_{r-1}$ and $S_{n,r}[B]$ is an independent set with $n-r+1$ vertices. Each vertex in $A$ is adjacent to all vertices in $B$. The quotient matrix of the partitions $A$ and $B$ is $$\begin{pmatrix} r-2 & n-r+1 \\ r-1 & 0 \end{pmatrix}.$$ The characteristic polynomial of the matrix is $x^2 -(r-2)x -(r-1)(n-r+1)=0$. Since the partition $V(S_{n,r})=A \cup B$ is equitable, $$\rho(S_{n,r})=\frac{r-2+\sqrt{(r-2)^2+4(r-1)(n-r+1)}}{2}.$$ This completes the proof of Proposition \ref{jk}. \qed \bigskip Note that $\rho(S_{n,2})=\sqrt{n-1}$. Thus for $r=2$, equality in Theorem~\ref{main2} holds if and only if $G$ is $S_{n,2}$ or a Moore graph. For $r \ge 3$, the bound in Theorem~\ref{main2} may be improved, and we guess that the spectral radius of $S_{n,r}$ is the minimum of $\rho(H)$ among all $n$-vertex $K_{r+1}$-saturated graphs $H$. \bigskip {\bf Acknowledgement.} We thank Xuding Zhu for helpful comments.

https://arxiv.org/abs/2210.05520

On the convexity of the quaternionic essential numerical range

The numerical range in the quaternionic setting is, in general, a non convex subset of the quaternions. The essential numerical range is a refinement of the numerical range that only keeps the elements that have, in a certain sense, infinite multiplicity. We prove that the essential numerical range of a bounded linear operator on a quaternionic Hilbert space is convex. A quaternionic analogue of Lancaster theorem, relating the closure of the numerical range and its essential numerical range, is also provided.

\section*{Introduction} Let $\mathbb{F}$ be the field of complex numbers or the skew field $\mathbb{H}$ of Hamilton quaternions. Let $\mathcal{H}$ be a Hilbert space over $\mathbb{F}$ and let $T$ be a bounded linear operator on $\mathcal{H}$. The numerical range of $T$ is the set \[ W(T)=W_{\mathbb{F}}(T)=\{\langle Tx,x\rangle:\|x\|=1, x\in\mathcal{H}\}, \] where $\langle \cdot\,,\cdot\,\rangle:\mathcal{H}\times\mathcal{H}\to\mathbb{F}$ is the inner product on $\mathcal{H}$. This subset of $\mathbb{F}$ was introduced and studied by Toeplitz in 1918, who proved that, when $\mathbb{F}=\mathbb{C}$, the outer boundary of $W(T)$ is a convex curve and conjectured that the whole numerical range was convex, see \cite{To}. Shortly after, in 1919, Hausdorff \cite{Ha} proved the conjecture. Since then, this result is known as the Toeplitz-Hausdorff Theorem. Over the years, the investigation of the numerical range continuously increased, including the cases of linear operators on infinite dimensional complex Hilbert spaces and complex Banach spaces. In 1951, Kippenhahn \cite{Ki} introduced the study of numerical range for quaternionic operators, i.e, when $\mathbb{F}=\mathbb{H}$. Soon it became evident that, although sharing many properties of its complex counterpart, the quaternionic numerical range was no longer always convex. The bild of an operator $T$, also introduced in \cite{Ki}, is the intersection $B(T)=W_{\mathbb{H}}(T)\cap\mathbb{C}$. Since every quaternion is, up to unitary equivalence, a complex number, many properties of the numerical range are encoded in the bild, including convexity. In fact, $W_{\mathbb{H}}(T)$ is convex if, and only if, $B(T)$ is convex, see \cite{CDM3}. However, the upper bild $B^+(T)$, which is the intersection of $W_{\mathbb{H}}(T)$ with the closure of the upper half-plane, is always convex. The pursuit of convexity remained an important issue in the quaternionic setting, with Au-Yeung establishing in \cite{Ye1} necessary and sufficient conditions for $W_{\mathbb{H}}(T)$ to be convex. In a series of recent papers \cite{CDM1} - \cite{CDM5} the convexity and shape of the numerical range of quaternionic matrices have been studied by the first three named autors. The notion of S-spectrum in \cite{CGSS} and its relation with the numerical range on infinite dimensional quaternionic Hilbert spaces was adressed in the recent preprint \cite{CDM6}. Another geometric object in the realm of infinite dimensional Hilbert spaces is the essential numerical range of an operator $T$. It is defined as the set \[ W_{e}(T)=W_{e,\mathbb{F}}(T)=\bigcap_{K\in\mathcal{K}(\mathcal{H})}\overline{W(T+K)}, \] where $\mathcal{K}(\mathcal{H})$ denotes the set of compact operators on the $\mathbb{F}$-Hilbert space $\mathcal{H}$. Taking $K$ to be the zero operator in the above definition, we see that $W_{e,\mathbb{F}}(T)\subseteq\overline{W_{\mathbb{F}}(T)}$. This paper is devoted to the study of the essential numerical range in the quaternionic setting. The main result is theorem \ref{thm convexity} where we show that, for $\mathbb{F}=\mathbb{H}$, the essential numerical range $W_{e}(T)=W_{e,\mathbb{H}}(T)$ is always a convex set. Thus, at least convexity of this essential part of the numerical range is guaranteed even in the quaternionic setting. We emphasize that this is a surprising and unexpected result since the essential numerical range is the intersection of non-convex sets and nothing indicates it is convex in its formulation. To secure this result we use a general property (lemma \ref{lemmasequence}): given a pair of unitary sequences $x_n^{(1)}, x_n^{(2)}$ and $T\in \mathcal{B}(\mathcal{H})$, a judiscious choice of $N$ and $M$ shows that the following vectors are close to orthogonal \[ \inn{x_N}{y_M} \approx \inn{Tx_N}{y_M} \approx \inn{T^*x_N}{y_M}\approx 0. \] We can then form an essential sequence (see definition \ref{def_ess_seq}) for the convex combination $\alpha^2 \omega^{(1)}+\beta^2\omega^{(2)}$, $\omega^{(1)}, \omega^{2}\in W_e(T)$, with elements of the form $\alpha x_N^{(1)}+\beta y_M^{(2)}$. The referred quasi orthogonality implies that \begin{align*} &\|\alpha x_N^{(1)}+\beta x_M^{(2)}\|^2\approx \alpha^2\inn{x_N^{(1)}}{x_N^{(1)}}+\beta^2 \inn{x_M^{(2)}}{x_M^{(2)}}=1\\ &\langle T(\alpha x_N^{(1)}+\beta x_M^{(2)}), \alpha x_N^{(1)}+\beta x_M^{(2)}\rangle\approx \alpha^2\langle T x_N^{(1)}, x_N^{(1)}\rangle+\beta^2 \langle T x_M^{(2)}, x_M^{(2)}\rangle\approx \alpha^2 \omega^{(1)}+\beta^2\omega^{(2)}. \end{align*} We finish the paper with theorem \ref{thm lancaster quaternion}, where we prove a quaternionic version of Lancaster theorem relating the numerical range and the essential numerical range, see \cite{L}. Due to the nonconvexity of the numerical range, we need to introduce the notion of inter-convex hull (see (\ref{intraconvex hull})). The result asserts that the closure of the quaternionic numerical range is precisely the inter-convex hull of the quaternionic essential numerical range and the quaternionic numerical range, \emph{i.e} $\overline{W(T)}={\rm iconv\,}\{W_e(T), W(T)\}$. In spite of the formal similarities with its complex counterpart, there are worth mentioning differences. Foremost we can not infer that the numerical range is closed when it contains the essential numerical range (see remark \ref{remark_lancaster}) as in complex Hilbert spaces \cite[Corollary $1$]{L}. This is because the quaternionic numerical range lacks convexity and the quaternionic Lancaster theorem uses the weaker notion of inter-convex hull. In addition, remark \ref{remark_lancaster} tells us that, even though the upper bild is convex, we still do not recover Lancaster theorem in its complex form. \bigskip \section{Notation and preliminaries} The division ring of real quaternions $\mathbb{H}$, also known as Hamilton quaternions, is an algebra over the field of real numbers with basis $\{1,i,j,k\}$ and product defined by $i^2=j^2=k^2=ijk=-1$. Given a quaternion $q=q_0+q_1i+q_2j+q_3k$, its conjugate is $q^*=q_0-q_1i-q_2j-q_3k$. We call ${\rm Re\,}(q)=\frac{q+q^*}{2}$ and ${\rm Im\,}(q)=\frac{q-q^*}{2}$ the real and imaginary parts of $q$, respectively. The norm of $q$ is the nonnegative real number $|q|=\sqrt{qq^*}$. Two quaternions $q,q'\in\mathbb{H}$ are similar if there is a unitary $u\in\mathbb{H}$ such that $u^*qu=q'$, in which case we write $q\sim q'$. This is an equivalence relation and we denote the equivalence class of $q$ by $[q]$. Let $\mathcal{H}$ denote an infinite dimensional two-sided Hilbert space over $\mathbb{H}$. In particular, the norm of $x\in\mathcal{H}$ is defined by the underlying $\mathbb{H}$-inner product as $\|x\|=\sqrt{\langle x,x\rangle}$. The inner product verifies the usual Cauchy-Schwartz inequality: $|\langle x,y\rangle|\leq\|x\|\|y\|$, for every $x,y\in\mathcal{H}$. The space of bounded, right $\mathbb{H}$-linear operators on $\mathcal{H}$ is denoted by $\mathcal{B}(\mathcal{H})$, its closed ideal of compact operators by $\mathcal{K}(\mathcal{H})$ and the group of invertible operators by $\mathcal{B}(\mathcal{H})^{-1}$. Every linear operator $T$ considered in the text will be a bounded linear operator in $\mathcal{B}(\mathcal{H})$. Given $q\in\mathbb{H}$ and $T \in \mathcal{B}(\mathcal{H})$, we define the operator $\Delta_q(T):\mathcal{H}\to\mathcal{H}$ by \[ \Delta_q(T)=T^2-2{\rm Re\,}(q)T+|q|^2I, \] where $I$ is the identity operator. Clearly, $\Delta_q(T)\in\mathcal{B}(\mathcal{H})$. The spherical spectrum of $T$, abbreviated S-spectrum, is the set \[ \sigma^S(T)=\left\{q\in\mathbb{H}:\Delta_q(T)\notin \mathcal{B}(\mathcal{H})^{-1}\right\}, \] which seems to be the appropriate notion for spectral analysis of linear operators on infinite dimensional quaternionic Hilbert spaces, see \cite{CGSS}. Let $\pi:\mathcal{B}(\mathcal{H})\rightarrow \mathcal{B}(\mathcal{H})/\mathcal{K}(\mathcal{H})$ denote the canonical quotient map and $\mathcal{C}(\mathcal{H})=\mathcal{B}(\mathcal{H})/\mathcal{K}(\mathcal{H})$ the Calkin algebra. Let $\pi(T)=[T]$ denote the equivalence class $T+\mathcal{K}(\mathcal{H})$, for $T\in \mathcal{B}(\mathcal{H})$. Then $\mathcal{C}(\mathcal{H})$ is a normed algebra with $\lVert [T]\rVert=\inf_{K\in\mathcal{K}(\mathcal{H})}\lVert T+K\rVert \leq \lVert T\rVert$. We say that $T$ is a Fredholm operator if the class $[T]$ is invertible in $\mathcal{C}(\mathcal{H})$. According to Atkinson Theorem, $T\in \mathcal{B}(\mathcal{H})$ is a Fredholm operator if and only if its range is closed and the kernels $\ker(T)$ and $\ker(T^*)$ are finite dimensional, where $T^*\in\mathcal{B}(\mathcal{H})$ is the adjoint of $T$. The set of all Fredholm operators in $\mathcal{B}(\mathcal{H})$ is denoted by $\mathcal{F}(\mathcal{H})$. The essential S-spectrum of $T\in \mathcal{B}(\mathcal{H})$, defined by \begin{equation*} \label{eq01} \sigma_{e}^S(T)=\left\{ q \in \mathbb{H} : \Delta_q(T) \notin \mathcal{F}(\mathcal{H}) \right\}, \end{equation*} is a non-empty compact subset of $\sigma^S(T)$, see \cite{MT}. In the sequel, we will be working in the quaternion setting, that is, the quaternions $\mathbb{H}$ are our ground field (skewfield to be more precise). Therefore, when we write $W(T)$ or $W_e(T)$, we always refer to the quaternionic numerical range or quaternionic essential numerical range. Finally, define the essential bild and the essentials upper and lower bilds to be, respectively, $B_e(T)=W_e(T)\cap\mathbb{C}$, $B_e^+(T)=W_e(T)\cap\mathbb{C}^+$, and $B_e^-(T)=W_e(T)\cap\mathbb{C}^-$, where $\mathbb{C}^{\pm}$ is the closure of the respective half-planes. \bigskip \section{Properties of the essential numerical range} This section is devoted to elementary properties of the essential numerical range and to prove some criteria for a quaternion to be in the essential numerical range of an operator. The results and their proofs are identical to the complex case with some adjustments. For the sake of completeness full proofs are provided. We start with an auxiliary result concerning compact operators. \begin{lemma}\label{lem compact} An operator $T$ is compact if and only if $\langle Te_n,e_n\rangle\to 0$ for every orthonormal set $(e_n)_{n}$. \end{lemma} \begin{proof} Let $T\in\mathcal{B}(\mathcal{H})$ be compact and let $(e_n)_{n}$ be an orthonormal set. Let $P_n$ be the projection onto $\text{span}\,\{e_1,\dots, e_n\}$. Since $T$ is compact, it is the limit of a sequence of finite rank operators, i.e., $\lim_{n\rightarrow\infty}\lVert P_nT-T\rVert=0$ (see \cite[Corollary 4.5]{C}). Then \[ \lim_{n\rightarrow\infty} \lVert (I-P_n)T(I-P_n)\rVert \leq\lim_{n\rightarrow\infty}\lVert T-P_nT\rVert \lVert I-P_n\rVert=0. \] Since $(I-P_n)e_{n+1}=e_{n+1}$, and using the Cauchy-Schwatz inequality, we have \begin{eqnarray*} \lvert \langle T e_{n+1}, e_{n+1}\rangle\rvert &=& \lvert \langle T(I-P_n) e_{n+1}, (I-P_n)e_{n+1}\rangle \rvert\\ &\leq & \lVert (I-P_n)T(I-P_n)\rVert. \end{eqnarray*} Hence, $\langle T e_{n}, e_{n}\rangle\to 0$. For the converse, suppose that $T\in \mathcal{B}(\mathcal{H})$ is such that $\langle T e_{n}, e_{n}\rangle\to 0$, for every orthonormal set $(e_n)_{n}$. From $\lVert T\rVert=\sup_{\|x\|=\|y\|=1}|\langle Tx, y\rangle|,$ there exist unit vectors $x_1,y_1\in\mathcal{H}$ such that \begin{equation* \lvert \langle Tx_1,y_1\rangle \rvert\geq\frac{\lVert T\rVert }{2}. \end{equation*} A straightforward computation shows the following ``polarization identity'', for all $x,y\in\mathcal{H}$: \begin{equation*} 4\langle Tx,y\rangle = {} \langle T(x+y), x+y \rangle-\langle T(x-y), x-y \rangle + \Big(\langle T(x+yi), x+yi \rangle- \langle T(x-yi), x-yi \rangle\Big)i \end{equation*} \[ +k \Big(\langle T(x+yk), x+yk \rangle- \langle T(x-yk), x-yk \rangle\Big) + k \Big(\langle T(x+yj), x+yj \rangle- \langle T(x-yj), x-yj \rangle\Big)i . \] In particular, it follows that \[ |\langle Tx_1,y_1\rangle|\leq \frac{1}{4} \sum_{u\in U} |\langle Tu,u\rangle|, \] where $U=\left\{x_1+\eta y_1 \, : \, \eta= \pm 1, \pm i, \pm j, \pm k\right\}$. More precisely, for some $u_0\in U$ we can write \begin{eqnarray*} |\langle Tx_1,y_1\rangle| & \leq & \frac{8}{4} |\langle Tu_0,u_0 \rangle| = 2 \Bigg| \langle T\left(\frac{u_0}{\|u_0\|}\right), \frac{u_0}{\|u_0\|} \rangle\Bigg| \, \|u_0\|^2 \\ & \leq & 8 \Bigg| \langle T\left(\frac{u_0}{\|u_0\|}\right), \frac{u_0}{\|u_0\|} \rangle\Bigg| , \end{eqnarray*} where in the last inequality we used the fact that $\|u_0\|\leq 2$. Set $\rho_1=u_0/ \|u_0\|\in \mathcal{H} $. Then, $\rho_1$ is a unit vector such that \begin{eqnarray*} \frac{\lVert T\rVert}{2} \leq \lvert \langle Tx_1,y_1\rangle\rvert \leq 8 \lvert \langle T\rho_1, \rho_1\rangle \rvert \Leftrightarrow \frac{\lVert T\rVert}{16} \leq \lvert \langle T\rho_1, \rho_1\rangle \rvert. \end{eqnarray*} Now, let $P_1$ be the orthogonal projection onto $\text{span}\,\{\rho_1\}$. By applying the above argument to the operator $(I-P_1)T(I-P_1)$, we can find a unit vector $\rho_2$ orthogonal to $\rho_1$ such that \begin{eqnarray*} \frac{\lVert (I-P_1)T(I-P_1)\rVert}{16} \leq \lvert \langle T\rho_2, \rho_2\rangle \rvert. \end{eqnarray*} Moreover, a recursive procedure allows us to construct an orthonormal sequence $(\rho_n)_{n}$ such that if $P_n$ is the projection onto the span of $\{\rho_1, \dots, \rho_{n}\}$ then \[ \frac{ \lVert (I-P_n)T(I-P_n) \rVert}{16} \leq \lvert \langle T\rho_{n+1}, \rho_{n+1}\rangle \rvert. \] Since $\rho_n$ is an orthonormal sequence, by assumption, we have $\lim_{n\rightarrow \infty}\langle T\rho_n, \rho_n\rangle=0$, so that \[ \lim_{n\rightarrow \infty} \lVert (I-P_n)T(I-P_n) \rVert=\lim_{n\rightarrow \infty} l\Vert (P_nT+TP_n-P_nTP_n)-T\rVert=0, \] and thus $T$ is compact (being the limit of the finite rank operators $P_nT+TP_n-P_nTP_n$). \end{proof} \bigskip Next result, well-known in the complex setting (see \cite[Corollary in page 189]{FSW}), gives necessary and sufficient conditions for an element $q\in\mathbb{H}$ to belong to $W_e(T)$, for some operator $T\in\mathcal{B}(\mathcal{H})$. A very important class of unitary vectors regarding the essential numerical range, portrayed bellow in condition $b)$, will be called an essential sequence, see definition \ref{def_ess_seq}. As usual, we write $x_n\rightharpoonup x$ if a sequence $(x_n)_n$ in $\mathcal{H}$ converges to $x\in\mathcal{H}$ in the weak topology. \begin{theorem}\label{theoEssNRseq} Let $q\in \mathbb{H}$. The following conditions are equivalent: \begin{enumerate}[a)] \item $q\in W_e(T)$. \item There exists a sequence of unit vectors $(x_n)_{n}$ in $\mathcal{H}$ such that $x_n\rightharpoonup 0$ and $\langle Tx_n,x_n\rangle\to q$. \item There exists an orthonormal sequence $(e_n)_{n}$ in $\mathcal{H}$ such that $\langle Te_n,e_n\rangle\to q$. \end{enumerate} \end{theorem} \begin{proof} $b)\Rightarrow a).\,$ Suppose b) holds. To see that $q\in \bigcap_{K\in\mathcal{K}(\mathcal{H})}\overline{W(T+K)}$, we will show that $\langle (T+K)x_n,x_n\rangle\to q$, for every compact operator $K$. At this point we need the following well-known result: if $K$ is compact and $x_n\rightharpoonup x$, then $Kx_n\to Kx$ strongly. In particular, if $x_n\rightharpoonup 0$ then $\|Kx_n\|\to 0$. It follows that \[ \langle (T+K)x_n,x_n\rangle=\langle Tx_n,x_n\rangle+\langle Kx_n,x_n\rangle\to q, \] since we have $\left|\langle Kx_n,x_n\rangle\right|\leq\|Kx_n\|$, for every $n$. $c)\Rightarrow b).\,$ The result follows from the fact that $e_n\rightharpoonup 0$ for every orthonormal sequence $(e_n)_n$. $a)\Rightarrow c)$. Since $W_e(T)=[B_e(T)]$, it is enough to prove the result for the essential upper bild. Let $q\in B^+_e(T)$. From ${B_{e}^+(T)}\subseteq \overline{B^+(T)}$, there is a sequence of unit vectors $(\xi_n)_n$ in $\mathcal{H}$ such that $\langle T\xi_n,\xi_n\rangle\in B^+(T)$ and $\lim_{n\rightarrow\infty}\langle T\xi_n,\xi_n\rangle=q$. Take $\xi_{M}$, which we call without loss of generality $\xi_1$, such that \[ |\langle T\xi_1,\xi_1\rangle-q|\leq \frac{1}{2}. \] Let $\mathcal{L}_1:=\text{span}\,\{\xi_1\}$ and write $\mathcal{H}=\mathcal{L}_1\oplus\mathcal{L}_1^\bot$. Denote $P_1:\mathcal{H}\rightarrow\mathcal{H}$ the orthogonal projection onto $\mathcal{L}_1$. From \cite[Corollary 3.3]{CDM3} we know that the quaternionic numerical range of an operator, and therefore its upper bild, always intersects the real line. So, we can take a real number $\mu_1\in B^+\Big((I-P_1)\restr{T}{\mathcal{L}_1^\bot}\Big)\cap\mathbb{R}$. Let $F_1$ be the finite rank operator such that $ T+F_1 =\mu_1P_1+(I-P_1)T(I-P_1)$. Then, $F_1$ compact and, since $q\in B_e^+(T)$, it follows that \[ q\in \overline{B^+(T+F_1)}=\overline{B^+(\mu_1P_1+(I-P_1)T(I-P_1))}. \] However, it is clear that \begin{eqnarray*} & &B^+\left(\mu_1P_1+(I-P_1)T(I-P_1)\right) =\\ &=& \left\{\langle (\mu_1P_1+(I-P_1)T(I-P_1)) (x_1+x_2), x_1+x_2 \rangle\, :\, (x_1,x_2)\in\Omega\right\} \cap \mathbb{C}^+ \\ &=& \left\{\mu_1\lVert x_1\rVert^2+\lVert x_2\rVert^2\langle (I-P_1)T \frac{x_2}{\|x_2\|},\frac{x_2}{\|x_2\|} \rangle\, :\, (x_1,x_2)\in\Omega\right\}\cap \mathbb{C}^+ , \end{eqnarray*} where $\Omega=\left\{(x_1,x_2):x_1\in \mathcal{L}_1, x_2\in\mathcal{L}_1^\bot, \, \lVert x_1\rVert^2+\lVert x_2\rVert^2=1\right\}$. Since $\mu_1\in B^+\Big((I-P_1)\restr{T}{\mathcal{L}_1^\bot}\Big)\cap \mathbb{R}$ and $B^+\Big((I-P_1)\restr{T}{\mathcal{L}_1^\bot}\Big)$ is convex (see \cite[Corollary 1]{Ye1}) we obtain \[ B^+(\mu_1P_1+(I-P_1)T(I-P_1)) = B^+\Big((I-P_1)T_{|\mathcal{L}_1^\bot}\Big). \] Hence, $q\in \overline{B^+\Big((I-P_1)\restr{T}{\mathcal{L}_1^\bot}\Big)}$. So there is a unit vector $\xi_2\in \mathcal{L}_1^\bot$ such that \begin{eqnarray*} \lvert \langle (I-P_1)\restr{T}{\mathcal{L}_1^\bot}\xi_2,\xi_2\rangle - q\rvert \leq \frac{1}{2^2} & \Leftrightarrow & |\langle T \xi_2,\xi_2\rangle - q| \leq \frac{1}{2^2}. \end{eqnarray*} If $\xi_1, \dots, \xi_n$ are orthonormal vectors such that $|\langle T \xi_n,\xi_n\rangle - q| \leq \frac{1}{2^n},$ we can repeat the above procedure with $\mathcal{L}_n:=\text{span}\,\{\xi_1, \dots, \xi_n\}$, $P_n$ the orthogonal projection onto $\mathcal{L}_n$, $\mu_n\in B^+\Big((I-P_n)\restr{T}{\mathcal{L}_n^\bot}\Big)\cap\mathbb{R}$ and $F_n$ such that $T+F_n=\mu_nP_n+(I-P_n)T(I-P_n)$. We thus obtain a unit vector $\xi_{n+1}$ orthogonal to each $\xi_k$ for $1\leq k\leq n$ such that \[ |\langle T \xi_{n+1},\xi_{n+1}\rangle - q| \leq \frac{1}{2^{n+1}}. \] By recursion, there exists an orthonormal sequence $(\xi_n)_{n}$ in $\mathcal{H}$ such that $\langle T\xi_n,\xi_n\rangle\to q$. \end{proof} We will call any sequence satisfying b) an essential sequence for $q$, as stated in the following definition. \begin{definition}\label{def_ess_seq} An essential sequence $(x_n)_{n}\subset \mathcal{H}$ for $q$ is a sequence of unit vectors such that $x_n\rightharpoonup 0$ and $\langle Tx_n,x_n\rangle\to q$. \end{definition} \bigskip An immediate consequence of theorem \ref{theoEssNRseq} is the non-emptiness of the essential numerical range. In fact, for any orthonormal sequence $(e_n)_{n}$, the sequence $\Big(\langle Te_n,e_n\rangle\Big)_{n}$ is bounded by $\|T\|$. Then, it has a convergent subsequence. By c) in theorem \ref{theoEssNRseq} we have that $W_e(T)$ is non-empty. Moreover, it is clear that $W_e(T)$ is a compact set since it is closed and bounded in $\mathbb{H}$. These properties are summarized in the corollary below. \medskip \begin{corollary} $W_e(T)$ is a non-empty and compact set. \end{corollary} The essential numerical range in the quaternionic setting shares many properties with either the complex essential numerical range or the quaternionic numerical range. We collect some of such properties below. The proofs are direct and for that reason only a short hint is provided. \bigskip \begin{proposition}\label{prop properties ess n.r} The following properties of the quaternionic essential numerical range hold. \begin{enumerate}[(i)] \item $W_e(T+K)=W_e(T)$, for all $K\in\mathcal{K}(\mathcal{H})$. \item $q\in W_e(T)$ if and only if $[q]\subseteq W_e(T)$. \item $W_e(T^*)=W_e(T)$. \item $W_e(T)\subseteq \overline{\mathbb{D}(0, \|\pi(T)\|)}$. \item If $a,b\in\mathbb{R}, W_e(aT+bI)=aW_e(T)+b$. \item $W_e(T+S)\subseteq W_e(T)+W_e(S)$. \item If $U\in\mathcal{B}(\mathcal{H})$ is unitary, then $W_e(UTU^*)=W_e(T)$. \item $W_e(T)$ contains all eigenvalues of $T$ of infinite multiplicity. \end{enumerate} \end{proposition} \begin{proof} $(i)$ follows from $K+\mathcal{K}(\mathcal{H})=\mathcal{K}(\mathcal{H}) $, for any $K \in\mathcal{K}(\mathcal{H})$; $(ii)$ results from $q\in {W(T)}$ if and only if $[q]\subseteq {W(T)}$, for every operator $T$; $(iii)$ is a consequence of $W(T^*)=W(T)$, for every $T \in \mathcal{B}(\mathcal{H})$; the inclusion $W(T)\subseteq \overline{\mathbb{D}(0, \|T\|)}$ implies $(iv)$; $(v)$ holds because $W(aT+bI)=aW(T)+b$, for $a,b\in\mathbb{R}$; from $\mathcal{K}(\mathcal{H})+\mathcal{K}(\mathcal{H})=\mathcal{K}(\mathcal{H})$ and $W(T+S)\subseteq W(T)+W(S)$ we obtain $(vi)$; $(vii)$ follows from $W(UTU^*)=W(T)$; for $(viii)$ note that the orthonormal set $(e_n)_{n}$ of eigenvectors satisfying $Te_n=e_n q$ is an essential sequence for $q$. \end{proof} \medskip From \cite[Theorem 2.9]{CDM6} we know that $\sigma^S(T+K) \subseteq \overline{W(T+K)}$, for every $K\in \mathcal{K}(\mathcal{H})$. Using the notion of Weyl S-spectrum, $\sigma_w^S(T):= \bigcap_{K\in \mathcal{K}(\mathcal{H})}\sigma^S(T+K) $, and that $\sigma_{e}^S(T)\subseteq \sigma_w^S(T)\subseteq \sigma^S(T)$ (see Definition 6.1 and Theorem 6.6 in \cite{MT}), we have the following result. \begin{theorem}\label{Theo_Sspecturm_NR} $\sigma_e^S(T)\subseteq W_e(T)$. \end{theorem} \section{Convexity} \bigskip In this section we establish the main result of the paper which asserts that the quaternionic essential numerical range is convex. To see this we will show that for any two elements $\w1, \w2$ in $W_e(T)$, their convex combination can be arbitrarily approximated by elements $\langle Tz, z\rangle$, where $z \in \mathcal{H}$ is generated by an essential sequence for $\w1$ and an essential sequence for $\w2$. To construct such elements $z\in\mathcal{H}$ we need a preparatory lemma which states a general property enjoyed by a pair of unitary sequences weakly vanishing and a bounded linear operator. \begin{lemma}\label{lemmasequence} Let $T\in B(\mathcal{H})$ and $\big(x^{(i)}_n\big)_n$, $i=1,2$, be unitary sequences in $\mathcal{H}$ such that $x_n^{(i)}\rightharpoonup 0$. For any $\varepsilon>0$ and $N \in \mathbb{N}$, there is $M \in \mathbb{N}$ such that $M\geq N $ and \[ \big|\inn{x^{(1)}_N}{x^{(2)}_M}\big|\leq \varepsilon, \quad \big|\inn{Tx^{(1)}_N}{x^{(2)}_M}\big|\leq \varepsilon, \quad\text{and} \quad \big|\inn{T^*x^{(1)}_N}{x^{(2)}_M}\big|\leq \varepsilon. \] \end{lemma} \begin{proof} Let $\delta>0$. Let $(e_k)_k$ be an orthonormal basis for $\mathcal{H}$ and $P_K$ be the projection onto $\text{span}\,\{e_1,\dots, e_K\}$. Since $(I-P_K)y\xrightarrow[K \to \infty]{}0$ for every $y \in \mathcal{H}$, then, for the above $\delta>0$ and $N\in\mathbb{N}$, we may find $K \in \mathbb{N}$ such that \begin{equation}\label{Kcondts} \|(I-P_K)x^{(1)}_N\|\leq \delta, \quad \|(I-P_K)Tx^{(1)}_N\|\leq \delta \quad\text{ and }\quad \|(I-P_K)T^* x^{(1)}_N\|\leq \delta. \end{equation} We can find an $M\in\mathbb{N}$ that depends on $\delta$, $N$, $K$, such that $M \geq N$ and \begin{equation}\label{Mcondts} |\langle x^{(2)}_M, e_k\rangle | \leq \frac{\delta}{2^{k/2}}, \text{ for every } 1 \leq k \leq K. \end{equation} Inequality (\ref{Mcondts}) follows from the fact that $\big(x^{(2)}_n\big)_n$ vanishes weakly, and that implies coordinatewise convergence to zero. It follows that \[ \Big\|\sum_{1\leq k \leq K} \langle x^{(2)}_M, e_k\rangle \;e_k\Big\|^2 = \sum_{1\leq k \leq K} \Big| \langle x^{(2)}_M, e_k\rangle \Big|^2 \leq \delta^2 \] and therefore, \begin{equation} \label{convlemma4} \big\|P_K x^{(2)}_M\big\|\leq \delta. \end{equation} Noting that $\Vert x^{(1)}_N\rVert=\lVert x^{(2)}_M\rVert=1$, we have \begin{align*} \Big|\langle x^{(1)}_N, x^{(2)}_M\rangle\Big| &\leq \Big| \langle x^{(1)}_N, (I-P_K)x^{(2)}_M\rangle \Big| + \Big| \langle x^{(1)}_N, P_K x^{(2)}_M\rangle \Big| \\ &\leq\big\| (I-P_K)x^{(1)}_N\big\| \,\,\big\| x^{(2)}_M \big\| + \big\| x^{(1)}_N\big\|\,\,\big\|P_K x^{(2)}_M\big\|\\ &\leq 2\delta \,\,\,\,\,\,\,(\textrm{from}\,\,(\ref{Kcondts})\,\,\text{and}\,\,(\ref{convlemma4})). \end{align*} Using a similar reasoning, we can show that \begin{align*} \big|\langle Tx^{(1)}_N, x^{(2)}_M\rangle\big|&\leq \| (I-P_K)Tx^{(1)}_N\|\,\, \|x^{(2)}_M\|+\|Tx^{(1)}_N\| \| P_K x^{(2)}_M\|\\ &\leq \| (I-P_K)Tx^{(1)}_N\|+\|T\| \,\, \delta \\ &\leq \delta+\|T\| \,\, \delta \end{align*} and $\big|\langle T^*x^{(1)}_N, x^{(2)}_M\rangle\big|\leq \delta+\|T\| \,\, \delta$. Letting $\delta$ be such that $\max\{2, 1+\|T\|\}\delta\leq \varepsilon$ the lemma follows. \end{proof} \begin{theorem}\label{thm convexity} $ W_e(T)$ is convex. \end{theorem} \begin{proof} Convexity of $W_e(T)$ will be proved by showing that $\alpha^2\w1+\beta^2\w2 \in W_e(T)$ for any $\alpha^2+\beta^2=1$, when $\w1, \w2$ lie in $W_e(T)$. For that we will prove there is an essential sequence $(\tilde z_p)_p$ for $\alpha^2\w1+\beta^2\w2$. Let $\big(x^{(i)}_n\big)_n$ be an essential sequence for $\omega^{(i)}$ and denote $\omega_n^{(i)}=\langle Tx^{(i)}_n, x^{(i)}_n\rangle$, for $i=1,2$. For any $p\in \mathbb{N}$ let $\varepsilon=1/p$. One of the conditions for the sequence $\big(x^{(i)}_n\big)_n$ to be essential for $\w{i}$ is that $\w{i}_n \to \w{i}$ when $n \to \infty$. Hence, for the given $\varepsilon$, there exists $N\geq p$ satisfying \begin{equation}\label{Ncondts} |\w1_n-\w1|\leq \varepsilon \text{ and } |\w2_n-\w2|\leq \varepsilon, \; \text{ for } n \geq N. \end{equation} Pick $M$ according to the previous lemma. For the fixed $\alpha$ and $\beta$, let $z=\alpha x^{(1)}_N+\beta x^{(2)}_M$. Since $\alpha^2+\beta^2=1$ and $\alpha \beta \leq \frac{1}{2}$, we easily verify that \begin{equation}\label{convlemma1}\\ \big|\|z\|^2-1\big|\leq \big|\langle x^{(1)}_N, x^{(2)}_M\rangle \big|\leq \varepsilon. \end{equation} A simple computation shows that \begin{align*} \Big|\langle T z, z\rangle - \big(\alpha^2\w1_N+\beta^2\w2_M\big)\Big| &= \alpha\beta\Big|\langle Tx^{(1)}_N, x^{(2)}_M\rangle + \overline{\langle T^*x^{(1)}_N, x^{(2)}_M\rangle} \Big| \leq \varepsilon.\nonumber \end{align*} From (\ref{Ncondts}), it follows that \begin{align} \Big|\langle T z, z\rangle - \big(\alpha^2\w1+\beta^2\w2\big)\Big|&\leq \Big| \big(\alpha^2\w1_N+\beta^2\w2_M\big)-\big(\alpha^2\w1+\beta^2\w2\big) \Big|\nonumber\\ & \,\,\,\,\,\,\,\,\,\,+ \Big|\langle T z, z\rangle - \big(\alpha^2\w1_N+\beta^2\w2_M\big)\Big|\nonumber\\ % &\leq \alpha^2 \Big|\w1_N - \w1\big|+ \beta^2 \Big|\w2_M - \w2\big| +\Big|\langle T z, z\rangle - \big(\alpha^2\w1_N+\beta^2\w2_M\big)\Big| \nonumber\\ &\leq 2\varepsilon.\label{convlemma2} \end{align} Observing that the fixed integers $N$ and $M$ depend on $\varepsilon$, that is on $p\in\mathbb{N}$, we denote them by $N_p$ and $M_p$; likewise, we denote $z$ by $z_p$. To get an essential sequence we have to normalize $(z_p)_p$. Write $\tilde z_p=\frac{z_p}{\|z_p\|}$. From (\ref{convlemma1}), $\|z_p\|\to 1\,\,(p\to\infty)$, and so $(\tilde z_p)_p$ is well defined. By definition, $z_p=\alpha x^{(1)}_{N_p}+\beta x^{(2)}_{M_p}$, and $x^{(1)}_{N_p}, x^{(2)}_{M_p} \rightharpoonup 0$, when $p \to \infty$. By linearity and since $\|z_p\|\to 1$, we have that $\tilde z_p \rightharpoonup 0$. Finally, from (\ref{convlemma2}) it follows that \[ \langle T \tilde z_p, \tilde z_p\rangle = \frac{1}{\|z_p\|^2} \langle T z_p, z_p\rangle \to \alpha^2\w1+\beta^2\w2. \] The sequence $(\tilde z_p)_p$ is essential for $\alpha^2\w1+\beta^2\w2$ and thus, by theorem \ref{theoEssNRseq}, $\alpha^2\w1+\beta^2\w2\in W_e(T).$ \end{proof} \bigskip Next result establishes the relation between the boundary of the numerical range and the essential numerical range. This is the quaternionic analogue of Lancaster's theorem for the complex numerical range, see \cite[Theorem 1]{L}. Since the quaternionic numerical range is not always convex, a modification is imposed and we need to introduce the notion of inter-convex hull of sets (see \cite[Definition 3.2]{CDM2}). The inter-convex hull of the sets $A$ and $B$, denoted by ${\rm iconv\,}\{A, B\}$, closes the set $A \cup B$ to the convex combinations with one element of each sets, \begin{equation}\label{intraconvex hull} {\rm iconv\,}\{A, B\}=\{\alpha a+(1-\alpha) b: a \in A, \, b \in B, \, 0\leq \alpha \leq 1\}. \end{equation} \begin{theorem}\label{thm lancaster quaternion} The closure of the numerical range is $\overline{W(T)}={\rm iconv\,}\{W_e(T), W(T)\}$. \end{theorem} \begin{proof} We start proving that ${\rm iconv\,}\{W_e(T), W(T)\} \subseteq\overline{W(T)}$. Let $\bar \omega \in {\rm iconv\,}\{W_e(T), W(T)\}$. Then $\bar{\omega}= \alpha^2 \omega+\beta^2 \omega_e$ with $\omega \in W(T), \omega_e \in W_e(T)$ and $\alpha^2+\beta^2=1$. In particular, we can take a unitary $y \in \mathcal{H}$ such that $\omega=\inn{Ty}{y}$ and an essential sequence $(y_n)_n$ for $\omega_e$. Since $y_n \rightharpoonup 0$, we have that $\lim \,\inn{y_n}{y}=\lim \, \inn{y_n}{Ty}=\lim\, \inn{y_n}{T^*y}=0$. Let $z_n=\alpha y+\beta y_n$. Then, \[ \innt{z_n}=\alpha^2 \innt{y}+\beta^2\innt{y_n} +\alpha\beta\Big(\inn{Ty}{y_n}+\inn{Ty_n}{y}\Big) \to \alpha^2 w+\beta^2w_e=\bar{\omega}. \] Furthermore, \[\|z_n\|^2=\alpha^2\|y\|^2 +\beta^2\|y_n\|^2 +\alpha\beta\Big(\inn{y}{y_n}+\inn{y_n}{y}\Big)\to 1. \] Thus $W(T) \ni \innt{\frac{z_n}{\|z_n\|}} \to \bar\omega$, and $\bar\omega \in \overline{W(T)}$. To prove the converse inclusion, take $\overline\omega \in \overline{W(T)}$. There is a sequence $\left(y_n\right)_n$ in $\mathcal{H}$ satisfying $\|y_n\|=1$ and $\omega_n=\langle T y_n, y_n\rangle \to \overline\omega $. Since this sequence is in the unit circle, there is an element $y \in \mathcal{H}$ in the unit disk such that $y_n$ converges weakly to $y$. If $y=0$, then $(y_n)_n$ is an essential sequence for $\overline\omega$. From theorem \ref{theoEssNRseq} we have $\overline\omega \in W_e(T)$. If $\|y\|=1$, we have that $y_n\rightharpoonup y$, with $\|y\|=1=\|y_n\|$. It is well-known that in this case $y_n \rightarrow y$ (strongly). Thus, $\innt{y_n} \to \innt{y}$, that is, $ \overline\omega = \innt{y} \in W(T)$. Assume now that $\|y\| \neq 0,1$. Using that $\langle y_n, h\rangle \to \langle y, h\rangle$ for any $ h \in \mathcal{H}$, we can prove that $\lim \;\langle Ty_n, y\rangle=\lim \;\langle Ty, y_n\rangle= \langle Ty, y\rangle$ and therefore \[ \lim \; \langle Ty_n, y_n\rangle =\lim \;\left[ \langle Ty, y\rangle + \langle T(y_n-y), y_n-y\rangle\right]. \] It is easy to see that $\lim \|y_n-y\|^2= 1-\|y\|^2$. Then \begin{align*} \overline\omega=\lim\;\langle Ty_n, y_n\rangle= & \lim \left[\|y\|^2 \big\langle T\frac{y}{\|y\|}, \frac{y}{\|y\|}\big\rangle + \|y_n-y\|^2\Big\langle T\frac{y_n-y}{\|y_n-y\|}, \frac{y_n-y}{\|y_n-y\|}\Big\rangle\right]\\ =& \|y\|^2 \big\langle T\frac{y}{\|y\|}, \frac{y}{\|y\|}\big\rangle + (1-\|y\|^2)\lim \Big\langle T\frac{y_n-y}{\|y_n-y\|}, \frac{y_n-y}{\|y_n-y\|}\Big\rangle . \end{align*} We have just written $\overline\omega$ as a convex combination of $\omega=\langle T\frac{y}{\|y\|}, \frac{y}{\|y\|}\big\rangle \in W(T)$ and $\omega_e= \lim \;\langle T\frac{y_n-y}{\|y_n-y\|}, \frac{y_n-y}{\|y_n-y\|}\Big\rangle$. We use theorem \ref{theoEssNRseq} again, observing that $\left(\frac{y_n-y}{\|y_n-y\|}\right)_n$ is an essential sequence for $w_e$, to conclude that $w_e\in W_e(T)$. Therefore, $\overline\omega \in {\rm iconv\,}\{W_e(T), W(T)\}$. \end{proof} As in other results concerning the quaternionic numerical range, next corollary shows that we can simply consider what happens in the complex plane. Given a quaternion $q=q_0+q_1i+q_2j+q_3k$, we define $\pi_{(1)}(q)={\rm Re\,}(q)=q_0$ and $\pi_{(i)}(q)=q_1$. \begin{corollary} Let $T\in\mathcal{B}(\mathcal{H})$. Then $\overline{B(T)}={\rm iconv\,}\{B_e(T), B(T)\}$. \end{corollary} \begin{proof} From theorem \ref{thm lancaster quaternion} we have \[ {\rm iconv\,}\{B_e(T), B(T)\}\cap \mathbb{C} \subseteq {\rm iconv\,}\{W_e(T), W(T)\}\cap \mathbb{C}=\overline{W(T)}\cap \mathbb{C}. \] We obtain that ${\rm iconv\,}\{B_e(T), B(T)\} \subseteq \overline{B(T)}$. For the converse inclusion, take an element $\bar\omega \in \overline{B(T)}$. According to theorem \ref{thm lancaster quaternion} there are $\omega \in W(T)$, $\omega_e \in W_e(T)$ and $\alpha \in [0,1]$, such that \begin{equation}\label{omega_bar} \bar\omega =\alpha \omega +(1-\alpha)\omega_e. \end{equation} Observe that when $\alpha=0$ or $\alpha=1$ the inclusion immediately follows. So suppose $\alpha\neq 0, 1$. We can write $\omega=a+bq$ and $w_e=c+dq_e$, where $a, b, c, d\in \mathbb{R}$, $q\in {\rm Im\,}(q)$, $q_e\in {\rm Im\,}(q_e)$ and $|q|=|q_e|=1$. Therefore, we have \[ \bar\omega = (\alpha a+(1-\alpha)c)+(\alpha bq+(1-\alpha)dq_e). \] Note that $\alpha bq+(1-\alpha)dq_e\in \text{span}\,\{i\}$, since $\bar\omega\in \mathbb{C}$. Assume that $\bar\omega\in \mathbb{C}^+$. If $\bar\omega\in \mathbb{C}^-$, the proof is analogous. By circularity of the bild, there are $\omega_{(i)} \in [\omega] \cap \mathbb{C}^+$ in the bild and $\omega_{e,(i)} \in [\omega_e] \cap \mathbb{C}^+$. We can write $\omega_i=a+|b|i$ and $\omega_{e,i}=c+|d|i$. Define $\overline{\omega}_i=\alpha\omega_i+(1-\alpha)\omega_{e,i}$, which can be written as \[ \overline{\omega}_i=(\alpha a+(1-\alpha)c)+(\alpha |b|+(1-\alpha)|d|)i. \] Clearly, $\pi_{(1)}(\bar\omega)=\pi_{(1)}(\bar\omega_i)$. On the other hand, since $\alpha bq+(1-\alpha)dq_e\in \text{span}\,\{i\}$ and $|q|=|q_e|=1$, we have \begin{align*} 0 \leq \pi_{(i)}(\bar\omega) &= \Big|\pi_{(i)}(\bar\omega)i\Big|\\ &=\Big|\alpha bq+(1-\alpha)dq_e\Big|\\ &\leq\alpha \lvert b\rvert+(1-\alpha)\lvert d\rvert\\ &=\pi_{(i)}(\bar\omega_i). \end{align*} Assuming that $\alpha |b|-(1-\alpha)|d|\geq 0$, let now $\widetilde{\omega}_i=\alpha\omega_i+(1-\alpha)\omega^*_{e,i}$; otherwise, define $\widetilde{\omega}_i=\alpha\omega_i^*+(1-\alpha)\omega_{e,i}$. Clearly, $\widetilde{\omega}_i\in \mathbb{C}^+$ and $\pi_{(1)}(\widetilde{\omega_i})=\pi_{(1)}(\overline{\omega})$. We have \begin{align*} 0 \leq \pi_{(i)}(\widetilde{\omega}_i) &= \Big|\alpha |b|-(1-\alpha)|d|\Big|\\ &=\Big|\alpha \lvert bq\rvert-(1-\alpha)\lvert dq_e^*\rvert\Big|\\ &\leq \Big|\alpha bq-(1-\alpha)dq_e^*\Big|\\ &= \Big|\alpha bq+(1-\alpha)dq_e\Big|\\ &= \pi_{(i)}(\bar\omega), \end{align*} since $\alpha bq+(1-\alpha)dq_e\in \text{span}\,\{i\}$ and $\overline{\omega}\in\mathbb{C}^+$. Then we have found two elements $\bar \omega_{i}$ and $\tilde \omega_{i}$, both in ${\rm iconv\,}\{B_e(T), B(T)\}$, such that \begin{align*} &\pi_{(1)}(\bar\omega)=\pi_{(1)} \big(\bar \omega_{i}\big)=\pi_{(1)}\big(\tilde \omega_{i}\big)\\ 0\leq &\pi_{(i)}\big(\tilde \omega_{i} \big)\leq \pi_{(i)}(\bar \omega) \leq \pi_{(i)} \big(\bar \omega_{i} \big). \end{align*} Now we will show that $\overline{\omega}$ is also in ${\rm iconv\,}\{B_e(T), B(T)\}$. Consider the affine transformation \begin{align*} & f:B_e(T)\longrightarrow \mathbb{C}\\ &f(z)=\alpha \omega_i+(1-\alpha)z. \, \end{align*} Since $B_e(T)$ is convex, $[\omega^*_{e,i}, \omega_{e,i}]\subset B_e(T)$. Affine transformations map lines into lines so we have \[ f([\omega^*_{e,i}, \omega_{e,i}])= [\widetilde{\omega}_i, \overline{\omega}_i]. \] Observe that $\widetilde{\omega}_i\neq \overline{\omega}_i$, since $\alpha\neq 1$. Since $\overline{\omega}\in [\widetilde{\omega}_i, \overline{\omega}_i]$, there exists $\eta\in [\omega^*_{e,i}, \omega_{e,i}]\subset B_e(T)$ such that $f(\eta)=\overline{\omega},$ that is, $\alpha \omega_i+(1-\alpha)\eta=\overline{\omega}$. We conclude that $\overline{\omega}\in {\rm iconv\,}\{B_e(T), B(T)\}$. \end{proof} \begin{remark}\label{remark_lancaster} In the complex setting \cite[Corollary $1$]{L} proves that the numerical range is closed if and only if the $W_{\mathbb{C},e}(T)$ is a subset of the $W_\mathbb{C}(T)$. The relation in the previous result induces the idea that the same result might hold for quaternions. However, that is not the case. Take the operator $T=\text{diag}\,\{-1+i, 1+i\}\oplus diag\{s_n\}$, where $s_n$ is a sequence that runs over $(-1/2,1/2)i\cap \mathbb{Q}i$. Applying theorem \ref{theoEssNRseq} and theorem \ref{thm convexity}, we have \[B_e(T)=[-i/2,i/2].\] From theorem 4.2 in \cite{CDM6}, it follows \[ \overline{B^+(T)}=\text{conv}\{-1+i,1+i, -1/3, 1/3] \}. \] Nevertheless, the upper bild, and therefore the bild, is not closed. For example, the boundary line segment joining $-1/3$ to $-1+i$ does not belong to $B(T)$. Thus we have $B_e(T)=[-i/2,i/2]\subseteq B(T)$ but $B(T)$ is not closed. \end{remark} \bigskip

https://arxiv.org/abs/1712.03568

On a Detail in Hales's "Dense Sphere Packings: A Blueprint for Formal Proofs"

In "Dense Sphere Packings: A Blueprint for Formal Proofs" Hales proves that for every packing of unit spheres, the density in a ball of radius $r$ is at most $\pi/\sqrt{18}+c/r$ for some constant $c$. When $r$ tends to infinity, this gives a proof to the famous Kepler conjecture. As formulated by Hales, $c$ depends on the packing. We follow the proofs of Hales to calculate a constant $c'$ independent of the sphere packing that exists as mentioned in "A Formal Proof of the Kepler Conjecture" by Hales et al..

\section{Introduction} \label{sec:Introduction} In \cite{Blueprint} Hales proves that for every packing of infinitely many unit spheres into three dimensional space, there exists a constant $c$ such that the \emph{density} inside a ball of radius $r$ is upper bounded by $\pi/\sqrt{18}+c/r$. Here, the density is defined as the volume of the intersection of the packed unit spheres with the container sphere of radius $r$ divided by the volume of the container sphere. The famous Kepler conjecture states that the density tends to $\pi/\sqrt{18}$ when $r$ tends to infinity which is implied by the density bound shown by Hales. To make, for example, statements about bounds for finite packings inside a container, we need $c$ to be independent of the packing, i.e., we want a statement of the form: There exists a constant $c'$ such that for every infinite sphere packing the density inside a ball of radius $r$ is upper bounded by $\pi/\sqrt{18}+c'/r$. If it holds for an infinite packing, then this statement would also hold for a finite packing. First, we give some definitions that are necessary to understand the crucial lemmas from \cite{Blueprint}. Then, we follow the proofs of the lemmas and calculate a constant $c'$ independent of the packing. Instead of using $\mathcal{O}$-Notation as in \cite{Blueprint}, we give more detailed calculations to be able to give an actual value for $c'$. When we cite lemmas or definitions from \cite{Blueprint} (with occasional slight modifications) we give the corresponding number of the lemma or definition in \cite{Blueprint} parenthesized. We try to give definitions as closely as possible to the point where we use them. \section{Main Lemmas} Let us denote by $\vol$ the Lebesgue measure on Euclidean space $\mathbb{R}^3$. In the sequel, let $V \subset \mathbb{R}^3$ be a point set that induces a packing of infinitely many unit spheres, i.e., the points in $V$ are the centers of the spheres and thus have pairwise distance at least 2. Let $\B(\mathbf{p},r)$ be the open sphere centered at point $\mathbf{p}$ with radius $r$. Let $V(\mathbf{p},r) = V \cap \B(\mathbf{p},r)$. The \emph{density} $\density (V,\mathbf{p},r)$ inside a container sphere centered at point $\mathbf{p}$ with radius $r$ is defined by \begin{equation*} \density(V,\mathbf{p},r)= \frac{\vol \left( \B(\mathbf{p},r)\cap \bigcup_{\mathbf{v}\in V} \B(\mathbf{v},1) \right) } {\vol\left( \B(\mathbf{p},r)\right) }. \end{equation*} In \cite{Blueprint}, it is shown that the face-centered cubic (FCC) packing of unit spheres has optimal density when $r$ tends to infinity. In the FCC-packing, the spheres are arranged in layers. In each layer, the spheres are arranged with there centers on a hexagonal grid where neighboring grid points have distance 2. The vertices of the second layer lie above centers of the triangles in the first layer. In the third layer, the vertices lie above centers of triangles in the first and in the second layer. These three layers are then repeated infinitely often. The layers are pushed together until the spheres touch. See the following picture for illustration. \begin{figure}[H] \centering \includegraphics[width = 0.3\textwidth]{Close_packing2.pdf} \caption{Three layers of the FCC-packing} \end{figure} Consider the Voronoi diagram of the circle centers in a FCC-packing. The Voronoi cells are dodecahedra with volume $4\sqrt{2}$ which yields a density of the packing of $\pi/\sqrt{18}$. Let $\voronoi(V,\mathbf{v})$ denote the cell of $\mathbf{v}$ in the Voronoi diagram of $V$ and, more generally, let $\voronoi(V,\underline{\mathbf{u}})$ be the intersection of the Voronoi cells for points in the list $\underline{\mathbf{u}}=[\mathbf{u}_0;\mathbf{u}_1, \dots]$. \begin{newdef}[Definition 6.11 ] A function $\G\colon V\to \mathbb{R}$ on a set $V \subset \mathbb{R}^3$ is \emph{negligible} if there is a constant $c_1$ such that for all $r \geq 1$, \begin{equation*} \sum_{\mathbf{v} \in \B(\mathbf{0},r) \cap V} \G(\mathbf{v}) \leq c_1 r^2. \end{equation*} A function $\G\colon V \to \mathbb{R}$ is \emph{FCC-compatible} if for all $\mathbf{v} \in V$ \begin{equation*} 4\sqrt{2} \leq \vol(\voronoi(V,\mathbf{v}))+\G({\mathbf{v}}). \end{equation*} \label{def:FCCnegl} \end{newdef} \emph{FCC-compatible} means that the volume of every Voronoi cell is close to the volume of those in the FCC-packing. Then, \emph{negligible} means, that the error is small. Now we are ready to look at a lemma in \cite{Blueprint} from which the Kepler conjecture follows under certain assumption. \begin{lem}[Lemma 6.13 from\cite{Blueprint}] If there exists a negligible FCC-compatible function $\G\colon V \to \mathbb{R}$ for a saturated packing $V$, then there exists a constant $c = c(V)$ such that for all $r \geq 1$ \begin{equation*} \density(V,\mathbf{0},r) \leq \frac{\pi}{\sqrt{18}}+\frac{c}{r}. \end{equation*} \label{lem:6.13} \end{lem} Since this lemma is not sufficient to show the Kepler conjecture, two more lemmas are required. Therefore, we need two more definitions. \begin{newdef}[Definition 6.34 ] If $\underline{\mathbf{u}}=\left[\mathbf{u}_0; \dots ; \mathbf{u}_k\right]$ is a list of points in $\mathbb{R}^n$, then let $\h(\underline{\mathbf{u}})$ be the circumradius of its point set $\left\lbrace \mathbf{u}_0; \dots ; \mathbf{u}_k \right\rbrace$. \end{newdef} \begin{newdef}[Definition 6.88 ] Set \begin{equation*} h_0 = 1.26. \end{equation*} Let $\funcL\colon\mathbb{R}\to \mathbb{R}$ be the piecewise linear function \begin{equation*} \funcL(h) = \begin{cases} \frac{h_0-h}{h_0-1} &,h\leq h_0 \\ 0 &,h \geq h_0. \end{cases} \end{equation*} \label{def:L} \end{newdef} We call a packing \emph{saturated} if no unit sphere can be added to the packing. \begin{lem}[Lemma 6.95] For any saturated packing $V$ and any $\mathbf{u}_0 \in V$, \begin{equation} \sum_{\mathbf{u}_1 \in V : \h\left(\mathbf{u}_0,\mathbf{u}_1\right)\leq h_0} \funcL\left(\h\left\lbrace\mathbf{u}_0,\mathbf{u}_1\right\rbrace\right)\leq 12. \label{eq:lem6.95} \end{equation} \label{lem:6.95} \end{lem} We will not discuss the proof of this lemma since it is irrelevant to calculate the constant $c'$ explained in Section~\ref{sec:Introduction}. It is a computer proof anyways. \begin{lem}[Lemma 6.97 ] Inequality~(\ref{eq:lem6.95}) implies that for every saturated packing $V$, there exists a negligible FCC-compatible function $\G\colon V \to \mathbb{R}$. \label{lem:6.97} \end{lem} Lemmas~\ref{lem:6.13},~\ref{lem:6.95}~and~\ref{lem:6.97} together imply the Kepler conjecture. As mentioned above, we are interested in replacing the constant $c$ in Lemma~\ref{lem:6.13} by a constant $c'$ that is independent of the packing $V$. In the proof of Lemma~\ref{lem:6.13} in \cite{Blueprint} it is shown that \begin{equation} \density(V,\mathbf{0},r)\leq \frac{\pi}{\sqrt{18}}\left(1+\frac{3}{r}\right)^3 + c_1\frac{\left(r+1\right)^2}{r^34\sqrt{2}}, \label{eq:lem:6.13} \end{equation} where $c_1$ is the constant from Definition~\ref{def:FCCnegl} for the FCC-compatible negligible function $\G\colon V\to \mathbb{R}$ that exists by assumption. So, what we need to show in order to find $c'$ is the following: There exists a constant $d$ independent of the function $\G$ and therefore of the packing $V$ such that we can set $c_1=d$ in Definition~\ref{def:FCCnegl} and Lemma~\ref{lem:6.97} still holds. In the following, we derive such a constant $d$, reproduce the proof of Lemma~\ref{lem:6.97}, and show that the constructed FCC-compatible function also fulfills the stronger definition of negligible for $c_1=d$. \section{Proof of Lemma~\ref{lem:6.97}} For the following function, we will show that it is FCC-compatible and negligible: \begin{equation} \G(\mathbf{u})=-\vol \left(\voronoi\left(V,\mathbf{u}\right)\right)+8m_1- \sum_{\mathbf{v}\in V\setminus\left\lbrace\mathbf{u}\right\rbrace} 8m_2 \funcL\left(\h\left(\left[\mathbf{u};\mathbf{v}\right]\right)\right) \end{equation} for constants $m_1$ and $m_2$ defined later and the function $\funcL$ as described in Definition~\ref{def:L}. To prove that this function is FCC-compatible and negligible, another lemma is used. For this lemma we need more definitions. In the following, if we talk about cells, we mean so called \emph{Marchal cells}. We refer the reader to Definition 6.51 in \cite{Blueprint} since we will not need most of the definition for the calculations made here. Cells are defined by four points in $V$ and a number $0\leq k \leq 4$ and are denoted by $\cell(\underline{\mathbf{u}},k)$ ("$k$-cell") for $\underline{\mathbf{u}}$ is a list of four points with some extra property to be explained later. 4-cells are always tetrahedra. The cells form a partition of $\mathbb{R}^3$. For a cell $X \neq \emptyset$, let $V(X) = X \cap V$ for a saturated packing $V$ (Definition 6.62 and Lemma 6.63 in \cite{Blueprint}). The following three definitions build upon each other. \begin{newdef}[Definition 3.7] A set $C$ is \emph{r-radial} at center $\mathbf{v}$ if the two conditions $C \subset \B(\mathbf{v},r)$ and $\mathbf{v}+ \mathbf{u} \in C$ imply $\mathbf{v}+t\mathbf{u} \in C$ for all $t$ satisfying $0 < \Vert\mathbf{u}\Vert t < r$. A set $C$ is \emph{eventually radial} at center $\mathbf{v}$ if $C \cap \B(\mathbf{v},r)$ is r-radial at center $\mathbf{v}$ for some $r>0$. \end{newdef} \begin{newdef}[Definition 3.11 ] When $C$ is measurable and eventually radial at center $\mathbf{v}$, define the \emph{solid angle} of $C$ at $\mathbf{v}$ to be \begin{equation*} \sol(C,\mathbf{v}) = 3\frac{\vol\left(C \cap \B\left(\mathbf{v},r\right)\right)}{r^3}, \end{equation*} where $r$ is as in the definition of eventually radial. By Lemma 3.10 in \cite{Blueprint}, this yields the same value when replacing r by $r'$ for any $0 \le r'\le r$. \end{newdef} \begin{newdef}[Definition 6.66] Define the \emph{total solid angle} of a cell $X$ to be \begin{equation*} \tsol(X)=\sum_{\mathbf{v}\in V(X)} \sol(X,\mathbf{v}). \end{equation*} \end{newdef} We will use this definition later. The following seven definitions build upon each other. \begin{newdef}[Definition 5.43] Recall that a set $A \subset \mathbb{R}^n$ is \emph{affine} if for every $\mathbf{v}, \mathbf{w} \in A$ and every $t \in \mathbb{R}$, \begin{equation*} t\mathbf{v} + \left( 1-t\right)\mathbf{w} \in A. \end{equation*} Recall that the \emph{affine hull} of $P \subset \mathbb{R}^n$ (denoted $\aff(P)$) is the smallest affine set containing $P$. The \emph{affine dimension} of $P$ (written $\dimaff (P)$ is $\card (S)-1$, where $S$ is a set of smallest cardinality such that $P \subset \aff(S)$. In particular, the affine dimension of the empty set is $-1$. [...] \end{newdef} \begin{newdef}[Definition 6.18] Let $V$ be a saturated packing. When $k=0,1,2,3$, let $\underline{V}(k)$ be the set of lists $\underline{\mathbf{u}}=\left[\mathbf{u}_0; \dots; \mathbf{u}_k\right]$ of length $k+1$ with $\mathbf{u}_i \in V$ such that \begin{equation*} \dimaff \left(\Omega\left(V,\left[\mathbf{u}_0; \dots; \mathbf{u}_j\right]\right)\right)= 3-j \end{equation*} for all $0<j\leq k$. Set $\underline{V}(k)=\emptyset$ for $k >3$. \label{def:6.18} \end{newdef} \begin{newdef}[Definition 6.24] Let $V$ be a saturated packing and let $\underline{\mathbf{u}}=\left[\mathbf{u}_0; \dots ;\mathbf{u}_k\right] \in \underline{V}(k)$ for some $k$. Define points $\omega_j = \omega_j(V, \underline{\mathbf{u}}) \in \mathbb{R}^3$ by recursion over $j \le k$ \begin{align*} \omega_0 &= \mathbf{u}_0\\ \omega_{j+1} &= \text{ the closest point to $\omega_j$ on $\voronoi\left(V, \left[\mathbf{u}_0; \dots ;\mathbf{u}_{j+1}\right]\right)$}, \end{align*} Set $w(V,\underline{\mathbf{u}}) =\omega_k(V,\underline{\mathbf{u}})$, when $\underline{\mathbf{u}} \in \underline{V}(k)$. The set $V$ is generally fixed and is dropped from the notation. \end{newdef} Let $\conv\lbrace \mathbf{v}_0,\dots, \mathbf{v_n}\rbrace$ denote the convex hull of the point set $\lbrace \mathbf{v}_0,\dots, \mathbf{v_n}\rbrace$. \begin{newdef}[Definition 6.51] Let $V$ be a saturated packing. Let $\underline{\mathbf{u}}=\left[\mathbf{u}_0; \dots ; \mathbf{u}_3\right] \in \underline{V}(3)$. Define $\xi(\underline{\mathbf{u}})$ as follows. If $\sqrt{2} \le \h\left(\left[\mathbf{u}_0;\dots ;\mathbf{u}_2\right]\right)$, then let $\xi(\underline{\mathbf{u}})=\omega\left(\left[\mathbf{u}_0;\dots ;\mathbf{u}_2\right]\right)$. If $\h\left(\left[\mathbf{u}_0;\dots ;\mathbf{u}_2\right]\right) < \sqrt{2} \le \h(\underline{\mathbf{u}})$, define $\xi(\underline{\mathbf{u}})$ to be the unique point in \begin{equation*} \conv \left\lbrace \omega\left(\left[\mathbf{u}_0;\dots ;\mathbf{u}_2\right]\right), \omega\left(\underline{\mathbf{u}}\right)\right\rbrace \end{equation*} at distance $\sqrt{2}$ from $\mathbf{u}_0$. [...] \end{newdef} \begin{newdef}[Definition 2.66] When $\mathbf{v}_0 \neq \mathbf{v}_1$, write $\dih_V\left(\left\lbrace\mathbf{v}_0, \mathbf{v}_1\right\rbrace,\left\lbrace\mathbf{v}_2, \mathbf{v}_3\right\rbrace\right)$ for the angle $\gamma \in \left[0,\pi\right]$ formed by \begin{align*} \overline{\mathbf{w}}_2 &= (\mathbf{w}_1 \cdot \mathbf{w}_1) \mathbf{w}_2 - (\mathbf{w}_1 \cdot \mathbf{w}_2)\mathbf{w}_1 \text{ and}\\ \overline{\mathbf{w}}_3 &= (\mathbf{w}_1 \cdot \mathbf{w}_1) \mathbf{w}_3 - (\mathbf{w}_1 \cdot \mathbf{w}_3)\mathbf{w}_1, \end{align*} where $\mathbf{w}_i = \mathbf{v}_i - \mathbf{v}_0$. \end{newdef} \begin{newdef}[Definition 6.67] Let $E(X)$ be the set of \emph{extremal edges} of the $k$-cell $X$ in a saturated packing $V$. More precisely, let \begin{equation*} E(X) = \left\lbrace\left\lbrace\mathbf{u}_i, \mathbf{u}_j\right\rbrace : \mathbf{u}_i \neq \mathbf{u}_j \in V(X) \right\rbrace. \end{equation*} In particular, $E(X)$ is empty for $0$ and $1$-cells and contains $\binom{k}{2}$ pairs when $2\le k \le 4$. \label{def:6.67} \end{newdef} \begin{newdef}[Definition 6.68] Let $V$ be a saturated packing. Let $X$ be a $k$-cell, where $2\le k \le 4$. Let $\varepsilon \in E(X)$. We define the dihedral angle $\dih(X, \varepsilon)$ of $X$ along $\varepsilon$ as follows. Explicitly, if $X$ is a null set, then set $\dih(X,\varepsilon)=0$. Otherwise, choose $\underline{\mathbf{u}}=[\mathbf{u}_0;\mathbf{u}_1;\mathbf{u}_2;\mathbf{u}_3] \in \underline{V}(3)$ such that $X=cell(\underline{\mathbf{u}},k)$ and $\varepsilon = \lbrace \mathbf{u}_0, \mathbf{u}_1\rbrace$. Set $\dih(X,\varepsilon)=\dih_V\left(\left\lbrace \mathbf{u}_0, \mathbf{u}_1\right\rbrace, \left\lbrace\mathbf{v}, \mathbf{w}\right\rbrace\right)$, where \begin{align*} \left\lbrace\mathbf{v},\mathbf{w}\right\rbrace = \begin{cases} \left\lbrace\xi(\underline{\mathbf{u}}), \omega(\underline{\mathbf{u}})\right\rbrace & ,k=2\\ \left\lbrace\mathbf{u}_2,\xi(\underline{\mathbf{u}})\right\rbrace & ,k=3\\ \left\lbrace\mathbf{u}_2, \mathbf{u}_3\right\rbrace &,k=4. \end{cases} \end{align*} This is independent of the choice of $\underline{\mathbf{u}}$ defining $X$. \end{newdef} Now, we need just two more definitions before we can state the lemma mentioned above. The second one builds upon all the previously given definitions. \begin{newdef}[Definition 6.70] Define the following constants [...]: \begin{align*} \sol_0 &= 3\arccos\left(\frac{1}{3}\right)-\pi\\ \tau_0 &= 4\pi - 20\sol_0\\ m_1 &= \sol_0 2 \frac{\sqrt{2}}{\tau_0} \approx 1.012\\ m_2 & = \left(6\sol_0 - \pi\right)\frac{\sqrt{2}}{6\tau_0} \approx 0.0254\\ \end{align*} [...] \end{newdef} \begin{newdef}[Definition 6.79] For any cell $X$ of a saturated packing, define the function $\gamma(X,\ast)$ on $\left\lbrace f \colon \mathbb{R} \to \mathbb{R}\right\rbrace$ by \begin{equation*} \gamma(X,f) = \vol\left(X\right) - \left(\frac{2m_1}{\pi}\right)\tsol\left(X\right) + \left(\frac{8m_2}{\pi}\right)\sum_{\varepsilon \in E(X)} \dih\left(X,\varepsilon\right)f\left(h\left(\varepsilon\right)\right). \end{equation*} \end{newdef} Now, we can state the lemma. \begin{lem}[Lemma 6.86] Let $f$ be any bounded, compactly supported function. Set \begin{equation*} \G\left(\mathbf{u}_0,f\right)=-\vol\left(\voronoi\left(V,\mathbf{u}_0\right)\right)+8m_1 - \sum_{\mathbf{u}\in V \setminus \left\lbrace \mathbf{u}_0\right\rbrace} 8m_2f\left(\h\left(\left[\mathbf{u}_0;\mathbf{u}\right]\right)\right). \end{equation*} If \begin{equation*} \sum_{\mathbf{v}\in V \setminus \left\lbrace\mathbf{u}\right\rbrace} f\left(\h\left(\left[\mathbf{u};\mathbf{v}\right]\right)\right) \le 12, \end{equation*} then $\G(\ast,f)$ is FCC-compatible. Moreover, if there exists a constant $c_0$ such that for all $r \ge 1$ \begin{equation*} \sum_{X \subset \B\left(\mathbf{0},r\right)} \gamma\left(X,f\right) \ge c_0 r^2, \end{equation*} then $\G(\ast,f)$ is negligible. \footnote{In the sequel, by slight abuse of notation $\sum_{X \subset \B(\mathbf{0},r)}$ refers to the summation only over all Marchal cells $X \subset \B(\mathbf{0},r)$.} \label{lem:6.86} \end{lem} This Lemma almost implies that there is a FCC-compatible negligible function. Using this Lemma, it is sufficient to show that Lemma~\ref{lem:6.95} implies the existence of suitable $\G$ and $f$ in Lemma~\ref{lem:6.86} to prove Lemma~\ref{lem:6.97}. We will turn to this later. Now, we focus on the proof of Lemma~\ref{lem:6.86}. Since we are only interested in the constant in the definition of negligible, we will only outline this part of the proof here. The idea of the proof of negligibility is as follows. If one can show that \begin{align} -\sum_{\mathbf{u} \in V(\mathbf{0},r)}\G\left(\mathbf{u},f\right) &\geq \sum_{X \subset \B(\mathbf{0},r)}\gamma\left(X,f\right) + c_2 r^2\label{eq:GandGamma}\\ \intertext{for some constant $c_2$. Then by the assumption in Lemma~\ref{lem:6.86}}\nonumber\\ -\sum_{\mathbf{u} \in V(\mathbf{0},r)}\G\left(\mathbf{u},f\right) &\geq \left(c_0 +c_2\right)r^2,\label{eq:negligible} \end{align} which directly implies, that $\G(\ast,f)$ is negligible for $c_1 = -(c_0 + c_2)$ in Definition~\ref{def:FCCnegl}. Since we are only interested in the independence of $c_1$ of the packing $V$, we will now focus on calculating the constant $c_2$ and thereby showing that $c_2$ is independent of the packing. We will get an estimate on $c_0$ later satisfying the conditions in Lemma~\ref{lem:6.86}. \subsection{Calculation of $c_2$} In this section we go through the proof of Lemma~\ref{lem:6.86} and show that \begin{equation*} -\sum_{\mathbf{u} \in V(\mathbf{0},r)}\G\left(\mathbf{u},f\right) \geq \sum_{X \subset \B(\mathbf{0},r)}\gamma\left(X,f\right) + c_2 r^2 \end{equation*} for some constant $c_2$ independent of the packing $V$. Observe that the equalities \begin{align*} &-\sum_{\mathbf{u} \in V(\mathbf{0},r)}\G\left(\mathbf{u},f\right) = \\ &\sum_{\mathbf{u} \in V(\mathbf{0},r)}\vol\left(\voronoi\left(V,\mathbf{u}\right)\right)-\sum_{\mathbf{u} \in V(\mathbf{0},r)}8m_1 + \sum_{\mathbf{u} \in V(\mathbf{0},r)}\sum_{\mathbf{v}\in V \setminus \left\lbrace \mathbf{u}\right\rbrace} 8m_2f\left(\h\left(\left[\mathbf{u};\mathbf{v}\right]\right)\right) \end{align*} and \begin{align*} &\sum_{X \subset \B(\mathbf{0},r)}\gamma\left(X,f\right) = \\ &\sum_{X \subset \B(\mathbf{0},r)}\vol\left(X\right) - \sum_{X \subset \B(\mathbf{0},r)}\left(\frac{2m_1}{\pi}\right)\tsol\left(X\right) + \sum_{X \subset \B(\mathbf{0},r)}\left(\frac{8m_2}{\pi}\right)\sum_{\varepsilon \in E(X)} \dih\left(X,\varepsilon\right)f\left(h\left(\varepsilon\right)\right) \end{align*} have each three summands and we will relate them in this order. So, we start by showing that \begin{align} \sum_{\mathbf{u} \in V(\mathbf{0},r)}\vol\left(\voronoi\left(V,\mathbf{u}\right)\right) \geq \sum_{X \subset \B(\mathbf{0},r)}\vol\left(X\right) - \frac{56}{3}\pi r^2.\label{eq:GandGammFirst} \end{align} By Lemma 6.7 in \cite{Blueprint}, $\voronoi(V,\mathbf{v}) \subset \B(\mathbf{v},2)$, so we have \begin{align*} \sum_{\mathbf{u} \in V(\mathbf{0},r)}\vol\left(\voronoi\left(V,\mathbf{u}\right)\right) & \geq \vol\left(\B\left(\mathbf{0}, r-2\right)\right)\\ &=\vol\left(\B\left(\mathbf{0},r\right)\right) - \vol\left(\B\left(\mathbf{0},r\right)\setminus\B\left(\mathbf{0},r-2\right)\right)\\ &\geq \sum_{X \subset \B\left(\mathbf{0},r\right)}\vol\left(X\right) - \left(\frac{4}{3}\pi r^3 - \frac{4}{3}\pi\left(r-2\right)^3\right),\\ \intertext{since Marchal cells form a partition of space as mentioned above. So we get } \sum_{\mathbf{u} \in V(\mathbf{0},r)}\vol\left(\voronoi\left(V,\mathbf{u}\right)\right) & \geq \sum_{X \subset \B\left(\mathbf{0},r\right)}\vol\left(X\right) - 8 \pi r^2 - \frac{32}{3}\pi.\\ \intertext{Since negligible is only defined for $r\geq 1$, we can assume $r\geq 1$ here. So we have}\\ \sum_{\mathbf{u} \in V(\mathbf{0},r)}\vol\left(\voronoi\left(V,\mathbf{u}\right)\right) &\geq \sum_{X \subset \B(\mathbf{0},r)}\vol\left(X\right) - \frac{56}{3}\pi r^2, \end{align*} which proves (\ref{eq:GandGammFirst}). Next, we will show that \begin{align} -\sum_{\mathbf{u}\in V(\mathbf{0},r)} 8 m_1 \ge -\left(\frac{2m_1}{\pi}\right)\sum_{X \subset \B(\mathbf{0},r)}\tsol(X)-m_1 \cdot 2240 r^2.\label{eq:GandGammaSecond} \end{align} First, we need to estimate the diameter of a Marchal cell. For a cell $X=\cell(\underline{\mathbf{u}},k)$ for $\underline{\mathbf{u}}=[\mathbf{u}_0;\dots] \in \underline{V}(3)$ it holds that \begin{align*} X = \cell(\underline{\mathbf{u}},k) &\subset \voronoi(V,\mathbf{u}_0) \cup \dots \cup \voronoi(V,\mathbf{u}_{k-1})\\ \intertext{(see the proof of Lemma 6.63 in \cite{Blueprint}). By Lemma 6.7, $\voronoi(V,\mathbf{v}) \subset \B(\mathbf{v},2)$, so we get} X &\subset \B(\mathbf{u}_0,2) \cup \dots \cup \B(\mathbf{u}_{k-1},2). \end{align*} By Definition~\ref{def:6.18}, the Voronoi cells $\voronoi(V,\mathbf{u}_0), \dots, \voronoi(V,\mathbf{u}_{k-1})$ share at least one point and so the spheres $\B(\mathbf{u}_0,2), \dots, \B(\mathbf{u}_{k-1},2)$ do. Therefore, we can conclude that the diameter of a Marchal cell is at most 4. Now, \begin{align*} \sum_{X\subset \B(\mathbf{0},r)}\tsol(X) &= \sum_{X\subset \B(\mathbf{0},r)}\sum_{\mathbf{v}\in V(X)} \sol(X,\mathbf{v})\\ \intertext{by definition. Since the diameter of each cell is at most 4, we can rewrite this as} \sum_{X\subset \B(\mathbf{0},r)}\tsol(X)&= \sum_{X \subset \B(\mathbf{0},r+4)}\sum_{\mathbf{u}\in V(X)}\sol(X,\mathbf{u}) - \sum_{X \subset \B(\mathbf{0},r+4):X \nsubseteq \B(\mathbf{0},r)}\sum_{\mathbf{u}\in V(X)}\sol(X,\mathbf{u})\\ &\ge \sum_{\mathbf{v}\in V(\mathbf{0},r)}\sum_{X:\mathbf{v}\in V(X)}\sol\left(X,\mathbf{v}\right) - \sum_{\mathbf{u}\in V(\mathbf{0},r+4)\setminus V(\mathbf{0},r-4)}\sum_{X:\mathbf{u}\in V(X)} \sol(X,\mathbf{u}),\\ \intertext{since $V(X) = X \cap V$ and a cell with diameter at most 4 that is not completely contained in a sphere of radius $r$ cannot intersect the sphere with the same center of radius $r-4$. Next, we use that the solid angles around one point sum up to $4\pi$. Furthermore, to estimate the number of points $\mathbf{u}\in V(\mathbf{0},r+4)\setminus V(\mathbf{0},r-4)$, we use the following volume argument. Each $\mathbf{u} \in V(\mathbf{0},r+4)\setminus V(\mathbf{0},r-4)$ is the center of a packed unit sphere and therefore the volume of those unit spheres sum up to less than the volume of $\B(\mathbf{0},r+5)\setminus \B(\mathbf{0},r-5)$. Vice versa dividing $\vol(\B(\mathbf{0},r+5)\setminus \B(\mathbf{0},r-5))$ by the volume of a unit sphere yields an upper bound on the number of points, so we get} \sum_{X\subset \B(\mathbf{0},r)}\tsol(X) &\geq \sum_{\mathbf{v}\in V(\mathbf{0},r)} 4\pi - \left(\frac{\frac{4}{3}\pi\left(r+5\right)^3-\frac{4}{3}\pi\left(r-5\right)^3}{\frac{4}{3}\pi}\right)4\pi\\ &= \sum_{\mathbf{v}\in V(\mathbf{0},r)} 4\pi - (30r^2 + 250)4\pi\\ &\ge \sum_{\mathbf{v}\in V(\mathbf{0},r)} 4\pi - 1120\pi r^2,\\ \intertext{since $r\geq 1$.} \end{align*} By rearranging, we obtain \begin{equation*} -\left(\frac{2m_1}{\pi}\right)\sum_{X\subset\B(\mathbf{0},r)}\tsol(X) - 2240m_1 r^2 \le -\sum_{\mathbf{v}\in V(\mathbf{0},r)} 8m_1, \end{equation*} proving (\ref{eq:GandGammaSecond}). Finally, we show that \begin{equation} \sum_{\mathbf{u}\in V(\mathbf{0},r)}\sum_{\mathbf{v}\in V \setminus \left\lbrace \mathbf{u}\right\rbrace} 8m_2f\left(\h\left(\left[\mathbf{u};\mathbf{v}\right]\right)\right) \ge \frac{8m_2}{\pi}\sum_{X\in\B(\mathbf{0},r)}\sum_{\varepsilon\in E(X)} \dih(X,\varepsilon)f\left(\h\left(\varepsilon\right)\right).\label{eq:GandGammThird} \end{equation} For any $\varepsilon \in E(X)$ it holds that $\varepsilon \subset V(X) \subset X$ by Definition~\ref{def:6.67}, and Lemma~6.63 in \cite{Blueprint}. This implies \begin{align*} \sum_{X\subset\B(\mathbf{0},r)}\sum_{\varepsilon\in E(X)} \dih(X,\varepsilon)f\left(\h\left(\varepsilon\right)\right) &\le \sum_{\varepsilon=\lbrace\mathbf{u},\mathbf{v}\rbrace \subset\B(\mathbf{0},r)}\sum_{X:\varepsilon\in E(X)} \dih\left(X,\varepsilon\right)f\left(\h\left(\varepsilon\right)\right)\\ &= \sum_{\varepsilon \subset \B(\mathbf{0},r)}2\pi f\left(\h\left(\varepsilon\right)\right),\\ \intertext{since the dihedral angle around an edge sums up to $2\pi$.\footnotemark When summing over ordered pairs, each edge appears twice, so we have to divide by two and get} \sum_{X\subset\B(\mathbf{0},r)}\sum_{\varepsilon\in E(X)} \dih(X,\varepsilon)f\left(\h\left(\varepsilon\right)\right) &\le \sum_{\mathbf{u}\in V(\mathbf{0},r)}\sum_{\mathbf{v}\in V(\mathbf{0},r)\setminus \lbrace \mathbf{u}\rbrace} \pi f\left(\h\left(\mathbf{u},\mathbf{v}\right)\right). \end{align*} \footnotetext{Here, we follow the argumentation of Hales \cite{Blueprint}. The author noticed that in Lemma 6.69 in \cite{Blueprint}, it is required that $\h(\varepsilon)<\sqrt{2}$ to have that the dihedral angles sum up to $2\pi$. Since Lemma~\ref{lem:6.86} is only used for $f=\funcL$ and $\funcL(\h(\varepsilon))=0$ for $\h(\varepsilon)\ge\sqrt{2}$, this does not cause any problems.} So, for the last summand we get by rearrangement \begin{equation*} \frac{8m_2}{\pi} \sum_{X\in\B(\mathbf{0},r)}\sum_{\varepsilon\in E(X)} \dih(X,\varepsilon)f\left(\h\left(\varepsilon\right)\right) \le \sum_{\mathbf{u}\in V(\mathbf{0},r)}\sum_{\mathbf{v}\in V(\mathbf{0},r)\setminus \lbrace \mathbf{u}\rbrace} 8m_2 f\left(\h\left(\mathbf{u},\mathbf{v}\right)\right), \end{equation*} showing (\ref{eq:GandGammThird}). Combining (\ref{eq:GandGammFirst}), (\ref{eq:GandGammaSecond}), and (\ref{eq:GandGammThird}), we have \begin{align*} &-\G\left(\mathbf{u},f\right) \\ &=\sum_{\mathbf{u} \in V(\mathbf{0},r)}\vol\left(\voronoi\left(V,\mathbf{u}\right)\right)-\sum_{\mathbf{u} \in V(\mathbf{0},r)}8m_1 + \sum_{\mathbf{u} \in V(\mathbf{0},r)}\sum_{\mathbf{v}\in V \setminus \left\lbrace \mathbf{u}\right\rbrace} 8m_2f\left(\h\left(\left[\mathbf{u};\mathbf{v}\right]\right)\right) \\ &\ge\sum_{X \subset \B(\mathbf{0},r)}\vol\left(X\right) - \frac{56}{3}\pi r^2 -\left(\frac{2m_1}{\pi}\right)\sum_{X \subset \B(\mathbf{0},r)}\tsol(X)-m_1 \cdot 2240 r^2 + \\ &\frac{8m_2}{\pi}\sum_{X\in\B(\mathbf{0},r)}\sum_{\varepsilon\in E(X)} \dih(X,\varepsilon)f\left(\h\left(\varepsilon\right)\right)\\ &=\sum_{X \subset\B(\mathbf{0},r)}\gamma(X,f) - \left(\frac{56}{3}+m_1\cdot 2240\right)r^2, \intertext{i.e. (\ref{eq:GandGamma}) holds with $c_2 = -\frac{56}{3}-m_1\cdot 2240$.} \end{align*} As mentioned above, we only need to show that the conditions in Lemma~\ref{lem:6.86} hold for some $f$ to prove Lemma~\ref{lem:6.97}. The first condition holds for $f=\funcL$ by Lemma~\ref{lem:6.95}. It remains to show that the second condition also holds for $f=\funcL$, namely $\sum_{X \subset \B\left(\mathbf{0},r\right)} \gamma\left(X,\funcL\right) \ge c_0 r^2$. Then, we will have showed that the function $\G(\ast,\funcL)$ is FCC-compatible and negligible for $c_1=-(c_0 + c_2)$ in Definition~\ref{def:FCCnegl}. For $c_2$ we already showed that it is independent of the packing, now, we will do so for $c_0$, i.e. prove that $\sum_{X \subset \B\left(\mathbf{0},r\right)} \gamma\left(X,\funcL\right) \ge c_0 r^2$ holds for a $c_0$ independent of the packing. \subsection{Calculation of $c_0$} In this section, we will show that \begin{equation} \sum_{X \subset \B\left(\mathbf{0},r\right)} \gamma\left(X,f\right) \ge c_0 r^2\label{eq:Gamma} \end{equation} holds for $f=\funcL$ for a constant $c_0$ independent of the packing that defines the cells $X$. We will need more definitions for this proof and we will try to give them as closely as possible to the point where they are used. \begin{newdef}[Definition 6.70 and 6.88] Set \begin{align*} h_+ &= 1.3254.\\ \intertext{Let $\M \colon \mathbb{R} \to \mathbb{R}$ be the piecewise polynomial function} \M (h)&= \begin{cases} \frac{\sqrt{2}-h}{\sqrt{2}-1}\frac{h_+-h}{h_+-1}\frac{17h-9h^2-3}{5} &,h \le \sqrt{2}\\ 0 &,h>\sqrt{2}. \end{cases}\\ \intertext{Let $h_-\approx1.23175$ be the unique root of the quadratic polynomial $\M(h)-\funcL(h)$ that lies in the interval $[1.231,1.232]$.} \end{align*} \end{newdef} \begin{newdef}[Definition 6.89] A \emph{critical edge} $\varepsilon$ of a saturated packing $V$ is an unordered pair that appears as an element of $E(X)$ for some $k$-cell $X$ of the packing $V$ such that $h(\varepsilon) \in [h_-,h_+]$. Let $\EC(X)$ be the set of critical edges that belong to $E(X)$. If $X$ is any cell such that $\EC(X)$ is not empty, let the \emph{weight} $\wt(X)$ of $X$ be $1/\operatorname{card}(\EC(X))$. \label{def:critEdge} \end{newdef} Now, we can start with estimating $\sum_{X \subset \B(\mathbf{0},r)} \gamma(X,\funcL)$. \begin{align} \sum_{X \subset \B(\mathbf{0},r)} \gamma(X,\funcL) &= \sum_{X \subset \B(\mathbf{0},r) : \EC(X) \neq \emptyset} \gamma(X,\funcL) + \underbrace{\sum_{X \subset \B(\mathbf{0},r): \EC(X) = \emptyset} \gamma(X,\funcL)}_{\geq 0 \text{ by Lemma 6.92 in \cite{Blueprint}}}\nonumber\\ &\ge \sum_{X \subset \B(\mathbf{0},r) : \EC(X) \neq \emptyset} \gamma(X,\funcL)\nonumber\\ &= \sum_{X \subset \B(\mathbf{0},r) : \EC(X) \neq \emptyset} \left(\gamma(X,\funcL) \underbrace{\sum_{\varepsilon \in \EC(X)}\wt(X)}_{=1 \text{ by Definition~\ref{def:critEdge}}}\right)\nonumber\\ &= \sum_{X \subset \B(\mathbf{0},r) : \EC(X) \neq \emptyset}\sum_{\varepsilon \in \EC(X)}\gamma(X,\funcL)\wt(X)\nonumber\\ &= \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X:\varepsilon \in \EC(X)} \gamma(X,\funcL)\wt(X) - \underbrace{\sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X:\varepsilon \in \EC(X), X \nsubseteq \B(\mathbf{0},r)} \gamma(X,\funcL)\wt(X)}_{=:\alpha}\label{eq:gammaIntermediate} \end{align} Next, we will show that the term $\alpha$ is upper bounded by a constant~times~$r^2$. \begin{align} \alpha &= \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X:\varepsilon \in \EC(X), X \nsubseteq \B(\mathbf{0},r)} \gamma(X,\funcL)\wt(X)\nonumber\\ &\le \sum_{X \subset \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4)} \sum_{\varepsilon \in \EC(X)} \gamma(X,\funcL)\wt(X),\label{eq:1} \intertext{since the diameter of $X$ is at most 4 and $X$ intersects the boundary of $\B(\mathbf{0},r)$. By (\ref{eq:1}),} \alpha &\le \sum_{X \subset \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4)}\left(\gamma(X,\funcL) \underbrace{\sum_{\varepsilon \in \EC(X)} \wt(X)}_{=1 \text{ by Definition~\ref{def:critEdge}}}\right).\label{eq:2} \intertext{Next, we plug in the definition of $\gamma(X,L)$. By (\ref{eq:2}),} \alpha &\le\sum_{X \subset \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4)} \left(\vol\left(X\right) - \underbrace{\left(\frac{2m_1}{\pi}\right)\tsol\left(X\right)}_{\ge 0 \text{ by definition}} + \left(\frac{8m_2}{\pi}\right)\sum_{\varepsilon \in E(X)} \dih\left(X,\varepsilon\right)\funcL\left(h\left(\varepsilon\right)\right)\right)\nonumber\\ &\le \sum_{X \subset \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4)} \vol\left(X\right)+ \sum_{X \in \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4)}\left(\frac{8m_2}{\pi}\right)\sum_{\varepsilon \in E(X)} \dih\left(X,\varepsilon\right)\funcL\left(h\left(\varepsilon\right)\right).\label{eq:3} \intertext{Since Marchal cells are a partition of the space, the volume of the cells contained in a spherical shell sums up to at most the volume of the spherical shell. Furthermore, it holds for any $\varepsilon \in E(X)$ that $\varepsilon \subset V(X) \subset X$, so we get from (\ref{eq:3})} \alpha &\le \frac{4}{3}\pi\left(r+4\right)^3-\frac{4}{3}\pi\left(r-4\right)^3 + \frac{8m_2}{\pi}\sum_{\varepsilon\subset \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4)}\sum_{X:\varepsilon\in E(X)}\dih\left(X,\varepsilon\right)\funcL\left(h\left(\varepsilon\right)\right).\label{eq:4} \intertext{Since the dihedral angle for edges $\varepsilon$ with $\h(\varepsilon)<\sqrt{2}$ sums up to $2\pi$ (see Lemma~6.69 in \cite{Blueprint}) and for $\h(\varepsilon)\ge \sqrt{2}$ the factor $\funcL(\h(\varepsilon))=0$, we obtain from (\ref{eq:4})} \alpha &\le 32\pi r^2 + \frac{512}{3} \pi + \sum_{\varepsilon\subset \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4)}16m_2\funcL\left(h\left(\varepsilon\right)\right).\label{eq:5} \intertext{Now, we want to estimate the number of edges such that $\funcL(\h(\varepsilon)) > 0$. First, we give an upper bound on the number of points in $V$ inside the spherical shell as follows. By decreasing the inner radius and increasing the outer radius by 1, all unit spheres packed with centers in the original shell are completely contained in the enlarged shell. Then, we divide the volume of the enlarged shell by the volume of a unit sphere. Next, we want to give an upper bound on the number of points in $V$ with distance at most 2.52 from a given point, since for longer edges $\funcL(\h(\varepsilon))=0$. We do this in a similar way as before by a volume argument. By multiplying these two values, we count each edge twice, so we have to divide by 2. In addition, the function $\funcL$ is upper bounded by $1.26/0.26$. So, by (\ref{eq:5})} \alpha&\le 32\pi r^2 + \frac{512}{3} \pi +\underbrace{\frac{\frac{4}{3}\pi\left(r+5\right)^3-\frac{4}{3}\pi\left(r-5\right)^3}{\frac{4}{3}\pi}}_{\ge | V \cap \left(\B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4)\right)|}\cdot\underbrace{\frac{\frac{4}{3}\pi \cdot 3.52^3}{\frac{4}{3}\pi}}_{\ge |V \cap \B(\mathbf{v},2.52)| \text{ for any }\mathbf{v}}\cdot \frac{1}{2}\cdot16m_2\cdot\frac{126}{26}\nonumber\\ &\le 32\pi r^2 + \frac{512}{3}\pi + \left(30r^2 + 250\right)\cdot 3.52^3\cdot 8 \cdot \frac{63}{13}\cdot 0.0255\nonumber\\ &\le 1394.1 r^2 + 11315.6\nonumber \intertext{As mentioned before, we can assume $r\ge 1$ and obtain} \alpha&\le 12710 r^2.\label{eq:alpha} \end{align} Now, we know from inequalities (\ref{eq:gammaIntermediate}) and (\ref{eq:alpha}) that \begin{align} \sum_{X \subset\B(\mathbf{0},r)} \gamma(X,L) \ge \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X:\varepsilon \in \EC(X)} \gamma(X,L)\wt(X) - 12710^2.\label{eq:gammaIntermediate2} \end{align} To proceed with the estimation, we need two more definitions. \begin{newdef}[Definition 6.90] Set \begin{equation*} \beta_0(h) = 0.005\left(1-\frac{\left(h- h_0\right)^2}{\left(h_+ - h_0\right)^2}\right). \end{equation*} If $X$ is a 4-cell with exactly two critical edges and if those edges are opposite, then set \begin{equation*} \beta(\varepsilon,X) = \beta_0\left(\h\left(\varepsilon\right)\right) - \beta_0\left(\h\left(\varepsilon'\right)\right), \text{ where } \EC(X) = \lbrace\varepsilon,\varepsilon'\rbrace. \end{equation*} Otherwise, for all other edges in all other cells, set $\beta(\varepsilon,X)=0$. \end{newdef} \begin{newdef}[Definition 6.91] Let $V$ be a saturated packing. Let $\varepsilon \in \EC(X)$ be a critical edge of a $k$-cell $X$ of $V$ for some $2\le k\le 4$. A \emph{cell cluster} is the set \begin{equation*} \CL(\varepsilon) = \left\lbrace X : \varepsilon \in \EC(X)\right\rbrace \end{equation*} of all cells around $\varepsilon$. Define \begin{equation*} \Gamma(\varepsilon) = \sum_{X \in \CL(\varepsilon)} \gamma(X,L)\wt(X) + \beta(\varepsilon,X). \end{equation*} \end{newdef} Using these definitions, we can rewrite inequality (\ref{eq:gammaIntermediate2}) as follows. \begin{align} \sum_{X \subset\B(\mathbf{0},r)} \gamma(X,L) &\ge \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X:\varepsilon \in \EC(X)} \gamma(X,L)\wt(X) - 12710^2\nonumber\\ &=\sum_{\varepsilon \subset \B(\mathbf{0},r)} \left(\Gamma(\varepsilon) - \sum_{X \in \CL(\varepsilon)}\beta(\varepsilon,X)\right) - 12710r^2\nonumber\\ &=\sum_{\varepsilon \subset \B(\mathbf{0},r)} \Gamma(\varepsilon) - \underbrace{\sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X \in \CL(\varepsilon)}\beta(\varepsilon,X)}_{=: \zeta}- 12710r^2\label{eq:gammaIntermediate3} \end{align} Next, we will upper bound the term $\zeta$. \begin{align} \zeta &= \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X \in \CL(\varepsilon)}\beta(\varepsilon,X)\nonumber\\ &= \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X \in \CL(\varepsilon): X \subset \B(\mathbf{0},r)}\beta(\varepsilon,X) + \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X \in \CL(\varepsilon):X \nsubseteq \B(\mathbf{0},r)}\beta(\varepsilon,X)\nonumber\\ &= \sum_{X \subset \B(\mathbf{0},r)}\sum_{\varepsilon \in \EC(X)}\beta(\varepsilon,X) + \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{X \in \CL(\varepsilon):X \nsubseteq \B(\mathbf{0},r)}\beta(\varepsilon,X)\nonumber\\ &= 0+\sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{\substack{X \in \CL(\varepsilon):X \nsubseteq \B(\mathbf{0},r)\\ \wedge X \text{ is 4-cell with}|\EC(X)|=2}}\beta(\varepsilon,X),\label{eq:6} \intertext{since $\beta(\varepsilon,X) > 0$ only for 4-cells with exactly two critical edges $\varepsilon, \varepsilon'$ and $\beta(\varepsilon,X) + \beta(\varepsilon',X)=0$. Furthermore, $\beta(\varepsilon,X) \le 0.005$, so we get from (\ref{eq:6})}\nonumber\\ \zeta &\le \sum_{\varepsilon \subset \B(\mathbf{0},r)}\sum_{\substack{X \in \CL(\varepsilon):X \nsubseteq \B(\mathbf{0},r)\\ \wedge X \text{ is 4-cell with }|\EC(X)|=2}} 0.005. \label{eq:7} \intertext{Next, we want to exchange the inner and outer sum. Since the inner sum is zero for non-critical edges and the cells intersect the boundary of $\B(\mathbf{0},r)$, we get from (\ref{eq:7})}\nonumber\\ \zeta &\le\sum_{\substack{X \subset \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4) :\\ X \text{ is a 4-cell with }|\EC(X)|=2}}\sum_{\varepsilon\in \EC(X)} 0.005\nonumber\\ &\le\sum_{\substack{X \subset \B(\mathbf{0},r+4) \setminus \B(\mathbf{0},r-4) : \\X \text{ is a 4-cell}}} 0.01.\label{eq:8} \intertext{Now, we will estimate the number of points in $V$ inside the sperical shell as before. A 4-cell is the convex hull of four points in $V$ with circumradius at most $\sqrt{2}$ (see Definition~6.51 in \cite{Blueprint}). Therefore, for a fixed point $\mathbf{v}\in V$ all points that can form a 4-cell with $\mathbf{v}$ must lie inside a ball of radius $2\sqrt{2}$ centered at $\mathbf{v}$. As before, we calculate an upper bound on the number of points in $V$ inside a ball of radius $2\sqrt{2}$. Then, the number of subsets of size three that can be formed with these points is an upper bound on the number of 4-cells a point in $V$ can be part of. By multiplying these numbers, we count each 4-cell 4 times, so we divide by 4. So, we get from (\ref{eq:8})}\nonumber\\ \zeta & \le \frac{\frac{4}{3}\pi(r+5)^3 - \frac{4}{3}\pi(r-5)^3}{\frac{4}{3}\pi} \cdot {\left\lfloor\frac{\frac{4}{3}\pi(2\sqrt{2}+1)^3}{\frac{4}{3}\pi}\right\rfloor \choose 3} \cdot \frac{1}{4}\cdot 0.01\nonumber\\ & = \frac{\left(30r^2 + 250\right)}{4}\cdot {\left\lfloor\left( 2\sqrt{2}+1\right)^3\right\rfloor \choose 3}\cdot 0.01\nonumber\\ &= \left(7.5 r^2 + 62.5\right)\cdot 27720\cdot 0.01\nonumber\\ &= 2079r^2 + 17325.\nonumber\\ \intertext{Again, we can assume $r\ge 1$, so}\nonumber \zeta &\le 19404 r^2. \end{align} Now, we plug this into inequality~(\ref{eq:gammaIntermediate3}) and obtain \begin{align*} \sum_{X \subset\B(\mathbf{0},r)} \gamma(X,L) &\ge \sum_{\varepsilon \subset \B(\mathbf{0},r)} \Gamma(\varepsilon) - 19404r^2- 12710r^2\\ &= \underbrace{\sum_{\varepsilon \subset \B(\mathbf{0},r)} \Gamma(\varepsilon)}_{\ge 0 \text{ by Theorem~6.93 in \cite{Blueprint}}}-32114r^2\\ &\ge - 32114r^2, \end{align*} i.e. (\ref{eq:Gamma}) holds for $c_0= -32114$. We showed that both conditions in Lemma~\ref{lem:6.86} hold and therefore the function $\G(\ast,\funcL)$ is FCC-compatible and negligible (see (\ref{eq:negligible}) and the explanation thereafter) for \begin{align*} c_1=-(c_0 + c_2) &= \frac{56}{3}+m_1\cdot 2240 + 32114\\ &\le \frac{56}{3}+1.013\cdot 2240 + 32114\\ &\le 34402. \end{align*} By (\ref{eq:lem:6.13}), the constant in Lemma~\ref{lem:6.13} only depends on constants and $c_1$. So it is independent of the packing since we showed that $c_1$ is. Therefore, we turn now to inequality (\ref{eq:lem:6.13}) and will later plug in $c_1$ as calculated above. \begin{align*} \density(V,\mathbf{0},r)&\leq \frac{\pi}{\sqrt{18}}\left(1+\frac{3}{r}\right)^3 + c_1\frac{\left(r+1\right)^2}{r^34\sqrt{2}}\\ &= \frac{\pi}{\sqrt{18}}\left(1+\frac{9}{r}+\frac{27}{r^2}+\frac{27}{r^3}\right)+c_1\frac{r^2+2r+1}{r^3 4\sqrt{2}}\\ &=\frac{\pi}{\sqrt{18}} + \frac{\pi}{\sqrt{18}}\left(\frac{9}{r}+\frac{27}{r^2}+\frac{27}{r^3}\right) + c_1\left(\frac{1}{r4\sqrt{2}}+\frac{2}{r^24\sqrt{2}}+\frac{1}{r^34\sqrt{2}}\right) \intertext{Since $r\ge 1$, we have $\frac{1}{r^3} \le \frac{1}{r^2} \le \frac{1}{r}$ and get} &\le \frac{\pi}{\sqrt{18}} + \frac{63\pi}{\sqrt{18}r}+\frac{c_1}{\sqrt{2}r}\\ &= \frac{\pi}{\sqrt{18}} + \frac{21\pi+c_1}{\sqrt{2}r}\\ &=\frac{\pi}{\sqrt{18}} + \frac{21\pi+34402}{\sqrt{2}}\cdot\frac{1}{r}. \end{align*} Summarizing, we showed that the constant in Lemma~\ref{lem:6.13} does not depend on the particular packing $V$ but only on the constant for the assumed existing FCC-compatible negligible function. Then, we showed that there is a FCC-compatible negligible function for which the definition for negligible holds for a constant independent of the packing. So, we can state the main result of this work as follows. \begin{thm} For a saturated packing $V$ and all $r\ge 1$ it holds that \begin{equation*} \density(V,\mathbf{0},r) \le \frac{\pi}{\sqrt{18}} + \frac{21\pi+34402}{\sqrt{2}}\cdot\frac{1}{r} \le \frac{\pi}{\sqrt{18}} + 24373\cdot\frac{1}{r}. \end{equation*} \end{thm} \bibliographystyle{hplain}

https://arxiv.org/abs/2003.07234

On the fixed volume discrepancy of the Korobov point sets

This paper is devoted to the study of a discrepancy-type characteristic -- the fixed volume discrepancy -- of the Korobov point sets in the unit cube. It was observed recently that this new characteristic allows us to obtain optimal rate of dispersion from numerical integration results. This observation motivates us to thoroughly study this new version of discrepancy, which seems to be interesting by itself. This paper extends recent results by V. Temlyakov and M. Ullrich on the fixed volume discrepancy of the Fibonacci point sets.

\section{Introduction} \label{I} This paper is a follow up to the recent paper \cite{VT172}. It is devoted to the study of a discrepancy-type characteristic -- the fixed volume discrepancy -- of a point set in the unit cube $\Omega_d:=[0,1)^d$. We refer the reader to the following books and survey papers on discrepancy theory and numerical integration \cite{BC}, \cite{Mat}, \cite{NW10}, \cite{VTbookMA}, \cite{Bi}, \cite{DTU}, \cite{T11}, and \cite{VT170}. Recently, an important new observation was made in \cite{VT161}. It claims that a new version of discrepancy -- the $r$-smooth fixed volume discrepancy -- allows us to obtain optimal rate of \emph{dispersion} from numerical integration results (see \cite{AHR,BH19,DJ,RT,Rud,Sos,Ull,U19,UV18} for some recent results on dispersion). This observation motivates us to thoroughly study this new version of discrepancy, which seems to be interesting by itself. The \emph{$r$-smooth fixed volume discrepancy} takes into account two characteristics of a smooth hat function $h^r_B$ -- its smoothness $r$ and the volume of its support $v:=\mathrm{vol}(B)$ (see the definition of $h^r_B$ below). We now proceed to a formal description of the problem setting and to formulation of the results. Denote by $\chi_{[a,b)}(x)$ a univariate characteristic function (on ${\mathbb R}$) of the interval $[a,b)$ and, for $r=1,2,3,\dots$, we inductively define $$ h^1_u(x):= \chi_{[-u/2,u/2)}(x) $$ and $$ h^r_u(x) := h^{r-1}_u(x)\ast h^1_u(x), $$ where $$ f(x)\ast g(x) := \int_{\mathbb R} f(x-y)g(y)dy. $$ Note that $h^2_u$ is the \emph{hat function}, i.e., $h_{u}^2(x) =\max\{u-|x|,0\}$. Let $\Delta_tf(x):=f(x)-f(x+t)$ be the first difference. We say that a univariate function $f$ has smoothness $1$ in $L_1$ if $\|\Delta_t f\|_1\le C|t|$ for some absolute constant $C<\infty$. In case $\|\Delta^r_t f\|_1 \le C|t|^r$, where $\Delta^r_t:= (\Delta_t)^r$ is the $r$th difference operator, $r\in {\mathbb{N}}$, we say that $f$ has smoothness $r$ in $L_1$. Then, $h^r_u(x)$ has smoothness $r$ in $L_1$ and has support $(-ru/2,ru/2)$. For a box $B$ of the form \begin{equation}\label{eq:B} B= \prod_{j=1}^d [z_j-ru_j/2,z_j+ru_j/2) \end{equation} define \begin{equation}\label{eq:hB} h^r_B(\mathbf x):= h^r_\mathbf u(\mathbf x-\mathbf z):=\prod_{j=1}^d h^r_{u_j}(x_j-z_j). \end{equation} We begin with the \emph{non-periodic} $r$-smooth fixed volume discrepancy introduced and studied in \cite{VT161}. \begin{Definition}\label{ED.1} Let $r\in{\mathbb{N}}$, $v\in (0,1]$ and $\xi:= \{\xi^\mu\}_{\mu=1}^m\subset [0,1)^d$ be a point set. We define the $r$-smooth fixed volume discrepancy with equal weights as \begin{equation}\label{E.1} D^r(\xi,v):= \sup_{B\subset \Omega_d:vol(B)=v}\left|\int_{\Omega_d} h_B^r(\mathbf x)d\mathbf x-\frac{1}{m}\sum_{\mu=1}^m h_B^r(\xi^\mu)\right|. \end{equation} \end{Definition} \noindent It is well known that the Fibonacci cubature formulas are optimal in the sense of order for numerical integration of different kind of smoothness classes of functions of two variables, see e.g.~\cite{DTU,TBook,VTbookMA}. We present a result from \cite{VT161}, which shows that the Fibonacci point set has good fixed volume discrepancy. Let $\{b_n\}_{n=0}^{\infty}$, $b_0=b_1 =1$, $b_n = b_{n-1}+b_{n-2}$, $n\ge 2$, be the Fibonacci numbers. Denote the $n$th {\it Fibonacci point set} by $$ \mathcal F_n:= \Big\{\big(\mu/b_n,\{\mu b_{n-1} /b_n \}\big)\colon\, \mu=1,\dots,b_n\Big\}. $$ In this definition $\{a\}$ is the fractional part of the number $a$. The cardinality of the set $\mathcal F_n$ is equal to $b_n$. In \cite{VT161} we proved the following upper bound. \begin{Theorem}\label{ET.1} Let $r\ge2$. There exist constants $c,C>0$ such that for any $v\ge c/b_n$ we have \begin{equation}\label{E.3} D^{r}(\mathcal F_n,v) \,\le\, C\,\frac{\log(b_n v)}{b_n^r}. \end{equation} \end{Theorem} \bigskip The main object of interest in the paper \cite{VT172} was the \emph{periodic} $r$-smooth $L_p$-discrepancy of the Fibonacci point sets. For this, we define the periodization $\tilde f$ (with period $1$ in each variable) of a function $f\in L_1({\mathbb R}^d)$ with a compact support by $$ \tilde f(\mathbf x) := \sum_{\mathbf m\in \mathbb Z^d} f(\mathbf m+\mathbf x) $$ and, for each $B\subset [0,1)^d$, we let $\tilde h^r_B$ be the periodization of $h^r_B$ from~\eqref{eq:hB}. We now define the periodic $r$-smooth $L_p$-discrepancy. \begin{Definition}\label{ID.1}For $r\in{\mathbb{N}}$, $1\le p\le \infty$ and $v\in(0,1]$ define the periodic $r$-smooth fixed volume $L_p$-discrepancy of a point set $\xi$ by \begin{equation}\label{1.4} \tilde D^{r}_p(\xi,v):= \sup_{B\subset \Omega_d:vol(B)=v} \left\|\int_{\Omega_d} \tilde h^r_B(\mathbf x-\mathbf z)d\mathbf x- \frac1m\sum_{\mu=1}^m \tilde h^r_B(\xi^\mu-\mathbf z)\right\|_p \end{equation} where the $L_p$-norm is taken with respect to $\mathbf z$ over the unit cube $\Omega_d=[0,1)^d$.\\ \end{Definition} In the case of $p=\infty$ this concept was introduced and studied in \cite{VT163}. The following upper bound for $1\le p<\infty$ was proved in \cite{VT172}. \begin{Theorem}\label{IT.1} Let $r\in{\mathbb{N}}$ and $1\le p<\infty$. There exist constants $c,C>0$ such that for any $v\ge c/b_n$ we have $$ \tilde D^{r}_p(\mathcal F_n,v) \,\le\, C\,\frac{\sqrt{\log(b_n v)}}{b_n^r}. $$ \end{Theorem} \medskip In the case $p=\infty$ a weaker upper bound was proved in \cite{VT172}. \begin{Theorem}\label{IT.2} Let $r\ge 2$. There exist constants $c,C>0$ such that for any $v\ge c/b_n$ we have $$ \tilde D^{r}_\infty(\mathcal F_n,v) \,\le\, C\,\frac{\log(b_n v)}{b_n^r}. $$ \end{Theorem} Some comments, which show that Theorems \ref{IT.1} and \ref{IT.2} cannot be improved in a certain sense were given in \cite{VT172}. Our main interest in this paper is to study the Korobov cubature formulas instead of the Fibonacci cubature formulas from the point of view of the fixed volume discrepancy. We prove a conditional result under assumption that the Korobov cubature formulas are exact on a certain subspace of trigonometric polynomials with frequencies from a hyperbolic cross. There are results that guarantee existence of such cubature formulas (see Section \ref{F} for a discussion). Let $m\in{\mathbb{N}}$, $\mathbf a := (a_1,\dots,a_d)$, $a_1,\dots,a_d\in\mathbb Z$. We consider the cubature formulas $$ P_m (f,\mathbf a):= m^{-1}\sum_{\mu=1}^{m}f\left (\left \{\frac{\mu a_1} {m}\right\},\dots,\left \{\frac{\mu a_d}{m}\right\}\right), $$ which are called the {\it Korobov cubature formulas}. In the case $d=2$, $m=b_n$, $\mathbf a = (1,b_{n-1})$ we have $$ P_m (f,\mathbf a) =\Phi_n (f):= \frac{1}{b_n}\sum_{\mathbf y\in \mathcal F_n} f(\mathbf y). $$ Denote $$ \mathbf y^{\mu}:=\left (\left \{\frac{\mu a_1} {m}\right\},\dots,\left \{\frac{\mu a_d}{m}\right\}\right), \quad \mu = 1,\dots,m,\quad \mathcal K_m(\mathbf a):=\{\mathbf y^\mu\}_{\mu=1}^m. $$ The set $\mathcal K_m(\mathbf a)$ is called the {\it Korobov point set}. Further, denote $$ S(\mathbf k, \mathbf a) \,:=\, P_m\left(e^{i2\pi(\mathbf k,\mathbf x)},\mathbf a\right) \,=\, m^{-1}\sum_{\mu=1}^{m}e^{i2\pi(\mathbf k,\mathbf y^{\mu})}. $$ Note that \begin{equation}\label{2.1} P_m (f,\mathbf a) =\sum_{\mathbf k}\hat f(\mathbf k)\, S(\mathbf k,\mathbf a),\quad \hat f(\mathbf k) := \int_{[0,1)^d} f(\mathbf x)\, e^{-i2\pi(\mathbf k,\mathbf x)}d\mathbf x, \end{equation} where for the sake of simplicity we may assume that $f$ is a trigonometric polynomial. It is clear that (\ref{2.1}) holds for $f$ with absolutely convergent Fourier series. It is easy to see that the following relation holds \begin{equation}\label{2.2} S(\mathbf k,\mathbf a)= \begin{cases} 1&\quad\text{ for }\quad \mathbf k\in L(m,\mathbf a),\\ 0&\quad\text{ for }\quad \mathbf k\notin L(m,\mathbf a), \end{cases} \end{equation} where $$ L(m,\mathbf a) := \bigl\{ \mathbf k:(\mathbf a,\mathbf k) \equiv 0 \qquad \pmod{m}\bigr\} . $$ For $N\in{\mathbb{N}}$ define the {\it hyperbolic cross} by $$ \Gamma(N,d):= \left\{\mathbf k= (k_1,\dots,k_d)\in\mathbb Z^d\colon \prod_{j=1}^d \max(|k_j|,1) \le N\right\}. $$ Denote $$ \mathcal T(N,d) := \left\{f\, :\, f(\mathbf x)= \sum_{\mathbf k\in \Gamma(N,d)} c_\mathbf k e^{i2\pi (\mathbf k,\mathbf x)}\right\}. $$ It is easy to see that the condition \begin{equation}\label{ex} P_m(f,\mathbf a) = \hat f(\mathbf 0), \quad f\in \mathcal T(N,d), \end{equation} is equivalent to the condition \begin{equation}\label{exs} \Gamma(N,d)\cap \bigl(L(m,\mathbf a)\backslash\{\mathbf 0\}\bigr) = \varnothing. \end{equation} \begin{Definition}\label{exact} We say that the Korobov cubature formula $P_m(\cdot,\mathbf a)$ is exact on $\mathcal T(N,d)$ if condition \eqref{ex} (equivalently, condition \eqref{exs}) is satisfied. \end{Definition} \begin{Theorem}\label{IT.3} Suppose that $P_m(\cdot,\mathbf a)$ is exact on $\mathcal T(L,d)$ with some $L\in{\mathbb{N}}$, $L\ge 2$. Let $r\in{\mathbb{N}}$ and $1\le p<\infty$. There exist constants $c(d),C(d,p)>0$ such that for any $v\ge c(d)/L$ we have $$ \tilde D^{r}_p(\mathcal K_m(\mathbf a),v) \,\le\, C(d,p)L^{-r} (\log (Lv))^{(d-1)/2}. $$ \end{Theorem} \medskip In the case $p=\infty$ we prove a weaker upper bound. \begin{Theorem}\label{IT.4} Let $r\ge 2$. Suppose that $P_m(\cdot,\mathbf a)$ is exact on $\mathcal T(L,d)$ with some $L\in{\mathbb{N}}$, $L\ge 2$. There exist constants $c(d),C(d)>0$ such that for any $v\ge c(d)/L$ we have $$ \tilde D^{r}_\infty(\mathcal K_m(\mathbf a),v) \,\le\, C(d)L^{-r}(\log (Lv))^{d-1}. $$ \end{Theorem} \section{Proofs of Theorems \ref{IT.3} and \ref{IT.4}} \label{Fib} The proofs of both theorems go along the same lines. We give a detailed proof of Theorem \ref{IT.3} and point out a change of this proof, which gives Theorem \ref{IT.4}. For continuous functions of $d$ variables, which are $1$-periodic in each variable, consider the Korobov cubature formula $P_m(\cdot,\mathbf a)$. For the univariate \emph{test functions} $h^r_u$ we obtain by the properties of convolution that $$ \hat h^r_u(y) = \hat h^{r-1}_u(y)\, \hat h^1_u(y),\qquad y\in{\mathbb R} $$ which implies for $y\neq 0$ $$ \hat h^r_u(y) = \left(\frac{\sin(\pi yu)}{\pi y}\right)^r. $$ Therefore, $$ \left|\hat {\tilde h}^r_u(k)\right| \,\le\, \min\left(|u|^r,\frac{1}{|k|^r}\right) \,=\, \left(\frac{|u|}{k'}\right)^{r/2}\min\left(|k' u|^{r/2},\frac{1}{|ku|^{r/2}}\right), $$ where $k':=\max\{1,|k|\}$. (Here, we used for a moment $\hat h$ for the Fourier transform of $h$ on ${\mathbb R}$. This should not lead to any confusion.) \bigskip It is convenient for us to use the following abbreviated notation for the product $$ pr(\mathbf u):= pr(\mathbf u,d) := \prod_{j=1}^d u_j. $$ For $B\subset\Omega_d$ of the form~\eqref{eq:B} and $\mathbf z\in\Omega_d$, we have \begin{equation}\label{eq:shift} \hat {\tilde h}^r_{B+\mathbf z}(\mathbf k) \,=\, e^{-i2\pi(\mathbf k,\mathbf z)}\, \hat {\tilde h}^r_B(\mathbf k), \end{equation} where $\tilde h^r_{B+\mathbf z}(\mathbf x):=\tilde h^r_{B}(\mathbf x-\mathbf z)$, see~\eqref{eq:hB}. Therefore, we obtain from the above that \[ \left|\hat {\tilde h}^r_B(\mathbf k)\right| \,\le\, \prod_{j=1}^d \left(\frac{|u_j|}{k_j'}\right)^{r/2}\min\left(|k_j' u_j|^{r/2},\frac{1}{|k_j u_j|^{r/2}}\right). \] For $\mathbf s\in {\mathbb{N}}_0^d$, we define $$ \rho(\mathbf s) := \Big\{\mathbf k \in \mathbb Z^d\colon [2^{s_j-1}] \le |k_j| < 2^{s_j}, \quad j=1,\dots,d \Big\}, $$ where $[a]$ denotes the integer part of $a$, and obtain, for $\mathbf k\in\rho(\mathbf s)$, that \begin{equation}\label{eq:HB} \left|\hat {\tilde h}^r_B(\mathbf k)\right| \,\le\, H_B^r(\mathbf s) \,:=\, \left(\frac{pr(\mathbf u)}{2^{\|\mathbf s\|_1}}\right)^{r/2}\,\prod_{j=1}^d \min\left((2^{s_j}u_j)^{r/2},\frac{1}{(2^{s_j}u_j)^{r/2}}\right). \end{equation} Later we will need certain sums of these quantities. First, consider $$ \sigma^r_\mathbf u(t) \,:=\, \sum_{\|\mathbf s\|_1=t}\prod_{j=1}^d \min\left((2^{s_j}u_j)^{r/2},\frac{1}{(2^{s_j}u_j)^{r/2}}\right),\quad t\in{\mathbb{N}}_0. $$ The following technical lemma is part~(I) from \cite[Lemma 6.1]{VT161}. \begin{Lemma}\label{L2.2} Let $r>0$, $t\in {\mathbb{N}}$ and $\mathbf u\in (0,1/2]^d$ be such that $pr(\mathbf u)\ge 2^{-t}$. Then, we have \[ \sigma^r_\mathbf u(t) \,\le\, C(d) \frac{\left(\log(2^{t+1}pr(\mathbf u))\right)^{d-1}}{(2^tpr(\mathbf u))^{r/2}}. \] \end{Lemma} \bigskip \noindent This lemma and \eqref{eq:HB} imply that \begin{equation}\label{eq:HB-bound} \sum_{\|\mathbf s\|_1=t} H_B^r(\mathbf s)^2 \,\le\, C_1\, 2^{-2rt}\, \Big(\log\big(2^{t}\, v\big)\Big)^{d-1}, \end{equation} where $v:=\mathrm{vol}(B)=r^d\, pr(\mathbf u)$, for all $r\ge1$ and all $t\in{\mathbb{N}}_0$ with $v\ge r^d\,2^{-t+1}$ and an absolute constant $C_1<\infty$. \bigskip Additionally, we need a result from harmonic analysis -- a corollary of the Littlewood-Paley theorem. Denote $$ \delta_\mathbf s(f,\mathbf x):= \sum_{\mathbf k\in\rho(\mathbf s)} \hat f(\mathbf k)e^{i2\pi(\mathbf k,\mathbf x)}. $$ Then it is known that for $p\in [2,\infty)$ one has \begin{equation}\label{eq:LP} \|f\|_p \le C(d,p) \left(\sum_{\mathbf s\in{\mathbb{N}}_0^d}\|\delta_\mathbf s(f)\|_p^2\right)^{1/2}. \end{equation} Note that in the proof of Theorem \ref{IT.4} we use the simple triangle inequality \begin{equation}\label{eq:triangle} \|f\|_\infty \le \sum_{\mathbf s}\|\delta_\mathbf s(f)\|_\infty. \end{equation} instead of \eqref{eq:LP}. \bigskip Let us define \begin{equation}\label{2.3} E^r_B(\mathbf z):= \frac{1}{m}\sum_{\mu=1}^{m} \tilde h^r_B(\mathbf y^\mu-\mathbf z)- \int_{[0,1)^d} \tilde h^r_B(\mathbf x)d\mathbf x \end{equation} such that \[ \tilde D^{r}_p(\mathcal K_m(\mathbf a),v) \,=\, \sup_{B\subset \Omega_d:vol(B)=v} \left\|E^r_B\right\|_p \] By formulas \eqref{2.1}, \eqref{2.2} and \eqref{eq:shift} we obtain \[ E^r_B(\mathbf z) \,=\, \sum_{\mathbf k\neq 0} \hat {\tilde h}^r_B(\mathbf k)\,S(\mathbf k,\mathbf a)\, e^{-i2\pi(\mathbf k,\mathbf z)} \,=\, \sum_{\mathbf k\in L(m,\mathbf a)\setminus\{0\}} \hat {\tilde h}^r_B(\mathbf k)\, e^{-i2\pi(\mathbf k,\mathbf z)}. \] It is apparent from~\eqref{eq:LP} that it remains to bound $\|\delta_s(E_B^r)\|_p$. If $t\neq 0$ is such that $2^t\le L$ then for $\mathbf s$ with $\|\mathbf s\|_1=t$ we have $\rho(\mathbf s)\subset \Gamma(L)$. Then our assumption \eqref{exs} implies that $S(\mathbf k,\mathbf a)=0$ for $\mathbf k\in\rho(\mathbf s)$ and, therefore, $\delta_s(E^r_B)=0$. Let $t_0\in {\mathbb{N}}$ be the smallest number satisfying $2^{t_0}>L$, i.e., $t_0>\log L$. Then, from~\eqref{eq:LP} for $p\in[2,\infty)$, we have \begin{equation}\label{eq:EB-bound} \|E^r_B\|_p \le C(p) \left(\sum_{t=t_0}^\infty \sum_{\|\mathbf s\|_1=t}\|\delta_s(E^r_B)\|_p^2\right)^{1/2}. \end{equation} Moreover, \eqref{exs} implies that for $t\ge t_0$ we have \begin{equation}\label{eq:number} \#\big(\rho(\mathbf s)\cap L(m,\mathbf a)\big) \le C_2\, 2^{t-t_0}, \quad \|\mathbf s\|_1=t. \end{equation} By Parseval's identity we obtain \[ \|\delta_s(E^r_B)\|_2 \,=\, \sqrt{\sum_{\mathbf k\in\rho(\mathbf s)\cap L(m,\mathbf a)}|\hat {\tilde h}^r_B(\mathbf k)|^2} \,\le\, \sqrt{\#\big(\rho(\mathbf s)\cap L(m,\mathbf a)\big)}\cdot H_B^r(\mathbf s) \] and, by the triangle inequality, \[ \|\delta_s(E_B)\|_\infty \,\le\, \#\big(\rho(\mathbf s)\cap L(m,\mathbf a)\big)\cdot H_B^r(\mathbf s) \] Hence, using the inequality $$ \|f\|_p \le \|f\|_2^{2/p}\|f\|_\infty^{1-2/p} $$ for $2\le p\le \infty$, we get \[ \|\delta_s(E^r_B)\|_p \,\le\, \Big(\#\big(\rho(\mathbf s)\cap L(m,\mathbf a)\big)\Big)^{1-1/p}\cdot H_B^r(\mathbf s). \] Combining this with \eqref{eq:HB-bound}, \eqref{eq:EB-bound} and \eqref{eq:number}, we finally obtain for all $v=\mathrm{vol}(B)\ge 2r^d 2^{-t_0}$ and $p\in[2,\infty)$ that \[\begin{split} \|E^r_B\|_p \,&\le\, C \left(\sum_{t=t_0}^\infty 2^{2(t-t_0)(1-1/p)} \sum_{\|\mathbf s\|_1=t}H_B^r(\mathbf s)^2\right)^{1/2} \\ \,&\le\, C' \left(\sum_{t=t_0}^\infty 2^{2(t-t_0)(1-1/p)}\, 2^{-2rt}\, \left(\log\big(2^{t}\, v\big)\right)^{d-1}\right)^{1/2} \\ \,&=\, C'\, 2^{-r t_0} \left(\sum_{t=0}^\infty 2^{2t(1-1/p-r)}\, \left(\log\big(2^{t+t_0}\, v\big)\right)^{d-1}\right)^{1/2} \\ \,&\le\, C''\, 2^{-r t_0}\, \left(\log\big(2^{t_0}\, v\big)\right)^{(d-1)/2} \left(\sum_{t=0}^\infty t\, 2^{2t(1-1/p-r)}\right)^{1/2}. \end{split}\] Using $t_0\ge\log L$ and that $\|E^r_B\|_p\le\|E^r_B\|_2$ for $p<2$, this implies Theorem~\ref{IT.3}. (Here, we used that clearly $r> 1-1/p$ for $p<\infty$.) As we pointed out above, in the proof of Theorem \ref{IT.4} we use inequality \eqref{eq:triangle} instead of \eqref{eq:LP}. Moreover we use \[ \sum_{\|\mathbf s\|_1=t} H_B^r(\mathbf s) \,\le\, C_1\, 2^{-rt}\,\left(\log\big(2^{t}\, v\big)\right)^{d-1}, \] for all $r\ge1$ instead of~\eqref{eq:HB-bound}. However, note that we need $r>1$ for the last series in the above computation to be finite. This implies $$ \|E_B^r\|_\infty \,\le\, C\, L^{-r}\left(\log\left(L\,v\right)\right)^{d-1}. $$ \section{Dispersion of the Korobov point sets} \label{F} We remind the definition of dispersion. Let $d\ge 2$ and $[0,1)^d$ be the $d$-dimensional unit cube. For $\mathbf x,\mathbf y \in [0,1)^d$ with $\mathbf x=(x_1,\dots,x_d)$ and $\mathbf y=(y_1,\dots,y_d)$ we write $\mathbf x < \mathbf y$ if this inequality holds coordinate-wise. For $\mathbf x<\mathbf y$ we write $[\mathbf x,\mathbf y)$ for the axis-parallel box $[x_1,y_1)\times\cdots\times[x_d,y_d)$ and define $$ \mathcal B:= \{[\mathbf x,\mathbf y): \mathbf x,\mathbf y\in [0,1)^d, \mathbf x<\mathbf y\}. $$ For $n\ge 1$ let $T$ be a set of points in $[0,1)^d$ of cardinality $|T|=n$. The volume of the largest empty (from points of $T$) axis-parallel box, which can be inscribed in $[0,1)^d$, is called the dispersion of $T$: $$ \text{disp}(T) := \sup_{B\in\mathcal B: B\cap T =\emptyset} vol(B). $$ An interesting extremal problem is to find (estimate) the minimal dispersion of point sets of fixed cardinality: $$ \text{disp*}(n,d) := \inf_{T\subset [0,1)^d, |T|=n} \text{disp}(T). $$ It is known that \begin{equation}\label{F.1} \text{disp*}(n,d) \le C^*(d)/n. \end{equation} Inequality (\ref{F.1}) with $C^*(d)=2^{d-1}\prod_{i=1}^{d-1}p_i$, where $p_i$ denotes the $i$th prime number, was proved in \cite{DJ} (see also \cite{RT}). The authors of \cite{DJ} used the Halton-Hammersly set of $n$ points (see \cite{Mat}). Inequality (\ref{F.1}) with $C^*(d)=2^{7d+1}$ was proved in \cite{AHR}. The authors of \cite{AHR}, following G. Larcher, used the $(t,r,d)$-nets (see \cite{NX}, \cite{Mat} for results on $(t,r,d)$-nets and Definition \ref{GD.1} below for the definition). \begin{Definition}\label{GD.1} A $(t,r,d)$-net (in base $2$) is a set $T$ of $2^r$ points in $[0,1)^d$ such that each dyadic box $[(a_1-1)2^{-s_1},a_12^{-s_1})\times\cdots\times[(a_d-1)2^{-s_d},a_d2^{-s_d})$, $1\le a_j\le 2^{s_j}$, $j=1,\dots,d$, of volume $2^{t-r}$ contains exactly $2^t$ points of $T$. \end{Definition} It was demonstrated in \cite{VT161} how good upper bounds on fixed volume discrepancy can be used for proving good upper bounds for dispersion. This fact was one of the motivation for studying the fixed volume discrepancy. Theorem \ref{FT.1} below was derived from Theorem \ref{ET.1} (see \cite{VT161}). The upper bound in Theorem \ref{FT.1} combined with the trivial lower bound shows that the Fibonacci point set provides optimal rate of decay for the dispersion. \begin{Theorem}\label{FT.1} There is an absolute constant $C$ such that for all $n$ we have \begin{equation}\label{F.2} \operatorname{disp} (\mathcal F_n) \le C/b_n. \end{equation} \end{Theorem} The reader can find further recent results on dispersion in \cite{Rud}, \cite{Sos}, \cite{Ull}, and \cite{BH19}. We now proceed to new results on dispersion of the Korobov point sets. \begin{Theorem}\label{FT.2} Suppose that $P_m(\cdot,\mathbf a)$ is exact on $\mathcal T(L,d)$ with some $L\in{\mathbb{N}}$, $L\ge 2$. There exists a constant $C_1(d)>0$ such that we have $$ \operatorname{disp}(\mathcal K_m(\mathbf a)) \,\le\, C_1(d)L^{-1}. $$ \end{Theorem} \begin{proof} The proof is based on Theorem \ref{IT.4}. Specify some $r\in {\mathbb{N}}$, $r\ge 2$. Then, by Theorem \ref{IT.4} we have $$ \tilde D^{r}_\infty(\mathcal K_m(\mathbf a),v) \,\le\, C(d)L^{-r}(\log (Lv))^{d-1}. $$ Let a box $B\in \mathcal B$ be an empty box $B\cap \mathcal K_m(\mathbf a) = \varnothing$. Denote $v=vol(B)$. Note that for $B\in \mathcal B$ we have $\tilde h^r_B(\mathbf x)=h^r_B(\mathbf x)$. Then, from the Definition \ref{ID.1} of the $\tilde D^{r}_\infty(\mathcal K_m(\mathbf a),v)$ it follows that \begin{equation}\label{F1} \left|\int_{\Omega_d} h^r_B(\mathbf x)d\mathbf x- \frac1m\sum_{\mu=1}^m h^r_B(\mathbf y^\mu)\right|\le \tilde D^{r}_\infty(\mathcal K_m(\mathbf a),v). \end{equation} Our assumption that $B$ is an empty box implies that $$ \frac1m\sum_{\mu=1}^m h^r_B(\mathbf y^\mu)=0 $$ and, therefore, by \eqref{F1} we obtain \begin{equation}\label{F2} \left|\int_{\Omega_d} h^r_B(\mathbf x)d\mathbf x\right|\le \tilde D^{r}_\infty(\mathcal K_m(\mathbf a),v). \end{equation} If $v\le c(d)L^{-1}$, where $c(d)$ is from Theorem \ref{IT.4}, then Theorem \ref{FT.2} with $C_1(d)\ge c(d)$ follows. Assume $v\ge c(d)L^{-1}$. Then we apply Theorem \ref{IT.4} and obtain from \eqref{F2} \begin{equation}\label{F3} \left|\int_{\Omega_d} h^r_B(\mathbf x)d\mathbf x\right|\le C(d)L^{-r}(\log (Lv))^{d-1}. \end{equation} It follows from the definition of functions $h^r_B(\mathbf x)$ that \begin{equation}\label{F4} \left|\int_{\Omega_d} h^r_B(\mathbf x)d\mathbf x\right|\ge c_1(d) v^r. \end{equation} Inequalities \eqref{F3} and \eqref{F4} imply that $Lv\le C'(d)$. \newline Setting $C_1(d) := \max(c(d),C'(d))$, we complete the proof of Theorem \ref{FT.2}. \end{proof} We now make some comments. {\bf Fibonacci point sets.} As we already mentioned above we have in the case $d=2$, $m=b_n$, $\mathbf a=(1,b_{n-1})$, that $P_m(f,\mathbf a)=\Phi_n(f)$. Denote in this case $L(n):=L(m,\mathbf a)$. In other words $$ L(n) :=\Bigl\{ \mathbf k = (k_1,k_2)\in\mathbb Z^2\colon\; k_1 + b_{n-1} k_2\equiv 0 \; \pmod {b_n}\Bigr\}. $$ The following lemma is well known (see, for instance, \cite{VTbookMA}, p.274). \begin{Lemma}\label{FL.1} There exists an absolute constant $\gamma > 0$ such that for any $n > 2$ for the $2$-dimensional hyperbolic cross we have $$ \Gamma(\gamma b_n,2)\cap \bigl(L(n)\backslash\{\mathbf 0\}\bigr) = \varnothing. $$ \end{Lemma} The combination of Theorem \ref{FT.2} with Lemma \ref{FL.1} implies Theorem \ref{FT.1}. {\bf Special Korobov point sets.} Let $L\in {\mathbb{N}}$ be given. Clearly, we are interested in as small $m$ as possible such that there exists a Korobov cubature formula, which is exact on $\mathcal T(L,d)$. In the case of $d=2$ the Fibonacci cubature formula is an ideal in a certain sense choice. There is no known Korobov cubature formulas in case $d\ge 3$, which are as good as the Fibonacci cubature formula in case $d=2$. We now formulate some known results in this direction. Consider a special case $\mathbf a = (1,a,a^2,\dots,a^{d-1})$, $a\in{\mathbb{N}}$. In this case we write in the notation of $\mathcal K_m(\mathbf a)$ and $P_m(\cdot,\mathbf a)$ the scalar $a$ instead of the vector $\mathbf a$, namely, $\mathcal K_m(a)$ and $P_m(\cdot,a)$. The following Lemma \ref{FL.2} is a simple well known result (see, for instance \cite{VTbookMA}, p.285). \begin{Lemma}\label{FL.2} Let $m$ and $L$ be a prime and a natural number, respectively, such that \begin{equation}\label{F5} \bigl|\Gamma(L,d)\bigr| < (m-1)/d . \end{equation} Then there is a natural number $a\in [1,m)$ such that for all $\mathbf k\in\Gamma(L,d)$, $\mathbf k\ne\mathbf 0$ \begin{equation}\label{F6} k_1 + ak_2 +\dots+a^{d-1}k_d\not\equiv 0\qquad \pmod {m}. \end{equation} \end{Lemma} The combination of Theorem \ref{FT.2} with Lemma \ref{FL.2} implies the following Proposition \ref{FP.1}. \begin{Proposition}\label{FP.1} There exists a positive constant $C_2(d)$, which depends only on $d$, with the following property. For any $L\in {\mathbb{N}}$, $L\ge 2$, there exist a prime number $m\le C_2(d)|\Gamma(L,d)|$ and a natural number $a\in[1,m)$ such that $$ \operatorname{disp}(\mathcal K_m(a)) \le C_1(d)L^{-1}. $$ \end{Proposition} \begin{Corollary} Let $C_1(d)$ be the number from Theorem \ref{FT.2} and $p$ be a prime number. There exist a natural number $a \in [1, p)$ such that for any segments of natural numbers $I_1, \dots, I_d \in [1, p]$ satisfying the condition $$\prod\limits_{j=1}^d |I_j| \ge C_1(d) p^{d-1}(\log{p})^{d-1},$$ there exists a natural number $\mu \in [1, p]$ such that \begin{equation} \label{system} \begin{cases} \mu \in I_1 \pmod{p} \\ \mu a \in I_2 \pmod{p}\\ \vdots \\ \mu a^{d-1} \in I_d \pmod{p} . \end{cases} \end{equation} \end{Corollary} \begin{proof} Take $$L = \frac{C(d)p}{(\log{p})^{d-1}},$$ where $C(d)$ is small enough to satisfy (\ref{F5}). By Lemma \ref{FL.2} there exists a natural number $a \in [1, p)$, such that $P_p(\cdot, a)$ is exact on $\mathcal T(L,d)$. Denote $ I_j := [x_j, y_j], \ j = 1, \dots, d$. Then for $\tilde{I}_j := [\frac{x_j}{p}, \frac{y_j}{p}]$ we have $$\prod\limits_{j=1}^d |\tilde{I}_j| \ge \frac{C_1(d) (\log{p})^{d-1}}{p}.$$ By Theorem \ref{FT.2} the set $\mathcal K_p(a)$ intersects the box $\tilde{I}_1 \times \dots \times \tilde{I}_d$ at least at one point. Then, there exists a natural number $\mu \in [1, p]$, such that one has \begin{equation}\notag \begin{cases} \big \{ \frac{\mu }{p} \} \in \tilde{I}_1, \\ \big \{ \frac{\mu a}{p} \} \in \tilde{I}_2, \\ \vdots \\ \big \{ \frac{\mu a^{d-1}}{p} \} \in \tilde{I}_d. \end{cases}, \end{equation} which implies (\ref{system}). \end{proof} {\bf Acknowledgment.} The work was supported by the Russian Federation Government Grant No. 14.W03.31.0031.

https://arxiv.org/abs/1503.02352

Infinite-dimensional $\ell^1$ minimization and function approximation from pointwise data

We consider the problem of approximating a smooth function from finitely-many pointwise samples using $\ell^1$ minimization techniques. In the first part of this paper, we introduce an infinite-dimensional approach to this problem. Three advantages of this approach are as follows. First, it provides interpolatory approximations in the absence of noise. Second, it does not require \textit{a priori} bounds on the expansion tail in order to be implemented. In particular, the truncation strategy we introduce as part of this framework is independent of the function being approximated, provided the function has sufficient regularity. Third, it allows one to explain the key role weights play in the minimization; namely, that of regularizing the problem and removing aliasing phenomena. In the second part of this paper we present a worst-case error analysis for this approach. We provide a general recipe for analyzing this technique for arbitrary deterministic sets of points. Finally, we use this tool to show that weighted $\ell^1$ minimization with Jacobi polynomials leads to an optimal method for approximating smooth, one-dimensional functions from scattered data.

\section{Introduction}\label{s:introduction} Many problems in science and engineering require the approximation of a smooth function from a finite set of pointwise samples. Although a classical problem in approximation theory, in the last several years there has been an increasing focus on the use of convex optimization techniques for this task \cite{DoostanOwhadiSparse,HamptonDoostanCSPCE,MathelinGallivanCSPDErandom,PengHamptonDoostantweighted,RauhutWardSpher,Rauhut,RauhutWardWeighted,YanGuoXui_l1UQ,KarniadakisUQCS}. This is driven in part by applications such as uncertainty quantification, wherein the dimension is typically high and the amount of data severely limited. As dimension increases, smooth multivariate functions are increasingly well-represented by their best $k$-term approximation in certain orthogonal expansions (e.g.\ multivariate Legendre polynomials). Hence the expectation is that these techniques will yield improvements over more standard approaches -- such as discrete least squares and interpolation -- at least when the dimension is sufficiently high and the data points arise from appropriate sampling distributions. A number of recent studies, such as those listed above, have shown this to be the case. \subsection{Current approaches}\label{ss:currentapproaches} Let $f$ be a function in $L^2(D)$, where $D \subseteq \bbR^d$, $\{ \phi_i \}_{i \in \bbN}$ be an orthonormal basis and write $f = \sum_{i \in \bbN} x_i \phi_i$, where $x = \{ x_i \}_{i \in \bbN} \in \ell^2(\bbN)$ is the infinite vector of coefficients of $f$. If $\{ t_n \}^{N}_{n=1}$ is a finite set of points, the problem is to approximate $x$, and therefore $f$, from the data $\{ f(t_n) \}^{N}_{n=1}$. Since $x$ is an infinite vector, in order to compute an approximation to $f$ it is necessary to truncate in some way. In the usual formulation (see \cite{DoostanOwhadiSparse,HamptonDoostanCSPCE,MathelinGallivanCSPDErandom,PengHamptonDoostantweighted,RauhutWardSpher,Rauhut,RauhutWardWeighted,YanGuoXui_l1UQ,KarniadakisUQCS}), one introduces a fixed $M \geq N$ and seeks to approximate the first $M$ coefficients $x_1,\ldots,x_M$ of $x$. If $A = \{ \phi_i(t_n) \}^{N,M}_{n=1,i=1} \in \bbC^{N \times M}$ then the standard (weighted) $\ell^1$ minimization problem is as follows: \be{ \label{intro_bad_min} \min_{z \in \bbC^M} \| z \|_{1,w}\ \mbox{subject to $\| A z - y \| \leq \delta$},\qquad y = \{ f(t_n) \}^{N}_{n=1}. } Here $\| z \|_{1,w} = \sum^{M}_{i=1} w_i | z_i|$ is the $\ell^1_w$-norm on $\bbC^M$ with weights $w_i > 0$. The parameter $\delta$ handles the truncation, and is normally chosen so that $\{ x_i \}^{M}_{i=1}$ is feasible for \R{intro_bad_min}. That is, \be{ \label{finite_bad} \max_{n=1,\ldots,N} \left | f(t_n) - \sum^{M}_{i=1} x_i \phi_i(t_n) \right | \leq \delta. } In other words, the error introduced by truncating the infinite expansion to a vector of length $M$ is viewed as noise in the data. Unfortunately, this formulation raises a number of issues, which we describe next. Overcoming these issues is the goal of this paper. \pbk (i) In order to choose $\delta$, one must have an \textit{a priori} estimate for the truncation error $|f(t) - \sum^{M}_{i=1} x_i \phi_i(t) |$. Note that the approximation error resulting from \R{intro_bad_min} is sensitive to the choice of $\delta$ \cite{KarniadakisUQCS}. In practice, cross-validation techniques have been proposed to empirically determine the truncation error \cite{DoostanOwhadiSparse,HamptonDoostanCSPCE,KarniadakisUQCS}. Yet such techniques are time-consuming, largely lack theoretical support and may not always result in an accurate estimation. \pbk (ii) The approximation $\tilde{f}$ of $f$ obtained from \R{finite_bad} does not interpolate the data. In the absence of noise, interpolatory solutions are often desirable in applications since they ensure that the approximation exactly fits the underlying function $f$ at the points at which $f$ is known. \pbk (iii) The approximation $\tilde{f}$ can be dependent on the choice of weights, and is prone to aliasing (also known as overfitting) if the weights are chosen inappropriately \cite{RauhutWardWeighted}. \pbk (iv) Besides some specific cases where such techniques are known to perform extremely well -- such as when the coefficients $\{ x_i \}^{M}_{i=1}$ are sparse and the data points $\{ t_n \}^{N}_{n=1}$ are chosen randomly according to the orthogonality measure of the basis $\{\phi_i\}_{i \in \bbN}$ \cite{AdcockCSFunInterp,HamptonDoostanCSPCE,RauhutWardSpher,Rauhut,RauhutWardWeighted} -- very little is known about the approximation error $| f(t) - \tilde{f}(t) |$. In particular, for general $f$ (not necessarily having sparse coefficients) and arbitrary \textit{deterministic} scattered points $\{ t_n \}^{N}_{n=1}$ the quality of the approximation $\tilde{f}$ to $f$ is largely unknown. \pbk Issue (iv) has ramifications for a variety of applications where the primary limitation is the availability of data -- that is, where it is time-consuming or expensive to acquire more samples -- as opposed to data-rich scenarios where processing speed is the key concern (in which case classical techniques such as least-squares fitting are likely superior). If weighted $\ell^1$ minimization techniques are to find wide use in practice, then it is beneficial to have error bounds for both ideal (i.e.\ random sample points) and non-ideal conditions (i.e.\ fixed, deterministic sample points). \subsection{Our contributions} The purpose of this paper is to address these issues. In \S \ref{s:minproblems} we first propose an infinite-dimensional weighted $\ell^1$ minimization problem which removes the need for \textit{a priori} knowledge of magnitude of the expansion tail. In the absence of noise, its solutions are exactly interpolatory, unlike solutions of \R{intro_bad_min}. As one might expect, however, such an infinite-dimensional minimization problem cannot be solved numerically. Hence we next introduce a truncation strategy based on a user-controlled parameter $K \in \bbN$. This leads to finite-dimensional minimization problem over $\bbC^K$, reminiscent of \R{intro_bad_min} but with a number of key differences. First, unlike \R{intro_bad_min}, it requires no knowledge of the expansion tail, and second, it retains the interpolatory property of the infinite-dimensional problem. In \S \ref{s:truncation} we show how to select the parameter $K$ in a manner independent of $f$, whenever $f$ has sufficient regularity, and dependent only on the basis $\{ \phi_i \}_{i \in \bbN}$ and data points $\{ t_n \}^{N}_{n=1}$. Formulating the minimization problem in an infinite-dimensional setting also allows us to address issue (iii). In \S \ref{s:need} we first show that unweighted $\ell^1$ minimization is largely unsuitable for the function interpolation from scattered, deterministic data, since it leads to an \textit{aliasing} phenomenon. Specifically, without weights it is possible for the optimization problem to have infinitely-many solutions which interpolate $f$ at the data points, but which do not approximate $f$ to any accuracy away from these points. Fortunately, this problem can be completely resolved by the introduction of slowly-growing weights. In effect, these weights regularize the optimization problem and ensure that such bad solutions of the unweighted problem, while still feasible, are no longer minimizers of the weighted problem. Through subsequent analysis we quantify how fast the weights need to grow to resolve this phenomenon, and demonstrate this result with numerical examples. Issues (i)--(iii) are the focus of the first half of this paper (\S \ref{s:prelim}--\ref{s:truncation}). In the second half, we consider (iv). More precisely, we pose and answer the following two questions: \pbk (a) How well can one approximate a function $f$ using weighted $\ell^1$ minimization from its samples taken on an arbitrary deterministic grid of $N$ points? \pbk (b) How does this approximation perform in comparison to other techniques, such as least squares? \pbk Note that we do not assume any sparsity of the coefficients $x = \{ x_i \}_{i \in \bbN}$ of $f$, although we do assume some mild decay of $x_i$ as $i \rightarrow \infty$ (otherwise the weighted $\ell^1$ problem does not make sense). We also do not assume any structure to the data: the points $\{ t_n \}^{N}_{n=1}$ are deterministic and can be arbitrarily distributed in the domain. As is standard in scattered data approximation, we classify the error in terms of its density (or fill distance) \cite{WendlandScatteredBook}. Our motivation for examining (a) and (b) is the following. Least-squares fitting is a classical and widely-used technique, but is well known to be intensive in the number of samples required to achieve stability and accuracy (see \cite{DavenportEtAlLeastSquares,MiglioratiNobileLowDisc,MiglioratiEtAlFoCM} and references therein). Conversely, under certain conditions (sparsity and random sampling) $\ell^1$ techniques are known to give very good approximations from relatively few samples. However, in certain practical scenarios -- such as when using legacy data -- one may not have the luxury to choose the data points in a way to deliver the best accuracy of the approximation. Moreover, while functions in high dimensions tend to have sparse coefficients in polynomial bases \cite{CohenDeVoreSchwabFoCM,CohenDeVoreSchwabRegularity,DoostanOwhadiSparse}, in low (in particular, one) dimensions polynomial coefficients usually exhibit rapid decay, but typically little sparsity. Since $\ell^1$-based techniques are computationally more intensive than classical methods such as least-squares fitting, this raises the following question: is it still worth using $\ell^1$ techniques even when the data points are scattered and sparsity is not assured? In \S \ref{s:lin_approx_err} we present a general mathematical framework for answering these questions. We introduce a linear approximation error analysis for the infinite-dimensional weighted $\ell^1$-minimization, which allows it to be compared directly to existing techniques. In particular, we reduce (b) to a question about the behaviour of three particular quantities that depend on the data points $\{ t_n \}^{N}_{n=1}$, the expansion basis $\{ \phi_i \}_{i \in \bbN}$ and the weights $\{ w_i \}_{i \in \bbN}$. Analyzing these quantities for each specific problem setup provides an answer to (b). To illustrate the various aspects of this framework, in the final part of this paper (\S \ref{s:AlgPoly_Examp} and \S \ref{s:TrigPoly_Examp}) we consider several examples, including one-dimensional Jacobi polynomial approximations from scattered data points. In this case, we prove the following: \thm{ \label{t:intro_thm} For $\alpha,\beta > -1$ let $\{ \phi_i \}_{i \in \bbN}$ be the orthonormal Jacobi polynomial basis \R{Jacobi_ON} on $[-1,1]$ and let $T = \{ t_n \}^{N}_{n=1}$ be a set of $N$ scattered points in $[-1,1]$. Let $h$ be the density of the points, defined by \R{h_def}, and suppose that the truncation parameter \be{ \label{trunc_choice} K \gtrsim h^{\frac{1}{2r}} \xi^{-1-\frac{1}{r}}, } for some $r \in \bbN$, where $\xi$ is the minimal separation between the points $T$. Fix weights $w = \{ w_i \}_{i \in \bbN}$ with $w_{i} = \| \phi_i \|_{L^\infty} i^{\gamma}$ for some $\gamma > \max\{1/2 - q,0\}$, where $q$ is as in \R{Jacobi_ON_growth}, and let $f = \sum_{i \in \bbN} x_i \phi_i$ with $x \in \ell^1_{\tilde{w}}(\bbN)$ for $\tilde{w}_i = \sqrt{i} (w_i)^2$. Then given measurements $y = \{ f(t_n) \}^{N}_{n=1}$ one can compute, via weighted $\ell^1$ minimization with weights $w$, an approximation $\hat{x}$ to the coefficients $x$ satisfying \be{ \label{intro_err} \| x - \hat{x} \| \lesssim \| x - P_M x \|_{1,w} + \| x - P_K x \|_{1,\tilde{w}}, } where $P_K x = \{ x_1,\ldots,x_K,0,0,\ldots \}$ and $P_M x = \{x_1,\ldots,x_M,0,0,\ldots\}$, provided \be{ \label{intro_h_scale} h^{-1} \gtrsim M^2 \log M. } Moreover, the approximation $\tilde{f} = \sum^{K}_{i=1} \hat{x}_i \phi_i$ exactly interpolates the data: $\tilde{f}(t_n) = f(t_n)$, $\forall n$. } This theorem demonstrates the key aspects of this paper. (i): the truncation parameter $K$ is determined independently of $f$, and its contribution to the overall error is clarified by \R{intro_err}. In particular, if the data is roughly equally-spaced, then $h, \xi = \ord{1/N}$ and it suffices to take $K = \ordu{N^{1+\frac{1}{2r}}}$ for any $r>0$. (ii): the approximation $\tilde{f}$ exactly interpolates $f$ in the absence of noise. Note that noise can also be dealt with within our framework; we exclude it here for ease of presentation. (iii): one gets an explicit criterion for how to choose the weights. (iv): the estimate \R{intro_err} for the approximation error depends only on the density $h$ of the deterministic points $T$ which can be arbitrarily distributed in the domain. As we discuss in \S \ref{s:AlgPoly_Examp}, the estimates \R{intro_err} and \R{intro_h_scale} demonstrate not just good performance of this approach for scattered data, but in fact near-optimal performance. As we explain, no stable method which is convergent as $h \rightarrow 0$ can exhibit an error bound depending on $x - P_M x$ measured in some norm with $M$ growing faster than $h^{-1/2}$ as $h \rightarrow \infty$. Our numerical results support this conclusion, and in fact show that weighted $\ell^1$ minimization performs rather better in practice and similarly to an oracle least-squares fit. \subsection{Relation to previous work}\label{ss:relation} A theory for reconstruction of sparse polynomials from random pointwise samples was developed in a series of papers by Rauhut \& Ward \cite{Rauhut,RauhutWardWeighted}. Extension and application of this work in uncertainty quantification has been considered in \cite{AdcockCSFunInterp,ChkifaDownwardsCS,DoostanOwhadiSparse,HamptonDoostanCSPCE,MathelinGallivanCSPDErandom,PengHamptonDoostantweighted,YanGuoXui_l1UQ,KarniadakisUQCS}. The use of weighted $\ell^1$ minimization was introduced in \cite{PengHamptonDoostantweighted,RauhutWardWeighted,KarniadakisUQCS}. We also use weighted minimization in this paper for approximation from determinstic samples, yet for rather different purposes. Namely, weights are chosen to regularize the minimization problem and remove the aliasing phenomenon. Typically, this requires only very slow growth of the weights, which we quantify in the paper. Unlike other works, we do not select weights based on \textit{a priori} information about decay of the polynomial coefficients. In fact, in \S \ref{s:need} we will show that choosing weights in this way leads to inconsistent and often negligible improvements when the samples are scattered and deterministic. The infinite-dimensional framework we introduce in this paper is inspired in part by the framework of infinite-dimensional compressed sensing in Hilbert spaces, due to A.\ C.\ Hansen and the present author \cite{BAACHShannon,BAACHGSCS,BAACHOptimality} (see also \cite{BAGSAIEP} for an overview). A key difference is the need for weighted minimization in the present setup, due to the lack of continuity of the sampling operator. We note also that our worst case analysis and comparison to least-squares fitting is similar to that presented in \cite{GSl1} for generalized sampling in the Hilbert space setting. The examples we use in this paper consist of algebraic and trigonometric polynomials respectively. Polynomial approximations (so-called polynomial chaos expansions) are popular in areas such as uncertainty quantification \cite{SMUQ,XiuKarniadakisPC}. However, we stress that the framework and analysis of \S \ref{s:prelim}--\ref{s:lin_approx_err} of this paper is completely general, and can be applied to other bases. We mention several other examples in \S \ref{s:conclusions}. Our examples are also one-dimensional. We do this so as to better elucidate the key ideas, without the notational complexities of the multivariate setting. On this topic, we wish to clarify that the aim of this paper is not to propose weighted $\ell^1$ minimization as a panacea for function approximation. In the one-dimensional setting especially there is a wealth of other techniques which are likely superior (see \cite{FEStability,AdcockPlatteMapped,BoydRunge,TrefPlatteIllCond} and references therein). The advantages of weighted $\ell^1$ minimization come to the fore as the dimension increases, as has been verified empirically in a number of works such as those mentioned previously. Instead, the purpose of this paper is to first propose a framework for weighted $\ell^1$ minimization that overcomes some existing issues, and second provide a more comprehensive analysis of its approximation capabilities for fixed samples. We use the one-dimensional case to this end primarily for illustrative purposes. \section{Preliminaries}\label{s:prelim} Let $D \subseteq \bbR^d$ be a domain and $\nu(t)$ an integrable nonnegative weight function satisfying $\int_D \nu(t) \D t = 1$. Let $L^2_{\nu}(D)$ be the space of complex-valued weighted square-integrable functions on $D$ with norm $\nm{\cdot}_{L^2_{\nu}}$ and inner product $\ip{\cdot}{\cdot}_{L^2_{\nu}}$, and suppose that $\{ \phi_i \}_{i \in \bbN} \subseteq L^2_{\nu}(D) \cap L^\infty(D)$ is a set of functions that are orthonormal with respect to $\nu$. Note that \be{ \label{phii_unif} 1 = \| \phi_i \|^2_{L^2_{\nu}} \leq \| \phi_i \|^2_{L^{\infty}},\qquad \forall i \in \bbN, } where $\nm{\cdot}_{L^{\infty}}$ is the uniform norm on $D$. \subsection{Scattered data} For $N \in \bbN$, let $T = \{ t_n \}^{N}_{n=1} \subseteq \overline{D}$ be a set of $N$ scattered data points. Our aim is to approximate a function $f : D \rightarrow \bbC$ from the values $\{ f(t_{n}) \}^{N}_{n=1}$. To ensure an accurate approximation, we require a notion of closeness of the points $T$. We quantify this by defining the \textit{density} \be{ \label{h_def} h = \sup_{t \in D} \min_{n=1,\ldots,N} | t - t_{n} |, } (also known as the fill distance \cite{WendlandScatteredBook}) where $\abs{\cdot}$ is the Euclidean distance. In our analysis later, we will present convergence rates of the various approximations in terms of $h \rightarrow 0$. Associated to the points $T$ will also be a set of values $\tau_{n} \geq 0$, $n = 1,\ldots,N$, which we refer to as \textit{quadrature weights}. This is not to be confused with the optimization weights $w_i$ introduced later. For simplicity, we define these as follows \be{ \label{quad_weights_1} \tau_{n} = \int_{V_{n}} \nu(t) \D t,\quad n=1,\ldots,N,\qquad V_{n} = \left \{ t \in D : | t - t_n | \leq | t - t_m | , \ \forall m \neq n \right \}, } where $V_{n}$ are the Voronoi cells of the points $T$ in $D$. Given such quadrature weights, we define the following sesquilinear form on $L^2_{\nu}(D) \cap L^\infty(D)$: \bes{ \ip{f}{g}_h = \sum^{N}_{n=1} \tau_{n} f(t_{n}) \overline{g(t_{n}) }, } and write $\nm{\cdot}_h = \sqrt{\ip{\cdot}{\cdot}_h}$ for the corresponding seminorm. Note that the quadrature weights $\tau_n$ are not strictly necessary at this stage, but will play a pivotal role later in the paper. \subsection{Weighted spaces}\label{ss:weighted_spaces} For the remainder of this paper, $w = \{ w_i \}_{i \in \bbN}$ will be a set of positive weights satisfying \be{ \label{weights_growth} w_i \geq \| \phi_i \|_{\infty} \geq 1,\quad \forall i \in \bbN, } where the latter inequality is due to \R{phii_unif}. Define the weighted $\ell^p$ spaces by \bes{ \ell^p_w(\bbN) = \left \{ x = \{ x_i \}_{i \in \bbN} : \ \| x \|_{p,w} : = \left ( \sum_{i \in \bbN} (w_i)^p | x_i |^p \right )^{1/p} < \infty \right \},\quad p > 0. } Note that $\| x \|_{p,w} = \| W x \|_{p}$, where \be{ \label{W_def} W = \mathrm{diag}(w_1,w_2,\ldots), } is the infinite diagonal matrix of weights. For the remainder of this paper, we will assume that the function we wish recover $f = \sum_{i \in \bbN} x_i \phi_i \in L^2_{\nu}(D)$ has coefficients $x = \{x_i \}_{i \in \bbN} \in \ell^1_w(\bbN)$. \subsection{Other notation} For $\Delta \subseteq \bbN$ we let $P_{\Delta} : \ell^2(\bbN) \rightarrow \ell^2(\bbN)$ be the projection defined by \bes{ (P_{\Delta} x)_j = x_j,\quad j \in \Delta,\qquad (P_{\Delta} x)_j =0,\quad j \notin \Delta. } If $\Delta = \{1,\ldots,K \}$ for some $K \in \bbN$, then we merely write $P_K$. We also let $\{ e_j \}_{j \in \bbN}$ denote the canonical basis of $\ell^2(\bbN)$, so that \bes{ P_{\Delta}(\cdot) = \sum_{j \in \Delta} \ip{\cdot}{e_j} e_j. } We will allow the slight abuse of notation throughout the paper in thinking of $P_{\Delta} x$ as both an element of $\ell^2(\bbN)$ and $\bbC^{|\Delta|}$. The intended meaning will be clear from the context. If $x \in \bbC$, we let $\mathrm{sign}(x) = x / | x |$ be its complex sign with the convention that $\mathrm{sign}(0) =0$. For $x \in \ell^\infty(\bbN)$ we let $\mathrm{sign}(x) = \{ \sgn(x_i) \}_{i \in \bbN} \in \ell^\infty(\bbN)$ be the corresponding sequence of complex signs of the entries of $x$. Finally, we use the notation $a \lesssim b$ to mean that there exists a constant $C$ independent of all relevant parameters such that $a \leq C b$. \subsection{Examples}\label{ss:examples} As mentioned in \S \ref{ss:relation}, the examples we consider in this paper consist of one-dimensional functions on bounded intervals, which we take to be $D=(-1,1)$ without loss of generality. In \S \ref{s:conclusions} we briefly discuss extensions to higher dimensions, unbounded intervals and other approximation systems. \examp{ \label{ex:Jacobi} If $f$ is smooth, then it is natural to approximate it using a basis of orthogonal polynomials. Let \bes{ \nu(t) = \nu^{(\alpha,\beta)}(t) = c^{(\alpha,\beta)}(1-t)^\alpha(1+t)^{\beta},\qquad \alpha,\beta > -1, } be the Jacobi weight function, where $c^{(\alpha,\beta)} = \left( \int^{1}_{-1} (1-t)^\alpha(1+t)^{\beta} \D t \right )^{-1}$ is a normalizing constant, $P^{(\alpha,\beta)}_j$ be the $j^{\rth}$ orthogonal polynomial with respect to this weight function, and \be{ \label{Jacobi_ON} \phi_j = \left ( \kappa^{(\alpha,\beta)}_{j-1} \right )^{-1/2} P^{\alpha,\beta}_{j-1}, } be the corresponding orthonormal polynomial, where $\kappa^{(\alpha,\beta)}_j$ is as in \R{Jacnorm_def}. One can show that \be{ \label{Jacobi_ON_growth} \| \phi_j \|_{L^\infty} = \ord{j^{q+1/2}},\quad j \rightarrow \infty,\qquad \mbox{where $q = \max \{ \alpha,\beta , -1/2 \}$}. } See Appendix \ref{a:Jacobi} (several other properties of Jacobi polynomials that will be needed later are also listed therein). Since the weights $\{ w_i \}_{i \in \bbN}$ introduced in \S \ref{ss:weighted_spaces} are required to satisfy \R{weights_growth}, this means that for this example they must grow at least as fast as $j^{q+1/2}$ as $j \rightarrow \infty$. } \examp{ \label{ex:Trigonometric} Functions that are smooth and periodic can be efficiently approximated using trigonometric polynomials. In this case, we have $\nu(t) = 1/2$ and define $\{ \phi_i \}_{i \in \bbN}$ to be the Fourier basis \be{ \label{fourier_basis} \phi_j(t) =\E^{\I j \pi t},\quad j \in \bbZ. } For convenience we index over $\bbZ$ rather than $\bbN$ in this example. Note that $\| \phi_j \|_{\infty}=1$ and therefore the weights $w_j$ in this example are required to satisfy $w_j \geq 1$, $\forall j \in \bbZ$. } \section{Minimization problems}\label{s:minproblems} Define the operator $U : \ell^1_{w}(\bbN) \rightarrow \bbC^N$ by $U x = \{ \sqrt{\tau_{n}} g(t_{n}) \}^{N}_{n=1}$ where $g = \sum_{i \in \bbN} x_i \phi_i$. Note that this operator is bounded. We shall also view $U \in \bbC^{N \times \infty}$ as the infinite matrix with entries \bes{ U_{n,i} = \sqrt{\tau_{n}} \phi_i(t_{n}),\quad n=1,\ldots,N,\ i \in \bbN. } From now on, we make no distinction between the operator $U$ and the infinite matrix. \subsection{Infinite-dimensional weighted $\ell^1$ minimization} Let $f = \sum_{i \in \bbN} x_i \phi_i$ be a function we wish to recover, where $x = \{ x_i \}_{i \in \bbN} \in \ell^1_w(\bbN)$. Suppose first that we are given noiseless measurements of $f$, that is, $f(t_{n})$, $n=1,\ldots,N$, and let \bes{ y = \left \{ \sqrt{\tau_{n}} f(t_{n}) \right \}^{N}_{n=1} \in \bbC^N, } be the vector of measurements normalized by the quadrature weights. To recover the infinite vector $x$ of coefficients, and therefore $f$, we shall use weighted $\ell^1$ minimization. In order to avoid issues of truncation (recall \S \ref{ss:currentapproaches}), we first formulate the following infinite-dimensional optimization problem: \be{ \label{inf_min_noiseless} \inf_{ z \in \ell^1_w(\bbN)} \| z \|_{1,w}\ \mbox{subject to $U z = y$}. } If $\hat{x} \in \ell^1_w(\bbN)$ is a minimizer of \R{inf_min_noiseless}, then the corresponding approximation $\tilde{f}$ to $f$ is given by \be{ \label{f_N_inf} \tilde{f} = \sum_{i \in \bbN} \hat{x}_i \phi_i. } In general, the measurements may be noisy. Suppose we are given \bes{ f(t_n) + e_{n},\quad n=1,\ldots,N, } where $|e_n| \leq \eta$, $n=1,\ldots,N$, for some known $\eta \geq 0 $. Write \be{ \label{y_N_def} y = \{ \sqrt{\tau_n} \left ( f(t_n) + e_n \right ) \}^{N}_{n=1} \in \bbC^N. } In this case, we solve the inequality-constrained optimization problem \be{ \label{inf_min} \inf_{ z \in \ell^1_w(\bbN)} \| z \|_{1,w}\ \mbox{subject to $\| U z - y \| \leq \eta$}. } Note that \R{inf_min_noiseless} is just a special case of \R{inf_min} corresponding to $\eta = 0$, and that both \R{inf_min_noiseless} and \R{inf_min} always have a solution, since the feasible set nonempty (specifically, $x$ is always feasible). Note also that solutions of \R{inf_min_noiseless} are interpolatory in the sense that $\tilde{f}(t_n) = f(t_n)$, $n=1,\ldots,N$, whenever $\tilde{f}$ is given by \R{f_N_inf} with $\hat{x}$ being a minimizer of \R{inf_min_noiseless}. Conversely, solutions of \R{inf_min} yield approximations $\tilde{f}$ that are interpolatory up to the noise magnitude $\eta$. Throughout, we shall assume that the noise bound $\eta$ is known. If $\eta$ is unknown, one may still solve the equality-constrained problem \R{inf_min_noiseless} in practice (or the inequality-constrained problem \R{inf_min} with some estimate of the noise). However, there are no known recovery guarantees for this problem. See \cite[Chpt.\ 11]{FoucartRauhutCSbook} for some work in this direction in the context of finite-dimensional compressed sensing with random Gaussian matrices. \subsection{Truncation} Unfortunately, neither problem \R{inf_min_noiseless} or \R{inf_min} is numerically solvable, since they require optimizing over an infinite-dimensional space. Let $K \in \bbN$ be a truncation parameter. To form a computable problem, we replace the space $\ell^1_{w}(\bbN)$ with $\bbC^K$ and truncate the $N \times \infty$ matrix $U$ to the $N \times K$ matrix $U P_K$ spanned by its first $K$ columns. Hence, we now consider the problem \be{ \label{fin_min_noiseless} \min_{ z \in \bbC^K} \| z \|_{1,w}\ \mbox{subject to $U P_{K} z = y $}, } in the noiseless case, as well as its noisy analogue \be{ \label{fin_min} \min_{ z \in \bbC^K} \| z \|_{1,w}\ \mbox{subject to $\| U P_{K} z - y \| \leq \eta $}. } Both of these problems are finite dimensional, and can be solved using standard algorithms. If $\hat{x} \in \bbC^K$ is a minimizer of either, then the approximation to $f$ is given by \be{ \label{f_N_K} \tilde{f} = \sum^{K}_{i=1} \hat{x}_i \phi_i. } Note that neither \R{fin_min_noiseless} nor \R{fin_min} modify the constraints of the infinite-dimensional problems \R{inf_min_noiseless} and \R{inf_min}. In particular, \R{fin_min_noiseless} remains interpolatory and \R{inf_min} is interpolatory up to the noise. With this in hand, the general idea is to choose $K$ in such a way to ensure closeness of the solutions of the finite-dimensional problems \R{fin_min_noiseless} and \R{fin_min} to those of the infinite-dimensional problems \R{inf_min_noiseless} and \R{inf_min}. In \S \ref{s:truncation} we shall show that it is possible to choose $K$ in a function-independent manner, thus overcoming issue (i) of \S \ref{ss:currentapproaches}. \rem{ It is important that $K$ be chosen suitably large. To see why, consider Example \ref{ex:Jacobi}. If $K=N$ then \R{fin_min_noiseless} has a unique solution and $\tilde{f}$ is just the polynomial interpolant of $f$ of degree $N-1$. However, for equispaced (or more generally, scattered) data this is well known to be a poor approximation to $f$, since it suffers from Runge's phenomenon. The approximations $\tilde{f}$ will generally diverge and the matrix $U P_N$ will have an exponentially-large condition number. On the other hand, if one replaces $U P_N$ by $U P_K$ with $K > N$, then provided $K$ is sufficiently large the singular values of $UP_K$ are provably bounded away from zero (see \S \ref{s:truncation}). } \subsection{Comparison to least-squares fitting}\label{ss:LS_comp} Classical least-squares fitting corresponds to the approximation $\tilde{f} = \sum^{M}_{i=1} \check{x}_i \phi_i$, where $\check{x}$ is the solution of the problem \be{ \label{LS_fit} \min_{z \in \bbC^M} \| U P_M z - y \|. } An important difference between least-squares fitting and weighted $\ell^1$ minimization is the choice of the truncation parameters. In the former, the parameter $M$ affects both the approximation error $\| f - \tilde{f} \|$ and the robustness of the approximation. In practice, $M$ must be chosen suitably small in relation to $1/h$ to ensure a stability and robustness, while also being sufficiently large to give a good approximation. The issue of how to best choose $M$, which we discuss further in \S \ref{ss:LS_fit_comp} for the specific case of polynomials, is nontrivial. While there are many known theoretical estimates for how $M$ should scale for different function systems and datasets (see, for example, \cite{AdcockPlatteMapped,AdcockNecSamp,DavenportEtAlLeastSquares,DemanetTownsendExtrap,MiglioratiNobileLowDisc,MiglioratiEtAlFoCM} and references therein), optimal selection of $M$ is difficult in practice. In particular, standard theoretical guarantees usually only determine the asymptotic order of $M$ with $1/h$. Constants, if known, tend to be overly pessimistic. This problem becomes more acute in multiple dimensions, since the ordering of the basis functions plays an increasingly important role. Conversely, the truncation parameter $K$ in the weighted $\ell^1$ minimization formulations \R{fin_min_noiseless} and \R{fin_min} plays a completely different role: namely, it allows one to approximately compute solutions of the infinite-dimensional problems \R{inf_min_noiseless} and \R{inf_min}. Once $K$ is large enough so that the truncation error when passing to the finite-dimensional problems \R{fin_min_noiseless} and \R{fin_min} is negligible, changing $K$ has little effect on the accuracy of the solution $\tilde{f}$. Note also that \R{fin_min_noiseless} leads to interpolatory approximations, which is not the case for the least-squares fit \R{LS_fit}. \section{The need for weights}\label{s:need} Before analyzing \R{inf_min_noiseless} and \R{fin_min} in detail, we first examine the role weights play in the minimization. In particular, we shall show that it is in general necessary for the ratios \be{ \label{weights_growth2} w_{i} / \| \phi_i \|_{\infty} \rightarrow \infty,\quad i \rightarrow \infty, } in order for the weighted minimization problems to give convergent approximations to $f$ in the case of fixed, deterministic data. If this is not the case, then the minimization problem can have multiple solutions which aliase the data, leading in general to poor approximations. In \S \ref{s:lin_approx_err} we shall prove that \R{weights_growth2} is sufficient to guarantee a good approximation. To demonstrate its general necessity, we consider Example \ref{ex:Trigonometric}. Recall that $\| \phi_j \|_{\infty}=1$ in this case. \prop{ \label{p:aliasing} Let $D$, $\nu$ and $\{ \phi_j \}_{j \in \bbZ}$ be as in \R{fourier_basis}. Let $T = \{ t_n \}^{N}_{n=1}$ be a set of $N$ data points such that $t_n P \in \bbZ$ for some $P \in \bbN$ and all $n=1,\ldots,N$. Suppose that $\hat{x} \in \ell^1(\bbZ)$ is a solution of \be{ \label{bad_prob} \inf_{z \in \ell^1(\bbZ)} \| z \|_{1}\ \mbox{subject to $U z = y$}, } where $U = \{ \phi_j(t_n) \}^{N,\infty}_{n=1,j=-\infty}$ and $y \in \bbC^N$. Then every shift of the entries of $\hat{x}$ by a multiple of $2P$ is also a solution of \R{bad_prob}. That is, for every $k \in \bbZ$, the element $z \in \ell^1(\bbZ)$ given by \be{ \label{aliased_solutions} z_{i} = \hat{x}_{i-2kP},\quad i \in \bbZ, } is a solution of \R{bad_prob}. } \prf{ Shifting the entries of $x$ does not affect its $\ell^1$ norm, therefore $\| z \|_1 = \| x \|_1$. Moreover, \bes{ (U z)_n = \sum_{j \in \bbZ} z_j \E^{\I j \pi t_n} = \sum_{j \in \bbZ} x_j \E^{\I (j+2k P) \pi t_n} = \sum_{j \in \bbZ} x_j \E^{\I j \pi t_n} = (U x)_n = y_n,\quad n=1,\ldots,N. } Hence $z$ is feasible for \R{bad_prob}, and therefore a minimizer. } The absence of quadrature weights $\tau_n$ does not change the conclusion here, since we consider the equality-constrained minimization. We could also consider the inequality-constrained problem with much the same result, but we present the equality-constrained problem to show that the phenomenon is not due in any way to the increased size of the feasible set when $\eta > 0$. Taken on its own, the fact that \R{bad_prob} has multiple solutions may not be alarming. After all, convex optimization problems often do. However, in this case the effect is catastrophic. Consider the problem where $f = \phi_0 = 1$ so that its coefficients are $x = e_0$. Since $y_n = f(t_n) = 1$ in this case, if $z \in \ell^1(\bbZ)$ is feasible for \R{bad_prob} then \bes{ 1 = |(U z)_n| = \left | \sum_{j \in \bbZ} z_j \E^{\I j \pi t_n } \right | \leq \| z \|_1. } Hence $x$ itself is a solution of \R{bad_prob}, and by Proposition \ref{p:aliasing} so is every shift $z = e_{2kP}$ of $x$ by a multiple of $2P$. However, for all these solutions one has $\| x - z \| = 2$. Thus, although there is one solution of \R{bad_prob} which recovers $x$ (and therefore $f$) exactly there are also infinitely many solutions of \R{bad_prob} that give meaningless approximations to $x$. \begin{figure}[t] \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{Fig1_N11} & \hspace{2pc} &\includegraphics[width=5.00cm]{Fig1_N21} \end{array}$ \caption{Aliasing in $\ell^1$ minimization. The black line is the function $f(t) = \phi_0 = 1$, the red dots are the data points with $N=11$ (left) and $N=21$ (right) and the blue curve is the aliased solution $\phi_{10}$ (left) and $\phi_{20}$ (right). Although these solutions interpolate $f$, they do not approximate $f$ in between the data points.} \label{f:FourAlias} \end{center} \end{figure} This effect is due to aliasing the data by higher-frequency Fourier modes. The shifted solutions of Proposition \ref{p:aliasing} correspond to the functions $\phi_{k P}$, $k \in \bbZ$, which interpolate $f$ at the data points but oscillate with frequency proportional to $k P$ in between the data points (see Fig.\ \ref{f:FourAlias}). Of course, in the simplified scenario described here the aliasing problem could have been avoided by solving a truncated problem with truncation $K = P$. However, as discussed, this will not work in the general case when truncation with $K \gg N$ is required in order to control the tail and ensure (in the noiseless case) an interpolatory solution. Now suppose that weights $w_i$ are added, and \R{bad_prob} is replaced by \be{ \label{good_problem} \inf_{z \in \ell^1_w(\bbZ)} \| z \|_{1,w}\quad \mbox{subject to $U z = y$}. } Assume the weights $w$ satisfy $w_{-i} = w_{i}$, $i \in \bbN$ and $1 \leq w_0 < w_1 < w_2< \ldots$, and consider the case of $f = \phi_0$ once more. Then none of the aliased solutions of \R{bad_prob} are solutions of \R{good_problem}, since they all have larger weighted $\ell^1$-norm: $\| e_{2kP} \| = w_{2kP} > w_0 = \| e_0 \|$. Hence, adding growing weights regularizes the problem \R{good_problem} and removes the bad, aliased solutions of \R{bad_prob}. This improvement is illustrated numerically in Fig.\ \ref{f:Alias2}. This figure also shows that this phenomenon is not limited to the equality-constrained minimization problem. \begin{figure}[t] \begin{center} $\begin{array}{ccc} \includegraphics[width=4.5cm]{Fig2_nonoise_1} & \includegraphics[width=4.5cm]{Fig2_nonoise_2} &\includegraphics[width=4.5cm]{Fig2_nonoise_3} \\ \mbox{\small $\| f - \tilde{f} \|_{L^\infty} =\mbox{5.32e-1}$} &\mbox{\small $ \| f - \tilde{f} \|_{L^\infty} =\mbox{1.63e-01}$} & \mbox{\small $\| f - \tilde{f} \|_{L^\infty} = \mbox{9.66e-04}$} \\\\ \includegraphics[width=4.5cm]{Fig2_noise_1} & \includegraphics[width=4.5cm]{Fig2_noise_2} &\includegraphics[width=4.5cm]{Fig2_noise_3} \\ \mbox{\small $\| f - \tilde{f} \|_{L^\infty} =\mbox{5.38e-1}$} & \mbox{\small $\| f - \tilde{f} \|_{L^\infty} =\mbox{1.64e-01}$} & \mbox{\small $\| f - \tilde{f} \|_{L^\infty} = \mbox{1.02e-2}$} \end{array}$ \caption{Recovery of the function $f(t) = \cos(\pi x) \exp(\sin(\pi x))$ (shown in black) from $N=20$ data points (shown in red). The blue curve is the function $\tilde{f}$ obtained from (weighted) $\ell^1$ minimization using the Fourier basis with weights $w_{i} = 1$ (left), $w_i =1+|i|^{1/10}$ (middle) and $w_i = 1+|i|^{1/2}$ (right). Top row: equality-constrained minimization \R{inf_min_noiseless}. Bottom row: inequality-constrained minimization \R{inf_min} with $\eta = 10^{-2}$. } \label{f:Alias2} \end{center} \end{figure} \rem{ The use of weighted minimization strategies has been occasionally motivated by the desire to match the decay of the true coefficients $x$ of the unknown function and thereby obtain better approximations \cite{PengHamptonDoostantweighted,RauhutWardWeighted}. However, this is not the primary role the weights in play in this setting. To demonstrate this point, in Fig.\ \ref{f:Alias3} we plot the error for weighted $\ell^1$ minimization using Chebyshev polynomials for a number of different test functions and weighting strategies. As can be seen, increasing the weights does not lead to a consistent improvement across all functions, even though all functions used (beside the final one) are infinitely smooth and thus have coefficients which decay superalgebraically fast. While weights might help in some small way by promoting smoothness, these results suggest that the effect on the approximation error is much less than the role they play in regularizing the problem. Furthermore, higher weights may well cause problems for numerical solvers, due to the increasing ill-conditioning of the $N \times K$ system matrix $U W^{-1} P_K$. We note in passing that this situation is quite unlike the case of weighted $\ell^2$ minimization (see, for example, \cite{MSNTrigPoly,MSNJCP}), in which case the error for a smooth function decays only algebraically fast at a rate dependent on the algebraic growth rate of the weights. Thus, for $\ell^2$ minimization, rapidly-growing weights promote smoothness. Conversely, we will prove later that for weighted $\ell^1$ minimization the error decays superalgebraically fast for all smooth functions whenever the weights meet a minimum growth condition (Theorem \ref{t:Leg_poly_full_l1}). } \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=4.5cm]{Fig3_f1} & \includegraphics[width=4.5cm]{Fig3_f2} & \includegraphics[width=4.5cm]{Fig3_f3} \\ \mbox{\small $f(t) = \frac{1}{1+25 t^2}$} & \mbox{\small $f(t) = \frac{1}{35-34t}$} & \mbox{\small $f(t) = \cos(30 t)$} \\ \includegraphics[width=4.5cm]{Fig3_f4} & \includegraphics[width=4.5cm]{Fig3_f5} & \includegraphics[width=4.5cm]{Fig3_f6} \\ \mbox{\small $f(t) = \frac{\cosh(30 t^2)}{\cosh(30)}$} & \mbox{\small $f(t) = \sqrt{1.01+t}$} & \mbox{\small $f(t) = t^5 \log(t^2)$} \end{array}$ \caption{Weighted $\ell^1$ minimization with equispaced data and Chebyshev polynomials. The error $\| f - \tilde{f} \|_{L^\infty}$ against $N$ is shown for the choice $w_{i} = i^{\gamma}$, where $\gamma = 0.0,0.05,0.1,0.25,0.5,0.75,1.0,1.5,2.0,2.5$ (thickest to thinnest). The truncation parameter $K=4N$ was used. As with all numerical results in this paper, the minimization problem \R{fin_min_noiseless} was solved using the CVX optimization package.} \label{f:Alias3} \end{center} \end{figure} \section{Approximation error of weighted $\ell^1$ minimization}\label{s:lin_approx_err} The remainder of this paper is devoted to the analysis of the problems \R{fin_min_noiseless} and \R{fin_min}. In this section, we present a linear approximation error analysis. Truncation and the choice of the parameter $K$ is addressed in \S \ref{s:truncation}, and in \S \ref{s:AlgPoly_Examp} and \S\ref{s:TrigPoly_Examp} we apply these results to the examples of \S \ref{ss:examples}. \subsection{A general recovery result} We first require the following result, which bounds the error of weighted $\ell^1$ minimization subject to the existence of a particular dual vector $u$: \lem{ \label{l:dual_certificate} Let $\Delta \subseteq \{1,\ldots,K\}$. Suppose that \bes{ (i): \| P_{\Delta} U^* U P_{\Delta} - P_{\Delta} \| \leq \alpha,\qquad (ii): \max_{i \notin \Delta} \left \{ \| U e_i \| / w_i \right \} \leq \beta, } and that there exists a vector $u = W^{-1} U^* u'$ for some $u' \in \bbC^N$, such that \bes{ (iii): \| W(P_{\Delta} u - \sgn(P_{\Delta} x)) \| \leq \gamma,\qquad (iv): \| P^{\perp}_{\Delta} u \|_{\infty} \leq \theta,\qquad (v): \| u' \| \leq L \sqrt{s}, } where $s = \sum_{i \in \Delta} (w_i)^2$, for constants $0 \leq \alpha , \theta < 1$ and $\beta , \gamma, L \geq 0$ satisfying \bes{ \frac{\sqrt{1+\alpha} \beta \gamma}{(1-\alpha)(1-\theta) } < 1. } Let $\hat{x}$ be a minimizer of \R{fin_min}. Then, if $\bar{x} \in \ell^1_w(\bbN)$ is feasible for \R{fin_min}, i.e. $\| U P_K \bar{x} - y \| \leq \eta$, the error estimate \be{ \label{l1_error_est} \| \hat{x} - x \| \leq 2\left ( C_1 + C_2 L \sqrt{s} \right ) \eta + C_2 \left (2 \| P^{\perp}_{\Delta} x \|_{1,w} + \| x - \bar{x} \|_{1,w} \right ), } holds, where $C_1 = \left ( 1 + \frac{\gamma}{1-\theta} \right ) C_0$, $C_2 = \frac{\beta}{1-\theta}\left ( 1 + \frac{\gamma}{1-\theta} \right ) C_0 + \frac{1}{1-\theta}$ and $C_0 = \left ( 1 - \frac{\sqrt{1+\alpha} \beta \gamma }{(1-\alpha)(1-\theta)} \right )^{-1} \frac{\sqrt{1+\alpha}}{1-\alpha}$. } Recall that the problem \R{fin_min_noiseless} can be viewed as a special case of \R{fin_min} corresponding to the case $\eta= 0$. Hence this result considers only \R{fin_min}. The proof of this lemma is given in \S \ref{ss:dual_certificate_proof}. In practice, we shall use following result, which is a straightforward consequence of Lemma \ref{l:dual_certificate}: \lem{ \label{l:Delta_recov} Let $\Delta \subseteq \{1,\ldots,K \}$. Suppose that there are constants $0 \leq \alpha,\theta < 1$ such that \bes{ (a): \| P_{\Delta} U^*U P_{\Delta} - P_{\Delta} \| \leq \alpha,\qquad (b): \| P^{\perp}_{\Delta} W^{-1} U^* U P_{\Delta} A^{-1} W P_{\Delta} \sgn(x) \|_{\infty} \leq \theta, } where $A = P_{\Delta} U^* U P_{\Delta}$. If $\hat{x}$ is a minimizer of \R{fin_min} and $s = \sum_{i \in \Delta} (w_i)^2$, then \be{ \label{err_bound} \| \hat{x} - x \| \leq 2 C_3 \left ( 1 + \frac{C_3+1}{1-\theta} \sqrt{s} \right ) \eta + \frac{C_3+1}{1-\theta} \left ( 2 \| P^{\perp}_{\Delta} x \|_{1,w} + \| x - \bar{x} \|_{1,w} \right ) , } where $\bar{x} \in \ell^1_{w}(\bbN)$ is any feasible solution of \R{fin_min} and $C_3 = \frac{\sqrt{1+\alpha}}{1-\alpha} $. } \prf{ We apply Lemma \ref{l:dual_certificate} with $u = W^{-1} U^* U P_{\Delta} A^{-1} W P_{\Delta} \sgn(x)$. Note that $(a)$ and $(b)$ imply $(i)$ and $(iv)$ respectively. Also, by construction, $P_{\Delta} u = P_{\Delta} \sgn(x)$ and therefore $(iii)$ holds with $\gamma = 0$. Now consider $(ii)$. By definition \bes{ \| U e_i \| = \| \phi_i \|_h \leq \| \phi_i \|_{\infty} \sqrt{\sum^{N}_{n=1} \tau_n } \leq w_i, } where the last inequality is due to \R{weights_growth} and the fact that $\sum^{N}_{n=1} \tau_n = 1$ when the weights $\tau_n$ are given by \R{quad_weights_1}. Hence $(ii)$ holds with $\beta = 1$. Finally, observe that \eas{ \| u' \| &= \| U P_{\Delta} A^{-1} W P_{\Delta} \sgn(x) \| \leq \| U P_{\Delta} \| \| A^{-1} \| \| W P_{\Delta} \sgn(x) \| \leq \frac{\sqrt{1+\alpha}}{1-\alpha} \sqrt{s}, } where the final inequality follows from $(a)$. Hence $(v)$ holds with $L = \frac{\sqrt{1+\alpha}}{1-\alpha}$. } \subsection{Linear approximation error preliminaries} Lemma \ref{l:Delta_recov} gives conditions under which $x$ is approximated with error depending on the magnitude of its coefficients $x_j$ outside some set $\Delta$. This depends on the conditioning of the corresponding submatrix (condition (a)) and the off-support magnitude of the vector $u$ (condition (b)). Under a sparsity condition on the coefficients $x$, and appropriate random choices of the points $T$, one may use this result to prove estimates relating the number of measurements to the sparsity \cite{AdcockCSFunInterp}. However, as discussed in \S \ref{s:introduction}, in practice the data points may not arise from such distributions. In this section, we consider arbitrary deterministic scattered data points and present a linear error analysis. This follows by setting $\Delta = \{1,\ldots,M \}$ for some $M \leq K$ and using Lemma \ref{l:Delta_recov} to determine how large $M$ can be chosen in relation to the density $h$ of the points. Doing this will allow us to make a direct comparison with other techniques (see Remark \ref{r:cooking_time}). Since such statements will be asymptotic in $h \rightarrow 0$, we first require an additional assumption. Let $H$ be a subspace of $L^2_{\nu}(D) \cap L^{\infty}(D)$ which is closed under multiplication and complex conjugation and such that $f \in H$ and $\{ \phi_i \}_{i \in \bbN} \subseteq H$. We now assume that the points $T$ satisfy \be{ \label{quad_conv} \sum^{N}_{n=1} \tau_n g(t_n) \rightarrow \int_{D} g(t) \nu(t) \D t,\quad h \rightarrow 0,\qquad \forall g \in H. } In particular, since $H$ is closed under multiplication and complex conjugation, one has that \bes{ \ip{f}{g}_{h} \rightarrow \ip{f}{g}_{L^2_{\nu}},\quad h \rightarrow 0,\qquad \forall f,g \in H. } Hence the discrete inner product is equivalent to $\ip{\cdot}{\cdot}_{L^2_{\nu}}$ on finite-dimensional subspaces of $H$ for sufficiently small $h$. Note that this assumption is by no means stringent. For example, if $D = (-1,1)$ we may take $H$ to be the space of functions for which $|f(t)|^2 \nu(t)$ is Riemann integrable. Before stating our main result (Theorem \ref{t:full_samp_recov}), we first require some additional notation and technical lemmas. For $h >0$ and $M,R \in \bbN$ we define the quantities \be{ \label{E_2NM} E_2(h,M) = \| P_M - P_M U^* U P_M \|,\quad E_{\infty}(h,M) = \| P_M - P_M U^* U P_M \|_{\infty} , } and \be{ \label{F_N_M_R} F(h,M,R) = \| P^{\perp}_R W^{-1} U^* U P_M \|_{\infty}. } We also set $E(h,M) = \max \{ E_2(h,M),E_{\infty}(h,M) \}$. \lem{ \label{E_decay} For fixed $M$, we have $E(h,M) \rightarrow 0$ as $h \rightarrow 0$. } \prf{ Since all norms on $\bbC^M$ are equivalent, it suffices to show that $(U^* U)_{i,j}\rightarrow \delta_{ij}$ for each $i,j=1,\ldots,M$ as $h \rightarrow 0$. However, by \R{quad_conv} and orthogonality of the $\phi_j$, we have $ (U^* U)_{i,j} = \ip{\phi_i}{\phi_j}_{h} \rightarrow \ip{\phi_i}{\phi_j}_{L^2_{\nu}} = \delta_{ij}, $ as required. } \lem{ \label{F_decay} Suppose that the weights $w_i = z_i \| \phi_i \|_{L^\infty}$ for some $z_i \geq 1$. Then \bes{ F(h,M,R) \leq \frac{\sqrt{M} \sqrt{1+E(h,M)}}{\inf_{i > R} \{ z_i \} }. } } \prf{ Let $x \in P_M(\ell^2(\bbN))$, $\| x \|_{\infty} = 1$ be arbitrary. Then $(W^{-1} U^* U P_M x)_{i} = \ip{g}{\phi_i}_{h}/w_i$, where $g = \sum^{M}_{j=1} x_j \phi_j$. By the Cauchy--Schwarz inequality, \bes{ \| P^{\perp}_R W^{-1} U^* U P_M x \|_{\infty} \leq \sup_{i > R} \left \{ \frac{1}{w_i} \| g \|_h \| \phi_i \|_h \right \}. } Now $\| g \|^2_h = \ip{x}{P_M U^* U P_M x} \leq \left ( 1 + E(h,M) \right ) \| x \|^2 \leq M \left ( 1 + E(h,M) \right )$. Also $\| \phi_i \|_h \leq \| \phi_i \|_{L^\infty} = w_i / z_i$ and therefore \bes{ \| P^{\perp}_R W^{-1} U^* U P_M \|_{\infty} \leq \frac{\sqrt{M} \sqrt{1+E(h,M)}}{\inf_{i > R} \{ z_i \} }, } as required. } \lem{ Suppose that the weights $w_i = z_i \| \phi_i \|_{L^\infty}$, where $z_i \geq 1$ and $z_i \rightarrow \infty$ as $i \rightarrow \infty$. Then for any $0 < \epsilon < 1/2$ and any $M \in \bbN$ there exists a $R \in \bbN$ and $h > 0$ depending only on $M$ and $\epsilon$ such that \be{ \label{NMR_condition} E(h,M) < \epsilon,\quad E(h,R) < \epsilon \frac{\min_{M < i \leq R} \{ w_i \} }{\max_{i=1,\ldots,M} \{ w_i \} }, \quad F(h,M,R) \leq \frac{\epsilon}{\max_{i=1,\ldots,M} \{ w_i \} }. } } \prf{ By Lemma \ref{E_decay} we can find an $h_1$ such that $E(h,M) < \epsilon$ for all $h \leq h_1$, thus satisfying the first condition in \R{NMR_condition}. Using Lemma \ref{F_decay}, we note that \bes{ F(h,M,R) \leq \frac{\sqrt{2M}}{\min_{i > R} \{ z_i \}},\quad \forall h \leq h_1. } To satisfy the third condition in \R{NMR_condition}, we pick $R$ sufficiently large so that \bes{ \inf_{i > R} \{ z_i \} > \sqrt{2M} \max_{i=1,\ldots,M} \{ w_i \} / \epsilon. } We now pick $h_2$ sufficiently small so that \bes{ E(h,R) < \epsilon \frac{\min_{M < i \leq R} \{ w_i \} }{\max_{i=1,\ldots,M} \{ w_i \} },\quad \forall h \leq h_2, } and then set $h = \min \{ h_1,h_2 \}$. } \subsection{Main result} In order to present our main result, we first introduce the following notation: \be{ \label{trunc} T_{h,K,\eta}(x) = \inf \left \{ \| x - \bar{x} \|_{1,w} : \bar{x} \in \bbC^K, \| U P_K \bar{x} - y \| \leq \eta \right \}. } \thm{ \label{t:full_samp_recov} Suppose that the weights $w_i = z_i \| \phi_i \|_{L^\infty}$, where $z_i \geq 1$ and $z_i \rightarrow \infty$ as $i \rightarrow \infty$. For $0 < \epsilon < 1/2$, let $h>0$ and $M,R \in \bbN$ be such that \R{NMR_condition} holds. Then there exists a constant $C(\epsilon)$ such that, whenever $\hat{x}$ is a minimizer of \R{fin_min}, \be{ \label{l1w_err_bd} \| x - \hat{x} \| \leq C(\epsilon) \left [ \left ( 1 + \| P_M w \| \right ) \eta + \| P^{\perp}_M x \|_{1,w} + T_{h,K, \eta }(x) \right ], } where $T_{h,K, \eta }(x)$ is as in \R{trunc}. Moreover, $\lim_{\epsilon \rightarrow 0^+} C(\epsilon) = 4$. } Note that the weights condition is equivalent to \R{weights_growth2} which was shown to be necessary in \S \ref{s:need}. Theorem \ref{t:full_samp_recov} shows that the same condition is also sufficient. We note also that the truncation parameter $K$ only influences the term $T_{h,K,\eta}(x)$. We defer the detailed analysis of this term to \S \ref{s:truncation}. \prf{ We use Lemma \ref{l:Delta_recov} with $\Delta = \{1,\ldots,M\}$. Note first that \be{ \label{alpha_bound} E(h,M) < \epsilon <1, } and therefore $(a)$ holds with $\alpha \leq \epsilon$. Now consider $(b)$. Write \be{ \label{u_full_def} u = W^{-1} U^* U P_M A^{-1} W P_M \sgn{(x)}, } so that $(b)$ is equivalent to $\| P^\perp_M u \|_{\infty} \leq \theta$. We have \be{ \label{R_splitting} \| P^{\perp}_M u \|_{\infty} = \max \left \{ \| P_R P^{\perp}_M u \|_{\infty} , \| P^{\perp}_R u \|_{\infty} \right \}. } We consider both terms separately. For $\| P_R P^{\perp}_M u \|_{\infty}$, \R{u_full_def} gives \eas{ \| P_R P^{\perp}_M u \|_{\infty} &\leq \| P_R P^{\perp}_M W^{-1} \|_{\infty} \| P_R P^{\perp}_M U^* U P_M \|_{\infty} \| A^{-1} \|_{\infty} \| P_M W \|_{\infty} \| \sgn{(x)} \|_{\infty} \\ & \leq \frac{\max_{1 \leq i \leq M} \{ w_i \} }{\min_{M<i\leq R} \{ w_i \}} \| P_R P^{\perp}_M U^* U P_M \|_{\infty} \| A^{-1} \|_{\infty}. } Note that $\| I - A \|_{\infty} \leq E(h,M)$ and therefore $\| A^{-1} \|_{\infty} \leq \frac{1}{1-E(h,M)}$. Moreover \eas{ \| P_R P^{\perp}_M U^* U P_M \|_{\infty} &= \| P_R P^{\perp}_M (I - U^* U ) P_M \|_{\infty} \leq \| P_R (I - U^* U ) P_R \|_{\infty} \leq E(h,R), } where the first equality is due to the fact that $P^{\perp}_M P_M = 0$. Therefore, we obtain \be{ \label{M_R_split_term} \| P_R P^{\perp}_M u \|_{\infty} \leq \frac{\max_{1 \leq i \leq M} \{ w_i \} }{\min_{M<i\leq R} \{ w_i \}} \left ( \frac{E(h,R)}{1-E(h,M)} \right ) \leq \frac{\epsilon}{1-\epsilon}. } Now consider the other term in \R{R_splitting}. By \R{u_full_def} and the definition of $F(h,M,R)$, \eas{ \| P^{\perp}_R u \|_{\infty} &\leq \| P^{\perp}_R W^{-1} U^* U P_M \|_{\infty} \| A^{-1} \|_{\infty} \| P_M W \|_{\infty} \leq \frac{\max_{i=1,\ldots,M} \{ w_i \} }{1-E(h,M)} F(h,M,R) \leq \frac{\epsilon}{1-\epsilon}. } Combining this with \R{M_R_split_term} and substituting into \R{R_splitting} yields $\theta \leq \frac{\epsilon}{1-\epsilon} < 1$. The result now follows immediately from Lemma \ref{l:Delta_recov}. } \subsection{Comparison with least-squares fitting}\label{ss:LS_fit_comp} The following result is standard (a proof is given for completeness): \thm{ \label{t:LS_err} For $0 < \epsilon < 1$ let $M \in \bbN$ and $h>0$ be such that $E_{2}(h,M) \leq \epsilon < 1$, where $E_2(h,M)$ is as in \R{E_2NM}. Then \R{LS_fit} has a unique solution $\check{x}$, and this satisfies \be{ \label{LS_err_bd} \| x - \check{x} \| \leq \left ( 1 + \frac{1}{\sqrt{1-\epsilon}} \right ) \| x - P_M x \|_{1,w} + \frac{1}{\sqrt{1-\epsilon}} \eta, } for any $w = \{ w_i \}_{i \in \bbN}$ with $w_i \geq \| \phi_i \|_{L^\infty}$. } \prf{ Observe that \eas{ \| U P_M z \|^2 &= z^* P_M U^* U P_M z = \| z \|^2 - z^* \left ( P_M - P_M U^* U P_M \right ) z \geq \left ( 1 - E_2(h,M) \right ) \| z \|^2. } Hence $U P_M$ has full rank since $E_{2}(h,M) \leq \epsilon < 1$, and its minimum singular value satisfies $\sigma_{\min} \geq \sqrt{1-\epsilon}$. This implies that $\check{x}$ is unique and is given by $\check{x} = (U P_M)^{\dag} y$. Hence \eas{ \check{x} - P_M x &= (U P_M)^{\dag} \left ( U x + \{ \sqrt{\tau_n} e_n \}^{N}_{n=1} \right ) - P_M x = (U P_M)^{\dag} \left ( U (x - P_M x) + \{ \sqrt{\tau_n} e_n \}^{N}_{n=1} \right ). } Therefore \bes{ \| \check{x} - P_M x \| \leq \frac{1}{\sqrt{1-\epsilon}} \left ( \| U(x-P_M x) \| + \eta \right ) \leq \frac{1}{\sqrt{1-\epsilon}} \left ( \| x - P_M x \|_{1,w} + \eta \right ) . } Since $\| x - \check{x} \| \leq \| x - P_M x \| + \| \check{x} - P_M x \| \leq \| x - P_M x \|_{1,w} + \| \check{x} - P_M x \|$, the result follows. } The error bound \R{LS_err_bd} is similar to the bound \R{l1w_err_bd} for weighted $\ell^1$ minimization. In the absence of noise and truncation error, both depend on the term $x-P_M x$, i.e.\ the tail of $x$ beyond its first $M$ coefficients. The primary difference is in the size of $M$, which is determined through the conditions of Theorems \ref{t:full_samp_recov} and \ref{t:LS_err} respectively. We shall discuss this point further in \S \ref{s:AlgPoly_Examp} and \S \ref{s:TrigPoly_Examp}. But we first reiterate (see also \S \ref{ss:LS_comp}) that $M$ is a fixed parameter for least squares, required for implementation, whereas for weighted $\ell^1$ it is introduced solely to provide an estimate for approximation error. Lemma \ref{l:dual_certificate} asserts that weighted $\ell^1$ minimization can recover coefficients of $x$ corresponding to other subsets $\Delta$, provided the various conditions hold. Note also that for weighted $\ell^1$ minimization the parameter $M$ appearing in Theorem \ref{t:full_samp_recov} is in no way related to the truncation parameter $K$, besides the relation $M \leq K$. \rem{ \label{r:LSl1weightsdiff} The error bound \R{LS_err_bd} also differs from \R{l1w_err_bd} in that it holds for the weights $w_i = \| \phi_i \|_{L^{\infty}}$ and its noise term does not involve the factor $\| P_M w \|$. See \S \ref{s:AlgPoly_Examp} for further discussion. We also remark that the term $\| x - P_M x \|_{1,w}$ in \R{l1w_err_bd} can be improved slightly to $\| f - \tilde{f} \|_{L^\infty}$, where $\tilde{f} = \sum^{M}_{i=1} x_i \phi_i$ is the projection of $f$ onto $\spn \{ \phi_1,\ldots,\phi_M \}$. We opt for \R{LS_err_bd} so that a direct comparison can be made with \R{l1w_err_bd}. Note that one also has $\| x - P_M x \| \leq \| f - \tilde{f} \|_{L^\infty} \leq \| x - P_M x \|_{1,w}$. } \rem{ \label{r:cooking_time} Theorems \ref{t:full_samp_recov} and \ref{t:LS_err} provide a recipe for determining the worst-case behaviour of weighted $\ell^1$ minimization for scattered data. This is as follows. Given a orthonormal system $\{ \phi_i \}_{i \in \bbN}$, an $h > 0$ and an $0 < \epsilon < 1/2$, determine: \bull{ \item[1.] the largest $M = M_1(h)$ such that $E_{2}(h,M) < \epsilon$, \item[2.] the largest $M = M_2(h)$ such that \R{NMR_condition} holds for some appropriate $R$ and $\{ w_i \}_{i \in \bbN}$. } In this case, the errors for both weighted $\ell^1$ minimization and least-squares fitting are determined by $\| x - P_{M} x \|_{1,w}$, where $M = M_1(h)$ for the former and $M = M_2(h)$ for the latter. Hence, if $M_1(h) \asymp M_2(h)$ as $h \rightarrow 0$ it follows that both least-squares fitting and weighted $\ell^1$ minimization are guaranteed to converge at roughly the same asymptotic rate as $h \rightarrow 0$. Since $M_1(h)$ and $M_2(h)$ are dependent on the system $\{ \phi_i \}_{i \in \bbN}$ a separate analysis must be carried out in each case. The last two sections of this paper will be devoted to doing this for the examples of \S \ref{ss:examples}. } \subsection{Proof of Lemma \ref{l:dual_certificate}} \label{ss:dual_certificate_proof} To complete this section we give the proof of Lemma \ref{l:dual_certificate}. \prf{[Proof of Lemma \ref{l:dual_certificate}] Let $v = \hat{x} - x$. Then $A v = P_{\Delta} U^* U v - P_{\Delta} U^* U P^{\perp}_{\Delta} v$, where $A$ is the restriction of $P_{\Delta} U^* U P_{\Delta} $ to $P_{\Delta}(\ell^2(\bbN))$. By $(i)$, we have $\| A^{-1} \| \leq \frac{1}{1-\alpha}$ and \bes{ \| P_{\Delta} U^* \|^2 = \| U P_{\Delta} \|^2 = \| P_{\Delta} U^* U P_{\Delta} \| \leq 1 + \alpha. } Thus \eas{ \| P_{\Delta} v \| &\leq \frac{1}{1-\alpha} \| P_{\Delta} U^* \| \| U v \| + \frac{1}{1-\alpha} \| P_{\Delta} U^* U P^{\perp}_{\Delta} v \| \leq \frac{\sqrt{1+\alpha}}{1-\alpha} \left ( \| U v \| + \| U P^{\perp}_{\Delta} v \| \right ). } Observe that \be{ \label{UN_v} \| U v \| = \| U \hat{x} - U x \| \leq 2 \eta . } Hence \bes{ \| P_{\Delta} v \| \leq \frac{\sqrt{1+\alpha}}{1-\alpha}\left ( 2 \eta + \| U P^{\perp}_{\Delta} v \| \right ). } The second term can be estimated as follows: \eas{ \| U P^{\perp}_{\Delta} v \| \leq \sum_{i \notin \Delta} | v_i | \| U e_i \| \leq \beta \| P^{\perp}_{\Delta} v \|_{1,w}, } where the latter inequality is due to $(ii)$. Hence we get \be{ \label{v_Delta_Delta_perp} \| P_{\Delta} v \| \leq \frac{\sqrt{1+\alpha}}{1-\alpha} \left ( 2 \eta +\beta \| P^{\perp}_{\Delta} v \|_{1,w} \right ). } We shall return to this inequality later, but let us now consider $\hat{x}$. \ea{ \nm{\hat{x}}_{1,w} &= \nm{P_\Delta\hat{x}}_{1,w} + \nmu{P^{\perp}_\Delta \hat{x} }_{1,w} \nn \\ & \geq \Re \ip{P_\Delta W \hat{x}}{\mathrm{sign}(P_\Delta x)} + \nmu{P^{\perp}_\Delta v }_{1,w} - \nmu{P^{\perp}_\Delta x}_{1,w} \nn \\ &= \Re \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x)} + \nm{P_\Delta x}_{1,w} + \nmu{P^{\perp}_\Delta v }_{1,w} - \nmu{P^{\perp}_\Delta x}_{1,w} \nn \\ & = \Re \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x)} + \nm{x}_{1,w} + \nmu{P^{\perp}_\Delta v }_{1,w} - 2 \nmu{P^{\perp}_\Delta x}_{1,w}. \label{sign_ineq} } Now let $\bar{x} \in \ell^1_w(\bbN)$ be any feasible solution for \R{fin_min}. Then $\| \hat{x} \|_{1,w} \leq \| \bar{x} \|_{1,w}$ and we get \bes{ \| \bar{x} \|_{1,w} \geq \Re \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x)} + \nm{x}_{1,w} + \nmu{P^{\perp}_\Delta v }_{1,w} - 2 \nmu{P^{\perp}_\Delta x}_{1,w}. } After rearranging this gives \be{ \label{2nd_bound} \| P^{\perp}_\Delta v \|_{1,w} \leq | \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x)} | + 2 \| P^{\perp}_\Delta x \|_{1,w} + \| x - \bar{x} \|_{1,w}. } We next estimate $| \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x)} |$. We have \bes{ | \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x)} | \leq | \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x) - P_{\Delta} u} | + | \ip{W v}{u} | + | \ip{P^{\perp}_{\Delta} W v}{P^{\perp}_{\Delta} u} |. } Note that $| \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x) - P_{\Delta} u} | \leq \gamma \| P_{\Delta} v \|$ by $(iii)$ and also that $\ip{W v}{u} = \ip{v}{W u} = \ip{v}{U^* u'} = \ip{ U v}{u'}$. Hence, \R{UN_v} and $(v)$ give \bes{ |\ip{W v}{u} | \leq \| U v \| L \sqrt{s} \leq 2 \eta L \sqrt{s}. } Finally, by $(iv)$, we have $| \ip{P^{\perp}_{\Delta} W v}{P^{\perp}_{\Delta} u} | \leq \| P^{\perp}_{\Delta} u \|_{\infty} \| P^{\perp}_{\Delta} v \|_{1,w} \leq \theta \| P^{\perp}_{\Delta} v \|_{1,w}$ and therefore \bes{ | \ip{P_\Delta W v}{\mathrm{sign}(P_\Delta x)} | \leq \gamma \| P_{\Delta} v \| +2 \eta L \sqrt{s} + \theta \| P^{\perp}_{\Delta} v \|_{1,w}. } Substituting into \R{2nd_bound} and rearranging now yields \bes{ (1-\theta) \| P^{\perp}_{\Delta} v \|_{1,w} \leq \gamma \| P_{\Delta} v \| + 2 \eta L \sqrt{s} + 2 \| P^{\perp}_{\Delta} x \|_{1,w}+ \| x - \bar{x} \|_{1,w}, } and applying \R{v_Delta_Delta_perp} gives \bes{ \| P_{\Delta} v \| \leq \frac{\sqrt{1+\alpha}}{1-\alpha} \left [ 2 \eta+ \frac{\beta}{1-\theta} \left ( \gamma \| P_{\Delta} v \| + 2 \eta L \sqrt{s} + 2 \| P^{\perp}_{\Delta} x \|_{1,w} + \| x - \bar{x} \|_{1,w}\right ) \right ]. } Hence \eas{ \| P_{\Delta} v \| &\leq \left ( 1 - \frac{\sqrt{1+\alpha} \beta \gamma}{(1-\alpha)(1-\theta)} \right )^{-1} \frac{2\sqrt{1+\alpha}}{1-\alpha} \left ( 1 + \frac{\beta}{1-\theta} L \sqrt{s} \right )\eta \\ & +\left ( 1 - \frac{\sqrt{1+\alpha} \beta \gamma q}{(1-\alpha)(1-\theta)} \right )^{-1} \frac{\sqrt{1+\alpha} \beta}{1-\alpha} \left ( 2 \| P^{\perp}_{\Delta} x \|_{1,w} + \| x - \bar{x} \|_{1,w} \right ) \\ & =2 C_0 \left ( 1 + \frac{\beta}{1-\theta} L \sqrt{s} \right )\eta + C_0 \beta \left ( 2 \| P^{\perp}_{\Delta} x \|_{1,w} + \| x - \bar{x} \|_{1,w} \right ). } Since $\| P^{\perp}_{\Delta} v \| \leq \| P^{\perp}_{\Delta} v \|_1 \leq \| P^{\perp}_{\Delta} v \|_{1,w}$ (recall \R{weights_growth}), we now get \eas{ \| v \| \leq& \| P_{\Delta} v \| + \| P^{\perp}_{\Delta} v \|_{1,w} \\ \leq& \left ( 1 + \frac{\gamma}{1-\theta} \right ) \| P_{\Delta} v \| + \frac{2\eta}{1-\theta} L \sqrt{s} + \frac{1}{1-\theta} \left ( 2 \| P^{\perp}_{\Delta} x \|_{1,w} + \| x - \bar{x} \|_{1,w} \right ) \\ \leq& \left [ \left ( 1 + \frac{\gamma}{1-\theta} \right ) C_0 \left (1 + \frac{\beta}{1-\theta} L \sqrt{s} \right ) + \frac{L\sqrt{s}}{1-\theta} \right ] (2 \eta + \delta ) \\ &+ \left [ C_0 \beta \left ( 1 + \frac{\gamma}{1-\theta} \right ) + \frac{1}{1-\theta} \right ] \left ( 2 \| P^{\perp}_{\Delta} x \|_{1,w} + \| x - \bar{x} \|_{1,w} \right ), } as required. } \section{Handling truncation: the choice of $K$}\label{s:truncation} We now consider the truncation parameter $K$. Due to Theorem \ref{t:full_samp_recov}, it suffices to estimate the quantity $T_{h,K,\eta}(x)$ defined in \R{trunc}. \thm{ \label{p:trunc_err} For all sufficiently large $K$ we have $\mathrm{Ran}(U) = \mathrm{Ran}(U P_K)$. In particular, the problems \R{fin_min_noiseless} and \R{fin_min} have solutions for all large $K$. Moreover, suppose that $r = \mathrm{rank}(U) \leq N$ and $K$ is sufficiently large so that $\mathrm{Ran}(U P_K) = \mathrm{Ran}(U)$. If $x \in \ell^1_w(\bbN)$ then \bes{ T_{h,K,\eta}(x) \leq \| x - P_K x \|_{1,w} + \| P_K w \| /\sigma_{\min} \| x - P_K x \|_{1,w}, } where $\sigma_{\min}$ is the minimum singular value of $U P_K$. Moreover, if $\{ w_i \}_{i \in \bbN}$ is nondecreasing, and $x \in \ell^1_{\tilde{w}}(\bbN)$, where $\tilde{w} = \{ \tilde{w}_i \}_{i \in \bbN}$ with $\tilde{w}_i \geq \sqrt{i} w^2_{i}$, $\forall i \in \bbN$, then \bes{ T_{h,K,\eta}(x) \leq \| x - P_K x \|_{1,w} + 1/\sigma_{\min} \| x - P_K x \|_{1,\tilde{w}}. } } \prf{ The first observation is immediate since $U$ has finite rank. Suppose now that $K$ is such that $\mathrm{Ran}(U P_K) = \mathrm{Ran}(U)$ and write $\bar{x} = P_K x + (U P_K)^{\dag} U (x-P_K x)$. Then \bes{ \| U P_K \bar{x} - y \| = \| U P_K x + U(x-P_K x) - U x - \{ \sqrt{\tau_n} e_n \}^{N}_{n=1} \| \leq \eta. } Hence $\bar{x}$ is feasible. Moreover, $\| x - \bar{x} \|_{1,w} \leq \| x - P_K x \|_{1,w} + \| (U P_K)^{\dag} U (x-P_K x) \|_{1,w}$ and \eas{ \| (U P_K)^{\dag} U (x-P_K x) \|^2_{1,w} & \leq \| P_K w \|^2 \| (U P_K)^{\dag} U (x-P_K x) \|^2 \\ & \leq \| P_K w \|^2 \| U(x-P_K x) \|^2 / \sigma^2_{\min} \\ & \leq \| P_K w \|^2 \| x - P_K x \|^2_{1,w} / \sigma^2_{\min}. } To obtain the final result, we note that \bes{ \| P_K w \| \| x - P_K x \|_{1,w} \leq w_{K} \sqrt{K} \sum_{i > K} w_i | x_i | \leq \sum_{i > K} \tilde{w}_i | x_i| = \| x - P_K x \|_{1,\tilde{w}}, } whenever the $w_i$'s are nondecreasing, as required. } Note that it need not be the case that \R{fin_min_noiseless} or \R{fin_min} have solutions for arbitrary $K \geq N$, since $\mathrm{Ran}(U P_K) \neq \mathrm{Ran}(U)$ in general. However, this theorem shows that this holds for all large $K$. Furthermore, this result shows that once $K$ is chosen so that $1/\sigma_{\min}$ is moderate in size, the effect of truncation is bounded by the decay of the coefficients $x_i$, $i > K$. \rem{ \label{r:trunc_weak} Theorem \ref{p:trunc_err} asserts that the truncation parameter $K$ can be chosen independently of the coefficients whenever $x \in \ell^1_{\tilde{w}}(\bbN)$. While it is possible to show that $T_{h,K,\eta}(x) \rightarrow 0$ as $K \rightarrow \infty$ for $x \in \ell^1_{w}(\bbN)$ \cite[Prop.\ 6.6]{BAACHGSCS}, it is currently unknown whether a bound for $T_{h,K,\eta}(x)$ involving only $\| x - P_{K} x \|_{1,w}$ holds. In other words, if $x \in \ell^1_{w}(\bbN)$ but $x \notin \ell^1_{\tilde{w}}(\bbN)$ the truncation strategy may no longer be independent of $x$. Improving this result is an open problem. } It is important to quantify precisely how large $K$ needs to be in relation to $N$ to ensure a small truncation error. For this, we shall use the following lemma: \lem{ \label{l:gamma} The minimum singular value $\sigma_{\min}$ of $U P_K$ satisfies \be{ \label{gamma_def} \sigma_{\min} \geq \inf_{\substack{y \in \bbC^N \\ \| y \| =1}} \sup_{g \in G_y} \sup_{\substack{\phi \in \Phi_K \\ \phi \neq 0}} \left \{ \frac{1-\| g - \phi \|_{\infty}}{\| g \|_{\nu} + \| g - \phi \|_{\infty}} \right \}, } where $\Phi_K = \spn \{ \phi_1,\ldots,\phi_K \}$ and $G_y = \left \{ g \in L^2_{\nu}(D) \cap L^\infty(D) : g(t_n) = y_{n} / \sqrt{\tau_n},\ n=1,\ldots,N \right \}$. } \prf{ The minimum singular value is given by $\sigma_{\min} = \inf_{\substack{y \in \bbC^N \\ \| y \| =1}} \| (U P_K)^* y \|$. Fix $y \in \bbC^N$, $\| y \| =1$ and observe that \bes{ \| (U P_K)^* y \| = \sup_{\substack{z \in \bbC^K \\ \| z \|=1}} \left | \ip{y}{U P_K z} \right |= \sup_{\substack{\phi \in \Phi_K \\ \| \phi \|_{\nu} = 1}} \left | \sum^{N}_{n=1} \sqrt{\tau_n} y_n \overline{\phi(t_n)} \right |. } Let $g \in G_y$ and $\phi \in \Phi_K$. Then \eas{ \| (U P_K)^* y \| &\geq \left | \sum^{N}_{n=1} \sqrt{\tau_n} y_n \overline{\phi(t_n) } \right | \geq \| y \|^2 - \left | \sum^{N}_{n=1} \sqrt{\tau_n} y_n \overline{g(t_n) - \phi (t_n)} \right | \geq \| y \|^2 - \| y \| \| g - \phi \|_{L^\infty}. } Since $\| \phi \|_{\nu} \leq \| g \|_{\nu} +\| g - \phi \|_{\nu} \leq \| g \|_{\nu} + \| g - \phi \|_{\infty}$ the result now follows immediately. } \rem{ \label{r:just_compute} In practice, rather than performing an analysis of $\sigma_{\min}$ via Lemma \ref{l:gamma}, one may choose $K$ simply by calculating the minimal singular value of $UP_K$. If this is sufficiently large, then Theorem \ref{p:trunc_err} guarantees that the truncation error is moderate. } \subsection{Examples}\label{ss:trunc_examples} Lemma \ref{l:gamma} allows one to determine the required condition on $K$. Note that this depends completely on the points $T$ and the basis $\{ \phi_i \}_{i \in \bbN}$. We now do this for the two examples of \S \ref{ss:examples}: \thm{ \label{t:Leg_trunc} For $\alpha,\beta > -1$ let $\{ \phi_i \}_{i \in \bbN}$ be the Jacobi polynomial basis \R{Jacobi_ON} and let $T = \{ t_n \}^{N}_{n=1}$ be an ordered set of points in $[-1,1]$. Then for every $r \in \bbN$ there exists a $C_r >0$ such that \bes{ \sigma_{\min} \geq 1 - \frac{C_r K^{-r} \xi^{-r-1/2}}{\sqrt{\xi / (2 h_N)}-C_r K^{-r} \xi^{-r-1/2}}, } where $\sigma_{\min}$ is the minimal singular value of $U P_K$, \bes{ \xi = \frac12\min_{n=0,\ldots,N} \{ t_{n+1}-t_n \},\qquad h = \sup_{-1 \leq t \leq 1} \min_{n=1,\ldots,N} | t - t_n |, } and $t_0 = -t_1-2$, $t_{N+1} = 2-t_N$. In particular, if \bes{ K \geq \left ( \sqrt{2}C_r(1+\epsilon^{-1})\right )^{\frac{1}{r}} h^{\frac{1}{2r}} \xi^{-1-\frac1r}, } for some $0 < \epsilon < 1$ then $\sigma_{\min} > 1- \epsilon$. } We defer the proof of this result until \S \ref{ss:Leg_proofs}. Note that in the case of equispaced data, we have $h = \xi = 1/N$ and so it suffices to take $K \gtrsim N^{1+\frac{1}{2r}}$ for any $r \in \bbN$. In practice, we have found that $K=4N$ is sufficient in all examples (recall also Remark \ref{r:just_compute}). On the other hand, if the data clusters severely, then this theorem suggests that a larger value of $K$ may be necessary. \thm{ \label{t:Trig_trunc} Let $\{ \phi_i\}_{i \in \bbZ}$ be the Fourier basis \R{fourier_basis} and let $T = \{ t_n \}^{N}_{n=1}$ be a set of $N$ ordered points in $[-1,1]$. Then for every $r \in \bbN$ there exists a $C_r >0$ such that \bes{ \sigma_{\min} \geq 1 - \frac{C_r K^{-r} \xi^{-r-1/2}}{\sqrt{\xi / (2 h_N)}-C_r K^{-r} \xi^{-r-1/2}}, } where $\sigma_{\min}$ is the minimal singular value of $U P_K$, \bes{ \xi = \frac12\min_{n=0,\ldots,N} \{ t_{n+1}-t_n \},\qquad h = \sup_{-1 \leq t \leq 1} \min_{n=1,\ldots,N} | t - t_n |, } and $t_0 = -1$, $t_{N+1} = 1$. In particular, if \bes{ K \geq \left ( \sqrt{2}C_r(1+\epsilon^{-1})\right )^{\frac{1}{r}} h^{\frac{1}{2r}} \xi^{-1-\frac1r}, } for some $0 < \epsilon < 1$ then $0 \leq \gamma < \epsilon$. } This result is exactly the same as that for Jacobi polynomials (Theorem \ref{t:Leg_trunc}), except up to a minor change in the definition of the values $t_0$ and $t_{N+1}$, and therefore $\xi$. Its proof is near identical, and hence is omitted. \section{Jacobi polynomials on the unit interval}\label{s:AlgPoly_Examp} We now consider Example \ref{ex:Jacobi}. For convenience, we recall the growth condition \R{Jacobi_ON_growth}: \be{ \label{Jacobi_ON_growth2} \| \phi_j \|_{L^\infty} = \ord{j^{q+1/2}},\quad j \rightarrow \infty,\qquad \mbox{where $q = \max \{ \alpha,\beta , -1/2 \}$}. } Our main result is the following: \thm{ \label{t:Leg_poly_full_l1} For $\alpha,\beta > -1$ let $\{ \phi_i \}_{i \in \bbN}$ be the orthonormal Jacobi polynomial basis \R{Jacobi_ON}, $T = \{ t_n \}^{N}_{n=1} \subseteq [-1,1]$ be a set of scattered points and suppose that $h$ is as in \R{h_def}. Suppose that the weights $w_i = i^{\gamma} \| \phi_i \|_{L^{\infty}}$ for some $\gamma > 0$. Then for $0 < \epsilon < 1/2$ there exists a $c(\epsilon) > 0$ such that if \be{ \label{hcondPoly} h \leq c(\epsilon) \left \{ \begin{array}{cl} \frac{1}{M^{2+2(q+1)/\gamma} \log M} & 0 < \gamma \leq 1/2-q \\ \frac{1}{M^2 \log M} & \gamma> 1/2-q \end{array} \right ., } where $q$ is given by \R{Jacobi_ON_growth2}, any minimizer $\hat{x}$ of \R{fin_min} satisfies \bes{ \| x - \hat{x} \| \leq C(\epsilon) \left ( M^{\gamma+q+1} \eta + \| x -P_M x \|_{1,w} + T_{N,K, \eta}(x) \right ), } for some constant $C$ depending on $\epsilon$ only, where $T_{N,K, \eta }(x)$ is as in \R{trunc}. } One also has a similar, albeit simpler, result for least squares: \thm{ \label{t:Jacobi_poly_LS} Let $\{ \phi_i \}_{i \in \bbN}$ be the orthonormal Jacobi polynomial basis \R{Jacobi_ON}, $T = \{ t_n \}^{N}_{n=1} \subseteq [-1,1]$ be a set of scattered points and suppose that $h$ is as in \R{h_def}. Then for each $0 < \epsilon < 1$ there exists a $c(\epsilon) > 0$ such that if \be{ \label{hcondPolyLS} h \leq c(\epsilon) M^{-2}, } then the solution $\check{x}$ of \R{LS_fit} exists uniquely and satisfies \bes{ \| x - \check{x} \| \leq \left ( 1 + \frac{1}{\sqrt{1-\epsilon}} \right ) \| x - P_M x \|_{1,w} + \frac{1}{\sqrt{1-\epsilon}} \eta, } for any $w = \{ w_i \}_{i \in \bbN}$ with $w_i \geq \| \phi_i \|_{L^\infty}$. } Theorems \ref{t:Leg_poly_full_l1} and \ref{t:Jacobi_poly_LS} assert that both techniques guarantee an approximation error that depends on $x - P_M x$ measured in some norm, for the same asymptotic scaling of $h$ with $M$ up to log factors. Hence, up to the choice of norm, weighted $\ell^1$ minimization with scattered data, Jacobi polynomials and sufficiently large weights $w_i$ is guaranteed a similar approximation rate as least-squares fitting. It is informative to now consider the following two cases: \pbk \textit{Smooth functions.} Let $f \in C^{\infty}([-1,1])$. Then the coefficients $x_{i} = \ord{i^{-k}}$ as $i \rightarrow \infty$ for any $k >0$. Hence $\| x - P_M x \|_{1,w} = \ord{M^{-k}}$ as $M \rightarrow \infty$ for any $k >0$ whenever the weights $w_i$ grow at most algebraically fast in $i$. By Theorems \ref{t:Leg_poly_full_l1} and \ref{t:Jacobi_poly_LS} the approximation errors for weighted $\ell^1$ minimization and least squares both decay superalgebraically fast in $h$ as $h \rightarrow 0$; that is, faster than $h^k$ for any $k > 0$. \pbk \textit{Analytic functions.} Suppose $f$ is analytic so that $x_i = \ord{\rho^{-i}}$ for some $\rho > 1$. For algebraic weights $w_i$ one has $\| x - P_M x \|_{1,w} = \ord{(\rho')^{-M}}$ as $M \rightarrow \infty$ for any $\rho' < \rho$. Therefore the approximation error for least squares behaves like $\| x - \check{x} \| = \ordu{(\rho')^{-1/\sqrt{h}}}$ as $h \rightarrow 0$, and for weighted $\ell^1$ minimization one has the marginally slower convergence $\| x - \hat{x} \| = \ordu{(\rho')^{-1/\sqrt{h \log(h)}}}$, provided the weights $w_{i} = i^{\gamma} \| \phi_i \|_{L^\infty}$ with $\gamma > 1/2 - q$. \rem{ On the other hand, for functions of finite smoothness, the need for more rapidly-growing weights in Theorem \ref{t:Leg_poly_full_l1} translates into slower algebraic convergence of the approximation than that of least squares. We expect that this may be an artifact of the proof, however. } Suppose now that the data $T$ is equispaced. We first note the following general result: \thm{ \label{t:PTK} Let $T = \{ t_n \}^{N}_{n=1}$ be an equispaced grid of $N$ points in $[-1,1]$, $E \subseteq \bbC$ be a compact set containing $[-1,1]$ in its interior and let $B(E)$ be the Banach space of functions continuous on $B$ and analytic in its interior, with norm $\| f \|_{B} = \sup_{z \in B} | f(z) |$. Let $F : B(E) \rightarrow L^\infty(-1,1)$ be a mapping such that for each $f \in B(E)$, $F(f)$ depends only on the data $\{ f(t_n) \}^{N}_{n=1}$, and suppose that, for constants $C>0$, $\sigma > 1$ and $1/2 < \tau \leq 1$, \bes{ \| f - F(f) \|_{L^\infty} \leq C \sigma^{-N^{\tau}} \| f \|_{E},\quad \forall f \in B(E). } If $\| f \|_{T,\infty} = \max_{n=1,\ldots,N} | f(t_n) |$ then there exists a constant $\nu > 1$ such that \be{ \label{condition_replacement} \Theta(F) = \sup_{\substack{f \in B(E) \\ \| f \|_{T,\infty} \neq 0} } \left \{ \frac{\| F(f) \|_{L^\infty}}{\| f \|_{T,\infty}} \right \} \geq \nu^{N^{2 \tau - 1}}. } } This theorem is due to Platte, Trefethen \& Kuijlaars \cite{TrefPlatteIllCond} (a minor modification is made in \R{condition_replacement} which is more suitable for our purposes). It states the following: for any method that achieves an error for all functions in $B(E)$ that is exponentially-decaying as $N \rightarrow \infty$ with rate $\tau$ it is possible to find a function $f \in B(E)$ which is bounded on the set $T$, but for which $\|F(f)\|_{L^\infty}$ is exponentially large with index $2 \tau - 1$. In particular, the best possible convergence rate for a robust method, i.e.\ one for which $\Theta(F) \leq C$ for all $N \in \bbN$, is root-exponential in $N$. Note that this theorem is very general: the method $F$ can be linear or nonlinear, and $F(f)$ only needs to be defined for extremely smooth (specifically, analytic) functions. Now consider the cases of weighted $\ell^1$ minimization and least-squares fitting with Jacobi polynomials. By earlier arguments, the corresponding approximation errors behave like $\ordu{(\rho')^{-\sqrt{N}}}$ and $\ordu{(\rho')^{-\sqrt{N/\log(N)}}}$ respectively as $N \rightarrow \infty$. Moreover, by setting $x = 0$ in Theorems \ref{t:Leg_poly_full_l1} and \ref{t:Jacobi_poly_LS} respectively, one deduces that $\Theta(F) \lesssim 1$ for the former and $\Theta(F) \lesssim (N/\log(N))^{(\gamma+q+1)/2}$ for the latter. According to Theorem \ref{t:PTK}, least-squares fitting attains the best possible convergence rate for a robust method, while weighted $\ell^1$ minimization (with sufficiently large weights) attains nearly the best possible convergence rate, with only slow, algebraic growth of $\Theta(F)$. \rem{ Although Theorem \ref{t:PTK} applies only to equispaced data, it can also be formulated for much more general data. Loosely speaking, similar conclusions apply unless the data clusters quadratically at the endpoints $x = \pm 1$ \cite{AdcockNecSamp}. } \rem{ \label{noise_growth} For general data, the constant $M^{\gamma+q+1}$ in Theorem \ref{t:Leg_poly_full_l1} implies mild ill-conditioning of weighted $\ell^1$ minimization as $h \rightarrow 0^{+}$. This is not seen in computations, and we expect it is also an artifact of the proof. Removing this factor is an open problem. } \subsection{Numerical examples}\label{ss:Polynumexamp} In Figs.\ \ref{f:LegEqui} and \ref{f:LegJitt} we give a results for polynomial approximations from equispaced and jittered data. Although it has only been proved that weighted $\ell^1$ minimization performs as well (up to a log factor) as least-squares fitting, these results show that it in fact exhibits rather better performance, similar to that of the best possible least-squares fit. Note that this oracle least squares cannot be implemented in practice since $M$ is calculated by minimizing the approximation error. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=4.5cm]{fig4_1} & \includegraphics[width=4.5cm]{fig4_3} & \includegraphics[width=4.5cm]{fig4_4} \\ f(x) = \frac{1}{50/49-\sin(\pi x)} & f(x) = \sin(50 x^2) & f(x) = \frac{1}{1+50 x^2} \\ \\ \includegraphics[width=4.5cm]{fig4_5} & \includegraphics[width=4.5cm]{fig4_6} & \includegraphics[width=4.5cm]{fig4_8} \\ f(x) = \frac{\cosh(100 x^2)}{\cosh(100)} & f(x) = |x|^3 & f(x) = \sin(80 x) \end{array}$ \caption{ Numerical comparison of weighted $\ell^1$ minimization and least-squares fitting for approximation from equispaced data using Legendre polynomials. The error against $N$ is plotted for each method. The solid black line is weighted $\ell^1$ minimization with $K=4N$ and weights $w_{i} = i$. The dashed lines are least squares with $M = c \sqrt{N}$ and $c=0.5,1.0,1.5,2,2.5,3.0$. The solid blue line is oracle least squares based on choosing $M$ between $1$ and $N$ which minimizes the error $\| f - \tilde{f} \|_{L^\infty}$ for a given $N$ and $f$. Random noise of magnitude $10^{-8}$ was added to the data. } \label{f:LegEqui} \end{center} \end{figure} \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=4.5cm]{fig5_1} & \includegraphics[width=4.5cm]{fig5_3} & \includegraphics[width=4.5cm]{fig5_4} \\ \mbox{\scriptsize $f(x) = \frac{1}{50/49-\sin(\pi x)}$} & \mbox{\scriptsize $f(x) = \sin(50 x^2) $} & \mbox{\scriptsize $f(x) = \frac{1}{1+50 x^2}$} \\ \\ \includegraphics[width=4.5cm]{fig5_5} & \includegraphics[width=4.5cm]{fig5_6} & \includegraphics[width=4.5cm]{fig5_8} \\ \mbox{\scriptsize $f(x) = \frac{\cosh(100 x^2)}{\cosh(100)}$} & \mbox{\scriptsize $f(x) = |x|^3$} & \mbox{\scriptsize $f(x) = \sin(80 x)$} \end{array}$ \caption{ The same as Fig.\ \ref{f:LegEqui} except for randomly jittered data. } \label{f:LegJitt} \end{center} \end{figure} \subsection{Proofs}\label{ss:Leg_proofs} The proof of Theorem \ref{t:Leg_poly_full_l1} relies on the following three lemmas, which provide estimates for the quantities $E_2(h,M)$, $E_{\infty}(h,M)$ and $F(h,M,R)$ respectively. \lem{ \label{l:E2_Leg} For $\alpha,\beta > -1$ let $\{ \phi_i \}_{i \in \bbN}$ be the orthonormal Jacobi polynomial basis \R{Jacobi_ON}, $T = \{ t_n \}^{N}_{n=1} \subseteq D$ be a set of scattered data points and suppose that $h$ is as in \R{h_def}. If $h M^2 \leq 1$ then $E_{2}(h,M) \lesssim \sqrt{h} M$, where $E_{2}(h,M)$ is as in \R{E_2NM}. } \prf{ By self-adjointness, \be{ \label{norm_identity} \| P_M - P_M U^* U P_M \| = \sup_{\substack{x \in P_M(\ell^2(\bbN)) \\ \| x \|=1}} | \ip{(P_M - P_M U^* U P_M) x}{x} |. } Let $x \in P_M(\ell^2(\bbN))$, $\| x \| =1$ be arbitrary and set $g = \sum^{M}_{j=1} x_j \phi_j \in \bbP_{M-1}$ so that $\| g \|_{L^2_{\nu^{(\alpha,\beta)}}} = 1$. Let $\{ V_n \}^{N}_{n=1}$ be the Voronoi cells of the points $\{ t_n \}^{N}_{n=1}$ and set $\chi(t) =\sum^{N}_{n=1} g(t_n) \bbI_{V_n}(t)$. Then, by the definition \R{quad_weights_1} of the weights $\tau_n$, we have $\| \chi \|^2_{L^2_{\nu^{(\alpha,\beta)}}} = \sum^{N}_{n=1} \tau_n | g(t_n) |^2$, and therefore \bes{ | \ip{(P_M - P_M U^* U P_M) x}{x} | = \left | 1 - \sum^{N}_{n=1}\tau_n | g(t_n) |^2 \right | = \left | \| g \|^2_{L^2_{\nu^{(\alpha,\beta)}}} - \| \chi \|^2_{L^2_{\nu^{(\alpha,\beta)}}} \right |. } Hence \be{ \label{diff_ineq} | \ip{(P_M - P_M U^* U P_M) x}{x} | \leq \| g - \chi \|_{L^2_{\nu^{(\alpha,\beta)}}} \left ( 2 \| g \|_{L^2_{\nu^{(\alpha,\beta)}}} + \| g - \chi \|_{L^2_{\nu^{(\alpha,\beta)}}} \right ), } and so it suffices to show that \be{ \label{desired_result} \| g - \chi \|_{L^2_{\nu^{(\alpha,\beta)}}} \lesssim \sqrt{h} M \| g \|_{L^2_{\nu^{(\alpha,\beta)}}}. } We have \bes{ \| g - \chi \|^2_{L^2_{\nu^{(\alpha,\beta)}}} = \sum^{N}_{n=1} \int_{V_n} \left | g(t) - g(t_n) \right |^2 \nu^{(\alpha,\beta)}(t) \D t = \sum^{N}_{n=1} I_n. } Let $n_0$ be the unique number such that $0 \in V_{n_0}$. Then we write this as \be{ \label{int_sum_split} \| g - \chi \|^2_{L^2_{\nu^{(\alpha,\beta)}}} = \sum^{N}_{n=n_0+1} I_n + \sum^{n_0-1}_{n=1} I_n + I_{n_0} = S_{1} + S_{-1} + S_{0}. } We shall address each term separately. Consider $S_1$. Since $\nu^{(\alpha,\beta)}(t) \lesssim (1-t)^{\alpha}$ on $[0,1]$, we have \bes{ I_n \lesssim \int_{V_n} | g(t) - g(t_n) |^2 (1-t)^{\alpha} \D t,\quad n=n_0,\ldots,N. } We now consider three cases: (i) $-1 < \alpha < 0$, (ii) $\alpha = 0$ and (iii) $\alpha > 0$. Consider case (i). Then \eas{ I_n \lesssim \int_{V_n} \left| \int_{V_n} |g'(s)| \D s \right |^2 (1-t)^{\alpha} \D t \lesssim \int_{V_n} (1-t)^{\alpha} \D t \int_{V_n} (1-t)^{-\alpha-1} \D t \int_{V_n} |g'(t)|^2 (1-t)^{\alpha+1} \D t } By construction, $V_n$ is of width at most $2h$. Hence, after a short calculation we get that \bes{ I_n \lesssim h \int_{V_n} |g'(t)|^2 (1-t)^{\alpha+1} \D t,\qquad -1 < \alpha < 0. } Now consider case (ii). We have \bes{ I_n \lesssim \int_{V_n} \D t \left| \int_{V_n} |g'(s)| \D s \right |^2 \leq \left ( \int_{V_n} \D t \right )^2 \int_{V_n} |g'(t)|^2 \D t \lesssim h^2 \int_{V_n} | g'(t) |^2 \D t,\qquad \alpha = 0. } Final, consider case (iii). By similar arguments, we get that \bes{ I_n \lesssim \left ( \int_{V_n} (1-t)^{\alpha} \int^{t}_{t_n} (1-s)^{-\alpha-1} \D s \D t \right ) \int_{V_n} |g'(t)|^2 (1-t)^{\alpha+1} \D t. } Write $V_n = (a,b)$ where $0 \leq a \leq t_n$ and $t_n \leq b \leq 1$. Then \eas{ \int_{V_n} (1-t)^{\alpha} \int^{t}_{t_n} (1-s)^{-\alpha-1} \D s \D t &= \frac{1}{\alpha} \int^{b}_{a} (1-t)^{\alpha} \left ( (1-t)^{-\alpha} - (1-t_n)^{-\alpha} \right ) \D t \\ & = \frac{1}{\alpha} \left [ (b-a) + \frac{1}{\alpha+1} \frac{(1-b)^{\alpha+1} - (1-a)^{\alpha+1}}{(1-t_n)^{\alpha}} \right ] } Note that $1-b \leq 1-t_n$ and $1-a \geq 1-t_n$. Therefore \bes{ \int_{V_n} (1-t)^{\alpha} \int^{t}_{t_n} (1-s)^{-\alpha-1} \D s \D t \leq \frac{1}{\alpha} \left [ b-a - \frac{b-a}{\alpha+1} \right ] \lesssim h. } Hence \bes{ I_n \lesssim h \int_{V_n} |g'(t)|^2 (1-t)^{\alpha+1} \D t,\qquad \alpha > 0. } With these estimates to hand, we now deduce the following bound for the term $S_1$ in \R{int_sum_split}: \be{ \label{S1_bd} S_1 \lesssim \left \{ \begin{array}{lc} h^2 \| g' \|^2_{L^2(0,1)} & \alpha = 0 \\ h \| g' \|^2_{L^2_{\nu^{(\alpha+1,\beta+1)}}} & \alpha \neq 0 \end{array} \right ., } This follows from the fact that the $V_n$ form a partition of $(-1,1)$, the definition of $n_0$ and the fact that $(1+t)^{\beta+1} \geq 1$ for $t \in [0,1]$. Identical arguments give a similar result for $S_{-1}$: \be{ \label{Sm1_bd} S_{-1} \lesssim \left \{ \begin{array}{lc} h^2 \| g' \|^2_{L^2(-1,0)} & \beta = 0 \\ h \| g' \|^2_{L^2_{\nu^{(\alpha+1,\beta+1)}}} & \beta \neq 0 \end{array} \right . . } Finally, consider $S_0$. Let $V_{n_0} = J_{-1} \cup J_{1}$, where $J_{1} \subseteq [0,1]$ and $J_{-1} \subseteq [-1,0]$. If $h M^2 \lesssim 1$ we may assume that $h \leq 1/2$ so that $|1-t| \geq 1/2$ and $|1+t| \geq 1/2$ for $t \in V_{n_0}$. Then we have \bes{ J_{\pm 1} \lesssim \left ( \int_{J_{\pm 1}} \D t \right )^2 \int_{J_{\pm 1}} |g'(t)|^2 \D t \lesssim h^2 \int_{J_{\pm 1}} |g'(t)|^2 \D t \lesssim h^2 \int_{J_{\pm 1}} |g'(t)|^2 \nu^{(\gamma,\delta)}(t)\D t, } for any $-1 < \gamma,\delta \leq 0$. It now follows that $J_{\pm1}$ satisfy exactly the same bounds as \R{S1_bd} and \R{Sm1_bd} for $S_{\pm 1}$. Therefore, in order to estimate the left-hand side of \R{int_sum_split} it suffices from now on to consider only $S_{\pm 1}$. For this, we shall use Markov's inequality: \be{ \label{Markov} \| g' \|_{L^2(I)} \lesssim M^2/|I| \| g \|_{L^2(I)},\quad \forall g \in \bbP_{M}, } where $I$ is an arbitrary bounded interval, as well as the following Markov-type inequality: \be{ \label{General_Markov} \| g' \|_{L^2_{\nu^{(\alpha+1,\beta+1)}}} \lesssim M \| g \|_{L^2_{\nu^{(\alpha,\beta)}}},\quad \forall g \in \bbP_{M}. } Markov's inequality \R{Markov} is well-known. A proof of \R{General_Markov} is given in Appendix \ref{a:Jacobi}. There are now four cases: (a) $\alpha = \beta = 0$, (b) $\alpha = 0$, $\beta \neq 0 $, (c) $\alpha \neq 0 $, $\beta = 0$ and (d) $\alpha \neq 0$, $\beta \neq 0$. Consider case (a). Then by \R{int_sum_split}, \R{S1_bd}, \R{Sm1_bd} and \R{Markov} we find that $\| g - \chi \|^2_{L^2} \lesssim h^2 M^4 \| g \|^2_{L^2}$. Since $h M^2 \lesssim 1$, this now gives \R{desired_result} for case (a). Now consider case (b). By \R{int_sum_split}, \R{S1_bd}, \R{Sm1_bd}, \R{Markov} and \R{General_Markov} we get that \bes{ \| g - \chi \|^2_{L^2_{\nu^{(0,\beta)}}} \lesssim h^2 M^4 \| g \|^2_{L^2(0,1)} + h M^2 \| g \|^2_{L^2_{\nu^{(0,\beta)}}} \lesssim h M^2 \| g \|^2_{L^2_{\nu^{(0,\beta)}}}. } Here in the final inequality we use the facts that $h M^2 \lesssim 1$ and $(1+t)^{\beta} \geq 1$ for $t \in [0,1]$. Thus we get \R{desired_result} in this case as well. Case (c) is near-identical to case (b). To complete the proof, we consider case (d). Using \R{int_sum_split}, \R{S1_bd}, \R{Sm1_bd} and \R{General_Markov}, we get \bes{ \| g - \chi \|^2_{L^2_{\nu^{(\alpha,\beta)}}} \lesssim h M^2 \| g \|^2_{L^2_{\nu^{(0,\beta)}}}, } which yields \R{desired_result}. } \lem{ \label{l:Einf_Leg} For $\alpha,\beta > -1$ let $\{ \phi_i \}_{i \in \bbN}$ be the orthonormal Jacobi polynomial basis \R{Jacobi_ON}, $T = \{ t_n \}^{N}_{n=1} \subseteq D$ be a set of scattered data points and suppose that $h$ is as in \R{h_def}. If $h M^2 \leq 1$ then \bes{ E_{\infty}(h,M) \lesssim h M^2 \log M, } where $E_{\infty}(h,M)$ is as in \R{E_2NM}. } \prf{ Let $x \in P_M(\ell^2(\bbN))$, $\| x \|_{\infty} = 1$ and set $g = \sum^{M}_{j=1} x_j \phi_j \in \bbP_{M-1}$ as in the previous proof. Then \be{ \label{unif_norm_identity} \| (P_M - P_M U^* U P_M )x \|_{\infty} = \max_{i=1,\ldots,M} \left | \ip{g}{\phi_i}_{L^2_{\nu^{(\alpha,\beta)}}} - \ip{g}{\phi_i}_h \right |. } Observe that \eas{ \left | \ip{g}{\phi_i}_{L^2_{\nu^{(\alpha,\beta)}}} - \ip{g}{\phi_i}_h \right | &= \left | \sum^{N}_{n=1} \int_{V_n} \left ( g(t) \phi_i(t) - g(t_n) \phi_i(t_n) \right ) \nu^{(\alpha,\beta)}(t) \D t \right | \\ & \leq \sum^{N}_{n=1} \int_{V_n} \left | \int_{t_n}^t \left ( g(s) \phi_i(s) \right )' \D s \right | \nu^{(\alpha,\beta)}(t) \D t \\ & \lesssim \sum^{N}_{n=1} \int_{V_n} \nu^{(\alpha,\beta)}(t) \D t \int_{V_n} | (g(s) \phi_i(s))' | \D s . } Since $\| x \|_{\infty} = 1$, we have $| (g(s) \phi_i(s))' | \leq \sum^{M}_{j=1} | ( \phi_i(s) \phi_j(s))' |$ and we now substitute into \R{unif_norm_identity} to deduce that \ea{ \label{unif_norm_ineq} \| (P_M - P_M U^* U P_M )x \|_{\infty} &\lesssim \max_{i=1,\ldots,M} \sum^{M}_{j=1} \sum^{N}_{n=1} \int_{V_n} \nu^{(\alpha,\beta)}(t) \D t \int_{V_n} | (\phi_i(s) \phi_j(s))' | \D s \nn \\ & = \max_{i=1,\ldots,M} \sum^{M}_{j=1} \sum^{N}_{n=1} I_n = \max_{i=1,\ldots,M} \sum^{M}_{j=1} \left ( S_{1}+S_{-1}+S_0 \right ), } where, as in the previous lemma, $S_1$ corresponds to $[0,1]$, $S_{-1}$ corresponds to $[-1,0]$ and $S_0$ corresponds to the term $I_{n_0}$ where $ 0 \in V_{n_0}$. Consider the term $S_1$. By \R{h_asymp} and \R{Jacobi_Local}, we have \bes{ | \phi^{(k)}_i(t) | \lesssim \min \left \{ (\sqrt{1-t^2})^{-\alpha-k-1/2} i^{k} , i^{2k+\alpha+1/2} \right \},\quad 0 \leq t \leq 1. } Since $1 \leq i,j \leq M$, we have that \bes{ S_1 \lesssim \sum^{N}_{n=n_0+1} \int_{V_n} (1-t)^{\alpha} \D t \int_{V_n}\min \left \{ (\sqrt{1-t^2})^{-2\alpha-2} M , M^{2\alpha+3} \right \} \D t. } Let $n_0+1 \leq n^* < N$ be arbitrary (its value will be chosen later) and split this sum into two according to $n^*$. Then \ea{ S_1 &\lesssim M \sum^{n^*}_{n=n_0+1} \int_{V_n} (1-t)^{\alpha} \D t \int_{V_n} (\sqrt{1-t^2})^{-2\alpha-2} \D t + M^{2\alpha+3} \sum^{N}_{n=n^*+1} \int_{V_n} (1-t)^{\alpha} \D t \int_{V_n}1 \D t \nn \\ & = M S^{-}_{1} + M^{2\alpha+3} S^{+}_{1}. \label{S1} } We consider $S^{\pm}_1$ separately. For $S^{+}_{1}$, we have \be{ \label{Splus1} S^{+}_{1} \lesssim h \int^{1}_{z^*} (1-t)^{\alpha} \D t \lesssim h (1-z^*)^{\alpha+1}, } where $z^*$ is the right endpoint of $V_{n^*}$. Now consider $S^{-}_{1}$: \bes{ S^{-}_{1} \lesssim h \sum^{n^*}_{n=n_0+1} \int_{V_n} (1-t)^{\alpha} \D t \sup_{t \in V_n}(1-t)^{-\alpha-1} = h \sum^{n^*}_{n=n_0+1} \int_{V_n} (1-t)^{\alpha} (1-z_n)^{-\alpha-1} \D t. } where $z_n$ is the right endpoint of $V_n$. However, if $t \in V_n$ is arbitrary then $z_n \leq t + h$. Thus $(1-z_n)^{-\alpha-1} \leq (1-t-h)^{-\alpha-1}$, and this gives \bes{ S^{-}_{1} \lesssim h \int^{z^*}_{0} (1-t)^{\alpha}(1-t-h)^{-\alpha-1} \D t \leq h \int^{1}_{1-z^*} s^{-1} (1-h/s)^{-\alpha-1} \D s } Suppose now that $n^*$ is chosen so that \be{ \label{nstar_choice} 1 - 2/M^2 - 2 h \leq z^* \leq 1 - 2/M^2, } (recall that the Voronoi cells are of width at most $2h$, hence such a choice is possible). Then, since $h M^2 \leq 1$, we find that $1-h/s \gtrsim 1$ for $1-z^* \leq s \leq 1$. Therefore, if \R{nstar_choice} holds we have \be{ \label{Sminus1} S^{-}_{1} \lesssim h \int^{1}_{1-z^*} s^{-1} \D s \lesssim h \log(1-z^*) \lesssim h \log (M). } Combining this with \R{Splus1} and \R{nstar_choice}, and using the fact that $h M^2 \leq 1$ once more, we now get that $S^{+}_{1} \lesssim h (M^{-2} + h )^{\alpha+1} \lesssim h M^{-2\alpha-2}$. Substituting this and \R{Sminus1} back into \R{S1} now gives \be{ \label{S1bound} S_1 \lesssim S^{-}_{1} \lesssim h M \log(M) + h M \lesssim h M \log(M), } which completes the estimate for $S_{1}$. The estimate for $S_{-1}$ is near-identical, except that we use \R{Jacobi_negative} as well as \R{Jacobi_Local} since $S_1$ sums over integrals contained in the negative portion of the interval. Hence we get \be{ \label{Sm1bound} S_{-1} \lesssim h M \log(M), } for this term as well. Next we need to estimate \bes{ S_0 = \int_{V_{n_0}} \nu^{(\alpha,\beta)}(t) \D t \int_{V_{n_0}} | (\phi_i(s) \phi_j(s))' | \D s } Since $V_{n_0} \subseteq [-2h,2h]$, we have that $\nu^{(\alpha,\beta)}(t) \lesssim 1$ for $t \in V_{n_0}$. Also, by \R{Jacobi_Local} and \R{Jacobi_negative}, we have $| (\phi_i(s) \phi_j(s))' | \lesssim M$, $s \in V_{n_0}$. Hence we get $S_0 \lesssim h M$. Combining this with \R{S1bound} and \R{Sm1bound} and substituting into \R{unif_norm_ineq} now gives \bes{ \| (P_M - P_M U^* U P_M )x \|_{\infty} \lesssim \sum^{M}_{j=1} h M \log M = h M^2 \log M, } from which the result follows immediately. } \lem{ \label{l:F_N_M_R_poly} For $\alpha,\beta > -1$ let $\{ \phi_i \}_{i \in \bbN}$ be the orthonormal Jacobi polynomial basis \R{Jacobi_ON}, $T= \{ t_n \}^{N}_{n=1} \subseteq D$ be a set of $N$ scattered data points and suppose that $h$ is as in \R{h_def}. Suppose that $h M^2 \leq 1$. If the weights $w_i \gtrsim i^{q+1/2}$, where $q$ is as in \R{Jacobi_ON_growth2}, then the quantity $F(h,M,R)$ defined by \R{F_N_M_R} satisfies \be{ \label{F_poly_gen} F(h,M,R) \lesssim \sqrt{M} \sup_{i > R} \left \{ i^{q+1/2} / w_i \right \}, } Moreover, if the weights $w_i \gtrsim i \log i$ and $R \geq M$ then \be{ \label{F_poly_log} F(h,M,R) \lesssim h M \sup_{i > R} \left \{ \frac{i \log i}{w_i} \right \}. } } \prf{ As before, let $x \in P_M(\ell^2(\bbN))$, $\| x \|_{\infty} = 1$ and set $g = \sum^{M}_{j=1} x_j \phi_j \in \bbP_{M-1}$. Then \bes{ \| P^{\perp}_R W^{-1} U^* U P_M x \|_{\infty} = \sup_{i > R} \left | \frac{1}{w_i} \ip{g}{\phi_i}_h \right | \leq \| g \|_h \sup_{i > R} \left \{ \frac{\| \phi_i \|_h}{w_i} \right \}. } Note that $\| \phi_i \|_h \leq \| \phi_i \|_{L^\infty} \lesssim i^{q+1/2}$ by \R{Jacobi_ON_growth} and also \eas{ \| g \|^2_h &= \ip{P_M U^* U P_M x}{x} \leq \left ( 1+\| P_M - P_M U^* U P_M \| \right ) \| x \|^2 \lesssim M, } where in the final inequality we use Lemma \ref{l:E2_Leg} and the fact that $\| x \| \leq \sqrt{M} \| x \|_{\infty} = \sqrt{M}$. This now gives \R{F_poly_gen}. For \R{F_poly_log} we use orthogonality and the fact that $R \geq M$ to get \be{ \label{FNMR_bison} \| P^{\perp}_R W^{-1} U^* U P_M x \|_{\infty} = \sup_{i > R} \left | \frac{1}{w_i} \ip{g}{\phi_i}_h \right | = \sup_{i > R} \left \{ \frac{1}{w_i} \left | \ip{g}{\phi_i}_{L^2_{\nu}} - \ip{g}{\phi_i}_h \right | \right \}, } We now proceed in a similar manner to the proof of Lemma \ref{l:Einf_Leg}. First, since $\| x \|_{\infty}=1$ we have \bes{ \left | \ip{g}{\phi_i}_{L^2_{\nu}} - \ip{g}{\phi_i}_h \right | \leq \sum^{M}_{j=1} \sum^{N}_{n=1} \int_{V_n} \nu^{(\alpha,\beta)}(t) \D t \int_{V_n} | (\phi_i(s) \phi_j(s))' | \D s. } We now argue in a similar way, using the fact that $j \leq M \leq R \leq i$. This gives \bes{ \left | \ip{g}{\phi_i}_{L^2_{\nu}} - \ip{g}{\phi_i}_h \right | \lesssim \sum^{M}_{j=1} h i \log(i) \lesssim M h i \log(i). } Substituting back into \R{FNMR_bison} now gives the required result. } We are now ready to prove Theorem \ref{t:Leg_poly_full_l1}: \prf{[Proof of Theorem \ref{t:Leg_poly_full_l1}] We use Theorem \ref{t:full_samp_recov} and the estimates of Lemmas \ref{l:E2_Leg}--\ref{l:F_N_M_R_poly}. Note that \bes{ \min_{M < i \leq R } \{ w_i \},\ \max_{i=1,\ldots,M} \{ w_i \} \asymp M^{\gamma+q+1/2},\quad M \rightarrow \infty. } Hence, we require $h$, $M$ and $R \geq M$ such that \bes{ E(h,M) < \epsilon,\quad E(h,R) \lesssim \epsilon,\quad F(h,M,R) \lesssim \epsilon M^{-\gamma-q-1/2}. } Since $R \geq M$, Lemmas \ref{l:E2_Leg} and \ref{l:Einf_Leg} give that the first two conditions are satisfied provided \bes{ h \lesssim \frac{\epsilon}{R^2 \log R}. } We now consider $F(h,M,R)$. Suppose first that $0 < \gamma \leq 1/2-q$. Then Lemma \ref{l:F_N_M_R_poly} gives that $F(N,M,R) \lesssim \epsilon M^{-\gamma-q-1/2}$ provided $R \gtrsim \epsilon^{-1/\gamma} M^{1+(q+1)/\gamma}$. Hence this results in the condition \be{ \label{hN_argh} h \lesssim \frac{c(\epsilon)}{M^{2+2(q+1)/\gamma} \log M}, } which gives the result for $0 < \gamma \leq 1/2-q$. Now suppose $\gamma > 1/2-q$. Then $w_i \gtrsim i^{\gamma+q+1/2} \gtrsim i \log i$ and therefore Lemma \ref{l:F_N_M_R_poly} gives $F(N,M,R) < \epsilon M^{-\gamma-q - 1/2}$ whenever \bes{ h \frac{\log R}{R^{\gamma+q-1/2}} < \epsilon M^{-3/2 - \gamma-q}. } Suppose that $R = c M$ for some $c$. Then the above condition holds, provided \bes{ h\lesssim \frac{\epsilon}{M^2 \log M}. } Moreover, substituting $R = c M$ into \R{hN_argh} gives precisely the same condition on $h$ in terms of $M$. Hence the result follows. } The proof of Theorem \ref{t:Jacobi_poly_LS} is straightforward: \prf{[Proof of Theorem \ref{t:Jacobi_poly_LS}] We use Theorem \ref{t:LS_err} in combination with Lemma \ref{l:E2_Leg}. } Finally, we also give the proof of Theorem \ref{t:Leg_trunc}: \prf{[Proof of Theorem \ref{t:Leg_trunc}] We shall use Lemma \ref{l:gamma}. Let $y \in \bbC^N$, $\| y \| =1$ be given. Let $\chi \in C^{\infty}_{c}(-1,1)$ be a smooth compactly-supported function in $(-1,1)$ with the properties \bes{ \| \chi \| = 1,\quad \chi(0) = 1,\quad 0 \leq \chi(x) \leq 1,\ x \in (-1,1). } Define \bes{ g(t) = \sum^{N}_{n=1} \frac{y_{n}}{\sqrt{\tau_n}} g_{n}(t),\quad g_{n}(t) = \chi \left ( \frac{x-t_n}{\xi_{n}} \right ), } where $\xi_{n} = \mathrm{dist}(t_n,\partial V_n)$ is the distance of the point $t_n$ from the boundary of its Voronoi cell (in this one-dimensional setting, $V_n$ is an interval and its boundary is the set of the two endpoints). Observe that \bes{ \xi_{n} = \frac12 \min \left \{ t_{n+1}-t_n , t_n - t_{n-1} \right \},\quad n=1,\ldots,N, } and therefore $\xi_{n} \geq \xi$ for each $n=1,\ldots,N$. By construction $\mathrm{supp}(g_{n}) \subseteq V_n$, $n=1,\ldots,N$, and therefore $\mathrm{supp}(g_{n}) \cap \mathrm{supp}(g_{m}) = 0$, $n \neq m$. It follows that $g \in G_y$. Since $g \in C^{\infty}[-1,1]$ a standard result in polynomial approximation gives that \bes{ \inf_{\substack{\phi \in \Phi_K \\ \phi \neq 0}} \| g - \phi \|_{\infty} \leq C_r K^{-r} \| g^{(r)} \|_{\infty},\quad K \geq r, } for some constant $C_r > 0$ independent of $K$ and $g$ (see \cite[(5.4.16)]{SMSD}). Observe that \bes{ \| g^{(r)} \|_{\infty} = \max_{n=1,\ldots,N} \frac{|y_{n}|}{\sqrt{\tau_n}} \| g^{(r)}_{n} \|_{\infty} \leq \| y \| \max_{n=1,\ldots,N} \frac{(\xi_{n})^{-r}}{\sqrt{\tau_n}} \leq \xi^{-r-1/2}, } where in the last inequality we use that fact that $\xi_{n} \leq \tau_n$. Hence \be{ \label{interp_poly_err} \inf_{\substack{\phi \in \Phi_K \\ \phi \neq 0}} \| g - \phi \|_{\infty} \leq C_r K^{-r} \xi^{-r-1/2},\quad K \geq r. } In order to apply \R{gamma_def}, it remains to estimate $\| g \|_{\nu} = \| g \|$. Since the $g_{n}$'s have disjoint supports, we have \bes{ \| g \|^2 = \sum^{N}_{n=1} \frac{|y_{n}|^2}{\tau_n} \| g_{n} \|^2 = \sum^{N}_{n=1} \frac{|y_{n}|^2}{\tau_n} \xi_{n} \geq \min_{n=1,\ldots,N} \{ \xi_{n} / \tau_n \} \geq \xi_{N} / \max_{n=1,\ldots,N} \tau_n. } Observe that $\tau_n \leq 2 h$. Therefore, $\| g \| \geq \sqrt{\xi / (2 h)}$. Substituting this and \R{interp_poly_err} into \R{gamma_def} now gives the result. } Finally, we now note that Theorem \ref{t:intro_thm} follows immediately from Theorems \ref{t:Leg_trunc} and \ref{t:Leg_poly_full_l1}. \section{Trigonometric polynomials on bounded intervals}\label{s:TrigPoly_Examp} We now consider Example \ref{ex:Trigonometric}. Note that in this case we define the projections $P_N : \ell^2(\bbZ) \rightarrow \ell^2(\bbZ)$ by $P_N x = \{\ldots,0,0,x_{-N},x_{-N+1},\ldots,x_{N-1},0,0,\ldots\}$. Our main result is as follows: \thm{ \label{t:Trig_Scatter} Let $\{ \phi_i\}_{i \in \bbZ}$ be the Fourier basis \R{fourier_basis}, $T = \{ t_n \}^{N}_{n=1} \subseteq [-1,1]$ be a set of $N$ scattered data points and suppose that $h$ is as in \R{h_def}. Suppose that the weights $w_{i} = 1+ |i|^{\gamma}$ for some $\gamma >0$. Then for each $0 < \epsilon < 1/2$ there exists a $c(\epsilon) > 0$ such that if \bes{ h \leq c(\epsilon) \left \{ \begin{array}{cc} M^{-3/2-3/(4\gamma)} & 0 < \gamma <1 \\ M^{-3/2} & \gamma > 1 \end{array} \right . , } then any minimizer $\hat{x}$ of \R{fin_min} satisfies \bes{ \| x - \hat{x} \| \leq C(\epsilon) \left [ M^{\gamma+1/2} \eta + \| P^{\perp}_M x \|_{1,w} + T_{N,K, \eta}(x) \right ). } for some constant $C$ depending on $\epsilon$ only, where $T_{N,K, \eta}(x)$ is as in \R{trunc}. } For least-squares fitting, we have the following: \thm{ \label{t:Trig_LS} Let $\{ \phi_i\}_{i \in \bbZ}$ be the Fourier basis \R{fourier_basis}, $T = \{ t_n \}^{N}_{n=1} \subseteq [-1,1]$ be a set of $N$ scattered data points and let $h$ be as in \R{h_def}. Then for each $0 < \epsilon < 1$ there exists a $c(\epsilon) > 0$ such that if \be{ \label{hcondTrigLS} h \leq c(\epsilon) M^{-1}, } then the solution $\check{x}$ of \R{LS_fit} exists uniquely and satisfies \bes{ \| x - \check{x} \| \leq \left ( 1 + \frac{1}{\sqrt{1-\epsilon}} \right ) \| x - P_M x \|_{1,w} + \frac{1}{\sqrt{1-\epsilon}} \eta, } for any $w = \{ w_i \}_{i \in \bbN}$ with $w_i \geq 1$. } Unlike the case of Jacobi polynomials, these results give a worse recovery guarantee for weighted $\ell^1$ minimization than that of least-squares fitting. However, we do not believe the scaling $h \lesssim M^{-3/2}$ is sharp, and instead we conjecture that the true scaling is $h \lesssim (M \log M)^{-1}$. Proving this conjecture is an open problem. We remark in passing that this conjecture holds in the special case where the data is equispaced (we omit the proof for brevity's sake). \subsection{Numerical examples} \begin{figure}[t] \begin{center} $\begin{array}{ccc} \includegraphics[width=4.5cm]{fig6_1.eps} & \includegraphics[width=4.5cm]{fig6_2.eps} & \includegraphics[width=4.5cm]{fig6_3.eps} \\ f(x) = \frac{1}{1+500 \cos^2(\pi x)} & f(x) = \cos(4 \pi x) \exp(\sin(40 \pi x)) & f(x) = \frac{1}{20/19-\sin(10 \pi x)} \end{array}$ \caption{ Numerical comparison of weighted $\ell^1$ minimization and least-squares fitting for approximation from jittered data. The error against $N$ is plotted for each method. The solid black line is weighted $\ell^1$ minimization with $K=4N$ and weights $w_{i} = \sqrt{i}$. The dashed lines are least squares with $M = c N$ and $c=\frac16,\frac14,\frac12,\frac23,\frac34,\frac56$. The solid blue line is oracle least squares based on choosing $M$ to minimize the error for a given $N$ and $f$. Random noise of magnitude $10^{-8}$ was added to the data. } \label{f:FourJitt} \end{center} \end{figure} In Fig.\ \ref{f:FourJitt} we give a comparison of the two techniques for jittered data. In alignment with the discussion above, these results suggest the scaling predicted by Theorem \ref{t:Trig_Scatter} is not optimal. In fact, weighted $\ell^1$ minimization performs better than least-squares fitting (including the oracle case) in all the examples. We suspect this strong performance is due in part to the presence of some sparsity in the functions considered, and the fact that jittered points are near-optimal points for the recovery of sparse trigonometric polynomials \cite{FoucartRauhutCSbook}. \subsection{Proofs} We first require the following lemma: \lem{ Let $\{ \phi_i\}_{i \in \bbZ}$ be the orthonormal Fourier basis \R{fourier_basis}, $T = \{ t_n \}^{N}_{n=1} \subseteq D$ be a set of $N$ scattered data points and suppose that $h$ is as in \R{h_def}. Suppose that $h M \leq 1$. If $E_2(h,M)$ and $E_{\infty}(h,M)$ are as in \R{E_2NM} and \R{E_2NM} respectively, then \bes{ E_{2}(h,M) \lesssim h M,\quad E_{\infty}(h,M) \lesssim h M^{3/2}. } } \prf{ Consider $E_{2}(h,M)$ first. As in the proof of Lemma \ref{l:E2_Leg}, let $x \in P_M(\ell^2(\bbZ))$, $\| x \|=1$ be arbitrary and set $g = \sum^{M-1}_{j=-M} x_j \phi_j$ so that $\| g \|_{L^2} = 1$. Arguing in an identical manner, we see that it suffices to show that \be{ \label{fourier_E2_desired} \sum^{N}_{n=1} \int_{V_n} | g(t) - g(t_n) |^2 \D t \lesssim h^2 M^2. } Observe that $| g(t) -g(t_n) |^2 \lesssim h \int_{V_n} |g'(s) |^2 \D s$ and therefore $\sum^{N}_{n=1} \int_{V_n} | g(t) - g(t_n) |^2 \D t \lesssim h^2 \| g' \|^2_{L^2}$. To get \R{fourier_E2_desired} we recall Bernstein's inequality $\| g' \|_{L^2} \lesssim M \| g \|_{L^2}$ for trigonometric polynomials. For $E_{\infty}(h,M)$ we let $x \in P_M(\ell^2(\bbZ))$ with $\| x \|_{\infty}=1$ and defined $g$ as before. As in the proof of Lemma \ref{l:Einf_Leg} it suffices to estimate \bes{ \max_{i=-M,\ldots,M-1} \left | \ip{g}{\phi_i}_{L^2} - \ip{g}{\phi_i}_h \right |. } Arguing in the standard way, we see that \bes{ \max_{i=-M,\ldots,M-1} \left | \ip{g}{\phi_i}_{L^2} - \ip{g}{\phi_i}_h \right | \lesssim h \| (g \phi_i)' \|_{L^1} \leq h\| (g \phi_i)' \|_{L^2} } Since $g \phi_{i}$ is a trigonometric polynomial of degree at most $2M$ Bernstein's inequality gives \bes{ \max_{i=-M,\ldots,M-1} \left | \ip{g}{\phi_i}_{L^2} - \ip{g}{\phi_i}_h \right | \lesssim h M \| x \|_{2} \leq h M^{3/2}, } where in the final inequality we use the Cauchy--Schwarz inequality the fact that $\| x \|_{\infty} = 1$. This now gives the estimate for $E_{\infty}(h,M)$. } \lem{ Let $\{ \phi_i\}_{i \in \bbZ}$ be the orthonormal Fourier basis \R{fourier_basis}, $T = \{ t_n \}^{N}_{n=1} \subseteq D$ be a set of $N$ scattered data points and suppose that $h$ is as in \R{h_def}. Suppose that $h M \leq 1$. Then the quantity $F(h,M,R)$ defined by \R{F_N_M_R} satisfies \bes{ F(h,M,R) \lesssim \sqrt{M} \sup_{i > R} \left \{ 1/w_{i} \right \}. } Moreover, if the weights $w_i \gtrsim i$ and $R \geq M$, then \bes{ F(h,M,R) \lesssim h \sqrt{M} \sup_{i>R} \left \{ i / w_i \right \}. } } \prf{ We argue as in the proof of Lemma \ref{l:F_N_M_R_poly}. In the first case, since the functions $\phi_i$ are uniformly bounded we have $\| P^{\perp}_R W^{-1} U^* U P_M x \|_{\infty} \leq \| g \|_{h} \sup_{i>R} \{ 1/w_{i} \}$, where $g = \sum^{M-1}_{j=-M} x_j \phi_j$ and $\| x \|_{\infty} = 1$. By the same argument, we find that $\| g \|_{h} \lesssim \| g \|_{L^2} \lesssim \sqrt{M}$, which gives the first result. Now consider the second. With $x$ and $g$ as before, we have \bes{ \| P^{\perp}_R W^{-1} U^* U P_M x \|_{\infty} = \sup_{i > R} \left \{ \frac{1}{w_i} \left | \ip{g}{\phi_i}_{L^2} - \ip{g}{\phi_i}_h \right | \right \}. } As in the previous lemma, we note that $\left | \ip{g}{\phi_i}_{L^2} - \ip{g}{\phi_i}_h \right | \leq h \| (g \phi_i)' \|_{L^2}$, and therefore by Bernstein's inequality $\left | \ip{g}{\phi_i}_{L^2} - \ip{g}{\phi_i}_h \right | \lesssim h i \| x \|_{2} \lesssim h i \sqrt{M}$. This gives the second result. } \prf{[Proof of Theorem \ref{t:Trig_Scatter}] With this lemma in hand, the proof is identical in manner to that of Theorem \ref{t:Leg_poly_full_l1}. We omit the details. } \section{Conclusions}\label{s:conclusions} We have presented an infinite-dimensional framework for weighted $\ell^1$ minimization. Its advantages are that it does not require \textit{a priori} knowledge of the expansion tail in order to be implemented and in the absence of noise it leads to interpolatory approximations. We have discussed the role weights play in the minimization in resolving the aliasing phenomenon, as opposed to promoting smoothness, and provided an explicit way to choose the truncation parameter. In the second half this paper we performed a linear error analysis for this framework valid for arbitrary scattered data, and used it to show near-optimal performance for Jacobi polynomial bases. There are several topics for future research. Three immediate problems are (i) to obtain a better scaling than \R{hcondTrigLS} in the trigonometric polynomial case, (ii) to estimate the truncation error $T_{h,K,\eta}(x)$ in a way that does not require additional regularity of $x$ (see Theorem \ref{p:trunc_err} and Remark \ref{r:trunc_weak}), and (iii) to improve the noise bound in Theorem \ref{t:Leg_poly_full_l1} (see Remark \ref{noise_growth}). Besides these, a question of singular importance are the extensions of \S \ref{s:AlgPoly_Examp} to higher dimensions and to unbounded intervals (using Laguerre and Hermite polynomials, for example). Other higher-dimensional problems can also be investigated, such as approximations in spherical harmonics (see also \cite{RauhutWardSpher}). Another topic is the optimal selection of weights. The results of this paper suggest that weights aid approximations from deterministic, scattered data by resolving the aliasing phenomenon and not necessarily by matching the decay of the expansion coefficients. In particular, slowly growing weights seem sufficient, at least in the one-dimensional setting. The situation may be different however in the multidimensional case when the samples are random. It has recently been shown in \cite{AdcockCSFunInterp,ChkifaDownwardsCS} that weighted $\ell^1$ minimization with a specific choice of weights leads to optimal approximation rates for certain classes of multivariate functions when the samples are drawn randomly from the orthogonality measure of the polynomial basis. We also note the possibility of using reweighted $\ell^1$ minimization, where weights iteratively updated to get a better estimation of the support set of $x$ \cite{PengHamptonDoostantweighted,KarniadakisUQCS}. We expect this technique can be combined with our framework. \section*{Acknowledgements} The work was supported by the Alfred P. Sloan Foundation and the Natural Sciences and Engineering Research Council of Canada through grant 611675. A preliminary version of this work was presented during the Research Cluster on ``Computational Challenges in Sparse and Redundant Representations'' at ICERM in November 2014. The author would like to thank the participants for the useful feedback received during the program. He would also like to thank Alireza Doostan, Anders Hansen, Rodrigo Platte, Aditya Viswanathan, Rachel Ward and Dongbin Xiu.

https://arxiv.org/abs/1506.01471

Minimal Gersgorin tensor eigenvalue inclusion set and its numerical approximation

For a complex tensor A, Minimal Gersgorin tensor eigenvalue inclusion set of A is presented, and its sufficient and necessary condition is given. Furthermore, we study its boundary by the spectrums of the equimodular set and the extended equimodular set for A. Lastly, for an irreducible tensor, a numerical approximation to Minimal Gersgorin tensor eigenvalue inclusion set is given.

\section{Introduction} Let $N=\{1,2,\ldots, n\}.$ For a complex (real) order $m$ dimension $n$ tensor $\mathcal {A}=(a_{i_1\cdots i_m})$ (written $\mathcal {A}\in \mathbb{C}^{[m,n]}$ ($\mathbb{R}^{[m,n]}$), respectively), where \[a_{i_1\cdots i_m}\in \mathbb{C}~ (\mathbb{R}), ~i_j=1,\ldots,n,~j=1,\ldots, m.\] we call a complex number $\lambda$ an eigenvalue of $\mathcal {A}$ and a nonzero complex vector $x$ an eigenvector of $\mathcal {A}$ associated with $\lambda$, if \begin{equation}\label{eigvalue}\mathcal {A}x^{m-1}=\lambda x^{[m-1]},\end{equation} where $\mathcal {A}x^{m-1}$ and $ x^{[m-1]}$ are vectors, whose $i$th components are \[(\mathcal {A}x^{m-1})_i=\sum\limits_{i_2,\ldots,i_m\in N} a_{ii_2\cdots i_m}x_{i_2}\cdots x_{i_m}\] and \[(x^{[m-1]})_i=x_i^{m-1},\] respectively. Note that there are other definitions of eigenvalue and eigenvectors, such as, H-eigenvalue, D-eigenvalue and Z-eigenvalue; see \cite{Qi3,Qi5,Qi2}. Obviously, the definition of eigenvalue for matrices follows from the case $m=2$. Tensor eigenvalues and eigenvectors have received much attention recently in the literatures \cite{Ch,La1,Ko1,Li,Ng,Qi3,Qi4,Qi1,Wa}. Many important results on the eigenvalue problem of matrices have been successfully extended to higher order tensors; see \cite{Ch,Ca1,Li1,Ng,Qi3,Qi4,Qi5,Ya,Ya1}. In \cite{Qi3}, Qi generalized Ger\v{s}gorin eigenvalue inclusion theorem from matrices to real supersymmetric tensors, which can be easily extended to generic tensors; see \cite{Ya}. \begin{theorem} \emph{\cite{Qi3}} \label{th 1.1} Let $\mathcal {A}=(a_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$ and $\sigma(\mathcal {A})$ be the spectrum of $\mathcal {A}$, that is, \[\sigma(\mathcal {A})=\{\lambda\in\mathbb{C}: \mathcal {A}x^{m-1}=\lambda x^{[m-1]},~ x\in \mathbb{C}^n\backslash\{0\}\}.\] Then \begin{equation}\label{eq 2.1} \sigma(\mathcal {A})\subseteq \Gamma(\mathcal {A})=\bigcup\limits_{i\in N} \Gamma_i(\mathcal {A}), \end{equation} where \[\Gamma_i(\mathcal {A})=\left\{z\in \mathbb{C}:|z-a_{i\cdots i}|\leq r_i(\mathcal {A})=\sum\limits_{i_2,\ldots,i_m\in N,\atop \delta_{ii_2\ldots i_m}=0} |a_{ii_2\cdots i_m}|\right\}\] and \[\delta_{i_1i_2\cdots i_m}= \left\{ \begin{array}{cc} \delta_{i_1i_2\cdots i_m}=1, & if ~i_1=\cdots =i_m,~\\ 0, & otherwise. \end{array} \right. \] \end{theorem} It is easy to see that when $m=2$, Theorem \ref{th 1.1} reduces to the well known Ger\v{s}gorin eigenvalue inclusion theorem of matrices \cite{Ger,Varga}. Here, we call $\Gamma_i(\mathcal {A})$ the $i$-th Ger\v{s}gorin tensor eigenvalue inclusion set. Note that $\Gamma_i(\mathcal {A})$ is a closed set in the complex plane $\mathbb{C}$, Hence, $\Gamma(\mathcal {A})$, which consists of the $n$ sets $\Gamma_i(\mathcal {A})$, is also closed and bounded in $\mathbb{C}$. Also in \cite{Qi3}, Qi obtain another interesting result on Ger\v{s}gorin tensor eigenvalue inclusion set $\Gamma(\mathcal {A})$. \begin{theorem} \emph{\cite{Qi3}} If $\Gamma_i(\mathcal {A})$ is disjoint with the other $\Gamma_j(\mathcal {A}),~j\neq i$, then there are exactly $(m-1)^{n-1}$ eigenvalues which lie in $\Gamma_i(\mathcal {A})$. Furthermore, if all of $\Gamma_i(\mathcal {A})$, $i=l_1,l_2,\ldots,l_k$ are connected but disjoint with the other $\Gamma_j(\mathcal {A})$, $j\neq i$, then there are exactly $k(m-1)^{n-1}$ eigenvalues which lie in $\bigcup\limits_{i=l_1,l_2, \ldots, l_k} \Gamma_i(\mathcal {A}) $. \end{theorem} In this paper, we also focus on Ger\v{s}gorin tensor eigenvalue inclusion set. And we present Minimal Ger\v{s}gorin tensor eigenvalue inclusion set, give its sufficient and necessary condition, and research its boundary. For an irreducible tensor, we give a set which approximates to Minimal Ger\v{s}gorin tensor eigenvalue inclusion set. \section{Minimal Ger\v{s}gorin tensor eigenvalue inclusion set} In this section, we present Minimal Ger\v{s}gorin tensor eigenvalue inclusion set and study its characteristic. First, a lemma is given. \begin{lemma} \emph{\cite{Ya}} \label{lemma 2.1} Let $\mathcal{A}=(a_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$ and $D=diag(d_1,d_2,\ldots,d_n)$ be a diagonal nonsingular matrix. If \[\mathcal{B}=(b_{i_1\cdots i_m})=\mathcal{A}D^{-(m-1)}\overbrace{DD\cdots D}\limits^{m-1},\] where \[b_{i_1\cdots i_m}=d_{i_1}^{-(m-1)}a_{i_1i_2\cdots i_m}d_{i_2}\cdots d_{i_m},~ i_1,\ldots, i_m\in N,\] then $ \mathcal{A}, ~\mathcal{B}$ have the same eigenvalues. \end{lemma} From Lemma \ref{lemma 2.1}, we obtain the following tensor eigenvalue inclusion set. \begin{theorem} \label{th 2.1} Let $\mathcal {A}=(a_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]} $ and $x=(x_1,x_2,\ldots,x_n)^T$ be an entrywise positive vector, i.e., $x=(x_1,x_2,\ldots,x_n)^T>0$. Then \begin{equation}\label{equation2.1}\sigma(\mathcal {A})\subseteq \Gamma^{r^x}(\mathcal {A})=\bigcup\limits_{i\in N}\Gamma^{r^x}_i(\mathcal {A}),\end{equation} where \[\Gamma^{r^x}_i(\mathcal {A})=\left\{z\in \mathbb{C}: |z-a_{i\cdots i}|\leq r^x_i(\mathcal {A})= \sum\limits_{i_2,\ldots,i_m\in N,\atop \delta_{ii_2\cdots i_m}=0} \frac{|a_{ii_2\cdots i_m}|x_{i_2}\cdots x_{i_m}}{x_{i}^{m-1}} \right\}.\] Furthermore, \begin{equation}\label{equation2.2}\sigma(\mathcal {A})\subseteq \bigcap\limits_{x>0}\Gamma^{r^x}(\mathcal {A}).\end{equation} \end{theorem} \begin{proof} Let $X=diag(x_1,x_2,\ldots,x_n)$ and $\mathcal {B}=\mathcal{A}X^{-(m-1)}\overbrace{XX\cdots X}\limits^{m-1}=(b_{i_1\cdots i_m})$. It is obvious that $X$ is nonsingular and $\sigma(\mathcal {B})=\sigma(\mathcal{A})$ from Lemma \ref{lemma 2.1}. From Theorem \ref{th 1.1}, we have \[\sigma(\mathcal {B})\subseteq \Gamma (\mathcal {B})=\bigcup\limits_{i\in N} \Gamma_i(\mathcal {B}), \] where \[\Gamma_i(\mathcal {B})=\left\{z\in \mathbb{C}:|z-b_{i\cdots i}|\leq r_i(\mathcal {B})=\sum\limits_{i_2,\ldots,i_m\in N,\atop \delta_{ii_2\ldots i_m}=0} |b_{ii_2\cdots i_m}|\right\}\] Note that $b_{i\cdots i}=a_{i\cdots i}$ and $r_i(\mathcal {B})=r^x_i(\mathcal {A})$. Hence, $\Gamma_i(\mathcal {B})=\Gamma^{r^x}_i(\mathcal {A})$, consequently, $\Gamma (\mathcal {B})=\Gamma^{r^x}(\mathcal {A})$. Therefore, \[ \sigma(\mathcal{A})=\sigma(\mathcal {B})\subseteq \Gamma (\mathcal {B})=\Gamma^{r^x}(\mathcal {A}). \] Furthermore, for any $x>0,~x\in \mathbb{R}^n$, (\ref{equation2.1}) also holds. Hence, (\ref{equation2.2}) follows. \end{proof} The set $\bigcap\limits_{x>0}\Gamma^{r^x}(\mathcal {A})$ in Theorem \ref{th 2.1} is of interest theoretically because it provides a set, containing all the eigenvalues of $\mathcal {A}$, which is called Minimal Ger\v{s}gorin tensor eigenvalue inclusion set defined as follows. \begin{definition} \label{minimal ger} Let $\mathcal {A}=(a_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$. Then \begin{equation} \Gamma^R(\mathcal {A})=\bigcap\limits_{x \in \mathbb{R}^n, \atop x>0}\Gamma^{r^x}(\mathcal {A}) \end{equation} is Minimal Ger\v{s}gorin tensor eigenvalue inclusion set of $\mathcal {A}$, related to the collection of all weighted sums, $r^x_i(\mathcal {A})$, where $x=(x_1,x_2,\ldots,x_n)^T>0$. \end{definition} Next, a sufficient and necessary condition for the elements belonging to Minimal Ger\v{s}gorin tensor eigenvalue inclusion set is provided. Before that, we give some results involving the $Perron-Frobenius$ theory of nonnegative tensors \cite{Ya}. Given a real tensor $\mathcal {A}$, we call $\mathcal {A}$ nonnegative, denoted by $\mathcal {A}\geq 0$, if every of its entries is nonnegative. \begin{definition} \cite[Definition 2.1]{Ch} A tensor $\mathcal {A}=(a_{i_1\cdots,i_m})\in \mathbb{C}^{[m,n]}$ is called reducible, if there exists a nonempty proper index subset $I\subset N$ such that \[|a_{i_1\cdots,i_m }|=0,~\forall i_1\in I, ~\forall i_2,\ldots,i_m\notin I.\] If $\mathcal {A}$ is not reducible, then we call $\mathcal {A}$ irreducible.\end{definition} \begin{lemma} \label{lemma 2.2}\emph{\cite{Ya}} \label{nonnegative-tensor-pf} If $\mathcal{A} \in \mathbb{R}^{[m,n]}$ is nonnegative, then the spectral radius \[\rho(\mathcal{A})=\max\{|\lambda|: \lambda \in \sigma(\mathcal {A})\}\] is an eigenvalue with a nonnegative eigenvector $x\neq 0$ corresponding to it. Moreover, if $\mathcal{A}$ is irreducible, then $\rho(\mathcal{A})>0$ and $x$ is positive. \end{lemma} \begin{lemma} \label{lemma 2.4} \emph{\cite[Theorem 5.3]{Ya}} Let $\mathcal {A} \in \mathbb{R}^{[m,n]}$ be nonnegative. Then \[\rho(\mathcal {A}) =\max\limits_{x\geq 0, x\neq 0} \min\limits_{x_i>0} \frac{(\mathcal {A}x^{m-1})_i} {x^{m-1}_i}= \min\limits_{x\geq 0, x\neq 0} \max\limits_{x_i>0} \frac{(\mathcal {A}x^{m-1})_i} {x^{m-1}_i}.\] \end{lemma} \begin{remark} The first equality is only given in Theorem 5.3 of \cite{Ya}. And similar to the proof of Theorem 5.3 of \cite{Ya}, the second is proved easily. \end{remark} From Lemma \ref{lemma 2.4}, we can obtain the following result. \begin{corollary} \label{corollary 2.1} Let $\mathcal {A} \in \mathbb{R}^{[m,n]}$ be nonnegative. Then \[\rho(\mathcal {A}) = \sup\limits_{x>0} \min\limits_{i\in N} \frac{(\mathcal {A}x^{m-1})_i} {x^{m-1}_i}=\inf\limits_{x>0} \max\limits_{i\in N} \frac{(\mathcal {A}x^{m-1})_i} {x^{m-1}_i}.\] \end{corollary} From Corollary \ref{corollary 2.1}, we have the following result. \begin{lemma} \label{lemma 2.5} Let $\mathcal {A}=(a_{i_1\cdots,i_m})\in \mathbb{C}^{[m,n]}$ and $z$ be any complex number. And let $\mathcal {B}(z)=(b_{i_1\cdots,i_m})\in \mathbb{R}^{[m,n]}$, where \[b_{i\cdots i}=-|z-a_{i\cdots i}|,~b_{ii_2\cdots,i_m}=|a_{ii_2\cdots,i_m}|~for ~ i\in N ~and~\delta_{ii_2\cdots,i_m}=0.\] Then $\mathcal {B}(z)$ possesses a real eigenvalue $v(z)$ which has the property that if $\lambda\in \sigma(\mathcal {B}(z))$, then \[Re(\lambda) \leq v(z).\] Furthermore, \begin{equation}\label{eq3.1}v(z)=\inf\limits_{x>0}\max\limits_{i\in N} \frac{(\mathcal {B}(z)x^{m-1})_i}{x_i^{m-1}}. \end{equation} \end{lemma} \begin{proof} Let $\mu=\max\limits_{i\in N}|z-a_{i\cdots i}|$, and $\mathcal {C}=(c_{i_1\cdots,i_m})\in \mathbb{R}^{[m,n]}$, where \begin{equation} \label{eq 2.2} c_{i\cdots i}=\mu-|z-a_{i\cdots i}|~and~c_{ii_2\cdots,i_m}=|a_{ii_2\cdots,i_m}|, ~i\in N, ~\delta_{ii_2\cdots,i_m}=0. \end{equation} Then $B(z)=-\mu\mathcal {I}+\mathcal {C}$, where $\mathcal {C}\geq 0$. From Equality (\ref{eigvalue}), it is obvious that $\lambda(\mathcal {C})\in \sigma (\mathcal {C})$ if and only if $-\mu+\lambda(\mathcal {C})\in \sigma (\mathcal {B}(z))$. Moreover, From Lemma \ref{lemma 2.2}, $\rho(\mathcal {C})$ is an eigenvalue of $\mathcal {C}$ with a nonnegative eigenvector. Hence, $-\mu+\rho(\mathcal {C})$ is an eigenvalue of $\mathcal {B}(z)$ with a nonnegative eigenvector. Let \begin{equation} \label{eq 2.3} v(z)=-\mu+\rho(\mathcal {C}).\end{equation} Obviously, $v(z)\in \sigma (\mathcal{B}(z))$ is real. And for any $\lambda \in \sigma(\mathcal {B}(z)),$ we have \[Re(\lambda)=Re(\lambda+\mu-\mu)=Re(\lambda+\mu)-\mu\leq |\lambda+\mu|-\mu\leq \rho(\mathcal {C})-\mu=v(z).\] Furthermore, by Corollary \ref{corollary 2.1}, we have \[\rho(\mathcal {C})=\inf\limits_{x>0} \max\limits_{i\in N} \frac{(\mathcal {C}x^{m-1})_i} {x^{m-1}_i}.\] Then \begin{eqnarray*}v(z)&=&-\mu+\rho(\mathcal {C})\\ &=&-\mu+\inf\limits_{x>0} \max\limits_{i\in N} \frac{(\mathcal {C}x^{m-1})_i} {x^{m-1}_i}\\ &=&\inf\limits_{x>0} \max\limits_{i\in N} \frac{((-\mu\mathcal {I}+\mathcal {C})x^{m-1})_i} {x^{m-1}_i}\\ &=&\inf\limits_{x>0} \max\limits_{i\in N} \frac{(\mathcal {B}(z)x^{m-1})_i} {x^{m-1}_i}.\end{eqnarray*} The conclusion follows. \end{proof} For $v(z)$ in Lemma \ref{lemma 2.5}, we give the following property. \begin{proposition}\label{lemma 4.2} Let $\mathcal {A}=(a_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$ and $v(z)$ be defined as Lemma \ref{lemma 2.5}. Then $v(z)$ is uniformly continuous on $\mathbb{C}$. \end{proposition} \begin{proof} Let $\mathcal {B}(z)\in \mathbb{R}^{[m,n]}$ be defined as Lemma \ref{lemma 2.5}. Then, similar to the proof of Theorem \ref{theorem3.1}, \[v(z)=\inf\limits_{x>0}\max\limits_{i\in N} \frac{(\mathcal {B}(z)x^{m-1})_i}{x_i^{m-1}}=\inf\limits_{x>0}\max\limits_{i\in N} \{r^{x}_i(\mathcal {A})-|z-a_{i\cdots i}|\}.\] Note that for any $z,~\tilde{z}\in \mathbb{C}$ and $x>0,~x\in \mathbb{R}^n$, \[\begin{array}{lll} \max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-|z-a_{i\cdots i}|\}& = & \max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-|z-\tilde{z}+\tilde{z}-a_{i\cdots i}|\} \\ &\geq&\max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-(|z-\tilde{z}|+|\tilde{z}-a_{i\cdots i}|)\}\\ &=&\max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-|\tilde{z}-a_{i\cdots i}|\}-|z-\tilde{z}|, \end{array}\] that is, \[\max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-|\tilde{z}-a_{i\cdots i}|\}\leq |z-\tilde{z}|+\max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-|z-a_{i\cdots i}|\}.\] Hence, \[\begin{array}{lll} \inf\limits_{x>0}\max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-|\tilde{z}-a_{i\cdots i}|\} &\leq& \inf\limits_{x>0} \left\{|z-\tilde{z}|+\max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-|z-a_{i\cdots i}|\}\right\}\\ &=&|z-\tilde{z}|+\inf\limits_{x>0}\max\limits_{i\in N}\{r^{x}_i(\mathcal {A})-|z-a_{i\cdots i}|\}, \end{array}\] which implies \[v(\tilde{z})-v(z)\leq |z-\tilde{z}|.\] Similarly, we can obtain \[v(z)-v(\tilde{z})\leq |z-\tilde{z}|.\] Therefore, \[|v(z)-v(\tilde{z})|\leq |z-\tilde{z}|.\] This implies that $v(z)$ is uniformly continuous on $\mathbb{C}$. \end{proof} We now establish the sufficient and necessary condition for $\Gamma^{R}(\mathcal {A})$. \begin{theorem} \label{theorem3.1} Let $\mathcal {A}, ~\mathcal {B}(z)$ and $v(z)$ be defined as Lemma \ref{lemma 2.5}. Then \begin{equation}\label{eq3.2}z\in \Gamma^{R}(\mathcal {A})~if ~and ~only ~if ~v(z)\geq 0.\end{equation} \end{theorem} \begin{proof} Assume that $z\in \Gamma^{R}(\mathcal {A})$. From Definition \ref{minimal ger}, we have that for each vector $x>0$, $z\in \Gamma^{r^x}(\mathcal {A})$. Hence, there is an $i_0\in N$ such that \[|z-a_{i_0\cdots i_0}|\leq r^{x}_{i_0}(\mathcal {A}),\] equivalently, \[r^{x}_{i_0}(\mathcal {A})-|z-a_{i_0\cdots i_0}|\geq 0.\] Note that for any $i\in N$, \[\frac{(\mathcal {B}(z)x^{m-1})_i}{x_i^{m-1}}=r^{x}_i(\mathcal {A})-|z-a_{i\cdots i}|.\] Then $\frac{\left(\mathcal {B}(z)x^{m-1}\right)_{i_0}}{x_{i_0}^{m-1}}\geq 0$, which implies that for each vector $x>0$, \[\max\limits_{i\in N} \frac{(\mathcal {B}(z)x^{m-1})_i}{x_i^{m-1}}\geq 0.\] By Lemma \ref{lemma 2.5}, \[v(z)=\inf\limits_{x>0}\max\limits_{i\in N} \frac{(\mathcal {B}(z)x^{m-1})_i}{x_i^{m-1}}\geq 0.\] Conversely, suppose that $v(z)\geq 0$. From Equality (\ref{eq3.1}), then for each vector $x>0$, there is an $i\in N$ such that \[0\leq v(z)\leq \frac{(\mathcal {B}(z)x^{m-1})_i}{x_i^{m-1}}=r^{x}_i(\mathcal {A})-|z-a_{i\cdots i}|.\]Hence, $|z-a_{i\cdots i}|\leq r^{x}_i(\mathcal {A})$, and then $z\in \Gamma^{r^x}_i(\mathcal {A})$ which implies $z\in \Gamma^{r^x}(\mathcal {A})$. Since this inclusion holds for each vector $x>0$, we have $z\in \Gamma^R(\mathcal {A})$ from Definition \ref{minimal ger}. \end{proof} \section{Boundary of $\Gamma^{R}(\mathcal {A})$} In Section 2, a sufficient and necessary condition for the elements belonging to Minimal Ger\v{s}gorin tensor eigenvalue inclusion set $\Gamma^{R}(\mathcal {A})$, is given. We in this section focus on the boundary of $\Gamma^{R}(\mathcal {A})$, and establish relationships between its boundary (see Definition \ref{boundary}), the spectrum of the equimodular set for $\mathcal {A}$ (see Definition \ref{equimodular}), the spectrum of the extended equimodular set for $\mathcal {A}$ (see Definition \ref{extended equimodular}) and $\Gamma^{R}(\mathcal {A})$. \begin{definition}\label{boundary} Let $\mathcal{A} \in \mathbb{C}^{[m,n]}$. The boundary of Minimal Ger\v{s}gorin tensor eigenvalue inclusion set $\Gamma^{R}(\mathcal {A})$, denoted by $\partial \Gamma^{R}(\mathcal {A})$, is defined by \[\partial\Gamma^{R}(\mathcal {A})=\overline{\Gamma^{R}(\mathcal {A})}\bigcap \overline{(\Gamma^{R}(\mathcal {A}))'},\] where $(\Gamma^{R}(\mathcal {A}))'$ is the complement of $\Gamma^{R}(\mathcal {A})$, and $\overline{\Gamma^{R}(\mathcal {A})}$ is the closure of $\Gamma^{R}(\mathcal {A})$. \end{definition} \begin{definition}\label{equimodular} Let $\mathcal{A}=(a_{i_1\cdots i_m}) \in \mathbb{C}^{[m,n]}$. The equimodular set of $\mathcal{A}$, denoted by $\Omega(\mathcal{A})$, is defined as \[\Omega(\mathcal{A})=\{\mathcal {Q}=(q_{i_1\cdots i_m}) \in \mathbb{C}^{[m,n]}: q_{i\cdots i}= a_{i\cdots i}, |q_{ii_2\cdots i_m}|= |a_{ii_2\cdots i_m}|, ~i\in N, \delta_{ii_2\cdots i_m}=0\}.\] \end{definition} \begin{definition}\label{extended equimodular} Let $\mathcal{A}=(a_{i_1\cdots i_m}) \in \mathbb{C}^{[m,n]}$. The extended equimodular set of $\mathcal{A}$, denoted by $\hat{\Omega}(\mathcal{A})$, is defined as \[\hat{\Omega}(\mathcal{A})=\{\mathcal {Q}=(q_{i_1\cdots i_m}) \in \mathbb{C}^{[m,n]}: q_{i\cdots i}= a_{i\cdots i}, |q_{ii_2\cdots i_m}|\leq |a_{ii_2\cdots i_m}|, i\in N, \delta_{ii_2\cdots i_m}=0\}.\] \end{definition} A sufficient and necessary condition for the point lying on the boundary of $\Gamma^{R}(\mathcal {A})$ is given as follows. \begin{proposition}\label{boundary nec} Let $\mathcal {A}$ and $v(z)$ be defined as Lemma \ref{lemma 2.5}. $z\in \partial \Gamma^{R}(\mathcal {A})$ if and only if $v(z)= 0$, and there is a sequence of $\{z_j\}_{j=1}^{\infty} \subseteq (\Gamma^{R}(\mathcal {A}))'$ (i.e., $v(z_j)<0$ for all $j\geq 1$) such that $\lim\limits_{j\rightarrow\infty} z_j=z$. \end{proposition} \begin{proof} Note that $\Gamma^{R}(\mathcal {A})$ is a compact set in the complex plane. Hence, from Definition \ref{boundary}, \[\partial\Gamma^{R}(\mathcal {A})=\overline{\Gamma^{R}(\mathcal {A})}\bigcap \overline{(\Gamma^{R}(\mathcal {A}))'}=\Gamma^{R}(\mathcal {A})\bigcap \overline{(\Gamma^{R}(\mathcal {A}))'}.\] Furthermore, by Theorem \ref{theorem3.1}, we have that \begin{equation}\label{eq4.1}z\in \overline{(\Gamma^{R}(\mathcal {A}))'}~if ~and ~only ~if ~v(z)\leq 0.\end{equation} Therefore, if $z\in \partial \Gamma^{R}(\mathcal {A})$, then, from (\ref{eq3.2}) and (\ref{eq4.1}), $v(z)= 0$. Note that $(\Gamma^{R}(\mathcal {A}))'$ is open and unbounded, then, by (\ref{eq4.1}) and $z\in \overline{(\Gamma^{R}(\mathcal {A}))'}$, there is a sequence of $\{z_j\}_{j=1}^{\infty} \subseteq (\Gamma^{R}(\mathcal {A}))' $ such that $\lim\limits_{j\rightarrow\infty} z_j=z$, where $v(z_j)<0$ for all $j\geq 1.$ Conversely, it is obvious by assumption that $z\in \overline{(\Gamma^{R}(\mathcal {A}))'}$ and $z\in \Gamma^{R}(\mathcal {A})=\overline{\Gamma^{R}(\mathcal {A})}$, that is, \[z\in \partial \Gamma^{R}(\mathcal {A}).\] The proof is completed. \end{proof} \begin{proposition}\label{pro 4.2} For any $\mathcal {A}=(a_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$ and any $z\in \mathbb{C}$ with $v(z)=0$, there is a tensor $\mathcal {Q}=(q_{i_1\cdots i_m})\in \Omega(\mathcal{A})$ for which $z$ is an eigenvalue of $\mathcal {Q}$. Then \[ \partial \Gamma^{R}(\mathcal {A})\subseteq \sigma (\Omega(\mathcal{A}))\subseteq \sigma (\hat{\Omega}(\mathcal{A}))\subseteq \Gamma^{R}(\mathcal {A}), \] where $\sigma (\Omega(\mathcal{A}))=\bigcup\limits_{\mathcal{D}\in \Omega(\mathcal{A})} \sigma(\mathcal{D})$ and $\sigma (\hat{\Omega}(\mathcal{A}))=\bigcup\limits_{\mathcal{D}\in \hat{\Omega}(\mathcal{A})} \sigma(\mathcal{D})$. \end{proposition} \begin{proof} From Definition \ref{minimal ger}, \ref{equimodular} and \ref{extended equimodular}, it is obvious that $\Omega(\mathcal{A})\subseteq \hat{\Omega}(\mathcal{A})$ and $\sigma (\Omega(\mathcal{A}))\subseteq \sigma (\hat{\Omega}(\mathcal{A}))\subseteq \Gamma^{R}(\mathcal {A})$. Hence, we next only prove $\partial \Gamma^{R}(\mathcal {A})\subseteq \sigma (\Omega(\mathcal{A}))$. If $z\in \mathbb{C}$ is such that $v(z)=0$, then, from the proof of Lemma \ref{lemma 2.5}, there is a vector $y=(y_1,\ldots, y_n)^T\geq 0$, $y\neq 0$ such that $\mathcal {B}(z)y=0,$ where $\mathcal {B}(z)$ is defined as Lemma \ref{lemma 2.5}. This implies that for any $k\in N$ \begin{equation}\label{eq4.2} |z-a_{k\cdots k}|y_k^{m-1} =\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}. \end{equation} Now, let $\psi_k$ satisfy \[ z-a_{k\cdots k}=|z-a_{k\cdots k}|e^{\emph{\textbf{i}}\psi_{k}},~ k\in N, \] and $\mathcal {Q}=(q_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$, where \[q_{k\cdots k}=a_{k\cdots k} ~and ~ q_{kk_2\cdots k_m}=|a_{kk_2\cdots k_m}|e^{\textbf{\emph{i}}\psi_{k}}, ~k\in N,~\delta_{kk_2\cdots k_m}=0.\] Hence, from Definition \ref{equimodular}, $\mathcal {Q}\in \Omega(\mathcal{A})$. Moreover, By considering the $k$-th entry $(\mathcal{Q}y^{m-1})_k$ of $\mathcal{Q}y^{m-1}$, we have that for any $k\in N$, \begin{eqnarray*}(\mathcal{Q}y^{m-1})_k&=&\sum\limits_{k_2,\ldots,k_m\in N}q_{kk_2\cdots k_m}y_{k_2}\cdots y_{k_m}\\ &=&q_{k\cdots k}y_{k}^{m-1}+ e^{\emph{\textbf{i}}\psi_k}\left( \sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}\right)\\ &=&(z-(z-a_{k\cdots k}))y_{k}^{m-1}+ e^{\textbf{\emph{i}}\psi_k}\left( \sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}\right)\\ &=&zy_{k}^{m-1}+ e^{\emph{\textbf{i}}\psi_k}\left( -|z-a_{k\cdots k}|y_{k}^{m-1}+ \sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}\right)\\ &=& zy_{k}^{m-1}, ~(By~ Equality~ (\ref{eq4.2})) \end{eqnarray*} that is, \[\mathcal{Q}y^{m-1}=zy^{[m-1]}.\] Note that $y\neq 0$, then $z$ is an eigenvalue of $\mathcal{Q}\in \Omega(\mathcal{A})$, which shows that $v(z)=0$ implies $z\in \sigma(\Omega(\mathcal{A}))$. From Proposition \ref{boundary nec}, we have that for each point $z\in \partial \Gamma^{R}(\mathcal {A})$, $v(z)=0$, consequently, $z\in \sigma(\Omega(\mathcal{A}))$. Hence, $\partial \Gamma^{R}(\mathcal {A})\subseteq \sigma (\Omega(\mathcal{A})) $. The proof is completed. \end{proof} For the sets $\sigma (\hat{\Omega}(\mathcal{A}))$ and $\Gamma^{R}(\mathcal {A})$, we have the following result. \begin{proposition}\label{pro 4.3} Let $\mathcal {A}=(a_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$. Then \[\sigma (\hat{\Omega}(\mathcal{A}))= \Gamma^{R}(\mathcal {A}).\] \end{proposition} \begin{proof} From Proposition \ref{pro 4.2}, we have $\sigma (\hat{\Omega}(\mathcal{A}))\subseteq \Gamma^{R}(\mathcal {A})$. Hence, we only prove $\Gamma^{R}(\mathcal {A})\subseteq \sigma (\hat{\Omega}(\mathcal{A}))$. Let $z\in \Gamma^{R}(\mathcal {A})$. Then, from Theorem \ref{theorem3.1}, $v(z)\geq 0$. And from the proof of Lemma \ref{lemma 2.5}, there is a vector $y=(y_1,\ldots, y_n)^T\geq 0$, $y\neq 0$ such that \[\mathcal {B}(z)y^{m-1}=v(z)y^{[m-1]},\] where $\mathcal {B}(z)$ is defined as Lemma \ref{lemma 2.5}. Hence for any $k\in N$, \begin{equation}\label{eq4.3} (|z-a_{k\cdots k}|+v(z))y_k^{m-1} =\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m} \end{equation} Now, let $\mathcal {Q}=(q_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$, with \[q_{k\cdots k}=a_{k\cdots k} ~and ~ q_{kk_2\cdots k_m}=\mu_k a_{kk_2\cdots k_m}, ~k\in N,~\delta_{kk_2\cdots k_m}=0,\] where \[\mu_k=\left\{\begin{array}{cc} \frac{\left( \sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}\right)-v(z)y_k^{m-1}}{\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}}, & if ~\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}>0, \\ 1, &if ~\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}=0. \end{array} \right.\] Furthermore, from Equality (\ref{eq4.3}) and the fact that both $|z-a_{k\cdots k}|y_k^{m-1}\geq 0$ and $v(z)y_k^{m-1}\geq 0$ hold for any $k\in N$, we easily obtain $0\leq \mu_k \leq 1$ for any $k\in N$. Hence, \[\mathcal {Q}\in \hat{\Omega}(\mathcal{A}).\] For the tensor $\mathcal {Q}$, we have from Equality (\ref{eq4.3}) that for any $k\in N$, \begin{eqnarray*}|z-q_{k\cdots k}|y_k^{m-1}&=&|z-a_{k\cdots k}|y_k^{m-1}\\ &=&\left(\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}\right)-v(z)y_k^{m-1}\\ &=&\mu_k\left(\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |a_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}\right)\\&=&\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |q_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m}, \end{eqnarray*} i.e., \[|z-q_{k\cdots k}|y_k^{m-1}=\sum\limits_{k_2,\dots,k_m\in N,\atop \delta_{kk_2\cdots k_m}=0} |q_{kk_2\cdots k_m}|y_{k_2}\cdots y_{k_m} ~for ~any~k\in N.\] Now, the above expression is exactly of the form of Equality (\ref{eq4.2}) in the proof of Proposition \ref{pro 4.2}. Hence, similar to the proof of Proposition \ref{pro 4.2}, we have that there is a tensor $\mathcal {P} \in \Omega(\mathcal{Q})$ such that $z\in \sigma (\mathcal {P})$. Note that $ \mathcal {P}\in \Omega(\mathcal{Q})$ and $\mathcal{Q}\in \hat{\Omega}(\mathcal{A})$. Therefore, $\mathcal {P}\in \hat{\Omega}(\mathcal{A})$, consequently, $z\in \sigma ( \hat{\Omega}(\mathcal{A}))$ and $\Gamma^{R}(\mathcal {A})\subseteq \sigma (\hat{\Omega}(\mathcal{A}))$. The proof is completed. \end{proof} From Propositions \ref{pro 4.2} and \ref{pro 4.3}, we can obtain the following relationships. \begin{theorem}\label{theorem 4.1} Let $\mathcal {A}\in \mathbb{C}^{[m,n]}$. Then \[\partial \Gamma^{R}(\mathcal {A})\subseteq \sigma (\Omega(\mathcal{A}))\subseteq \Gamma^{R}(\mathcal {A}).\] \end{theorem} \begin{remark} From Proposition \ref{pro 4.3}, we known that $\sigma (\hat{\Omega}(\mathcal{A}))$ completely fills out $\Gamma^R(\mathcal {A})$, that is, $\sigma (\hat{\Omega}(\mathcal{A}))= \Gamma^{R}(\mathcal {A})$. And from Theorem \ref{theorem 4.1}, we known that if $\sigma (\Omega(\mathcal{A}))$ is a proper subset of $\Gamma^{R}(\mathcal {A})$ with $\sigma (\Omega(\mathcal{A}))\neq \partial \Gamma^{R}(\mathcal {A})$, then $\sigma (\Omega(\mathcal{A}))$ necessarily have internal boundaries in $\Gamma^{R}(\mathcal {A})$. \end{remark} \section{A numerical approximation of Minimal Ger\v{s}gorin tensor eigenvalue inclusion set} Unlike Ger\v{s}gorin tensor eigenvalue inclusion set $\Gamma(\mathcal {A})$, or $\Gamma^{r^x}(\mathcal {A})$ of (\ref{equation2.1}), Minimal Ger\v{s}gorin tensor eigenvalue inclusion set $\Gamma^{R}(\mathcal {A})$ of a complex tensor $\mathcal {A}$ is not easy to determine numerically generally. In this section, for an irreducible tensor $\mathcal {A}$, we give a numerical approximation of $\Gamma^{R}(\mathcal {A})$, which contains $\Gamma^{R}(\mathcal {A})$. We now give a lemma, which will be used below. \begin{lemma} \label{lemma 4.1} Let $\mathcal {A}=(a_{i_1\cdots i_m})\in \mathbb{C}^{[m,n]}$ be irreducible, and $v(z)$ be defined as Lemma \ref{lemma 2.5}. Then for each $i\in N$, \[v(a_{i\cdots i})>0.\] Furthermore, for each $a_{i\cdots i}$ and for each real $\theta$ with $0\leq \theta < 2\pi$, let $\tilde{\gamma}_i(\theta)>0$ be the smallest $\gamma>0$ for which \begin{equation}\label{eq5.0}\left\{\begin{array}{l} v(a_{i\cdots i}+\tilde{\gamma}_i(\theta)e^{\textbf{i}\theta})=0, \\ ~ and~there ~is~ a~ sequence~ of ~\{z_j\}_{j=1}^\infty ~with\\ \lim\limits_{j\rightarrow\infty} z_j=a_{i\cdots i}+\tilde{\gamma}_i(\theta)e^{\textbf{i}\theta}~ such ~that~ v(z_j)<0~ for~ all~ j\geq 1. \end{array} \right.\end{equation} Then, the complex interval $[ a_{i\cdots i}+te^{\textbf{i}\theta}]$, for $0\leq t\leq \tilde{\gamma}_i(\theta) $, is contained in $\Gamma^{R}(\mathcal {A})$ for each real $\theta\in [0,2\pi)$, that is, the set \[\bigcup\limits_{\theta~\in [0,2\pi) } [ a_{i\cdots i}+te^{\textbf{i}\theta}]_{t=0}^{\tilde{\gamma}_i(\theta)},~i\in N,\] is a subset of $\Gamma^{R}(\mathcal {A})$. \end{lemma} \begin{proof} Let $\mathcal {B}(z)=(b_{i_1\cdots i_m})\in \mathbb{R}^{[m,n]}$ be defined as Lemma \ref{lemma 2.5}, where, \[b_{i\cdots i}=-|z-a_{i\cdots i}|~and~b_{ii_2\cdots,i_m}=|a_{ii_2\cdots,i_m}|, ~ i\in N, ~\delta_{ii_2\cdots,i_m}=0.\] And let $\mathcal {C}=(c_{i_1\cdots,i_m})\in \mathbb{R}^{[m,n]}$, where \[c_{i\cdots i}=\mu-|z-a_{i\cdots i}|~and~c_{ii_2\cdots,i_m}=|a_{ii_2\cdots,i_m}|, ~i\in N, ~\delta_{ii_2\cdots,i_m}=0,\] $\mu=\max\limits_{i\in N}|z-a_{i\cdots i}|$. Then $B(z)=-\mu\mathcal {I}+\mathcal {C}$. Since $\mathcal {A}$ is irreducible, $\mathcal {B}(z)$, also $\mathcal {C}$, is irreducible. Hence, $\mathcal {C}$ is an irreducible nonnegative tensor. From Lemma \ref{lemma 2.2}, there is a positive eigenvector of $\mathcal {C}$ corresponding to $\rho (\mathcal {C})$. Therefore, similar to the proof of \ref{lemma 2.5}, there is a positive eigenvector of $\mathcal {B}(z)$ corresponding to $v(z)$. Moreover, by Equality (\ref{eq3.1}), that is, \[v(z)=\inf\limits_{x>0}\max\limits_{i\in N} \frac{(\mathcal {B}(z)x^{m-1})_i}{x_i^{m-1}},\] we get that for any $z$, there exists a positive vector $y$ such that for all $j\in N$, \begin{equation} \label{eq5.1} v(z)=\frac{(\mathcal {B}(z)y^{m-1})_j}{y_j^{m-1}}.\end{equation} Then for any $i\in N$, take $z=a_{i\cdots i}$ and let $x=(x_1,\ldots,x_n)^T>0$ be such that for any $j\in N$, $v(a_{i\cdots i})=\frac{(\mathcal {B}(z)x^{m-1})_j}{x_j^{m-1}}$. In particular, \[v(a_{i\cdots i})=\frac{(\mathcal {B}(z)x^{m-1})_i}{x_i^{m-1}}=r^{x}_i(\mathcal {A})-|a_{i\cdots i}-a_{i\cdots i}|=r^{x}_i(\mathcal {A}).\] Since $\mathcal {A}$ is irreducible, we have $r^{x}_i(\mathcal {A})>0$, consequently, $v(a_{i\cdots i})>0$. Furthermore, for each $\theta$ with $0\leq \theta< 2\pi$, $a_{i\ldots i}+te^{\emph{\textbf{i}}\theta}$, for all $t\geq 0$, is the semi-infinite complex line. Obviously, the function $ v(a_{i\ldots i}+te^{\emph{\textbf{i}}\theta})$ is positive at $t=0$, is continuous on this line, and is negative outside $\Gamma^R(\mathcal {A})$ from Theorem \ref{theorem3.1}. Hence, there is a smallest $\tilde{\gamma}_i(\theta)>0$ satisfying (\ref{eq5.0}). And by Proposition \ref{boundary nec}, we get that $a_{i\ldots i}+\tilde{\gamma}_i(\theta)e^{\textbf{\emph{i}}\theta} \in \partial \Gamma^R(\mathcal {A}).$ This implies that for each real $\theta$, [$a_{i\ldots i}$, $a_{i\ldots i}+\tilde{\gamma}_i(\theta)e^{\textbf{\emph{i}}\theta}$] is a subset of $\Gamma^R(\mathcal {A})$. The conclusion follows. \end{proof} We now give the following procedure for approximating Minimal Ger\v{s}gorin tensor eigenvalue inclusion set $\Gamma^R(\mathcal {A})$ for an irreducible tensor $\mathcal {A}=(a_{i_1\cdots,i_m})\in \mathbb{C}^{[m,n]}$. \textbf{Procedure of numerical approximation } Step 1. determine the positive numbers $\{ v(a_{j\cdots j})\}_{j\in N}$; Step 2. determine the largest $\tilde{\gamma}_j(\theta)$ for $0\leq \theta< 2\pi$, such that $a_{j\cdots j}+\tilde{\gamma}_j(\theta)e^{\emph{\textbf{i}}\theta} \in \partial \Gamma^R(\mathcal {A})$, i.e., \begin{equation}\label{eq4.4} v(a_{j\cdots j}+\tilde{\gamma}_j(\theta)e^{\emph{\textbf{i}}\theta})=0, ~with~ v(a_{j\cdots j}+(\tilde{\gamma}_j(\theta)+\varepsilon)e^{\emph{\textbf{i}}\theta})< 0 \end{equation} for all sufficiently small $\varepsilon>0$; Step 3. take the $m$ points $w_{k_{j_\theta}}=a_{j\cdots j}+\tilde{\gamma}_j(\theta)e^{\textbf{\emph{i}}\theta} \in \partial \Gamma^R(\mathcal {A})$ for $k_{j_\theta}=1,2,\ldots, m$, and detemine the set $ \bigcap\limits_{k_{j_\theta}=1}^m \Gamma^{w_{k_{j_\theta}}}(\mathcal {A})$ approximating to $\Gamma^R(\mathcal {A})$, where \[\Gamma^{w_{k_{j_\theta}}}(\mathcal {A})=\bigcup\limits_{i\in N}\{z\in \mathbb{C}:|z-a_{i\cdots i}|\leq |w_{k_{j_\theta}}-a_{i\cdots i}|\}.\] \begin{remark} \label{remark1} (i) To determine $v(a_{j\cdots j})$, we use Equality (\ref{eq 2.3}) with $z=a_{j\cdots j}$, i.e., \[v(a_{j\cdots j})=-\mu+\rho(\mathcal {C}),\] where $\mu= \max\limits_{i\in N}|a_{j\cdots j}-a_{i\cdots i}|$ and $\rho(\mathcal {C})$ is an eigenvalue of the irreducible nonnegative tensor $\mathcal {C}$ defined as (\ref{eq 2.2}) (note that the irreducibility of $\mathcal {C}$ is deduced by that of $\mathcal {A}$), and the following method for calculating $\rho(\mathcal {C})$ (see \cite{Li1,Ng}), i.e., if for an vector $x^{(0)}> 0$, let $\mathcal {D}=\mathcal {C}+h\mathcal {I}$, where $h>0$, and let $y^{(0)} = \mathcal {D}\left(x^{(0)}\right)^{m-1}$, \[\begin{array}{ll} x^{(1)}=\frac{\left(y^{(0)}\right)^{[\frac{1}{m-1}]}}{\parallel\left(y^{(0)}\right)^{[\frac{1}{m-1}]}\parallel},& y^{(1)} = \mathcal {D}\left(x^{(1)}\right)^{m-1},\\ x^{(2)}=\frac{\left(y^{(1)}\right)^{[\frac{1}{m-1}]}}{\parallel\left(y^{(1)}\right)^{[\frac{1}{m-1}]}\parallel},& y^{(2)} = \mathcal {D}\left(x^{(2)}\right)^{m-1},\\ \vdots &\vdots\\ x^{(k+1)}=\frac{\left(y^{(k)}\right)^{[\frac{1}{m-1}]}}{\parallel\left(y^{(k)}\right)^{[\frac{1}{m-1}]}\parallel},& y^{(k+1)} = \mathcal {D}\left(x^{(k+1)}\right)^{m-1},~k\geq 2\\ \vdots &\vdots\\ \end{array}\] and let \[\underline{\lambda}_k=\min\limits_{x_i^{(k)}>0} \frac{\left(y^{(k)}\right)_i}{\left(x_i^{(k)}\right)^{m-1}},~ \overline{\lambda}_k=\max\limits_{x_i^{(k)}>0} \frac{\left(y^{(k)}\right)_i}{\left(x_i^{(k)}\right)^{m-1}},~k=1,2,\ldots,\] then \[\underline{\lambda}_1\leq \underline{\lambda}_2\leq \cdots \leq \rho(\mathcal {D})=\rho(\mathcal {C})+h \leq \cdots \leq \overline{\lambda}_2\leq \overline{\lambda}_1.\] moreover, \begin{equation} \lim\limits_{k\longrightarrow \infty}\underline{\lambda}_k= \rho(\mathcal {D})=\rho(\mathcal {C})+h=\lim\limits_{k\longrightarrow \infty}\overline{\lambda}_k. \end{equation} Hence, we can obtain convergent upper and lower estimates of $v(a_{j\cdots j})$, which do not need great accuracy for graphing purpose, as Example \ref{example4.1} shows. (ii) The numerical estimation of $\tilde{\gamma}_j(\theta)$. From Lemma \ref{lemma 4.1}, there is $\tilde{\gamma}_j(\theta)$ such that (\ref{eq4.4}) holds. Now, let $z=a_{j\cdots j}$ and $\tilde{z}=a_{j\cdots j}+\tilde{\gamma}_j(\theta)e^{\textbf{i}\theta}$, we have from Proposition \ref{lemma 4.2} that \[\tilde{\gamma}_j(\theta)\geq v(a_{j\cdots j})>0.\] Hence, $v(a_{j\cdots j}+v(a_{j\cdots j})e^{\textbf{i}\theta})\geq 0$. If $v(a_{j\cdots j}+v(a_{j\cdots j})e^{\textbf{i}\theta})=0$, then take \[\tilde{\gamma}_j(\theta)=v(a_{j\cdots j})\] for which $v(a_{j\cdots j})$ can be determined by the method of (i), otherwise, $v(a_{j\cdots j}+v(a_{j\cdots j})e^{\textbf{i}\theta})>0$, then we increase the number $v(a_{j\cdots j})$ to $v(a_{j\cdots j})+\Delta$, $\Delta>0$, until $v(a_{j\cdots j}+(v(a_{j\cdots j})+\Delta)e^{\textbf{i}\theta})<0$, and apply a bisection search to the interval $[v(a_{j\cdots j}), ~v(a_{j\cdots j})+\Delta]$ to determine $\tilde{\gamma}_j(\theta)$ satisfying (\ref{eq4.4}). Note here that estimates of $\tilde{\gamma}_j(\theta)$ also do not need great accuracy for graphing purpose. (iii) It is obvious that $w_{k_{j_\theta}}$ in Step 3, $k_{j_\theta}=1,2,\ldots, m$, are not only boundary points of $\Gamma^R(\mathcal {A})$, but boundary points of $\Gamma^{w_{k_{j_\theta}}}(\mathcal {A})$, and that \[\Gamma^R(\mathcal {A})\subseteq \Gamma^{w_{k_{j_\theta}}}(\mathcal {A})\] which shows that the larger $m$ is, the better $\Gamma^{w_{k_{j_\theta}}}(\mathcal {A})$ approximates to $\Gamma^R(\mathcal {A})$. \end{remark} \begin{example} \label{example4.1} Consider the irreducible tensor \[\mathcal {A} = [A(1,:,:),A(2,:,:),A(3,:,:)]\in \mathbb{C}^{[3,3]},\] where \[A(1,:,:)=\left(\begin{array}{cccc} 2 & 0 &0 \\ 0& 0 &1 \\ 0& 0& 1 \\ \end{array} \right),~ A(2,:,:)=\left(\begin{array}{cccc} 0 & 0 &0 \\ 0& 2 &0 \\ 1& 0& 0 \\ \end{array} \right), ~A(3,:,:)=\left(\begin{array}{cccc} 1 & 1 &0 \\ 0& 1 &0 \\ 0& 0& 1 \\ \end{array} \right).\] We next give a numerical approximation to $\Gamma^R(\mathcal {A})$. By the part (i) of Remark \ref{remark1}, we compute $v(a_{iii})$ for $i=1,2,3$, and get \[v(a_{111})=v(a_{222})=1.62019803,~v(a_{333})=1.43720383.\] Furthermore, based on the entries $a_{111}=a_{222}=2$ and $a_{333}=1$, we look for six points \[\begin{array}{ll} w_1=a_{111}+\tilde{\gamma}_1(0),& w_2=a_{333}-\tilde{\gamma}_3(\pi),\\ w_3=a_{111}+\textbf{i}~\tilde{\gamma}_1(\frac{\pi}{2}),&w_4=a_{111}-\textbf{i}~\tilde{\gamma}_1(\frac{3\pi}{2}),\\ w_5=a_{333}+\textbf{i}~\tilde{\gamma}_3(\frac{\pi}{2}),& w_6=a_{333}-\textbf{i}~\tilde{\gamma}_3(\frac{3\pi}{2}) \end{array}\] of $\partial\Gamma^R(\mathcal {A})$, which are found by the method proposed in the part (ii) of Remark \ref{remark1}, that is, \[\begin{array}{lll} w_{1}=3.62019802,& w_{2}=-0.43720383,&w_3=2+\textbf{i}1.86790935,\\ w_4=2-\textbf{i}1.86790935,& w_5=1+\textbf{i}1.81661895,&w_6=1-\textbf{i}1.81661895. \end{array}\] And now $\Gamma(\mathcal {A})$, $\Gamma^{w_{1}}(\mathcal {A})$, $\Gamma^{w_{2}}(\mathcal {A})$, $\Gamma^{w_{3}}(\mathcal {A})$ $(\Gamma^{w_{3}}(\mathcal {A})=\Gamma^{w_{4}}(\mathcal {A}))$ and $\Gamma^{w_{5}}(\mathcal {A})$ $(\Gamma^{w_{5}}(\mathcal {A})=\Gamma^{w_{6}}(\mathcal {A}))$ are given by Figures.1, 2, 3, 4 and 5, respectively. In Figure 6, the set $\left(\bigcap\limits_{k=1}^6 \Gamma^{w_{k}}(\mathcal {A})\right)$, which approximates to $\Gamma^R(\mathcal {A})$, is shown with the inner boundary, and the boundary of $\Gamma(\mathcal {A})$ is shown with the outside. The six points $\{ w_k\}_{k=1}^6$ are plotted with asterisks. As we can see, $\left(\bigcap\limits_{k=1}^6\Gamma^{w_{k}}(\mathcal{A})\right) \subset \Gamma(\mathcal {A})$, that is, the set which approximates to Minimal Ger\v{s}gorin tensor eigenvalue inclusion set is also contained in Ger\v{s}gorin tensor eigenvalue inclusion set. Also, it is easy to see that more points of $\partial\Gamma^R(\mathcal {A})$ are given, the better $\Gamma^{w_{k}}(\mathcal {A})$ approximates to $\Gamma^R(\mathcal {A})$. \end{example} \section{Conclusions} In this paper, we present Minimal Ger\v{s}gorin tensor eigenvalue inclusion set $\Gamma^{R}(\mathcal {A})$, give a sufficient and necessary condition for $\Gamma^{R}(\mathcal {A})$ by using the $Perron-Frobenius$ theory of nonnegative tensors, and establish the relationships between $\partial \Gamma^{R}(\mathcal {A})$, $\sigma (\Omega(\mathcal{A}))$, $\sigma (\hat{\Omega}(\mathcal{A}))$ and $\Gamma^{R}(\mathcal {A})$, i.e., \[\partial \Gamma^{R}(\mathcal {A})\subseteq \sigma (\Omega(\mathcal{A}))\subseteq \sigma (\hat{\Omega}(\mathcal{A}))= \Gamma^{R}(\mathcal {A}).\] These results obtained are generalizations of the corresponding results of matrices \cite {Varga} to higher order tensors. In \cite{Li-Li}, Li et al. provided two new eigenvalue inclusion sets which are contained in Ger\v{s}gorin eigenvalue inclusion set for tensors. An interesting problem arises: what's the relationship between Minimal Ger\v{s}gorin tensor eigenvalue inclusion set and the sets in \cite{Li-Li}? In the future, we will research this problem. \bigskip \noindent {\bf Acknowledgments.} This work is supported by National Natural Science Foundations of China (10961027, 71161020, 71162005) and IRTSTYN.

https://arxiv.org/abs/1405.4320

Parameter selection and numerical approximation properties of Fourier extensions from fixed data

Fourier extensions have been shown to be an effective means for the approximation of smooth, nonperiodic functions on bounded intervals given their values on an equispaced, or in general, scattered grid. Related to this method are two parameters. These are the extension parameter $T$ (the ratio of the size of the extended domain to the physical domain) and the oversampling ratio $\eta$ (the number of sampling nodes per Fourier mode). The purpose of this paper is to investigate how the choice of these parameters affects the accuracy and stability of the approximation. Our main contribution is to document the following interesting phenomenon: namely, if the desired condition number of the algorithm is fixed in advance, then the particular choice of such parameters makes little difference to the algorithm's accuracy. As a result, one is free to choose $T$ without concern that it is suboptimal. In particular, one may use the value $T=2$ - which corresponds to the case where the extended domain is precisely twice the size of the physical domain - for which there is known to be a fast algorithm for computing the approximation. In addition, we also determine the resolution power (points-per-wavelength) of the approximation to be equal to $T \eta$, and address the trade-off between resolution power and stability.

\section{Introduction}\label{s:introduction} In many problems, one is faced with the task of recovering a smooth function $f: [-1,1] \rightarrow \bbC$ to high accuracy from its pointwise samples on an equispaced, or in general, scattered grid. This problem is challenging, unless the grid points have a specific distribution, since it is difficult to simultaneously ensure both rapid convergence and numerical stability. In particular, for equispaced data a result of Trefethen, Platte \& Kuijlaars states that no stable method can converge faster than root-exponentially in the number of data points \cite{TrefPlatteIllCond}, and that any method with more rapid convergence must necessarily be unstable. Nevertheless, it has been widely reported that so-called \textit{Fourier extensions} (also known as \textit{Fourier continuations}) lead to effective methods in practice for reconstructions from equispaced or scattered data \cite{FEStability,BoydFourCont,BoydRunge,brunoFEP,DHFEP,LyonFESVD}. This was confirmed recently in \cite{FEStability} wherein it was shown that Fourier extensions (henceforth abbreviated to FEs) circumvent the stability barrier of \cite{TrefPlatteIllCond} in a certain sense. Specifically, they converge down to a finite, but user-controlled, maximal accuracy. In the method of FEs a function $f : [-1,1] \rightarrow \bbC$ is approximated by a Fourier series of degree $N$ defined on a larger domain $[-T,T]$, where $T>1$ is a user-controlled parameter. Suppose now that the number of equispaced data points is equal to $2M+1$ for some $M \geq N$. As shown in \cite{FEStability,LyonFESVD}, if $M = \eta N$ for some fixed \textit{oversampling} ratio $\eta \geq 1$, then one may compute an accurate and stable FE approximation of $f$ from this data by a simple least-squares fit (see also \S \ref{s:FE} for details). Moreover, when the extension parameter $T$ is equal to $2$ -- that is, the extended domain $[-T,T]$ is precisely twice the size of the physical domain $[-1,1]$ -- an algorithm developed by M.\ Lyon allows for the computation of the FE approximation in only $\ord{M (\log M)^2}$ operations \cite{LyonFast}. Note that this fast algorithm relies on the particular symmetries of FEs when $T=2$. From a practical standpoint, it is vitally important to understand how to choose the parameters $T$ and $\eta$. The purpose of this paper is to address this issue. In particular, we seek to determine how choices of these parameters affect both the stability and the accuracy of the FE approximation. Note that, due to the aforementioned fast algorithm, there is a seeming advantage to using the value $T=2$. However, is choice optimal vis-a-vis the other properties of the algorithm, namely, convergence and stability? Or does another choice (albeit lacking a fast algorithm) give better numerical performance in these respects? We shall provide answers to these questions. It is known that different choices of $T$ affect the intrinsic approximation properties of the FE approximation system, i.e.\ the space of trigonometric polynomials of degree $N$ on the extended domain $[-T,T]$ \cite{BADHFEResolution,FEStability,BoydFourCont}. For example, when $T$ is close to $1$, the FE approximation system possesses better \textit{resolution} power for oscillatory functions \cite{BADHFEResolution}. However, it is perhaps not surprising that such choices also require larger oversampling parameters $\eta$ to maintain the algorithm's stability. When $M$ is fixed, a larger $\eta$ means a smaller parameter $N$, and therefore the best approximation error in the above subspace, which is determined by the size of $N$, is correspondingly larger. From this argument, it is apparent that a balance must be struck between $T$ and $\eta$ so as to preserve accuracy and stability. The main result we obtain in this paper through numerical experiment is that these two effects precisely counteract each other. Specifically, if the desired condition number of the algorithm is fixed in advance, then, provided $T$ is not too large, the precise choice of $T$ makes no substantial difference to the accuracy of the Fourier extension algorithm. Smaller $T$ is exactly offset by the requirement of a larger value of $\eta$ to preserve stability. From this result, we are able to draw two main conclusions. First, any attempt to optimize $T$ will only bring limited, and most likely highly function-dependent, success. Second, since the choice of $T$ makes little difference, one may safely use $T=2$, and the resulting fast algorithm, without worrying that this choice may be suboptimal. We remark in passing that some previous insight into the effect of the parameters was given by Bruno et al.\ in \cite{brunoFEP}. However, this was largely carried out for specific functions. On the face of it, the conclusion we draw may appear surprising, or at the very least, a peculiar phenomenon isolated to the particular choice of equispaced data. After further numerical experiments, we conclude that this phenomenon is actually quite widespread. Specifically, we show exactly the same results for both scattered nonequispaced data, as well as Fourier data. Hence we conclude that unless the data is chosen specifically to favour a particular choice of $T$ (see \S \ref{s:other_data} for an example of such data), the value of $T$ makes little difference to the algorithm. We note here that, much as in the case of equispaced data, a fast Fourier extension algorithm for scattered data has also been developed in the case $T=2$ \cite{LyonFENonuniform}. In some applications, including the numerical solution of PDEs, an important question about an approximation algorithm is that of resolution power. Specifically, how many measurements (e.g.\ equispaced function samples) are required to recover an oscillation of frequency $\omega$. This topic was first investigated rigorously by Gottlieb \& Orszag \cite{naspec}, who popularized the concept of \textit{points-per-wavelength}. Through our experiments we establish that this quantity for FE approximations is given by the product of $T$ and $\eta$. We give theoretical arguments as to why this should be the case, and in the case $T=2$ (and therefore, by the above discussion, all values of $T$) provide numerical results assessing the tradeoff between resolution power and numerical stability. As the reader will have noticed, our aim in this paper is to investigate FEs through numerical experiment. The main conclusion we draw is based solely on the result of these experiments. Although we do present some mathematical insight as to why it should hold, this is a ways short of a proof. We leave this as a topic for future work. As we explain in \S \ref{s:conclusion}, this will likely require an intricate analysis of the singular values and singular vectors of a certain matrix related to Slepian's prolate matrix \cite{SlepianV,Varah}, which is beyond the scope of this paper. Nevertheless, we feel the conclusion we draw, albeit without a proof, is of substantial independent interest for anyone seeking to use FEs in practice. \section{Fourier extensions}\label{s:FE} Our concern in this paper is the approximation of functions defined on compact intervals, which without loss of generality we take to be $[-1,1]$. The method of Fourier extensions (FEs) is based on approximating such functions using a Fourier series defined on an extended interval $[-T,T]$, where $T>1$ is the so-called extension parameter. In other words, given $N \in \bbN$ we compute an approximation to $f$ from the subspace \bes{ \cG^{(T)}_N : = \spn \left \{ \phi_n : |n| \leq N \right \},\qquad \phi_n(x) = \E^{\I \frac{n \pi}{T} x }. } \subsection{Approximation properties of the subspace $\cG^{(T)}_N$} We now present several results concerning the intrinsic approximation properties of the subspace $\cG^{(T)}_{N}$. We use the notation $\rH^{k}(-1,1)$ for the standard Sobolev space on an interval $(-1,1)$, where $k \geq 0$. We denote the corresponding norm by $\nm{\cdot}_{\rH^k(-1,1)}$. \thm{[\cite{BADHFEResolution}] \label{t:error_alg} Let $T_0 > 1$ and suppose that $f \in \rH^{k}(-1,1)$ for some $k \geq 0$. Then, for each $N \in \bbN$ and $T \geq T_0$, there exists a $\phi \in \cG^{(T)}_{N}$ such that \bes{ \| f - \phi \|_{\rL^2(-1,1)} \leq C(k,T_0) \left ( \frac{N \pi}{T} \right )^{-k} \| f \|_{\rH^k(-1,1)},\qquad \| \phi \|_{\rL^2(-T,T)} \leq C(k,T_0) \| f \|_{\rH^k(-1,1)}, } for some constant $C(k,T_0)$ depending on $k$ and $T_0$ only. } This theorem asserts \textit{algebraic} convergence of the best approximations in $\cG^{(T)}_{N}$ when $f$ has $k$ derivatives, and \textit{superalgebraic} convergence whenever $f$ is smooth. Note that it also implies the existence of a function $\phi \in \cG^{(T)}_{N}$ which gives such convergence rates, and which cannot grow too large on the extended domain $[-T,T]$. This will be of significance in \S \ref{ss:CondErrFE}. Our next result confirms \textit{geometric} convergence of best approximations in $\cG^{(T)}_{N}$ in the case that $f$ is analytic. To state this result, we first recall the definition of a Bernstein ellipse: \bes{ \cB(\rho) = \left \{ \tfrac12 \left ( \rho^{-1} \E^{\I \theta} + \rho \E^{-\I \theta} \right ) : \theta \in [-\pi,\pi] \right \} \subseteq \bbC,\quad \rho > 1. } As discussed in \cite{FEStability,DHFEP}, Fourier extensions can be viewed as polynomial approximations in the mapped variable $z = m(x)$, where \be{ \label{mapping} m(x) = 2 \frac{\cos \frac{\pi}{T} x - \cos \frac{\pi}{T} }{1- \cos \frac{\pi}{T}} - 1. } Note that $m$ maps $[0,1]$ to $[-1,1]$ bijectively. Since the convergence of polynomial approximations of analytic functions is determined by Bernstein ellipses, it makes sense to introduce the new regions \bes{ \cD(\rho) = m^{-1} (\cB(\rho)),\quad \rho > 1, } We now have the the following theorem: \thm{[\cite{DHFEP}] \label{t:error_geo} Let $T>1$ be given and suppose that $f$ is analytic in $\cD(\rho')$ for some $\rho' > 1$ and continuous on its boundary. Then, for each $N \in \bbN$, there exists a $\phi \in \cG^{(T)}_{N}$ such that \be{ \label{geo_err} \| f - \phi \|_{\rL^{\infty}(-1,1)} \leq \frac{c_f(T)}{1-\rho} \rho^{-N}, } where $c_f(T) > 0$ is proportional to $\max_{z \in D(\rho) } | f(z) |$, \bes{ \rho = \min \left \{ \rho' , E(T) \right \}, } and $E(T) = \cot^2 \left ( \frac{\pi}{4 T} \right )$. Moreover, $\phi$ satisfies \be{ \label{geo_norm} \| \phi \|_{\rL^{\infty}(-T,T)} \leq c_f(T) \left ( E(T) / \rho \right )^N. } } This theorem establishes geometric convergence of best approximations in $\cG^{(T)}_{N}$. However, \R{geo_norm} suggests that in order to obtain such a convergence rate, one may have to allow for exponential growth of the corresponding $\phi$ in the extended domain $[-T,T]$ whenever $\rho < E(T)$. We shall return to this observation in \S \ref{ss:CondErrFE}. We remark also that Theorem \ref{t:error_geo} asserts that the maximal rate of geometric convergence is limited to $E(T)$, even if $f$ is entire. This is due to the mapping $m^{-1}$ which introduces a square-root type singularity and thereby limits the overall rate of convergence. See \cite{FEStability,DHFEP} for a discussion. \subsection{Fourier extensions from equispaced data}\label{ss:FE_equi} The concern of the majority of this paper is the approximation of a function $f$ from its values \bes{ f(m/M),\quad m=-M,\ldots,M, } on an equispaced grid of $2M+1$ points. For convenience, let us define the operator \bes{ S_M : \rL^{\infty}(-1,1) \rightarrow \bbC^{2M+1},\ f \mapsto \frac{1}{\sqrt{M}}(f(m/M))^{M}_{m=-M}. } We refer to $S_M$ as the \textit{sampling} operator. Given the vector $S_M (f)$ of samples of $f$, we construct its FE approximation in the standard way via a least-squares data fit \cite{FEStability,BoydFourCont,brunoFEP,DHFEP,LyonFESVD}. Let $N \leq M$ be given. Then we define the FE approximation as follows: \be{ \label{FE_LS_fn} F^{(T)}_{N,M}(f) : = \underset{\phi \in \cG^{(T)}_N}{\operatorname{argmin}} \sum_{|m| \leq M} \left | f(m/M) - \phi(m/M) \right |^2, } or more succinctly, \bes{ F^{(T)}_{N,M}(f) : =\underset{\phi \in \cG^{(T)}_N}{\operatorname{argmin}} \left | S_M(f-\phi) \right |, } where $|\cdot|$ denotes the usual Euclidean norm on $\bbC^{2M+1}$. Note that $F_{N,M}$ is an operator with domain $\rL^\infty(-1,1)$ and range $\cG^{(T)}_N$. Moreover, if we denote \bes{ F^{(T)}_{N,M}(f) = \sum_{|n| \leq N} a_n \E^{\I \frac{n \pi}{T} x}, } then the vector $\mathbf{a} = (a_n)^{N}_{n=-N}$ of FE coefficients is the solution of the least squares problem \be{ \label{FE_LS_coeff} \mathbf{a} = \underset{\mathbf{c} \in \bbC^{2N+1}}{\operatorname{argmin}} \left |A^{(T)} \mathbf{c} - S_M(f) \right |, } where $A^{(T)} \in \bbC^{(2M+1) \times (2N+1)}$ has entries \bes{ (A^{(T)})_{m,n} = \frac{1}{\sqrt{M}} \E^{\I \frac{n m \pi}{T}},\quad |m| \leq M,\ |n| \leq N. } Note that the normalization $1/\sqrt{M}$ in both $S_M$ and $A^{(T)}$ means that the entries of the normal matrix $(A^{(T)})^* A^{(T)}$ are Riemann sum approximations to the Gram matrix of the functions $\phi_n(x)$. This ensures that the singular values of $A^{(T)}$ lie between $0$ and $1$ for large $M$, which, since we typically consider truncated SVDs with a fixed truncation parameter (see later), ensures that there is no linear drift in the error for large $M$. For convenience, let us now introduce some additional notation. Let \bes{ L^{(T)}_{N,M} : \bbC^{2M+1} \rightarrow \bbC^{2N+1}, } be defined by \be{ \label{LNM_def} L^{(T)}_{N,M}(\mathbf{b}) = \underset{\mathbf{c} \in \bbC^{2N+1}}{\operatorname{argmin}} \left | A^{(T)} \mathbf{c} - \mathbf{b} \right |, } and let \bes{ R^{(T)}_{N} : \bbC^{2N+1} \rightarrow \cG^{(T)}_N,\ \mathbf{a} = (a_n)^{N}_{n=-N} \mapsto \sum_{|n| \leq N} a_n \E^{\I \frac{n \pi}{T} x}. } Note that $F^{(T)}_{N,M} = R^{(T)}_{N} \circ L^{(T)}_{N,M} \circ S_M$. As discussed in \cite{FEStability}, the algebraic least-squares problem to be solved in \R{LNM_def} is highly ill-conditioned. When applied to \R{LNM_def}, different numerical algorithms may consequently give somewhat different results. For this reason, it is important to specify the solver used. In the majority of this paper, as has been previously considered in \cite{FEStability,BoydFourCont,LyonFESVD}, we solve \R{LNM_def} by using truncated singular value decompositions (SVDs). If $U \Sigma V^*$ denotes the SVD of $A^{(T)}$, where $\Sigma$ is the diagonal matrix of singular values $\sigma_1 \geq \sigma_2 \geq \ldots$, then we correspondingly define \bes{ L^{(T,\epsilon)}_{N,M}(\mathbf{b}) = V \Sigma^{(\epsilon)} U^* \mathbf{b}, } where $\Sigma^{(\epsilon)}$ is the diagonal matrix with $n^{\rth}$ entry $1/\sigma_n$ if $\sigma_n > \epsilon$ and $0$ otherwise. Here $\epsilon > 0$ is the truncation parameter, which we take to be $10^{-13}$ unless specified otherwise. We denote the corresponding FE by $F^{(T,\epsilon)}_{N,M}(f)$. Having said this, we note that the quantities introduced below for studying equispaced FEs -- namely, the condition number and numerical defect constant -- are not specific to the SVD algorithm. In particular, one can compute such quantities for each different numerical solver and thereby directly compare the effectiveness of an equispaced FE resulting from an SVD with an equispaced FE computed using \textit{Matlab}'s $\backslash$ or \textit{Mathematica's} \texttt{LeastSquares} commands, for example. We return to this briefly in \S \ref{ss:other_solvers}. \rem{ Regardless of the solver used, it is important that the least-squares \R{LNM_def} is regularized when solved numerically. This is done by the parameter $\epsilon$ with the SVD approach, or automatically when using an blackbox least-squares solver such as \textit{Matlab}'s $\backslash$ or \textit{Mathematica's} \texttt{LeastSquares}. As shown in \cite{FEStability}, the `exact' FE mapping $f \mapsto F^{(T)}_{N,M}(f)$, i.e.\ that obtained by solving \R{LNM_def} in infinite precision, is ill-conditioned and suffers from a Runge phenomenon unless the number of equispaced points $M$ scales quadratically with $N$. Such severe scaling is undesirable, and is due solely to the behaviour of the Fourier series corresponding to singular vectors with small singular values. Fortunately, when the system \R{LNM_def} is regularized and $F^{(T,\epsilon)}_{N,M}$ is computed, this scaling drops to linear in $N$. Moreover, the Fourier series of the excluded singular values are precisely those which are small on the domain $[-1,1]$ but large on $[-T,T] \backslash [-1,1]$. Thus, their exclusion has little effect on the approximation of $f$. Note that a similar behaviour is also witnessed when different solvers are used for \R{LNM_def}, such as those listed above. } \subsection{Condition number and error bounds for equispaced FE approximations}\label{ss:CondErrFE} We now provide estimates for the accuracy and stability of $F^{(T)}_{N,M}(f)$. The key point is that these formulae involve constants which can be computed numerically. This will be discussed in the next section. First, however, we require the following assumption: \vspace{1pc} \noindent \textbf{Assumption.} The operator $L^{(T)}_{N,M}$ defined by solving \R{LNM_def} with a standard numerical solver (e.g.\ truncated SVDs) is approximately a linear operator. \vspace{1pc} Note that the exact, i.e.\ infinite precision, version of $L^{(T)}_{N,M}$ is of course a linear operator. Hence it is not unreasonable that its finite precision counterpart acts in the same way. Observe also that this assumption implies that the overall numerical FE operator $F^{(T)}_{N,M}$ is also linear. With this in hand, we can now define the condition number in the usual way: \defn{[Condition number] The (absolute) condition number of the equisapced FE approximation $F^{(T)}_{N,M}$ is given by \be{ \label{kappa_def} \kappa^{(T)}_{N,M} = \max_{\substack{\mathbf{b} \in \bbC^{2M+1} \\ \mathbf{b} \neq 0}} \left \{ \frac{\nmu{R^{(T)}_N \circ L^{(T)}_{N,M}(\mathbf{b})}_{\infty} }{|\mathbf{b}|_{\infty}} \right \}. } Here $\nm{g}_{\infty} = \sup_{x \in [-1,1]} |g(x)|$ is the uniform norm on $[-1,1]$ for $g \in \rL^\infty(-1,1)$ and $|\mathbf{b}|_{\infty} = \max_{|m| \leq M} | b_m|$ for $\mathbf{b} = (b_m)_{|m| \leq M} \in \bbC^{2M+1}$. } We remark that $\kappa^{(T)}_{N,M}$ is the absolute condition number, as opposed to the more standard relative condition number \cite{TrefethenBau}. It measures the absolute sensitivity of the FE to perturbations in the samples of $f$, and transpires to be substantially easier to compute in practice. Note also that the definition implicitly assumes linearity of the mapping $L^{(T)}_{N,M}$. We now consider the approximation error. For this we require the following definition: \defn{[Numerical defect constant] The numerical defect constant of the equispaced FE is given by \be{ \label{lambda_def} \lambda^{(T)}_{N,M} = \max_{\substack{\mathbf{a} \in \bbC^{2N+1} \\ \mathbf{a} \neq 0}} \left \{ \frac{\nm{R^{(T)}_N \left ( \mathbf{a} - L^{(T)}_{N,M} \circ S_M \circ R^{(T)}_N (\mathbf{a}) \right )}_{\infty}}{| \mathbf{a} |_{\infty}} \right \}. } } Before showing the relevance of this constant to error bounds, let us first consider its meaning. Recall that each vector $\mathbf{a}$ corresponds uniquely to a function $\phi \in \cG^{(T)}_{N}$ given by $\phi = R^{(T)}_{N} \mathbf{a}$. Thus the numerator in \R{lambda_def} reads \bes{ \nm{\phi - F^{(T)}_{N,M}(\phi)}_{\infty}. } In infinite precision, the FE operator $F^{(T)}_{N,M}$ satisfies $F^{(T)}_{N,M}(\phi) = \phi$ for $\phi \in \cG^{(T)}_{N}$. In other words, it is a projection. Hence the numerical defect constant measures how close the numerical, i.e.\ finite precision, FE operator is to possessing this property. We are now able to provide an error bound for $F^{(T)}_{N,M}(f)$: \lem{ \label{l:error_numerical} Let $f \in \rL^\infty(-1,1)$ and suppose that $F^{(T)}_{N,M}(f)$ is given by \R{FE_LS_fn}. Then \be{ \label{error_numerical} \| f - F^{(T)}_{N,M}(f) \|_{\infty} \leq \inf_{\mathbf{a} \in \bbC^{2N+1}} \left \{ \left ( 1 + \kappa^{(T)}_{N,M} \right )\| f - R^{(T)}_N (\mathbf{a}) \|_{\infty} + \lambda^{(T)}_{N,M} | \mathbf{a} |_{\infty} \right \}, } where $\lambda^{(T)}_{N,M}$ is as in \R{lambda_def}. } \prf{ Let $\mathbf{a} \in \bbC^{2N+1}$ be arbitrary and write $\phi = R^{(T)}_{N}(\mathbf{a})\in \cG^{(T)}_N$. Then, using linearity of $F^{(T)}_{N,M}$ (Assumption 1), we obtain \bes{ \| f - F^{(T)}_{N,M}(f) \|_{\infty} \leq \| f - \phi \|_{\infty}+ \| F^{(T)}_{N,M}(f-\phi) \|_{\infty} + \| \phi - F^{(T)}_{N,M}(\phi) \|_{\infty} . } We consider the latter two terms separately. For the first, note that \eas{ \| F^{(T)}_{N,M}(f-\phi) \|_{\infty} = \| R^{(T)}_N \circ L^{(T)}_{N,M} \circ S_M ( f - \phi) \|_{\infty} \leq \kappa^{(T)}_{N,M} | S_M(f-\phi) |_{\infty} \leq \kappa^{(T)}_{N,M} \| f - \phi \|_{\infty}. } This gives the corresponding second term in \R{error_numerical}. We now consider the other term. We have \bes{ \nm{\phi - F^{(T)}_{N,M}(\phi)}_{\infty} = \nm{R^{(T)}_N \left ( \mathbf{a} - L^{(T)}_{N,M} \circ S_M \circ R^{(T)}_M (\mathbf{a}) \right )}_{\infty} \leq \lambda^{(T)}_{N,M} | \mathbf{a} |_{\infty}, } as required. } Let us now interpret this error bound. In \S \ref{ss:computing} we shall observe numerically that \be{ \label{kappa_lambda_growth} \kappa^{(T)}_{M/\eta,M} \sim \tilde{\kappa}_{T,\eta} \log M,\qquad \lambda^{(T)}_{M/\eta,M} \sim \tilde{\lambda}_{T,\eta} M,\qquad M \rightarrow \infty, } where $\tilde{\kappa}_{T,\eta}$ and $\tilde{\lambda}_{T,\eta}$ are independent of $M$. Furthermore, the ratio $\mu = \tilde{\lambda}_{T,\eta}/\tilde{\kappa}_{T,\eta}$ is roughly $10^{-13}$ in magnitude, regardless of the choice of $T$ or $\eta$. Hence, one has the estimate \be{ \label{error_numerical_2} \| f - F^{(T)}_{N,M}(f) \|_{\infty} \leq \tilde{\kappa}_{\eta,T} M \inf_{\mathbf{a} \in \bbC^{2N+1}} \left \{ \| f - R^{(T)}_N (\mathbf{a}) \|_{\infty} + \mu | \mathbf{a} |_{\infty} \right \},\qquad N = M/\eta. } The key aspect of this bound is that it separates the error into two component. The first, namely, $\tilde{\kappa}_{\eta,T}$, is determined by the parameters $\eta = M/N$ and $T$, and is independent of the function $f$. Moreover, as we see next, it can be computed numerically. The second, i.e.\ the term \be{ \label{EN_def} E_{N}(f) : = \inf_{\mathbf{a} \in \bbC^{2N+1}} \left \{ \| f - R^{(T)}_N (\mathbf{a}) \|_{\infty} + \mu | \mathbf{a} |_{\infty} \right \}, } is crucially independent of $\eta$ and depends only on the intrinsic approximation properties of the subspace $\cG^{(T)}_N$ and the smoothness of $f$. In particular, combining Lemma \ref{l:error_numerical} with Theorems \ref{t:error_alg} and \ref{t:error_geo}, we immediately obtain the following: \cor{ \label{c:EN_rate} Let $E_N(f)$ be given by \R{EN_def}. If $f \in \rH^k(-1,1)$ then \be{ \label{alg_conv} E_N(f) \leq \min_{0 \leq l \leq k} \left \{ C(l,T_0) \| f \|_{\rH^l(-1,1)} \left ( \left ( \frac{N \pi}{T} \right )^{-l} + \mu \right ) \right \}, } for each $T \geq T_0$, where $T_0$ and $C(l,T_0)$ are as in Theorem \ref{t:error_alg}. Moreover if $f$ is analytic in $\cD(\rho)$ and continuous on its boundary, then one has \be{ \label{exp_conv} E_N(f) \leq c_f(T) \rho^{-N} \left (1 + \mu E(T)^N \right ) , } where $c_f(T)$ is as in Theorem \ref{t:error_geo}. } This corollary explains the behaviour of $E_N(f)$ in both finite and infinite precision. In infinite precision, where $\mu = 0$, the bound \R{exp_conv} shows geometric decay of $E_N(f)$ for all $N$ at a rate equal to $\rho$. In finite precision, however, the small, but nonzero constant $\mu$ dramatically alters the convergence. Geometric decay still occurs for small $N$, when the term $\mu E(T)^N$ in \R{exp_conv} is small, but once $N \geq N_0 = - \log \mu / \log E(T)$ the right-hand side of \R{exp_conv} begins to increase. For $N \geq N_0$, $E_N(f)$ no longer decays geometrically. Instead, its decay is described by the bound \R{alg_conv}. Specifically, algebraic decay in $N$ occurs down to a maximal achievable accuracy on the order of $\mu$. Note that the constant term $C(l,T) \| f \|_{\rH^l(-1,1)}$ usually grows with $l$, thus as $E_N(f)$ approaches $\mu$ the effective rate of decay usually lessens. This behaviour is illustrated in Figure \ref{f:EN_Plot}. As we see, geometric convergence in infinite precision requires geometric growth of the coefficient vector $\mathbf{a}$. Conversely, in finite precision, such convergence is sacrificed for algebraic convergence whilst maintaining a bounded coefficient norm. We refer to \cite{FEStability} for a more detailed discussion. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=6.00cm]{EN_Plot} & \hspace{2pc} &\includegraphics[width=6.00cm]{EN_Coeff} \end{array}$ \caption{Best approximation error (left) and coefficient norm (right) for the function $f(x) = \frac{1}{40-39 x}$ in finite precision (thicker line) and infinite precision (thinner line). The latter was computed in \textit{Mathematica} using additional precision.} \label{f:EN_Plot} \end{center} \end{figure} \subsection{Computing the condition number and numerical defect constant}\label{ss:computing} Whilst Corollary \ref{c:EN_rate} explains the decay of $E_N(f)$, in order to understand the error of the equispaced FE $F^{(T)}_{N,M}(f)$ we need to determine the magnitudes of $\kappa^{(T)}_{N,M}$ and $\lambda^{(T)}_{N,M}$. The former also determines the stability of $F^{(T)}_{N,M}$. As we discuss further in \S \ref{s:conclusion}, it is as of yet unknown how to do this analytically, hence we now resort to numerical investigations. For this we need a means of computing $\kappa^{(T)}_{N,M}$ and $\lambda^{(T)}_{N,M}$. This follows from the next two lemmas: \lem{ \label{l:kappa_equals} Let $\mathbf{e}_m$, $|m| \leq M$, be the canonical basis for $\bbC^{2M+1}$. Then \bes{ \kappa^{(T)}_{N,M} = \sup_{x \in [-1,1]} \sum_{|m| \leq M} \left | R^{(T)}_N \circ L^{(T)}_{N,M}(\mathbf{e}_m)(x) \right |. } } \prf{ If $\mathbf{b} = (b_m)_{|m| \leq M} \in \bbC^{2M+1}$, we may write $\mathbf{b} = \sum_{|m| \leq M} b_m \mathbf{e}_m$. By linearity \bes{ R^{(T)}_{N} \circ L^{(T)}_{N,M}(\mathbf{b}) = \sum_{|m| \leq M} b_m R^{(T)}_{N} \circ L^{(T)}_{N,M}(\mathbf{e}_m), } and therefore \bes{ \kappa^{(T)}_{N,M} \leq \sup_{x \in [-1,1]} \sum_{|m| \leq M} | R^{(T)}_{N} \circ L^{(T)}_{N,M}(\mathbf{e}_m) | . } Conversely, \bes{ \kappa^{(T)}_{N,M} = \sup_{x \in [-1,1]} \max_{\substack{\mathbf{b} \in \bbC^{2M+1} \\ \mathbf{b} \neq 0}} \frac{\left | \sum_{|m| \leq M} b_m R^{(T)}_{N} \circ L^{(T)}_{N,M}(\mathbf{e}_m)(x) \right |}{| \mathbf{b} |_{\infty} }. } We now set $b_m$ equal to the complex sign of $R^{(T)}_N \circ L^{(T)}_{N,M}(\mathbf{e}_m)(x)$ to deduce the lower bound. } This lemma allows for approximate computation of $\kappa^{(T)}_{N,M}$. Let $K \in \bbN$ be given and define \bes{ x_k = \frac{T(k-1)}{K}-1,\quad k=1,\ldots,K_T, } where \bes{ K_T = \left \lfloor \frac{2 K}{T} +1 \right \rfloor . } Note that $\{ x_k \}^{K_T}_{k=1}$ is a set of $K_T$ equispaced nodes in $[-1,1]$. Therefore \bes{ \kappa^{(T)}_{N,M} = \lim_{K \rightarrow \infty} \kappa^{(T)}_{N,M,K}, } where \bes{ \kappa^{(T)}_{N,M,K} = \max_{k=1,\ldots,K_T} \sum_{|m| \leq M} \left | R^{(T)}_N \circ L^{(T)}_{N,M}(\mathbf{e}_m)(x_k) \right |, } is a computable quantity. We remark also $\kappa^{(T)}_{N,M,K}$ can be computed efficiently using Fast Fourier Transforms (FFTs), since the functions $R^{(T)}_N \circ L^{(T)}_{N,M}(\mathbf{e}_m)(x)$ are Fourier series and $\{ x_k \}^{K_T}_{k=1}$ are appropriately constructed equispaced nodes. Throughout this paper we shall consistently use the value $K=2^{15}$ in our numerical experiments. We use a similar approach in order to compute the numerical defect constant. Analogously to Lemma \ref{l:kappa_equals}, we have the following: \lem{ Let $\mathbf{e}_n$, $|n| \leq N$, be the canonical basis for $\bbC^{2N+1}$. Then \bes{ \lambda^{(T)}_{N,M} = \sup_{x \in [-1,1]} \sum_{|n| \leq N} \left | W^{(T)}_{N,M}(\mathbf{e}_n)(x) \right |, } where $W^{(T)}_{N,M} = R^{(T)}_{N} - R^{(T)}_{N} \circ L^{(T)}_{N,M} \circ S_M \circ R^{(T)}_{N}$. } Much as before, we may now write \bes{ \lambda^{(T)}_{N,M} = \lim_{K \rightarrow \infty} \lambda^{(T)}_{N,M,K},\qquad \lambda^{(T)}_{N,M,K} = \max_{k=1,\ldots,K_T} \sum_{|n| \leq N} \left | W^{(T)}_{N,M}(\mathbf{e}_n)(x_k) \right |. } where the latter can once more be computed efficiently using FFTs. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{T2_Kappa_Plot} & \includegraphics[width=5.00cm]{T2_Lambda_Plot} &\includegraphics[width=5.00cm]{T2_Mu_Plot} \\ \includegraphics[width=5.00cm]{T2_Kappa_Plot_Scaled} & \includegraphics[width=5.00cm]{T2_Lambda_Plot_Scaled} & \includegraphics[width=5.00cm]{T2_Mu_Plot_Scaled} \\ \end{array}$ \caption{Top row: the quantities $\kappa^{(T,\epsilon)}_{M/\eta ,M}$ (left), $\lambda^{(T,\epsilon)}_{M/\eta ,M}$ (middle) and $\mu^{(T,\epsilon)}_{M/\eta , M} : = \lambda^{(T)}_{M/\eta,M} / \kappa^{(T)}_{M/\eta ,M}$ (right) against $M$ for $\eta = 1,1.125,1.25,1.5,2,2.5,3,4,5$ (thickest to thinnest), $\epsilon = 10^{-13}$ and $T=2$. Bottom row: the same quantities scaled by $\log M$, $M$ and $M / \log M$ respectively.} \label{f:T_2_kappa_lambda} \end{center} \end{figure} \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{T125_Kappa_Plot} & \includegraphics[width=5.00cm]{T125_Lambda_Plot} & \includegraphics[width=5.00cm]{T125_Mu_Plot} \\ \includegraphics[width=5.00cm]{T125_Kappa_Plot_Scaled} & \includegraphics[width=5.00cm]{T125_Lambda_Plot_Scaled} & \includegraphics[width=5.00cm]{T125_Mu_Plot_Scaled} \\ \end{array}$ \caption{Top row: the quantities $\kappa^{(T,\epsilon)}_{M/\eta ,M}$ (left), $\lambda^{(T,\epsilon)}_{M/\eta ,M}$ (middle) and $ \lambda^{(T)}_{M/\eta,M} / \kappa^{(T)}_{M/\eta ,M}$ (right) against $M$ for $\eta = 1,1.125,1.25,1.5,2,2.5,3,4,5$ (thickest to thinnest), $\epsilon = 10^{-13}$ and $T=1.25$. Bottom row: the same quantities scaled by $\log M$, $M$ and $M / \log M$ respectively.} \label{f:T_125_kappa_lambda} \end{center} \end{figure} Having demonstrated how to compute $\kappa^{(T)}_{N,M}$ and $\lambda^{(T)}_{N,M}$, in Figures \ref{f:T_2_kappa_lambda} and \ref{f:T_125_kappa_lambda} we present the result of such computations for the case where the FE is computed using an SVD with tolerance $\epsilon = 10^{-13}$. The results confirm the scaling \R{kappa_lambda_growth} for these quantities. Moreover, these results also show that the quantity \be{ \label{mu_behaviour} \mu^{(T)}_{M/\eta ,M} = \left (\frac{\lambda^{(T)}_{M/\eta,M}}{\kappa^{(T)}_{M/\eta ,M}} \right ) \left ( \frac{\log M}{M} \right ) \leq \mu,\quad \forall M \gg 1, } where $\mu$ is roughly $10^{-13}$ in magnitude, regardless of the choice of $T$ and $\eta$. Note that in both cases a larger oversampling ratio $\eta$ leads to a smaller condition number $\kappa^{(T,\epsilon)}_{M/\eta,M}$ and numerical defect constant $\lambda^{(T,\epsilon)}_{M/\eta,M}$. Moreover, the larger value of $T$, in this case, $T=2$, has a smaller condition number for the same oversampling value than the smaller value $T=1.25$. The purpose of the next section is to investigate the exact nature of these relative scalings. \rem{ Several previous papers have investigated quantities similar to $\kappa^{(T)}_{N,M}$ and $\lambda^{(T)}_{N,M}$. In \cite{FEStability} and \cite{LyonFESVD}, quantities based on the exact singular values and vectors of the matrix $A^{(T)}$ were investigated using high-precision numerical computations. The approach we take above differs from these studies in two aspects. First, $\kappa^{(T)}_{N,M}$ and $\lambda^{(T)}_{N,M}$ can be formulated for any numerical solver, not just truncated SVDs. Second, when truncated SVDs are used, they incorporate the numerical errors in the calculation of the singular values and vectors. Since the matrix $A^{(T)}$ is ill-conditioned, these errors cannot be assumed to be insignificant. } \section{Numerical investigation} We now suppose that the FE is computed using an SVD as described in \S \ref{ss:FE_equi}. Our aim is to examine the behaviour of $\kappa^{(T,\epsilon)}_{N,M}$, and later $\lambda^{(T,\epsilon)}_{N,M}$, with respect to $T$ and the \textit{oversampling} ratio $\eta = M/N$, and how this affects the accuracy of the corresponding FE approximation. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=8.00cm]{ContourPlot_SVD_1e13} \end{array}$ \caption{\small Contour plot of $\kappa^{(T,\epsilon)}_{N,\eta N}$ against $1 < T \leq 6$ (horizontal axis) and $1 \leq \eta \leq 6$ (vertical axis) for $N=250$ and $\epsilon = 10^{-13}$.} \label{f:ContourPlot} \end{center} \end{figure} In Figure \ref{f:ContourPlot} we give a contour plot of $\kappa^{(T,\epsilon)}_{N,\eta N}$ as function of $T$ and $\eta$. As is evident, increasing either $\eta$ or $T$ leads to a smaller condition number. This suggests that in practice, a balance must be struck between $\eta$ and $T$ so as to get a good condition number whilst retaining a good approximation properties (recall that larger values of $T$ possess worse resolution power \cite{FEStability}, whereas larger values of $\eta$ yield worse approximations for a fixed budget of $M$ equispaced data points). Before investigating this interplay further, let us briefly explain why the condition number behaves in this way. Clearly, increasing $\eta$ results in a smaller value of $N = M/\eta$, and therefore an approximation space $\cG^{(T)}_{N}$ of smaller dimension. The condition number, the maximum taken over this space, therefore decreases. Conversely, when $T$ is decreased, this means that the Fourier basis is defined over a smaller domain. But the Fourier series $\phi = R^{(T)}_{N} \circ L^{(T)}_{N,M}(\mathbf{b})$ must fit the nonperiodic data $\mathbf{b}$ in a least-squares sense on the original domain $[-1,1]$ and must be periodic on the extended domain. To do this, $\phi$ will be required to take increasingly larger values between data points as $T \rightarrow 1^+$, giving it a bigger uniform norm in comparison to the data norm $| \mathbf{b} |_{\infty}$. \subsection{Setup} In order to make a comparison, for each different value of $T$ used we shall choose the ratio $\eta = M/N$ in such a way that the condition number is the same. Specifically, let $\kappa^* >1$ be fixed. Then for each $T$ and each $M$ in some specified range, we numerically compute the maximum $N$ such that the condition number is no more than $\kappa^* \log M$. In other words, we compute the function \be{ \label{Theta_def} \Theta^{(T)}(M;\kappa^*) = \max \left \{ N : \kappa^{(T)}_{N,M} \leq \kappa^* \log M \right \},\quad M \in \bbN. } Note that we allow $\log M$ factor here since $\kappa^{(T)}_{M/\eta,M}$ grows like $\log M$ as $M \rightarrow \infty$. With this scaling, $\Theta^{(T)}(M;\kappa^*)$ will be linear in $M$ (see later). The function \R{Theta_def} can be computed numerically for each $M$. This follows from the fact that $\kappa^{(T)}_{N,M}$ can be computed (see \S \ref{ss:computing}). Note also that $\Theta^{(T)}(M;\kappa^*)$ can be computed for any particular numerical solver used to solve the least squares \R{LNM_def}, and thus allows a comparison between different methods. We consider this further in \S \ref{ss:other_solvers}. Having computed \R{Theta_def} for each value of $T = T_1,\ldots,T_r$ and some range of $M$, we next use these values to compare approximation properties of the corresponding equispaced FEs \bes{ F^{(T_j)}_{N,M}(\cdot),\quad \mbox{where $N = \Theta^{(T_j)}(M;\kappa^*)$},\qquad j=1,\ldots,r. } When doing this, we shall consider a suite of different test functions, described further below. Observe that \R{Theta_def} determines the largest value of $N$ for which the condition number is at most $\kappa^* \log M$. Thus, setting $N = \Theta^{(T)}(M;\kappa^*)$ when computing the FE $F^{(T)}_{N,M}(f)$ ensures the best approximation properties for each value of $T$ (since $N$ is maximal) whilst retaining the same condition number for the different choices $T=T_1,\ldots,T_r$. Thus a comparison between these different values of $T$ can be made, using the condition number as the common fixed point. We now require appropriate test functions. Our first three functions are as follows: \eas{ f_1(x) = \E^{230 \sqrt{2} \I \pi x}\qquad f_2(x) = \sin(400 x^2)\qquad f_3(x) = \mathrm{Ai}(-66-70 x) } These functions all exhibit oscillations, which make them challenging to approximate from equispaced data. Plots of $f_2$ and $f_3$ are given in Figure \ref{f:functions}. Our next collection of test functions feature singularities in the complex plane near $[-1,1]$, again making them difficult to approximate: \eas{ f_4(x) = \frac{1}{1+1500 x^2}\qquad f_5(x) = \frac{1}{60-59 x}\qquad f_6(x) = \frac{1}{1+25 \sin^2 8 x}. } A plot of $f_6$ is shown in Figure \ref{f:functions}. Our final collection of functions, also displayed in Figure \ref{f:functions}, is \bes{ \quad f_7(x) = \E^{\sin (21.6 \pi x - 10.8 \pi )-\cos 8 \pi x}\qquad f_8(x) = \E^{-1/(8 x)^2}\qquad f_9(x) = s(x). } Note that $f_7$ is often used in testing algorithms for recovering functions to high accuracy from equispaced data, and $f_8$ is made challenging by its lack of analyticity and the flat region near $x=0$. The function $f_9$ is similar to that introduced in \cite{TrefethenFunctions}. It is obtained by the following iteration: \bull{ \item $s(x) = \sin \pi x$, $f(x) = s(x)$ \item For $j=1,2,\ldots,10$, $s(x) = 3/4 (1-2 s(x)^4)$, $f(x) = f(x) + s(x)$. } \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{f2_plot} & \includegraphics[width=5.20cm]{f3_plot} & \includegraphics[width=5.20cm]{f6_plot} \\ \includegraphics[width=5.00cm]{f7_plot} & \includegraphics[width=5.20cm]{f8_plot} & \includegraphics[width=5.20cm]{f9_plot} \end{array}$ \caption{Top row: the functions $f_2$, $f_3$ and $f_6$. Bottom row: the functions $f_7$, $f_8$ and $f_9$.} \label{f:functions} \end{center} \end{figure} \subsection{Numerical results} \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{Theta_SVD_1e13_kappa10_logscaled_2} & \includegraphics[width=5.00cm]{Theta_SVD_1e13_kappa25_logscaled_2} & \includegraphics[width=5.00cm]{Theta_SVD_1e13_kappa100_logscaled_2} \\ \includegraphics[width=5.00cm]{Theta_SVD_1e13_kappa10_logscaled_2_scaled} & \includegraphics[width=5.00cm]{Theta_SVD_1e13_kappa25_logscaled_2_scaled} & \includegraphics[width=5.00cm]{Theta_SVD_1e13_kappa100_logscaled_2_scaled} \end{array}$ \caption{\small Plots of $\Theta^{(T,\epsilon)}(M;\kappa^*)$ (top) and $\Theta^{(T,\epsilon)}(M;\kappa^*)/M$ (bottom) against $M$ for $\kappa^* = 10$ (left), $\kappa^* =25$ (middle) and $\kappa^* = 100$ (right) using $\epsilon = 10^{-13}$. The values of $T$ used (in order of increasing thickness) were $T=1.125,1.25,1.5,2,2.5,3.0,4.0,5.0,6.0$. Note that in the middle plots the $T=5.0$ line is identical to the $T=6.0$ line, and for the right plots the $T=4.0$, $T=5.0$ and $T=6.0$ lines are identical.} \label{f:ThetaPlot} \end{center} \end{figure} In Figure \ref{f:ThetaPlot} we plot the function $\Theta^{(T)}(M;\kappa^*)$ against $M$ for various values of $T$ and $\kappa^*$. Note that this function is approximately linear in $M$. Moreover, its gradient is larger for bigger values of $T$, as expected from the results given in Figure \ref{f:ContourPlot}. The approximate linear rate of growth of $\Theta^{(T)}(M;\kappa^*)$ is shown in Table \ref{tab:ThetaGrowth}. \begin{table} \begin{center} \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|} \hline $T$ & $1.125$ & $1.25$ & $1.5$ & $2.0$ & $2.5$ & $3.0$ & $4.0$ & $5.0$ & $6.0$ \\ \hline $\kappa^* = 10$ & 0.21 & 0.23 & 0.28 & 0.37 & 0.46 & 0.55 & 0.73 & 0.91 & 1.00 \\ \hline $\kappa^* = 25$ & 0.25 & 0.28 & 0.33 & 0.45 & 0.55 & 0.66 & 0.89 & 1.00 & 1.00 \\ \hline $\kappa^* = 100$ & 0.31 & 0.35 & 0.42 & 0.55 & 0.69 & 0.82 & 1.00 & 1.00 & 1.00 \\ \hline \end{tabular} \caption{The approximate linear scaling of $\Theta^{(T)}(M;\kappa^*)$ with $M$. Values were computed using linear regression on the data obtained in Figure \ref{f:ThetaPlot}.}\label{tab:ThetaGrowth} \end{center} \end{table} Next, in Figure \ref{f:FnApp} we consider the approximation of the functions $f_1,\ldots,f_9$ using the derived values for $\Theta^{(T)}(M;\kappa^*)$. These results point towards a surprising phenomenon. Besides the choices $T=5.0$ and $T=6.0$, all values of $T$ used lead to near-identical approximation errors, regardless of the function considered. Thus, seemingly the value of $T$, unless taken to be either $5.0$ or $6.0$ in this case, makes little or no difference to the FE approximation. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{f1app_SVD_1e13_kappa25_logscaled} & \includegraphics[width=5.00cm]{f2app_SVD_1e13_kappa25_logscaled} & \includegraphics[width=5.00cm]{f3app_SVD_1e13_kappa25_logscaled} \\ \includegraphics[width=5.00cm]{f4app_SVD_1e13_kappa25_logscaled} & \includegraphics[width=5.00cm]{f5app_SVD_1e13_kappa25_logscaled} & \includegraphics[width=5.00cm]{f6app_SVD_1e13_kappa25_logscaled} \\ \includegraphics[width=5.00cm]{f7app_SVD_1e13_kappa25_logscaled} & \includegraphics[width=5.00cm]{f8app_SVD_1e13_kappa25_logscaled} & \includegraphics[width=5.00cm]{f9app_SVD_1e13_kappa25_logscaled} \\ \end{array}$ \caption{\small Approximation errors for the functions $f_i$, $i=1,\ldots,9$ using the values $\Theta^{(T,\epsilon)}(M;\kappa^*)$ computed in Figure \ref{f:ThetaPlot} for $\kappa^* = 25$ and $\epsilon = 10^{-13}$.} \label{f:FnApp} \end{center} \end{figure} Let us explain why the two values $T=5.0$ and $T=6.0$ lead to worse approximations. This is due to a phenomenon we refer to as \textit{saturation}: \defn{ For a given $T >1$ and $\kappa^* > 1$, saturation occurs if \bes{ \limsup_{M \rightarrow \infty} \kappa^{(T)}_{M,M} / \log M < \kappa^* . } } Saturation means that the corresponding FE is too stable to take advantage of the allowed condition number $\kappa^*$. In other words, the maximal value of $N$ permitted is limited by the fact that $N \leq M$ in the FE approximation, and not by the condition number constraint $\kappa^{(T)}_{N,M} \leq \kappa^*$. When saturation occurs, the resulting FE approximation performs worse in terms of approximation than that corresponding to a value of $T$ for which saturation does not occur. Figure \ref{f:ThetaPlot} illustrates that the two values of $T$ which give worse approximations in Figure \ref{f:FnApp}, i.e.\ $T=5.0$ and $T=6.0$, do indeed saturate. This is further demonstrated in Figure \ref{f:FnApp2}, where we consider the approximation of the function $f_1$ using different values of $\kappa^*$. Figure \ref{f:ThetaPlot} shows that for $\kappa^* =10$ only $T=6.0$ saturates, whereas $T=5.0$ and $T=6.0$ both saturate for $\kappa^* = 25$, and for $\kappa^* = 100$ the values $T=4.0$, $T=5.0$ and $T=6.0$ all saturate. In Figure \ref{f:FnApp2} the FE approximations with these values of $T$ perform worse than the FEs corresponding to values of $T$ which do not saturate. Note that when $\kappa^* = 10$ the effect of the saturation for $T=6.0$ has less impact, since $\kappa^{(6,\epsilon)}_{M,M} /\log M \approx 5$ is reasonably close to $\kappa^*$ in this case. Hence saturation occurs, but to a lesser extent. Similarly, since $\kappa^{(6)}_{M,M} < \kappa^{(5)}_{M,M} < \kappa^{(4)}_{M,M}$ the effect of the saturation on the FE approximation when $\kappa^* = 100$ is less for $T=4.0$ than it is for $T=5.0$ and $T=6.0$. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{f1app_SVD_1e13_kappa10_logscaled} & \includegraphics[width=5.00cm]{f1app_SVD_1e13_kappa25_logscaled} & \includegraphics[width=5.20cm]{f1app_SVD_1e13_kappa100_logscaled} \end{array}$ \caption{\small Approximation errors for the function $f_1$ using the values $\Theta^{(T)}(M;\kappa^*)$ computed in Figure \ref{f:ThetaPlot} for $\kappa^* = 10$ (left), $\kappa^* = 25$ (middle) and $\kappa^* = 100$ (right).} \label{f:FnApp2} \end{center} \end{figure} With this in hand, we are now able to state the main empirical conclusion of this section: \textit{unless saturation occurs, the choice of $T$ makes little difference to the FE approximation}. In particular, one may use the value $T=2$ provided it does not saturate for given value of $\kappa^*$. We note in passing that this phenomenon is of course asymptotic in $M$, and relies on the fact that the functions under consideration require large numbers of equispaced points to be approximated to any accuracy. For smooth function lacking unpleasant features such as close singularities or oscillations, i.e.\ functions that can be resolved with small $M$, there will be slight discrepancies for different values of $T$. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{eta_kappa_10} & \includegraphics[width=5.00cm]{eta_kappa_25}& \includegraphics[width=5.00cm]{eta_kappa_100} \\ \kappa^* = 10,\ \tau(\kappa^*) \approx 0.18 & \kappa^* = 25,\ \tau(\kappa^*) \approx 0.22 & \kappa^* = 100,\ \tau(\kappa^*) \approx 0.27 \end{array}$ \caption{\small The approximate values $\nu^{(T)}(\kappa^*)$ from the Table \ref{tab:ThetaGrowth} plotted against $T$. The solid curve is the line with slope $\tau(\kappa^*)$, where $\tau(\kappa^*)$ is computed using linear regression on the data $\{ T , \nu^{(T)}(\kappa^*) \}$ corresponding to those nonsaturating values of $T$. } \label{f:eta_nu} \end{center} \end{figure} Let us give some explanation for why this conclusion should hold. Figure \ref{f:ThetaPlot} shows that the function $\Theta^{(T)}(M;\kappa^*)$ is approximately linear in $M$, i.e. $\Theta^{(T)}(M;\kappa^*) \approx \nu^{(T)}(\kappa^*) M$ for large $M$ for some constant $\nu^{(T)}(\kappa^*) >0$. Explicit values of the quantity $\nu^{(T)}(\kappa^*)$ are given in Table \ref{tab:ThetaGrowth}, and these are plotted against $T$ in Figure \ref{f:eta_nu}. As is evident from this figure, for each each fixed $\kappa^*$, the quantity $\nu^{(T)}(\kappa^*)$ is approximately linear in $T$ up to the point at which saturation occurs. Write $\nu^{(T)}(\kappa^*) \approx \tau(\kappa^*) T$ for some $\tau(\kappa^*) > 0$ and now consider the FE approximation $F^{(T)}_{N,M}$ where $N = \tau(\kappa^*) T M$. By \R{error_numerical_2} and \R{EN_def}, we see that the error is determined up to a mildly growing factor in $M$ by the quantity $E_{N}(f) = E_{\tau(\kappa^*) T M}(f)$. However, the error bound \R{alg_conv} depends on the ratio $N/T$, which in this case is equal to $\tau(\kappa^*) M$ regardless of the choice of $T$. Hence the bound \R{alg_conv} is independent of $T$ in the nonsaturating case for this choice of $N$, and therefore we expect the same $T$-independence of the FE approximation whenever $N$ is taken according to $\Theta^{(T)}(M;\kappa^*)$. This explanation aside, it is also of interest to numerically determine the so-called \textit{saturation point}: that is to say, the maximal value of $T$ (for a given $\kappa^*$) above which saturation occurs. This is shown in Figure \ref{f:Saturation_Compute}, where we plot the quantity $\kappa^{(T)}_{M,M}$ against $T$ for $M=500$. Note that this quantity is decreasing in $T$ (recall Figure \ref{f:ContourPlot}). Using this figure, we make the following observation. For the case $T=2$, saturation occurs when the $\kappa^* \approx 5e4$ or greater. Hence, we now conclude the following: \textit{if one implements the equispaced FE with $T=2$, $\epsilon = 10^{-13}$ and $N = \Theta^{(2,\epsilon)}(M;\kappa^*)$ chosen such that the $\kappa^*$ is less than $\approx 5e4$, then no other value of $T$ will asymptotically give a better approximation.} This establishes one of the main aims of this paper: namely, determining when the value $T = 2$, and hence the associated fast algorithm, can be used without concern that another value gives better accuracy. Note that this conclusion is very reasonable. In practice, we usually want the condition number to be much smaller than $10^4$ in magnitude. \begin{figure}\centering \includegraphics[width=7.50cm]{SaturationGraph_logscaled} \caption{The quantity $\kappa^{(T,\epsilon)}_{M,M} / \log M$ against $1 < T \leq 6$ for $M=500$ and $\epsilon = 10^{-13}$.} \label{f:Saturation_Compute} \end{figure} Having now ascertained that one may use $T=2$ without worry in most cases, we end this subsection by providing numerical values for the constants $\kappa^{(2,\epsilon)}_{M / \eta , M}$ and $\lambda^{(2,\epsilon)}_{M/\eta,M}$. These are shown in Figure \ref{f:T2_values}. As we see, the choices $\eta = 1.5$ or $\eta=2.0$ seem reasonable in practice. Increasing $\eta$ beyond this point brings only marginal benefits in stability. \begin{figure}\centering \begin{tabular}{|c|c|c|c|c|c|} \hline $\eta/M$ & 250 & 500 & 750 & 1000 \\ \hline 1.00 & 2.51e5 & 2.86e5 & 2.65e5 & 3.14e5 \\ \hline 1.25 & 1.01e4 & 1.25e4 & 1.72e4 & 1.99e4 \\ \hline 1.50 & 2.16e3 & 2.39e3 & 2.41e3 & 2.84e3 \\ \hline 2.00 & 1.88e2 & 2.25e2 & 2.89e2 & 3.27e2 \\ \hline 3.00 & 2.70e1 & 3.29e1 & 3.94e1 & 3.94e1 \\ \hline 4.00 & 1.18e1 & 1.53e1 & 1.67e1 & 1.84e1 \\ \hline \end{tabular} \hspace{1pc} \begin{tabular}{|c|c|c|c|c|c|} \hline $\eta/M$ & 250 & 500 & 750 & 1000 \\ \hline 1.00 & 6.16e-7 & 1.48e-6 & 2.76e-6 & 3.49e-6 \\ \hline 1.25 & 3.86e-8 & 8.82e-8 & 9.98e-8 & 1.33e-7 \\ \hline 1.50 & 3.04e-9 & 7.36e-9 & 1.56e-8 & 1.82e-8 \\ \hline 2.00 & 4.26e-10 & 1.02e-9 & 1.00e-9 & 1.23e-9 \\ \hline 3.00 & 2.97e-11& 8.20e-11 & 8.37e-11 & 1.97e-10 \\ \hline 4.00 & 1.71e-11 & 1.51e-11 & 3.59e-11 & 5.24e-11 \\ \hline \end{tabular} \caption{Values of $\kappa^{(T,\epsilon)}_{M/\eta,M}$ (left) and $\lambda^{(T,\epsilon)}_{M//\eta,M}$ (right) for $T=2$ and $\epsilon = 10^{-13}$.} \label{f:T2_values} \end{figure} \subsection{Other solvers}\label{ss:other_solvers} The phenomenon described above also occurs when a different numerical solver is used. We illustrate this in Figure \ref{f:othersolvers} for \textit{Matlab}'s $\backslash$ and \textit{Mathematica}'s \texttt{LeastSquares}. Interestingly, the effect is much more pronounced for the latter than for the former, although we believe it would become more apparent in the former for larger $M$ (recall that this phenomenon is asymptotic in $M$). The discrepencies in the results are likely due to the different algorithms having somewhat different default tolerances for solving ill-conditioned least-squares problems. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{f1app_BS_kappa25} & \includegraphics[width=5.00cm]{f5app_BS_kappa25} & \includegraphics[width=5.00cm]{f8app_BS_kappa25} \\ \includegraphics[width=5.00cm]{f1app_LS_kappa25} & \includegraphics[width=5.00cm]{f5app_LS_kappa25} & \includegraphics[width=5.00cm]{f8app_LS_kappa25} \end{array}$ \caption{\small Approximation errors for the functions $f_1(x)$, $f_5(x)$ and $f_8(x)$ (left to right) using \textit{Matlab}'s $\backslash$ (top row) and \textit{Mathematica}'s \texttt{LeastSquares} (bottom row). For each solver, the values $\Theta^{(T)}(M;\kappa^*)$ were computed for $\kappa^* = 25$.} \label{f:othersolvers} \end{center} \end{figure} \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{Theta_epsilon_logscaled_2} & \includegraphics[width=5.00cm]{Theta_epsilon_logscaled_2_scaled} & \includegraphics[width=5.00cm]{Fn_App_epsilon_logscaled_2} \end{array}$ \caption{\small Left and Middle: the functions $\Theta^{(T,\epsilon)}(M;\kappa^*)$ and $\Theta^{(T,\epsilon)}(M;\kappa^*)/M$ against $M$ for $T=2$, $\kappa^* = 10$ and $\epsilon = 10^{-2},10^{-3},\ldots,10^{-16}$ (thickest to thinnest). Note that the graphs for $\epsilon = 10^{-2},10^{-3},10^{-4}$ are the same. Right: approximation of $f(x) = \exp(250 \sqrt{2} \pi \I x)$ using these values.} \label{f:FnAppLog} \end{center} \end{figure} \subsection{Influence of the SVD tolerance $\epsilon$} Let us now return to the SVD algorithm. Thus far, we have taken the tolerance $\epsilon$ to be equal to $10^{-13}$. However, other values are possible, and it is of interest to determine how this affects the approximation. Note that some previous insight in this problem was given in \cite{BoydFourCont}, where the effect of $\epsilon$ was considered for specific functions. In Figure \ref{f:FnAppLog} we plot $\Theta^{(T,\epsilon)}(M;\kappa^*)$ for different values of $\epsilon$ using $T=2.0$. As is evident, $\Theta^{(T,\epsilon)}(M;\kappa^*)$ grows more rapidly for increasing $\epsilon$. This should come as no surprise, since it is the small singular vectors that cause ill-conditioning. However, the main purpose of this figure is to show the specific improvement that is possible by changing $\epsilon$. For example, when $\epsilon = 10^{-13}$ the function $\Theta^{(T,\epsilon)}(M;\kappa^*)$ is approximately $0.4 M$. Conversely, for $\epsilon = 10^{-6}$ it is approximately $0.8 M$, i.e.\ it scales roughly twice as quickly. If $\delta$ is some finite tolerance greater than $10^{-6}$, this means that the FE approximation with $\epsilon = 10^{-6}$ will approximate a given function to accuracy $\delta$ using roughly half the number of equispaced points as is required when $\epsilon = 10^{-13}$. This fact is also confirmed in the right panel of Figure \ref{f:FnAppLog}, where the oscillatory function $f(x) = \exp(250 \sqrt{2} \pi \I x)$ is approximated using different values of $\epsilon$. Resolving this function using $\epsilon = 10^{-13}$ requires roughly $3700$ equispaced data points, whereas when $\epsilon = 10^{-6}$ this value drops to around $1700$. Of course, the downside of a larger $\epsilon$ is that the minimal error is limited to approximately $\epsilon$, as can be seen in Figure \ref{f:FnAppLog}. Nonetheless, the conclusion we draw from this section is that if accuracy close to machine precision is not required -- as is typically the case in practice, where three to six digits is often acceptable -- then a viable way to increase the performance of the FE approximation whilst maintaining the condition number is to use a larger value of $\epsilon$. \section{Resolution power} In many applications, it is important to have an approximation algorithm with good \textit{resolution power}. Loosely speaking, this means that oscillatory functions are recovered using using a number of measurements that scales linearly with the frequency of oscillation with a constant that is as small as possible. Formally, let $\{ F_M \}_{M \in \bbN}$ be a sequence of approximations such that $F_M(f)$ depends only on the values of $f$ on an equispaced grid of $2M+1$ points. Let \bes{ \cR(\omega,\delta) = \min \left \{ M \in \bbN : \| \E^{ \I \pi \omega \cdot} - F_M(\E^{\I \pi \omega \cdot}) \|_{\infty} < \delta \right \},\quad \omega > 0,\ 0 < \delta < 1, } then we say that $F_M$ has resolution constant $0 < r <\infty$ if \be{ \label{} \cR (\omega,\delta) \sim r \omega,\quad \omega \rightarrow \infty, } for any fixed $\delta$. The approximation $F_M$ has good resolution power if $r$ is small, and bad resolution power if $r$ is large. For periodic oscillations, the Fourier series approximation on $[-1,1]$ of degree $M$ has optimal resolution constant $r=1$. Since the number equispaced points is $2M+1$, this corresponds to two \textit{points-per-wavelength}. Of course, Fourier series do not converge uniformly for nonperiodic oscillations, which is why we resort to alternative algorithms such as FEs. Naturally, though, it is desirable that the FE resolution constant be as close to the optimal value $r=1$ as possible. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{Resn_eta1_delta10e3} & \includegraphics[width=5.00cm]{Resn_eta32_delta10e3} & \includegraphics[width=5.20cm]{Resn_eta2_delta10e3} \\ \eta = 1 & \eta = 1.5 & \eta = 2 \end{array}$ \caption{\small The quantity $\cR(\omega,\delta ; T,\eta) / \omega$ against $\omega$ for $T=3.0,2.5,2.0,1.5,1.25,1.125$ and $\delta = 10^{-3}$. The solid lines indicate the values $T \eta$.} \label{f:ResnPower} \end{center} \end{figure} In Figure \ref{f:ResnPower} we numerically determine the resolution constant for the FE approximation $F^{(T,\epsilon)}_{M/\eta,M}$ by computing the function \bes{ \cR(\omega,\delta) =\cR(\omega,\delta ; T , \eta). } These results suggest that the resolution constant \be{ \label{r_prod} r \approx T \eta, } is approximately the product of the extension parameter $T$ and the oversampling ratio $\eta$. We shall discuss the consequences of this observation in a moment, but we first wish to explain why $r$ should be at most $T \eta$. Recall from \R{error_numerical_2} that the error of the FE approximation is determined (up to a small linear constant in $M$) by the decay of the factor $E_N(f)$, given by \R{EN_def}, where $N = M/\eta$. Now let \be{ \label{R0} \cR_0(\omega,\delta) = \min \left \{ N \in \bbN : E_N(\E^{\I \pi \omega \cdot}) < \delta \right \},\quad \omega > 0, \sqrt{2T} \mu < \delta < 1. } Then to show that $r = T \eta$, it suffices to prove that \bes{ \cR_0(\omega,\delta) \sim T \omega,\quad \omega \rightarrow \infty. } We now note the following: \lem{ Let $\cR_0(\omega,\delta)$ be as in \R{R0}. Then \bes{ \limsup_{\omega \rightarrow \infty} \cR_0(\omega,\delta)/\omega \leq T. } } \prf{ Let $f(x) = \E^{\I \pi \omega x}$ and write $\omega = P/T + z$ where $0 \leq z < 1/T$ and $P \in \bbN$. Then \bes{ f(x) = \E^{\I \frac{P \pi}{T} x} g(x),\qquad g(x) = \E^{\I \frac{z \pi}{T} x}. } Let $K \in \bbN$ and let $\phi \in \cG^{(T)}_{P+K}$ be given by $\phi(x) = \E^{\I \frac{P \pi}{T} x} \tilde{\phi}(x)$, where $\tilde{\phi} \in \cG^{(T)}_{K}$ is arbitrary. Define $N = P+K$, and let $\mathbf{a}$ be the vector of coefficients of $\phi$. Then by definition, \bes{ E_N(f) \leq \| f -\phi \|_{\infty} + \mu | \mathbf{a} |_{\infty} = \| g - \tilde{\phi} \|_{\infty} + \mu | \mathbf{\tilde{a}} |_{\infty}, } where $\mathbf{\tilde{a}}$ is the vector of coefficients of $\tilde{\phi}$. Since $\tilde{\phi}$ is arbitrary, we deduce that $E_N(f) \leq E_{K}(g)$. Thus $E_N(f) \leq \delta$ provided $E_K(g) \leq \delta$. Using Corollary \ref{c:EN_rate}, we see that $E_K(g) \leq \delta$ provided $K \geq K_0(\delta)$, where $K_0(\delta)$ is independent of $\omega$. Thus, $E_N(f) \leq \delta$ whenever $N = P + K_0(\delta) = (\omega - z)T + K_0(\delta)$. Since $K_0(\delta)$ is independent of $\omega$, we obtain the result. } From this, we deduce that $\cR_0(\omega,\delta) \leq T \omega + \ord{1}$ for large $\omega$ and hence $r \leq T \omega$. Unfortunately, we have no proof of the lower bound, although it is supported by the results in Figure \ref{f:ResnPower}. Let us now discuss the consequences of \R{r_prod}. Suppose we return to the earlier experiment where $\kappa^*$ is fixed and $N$ is chosen according to $N = \Theta^{(T)}(M;\kappa^*)$. For large $M$, Figure \ref{f:ThetaPlot} shows that \be{ \label{etaTkappa} \Theta^{(T)}(M;\kappa^*) \approx M/\eta^{(T)} ,\qquad \eta^{(T)} = \eta^{(T)}(\kappa^*), } for some fixed $\eta^{(T)}$ depending on $T$ and $\kappa^*$. In our previous observation, it was found that the resulting FE approximation was independent of $T$. In particular, this holds for the oscillatory exponential $\exp(\I \pi \omega x)$. But since the resolution constant $r$, which describes the point after which the approximation error begins to decay, is equal to $\eta T$, we therefore deduce the following: \bes{ \eta^{(T)} T = \eta^{(T')} T', } for any $T$ and $T'$ that do not saturate. This implies that the level curves of the condition number $\kappa = \kappa^{(T)}_{M/\eta,M}$ are approximately given by $T \eta = \mbox{constant}$, for sufficiently large $M$ and non-saturating $T$, which is in good agreement with the contour plot given in Figure \ref{f:ContourPlot}. In addition to this, another important implication of \R{r_prod} is that to get better resolution power one must necessarily worsen the stability of the algorithm. In other words, there is a direct relationship between $\kappa^*$ and $r$, regardless of the choice of $T$ made. In Table \ref{tab:ResnConstT2} we give numerical results for the resolution constant when $T=2$ (and therefore all nonsaturating $T$) for different values of $\kappa^*$. As we see, by allowing $\kappa^*$ to increase to roughly $500$ we get a marked improvement over when $\kappa^* = 10$. Beyond this point, further increases give only marginal gains at the expense of much larger condition numbers. \begin{table} \begin{center} \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|} \hline $\kappa^*$ & $10$ & $25$ & $100$ & $500$ & $1000$ & $5000$ & $10000$ \\ \hline $r$ & 5.41 & 4.44 & 3.64 & 2.98 & 2.77 & 2.38 & 2.25\\ \hline \end{tabular} \caption{The resolution constant $r = \eta^{(T)} T$ for $T=2$, where $\eta^{(T)} = \eta^{(T)}(\kappa^*)$ is given by \R{etaTkappa}.}\label{tab:ResnConstT2} \end{center} \end{table} \section{Other data}\label{s:other_data} In this final section, we illustrate that the main phenomena observed for equispaced data are also witnessed for numerous other types of data. To this end, we consider the following four examples: \begin{itemize} \item \textit{Jittered pointwise data}. Here the measurements of $f$ are pointwise samples at the jittered locations \bes{ f(x_m),\qquad x_m = \frac{m}{M} + z_m, } where $z_m \in (-\delta/M,\delta/M)$ and $0 < \delta < 1$. This is a typical example of a nonuniform sampling pattern for scattered data approximation. We shall choose the $z_m$'s as follows: \be{ \label{jittered_data} z_m = \frac{\delta}{M} \sin(M^2/m),\quad m \neq 0,\quad z_0 = 0. } \end{itemize} \begin{itemize} \item \textit{Logarithmic pointwise data}. Here we again sample $f$ pointwise at nodes $x_m$, but in this case the nodes are logarithmically distributed: \be{ \label{log_data} x_m = - x_{-m} = 10^{\left(\frac{m-1}{M-1} - 1\right) \log_{10}(c M)},\quad x_0 = 0, } where $c>0$ is a fixed, user-controlled parameter (we take $c=2$ in our results). This sampling pattern corresponds to a nonuniform sampling scenario where data is collected more densely at the origin. \end{itemize} \begin{itemize} \item \textit{Fourier data}. In some applications, rather than pointwise samples, we may wish to reconstruct $f$ from its Fourier coefficients: \be{ \label{Four_data} \hat{f}(m) = \int^{1}_{-1} f(x) \E^{-\I \pi m x} \D x,\quad |m| \leq M. } As discussed in \cite{BAACHOptimality}, this can be seen as a continuous analogue of the equispaced data recovery problem. In particular, there is a completely analogous result result to that of Platte, Trefethen \& Kuijlaars regarding stability and convergence \cite{AdcockHansenShadrinStabilityFourier}. Note that for this data we replace the uniform norms used in the error estimates and the definitions of $\kappa^{(T)}_{N,M}$ and $\lambda^{(T)}_{N,M}$ by the $\rL^2$ and $\ell^2$ norms. This is natural in view of Parseval's identity for Fourier coefficients. \end{itemize} \begin{itemize} \item \textit{Optimal pointwise data}. Finally, in order to show that the phenomenon is not witnessed for all data, we consider pointwise samples taken at the so-called mapped symmetric Chebyshev nodes (see \cite{BADHFEResolution}): \be{ \label{mapped_Cheby} x_m = - x_{-m-1} = m^{-1} \left ( \cos \frac{(2m+1) \pi}{2M+2} \right ),\ m=0,\ldots,M, } where $m$ is given by \R{mapping}. These nodes are derived from the observation that FE approximations correspond to algebraic polynomial approximations in the mapped co-ordinate $z = m(x)$ \cite{FEStability}. Chebyshev nodes provide optimal nodes for polynomial interpolation. Therefore, under the inverse mapping $m^{-1}$, they provide the optimal nodes \R{mapped_Cheby} for FE approximations. Note that since the nodes arise in this way, no oversampling is required in the FE approximation, i.e.\ we let $\eta = 1$ in this case. \end{itemize} \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{Theta_25_Jitter_logscaled_2} & \includegraphics[width=5.00cm]{Theta_25_Log_logscaled_2} & \includegraphics[width=5.00cm]{Theta_10_Fourier_logscaled_2} \\ \includegraphics[width=5.00cm]{Theta_25_Jitter_logscaled_2_scaled} & \includegraphics[width=5.00cm]{Theta_25_Log_logscaled_2_scaled} & \includegraphics[width=5.00cm]{Theta_10_Fourier_logscaled_2_scaled} \end{array}$ \caption{\small Plots of $\Theta^{(T)}(M;\kappa^*)$ (top row) and $\Theta^{(T)}(M;\kappa^*) / S(M)$ (bottom row) against $M$ for jittered \R{jittered_data} (left), logarithmic \R{log_data} (middle) and Fourier \R{Four_data} (right) data. Here $S(M) = M/\log(M)$ for the logarithmic data and $S(M) = M$ otherwise, and $\kappa^* = 25$ (jittered, logarithmic) or $\kappa^* = 10$ (Fourier).} \label{f:ThetaPlot_Data} \end{center} \end{figure} In Figure \ref{f:ThetaPlot_Data} we give plots of the function $\Theta^{(T)}(M;\kappa^*)$ for the first three data types. For jittered and Fourier data the scaling is linear, whereas for the logarithmic data $\Theta^{(T)}$ scales like $M / \log M$. This scaling is proportional to the reciprocal of the maximal spacing between nodes in the case, and hence is completely expected. Note also that no values of $T$ saturate for the logarithmic data, whereas values $T=5.0$ and $T=6.0$ saturate for the jittered data, and for the Fourier data the values $T=3.0$, $T=4.0$, $T=5.0$ and $T=6.0$ all saturate. The lower saturation point for the latter is due to the fact that the condition number is measured in the weaker $\rL^2$ norm. \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.20cm]{f1app_SVD_1e13_kappa25_Jitter_logscaled_2} & \includegraphics[width=5.20cm]{f4app_SVD_1e13_kappa25_Jitter_logscaled_2} & \includegraphics[width=5.20cm]{f8app_SVD_1e13_kappa25_Jitter_logscaled_2} \\ f(x) = \E^{230 \sqrt{2} \pi \I x} & f(x) = \frac{1}{1+1500 x^2} & f(x) = \E^{-1/(8 x)^2} \end{array}$ \caption{\small Approximation errors for jittered data \R{jittered_data} using the values $\Theta^{(T)}(M;\kappa^*)$ from Figure \ref{f:ThetaPlot_Data}.} \label{f:FnApp_Data} \end{center} \end{figure} \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.20cm]{f1app_SVD_1e13_kappa25_Log_logscaled_2} & \includegraphics[width=5.20cm]{f4app_SVD_1e13_kappa25_Log_logscaled_2} & \includegraphics[width=5.20cm]{f8app_SVD_1e13_kappa25_Log_logscaled_2} \\ f(x) = \E^{50 \sqrt{2} \pi \I x} & f(x) = \frac{1}{1+65 x^2} & f(x) = \E^{-2/(3x^2)} \end{array}$ \caption{\small Approximation errors for logarithmic data \R{log_data} using the values $\Theta^{(T)}(M;\kappa^*)$ from Figure \ref{f:ThetaPlot_Data}. } \label{f:FnApp_Data2} \end{center} \end{figure} \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{f1app_SVD_1e13_kappa10_logscaled_Fourier} & \includegraphics[width=5.00cm]{f2app_SVD_1e13_kappa10_logscaled_Fourier} & \includegraphics[width=5.00cm]{f3app_SVD_1e13_kappa10_logscaled_Fourier} \\ f(x) = \E^{175 \sqrt{2} \pi \I x} & f(x) = \frac{1}{1+1500 x^2} & f(x) = \frac{1}{1+25 \sin^2 8x} \end{array}$ \caption{\small Approximation errors for Fourier data \R{Four_data} using the values $\Theta^{(T)}(M;\kappa^*)$ from Figure \ref{f:ThetaPlot_Data}. } \label{f:FnApp_Data3} \end{center} \end{figure} \begin{figure} \begin{center} $\begin{array}{ccc} \includegraphics[width=5.00cm]{f1app_SVD_1e13_Optimal} & \includegraphics[width=5.00cm]{f2app_SVD_1e13_Optimal} & \includegraphics[width=5.00cm]{f3app_SVD_1e13_Optimal} \\ f(x) = \E^{200 \sqrt{2} \pi \I x} & f(x) = \frac{1}{1+1000 x^2} & f(x) = \E^{-1/(8x)^2} \end{array}$ \caption{\small Approximation errors for the optimal pointwise data \R{mapped_Cheby} using $M=N$.} \label{f:FnApp_Data4} \end{center} \end{figure} Next, in Figures \ref{f:FnApp_Data}--\ref{f:FnApp_Data3} we compare approximation errors using these values. For the jittered and Fourier data we see exactly the same phenomenon as before: namely, the approximation errors are roughly independent of $T$. A similar phenomenon is witnessed for the logarithmic data, although it is slightly weaker: $T=1.125$ gives somewhat better errors than the other values. This is due to the much more severe scaling of $\Theta^{(T)}(M;\kappa^*)$ with $M$ in this case, which means the asymptotic regime takes longer to set in. Finally, in Figure \ref{f:FnApp_Data4} we display approximation errors for the optimal pointwise data \R{mapped_Cheby}. As is evident, the phenomenon does not occur in this case. The reason for this is due to the choice of the data, which means that no oversampling is required. Thus the FE approximation error is proportional to $E_N(f)$, where $N = M$, multiplied by a mildly growing factor. Figure \ref{f:FnApp_Data4} therefore serves as a reminder that smaller values of $T$ are intrinsically better than larger values, provided one has freedom to pick ideal data. On the other hand, for nonideal data -- such as equispaced, jittered, logarithmic or Fourier data -- this effect is nullified by a worse scaling of $\Theta^{(T)}(M;\kappa^*)$. \section{Conclusions and open problems}\label{s:conclusion} The purpose of this paper was to document an interesting phenomenon in FE approximations from equispaced data. Namely, when the desired condition number is fixed, the choice of the extension parameter $T$ has no substantial effect on the approximation. This is on the proviso that saturation does not occur, which we have shown to be the case for moderate values of $T$ and $\kappa^*$. In particular, one may use $T=2$, and the associated fast algorithm, without concern that it is suboptimal. The main open problem is to provide mathematical analysis for the empirical conclusions drawn. We believe this is possible, although not straightforward. One possible approach towards this is to conduct an asymptotic analysis of the singular values and vectors of the matrix $A^{(T)}$. Recall that the normal form $(A^{(T)})^* A^{(T)}$ is a Riemann sum approximation to the Gram matrix $G$ of the FE basis functions $\phi_n(x) = \E^{\I \frac{n \pi}{T} x}$. As discussed in \cite{FEStability}, the matrix $G$ is precisely the prolate matrix. The eigenvalues and eigenvectors of this matrix were analyzed in detail by Slepian \cite{SlepianV} (see also \cite{Varah}). It may be possible to do the same for the discretized version $(A^{(T)})^* A^{(T)}$, and this is an important topic for future work. Another question raised by this work is that of whether it might be possible to vary $T$ with $M$ to achieve better results; in particular, improved resolution power. We believe this may be the case, the caveat being that there is currently no fast algorithm for $T \neq 2$. Some potential choices for varying $T$ with $M$ were considered previously in \cite{BADHFEResolution,FEStability}. But it may also be possible using the approach of this paper to numerically compute an optimal (in some sense) value of $T$ for each $M$. We leave this for future work. \section*{Acknowledgments} The authors would like to thank Daan Huybrechs, Mark Lyon and Rodrigo Platte for useful discussions. BA acknowledges support from the NSF DMS grant 1318894. \bibliographystyle{abbrv} \small

https://arxiv.org/abs/1402.3413

The colourful simplicial depth conjecture

Given $d+1$ sets of points, or colours, $S_1,\ldots,S_{d+1}$ in $\mathbb R^d$, a colourful simplex is a set $T\subseteq\bigcup_{i=1}^{d+1}S_i$ such that $|T\cap S_i|\leq 1$, for all $i\in\{1,\ldots,d+1\}$. The colourful Carathéodory theorem states that, if $\mathbf 0$ is in the convex hull of each $S_i$, then there exists a colourful simplex $T$ containing $\mathbf 0$ in its convex hull. Deza, Huang, Stephen, and Terlaky (Colourful simplicial depth, Discrete Comput. Geom., 35, 597--604 (2006)) conjectured that, when $|S_i|=d+1$ for all $i\in\{1,\ldots,d+1\}$, there are always at least $d^2+1$ colourful simplices containing $\mathbf 0$ in their convex hulls. We prove this conjecture via a combinatorial approach.

\section{Introduction} A {\em colourful point configuration} is a collection of $d+1$ sets of points $\mathbf{S}_1,\ldots,\mathbf{S}_{d+1}$ in $\mathbb R^d$. A {\em colourful simplex} is a subset $T$ of $\bigcup_{i=1}^{d+1}\mathbf{S}_i$ such that $|T\cap \mathbf{S}_i|\leq 1$. The colourful Carath\'eodory theorem, proved by B\'ar\'any in 1982~\cite{Bar82}, states that, given a colourful point configuration $\mathbf{S}_1,\ldots,\mathbf{S}_{d+1}$ in $\mathbb R^d$ such that ${\bf 0}\in\bigcap_{i=1}^{d+1}\operatorname{conv}(\mathbf{S}_i)$, there exists a colourful simplex $T$ containing ${\bf 0}$ in its convex hull. In the same paper, B\'ar\'any uses this theorem combined with Tverberg's theorem to give a bound on simplicial depth. His argument motivated the following question: how many colourful simplices, at least, contain ${\bf 0}$ in their convex hulls? Let $\mu(d)$ denote the minimal number of colourful simplices containing ${\bf 0}$ in their convex hulls over all colourful point configurations $\mathbf{S}_1,\ldots,\mathbf{S}_{d+1}$ in $\mathbb R^d$ such that ${\bf 0}\in\operatorname{conv}(\mathbf{S}_i)$ and $|\mathbf{S}_i|=d+1$ for $i=1,\ldots,d+1$. The colourful Carath\'eodory theorem states that $\mu(d)\geq 1$. The quantity $\mu(d)$ has been investigated by~\citet*{DHST06}. They proved that $2d\leq\mu(d)\leq d^2+1$ and conjectured that $\mu(d)=d^2+1$. Later \citet*{BM06} proved that $\mu(d)\geq\max\left(3d,\left\lceil\frac{d(d+1)}5\right\rceil\right)$ for $d\geq 3$, \citet*{ST06} proved that $\mu(d)\geq\left\lfloor\frac{(d+2)^2}4\right\rfloor$, and \citet*{DSX11} showed that $\mu(d)\geq\left\lceil\frac{(d+1)^2}2\right\rceil$. \citet*{DMS14} improved the bound to $\frac 12d^2+\frac 72d-8$ for $d\geq 4$. This latter result was obtained using a combinatorial generalization of the colourful point configurations suggested by B\'ar\'any{} and known as {\em octahedral systems}, see~\cite{DSX11}. We use this combinatorial approach to prove the conjecture. \begin{theorem}\label{thm:mainGeom} The equality $\mu(d)=d^2+1$ holds for every integer $d\geq 1$. \end{theorem} The outline of the paper goes as follows. Section~\ref{sec:preliminary} is divided into two parts. First we define the octahedral systems and show their link with the colourful point configurations. Second, we introduce one of our main tools: the decomposition of an octahedral system over some elementary octahedral systems called umbrellas. Section~\ref{sec:proof} is devoted to the proof of Theorem~\ref{thm:mainGeom}. \section{Preliminaries}\label{sec:preliminary} \subsection{Octahedral systems}\label{subsec:oct_sys} Let $V_1,\ldots,V_n$ be $n$ pairwise disjoint finite sets, each of size at least $2$. An {\em octahedral system} is a set $\Omega\subseteq V_1\times\cdots\times V_n$ satisfying the {\em parity condition}: the cardinality of $\Omega\cap (X_1\times\cdots\times X_n)$ is even if $X_i\subseteq V_i$ and $|X_i|=2$ for all $i\in\{1,\ldots,n\}$. We use the terminology of hypergraphs to describe an octahedral system: the sets $V_i$ are the {\em classes}, the elements in $V_i$ are the {\em vertices}, and the $n$-tuples in $V_1\times\cdots\times V_n$ are the {\em edges}. An edge whose $i$th component is a vertex $x\in V_i$ is {\em incident with the vertex $x$}, and conversely. A vertex $x$ incident with no edges is {\em isolated}. A class $V_i$ is {\em covered} if each vertex of $V_i$ is incident with at least one edge. Finally, the set of edges incident with $x$ is denoted by $\delta_\Omega(x)$ and the {\em degree of $x$}, denoted by $\deg_\Omega(x)$, refers to $|\delta_\Omega(x)|$. \begin{lemma}\label{lem:nonEmpty} In every nonempty octahedral system, at least one class is covered. \end{lemma} \begin{pf} Consider an octahedral system $\Omega\subseteq V_1\times\cdots\times V_n$. Suppose that no classes are covered. There is at least one isolated vertex $x_i$ in each $V_i$. Hence, if there were an edge $(y_1,\ldots,y_n)$ in $\Omega$, then the parity condition would not be satisfied for $X_i=\{x_i,y_i\}$. \end{pf} Given a colourful point configuration $\mathbf{S}_1,\ldots,\mathbf{S}_{d+1}$, the Octahedron Lemma~\cite{BM06,DHST06} states that, for any $\mathbf{S}'_1\subseteq \mathbf{S}_1,\ldots,\mathbf{S}'_{d+1}\subseteq \mathbf{S}_{d+1}$, with $|\mathbf{S}'_1|=\cdots=|\mathbf{S}'_{d+1}|=2$, the number of colourful simplices generated by $\bigcup_{i=1}^{d+1}\mathbf{S}'_i$ and containing ${\bf 0}$ in their convex hulls is even. The hypergraph over $V_1\times\cdots\times V_n$ where $V_i$ is identified with $\mathbf{S}_i$ and whose edges are identified with the colourful simplices containing ${\bf 0}$ in their convex hulls is therefore an octahedral system. Furthermore, a strengthening of the colourful Carath\'eodory{} Theorem, given in~\cite{Bar82}, states that if ${\bf 0}\in\bigcap_{i=1}^{d+1}\operatorname{conv}(\mathbf{S}_i)$, then each point of the colourful point configuration is in some colourful simplices containing ${\bf 0}$ in their convex hulls. Hence, in an octahedral system $\Omega$ arising from such a colourful point configuration, each class $V_i$ is covered. \subsection{Decompositions} The following proposition, proved in~\cite{DMS14}, states that the set of all octahedral systems is stable under the ``symmetric difference'' operation. \begin{proposition}\label{prop:stable} Let $\Omega$ and $\Omega'$ be two octahedral systems over the same vertex set. $\Omega\triangle\Omega'$ is an octahedral system. \end{proposition} \begin{pf} Let $\Omega''=\Omega\triangle\Omega'$. As $\Omega''$ is a subset of $V_1\times\cdots\times V_n$, we simply check that the parity condition is satisfied. Consider $X_1\subseteq V_1,\ldots,X_n\subseteq V_n$ with $|X_i|=2$ for $i=1,\ldots,n$. We have $$ |\Omega''\cap (X_1\times\cdots\times X_n)|=|\Omega\cap (X_1\times\cdots\times X_n)|+|\Omega'\cap (X_1\times\cdots\times X_n)|-2|\Omega\cap\Omega'\cap (X_1\times\cdots\times X_n)|.$$ All the terms of the sum are even, which allows to conclude. \end{pf} We now present a family of specific octahedral systems we call {\em umbrellas}. An umbrella $U$ is a set of the form $\{x^{(1)}\}\times\cdots\times\{x^{(i-1)}\}\times V_i\times\{x^{(i+1)}\}\times\cdots\times\{x^{(n)}\}$, with $x^{(j)}\in V_j$ for $j\neq i$. The class $V_i$ covered in $U$ is called its {\em colour}. $T=(x^{(1)},\ldots,x^{(i-1)},x^{(i+1)},\ldots,x^{(n)})$ is its {\em transversal}. An umbrella is clearly an octahedral system over $V_1\times\cdots\times V_n$ and we have the following proposition. \begin{proposition}\label{prop:propCol} Two umbrellas of the same colour have an edge in common if and only if they are equal. \end{proposition} \begin{pf} An umbrella is entirely determined by its colour $V_i$ and its transversal $T$. Therefore, if two umbrellas of the same colour have an edge in common, they necessarily have the same transversal, which implies that they are equal. \end{pf} It was implicitly proved in Section 3 of~\cite{DMS14} that any octahedral system can be described as a symmetric difference of umbrellas. In this paper, we describe an octahedral system as a symmetric difference of other octahedral systems to bound its cardinality. We now focus on octahedral systems where the size of each class is equal to the number of classes. Consider a nonempty octahedral system $\Omega\subseteq V_1\times\cdots\times V_n$ with $|V_i|=n$ for all $i\in\{1,\ldots,n\}$. Denote by $i_1$ the smallest $i\in\{1,\ldots,n\}$ such that $V_{i}$ is covered in $\Omega$ and order the vertices $\{x_1,\ldots,x_n\}$ of $V_{i_1}$ by increasing degree: $\deg_\Omega(x_1)\leq\cdots\leq\deg_\Omega(x_n)$. We define $\mathcal U$ to be the set of umbrellas of colour $V_{i_1}$ containing an edge of $\Omega$ incident with $x_1$ and $W=\triangle_{U\in\mathcal U}U$. Let $\Omega_j$ be the set of all edges in $\Omega\triangle W$ incident with $x_j$. Formally, $$\mathcal U=\{U: U\mbox{ umbrella of colour }V_{i_1}\mbox{ and }U\cap\delta_\Omega(x_1)\neq\emptyset\}\mbox{ and }\Omega_j=\delta_{\Omega\triangle W}(x_j).$$ Note that $|\mathcal U|=\deg_\Omega(x_1)$. In the remaining of the paper we refer to $(\mathcal U,\Omega_2,\ldots,\Omega_n)$ as a {\em suitable decomposition}. \begin{lemma}\label{lem:decompo} Let $(\mathcal U,\Omega_2,\ldots,\Omega_n)$ be a suitable decomposition and $W=\triangle_{U\in\mathcal U}U$. We have \begin{enumerate} \item[\textup{(i)}] $\Omega_j\cap\Omega_\ell=\emptyset$, for all $j\neq \ell$ (they have no edge in common), \item[\textup{(ii)}] $\Omega=W\triangle \Omega_2\triangle\cdots\triangle\Omega_n$, \item[\textup{(iii)}] $\Omega_j$ is an octahedral system, for all $j$, \item[\textup{(iv)}] $\deg_{\Omega}(x_j)\geq\max(|\mathcal U|,|\Omega_j|-|\Omega_j\cap W|)$ for all $j$. \item[\textup{(v)}] If $V_i$ is not covered in $\Omega$, then $V_i$ is neither covered in $\Omega\triangle W$ nor in any $\Omega_j$. \end{enumerate} \end{lemma} The terminology suitable decomposition is due to point (ii) of Lemma~\ref{lem:decompo}. \begin{pf}[Proof of Lemma~\ref{lem:decompo}] We first prove (i). The $i_1$th component of any edge in $\Omega_j$ is $x_j$. Therefore, $\Omega_j$ and $\Omega_\ell$ have no edge in common if $j\neq\ell$. We then prove (ii). There are exactly $\deg_\Omega(x_1)$ umbrellas of colour $V_{i_1}$ containing an edge of $\Omega$ incident with $x_1$. As $W$ is the symmetric difference of these umbrellas, $x_1$ is isolated in $\Omega\triangle W$. Thus, $\Omega_2,\ldots,\Omega_n$ form a partition of the edges in $\Omega\triangle W$ and $\Omega\triangle W=\Omega_2\triangle\cdots\triangle\Omega_n$. Taking the symmetric difference of this equality with $W$ we obtain $\Omega=W\triangle\Omega_2\triangle\cdots\triangle\Omega_n$. We now prove (iii). By definition, the $\Omega_j$'s are subsets of $V_1\times\cdots\times V_n$. It remains to prove that they satisfy the parity condition. Consider $X_i\subseteq V_i$ with $|X_i|=2$ for $i=1,\ldots,n$. If $X_{i_1}$ does not contain $x_j$, there are no edges in $\Omega_j$ induced by $X_1\times\cdots\times X_n$. If $X_{i_1}$ contains $x_j$, the edges in $\Omega_j$ induced by $X_1\times\cdots\times X_n$ are the ones induced by $X_1\times\cdots\times X_{i_1-1}\times\{x_j\}\times X_{i_1+1}\times\cdots\times X_n$. As $x_1$ is isolated in $\Omega\triangle W$, those edges are exactly the edges in $\Omega\triangle W$ induced by $X_1\times\cdots\times X_{i_1-1}\times\{x_1,x_j\}\times X_{i_1+1}\times\cdots\times X_n$. According to Proposition~\ref{prop:stable}, $W$ is an octahedral system and $\Omega\triangle W$ as well, hence there is an even number of edges. We prove (iv). We have $|\mathcal U|=\deg_\Omega(x_1)\leq\deg_\Omega(x_j)$ for all $j\in\{1,\ldots,n\}$. Furthermore, by definition of the symmetric difference, we have $(\Omega_2\triangle\cdots\triangle\Omega_n)\setminus W\subseteq \Omega$. This inclusion becomes $(\Omega_2\setminus W)\triangle\cdots\triangle(\Omega_n\setminus W)\subseteq \Omega$. As two $\Omega_\ell$'s share no edges, $\Omega_j\setminus W\subseteq\Omega$ and thus $\Omega_j\setminus W\subseteq\delta_\Omega(x_j)$ for all $j\in\{2,\ldots,n\}$. We obtain $$|\Omega_j|-|\Omega_j\cap W|\leq\deg_\Omega(x_j).$$ Finally to prove (v) it suffices to prove that a class $V_i$ not covered in $\Omega$ remains not covered in $\Omega\triangle W$. Indeed, if a class is covered in an $\Omega_j$, it is also covered in $\Omega\triangle W$, as no two $\Omega_\ell$'s have an edge in common. Consider $V_i$ not covered in $\Omega$. There is a vertex $x\in V_i$ incident with no edges in $\Omega$. In particular, there are no edges in $\Omega$ incident with $x_1$ and $x$. Therefore, the umbrellas in $\mathcal U$, which are defined by the edges incident with $x_1$, contain no edges incident with $x$. Hence, $x$ is isolated in $W=\triangle_{U\in\mathcal U}U$ and in $\Omega$. Finally, $x$ remains isolated in $\Omega\triangle W$. \end{pf} Unlike the suitable decomposition of $\Omega$, which is a decomposition over general octahedral systems, the decomposition given in the following lemma is over umbrellas. \begin{lemma}\label{lem:elDecompo} Consider an octahedral system $\Omega\subseteq V_1\times\cdots\times V_n$ with $|V_i|=n$ for all $i\in\{1,\ldots,n\}$. There exists a set of umbrellas $\mathcal D$, such that $\Omega=\triangle_{U\in\mathcal D}U$ and such that the following implication holds: \begin{center} $V_i$ is the colour of some $U\in\mathcal D$ $\Longrightarrow$ $V_i$ is covered in $\Omega$. \end{center} \end{lemma} \begin{pf} The proof works by induction on the number of covered classes in $\Omega$. If no classes are covered, then, according to Lemma~\ref{lem:nonEmpty}, $\Omega$ is empty. Suppose now that $k$ classes are covered, with $k\geq 1$, and consider a suitable decomposition $(\mathcal U,\Omega_2,\ldots,\Omega_n)$ of $\Omega$. Denote by $W$ the symmetric difference $W=\triangle_{U\in\mathcal U}U$. According to Proposition~\ref{prop:stable}, $W$ is an octahedral system, and so is $\Omega\triangle W$. There are stricly fewer covered classes in $\Omega\triangle W$ than in $\Omega$. Indeed, in $\Omega\triangle W$, the class $V_{i_1}$ is no longer covered, since $x_1$ is isolated, and according to (v) of Lemma~\ref{lem:decompo}, a class not covered in $\Omega$ remains not covered in $\Omega\triangle W$. By induction, there exists a set $\mathcal D'$ of umbrellas such that $\Omega\triangle W=\triangle_{U\in\mathcal D'}U$, and such that if there is an umbrella of colour $V_i$ in $\mathcal D'$, then $V_i$ is covered in $\Omega\triangle W$. As the umbrellas in $\mathcal D'$ are not of colour $V_{i_1}$, we have $\mathcal U\cap\mathcal D'=\emptyset$. Therefore, $\Omega=(\triangle_{U\in\mathcal U}U)\triangle(\triangle_{U\in\mathcal D'}U)$ and the set $\mathcal D=\mathcal U\cup\mathcal D'$ satisfies the statement of the lemma. \end{pf} \section{Proof of the main result}\label{sec:proof} The following theorem gives a general lower bound on the cardinality of an octahedral system. Our main theorem is a corollary of it. \begin{theorem}\label{thm:main} Let $\Omega\subseteq{V_1\times\cdots\times V_n}$ be an octahedral system with $|V_1|=\cdots=|V_n|=n\geq 2$. If $k\geq 1$ classes among the $V_i$'s are covered, then $$|\Omega|\geq k(n-2)+2.$$ \end{theorem} Before proving this theorem, we show how the main theorem can be deduced from it. \begin{pot} The inequality $\mu(d)\leq d^2+1$ is proved in~\cite{DHST06}. Let $\mathbf{S}_1,\ldots,\mathbf{S}_{d+1}$ be a colourful point configuration in $\mathbb{R}^d$. As explained in Section~\ref{subsec:oct_sys}, the set $\Omega\subseteq{V_1\times\cdots\times V_{d+1}}$, with $V_i=\mathbf{S}_i$ for $i=1,\ldots,d+1$ and whose edges correspond to the colourful simplices containing ${\bf 0}$ in their convex hulls, is an octahedral system. According to~\cite[Theorem 2.3.]{Bar82}, all the classes are covered in this octahedral system. Applying Theorem~\ref{thm:main} with $k=n=d+1$ gives the lower bound: $\mu(d)\geq d^2+1$. \end{pot} The remainder of the section is devoted to the proof of Theorem~\ref{thm:main}. The proof distinguishes two cases, corresponding to the following Propositions~\ref{prop:cas1} and~\ref{prop:cas2}. We first prove these propositions. \begin{proposition} \label{prop:cas1} Consider an octahedral system $\Omega\subseteq V_1\times\cdots\times V_n$ with $|V_i|=n$ for all $i\in\{1,\ldots,n\}$ and a class $V_i$ covered in $\Omega$. If $\Omega$ can be written as a symmetric difference of umbrellas, none of them being of colour $V_i$, then $|\Omega|\geq n^2$. \end{proposition} \begin{pf} Let $\mathcal D$ be a set of umbrellas such that there are no umbrellas of colour $V_i$ in $\mathcal D$ and $\Omega=\triangle_{U\in\mathcal D}U$. Denote by $y_1,\ldots,y_n$ the vertices of $V_i$, and by $\mathcal Q_j$ the set of umbrellas in $\mathcal D$ incident with $y_j$ for each $j\in\{1,\ldots,n\}$. As $\mathcal D$ does not contain any umbrellas of colour $V_i$, the umbrellas in $\mathcal Q_j$ all have transversals with $i$th component equal to $y_j$. Denote by $Q_j$ the symmetric difference of the umbrellas in $\mathcal Q_j$. We have that $Q_j$ is an octahedral system, according to Proposition~\ref{prop:stable}, and that $\delta_{\Omega}(y_j)=Q_j$, $Q_j\neq\emptyset$, and $Q_j\cap Q_\ell=\emptyset$ for all $j\neq \ell$. According to Lemma~\ref{lem:nonEmpty}, at least one class is covered in $Q_j$ and hence $|Q_j|\geq n$. Therefore, we have $$|\Omega|=\sum_{j=1}^n\deg_{\Omega}(y_j)= \sum_{j=1}^n|Q_j|\geq n^2$$ \end{pf} \begin{proposition}\label{prop:cas2} Consider an octahedral system $\Omega\subseteq V_1\times\cdots\times V_n$ with $|V_i|=n$ for all $i\in\{1,\ldots,n\}$ and a suitable decomposition $(\mathcal U,\Omega_2,\ldots,\Omega_n)$ of $\Omega$. Consider $\mathcal O\subseteq\{\Omega_2,\ldots,\Omega_n\}$ such that for each $\Omega_j\in\mathcal O$ there is a class $V_i$ covered in $\Omega_j$ and in no other $\Omega_\ell\in\mathcal O$. Denote by $\mathcal P\subseteq\mathcal O$ the set of umbrellas in $\mathcal O$. We have $$|\Omega|\geq |\mathcal U|(n-|\mathcal O|)+\sum_{\Omega_j\in\mathcal O}|\Omega_j|-|\mathcal U|(|\mathcal O|-|\mathcal P|)-|\mathcal U|-|\mathcal P|+1.$$ \end{proposition} \begin{pf} Let $W=\triangle_{U\in\mathcal U}U$. The number of edges in $\Omega$ is equal to $\sum_{j=1}^n\deg_\Omega(x_j)$. We bound $\deg_\Omega(x_j)$ by $|\mathcal U|$ for $j=1$ and if $\Omega_j\notin\mathcal O$ and by $|\Omega_j|-|\Omega_j\cap W|$ otherwise, see (iv) in Lemma~\ref{lem:decompo}. We obtain $$|\Omega|\geq|\mathcal U|(n-|\mathcal O|)+\sum_{\Omega_j\in \mathcal O}\left(|\Omega_j|-|\Omega_j\cap W|\right).$$ We introduce a graph $G=(\mathcal V,\mathcal E)$ defined as follows. We use the terminology {\em nodes} and {\em links} for $G$ in order to avoid confusion with the vertices and edges of $\Omega$. The nodes in $\mathcal V$ are identified with the umbrellas in $\mathcal U$ and the $\Omega_j$'s in $\mathcal O$: $\mathcal V=\mathcal U\cup\mathcal O$. There is a link in $\mathcal E$ between two nodes if the corresponding octahedral systems have an edge in common. $G$ is bipartite: indeed, two umbrellas in $\mathcal U$ are of the same colour $V_{i_1}$ and, according to Proposition~\ref{prop:propCol}, they do not have an edge in common. According to Lemma~\ref{lem:decompo}, two $\Omega_j$'s do not have an edge in common either. For $\Omega_j$ in $\mathcal O$, we have $|\Omega_j\cap W|=\sum_{U\in\mathcal U}|\Omega_j\cap U|=\deg_G(\Omega_j)$, note that here the degree is counted in $G$. The fact that the umbrellas in $\mathcal U$ are disjoint proves the first equality. The second equality is deduced from the facts that $\Omega_j$ has at most one edge in common with each umbrella in $\mathcal U$, the one incident with $x_j$, and that $\Omega_j$ has no neighbours in $\mathcal O$. We obtain the following bound \begin{eqnarray*} |\Omega|&\geq &|\mathcal U|(n-|\mathcal O|)+\sum_{\Omega_j\in\mathcal O}\left(|\Omega_j|-\deg_G(\Omega_j)\right)\\ &=& |\mathcal U|(n-|\mathcal O|)+\sum_{\Omega_j\in\mathcal O}|\Omega_j|-\deg_G(\mathcal O\setminus\mathcal P)-\deg_G(\mathcal P). \end{eqnarray*} Again, for the equality, we use the fact that $G$ is bipartite. The number of links in $\mathcal E$ incident with a node in $\mathcal O\setminus\mathcal P$ is at most $|\mathcal U|$. Hence, $\deg_G(\mathcal O\setminus\mathcal P)\leq |\mathcal U|(|\mathcal O|-|\mathcal P|)$. It remains to bound $\deg_G(\mathcal P)$. Note that if $U$ is an umbrella in $\mathcal P$, it is the only umbrella of its colour in $\mathcal P$, otherwise it would contradict the property of $\mathcal O$. We now prove that there are no cycles induced by $\mathcal P\cup\mathcal U$ in $G$. Suppose there is such a cycle $\mathcal C$ and consider an umbrella $U$ of $\mathcal P$ in this cycle. Denote its colour by $V_i$ and its neigbours in $\mathcal C$ by $L$ and $R$. As $G$ is simple, $L$ and $R$ are distinct. $L$ and $R$ are both in $\mathcal U$, and hence are of colour $V_{i_1}$ and do not have an edge in common. Therefore $U\cap L$ and $U\cap R$ do not have an edge in common either, which implies that the $i$th component of the transversals of $L$ and $R$ are distinct. Note that two umbrellas adjacent in $\mathcal C$, both of colour distinct from $V_i$, have necessarily transversals with the same $i$th component. Hence there must be another umbrella of colour $V_i$ in the path in $\mathcal C$ between $L$ and $R$ not containing $U$. This is a contradiction since $U$ is the only umbrella in $\mathcal P$ of colour $V_i$. The number of links in $\mathcal E$ incident with $\mathcal P$ is then at most $|\mathcal U|+|\mathcal P|-1$. This allows us to conclude. \end{pf} \begin{pf}[Proof of Theorem~\ref{thm:main}] Let $\Omega\subseteq{V_1\times\cdots\times V_n}$ be an octahedral system with $|V_1|=\cdots=|V_n|=n\geq 2$, and suppose that $k\geq 1$ classes $V_{i_1},\ldots,V_{i_k}$, with $i_1<\cdots<i_k$, are covered in $\Omega$. The proof works by induction on $k$.\\ If $k=1$, then $\Omega$ must contain at least $n$ edges for one class to be covered.\\ Assume now that $k>1$. If $|\mathcal U|\geq n-1$, then, according to (iv) of Lemma~\ref{lem:decompo}, $|\Omega|=\sum_{j=1}^n\deg_\Omega(x_j)\geq n|\mathcal U|\geq k(n-2)+2$ and we are done. Assume now that $|\mathcal U|\leq n-2$. We consider a suitable decomposition $(\mathcal U,\Omega_2,\ldots,\Omega_n)$ of $\Omega$ and distinguish two cases. Case $1$: {\em One of the covered classes $V_i$, for $i\in\{i_2,\ldots,i_k\}$, is not covered in any $\Omega_j$}. Let $V_i$ be a covered class in $\Omega$, which is not covered in any $\Omega_j$. For each $j\in\{2,\ldots,n\}$, applying Lemma~\ref{lem:elDecompo} on $\Omega_j$ gives a set $\mathcal D_j$ of umbrellas, all of colour distinct from $V_i$, such that $\Omega_j=\triangle_{U\in\mathcal D_j}U$. We obtain $\Omega=(\triangle_{U\in\mathcal U}U)\triangle(\triangle_{j=2}^n\triangle_{U\in\mathcal D_j}U)$, according to (ii) of Lemma~\ref{lem:decompo}. Thus, we can apply Proposition~\ref{prop:cas1} which ensures that $$|\Omega|\geq n^2\geq k(n-2)+2.$$ Case $2$: {\em Each covered class $V_i$, for $i\in\{i_2,\ldots,i_k$\}, is covered in at least one of the $\Omega_j$.} Choose a set $\mathcal O\subseteq\{\Omega_2,\ldots,\Omega_n\}$, minimal for inclusion, such that each covered class $V_i$, for $i\in\{i_2,\ldots,i_k\}$, is covered in at least one of the $\Omega_j\in\mathcal O$. Such a set $\mathcal O$ satisfies the statement of Proposition~\ref{prop:cas2}. Applying this proposition, we obtain $$|\Omega|\geq |\mathcal U|(n-|\mathcal O|)+\sum_{\Omega_j\in\mathcal O}|\Omega_j|-|\mathcal U|(|\mathcal O|-|\mathcal P|)-|\mathcal U|-|\mathcal P|+1.$$ We now bound $\sum_{\Omega_j\in\mathcal O}|\Omega_j|$. Let $k_j$ be the number of classes covered in $\Omega_j$. By minimality of $\mathcal O$, there is at least one class covered in each $\Omega_j\in\mathcal O$, and according to (v) of Lemma~\ref{lem:decompo} we have $k_j<k$, hence $1\leq k_j< k$. By induction, the cardinality of $\Omega_j$ is at least $k_j(n-2)+2$. This lower bound is not good enough for the $\Omega_j\notin\mathcal P$ such that $k_j=1$. We denote by $\mathcal A$ those $\Omega_j$'s. We explain now how to improve the lower bound for $\Omega_j\in\mathcal A$. Only one class is covered in $\Omega_j$ and $\Omega_j\notin \mathcal P$. According to Lemma~\ref{lem:elDecompo}, $\Omega_j$ can be written as a symmetric difference of distinct umbrellas of the same colour. According to Proposition~\ref{prop:propCol}, these umbrellas are pairwise disjoint and $|\Omega_j|$ is equal to $n$ times the number of umbrellas in this decomposition. Since $\Omega_j$ is not an umbrella itself, otherwise $\Omega_j$ would have been in $\mathcal P$, there are at least two umbrellas in this decomposition. We obtain $$\sum_{\Omega_j\in\mathcal O}|\Omega_j|\geq\left(\sum_{\Omega_j\in\mathcal O\setminus\mathcal A}k_j\right)(n-2)+2|\mathcal O\setminus\mathcal A|+2n|\mathcal A|=\left(\sum_{\Omega_j\in\mathcal O}k_j\right)(n-2)+2|\mathcal O|+n|\mathcal A|$$ We have thus $$|\Omega|\geq|\mathcal U|(n-|\mathcal O|)+\left(\sum_{\Omega_j\in\mathcal O}k_j\right)(n-2)+2|\mathcal O|+n|\mathcal A|-|\mathcal U|(|\mathcal O|-|\mathcal P|)-|\mathcal U|-|\mathcal P|+1.$$ \noindent Finally, we have \begin{eqnarray} 2|\mathcal O|-|\mathcal P|-|\mathcal A| &\leq &\sum_{\Omega_j\in\mathcal O}k_j\label{eq:3}\\ k-1 &\leq & \sum_{\Omega_j\in\mathcal O}k_j\label{eq:2} \end{eqnarray} Equation~\eqref{eq:3} is obtained by distinguishing the $\Omega_j$ with $k_j=1$ from those with $k_j\geq 2$. Equation~\eqref{eq:2} results from the fact that each class $V_{i_2},\ldots,V_{i_k}$ is covered in at least one $\Omega_j$ in $\mathcal O$. Thus, \begin{eqnarray*} |\Omega|&\geq& |\mathcal U|(n-|\mathcal O|)+\left(\sum_{\Omega_j\in\mathcal O}k_j\right)(n-2)+2|\mathcal O|+|\mathcal U||\mathcal A|-|\mathcal U|(|\mathcal O|-|\mathcal P|)-|\mathcal U|-|\mathcal P|+1\\ &\geq &(k-1)(n-2)+2|\mathcal O|-|\mathcal P|+1+\left(\sum_{\Omega_j\in\mathcal O}k_j-k+|\mathcal A|+n-2|\mathcal O|+|\mathcal P|\right)|\mathcal U| \end{eqnarray*} \noindent where we only used the inequalities $n\geq n-2\geq|\mathcal U|$ and~\eqref{eq:2}. According to~\eqref{eq:3}, the expression $$\left(\sum_{\Omega_j\in\mathcal O}k_j-k+|\mathcal A|+n-2|\mathcal O|+|\mathcal P|\right)$$ is nonnegative. Moreover, we have already noted that $|\mathcal U|=\deg_\Omega(x_1)$, which is at least $1$. Therefore, $$|\Omega|\geq (k-1)(n-2)+2|\mathcal O|-|\mathcal P|+1+\sum_{\Omega_j\in\mathcal O}k_j-k+|\mathcal A|+n-2|\mathcal O|+|\mathcal P|.$$ Using~\eqref{eq:2} again, we obtain $$|\Omega|\geq k(n-2)+2.$$ \end{pf} \section*{Aknowlegement} The author thanks Antoine Deza for introducing her to the colourful simplicial depth conjecture and Fr\'ed\'eric Meunier for his thorough reading of the manuscript and his helpful comments.

https://arxiv.org/abs/0809.0022

Lagrangians for dissipative nonlinear oscillators: the method of Jacobi Last Multiplier

We present a method devised by Jacobi to derive Lagrangians of any second-order differential equation: it consists in finding a Jacobi Last Multiplier. We illustrate the easiness and the power of Jacobi's method by applying it to several equations and also a class of equations studied by Musielak with his own method [Musielak ZE, Standard and non-standard Lagrangians for dissipative dynamical systems with variable coefficients. J. Phys. A: Math. Theor. 41 (2008) 055205 (17pp)], and in particular to a Liènard type nonlinear oscillator, and a second-order Riccati equation.

\section{Introduction} It should be well-known that the knowledge of a Jacobi Last Multiplier always yields a Lagrangian of any second-order ordinary differential equation \cite{JacobiVD}, \cite{Whittaker}. Yet many distinguished scientists seem to be unaware of this classical result. In this paper we present again the method of the Jacobi Last Multiplier in order to compare the easiness and the power of Jacobi's method with that proposed by Museliak et al \cite{musetal08} for the same purpose. We have already presented the properties of the Jacobi Last Multiplier in \cite{jlm05}. Some of our papers \cite{ennity}-\cite{nuctam_1lag} and the references within may give an idea of the many fields of applications yielded by Jacobi Last Multiplier. In \cite{musetal08} the authors searched for a Lagrangian of the following second-order ordinary differential equation \begin {equation} \ddot x+b(x)\dot x^2+c(x)x=0 \label{eq1} \end{equation} with $b(x),c(x)$ arbitrary functions of the dependent variable $x=x(t)$. After some lengthy calculations they found one Lagrangian. In \cite{nuctam_1lag} we show that (\ref{eq1}) is a subcase of a more general class of equations studied by Jacobi \cite{Jacobi 45}, i.e. \begin{equation} \ddot x+\frac{1}{2}\frac{\partial \varphi}{\partial x}\dot x^2+\frac{\partial \varphi}{\partial t}\dot x+B=0\label{Jeq} \end{equation} with $\varphi, B$ arbitrary functions of $t$ and $x$. We applied Jacobi's method to equation (\ref{eq1}) and showed that many (an infinite number of) Lagrangians can be easily derived. In the present paper we show how to obtain many (an infinite number of) Lagrangians for the class of equations \begin {equation} \ddot x+f(x)\dot x+g(x)=0 \label{2} \end{equation} with $f(x)$ and $g(x)$ arbitrary functions of the dependent variable $x(t)$. With the help of the Jacobi Last Multiplier standard and nonstandard Lagrangians can be derived without much effort. In \cite{musetal08_2} the author applied his lengthy method to (\ref{2}) in order to obtain at least one Lagrangian. This paper is organized in the following way. In section 2, we illustrate the Jacobi Last Multiplier and its properties \cite{Jacobi 42}-\cite{JacobiVD}, its connection to Lie symmetries \cite{Lie1874}, \cite{Lie 12}, and its link to the Lagrangian of any second-order differential equations \cite{JacobiVD}, \cite{Whittaker}. We also exemplify Jacobi's method with an equation \cite{MatLak74} of the class of equations (\ref{eq1}), i.e.: \begin {equation} \ddot x=x\frac{ - a + \lambda \dot x^2}{\lambda x^2 + 1} \label{careq} \end{equation} and a nonautonomous equation \cite{NorMar} of the more general class of equations (\ref{Jeq}), i.e.: \begin {equation} \ddot x=-\frac{\dot x^2}{x}+\frac{\dot x}{t}\label{NMeq}. \end{equation} Both examples were not included in \cite{nuctam_1lag}. In section 3, we apply Jacobi's method to the class of equations (\ref{2}), and show some particular examples such as a Li\`enard type nonlinear oscillator, and a second-order Riccati equation. In section 4, we conclude with some final remarks. In this paper we employ ad hoc interactive programs \cite{man2} written in REDUCE language to calculate the Lie symmetry algebra of the equations we study.\\ \section{The method by Jacobi} The method of the Jacobi Last Multiplier \cite{Jacobi 44 a}-\cite{JacobiVD}) provides a means to determine all the solutions of the partial differential equation \begin {equation} \mathcal{A}f = \sum_{i = 1} ^n a_i(x_1,\dots,x_n)\frac {\partial f} {\partial x_i} = 0 \label {2.1} \end {equation} or its equivalent associated Lagrange's system \begin {equation} \frac {\mbox{\rm d} x_1} {a_1} = \frac {\mbox{\rm d} x_2} {a_2} = \ldots = \frac {\mbox{\rm d} x_n} {a_n}.\label {2.2} \end {equation} In fact, if one knows the Jacobi Last Multiplier and all but one of the solutions, then the last solution can be obtained by a quadrature. The Jacobi Last Multiplier $M$ is given by \begin {equation} \frac {\partial (f,\omega_1,\omega_2,\ldots,\omega_{n- 1})} {\partial (x_1,x_2,\ldots,x_n)} = M\mathcal{A}f, \label {2.3} \end {equation} where \begin {equation} \frac {\partial (f,\omega_1,\omega_2,\ldots,\omega_{n- 1})} {\partial (x_1,x_2,\ldots,x_n)} = \mbox {\rm det}\left [ \begin {array} {ccc} \displaystyle {\frac {\partial f} {\partial x_1}} &\cdots &\displaystyle {\frac {\partial f} {\partial x_n}}\\ \displaystyle {\frac {\partial\omega_1} {\partial x_1}} & &\displaystyle {\frac {\partial\omega_1} {\partial x_n}}\\ \vdots & &\vdots\\ \displaystyle {\frac {\partial\omega_{n- 1}} {\partial x_1}} &\cdots &\displaystyle {\frac {\partial\omega_{n- 1}} {\partial x_n}} \end {array}\right] = 0 \label {2.4} \end {equation} and $\omega_1,\ldots,\omega_{n- 1} $ are $n- 1 $ solutions of (\ref {2.1}) or, equivalently, first integrals of (\ref {2.2}) independent of each other. This means that $M$ is a function of the variables $(x_1,\ldots,x_n)$ and depends on the chosen $n-1$ solutions, in the sense that it varies as they vary. The essential properties of the Jacobi Last Multiplier are: \begin{description} \item{ (a)} If one selects a different set of $n-1$ independent solutions $\eta_1,\ldots,\eta_{n-1}$ of equation (\ref {2.1}), then the corresponding last multiplier $N$ is linked to $M$ by the relationship: $$ N=M\frac{\partial(\eta_1,\ldots,\eta_{n-1})}{\partial(\omega_1, \ldots,\omega_{n-1})}. $$ \item{ (b)} Given a non-singular transformation of variables $$ \tau:\quad(x_1,x_2,\ldots,x_n)\longrightarrow(x'_1,x'_2,\ldots,x'_n), $$ \noindent then the last multiplier $M'$ of $\mathcal{A'}F=0$ is given by: $$ M'=M\frac{\partial(x_1,x_2,\ldots,x_n)}{\partial(x'_1,x'_2,\ldots,x'_n)}, $$ where $M$ obviously comes from the $n-1$ solutions of $\mathcal{A}F=0$ which correspond to those chosen for $\mathcal{A'}F=0$ through the inverse transformation $\tau^{-1}$. \item{ (c) } One can prove that each multiplier $M$ is a solution of the following linear partial differential equation: \begin {equation} \sum_{i = 1} ^n \frac {\partial (Ma_i)} {\partial x_i} = 0; \label {2.5} \end {equation} \noindent viceversa every solution $M$ of this equation is a Jacobi Last Multiplier. \item{ (d) } If one knows two Jacobi Last Multipliers $M_1$ and $M_2$ of equation (\ref {2.1}), then their ratio is a solution $\omega$ of (\ref {2.1}), or, equivalently, a first integral of (\ref {2.2}). Naturally the ratio may be quite trivial, namely a constant. Viceversa the product of a multiplier $M_1$ times any solution $\omega$ yields another last multiplier $M_2=M_1\omega$.\end{description} Since the existence of a solution/first integral is consequent upon the existence of symmetry, an alternate formulation in terms of symmetries was provided by Lie \cite {Lie 12}. A clear treatment of the formulation in terms of solutions/first integrals and symmetries is given by Bianchi \cite {Bianchi 18}. If we know $n- 1 $ symmetries of (\ref {2.1})/(\ref {2.2}), say \begin {equation} \Gamma_i = \sum_{j=1}^{n}\xi_{ij}(x_1,\dots,x_n)\partial_{x_j},\quad i = 1,n- 1, \label {2.6} \end {equation} Jacobi's last multiplier is given by $M =\Delta ^ {- 1} $, provided that $\Delta\not = 0 $, where \begin {equation} \Delta = \mbox {\rm det}\left [ \begin {array} {ccc} a_1 &\cdots & a_n\\ \xi_{1,1} & &\xi_{1,n}\\ \vdots & &\vdots\\ \xi_{n- 1,1}&\cdots &\xi_{n- 1,n} \end {array}\right]. \label {2.8} \end {equation} There is an obvious corollary to the results of Jacobi mentioned above. In the case that there exists a constant multiplier, the determinant is a first integral. This result is potentially very useful in the search for first integrals of systems of ordinary differential equations. In particular, if each component of the vector field of the equation of motion is missing the variable associated with that component, i.e., $\partial a_i/\partial x_i = 0 $, the last multiplier is a constant, and any other Jacobi Last Multiplier is a first integral. Another property of the Jacobi Last Multiplier is its (almost forgotten) relationship with the Lagrangian, $L=L(t,x,\dot x)$, for any second-order equation \begin{equation} \ddot x=F(t,x,\dot x) \label{geno2} \end{equation} is \cite{JacobiVD}, \cite{Whittaker} \begin{equation} M=\frac{\partial^2 L}{\partial \dot x^2} \label{relMLo2} \end{equation} where $M=M(t,x,\dot x)$ satisfies the following equation \begin{equation} \frac{{\rm d}}{{\rm d} t}(\log M)+\frac{\partial F}{\partial \dot x} =0.\label{Meq} \end{equation} Then equation (\ref{geno2}) becomes the Euler-Lagrangian equation: \begin{equation} -\frac{{\rm d}}{{\rm d} t}\left(\frac{\partial L}{\partial \dot x}\right)+\frac{\partial L}{\partial x}=0. \label{ELo2} \end{equation} The proof is given by taking the derivative of (\ref{ELo2}) by $\dot x$ and showing that this yields (\ref{Meq}). If one knows a Jacobi Last Multiplier, then $L$ can be easily obtained by a double integration, i.e.: \begin{equation} L=\int\left (\int M\, {\rm d} \dot x\right)\, {\rm d} \dot x+f_1(t,x)\dot x+f_2(t,x), \label{lagrint} \end{equation} where $f_1$ and $f_2$ are functions of $t$ and $x$ which have to satisfy a single partial differential equation related to (\ref{geno2}) \cite{laggal}. As it was shown in \cite{laggal}, $f_1, f_2$ are related to the gauge function $G=G(t,x)$. In fact, we may assume \begin{eqnarray} f_1&=& \frac{\partial G}{\partial x}\nonumber\\ f_2&=& \frac{\partial G}{\partial t} +f_3(t,x) \label{gf1f2o2} \end{eqnarray} where $f_3$ has to satisfy the mentioned partial differential equation and $G$ is obviously arbitrary. \\ In \cite{Jacobi 45} Jacobi himself found his ``new multiplier'' for the class of second-order ordinary differential equations\footnote{This is not Jacobi's original notation.} studied by Euler \cite{Euler} [Sect. I, Ch. VI, \S\S 915 ff.] (\ref{Jeq}). Indeed Jacobi derived that the multiplier of equation (\ref{Jeq}) is given by: \begin{equation} M=e^{\varphi(t,x)}, \label{JM} \end{equation} as it is obvious from (\ref{Meq}). Consequently in our previous paper \cite{nuctam_1lag} we derived a Lagrangian of the class of equations (\ref{Jeq}) by means of (\ref{relMLo2}), i.e. \begin {equation} L=\frac{1}{2} e^{\varphi(t,x)}\dot x^2+f_3(t,x)+\frac{{\rm d}}{{\rm d}t}G(t,x) \label{LagJ} \end{equation} with $f_3$ a function of $t$ and $x$ satisfying the following equation: \begin {equation} \frac{\partial f_3}{\partial x} +e^{\varphi(t,x)} B(t,x)=0. \end{equation} Equation (\ref{careq}) is a particular example of the equation considered by Jacobi. In fact from (\ref{JM}) and (\ref{LagJ}) we derive: \begin {equation} M=\frac{1}{\lambda x^2 + 1}, \end{equation} and consequently \begin {equation} L=\frac{\dot x^2}{2(\lambda x^2 + 1)} -\frac{a x^2}{2(\lambda x^2 + 1)}+\frac{{\rm d}}{{\rm d}t}G(t,x).\label{Lcareq} \end{equation} This Lagrangian is known \cite{MatLak74}. Equation (\ref{careq}) does not possess any Lie point symmetries apart translation in $t$. Therefore Noether's theorem applied to the autonomous Lagrangian $L$ in (\ref{Lcareq}) yields the following first integral: \begin {equation} I=\frac{a x^2 + \dot x^2}{2(\lambda x^2 + 1)}. \end{equation} Jacobi proved that in the case of a second-order differential equation if one knows a first integral and a last multiplier then the equation can be integrated by quadrature (a new Principle of Mechanics, indeed) \cite{Jacobi 42}, \cite{Jacobi 44 a}. Equation (\ref{NMeq}) is obtained by the symmetry reduction transformation $x=u', t=u$ of the third-order equation: \begin {equation} u'''=-\frac{u'u''}{u}, \end{equation} where $u(T)$ is a function of $T$. Equation (\ref{NMeq}) admits an eight-dimensional Lie point symmetry algebra and therefore is linearizable. In \cite{laggal} it was shown that if one knows several (at least two) Lie symmetries of the second-order differential equation (\ref{geno2}), i.e. \begin {equation} \Gamma_j =V_j(t,x)\partial_t+G_j(t,x)\partial_x, \quad j = 1,r, \label {gensym} \end {equation} then many Jacobi Last Multipliers could be derived by means of (\ref{2.8}), i.e. \begin {equation} {\displaystyle{\frac{1}{M_{nm}}}}=\Delta_{nm} = \mbox {\rm det}\left [ \begin {array} {ccc} 1 &\dot x & F(t,x,\dot x)\\[0.2cm] V_n &G_n &{\displaystyle{\frac{\mbox{\rm d} G_n}{\mbox{\rm d} t} -\dot x\frac{\mbox{\rm d} V_n}{\mbox{\rm d} t}}} \\[0.2cm] V_m &G_m &{\displaystyle{\frac{\mbox{\rm d} G_m}{\mbox{\rm d} t} -\dot x\frac{\mbox{\rm d} V_m}{\mbox{\rm d} t}}}\\ \end {array}\right],\label {Mnm} \end {equation} with $(n,m=1,r)$, and therefore many Lagrangians can be obtained by means of (\ref{lagrint}). In particular, fourteen different Lagrangians can be obtained if the equation admits an eight-dimensional Lie point symmetry algebra. We do not look for the fourteen Lagrangians of equation (\ref{NMeq}). Instead we use equation (\ref{JM}) to find a Jacobi Last Multiplier and consequently a Lagrangian. In fact from (\ref{JM}) and (\ref{LagJ}) we derive: \begin {equation} JLM=\frac{x^2}{t}, \end{equation} and consequently \begin {equation} Lag=\frac{\dot x^2 x^2}{2 t} +\frac{{\rm d}}{{\rm d}t}G(t,x).\label{LNMeq} \end{equation} If one applies Noether's theorem to $Lag$ then the following five first integrals of equation (\ref{NMeq}) can be derived: \begin{eqnarray} FI_1&=& x^2 (x^2 - 2 x \dot x y + \dot x^2 t^2) \nonumber\\ FI_2&=&\frac{x^2 \dot x ( - x + \dot x t)} {2 t} \nonumber\\ FI_3&=&\frac{x^2 \dot x^2}{2 t^2}\label{NMI} \\ FI_4&=& x (x - \dot x t)\nonumber\\ FI_5&=&\frac{x \dot x}{t}\nonumber \end{eqnarray} We underline that the first integrals $FI_1$ and $FI_4$ could not be derived if the gauge function $G(t,x)$ was assumed to be equal to zero. \section{Equations with space-dependent coefficients} The class of equations (\ref{2}) has an obvious Jacobi Last Multiplier and therefore Lagrangian if the following relationship holds between $f(x)$ and $g(x)$: \begin {equation} \frac{{\rm d}}{{\rm d}x}\left(\frac{g(x)}{f(x)}\right)=\alpha(1-\alpha)f(x) \label{3} \end{equation} where $\alpha$ is any constant $\neq 1$. In fact if (\ref{3}) holds then equation (\ref{2}) can be written as \begin {equation} \dot u+\alpha f(x) u=0 \label{ueq} \end{equation} i.e. \begin {equation} f(x)=-\frac{1}{\alpha}\frac{{\rm d}}{{\rm d}t}(\log u),\quad\quad {\rm with}\quad u=\dot x+\frac{g(x)}{\alpha f(x)}, \end{equation} and thus a Jacobi Last Multiplier for equation (\ref{2}) is \begin {equation} M=\exp\left(\int f(x) dt\right) = \exp\left(-\frac{1}{\alpha} \int d(\log u)\right)=u^{-1/\alpha} \label{Ms}\end{equation} and the corresponding Lagrangian is \begin {equation} L=u^{2-1/\alpha}+\frac{{\rm d}}{{\rm d}t}G(t,x)=\left(\dot x+\frac{g(x)}{\alpha f(x)}\right)^{2-1/\alpha}+\frac{{\rm d}}{{\rm d}t}G(t,x), \label{Ls} \end{equation} with $G(t,x)$ an arbitrary gauge function.\\ We note that this Lagrangian is autonomous and therefore it admits at least the Noether point symmetry of translation in $t$, and consequently the following first integral \begin {equation} In=\left(\dot x + \frac{g(x)}{\alpha f(x)}\right)^{1-1/ \alpha}\frac{\alpha f(x)\dot x - f(x)\dot x - g(x)}{\alpha^2 f(x)^2}\label{I2} \end{equation} We would like to remark that because of property (d) of the Jacobi Last Multiplier, then we can obtain another Jacobi Last Multiplier $\overline M=In M$ and consequently another Lagrangian of equation (\ref{2}). We will not pursue it here any further but one can envision a deluge of Lagrangians obtained by simply taking any function of the first integral $In$ in (\ref{I2}) and multiplying it for either $M$ in (\ref{Ms}) or $\overline M$, and so on ad libitum. \subsection{Examples} It is very easy to obtain a Jacobi Last Multiplier and therefore a Lagrangian for the following Li\`enard type nonlinear oscillator: \begin {equation} \ddot x+kx\dot x+\frac{k^2}{9}x^3+\lambda x=0. \label{1} \end{equation} In fact we know that $M=\exp(\int kx dt)$. Then if one can put the equation in the form \begin {equation} \dot u_1+\alpha k x u_1=0,\label{1u1eq}\end{equation} with $\alpha$ a constant to be determined, i.e. \begin {equation} kx=-\frac{1}{\alpha}\frac{{\rm d}}{{\rm d}t}(\log u_1) \end{equation} then \begin {equation} M=u_1^{-1/\alpha}. \end{equation} In the case of equation (\ref{1}) we have \begin {equation} \dot u_1+\frac{1}{3}k x u_1=0, \quad \quad {\rm with}\quad u_1=\dot x +\frac{k}{3}x^2+\frac{3}{k}\lambda, \quad ({\rm i.e.,}\quad \alpha=1/3).\end{equation} Therefore $M_1=u_1^{-3}$ and consequently\footnote{We do not take into consideration any nonessential multiplicative constant.}: \begin {equation} L_1=\frac{1}{u_1}+\frac{{\rm d}}{{\rm d}t}G(t,x)=\frac{1}{\dot x +\frac{k}{3}x^2+\frac{3}{k}\lambda}+\frac{{\rm d}}{{\rm d}t}G(t,x). \end{equation} Actually, we can derive another Lagrangian because substituting $f(x)=kx, g(x)=\frac{k^2}{9}x^3+\lambda x $ into equation (\ref{3}) yields two different $\alpha$, i.e.: \begin {equation} \frac{{\rm d}}{{\rm d}t}\left(\frac{g}{f}\right)-\alpha(1-\alpha)f=0\quad \Longrightarrow \quad 9\alpha^2 - 9\alpha + 2=0\quad \Longrightarrow \quad \alpha_{1,2}=\frac{1}{3}, \frac{2}{3} \end{equation} The case $\alpha=\frac{1}{3}$ has been considered above. If we substitute $\alpha=\frac{2}{3}$ into equation (\ref{ueq}) then we obtain $M_2=u^{-3/2}$ and consequently \begin {equation} L_2=\sqrt{u}+\frac{{\rm d}}{{\rm d}t}G(t,x)=\sqrt{\dot x +\frac{k}{6}x^2+\frac{3}{2 k}\lambda}+\frac{{\rm d}}{{\rm d}t}G(t,x). \end{equation} Equation (\ref{1}) admits an eight-dimensional Lie symmetry algebra and therefore is linearizable. Moreover one can determine at least twelve more Lagrangians \cite{CP07jlmmech}. We note that $L_1$ admits one Noether point symmetry while $L_2$ admits three Noether point symmetries. A particular case of (\ref{1}) is the second-order Riccati equation: \begin {equation} \ddot x+3x\dot x+x^3=0, \label{riceq} \end{equation} a member of the Riccati-chain \cite{Ames}. Equation (\ref{riceq}) is linearizable to a third-order linear equation by the transformation $x=\dot V(t)/V(t)$, namely (\ref{riceq}) transforms into $\ddot V=0$. It also well-known that equation (\ref{riceq}) is linearizable by means of a point transformation because it admits an eight-dimensional Lie symmetry algebra generated by the following operators: \begin{eqnarray} \Gamma_1&=&t^3(t x-2)\partial_t-t(x t-2) (x^2 t^2+2-2 x t)\partial_x\nonumber\\ \Gamma_2&=&x t^3\partial_t-(x t-1) (x^2 t^2+4-2 x t)\partial_x\nonumber\\ \Gamma_3&=&x t^2\partial_t-x (x^2 t^2+2-2 x t)\partial_x\nonumber\\ \Gamma_4&=&x t\partial_t-x^2 (x t-1)\partial_x\nonumber\\ \Gamma_5&=&x\partial_t-x^3\partial_x\\ \Gamma_6&=&\partial_t\nonumber\\ \Gamma_7&=&t\partial_t-x\partial_x\nonumber\\ \Gamma_8&=&t^2\partial_t-2(x t-1)\partial_x\nonumber \end{eqnarray} In order to find the linearizing transformation we have to look for a two-dimensional abelian intransitive subalgebra \cite{Lie 12}, and, following Lie's classification of two-dimensional algebras in the real plane \cite{Lie 12}, we have to transform it into the canonical form $$\partial_{\tilde x},\;\;\;\;\;\tilde t\partial_{\tilde x}$$ with $\tilde t$ and $\tilde x$ the new independent and dependent variables, respectively. We found that one such subalgebra is that generated by $\Gamma_1$ and $\Gamma_9\equiv\Gamma_2-\Gamma_8$. Then, it is easy to derive that $$ \tilde t= \frac{tx-1}{x(t x-2)},\;\;\;\;\;\tilde x=-\frac{x}{2t(t x-2)}$$ and equation (\ref{riceq}) becomes \begin{equation} {{\rm d}^2 \tilde x \over {\rm d}\tilde t^2}=0 \end{equation} We can derive fourteen different Lagrangians by using (\ref{Mnm}) and (\ref{lagrint}). Two of these Lagrangians admit five Noether symmetries, i.e.: \begin {equation} L_{56}=-\frac{1}{2(\dot x+x^2)}+\frac{{\rm d}}{{\rm d}t}G(t,x) \label{rL56} \end{equation} and \begin {equation} L_{19}=-\frac{1}{2 t^4 (x^2 t^2+\dot x t^2-2 x t+2)}+\frac{{\rm d}}{{\rm d}t}G(t,x). \end{equation} which are derived from \begin {equation} JLM_{56}=-\frac{1}{(\dot x +x^2)^3}, \end{equation} and \begin {equation} JLM_{19}=-\frac{1}{(t^2 x^2+t^2\dot x-2t x+2)^3}, \end{equation} respectively. We remark that $JLM_{56}$ can be also obtained from (\ref{ueq}). In fact equation (\ref{riceq}) can be written as \begin {equation} \dot u+ 3 x \alpha u=0, \quad \quad u=\dot x+x^2,\quad \alpha=\frac{1}{3}. \end{equation} If one applies Noether's theorem to $L_{56}$ then the following five first integrals of equation (\ref{riceq}) can be derived: \begin{eqnarray} I_1&=&\frac{(x^2 t^2 - 2 x t + \dot x t^2 + 2)^2}{4 (x^2 +\dot x)^2} \nonumber\\ I_2&=&\frac{(x^2 t^2 - 2 x t + \dot x t^2 + 2) (x^2 t - x +\dot x t)} {2 (x^2 + \dot x)^2} \nonumber\\ I_3&=&\frac{x^2 + 2 \dot x}{2 (x^2 + \dot x)^2}\\ I_4&=&\frac{(x^2 t - x + \dot x t)^2}{(x^2 + \dot x)^2}\nonumber\\ I_5&=&\frac{x^2 t - x +\dot x t}{x^2 + \dot x}\nonumber \label{ricI} \end{eqnarray} while Noether's theorem applied to $L_{19}$ yields: \begin{eqnarray} In_1&=&\frac{x^2 t - x + \dot x t}{x^2 t^2 - 2 x t + \dot x t^2 + 2}\nonumber\\ In_2&=&\frac{(x^3 t - 2 x^2 - 2 \dot x) x t + (\dot x t^2 + 4) \dot x + (2 \dot x t^2 + 3) x^2}{(x^2 t^2 - 2 x t + \dot x t^2 + 2)^2}\nonumber \\ In_3&=&\frac{(x^2 t - x + \dot x t) (x^2 + \dot x)}{(x^2 t^2 - 2 x t + \dot x t^2 + 2)^2} \\ In_4&=&\frac{(x^2 + \dot x)^2}{(x^2 t^2 - 2 x t + \dot x t^2 + 2)^2}\nonumber \\ In_5&=&\frac{x^2 + 2 \dot x}{2 (x^2 t^2 - 2 x t + \dot x t^2 + 2)^2}\nonumber\label{ricIn} \end{eqnarray} We remark the importance of the gauge function $G(t,x)$. None of the first integrals above, apart $I_3$ and $In_5$, could be derived if the gauge function was assumed to be equal to zero. \section{Final remarks} The Lagrangian which admits the maximum number of Noether point symmetries is that obtained by means of the Jacobi Last Multiplier which comes from the determinant (\ref{Mnm}) with the two solution symmetries, namely the two-dimensional abelian intransitive subalgebra which yields the linearizing transformation. We may infer that this is the physical Lagrangian. Also the Lagrangian obtained by using the Jacobi Last Multiplier (\ref{Ms}) possesses the maximum number of Noether point symmetries as shown in the case of equation (\ref{3}). This may explain why Lagrangian (\ref{Ls}) possesses nice physical properties as shown in \cite{Caretal05}. In the present paper, we do not claim to have been exhaustive in our presentation of the application of the Jacobi Last Multiplier for finding Lagrangians of any second-order differential equation. Indeed we would like to encourage other authors to apply Jacobi's method to their preferred equation. \section*{Acknowledgements} This work was initiated while K.M.T. was enjoying the hospitality of Professor M.C. Nucci and the facilities of the Dipartimento di Matematica e Informatica, Universit\`a di Perugia. K.M.T. gratefully acknowledges the support of the Italian Istituto Nazionale Di Alta Matematica ``F. Severi'' (INDAM), Gruppo Nazionale per la Fisica Matematica (GNFM), Programma Professori Visitatori.

https://arxiv.org/abs/1710.10875

Shifts of the prime divisor function of Alladi and Erdős

We introduce a variation on the prime divisor function $B(n)$ of Alladi and Erdős, a close relative of the sum of proper divisors function $s(n)$. After proving some basic properties regarding these functions, we study the dynamics of its iterates and discover behaviour that is reminiscent of aliquot sequences. We prove that no unbounded sequences occur, analogous to the Catalan-Dickson conjecture, and give evidence towards the analogue of the Erdős-Granville-Pomerance-Spiro conjecture on the pre-image of $s(n)$.

\section{Introduction} Let $n$ be a positive integer with prime factorization $n=p_1^{r_1}\dots p_k^{r_k}$. Consider the sum of prime divisors function $B(n) = \sum_{i=1}^kr_ip_i$ and the sum of distinct prime divisors function $\beta(n) = \sum_{i=1}^k p_i$. They can be viewed as variants of the sum of proper divisors function $s(n)=\sum_{d||n}d$, discussed since Pythagoras. The arithmetic properties of $B(n)$, for example, were studied by Alladi and Erd\"os \cite{AE}, and $\beta(n)$ studied by Hall \cite{H} since the 1970s.\footnote{In \cite{AE}, the authors denote $B(n),\beta(n)$ by $A(n), A^*(n)$. In this paper we follow the former convention, consistent with more recent literature.} These are large additive functions in the sense that they have the same average order as the largest prime factor of $n$, which is expected by the well-known result of Hardy and Ramanujan that $\Omega(n)$ and $\omega(n)$ have the same average order $\log\log n$. In this paper, we introduce perturbations of $\beta(n)$ and $B(n)$, by shifting the values at certain fixed points. Clearly $B(n)=\beta(n)=n$ if $n$ is prime. Then define for a fixed positive integer $n$, the functions \be B_a(n) = \begin{cases} n+a & n \text{ prime}\\ B(n) & n \text{ otherwise} \end{cases} \ee and similarly $\beta_a(n)$. They are no longer additive, and behave in similar and different ways in comparison to the original functions, as we discuss in Section \ref{shifts}. Our main interest will lie in their iterates, denoting $B^2_a(n)= B_a(B_a(n))$, and similarly for $\beta_a(n)$. The aliquot sequence $n, s(n), s(s(n)),\dots$ stops at either primes, or cycles of length two (amicable pairs) or longer (sociable numbers). Iterating $B_a(n)$ and $\beta_a(n)$, we encounter similar phenomena, detailed in Section \ref{iterates}. Surprisingly, for fixed $a$ only a handful of cycles are observed; it is in varying $a$ we find many different cycles. This suggests a version of the Catalan-Dickson conjecture, which states that there does not exist unbounded aliquot sequences (or, alternatively, the Guy-Selfridge counter-conjecture \cite{GS} which gives certain candidate counterexamples). It is obvious that $B_0^k(n)$ and $\beta_0^k(n)$ all eventually reach a fixed point, but for general $a$ there may be exceptional cases in which the sequences escape to infinity. In support of this, the Green-Tao theorem guarantees that there exists integers $n$ and $a$ such that \be n<B_a(n) < \dots < B_a^k(n), \text{ and }\ n<\beta_a(n) < \dots < \beta_a^k(n). \ee that for any $k>0$ (Corollary \ref{GT}). Nonetheless, we show in Theorem \ref{finite}, that in fact no unbounded sequences occur. \begin{thm} There exists finitely many cycles for each $a$, and all integers iterate to cycles. \end{thm} We conclude with evidence towards a variant of a conjecture of Erd\H{o}s, Granville, Pomerance, and Spiro \cite{EGPS}, which states that the pre-image of any set of asymptotic density zero under $s(n)$ is again of asymptotic density zero. Namely, we prove as Theorem \ref{EGPS}, using previous work of Pollack, Pomerance, and Thompson \cite{PPT} on divisor-sum fibres: \begin{thm} Let $\epsilon=\epsilon(x)$ be a fixed function tending to 0 as $x\to\infty$, and let $\mathscr A$ be a set of at most $x^{\frac12+\epsilon(x)}$ positive integers. Then \be \#\{n\leq x : B(n)\in\mathscr A\} = o_\epsilon(x) \ee as $x\to\infty$, uniformly in the choice of $\mathscr A$. \end{thm} \subsubsection*{Acknowledgments} The authors would like to thank Jeff Lagarias for suggesting the proof of Theorem \ref{finite}. \section{Shifts} \label{shifts} In this section, we prove some basic results of $B_a(n)$ and $\beta_a(n)$. First, from the fact that $B(n^k)=kn$ it follows that if $n$ is composite then $B_a(n^k)=kn$. Also, $B_a(n)$ is close to additive in the sense that $B_a(mn)=B_a(n)+B_a(m)$ if and only if $m,n$ are not prime. The following result shows that $B_a(n)$ is in certain ways similar to $B_0(n)$, independently of $a$. \begin{prop} \label{order} (a) The average order $B_a(n)$ is $\dfrac{\pi^2 n}{6\log n}$. In other words, \be \sum_{n\leq x} B_a(n) \sim \frac{\pi^2 x^2}{12\log x}. \ee (b) The average order of $B_a(n) - \beta_a(n)$ is $\log\log n$. In other words, \be \sum_{n\leq x}(B_a(n) - \beta_a(n)) = x \log\log x + O(x). \ee (c) For any fixed integer $N$, the set $\#|B_a(n) - \beta_a(n) = N|$ has positive natural density. In other words, \be \lim_{n\to\infty} \frac{n}{B_a(n) - \beta_a(n)} > 0. \ee \end{prop} \begin{proof} These follow simply from the fact that \be \sum_{n\leq x} B_a(n) = \sum_{n\leq x} B(n) + a \sum_{p\leq x} 1 \ee and similarly for $\beta_a(n)$, then applying the corollaries to \cite[Theorem 1.1]{AE}. The additional sum over $p$ is $a\pi(x)$, which gives a smaller error $ax/\log x$. For the last two claims, we also note that $B_a(n) - \beta_a(n)=B(n) - \beta(n).$ \end{proof} Let $f(n)$ be an arithmetic function taking positive integer values. Then for any fixed $N\geq 0$, the local density of $f(n)$ is defined to be \be d_N = \lim_{x\to \infty} \frac{1}{x}\#|n\leq x : f(n) = N|. \ee Then we may deduce the following result from \cite[\S3]{Iv}: \begin{cor} There exists $d_N$ such that \be \#\{n\leq x : B_a(n)-\beta_a(n) = N\} = d_N x + O(x^\frac12\log x) \ee and \be \#\{n\leq x : B_a(n)-\beta_a(n) \geq N\} \ll \frac{x}{N}. \ee uniformly in $N>0$. \end{cor} The next theorem shows a departure from the unshifted case. In fact, we find a qualitative difference depending on whether $a$ is 0, even, or odd. Recall that in the case of $a=0$, we have that $B_0(n)$ and $\beta_a(n)$ are additive, and therefore uniformly distributed for all $q$ by Delange's theorem (c.f. \cite{G}). \begin{thm} \label{dist} (a) Let $q=2$. Then $B_a(n)$ is uniformly distributed \textnormal{mod} $q$ if $a$ is even. That is, there is a constant $c_0>0$ such that \be \sum_{n=1}^\infty (-1)^{B_a(n)} = O(x\exp(c_0(\log x)^\frac23). \ee If $a$ is odd, we have instead $O(x)$, as $x\to\infty$. (b) Let $q,h$ be fixed integers with $q>2$. Then for any $a$ such that $q|ha$, $B_a(n)$ is uniformly distributed mod $q$. \end{thm} \begin{proof} First consider $q=2$. Notice that $(-1)^{B_a(p)}= (-1)^{B(p)}(-1)^a$, so the even case follows directly from \cite[Theorem 3.1]{AE}, so it suffices to consider the case with $a$ odd. We have the Dirichlet series \be \sum_{n=1}^\infty\frac{(-1)^{B_a(n)}}{n^s} = \sum_{n=1}^\infty\frac{(-1)^{B(n)}}{n^s} - 2\sum_{p} \frac{1}{p^s} = \frac{2^s+1}{2^s-1}\frac{\zeta(2s)}{\zeta(s)} - 2P(s) \ee where $P(s)=\sum_{p} p^{-s}$ denotes the so-called prime zeta function, which is essentially $\log\zeta(s)$ as $s\to1^+$. The claim for $q>2$ follows from a recent result of Goldfeld: by assumption the character $\exp(2\pi i \frac{hB_a(n)}{q}) = \exp(2\pi i \frac{hB(n)}{q})$ is completely multiplicative, and thus we may apply \cite[Theorem 1.2]{G} to get \be \sum_{\substack{n\leq x\\ B(n) \equiv h\textnormal{ mod } q}}1 = \frac{x}{q} + O\left(\frac{x}{(\log x)^\frac12}\right) \ee as $x\to\infty$, which gives the result. \end{proof} From the above we observe that the lack of additivity leads to obstructions to uniform distribution, which in turn leads to a departure from certain proofs related to iterates of $B_a(n)$ and $\beta_a(n)$, which we will discuss next. Nonetheless, we can conclude: \begin{cor} If $a$ is even, then \be \sum_{n=1}^\infty\frac{(-1)^{B_a(n)}}{n}=0. \ee \end{cor} \section{Iterates} \label{iterates} Now we turn to the dynamics of iterates of $B_a(n)$ and $\beta_a(n)$. While most of the results below concern $B_a(n)$, many of them can be carried over to $\beta_a(n)$ without difficulty, which we leave to the interested reader. Throughout we will also raise several problems regarding the sequences $B^k_a(n)$, which can also be posed for $\beta^k_a(n)$, suggested by the numerical computations. \subsection{Cycles} \label{cycles} We call $n$ a periodic point if $B_a^{k+l}(n)=B_a^k(n)$ for some $l$, and eventually periodic if $B_a^k(n)$ is periodic for some $k$. Define a cycle to be the orbit of a periodic point $n$. We will sometimes refer to the fixed points $B_a(n)=n$ as trivial cycles. \begin{lem} For any $a$ nonzero, $B_a(n) = n$ if and only if $n=4$. \end{lem} \begin{proof} Primes are not fixed points by definition. So the only possible fixed points are those of $B(n)$ for $n$ composite, which is $n=4$, as $B(n)<n$ for any composite $n>4$. \end{proof} \subsubsection{The first case} Recall that the stopping time for $n$, denoted $\sigma(n)$ is defined to be $\inf\{k: B^k_a(n)<n\}$. We will also define the total stopping time for $n$, denoted $\sigma_\infty(n)$ to be the number of iterates required for $B^k_a(n)$ to enter a cycle. \begin{prop} $B_1^k(n)$ is eventually periodic for all $n$, with cycles $(4)$ and $(5,6)$. \end{prop} \begin{proof} This is easy to check the orbits for small $n$, say $n\leq 6$, so we may assume that $n>6$. We first claim that $B^k_1(n)$ has stopping time $\sigma(n)=2$ if $n$ is prime, otherwise $\sigma(n)=1$. It suffices to check for $n=p$. Set $p+1=2m$. Then \be B^2_1(p)=B_1(p+1)=2 + B_1(m)\leq 2+m < p. \ee It follows from the claim that $B^k_1(n) < n$ for all $k>1$, thus $B^{2k}_1(n)$ is strictly decreasing until it reaches the cycle $(5,6)$. \end{proof} \begin{rem} The proof above fails for $a>1$, since $p+a$ may be prime if $a$ is even, while the final inequality $2+m<p$ is no longer guaranteed if $a$ is odd. \end{rem} \subsubsection{Amicable pairs} Do all primes above 3 occur in some amicable pair, i.e., 2-cycle for some $a$? We can answer this in the affirmative. \begin{prop} \label{amicable} Every $p>3$ occurs in a $2$-cycle $(p,B_a(p))$ for some integer $a$. \end{prop} \begin{proof} We have to produce an $a$ such that $B_a(p+a)=p$. This forces $p+a$ to be composite, so it suffices to ask if $B_0(p+a)=p$ for some $a$. Let $n$ be a composite solution to $B_0(n)=p$. Then setting $a=n-p$, we have that $B_a: p \mapsto n \mapsto p$ as desired. It remains to show that $B_0(n)=p$ always has a composite solution. Let $q$ be the largest prime less than $p$. If $p-q$ is prime, then choose $n=q(p-q)$. If not, then let $a$ be a prime dividing $p-q$, and write $p-q=ab$. Then choose $n=qa^b$. \end{proof} \begin{rem} Observe that if $a$ was chosen to be the largest prime dividing $p-q$, then the construction will in fact produce the smallest composite solution $n$. \end{rem} \subsubsection{Finite cycles} How many kinds of cycles appear for fixed $a$? Up to $a\leq 200$, we found at most 4 distinct nontrivial cycles, at $a=39$, we have: \[ (43, 82),(13,52,17,56),(7,46,25,10),(5, 44, 15, 8, 6) \] Note that from the above, we see that different $k$-cycles can occur for a fixed $k$ and $a$. It is natural to consider a variant of the Catalan-Dickson conjecture for $s(n)$: Does the sequence $B^k_a(n)$ ever escape to infinity as $k\to\infty$? Heuristically, the sequence $B_a(n)/n$ has average order $\pi^2/6\log n < 1$ by Proposition \ref{order}(a),\footnote{In contrast to $s(n)/n$ which is slightly greater than 1 on average.} which suggests that an unbounded sequence should not exist. We show that this is indeed true: \begin{thm} \label{finite} There exists finitely many cycles for each $a$, and all integers iterate to cycles. \end{thm} \begin{proof} It suffices to show that for any shift $a$ that there is a bound $C(a)$ such that all orbits of iterating $B_a$ from any starting point $n$ enter the range $[1, C(a)]$. Without loss of generality take $n$ be a prime $p$. We will show that for $p$ large enough, the next iterate of $p$ which goes downhill will go to a number less than $p$. Let $s$ be the smallest prime that does not divide $a$, which is less than $2a$ by Bertrand's postulate. Then $a$ mod $s$ is nonzero, hence relatively prime to it. On the other hand, at least one of $p+a,p+2a,\dots, p+sa$ is divisible by $s$, and therefore composite. Denote it by $p+ka$. We claim that for composite $n$, $B_a(n) \le 2 + \frac{n}{2}$. Assuming this, then whenever $p$ is such that \be B_a(p+ka)\le 2+ \frac{p+ sa}{2}< p \ee then we will go downhill. Now since $s < 2p$, we may choose $C(a)=2a^2 + 10 < p.$ It remains to prove the claim. Since $n$ is composite, it suffices to prove it for $a=0$. Let $n = p^kq$ where $p$ is the smallest prime factor of $n$, with $(p,q)=1, q>1$. (If $q=1$ then $B(p^k) = kp$ and we are done.) By additivity, we have then: \begin{align} B(n) &= kp + B(q) \\ & \le kp + q \\ & \le 2 + (\frac{p^k}{2}-1)q + q = 2 + \frac{n}{2}. \end{align} The last inequality follows since $2 + (\frac{p^k}{2}-1)q \ge p^k$, and $q >1$. \end{proof} The following table lists the distinct nontrivial cycles found for small $a$ and checking $n$ up to $10^6$. \begin{table}[ht!] \centering \caption{Nontrivial cycles for $n\leq 10^6$.} \label{tab:table1} \begin{tabular}{c| l | c| l} \toprule $a$ & cycles & $a$& cycles \\ \midrule 1 & $(5,6)$ &11&$(5,15,8,6)$\\ 2 & $(5,7,9,6)$ &12&$(5, 17, 29, 41, 53, 65, 18, 8, 6)$\\ 3 & $(5,8,6),(7,10)$ &13&$(5,16,8,6)$\\ 4 & $(5,9,6)$ &14&$(5, 19, 33, 14, 9, 6), (7, 21, 10)$ \\ 5 & $(7,12)$ &15&$(5,20,9,6),(19,34)$\\ 6 & $(7,13,19,25,10)$ &16&$(7, 23, 39, 16, 8, 6, 5, 21, 10)$\\ 7 & $(5,12,7,14,9,6)$ &17&$(7, 24, 9, 6, 5, 22, 13, 30, 10),(11,28)$\\ 8 & $(5,13,21,10,7,15,8,6)$ &18&$(5, 23, 41, 59, 77, 18, 8, 6),(7, 25, 10)$ \\ 9 & $(5,15,9,6),(13,22)$ &19&$(5, 24, 9, 6)$ \\ 10 & $(5,15,8,6)$ &20&$(5, 25, 10, 7, 27, 9, 6)$\\ \bottomrule \end{tabular} \end{table} \subsubsection{Cycle length} What are the lengths of cycles, and how do they depend on $a$ and $n$? What are the stopping times $\sigma(n)$ and $\sigma_\infty(n)$? We have not studied this question in detail, but numerical experiments suggest that both times are small, relative to $s(n)$, for example. \subsubsection{Sign patterns} Any $k$-cycle $(n,B_a(n),\dots,B^k_a(n))$ can be ordered so that $n$ is the least term in the sequence, making $n$ prime and $B^k_a(n)$ composite. We adopt the following notation: we will assign $\{+,-\}$ to denote in a cycle whether a number is prime or composite. For example, the cycle $(5,7,9,6)$ in $a=2$ has sign pattern $(+,+,-,-)$. We can now pose the following question: what are the possible sign patterns allowed in a $k$-cycle? All non-trivial cycles of length $k>2$ must have sign pattern of the form $(+,\dots,-)$. Do all combinations occur in between? For example, with $k=3$ we find both combinations $(+,+,-)$ and $(+,-,-)$ to occur. \subsection{Ascending chains} Looking in the other direction, a number $n$ is called abundant if the sum of proper divisors $s(n)>n$, and it was shown in an unpublished work of Lenstra, and later improved by Erd\H{os} \cite{E}, that for every $K$ there is an $n$ for which \be \label{abundantchain} n<s(n) < \dots < s^K(n). \ee Our $B_a(n)$ are simpler in the sense that $B_a(n)>n$ if and only if $n$ is prime and $a>1$, but on the other hand the analog of Lenstra's result requires the Green-Tao theorem \cite[Theorem 1.1]{GT}, which implies that there are arbitrarily long arithmetic progressions of primes, since such a progression yields primes $B_a^k(p) = p+ka$. Nonetheless we can immediately conclude: \begin{cor} \label{GT} For any $k>0$, there exists integers $n$ and $a$ such that \be n<B_a(n) < \dots < B_a^k(n), \text{ and }\ n<\beta_a(n) < \dots < \beta_a^k(n). \ee \end{cor} \begin{rem} Note that there is a natural extension of $B_a(n)$ to $\ensuremath{{\bf Z}}_{\geq0}$ by setting $B(1)=1$ and $B(0)=0$, and to negative integers by setting $B(-n)=B(n)$, similarly for $\beta_a(n)$. Thus iterating our functions can be viewed as studying dynamics on $\ensuremath{{\bf Z}}$ itself. We may also extend to ${\bf Q}$ by defining $B(\frac{x}{y}):=B(x)-B(y)$ for reduced fractions, though upon iterating once we return to $\ensuremath{{\bf Z}}$, and from there $\ensuremath{{\bf Z}}_{\geq0}$. One possible way of producing more interesting extensions is by setting $B(-n) = -B(n)$, and $B(\frac{x}{y})=B(x)/B(y).$ \end{rem} \subsection{Prime-divisor fibres} A question related to cycles is: For a fixed $p$, what is the set of solutions $\#\{n : B_a(n)=p\}$? The solution sets are the same for every $a$, except possibly the pre-image $\{p-a\}$, which will be counted if it is prime. From the proof of Theorem \ref{amicable} we have already found the smallest composite solution. In fact, the solutions to $B_a(n) = m$ for a fixed $m$ are given precisely by the prime partitions $\kappa(m)$ of $m$, as was already observed in \cite[Theoren 2.7]{J}, and there is at least one composite solution for all $n\ge 5$. From this fact we can immediately deduce from \cite[VIII.26]{FS} the following asymptotic: \begin{prop} We have \be \log(\#\{n : B_a(n)=m\}) \sim 2\pi\sqrt{\frac{m}{3\log m}}. \ee as $m\to\infty$. \end{prop} \begin{rem} Indeed, one even has a recursive definition for $\kappa(n)$ in terms of $\beta(n)$, \be \kappa(n)=\frac{1}{n}\big(\beta(n)+\sum_{i=1}^{n-1}\kappa(n-i)\beta(i)\big). \ee with the initial condition $\kappa(1)=0$. In other words, the number solutions of $B(n)=m$ are determined by the values of $\beta(i)$ for $i\leq m$. \end{rem} More generally, this can be phrased in terms of prime-divisor sum fibres, with reference to \cite{PPT}: Let $\mathscr{A}\subset {\bf N}$ be a set of asymptotic density zero (for example, the set of prime numbers). What is the preimage $B^{-1}_a(p)$ for a fixed $a,p$? Erd\H{o}s, Granville, Pomerance, and Spiro (EGPS) conjecture that the fibre $s^{-1}(\mathscr{A})$ also has asymptotic density zero. The following theorem shows that evidence towards the analogous conjecture also holds for $B(n) = B_0(n)$. \begin{thm} \label{EGPS} Let $\epsilon=\epsilon(x)$ be a fixed function tending to 0 as $x\to\infty$, and let $\mathscr A$ be a set of at most $x^{\frac12+\epsilon(x)}$ positive integers. Then \be \#\{n\leq x : B(n)\in\mathscr A\} = o_\epsilon(x) \ee as $x\to\infty$, uniformly in the choice of $\mathscr A$. \end{thm} \begin{proof} Let $a\in\mathscr A$, and $n$ a non-exceptional preimage of $a$ in the sense of \cite[\S2]{PPT}. Write $n=de$ where $d$ is the largest divisor of $n$ not exceeding $\sqrt{x}$. Then following the proof of \cite[Theorem 1.2]{PPT}, we observe firstly that from $B(n)\leq s(n)$ that \be B(n) \ll x^{\frac12-10\epsilon(x)}\log x. \ee Secondly, $B(de) = B(d) + B(e)$ let $g=\gcd(B(d),B(e))$, then \be \frac{a}{g} \equiv \frac{B(e)}{g} \pmod{B(d)/g}, \ee so that given $d$ we see that $B(e)$ lies in a uniquely determined residue class mod $B(d)/g$. Hence the number of choices of $B(e)$ given $d$ is \be \ll 1+x^{\frac12-10\epsilon(x)}\log x\frac{g}{B(d)}\le 1+x^{\frac12-10\epsilon(x)}\frac{\log x}{\log d} \ee where the second inequality follows from $B(d) \ge \log d $. Now we sum over possible values of $g$ and $d$. We can write each $d$ as $gh$ where $h\leq x^{1/2-10\epsilon(x)}/g$. Thinking of $g$ as fixed and summing on $d = gh$ gives a bound of $\ll (\log x)^2x^{1/2-10\epsilon(x)}$. Whereas summing over the $\tau(a)$ divisors $g$ of $a$ bounds the number of possibilities for $n$ by \be \ll \tau(a)(\log x)^2x^{1/2-10\epsilon(x)}<x^{1/\log\log x}x^{1/2-10\epsilon(x)}\leq x^{1/2-9\epsilon(x)}, \ee bounding the number of $n$ that arise in this way as desired. \end{proof} \begin{rem} Since $B_a(n)\leq s(n)$ is no longer true for $a>1$, it remains to ask whether the above holds for general $a$, and if moreover the EGPS conjecture should also hold. \end{rem}

https://arxiv.org/abs/2107.06965

Connections Between Finite Difference and Finite Element Approximations

We present useful connections between the finite difference and the finite element methods for a model boundary value problem. We start from the observation that, in the finite element context, the interpolant of the solution in one dimension coincides with the finite element approximation of the solution. This result can be viewed as an extension of the Green function formula for the solution at the continuous level. We write the finite difference and the finite element systems such that the two corresponding linear systems have the same stiffness matrices and compare the right hand side load vectors for the two methods. Using evaluation of the Green function, a formula for the inverse of the stiffness matrix is extended to the case of non-uniformly distributed mesh points. We provide an error analysis based on the connection between the two methods, and estimate the energy norm of the difference of the two solutions. Interesting extensions to the 2D case are provided.

\section{Introduction} When studying basic numerical methods for solving boundary value problems (BVP) many resources start with a Finite Difference (FD) approach followed by a separate Finite Element (FE) approach. In this paper, we adopt a new point of view that emphasizes on the connections between the FE and FD methods by solving a standard two point boundary value problem discretized on the same nodes through both FD and the FE methods. We present the connections that help simplify certain proofs for FD error approximation, and also, that lead to a better understanding of the advantages of each of these two methods. Consider the two-point boundary value problem \begin{equation}\label{2pBVP} -u''(x) =f(x), \ x \in (0, 1), \ u(0)=u(1)=0. \end{equation} It is known that, if $u \in C^2([0, 1])$ is the unique solution of \eqref{2pBVP}, then \begin{equation}\label{GF} u(x) =\int_0^1 G(x,s) f(s)\, ds, \ \text{for all} \ x \in(0,1), \end{equation} where, \begin{equation}\label{Gfunc} G(x,s) = \left\{ \begin{array}{rcl} s(1-x) & \mbox{if } \ 0\leq s\leq x,\\ x(1-s) & \mbox{if } \ x<s\leq 1.\\ \end{array} \right. \end{equation} For the discretization of \eqref{2pBVP}, we divide the interval $[0,1]$ into $n$ subintervals, using the nodes $0=x_0<x_1<\cdots < x_n=1$ and denote \\ $h_j:=x_j - x_{j-1}, j=1, 2, \cdots, n$. First, consider the finite difference approximation of $u''(x_j)$ that uses the quadratic polynomial interpolation of the solution $u$ at three nodes: $x_{j-1}, x_j$, and $x_{j+1}$: \begin{equation}\label{FD1} u^{''}(x_j) \approx \frac{2 u(x_{j-1}) }{h_j(h_j +h_{j+1})} + \frac {-2 u(x_j)}{h_j\, h_{j+1}} + \frac{2 u(x_{j+1})}{h_{j+1}(h_j +h_{j+1})}. \end{equation} For the uniform distribution of the nodes $x_j=h j$, $j=0,1, \ldots, n$, where $h=\frac{1}{n}$, the approximation \eqref{FD1} becomes the second order standard centered difference approximation \[ u^{''}(x_j) \approx \frac{u(x_{j-1}) -2 u(x_j) + u(x_{j+1})}{h^2}. \] For the general case of non-uniform distributed nodes, we let $f_j:=f(x_j)$, and for $u_j\approx u(x_j)$, we solve the system \begin{equation}\label{FDs1} \left\{ \begin{array}{rcl} u_0 & =0&\\ \\ \displaystyle \frac{-2 u_{j-1} }{h_j(h_j +h_{j+1})} + \frac {2 u_j}{h_j\, h_{j+1}} + \frac{-2 u_{j+1}}{h_{j+1}(h_j +h_{j+1})} & =f_j& \ j=\overline{1,n-1}.\\ \\ u_n & =0.& \end{array} \right. \end{equation} By multiplying the generic equation in \eqref{FDs1} with $\displaystyle \frac{h_j +h_{j+1}}{2}$, we get \begin{equation}\label{FDs2} \frac{-1}{h_j}\, u_{j-1} + \left (\frac {1}{h_j} + \frac{1}{h_{j+1}} \right ) u_j + \frac{-1} {h_{j+1}} u_{j+1} = \frac{h_j +h_{j+1}}{2}\, f_j. \end{equation} Denote ${u}^{FD} := [ u_1, u_2, \cdots, u_{n-1}]^T$, and $\tilde{f}:=[ f_1, f_2, \cdots, f_{n-1}]^T$. Let $W$ be the $(n-1)\times(n-1)$ diagonal matrix with entries: $\{\frac{h_1 +h_{2}}{2}, \cdots, \frac{h_{n-1}+h_{n}}{2} \}$ and let $S$ be the stiffness tridiagonal $(n-1)\times(n-1)$ matrix \begin{equation} \label{S} S=\left [\begin{matrix} \frac{1}{h_1}+ \frac{1}{h_2},& -\frac{1}{h_2} & & & & \\ -\frac{1}{h_2} & \frac{1}{h_2}+ \frac{1}{h_3} & -\frac{1}{h_3} & & & \\ \ & \ddots & \ddots & \ddots & & \\ & & & & & \\ & & & -\frac{1}{h_{n-2}} & \frac{1}{h_{n-2}}+ \frac{1}{h_{n-1}} & -\frac{1}{h_{n}} \\ & & & & -\frac{1}{h_{n-1}} & \frac{1}{h_{n-1}}+ \frac{1}{h_n} \\ \end{matrix} \right]. \end{equation} Then, using \eqref{FDs2} and the fact that $u_0=u_n=0$, the system \eqref{FDs1} is equivalent to \begin{equation}\label{FDs3} S\, {u}^{FD} = W\, \tilde{f}. \end{equation} For a better comparison of the finite difference solution $ {u}^{FD}$ to the finite element approximation $ {u}^{FE}$, we prefer to write the linear system in the form \eqref{FDs3}. Secondly, for the FE discretization, we will use the standard notation \[ a(u, v) = \int_0^1 u'(x) v'(x) \, dx, \ \text{and} \ (f, v) = \int_0^1 f(x) v(x) \, dx. \] Then, the variational formulation of \eqref{2pBVP} is : Find $u \in V:= H^1_0(0,1)$ such that \begin{equation}\label{VF} a(u, v) = (f, v), \ \text{for all} \ v \in V. \end{equation} Also, we consider the mesh nodes on $[0,1]$ as being the same as those used for the finite difference discretization: $0=x_0<x_1<\cdots < x_n=1$, and define the corresponding discrete space $V_h$ as the subspace of $ V = H^1_0(0,1)$, given by \[ V_h = \{ v_h \in V \mid v_h \text{ is linear on each } [x_j, x_{j + 1}]\}, \] i.e., $V_h$ is the space of all {\it piecewise linear continuous functions} with respect to the given nodes that are zero at $x=0$ and $x=1$. We consider the nodal basis $\{ \varphi_j\}_{j = 1}^{n-1} \subset V_h$ such that $\varphi_i(x_j ) = \delta_{ij}$. Thus, the discrete variational formulation of \eqref{2pBVP} is: Find $u_h \in V:= H^1_0(0,1)$ such that \begin{equation}\label{dVF} a(u_h, v_h) = (f, v_h), \ \text{for all} \ v_h \in V_h. \end{equation} We look for $u_h=u_h^{FE}$ with the nodal basis expansion \[ u_h := \sum_{i=1}^{n-1} u_i \varphi_i, \ \text{where} \ \ u_i=u_h(x_i). \] If we consider the test functions $v_h=\varphi_j, j=1,2,\cdots,n-1$ in \eqref{dVF}, we obtain the system \begin{equation}\label{eq:FEsol} S\, u^{FE} = \tildeu{f}, \end{equation} for finding the coefficient vector $u^{FE}=[u_1, \cdots, u_{n-1}]^T$, where \\ $\tildeu{f}:=[(f, \varphi_1), \cdots (f, \varphi_{n-1})]^T$ and $S$ is the $(n-1)\times (n-1)$ {\it tridiagonal} matrix with entries $S_{ij}=a(\varphi_i', \varphi_j')$. It is an easy exercise to verify that $S$ defined for the finite element discretization, coincides with the matrix defined in \eqref{S} for the finite difference discretization. We note that for any $\alpha \in \mathbb{R}^{n-1}, \alpha \neq 0$ \[ \alpha^T S \, \alpha = a(v_h, v_h) = \int_0^1 (v'_h(x))^2 \, dx>0, \] where $v_h = \sum_{i=1}^{n-1} \alpha_i \varphi_i$. Consequently, the stifnees matrix $S$ is invertible. We can also define the $(n-1)\times(n-1)$ Green matrix \[ \tilde{G} =\left [G(x_i,x_j) \right ]_{i,j=\overline{1,n-1}}. \] The purpose of the paper is to show how the two types of discretization are connected. Using that the Green function \eqref{Gfunc} can be viewed as a scaled finite element function, we will prove that the common stiffness matrix $S$ satisfies $ S^{-1}=\tilde{G}$ in the general non-uniform case. While the identity is well known for the uniform distributed nodes, see Section 12.2.2 in \cite{QSS}, by the best knowledge of the authors, the identity seems to be not available for the non-uniform case or in the finite element context. We will also show that the right hand side vector of \eqref{FDs3}, that defines the finite difference solution, is obtained by the Composite Trapezoid Rule (CTR) approximation of the entries of the dual vector $\tildeu{f}$. As a consequence, we can compare the $u^{FD}$ and $u^{FE}$ and estimate the energy norm of the error $u^{FD} -u^{FE}$. The rest of the paper is organized as follows: In Section \ref{sec:FDvsFE}, we present the similarity between the finite element and the finite difference linear systems for the model problem \eqref{2pBVP}. In Section \ref{sec:error}, we use the formulas for the FD and the FE solutions in order to compare the two discretizations and do an error analysis. Two interesting extensions to the 2D case are presented in Section \ref{sec:2D}. \section{The connections between the finite element and the finite difference of the 1D model problem}\label{sec:FDvsFE} In this section we investigate the systems for the FE and FD discretizations of \eqref{2pBVP} by relating the matrices and the load vectors for the two systems. The following lemma gives a formula for the general component of the dual vector $\tildeu{f}$ and the evaluation of the solution $u$ at the neighboring nodes viewed as degrees of freedom for the finite element approximation. In addition, the formula connects with the FD discretization via the formula \eqref{FDs2}. \subsection{A formula for the finite element solution} \begin{lemma}\label{fvarphi} Let $u$ be the solution of \eqref{2pBVP}, and $\{\varphi_j\}_{j=1,2,\cdots, n-1}$ be the nodal basis for $V_h$. Then, for $j=1,2,\cdots,n-1$, we have \begin{equation}\label{lemma1fphij} (f,\varphi_j)=\displaystyle -\frac{1}{h_j} u(x_{j-1})+\left(\frac{1}{h_j}+\frac{1}{h_{j+1}}\right) u(x_j)-\frac{1}{h_{j+1}} u(x_{j+1}) \end{equation} \end{lemma} \begin{proof} From the variational formulation \eqref{dVF}, we have $(f,\varphi_j)=a(u,\varphi_j)=\displaystyle \int_0^1u'(x)\varphi'_j(x) dx=\frac{1}{h_j}\int_{x_{j-1}}^{x_j} u'(x) dx+\frac{-1}{h_{j+1}}\int_{x_{j}}^{x_{j+1}} u'(x) dx=-\frac{1}{h_j} u(x_{j-1})+\left(\frac{1}{h_j}+\frac{1}{h_{j+1}}\right) u(x_j)-\frac{1}{h_{j+1}} u(x_{j+1})$. \end{proof} Next, for any $s\in (0,1)$, we define the function \begin{equation} \phi_s(x) = \left\{ \begin{array}{ccl} \displaystyle \frac{x}{s} & \mbox{if } \ 0\leq x\leq s,\\ \\ \displaystyle \frac{1-x}{1-s} & \mbox{if } \ s<x\leq 1.\\ \end{array} \right. \end{equation} Note that, if $s=x_j$, $j=1,2,\cdots,n-1$, then $\phi_{x_j}$ is a piecewise linear function that belongs to $V_h$. \begin{corollary} (Green's Formula) For any $s\in (0,1)$, we have \begin{equation}\label{fBigphij} (f,\phi_s)=\displaystyle \frac{1}{s(1-s)} u(s) \end{equation} \end{corollary} \begin{proof} In the Lemma \eqref{fvarphi}, take $n=2$, $x_0=0$, $x_2=1$, $x_1=s$, $h_1=s$, $h_2=1-s$. \end{proof} Note that the equation \eqref{fBigphij} is in fact the Green's formula \eqref{dVF}. This is because \begin{equation}\label{GBigphis} G(x,s)=s(1-s) \phi_s(x)=G(s,x), \ \text{for all} \ x, s \in (0, 1). \end{equation} This formula shows that the Green's Function can be interpreted as a scaled finite element function. \begin{corollary}\label{uFF} Let $\displaystyle u_h := \sum_{i=1}^{n-1} u_i \varphi_i$ be the finite element solution of \eqref{dVF}, and let ${u}^{FE} := [ u_1, u_2, \cdots, u_{n-1}]^T$. Then $u_h$ coincides with the linear interpolant of $u$ on the nodes $x_0,x_1, \ldots, x_n$. In other words, $u_j=u(x_j)$, $j=1,2,\cdots,n-1$. \end{corollary} \begin{proof} Since the stiffness matrix $S$ for the system \eqref{dVF} is given by \eqref{S}, from the equation \eqref{lemma1fphij} we have that $\tilde{u}=[u(x_1), u(x_2), \ldots, u(x_{n-1})]^T$ solves $S\tilde{u}=\tildeu{f}$. On the other hand, from \eqref{eq:FEsol} we have that $S\, u^{FE}=\tildeu{f}$. Using that the matrix $S$ is invertible, we conclude that $u^{FE}=\tilde{u}$. \end{proof} \begin{corollary} If $S$ is the stiffness matrix defined in \eqref{S}, then \begin{equation}\label{SinverseG} S^{-1}=\tilde{G}, \end{equation} consequently, $u^{FE}=S^{-1}\tildeu{f}= \tilde{G} \tildeu{f}$. \end{corollary} \begin{proof} Apply \eqref{fBigphij} for $s=x_j$, $j=1,2,\cdots,n-1$ to get \begin{equation}\label{uxPhij} u(x_j)=x_j(1-x_j)\int_0^1\phi_{x_j}(x)f(x) dx. \end{equation} Obviously, by the definition of $\phi_{x_j}$ we have $\phi_{x_j}\in V_h$. Hence, $\phi_{x_j}$ coincides with its interpolant on the nodes $x_0,x_1, \ldots, x_n$. Thus, \begin{equation}\label{Phivarphi} \phi_x(x)=\sum_{i=1}^{n-1} \phi_{x_j}(x_i) \varphi_i(x). \end{equation} For the equations \eqref{uxPhij} and \eqref{Phivarphi}, we get \begin{equation}\label{eq:FESf} u(x_j)=x_j(1-x_j)\sum_{i=1}^{n-1} \phi_{x_j}(x_i) \int_0^1f(x)\varphi_i(x) dx. \end{equation} Since $x_j(1-x_j)\phi_{x_j}(x_i)=G(x_j,x_i)$, we obtain $\displaystyle u(x_j)=\sum_{i=1}^{n-1} G(x_j,x_i)(f,\varphi_i)$, $j=1,2,\cdots,n-1$, which combined with the fact that the finite element solution coincides with the interpolant as given by Corollary \ref{uFF}, leads to $u^{FE}=\tilde{G}\tildeu{f}$. Since $u^{FE}=S^{-1}\tildeu{f}$ for any $\tildeu{f} \in \mathbb{R}^{n-1}$, we conclude that the formula \eqref{SinverseG} holds. \end{proof} In order to obtain the formula \eqref{eq:FESf}, it is essential that $\phi_{x_j}\in V_h$, and that $\phi_{x_j}$ can be expanded in the basis $\{ \varphi_i\}$. \subsection{A formula for the finite difference solution} Let us consider the finite difference discretization \eqref{FDs1} that is equivalent to \eqref{FDs2} and builds the system \eqref{FDs3}. For a continuous function $\theta : [0,1]\to \mathbb{R}$ such that $\theta(0)=\theta(1)=0$, the composite trapezoid rule (CTR) on the nodes $x_0,x_1, \ldots, x_n$ is \begin{equation}\label{CTR} \int_0^1\theta(x)dx \approx T_n(\theta):=\sum_{i=1}^{n-1} \theta(x_i) \frac{h_i+h_{i+1}}{2}. \end{equation} Note that if $\displaystyle I(\theta)=\sum_{i=1}^{n-1} \theta(x_i)\, \varphi_i$ is the interpolant of $\theta$, and by using that $\int_0^1 \varphi_i(x)dx=\displaystyle \frac{h_i+h_{i+1}}{2}$, then the CTR formula becomes \[ T_n(\theta)=\int_0^1I(\theta)(x)dx. \] We observe that the diagonal entries of $W$ in equation \eqref{FDs3} are exactly the weights of the quadrature formula \eqref{CTR}. Next, we will show that, componentwise, the finite difference solution $u_j$ is the CTR approximation of the solution $u(x_i)$ given by the Green's formula: \begin{equation}\label{Guj} u(x_j)=\int_0^1 G(x_j,x)f(x)dx. \end{equation} For each $j=1,2,\cdots,n-1$, we define \[ \theta_j(x)=G(x_j,x)f(x), \ x \in [0,1], \ \text{and} \] \[ \begin{aligned} w_j & :=T_n(\theta_j)=\displaystyle \sum_{i=1}^{n-1}\frac{h_i+h_{i+1}}{2}\theta_j(x_i) =\displaystyle \sum_{i=1}^{n-1} G(x_j,x_i)\frac{h_i+h_{i+1}}{2}f(x_i) \\ & =(\tilde{G}W\tilde{f})_j. \end{aligned} \] Consequently, we have \begin{equation}\label{eq:w} w= \tilde{G}W\tilde{f}. \end{equation} \begin{prop}\label{WT} The finite difference solution $u^{FD}$ of \eqref{FDs3} coincides with the vector $w=[w_1, w_2, \ldots, w_{n-1}]^T$. \end{prop} \begin{proof} From the formula \eqref{SinverseG}, we have that $u^{FD}=\tilde{G} W\tilde{f}$ which compared to \eqref{eq:w}, leads to $u^{FD}=w$. \end{proof} \begin{remark} We can prove the Proposition \ref{WT} without using finite element arguments needed to prove \eqref{SinverseG}. We can just verify that the vector $w$ satisfies the system $Sw=W\tilde{f}$ by essentially using that $\phi_{x_j}$ is linear on the intervals $[0, x_j]$ and $[x_j, 1]$. \end{remark} \subsection{Comparison} By rewriting the FD system, we have that the FD and the FE linear systems have the same stiffness matrices. The corresponding solutions are given by \[ u^{FE}=S^{-1}\tildeu{f}, \ \text{and} \ u^{FD}=S^{-1}W\tilde{f},\ \text{where} \] \[ (\tildeu{f})_j=(f,\varphi_j)=\int_0^1 f(x)\varphi_j(x)\, dx, \ \text{and} \ (W\tilde{f})_j=\frac{h_j+h_{j+1}}{2}f(x_j) . \] The right hand sides of the two systems are componentwise related by \begin{equation}\label{fTn} (f,\varphi_j)=T_n(f\varphi_j). \end{equation} \section{Error Estimates} \label{sec:error} As a consequence of the previous section, we provide error estimates for $u -u_h^{FD}$ in both the discrete infinity norm and the energy norm, and estimate the energy norm of the difference $u_h^{FE} -u_h^{FD}$. \subsection{The infinity norm estimate for the FD solution} From Lemma \ref{WT}, we can estimate the error $$ \displaystyle \max_{j=\overline{1,n-1}} |u(x_j)-u_j^{FD}| $$ in the general non-uniform case. For the uniform case, it is well known (see e.g. \cite{QSS}), that if $ f\in {\mathcal{C}}^2([0,1])$, then \begin{equation}\label{maxerrc} \displaystyle \max_{1\leq j\leq n} |u(x_j)-u_j^{FD}|\leq \frac{h^2}{96}\|f''\|_{\infty}, \end{equation} where for a continuous function $\theta$ on $[a,b]$, we define $\displaystyle \|\theta\|_{\infty, [a, b]}:=\max_{x\in[a,b]}|\theta(x)|$ and $\|\theta\|_{\infty}:=\|\theta\|_{\infty, [0, 1]}$. For extending \eqref{maxerrc} to the general nonuniform case, we will use the known formula for the CTR error as follows: For a function $\theta:[a,b]\to \mathbb{R}$, \\ $\theta \in {\mathcal{C}}^2([a,b])$, and the nodes $a\leq x_0 <x-1<\ldots <x_n=b$, we have \begin{equation}\label{CTRE} \displaystyle \left|\int_a^b\theta(x)dx - T_{n,[a,b]}\right|\leq\frac{b-a}{12} h^2\|\theta''\|_{\infty,[a,b]}, \end{equation} where $\displaystyle T_{n,[a,b]}(\theta)=\sum_{i=1}^n (x_i-x_{i-1})\frac{\theta(x_i)+\theta(x_{i-1})}{2}$ and $h=\displaystyle \max_{i=\overline{1,n}}(x_i-x_{i-1})$. \begin{theorem} Let $f\in {\mathcal{C}}^2[0,1])$, and let $u$ be the solution of the boundary value problem \eqref{2pBVP}, and $u^{FD}=[u_1,u_2,\ldots, u_{n-1}]^T$ be the finite difference solution of \eqref{FDs3}. If $h_j=x_j-x_{j-1}$ for $j=1,2,\ldots, n$ and $\displaystyle h:=\max_{j=\overline{1,n}} |h_j|$, then, we have \begin{equation}\label{EDEinf} \displaystyle \max_{1\leq j \leq n-1}|u(x_j)-u_j|\leq \frac{h^2}{48}(4\|f'\|_{\infty}+\|f''\|_{\infty}). \end{equation} \end{theorem} \begin{proof} From the proof of Corollary \ref{SinverseG} and the equation \eqref{uxPhij}, we have \[ u(x_j)=x_j(1-x_j)\int_0^1 \phi_{x_j}(x) f(x)dx. \] From Proposition \ref{WT}, equations \eqref{CTR} and \eqref{fTn}, we also have that \begin{equation}\label{uFDf} \begin{aligned} u_j&=(\tilde{G}W\tilde{f})_j= \displaystyle \sum_{i=1}^{n-1} G(x_j,x_i)\frac{h_i+h_{i+1}}{2}f(x_i) \\ &= \displaystyle x_j(1-x_j) \sum_{i=1}^{n-1} \phi_{x_j}(x_i)\frac{h_i+h_{i+1}}{2} = \displaystyle x_j(1-x_j) T_n(\phi_{x_j}f). \end{aligned} \end{equation} Consequently, the error $E_j=u(x_j)-u_j$ satisfies \begin{equation} \begin{aligned} E_j&= \displaystyle x_j(1-x_j)\left( \int_0^1\phi_{x_j}(x)f(x)dx- T_n(\phi_{x_j}f)\right)=\\ &=x_j(1-x_j)\left(\int_0^{x_j} \phi_{x_j}(x)f(x)dx - T_{n, [0, x_j]}(\phi_{x_j}f)\right) +\\ & + x_j(1-x_j)\left(\int_{x_j}^1 \phi_{x_j}(x)f(x)dx- T_{n, [ x_j,1]}(\phi_{x_j}f)\right). \end{aligned} \end{equation} By noting that $\phi_xf$ is ${\mathcal{C}}^2$ on $[0,x_j]$ and $[x_j,1]$, applying the estimate \eqref{CTRE} on the intervals $[0,x_j]$ and $[x_j,1]$ for $\theta(x)=\phi_{x_j}(x)f(x)$, and using \[ \displaystyle \max_{x\in [0,x_j]}|\theta''(x)|\leq \frac{2}{x_j}\|f'\|_{\infty}+\|f''\|_{\infty}, \ \text{and} \] \[ \displaystyle \max_{x \in [x_j,1]}|\theta''(x)|\leq \frac{2}{1-x_j}\|f'\|_{\infty}+\|f''\|_{\infty}, \] we obtain \begin{equation*} \begin{aligned} \displaystyle |E_j| & \leq x_j(1-x_j) \frac{h^2}{12}x_j\left(\frac{2}{x_j}\|f'\|_{\infty,[0,x_j]} +\|f''\|_{\infty}\right)+ \\ &+x_j(1-x_j) \frac{h^2}{12}(1-x_j)\left(\frac{2}{1-x_j}\|f'\|_{\infty,[x_j,1]}+\|f''\|_{\infty}\right). \end{aligned} \end{equation*} Since $x_j(1-x_j) \leq \frac{1}{4}$, the above estimate leads to \eqref{EDEinf}. \end{proof} \begin{remark} It is known that \eqref{FD1} is only first order accurate if $h_j\neq h_{j+1}$, see e.g., Section 1.3 in \cite{LeVeque07}. The estimate \eqref{EDEinf} shows that the FD method based on \eqref{FDs1} or \eqref{FDs2} is globally $O(h^2)$. We also emphasize that our proof is done using the formula of the solution \eqref{uFDf}, and it is not based on \\ $\delta$-functional arguments as usually done for proving \eqref{maxerrc}, see e.g., \cite{QSS}. \end{remark} Based on the CTR approximation, the formula \eqref{uFDf} allows for further error analysis in the energy norm. \subsection{Energy Norm Errors} For $v\in V=H_0^1(0, 1)$, define the energy norm \begin{equation} |v|=|v|_a=\left(\int_0^1 (v'(x))^2dx\right)^{\frac{1}{2}}=(a(v,v))^{\frac{1}{2}}. \end{equation} If the vector $u^{FD}=[u_1, \ldots, u_{n-1}]^T$ is the solution of the finite difference system \eqref{FDs3}, then we define the corresponding function in $V_h$ as $u_h^{FD}:=\displaystyle \sum_{i=1}^{n-1} u_i\varphi_i$. From the previous sections, we know that $u^{FD}$ satisfies the equation $Su^{FD}=\tilde{G}(W\tilde{f})$, where $W\tilde{f}$ is defined by \eqref{S} and \eqref{FDs3}, and $\displaystyle \frac{h_i+h_{i+1}}{2}f(x_i)=T_n(f\varphi_j)$. Thus, since the finite element stiffness matrix is still $S$, we have that \[ a(u_h^{FD}, \varphi_j)=T_n(f\varphi_j), \ \ \text{for all} \ j=1,2,\ldots, n-1. \] Using the linearity of $a(u_h^{FD},\cdot)$ and $T_n(f,\cdot)$, we have that \begin{equation}\label{uhFD} a(u_h^{FD},v_h)=T_n(fv_h), \ \text{for all} \ v_h \in V_h. \end{equation} On the other hand, we have \begin{equation}\label{uhFE} a(u_h^{FE}, v_h)=(f,v_h)=\int_0^1fv_h \ dx, \ \text{for all} \ v_h\in V_h. \end{equation} From the equations \eqref{uhFD} and \eqref{uhFE}, we obtain \[ a(u_h^{FE}-u_h^{FD}, v_h)=F_h(v_h):=\int_0^1fv_h \ dx-T_n(fv_h). \] Consequently, \begin{equation}\label{Fh} |u_h^{FE}-u_h^{FD}|=\sup_{v_h \in V_h} \frac{F_h(v_h)}{|v_h|}:=\|F_h\|_{V_h^*} \end{equation} where the supremum is taken over all non-zero vectors. \begin{theorem}\label{th:FED} Assume that $f\in{\mathcal{C}}^2([0,1])$, and that $u_h^{FE}$ and $u_h^{FD}$ are the finite element and the finite difference corresponding solutions of the boundary value problem \eqref{2pBVP} on the nodes $x_0=0<x_1<\ldots<x_n=1$. Let $\displaystyle h=\max_{i=\overline{1,n}} (x_i-x_{i-1})=\displaystyle \max_{i=\overline{1,n}}(h_i)$. Then, \begin{equation}\label{udiff} |u_h^{FE}-u_h^{FD}|\leq h^2\left( \frac{\|f''\|_{\infty}}{12}+\frac{\|f'\|_{\infty}}{6}\right). \end{equation} \end{theorem} \begin{proof} According to the equation \eqref{Fh}, we just have to find an upper bound for $\|F_h\|_{V_h^{\star}}$. Let $\theta_{i}(x):=f(x)v_h(x)$, for $x\in [x_{i-1},x_i]$, $i=1,2,\ldots, n$. Using the trapezoid rule error formula for $\theta_i$, we have that \begin{equation}\label{Fvh} F_h(v_h)=-\frac{1}{12}\sum_{i=1}^n h_i^3 \theta_i ''(\xi_i), \ \xi_i \in [x_{i-1},x_i]. \end{equation} To simplify the notation, we assume that $\displaystyle v_h=\sum_{i=1}^{n-1} \alpha_i \varphi_i$. Then, on the interval $(x_{i-1},x_i)$, $v'_h= \frac{\alpha_i - \alpha_{i-1}}{h_i}$, and from the equation \eqref{Fvh}, we have \[ F_h(v_h)=\displaystyle -\frac{1}{12}\left(\sum_{i=1}^n h_i^3 f''(\xi_i)v_h(\xi_i)+2\sum_{i=1}^n h_i^3\, \frac{\alpha_i-\alpha_{i-1}}{h_i}f'(\xi_i) \right). \] We have $h_i\leq h$ and $|v_h(x)| \leq \|v_h\|_{\infty}\leq |v_h|$, for all $v_h \in V_h$. Thus, \begin{equation}\label{Fh2} |F_h(v_h)| \leq \frac{1}{12}h^2|v_h|\sum_{i=1}^n h_i |f''(\xi_i)|+\frac{1}{6} h^2 \|f'\|_{\infty}\sum_{i=1}^n|\alpha_i - \alpha_{i-1}|. \end{equation} Using the discrete mean value theorem, (see e.g., Theorem 9.1 in \cite{QSS}), for some $\xi \in (0, 1)$, we have \[ \sum_{i=1}^n h_i |f''(\xi_i)|=|f''(\xi)| \sum_{i=1}^n h_i\leq \|f''\|_{\infty}. \] For the second sum in \eqref{Fh2}, the Cauchy-Schwarz inequality gives \[ \begin{aligned} \sum_{i=1}^n|\alpha_i-\alpha_{i-1}|& =\sum_{i=1}^n h_i^{\frac{1}{2}} \frac{|\alpha_i-\alpha_{i-1}|}{h_i^{\frac{1}{2}}} \leq \left(\sum_{i=1}^n h_i\right)^{1/2}\left(\sum_{i=1}^n \frac{|\alpha_i-\alpha_{i-1}|^2}{h_i}\right)^{1/2}\\ &=\left(\sum_{i=1}^n \int_{x_{i-1}}^{x_i} (v'_h)^2 dx\right )^{1/2}=|v_h|. \end{aligned} \] Combining the last two estimates with \eqref{Fh2}, gives \[ |F_h(v_h)|\leq \frac{1}{12} h^2 |v_h| \|f''\|_{\infty}+\frac{1}{6}h^2\|f'\|_{\infty} |v_h|. \] Hence, \[ \|F_h\|_{V_h^*} \leq h^2\left( \frac{\|f''\|_{\infty}}{12}+\frac{\|f'\|_{\infty}}{6}\right), \] which, together with \eqref{Fh} proves the statement of the theorem. \end{proof} \begin{corollary} Under the same assumptions as in the Theorem \ref{th:FED}, we have \begin{equation}\label{ChFD} |u-u_h^{FD} |\leq Ch(\|f''\|_\infty + \|f\|_{L^2(0, 1)}), \end{equation} where $C$ is a constant independent of $h$ or $f$. \end{corollary} \begin{proof} It is known that the following estimate for the finite element error in the energy norm holds \begin{equation}\label{ChFE} |u-u_h^{FE} |\leq Ch \|f\|_{L^2(0, 1)}. \end{equation} where $C$ is a constant independent of $h$ or $f$, see e.g. \cite{braess, brenner-scott, ern-guermond}. Now, the inequality \eqref{ChFD} follows using \eqref{ChFE}, \eqref{udiff}, and the triangle inequality \[ |u-u_h^{FD}| \leq |u-u_h^{FE}| + |u_h^{FE}-u_h^{FD}|. \] \end{proof} \section{Interesting extensions to the 2D case}\label{sec:2D} Let $\Omega\subset\mathbb{R}^2$ be a polygonal domain, and $f \in L^2(\Omega)$ or $f$ is a continuous functional on $H^1_0(\Omega)$. Consider the model problem: Find $u$ such that \begin{equation}\label{2DBVP} \left\{ \begin{array}{rccl} -\Delta u = & f & \text{in } \ \Omega,\\ u =& 0 & \text{on} \ \partial \Omega. \end{array} \right. \end{equation} The corresponding variational or weak formulation of \eqref{2DBVP} is: Find $u \in H^1_0(\Omega)$ such that \begin{equation}\label{2Dvf} a(u,v) = (f,v) ,\ \text{for all} \ v \in H^1_0(\Omega), \end{equation} where, for any $u,v\in H^1_0(\Omega)$, and for $f \in L^2(\Omega)$, \[ a(u,v): =\int_\Omega \nabla u\cdot \nabla v \, dx, \ \text{and}\ (f,v):=\int_\Omega f(x) v(x) \, dx. \] The existence and uniqueness of the solution of \eqref{2Dvf} is well know, see e.g., \cite{braess, brenner-scott, ern-guermond, demko-oden}. For the discretization of \eqref{2DBVP}, we let ${\mathcal{T}}_h$ be a triangulation of $\Omega$ and consider $V_h$ a conforming finite element space with a nodal basis $\{\varphi_1, \cdots, \varphi_n\}$ associated with the mesh ${\mathcal{T}}_h$. If $\varphi_j$ is a generic basis function with support $D_j \subset \Omega$, we assume \begin{equation*} \displaystyle D_j= \bigcup_{i=1}^{n_j} T_i, \ \text{with} \ T_i\ \text{being a triangle in} \ {\mathcal{T}}_h. \end{equation*} The formula \eqref{lemma1fphij} is a formula of significant importance because it shows the strong connection between the FE and the FD discretizations in 1D. A corresponding formula for the 2D case can be obtained by applying the Green's formula for the function $f\varphi_j = -\Delta u \varphi_j$ to get \begin{equation}\label{2DF} \begin{aligned} (f,\varphi_j) & = \int_\Omega -\Delta u \varphi_j \, dx = \int_\Omega \nabla u\cdot \nabla \varphi_j\, dx =\\ & = -\sum_{T_i\subset D_j}\int_{T_i} (\Delta \varphi_j)\, u\, dx +\sum_{T_i\subset D_j} \int_{\partial T_i} (\nabla u\cdot {\bf n}) \, u \, ds. \end{aligned} \end{equation} where ${\bf{n}}$ is the outer normal vector to the integration domain's boundary. Similary to the 1D case, the components $(f,\varphi_j)$ of the dual vector $f$ relative to the basis $\{\varphi_1, \cdots, \varphi_n\}$ can be written as linear combinations of degrees of freedom acting on $u$. However, in this case, the degrees of freedom have to involve integrations on triangles and on the edges of ${\mathcal{T}}_h$. We can redefine an interpolant using the new degrees of freedom in order to have the FE solution agree with the interpolant. However, the main challenge in this case is to match the number of degrees of freedom with the number $n$ of basis functions when considering the system given by \eqref{2DF}. Except for very simple meshes, this seems to be difficult to achieve. Next, we consider two special 2D cases, leaving the challenge to extend this idea to the more general cases of meshes in 2D or 3D for future work. \subsection{The Green function for the 1D BVP as a 2D discretization function fo the 2D BVP} Consider the problem \eqref{2DBVP} on the unit square $\Omega=(0, 1) \times (0, 1)$, and define the mesh ${\mathcal{T}}_1=T_1\cup T_2$, where $T_1, T_2$ are the two triangles determined by the positive slope diagonal $\Gamma$ of $\Omega$. We define $V_h=V_1= span\{G\}$, where $G=G(x,y)$ is the function defined in \eqref{Gfunc}. We have $G\in H^1_0(\Omega)$, and if $u$ is the solution of \eqref{2DBVP}, then \begin{equation}\label{2DG} \begin{aligned} (f,G) & = \int_\Omega -\Delta u\, G\, dxdy = \int_\Omega \nabla u\cdot \nabla G\, dxdy =\\ & = \int_{T_1} -\Delta G \, u\, dxdy \int_{\partial T_1} (\nabla G\cdot {\bf n}_{T_1})\, u \, ds +\\ & + \int_{T_2} -\Delta G \, u\, dxdy \int_{\partial T_2} (\nabla G\cdot {\bf n}_{T_2})\, u \, ds. \end{aligned} \end{equation} Using that $\Delta G= 0 \ \ \text{on} \ \Omega \backslash \Gamma$ and on $\Gamma= \{(x,y) \in\Omega |\ x=y\}$, and that \[ \nabla G _{|_{T_1}} \cdot {\bf n}_{T_1}+ \nabla G _{|_{T_2}} \cdot {\bf n}_{T_2}=\sqrt{2}, \] from \eqref{2DG}, we obtain a formula for $\int_\Gamma u\, ds $. \begin{theorem} Let $u$ be the solution of \eqref{2pBVP}. Then, \begin{equation}\label{2DGF} \int_\Gamma u\, ds = \frac{1}{\sqrt{2}} \int_\Omega f G\, dx dy = \frac{1}{\sqrt{2}} \int_\Omega \nabla u \cdot \nabla G \, dx dy. \end{equation} \end{theorem} \begin{remark} The simple calculations leading to formula \eqref{2DGF} have the following consequences: \begin{enumerate} \item [i)] The function $G=G(x,y)$ is the unique solution of \begin{equation}\label{2Ddelta} -\Delta u = \sqrt{2} \, \delta_\Gamma \ \text{on} \ \Omega =(0, 1)\times (0, 1), \end{equation} where \[ \delta_\Gamma (\varphi) = \int_\Gamma \varphi\, ds, \ \text{for all} \ \varphi \in C_0^\infty (\Omega). \] \item [ii)] The function $G=G(x,y)$ is the Riesz representation of \\ the functional $ \sqrt{2} \, \delta_\Gamma :H^1_0(\Omega)\to \mathbb{R}$. \item [iii)] Part ii) can be viewed as an extension of the 1D problem on $(0,1)$, where, for each $s \in (0, 1)$, the function \[ x \to s(1-s) \Phi_s(x) = G(s,x) \] is the Riesz representation of $\delta_s :H^1_0(\Omega)\to \mathbb{R}$, \[ \delta_s(\varphi) = \varphi(s), \ \text{for any} \ \varphi \in C_0^\infty (\Omega). \] \item [iv)] The function $G=G(x,y)$ can be viewed also as the finite element discretization of \eqref{2pBVP}, using $C^0-P^2$ discretization on the space $V_1$. \item [v)] In light of the FD versus FE connections presented here, any 2D quadrature approximation of the 2D integral $\displaystyle \frac{1}{\sqrt{2}} \int_\Omega f G$ can be viewed as a ``finite difference'' approximation of $\displaystyle \int_\Gamma u\ ds$, where $u$ is the solution of \eqref{2pBVP}. \end{enumerate} \end{remark} \subsection{ A bubble function for the 2D discretization on a special domain} Consider that $\Omega$ is the domain defined by one equilateral triangle $T$ with vertices $z_1, z_2, z_3$. We define the bubble function \[ B= \lambda_1 \lambda_2 \lambda_3, \] where $ \lambda_1, \lambda_2$ and $\lambda_3$ are the linear functions on $T$ with the property that $\lambda_i(z_j)=\delta_{ij},\ i, j=1,2,3$. We denote the area of the triangle $T$ by $|T|$. It is easy to check that $B\in H_0^1(\Omega)$ and \begin{equation}\label{B} -\Delta B= \frac{1}{\sqrt{3}\, |T|}. \end{equation} Then, for any function $\varphi \in H_0^1(\Omega)$, we have \begin{equation}\label{B2} \frac{1}{\sqrt{3}\, |T|} \int_\Omega \varphi = \int_\Omega (-\Delta B) \varphi = \int_\Omega \nabla B\cdot \nabla \varphi . \end{equation} In particular, if $u\in H_0^1(\Omega)$ is the solution of \eqref{2DBVP} for a given $f$, we have \begin{equation}\label{B3} \frac{1}{\sqrt{3}\, |T|} \int_\Omega u = \int_\Omega \nabla B\cdot \nabla u = \int_\Omega f\, B = (f, B). \end{equation} \begin{theorem} Let $\Omega$ be an equilateral triangle, and let $u$ be the solution of \eqref{2pBVP}, then \begin{equation}\label{2DBF} \int_\Omega u = \sqrt{3}|\Omega| \int_\Omega f B. \end{equation} \end{theorem} \begin{remark} As in the previous case, we point out the following consequences: \begin{enumerate} \item [i)] The function $\sqrt{3}\, |\Omega|\, B$ is the unique solution of \[ \left\{ \begin{array}{rccl} -\Delta u = & 1 & \text{in } \ \Omega,\\ u =& 0 & \text{on} \ \partial \Omega. \end{array} \right. \] \item [ii)] From \eqref{B2}, the function $\sqrt{3}\, B$ is the Riesz representation of the functional $ F :H^1_0(\Omega)\to \mathbb{R}$, \[ F(\varphi)= \frac{1}{ |\Omega|} \int_\Omega \varphi. \] \item [iii)] The function $\sqrt{3}\, |\Omega|\, B$ can be viewed also as the finite element discretization of \eqref{2pBVP}, using $C^0-P^3$ discretization on the space \\ $V_0:= span\{B\}$. \item [iv)] Any 2D quadrature approximation of $\sqrt{3} \int_\Omega f B\, dx dy$ can be viewed as a ``finite difference'' approximation of $\frac{1}{|\Omega|} \int_\Omega u$, where $u$ is the solution of \eqref{2pBVP}. \end{enumerate} \end{remark} We note that the assumption that $\Omega$ be an equilateral triangle is needed in order to have that the Laplacian operator acting on the bubble function is a constant function. \section{Conclusion}\label{sec:conclusion} We considered the finite difference and the finite element methods for a model boundary value problem. We emphasized the connections between the two methods by rewriting the FD system such that the matrix of the system coincides with the stiffness matrix of the FE discretization. In this reformulation, the right hand side vector for the FD system can be also viewed as a dual vector obtained componentwise from the composite trapezoid rule approximation of the FE dual vector. Using the connection between the Green function and the $C^0-P^1$ finite element basis functions, we found that the inverse of the stiffness matrix $S$ is exactly the matrix obtained by evaluating the Green function at the interior grid nodes $\{(x_i, x_j)\} \subset (0, 1) \times (0, 1)$. Consequently, we found simplified proofs for the standard FD error analysis, and provided an energy estimate for the difference between the FE and FD solutions. We presented interesting possible extensions to the 2D case based on the fact that the components of the FE dual vector associated with the given data $f$ acting on conforming finite element basis functions, can be written as linear combinations of various degrees of freedom acting on the solution $u$. The challenges of extending these ideas to the more general cases of discrete spaces and meshes in 2D or 3D, remain to be addressed in our future work. \vspace{0.4in} \paragraph*{\bf Acknowledgment.} The authors are grateful for an Oden Institute fellowship which allowed them to visit UT Austin in the Fall of 2019. \bibliographystyle{plain}

https://arxiv.org/abs/0706.4100

Embedding nearly-spanning bounded degree trees

We derive a sufficient condition for a sparse graph G on n vertices to contain a copy of a tree T of maximum degree at most d on (1-\epsilon)n vertices, in terms of the expansion properties of G. As a result we show that for fixed d\geq 2 and 0<\epsilon<1, there exists a constant c=c(d,\epsilon) such that a random graph G(n,c/n) contains almost surely a copy of every tree T on (1-\epsilon)n vertices with maximum degree at most d. We also prove that if an (n,D,\lambda)-graph G (i.e., a D-regular graph on n vertices all of whose eigenvalues, except the first one, are at most \lambda in their absolute values) has large enough spectral gap D/\lambda as a function of d and \epsilon, then G has a copy of every tree T as above.

\section{Introduction} In this paper we obtain a sufficient condition for a sparse graph $G$ to contain a copy of every nearly-spanning tree $T$ of bounded maximum degree, in terms of the expansion properties of $G$. The restriction on the degree of $T$ comes naturally from the fact that we consider graphs of constant degree. Two important examples where our condition applies are random graphs and graphs with a large spectral gap. The random graph $G(n,p)$ denotes the probability space whose points are graphs on a fixed set of $n$ vertices, where each pair of vertices forms an edge, randomly and independently, with probability $p$. We say that the random graph $G(n,p)$ possesses a graph property $\cal P$ {\em almost surely}, or a.s. for short, if the probability that $G(n,p)$ satisfies $\cal P$ tends to 1 as the number of vertices $n$ tends to infinity. The problem of existence of large trees with specified shape in random graphs has a long history with most of the results being devoted to finding a long path. Erd{\cal H}{o}s conjectured that a random graph $G(n,c/n)$ a.s. contains a path of length at least $(1-\alpha(c))n$, where $\alpha(c)$ is a constant smaller than one for all $c>1$ and $\lim_{c \rightarrow \infty} \alpha(c)=0$. This conjecture was proved by Ajtai, Koml\'os and Szemer\'edi \cite{AKS} and, in a slightly weaker form, by Fernandez de la Vega \cite{FD1}. These results were significantly improved by Bollob\'as \cite{B}, who showed that $\alpha(c)$ decreases exponentially in $c$. Finally Frieze \cite{F} determined the correct speed of convergence of $\alpha(c)$ to zero and proved that $\alpha(c)=(1+o(1))ce^{-c}$. The question of existence of large trees of bounded degree other than paths in sparse random graphs was studied by Fernandez de la Vega in \cite{FD2}. He proved that there exist two constants $a_1>0$ and $a_2>0$ such that for fixed tree $T$ of order $n/a_1$ with maximum degree at most $d$ the random graph $G(n, c/n)$ with $c=a_2d$ almost surely contains $T$. The constant $a_1$ in this result is rather large and allows to embed only trees that occupy a small proportion of the random graph. Also, observe that Fernandez de la Vega's result gives the almost sure existence of a {\em fixed} tree $T$, and not of all such trees simultaneously. Our first theorem improves the result of Fernandez de la Vega and generalizes the above mentioned results on the existence of long paths. It shows that the sparse random graph contains almost surely every nearly-spanning tree of bounded degree. \begin{theorem} \label{random} Let $d\geq 2$, $0<\varepsilon<1/2$ and let $$c \geq \frac{10^6 d^3 \log d \log^2 (2/\varepsilon)}{\varepsilon}.$$ Then almost surely the random graph $G(n, c/n)$ contains every tree of maximum degree at most $d$ on $(1-\varepsilon)n$ vertices. \end{theorem} Results guaranteeing the existence of a long path in a sparse graph can be obtained in a more general situation when the host graph has certain expansion properties. Given a graph $G=(V,E)$ and a subset $X \subset V$ let $N_G(X)$ denote the set of all neighbors of vertices of $X$ in $G$. Using his celebrated rotation-extension technique, P\'osa \cite{P} proved that if for every $X$ in $G$ with $|X|\leq k$ we have that $|N_G(X)\setminus X| \geq 2|X|-1$, then $G$ contains a path of length $3k-2$. A remarkable generalization of this result from paths to trees of bounded degree was obtained by Friedman and Pippenger \cite{FP}. They proved that if $|N_G(X)| \geq (d+1)|X|$ for every set $X$ in $G$ with $|X|\leq 2k-2$, then $G$ contains every tree with $k$ vertices and maximum degree at most $d$. Note that this result allows to embed only trees whose size is relatively small compared to the size of $G$. What if we want to embed trees which are nearly-spanning? It turns out that a slightly stronger expansion property, based on the spectral gap condition, is already enough to attain this goal. For a graph $G$ let $\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_n$ be the eigenvalues of its adjacency matrix. The quantity $\lambda(G)= \max_{i \geq 2} |\lambda_i|$ is called the {\em second eigenvalue} of $G$. A graph $G=(V,E)$ is called an {\em $(n,D,\lambda)$-graph} if it is $D$-regular, has $n$ vertices and the second eigenvalue of $G$ is at most $\lambda$. It is well known (see, e.g., \cite{AloSpe00} for more details) that if $\lambda$ is much smaller than the degree $D$, then $G$ has strong expansion properties, so the ratio $D/\lambda$ could serve as some kind of measure of expansion of $G$. Our next result shows that an $(n,D,\lambda)$-graph $G$ with large enough spectral gap $D/\lambda$ contains a copy of every nearly-spanning tree with bounded degree. \begin{theorem} \label{pseudo-random} Let $d\geq 2$, $0<\varepsilon<1/2$ and let $G$ be an $(n,D,\lambda)$-graph such that $$\frac{D}{\lambda} \geq \frac{160d^{5/2}\log (2/\varepsilon)}{\varepsilon}.$$ Then $G$ contains a copy of every tree $T$ with $(1-\varepsilon)n$ vertices and with maximum degree at most $d$. \end{theorem} Our main results are tightly connected to the notion of universal graphs. For a family ${\cal H}$ of graphs, a graph $G$ is ${\cal H}$-universal if $G$ contains every member of ${\cal H}$ as a (not necessarily induced) subgraph. The construction of sparse universal graphs for various families arises in several fields such as VLSI circuit design, data representation and parallel computing (see, e.g., the introduction of \cite{ACKRRS} for a short survey and relevant references). Our two main results show that sparse random graphs and pseudo-random graphs on $n$ vertices are universal graphs for the family of bounded-degree trees on $(1-\varepsilon)n$ vertices. Quite an extensive research exists on universal graphs for trees \cite{BCLR}, \cite{CG1}, \cite{CG2}, \cite{CG3}, \cite{CGP}, \cite{FP}. The most interesting result is that of Bhatt et al. who showed in \cite{BCLR} that there exists a universal graph $G$ on $n$ vertices for the family of trees on $n$ vertices with maximum degree $d$, whose maximum degree is bounded by a function of $d$. It is instructive to compare our results with those of \cite{BCLR}: they succeed in embedding spanning trees as opposed to nearly-spanning in our case; on the other hand, their universal graph is a concrete carefully constructed graph that has very dense pieces locally, while we provide a very large family of universal graphs possessing many additional properties that can be useful for obtaining further results on universal graphs. The results of Theorem \ref{random} and \ref{pseudo-random} can be deduced from a more general statement which we present next. We need the following definition. \begin{defin}\label{def1} Given two positive numbers $c$ and $\alpha<1$, a graph $G=(V,E)$ is called an $(\alpha,c)$-expander if every subset of vertices $X\subset V(G)$ of size $|X|\le \alpha |V(G)|$ satisfies: $$ |N_G(X)|\ge c |X|\ . $$ \end{defin} \begin{theorem}\label{th1} Let $d\ge 2$, $0<\varepsilon<1/2$. Let $G=(V,E)$ a graph on $n$ vertices of minimum degree $\delta=\delta(G)$ and maximum degree $\Delta=\Delta(G)$. Let $n, \delta, \Delta$ satisfy: \begin{enumerate} \item (order of graph is large enough) $$n \ge \frac{480d^3\log (2/\varepsilon)}{\varepsilon};$$ \item (maximum degree is not too large compared to the minimum degree) $$\Delta^2 \le \frac{1}{K} e^{\delta/(8K)-1}~~ \mbox{where} ~~K=\frac{20 d^2 \log (2/\varepsilon)}{\varepsilon}.$$ \item (local expansion) Every induced subgraph $G_0$ of $G$ with minimum degree at least $\frac{\varepsilon \delta}{40d^2 \log (2/\varepsilon)}$ is a $(\frac{1}{2d+2},d+1)$-expander. \end{enumerate} Then $G$ contains a copy of every tree $T$ on at most $(1-\varepsilon)n$ vertices of maximum degree at most $d$. \end{theorem} The rest of this paper is organized as follows. In the next two sections we show how Theorem \ref{th1} can be used to embed nearly-spanning trees of bounded degree in random and pseudo-random graphs. We present the proof of Theorem \ref{pseudo-random} first, since it is short and less technical, and then prove Theorem \ref{random}. In Section 4 we describe the plan of the proof of Theorem \ref{th1} and discuss some technical tools needed to fulfill this plan. The proof of this theorem appears in Section 5. The last section of the paper contains several concluding remarks and open problems. Throughout the paper we make no attempts to optimize the absolute constants. To simplify the presentation, we often omit floor and ceiling signs whenever these are not crucial. Throughout the paper, $\log$ denotes logarithm in the natural base $e$. \section{Embedding in pseudo-random graphs} In this section we prove Theorem \ref{pseudo-random}. First we need the following lemma that shows that an $(n,D,\lambda)$-graph has the local expansion property required by Condition 3 of Theorem \ref{th1}. \begin{lemma}\label{prg} Let $d\ge 2$. Let $G=(V,E)$ be an $(n,D,\lambda)$-graph. Denote $$ D_0=\frac{2\lambda (d+1)}{\sqrt{d}}\ . $$ Then every induced subgraph $G_0$ of $G$ of minimum degree at least $D_0$ is a $\big(\frac{1}{2d+2},d+1\big)$-expander. \end{lemma} {\bf Proof.\hspace{1.5em}} We will use the following well known estimate on the edge distribution of an $(n,D,\lambda)$-graph $G$ (see, e.g., \cite{AloSpe00}, Corollary 9.2.5). For every two (not necessarily disjoint) subsets $B,C\subseteq V$, let $e(B,C)$ denote the number of ordered pairs $(u,v)$ with $u \in B, v \in C$ such that $uv$ is an edge. Note that if $u,v \in B \cap C$, then the edge $uv$ contributes $2$ to $e(B,C)$. In this notation, $$ \left|e(B,C)-\frac{|B||C|D}{n}\right|\le \lambda\sqrt{|B||C|}\ . $$ Let $U$ be a subset of vertices of $G$ such that the induced subgraph $G_0=G[U]$ has minimum degree at least $D_0$. Suppose that the claim is false. Then there exists a subset $X\subset U$ of size $|X|=t \leq |U|/(2d+2)$ satisfying $|N_{G_0}(X)|< (d+1)|X|$. By the above estimate with $B=X$ and $C=N_{G_0}(X)$ we have: $$ D_0t\le e(B,C)\le \frac{t(d+1)tD}{n} +\lambda t \sqrt{ d+1}\, $$ and therefore \begin{equation}\label{eq1} \frac{t}{n} \ge \frac{D_0}{(d+1)D}-\frac{\lambda}{\sqrt{d+1}D}\ . \end{equation} Also, note that there are no edges of $G$ from $X$ to $Y=U-(X \cup N_{G_0}(X))$ as $G_0$ is an induced subgraph of $G$. From $t=|X|\le |U|/(2d+2)$ and $|N_{G_0}(X)| \leq (d+1)t$ it follows that $|Y|\ge dt$. Thus $$ 0=e(X,Y)\ge \frac{t (dt) D}{n}-\lambda \sqrt{t (dt)}\,, $$ implying \begin{equation}\label{eq2} \frac{t}{n}\le \frac{\lambda}{\sqrt{d} D}\ . \end{equation} Comparing (\ref{eq1}) and (\ref{eq2}) we obtain $$ \frac{D_0}{(d+1)D}-\frac{\lambda}{\sqrt{d+1}D}\le \frac{\lambda}{\sqrt{d}D}\ . $$ Plugging in the definition of $D_0$ we derive a contradiction. \hspace*{\fill}\mbox{$\Box$} \vspace{0.25cm} \noindent {\bf Proof of Theorem \ref{pseudo-random}.}\hspace{1.5em} Since every graph with minimal degree $k$ contains all the trees on $k$ vertices we can assume that $D \leq (1-\varepsilon)n$. Let $A$ be the adjacency matrix of $G$. The trace of $A^2$ equals the number of ones in $A$, which is exactly $2|E(G)|=nD$. We thus obtain that $$ nD=Tr(A^2)=\sum_{i=1}^n\lambda_i^2\le D^2+(n-1)\lambda^2 $$ and therefore $\lambda^2 \geq \frac{D(n-D)}{n-1} \geq \varepsilon D$. This together with our assumption on $D/\lambda$ implies $$n \geq D \geq \varepsilon\left(\frac{ D}{\lambda}\right)^2 \geq \frac{160^2 d^5 \log^2 (2/\varepsilon)}{\varepsilon}. $$ Since $\Delta(G)=\delta(G)=D$, from this inequality it follows that $G$ satisfies Conditions 1, 2 of Theorem \ref{th1}. Finally since $$ \frac{\varepsilon D}{40 d^2 \log (2/\varepsilon)} \geq \frac{2(d+1)\lambda}{\sqrt{d}}$$ we can conclude using Lemma \ref{prg} that $G$ also satisfies the last condition of Theorem \ref{th1}. Thus $G$ contains every tree of size $(1-\varepsilon)n$ with maximum degree at most $d$. \hfill $\Box$ \section{Embedding in random graphs} To prove Theorem \ref{random} we first need to show that a sparse random graph contains a.s. a nearly spanning subgraph with good local expansion properties. \begin{lemma} \label{rg} For every integer $d\ge 2$, real $0<\theta<1/2$ and $D \geq 50\theta^{-1}$ the random graph $G\big(n,\frac{4D}{n}\big)$ almost surely contains a subgraph $G^*$ having the following properties: \begin{enumerate} \item $|V(G^*)|\ge (1-\theta)n$; \item $D\le d_{G^*}(v)\le 10D$ for every $v\in V(G^*)$; \item every induced subgraph $G_0$ of $G^*$ of minimum degree at least $D_0=100d\log D$ is a $\big(\frac{1}{2d+2},d+1\big)$-expander. \end{enumerate} \end{lemma} The following statement contains a few easy facts about random graphs. \begin{propos} \label{properties} Let $G(n,p)$ be a random graph with $np>20$, then almost surely \noindent $(i)$\, The number of edges between any two disjoint subsets of vertices $A, |A|=a$ and $B, |B|=b$ with $abp \geq 32n$ is at least $abp/2$ and at most $3abp/2$. \noindent $(ii)$\, Every subset of vertices of size $a \leq n/4$ spans less than $anp/2$ edges. \end{propos} {\bf Proof.\hspace{1.5em}} (i)\, Since the number of edges between $A$ and $B$ is a binomially distributed random variable with parameters $ab$ and $p$, it follows by the standard Chernoff-type estimates (see, e.g., \cite{AloSpe00}) that (denoting $t=abp/2$) $$\mathbb{P}\Big[e(A,B)-abp < -t\Big] \leq e^{-\frac{t^2}{2abp}}=e^{-abp/8}$$ and $$\mathbb{P}\Big[e(A,B)-abp > t\Big] \leq e^{-\frac{t^2}{2abp}+\frac{t^3}{2(abp)^2}}=e^{-abp/16}.$$ Using that $abp\geq 32n$ we can bound the probability that there are sets $A, B$ with $|e(A,B)-abp|>abp/2$ by $2^n \cdot 2^n \cdot \big(2e^{-2n}\big)=o(1)$. \noindent Since $np/2 \geq 10$ and $n/a \geq 4$, the probability that there is a subset of size $a$ which violates the assertion (ii) is at most $$\mathbb{P}_a \leq {n \choose a} {a^2/2 \choose anp/2}p^{anp/2} \leq \left(\frac{en}{a} \Big(\frac{ea}{np}\Big)^{np/2} p^{np/2}\right)^a= \left(\frac{e^{np/2+1}}{(n/a)^{np/2-1}}\right)^a \leq \left(\frac{e^{11}}{(n/a)^9}\right)^a.$$ It is easy to see that $\mathbb{P}_a \ll n^{-1}$ for all $a \leq n/4$ and so $\sum_a\mathbb{P}_a=o(1)$. \hspace*{\fill}\mbox{$\Box$} \vspace{0.25cm} \noindent {\bf Proof of Lemma \ref{rg}.}\hspace{1.5em} Let $G=G(n,p)$ be a random graph with $p=\frac{4D}{n}$ and let $X$ be the set of $\theta n/2$ vertices of largest degrees in $G$. By Part (ii) of Proposition \ref{properties}, a.s. this set spans less than $|X|np/2=2D|X|$ edges. Also, since $\frac{4D}{n}|X|(n-|X|) \geq 2D \theta (n/2)\geq 50 n$, Part (i) of this proposition implies that a.s. the number of edges between $X$ and $V(G)-X$ is at most $3|X|np/2=6D|X|$. Therefore the sum of the degrees of the vertices in $X$ is bounded by $10D|X|$ and hence there is a vertex in $X$ with degree at most $10D$. By definition of $X$, this implies that there are at most $\theta n/2$ vertices in $G$ with degree larger than $10D$. Delete these vertices and denote the remaining graph by $G'$. Next as long as $G'$ contains a vertex $v$ of degree less than $D$, delete it. If we deleted more than $\theta n/2$ vertices, then the original random graph contains two sets $Y$ and $V(G')-Y$ such that $|Y|=\theta n/2$, $|V(G')-Y|\geq (1-\theta)n\geq n/2$ and there are less than $D|Y|\leq p|Y||V(G')-Y|/2$ edges between them. Since $\frac{4D}{n}|Y||V(G')-Y| \geq \theta D n\geq 50 n$, again by Part (i) of the previous statement this a.s. does not happen. Denote the resulting graph by $G^*$. Then it satisfies the first two conditions of the lemma and it remains to verify the third condition. Suppose to the contrary that $G^*$ contains a subset of vertices $U$ such that the induced subgraph $G_0=G^*[U]$ has minimum degree at least $D_0=100d\log D$ and is not a $\big(\frac{1}{2d+2},d+1\big)$-expander. Then there exists a set $X\subset U$ of size $|X|=t$ such that the set $C= N_{G_0}(X)$ has size at most $(d+1)t$ and there are at least $D_0|X|/2=50dt\log D$ edges with an end in $X$ and another end in $C$. If $t\le \frac{\log D}{D}n$, then the probability that $G(n,p)$ contains such sets is at most \begin{eqnarray*} \mathbb{P}_t &\leq& {n \choose t}{n\choose {(d+1)t}}{{{t(d+1)t}}\choose{50d t\log D}} p^{50dt \log D }\\ &\le& \left[\left(\frac{en}{t} \right) \left( \frac{en}{(d+1)t}\right)^{d+1} \left(\frac{e(d+1)tp}{50d\log D}\right)^{50d\log D}\right]^t\\ &\le&\left[e\,\left(\frac{n}{t}\right)^{2d} \left(\frac{e}{8}\cdot\frac{Dt}{n\log D}\right)^{50d \log D} \right]^t\\ &=& \left[e\,\left(\frac{e}{8}\right)^{50d \log D} \left(\frac{D}{\log D}\right)^{2d} \left(\frac{Dt}{n\log D} \right)^{50d \log D-2d}\right]^t \\ &<& \left[e^{-25d \log D+2d\log D+1} \left(\frac{t}{n\log D/D} \right)^{40d \log D}\right]^t \\ &\leq& \left[D^{-20d} \left(\frac{t}{n\log D/D} \right)^{40d\log D}\right]^t \ . \end{eqnarray*} Checking separately two cases $t < {\log n}$ and ${\log n} \leq t < \frac{\log D}{D}n$ it is easy to see that in both $\mathbb{P}_t \ll n^{-1}.$ If $t\ge \frac{\log D}{D}n$ we apply a different argument. Note that there are no edges of $G(n,p)$ from $X$ to $C=U-(X \cup N_{G_0}(X))$ since $G_0$ is an induced subgraph. From $t=|X|\le |U|/(2d+2)$ and $|N_{G_0}(X)| \leq (d+1)t$ it follows that $|C|\ge dt$ and therefore the probability of such event in $G(n,p)$ is at most \begin{eqnarray*} \mathbb{P}_t&\leq& {n\choose t}{{n}\choose{dt}}(1-p)^{dt^2}\le \left[\frac{en}{t}\,\cdot\, \left(\frac{en}{dt}\right)^de^{-pdt} \right]^t \\ &\leq& \left[\left(\frac{en}{t}\right)^{2d}e^{-pdt}\right]^t = \left[\left(\frac{en}{t}\right)^2\,\cdot\,e^{-pt}\right]^{dt}\\ &\leq& \left[\left(\frac{en}{n\log D/D}\right)^2\,\cdot\, e^{-\frac{4D}{n}\,\cdot\,\frac{\log D}{D}n}\right]^{dt}\\ &\leq& \left( D^2 D^{-4}\right)^{dt}=o(n^{-1})\ . \end{eqnarray*} Thus the probability that $G^*$ fails to satisfy the third condition is at most $\sum_{t=1}^n \mathbb{P}_t=o(1)$. \hfill $\Box$ \vspace{0.25cm} \noindent {\bf Proof of Theorem \ref{random}.} \hspace{1.5em} Let $d \geq 2$, $0<\varepsilon<1/2$ and $c$ satisfy the assumption of Theorem \ref{random}. Set $\theta=0.01\varepsilon$, $D=c/4$ and let $\varepsilon_1=\frac{\varepsilon-\theta}{1-\theta}\geq 0.99\varepsilon$. Then by Lemma \ref{rg} $G(n, c/n)$ almost surely contains a subgraph $G^*$ of order $n_1\geq (1-\theta)n$ such that $D \leq \delta(G^*)\leq \Delta(G^*) \leq 10D$ and every induced subgraph of $G^*$ with minimum degree at least $100d \log D$ is an $\big(\frac{1}{2d+2},d+1\big)$-expander. Using that $\Delta(G^*) \leq 10\delta(G^*)$ and $$n_1 \geq \delta(G^*) \geq D \geq \frac{10^6d^3\log d \log^2 (2/\varepsilon)}{4\varepsilon}> \frac{480 d^3 \log(2/\varepsilon_1)}{\varepsilon_1}$$ we conclude that $G^*$ satisfies Conditions 1 and 2 of Theorem \ref{th1} (with $\varepsilon_1$). To verify the third condition it is enough to check that the assumptions in Theorem \ref{random} imply that $$100d\log D \leq \frac{\varepsilon_1 D}{40d^2 \log (2/\varepsilon_1)}.$$ Note that one can simply substitute the lower bound for $D$ in the above expression, since $x/\log x$ is an increasing function for $x>3$. Therefore by Theorem \ref{th1}, $G^*$ contains every tree of size $(1-\varepsilon_1)n_1 \geq (1-\varepsilon_1)(1-\theta)n=(1-\varepsilon)n$ with maximum degree at most $d$. \hfill $\Box$ \section{Embedding plan and main tools} To prove Theorem \ref{th1} we will use the following framework. Given a tree $T$, we first cut it into subtrees $T_1, T_2, \ldots ,T_s$ of carefully chosen sizes, so that the number $s$ of these subtrees satisfies $s \leq 10d^2 \log(2/\varepsilon)$, and each subtree $T_i$ is connected by a unique edge to the union of all previous subtrees. The subtrees $T_i$ will be embedded sequentially in order, starting from $T_1$. We then choose $s$ pairwise disjoint sets of vertices $S_1,S_2, \ldots, S_s$ whose total size is at most $\varepsilon n/2$, such that each vertex of the graph has many neighbors in each set $S_i$. The set $S_i$ will be used only when embedding the subtree $T_i$, and will not be touched before that step. (It can be used later, but we will not do it here, as it complicates matters and does not improve the estimates in any essential way). During the embedding process we maintain a set $R$ of at most $s$ vertices, which will consist of all roots of the trees $T_i$ that still have to be embedded, and will not contain any vertex of the sets $S_i$. At the $i$-th step we are to embed the tree $T_i$ starting from a given root $x_i\in R$. (At the first step a root is chosen arbitrarily.) Suppose that the current set of unused vertices of $G$ is $V_{i-1}$. We take an arbitrary subset $U_i$ of size $|U_i|=\Theta(|V(T_i)|d)$ which contains the vertex $x_i$ that will be the root of $T_i$ (but contains no other members of $R$), contains the set $S_i$, and contains no member of $S_j$ for $j \neq i$. Note that as each vertex has many neighbors in $S_i$, the minimum degree in the induced subgraph of $G$ on $U_i$ is large. Since the minimum degree of $G[U_i]$ is large enough, we can use the result of Friedman and Pippenger to embed a copy of $T_i$ in $U_i$, rooting it at $x_i$. (We actually need a slightly modified, rooted version, of their result). All vertices of $U_i$ unused when embedding $T_i$ are recycled, and we thus get $V_i$ by deleting from $V_{i-1}$ only the vertices used for embedding $T_i$. The final step of embedding $T_i$ is to embed the edges crossing from $T_i$ to yet unembedded pieces $T_j$, with $j>i$ (using vertices of $U_i$). We then add the endpoints of those edges outside $T_i$ to the list $R$ of special vertices, and delete $x_i$ from $R$. Each of the newly added special vertices will serve as a root for embedding the corresponding piece $T_j$. Observe that the number of special vertices is at most $s$ at any stage of the embedding. The precise technical details are described in what follows. \subsection{The result of Friedman and Pippenger} The cornerstone of our proof is the embedding result of Friedman and Pippenger. In fact, we need a slightly stronger version of it -- they showed the existence of a tree $T$, while we need to embed a rooted version of $T$ in $G$ starting from a fixed vertex $v\in V(G)$ as its root. Luckily, a careful reading of \cite{FP} reveals that the following holds as well. \begin{theorem}[\cite{FP}]\label{FP} Let $T$ be a tree on $k$ vertices of maximum degree at most $d$ rooted at $r$. Let $H=(V,E)$ be a non-empty graph such that, for every $X\subset V(H)$ with $|X|\le 2k-2$, $$ |N_H(X)|\ge (d+1) |X|\ . $$ Let further $v\in V(H)$ be an arbitrary vertex of $H$. Then $H$ contains a copy of $T$, rooted at $v$. \end{theorem} \subsection{Cutting the tree into pieces} \begin{propos}\label{tree-cut1} Let $d \geq 2$ and $k$ be positive integers. Let $T$ be a tree on at least $k+1$ vertices with maximum degree at most $d$. Then there exists an edge $e\in E(T)$ such that at least one of the two trees obtained from $T$ by deleting $e$ has at least $k$ and at most $(d-1)(k-1)+1$ vertices. \end{propos} {\bf Proof.\hspace{1.5em}} Choose a leaf $r$ of $T$ arbitrarily and root $T$ at $r$. For $i\ge 0$ denote by $L_i$ the set of vertices of $T$ at distance $i$ from $r$. For a vertex $v\in V(T)$ let $t(v)$ be the number of vertices in the subtree of $T$ rooted at $v$. Now, let $$ i_0=\max\{i: L_i\mbox{ contains a vertex $v$ with $t(v)\ge k$}\}\ . $$ As $L_1$ has only one vertex $v$ with $t(v)=|V(T)|-1\ge k$, it follows that $i_0\ge 1$. Choose a vertex $u\in L_{i_0}$ such that $t(u)\ge k$. Then by the definition of $i_0$ all sons $w$ of $u$ in $T$ satisfy: $t(w)\le k-1$, the number of sons does not exceed $d-1$, and therefore $t(u)\le (d-1)(k-1)+1$. Let now $x$ be the father of $u$ in $T$. Then $e=(x,u)$ is the required edge. \hspace*{\fill}\mbox{$\Box$} \begin{corollary}\label{tree-cut2} Suppose $0 < \varepsilon <1/2$, and let $T$ be an arbitrary tree on $(1-\varepsilon)n$ vertices, with maximum degree at most $d$. Then one can cut $T$ into $s$ subtrees $T_1,T_2, \ldots ,T_s$, so that each tree $T_i$ is connected by a unique edge to the union of all trees $T_j$ with $j<i$, and such that for every $i>1$, $$ \frac{\varepsilon n/2 +\sum_{j>i} |V(T_j)|}{8d^2} \leq |V(T_i)| \leq \frac{\varepsilon n/2 +\sum_{j>i} |V(T_j)|}{8d}. $$ For $i=1$, the upper bound holds, but the lower bound may fail. Moreover, $s \leq 10d^2 \log (2/\varepsilon).$ \end{corollary} {\bf Proof.\hspace{1.5em}} We choose the trees $T_i$ one by one, starting from the {\em last} one. By Proposition \ref{tree-cut1} we first find a tree $T'_1$ of size at least $\frac{\varepsilon n}{16d^2}$ and at most $\frac{\varepsilon n}{16d}$, and omit it from $T$. Suppose we have already chosen $T'_1, T'_2, \ldots ,T'_{i-1}$ such that for every $j<i$ $$ \frac{\varepsilon n/2 +\sum_{r<j} |V(T'_r)|}{8d^2} \leq |V(T'_j)| \leq \frac{\varepsilon n/2 +\sum_{r<j} |V(T'_r)|}{8d}, $$ and each $T'_j$ has a unique edge joining it to $V(T)\setminus \cup_{r<j}V(T'_r)$. Let $T'$ be the tree obtained from $T$ by omitting all the vertices of all subtrees $T'_j$, $j<i$. If the number of vertices of $T'$ is at most $\frac{\varepsilon n/2 +\sum_{r<i} |V(T'_r)|}{8d},$ then define $T'_i=T'$ and $s=i$. Else, apply Proposition \ref{tree-cut1} to find a tree $T'_i$ in $T'$ whose size is at least $\frac{\varepsilon n/2 +\sum_{r<i} |V(T'_r)|}{8d^2}$ and at most $\frac{\varepsilon n/2 +\sum_{r<i} |V(T'_r)|}{8d}$, and continue. To estimate the number of steps in this process define $a_i=\varepsilon n/2+\sum_{j\leq i}|V(T'_j)|$. Observe that $a_0=\varepsilon n/2, a_i\leq \varepsilon n/2+|V(T)|\leq n$ and $a_i=a_{i-1}+|V(T'_i)| \geq \left(1+\frac{1}{8d^2}\right)a_{i-1}$. From this it follows that $$\frac{2}{\varepsilon} \geq \frac{a_i}{a_0} \geq \left(1+\frac{1}{8d^2}\right)^i,$$ and hence this process terminates after at most $10 d^2 \log (2/\varepsilon)$ steps. Finally, for $1\le i\le s$ define $T_i=T'_{s-i+1}$. \hspace*{\fill}\mbox{$\Box$} \subsection{Splitting vertex degrees} \begin{lemma}\label{l44} Let numbers $K, \delta,\Delta$ satisfy $$ K\Delta^2 e^{-(\delta/8K)+1} <1. $$ Then the following holds. Let $H=(V,E)$ be a graph in which $\delta \le d(v)\le \Delta$ for each $v\in V$. Then $H$ contains $K$ pairwise disjoint sets of vertices $S_1, S_2, \ldots ,S_K$ such that every vertex of $H$ has at least $\frac{\delta }{2K}$ neighbors in each set $S_i$. \end{lemma} {\bf Proof.\hspace{1.5em}} This is a simple consequence of the Lov\'asz Local Lemma (c.f., e.g., \cite{AloSpe00}, Chapter 5). Color the vertices of $H$ randomly and independently by $K$ colors. For each vertex $v$ and color $i$, $1 \leq i \leq K$, let $A_{v,i}$ be the event that $v$ has less than $\frac{\delta }{2K}$ neighbors of color $i$. By Chernoff's Inequality the probability of each event $A_{v,i}$ is at most $e^{-\delta/(8K)}$. In addition, each event $A_{v,i}$ is mutually independent of all events but the events $A_{u,j}$ where either $u=v$ or $u$ and $v$ have common neighbors in $H$. As there are less than $K\big(\Delta(\Delta-1)+1\big)\leq K\Delta^2$ such events, it follows by the Local Lemma that with positive probability none of the events $A_{v,i}$ holds. The desired result follows, by letting $S_i$ denote the set of all vertices of color $i$. \hspace*{\fill}\mbox{$\Box$} \section{Proof of Theorem \ref{th1}} Let $G=(V,E)$ be a graph satisfying the assumptions of the theorem. Let $T$ be a tree on at most $(1-\varepsilon)n$ vertices with maximum degree at most $d$. By Corollary \ref{tree-cut2} the tree $T$ can be partitioned into subtrees $T_1, T_2, \ldots, T_s$ satisfying the conditions of the Corollary, where $s \leq 10 d^2 \log (2/\varepsilon).$ Choose an arbitrary root for $T_1$. For $i>1$, the root of $T_i$ is the vertex incident with the unique edge that connects $T_i$ to the union of the previous trees. Put $K=2s/\varepsilon$. By Condition 2 in the theorem, and Lemma \ref{l44} there are $K$ pairwise disjoint sets of vertices $S_i$ of $G$ such that every vertex of $G$ has at least $ \delta/(2K) \geq \frac{\varepsilon \delta}{40 d^2 \log (2/\varepsilon)}$ neighbors in each set $S_i$. Take the $s$ smallest sets $S_i$ and renumber them so that they are denoted by $S_1,S_2,\ldots ,S_s$. Obviously, their total size is at most $\frac{ns}{K} =\frac{\varepsilon n}{2}$. We will not use all the other sets $S_i, i>s$ in the rest of the proof. Let $x_1$ be an arbitrary vertex of $G$ that does not lie in any of the sets $S_i$. Define $R=\{x_1\}$ and let $U_1$ denote the set of all vertices of $G$ besides those in $\cup_{j \neq 1} S_j$. As $U_1$ contains $S_1$, every vertex in the induced subgraph $G[U_1]$ of $G$ on $U_1$ has degree at least $\frac{\varepsilon \delta}{40 d^2 \log (2/\varepsilon)}$. Therefore, by Condition 3 in Theorem \ref{th1}, and by Theorem \ref{FP} there is a copy of $T_1$ in $G[U_1]$ rooted at $x_1$. (Note that by Corollary \ref{tree-cut2}, the size of $T_1$ is at most $\frac{|U_1|}{8d}$ and hence indeed one can apply here Theorem \ref{FP}.) Moreover, we can in fact embed in $U_1$ the required tree $T_1$ together with the edges connecting it to the trees $T_j$ with $j>1$. Add the endpoints of these edges to the list $R$ of planned roots for the trees $T_j$, and delete $x_1$ from $R$. This completes the embedding of $T_1$. Assume that we have already embedded the first $i-1$ trees $T_r$, where each $T_r$ has been rooted in the vertex of $R$ specified as its root, and where in step number $r$ the tree $T_r$ has been embedded using no vertices of $R$ besides its root, and no vertices of $\cup_{j \neq r} S_j$, we proceed to the $i$-th step, in which we are to embed the tree $T_i$ starting from a given root $x_i \in R$. Let $U_i$ be the set of all vertices of $G$ that have not been used for embedding the part of $T$ embedded so far, besides the vertices in $R-\{x_i\}$ and besides the vertices in $\cup_{j \neq i} S_j$. As before, since $U_i$ contains $S_i$, every vertex in the induced subgraph $G[U_i]$ of $G$ on $U_i$ has degree at least $\frac{\varepsilon \delta}{40 d^2 \log (2/\varepsilon)}$. Therefore, by Condition 3 in Theorem \ref{th1}, it is a $(\frac{1}{2d+2},d+1)$-expander. By Corollary \ref{tree-cut2}, the size of $T_i$ is at most $|U_i|/8d$ and the number of edges connecting $T_i$ to the trees $T_j$ for $j>i$ is bounded by $$s \leq 10 d^2 \log (2/\varepsilon) \leq \frac{\varepsilon n}{48d} \leq \frac{|U_i|}{24d}.$$ Therefore the size of $T_i$ together with the vertices in $T_j$ for $j>i$ that are connected to it is less than a fraction $\frac{1}{6d} \leq \frac{1}{2(2d+2)}$ of the size of $U_i$. Therefore, by Theorem \ref{FP} we can embed $T_i$ including the edges connecting it to the trees $T_j$ with $j>i$ in $U_i$, rooting it at $x_i$, and add the endpoints of the edges from $T_i$ to future $T_j$'s to $R$. As this process can be carried out until we finish the embedding of $T_s$, the assertion of the theorem follows. \hspace*{\fill}\mbox{$\Box$} \section{Concluding remarks} \begin{itemize} \item Our lower bound on the edge probability of a random graph in Theorem \ref{random} seems far from being best possible, and the correct order of magnitude should probably be more similar to the case of a longest path. Hence it is likely that already when $c=O(d\log (1/\epsilon))$ the random graph $G(n,c/n)$ contains a.s. every tree on $(1-\epsilon)n$ vertices with maximum degree at most $d$. \item Embedding {\em spanning} trees of bounded degree in sparse random graphs is an intriguing question which is completely open. In case of the path, this question is very well understood (see, e.g., Chapter 8 of \cite{Bol}) and it is known that for $p=O(\log n/n)$ the random graph $G(n,p)$ a.s. contains a Hamiltonian path. We believe that a more general result should be true, i.e., such a random graph should already contain a.s. every tree on $n$ vertices with maximum degree at most $d$. Our methods are insufficient to attack this problem. Using Theorem \ref{random} we can only prove the following much weaker result. Let $T$ be a tree on $n$ vertices with at least $\epsilon n$ leaves, then there exists a constant $a(\epsilon,d)$ such that the random graph $G\big(n,\frac{a\log n}{n}\big)$ a.s. contains $T$. Here is a brief sketch of the proof: we split $G(n,p)$ into two random graphs $G(n,p_1)$ and $G(n,p_2)$, where $1-p=(1-p_1)(1-p_2)$, $p_1=\Theta(1/n)$, $p_2=\Theta(\log n/n)$. Let $T'$ be the tree obtained from $T$ by deleting its leaves. We use Theorem \ref{random} to embed a copy of $T'$ in $G(n,p_1)$. Then we expose the edges of $G(n,p_2)$ between the set of vertices $V_0$ of $G$, not occupied by a copy of $T'$, and the rest of the graph, and embed the leaves of $T$ in $V_0$ using matching-type results. \item Besides the model $G(n,p)$, another model of random graphs, drawing a lot of attention is the model of random regular graphs. A random regular graph $G_{n,D}$ is obtained by sampling uniformly at random over the set of all simple $D$-regular graphs on a fixed set of $n$ vertices. By the result of Friedman, Kahn and Szemer\'edi \cite{FKS} the second eigenvalue of $G_{n,D}$ is almost surely at most $O(\sqrt{D})$ (see \cite{Fr} for a more precise result). Therefore our Theorem \ref{pseudo-random} immediately implies that if $D=D(d,\epsilon)$ is sufficiently large then $G_{n,D}$ a.s. contains every tree on $(1-\epsilon)n$ vertices with maximum degree $d$. This result as well as the result of Theorem \ref{pseudo-random} are probably not optimal. We suspect that sufficiently large spectral gap (as a function of $d$ only) already suffices to guarantee the embedding of every spanning tree of bounded degree in a graph $G$ of order $n$. This is not known even for the Hamiltonian path, and the best result in this case, obtained in \cite{KS}, requires spectral gap of order roughly $\log n$. \end{itemize} \vspace{0.15cm} \noindent {\bf Acknowledgment.}\, A major part of this work was carried out when the authors were visiting Microsoft Research at Redmond, WA. We would like to thank the members of the Theory Group at Microsoft Research for their hospitality and for creating a stimulating research environment.

https://arxiv.org/abs/1803.09633

Direct Proofs of the Fundamental Theorem of Calculus for the Omega Integral

When introduced in a 2018 article in the American Mathematical Monthly, the omega integral was shown to be an extension of the Riemann integral. Although results for continuous functions such as the Fundamental Theorem of Calculus follow immediately, a much more satisfying approach would be to provide direct proofs not relying on the Riemann integral. This note provides those proofs.

\section{Introduction} The omega integral, which makes use of the hyperreals to integrate real functions, was introduced in \cite{monthly} and proven to be an extension of the Riemann integral. Theorems relating to continuous functions, such as integrability, additivity, and the fundamental theorem follow immediately. It is natural to seek direct proofs of such theorems (that is, proofs that do not rely on the Riemann integral). The purpose of this note is to provide these proofs. We will make use of the following notational conventions from \cite{monthly}: 0 will not be considered to be an infinitesimal and $\Omega$ is often used to represent a positive unlimited hyperreal integer and $\omega:=1/\Omega$ for its infinitesimal reciprocal. In addition, capital Greek letters generally represent unlimited hyperreals, and lowercase Greek letters generally represent infinitesimals. Otherwise, notation and basic facts about the hyperreals will follow \cite{goldblatt}, although we write the standard part operator $\mathop{\rm st}(\cdot)$ in place of the shadow. All references to integrability in this article will refer to the omega integral. For convenience we restate below the definition of the omega integral from \cite{monthly}, which is based on right-hand sums for equal-width partitions using an unlimited number of subintervals. First, we remind the reader that any function $f:[a,b]\rightarrow {\mathbb R}$ defined on a subinterval of ${\mathbb R}$ admits a natural extension to a function ${}^*f:{}^*[a,b]\rightarrow {}^*{\mathbb R}$ defined on the hyperreal interval ${}^*[a,b] = \{x\in {}^*{\mathbb R} \mid a\leq x\leq b\}$ (see \cite{goldblatt}). For convenience, we will refer to both the original function $f$ and its hyperreal extension ${}^*f$ by the same name $f$. The omega integral is then defined as follows: \begin{definition}[Omega integral] Let $f:[a,b]\rightarrow{\mathbb R}$ be a function. If there exists $L\in \ $ such that $\mathop{\rm st}\left(\sum_{k=1}^\Omega f(x_k)\Delta x\right)=L$ for every positive unlimited hyperreal integer $\Omega$ (where $\Delta x=\frac{b-a}{\Omega}$ and $x_k=a+k\Delta x$ for $k=1,2,\dots,\Omega$), then we write $$\int_a^b f(x)\,dx =L$$ and call $f$ integrable. If $\sum_{k=1}^\Omega f(x_k)\Delta x$ is positive infinite for every $\Omega$ we write $\int_a^b f(x)\,dx =\infty$, and for negative infinite sums we use $-\infty$. \end{definition} We call $\sum_{k=1}^\Omega f(x_k)\Delta x$ an \emph{omega sum} irregardless of the name of the unlimited hyperreal integer used therein. The following lemma will also be helpful. \begin{lemma}\label{L1} Let $\Omega$ be a positive unlimited integer, let $M$ be a positive real number, and let $\gamma_k$ be 0 or infinitesimal for $k=1,\dots,\Omega$. Then $\sum_{k=1}^\Omega \gamma_k M\frac{1}{\Omega}$ is 0 or infinitesimal.\end{lemma} \begin{proof} Let $\varepsilon>0$ be real. Since $\gamma_k$ is 0 or infinitesimal, $-\frac{\varepsilon}{M}<\gamma_k<\frac{\varepsilon}{M}$. Hence $\sum_{k=1}^{\Omega}-\frac{\varepsilon}{M}< \sum_{k=1}^{\Omega}\gamma_k<\sum_{k=1}^{\Omega}\frac{\varepsilon}{M}$, and multiplying by the positive constant $M\frac{1}{\Omega}$ gives $\sum_{k=1}^{\Omega}\frac{-\varepsilon}{\Omega}< \sum_{k=1}^{\Omega}\gamma_k M\frac{1}{\Omega}<\sum_{k=1}^{\Omega}\frac{\varepsilon}{\Omega}=\frac{\varepsilon}{\Omega}\Omega=\varepsilon$. Since $\varepsilon$ was arbitrary, $\sum_{k=1}^\Omega \gamma_k M\frac{1}{\Omega}$ is 0 or infinitesimal. \end{proof} \section{Integrability of Continuous Functions} Intuitively, a partition using an unlimited number of subintervals only leaves room for an approximation to a continuous function to stray infinitesimally far from the value of the function on real numbers, and on a finite-length interval the resulting approximation is within an infinitesimal of the value we seek. To prove that any two such sums have the same standard part, we show that the sum for a common refinement of the partitions is within an infinitesimal of each sum. \begin{theorem}\label{A} Let $f$ be continuous on $[a,b]$. Then $f$ is integrable. \end{theorem} \begin{proof} Let $f$ be continuous on $[a,b]$. Then $f$ is bounded and each omega sum is limited. Let $\Omega$ be a positive unlimited hyperreal integer and consider its omega sum: \begin{equation}\label{1} \sum_{k=1}^\Omega f(x_k)\Delta x, \text{ where } \Delta x=(b-a)\cdot\frac{1}{\Omega} \text{ and }x_k=a+k(b-a)\cdot\frac{1}{\Omega}. \end{equation} Then for any positive unlimited hyperreal integer $B$, we form the omega sum based upon $B\Omega$ subintervals, calling the partition points $y_k$ to distinguish them from the partition points of the previous sum: \begin{equation}\label{2} \sum_{k=1}^{B\Omega}f(y_k)\Delta y,\text{ where } \Delta y=(b-a)\cdot\frac{1}{B\Omega} \text{ and } y_k=a+k(b-a)\cdot\frac{1}{B\Omega}.\end{equation} We wish to show that the omega sums in \eqref{1} and \eqref{2} differ by at most an infinitesimal. To that end, fix $n$ with $0\le n\le\Omega-1$. If $1\le p\le B$, then \begin{align*} x_n=y_{Bn}<y_{Bn+p} &=a+(Bn+p)(b-a)\frac{1}{B\Omega}\\ &=a+n(b-a)\frac{1}{\Omega}+p(b-a)\frac{1}{B\Omega}\\ &=x_n+(b-a)\frac{p}{B}\frac{1}{\Omega}\\ &\le x_n+(b-a)\frac{1}{\Omega}=x_{n+1}.\end{align* Since $x_n$ and $x_{n+1}$ differ by an infinitesimal and $f$ is continuous, $$f(y_{Bn+p})=f(x_{n+1})+\gamma_{n,p}$$ for each such $n$ and $p$, where $\gamma_{n,p}$ is 0 or infinitesimal (see \cite{goldblatt}). Then \begin{align*} \sum_{k=1}^{B\Omega}f(y_k)\Delta y&=\sum_{n=0}^{\Omega-1}\sum_{p=1}^B f(y_{Bn+p})(b-a)\frac{1}{B\Omega}\\ &=\sum_{n=0}^{\Omega-1}\sum_{p=1}^B\left(f(x_{n+1})+\gamma_{n,p}\right)(b-a)\frac{1}{B\Omega}\\ &=\sum_{n=0}^{\Omega-1}\sum_{p=1}^Bf(x_{n+1})(b-a)\frac{1}{B\Omega}+\sum_{n=0}^{\Omega-1}\sum_{p=1}^B\gamma_{n,p}(b-a)\frac{1}{B\Omega}\\ &=\sum_{n=0}^{\Omega-1}B\cdot f(x_{n+1})(b-a)\frac{1}{B\Omega}+\sum_{n=0}^{\Omega-1}\sum_{p=1}^B\gamma_{n,p}(b-a)\frac{1}{B\Omega}\\ &=\left(\sum_{n=1}^\Omega f(x_n)(b-a)\frac{1}{\Omega}\right)+\sum_{n=0}^{\Omega-1}\sum_{p=1}^B\gamma_{n,p}(b-a)\frac{1}{B\Omega}\\ &=\sum_{k=1}^\Omega f(x_n)\Delta x+\gamma \end{align* where $\gamma$ is 0 or infinitesimal, the last step by lemma \ref{L1}. Now consider any two omega sums \begin{equation}\label{7} \sum_{k=1}^{B_1}f(x_k)\Delta x\end{equation} and \begin{equation}\label{8}\sum_{k=1}^{B_2}f(y_k)\Delta y,\end{equation} where $B_1$ and $B_2$ are arbitrary positive unlimited integers. By the previous work, the omega sum based on the common partition refinement for $B_1B_2$, $$\sum_{k=1}^{B_1B_2}f(z_k)\Delta z,$ has the same standard part as each of the omega sums in \eqref{7} and \eqref{8}. Consequently all omega sums have the same standard part and $f$ is integrable.\end{proof} \section{Additivity for Continuous Functions} In \cite{monthly} it is shown that the omega integral is not additive in general, but through the Riemann integral we know that the omega integral is additive for continuous functions. The difficulty presented in a direct proof using omega sums is that the partition points of an omega sum for the integral on the right in Theorem \ref{B} might not match the partition points for any omega sum for one or both integrals on the left, a situation that occurs whenever $\frac{c-a}{b-a}$ is irrational. In a similar manner to the proof of Theorem \ref{A}, as long as our partition points are infinitesimally close then the corresponding values of $f$ will be as well, allowing us to show the appropriate omega sums are close enough. \begin{theorem}\label{B} If $f$ is continuous on $[a,c]$ and $a<b<c$ then $$\int_a^b f(x)\,dx+\int_b^c f(x)\,dx=\int_a^c f(x)\,dx.$$ \end{theorem} \begin{proof} Let $\Omega$ be a positive unlimited integer and form the associated omega sum for $f$ on $[a,c]$: $$\sum_{k=1}^\Omega f(x_k)\Delta x, \text{ where } \Delta x=(c-a)\cdot\frac{1}{\Omega} \text{ and }x_k=a+k(c-a)\cdot\frac{1}{\Omega}.$$ Since $a<b<c$, there exists a positive unlimited integer $B<\Omega$ such that \begin{equation}\label{12} x_B=a+B\Delta x\le b<a+(B+1)\Delta x=x_{B+1}.\end{equation} Using $B$ subintervals we may form an omega sum for $f$ on $[a,b]$: $$\sum_{k=1}^B f(y_k)\Delta y, \text{ where } \Delta y=(b-a)\cdot\frac{1}{B} \text{ and }y_k=a+k(b-a)\cdot\frac{1}{B}.$$ Then although the partition points $x_k$ and $y_k$ may not be equal, they differ by at most an infinitesimal; in fact, for each $k\in\{1,2,\dots,B\}$, $$\left|x_k-y_k\right|\le\left|x_B-y_B\right|=\left|x_B-b\right|<\Delta x,$ the last inequality stemming from \eqref{12}. Then $|x_k-y_k|$ is at most infinitesimal, and since $f$ is continuous $|f(x_k)-f(y_k)|$ is also at most infinitesimal. Therefore \begin{equation}\label{15} \begin{split} &\left|\sum_{k=1}^B f(y_k)\Delta y-\sum_{k=1}^B f(x_k)\Delta x\right|\\ &\qquad=\left|\sum_{k=1}^B f(y_k)\Delta y-\left(\sum_{k=1}^B f(x_k)\Delta y+\sum_{k=1}^Bf(x_k)(\Delta x-\Delta y)\right)\right|\\ \qquad&\le\left|\sum_{k=1}^B\left(f(y_k)-f(x_k)\right)\Delta y\right|+\left|\sum_{k=1}^B f(x_k)(\Delta x-\Delta y)\right|.\end{split}\end{equation} The term on the left is infinitesimal by Lemma \ref{L1}. To see how $\Delta x$ and $\Delta y$ compare, use the fact that $b=y_B=a+B\Delta y$ combined with \eqref{12}, $$a+B\Delta x\le a+B\Delta y<a+(B+1)\Delta x,$ to conclude that $$0\le\Delta y-\Delta x<\frac{\Delta x}{B}.$ Then the remaining term from \eqref{15} yields \begin{align*} \left|\sum_{k=1}^B f(x_k)(\Delta x-\Delta y)\right| &=\left|\Delta x-\Delta y\right|\cdot\left|\sum_{k=1}^B f(x_k)\right|\\ &<\frac{\Delta x}{B}\left|\sum_{k=1}^B f(x_k)\right|\\ &=\frac 1B\left|\sum_{k=1}^B f(x_k)\Delta x\right|\\ &\le \frac1B\sum_{k=1}^B|f(x_k)|\Delta x\\ &\le\frac1B\sum_{k=1}^\Omega|f(x_k)|\Delta x\\ &=\frac 1B\left(\int_a^c |f(x)|\,dx+\text{infinitesimal}\right) \end{align*} since $|f|$ is continuous and hence integrable by Theorem \ref{A}. Therefore the last quantity in the above inequalities is also at most infinitesimal, allowing the conclusion that \begin{equation}\label{20} \left|\sum_{k=1}^B f(y_k)\Delta y-\sum_{k=1}^B f(x_k)\Delta x\right| \end{equation} is infinitesimal. Now form the omega sum for $f$ on $[b,c]$ using $\Omega-B$ subintervals: $$\sum_{k=1}^{\Omega-B} f(z_k)\Delta z, \text{ where } \Delta z=(c-b)\cdot\frac{1}{\Omega-B} \text{ and }z_k=b+k(c-b)\cdot\frac{1}{\Omega-B}.$$ A similar argument to the one above shows that \begin{equation}\label{22} \left|\sum_{k=1}^{\Omega-B}f(z_k)\Delta z-\sum_{k=B+1}^{\Omega}f(x_k)\Delta x\right|\end{equation} is infinitesimal. Combining \eqref{20} and \eqref{22}, \begin{align*} \sum_{k=1}^\Omega f(x_k)\Delta x&=\sum_{k=1}^Bf(x_k)\Delta x+\sum_{k=B+1}^\Omega f(x_k)\Delta x\\ &=\sum_{k=1}^Bf(y_k)\Delta y+\gamma_1+\sum_{k=1}^{\Omega-B}f(z_k)\Delta z+\gamma_2\end{align* where $\gamma_1$ and $\gamma_2$ are 0 or infinitesimal. But since $f$ is continuous, by Theorem \ref{A} each of these omega sums is within an infinitesimal of its associated integral and the definition of omega integral yields $$\int_a^c f(x)\,dx=\int_a^b f(x)\,dx+\int_b^c f(x)\,dx.$ \end{proof} \section{Fundamental Theorem of Calculus, Part I} The usual proof of the fundamental theorem given in a calculus course still works for the omega integral. There are, however, a few details to be careful about. \begin{theorem}\label{C} If $f$ is continuous on $[a,b]$, then the function $$F(x)=\int_a^x f(t)\,dt$$ is differentiable on $[a,b]$ and $F'(x)=f(x)$. \end{theorem} \begin{proof} Let $f$ be continuous on $[a,b]$. Then for any $c\in[a,b]$, by Theorem \ref{A} we may define $F_c:[a,b]\to{\bold R}$ by $$F_c(x)=\int_c^x f(t)\,dt.$ As with any real function, $F_c$ extends to the domain ${}^*[a,b]$ consisting of hyperreals between $a$ and $b$ inclusive. We can then associate the integral notation with the extended function; that is, for any hyperreal $x\in{}^*[a,b]$, $$\int_c^x f(t)\,dt=F_c(x).$ The transfer property allows us to extend the properties of integrals as well. For instance, for each real $c\in[a,b]$, we write $$\forall x\in[a,b] \left(m\le f(t)\le M\,\, \forall t\in[c,x]\right)\Longrightarrow m(x-c)\le F_c(x)\le M(x-c)$ transfers to \begin{equation}\label{34} \forall x\in{}^*[a,b] \left(m\le f(t)\le M\,\, \forall t\in{}^*[c,x]\right)\Longrightarrow m(x-c)\le F_c(x)\le M(x-c).\end{equation} The additivity we will also need transfers. Fixing $x\in[a,b]$ for each real $\alpha$ such that $x+\alpha\in[a,b]$, we have \begin{equation}\label{35} \int_a^x f(t)\,dt+\int_x^{x+\alpha}f(t)\,dt=\int_a^{x+\alpha}f(t)\,dt.\end{equation} By transfer the same is true for any hyperreal $\alpha$ such that $x+\alpha\in{}^*[a,b]$. Fix $x\in[a,b]$. In order to prove that $F'(x)=f(x)$ we need to show that for any infinitesimal $\alpha$, $$\mathop{\rm st}\left(\frac{F(x+\alpha)-F(x)}{\alpha}\right)=f(x).$ To that end, let $\alpha$ be an infinitesimal such that $x+\alpha\in {}^*[a,b]$. Using the definition of $F$ and \eqref{35}, \begin{equation}\label{37} \frac{F(x+\alpha)-F(x)}{\alpha}=\frac{\int_a^{x+\alpha}f(t)\,dt-\int_a^x f(t)\,dt}{\alpha}=\frac{\int_x^{x+\alpha}f(t)\,dt}{\alpha}.\end{equation} Since $f$ is continuous, $|f(t)-f(x)|$ is infinitesimal for each $t$ between $x$ and $x+\alpha$. Thus for each positive real $\varepsilon$, $$f(x)-\varepsilon<f(t)<f(x)+\varepsilon$ for each such $t$, and by \eqref{34} $$\left(f(x)-\varepsilon\right)\alpha\le\int_x^{x+\alpha}f(t)\,dt\le\left(f(x)+\varepsilon\right)\alpha.$ Dividing by $\alpha$ and combining with \eqref{37} shows that $$f(x)-\varepsilon<\frac{F(x+\alpha)-F(x)}{\alpha}<f(x)+\varepsilon,$ or $$\left|\frac{F(x+\alpha)-F(x)}{\alpha}-f(x)\right|<\varepsilon.$ As $\varepsilon$ is arbitrary, $\left|\frac{F(x+\alpha)-F(x)}{\alpha}-f(x)\right|$ is infinitesimal for each $\alpha$ and we conclude that $F'(x)=f(x)$. \end{proof} Notice that throughout the proof the lower limits of integration are always real, so that there is no need to extend to functions $F_c$ for $c\in{}^*[a,b]$. \section{Fundamental Theorem of Calculus, Part II} The usual calculus-level proof of FTC~II from FTC~I applies. However, just as with the Riemann integral, it is also possible to give a direct proof of FTC~II, one that does not require proving FTC~I first. We begin with a lemma. \begin{lemma}\label{L2} If $H'=f$ is continuous on $[a,b]$, then for any hyperreal $x\in{}^*[a,b]$ and infinitesimal $\alpha$ for which $x+\alpha\in{}^*[a,b]$, one has $$H'(x)=\frac{H(x+\alpha)-H(x)}{\alpha}+\gamma$$ where $\gamma$ is 0 or infinitesimal. \end{lemma} \begin{proof} Let $x,x_0\in{\mathbf R}$ such that both $x$ and $x+x_0$ are in $[a,b]$. Then the mean value theorem may be applied to $H$ since $H$ is differentiable on $[a,b]$. Thus there is some $c\in{\mathbf R}$ such that $c$ is between $x$ and $x+x_0$ and $x_0H'(c)=H(x+x_0)-H(x)$. Then by transfer, for all $x,x_0\in{}^*{\mathbf R}$ such that $x,x+x_0\in {}^*[a,b]$, there exists $c\in{}^*{\mathbf R}$ such that $c$ is between $x$ and $x+x_0$ and $x_0H'(c)=H(x+x_0)-H(x)$. Let $x$ be a hyperreal in ${}^*[a,b]$ and let $\alpha$ be an infinitesimal such that $x+\alpha\in{}^*[a,b]$. Then there is some hyperreal $c$ between $x$ and $x+\alpha$ such that $\alpha H'(c)=H(x+\alpha)-H(x)$. But since $H'=f$ is continuous on $[a,b]$ and $c$ differs from $x$ by an infinitesimal, $H'(c)=H'(x)+\gamma$ where $\gamma$ is 0 or infinitesimal. Hence $H(x+\alpha)-H(x)=\alpha H'(c)=\alpha(H'(x)+H'(c)-H'(x))=\alpha H'(x)+\alpha\gamma$ and the conclusion follows. \end{proof} \begin{theorem}\label{D} If $f$ is continuous on $[a,b]$ and $H$ is any antiderivative of $f$ on $[a,b]$, then $$\int_a^b f(x)\,dx=H(b)-H(a).$$ \end{theorem} \begin{proof} We are given that $H'(x)=f(x)$ on $[a,b]$. Given a positive unlimited integer $\Omega$, its associated omega sum is \begin{equation}\label{51}\sum_{k=1}^\Omega H'(x_k)\Delta x= \sum_{k=1}^\Omega H'(a+k(b-a)\omega)\cdot (b-a)\omega,\end{equation} writing $\omega$ for $\frac{1}{\Omega}$. Then using Lemma \ref{L2} with $x=a+k(b-a)\omega$ and $\alpha=(b-a)\omega$, \begin{eqnarray*} & & \sum_{k=1}^\Omega H'(a+k(b-a)\omega)\cdot (b-a)\omega\\ & = & \left(\sum_{k=1}^{\Omega-1}\left(\frac{H(a+k(b-a)\omega+(b-a)\omega)-H(a+k(b-a)\omega)}{(b-a)\omega}+\gamma_k\right)(b-a)\omega\right)\\ & & +H'(b)\cdot(b-a)\omega \end{eqnarray*} where $\gamma_k$ is 0 or infinitesimal for each $k$. Writing $\beta$ for the $H'(b)(b-a)\omega$, which is 0 or infinitesimal, and simplifying yields the telescoping sum \begin{eqnarray*} & = & \beta+\sum_{k=1}^{\Omega-1}\left(H(a+(k+1)(b-a)\omega)-H(a+k(b-a)\omega)+\gamma_k(b-a)\omega\right)\\ & = & \beta + H(a+2(b-a)\omega)-H(a+1(b-a)\omega)+\gamma_1(b-a)\omega\\ & & +H(a+3(b-a)\omega)-H(a+2(b-a)\omega)+\gamma_2(b-a)\omega\\ & & +H(a+4(b-a)\omega)-H(a+3(b-a)\omega)+\gamma_3(b-a)\omega\\ & & +\dots\\ & & +H(a+\Omega(b-a)\omega)-H(a+(\Omega-1)(b-a)\omega)+\gamma_{\Omega-1}(b-a)\omega\\ & = & \beta+H(a+\Omega(b-a)\omega)-H(a+1(b-a)\omega)+\sum_{k=1}^{\Omega-1}\gamma_k(b-a)\omega. \end{eqnarray*} By Lemma \ref{L1}, $\gamma=\sum_{k=1}^{\Omega-1}\gamma_k(b-a)\omega$ is 0 or infinitesimal. Since $H$ is continuous at $a$ (it is differentiable there), $H(a+(b-a)\omega)=H(a)+\eta$ where $\eta$ is 0 or infinitesimal. We then arrive at $$\sum_{k=1}^\Omega H'(a+k(b-a)\omega)\cdot (b-a)\omega=H(b)-H(a)-\eta+\beta+\gamma,$ the standard part of which is $H(b)-H(a)$. Since $\Omega$ was arbitrary the proof is complete. \end{proof}

https://arxiv.org/abs/0802.3523

On linear versions of some addition theorems

Let K \subset L be a field extension. Given K-subspaces A,B of L, we study the subspace spanned by the product set AB = {ab | a \in A, b \in B}. We obtain some lower bounds on the dimension of this subspace and on dim B^n in terms of dim A, dim B and n. This is achieved by establishing linear versions of constructions and results in additive number theory mainly due to Kemperman and Olson.

\section{Introduction} Let $G$ be a group, written multiplicatively. Given subsets $A,B \subset G$, we denote by \begin{equation*} AB = \{ab \mid a \in A, b\in B \} \end{equation*} the \textit{product set} of $A,B$. For $A,B$ finite, several results in additive number theory give estimates on the cardinality of $AB$ as a function of $|A|,|B|$. Instances of such results are the Cauchy-Davenport theorem for cyclic groups of prime order, a theorem of Kneser for abelian groups, and theorems of Kemperman and of Olson for possibly nonabelian groups. In this paper, we consider the following analogous question. Given a field extension $K\subset L$, and finite-dimensional $K$-subspaces $A,B$ of $L$, define \begin{equation*} \langle AB\rangle =\text{ the $K$-subspace spanned by the product set $AB$ in $L$.} \end{equation*} What can be said, then, about the dimension of the subspace $\langle AB\rangle $? The main object of this paper is to establish linear analogues, in this new setting, of several results in additive number theory. In particular, we shall obtain nontrivial lower bounds on $\dim \langle AB\rangle$ in terms of $\dim A$, $\dim B$. This question has barely been addressed in the literature. What seems to be the so far unique result of this type is due to Hou, Leung and Xiang \cite {Xian} and is given below. We first recall Kneser's theorem from additive number theory \cite{Kn53, Kn55}. \begin{theorem}[Kneser] \label{Kneser} Let $G$ be an abelian group and let $A,B\subset G$ be finite nonempty subsets. Then \begin{equation*} |AB|\geq |AH|+|BH|-|H|, \end{equation*} where $H=\{g\in G\mid gAB=AB\}$ is the stabilizer of $AB$. \end{theorem} The following linear version has been obtained in \cite{Xian}, and has motivated us to further explore the links between additive number theory, linear algebra and field extensions. \begin{theorem}[Hou, Leung and Xiang] \label{linear Kneser} \label{TH_HLX} Let $K\subset L$ be a commutative field extension, and let $A,B\subset L$ be nonzero finite-dimensional $K$-subspaces of $L$. Suppose that every algebraic element in $L$ is separable over $K.$ Let $H$ be the subfield of $L$ which stabilizes $\langle AB\rangle.$ Then \begin{equation*} \dim _{K}\langle AB\rangle \geq \dim _{K}A+\dim _{K}B-\dim _{K}H. \end{equation*} \end{theorem} The paper is organized as follows. Several classical results and constructions in additive number theory are recalled in Section~\ref{sec-kem}. In Section 3, we give a new variant of a theorem of Olson. The switch to the field extension setting is performed from Section~\ref{sec_linear setting} on, where we give linear analogues of the additive results stated in the preceding sections. \section{Classical addition theorems\label{sec-kem}} \subsection{A basic result} One of the simplest results on the product of two sets in a finite group is the following. Given a subset $X$ of a group $G$, we denote $X^{-1}=\{x^{-1} \mid x \in X \}$. \begin{proposition} \label{basic} Let $G$ be a finite group. Let $A,B$ be nonempty subsets of $G$. If $|A|+|B|>|G|$, then $AB=G$. \end{proposition} The proof, informally, goes as follows. For $x \in G$, we have $A^{-1}x \cap B \neq \emptyset$, since $|A^{-1}x|+|B|$ = $|A|+|B| > |G|$. It follows that $x \in AB$. \smallskip Linearizing this result requires a little more work. This is done in Section~\ref{lin ols}, with various proofs. \subsection{Theorems of Kemperman and Olson} Here are a few classical results in additive number theory, to be linearized in subsequent sections. We refer the reader to \cite{ols} for proofs. \begin{theorem}[Kemperman] \label{th_ke}\label{th_kem}\label{THK} Let $A,B$ be two finite subsets of a group $G.\;$Assume there exists an element $c\in AB$ appearing exactly once as a product $c=ab$ with $a\in A$, $b\in B$. Then \begin{equation*} \left| AB\right| \geq \left| A\right| +\left| B\right| -1. \end{equation*} \end{theorem} The next result is a nonabelian analogue of Kneser's Theorem~\ref{Kneser}. \begin{theorem}[Olson] \label{thOl} Let $A,B$ be two finite subsets of a group $G.\;$There exists a nonempty subset $S$ of $AB$ and a finite subgroup $H$ of $G$ such that \begin{equation*} \left| AB\right| \geq \left| S\right| \geq \left| A\right| +\left| B\right| -\left| H\right| \end{equation*} and either $HS=S$ or $SH=S$. \end{theorem} Here is an immediate corollary. \begin{theorem}[Kemperman] \label{torsion-free} Let $G$ be a torsion-free group. Let $A,B$ be nonempty finite subsets of $G$. Then \begin{equation*} \left| AB\right| \geq \left| A\right| +\left| B\right| -1. \end{equation*} \end{theorem} \noindent Olson \cite{ols} also derived, as a consequence of Theorem~\ref {th_kem} and \ref{thOl}, the following two interesting results. Linear versions are given in Section~\ref{sec: powers}. \begin{theorem}[Olson] \label{thOl2} Let $A,B$ be finite subsets of a group $G$ with $1\in B$. Then \begin{equation*} AB^{2}=AB\;\text{ or }\;\left| AB\right| \geq \left| A\right| +\frac{1}{2}\left| B\right|. \end{equation*} \end{theorem} \begin{theorem}[Olson] \label{Th_ol3} Let $B$ be a finite subset of a group $G$. Then \begin{equation*} \left| B^{n}\right| =\left| B^{n+1}\right| \;\text{ or }\;\left| B^{n}\right| \geq \left| B^{n-1}\right| +\frac{1}{2}\left| B\right| . \end{equation*} \end{theorem} \subsection{A main tool: the Kemperman transform}\label{tool} The above results are obtained by cleverly iterating Kemperman transforms, which we now recall. Let $G$ be a (possibly nonabelian) multiplicative group and $(A,B)$ a pair of nonempty finite subsets of $G.\;$Let $x$ be any element in $G$.\ The Kemperman transformed pairs $(A^{\prime },B^{\prime })$ and $(A^{\prime \prime },B^{\prime \prime })$ with respect to $x$ are defined by \begin{equation*} (A^{\prime} = A\cup Ax, \, B^{\prime} = B\cap x^{-1}B) \end{equation*} and \begin{equation*} (A^{\prime\prime} = A \cap Ax^{-1}, \, B^{\prime\prime} = B\cup xB). \end{equation*} The following properties are straightforward to check. \begin{enumerate} \item $A^{\prime }B^{\prime }\subset AB$ \ and\ $A^{\prime \prime }B^{\prime \prime }\subset AB$; \item $B^{\prime }\neq \emptyset $ $\Longleftrightarrow $ $x\in BB^{-1}$\ and\ $A^{\prime \prime }\neq \emptyset $ $\Longleftrightarrow $ $x\in A^{-1}A$; \item $\left| A^{\prime }\right| >\left| A\right| $ $\Longleftrightarrow $ $Ax\neq A$ \ and \ $\left| B^{\prime \prime }\right| >\left| B\right| $ $\Longleftrightarrow $ $xB\neq B$. \end{enumerate} \begin{proposition}[Kemperman \protect\cite{Kem}] Let $(A,B)$ be a pair of nonempty finite subsets in a group $G.\;$Assume that there is an element $d\in A^{-1}A\cap BB^{-1}$ such that either $Ad\neq A$ or $dB\neq B$. Then there is a pair of nonempty subsets $(A_{1},B_{1})$, obtained as a Kemperman transform of $(A,B)$, satisfying: \begin{enumerate} \item[(1)] $A_{1}B_{1}\subset AB,$ \item[(2)] either $\left| A_{1}\right| +\left| B_{1}\right| =\left| A\right| +\left| B\right| $ and $\left| A_{1}\right| >\left| A\right| ,$ or $\left| A_{1}\right| +\left| B_{1}\right| >\left| A\right| +\left| B\right| .$ \end{enumerate} \end{proposition} \section{A variant of a theorem of Olson} \label{variant olson} In this section we give a variant of Theorem~\ref{thOl2} which is neither stronger nor weaker, in the sense that neither result implies the other one. A linearized version of this variant will be presented in Section~\ref{sec: lin ABC}. \begin{theorem} \label{ABC} Let $A,B,C$ be nonempty finite subsets of a group $G$. Suppose $B\subset C$ and $1\in C.$ Then \begin{equation*} ABC=AB\;\text{ or }\;\left| ABC\right| \geq \left| A\right| +\left| B\right| . \end{equation*} \end{theorem} \begin{proof} The proof is close to that of Theorem \ref{thOl2} in \cite{ols}. If $C=\{1\}$ then $ABC=AB.$ We now assume $|C|\geq 2$ and proceed by induction on $\left| AB\right| .$ If $\left| AB\right| =1$ then $\left| A\right| =\left| B\right| =1,$ and $|ABC|=|C|\geq 2\geq |A|+|B|$. Assume now $\left| AB\right| >1$ and our statement true for any pair $(A_{1},B_{1})$ such that $B_{1}\subset C$ and $\left| A_{1}B_{1}\right| <\left| AB\right| .$ As in Olson's original proof of Theorem \ref{thOl2}, we consider two cases. \medskip \noindent $(i)$ Assume there exists a nonempty subset $S\subset AB$ such that $SC=S$. Since $S$ is finite, we have $Sc=S$ for all $c\in C$, whence $S{C}^{-1}=S.$ Set $A_{0}=A\cap S$ and $A_{1}=A\setminus S,$ so that $A=A_{0}\sqcup A_{1}$. We claim that \begin{equation} A_{0}B=S. \label{eq: A0B} \end{equation} Indeed, as $B\subset C$, we have $A_{0}B\subset SC=S$. Conversely, let $s\in S$. Then $s=ab$ for some $a\in A$ and $b\in B$. Since $a=sb^{-1}\in SC^{-1}=S,$ it follows that $a\in A_{0}$, whence $s\in A_{0}B$ as desired. As a first consequence, we obtain \begin{equation} AB=S\sqcup A_{1}B. \label{eq: AB} \end{equation} Indeed, we have $AB=A_{0}B\cup A_{1}B=S\cup A_{1}B$. Moreover, as $S\cap A_{1}=\emptyset $, we have $SB^{-1}\cap A_{1}=\emptyset $ since $SB^{-1}\subset SC^{-1}=S$. It follows that $S\cap A_{1}B=\emptyset .$ Our next claim is that \begin{equation} ABC=S\sqcup A_{1}BC. \label{eq: ABC} \end{equation} Indeed, we have $ABC=SC\cup A_{1}BC=S\cup A_{1}BC$. The intersection $S\cap A_{1}BC$ is indeed empty, since $S=SC^{-1}$ and so $SC^{-1}\cap A_{1}B=S\cap A_{1}B=\emptyset $. Consequently, we derive \begin{equation} |ABC|\geq |S|+|A_{1}BC|\geq |A_{0}|+|A_{1}BC|, \label{eq: cardABC} \end{equation} where the estimate $|S|\geq |A_{0}|$ follows from the equality $S=A_{0}B$ in (\ref{eq: A0B}). We may assume $A_{1}\not=\emptyset $, for otherwise $S=A_{0}B=AB$, whence $ABC=SC=S=AB$ and we are done. We have $|A_{1}B|<|AB|$ by (\ref{eq: AB}). By our induction hypothesis, we either have $A_{1}BC=A_{1}B$ or $|A_{1}BC|\geq |A_{1}|+|B|$. If $A_{1}BC=A_{1}B$, then by (\ref{eq: ABC}) and (\ref{eq: AB}), we obtain $ABC=S\cup A_{1}BC=S\cup A_{1}B=AB$ and we are done in this case. If now $|A_{1}BC|\geq |A_{1}|+|B|$, then by (\ref{eq: cardABC}) and $|A|=|A_{0}|+|A_{1}|$, we get \begin{equation*} |ABC|\geq |A_{0}|+|A_{1}BC|\geq |A_{0}|+|A_{1}|+|B|=|A|+|B|, \end{equation*} and we are done in this case as well. \medskip \noindent $(ii)$ Assume now $XC\neq X$ for every nonempty subset $X\subset AB.$ By Theorem \ref{thOl}, there exists a nonempty subset $S\subset AB$ and a subgroup $H$ such that \begin{equation} \left| S\right| \geq \left| A\right| +\left| B\right| -\left| H\right| \label{ineqO} \end{equation} and $HS=S$ or $SH=S.$ Let $c_{0}\in C$ be such that $Sc_{0}\neq S.\;$ We claim that \begin{equation} |S\cup Sc_{0}|\geq |S|+|H|. \label{ineq} \end{equation} Indeed, if $HS=S,$ then $H(S\cup Sc_{0})=S\cup Sc_{0},$ and hence $S\cup Sc_{0}$ is a disjoint union of left cosets of $H.$ This implies (\ref{ineq}) in this case. On the other hand, if $SH=S$, pick $s_{0}\in S$ such that $s_{0}c_{0}\notin S $, and set $H^{\prime }=H\cup \{c_{0}\}$. In the set $SH^{\prime }$, the element $s_{0}c_{0}$ appears exactly once, since $SH\subset S.$ Applying Theorem \ref{th_kem}, it follows that \begin{equation} \left| SH^{\prime }\right| \geq \left| S\right| +\left| H^{\prime }\right| -1=\left| S\right| +\left| H\right| . \label{ineqST} \end{equation} This implies (\ref{ineq}) again, since $SH^{\prime }=S\cup Sc_{0}$. \smallskip Now, as $\{1,c_{0}\}\subset C$, we have $S\cup Sc_{0}\subset SC$, and therefore \begin{equation*} \left| ABC\right| \geq |SC|\geq \left| S\cup Sc_{0}\right| . \end{equation*} Combined with (\ref{ineqO}) and (\ref{ineq}), this gives \begin{equation*} \left| ABC\right| \geq \left| S\right| +\left| H\right| \geq \left| A\right| +\left| B\right| . \end{equation*} \end{proof} In the abelian case, the previous theorem is much shorter to prove, using Kneser's theorem, and remains true without the assumption $B\subset C$. \begin{theorem} Let $A,B,C$ be nonempty finite subsets of an abelian group $G$. Suppose $1\in C.$ Then \begin{equation*} ABC=AB\;\text{ or }\;\left| ABC\right| \geq \left| A\right| +\left| B\right| . \end{equation*} \end{theorem} \begin{proof} By Kneser's theorem, we obtain \begin{equation*} \left| AB\right| \geq \left| A\right| +\left| B\right| -\left| H\right| , \end{equation*} where $H$ is the stabilizer of $AB$ in $G$.\ We have $HAB=AB$ and hence $HABC=ABC$. In particular, both $AB$ and $ABC$ are disjoint unions of cosets of $H$. We have $AB\subset ABC$ since $1\in C.\;$ Therefore, if $ABC\neq AB$, it follows from the above that \begin{equation*} \left| ABC\right| \geq \left| AB\right| +\left| H\right| \geq \left| A\right| +\left| B\right| . \end{equation*} \end{proof} \section{The linear setting\label{sec_linear setting}} From now on, and for the remainder of this paper, $K$ is a commutative field and $L$ a (possibly skew) field extension containing $K$ in its center.\ Given any subset $S\subset L,$ we write $\langle S\rangle $ for the $K$-subspace of $L$ generated by $S.$ For subsets $S_{1},S_{2}$ of $L$, we consider the product set \begin{equation*} S_{1}S_{2}=\{s_{1}s_{2}\mid s_{1}\in S_{1},s_{2}\in S_{2}\} \end{equation*} and the $K$-subspace $\langle S_{1}S_{2}\rangle $ of $L$ spanned by $S_{1}S_{2}$. Note, for later use, the equality \begin{equation*} \langle \, \langle S_{1}\rangle \langle S_{2}\rangle \, \rangle =\langle S_{1}S_{2}\rangle . \end{equation*} If $S_1 = \{x\}$, we simply write $xS_2$ instead of $\{x\}S_2$. When $S_{1}=A$, $S_{2}=B$ are finite-dimensional $K$-subspaces of $L,$ it is easy to verify that $\langle AB\rangle $ is finite-dimensional, with \begin{equation*} \dim _{K}\langle AB\rangle \leq (\dim _{K}A)(\dim _{K}B). \end{equation*} Also, for any nonzero $x\in L$, the sets $xA$ and $Ax$ are $K$-subspaces of $L$, with $\dim _{K}(xA)=\dim _{K}(Ax)=\dim _{K}A.$ \medskip \noindent \textbf{Notation.} For any subset $X \subset L$, we denote $X_{*} = X \setminus \{0\}$ and \begin{equation*} {X}_{*}^{-1} = \{x^{-1} \mid x \in X_{*} \}, \end{equation*} the set of inverses of the nonzero elements of $X$. \medskip Note that ${X}_{*}^{-1}\cup \{0\}$ is not a $K$-subspace of $L$ in general, even if $X$ is. \bigskip In subsequent sections we establish linear versions of the addition theorems recalled above. In particular, we obtain lower bounds on $\dim _{K}\langle AB \rangle$ in terms of $\dim _{K}A$, $\dim _{K}B$. As for the groups setting, our main tool will be a linear version of the Kemperman transform. \section{A linear Kemperman transform} Let $(A,B)$ be a pair of finite-dimensional $K$-subspaces of $L$. Let $x\in L\backslash \{0\}$.\ We define the transformed pairs $(A^{\prime },B^{\prime })$ and $(A^{\prime \prime},B^{\prime \prime })$ with respect to $x$ as follows: \begin{equation*} (A^{\prime} = A+Ax, \, B^{\prime} = B\cap x^{-1}B) \end{equation*} and \begin{equation*} (A^{\prime\prime} = A \cap Ax^{-1}, \, B^{\prime\prime} = B+ xB), \end{equation*} where $+$ denotes the usual sum of vector subspaces. Since $A+Ax = \langle A \cup Ax \rangle$, we may view them as linear analogues of the classical Kemperman transforms. (Compare with Section~\ref{tool}.) They satisfy the analogous properties below: \begin{enumerate} \item $A^{\prime },B^{\prime },A^{\prime \prime },B^{\prime \prime }$ are $K $-vector subspaces of $L,$ \item $\langle A^{\prime }B^{\prime }\rangle \subset \langle AB\rangle $ \ and \ $\langle A^{\prime \prime }B^{\prime \prime }\rangle \subset \langle AB\rangle ,$ \item $B^{\prime }\neq \{0\}$ $\Longleftrightarrow $ $x\in B{B}_{\ast }^{-1}$ \ and \ $A^{\prime \prime }\neq \{0\}$ $\Longleftrightarrow $ $x\in {A}_{\ast }^{-1}A,$ \item $\dim _{K}A^{\prime }>\dim _{K}A$ $\Longleftrightarrow $ $Ax\neq A$ \ and \ \hfill \newline $\dim _{K}B^{\prime \prime }>\dim _{K}B$ $\Longleftrightarrow $ $xB\neq B.$ \end{enumerate} \begin{proposition}\label{one step} \label{prop_ols} With the same notation as above, set $D={A}_{\ast }^{-1}A\cap B{B}_{\ast }^{-1}$. Suppose that either $AD\not\subset A$ or $DB\not\subset B$. Then there exists a pair $(A_{1},B_{1})$ of $K$-subspaces of $L$ satisfying: \begin{enumerate} \item[(1)] $A_{1}\neq \{0\}$ and $B_{1}\neq \{0\},$ \item[(2)] $\langle A_{1}B_{1} \rangle \subset \langle AB \rangle,$ \item[(3)] either $\dim_{K}A_{1}+\dim_{K}B_{1} = \dim_{K}A+\dim_{K}B$ and $\dim_{K}A_{1}>\dim_{K}A$, or $\dim_{K}A_{1}+\dim_{K}B_{1}>\dim_{K}A+\dim_{K}B.$ \end{enumerate} \end{proposition} \noindent Point (3) above looks somewhat technical, but it has a very interesting and useful meaning. It says that, in $\mathbb{N}^2$ ordered lexicographically, one has $$ \big(\dim _{K}A_{1}+\dim _{K}B_{1},\, \dim _{K}A_{1}\big) > \big(\dim _{K}A+\dim _{K}B,\, \dim _{K}A\big). $$ This is used, for instance, in the proof of Corollary~\ref{cor2} below. Here again, note that the set $D={A}_{*}^{-1}A\cap B{B}_{*}^{-1}$ needs not be a $K$-subspace of $L$. \bigskip \begin{proof} Since $AD\not\subset A$ or $DB\not\subset B,$ there is an element $d\neq 0$ in $D$ such that $Ad\not\subset A$ or $dB\not\subset B$.\ Denote $p,q$ the dimensions of the following quotient spaces: $$ p = \dim_K(A+Ad)/A, \; \; q = \dim_K(B+dB)/B. $$ We have $\max (p,q)\geq 1$, and we need to distinguish two cases. \smallskip \noindent $\bullet{}$ Assume first $p\geq q.$ In this case, we make the linear Kemperman transform $$ A_{1}=A+Ad, \;\; B_{1}=B\cap d^{-1}B. $$ We have $A \subsetneq A_1$ since $p \ge 1$, and $B_{1}\neq \{0\}$ since $d \in D \subset B{B}_{*}^{-1}$. Moreover $\langle A_{1}B_{1} \rangle \subset \langle AB \rangle$. This settles the first two requirements on $(A_1,B_1)$. As for the third one, we claim that $$ \dim_K A_1 = \dim_K A+p, \;\; \dim_K B_1 = \dim_K B -q. $$ Indeed, our definitions imply $p = \dim_K A_1 - \dim_K A$, and we have \begin{eqnarray*} \dim_K B_1 & = & \dim_K (B \cap d^{-1}B)\\ & = & \dim_K (dB \cap B)\\ & = & \dim_K dB + \dim_K B - \dim_K (dB + B)\\ & = & \dim_K B -q. \end{eqnarray*} It follows that $$ \dim_{K}A_{1}+\dim_{K}B_{1}=\dim_{K}A+\dim_{K}B+p-q, $$ whence $\dim_{K}A_{1}+\dim_{K}B_{1} \ge \dim_{K}A+\dim_{K}B$, and $\dim _{K}A_{1}>\dim _{K}A$ since $p \ge 1$. This yields the third requirement on $(A_1,B_1)$. \smallskip \noindent $\bullet{}$ Assume now $p<q.$ Here we use the other linear Kemperman transform and set $$ A_{1}=A\cap Ad^{-1}, \;\; B_{1}=B+dB. $$ This time we have $B \subsetneq B_1$ since $q \ge 1$, and $A_{1}\neq \{0\}$ since $d \in D \subset {A}_{*}^{-1}A$. We have $\langle A_{1}B_{1} \rangle \subset \langle AB \rangle$, and a similar calculation as above yields $$ \dim_K A_1 = \dim_K A-p, \;\; \dim_K B_1 = \dim_K B+q. $$ It follows that $$ \dim_{K}A_{1}+\dim_{K}B_{1}=\dim_{K}A+\dim_{K}B+q-p, $$ implying $\dim_{K}A_{1}+\dim_{K}B_{1} > \dim_{K}A+\dim_{K}B$, as desired. \end{proof} \begin{corollary} \label{cor2} Let $A,B$ be nonzero finite-dimensional $K$-subspaces of $L$. Then there exist nonzero finite-dimensional $K$-subspaces $E,F$ of $L$ satisfying \begin{enumerate} \item[(1)] $\langle EF\rangle \subset \langle AB\rangle ,$ \item[(2)] $\dim _{K}E+\dim _{K}F\geq \dim _{K}A+\dim _{K}B$, \item[(3)] $ED=E$ and $DF=F$, where $D={E}_{\ast }^{-1}E\cap F{F}_{\ast }^{-1}.$ \end{enumerate} \end{corollary} \begin{proof} By successive applications of the previous proposition, we get a sequence \begin{equation*} (A,B)=(A_{0},B_{0}),(A_{1},B_{1}),\dots ,(A_{i},B_{i}),\dots \end{equation*} of pairs of $K$-subspaces of $L$ satisfying the following properties, for $i \ge 1$: \begin{itemize} \item $\langle A_{i}B_{i}\rangle \subset \langle A_{i-1}B_{i-1}\rangle ,$ \item $\big(\dim _{K}A_{i}+\dim _{K}B_{i},\, \dim _{K}A_{i}\big) > \big(\dim _{K}A_{i-1}+\dim _{K}B_{i-1},\, \dim _{K}A_{i-1}\big)$ \\ in $\mathbb{N}^{2}$ ordered lexicographically. \end{itemize} Moreover, these spaces have bounded dimension, since \begin{eqnarray*} \max \{\dim _{K}A_{i},\dim _{K}B_{i}\} & \leq & \dim _{K}\langle A_{i}B_{i}\rangle \\ & \leq & \dim _{K}\langle AB\rangle \\ & \leq & (\dim _{K}A)(\dim _{K}B). \end{eqnarray*} It follows that the above sequence must be finite. By Proposition~\ref{one step}, there is an index $n\geq 0$ such that the set $D=(A_{n})_{\ast }^{-1}A_{n}\cap B_{n}(B_{n})_{\ast }^{-1}$ satisfies $A_{n}D\subset A_{n}$ and $DB_{n}\subset B_{n}$. In fact $A_{n}D = A_{n}$ and $DB_{n} = B_{n}$, since $1 \in D$. Setting $E=A_n$, $F=B_n$, it follows from the properties of the sequence of $(A_i,B_i)$ that $\langle EF\rangle \subset \langle AB\rangle$ and $\dim _{K}E+\dim _{K}F\geq \dim _{K}A+\dim _{K}B$. \end{proof} \section{Linearizing Theorem~\ref{thOl} of Olson} \label{lin ols} With the linearized Kemperman transform at hand, we are now ready to establish linear versions of the addition theorems of Section~\ref{sec-kem} and \ref{variant olson}. We start with Theorem~\ref{thOl} of Olson. The following easy lemma will be needed in the process. \begin{lemma} \label{lem_util} Let $D$ be a finite-dimensional $K$-subspace of $L$. Then $D $ is a (possibly skew) field if and only if \thinspace\ $1\in D$ and $D^{2}\subset D$. \end{lemma} \begin{proof} It suffices to show that every nonzero element $d\in D$ is invertible in $D$. Now the map $L_{d}:D\rightarrow D$ defined by $L_{d}(x)=dx$ for all $x\in D $, is linear and injective. Hence $L_{d}$ is bijective, and therefore there is some $d^{\prime }\in D$ with $L(d^{\prime })=1$. Obviously $d^{\prime }$ is the inverse of $d$ in $L$ and it does live in $D$. \end{proof} \begin{theorem} \label{TH-Ol8lin}Let $K$ be a commutative field and $L$ a field extension of $K$. Let $A,B$ be finite-dimensional $K$-vector spaces in $L$ distinct from $\{0\}$.\ Then there exist a $K$-vector subspace $S$ of $\langle AB\rangle $ and a subfield $H$ of $L$ such that \begin{enumerate} \item[(1)] $K\subset H\subset L,$ \item[(2)] $\dim _{K}S\geq \dim _{K}A+\dim _{K}B-\dim _{K}H$, \item[(3)] $HS=S$ or $SH=S$. \end{enumerate} \end{theorem} \begin{proof} By Corollary~\ref{cor2}, there are subspaces $E,F$ such that $$\langle EF\rangle \subset \langle AB\rangle, \; \; \dim_K E + \dim_K F \ge \dim_K A + \dim_K B, $$ and $ED=E$, $DF=F$ where $D={E}_{\ast }^{-1}E\cap F{F}_{\ast }^{-1}$. We start by assuming $\dim_K E\geq \dim _{K}F$ and, as in the proof of Olson's theorem, we distinguish two cases: \smallskip \noindent $(i)$ $F{F}_{\ast }^{-1}\not\subset {E}_{\ast }^{-1}E.$ Then there exist $x_{1},x_{2}\in F_\ast$ such that $x_{1}x_{2}^{-1}\notin {E}_{\ast }^{-1}E$. Therefore $Ex_{1}\cap Ex_{2}=\{0\}$ and $Ex_{1}\oplus Ex_{2}\subset \langle EF\rangle .\;$This gives $\dim _{K}\langle EF\rangle \geq 2\dim _{K}E\geq \dim _{K}E+\dim _{K}F.$ Let $H=K$ and $S=\langle EF\rangle $. Then $H$ is a subfield of $L$ stabilizing $S$, and $\dim _{K}\langle S\rangle \geq \dim _{K}A+\dim _{K}B-1$, as desired. \smallskip \noindent $(ii)$ $F{F}_{\ast }^{-1}\subset {E}_{\ast }^{-1}E.$ Then $D=F{F}_{\ast }^{-1}$ and $1\in D$. Moreover, we have \begin{equation*} DD=DF{F}_{\ast }^{-1}=F{F}_{\ast }^{-1}=D \end{equation*} since $DF=F.\;$So $D^{2}\subset D$.\ Let $z$ be a nonzero element of $F$.\ We have \begin{equation*} F=Fz^{-1}z\subset Dz\subset DF=F \end{equation*} because $Fz^{-1}\subset D$ and $DF=F$.\ This implies that $Dz=F$ and thus $D=Fz^{-1}$ is a finite-dimensional $K$-vector space in $L$.\ Since $1\in D$ and $D^{2}=D$, we derive from Lemma \ref{lem_util} that $D$ is a subfield of $L$ containing $K$.\ Moreover $\dim _{K}D=\dim _{K}F$.\ Set $S=\langle EF\rangle $ and $H=z^{-1}Dz$.\ Since $D$ is a field, $H$ is also a field and $\dim _{K}H=\dim _{K}D=\dim _{K}F.\;$We have \begin{equation*} SH\subset \langle EFH\rangle =\langle EFz^{-1}Dz\rangle =\langle EDzz^{-1}Dz\rangle =\langle EDz\rangle =\langle EF\rangle =S \end{equation*} since $Dz=F$ and $D^{2}=D.$ It follows that $SH=S$. We have $\dim _{K}S=\dim _{K}\langle EF\rangle \geq \dim _{K}E$ since $F\neq \{0\}$.\ Finally, this gives \begin{equation*} \dim _{K}S\geq \dim _{K}E+\dim _{K}F-\dim _{K}H \end{equation*} because $\dim _{K}H=\dim _{K}D=\dim _{K}F.$ This settles the case $\dim_K E\geq \dim _{K}F$. The case $\dim_K F\geq \dim _{K}E$ can be treated in a similar way. \end{proof} \bigskip \noindent \textbf{Remarks.} \begin{enumerate} \item[$(i)$] In contrast to the linear version of Kneser's Theorem~\ref{linear Kneser}, our linear version of Olson's Theorem does not require any separability hypothesis. \item[$(ii)$] Assume that $L$ is commutative and $\dim_K L =p$, a prime number. In that case, there is no intermediate field $K \subset H \subset L$ besides $K, L$. It follows from the above theorem that either $\langle AB \rangle = L$ or $\dim_K \langle AB \rangle \ge \dim_K S \ge \dim_K A + \dim_K B -1$. This is also a consequence of Theorem~\ref{linear Kneser}, provided $L$ is further assumed to be separable over $K$. \item[$(iii)$] Suppose that $K$ is a finite field of cardinality $q$.\ Then $A^{\ast }=A\backslash \{0\}$ and $B^{\ast }=B\backslash \{0\}$ are finite subsets of the group $L^{\ast }=L\backslash \{0\}$, and by Theorem~\ref{thOl}, there exist a subset $\mathcal{S}$$^{\ast }$ of $A^{\ast }B^{\ast}$ and a subgroup $\mathcal{H}^{\ast }$ of $L^{\ast }$ such that \begin{equation*} \left| \text{$\mathcal{S}$}^{\ast }\right| \geq \left| A^{\ast }\right| +\left| B^{\ast }\right| -\left| \text{$\mathcal{H}$}^{\ast }\right| . \end{equation*} Thus, we get \begin{equation*} \left| A\right| +\left| B\right| \leq \left| \text{$\mathcal{S}$}^{\ast }\right| +\left| \text{$\mathcal{H}$}^{\ast }\right| +2. \end{equation*} However, with Theorem \ref{TH-Ol8lin}, we obtain \begin{equation*} q^{\dim _{K}A}q^{\dim _{K}B}\leq q^{\dim _{K} S}q^{\dim _{K}H}, \end{equation*} and since $\left| K\right| =q$, this gives \begin{equation*} \left| A\right| \left| B\right| \leq \left| S\right| \left| H\right|. \end{equation*} Thus, Theorem \ref{thOl} gives an upper bound for $\left| A\right| +\left| B\right|$, whereas Theorem \ref{TH-Ol8lin} gives an upper bound for $\left| A\right| \left| B\right| $. Note that $\mathcal{S}^{\ast }\neq S\backslash \{0\}$ and $\mathcal{H}^{\ast }\neq H\backslash \{0\}$ in general. \end{enumerate} \bigskip One first easy consequence is a linear version of Kemperman's Theorem~\ref {torsion-free} on torsion-free groups. \begin{theorem} \label{linear torsion-free} Let $K$ be a commutative field and $L$ a (possibly skew) purely transcendental extension of $K$. Let $A,B$ be nonzero finite-dimensional $K$-subspaces of $L$. Then \begin{equation*} \dim _{K}\langle AB\rangle \geq \dim _{K}A+\dim _{K}B-1. \end{equation*} \end{theorem} \begin{proof} By Theorem~\ref{TH-Ol8lin}, there is a subspace $S\subset \langle AB\rangle $ and an intermediary field $K\subset H\subset L$ such that \begin{equation*} \dim _{K}\langle AB\rangle \geq \dim _{K}A+\dim _{K}B-\dim _{K}H, \end{equation*} and $HS=S$ or $SH=S$. As $\dim _{K}S$ is finite, it follows that $\dim _{K}H$ must be finite as well. But $K$ is the only finite-dimensional subfield of $L$. It follows that $H=K$ and hence $\dim _{K}H=1$. \end{proof} \bigskip \noindent \textbf{Remark.} The above lower bound is sharp. Indeed, let $x\in L\setminus K$. Then $x$ is transcendental over $K$. Fix positive integers $r,s$. Let $A,B$ be the subspaces of $L$ generated by $\{1,x,\ldots ,x^{r-1}\} $, $\{1,x,\ldots ,x^{s-1}\}$, respectively. Then $\dim _{K}A=r$, $\dim _{K}B=s$ and $\dim _{K}\langle AB\rangle =r+s-1$, since $\langle AB\rangle $ is the subspace spanned by the basis $\{1,x,\ldots ,x^{r+s-2}\}$. \bigskip We now derive, from Theorem~\ref{TH-Ol8lin} again, a linear version of the basic Proposition~\ref{basic}. \begin{proposition} \label{linear basic noncomm} Let $K$ be a commutative field and $L$ a field extension containing $K$ in its center, with $\dim _{K}L$ finite. Let $A,B$ be nonzero subspaces of $L$ satisfying $\dim _{K}A+\dim _{K}B>\dim _{K}L$. Then $\langle AB\rangle =L$. \end{proposition} \smallskip Note that the hypothesis that $L$ is a field is essential. For otherwise, a counterexample would be provided by $A=L$ and $B=$ a proper nonzero left ideal of $L$, yielding $\langle AB \rangle = B$. \medskip \begin{proof} By Theorem~\ref{TH-Ol8lin}, there is a subspace $S\subset \langle AB\rangle $ and an intermediate field $K\subset H\subset L$ such that \begin{equation} \dim _{K}S\geq \dim _{K}A+\dim _{K}B-\dim _{K}H, \label{olson} \end{equation} and either $HS=S$ or $SH=S$. We claim that $HS=L=SH$. Indeed, fix any nonzero element $x\in L$. It follows from (\ref{olson}) and the hypothesis $\dim _{K}A+\dim _{K}B>\dim _{K}L$, that \begin{equation*} \dim _{K}S+\dim _{K}(xH) >\dim _{K}L. \end{equation*} Therefore $S\cap xH\neq \{0\}$. Hence, there are nonzero elements $s\in S$ and $h\in H$ such that $s=xh$. But then $x=sh^{-1}$ belongs to $SH$, since $h^{-1}\in H$. It follows that $L=HS$. The same argument, with $xH$ replaced by $Hx$, yields $L=SH$. We conclude $L=S=\langle AB\rangle $, since $S\subset \langle AB\rangle $ and $S=HS$ or $SH$. \end{proof} When $L$ is commutative, the above result admits a much simpler proof, which does not require Theorem~\ref{TH-Ol8lin}. \begin{proposition} \label{linear basic comm} Let $K\subset L$ be a commutative field extension, with $\dim _{K}L$ finite. Let $A,B$ be nonzero subspaces of $L$ satisfying $\dim _{K}A+\dim _{K}B>\dim _{K}L$. Then $\langle AB\rangle =L$. \end{proposition} \begin{proof} We may assume $A,B\neq L$ and proceed by induction on $\dim_K B$. If $\dim_K B=1$, then $\dim_K A=\dim_K L$ and $\langle AB\rangle =\langle LB\rangle =L$. Assume now $\dim_K B\geq 2$. Since $\langle LB\rangle =L$, there must be a nonzero element $x\in L$ such that $xB\not\subset A$. Set \begin{equation*} A^{\prime }=A+xB,\; B^{\prime }=Ax^{-1}\cap B. \end{equation*} We have $A^{\prime }B^{\prime }\subset \langle AB\rangle $ by construction and the commutativity of $L$. Moreover, the subspace $B^{\prime }$ is nonzero, since $\dim _{K}Ax^{-1}+\dim _{K}B$ = $\dim _{K}A+\dim _{K}B $ $>\dim _{K}L$. Finally, $\dim _{K}A^{\prime }+\dim _{K}B^{\prime} =\dim _{K}A+\dim _{K}B$, since \begin{eqnarray*} \dim _{K}A^{\prime }=\dim _{K}(A+xB) &=&\dim _{K}A+\dim _{K}xB-\dim _{K}(A\cap xB) \\ &=&\dim _{K}A+\dim _{K}B-\dim _{K}(Ax^{-1}\cap B) \\ &=&\dim _{K}A+\dim _{K}B-\dim _{K}B^{\prime }. \end{eqnarray*} By the induction hypothesis, we conclude $\langle A^{\prime }B^{\prime }\rangle =L=\langle AB\rangle $. \end{proof} \bigskip \noindent \textbf{Remark.} We are grateful to Joseph Oesterl\'{e} for providing us with the following alternative proof of Proposition~\ref{linear basic noncomm}, which only uses duality in vector spaces. \medskip \begin{proof} Let $H$ be any hyperplane in $L$, and let $\varphi :L\rightarrow K$ be a linear form with kernel $H$. It suffices to show that there exist $a\in A$, $b\in B$ such that $\varphi (ab)\not=0$, implying $\langle AB\rangle \not\subset H$. The map $\beta :L\times L\rightarrow K$ defined by $\beta (x,y)=\varphi (xy)$ for all $x,y\in L$ is a $K$-bilinear form, and it is non-degenerate: if $x\not=0$, then $xL$ is equal to $L$ and hence $\varphi (xL)\not=\{0\}$. Therefore $\beta $ induces an isomorphism $\gamma :L\rightarrow L^{\ast }$, where $L^{\ast }$ is the dual of $L$, defined by the formula $\gamma (x)(y)=\beta (x,y)$ for all $x,y\in L$. From the inclusion map $j:B\rightarrow L$ we deduce, by transposition, a surjection $j^{t}:L^{\ast }\rightarrow B^{\ast }$, defined as usual by $j^{t}(\psi )=\psi \circ j$ for all $\psi \in L^{\ast }$. The composition $j^{t}\circ \gamma :L\rightarrow B^{\ast }$ is then also surjective, and therefore \begin{equation*} \dim_K \ker (j^{t}\circ \gamma )=\dim_K L-\dim_K B<\dim_K A. \end{equation*} Thus, there is some $a\in A$ satisfying $j^{t}(\gamma (a))\not=0$. Since the linear form $\gamma (a)\circ j:B\rightarrow K$ does not vanish, there must be some $b\in B$ satisfying $\gamma (a)(b)\not=0$, i.e. $\beta (a,b)=\varphi (ab)\not=0$. \end{proof} \section{Linearizing Theorem~\ref{THK} of Kemperman} We shall now establish a linear analogue of Kemperman's Theorem~\ref{THK}, according to which, for subsets $A,B$ in a group $G$, one has $|AB| \ge |A|+|B|-1$ provided there is an element $c \in AB$ with a \textit{unique representation} of the form $c=ab$ with $a\in A, b\in B$. In order to properly linearize this result, we need to rephrase the above unicity condition on $c$. First, up to translation of $A,B$, we may assume that $1 \in A \cap B$ and that $c=1$ admits the unique representation $1 = ab $ with $a=b=1$ as a product in $AB$. If we write \begin{eqnarray*} A & = & \{1\} \sqcup \overline{A} \\ B & = & \{1\} \sqcup \overline{B} \end{eqnarray*} with $\overline{A}, \overline{B}$ the respective complements of $\{1\}$ in $A,B$, then the unicity of 1 as a product in $AB$ is equivalent to the disjointness condition \begin{equation*} \{1\} \cap (\overline{A} \cup \overline{B} \cup \overline{A}\; \overline{B}) = \emptyset, \end{equation*} which may also be written as $AB = \{1\} \sqcup (\overline{A} \cup \overline{B} \cup \overline{A}\; \overline{B}).$ These equivalent conditions motivate the following formulation of our sought-for linearization. \begin{theorem} \label{th_AB}Let $K$ be a commutative field and $L$ a field extension of $K$. Let $A,B$ be finite-dimensional $K$-vector spaces in $L$ such that $K\subset A\cap B$. Suppose there exist subspaces $\overline{A},\overline{B} \subset L$ such that \begin{equation} A=K\oplus \overline{A},\text{ }B=K\oplus \overline{B}\text{ and }K\cap (\overline{A}+\overline{B}+\langle \overline{A}\,\overline{B}\rangle )=\{0\}. \label{Cond} \end{equation} Then \begin{equation*} \dim _{K}\langle AB\rangle \geq \dim _{K}A+\dim _{K}B-1. \end{equation*} \end{theorem} \begin{proof} Observe first that, if $A=K+\overline{A}$ and $B=K+\overline{B}$, then $\langle AB\rangle =K+(\overline{A}+\overline{B}+\langle \overline{A}\, \overline{B}\rangle )$. Therefore, condition~(\ref{Cond}) is equivalent to \begin{equation*} A=K\oplus \overline{A},\text{ }B=K\oplus \overline{B}\text{ and }\langle AB\rangle =K\oplus (\overline{A}+\overline{B}+\langle \overline{A}\,\overline{B}\rangle ). \end{equation*} Since $1\in A\cap B$, we have $A+B\subset \langle AB\rangle $. We now distinguish two cases. \smallskip \noindent \textbf{Case 1.} Assume $A\cap B=K$. Then $\dim _{K}(A+B)=\dim _{K}A+\dim _{K}B-1$, and we are done since $\dim _{K}\langle AB\rangle \geq \dim _{K}(A+B)$. \smallskip \noindent \textbf{Case 2.} Assume now $A\cap B\not=K$, i.e. $\dim _{K}A\cap B\geq 2$. We first claim that \begin{equation} A\cap B=K\oplus (\overline{A}\cap \overline{B}). \end{equation} Indeed, let $x\in A\cap B$. Then there are $\lambda ,\mu \in K$ and $\overline{x}\in \overline{A},\overline{y}\in \overline{B}$ such that \begin{equation*} x=\lambda +\overline{x}=\mu +\overline{y}. \end{equation*} Since $K\cap (\overline{A}+\overline{B})=\{0\}$, it follows that $\lambda =\mu $ and $\overline{x}=\overline{y}$, so that $x\in K\oplus (\overline{A} \cap \overline{B})$, as claimed. The reverse inclusion is immediate. Observe that in the present case, we have $\overline{A}\cap \overline{B}\not=\{0\}$. We shall perform a suitable sequence of linear Kemperman transforms on the pair $(A,B)$ and eventually reach Case 1 again, thereby concluding the proof. Let $0\not=d\in \overline{A}\cap \overline{B}$ be any nonzero element. We perform a Kemperman transform on $(A,B)$ relative to $d$, and get a new pair $(A_{1},B_{1})$, by defining either \begin{eqnarray} (i) &A_{1}=A+Ad,&B_{1}=B\cap d^{-1}B,\text{ or } \\ (ii) &A_{1}=A\cap Ad^{-1},&B_{1}=B+dB. \end{eqnarray} In either case, we have $\langle A_{1}B_{1}\rangle \subset \langle AB\rangle $, and one of $(i)$ or $(ii)$ will yield the estimate \begin{equation*} \dim _{K}A_{1}+\dim _{K}B_{1}\geq \dim _{K}A+\dim _{K}B. \end{equation*} \noindent Case $(i)$. For any nonzero element $d\in \overline{A}\cap \overline{B}$, define \begin{eqnarray*} A_{1} &=&A+Ad, \\ B_{1} &=&B\cap d^{-1}B. \end{eqnarray*} We shall show that this new pair satisfies the hypotheses of the theorem. Indeed, define \begin{eqnarray*} \overline{A_{1}} &=&\overline{A}+Ad, \\ \overline{B_{1}} &=&\overline{B}\cap d^{-1}\overline{B}. \end{eqnarray*} We claim that conditions (\ref{Cond}) are satisfied, i.e. \begin{eqnarray} A_{1} &=&K\oplus \overline{A_{1}}, \label{cond1} \\ B_{1} &=&K\oplus \overline{B_{1}}, \label{cond2} \\ \{0\} &=&K\cap (\overline{A_{1}}+\overline{B_{1}}+\langle \overline{A_{1}}\, \overline{B_{1}}\rangle ). \label{cond3} \end{eqnarray} \noindent \textbf{For (\ref{cond1}):} We have $K+\overline{A_{1}}=K+\overline{A}+Ad=A+Ad=A_{1}$. The sum is direct, since $\overline{A_{1}} \subset \overline{A}+Kd+\overline{A}d\subset \overline{A}+\overline{B} +\langle \overline{A}\,\overline{B}\rangle $, whence $K\cap \overline{A_{1}}=\{0\}$ by (\ref{Cond}). \smallskip \noindent \textbf{For (\ref{cond2}):} To start with, we have $K\cap \overline{B_{1}}\subset K\cap \overline{B}=\{0\}$ by (\ref{Cond}). We next verify the inclusion $K+\overline{B_{1}}\subset B_{1}$. We have $K\subset B$, and $K\subset d^{-1}B$ since $d\in B$, whence $K\subset B_{1}$. Moreover, we have $\overline{B_{1}}=\overline{B}\cap d^{-1}\overline{B} \subset B\cap d^{-1}B=B_{1}$. This establishes the desired inclusion. It remains to prove the reverse inclusion $B_{1}\subset K+\overline{B_{1}}$. Let $x\in B_{1}=B\cap d^{-1}B$. Since $B=K\oplus \overline{B}$, we may write $x=\lambda +\overline{x}$ for some $\lambda \in K$ and $\overline{x}\in \overline{B}$. It remains to show that $\overline{x}\in \overline{B_{1}}= \overline{B}\cap d^{-1}\overline{B}$, i.e. that $\overline{x}\in d^{-1} \overline{B}$. Since $x\in d^{-1}B$, there are elements $y\in B$, $\mu \in K$ and $\overline{y}\in \overline{B}$ such that $x=d^{-1}y$ and $y=\mu + \overline{y}$. Hence $dx=y$, and therefore \begin{equation*} d\lambda +d\overline{x}=\mu +\overline{y}. \end{equation*} It follows that $\mu =d\lambda +d\overline{x}-\overline{y}$, whence $\mu \in K\cap (\overline{B}+\langle \overline{A}\,\overline{B}\rangle )$. Therefore $\mu =0$ by (\ref{Cond}), whence $d\overline{x}=\overline{y}-d\lambda \in \overline{B}$. This shows that $\overline{x}\in d^{-1}\overline{B}$, implying $\overline{x}\in \overline{B_{1}}$ and finally $x\in K+\overline{B_{1}}$, as desired. \smallskip \noindent \textbf{For (\ref{cond3}):} By (\ref{Cond}), it suffices to show the inclusion \begin{equation*} (\overline{A_{1}}+\overline{B_{1}}+\langle \overline{A_{1}}\,\overline{B_{1}} \rangle )\subset (\overline{A}+\overline{B}+\langle \overline{A}\,\overline{B}\rangle ). \end{equation*} \smallskip Considering each summand at a time in the left-hand side, we have: \newline \smallskip $\bullet $ $\overline{A_{1}}=\overline{A}+Ad=\overline{A}+Kd+ \overline{A}d\subset (\overline{A}+\overline{B}+\langle \overline{A}\,\overline{B}\rangle )$,\newline \smallskip $\bullet $ $\overline{B_{1}}=\overline{B}\cap d^{-1}\overline{B}\subset \overline{B},$\newline \smallskip $\bullet $ $\overline{A_{1}}\,\overline{B_{1}}=(\overline{A}+Ad) (\overline{B}\cap d^{-1}\overline{B})\subset \overline{A}\,\overline{B}+A\, \overline{B}\subset \overline{A}\,\overline{B}+\overline{B}+\overline{A}\,\overline{B}$, \newline and we are done. \medskip \noindent Case $(ii)$. For any nonzero element $d\in \overline{A}\cap \overline{B}$, define \begin{eqnarray*} A_{1} &=&A\cap Ad^{-1}, \\ B_{1} &=&B+dB. \end{eqnarray*} This time, we set \begin{eqnarray*} \overline{A_{1}} &=&\overline{A}\cap \overline{A}d^{-1}, \\ \overline{B_{1}} &=&\overline{B}+dB. \end{eqnarray*} With arguments similar to those of Case $(i)$, we can prove that conditions (\ref{Cond}) are satisfied again, i.e. \begin{eqnarray*} A_{1} &=&K\oplus \overline{A_{1}}, \\ B_{1} &=&K\oplus \overline{B_{1}}, \\ \{0\} &=&K\cap (\overline{A_{1}}+\overline{B_{1}}+\langle \overline{A_{1}}\,\overline{B_{1}}\rangle ). \end{eqnarray*} Now, as in the proof of Corollary \ref{cor2}, we iterate the above Kemperman transforms as long as possible. At each step, we get new subspaces $A_{i}$, $B_{i}$ satisfying (\ref {Cond}) and such that $$ \big(\dim _{K}A_{i}+\dim _{K}B_{i},\, \dim _{K}A_{i}\big) > \big(\dim _{K}A_{i-1}+\dim _{K}B_{i-1},\, \dim _{K}A_{i-1}\big) $$ in $\mathbb{N}^{2}$ ordered lexicographically. Since these subspaces have bounded dimension, the iteration cannot continue indefinitely and Case 1 must eventually be reached. This means that there exist subspaces $E,F$ of $L$ such that \begin{equation} \langle EF\rangle \subset \langle AB\rangle , \label{H1} \end{equation} \begin{equation} \dim _{K}E+\dim _{K}F\geq \dim _{K}A+\dim _{K}B, \label{H2} \end{equation} and satisfying $E\cap F=K$ together with the hypotheses of the theorem.\ By Case 1, we have \begin{equation*} \dim _{K}\langle EF\rangle \geq \dim _{K}E+\dim _{K}F-1. \end{equation*} The desired inequality, namely \begin{equation*} \dim _{K}\langle AB\rangle \geq \dim _{K}A+\dim _{K}B-1, \end{equation*} now follows from (\ref{H1}) and (\ref{H2}). \end{proof} \section{Linearizing Theorem~\ref{ABC}} \label{sec: lin ABC} Throughout the last two sections, we shall assume that $L$ is a commutative field extension of $K$ and that every algebraic element of $L$ is separable over $K$. This allows us to use Theorem~\ref{linear Kneser}, the linear version of Kneser's Theorem. Our results below probably remain true in the more general setting of the preceding sections, where $L$ is only assumed to be a field containing $K$ in its center. But we have no proof of this so far. \begin{lemma} \label{Lem_util2}Let $K$ be a commutative field, $H$ a field extension of $K$ and $L$ a field extension of $H$.\ Let $V$ be a finite-dimensional $K$-vector space in $L$ such that $HV=V$.\ Then there exists a finite subset $R_{V}\subset V$ such that \begin{equation} V=\bigoplus_{v\in R_{V}}Hv. \label{dec} \end{equation} In particular $\dim _{K}H$ is finite and divides $\dim _{K}V$. \end{lemma} \begin{proof} For any $v\in V$, $Hv$ is vector subspace of $V$.\ Moreover for any $v,v^{\prime }$ in $V$, one has $Hv=Hv^{\prime }$ or $Hv\cap Hv^{\prime }=\{0\}.\;$The lemma follows immediately since $\dim _{K}V$ is finite. \end{proof} \bigskip \noindent \textbf{Remark.} If $HV=V,$ then $V$ can be interpreted as a finite-dimensional left $H$-module, and (\ref{dec}) gives its decomposition into irreducible components. \bigskip We now give a linear version of Theorem~\ref{ABC}. As mentioned above, the hypotheses on $L$ are probably more restrictive than actually necessary. \begin{theorem} \label{linear ABC} \label{Th_L}Let $K\subset L$ be commutative fields such that every algebraic element in $L$ is separable over $K.$ Let $A,B,C\subset L$ be finite-dimensional $K$-subspaces of $L$ such that $A,B\neq \{0\}$ and $K\subset C.$ Then either \begin{equation*} \langle ABC\rangle =\langle AB\rangle \;\text{ or }\;\dim _{K}\langle ABC\rangle \geq \dim _{K}A+\dim _{K}B. \end{equation*} \end{theorem} \begin{proof} We shall apply Theorem~\ref{linear Kneser}. The stabilizer $H$ of $\langle AB\rangle $ is a field extension of $K$, and we have \begin{equation} \dim _{K}\langle AB\rangle \geq \dim _{K}A+\dim _{K}B-\dim _{K}H. \label{ineq_kneserV} \end{equation} Then $H$ stabilizes $\langle ABC\rangle $, since $H\langle AB\rangle c=\langle AB\rangle c$ for all $c\in C$. Of course $\langle AB\rangle $ is a subspace of $\langle ABC\rangle $, since $1\in C$. Assume $\langle ABC\rangle \neq \langle AB\rangle $. Since $H\langle ABC\rangle =\langle ABC\rangle $, Lemma~\ref{Lem_util2} implies the existence of a nonempty finite subset $R\subset \langle ABC\rangle $ such that $R\cap \langle AB\rangle =\emptyset $ and \begin{equation*} \langle ABC\rangle =\langle AB\rangle \oplus \bigoplus_{v\in R}Hv. \end{equation*} In particular, we have \begin{equation*} \dim _{K}\langle ABC\rangle =\dim _{K}\langle AB\rangle +\left| R\right| \dim _{K}H. \end{equation*} Since $\left| R\right| >0,$ this gives \begin{equation*} \dim _{K}\langle ABC\rangle \geq \dim _{K}\langle AB\rangle +\dim _{K}H. \end{equation*} Finally, by (\ref{ineq_kneserV}), we obtain the desired inequality \begin{equation*} \dim _{K}\langle ABC\rangle \geq \dim _{K}A+\dim _{K}B. \end{equation*} \end{proof} \begin{corollary} \label{cor3}Let $K\subset L$ be commutative fields such that every algebraic element in $L$ is separable over $K.$ Let $A,B$ be nonzero finite-dimensional $K$-subspaces of $L$. Then either \begin{equation*} \langle AB{B}_{\ast }^{-1}B\rangle =\langle AB\rangle \;\text{ or }\;\dim _{K}\langle AB^{2}\rangle \geq \dim _{K}A+\dim _{K}B. \end{equation*} \end{corollary} \begin{proof} Assume $\langle AB{B}_{\ast }^{-1}B\rangle \neq \langle AB\rangle $. Then there exists $b_{0}\in B$ such that $\langle ABb_{0}^{-1}B\rangle \neq \langle AB\rangle $.\ Thus $\langle ABb_{0}^{-1}Bb_{0}^{-1}\rangle \neq \langle ABb_{0}^{-1}\rangle .$ We have $1\in Bb_{0}^{-1}.$ Thus, applying the above theorem to $A$ and $Bb_{0}^{-1}$, we get \begin{equation*} \dim _{K}\langle AB^{2}\rangle =\dim _{K}\langle ABb_{0}^{-1}Bb_{0}^{-1}\rangle \geq \dim _{K}A+\dim _{K}B. \end{equation*} \end{proof} \section{Powers of subspaces} \label{sec: powers} As in the preceding section, we assume that $L$ is a commutative field extension of $K$ in which every algebraic element is separable over $K$. If $B$ is a nonzero finite-dimensional $K$-subspace of $L$, we shall consider the sequence of powers $B$, $\langle B^2 \rangle, \langle B^3 \rangle, \ldots $ and analyze the evolution of the nondecreasing sequence $\dim_K\langle B^i \rangle, i \geq 1$. Without loss of generality, replacing $B$ by $b^{-1}B$ for some $b \in B \setminus \{0\}$, we may and will assume that $B$ contains 1. Under this hypothesis, the sequence $\langle B^i \rangle$ turns into an ascending chain \begin{equation*} B \subset \langle B^2 \rangle \subset \langle B^3 \rangle \subset \ldots. \end{equation*} This chain may eventually stabilize at $\langle B^n \rangle$ for some $n \ge 1$, for instance if $L$ is finite-dimensional over $K$. We start by analyzing the least such exponent $n$, if any. \begin{proposition} \label{chain} Let $K$ be a commutative field and $L$ a field extension of $K$ containing $K$ in its center. Let $B$ be a finite-dimensional $K$-subspace of $L$ containing 1. Let $n\geq 1$. The following are equivalent: \begin{enumerate} \item[(1)] $\langle B^{n+1}\rangle =\langle B^{n}\rangle $; \item[(2)] $\langle B^{2n}\rangle =\langle B^{n}\rangle $; \item[(3)] $\langle B^{n}\rangle $ is a field. \end{enumerate} \end{proposition} \begin{proof} First observe that, if $U,V$ are any subsets of $L$, then \begin{equation*} \langle UV\rangle =\langle \langle U\rangle \langle V\rangle \rangle . \end{equation*} Indeed, both $K$-subspaces are generated by the subset $UV$. In particular, we may and will freely use formulas such as $\langle B^{m}\rangle =\langle \langle B^{i}\rangle \langle B^{m-i}\rangle \rangle $ for integers $0\leq i\leq m$. Assume first $\langle B^{n+1}\rangle =\langle B^{n}\rangle $. Then we claim that $\langle B^{n+i}\rangle =\langle B^{n}\rangle $ for all $i\geq 1$. Indeed, proceeding by induction on $i$, we have \begin{equation*} \langle B^{n+i}\rangle =\langle \langle B^{n+i-1}\rangle B\rangle =\langle B^{n+1}\rangle =\langle B^{n}\rangle . \end{equation*} For $i=n$, this gives $\langle B^{2n}\rangle =\langle B^{n}\rangle $. In turn, this equality is equivalent to $\langle \langle B^{n}\rangle \langle B^{n}\rangle \rangle =\langle B^{n}\rangle $. By Lemma \ref{lem_util}, it follows that $\langle B^{n}\rangle $ is a field. Conversely, if $\langle B^{n}\rangle $ is a field, then $\langle B^{2n}\rangle =\langle \langle B^{n}\rangle \langle B^{n}\rangle \rangle =\langle B^{n}\rangle $. As $\langle B^{n}\rangle \subset \langle B^{n+1}\rangle \subset \langle B^{2n}\rangle $, this implies that $\langle B^{n+1}\rangle =\langle B^{n}\rangle $. \end{proof} \medskip In particular, the smallest integer $n \ge 1$, if any, such that $\langle B^{n}\rangle =\langle B^{n+1}\rangle$ coincides with the smallest integer $n \ge 1$, if any, such that $\langle B^{n}\rangle$ is a field. \begin{theorem} \label{TH_last}Let $K$ be a commutative field and $L$ a commutative field extension of $K$.$\;$Suppose that every algebraic element in $L$ is separable over $K.$ Let $B$ be a finite-dimensional $K$-vector space in $L$ containing 1, and let $n\geq 1$. Then either \begin{equation*} \langle B^{n+1}\rangle =\langle B^{n}\rangle \end{equation*} or \begin{equation*} \dim _{K}\langle B^{n+1}\rangle \geq \dim _{K}\langle B^{n-1}\rangle +\dim _{K}B. \end{equation*} \end{theorem} \begin{proof} Observe first that $\langle B^{n+1}\rangle =\langle B^{n}\rangle $ if and only if $\langle B^{n+1}\rangle =\langle B^{n}b\rangle $ for all $b\in B\setminus \{0\}.$ Indeed, we have $\langle B^{n}\rangle \subset \langle B^{n}b\rangle \subset \langle B^{n+1}\rangle $, and $\langle B^{n+1}\rangle $ is the sum of all the subspaces $\langle B^{n}b\rangle $ where $b$ runs over $B\setminus \{0\}$. It follows that $\langle B^{n+1}\rangle =\langle B^{n}\rangle $ is equivalent to $\langle B^{n}b_{1}\rangle =\langle B^{n}b_{2}\rangle $ for all $b_{1},b_{2}\in B\setminus \{0\}$, which in turn is equivalent to $\langle B^{n}(B{B}_{\ast }^{-1})\rangle =\langle B^{n}\rangle $. Assume now $\langle B^{n}(B{B}_{\ast }^{-1})\rangle \neq \langle B^{n}\rangle $. Since $L$ is a commutative field, this is equivalent to $\langle B^{n}({B}_{\ast }^{-1}B)\rangle \neq \langle B^{n}\rangle $. Applying Corollary \ref{cor3} to $B^{n-1}$ and $B$, this gives \begin{equation*} \dim _{K}\langle B^{n+1}\rangle \geq \dim _{K}\langle B^{n-1}\rangle +\dim _{K}B, \end{equation*} as required. \end{proof} \bigskip \begin{corollary} Let $K$ be a commutative field and $L$ a finite separable commutative extension of $K$. Let $B$ be a $K$-vector space in $L$ containing 1. Then the smallest integer $n\geq 1$ such that $\langle B^{n}\rangle $ is a field satisfies \begin{equation*} n\leq 2\,\dim _{K}L/\dim _{K}B. \end{equation*} \end{corollary} \begin{proof} By Proposition~\ref{chain}, we have $B\subsetneqq \langle B^{2}\rangle \subsetneqq \cdot \cdot \cdot \subsetneqq \langle B^{n}\rangle $. It follows from Theorem~\ref{TH_last} that $$ \dim _{K}\langle B^{n}\rangle \geq \left\{ \begin{array}{rcl} \frac{(n+1)}{2} \dim _{K}B & \; & \text{ if }n\text{ is odd}, \\ & & \\ \frac{n}{2} \dim _{K}B & \;& \text{ if }n \text{ is even}. \end{array} \right. $$ Since $\dim _{K}\langle B^{n}\rangle \leq \dim _{K}L$, this imposes the inequality $$ n \le \left\{ \begin{array}{lll} 2\, {\dim _{K}L}/{\dim _{K}B}-1 &\; \; & \text{ if }n\text{ is odd,}\\ & & \\ 2\, {\dim _{K}L}/{\dim _{K}B}&\; \; & \text{ if }n\text{ is even.}\\ \end{array} \right. $$ \end{proof} \bigskip \noindent \textbf{Remark.} From the previous Corollary, we deduce \begin{equation*} n\leq \left\lfloor \frac{2\dim _{K}L}{\dim _{K}B}\right\rfloor. \end{equation*} This upper bound is sharp.\ This can easily be seen, for example by choosing for $B$ a supplementary space of $K$ in $L.$ \bigskip \smallskip \noindent \textbf{Acknowledgment.} The authors thank Vincent Fleckinger, Joseph Oesterl\'{e} and Surya D. Ramana for stimulating discussions concerning Section~\ref{lin ols} of this paper.

https://arxiv.org/abs/1708.08085

Valuations, arithmetic progressions, and prime numbers

In this short note, we give two proofs of the infinitude of primes via valuation theory and give a new proof of the divergence of the sum of prime reciprocals by Roth's theorem and Euler-Legendre's theorem for arithmetic progressions.

\section{The infinitude of primes via valuation theory} We cite Neukirch's book \cite{N} for the facts in valuation theory. \begin{Ostrowski}[{\cite[p.~119, (3.7)]{N}}] Every non-trivial valuation on the field of rational numbers is equivalent to either the usual absolute value or the $p$-adic valuation for some prime number $p$. \end{Ostrowski} By this beautiful theorem, we see that the infinitude of primes is equivalent to the infinitude of equivalence classes of non-trivial valuations on the field of rational numbers. In this section, we give two valuation theoretic proofs of Euclid's theorem by the approximation theorem. Note that we don't use Ostrowski's theorem, but we have to consider the infinite place. \begin{app}[{\cite[p.~117, (3.4)]{N}}] Let $| \ |_1, \dots, | \ |_n$ be pairwise inequivalent non-trivial valuations of the field of rational numbers and let $a_1, \dots, a_n$ be given rational numbers. Then, for every $\varepsilon > 0$, there exists a rational number $q$ such that \[ |q-a_i|_i < \varepsilon \] for all $i=1, \dots, n$. \end{app} Let $| \ |_p$ (resp.\ $| \ |_{\infty}$) be the $p$-adic valuation normalized as $|p|_p=p^{-1}$ (resp.\ the usual absolute value) and $\mathbb{Q}_p$ (resp.\ $\mathbb{Q}_{\infty}$) the field of $p$-adic numbers for a prime number $p$ (resp.\ the field of real numbers). In the following two proofs only, we denote $p$ as a prime number or the symbol $\infty$. \begin{proof}[Proof of Euclid's theorem by the product formula] We assume that there are only finitely many primes. For each $p$, take a rational number $a_p$ such that $|a_p|_p > 1$. By the approximation theorem, we can take a rational number $q$ such that $|q|_p > 1$ for every $p$. Then, $\prod_p|q|_p > 1$ holds. On the other hand, by the product formula (\cite[p.~108, (2.1)]{N}), $\prod_p|q|_p$ must be equal to $1$. This is a contradiction. \end{proof} \begin{proof}[Proof of Euclid's theorem by the topology of the adele ring] We assume that there are only fin- \noindent itely many primes. Then, the adele ring $\mathbb{A}_{\mathbb{Q}}$ (\cite[p.~357]{N}) over the field of rational numbers $\mathbb{Q}$ is just the direct product $\prod_p\mathbb{Q}_p$ and the topology coincides with the product topology. Since $\mathbb{Q}$ is dense in $\mathbb{Q}_p$ for each $p$, for the diagonal embedding, $\mathbb{Q}$ is also dense in $\mathbb{A}_{\mathbb{Q}}$ by the approximation theorem. On the other hand, $\mathbb{Q}$ is discrete in $\mathbb{A}_{\mathbb{Q}}$ because of $\{(-1/2, 1/2) \times \prod_{p \neq \infty}\mathbb{Z}_p\}\cap \mathbb{Q} = \{0\}$. Since every discrete subgroup of a Hausdorff topological group is closed, we have that $\mathbb{Q}$ is equal to $\mathbb{A}_{\mathbb{Q}}$. This is clearly impossible. \end{proof} Furstenberg \cite{F} found a beautiful topological proof of Euclid's theorem. The above second proof also gives a topological proof. \section{The divergence of the sum of prime reciprocals via arithmetic progressions} Euler gave an analytic proof of Euclid's theorem by using the Euler product formula of the Riemann zeta function and the divergence of the harmonic series. Euler proved not only Euclid's theorem, but also the following stronger fact \cite[Theorema 19]{Eu}: \begin{Euler} The sum of prime reciprocals diverges$:$ \[ \sum_p\frac{1}{p}=\infty. \] \end{Euler} Erd\H os \cite{E} gave another proof of Euler's theorem by combining his combinatorial counting proof of Euclid's theorem and some additional estimate (Lemma \ref{lem2}). Most of other proofs of Euclid's theorem seem to have no potential to be extended to a proof of Euler's theorem. However, if we use Erd\H os' estimate to ensure that a certain set has positive upper density and use the celebrated theorem by Szemer\'edi instead of using van der Waerden's theorem in the proof of Euclid's theorem by Alpoge \cite{A}, we can give a new proof of Euler's theorem \emph{without the divergence of the harmonic series or quantitative argument}. We recall Szemer\'edi's theorem: \begin{definition}[upper density] Let $A$ be a set of positive integers. Then, we define \emph{the upper density} $\overline{d}(A)$ of $A$ by \[ \overline{d}(A) := \limsup_{N \to \infty}\frac{\#(A \cap \{1, 2, \dots, N\})}{N}. \] \end{definition} \begin{Szemeredi} Let $A$ be a set of positive integers which has positive upper density. Then, $A$ contains an arithmetic progression of length $k$, for every positive integer $k$. \end{Szemeredi} The case $k=3$ was proved by Roth \cite{R} in 1953 and the case $k=4$ was proved by Szemer\'edi \cite{S1} in 1969. Finally, the general case was established by Szemer\'edi \cite{S2} in 1975. Note that the proofs become much more complicated as $k$ is larger. Actually, we need not to use Szemer\'edi's theorem, and Roth's theorem is sufficient to deduce Euler's theorem. After the proof by Alpoge, Granville \cite{G} found another way of deducing Euclid's theorem by van der Waerden's theorem. Although we use an existence of an arithmetic progression of sufficiently large length in Alpoge's method, we use only an arithmetic progression of length four in Granville's proof by Fermat's theorem for squares in an arithmetic progression. Furthermore, it is enough in the length-three case if we replace Fermat's theorem with the following Euler--Legendre's theorem \cite[Vol.\ II.\ 572--573]{D}: \begin{Euler--Legendre} There are no length-three arithmetic progressions whose terms are cubes of positive integers. \end{Euler--Legendre} Based on the above observation, we give a new proof of the divergence of the sum of prime reciprocals by Roth's theorem and Euler--Legendre's theorem. This is still \emph{overkill}, but we see the power of Roth's theorem. We need the following two lemmas. \begin{lemma}[Pigeonhole principle for upper density] Let $A$ be a set of positive integers with $\overline{d}(A) > 0$. If $A$ is partitioned into finitely many classes, then there is at least one class which has positive upper density. \label{lem1}\end{lemma} \begin{proof} This is clear by definition. \end{proof} \begin{lemma} Let $p_j$ denote the $j$th prime number and $P_r$ be the set of positive integers which do not have $p_{r+1}, p_{r+2}, \dots$ as their prime factors. We assume that the sum of prime reciprocals converges. Then, there exists some positive integer $r$ such that $P_r$ has positive upper density. \label{lem2}\end{lemma} This is a rephrasing of Erd\H os' estimate in \cite{E}. \begin{proof} Under the assumption, we take a positive integer $r$ as $\sum_{j > r}\frac{1}{p_j} \leq 1/2$ holds. Since positive integers less than or equal to $N$ which are not contained in $P_r$ are divided by at least one of $p_{r+1}, p_{r+2}, \dots$, we have \[ N-\#(P_r \cap \{1, 2, \dots, N\}) \leq \sum_{j > r}\left\lfloor\frac{N}{p_j}\right\rfloor \leq \sum_{j > r}\frac{N}{p_j} \leq \frac{N}{2} \] for any positive integer $N$. Thus, we have \[ \#(P_r \cap \{1, 2, \dots, N\}) \geq \frac{N}{2} \] and $\overline{d}(P_r) \geq 1/2$. \end{proof} \begin{proof}[New proof of Euler's theorem] We assume that the sum of prime reciprocals converges. Let $r$ and $P_r$ be as in Lemma \ref{lem2}. For a tuple $v \in \{0, 1, 2\}^r$, we define a subset $P_r^{(v)}$ of $P_r$ by \[ P_r^{(v)} := \{n \in P_r \mid n=p_1^{e_1}\cdots p_r^{e_r}, \ (e_1, \dots, e_r) \equiv v \pmod{3}\}. \] Then, by Lemma \ref{lem1} and Lemme \ref{lem2}, there exists a tuple $v \in \{0, 1, 2\}^r$ such that the upper density of $P_r^{(v)}$ is positive. Hence, by Roth's theorem, there are positive integers $A$ and $D$ satisfying $A, A+D, A+2D \in P_r^{(v)}$. Let $R$ be the unique cubefree integer in $P_r^{(v)}$. Then, $A$ and $D$ are divided by $R$ and all $a, a+d, a+2d$ are cubes for $a:=A/R$ and $d:=D/R$. This contradicts to Euler--Legendre's theorem. Therefore, the sum of prime reciprocals diverges. \end{proof} \begin{remark} Darmon and Merel \cite{DM} proved that there are no non-trivial length-three arithmetic progressions whose terms are $n$th powers for $n \geq 3$. \end{remark} \section*{Acknowledgment.} The author would like to thank Junnosuke Koizumi for letting the author know Alpoge's work \cite{A}. He also would like to thank Kenji Sakugawa, Masataka Ono, Yuta Suzuki, and Toshiki Matsusaka for their valuable comments.

https://arxiv.org/abs/1707.04926

Theoretical insights into the optimization landscape of over-parameterized shallow neural networks

In this paper we study the problem of learning a shallow artificial neural network that best fits a training data set. We study this problem in the over-parameterized regime where the number of observations are fewer than the number of parameters in the model. We show that with quadratic activations the optimization landscape of training such shallow neural networks has certain favorable characteristics that allow globally optimal models to be found efficiently using a variety of local search heuristics. This result holds for an arbitrary training data of input/output pairs. For differentiable activation functions we also show that gradient descent, when suitably initialized, converges at a linear rate to a globally optimal model. This result focuses on a realizable model where the inputs are chosen i.i.d. from a Gaussian distribution and the labels are generated according to planted weight coefficients.

\section{Numerical experiments} \begin{figure}[!t] \centering \begin{subfigure}[b]{.43\textwidth} \includegraphics[width=\linewidth]{SoftPlus-varyK.pdf} \hspace{1cm} \caption{$d =10$}\label{SoftPlus-varyK} \end{subfigure} \hspace{1cm} \begin{subfigure}[b]{.43\textwidth} \includegraphics[width=\linewidth]{SoftPlus-varyD.pdf} \hspace{1cm} \caption{$k=10$}\label{SoftPlus-varyD} \end{subfigure} \caption{Empirical probability of having no spurious local minimizers for a one-hidden layer neural network with softplus activation. Dotted points depict the empirical probabilities, the solid line is obtained by fitting a logistic model to the dotted points, and the dashed line depicts the point where the probability is $1/2$. The number of samples are equal to $n=100$. In (a) we fix $d=10$ and vary $k$ and in (b) we fix $k=10$ and vary $d$. In both experiments as the number of parameters $(kd)$ increases beyond the sample size $n$, the probability of having no spurious local minimizer increases.} \label{fig:softplus} \end{figure} { In this section, we numerically investigate whether gradient descent finds the global optimum for various configurations of $n,k,d$. In our first experiment, the data is generated from a planted Gaussian model of the form \begin{align*} y_i=\langle\vct{v}^*,\phi(\mtx{W}^*\vct{x}_i)\rangle\quad\text{with}\quad\vct{x}_i \sim \mathcal{N}(\vct{0},\mtx{I}_d), \end{align*} and $\vct{v}^*$ the all-one vector. We consider two activations, namely the softplus $\phi(z)=\frac{1}{10}\log\left(1+e^{10z}\right)$ and the quadratic activation $\phi(z) =z^2$.} {In Figure \ref{fig:softplus}, we show the results for the softplus activation with $n=100$. In Figure~\ref{SoftPlus-varyK}, we fix the input dimension at $d = 10$ and vary the number of hidden nodes $k$. For each value of ($n$, $k$, $d$) we carryout $10$ experiments. In each experiment, we generate $\mtx{W}^*$ at random with i.i.d.~$\mathcal{N}(0,1)/\sqrt{d}$ entries. For each value of $\mtx{W}^*$ we carry out $10$ trials where in each trial we run gradient descent starting from a random initialization pair $(\vct{v}_0,\mtx{W}_0)$ generated at random with $\vct{v}_0$ consisting of i.i.d.~Rademacher $\pm 1$ entries and $\mtx{W}_0$ with i.i.d.~$\mathcal{N}(0,1)/\sqrt{d}$ entries. If a global minimum was obtained for every single initialization, we declare that the loss function has no spurious local minima for the corresponding $\mtx{W}^*$. In Figure \ref{fig:softplus} we plot the empirical probability that the loss function has no spurious local optima for the softplus activation. Dots correspond to the simulation results and the solid curve is obtained by fitting a logistic model to the results. In Figure \ref{SoftPlus-varyK} we focus on different number of hidden units, with the input dimension fixed at $d=10$. The dashed line indicates the value of $k$ ($k= 11.40$) at which the fitted logistic model crosses the probability $1/2$. As $k$ grows we enter a region that the gradient descent frequently converges to global minimizers, and for $k\ge 13$ every single local minimizer found by gradient descent is a global minimizer. Note that in this regime the number of parameters $(kd)$ exceeds the sample size $n$. This suggests that a modest amount of over-parameterization is sufficient for the no spurious local optima conclusion to hold and the global convergence to occur. In Figure~\ref{SoftPlus-varyD}, we fix the number of hidden nodes at $k = 10$ and vary the input dimension $d$. We observe a similar trend: As $d$ grows we enter a region where every single local minimizer found by gradient descent is a global minimizer. Here, the dashed line is located at $d = 10.70$, corresponding to a $1/2$ probability of converging to a global minima. For $d \ge12 $, gradient descent always finds a global minimizer.} {We also repeat a similar experiment with quadratic activations and report the results in Figure~\ref{fig:quad}. In Figure \ref{Quad activation-varyK}, we fix the input dimension at $d = 20$ and the number of samples at $n=100$ and vary $k$. The dashed line here is located at $k= 5.5$. In Figure \ref{Quad activation-varyD}, we fix the number of hidden units at $k = 5$ and vary $d$. The dashed line in this experiment is located at $d= 25.40$. These experiments further corroborate our theory that over-parameterization helps gradient descent find a global minimizer with quadratic activations.} \begin{figure}[!t] \centering \begin{subfigure}[b]{.43\textwidth} \includegraphics[width=\linewidth]{Quad-varyK.pdf} \hspace{1cm} \caption{$d = 20$}\label{Quad activation-varyK} \end{subfigure} \hspace{1cm} \begin{subfigure}[b]{.43\textwidth} \includegraphics[width=\linewidth]{Quad-varyD.pdf} \hspace{1cm} \caption{$k=5$}\label{Quad activation-varyD} \end{subfigure} \caption{Empirical probability of having no spurious local minimizers for a one-hidden layer neural network with quadratic activations. Dotted points depict the empirical probabilities, the solid line is obtained by fitting a logistic model to the dotted points, and the dashed line depicts the point where the probability is $1/2$. The number of samples is $n=100$. In (a) we fix $d=10$ and vary $k$ and in (b) we fix $k=10$ and vary $d$. In both experiments as the number of parameters $(kd)$ increases beyond the sample size $n$, the probability of having no spurious local minimizer increases.} \label{fig:quad} \end{figure} {In the next set of experiments, we generate the labels and features i.i.d.~at random \begin{align*} \vct{x}_i \sim \mathcal{N}(\vct{0},\mtx{I}_d)\quad\text{and}\quad y_i\sim \mathcal{N}( 0, 1). \end{align*} We fit a softplus one-hidden layer network of the form $\vct{x}\rightarrow\vct{1}^T\phi\left(\mtx{W}\vct{x}\right)$ to this data with varying number of hidden units $k$. In Figure \ref{fig:softplus-quad-rand-data} we plot the square root of the objective function $\mathcal{L}$ as a function of iterations. Interestingly, we observe that even for randomly labeled data when $kd>1.5n$, gradient descent is able to find a global minimizer. These experiments show that simple gradient descent is almost always able to find a global minimizer in the sufficiently over-parametrized regime, regardless of whether the data is generated from a planted model.} \begin{figure}[!t] \centering \begin{subfigure}[b]{.49\textwidth} \includegraphics[width=\linewidth]{arb_data_data_relu.pdf} \hspace{1cm} \caption{Softplus with random data $(d,n)=(10,100)$.} \end{subfigure} \begin{subfigure}[b]{.49\textwidth} \includegraphics[width=\linewidth]{arb_data_data_quad.pdf} \hspace{1cm} \caption{Quadratic with random data $(d,n)=(20,100)$.} \end{subfigure} \caption{Plot of RMSE=$\sqrt{\mathcal{L}(\vct{v},\mtx{W})}$ vs. the number of gradient descent iterations. As the number of hidden units $k$ increases, gradient descent converges to a solution with zero RMSE. } \label{fig:softplus-quad-rand-data} \end{figure} \section{Introduction} \subsection{Motivation} Neural network architectures (a.k.a.~deep learning) have recently emerged as powerful tools for automatic knowledge extraction from raw data. These learning architectures have lead to major breakthroughs in applications such as visual object classification \cite{krizhevsky2012imagenet}, speech recognition \cite{mohamed2012acoustic} and natural language processing \cite{collobert2008unified}. Despite their wide empirical use the mathematical success of these architectures remains a mystery. Although the expressive ability of neural networks is relatively well-understood \cite{barron1994approximation}, computational tractability of training such networks remains a major challenge. In fact, training neural nets is known to be NP-hard even for very small networks \cite{blum1988training}. The main challenge is that training neural networks correspond to extremely high-dimensional and nonconvex optimization problems and it is not clear how to provably solve them to global optimality. Worst-case pessimism aside, these networks are trained successfully in practice via local search heuristics on real or randomly generated data. In particular, over-parameterized neural networks-where the number of parameters exceed the number of data samples-can be optimized to global optimality using local search heuristics such as gradient or stochastic gradient methods \cite{zhang2016understanding}. In this paper we wish to provide theoretical insights into this phenomenon by developing a better understanding of optimization landscape of such over-parameterized shallow neural networks. \subsection{Problem formulation and models} A fully connected artificial neural network is composed of computational units called neurons. The neurons are decomposed into layers consisting of one input layer, one output layer and a few hidden layers with the output of each layer fed in (as input) to the next layer. In this paper we shall focus on neural networks with only a single hidden layer with $d$ inputs, $k$ hidden neurons and a single output. The overall input-output relationship of the neural network in this case is a function $f:\mathbb{R}^d\rightarrow\mathbb{R}$ that maps the input vector $\vct{x}\in\mathbb{R}^d$ into a scalar output via the following equation \begin{align} \label{model} \vct{x}\mapsto f_{\vct{v},\mtx{W}}(\vct{x}):=\sum_{\ell=1}^k\vct{v}_\ell\phi(\langle\mtx{w}_\ell,\vct{x}\rangle). \end{align} In the above the vectors $\vct{w}_\ell\in\mathbb{R}^d$ contains the weights of the edges connecting the input to the $\ell$th hidden node and $\vct{v}_\ell\in\mathbb{R}$ is the weight of the edge connecting the $\ell$th hidden node to the output. Finally, $\phi:\mathbb{R}\rightarrow\mathbb{R}$ denotes the activation function applied to each hidden node. For more compact notation we gathering the weights $\vct{w}_\ell/\vct{v}_\ell$ into larger matrices $\mtx{W}\in\mathbb{R}^{k\times d}$ and $\vct{v}\in\mathbb{R}^k$ of the form \begin{align*} \mtx{W}=\begin{bmatrix}\vct{w}_1^T\\\vct{w}_2^T\\\vdots\\\vct{w}_k^T\end{bmatrix}\quad\text{and}\quad\vct{v}=\begin{bmatrix}{v}_1\\{v}_2\\\vdots\\ {v}_k\end{bmatrix}. \end{align*} We can now rewrite our input-output model \eqref{model} in the more succinct form \begin{align} \label{model2} \vct{x}\mapsto f_{\vct{v},\mtx{W}}(\vct{x}):=\vct{v}^T\phi(\mtx{W}\vct{x}). \end{align} Here, we have used the convention that when $\phi$ is applied to a vector is corresponds to applying $\phi$ to each entry of that vector. When training a neural network, one typically has access to a data set consisting of $n$ feature/label pairs $(\vct{x}_i,y_i)$ with $\vct{x}_i\in\mathbb{R}^d$ representing the feature and $y_i$ the associated label. We wish to infer the best weights $\vct{v},\mtx{W}$ such that the mapping $f_{\vct{v},\mtx{W}}$ best fits the training data. More specifically, we wish to minimize the misfit between $f_{\vct{v},\mtx{W}}(\vct{x}_i)$ and $y_i$ via an optimization problem of the form \begin{align} \underset{\vct{v}, \mtx{W}}{\min}\text{ }\mathcal{L}(\vct{v},\mtx{W})\,, \end{align} where \begin{align} \label{landscape} \mathcal{L}(\vct{v},\mtx{W}):=\frac{1}{2n}\sum_{i=1}^n \left(y_i-f_{\vct{v},\mtx{W}}(\vct{x}_i)\right)^2=\frac{1}{2n}\sum_{i=1}^n\left(y_i-\vct{v}^T\phi(\mtx{W}\vct{x}_i)\right)^2. \end{align} In this paper we wish to understand the landscape of optimization problems of the form \eqref{landscape}. \section{Main results} In this section we discuss the main results of this paper. The first set of results focus on understanding the global landscape of neural network optimization with one hidden layer with a particular activation function. We also discuss how this landscape characterization enables algorithms to find a global optima in polynomial time. The second set of results focuses on the local convergence of gradient descent but applies to a broad set of activation functions. \subsection{Global landscape analysis with quadratic activations}\label{subsec:global} We begin by discussing some global properties of the loss function of training neural networks. \begin{theorem}\label{globthm} Assume we have an arbitrary data set of input/label pairs $\vct{x}_i\in\mathbb{R}^d$ and $y_i$ for $i=1,2,\ldots,n$. Consider a neural network of the form \begin{align*} \vct{x}\mapsto \vct{v}^T\phi(\mtx{W}\vct{x}), \end{align*} with $\phi(z)=z^2$ a quadratic activation and $\mtx{W}\in\mathbb{R}^{k\times d}$, $\vct{v}\in\mathbb{R}^k$ denoting the weights connecting input to hidden and hidden to output layers. We assume $k\ge 2d$ and set the weights of the output layer $\vct{v}$ so as to have at least $d$ positive entries and at least $d$ negative entries. Then, the training loss as a function of the weights $\mtx{W}$ of the hidden layer \begin{align*} \mathcal{L}(\mtx{W})=\frac{1}{2n}\sum_{i=1}^n\left(y_i-\vct{v}^T\phi(\mtx{W}\vct{x}_i)\right)^2, \end{align*} obeys the following two properties. \begin{itemize} \item There are no spurious local minima, i.e.~all local minima are global. \item All saddle points have a direction of strictly negative curvature. That is, at a saddle point $\mtx{W}_s$ there is a direction $\mtx{U}\in\mathbb{R}^{k\times d}$ such that \begin{align*} \emph{vect}(\mtx{U})^T\nabla^2\mathcal{L}(\mtx{W}_s)\emph{vect}(\mtx{U})<0. \end{align*} \end{itemize} Furthermore, for almost every data inputs $\{\vct{x}_i\}_{i=1}^n$, as long as \begin{align*} d\le n\le cd^2, \end{align*} the global optimum of $\mathcal{L}(\mtx{W})$ is zero. Here, $c>0$ is a fixed numerical constant. \end{theorem} The above result states that given an arbitrary data set, the optimization landscape of fitting neural networks have favorable properties that facilitate finding globally optimal models. In particular, by setting the weights of the last layer to have diverse signs all local minima are global minima and all saddles have a direction of negative curvature. This in turn implies that gradient descent on the input-to-hidden weights, when initialized at random, converges to a global optima. All of this holds as long as the neural network is sufficiently wide in the sense that the number of hidden units exceed the dimension of the inputs by a factor of two ($k\ge 2d$). An interesting and perhaps surprising aspect of the first part of this theorem is its generality: it applies to any arbitrary data set of input/label pairs of any size! However, this result only guarantees convergence to a global optima but does not explain how good this global model is at fitting the data. Stated differently, it does not provide a bound on the optimal value. Such a bound may not be possible with adversarial data. However, at least intuitively, one expects the optimal value to be small when the input data $\vct{x}_i$ are sufficiently diverse and the neural network is sufficiently over-parameterized. The second part of Theorem \ref{globthm} confirms that this is indeed true. In particular assuming the input data $\vct{x}_i$ are distributed i.i.d.~$\mathcal{N}(0,\mtx{I}_d)$, and the total number of weights ($kd$) exceeds the number of data samples ($n$), the globally optimal model perfectly fits the labels (has optimal value of $0$). The interesting part of this result is that it still holds with arbitrary labels. Thus, this theorem shows that with random inputs, and a sufficiently diverse choice of the hidden-output weights, using gradient descent to iteratively update the input-hidden layer weights converges to a model that provably fits arbitrary labels! This result is also inline with recent numerical evidence in \cite{zhang2016understanding} that demonstrates that stochastic gradient descent learns deep, over-parametrized models with zero training error even with an arbitrary choice of labels. While the above theorem shows that the saddles are strict and there is a direction of negative curvature at every saddle point, the margin of negativity is not quantified. Thus, the above theorem does not provide explicit convergence guarantees. In the theorem below we provide such a guarantee albeit for a more restrictive data model. \begin{theorem}\label{gthmrand} Assume we have a data set of input/label pairs $\{(\vct{x}_i,y_i)\}_{i=1}^n$ with the labels $y_i\in\mathbb{R}$ generated according to a planted two layer neural network model of the form \begin{align*} y_i=\langle\vct{v}^*,\phi(\mtx{W}^*\vct{x}_i)\rangle, \end{align*} with $\phi(z)=z^2$ a quadratic activation and $\mtx{W}^*\in\mathbb{R}^{k\times d}$, $\vct{v}^*\in\mathbb{R}^k$ the weights of the input-hidden and hidden-output layer with $\sigma_{\min}(\mtx{W}^*)>0$. Furthermore, assume $k\ge d$ and that all the non-zero entries of $\vct{v}^*$ have the same sign (positive or negative). We set the hidden-output weights to a vector $\vct{v}\in\mathbb{R}^k$ with non-zero entries also having the same sign with at least $d$ entries strictly nonzero (positive if the nonzero entries of $\vct{v}^*$ are positive and negative otherwise). Then, as long as \begin{align*} d\le n\le cd^2, \end{align*} with $c$ a fixed numerical constant, then the training loss as a function of the input-output weights $\mtx{W}$ of the hidden layer \begin{align*} \mathcal{L}(\mtx{W})=\frac{1}{2n}\sum_{i=1}^n\left(y_i-\vct{v}^T\phi(\mtx{W}\vct{x}_i)\right)^2, \end{align*} obeys the following two properties, for almost every input data $\{\vct{x}_i\}_{i=1}^n$. \begin{itemize} \item There are no spurious local minima, i.e.~all local minima are global optima. \item All saddle points have a direction of negative curvature. That is, at a saddle point $\mtx{W}_s$ there is a direction $\mtx{U}\in\mathbb{R}^{k\times d}$ such that \begin{align*} \emph{vect}(\mtx{U})^T\nabla^2\mathcal{L}(\mtx{W}_s)\emph{vect}(\mtx{U})<0. \end{align*} \item The global optima has loss value equal to zero ($\mathcal{L}(\mtx{W})=0$). \end{itemize} Furthermore, suppose that the inputs $\vct{x}_i\in\mathbb{R}^d$ are distributed i.i.d.~$\mathcal{N}(\vct{0},\mtx{I}_d)$. Assume $k>d$ and $cd\log d\le n \le Cd^2$ for fixed constants $c$ and $C$. Also, assume we set all entries of $\vct{v}$ to $\nu$ (i.e.~$\vct{v}=\nu\vct{1}$) with $\nu$ having the same sign as the nonzero entries of $\vct{v}^*$. \begin{itemize} \item Then all points $\mtx{W}$ satisfying the approximate local minima condition \begin{align} \label{approxmin} \fronorm{\nabla \mathcal{L}(\mtx{W})}\le \epsilon_g\quad\text{and}\quad\nabla^2\mathcal{L}(\mtx{W})\succeq -\epsilon_{H}\mtx{I}, \end{align} obey \begin{align*} \mathcal{L}(\mtx{W})\le\frac{\epsilon_g}{\sqrt{\nu}}\max\left(\sqrt{1+14\nucnorm{{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*}},4\frac{\epsilon_g}{\sqrt{\nu}}\right)+\frac{\epsilon_{H}}{2\nu}\nucnorm{{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*}, \end{align*} with probability at least $1-12de^{-\gamma d}-8/d$. Here $\gamma$ is a fixed numerical constant. \end{itemize} \end{theorem} The above result considers the setting where the input-output data set is generated according to a neural network model with Gaussian random input vectors. We show that if the data is generated according this model, then as long as the neural network is over-parameterized ($n\lesssim kd$) then all points obeying condition \eqref{approxmin} have small objective value. The reason this result is useful is that points obeying the approximate local minimum condition \eqref{approxmin} can be found in time depending polynomially on $\frac{1}{\epsilon_g}$ and $\frac{1}{\epsilon_{H}}$. Algorithms that provably work in this setting include cubic regularization \cite{nesterov2006cubic}, trust region methods \cite{curtis2014trust}, approximate cubic regularization schemes \cite{agarwal2016finding, carmon2016gradient, carmon2016accelerated}, randomly initialized and perturbed (stochastic) gradients \cite{jin2017escape, ge2015escaping, lee2016gradient, levy2016power, jin2017escape}. Therefore, the above theorem demonstrates that a weight matrix with small training error(i.e.~$\mathcal{L}(\mtx{W})\le\epsilon$) can be found in a computationally tractable manner (specifically with poly$(\frac{1}{\epsilon})$ computational effort). \subsection{Local convergence analysis with general activations} We now extend our analysis to general activations functions. However, our results in this section require a sufficiently close initialization to the global optimum. Starting from such a sufficiently close initialization we apply gradient descent updates based on the loss function \eqref{landscape}. Our results apply to any differentiable activation function with bounded first and second order derivatives. We believe our result will eventually extend also to non-differentiable activations by smoothing techniques but we do not pursue this direction in this paper. Before we state our theorem, we make a technical assumption regarding the activation $\phi$. \begin{assumption}\label{ass-phi} For $\sigma\in \mathbb{R}_{\ge 0 }$, define $\mu(\sigma) = E[\phi'(\sigma g)]$ and $\gamma(\sigma) = E[\phi''(\sigma g)]$ where $g\sim\mathcal{N}(0,1)$. We consider activations $\phi$ such that $\mu(\sigma)$ is zero/nonzero everywhere. Likewise, we assume that one of the following holds true for $\gamma(\sigma)$: \begin{itemize} \item [$(a)$] Function $\gamma(\sigma)$ is nonzero everywhere. \item [$(b)$] Function $\gamma(\sigma)$ is identical to zero. \end{itemize} \end{assumption} We note that $\mu(\sigma)$ and $\gamma(\sigma)$ can be thought of as the average slope and curvature of the activation function. Thus the assumption on $\mu(\sigma)$ can be interpreted as the activation should have a nonzero average slope for all $\sigma>0$ (i.e.~the mapping is somewhat correlated with the identity mapping $\phi(x)=x$, under any gaussian measure) or has average slope equal to zero for all $\sigma>0$ (i.e.~the mapping has no correlation with the identity mapping, under any gaussian measure). \begin{example}\label{myex} We provide several examples of an activation function that satisfy Assumption~\ref{ass-phi}. \begin{enumerate} \item (Softplus) $\phi_b(z) = \frac{1}{b} \log(1+e^{bz})$, for a fixed $b >0$. \item (Sigmoid) $\phi_b(z) = \frac{1}{1+e^{-bz}}$, for a fixed $b>0$. \item (Erf) $\phi(z) =\frac{2}{\sqrt{\pi}} \int_0^z e^{-t^2/2} {\rm d} t$. \item (Hyperbolic Tangent) $\phi(z) = \tanh(z)$. \end{enumerate} Note that all of these activations obey $\mu(\sigma) > 0$ as they are all strictly increasing. Furthermore, the Softplus activation satisfies Assumption~\ref{ass-phi} (a) because it is strictly convex, while the other three activations satisfy Assumption~\ref{ass-phi} (b), because $\phi''$ is an odd function in these cases. These activations along with the popular ReLU activation ($\phi(z)=\max(0,z)$) are depicted in Figure \ref{fig:softplus-relu}. \end{example} \begin{figure} \centering \begin{tikzpicture}[scale=1] \begin{groupplot}[scale=1,group style={group size=2 by 1,horizontal sep=1cm,xlabels at=edge bottom, ylabels at=edge left,xticklabels at=edge bottom},xlabel=$z$, ylabel=$\phi(z)$, legend style={font=\tiny,at={(0.25,0.95)},anchor=north,legend cell align=left}] \nextgroupplot \addplot [blue,line width=2pt] table[x index=0,y index=1]{./Activations}; \addlegendentry{ReLU} \addplot [magenta,line width=2pt] table[x index=0,y index=2]{./Activations}; \addlegendentry{Softplus} \nextgroupplot \addplot [red,line width=2pt] table[x index=0,y index=3]{./Activations}; \addlegendentry{Sigmoid} \addplot [black,line width=2pt] table[x index=0,y index=4]{./Activations}; \addlegendentry{Erf} \addplot [teal,line width=2pt] table[x index=0,y index=5]{./Activations}; \addlegendentry{Hyperbolic tangent} \end{groupplot} \end{tikzpicture} \caption{Different activations from Example \ref{myex} with $b=4$ along with the ReLU activation.} \label{fig:softplus-relu} \end{figure} We now have all the elements to state our local convergence result. \begin{theorem}\label{localthm} Assume we have a data set of input/label pairs $\{(\vct{x}_i,y_i)\}_{i=1}^n$ with the inputs $\vct{x}_i\in\mathbb{R}^d$ distributed i.i.d.~$\mathcal{N}(\vct{0},\mtx{I}_d)$ and the labels $y_i\in\mathbb{R}$ generated according to a planted two layer neural network model of the form \begin{align*} y_i=\langle\vct{v}^*,\phi(\mtx{W}^*\vct{x}_i)\rangle, \end{align*} Here, $\phi:\mathbb{R}\rightarrow \mathbb{R}$ is any activation function with bounded first and second derivatives, satisfying Assumption~\ref{ass-phi}($a$). Further, $\mtx{W}^*\in\mathbb{R}^{k\times d}$, $\vct{v}^*\in\mathbb{R}^k$ are the weights of the input-hidden and hidden-output layer. We also assume that the planted weight vector/matrix $\vct{v}^*/\mtx{W}^*$ obey \begin{align} \label{boundassthm} 0<v_{\min} \le |v^*_\ell| \le v_{\max}\quad0<w_{\min} \le \twonorm{\vct{w}^*_\ell} \le w_{\max}\quad\text{for }\ell=1,2,\ldots,k\,, \end{align} for some fixed constants $v_{\min}, v_{\max}, w_{\min}$, and $w_{\max}$. Further, assume that $k\ge d$ and \begin{align} d \le n \le c_0 \frac{\sigma_{\min}^4(\mtx{W}^*)}{\sigma_{\max}^8(\mtx{W}^*)} d^2 \label{samplesize} \end{align} for a sufficiently small constant $c_0>0$. To estimate the weights $\vct{v}^*$ and $\mtx{W}^*$, we start from initial weights $\vct{v}_0$ and $\vct{W}_0$ \begin{align} \fronorm{\mtx{W}_0-\mtx{W}^*} &\le C_0\frac{\sigma_{\min}^3(\mtx{W}^*)}{\sigma_{\max}(\mtx{W}^*)}\cdot \frac{d^{2.5}}{n^{1.5} k} \,, \label{eq:initW0}\\ \opnorm{{\mtx{v}}_0-{\mtx{v}}^*}_\infty &\le C_0\frac{\sigma_{\min}^3(\mtx{W}^*)}{\sigma_{\max}(\mtx{W}^*) }\cdot \frac{d^{2.5}}{n^{1.5} k^{1.5}} \,, \label{eq:initv0} \end{align} and apply gradient descent updates of the form \begin{align} \label{gradd} \vct{v}_{\tau+1} &= \vct{v}_{\tau}-\alpha\nabla_{\vct{v}} \mathcal{L}(\vct{v}_\tau,\mtx{W}_\tau),\nonumber\\ \vct{W}_{\tau+1} &=\vct{W}_{\tau}-\alpha\nabla_{\vct{W}} \mathcal{L}(\vct{v}_\tau,\mtx{W}_\tau), \end{align} with the learning rate obeying $\alpha\le 1/\beta$. Then there is an event of probability at least $1 - n^{-1} - 2ne^{-b \sqrt{d}}$, such that on this event starting from any initial point obeying \eqref{eq:initW0}-\eqref{eq:initv0} the gradient descent updates \eqref{gradd} satisfy \begin{align*} \mathcal{L}({\mtx{v}}_\tau,\mtx{W}_\tau) \le \left(1- c \alpha \frac{d}{n} \right)^\tau \mathcal{L}({\mtx{v}}_0,\mtx{W}_0)\,. \end{align*} Here, $\beta$ is given by \begin{align} \beta: = C \left(\sigma_{\max}^2(\mtx{W}^*) + 1\right) k\,, \end{align} and $b, C_0, c_0, C, c>0$ are fixed numerical constants. \end{theorem} \begin{remark} We note that the above theorem is still valid if Assumption~\ref{ass-phi} (b) holds in lieu of Assumption~\ref{ass-phi} (a). The only change is that the term $\sigma_{\min}(\mtx{W})$ should be now replaced by one in Equations~\eqref{samplesize}, \eqref{eq:initW0} and \eqref{eq:initv0}. \end{remark} The above result considers the setting where the input-output data set is generated according to a neural network model with a general activation and Gaussian random input vectors. We show that if the data is generated according this model, then as long as the neural network is over-parameterized ($n\lesssim kd$) then gradient descent converges to the planted model when initialized close to this planted model. This implies that a training error of zero can be achieved locally, using gradient descent. We would like to note that assumptions \eqref{boundassthm} on the weights of the planted neural networks are only made to avoid unnecessarily complicated expressions in the statement of the theorem. As it will become clear in the proofs (Section \ref{proofs}) our result continues to hold without these assumptions. \input{Num2} \section{Related Work} \label{sec:related} As mentioned earlier, neural networks have enjoyed great empirical success \cite{krizhevsky2012imagenet, mohamed2012acoustic, collobert2008unified}. To explain this success, many papers have studied the expressive ability of shallow neural networks dating back to the 80s (e.g.~see \cite{barron1994approximation}). More recently, interesting results have focused on the expressive ability of deeper and sparser architectures \cite{telgarsky2016benefits, bolcskei2017optimal, kuo2017cnn, wiatowski2017energy}. Computational tractability of training networks however is still a major challenge. In fact, training neural nets is known to be NP-hard even for very small networks \cite{blum1988training}. Despite this worst-case pessimism, local search heuristics such as gradient descent are surprisingly effective. For instance, \cite{zhang2016understanding} empirically demonstrated that sufficiently over-parametrized networks can be efficiently optimized to near global optimality with stochastic gradient descent. For a neural network with zero hidden units and a single output with a monotonic activation function $\sigma$, numerous authors \cite{mei2016landscape,hazan2015beyond,kakade2011efficient,kalai2009isotron, soltanolkotabi2017learning,soltanolkotabi2017theoretical} have shown that gradient-based methods converge to the global optimum under various assumptions and models. As a result of this literature a good understanding of the optimization landscape of learning single activations have emerged (at least when the data set is generic). Roughly stated, in this case the loss function only has a single local optima that coincides with the global optima and the gradient descent direction is sufficiently correlated with the direction pointing to the global optimum. Thus these results are able to explain the success of gradient-based methods. However, when there are hidden units, the loss function exhibits a completely different landscape, so that such analyses do not easily generalize. We now turn our attention to global convergence results known for deeper architectures. For a 2-layered network with leaky ReLU activation,~\cite{soudry2016no} showed that gradient descent on a modified loss function can obtain a global minimum of the modified loss function; however, this does not imply reaching a global minimum of the original loss function. Under the same setting, \cite{xie2016diversity} showed that critical points with large ``diversity" are near global optimality. \cite{choromanska2015loss} used several assumptions to simplify the loss function to a polynomial with i.i.d.~Gaussian coefficients. They then showed that every local minima of the simplified loss has objective value comparable to the global minima. \cite{kawaguchi2016deep} used similar assumptions to show that all local minimum are global minimum in a nonlinear network. However, \cite{choromanska2015loss,kawaguchi2016deep} require an \emph{independent activations} assumption meaning the activations of the hidden units are independent of the input and/or mutually independent, which is violated in practice. In comparison, our global convergence results hold for any arbitrary data set, without any additional assumptions. However, our global convergence results (see Section~\ref{subsec:global}) only focus on quadratic activations with a single hidden layer. We would like to mention a few interesting results regarding asymptotic characterizations of the dynamics of training shallow neural networks with Gaussian inputs using ideas from statistical physics \cite{saad1995line, biehl1996transient, vicente1997functional}. Saad and Solla \cite{saad1995line} study the online dynamics of learning a fully connected neural network with one-hidden layer in a Gaussian planted model (or Gaussian student-teacher model). In this model the output labels are generated according to planted weight vectors (a.k.a.~a teacher network) with Gaussian input data. Then a new ``student" network is trained to find the teacher weights. This result focuses on the case where the hidden-to-output weights are fixed to $+1$ (a.k.a.~soft committee machines) and the asymptotic regime where the number of inputs and the iterations tend to infinity ($d\rightarrow +\infty$ and $\tau\rightarrow \infty$). The authors provide certain dynamical equations that describes the asymptotic behavior of the training process for one-hidden layer neural networks using an ``online SGD" or ``resampled SGD" procedure where the SGD update is performed using a fresh and independent data point per new iteration. {These dynamical equations do not admit a closed form solution, but do enable the analysis of the convergence behavior close to stationarity.} Furthermore, by simulating these dynamical equations it is possible to gain some insights about the evolution of the online SGD updates. In particular, the authors use such simulations to gain insights into the convergence/divergence of such neural networks in the special case where diagonal input-hidden weights ($\mtx{W}$) are used. The paper \cite{biehl1996transient} also studies studies soft-committee machines under similar assumptions to those of \cite{saad1995line}. The authors of \cite{biehl1996transient} use tools from statistical physics (and in particular the aforementioned dynamical equations) to demonstrate that the dynamics of training such neural networks has several fixed points and find learning-rate dependent phenomena, such as splitting and disappearing of fixed points. This paper also differs from \cite{saad1995line} in that it also applies to transient iterations (finite $\tau$). Finally, the paper \cite{vicente1997functional} also studies soft committee machines but instead of online SGD, they present the analysis for a locally optimized online learning algorithm that focuses on extracting the largest possible amount of information from each new sample. The authors also demonstrate numerically that this choice leads to faster escape from the plateaux and hence faster decay of the generalization error. Our results are not directly comparable to this literature as it differs in terms of assumptions and conclusions in a variety of ways. In particular we focus on (1) analytic formulas, (2) gradient descent without resampling, (3) a non-asymptotic regime (finite $d$) and transient dynamics (finite $\tau$), (4) the over-parameterized case $n<kd$, and (5) general hidden-output weights. Some of the results presented in this paper also apply without assuming a planted model or random/Gaussian input data (e.g.~see Theorem \ref{globthm}). We would like to note that there is also interesting growing literature on learning shallow neural networks with a single hidden layer with i.i.d.~inputs, and under a realizable model (i.e.~the labels are generated from a network with planted weights) \cite{tian2017analytical, brutzkus2017globally, zhang2017electron, li2017convergence, janzamin2015beating, zhong2017recovery}. For isotropic Gaussian inputs, \cite{tian2017analytical} shows that with two hidden unites ($k=2$) there are no critical points for configurations where both weight vectors fall into (or outside) the cone of ground truth weights. With the same assumptions, \cite{brutzkus2017globally} proves that for a single-hidden ReLU network with a single non-overlapping convolutional filter, all local minimizers of the population loss are global; they also give counter-examples in the overlapping case and prove the problem is NP-hard when inputs are not Gaussian. \cite{zhang2017electron} show that for single-hidden layer networks with non-standard activation functions gradient descent converges to global minimizers. \cite{li2017convergence} focuses on a Recursive Neural Net (RNN) model where the hidden to output weights are close to the identity and shows that stochastic methods converge to the planted model over the population objective. \cite{zhong2017recovery} studies general single-hidden layer networks and shows that a version of gradient descent which uses a fresh batch of samples in each iteration converges to the planted model. This holds using an initialization obtained via a tensor decomposition method. Our approach and local convergence results differ from this literature in a variety of different ways. First, unlike some of these results such as \cite{brutzkus2017globally,li2017convergence}, we study the optimization properties of the empirical function, not its expected value. Second, we focus on the over-parametrized regime ($n<<kd$) which is the regime where most popular neural networks are trained. Mathematically, this is an important distinction as in this data-poor regime the empirical loss is no longer close to the population loss and therefore one can no longer infer the convergence behavior of gradient descent based on connecting the empirical loss to the population loss. Third, we optimize over both weights $\vct{v}$ and $\mtx{W}$, while most of this literature assumes $\vct{v}=\vct{v}^*=\vct{1}$ \cite{li2017convergence, zhong2017recovery}. Finally, our framework does not require a fresh batch of samples per new gradient iteration as in \cite{zhong2017recovery}. Several publications study the effect of over-parametrization on the training of neural networks \cite{poston1991local,haeffele2015global, nguyen2017loss}. These results require a large amount of over-parametrization, mainly that the width of one of the hidden layers to be greater than the number of training samples, which is unrealistic for commonly used networks. For instance, for the case of a single hidden layer using the notation of this paper these publications require $n\lesssim k$. These results also require additional technical assumptions which we do not detail here. However, these results work for fairly general activations and also deep architectures. In comparison, our global optimality results only allow for quadratic activations and a single hidden layer. However, our results are completely deterministic and allow for modest over-parameterization ($n\lesssim kd$). Generalizing these result to other activations and deeper architectures is an interesting future direction. While discussing quadratic activations we would like to mention the interesting paper \cite{livni2014computational}. This paper is perhaps the first result to clearly state that over-parameterization is helpful for optimization purposes. The authors further showed that the Frank-Wolfe algorithm can optimize a neural net with quadratic activations of the form $\phi(\vct{x})=b+\langle \vct{w}_0, x\rangle + \sum_{\ell=1}^K\alpha_k (\langle w_\ell, x\rangle)^2$ under norm constraints of the form $\abs{\alpha_\ell}\le1$ and $\twonorm{\vct{w}_\ell}=1$. In comparison we focus on slightly less general activations (no linear term) and study landscape properties that are directly useful for analyzing gradient descent/local search methods in lieu of Franke-Wolfe type schemes. Finally, in the case of over-parameterized neural networks not all global optima may generalize to new data instances. Understanding the generalization capability of the solutions reached by (stochastic) gradient descent is an important research direction. See \cite{bartlett2017spectrally} for an interesting result in this direction under margin assumptions. Another line of research \cite{goel2016reliably, shalev2011learning, zhang2016l1} focuses on improper learning models using kernel-based approaches. These results hold under much more general data models than the realizable case. However, the practical success of deep learning is intricately tied to using gradient-based training procedures, and the learnability of these networks using improper learning does not explain the success of gradient-based methods. Related, \cite{janzamin2015beating} proposes a method of moments estimator using tensor decomposition. Finally, we would like to mention a few interesting recent developments that appeared after the initial arxiv submission of this paper. The authors of \cite{safran2017spurious} demonstrated analytically (with variable precision arithmetics) that without over-parameterization the landscape of one-hidden layer neural networks has bad local minima as soon as the number of hidden unites exceeds six ($k\ge 6$). This interesting result clearly demonstrates that the over-parametrization is necessary to ensure a favorable optimization landscape when training one-hidden layer neural networks. Another recent paper \cite{soudry2017exponentially} by Soudry and Hoffer suggests that ``most" local minima are global in the over-parameterized regime. This paper provides further evidence that over-parametrizing neural networks leads to more favorable optimization landscapes when training such neural nets. Finally, the recent paper \cite{li2017algorithmic} proves that gradient descent when initialized at zero can lead to training errors close to (but not equal to) zero. This result holds for one-hidden layer neural networks with quadratic activations and Gaussian inputs. Developing rigorous convergence guarantees to a global optima with training error equal to zero which also applies to other activations is an important future research direction. \section{Preliminaries and notations} Before we dive into the proofs in the next two sections we collect some useful results and notations in this section. \subsection{Notations} For two matrices $\mtx{A} = [\mtx{A}_1,\dotsc, \mtx{A}_p]\in\mathbb{R}^{m\times p}$ and $\mtx{B} =[\mtx{B}_1,\dotsc, \mtx{B}_p]\in \mathbb{R}^{n\times p}$, we define their Khatri-Rao product as $\mtx{A} * \mtx{B} = [\mtx{A}_1\otimes \mtx{B}_1,\dotsc, \mtx{A}_p\otimes \mtx{B}_p]\in\mathbb{R}^{mn\times p}$, where $\otimes$ denotes the Kronecher product. For two matrices $\mtx{A}, \mtx{B}$, we denote their Hadamard (entrywise) product by $\mtx{A}\circ\mtx{B}$. For a matrix $\mtx{A}$, we denote its maximum and minim singular values by $\sigma_{\max}(\mtx{A})$ and $\sigma_{\min}(\mtx{A})$, respectively. For a random variable $Y$, its sub-exponential norms is defined as \[ \|Y\|_{\psi_1} = \inf\{C>0:\; \operatorname{\mathbb{E}} \exp(|Y|/C) \le 2\}\,. \] Further, for a centered random vector $\vct{x}\in \mathbb{R}^d$, we define its sub-exponential norm as \begin{align} \|\vct{x}\|_{\psi_1} = \sup_{\vct{y}\in S^{d-1}} \|\<\vct{x},\vct{y}\>\|_{\psi_1}\label{def2:subexp} \end{align} Throughout the paper we use $c, C$ to refer to constants whose values may change from line to line. \subsection{Derivative calculations} In this section we gather some derivative calculations that we will use throughout the proof. As a reminder the loss function is given by \begin{align*} \mathcal{L}(\vct{v},\mtx{W})=\frac{1}{2n}\sum_{i=1}^n\left(\vct{v}^T\phi(\mtx{W}\vct{x}_i)-y_i\right)^2. \end{align*} To continue let us define the residual vector $\vct{r}\in\mathbb{R}^n$ with the $i$th entry given by \begin{align*} r_i=\vct{v}^T\phi(\mtx{W}\vct{x}_i)-y_i. \end{align*} \subsubsection{First order derivatives} We begin by calculating the gradient with respect to the weight matrix $\mtx{W}$. To this aim we begin by calculating the gradient with respect to the $q$th row of $\mtx{W}$ denoted by $\vct{w}_q$. This is equal to \begin{align} \label{gradw} \nabla_{\vct{w}_q} \mathcal{L}(\vct{v},\mtx{W})=\frac{v_q}{n}\sum_{i=1}^n\left(\vct{v}^T\phi(\mtx{W}\vct{x}_i)-y_i\right)\phi'(\langle\vct{w}_q,\vct{x}_i\rangle)\vct{x}_i=\frac{v_q}{n}\sum_{i=1}^nr_i\phi'(\langle\vct{w}_q,\vct{x}_i\rangle)\vct{x}_i. \end{align} Aggregating these gradients as a row vector and setting $\mtx{D}_{\vct{v}}={\rm diag}(v_1,\dotsc,v_k)$ we arrive at \begin{align} \label{gradmat} \nabla_{\mtx{W}} \mathcal{L}(\vct{v},\mtx{W})=\mtx{D}_{\vct{v}}\left(\frac{1}{n}\sum_{i=1}^nr_i\phi'(\mtx{W}\vct{x}_i)\vct{x}_i^T\right). \end{align} We also define the Jacobian matrix $\mtx{J}=\begin{bmatrix}\mtx{J}_1 & \mtx{J}_2 & \ldots & \mtx{J}_n\end{bmatrix}\in\mathbb{R}^{kd\times n}$ with \begin{align*} \mtx{J}_i=\begin{bmatrix} v_1 \phi'(\mtx{w}_1^T\vct{x}_i) \vct{x}_i \\ v_2 \phi'(\mtx{w}_2\vct{x}_i) \vct{x}_i \\ \vdots\\ v_k \phi'(\mtx{w}_k^T\vct{x}_i) \vct{x}_i \end{bmatrix}. \end{align*} Let $\mtx{X}\in\mathbb{R}^{d\times n}$ be the data matrix with the $i$th column equal to $\vct{x}_i$. Using the Khatri-Rao product we can rewrite the Jacobian matrix in the form \begin{align} \label{myJ} \mtx{J}=\mtx{D}_{\vct{v}}\phi'(\mtx{W}\mtx{X})\ast\mtx{X}. \end{align} Using this Jacobian matrix we can write the vectorized version of the gradient, i.e.~gradient with respect to vect$(\mtx{W})$ as \begin{align} \label{gradvec} \nabla_{\text{vect}(\mtx{W})} \mathcal{L}(\vct{v},\mtx{W})=\frac{1}{n} \mtx{J} \vct{r}. \end{align} Taking the derivative with respect to $v_q$ we arrive at \begin{align*} \frac{\partial}{\partial v_q}\mathcal{L}(\vct{v},\mtx{W})=\frac{1}{n}\sum_{i=1}^n\left(\vct{v}^T\phi(\mtx{W}\vct{x}_i)-y_i\right)\phi(\langle \vct{w}_q,\vct{x}_i\rangle)=\frac{1}{n}\sum_{i=1}^nr_i\phi(\langle \vct{w}_q,\vct{x}_i\rangle). \end{align*} Thus the gradient with respect to $\vct{v}$ is equal to \begin{align} \label{gradv} \nabla_{\vct{v}}\mathcal{L}(\vct{v},\mtx{W})=\frac{1}{n}\phi(\mtx{W}\mtx{X})\vct{r}. \end{align} \subsubsection{Second order derivatives} Using \eqref{gradw} we have \begin{align*} \frac{\partial^2}{\vct{w}_p^2} \mathcal{L}(\vct{v},\mtx{W})=\frac{v_p}{n}\sum_{i=1}^n\left(\vct{v}^T\phi(\mtx{W}\vct{x}_i)-y_i\right)\phi''(\langle\vct{w}_p,\vct{x}_i\rangle)\vct{x}_i\vct{x}_i^T+\frac{v_p^2}{n}\sum_{i=1}^n\left(\phi'(\langle\vct{w}_p,\vct{x}_i\rangle)\right)^2\vct{x}_i\vct{x}_i^T. \end{align*} Also using \eqref{gradw}, for $p\neq q$ \begin{align*} \frac{\partial^2}{\partial\vct{w}_p\vct{w}_q} \mathcal{L}(\vct{v},\mtx{W})=\frac{v_pv_q}{n}\sum_{i=1}^n\phi'(\langle\vct{w}_p,\vct{x}_i\rangle)\phi'(\langle\vct{w}_q,\vct{x}_i\rangle)\vct{x}_i\vct{x}_i^T. \end{align*} Thus \begin{align} \label{partHess} \left(\text{vect}(\mtx{U})\right)^T&\nabla_{\mtx{W}}^2\mathcal{L}(\vct{v},\mtx{W}) \left(\text{vect}(\mtx{U})\right)\nonumber\\&=\frac{1}{n}\sum_{i=1}^n\left(\vct{v}^T\phi(\mtx{W}\vct{x}_i)-y_i\right)\left(\sum_{p=1}^kv_p\phi''(\langle\vct{w}_p,\vct{x}_i\rangle)(\langle\vct{u}_p,\vct{x}_i\rangle)^2\right)\nonumber\\ &+\frac{1}{n}\sum_{i=1}^n \left(\sum_{p=1}^kv_p\phi'(\langle\vct{w}_q,\vct{x}_i\rangle)\langle\vct{u}_p,\vct{x}_i\rangle\right)^2\nonumber\\ =&\frac{1}{n}\sum_{i=1}^nr_i\left(\vct{x}_i^T\mtx{U}^T\mtx{D}_{\vct{v}}\mtx{D}_{\phi''(\mtx{W}\vct{x}_i)}\mtx{U}\vct{x}_i\right)+\frac{1}{n}\sum_{i=1}^n \left( \vct{v}^T\mtx{D}_{\phi'(\mtx{W}\vct{x}_i)}\mtx{U}\vct{x}_i\right)^2. \end{align} Using the expression for the Jacobian matrix the latter can also be written in the form \begin{align} \label{partHess2} \left(\text{vect}(\mtx{U})\right)^T&\nabla_{\mtx{W}}^2\mathcal{L}(\vct{v},\mtx{W}) \left(\text{vect}(\mtx{U})\right)\nonumber\\ &=\frac{1}{n}\sum_{i=1}^n\langle \mtx{U}^T\mtx{D}_{\vct{v}}\mtx{D}_{\phi''(\mtx{W}\vct{x}_i)}\mtx{U}, r_i\vct{x}_i\vct{x}_i^T\rangle+\frac{1}{n}\text{vect}(\mtx{U})^T\mtx{J}\mtx{J}^T\text{vect}(\mtx{U}). \end{align} Finally, note that \begin{align} \label{Hessianv} \nabla_{\vct{v}}^2\mathcal{L}(\vct{v},\mtx{W})=\phi\left(\mtx{W}\mtx{X}\right) \phi\left(\mtx{W}\mtx{X}\right)^T. \end{align} \subsection{Useful identities involving matrix products} In this section we gather a few preliminary results regarding matrix products that will be useful throughout our proofs. Most of these identities are trivial and we thus skip their proof. The only exception is Lemma \ref{Pre4} which is proved in Appendix \ref{pfprem}. \begin{lemma}\label{Pre1} For two vectors $\mtx{u}\in\mathbb{R}^{m\times 1}$ and ${\mtx{v}}\in \mathbb{R}^{n\times 1}$, and a matrix $\mtx{M}\in\mathbb{R}^{k\times n}$, we have \[\mtx{u}\otimes \mtx{M}{\mtx{v}} = \mtx{D_M} (\mtx{u}\otimes {\mtx{v}})\,,\] where $\mtx{D_M}\in \mathbb{R}^{mk\times mn}$ is the block diagonal matrix with $m$ copies of $\mtx{M}$ in the diagonal blocks. \end{lemma} \begin{lemma}\label{Pre2} For vectors $\vct{a}\in\mathbb{R}^{m\times 1}$ and $\vct{b}\in \mathbb{R}^{n\times 1}$, and a matrix $\mtx{M}\in\mathbb{R}^{n\times m}$, we have \[(\vct{a}\otimes \vct{b})^T {\rm vec}(\mtx{M}) = \vct{b}^T \mtx{M} \vct{a}\,,\] where ${\rm vec}(M)$ denotes the vectorization of the matrix $\mtx{M}$ formed by stacking the columns of $\mtx{M}$ into a single column vector. \end{lemma} \begin{lemma}\label{Pre4} Consider an arbitrary matrix $\mtx{A}\in \mathbb{R}^{k\times d}$ and an arbitrary vector ${\mtx{v}} \in \mathbb{R}^k$ and define $\mtx{D_v}:={\rm diag}({\mtx{v}})$. Then the following identity holds \begin{align} \|{\mtx{v}}\|_\infty \sigma_{\min}(\mtx{A}) \le \sigma_{\max}(\mtx{D_v}\mtx{A})\,. \end{align} \end{lemma} \begin{lemma}\label{Pre5} For any two matrices $\mtx{A}$ and $\mtx{B}$, we have \begin{align} (\mtx{A}\ast \mtx{B})^T (\mtx{A}\ast \mtx{B}) = (\mtx{A}^T \mtx{A}) \circ (\mtx{B}^T\mtx{B})\,. \end{align} \end{lemma} \subsection{Useful probabilistic identities involving random vectors and matrices} In this Section we gather some useful probabilistic identities that will be used throughout our proofs. We defer the proof of these results to Appendix \ref{pfprem}. The first result, proven in Appendix \ref{prfPre3}, relates the tail of a random variable to a proper Orlicz norm. \begin{lemma}\label{Pre3} Suppose that a random variable $Y$ satisfies the following tail bound \begin{align} \P(|Y|\ge t) \le 2e^{-c\min(t^2/A, t/B)}\,. \end{align} Then, $\|Y\|_{\psi_1} \le 9\max(\sqrt{A/(2c)},B/c)$. \end{lemma} The second result concerns the minimum eigenvalue of the Khatrio-Rao product of a generic matrix with itself. \begin{lemma}\label{XkrX} For almost every data input matrix $\mtx{X}\in \mathbb{R}^{d\times n}$, as long as $d\le n\le d(d+1)/2$, the following holds \begin{align*} \sigma_{\min}(\mtx{X}\ast\mtx{X})>0. \end{align*} \end{lemma} We refer to Appendix~\ref{app:XkrX} for the proof of Lemma~\ref{XkrX}. Finally, we state a Lemma based on \cite{WF, soltanolkotabi2014algorithms, soltanolkotabi2017structured} that allows us to lower bound the loss function $\mathcal{L}(\vct{v},\mtx{W})$ in terms of how close $(\vct{v},\mtx{W})$ is to $(\vct{v}^*,\mtx{W}^*)$. We prove this Lemma in Appendix \ref{PRlemmapf}. \begin{lemma}\label{PRlemma} Let $\mtx{A}\in\mathbb{R}^{d\times d}$ be a symmetric positive semidefinite matrix. Also for $i=1,2,\ldots,n$, let $\vct{x}_i$ be i.i.d.~random vectors distributed as $\mathcal{N}(\vct{0},\mtx{I}_d)$. Furthermore, assume \begin{align*} n\ge c(\delta)d \log d, \end{align*} with $c(\delta)$ a constant only depending on $\delta$. Then for all matrices $\mtx{M}\in\mathbb{R}^{d\times d}$ we have \begin{align*} \nucnorm{\frac{1}{n}\sum_{i=1}^n \left(\vct{x}_i^T\mtx{A}\vct{x}_i\right)\vct{x}_i\vct{x}_i^T-\left(2\mtx{A}+\emph{trace}(\mtx{A})\mtx{I}\right)}\le\delta\cdot\nucnorm{\mtx{A}}, \end{align*} holds with probability at least $1-10de^{-\gamma d}-8/d$. Here, $\nucnorm{\cdot}$ denotes the nuclear norm i.e.~sum of singular values of the input matrix. \end{lemma} \section{Proof of global landscape results} \label{proofs} In this section we prove the two global theorems. \subsection{Derivative calculations for quadratic activations} We begin by gathering the derivatives of the loss function for quadratic activations. We note that for quadratic activations the loss function \eqref{landscape} as a function of $\mtx{W}$ takes the form \begin{align*} \mathcal{L}(\mtx{W})=&\frac{1}{2n}\sum_{i=1}^n\left(y_i-\sum_{\ell=1}^k\vct{v}_\ell\abs{\langle\mtx{w}_\ell,\vct{x}_i\rangle}^2\right)^2,\\ =&\frac{1}{2n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T{\rm diag}(\vct{v})\mtx{W}\vct{x}_i-y_i\right)^2,\\ :=&\frac{1}{2n}\sum_{i=1}^nr_i^2. \end{align*} Using \eqref{gradmat} we have \begin{align} \label{gradmat2} \nabla_{\mtx{W}} \mathcal{L}(\vct{v},\mtx{W})=2\mtx{D}_{\vct{v}}\mtx{W}\left(\frac{1}{n}\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T\right). \end{align} Using \eqref{myJ} the Jacobian matrix is given by \begin{align} \label{myJ2} \mtx{J}=2\mtx{D}_{\vct{v}}(\mtx{W}\mtx{X})\ast\mtx{X}. \end{align} Using this Jacobian matrix we can write the vectorized version of the gradient, i.e.~gradient with respect to vect$(\mtx{W})$ as \begin{align} \label{gradvec2} \nabla_{\text{vect}(\mtx{W})} \mathcal{L}(\vct{v},\mtx{W})=\frac{1}{n} \mtx{J} \vct{r}. \end{align} Also \eqref{partHess2} specialized to quadratic activations results in the partial Hessian \begin{align} \label{partHess3} \left(\text{vect}(\mtx{U})\right)^T\nabla_{\mtx{W}}^2\mathcal{L}(\vct{v},\mtx{W}) \left(\text{vect}(\mtx{U})\right)=\frac{2}{n}\sum_{i=1}^nr_i\left(\vct{x}_i^T\mtx{U}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)+\frac{2}{n}\sum_{i=1}^n \left( \vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)^2. \end{align} \subsection{Proof of Theorem \ref{globthm}} We begin by proving a simple lemma. \begin{lemma}\label{globalMlem} Any point $\widetilde{\mtx{W}}\in\mathbb{R}^{k\times d}$ obeying \begin{align} \label{gradcondW} \frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\widetilde{\mtx{W}}^T{\rm diag}(\vct{v})\widetilde{\mtx{W}}\vct{x}_i-y_i\right)\vct{x}_i\vct{x}_i^T=0, \end{align} is a global optimum of the loss function \begin{align*} \mathcal{L}(\mtx{W})=\frac{1}{2n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T{\rm diag}(\vct{v})\mtx{W}\vct{x}_i-y_i\right)^2. \end{align*} \end{lemma} \begin{proof} Consider the optimization problem \begin{align*} \underset{\mtx{M}\in\mathbb{R}^{d\times d}}{\min}\text{ }f(\mtx{M}):=\frac{1}{2n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{M}\vct{x}_i-y_i\right)^2. \end{align*} The objective value is convex in terms of $\mtx{M}$ (in fact its a quadratic). Therefore, at any point where the gradient with respect to $\widetilde{\mtx{M}}$ is zero i.e.~$\nabla f(\widetilde{\mtx{M}})=0$, that point must be a global optimum of $f$. More specifically, any point $\widetilde{\mtx{M}}$ obeying \begin{align} \label{gradM} \frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\widetilde{\mtx{M}}\vct{x}_i-y_i\right)\vct{x}_i\vct{x}_i^T=0, \end{align} is a global optimum of $f(\mtx{M})$. That is, for any arbitrary matrix $\mtx{M}\in\mathbb{R}^{d\times d}$ and any point $\widetilde{\mtx{M}}\in\mathbb{R}^{d\times d}$ obeying \eqref{gradM} we have \begin{align} \label{globM} f(\mtx{M})\ge f(\widetilde{\mtx{M}}). \end{align} Now note that if \eqref{gradcondW} holds for some $\widetilde{\mtx{W}}$, then \eqref{gradM} holds with $\widetilde{\mtx{M}}=\widetilde{\mtx{W}}^T$diag($\vct{v})\widetilde{\mtx{W}}$. Thus for any $\mtx{W}\in\mathbb{R}^{k\times d}$, using \eqref{globM} with $\mtx{M}=\mtx{W}^T$diag$(\vct{v})\mtx{W}$ we have \begin{align*} \mathcal{L}(\mtx{W})=f\left(\mtx{W}^T{\rm diag}(\vct{v})\mtx{W}\right)\ge f(\widetilde{\mtx{M}})=\mathcal{L}(\widetilde{\mtx{W}}). \end{align*} Thus any $\widetilde{\mtx{W}}$ obeying \eqref{gradcondW} is a global optima of $\mathcal{L}(\mtx{W})$, concluding the proof. \end{proof} \subsubsection{Proof of no spurious local minima and strict saddle property} We will prove the first two conclusions of the theorem (i.e.~no spurious local optima and saddles have a direction of negative curvature) simultaneously. To show this note that since the function $\mathcal{L}$ is twice differentiable all local optima or saddles that do not have a direction of negative curvature obey \begin{align} \label{secord} \nabla \mathcal{L}(\mtx{W})=\mtx{0}\quad\text{and}\quad \nabla^2\mathcal{L}(\mtx{W})\succeq \mtx{0}. \end{align} We will prove that all points obeying \eqref{secord} are a global optima of $\mathcal{L}(\mtx{W})$. To this aim let us define the sets \begin{align*} \mathcal{S}_{+}=\{i: \vct{v}_i>0\}\quad\text{and}\quad\mathcal{S}_{-}=\{i: \vct{v}_i<0\}, \end{align*} i.e.~the indices of the positive and negative entries of $\vct{v}$. Now note that two cases are possible \begin{itemize} \item \textbf{Case I:} At least one of the sub-matrices $\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}_{+}}$ and $\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}_{-}}$ is not rank deficient. That is, \begin{align} \label{fullrank} \text{rank}\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}_{+}}=d\quad\text{or}\quad \text{rank}\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}_{-}}=d. \end{align} \item \textbf{Case II:} Both of the sub-matrices $\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}_{+}}$ and $\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}_{-}}$ are rank deficient. That is, \begin{align} \label{cond2} \text{rank}\left(\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}_{+}}\right)<d\quad\text{or}\quad \text{rank}\left(\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}_{-}}\right)<d. \end{align} \end{itemize} In both cases we will show that for any $\mtx{W}$ obeying \eqref{secord}, we have \begin{align*} \sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T=\frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T{\rm diag}(\vct{v})\mtx{W}\vct{x}_i-y_i\right)\vct{x}_i\vct{x}_i^T=0. \end{align*} The latter together with Lemma \ref{globalMlem} immediately implies that any $\mtx{W}$ obeying \eqref{secord} is a global optimum, proving that there are no spurious local minima and no saddles which do not have a direction of strict negative curvature. We now proceed to show that indeed $\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T=0$ holds in both cases mentioned above. \noindent\textbf{Case I: } Note that \eqref{fullrank} together with the fact that $k\ge d$ implies that the matrix $\mtx{D}_{\vct{v}}\mtx{W}$ has a left inverse i.e.~there exists a matrix $\mtx{M}\in\mathbb{R}^{d\times k}$ such that $\mtx{M}\mtx{D}_{\vct{v}}\mtx{W}=\mtx{I}$. Furthermore, note that by \eqref{gradmat2}, $\nabla \mathcal{L}(\mtx{W})=\mtx{0}$ is equivalent to \begin{align*} \mtx{D}_{\vct{v}}\mtx{W}\left(\frac{1}{n}\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T\right)=0. \end{align*} Multiplying both sides of the above equality on the left by $\mtx{M}$ we conclude that \begin{align*} \frac{1}{n}\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T=0, \end{align*} concluding the proof in case I. \noindent\textbf{Case II: } First, note that for any $\mtx{W}$ obeying $\nabla^2\mathcal{L}(\mtx{W})\succeq \mtx{0}$ and any $\mtx{U}\in\mathbb{R}^{k\times d}$ by \eqref{partHess2} we have \begin{align} \label{69} 0\le&\frac{1}{2} \left(\text{vect}(\mtx{U})\right)^T\nabla_{\mtx{W}}^2\mathcal{L}(\vct{v},\mtx{W}) \left(\text{vect}(\mtx{U})\right),\nonumber\\ =&\frac{1}{n}\sum_{i=1}^nr_i\left(\vct{x}_i^T\mtx{U}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)+\frac{1}{n}\sum_{i=1}^n \left( \vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)^2. \end{align} Let us choose $\mtx{U}$ of the form $\mtx{U}=\vct{a}\vct{b}^T$ with $\vct{a}\in\mathbb{R}^k$ obeying $\mtx{W}^T\mtx{D}_{\vct{v}}\vct{a}=0$ (i.e.~$\mtx{D}_{\vct{v}}\vct{a}\in$Null($\mtx{W}^T$)) and $\vct{b}\in\mathbb{R}^d$ an arbitrary vector. Plugging such a $\mtx{U}$ into \eqref{69} we conclude that \begin{align} \label{ineq} 0\le&\frac{1}{n}\sum_{i=1}^nr_i\left(\vct{x}_i^T\mtx{U}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)+\frac{1}{n}\sum_{i=1}^n \left( \vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)^2,\nonumber\\ =&(\vct{a}^T\mtx{D}_{\vct{v}}\vct{a})\vct{b}^T\left(\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T\right)\vct{b}+\frac{1}{n}\sum_{i=1}^n \left( \vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\vct{a}\vct{b}^T\vct{x}_i\right)^2,\nonumber\\ =&\left(\vct{a}^T\mtx{D}_{\vct{v}}\vct{a}\right)\vct{b}^T\left(\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T\right)\vct{b}, \end{align} holds for all $\vct{b}\in\mathbb{R}^d$ and all $\vct{a}\in\mathbb{R}^k$ obeying $\mtx{W}^T\mtx{D}_{\vct{v}}\vct{a}=0$. Next, we note that by the assumptions of Theorem \ref{globthm} $\abs{\mathcal{S}_{+}}\ge d$ and $\abs{\mathcal{S}_{+}}\ge d$. This together with \eqref{cond2} implies that there exists non-zero vectors $\vct{u},\vct{w}\in\mathbb{R}^k$ with non-zero entries belonging to $\mathcal{S}_{+}$ and $\mathcal{S}_{-}$ such that \begin{align*} \mtx{W}^T\mtx{D}_{\vct{v}}\vct{u}=\mtx{W}^T\mtx{D}_{\vct{v}}\vct{w}=0. \end{align*} Furthermore, \begin{align*} \vct{u}^T\mtx{D}_{\vct{v}}\vct{u}=\sum_{i\in\mathcal{S}_{+}} v_iu_i^2>0. \end{align*} Thus using \eqref{ineq} with $\vct{a}=\vct{u}$ we conclude that for all $\vct{b}\in\mathbb{R}^d$ we have \begin{align} \label{myeq1} \vct{b}^T\left(\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T\right)\vct{b}\ge 0. \end{align} Similarly, \begin{align*} \vct{w}^T\mtx{D}_{\vct{v}}\vct{w}=\sum_{i\in\mathcal{S}_{-}} v_iw_i^2< 0. \end{align*} Thus using \eqref{ineq} with $\vct{a}=\vct{w}$ we conclude that for all $\vct{b}\in\mathbb{R}^d$ we have \begin{align} \label{myeq2} \vct{b}^T\left(\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T\right)\vct{b}\le 0. \end{align} Equations \eqref{myeq1} and \eqref{myeq2} together imply that \begin{align*} \sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T=0, \end{align*} concluding the proof in case II. \subsubsection{Proof of zero training error} Note that in the previous section we proved that all global optima of $\mathcal{L}(\mtx{W})$ obey \begin{align*} \sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T=0. \end{align*} The latter identity can be rewritten in the form \begin{align} \label{tempglob} \left(\mtx{X}\ast\mtx{X}\right)\vct{r}=0. \end{align} Using Lemma \ref{XkrX}, $\sigma_{\min}(\mtx{X}\ast\mtx{X})>0$ for almost every data input matrix $\mtx{X}\in \mathbb{R}^{d\times n}$, as long as $n\le c d^2$ with $c$ a fixed numerical constant. This combined with \eqref{tempglob} implies that $\vct{r}=\vct{0}$, completing the proof for zero training error. \subsection{Proof of Theorem \ref{gthmrand}} First note that since we assume both $\vct{v}$ and $\vct{v}^*$ have the same sign pattern we can without loss of generality assume both $\vct{v}$ and $\vct{v}^*$ have non-negative entries. We will first prove that there are no spurious local minima, and all saddles have a direction of negative curvature, and all global optima have zero loss value. We then proceed to show all approximate local minima have small objective value in Section \ref{pfneg}. \subsubsection{Proof of no spurious local minima, zero training error, and strict saddle} We begin by noting that the function $\mathcal{L}$ is twice differentiable and thus all local optima or saddle points that do not have a direction of negative curvature obey \begin{align} \label{secordp} \nabla \mathcal{L}(\mtx{W})=\mtx{0}\quad\text{and}\quad \nabla^2\mathcal{L}(\mtx{W})\succeq\mtx{0}. \end{align} Note that $\nabla \mathcal{L}(\mtx{W})=\mtx{0}$ implies that \begin{align} \label{tmppp2} \mtx{D}_{\vct{v}}\mtx{W}\left(\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\right)=\mtx{0}. \end{align} Define the set \begin{align*} \mathcal{S}=\{i: v_i>0\}. \end{align*} Since we assume $k\ge d$, two cases can occur. \begin{itemize} \item \textbf{Case I:} rank($\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}})=d$. \item \textbf{Case II:} rank($\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}})<d$. \end{itemize} In both cases we will show that any $\mtx{W}$ obeying \eqref{secord}, must be a global optimum. \noindent\textbf{Case I: }rank($\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}})=d$.\\ In this case the matrix $\mtx{D}_{\vct{v}}\mtx{W}$ has a left inverse, i.e.~there exists a $\mtx{M}\in\mathbb{R}^{d\times k}$ such that $\mtx{M}\mtx{D}_{\vct{v}}\mtx{W}=\mtx{I}$. Multiplying both sides of \eqref{tmppp2} on the left by $\mtx{M}$ yields \begin{align} \label{tmpthm22} \sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T=\left(\mtx{X}\ast\mtx{X}\right)\vct{r}=\frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T{\rm diag}(\vct{v})\mtx{W}\vct{x}_i-y_i\right)\vct{x}_i\vct{x}_i^T=0. \end{align} The latter together with Lemma \ref{globalMlem} immediately implies that any $\mtx{W}$ obeying \eqref{secordp} and the condition of Case I is a global optimum, proving that there are no spurious local minima. Furthermore using Lemma \ref{XkrX}, $\sigma_{\min}(\mtx{X}\ast\mtx{X})>0$ as long as $n\le c d^2$ with $c$ a fixed numerical constant. This combined with \eqref{tmpthm22} implies that $\vct{r}=\vct{0}$, completing the proof for zero training error in this case. \noindent\textbf{Case II: }rank($\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}})<d$.\\ First, note that for any $\mtx{W}$ obeying $\nabla^2\mathcal{L}(\mtx{W})\succeq \mtx{0}$ and any $\mtx{U}\in\mathbb{R}^{k\times d}$ by \eqref{partHess2} we have \begin{align} \label{692} 0\le&\frac{1}{2} \left(\text{vect}(\mtx{U})\right)^T\nabla_{\mtx{W}}^2\mathcal{L}(\vct{v},\mtx{W}) \left(\text{vect}(\mtx{U})\right),\nonumber\\ =&\frac{1}{n}\sum_{i=1}^nr_i\left(\vct{x}_i^T\mtx{U}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)+\frac{1}{n}\sum_{i=1}^n \left( \vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)^2. \end{align} In this case since $\left(\mathcal{D}_{\vct{v}}\right)_{\mathcal{S}}$ is invertible and $\left(\mtx{D}_{\vct{v}}\mtx{W}\right)_{\mathcal{S}}$ is rank deficient there exists a nonzero vector $\vct{a}\in\mathbb{R}^k$ supported on $\mathcal{S}$ obeying $\mtx{W}^T\mtx{D}_{\vct{v}}\vct{a}=0$ (i.e.~$\mtx{D}_{\vct{v}}\vct{a}\in$Null($\mtx{W}^T$)). Thus plugging $\mtx{U}$ of the form $\mtx{U}=\vct{a}\vct{b}^T$ with $\vct{b}\in\mathbb{R}^d$ in \eqref{692} we conclude that \begin{align} \label{ineqthm22} 0\le&\frac{1}{n}\sum_{i=1}^nr_i\left(\vct{x}_i^T\mtx{U}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)+\frac{1}{n}\sum_{i=1}^n \left( \vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)^2,\nonumber\\ =&(\vct{a}^T\mtx{D}_{\vct{v}}\vct{a})\vct{b}^T\left(\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T\right)\vct{b}+\frac{1}{n}\sum_{i=1}^n \left( \vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\vct{a}\vct{b}^T\vct{x}_i\right)^2,\nonumber\\ =&\left(\vct{a}^T\mtx{D}_{\vct{v}}\vct{a}\right)\vct{b}^T\left(\sum_{i=1}^nr_i\vct{x}_i\vct{x}_i^T\right)\vct{b}, \end{align} holds for all $\vct{b}\in\mathbb{R}^d$. Also note that \begin{align*} \vct{a}^T\mtx{D}_{\vct{v}}\vct{a}=\sum_{i\in\mathcal{S}} v_ia_i^2>0. \end{align*} The latter together with \eqref{ineqthm22} implies that \begin{align} \label{rxxpos} \sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\succeq \mtx{0}. \end{align} Thus for any $\mtx{W}$ obeying \eqref{secordp} and the condition of Case II we have \begin{align*} \mathcal{L}(\mtx{W})&=\frac{1}{2n}\langle \mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}-{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*,\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\rangle,\\ &\overset{(a)}{=}-\frac{1}{2n}\langle{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*,\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\rangle,\\ &\overset{(b)}{\le} 0. \end{align*} where (a) follows from \eqref{tmppp2} and (b) follows from \eqref{rxxpos} combined with the fact that $\vct{v}^*$ has nonnegative entries. Thus, $\mathcal{L}(\mtx{W})=0$ and therefore such a $\mtx{W}$ must be a global optimum. This completes the proof in Case II. \subsubsection{Bounding the objective value for approximate local minima} \label{pfneg} Assume $\mtx{W}$ is an approximate local minima. That is it obeys \begin{align} \label{secordp3} \fronorm{\nabla \mathcal{L}(\mtx{W})}\le\epsilon_g\quad\text{and}\quad \nabla^2\mathcal{L}(\mtx{W}_s)\succeq -\epsilon_{H}\mtx{I}. \end{align} For such a point by Cauchy-Schwarz we have \begin{align} \label{trivial} \langle\mtx{W},\nabla \mathcal{L}(\mtx{W})\rangle\le \epsilon_g\fronorm{\mtx{W}}. \end{align} Furthermore, for such a point using \eqref{partHess2} \begin{align} \label{6923} -\epsilon_{H}\fronorm{\mtx{U}}^2\le&\frac{1}{2} \left(\text{vect}(\mtx{U})\right)^T\nabla_{\mtx{W}}^2\mathcal{L}(\vct{v},\mtx{W}) \left(\text{vect}(\mtx{U})\right),\nonumber\\ =&\frac{1}{n}\sum_{i=1}^nr_i\left(\vct{x}_i^T\mtx{U}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)+\frac{1}{n}\sum_{i=1}^n \left( \vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{U}\vct{x}_i\right)^2, \end{align} holds for all $\mtx{U}\in\mathbb{R}^{k\times d}$. Since $\vct{v}=\nu\vct{1}$ there exists a unit norm, non-zero vector $\vct{a}\in\mathbb{R}^k$ supported on $\mathcal{S}$ such that $\mtx{W}^T\mtx{D}_{\vct{v}}\vct{a}=0$. We now set $\mtx{U}=\vct{a}\vct{b}^T$ in \eqref{6923} with $\vct{b}\in\mathbb{R}^d$. We conclude that \begin{align} \label{tmpsecthm} \left(\vct{a}^T\mtx{D}_{\vct{v}}\vct{a}\right)\left(\vct{b}^T\left(\frac{1}{n}\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\right)\vct{b}\right)\ge -\epsilon_{H}\twonorm{\vct{b}}^2. \end{align} holds for all $\vct{b}\in\mathbb{R}^d$. Also note that \begin{align*} \vct{a}^T\mtx{D}_{\vct{v}}\vct{a}=\nu\twonorm{\vct{a}}^2=\nu. \end{align*} Thus, \eqref{tmpsecthm} implies that \begin{align} \label{condrxx} \frac{1}{n}\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\succeq -\frac{\epsilon_{H}}{\nu}\mtx{I}. \end{align} Now note that the gradient is equal to \begin{align} \label{gradthm22} \nabla\mathcal{L}(\mtx{W})=\mtx{D}_{\vct{v}}\mtx{W}\left(\sum_{i=1}^n r_i\vct{x}_i\vct{x}._i^T\right). \end{align} Thus using \eqref{gradthm22}, \eqref{trivial}, and \eqref{condrxx} we conclude that for any point $\mtx{W}$ obeying \eqref{secordp3} we have \begin{align} \label{tmpbndLW} \mathcal{L}(\mtx{W})&=\frac{1}{2n}\langle \mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}-{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*,\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\rangle,\nonumber\\ &\overset{(a)}{=}\langle\mtx{W},\nabla\mathcal{L}(\mtx{W})\rangle-\frac{1}{2n}\langle{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*,\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\rangle,\nonumber\\ &\overset{(b)}{\le} \epsilon_g\fronorm{\mtx{W}}-\frac{1}{2}\langle{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*,\frac{1}{n}\sum_{i=1}^n r_i\vct{x}_i\vct{x}_i^T\rangle,\nonumber\\ &\overset{(c)}{\le}\epsilon_g\fronorm{\mtx{W}}+\frac{\epsilon_{H}}{2\nu}\nucnorm{{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*}. \end{align} Here, (a) follows from \eqref{gradthm22}, (b) from \eqref{trivial} and (c) from Holder's inequality combined with \eqref{condrxx}. We proceed by showing that for a point $\mtx{W}$ obeying $\fronorm{\nabla \mathcal{L}(\mtx{W})}\le\epsilon_g$ it's Frobenius norm norm is also sufficiently small. To this aim note that for such a point by Cauchy-Schwarz have \begin{align} \label{simple} \langle\nabla \mathcal{L}(\mtx{W}),\mtx{W}\rangle\le \Delta\fronorm{\mtx{W}}. \end{align} Also \begin{align} \label{myineqgg} \langle\nabla \mathcal{L}(\mtx{W}),\mtx{W}\rangle=&\frac{1}{n}\langle\mtx{D}_{\vct{v}}\mtx{W}\sum_i \left(\vct{x}_i^T\left(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}-{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\vct{x}_i\right)\vct{x}_i\vct{x}_i^T,\mtx{W}\rangle\nonumber\\ =&\frac{1}{n}\sum_{i=1}^n \left(\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i\right)^2-\frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i\right)\left(\vct{x}_i^T{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\vct{x}_i\right)\nonumber\\ \ge&\left(\frac{1}{n}\sum_{i=1}^n\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i\right)^2-\frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i\right)\left(\vct{x}_i^T{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\vct{x}_i\right). \end{align} Using Holder's inequality for matrices we have \begin{align*} \frac{1}{n}\sum_{i=1}^n\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i-\text{trace}\left(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\right)=&\langle \mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W},\frac{1}{n}\sum_{i=1}^n\vct{x}_i\vct{x}_i^T-\mtx{I}\rangle\\ \le&\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W})\opnorm{\frac{1}{n}\sum_{i=1}^n\vct{x}_i\vct{x}_i^T-\mtx{I}}\\ =&\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W})\opnorm{\frac{1}{n}\sum_{i=1}^n\vct{x}_i\vct{x}_i^T-\mtx{I}}. \end{align*} Thus by standard concentration of sample covariance matrices we conclude that for any $\delta>0$, as long as $n\ge \frac{c}{\delta^2} d$ for a fixed numerical constant $c$, then for all $\mtx{W}$ \begin{align*} \abs{\frac{1}{n}\sum_{i=1}^n\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i-\text{trace}\left(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\right)}\le \delta\cdot\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}), \end{align*} holds with probability at least $1-2e^{-b d}$, for a fixed constant $b > 0$. Thus with high probability for all $\mtx{W}$ we have \begin{align*} \frac{1}{n}\sum_{i=1}^n\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i\ge (1-\delta)\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}). \end{align*} Plugging the latter into \eqref{myineqgg} we conclude that \begin{align} \label{myineq2gg} \langle\nabla \mathcal{L}(\mtx{W}),\mtx{W} \rangle \ge (1-\delta)^2\left(\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W})\right)^2-\frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i\right)\left(\vct{x}_i^T{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\vct{x}_i\right). \end{align} We now apply Lemma \ref{PRlemma} to conclude that as long as $n\ge c(\delta) d\log d$ \begin{multline*} \opnorm{\frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\vct{x}_i\right)\vct{x}_i\vct{x}_i^T-\left(2{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*+\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\mtx{I}\right)}\\\le\delta\cdot\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right), \end{multline*} holds with probability at least $1-10de^{-b d}-8/d$. The latter implies that with high probability for all $\mtx{W}$ we have \begin{align*} \Bigg|\frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i\right)&\left(\vct{x}_i^T{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\vct{x}_i\right)\\ &-\left(2\langle{\mtx{W}}^T\mtx{D}_{\vct{v}}\mtx{W} ,{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^* \rangle+\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\text{trace}\left({\mtx{W}}^T\mtx{D}_{\vct{v}}\mtx{W}\right)\right)\bigg|\\ \le& \delta\cdot\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\text{trace}\left({\mtx{W}}^T\mtx{D}_{\vct{v}}\mtx{W}\right). \end{align*} Plugging the latter into \eqref{myineq2gg} and we conclude that with high probability for all $\mtx{W}$ we have \begin{align*} \langle\nabla \mathcal{L}(\mtx{W}),\mtx{W} \rangle \ge& (1-\delta)^2\left(\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W})\right)^2-\frac{1}{n}\sum_{i=1}^n\left(\vct{x}_i^T\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W}\vct{x}_i\right)\left(\vct{x}_i^T{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\vct{x}_i\right).\\ \ge&(1-\delta)^2\left(\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W})\right)^2-\delta\cdot\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\text{trace}\left({\mtx{W}}^T\mtx{D}_{\vct{v}}\mtx{W}\right)\\ &-\left(2\langle{\mtx{W}}^T\mtx{D}_{\vct{v}}\mtx{W} ,{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^* \rangle+\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\text{trace}\left({\mtx{W}}^T\mtx{D}_{\vct{v}}\mtx{W}\right)\right)\\ \ge&(1-\delta)^2\left(\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W})\right)^2-(3+\delta)\cdot\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\text{trace}\left({\mtx{W}}^T\mtx{D}_{\vct{v}}\mtx{W}\right). \end{align*} Plugging the latter into \eqref{simple} implies that \begin{align*} (1-\delta)^2\left(\text{trace}(\mtx{W}^T\mtx{D}_{\vct{v}}\mtx{W})\right)^2-(3+\delta)\cdot\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\text{trace}\left({\mtx{W}}^T\mtx{D}_{\vct{v}}\mtx{W}\right)\le \epsilon_g\fronorm{\mtx{W}}. \end{align*} Using the assumption that $\vct{v}=\nu\vct{1}$, the latter is equivalent to \begin{align*} &\fronorm{\mtx{W}}\left((1-\delta)^2\nu^2\fronorm{\mtx{W}}^2-\nu(3+\delta)\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\right)\\ &=(1-\delta)^2\nu^2\fronorm{\mtx{W}}^3-(3+\delta)\nu\fronorm{\mtx{W}}\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)\le \epsilon_g. \end{align*} From the latter we can conclude that for all $a>0$ we have \begin{align*} \fronorm{\mtx{W}}\le& \max\left(\sqrt{\frac{\epsilon_g}{a\nu^2(1-\delta)^2}+\frac{3+\delta}{\nu(1-\delta)^2}\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)}, a\right) \end{align*} Setting $a=\frac{\epsilon_g}{\nu(1-\delta)^2}$ and $\delta=1/2$, the latter inequality implies that \begin{align*} \fronorm{\mtx{W}}\le \max\left(\frac{1}{\sqrt{\nu}}\sqrt{1+14\text{trace}\left({\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*\right)},\frac{4}{\nu}\epsilon_g\right). \end{align*} Plugging the latter into \eqref{tmpbndLW} we conclude that for all $\mtx{W}$ obeying \eqref{secordp3} \begin{align*} \mathcal{L}(\mtx{W})\le&\epsilon_g\fronorm{\mtx{W}}+\frac{\epsilon_{H}}{2\nu}\nucnorm{{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*},\\ \le&\frac{\epsilon_g}{\sqrt{\nu}}\max\left(\sqrt{1+14\nucnorm{{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*}},4\frac{\epsilon_g}{\sqrt{\nu}}\right)+\frac{\epsilon_{H}}{2\nu}\nucnorm{{\mtx{W}^*}^T\mtx{D}_{\vct{v}^*}\mtx{W}^*}, \end{align*} holds with high probability, concluding the proof. \section{Proof of local convergence results} In this section we prove Theorem \ref{localthm}. We begin by explaining a crucial component of our proof which involves bounding the spectrum of the Jacobian matrix. These results are the subject of the next section with the corresponding proofs appearing in Section \ref{proofJ}. We then utilize these intermediate results to finalize the proof of Theorem \ref{localthm} in Section \ref{pfthmloc}. \subsection{Key lemmas on spectrum of the Jacobian Matrix}\label{sec:key-lemma} We start with a few definitions. As defined previously in Assumption \ref{ass-phi} for $\sigma\in \mathbb{R}$ and $g\sim\mathcal{N}(0,\sigma^2)$ we define $\mu_\phi(\sigma):= \operatorname{\mathbb{E}}[\phi'(g)]$, and let \begin{align}\label{mub} \vct{\mu}=\left(\mu_\phi(\twonorm{\vct{w}_1}),\mu_\phi(\twonorm{\vct{w}_2}),\ldots,\mu_\phi(\twonorm{\vct{w}_k})\right)\,. \end{align} Similarly, \[\gamma_\phi(\sigma):= \frac{1}{\sigma^2} \operatorname{\mathbb{E}}[\phi'(g)g] \,, \] and note that by Stein's lemma we have \begin{align*} \gamma_\phi(\sigma)=\frac{1}{\sigma^2}\operatorname{\mathbb{E}}[\phi'(g)g]=\operatorname{\mathbb{E}}[\phi''(g)]\,. \end{align*} Also define \begin{align}\label{Gammab} \mtx{\Gamma}=\text{diag}\left(\gamma_\phi(\twonorm{\vct{w}_1}),\gamma_\phi(\twonorm{\vct{w}_2}),\ldots,\gamma_\phi(\twonorm{\vct{w}_k})\right)\, \end{align} Recall that the Jacobian matrix is given by $\mtx{J} = \mtx{D_v} \phi'(\mtx{W}\mtx{X}) \ast \mtx{X}$. To bound the spectrum of the Jacobian we first introduce a new matrix obtained by centering the $\phi'(\mtx{W}\mtx{X})$ component. Precisely, we let \begin{align}\label{mycenter} \tilde{\J} = \mtx{X} \ast \mtx{D_v}(\phi'(\mtx{W}\mtx{X}) - \operatorname{\mathbb{E}}[\phi'(\mtx{W}\mtx{X})])\,. \end{align} Our first result, proven in in Appendix~\ref{app:J-Jt}, relates the spectrum of $\mtx{J}$ to that of $\tilde{\J}$. Note that in case of $\mu(\sigma) = 0$ everywhere (See Assumption~\ref{ass-phi}), then $\operatorname{\mathbb{E}}[\phi'(\mtx{W}\mtx{X})] = 0$ and $\mtx{J} = \mtx{\tilde{J}}$, up to a permutation of the rows. Therefore, this step becomes superfluous in this case. \begin{proposition}\label{pro:J-Jt} Let $\phi:\mathbb{R}\rightarrow \mathbb{R}$ be a nonlinear activation obeying $\abs{\phi''}\le L$. Assume the inputs $\vct{x}_i\in\mathbb{R}^d$ are distributed i.i.d.~$\mathcal{N}(\vct{0},\mtx{I}_d)$ for $i=1,2,\ldots,n$. Then, \begin{align} \sigma_{\min}(\mtx{J}) &\ge \sigma_{\min}(\tilde{\J})\,,\\ \sigma_{\max}(\mtx{J}) &\le \sigma_{\max}(\tilde{\J}) + 3\sqrt{n}\twonorm{\mtx{D_v}\vct{\mu}}\,,\label{eq:J-JtB} \end{align} holds with probability at least $$1-2e^{-n/2} -2n\cdot\exp\left(-c\frac{\|\vct{\mu}\|^2}{\sigma_{\max}^2(\mtx{W})}\right)\,,$$ with $c = v_{\min}^2/(32L^2 v_{\max}^2)$. \end{proposition} Now that we have established a connection between the spectrum of $\mtx{J}$ and $\tilde{\mtx{J}}$, we focus on bounding $\sigma_{\min}(\tilde{\J})$ and $\sigma_{\max}(\tilde{\J})$. The following proposition is at the core of our analysis and may be of independent interest. We defer the proof of this result to Section~\ref{proofJ}. \begin{proposition}\label{pro:J-eig} Let $\phi:\mathbb{R}\rightarrow \mathbb{R}$ be a general activation obeying $\abs{\phi''}\le L$. Assume the inputs $\vct{x}_i\in\mathbb{R}^d$ are distributed i.i.d.~$\mathcal{N}(\vct{0},\mtx{I}_d)$ for $i \in [n]$. Suppose that $n\ge d$ and $k\ge d$. Furthermore, assume \begin{align} &\sqrt{n} \le c\frac{\sigma_{\min}^2\left(\mtx{\Gamma}\mtx{W}\right)}{L^4 \sigma^4_{\max}\left(\mtx{W}\right)+1} d \label{ass1-1}\,, \end{align} holds for a sufficiently small constant $c>0$. Then, there exist constants $C>0$, such that, \begin{align} \sigma_{\min}(\tilde{\J}) &\ge \frac{d}{2} \, \sigma_{\min} (\mtx{D_v}\mtx{\Gamma} \mtx{W})\,,\label{eq:sigminJ1}\\ \sigma_{\max}(\tilde{\J}) &\le C\left(d+ \sqrt{nd}\, \sigma_{\max}(\mtx{D_v}\mtx{\Gamma} \mtx{W})\right) \,,\label{eq:sigmaxJ1} \end{align} holds with probability at least $1 - ne^{-b_1\sqrt{n}} - n^{-1} - 2ne^{-b_2 d}$ with $b_1,b_2>0$ fixed numerical constants. \end{proposition} \begin{remark}\label{rem:assb} As discussed in the proof of Proposition~\ref{pro:J-eig}, when Assumption~\ref{ass-phi} (b) holds in lieu of Assumption~\ref{ass-phi} (a), then $\mtx{\Gamma} = \boldsymbol{0}$. In this case the statement of the above proposition must be modified as follows. Equation~\eqref{ass1-1} should be replaced with \begin{align} &\sqrt{n} \le c\frac{1}{L^4 \sigma^4_{\max}\left(\mtx{W}\right)+1} d\,. \end{align} Further, we have the following bounds in lieu of equations~\eqref{eq:sigminJ1} and~\eqref{eq:sigmaxJ1}. \begin{align*} \sigma_{\min}(\mtx{\tilde{J}}) &\ge \frac{d}{2} \,.\\ \sigma_{\max}(\mtx{\tilde{J}}) &\le C d\,. \end{align*} (See Equations~\eqref{assb-min} and \eqref{assb-max}). \end{remark} We now combine Propositions~\ref{pro:J-Jt} and~\ref{pro:J-eig} to characterize the spectrum of $\mtx{J}$. We defer the proof of this result to Appendix \ref{app:log-eig}. \begin{proposition}\label{cor:log-eig} Let $\phi:\mathbb{R}\rightarrow \mathbb{R}$ be a general activation obeying Assumption~\ref{ass-phi} $(a)$ and $ |\phi'' |\le L$ . Assume the inputs $\vct{x}_i\in\mathbb{R}^d$ are distributed i.i.d.~$\mathcal{N}(\vct{0},\mtx{I}_d)$ for $i\in [n]$. Suppose that \begin{align} \label{boundass} 0<v_{\min} \le |v^*_\ell| \le v_{\max}\quad0<w_{\min} \le \twonorm{\vct{w}^*_\ell} \le w_{\max}\quad\text{for }\ell=1,2,\ldots,k, \end{align} for some fixed constants $v_{\min}, v_{\max}, w_{\min}, w_{\max}$. Further, assume that $k\ge d$ and \begin{align} d \le n \le \frac{c_0 \sigma_{\min}^4(\mtx{W})}{\sigma_{\max}^8(\mtx{W})} d^2\label{ass1-rep} \end{align} for a sufficiently small constant $c_0>0$. Then, there exist constants $C\ge c>0$, such that, \begin{align} \sigma_{\min}(\mtx{J}) &\ge c \, \sigma_{\min}(\mtx{W})\,d \,,\label{eq:sigminJ} \\ \sigma_{\max}(\mtx{J}) &\le C\, \sigma_{\max}( \mtx{W})\sqrt{nk} \,,\label{eq:sigmaxJ} \end{align} holds with probability at least $1 - n^{-1} - 2ne^{-b \sqrt{d}}$ with $b>0$ a fixed numerical constant. Also when Assumption~\ref{ass-phi} $(b)$ holds in lieu of Assumption~\ref{ass-phi} $(a)$, the term $\sigma_{\min}(\mtx{W})$ should be replaced by one in Equations~\eqref{ass1-rep} and \eqref{eq:sigminJ}. \end{proposition} A noteworthy case that will be used in our analysis for both local and global convergence results is $\phi(z) =z^2/2$. Note that for this choice of $\phi$, we have $\operatorname{\mathbb{E}}[\phi'(\mtx{WX})] = 0$ and therefore the centering step~\eqref{cor:log-eig} becomes superfluous as $\tilde{\J} = \mtx{J}$. Further, $\mtx{\Gamma} = \mtx{I}$. Applying Proposition~\ref{cor:log-eig} to this case with $\mtx{W} = \mtx{I}$ and $\vct{v}= (1,1,\dotsc, 1,1)$, we obtain the following result. \begin{corollary}\label{cor:ldentity} Let $\mtx{X}\in\mathbb{R}^{d\times n}$ be a matrix with i.i.d. $\mathcal{N}(0,1)$ entries. Suppose that $d\le n \le c_1 d^2$, for a sufficiently small constant $c_1>0$. Then, there exist constants $C\ge c>0$, such that, \begin{align} \sigma_{\min}(\mtx{X} \ast \mtx{X}) &\ge cd \,,\label{eq:sigminJ-I} \\ \sigma_{\max}(\mtx{X} \ast \mtx{X}) &\le C\sqrt{nd} \,,\label{eq:sigmaxJ-I} \end{align} holds with probability at least $1 - ne^{-b_1\sqrt{n}} - n^{-1} - 2ne^{-b_2 d}$ with $b_1,b_2>0$ fixed numerical constants. \end{corollary} \subsection{Proof of Proposition~\ref{pro:J-eig}}\label{proofJ} Bounding $\sigma_{\min}(\tilde{\J})$ is involved and requires several technical lemmas that are of independent interest. We first present an outline of our approach and then discuss the steps in details. We then focus on bounding $\sigma_{\max}(\tilde{\J})$, which follows along the same lines. \noindent Our proof strategy consists of four main steps: \begin{enumerate} \item \emph{(Whitening).} We let $\mtx{M}\in \mathbb{R}^{d\times k}$ be the left inverse of $\mtx{D_v}\mtx{\Gamma}\mtx{W}$ and construct a block-diagonal matrix $\mtx{D_M}$ with $d$ copies of $\mtx{M}$ on its diagonal. We construct the whitened matrix $\mtx{D_M}\mtx{\tilde{J}}\in \mathbb{R}^{d^2\times n}$. The reason this operation is useful is that it acts as a partial centering of the entries of $\mtx{\tilde{J}}$ so that almost all entries of the resulting matrix have zero mean. \item \emph{(dropping rows)} Let us index the rows of $\mtx{D_M}\mtx{\tilde{J}}\in \mathbb{R}^{d^2\times n}$ by $(i,j)$ for $1\le i\le j\le d$. We next construct the matrix $\mtx{\tilde{J}^c}$ which is obtained from $\mtx{D_M}\mtx{\tilde{J}}$ by dropping rows indexed by $(i,i)$. The reason this operation is useful is that it removes the entries of $\mtx{\tilde{J}}$ that are not centered. Indeed, the notation $\mtx{\tilde{J}^c}$ is used to cue that the columns of $\mtx{\tilde{J}^c}$ are centered (have zero mean). \item \emph{(Bounding sub-exponential norms).} In this step we show that the columns of $\mtx{\tilde{J}^c}$ have bounded sub-exponential norm. \item \emph{(Bounding singular values).} We prove bounds on singular values of a matrix with independent columns and bounded sub-exponential norms. \end{enumerate} Steps one and two collectively act as a nontrivial ``centering" of the columns of $\mtx{\tilde{J}}$. The reason this centering is required is that the columns of the matrix $\mtx{\tilde{J}}$ have non-zero mean which lead to rather large sub-exponential norms. By centering the columns we are able to reduce the sub-exponential norm. The reader may be puzzled as to why we do not use the trivial centering of subtracting the mean from each column of $\mtx{\tilde{J}}$. The reason we do not pursue this path is that we can not directly relate the minimum eigenvalues of the resulting matrix to that of $\mtx{\tilde{J}}$ in a useful way. Steps one and two allow us to center the columns of $\mtx{\tilde{J}}$, while being able to relate the minimum eigenvalue of $\mtx{\tilde{J}}$ to that of the centered matrix $\mtx{\tilde{J}^c}$. We note that under Assumption~\ref{ass-phi} (b), the whitening and droppings step are superfluous because the matrix $\mtx{\tilde{J}}$ is centered. We are now ready to discuss each of these steps in greater detail. \subsubsection{Whitening $\tilde{\J}$} In this step we whiten the matrix $\mtx{\tilde{J}}$ in such a way as most of the entries of the corresponding matrix have zero mean. To explain our whitening procedure we begin this section by computing the expectation of $\mtx{\tilde{J}}$. \begin{lemma}\label{lem:center1} Let $\mtx{\Gamma}$ be given by~\eqref{Gammab} and set $\mtx{A}:= \mtx{D_v}\mtx{\Gamma} \mtx{W}$. Construct the block diagonal matrix $\mtx{D_A}\in \mathbb{R}^{kd\times d^2}$ with $d$ copies of $\mtx{A}$ on its diagonal. Further, consider $\mtx{Q}\in \mathbb{R}^{d^2\times n}$, where its rows are indexed by $(i,j)$, for $1\le i\le j \le d$. The rows $(i,i)$ of $\mtx{Q}$ are all-one, while the other rows are all-zero. Then, \begin{align} \operatorname{\mathbb{E}}[\mtx{\tilde{J}}] = \mtx{D_A}\mtx{Q} \,, \end{align} where the expectation is with respect to the randomness in the inputs $\vct{x}_i$. \end{lemma} Lemma~\ref{lem:center1} above is proven in Appendix~\ref{app:center1}. To center, most of the entries of $\mtx{\tilde{J}}$ we use the structure of $\mtx{Q}$ and whiten it from the left-hand side. We will also show that this whitening procedure will not significantly decrease the minimum eigenvalue. Note that we can assume that $\sigma_{\min}(\mtx{D_v}\mtx{\Gamma}\mtx{W})>0$ otherwise, Claim~\eqref{eq:sigminJ1} becomes trivial. Now let $\mtx{M}\in R^{d\times k}$ be the left inverse of $\mtx{D_v}\mtx{\Gamma}\mtx{W}$ and construct a block-diagonal matrix $\mtx{D_M}$ with $d$ copies of $\mtx{M}$ on its diagonal.The lemma below, relates the minimum singular value of $\mtx{\tilde{J}}$ to that of the whitened matrix $\mtx{D}_{\mtx{M}}\mtx{\tilde{J}}$. \begin{lemma}\label{lem:center2} We have \[\sigma_{\min}(\tilde{\J}) \ge \sigma_{\min}(\mtx{D}_{\mtx{M}}\tilde{\J}) \sigma_{\min}(\mtx{D_v}\mtx{\Gamma}\mtx{W})\,.\] \end{lemma} \begin{proof} We prove this lemma by contradiction i.e.~assume \[\sigma_{\min}(\tilde{\J}) < \sigma_{\min}(\mtx{D}_{\mtx{M}}\tilde{\J}) \sigma_{\min}(\mtx{D_v}\mtx{\Gamma}\mtx{W})\,.\] Now let ${\mtx{v}}$ be the bottom singular vector of $\tilde{\J}$. Then, \[ \twonorm{\tilde{\J}{\mtx{v}}} = \sigma_{\min}(\tilde{\J}) \twonorm{{\mtx{v}}} < \sigma_{\min}(\mtx{D_v}\mtx{\Gamma}\mtx{W}) \sigma_{\min}(\mtx{D_M}\tilde{\J})\twonorm{{\mtx{v}}} \le \frac{\twonorm{\mtx{D_M}\tilde{\J}{\mtx{v}}}}{\sigma_{\max}(\mtx{M})} \le\twonorm{\tilde{\J}{\mtx{v}}}\,, \] which is a contradiction. Note that in the penultimate inequality we used the fact that $\mtx{M}\mtx{D_v}\mtx{\Gamma}\mtx{W} = \mtx{I}$ and the last inequality holds because $\|\mtx{D_M}\|=\sigma_{\max}(\mtx{M})$. \end{proof} \subsubsection{Dropping rows} By Lemma~\ref{lem:center1}, we have \[\operatorname{\mathbb{E}}[\mtx{D_M}\mtx{\tilde{J}}] = \mtx{D_M}\mtx{D_A}\mtx{Q} = \mtx{Q}\,. \] Therefore, the whitened matrix $\mtx{D_M}\mtx{\tilde{J}}$ is almost centered. To reach a completely centered matrix we drop the rows corresponding to the nonzero rows of $\mtx{Q}$. Specifically, let $\mtx{\tilde{J}^c}$ be the matrix obtained after dropping rows $(i,i)$ from $\mtx{D_M}\mtx{\tilde{J}}$, for $1\le i\le d$. By dropping these rows, $\mtx{Q}$ vanishes and hence, $\operatorname{\mathbb{E}}[\mtx{\tilde{J}^c}] = \vct{0}$. Note that $\mtx{\tilde{J}^c}$ is obtained by dropping $d$ of the rows from $\mtx{D_M}\tilde{\J}$. Hence, $\sigma_{\min}(\mtx{D_M}\tilde{\J}) \ge \sigma_{\min}(\mtx{\tilde{J}^c})$. Combining the latter with Lemma \ref{lem:center2} we arrive at \begin{align} \label{mahdicentered} \sigma_{\min}(\tilde{\J}) \ge \sigma_{\min}(\mtx{\tilde{J}^c}) \sigma_{\min}(\mtx{D_v}\mtx{\Gamma}\mtx{W}). \end{align} Thus, in the remainder we shall focus on lower bounding $\sigma_{\min}(\mtx{\tilde{J}^c})$. \subsubsection{Bounding sub-exponential norms} Let $\mtx{\tilde{J}}_{\vct{x}}$ be the column of $\mtx{\tilde{J}}$ corresponding to data point $\vct{x}$. By Lemma~\ref{Pre1}, we can write \[ \mtx{D_M}\mtx{\tilde{J}}_{\vct{x}} = \vct{x} \otimes \mtx{M}\mtx{D_v} (\phi'(\mtx{W}\vct{x}) - \operatorname{\mathbb{E}}[\phi'(\mtx{W}\vct{x})])\,. \] We recall that $\mtx{\tilde{J}^c}$ is obtained by dropping $d$ of the rows from $\mtx{D_M}\mtx{\tilde{J}}$. By the structure of $\mtx{Q} = \operatorname{\mathbb{E}}[\mtx{D_M}\mtx{\tilde{J}}]$, we can alternatively obtain $\mtx{\tilde{J}^c}$ by dropping the same set of rows from $\mtx{D_M}\mtx{\tilde{J}} - \operatorname{\mathbb{E}}[\mtx{D_M}\mtx{\tilde{J}}]$. Note that dropping entries from a vector can only reduce the Orlicz norm. Therefore, \begin{align}\label{eq:subexp1} \|\mtx{\tilde{J}^c}_{\vct{x}}\|_{\psi_1} \le \Big\|\mtx{D_M}\mtx{\tilde{J}}_{\vct{x}} - \operatorname{\mathbb{E}}[\mtx{D_M}\mtx{\tilde{J}}_{\vct{x}}] \Big\|_{\psi_1} = \Big\|\vct{x}\otimes \mtx{M}\mtx{D_v} \vct{z} - \operatorname{\mathbb{E}}[\vct{x}\otimes \mtx{M}\mtx{D_v} \vct{z} ]\Big\|_{\psi_1}. \end{align} In order to bound the right-hand-side of~\eqref{eq:subexp1}, we need to study the tail of the random variable \begin{align*} \bigg\langle\vct{x}\otimes \mtx{M}\mtx{D_v} \vct{z} - \operatorname{\mathbb{E}}[\vct{x}\otimes \mtx{M}\mtx{D_v} \vct{z}],\vct{u}\bigg\rangle \end{align*} for any unit-norm vector $\vct{u}\in\mathbb{R}^{kd}$. To simplify this random variable define $\widetilde{\mtx{U}}\in\mathbb{R}^{d\times d}$ by arranging every $k$ entries of $\vct{u}$ as the rows of $\widetilde{\mtx{U}}$. Note that $\vct{u}=$vect$(\widetilde{\mtx{U}})$. Thus Lemma~\ref{Pre2} allows us to rewrite this random variable in the simplified form \begin{align} \label{tempsubexp} \langle\vct{x}\otimes \mtx{M}\mtx{D_v} \vct{z} - \operatorname{\mathbb{E}}[\vct{x}\otimes \mtx{M}\mtx{D_v} \vct{z}],\vct{u}\rangle=\vct{x}^T\mtx{\widetilde{U}} \mtx{M} \mtx{D_v}\vct{z}-\operatorname{\mathbb{E}}[\vct{x}^T\mtx{\widetilde{U}} \mtx{M} \mtx{D_v}\vct{z}], \end{align} with $\vct{z}\in\mathbb{R}^k$ defined as $\vct{z}:= \phi'(\mtx{W}\vct{x}) - \operatorname{\mathbb{E}}[\phi'(\mtx{W}\vct{x})]$. To characterize the tails of the latter random variable we use the following lemma proven in Appendix~\ref{proof:subexp2}. \begin{proposition}\label{pro:subexp2} Let $\vct{x}\in\mathbb{R}^d$ be a random Gaussian vector distributed as $\mathcal{N}(\vct{0},\mtx{I}_d)$. Also define $\vct{z}:= \phi'(\mtx{W}\vct{x}) - \operatorname{\mathbb{E}}[\phi'(\mtx{W}\vct{x})]$. Then, \begin{align} \label{HW} \mathbb{P}\bigg\{\abs{\vct{x}^T\mtx{U}\vct{z}-\operatorname{\mathbb{E}}[\vct{x}^T\mtx{U}\vct{z}]}\ge t\bigg\}\le 2\exp\left(-\frac{1}{C}\min\left(\frac{t^2}{K^4\fronorm{\mtx{U}}^2},\frac{t}{K^2\opnorm{\mtx{U}}}\right)\right), \end{align} holds with $C$ a fixed numerical constant and $K= \sqrt{L^2\sigma_{\max}^2(\mtx{W})+1}$. \end{proposition} We apply Proposition~\ref{pro:subexp2}, with $\mtx{U} = \mtx{\widetilde{U}} \mtx{M} \mtx{D_v}$ and use Lemma~\ref{Pre3}, to conclude that \begin{align} \Big\|\vct{x}^T \mtx{\widetilde{U}} \mtx{M} \mtx{D_v} \vct{z} - \operatorname{\mathbb{E}}[\vct{x}^T \mtx{\widetilde{U}} \mtx{M} \mtx{D_v} \vct{z} ]\Big\|_{\psi_1}&\le C(L^2\sigma^2_{\max}(\mtx{W})+1) \fronorm{\mtx{\widetilde{U}} \mtx{M} \mtx{D_v}},\nonumber\\ &\le C(L^2\sigma_{\max}^2(\mtx{W})+1)\, \sigma_{\max}(\mtx{M} \mtx{D_v}) \fronorm{\mtx{\widetilde{U}}} ,\nonumber\\ &\le \frac{C}{\sigma_{\min}(\mtx{\Gamma}\mtx{W})} (L^2\sigma_{\max}^2(\mtx{W})+1)\,.\label{eq:subexp3} \end{align} In the last inequality we used the fact that $\fronorm{\widetilde{\mtx{U}}}=\twonorm{\vct{u}}=1$. Now since $\vct{u}$ was arbitrary, recalling Definition~\ref{def2:subexp} and combining equations \eqref{eq:subexp1}, \eqref{tempsubexp} and \eqref{eq:subexp3} we conclude that \begin{align}\label{eq:subexp-impt} \|\mtx{\tilde{J}^c}_{\vct{x}}\|_{\psi_1} \le \frac{C}{\sigma_{\min}(\mtx{\Gamma}\mtx{W})} (L^2\sigma_{\max}^2(\mtx{W})+1)\,. \end{align} \subsubsection{Bounding minimum singular value} Now that we have a bound on the sub-exponential norm of the columns of $\mtx{\tilde{J}^c}$, we are now ready to bound $\sigma_{\min}(\mtx{\tilde{J}^c})$. To this aim we first state a lemma on the spectrum of matrices with independent sub-exponential columns. The Lemma is similar to~\cite[Theorem 3.2]{adamczak2011restricted} and we give a proof in Appendix~\ref{app:mineig}. \begin{proposition}\label{pro:mineig} Let $\vct{u_1}, \vct{u_2}, \dotsc, \vct{u_n}$ be independent sub-exponential random vectors and also let $\psi = \max_{1\le i\le n}\, \|\vct{u_i}\|_{\psi_1}$. Furthermore, let $\mtx{U}$ be a matrix with $\vct{u_1}, \dotsc, \mtx{u_n}$ as columns. Define $\eta_{\min} = \min_{1\le i\le n} \|\vct{u_i}\|$ and $\eta_{\max} = \max_{1\le i\le n} \|\vct{u_i}\|$. Further, set $\xi = \psi K+K'$, where $K,K'\ge 1$ are arbitrary but fixed. Also define \begin{align*} \Delta = C\xi^2 \eta_{\min} \sqrt{n} \log \Big({\frac{2\eta_{\min}}{\sqrt{n}}}\Big). \end{align*} Then \begin{align*} \sigma_{\min}^2(\mtx{U}) \ge \eta_{\min}^2 - \Delta\quad\text{and}\quad\sigma_{\max}^2(\mtx{U}) \le \eta_{\max}^2 + \Delta, \end{align*} hold with probability larger than \[ 1 - C\exp\Big(-cK\sqrt{n}\log \Big({\frac{2\eta_{\min}}{\sqrt{n}}} \Big) \Big) -\P\Big(\eta_{\max} \ge K'\eta_{\min} \Big)\,. \] Here, $c,C>0$ are universal constants. \end{proposition} To apply Proposition~\ref{pro:mineig}, we only need to control norms of columns $\mtx{\tilde{J}^c}_{\vct{x}}$ (and thus the parameters $\eta_{\min}$ and $\eta_{\max}$ in the proposition statement). To lighten the notation, we let $\vct{z} = \mtx{M} \mtx{D_v} (\phi'(\mtx{W}\vct{x}) - \operatorname{\mathbb{E}}[\phi'(\mtx{W}\vct{x})])$. The column of $\mtx{D_M}\mtx{\tilde{J}}$ corresponding to data point $\vct{x}$ is given by $\vct{x}\otimes \vct{z}$ and the column $\mtx{\tilde{J}^c}_{\vct{x}}$ is obtained after dropping the entries $x_i z_i$, for $1\le i\le d$. Hence, \begin{align} \label{temp1} \twonorm{\mtx{\tilde{J}^c}_{\vct{x}}}^2= \sum_{i=1}^d z_i^2 (\sum_{j\neq i} x_j^2) \,. \end{align} We continue by stating a lemma that bounds the Euclidean norm of the random vector $\vct{z}$. We defer the proof of this lemma to Appendix~\ref{app:twonormconc}. \begin{lemma}\label{twonormconc} Assume $\vct{x}$ is a Gaussian random vector distributed as $\mathcal{N}(\vct{0},\mtx{I})$. Furthuremore, assume $\phi:\mathbb{R}\to \mathbb{R}$ is an activation function with bounded second derivative, i.e.~$\abs{\phi''}\le L$. Define \begin{align*} \rho(\mtx{W}):=L\frac{\sigma_{\max}\left(\mtx{W}\right)}{\sigma_{\min}\left(\mtx{\Gamma}\mtx{W}\right)}. \end{align*} Then for $\vct{z} = \mtx{M} \mtx{D_v}(\phi'(\mtx{W}\vct{x}) -\operatorname{\mathbb{E}}[\phi'(\mtx{W}\vct{x})])$, \begin{align} \label{lbnd} \sqrt{d}-t\rho(\mtx{W}) \le \twonorm{\vct{z}} \le (\sqrt{d} +t )\rho(\mtx{W})\,, \end{align} holds with probability at least $1-2e^{-\frac{t^2}{2}}$. \end{lemma} Further, by concentration of $\chi^2$ random variables, with probability at least $1-2e^{-(d-1)/32}$, the following bounds hold. For any fixed $1\le i\le n$, \begin{align}\label{eq:chi2} \frac{d-1}{2}\le \Big|\sum_{j\neq i} x_j^2 \Big| \le \frac{3(d-1)}{2}\,. \end{align} Plugging bound \eqref{lbnd}, with $t=2\sqrt{\log n}$, and bound~\eqref{eq:chi2} into \eqref{temp1}, the followings inequalities hold with probability at least $1-2n^{-2} - 2e^{-(d-1)/32}$, \begin{align} \twonorm{\mtx{\tilde{J}^c}_{\vct{x}}} &\ge \sqrt{\frac{d-1}{2}} (\sqrt{d}-2\rho(\mtx{W}) \sqrt{\log n})\ge \frac{1}{3}\sqrt{d(d-1)}\,,\label{tmp0}\\ \twonorm{\mtx{\tilde{J}^c}_{\vct{x}}} &\le \sqrt{\frac{3(d-1)}{2}} (\sqrt{d}+2\sqrt{\log n})\rho(\mtx{W}) \le 3\sqrt{{d(d-1)}}\rho(\mtx{W})\,. \end{align} Note that the second inequality in~\eqref{tmp0} holds because by our assumption~\eqref{ass1-1}, \begin{align} \label{cond} \rho^2(\mtx{W}) <\frac{L^2\sigma^2_{\max}\left(\mtx{W}\right)+1}{\sigma^2_{\min}\left(\mtx{\Gamma}\mtx{W}\right)}\le c_1\frac{d}{\sqrt{n}} < \frac{d}{16\log n}\,. \end{align} Therefore, by union bounding over $n$ columns with probability at least $1-2n^{-1}-2n e^{-(d-1)/32}$ we have \begin{align} \eta_{\min} &\ge \frac{1}{3}\sqrt{d(d-1)}\,,\label{eq:eta-min}\\ \eta_{\max}&\le 3\sqrt{d(d-1)}\rho(\mtx{W})\,. \label{eq:eta-max} \end{align} Also, by~\eqref{eq:subexp-impt}, the maximum sub-exponential norms of the columns is bounded by \begin{align}\label{eq:subexp-impt2} \psi \le \frac{C}{\sigma_{\min}(\mtx{\Gamma}\mtx{W})} (L^2\sigma_{\max}^2(\mtx{W})+1)\,. \end{align} We now have all the elements to apply Proposition~\ref{pro:mineig}. In particular we use $K = 1$ and $K' = 10 \rho(\mtx{W})$, to conclude that \begin{align} \sigma_{\min}^2(\mtx{\tilde{J}^c}) \ge \eta_{\min}^2 \left(1 - C (\psi + \rho(\mtx{W}))^2 \frac{\sqrt{n}}{\eta_{\min}} \log\left(\frac{2\eta_{\min}}{\sqrt{n}} \right) \right)\,, \end{align} holds with probability at least \[ 1- C\exp\left(-c\sqrt{n} \log \left(\frac{2\eta_{\min}}{\sqrt{n}}\right) \right)\,. \] Using the derived lower bound on $\eta_{\min}$ and the upper bound on $\psi$, we obtain that with high probability $\sigma_{\min}^2(\mtx{\tilde{J}^c}) \ge d/2$, as long as \begin{align} \frac{L^4 \sigma^4_{\max}\left(\mtx{W}\right)+1}{\sigma_{\min}^2\left(\mtx{\Gamma}\mtx{W}\right)}\le c_1 \frac{d}{\sqrt{n}} \,, \end{align} which holds true by the assumption given in Equation~\eqref{ass1-1}. Finally by \eqref{mahdicentered}, we conclude that \begin{align} \sigma_{\min}(\mtx{\tilde{J}}) \ge \frac{d}{2} \sigma_{\min}(\mtx{D_v}\mtx{\Gamma}\mtx{W})\,. \end{align} Note that under Assumption~\ref{ass-phi} (b), we skip the whitening step and by following the same proof, we obtain \begin{align} \sigma_{\min}(\mtx{\tilde{J}}) \ge \frac{d}{2} \,.\label{assb-min} \end{align} \subsubsection{Bounding maximum singular value} The argument we used for the minimum eigenvalue does not apply to the largest eigenvalue. This is because $\mtx{D_M}$ is a fat matrix and the maximum eigenvalue of $\mtx{D_M} \tilde{\J}$ does not provide any information about the maximum eigenvalue of $\tilde{\J}$. To see this clearly, consider the case where the column space of $\tilde{\J}$ intersects with the null space of $\mtx{D_M}$ e.g.~when maximum eigenvector of $\tilde{\J}$ is in the null space of $\mtx{D_M}$. For bounding $\sigma_{\max}(\mtx{\tilde{J}})$, instead of whitening and dropping some of the row, we center $\mtx{\tilde{J}}$ in the most natural way. Specifically, in this section we let $\mtx{\tilde{J}^c}:= \mtx{\tilde{J}} - \operatorname{\mathbb{E}}[\mtx{\tilde{J}}]$. Then, \[\mtx{\tilde{J}^c}_{\vct{x}} = \tilde{\J}_{\vct{x}} - \operatorname{\mathbb{E}}[\tilde{\J}_{\vct{x}}] = \vct{x}\otimes \mtx{D_v}\vct{z} - \operatorname{\mathbb{E}}[\vct{x}\otimes \mtx{D_v}\vct{z}]\,, \] with $\vct{z} = \phi'(\mtx{W}\vct{x})-\operatorname{\mathbb{E}}[\phi'(\mtx{W}\vct{x})]$. Applying Proposition~\ref{pro:subexp2}, we get \begin{align} \|\mtx{\tilde{J}^c}_{\vct{x}}\|_{\psi_1} \le {C}\|{\mtx{v}}\|_\infty (L^2\sigma_{\max}^2(\mtx{W})+1)\,. \end{align} Similar to Section~\ref{pro:mineig}, we can bound the maximum and the minimum norms of the columns of $\mtx{\tilde{J}^c}$. With probability at least $1-2n^{-1}-2n e^{-(d-1)/32}$, \begin{align} \eta_{\min}\ge d/3\,,\quad \eta_{\max}\le 3d\rho(\mtx{W})\,. \end{align} By employing Proposition~\ref{pro:mineig}, \begin{align} \sigma_{\max}^2(\mtx{\tilde{J}^c}) &\le C \left (d^2+ \sqrt{n} d\|{\mtx{v}}\|_\infty^2\, (L^4\sigma_{\max}^4(\mtx{W})+1) + \sqrt{n}d\rho^2(\mtx{W}) \right) \nonumber \\ &\le Cd^2\left(1+ c_1\|{\mtx{v}}\|_\infty^2\, \sigma_{\min}^2(\mtx{\Gamma} \mtx{W})+c_1\right)\nonumber\\ &\le Cd^2\left(1+ \|{\mtx{v}}\|_\infty^2\, \sigma_{\min}^2(\mtx{\Gamma} \mtx{W})\right)\,,\label{sigmaxJc} \end{align} where the second inequality follows from our assumption given by Equation~\eqref{ass1-1}. We next bound the maximum eigenvalue of $\operatorname{\mathbb{E}}[\mtx{\tilde{J}}]$. Invoking Lemma~\ref{lem:center1}, we have \[ \operatorname{\mathbb{E}}[\mtx{\tilde{J}}] = \mtx{D_A}\mtx{Q} = {\rm vec}(\mtx{A}) \mathbf{1}^T_n\,, \] where $\mathbf{1}_n\in \mathbb{R}^n$ is the all-one vector. Therefore, \begin{align} \sigma_{\max}(\operatorname{\mathbb{E}}[\mtx{\tilde{J}}]) \le \sqrt{n} \|\mtx{A}\|_F \le \sqrt{n d}\, \sigma_{\max}(\mtx{D_v}\mtx{\Gamma} \mtx{W})\,.\label{sigmaxQ} \end{align} Combining Equations~\eqref{sigmaxJc} and \eqref{sigmaxQ} via the triangular inequality we arrive at: \begin{align*} \sigma_{\max}(\mtx{\tilde{J}}) &\le Cd\Big(1+ \|{\mtx{v}}\|_\infty\, \sigma_{\min}(\mtx{\Gamma} \mtx{W})\Big) + \sqrt{n d}\, \sigma_{\max}(\mtx{D_v}\mtx{\Gamma} \mtx{W}),\\ &\le C\left(d+ \sqrt{nd}\, \sigma_{\max}(\mtx{D_v}\mtx{\Gamma} \mtx{W})\right)\,, \end{align*} where the second inequality holds because $n\ge d$ and $\|{\mtx{v}}\|_\infty \sigma_{\min}(\mtx{\Gamma} \mtx{W}) \le \sigma_{\max}(\mtx{D_v}\mtx{\Gamma} \mtx{W})$ as per Lemma~\ref{Pre4}. Note that under Assumption~\ref{ass-phi} (b), matrix $\mtx{\tilde{J}}$ is centered and the whitening step should be skipped. Indeed, by a similar argument for inequality~\eqref{sigmaxJc}, we have \begin{align} \sigma_{\max}^2(\mtx{\tilde{J}}) &\le C \left (d^2+ \sqrt{n} d v_{\max}^2 (L^4\sigma_{\max}^4(\mtx{W})+1) + \sqrt{n}d\sigma_{\max}^2(\mtx{W}) \right),\nonumber\\ &\le Cd^2 (1+c_1 v_{\max}^2+c_1),\nonumber\\ &\le C d^2\,. \label{assb-max} \end{align} \subsection{Proof of Theorem~\ref{localthm}}\label{pfthmloc} We begin by stating a few useful lemmas. We defer the proofs of these lemmas to the Appendices. Throughout, we assume $|\phi'| < B$ and $|\phi''| <L$. We present the proof under Assumption~\ref{ass-phi} (a). The proof under Assumption~\ref{ass-phi} (b) follows by an analogous argument. We begin by a lemma that bound the perturbation of the Jacobian matrix as a function of its inputs. We defer the proof to Appendix~\ref{proof:Jpert}. \begin{lemma}\label{Jpert} Assume $\vct{x}_i$ are i.i.d.~Gaussian random vectors distributed as $\mathcal{N}(\vct{0},\mtx{I})$. Further, suppose that $n\le c d^2$, and let $\mtx{J}({\mtx{v}},\mtx{W}) = \mtx{D_v}\phi'(\mtx{W}\mtx{X})\ast\mtx{X}$ be the Jacobian matrix associated to weights ${\mtx{v}}$ and $\mtx{W}$. Then, for any two fixed matrices $\widetilde{\mtx{W}},\mtx{W}\in\mathbb{R}^{k\times d}$ and any two fixed vectors ${\mtx{v}},\widetilde{\vct{v}} \in \mathbb{R}^{k}$, \begin{align}\label{eq:Jpert} \opnorm{J(\widetilde{\vct{v}},\widetilde{\mtx{W}})-J({\mtx{v}},\mtx{W})}\le C'\sqrt{nd}\, \Big(\|{\mtx{v}}\|_\infty \opnorm{\widetilde{\mtx{W}}-\mtx{W}} + \|\widetilde{\vct{v}}-{\mtx{v}}\|_\infty\Big) , \end{align} holds with probability at least $1 - ne^{-b_1\sqrt{n}} - n^{-1} - 2ne^{-b_2 d}$. Here, $b_1,b_2>0$ are fixed numerical constants. \end{lemma} Now note that by Proposition~\ref{cor:log-eig} for ${\mtx{v}}^*$ and $\mtx{W}^*$, we have \begin{align} \sigma_{\min}(\mtx{J}({\mtx{v}}^*,\mtx{W}^*)) \ge c\,\sigma_{\min}(\mtx{W}^*) d\,.\label{sminW*} \end{align} Using this value of $c$ we define the radius $$R: = \frac{c}{4C'} \sigma_{\min}(\mtx{W}^*) \sqrt{\frac{d}{n}}\,,$$ where $C'$ is the same quantity that appears in Equation~\eqref{eq:Jpert}. Without loss of generality, we can assume $c < 4C' w_{\max}v_{\max}$. Plugging this into the definition of $R$ together with the fact that $d \le n$, allows us to conclude that $R\le v_{\max}$. Define the set $\Omega\subseteq \mathbb{R}^{k\times d} \times \mathbb{R}^k$ as follows: \begin{eqnarray} \Omega := \Big\{({\mtx{v}},\mtx{W}):\, \|{\mtx{v}}-{\mtx{v}}^*\|_\infty \le R,\,\, \fronorm{\mtx{W}-\mtx{W}^*} \le R/\|{\mtx{v}}^*\|_\infty\Big\} \end{eqnarray} In the next lemma, proven in Appendix~\ref{proof:aux1}, we relate the gradients $\nabla_{(\mtx{W})} \mathcal{L}(\vct{v},\mtx{W})$ and $\nabla_{\vct{v}} \mathcal{L}(\vct{v},\mtx{W})$ to the function value $ \mathcal{L}(\vct{v},\mtx{W})$. \begin{lemma}\label{lem:aux1} For $({\mtx{v}},\mtx{W})\in \Omega$, the following inequalities hold with probability at least $1 - n^{-1} - 2ne^{-b \sqrt{d}}$ for some constant $b>0$. \begin{align} \fronorm{\nabla_{\mtx{W}} \mathcal{L}(\vct{v},\mtx{W})}^2\ge m_{\rm{L}} \mathcal{L}(\vct{v},\mtx{W})\,\quad \,\,&m_{\rm{L}}:= \frac{1}{2} c^2 \sigma_{\min}^2(\mtx{W}^*) \frac{d^2}{n}\,,\label{eq:aux1}\\ \fronorm{\nabla_{\mtx{W}} \mathcal{L}(\vct{v},\mtx{W})}^2\le m_{\rm{U}} \mathcal{L}(\vct{v},\mtx{W})\,\quad \,\,&m_{\rm{U}}:= \frac{9}{2} C^2 \sigma_{\max}^2(\mtx{W}^*) k\,,\label{eq:aux2}\\ \opnorm{\nabla_{\vct{v}} \mathcal{L}(\vct{v},\mtx{W})}_\infty^2\le \widetilde{m}_{\rm{U}} \mathcal{L}(\vct{v},\mtx{W})\,\quad \,\,&\widetilde{m}_{\rm{U}}:= \left(4\phi^2(0)+ 128 B^2 + 128B^2 w_{\max}^2\right)\,.\label{eq:aux3} \end{align} \end{lemma} Our next lemma, proven in Appendix~\ref{proof:aux2}, upper bounds the function value $\mathcal{L}(\vct{v},\mtx{W})$ in terms of distance of $(\vct{v},\mtx{W})$ to the planted solution $(\vct{v},\mtx{W})$. \begin{lemma}\label{lem:aux2} The following bound holds with probability at least $1 - 2e^{-b_0d}$ for some constant $b_0>0$. \begin{align} \mathcal{L}({\mtx{v}},\mtx{W}) \le \twonorm{{\mtx{v}}^*}^2 \fronorm{\mtx{W}-\mtx{W}^*}^2 + 2\left( \phi^2(0)+2B^2 w_{\max}^2 \right) k^2 \opnorm{{\mtx{v}}-{\mtx{v}}^*}_\infty^2\,. \end{align} \end{lemma} Finally, the lemma below, proven in Appendix~\ref{proof:aux3}, controls the second order derivative of the loss function $\mathcal{L}({\mtx{v}},\mtx{W})$. \begin{lemma}\label{lem:aux3} The function $\mathcal{L}({\mtx{v}},\mtx{W})$ is $\beta$-smooth on $\Omega$. Namely, there exists an event of probability at least $1 - n^{-1} - 2ne^{-b \sqrt{d}}$, such that on this event, for any $({\mtx{v}},\mtx{W})\in \Omega$, we have \begin{align} \nabla^2 \mathcal{L}({\mtx{v}},\mtx{W}) \le \beta \mtx{I}\,, \end{align} where $\nabla^2\mathcal{L}({\mtx{v}},\mtx{W}) \in \mathbb{R}^{(kd+k)\times (kd+k)}$ denotes the Hessian w.r.t both ${\mtx{v}}$, $\mtx{W}$. Further, the smoothness parameter $\beta$ is given by \begin{align}\label{eq:beta} \beta: = \left( {3C^2}\sigma_{\max}^2(\mtx{W}^*) + 8 v_{\max}^2 BL+ 4B^2w_{\max}^2+2\phi^2(0)\right) k\,. \end{align} \end{lemma} With this lemmas in place we are now ready to present our local convergence analysis. To this aim note that Lemma~\ref{lem:aux1} (Equation~\eqref{eq:aux1}) implies that for $({\mtx{v}},\mtx{W})\in \Omega$, the function $\mathcal{L}({\mtx{v}},\mtx{W})$ satisfies the Polyak-Lojasiewicz (PL) inequality~\cite{polyak1963gradient} \begin{align} \label{polyak} \fronorm{\nabla_{\mtx{W}} \mathcal{L}(\vct{v},\mtx{W})}^2 + \fronorm{\nabla_{{\mtx{v}}} \mathcal{L}(\vct{v},\mtx{W})}^2 \ge m_{\rm{L}} \mathcal{L}(\vct{v},\mtx{W})\,. \end{align} We shall now focus on how the loss function value changes in one iteration. Using the $\beta$-smoothness condition for $\alpha\le 1/\beta$ we have \begin{align} &\mathcal{L}\left(\vct{v}-\alpha \nabla_{\vct{v}} \mathcal{L}\vct{v},\mtx{W}-\alpha\nabla_{\mtx{W}} \mathcal{L}(\mtx{W})\right),\nonumber\\ &\le \mathcal{L}({\mtx{v}},\mtx{W})-\alpha\left(\twonorm{\nabla_{\vct{v}} \mathcal{L}(\vct{v},\mtx{W})}^2+\fronorm{\nabla_{\mtx{W}} \mathcal{L}(\vct{v},\mtx{W})}^2\right)+\frac{\alpha^2\beta}{2}\left(\twonorm{\nabla_{\vct{v}} \mathcal{L}(\vct{v},\mtx{W})}^2+\fronorm{\nabla_{\mtx{W}} \mathcal{L}(\vct{v},\mtx{W})}^2\right),\nonumber\\ &= \mathcal{L}({\mtx{v}},\mtx{W})-\alpha\left(1-\frac{\alpha\beta}{2}\right)\left(\twonorm{\nabla_{\vct{v}} \mathcal{L}(\vct{v},\mtx{W})}^2 + \fronorm{\nabla_{\mtx{W}} \mathcal{L}(\vct{v},\mtx{W})}^2\right),\nonumber\\ &\le \left(1-\sm_{\rm{L}}\left(1-\frac{\alpha\beta}{2}\right)\right)\mathcal{L}(\vct{v},\mtx{W})\le \left(1-\frac{\alpha}{2}m_{\rm{L}}\right)\mathcal{L}(\vct{v},\mtx{W})\,,\label{eq:aux4} \end{align} where in the last inequality we used \eqref{polyak}. Our assumptions on the initial weights $\vct{v}_0$ and $\mtx{W}_0$ in equations \eqref{eq:initv0} and \eqref{eq:initW0} imply the following identities. \begin{align} \fronorm{\mtx{W}_0-\mtx{W}^*} &\le \frac{R m_{\rm{L}}}{4\sqrt{2 m_{\rm{U}}}\, v_{\max} \twonorm{{\mtx{v}}^*}}\,, \label{eq:initW}\\ \opnorm{{\mtx{v}}_0-{\mtx{v}}^*}_\infty &\le \frac{Rm_{\rm{L}}}{4\sqrt{2m_{\rm{U}}} v_{\max}\sqrt{2(\phi^2(0)+2B^2 w_{\max}^2)}\,k }\,. \label{eq:initv} \end{align} We next show that starting with ${\mtx{v}}_0,\mtx{W}_0$ obeying \eqref{eq:initW} and~\eqref{eq:initv}, the entire trajectory of gradient descent remains in the set $\Omega$. To establish this we use induction on $\tau$. The induction basis $\tau=0$ is trivial. Assuming the induction hypothesis for $0\le t\le \tau-1$, we show that the result continues to hold for $t=\tau$. By our induction hypothesis ($(\vct{v}_t,\mtx{W}_t)\in\Omega$ for $t=0,1,\ldots,\tau-1$) and \eqref{eq:aux4}, we have \begin{align*} \mathcal{L}(\vct{v}_{\tau},\mtx{W}_{\tau})&= \mathcal{L}\left(\vct{v}_{\tau-1}-\alpha\nabla_{\vct{v}} \mathcal{L}(\vct{v}_{\tau-1},\mtx{W}_{\tau-1}),\mtx{W}_{\tau-1}-\alpha\nabla_{\mtx{W}} \mathcal{L}(\mtx{v}_{\tau-1},\mtx{W}_{\tau-1})\right),\\ &\le\left(1-\frac{\alpha}{2}m_{\rm{L}}\right)\mathcal{L}(\vct{v}_{\tau-1},\mtx{W}_{\tau-1}). \end{align*} By iterating the above identity we arrive at \begin{align}\label{eq:gradient-tau} \mathcal{L}(\vct{v}_{\tau},\mtx{W}_{\tau}) &\le\left(1-\frac{\alpha}{2}m_{\rm{L}}\right)^\tau \mathcal{L}(\vct{v}_0,\mtx{W}_0). \end{align} To show that $(\vct{v}_\tau,\mtx{W}_\tau)\in\Omega$ we proceed by quantifying how far the gradient descent trajectory can get from $(\vct{v}_0,\mtx{W}_0)$. \begin{align} \fronorm{\mtx{W}_\tau-\mtx{W}_0}=&\fronorm{\sum_{t=1}^{\tau} \left(\mtx{W}_t-\mtx{W}_{t-1}\right)} \le\sum_{t=1}^{\tau}\fronorm{\mtx{W}_t-\mtx{W}_{t-1}}\nonumber\\ \le& \alpha \sum_{t=1}^{\tau}\fronorm{\nabla_{\mtx{W}} \mathcal{L}(\vct{v}_{t-1},\mtx{W}_{t-1})} \stackrel{(a)}{\le} \alpha\sqrt{m_{\rm{U}}} \sum_{t=1}^{\tau}\sqrt{\mathcal{L}(\vct{v}_{t-1},\mtx{W}_{t-1})}\nonumber\\ \stackrel{(b)}{\le}& \alpha\sqrt{m_{\rm{U}}} \sum_{t=1}^{\tau}\sqrt{\left(1-\frac{\alpha}{2}m_{\rm{L}}\right)^{t-1}\mathcal{L}(\vct{v}_0,\mtx{W}_0)}\nonumber\\ \le& \alpha\sqrt{m_{\rm{U}} \mathcal{L}(\vct{v}_0,\mtx{W}_0)} \sum_{t=1}^{\tau}\left(1-\frac{\alpha}{4}m_{\rm{L}}\right)^{t-1}\nonumber\\ \le&\frac{\alpha\sqrt{m_{\rm{U}} \mathcal{L}(\vct{v}_0,\mtx{W}_0)}}{\alpha\frac{m_{\rm{L}}}{4}}\nonumber\\ =&\frac{4}{m_{\rm{L}}}{\sqrt{m_{\rm{U}} \mathcal{L}(\vct{v}_0,\mtx{W}_0)}}\,,\label{W-dist} \end{align} where $(a)$ follows from $({\mtx{v}}_{t-1},\mtx{W}_{t-1})\in \Omega$ and Lemma \ref{lem:aux1} equation \eqref{eq:aux2} and $(b)$ follows from~\eqref{eq:gradient-tau}. Likewise, we obtain \begin{align} \opnorm{\vct{v}_\tau-\vct{v}_0}_\infty \le \frac{4}{m_{\rm{L}}} {\sqrt{\widetilde{m}_{\rm{U}} \mathcal{L}(\vct{v}_0,\mtx{W}_0)}}\label{v-dist} \end{align} Using bounds~\eqref{W-dist} and~\eqref{v-dist}, in order to show that $({\mtx{v}}_\tau,\mtx{W}_\tau)\in \Omega$, it suffices to show that \begin{align} \mathcal{L}({\mtx{v}}_0,\mtx{W}_0) \le \frac{R^2 m_{\rm{L}}^2}{16\max(\widetilde{m}_{\rm{U}} , m_{\rm{U}} v_{\max}^2 )}\,.\label{eq:LB} \end{align} The right-hand sides of Equation~\eqref{eq:LB} depends on $\max(\widetilde{m}_{\rm{U}},m_{\rm{U}} v_{\max}^2)$. The dominant term is $m_{\rm{U}} v_{\max}^2$ because it is of order at least $k$, while $m_{\rm{U}}$ is $O(1)$. Therefore, the desired bound in~\eqref{eq:LB} is equivalent to \begin{align} \mathcal{L}({\mtx{v}}_0,\mtx{W}_0) \le \frac{R^2 m_{\rm{L}}^2}{16m_{\rm{U}} v_{\max}^2}\,.\label{eq:LB2} \end{align} We can verify Equation~\eqref{eq:LB2} by using Lemma~\ref{lem:aux2} combined with \eqref{eq:initW} and~\eqref{eq:initv}. This completes the induction argument and shows that $({\mtx{v}}_\tau,\mtx{W}_\tau)\in \Omega$ for all $\tau\ge 1$. Finally, since $({\mtx{v}}_\tau,\mtx{W}_\tau)\in \Omega$ for all $\tau\ge 0$, \eqref{eq:gradient-tau} holds for all $\tau\ge 0$. Substituting for $m_{\rm{L}}$, we obtain \begin{align} \mathcal{L}(\vct{v}_{\tau},\mtx{W}_{\tau}) &\le\left(1-\frac{\alpha c^2}{4} \sigma_{\min}^2( \mtx{W}^*) \frac{d}{n}\right)^\tau \mathcal{L}(\vct{v}_0,\mtx{W}_0),\nonumber\\ &\le\left(1-{ c^2} w_{\max} \frac{\alpha d}{4n}\right)^\tau \mathcal{L}(\vct{v}_0,\mtx{W}_0)\,, \end{align} where the last step holds because $\sigma_{\min}(\mtx{W}^*) < w_{\max}$. This concludes the proof under Assumption~\ref{ass-phi} (a). The claim under Assumption~\ref{ass-phi} (b) can be proven by a similar argument. The only required adjustment is that the initial radius $R$ and the term $m_{\rm{L}}$ should now be defined via $R= \frac{c}{4C'} \sqrt{{d}/{n}}$ and $m_{\rm{L}}= c^2 {d^2}/(2n)$. \subsection*{Acknowledgements} This work was done in part while M.S. was visiting the Simons Institute for the Theory of Computing. A.J.~was partially supported by a Google Faculty Research Award. A.J.~would also like to acknowledge the financial support of the Office of the Provost at the University of Southern California through the Zumberge Fund Individual Grant Program. M.S.~would like to thank Peter Bartlett for discussions related to \cite{zhong2017recovery}.

https://arxiv.org/abs/2105.07472

Lexicographic Enumeration of Set Partitions

In this report, we summarize the set partition enumeration problems and thoroughly explain the algorithms used to solve them. These algorithms iterate through the partitions in lexicographic order and are easy to understand and implement in modern high-level programming languages, without recursive structures and jump logic. We show that they require linear space in respect to the set cardinality and advance the enumeration in constant amortized time. The methods discussed in this document are not novel. Our goal is to demonstrate the process of enumerating set partitions and highlight the ideas behind it. This work is an aid for learners approaching this enumeration problem and programmers undertaking the task of implementing it.

\section{Introduction} \label{sec:introduction} \citet[Section 7.2.1.5]{knuth2014art} discusses the problem of partition enumeration, which consists in enumerating the number of ways that a set can be partitioned into non-empty subsets (or blocks). Following Knuth's notation, each partition of a set $U=\{u_1,u_2,\dots,u_n\}$ can be represented by a \textit{restricted growth string} $a_1 a_2 \dots a_n$ such that \begin{equation}\label{eq:restricted-growth-string} a_1 = 0 \text{~~~and~~~} a_{j+1} \le 1 + \text{max}(a_1,\dots,a_j) \text{~for~} 1 \le j < n, \end{equation} where $a_z$ shows the block at which $u_z$ is in. Hence, two elements $u_i$ and $u_j$ are in the same block if and only if $a_i = a_j$. For example, the 15 possible partitions of the set with 4 elements are \begin{quote} 0000, 0001, 0010, 0011, 0012, 0100, 0101, 0102, 0110, 0111, 0112, 0120, 0121, 0122, 0123. \end{quote} These partitions are also presented here in \textit{lexicographic order} in respect to their restricted growth string representation. The partitions in this enumeration can be grouped based on the number of blocks that they have, where a $k$-partition is a partition with exactly $k$ blocks. For example, the above set has a single 1-partition which is the 0000 and a single 4-partition which is the 0123. The 2-partitions, in lexicographic order, are \begin{quote} 0001, 0010, 0011, 0100, 0101, 0110, 0111 \end{quote} and the 3-partitions are, in the same order, \begin{quote} 0012, 0102, 0112, 0120, 0121, 0122. \end{quote} The analogy of a partition with its restricted growth string is illustrated in Figure~\ref{fig:growth-string}. Although it is possible to set $a_1$ as any arbitrary value higher than 0, such change is inconsequential and, for simplicity, we study the $a_1=0$ case. The restrictions in Equation~\ref{eq:restricted-growth-string} are necessary in order to ensure that each partition has a unique restricted growth string representation; for example 0012 and 0021 that point to the same partition would otherwise both be traversed as part of the enumeration. \begin{figure} \centering \small \begin{tikzpicture} \node (A) at (-3,1) {$\left\{ u_1,u_2,u_3,u_4,u_5 \right\}$}; \node (Adesc) at (-3,1.75) {The initial set $U$}; \node (B) at (3,1) {$\big\{ \left\{u_2,u_3\right\},\left\{u_5\right\},\left\{u_4,u_1\right\} \big\}$}; \node (Bdesc) at (3,1.75) {A partition of $U$}; \node (C) at (-1,-1) {$\big\{ \overbrace{\left\{u_1,u_4\right\}}^0,\overbrace{\left\{u_2,u_3\right\}}^1,\overbrace{\left\{u_5\right\}}^2 \big\}$}; \node (Cdesc) at (-1,-1.9) {Sorted partition}; \node (D) at (5,-1) {$\overbrace{0}^{u_1} \overbrace{1}^{u_2} \overbrace{1}^{u_3} \overbrace{0}^{u_4} \overbrace{2}^{u_5}$}; \node (Ddesc) at (5,-1.9) {Restricted growth string}; \path [->](A) edge[line width=0.3mm] node[left] {} (B); \path [->](B) edge[line width=0.3mm] node[left] {} (C); \path [->](C) edge[line width=0.3mm] node[left] {} (D); \end{tikzpicture} \caption{The representation of a partition as a growth string. The blocks of the partition are sorted based on the least element in each block, in this example $u_1$, $u_2$ and $u_5$. Each number $a_i$ in the growth string shows the block containing $u_i$.} \label{fig:growth-string} \end{figure} It is immediately evident that the problem of enumerating all partitions of a set is a special case of another problem: enumerating all partitions of a set with at most $k$ blocks; if $k=n$, then the two problems are equivalent. Conversely, the second problem is a special case of a third problem: enumerating all partitions of a set with the number of blocks residing between $k_{min}$ and $k_{max}$, from which it is reduced for $k_{min}=1$ and $k_{max}=k$. The latter problem is also a more general case of enumerating all partitions with exactly $k$ blocks. The process can be further generalized by the problem of enumerating partitions of $K=\{k_1,k_2,k_3,\dots\}$ blocks, where $k_i$ are arbitrary integer block counts greater than zero. More formally, we define 5 lexicographic set partition enumeration problems: \begin{description} \item [Problem A] Enumerate all partitions of a set. \item [Problem B] Enumerate all partitions of a set with at most $k$ blocks. \item [Problem C] Enumerate all partitions of a set with exactly $k$ blocks. \item [Problem D] Enumerate all partitions of a set with the number of blocks between $k_{min}$ and $k_{max}$. \item [Problem E] Enumerate all partitions of a set with the number of blocks in the $K=\{k_1,k_2,k_3,\dots\}$. \end{description} These problems are stated, and later studied, in this order, as it appears to be a natural sequence of the algorithms involved. An overview of these problems can be seen in Figure~\ref{fig:reductions}. \begin{figure} \centering \small \begin{tikzpicture} \node[shape=circle,draw=black,line width=0.25mm] (A) at (0,0) {A}; \node[shape=circle,draw=black,line width=0.25mm] (B) at (0.6,1) {B}; \node[shape=circle,draw=black,line width=0.25mm] (C) at (1.8,1) {C}; \node[shape=circle,draw=black,line width=0.25mm] (D) at (1.2,2) {D}; \node[shape=circle,draw=black,line width=0.25mm] (E) at (1.2,3) {E}; \node[align=left] at (0.9,4) {\textbf{Reductions}}; \path [->,line width=0.25mm](A) edge node[left] {} (B); \path [->,line width=0.25mm](B) edge node[left] {} (D); \path [->,line width=0.25mm](C) edge node[left] {} (D); \path [->,line width=0.25mm](D) edge node[left] {} (E); \node[align=left] at (4.5,0) {$n$}; \node[align=left] at (4.5,1) {$n,k$}; \node[align=left] at (4.5,2) {$n,k_{min},k_{max}$}; \node[align=left] at (4.5,3) {$n,K$}; \node[align=left] at (4.5,4) {\textbf{Parameters}}; \node[align=left] at (9,0) {without restriction}; \node[align=left] at (9,1) {at most $k$ -- exactly $k$}; \node[align=left] at (9,2) {between $k_{min}$ and $k_{max}$}; \node[align=left] at (9,3) {any $k \in K$}; \node[align=left] at (9,4) {\textbf{Behavior}}; \end{tikzpicture} \caption{Overview of the set partition enumeration problems. The figure shows the reductions among Problems A, B, C, D and E, their parameters and their behavior. An edge $U \to V$ shows that problem $U$ can be reduced to problem $V$ with parameter adjustment. Problem E is the most general problem discussed in this document.} \label{fig:reductions} \end{figure} Knuth describes a classic set partition enumeration algorithm which is due to \citet{10.1145/367651.367661} and enumerates all partitions of a set in lexicographic order (Problem A). Other authors have since improved Hutchinson's work and incrementally achieved better running times \citep{nijenhuis2014combinatorial,110002673365,10.1093/comjnl/31.3.283,10.1093/comjnl/32.3.281}. These lexicographic enumeration algorithms can be iterative or recursive and are often based on the discovery tree formed by increasing number of elements $n$. In this document, we additionally articulate Problems B, C, D and E and adopt more specialized algorithms for the solution of these problems. The algorithms presented are formulated without \textit{jump} logic (e.g., \textit{goto}, \textit{break}) or recursive structures, as these are often difficult to understand and can lead to error-prone code. They abide by modern development practices, are easy to implement and, if implemented naturally, don't lead to write-only code. For this reason, we reduce the definition of each algorithm into the definition of two components: \begin{enumerate} \item The first partition that is returned by the enumerator. \item The function \texttt{next(p,s)} which determines the next partition that is visited, given the previous partition \texttt{p} and any auxiliary state \texttt{s} that the algorithm uses. \end{enumerate} Using these components, a full lexicographic enumeration is possible via an iterative approach, where in each iteration a new restricted growth string is generated. We prove that all presented algorithms use extra space $\Theta(n)$ and the \texttt{next} method runs in constant amortized time. We use the restricted growth string notation for the algorithms as it is easier to understand and can be implemented simply with an integer array. This document is inspired by the work of \citet{orlov2002efficient}, which is aimed at programmers undertaking the implementation of partition enumeration algorithms. Additionally, we describe the problems and their solutions in a learner's perspective as well, with formal complexity analysis. Finally, we introduce the additional problems B, D and E that are not mentioned in that report, and form the generalizations of the simpler problems A and C. This document is structured as follows: Section~\ref{sec:preliminaries} is a short preliminaries overview and the Problems A, B, C, D and E are presented individually in Sections~\ref{sec:problem-a} through~\ref{sec:problem-e}. \section{Preliminaries} \label{sec:preliminaries} It has been previously shown that the number of possible ways that a set of $n$ elements can be partitioned is defined by the $n$-th \textit{Bell number}, denoted by $\mathcal{B}_n$. Likewise, $\mathcal{S}(n,k)$ is the \textit{Stirling number of the second kind} and represents the number of ways to partition a set of $n$ objects into $k$ non-empty blocks. Bell numbers and Stirling numbers of the second kind are related via the following equation: \begin{equation} \mathcal{B}_n = \sum_{k=0}^n \mathcal{S}(n,k). \end{equation} We also note the following asymptotic relations that will be used for the analysis of the algorithms in the following sections: \begin{equation} \label{eq:asymptotic} \frac{\mathcal{B}_{n+1}}{\mathcal{B}_n} \sim \frac{n}{\ln n} \text{~~~and~~~} \mathcal{S}(n,k) \sim \frac{k^n}{k!}. \end{equation} The notation we use throughout this document to describe asymptotic equivalence is defined in~\citet[Section 1.4]{de1981asymptotic}. In particular, two functions $f(x)$ and $g(x)$ are asymptotically equivalent as $x \to \infty$ if the quotient $f(x)/g(x)$ tends to unity and we note this relation as $f(x) \sim g(x)$. Finally, we define $\mathcal{S}^+(n,k)$ as \begin{equation} \mathcal{S}^+(n,k) = \sum_{i=0}^k \mathcal{S}(n,i). \end{equation} Regarding the quantity $\mathcal{S}^+(n+1,k)/\mathcal{S}^+(n,k)$, because $\mathcal{S}(n,i) = i^n/i!$, the limit of the fraction of sums (asymptotically) only depends on the most dominant factor, which is $k \cdot \mathcal{S}(n,k) = k^{n+1}/k!$ for the numerator and $\mathcal{S}(n,k) = k^n/k!$ for the denominator. Hence, it also holds that asymptotically \begin{equation} \label{eq:asymptotic-2} \frac{\mathcal{S}^+(n+1,k)}{\mathcal{S}^+(n,k)} \sim \frac{\mathcal{S}(n+1,k)}{\mathcal{S}(n,k)} \sim k. \end{equation} \section{Problem A} \label{sec:problem-a} Problem A is the most basic problem of the set partition enumeration problem and can be solved using Hutchinson's algorithm. In this section, we present a modified version of this algorithm that is easier to be implemented in modern, high-level programming languages. The partitions are returned in the lexicographic order of their respective restricted growth strings. For example, for $n=4$, \texttt{next(0101) = 0102} and \texttt{next(0112) = 0120}. The full lexicographic order of the $n=4$ partitions were shown earlier in Section~\ref{sec:introduction}. Similar to the original algorithm, the modification uses auxiliary state array $b = [b_1, b_2, \dots, b_n]$, where $b_i = \text{max}(a_1, a_2, \dots, a_{i-1})$. If adjusted successively, then $b_i = \text{max}(a_{i-1}, b_{i-1})$. Hence, the restrictions of Equation~\ref{eq:restricted-growth-string} can be simplified to \begin{equation}\label{eq:conditions-algorithm-a} a_1 = 0 \text{~~~and~~~} a_{j} \le 1 + b_{j} \text{~for~} 1 \le j < n. \end{equation} The arrays $a$ and $b$ are initialized with zeroes. Starting from position $n$ of the $a$ array and moving towards position $1$, we find the right-most digit of the restricted growth string that can be incremented without violating the conditions of Equation~\ref{eq:conditions-algorithm-a}. The one-based index $c$ that will be incremented is, thus, the first to satisfy both $a_c < n - 1$ and $a_c \le b_c$, otherwise after the increment we would have an invalid partition. After finding such index $c$, we move in the opposite direction from $c+1$ to $n$ and zero out the digits (since we have performed an increment on a more significant digit) while adjusting the values of $b_i = \text{max}(a_{i-1}, b_{i-1})$. The process is repeated until no index can be incremented (except the first). More formally, the \texttt{next} method of this algorithm is described by the following steps (one-based indices notation). \begin{description} \item [Algorithm V] Accepts the array $a$ representing the previous partition and the auxiliary state variable $b$ representing the prefix maxima of array $a$. Changes the array $a$ in-place to point to the next partition and updates the array $b$ to reflect the changes. \item [V1] Set [$c \gets n$]. \item [V2] While [$a_c = n - 1$ Or $a_c > b_c$] Do [$c \gets c - 1$]. \item [V3] If [$c = 1$] Report the end of enumeration. \item [V4] Set [$a_c \gets a_c + 1$]. \item [V5] For [$i$] From [$c + 1$] To [$n$], execute the following steps: \begin{description} \item [V5.1] Set [$a_i \gets 0$]. \item [V5.2] Set [$b_i \gets \text{max}(a_{i-1},b_{i-1})$].~~~$\blacksquare$ \end{description} \end{description} The variable $a$ can be implemented using either an array or a linked list and there is theoretically no benefit in using either of these data structures since random access is not required. In practice, however, arrays are preferable due to their ubiquity, smaller overhead and cache efficiency. The same arguments can be made for the $b$ array. We estimate the complexity of Algorithm V via a recurrence formula. Assume $f(n)$ is the average number of steps (array changes) involved in a full enumeration of all partitions, which is equal to the total number of steps over $\mathcal{B}_n$. The total number of steps is equal to the number of advances where only the least significant digit changes plus the complexity of the rest of the cases, which can be reduced to the complexity of the same problem with $n-1$ elements where all steps are increased by 1 (because the least significant digit has to be traversed). The cases of the latter scenario are then $\mathcal{B}_{n-1}$, via which we can also say that the number of least significant digit advances are $\mathcal{B}_n - \mathcal{B}_{n-1}$ because both cases need to sum to $\mathcal{B}_n$. Thus, \begin{align*} f(n) &= \frac{\mathcal{B}_n - \mathcal{B}_{n-1} + \mathcal{B}_{n-1} (f(n-1)+1)}{\mathcal{B}_n} \\ &= 1 - \frac{\mathcal{B}_{n-1}}{\mathcal{B}_{n}} + \frac{\mathcal{B}_{n-1}}{\mathcal{B}_{n}} f(n-1) + \frac{\mathcal{B}_{n-1}}{\mathcal{B}_{n}}. \end{align*} Considering Equation~\ref{eq:asymptotic}, $f(n)$ can asymptotically be written as \begin{equation*} f(n) \sim 1 + \frac{\ln n}{n} f(n-1) \end{equation*} and it can easily be shown that $f(n)$ is decreasing and $\lim_{n \to \infty} f(n) = 1$. Hence, the amortized number of primitive array operations are $\Theta(1)$. Another perspective of this result is that the probability of the \texttt{next} method returning after a single change to the least significant digit is $1-\ln n / n$, which increases with $n$ and tends towards 1. Hence, it becomes increasingly likely that the \texttt{next} method will only have to perform a single change to the $a$ array. The extra space used is the $b$ array which corresponds to $\Theta(n)$ space requirement. It can be easily shown that the worst case complexity of the \texttt{next} method is linear in respect to $n$ and occurs for $c=1$, in which case the $a$ array will be traversed twice. \section{Problem B} \label{sec:problem-b} Problem B consists in enumerating the partitions of a set with at most $k$ blocks. As a result, Problem A can reduce to Problem B if we set $k=n$. However, the enumerations of problem B are a subset of those of Problem A and, thus, it would appear that Problem B can simply be solved by using Algorithm V and filtering only the $k$-partitions. Such filtering can be performed in constant time; the quantity $\text{max}(a_n,b_n)$ shows how many blocks the current partition has. However, despite the filtering process being constant, this algorithm would still not have constant amortized advance, as shown in the following paragraph. Suppose such algorithm exists and solves Problem B. Then, the iterations that would have to be performed are $\mathcal{B}_n$, and the enumerations would be $\mathcal{S}^+(n,k)$. Thus, the amortized number of steps required to advance the enumeration is \begin{equation*} \mathcal{T}_k(n) = \frac{\mathcal{B}_n}{\mathcal{S}^+(n,k)}. \end{equation*} Hence, due to Equations~\ref{eq:asymptotic} and~\ref{eq:asymptotic-2} \begin{align*} \frac{\mathcal{T}_k(n+1)}{\mathcal{T}_k(n)} &=\frac{\mathcal{B}_{n+1}}{\mathcal{B}_n} \cdot \frac{\mathcal{S}^+(n,k)}{\mathcal{S}^+(n+1,k)} \\ &\sim \frac{n}{\ln n} \cdot \frac{1}{k}, \end{align*} which is increasing with $n$ and tends towards infinity. Hence, $\mathcal{T}_k(n)$ cannot represent constant time advance as in the latter case the fraction should tend to a constant. As a result, Problem B cannot be solved in constant time via filtering the partitions of Algorithm V. Instead, we perform a simple modification to Algorithm V that guarantees that there will never be any partition with more than $k$ blocks. In the terminology of the restricted growth strings, the constraint can be imprinted as the absence of any number that is greater or equal to $k$. Such algorithm is very simple to design with a small modification on step V2: \begin{description} \item [Algorithm W] Accepts the array $a$ representing the previous partition, the setting $k$ that represents the \textit{maximum} number of blocks in any partition, and the auxiliary state variable $b$ representing the prefix maxima of array $a$. Changes the array $a$ in-place to point to the next partition and updates the array $b$ to reflect the changes. No change is performed on $k$. \item [W1] Set [$c \gets n$]. \item [W2] While [$a_c = k-1$ Or $a_c > b_c$] Do [$c \gets c - 1$]. \item [W3] If [$c = 1$] Report the end of enumeration. \item [W4] Set [$a_c \gets a_c + 1$]. \item [W5] For [$i$] From [$c + 1$] Up To [$n$], execute the following steps: \begin{description} \item [W5.1] Set [$a_i \gets 0$]. \item [W5.2] Set [$b_i \gets \text{max}(a_{i-1},b_{i-1})$].~~~$\blacksquare$ \end{description} \end{description} The only step that has changed is V2 where the condition $a_c = n-1$ is transformed into $a_c = k-1$. This change guarantees that the maximum number of blocks on any partition will be $k$, while maintaining the lexicographic order of the partitions. Although Algorithm W is only a slight modification of Algorithm V, the complexity analysis differs. The overall number of iterations is $\mathcal{S}^+(n,k)$ and in each one of them, we perform $n-c$ steps. The probability of performing more than 1 step is $\mathcal{S}^+(n-1,k)/\mathcal{S}^+(n,k)$. Thus, for the average number of steps $f(n)$ it holds that \begin{align*} f(n) &= \frac{ \mathcal{S}^+(n,k) - \mathcal{S}^+(n-1,k) + \mathcal{S}^+(n-1,k) \cdot (f(n-1) + 1) }{\mathcal{S}^+(n,k)} \\ &= 1 - \frac{1}{k} + \frac{1}{k} \cdot f(n-1) + \frac{1}{k} = 1 + \frac{1}{k} \cdot f(n-1) \\ &= 1 + \frac{1}{k} + \frac{1}{k^2} + \dots + \frac{1}{k^n} \sim 1 + \frac{1}{k-1}, \end{align*} which is independent of $n$ and, as a result, dictates a $\Theta(1)$ amortized advance. An interesting observation is that the average number of array operations of Algorithm W tends towards $1+1/(k-1)$ which, unlike Algorithm V, is higher than 1 for finite $k$. Experimental observations suggest that the number of steps converge rapidly, even for $n<15$. If $k=n$, then the number of steps performed by Algorithm W tends towards 1 and reduces to Algorithm V. Hence, Algorithm W becomes faster (in respect to the average \texttt{next} call) as $k$ increases, with the amortized number of steps per call to the \texttt{next} method ranging from 1 in the best case to 2 in the worst case ($k=2$). The extra memory of Algorithm W is $\Theta(n)$ due to the $b$ array. Similar to Algorithm V, the worst case complexity of the \texttt{next} method is also linear in respect to $n$ as the $a$ array is traversed twice. \section{Problem C} \label{sec:problem-c} Problem C consists in enumerating the partitions of a set with exactly $k$ blocks. Similar to Algorithm W, an algorithm for Problem C would be inefficient by a simple filtering of Algorithm V, despite the enumerations of Problem C being a subset of the enumerations of Problem A. The iterations that would have to be performed are $\mathcal{B}_n$, and the enumerations would be $\mathcal{S}(n,k)$. Thus, the amortized number of steps required to advance the enumeration is \begin{equation*} \mathcal{T}_k(n) = \frac{\mathcal{B}_n}{\mathcal{S}(n,k)}. \end{equation*} Hence, due to Equation~\ref{eq:asymptotic} \begin{align*} \frac{\mathcal{T}_k(n+1)}{\mathcal{T}_k(n)} &=\frac{\mathcal{B}_{n+1}}{\mathcal{B}_n} \cdot \frac{\mathcal{S}(n,k)}{\mathcal{S}(n+1,k)} \\ &\sim \frac{n}{\ln n} \cdot \frac{k^n}{k^{n+1}} = \frac{1}{k} \cdot \frac{n}{\ln n}, \end{align*} which is increasing with $n$ and tends towards infinity. Hence, $\mathcal{T}_k(n)$ cannot represent constant time advance. A constant time filtering via Algorithm W, however, appears to be possible; this property will allow us to design a constant time algorithm for Problem C. Consider, for example, a modification to Algorithm W that filters only the $k$-partitions (i.e., only those that have the number $k$ in the restricted growth string). The filtering operation is trivial to design in constant time by leveraging the maximum value of the $a$ array. Hence, the iterations of this algorithm would be $\mathcal{S}^+(n,k)$ and the enumerations of Problem C are $\mathcal{S}(n,k)$. As a result, the amortized number of steps required to advance the enumeration is \begin{equation*} \mathcal{T}_k(n) = \frac{\mathcal{S}^+(n,k)}{\mathcal{S}(n,k)}. \end{equation*} It can be shown that $\mathcal{T}_k(n)$ is asymptotically a constant time operation by considering that \begin{equation*} \frac{\mathcal{S}(n,k)}{\mathcal{S}(n,k-1)} = \left( \frac{k}{k-1} \right) ^ n \cdot \frac{1}{k}, \end{equation*} which approaches infinity as $n \to \infty $. Thus, $\mathcal{S}(n,k)$ becomes infinitely bigger than $\mathcal{S}(n,k-1)$ and, therefore, all the mass of the $\mathcal{S}^+(n,k)$ sum is being concentrated to its last term $\mathcal{S}(n,k)$. Therefore, for a fixed $k$: \begin{equation}\label{eq:limit-striling-sum} \mathcal{S}^+(n,k) \sim \mathcal{S}(n,k), \end{equation} which leads to $\mathcal{T}_k(n) = \Theta(1)$. This process implies the existence of Algorithm X which can be developed by simply filtering the partitions returned by Algorithm W and asymptotically perform $1+1/(k-1)$ operations and an extra two operations for the filtering process: the access of $a_n$ and $b_n$. Naturally, Algorithm X also enumerates the partitions in their lexicographic order. \begin{description} \item [Algorithm X] Accepts the array $a$ representing the previous partition, the setting $k$ that represents the \textit{exact} number of blocks in any partition, and the auxiliary state variable $b$ representing the prefix maxima of array $a$. Changes the array $a$ in-place to point to the next partition and updates the array $b$ to reflect the changes. No change is performed on $k$. \item [X1] Set [$m \gets 0$]. \item [X2] Execute the following statements While [$m \ne k$]. \begin{description} \item [X2.1] Invoke Algorithm W. \item [X2.2] Set [$m \gets \text{max}(a_n, b_n)$].~~~$\blacksquare$ \end{description} \end{description} The algorithm includes the $m$ variable that keeps track of the maximum number in the $a$ array which corresponds to the number of blocks in the next partition. The initial value of the $a$ array is $0^{n-k} 01 \dots (k-1)$ as this is the first lexicographic $k$-partition. As proven earlier, the amortized running time of an advance is $\Theta(1)$ and it uses extra $\Theta(n)$ memory. From the arguments leading to Equation~\ref{eq:limit-striling-sum}, it is implied that, asymptotically, step X2 will only get executed once and, as a result, the worst case complexity of Algorithm X is the same as Algorithm W, which is linear in respect to $n$. Algorithm X achieves a constant time advance via an extension to Algorithm W that skips the invalid partitions one at a time. A natural extension to Algorithm X can easily be stated by observing the pattern of partitions returned by Algorithm W. Below is a small contiguous subset for $n=7,k=3$ where the bold partitions indicate the (valid for Problem C) $3$-partitions. \begin{quote} \centering \textbf{100002} 0100010 0100011 \textbf{0100012} \textbf{0100020} \textbf{0100021} \textbf{0100022} 0100100 0100101 \textbf{0100102} 0100110 0100111 \textbf{0100112} \textbf{0100120} \textbf{0100121} \textbf{0100122} \textbf{0100200} \textbf{0100201} \textbf{0100202} \textbf{0100210} \textbf{0100211} \textbf{0100212} \textbf{0100220} \textbf{0100221} \textbf{0100222} 0101000 0101001 \textbf{0101002} 0101010 0101011 \textbf{0101012} 0101020 \textbf{0101021} \textbf{0101022} 0101100 0101101 \textbf{0101102} 0101110 0101111 \textbf{0101112} \textbf{0101120} \textbf{0101121} \textbf{0101122} \textbf{0101200} \textbf{0101201} \textbf{0101202} \end{quote} The valid partitions are not randomly distributed but instead usually come in sequences. In fact, the enumerations can be reduced to alternating sequences of valid and invalid partitions. This phenomenon is more obvious for larger values of $k$ where there are longer invalid sequences. Careful inspection of the enumerations reveals that all invalid sequences begin with a partition for which $a_n=0$, because the lexicographically previous partition has $a_n=k-1$ and is, thus, valid. Based on this observation, we define Algorithm Y as another modification of Algorithm W that skips sequences of invalid partitions instead of determining whether to visit the partitions one by one. Thus, the problem reduces to finding the next valid partition, given any invalid partition. For example, for the sequence above, the next valid partition from $0100010$ is $0100012$. The way to achieve this is to move a pointer $i$ from right to left (from LSD to MSD) and set $a_i$ as the highest number, lower than $k$, that is not in the array, until this value is found in the array. This process is depicted in the description below. \begin{description} \item [Algorithm Y] Accepts the array $a$ representing the previous partition, the setting $k$ that represents the \textit{exact} number of blocks in any partition, and the auxiliary state variable $b$ representing the prefix maxima of array $a$. Changes the array $a$ in-place to point to the next partition and updates the array $b$ to reflect the changes. No change is performed on $k$. \item [Y1] Execute Algorithm W. \item [Y2] If [$\text{max}(a_n,b_n) \ne k-1$]: \begin{description} \item [Y2.1] For [$(i,k_0)$] From [$(n,k-1)$] Down To [$(1,1)$] As Long As [$k_0 > b_i$]. \begin{description} \item [Y2.1.1] Set [$a_i \gets k_0$]. \item [Y2.1.2] Set [$b_i \gets k_0 - 1$].~~~$\blacksquare$ \end{description} \end{description} \end{description} The initial value of Algorithm Y is the same is Algorithm X. Algorithm Y performs as many steps as Algorithm W because it is easy to show that step Y2.1 will asymptotically not get executed. Specifically, step Y2.1 will be executed only if the generated partition is not a $k$-partition. The probability of a partition which is at most a $k$-partition being exactly a $k$-partition is asymptotically \begin{equation*} p = \lim_{n \to \infty} \frac{\mathcal{S}(n,k)}{\mathcal{S}^+(n,k)} = 1 \end{equation*} because of Equation~\ref{eq:limit-striling-sum}. Thus, asymptotically, the frequency at which step Y2.1 is executed is approaching 0 and, as a result, Algorithm Y performs as many steps as Algorithm W plus the comparison step in Y2. Similar to algorithm X, the worst case complexity of Algorithm Y is also linear in respect to $n$ since step Y2.1 will asymptotically not get executed. \begin{figure} \centering \begin{tikzpicture} \begin{axis}[ no markers, legend pos=north west, xlabel=$k$, ylabel=Time per \texttt{next} call, width=0.48\textwidth, ymode=log, enlargelimits=false, clip=false, axis on top, grid = major ] \addplot[very thick,black] table[x=k,y=Xpn] {comparison.csv}; \addplot[very thick,black,dashed] table[x=k,y=Ypn] {comparison.csv}; \legend{X,Y} \end{axis} \end{tikzpicture} \quad \begin{tikzpicture} \begin{axis}[ no markers, legend pos=north west, xlabel=$k$, ylabel=Time for enumeration, width=0.48\textwidth, enlargelimits=false, clip=false, axis on top, grid = major ] \addplot[very thick,black] table[x=k,y=X] {comparison.csv}; \addplot[very thick,black,dashed] table[x=k,y=Y] {comparison.csv}; \legend{X,Y} \end{axis} \end{tikzpicture} \caption{Amortized time in seconds per \texttt{next} call (left) and full enumeration (right) for Algorithms X and Y measured experimentally. The left y-axis is logarithmic for aesthetic reasons as the curve is quite steep for large $k$ in linear scale.} \label{fig:comparison} \end{figure} Experimental observations suggest that Algorithm Y is much faster than Algorithm X. Figure~\ref{fig:comparison} shows plots of their execution times in seconds for $n=16$ and values of $k$ ranging from 2 to 15. The left plot shows the amortized time per call to the \texttt{next} function while the plot to the right shows the time in seconds for the complete enumeration. The differences scale profoundly as $k$ increases and this is because the number of skips that Algorithm Y takes advantage of are dramatically increased as $k$ increases. In other terms, Algorithm X performs operations proportional to $\mathcal{S}^+(n,k)$, which is the amount of times that Algorithm W has to be invoked, while Algorithm Y only performs operations proportional to $\mathcal{S}(n,k)$. The difference between these two quantities increases with $k$ because Stirling numbers of the second kind are decreasing after their peak for specific $n$ (which is $k=7$ in this case), despite both algorithms running in constant amortized time. For this reason, we recommend Algorithm Y over Algorithm X despite them both having constant complexity. \section{Problem D} \label{sec:problem-d} Problem D is the generalization of problems A, B and C. It consists in enumerating the partitions of a set, whose number of blocks are between $k_{min}$ and $k_{max}$. For this problem, we define a new algorithm, Algorithm Z, that builds upon the concept of Algorithm Y where we skip the whole invalid sequence before generating the next partition. Instead of backtracking, however, after an invalid partition has been generated (Y2.1), Algorithm Z has a return process on par with step W5 which is split into two parts and, thus, no backtracking is necessary. The return process of Algorithm Z aims at replacing the subarray $a_{c+1} a_{c+2} \dots a_n$ with the lexicographically smallest string that guarantees a $k_{min}$-partition. The first part of the return process consists in filling the longest possible sequence with zeroes (Z7). For the second part of the return, Algorithm Z fills the remaining places so that $a_n=k_{min}-1$, $a_{n-1}=k_{min}-2$ or simply $a_{n-i}=k_{min}-i-1$ (Z8). \begin{description} \item [Algorithm Z] Accepts the array $a$ representing the previous partition, the setting $k_{min}$ that represents the \textit{minimum} number of blocks in any partition, the setting $k_{max}$ that represents the \textit{maximum} number of blocks in any partition, and the auxiliary state variable $b$ representing the prefix maxima of array $a$. Changes the array $a$ in-place to point to the next partition and updates the array $b$ to reflect the changes. No change is performed on $k_{min}$ or $k_{max}$. \item [Z1] Set [$c \gets n$]. \item [Z2] While [$a_c = k_{max} - 1$ Or $a_c > b_c$] Do [$c \gets c - 1$]. \item [Z3] If [$c = 1$] Report the end of enumeration. \item [Z4] Set [$a_c \gets a_c + 1$]. \item [Z5] Set [$b_{c+1} \gets \text{max}(a_c, b_c)$]. \item [Z6] Set [$z \gets b_{c+1} + n - c - k_{min}$]. \item [Z7] Execute [$z$] times and as long as $c \le n$: \begin{description} \item [Z7.1] Set [$a_c \gets 0$]. \item [Z7.2] Set [$b_{c+1} \gets b_c$]. \item [Z7.3] Set [$c \gets c + 1$]. \end{description} \item [Z8] For [$i$] From [$c$] To [$n$], execute the following steps: \begin{description} \item [Z8.1] Set [$a_i \gets b_i + 1$]. \item [Z8.2] Set [$b_{i+1} \gets b_i$].~~~$\blacksquare$ \end{description} \end{description} The algorithm includes an extra index in the $b$ array for convenience; it's not utilized but it prevents unnatural \textit{if} statements during steps Z7.2 and Z8.2 to check for array boundaries. Steps Z1-Z5 are performed in the same way as Algorithm W as the algorithm seeks the right-most element that can be incremented. During step Z7, the $z$ digits that follow are set to 0 after the $z$ value is being calculated in step Z6. This value is computed by considering what the value of $a_n$ would be if we kept increasing the $a_c$ values until the end of the array. If this value is higher than $k_{min}$, it means that we can still put an extra 0, as our goal, in the worst case, is to terminate with $a_n=k_{min}-1$. If $z$ is zero, step Z7 is skipped and step Z8 is executed immediately. The value $z$ cannot be negative as that would mean that creating a partition that has over $k_{min}$ number of blocks is impossible, which contradicts the fact that the previous partition was at least a $k_{min}$-partition. Finally, during step Z8 we fill up the rest of the array so that $a_i = k_{min}-i-1$. In conclusion, steps Z7 and Z8 guarantee that the next partition formed by the digits $c+1$ through $n$ will be the lexicographically smallest that guarantees a $k_{min}$-partition. The initial value of the $a$ array is $0^{n-k_{min}} 012 \dots (k_{min}-1)$ as this is the first lexicographic $k_{min}$-partition. A diagram of the operation of Algorithm Z is shown in Figure~\ref{fig:algorithm-z}. \begin{figure} \centering \small \begin{tikzpicture} \draw [line width=0.5mm] (-6,-0.4) -- (-6,0.4) -- (6,0.4) -- (6,-0.4) -- cycle; \draw [line width=0.25mm] (-3,0.4) -- (-3,0.3); \draw [line width=0.25mm] (-3,-0.4) -- (-3,-0.3); \draw [line width=0.25mm] (0,0.4) -- (0,0.3); \draw [line width=0.25mm] (0,-0.4) -- (0,-0.3); \node [inner sep=0cm] (A) at (6,0.7) {}; \node [inner sep=0cm] (B) at (-3,0.7) {}; \node [inner sep=0cm] (C) at (-3,-0.7) {}; \node [inner sep=0cm] (D) at (0,-0.7) {}; \node [inner sep=0cm] (E) at (6,-0.7) {}; \node [inner sep=0.5mm] (F) at (-6,0) {}; \node [inner sep=0cm] (G) at (-3,0) {}; \node [inner sep=0cm] (H) at (0,0) {}; \node [inner sep=0.5mm] (I) at (6,0) {}; \path [<->,line width=0.25mm] (F) edge [] node [fill=white] {$c$} (G); \path [<->,line width=0.25mm] (G) edge [] node [fill=white] {$z$} (H); \path [<->,line width=0.25mm] (H) edge [] node [fill=white] {$n-c-z$} (I); \path [->,line width=0.25mm] (A) edge [bend right=8, above, dashed] node {Z2} (B); \path [->,line width=0.25mm] (C) edge [bend right=20, below, dashed] node {Z7} (D); \path [->,line width=0.25mm] (D) edge [bend right=10, below, dashed] node {Z8} (E); \end{tikzpicture} \caption{The operation of Algorithm Z. Initially, step Z2 searches for the largest incrementable index $c$. Direction is then switched to continue with step Z7 that fills the region with $z$ zeroes. During the final step Z8, the algorithm fills the remaining places with the lexicographically minimum string that guarantees a $k_{min}$-partition.} \label{fig:algorithm-z} \end{figure} An interesting property of Algorithm Z is that, in certain cases, step Z8 is not required. Skipping step Z8 will result in a restricted growth string that doesn't respect Equation~\ref{eq:restricted-growth-string} and, as a result, it would appear that such a modification would result in wrong enumeration. This is, however, not always the case as Algorithm Z never returns isomorphic partitions, even with the absence of step Z8, for example it cannot return both 00100023 and 00100067. Thus, when the growth string is transformed into a partition (for example an object of a \texttt{Partition} class), these two growth strings are likely to be pointing to equal objects; this property also depends on the actual transformation implementation. Although this property, if leveraged, can lead to faster code, it is not recommended to do so as it will give rise to new issues and the benefits are often outweighed by complications, for example the possible unchecked increase of values in the $a$ array. Algorithm Z can be analyzed in the same way as the rest of the algorithms. For convenience, we define \begin{equation*} \mathcal{Q}(n,u,v) = \sum_{k=u}^v \mathcal{S}(n,k). \end{equation*} Then the probability of performing more than one step is \begin{equation*} q = \frac{\mathcal{Q}(n-1,k_{min},k_{max})}{\mathcal{Q}(n,k_{min},k_{max})} = \frac{\mathcal{S}(n-1,k_{max})}{\mathcal{S}(n,k_{max})} = \frac{1}{k_{max}} \end{equation*} for the same arguments that allowed us to formulate Equation~\ref{eq:asymptotic-2} and Equation~\ref{eq:limit-striling-sum}. Given that $k_{min}$ does not have an impact on the probability $q$ we might remove it: \begin{equation*} q = \frac{\mathcal{Q}(n-1,k_{max})}{\mathcal{Q}(n,k_{max})} = \frac{1}{k_{max}}. \end{equation*} Thus, \begin{align*} f(n) &= \frac{\mathcal{Q}(n,k_{max})-\mathcal{Q}(n-1,k_{max})+\mathcal{Q}(n-1,k_{max}) \cdot (f(n-1)+1)}{\mathcal{Q}(n,k_{max})} \\ &= 1 - \frac{1}{k_{max}} + \frac{1}{k_{max}} \cdot f(n-1) + \frac{1}{k_{max}} = 1 + \frac{1}{k_{max}} \cdot f(n-1). \end{align*} This is the same result as Algorithm W. Thus, the average number of operations that Algorithm Z performs are proportional to \begin{equation*} 1 + \frac{1}{k_{max}-1}, \end{equation*} which is independent of $n$ and implies that algorithm Z runs in constant amortized time. Initially, we removed $k_{min}$ because it was inconsequential, considering the reasoning of Equation~\ref{eq:asymptotic-2} and Equation~\ref{eq:limit-striling-sum}: in a sum of contiguous Stirling numbers of the second kind, only the highest term is relevant in an asymptotic perspective. Another way to look at the absence of $k_{min}$ is because as $n$ increases, the percentage of partitions with smaller number of blocks become less common. For example for $k_{min}=2$ and $k_{max}=3$, the percentages of 2-partitions in the enumeration as $n$ goes from 3 to 10 are 0.75, 0.54, 0.38, 0.26, 0.17, 0.12, 0.08, 0.05. Similar to the previous algorithms, the worst case complexity of Algorithm Z occurs when $c$ is lowest, in which case the $a$ array is traversed twice. Before we close this section, we would like to make a short mention to the problem of lexicographically enumerating the partitions of a set in reverse order. Sometimes the requirements of a problem dictate the enumeration be executed in reverse order; such operation can easily be produced from the existing algorithms mentioned in this work. As an example, Algorithm Z can be transformed to an algorithm that enumerates the partitions in reverse lexicographic order by inverting its logic. Specifically, during step Z2, instead of searching for the first index that can be incremented, we will search for the first index that can be decremented; this is the first index that is not 0 and at the same time, after its reduction, there is enough ``space'' to reach $k_{min}$ at the end of the string. During the respective steps Z7 and Z8, we create the lexicographically maximum string by increasing the values of $a_n$ until $k_{max}-1$ has been reached, and fill the remaining places with $k_{max}-1$. The reverse version of Algorithm Z is formulated below. \begin{description} \item [Algorithm Z*] Accepts the array $a$ representing the next partition, the setting $k_{min}$ that represents the \textit{minimum} number of blocks in any partition, the setting $k_{max}$ that represents the \textit{maximum} number of blocks in any partition, and the auxiliary state variable $b$ representing the prefix maxima of array $a$. Changes the array $a$ in-place to point to the previous partition and updates the array $b$ to reflect the changes. No change is performed on $k_{min}$ or $k_{max}$. \item [Z*1] Set [$c \gets n$]. \item [Z*2] While [($c > 1$) And ($a_c = 0$ Or $k_{min} - b_i > n - i$)] Do [$c \gets c - 1$]. \item [Z*3] If [$c = 1$] Report the end of enumeration. \item [Z*4] Set [$a_c \gets a_c - 1$]. \item [Z*5] Set [$b_{c+1} \gets \text{max}(a_c, b_c)$]. \item [Z*6] For [$i$] From [$c+1$] To [$n$] As Long As [$b_i < k_{max}-1$]: \begin{description} \item [Z*6.1] Set [$a_i \gets b_i + 1$]. \item [Z*6.2] Set [$b_{i+1} \gets a_i$]. \end{description} \item [Z*7] For [$i$] From [$c$] To [$n$], execute the following steps: \begin{description} \item [Z*7.1] Set [$a_i \gets k_{max} - 1$]. \item [Z*7.2] Set [$b_i \gets k_{max} - 1$].~~~$\blacksquare$ \end{description} \end{description} It is trivial to see that Algorithm Z* does the same number of steps as Algorithm Z and is bound by the same number of operations. Algorithms V, W, X and Y can easily be reversed using the same logic too but, as previously argued, this is not deemed necessary as Algorithm Z* supersedes them. \section{Problem E} \label{sec:problem-e} Problem E is the last problem that we define in this document and the most general one; any of the Problems A -- D can be reduced to Problem E and any algorithm that solves Problem E can also solve any other problem by simple adjustment of the parameter $K$. It consists in enumerating the partitions of a set, whose number of blocks are present in the set $K=\{k_1,k_2,k_3,\dots\}$. If the values of $k_i$ themselves are continuous, for example $\{k_{min},k_{min}+1,k_{min}+2,\dots,k_{max}\}$, then the problem is equivalent to Problem D for parameters $k_{min}$ and $k_{max}$. Using the methods mentioned in Sections~\ref{sec:problem-b} and~\ref{sec:problem-c}, it can be shown that Problem E can easily be solved using Algorithm Z in constant amortized time by simple filtering of the enumerated partitions. This can be achieved using an additional boolean array $c$, where $c_i$ shows whether $i$ exists in the $K$ set, and can be created in time independent of $n$. Once again, however, in this section we present an algorithm for solving Problem E without backtracking, using a single (two-way) pass over the $a$ array; we call this Algorithm U. Initially, regarding the parameter $K$, we sort it's array representation in ascending order so that $K=[k_{min},\dots,k_{max}]$. Then, we create the additional array $m$, where $m_i$ is equal to the lowest value in $K$ that is greater or equal to $i$. As a result, $m$ needs to hold at least $k_{max}$ elements. For example, if $K=[2,5]$, then $m=[2,2,5,5,5]$. This type of initialization is performed in linearithmic time in respect to $k_{max}$, which is independent of $n$ and, hence, the impact of the initialization in the amortized cost of each \texttt{next} call is zero. If $k_{max}$ is known in advance, it can also be done in linear time. The difference of Problem E with Problem D can be shown with an example. Consider $K=[2,5]$ and the partition 011111. The next partition cannot be 011112 because it is a 3-partition and 3 is not in the $K$ set. Incrementing the second digit is also not possible because that would lead to 01112{-} and there is no suitable value for the LSD that leads to a valid partition; the minimum is 0 which leads to a 3-partition and the maximum is 3 which leads to a 4-partition. It can be shown that incrementing the third digit is possible in this case as it leads to 0112{-}{-}, which can hold a valid partition, e.g., the 5-partition 011234. By considering the previous example we can induce a rule that dictates the first digit $a_c$ that can be incremented. After the index $c$ takes the value $a_c+1$, it is clear that the next partition cannot have less number of blocks than $t=\text{max}(a_1,a_2,\dots,a_c)+1$ and, as a result, the algorithm, using the $m_t$ value, determines the least amount of blocks that can be created; this process guarantees that the partition $a_{c+1} a_{c+2} \dots a_n$ will be either lexicographically minimal or invalid. The partition would be invalid if the remaining digits ($n-c$) are not enough to trigger an increase from $t$ number of blocks to $m_t$ number of blocks. In essence, the partition is (valid) lexicographically minimal if $n-c \ge m_t-t$. If the partition is invalid, the algorithm tries the next $c$ (the $c-1$ digit). In the edge case where $m_t=t$, it is implied that the single increment on the index $c$ results in a valid number of blocks and as a result the rest of the digits to the right can be filled with zeroes ($m_t-t=0$). Algorithm U is formulated below. \begin{description} \item [Algorithm U] Accepts the array $a$ representing the previous partition, the auxiliary state variable $b$ representing the prefix maxima of array $a$ and the sorted arrays $k$ and $m$. Changes the array $a$ in-place to point to the next partition and updates the array $b$ to reflect the changes. No change is performed on $k$ or $m$. \item [U1] Set [$c \gets n + 1$], [$k_{max} = \text{max}(k)$], [$k_{min} = \text{min}(k)$]. \item [U2] Do the following While [$a_c = k_{max} - 1$ Or $a_c > b_c$ Or $m_{t+1}-t>n-c$] \begin{description} \item [U2.1] Set [$c \gets c - 1$]. \item [U2.2] Set [$t \gets \text{max}(a_c + 1, b_c)$]. \end{description} \item [U3] If [$c = 1$] Report the end of enumeration. \item [U4] Set [$a_c \gets a_c + 1$]. \item [U5] Set [$b_{c+1} \gets \text{max}(a_c, b_c)$]. \item [U6] Set [$z \gets b_{c+1} + n - c - m_{b_{c+1}+1}$]. \item [U7] Execute [$z$] times and as long as $c \le n$: \begin{description} \item [U7.1] Set [$a_c \gets 0$]. \item [U7.2] Set [$b_{c+1} \gets b_c$]. \item [U7.3] Set [$c \gets c + 1$]. \end{description} \item [U8] For [$i$] From [$c$] To [$n$], execute the following steps: \begin{description} \item [U8.1] Set [$a_i \gets b_i + 1$]. \item [U8.2] Set [$b_{i+1} \gets a_i$].~~~$\blacksquare$ \end{description} \end{description} The process is similar to the other one pass algorithms and can be divided into 3 steps: find the index $c$ to increment (U2), switch direction and place as many zeroes as possible (U7) and, finally, fill up the rest of the $a$ array to guarantee the $m_t$-partition (U8). The steps U7 and U8 are almost identical to steps Z7 and Z8 but instead of $k_{min}$ we consider $m_t$ as shown earlier. The number of enumerations performed in Problem E are \begin{equation*} \sum_{k \in K} \mathcal{S}(n,k), \end{equation*} while in each one the probability of $c$ not being the LSD is \begin{equation*} \frac{\sum_{k \in K} \mathcal{S}(n-1,k)}{\sum_{k \in K} \mathcal{S}(n,k)} \end{equation*} which for the reasons explained in Equations~\ref{eq:asymptotic-2} and~\ref{eq:limit-striling-sum} is asymptotically equal to \begin{equation*} \frac{\mathcal{S}(n-1,\text{max}(K))}{\mathcal{S}(n,\text{max}(K))} = \frac{1}{\text{max}(K)} \end{equation*} and equivalent to the respective probability of Algorithm Z. As a result, the number of operations are asymptotically proportional to \begin{equation*} \frac{1}{\text{max}(k)-1} = \frac{1}{k_{max}-1}. \end{equation*} Furthermore, Algorithm U uses arrays of maximum size $n$ and, thus, the extra memory required is proportional to $n$. It is easy to show that the maximum number of steps of the \texttt{next} method of Algorithm U is also linear in respect to $n$. Regarding the reverse lexicographic version of Algorithm U, it is easy to produce Algorithm U* with the extra parameter $r$ which is, in essence, the inverted $m$ array; the value $r_i$ shows the maximum $k$ which is less than or equal to $i$: \begin{description} \item [Algorithm U*] Accepts the array $a$ representing the previous partition, the auxiliary state variable $b$ representing the prefix maxima of array $a$ and the sorted arrays $k$, $m$ and $r$. Changes the array $a$ in-place to point to the next partition and updates the array $b$ to reflect the changes. No change is performed on $k$, $m$ or $r$. \item [U*1] Set [$c \gets n + 1$], [$k_{max} = \text{max}(k)$], [$k_{min} = \text{min}(k)$]. \item [U*2] Do the following While [$a_c = 0$ Or $m_{t+1}-t>n-c$] \begin{description} \item [U*2.1] Set [$c \gets c - 1$]. \item [U*2.2] Set [$t \gets \text{max}(a_c - 1, b_c)$]. \end{description} \item [U*3] If [$c = 1$] Report the end of enumeration. \item [U*4] Set [$a_c \gets a_c - 1$]. \item [U*5] Set [$b_{c+1} \gets \text{max}(a_c, b_c)$]. \item [U*6] Set [$z \gets r_{b_{c+1} + n - c}$]. \item [U*7] Set [$c=c+1$]. \item [U*8] While [$b_i < z - 1$] \begin{description} \item [U*8.1] Set [$a_c \gets b_c + 1$]. \item [U*8.1] Set [$b_{c+1} \gets a_c$]. \end{description} \item [U*9] For [$i$] From [$c$] To [$n$], execute the following steps: \begin{description} \item [U*9.1] Set [$a_i \gets z-1$]. \item [U*9.2] Set [$b_{i+1} \gets z-1$].~~~$\blacksquare$ \end{description} \end{description} In conclusion, Algorithm U is the most general algorithm for solving the set partition enumeration problem. Its parameter $K$ dictates the desired number of blocks of the partitions and it can contain arbitrary values in $[1,n]$. It operates in constant amortized time per \texttt{next} call using, in the worst case, a single two-way pass over the restricted growth string array and is optimal under this condition. \section{Summary} This document brings attention to fundamental lexicographic set partition enumeration problems in a perspective that is suitable for learners and programmers. We approach the presentation of the problems in a progressive way that is easy to understand, and thoroughly explain the algorithms used to solve them. We show how these algorithms work and demonstrate their amortized constant time complexity. These algorithms are straightforward to implement in modern, high-level programming languages and don't have any complicated \textit{jump} logic or recursive structures. Table~\ref{tab:algorithms} presents a summary of the algorithms discussed in this document. Reference implementations of these algorithms in C++ are available online\footnote{https://github.com/gstamatelat/partitions-enumeration}. \begin{table} \centering \setstretch{0.5} \begin{tabular}{cccccc} \hline\\ Algorithm & Problem & Time & Space & Order & Iterations \\\\ \hline\\ V & A & $\displaystyle\sim 1$ & $\sim n$ & Lexicographic & $\displaystyle\mathcal{B}_n$ \\\\ W & B & $\displaystyle\sim \left( 1 + \frac{1}{k-1} \right)$ & $\sim n$ & Lexicographic & $\displaystyle\mathcal{S}^+(n,k)$ \\\\ X & C & $\displaystyle\sim \left( 1 + \frac{1}{k-1} \right)$ & $\sim n$ & Lexicographic & $\displaystyle\mathcal{S}(n,k)$ \\\\ Y & C & $\displaystyle\sim \left( 1 + \frac{1}{k-1} \right)$ & $\sim n$ & Lexicographic & $\displaystyle\mathcal{S}(n,k)$ \\\\ Z & D & $\displaystyle\sim \left( 1 + \frac{1}{k_{max}-1} \right)$ & $\sim n$ & Lexicographic & $\displaystyle\sum_{k=k_{min}}^{k_{max}} \mathcal{S}(n,k)$ \\\\ Z* & D & $\displaystyle\sim \left( 1 + \frac{1}{k_{max}-1} \right)$ & $\sim n$ & Reverse Lex. & $\displaystyle\sum_{k=k_{min}}^{k_{max}} \mathcal{S}(n,k)$ \\\\ U & E & $\displaystyle\sim \left( 1 + \frac{1}{\text{max}(K)-1} \right)$ & $\sim n$ & Lexicographic & $\displaystyle\sum_{k \in K} \mathcal{S}(n,k)$ \\\\ U* & E & $\displaystyle\sim \left( 1 + \frac{1}{\text{max}(K)-1} \right)$ & $\sim n$ & Reverse Lex. & $\displaystyle\sum_{k \in K} \mathcal{S}(n,k)$ \\\\ \hline \end{tabular} \caption{Summary of the algorithms discussed in this work.} \label{tab:algorithms} \end{table} This work can be extended by including various algorithms for the generation of random set partitions; in the general case a random partition with the required number of blocks inside an integer set $K$. Another line of future work is the restatement of Problems A, B, C, D and E for gray code enumerations instead of lexicographic. \bibliographystyle{apalike}

https://arxiv.org/abs/0905.2563

Stationary map coloring

We consider a planar Poisson process and its associated Voronoi map. We show that there is a proper coloring with 6 colors of the map which is a deterministic isometry-equivariant function of the Poisson process. As part of the proof we show that the 6-core of the corresponding Delaunay triangulation is empty.Generalizations, extensions and some open questions are discussed.

\section{Introduction} The Poisson-Voronoi map is a natural random planar map. Being planar, a specific instance can always be colored with 4 colors with adjacent cells having distinct colors. The question we consider here is whether such a coloring can be realized in a way that would be isometry-equivariant, that is, that if we apply an isometry to the underlying Poisson process, the colored Poisson-Voronoi map is affected in the same way. In other words, can a Poisson process be equivariantly extended to a colored Poisson-Voronoi map process? How many colors are needed? Can such an extension be deterministic? Extension of spatial processes, particularly of the Poisson process, have enjoyed a surge of interest in recent years. The general problem is to construct in the probability space of the given process, a richer process that (generally) contains the original process. Notable examples include allocating equal areas to the points of the Poisson process \cite{HP05, HHP06,HHP09,Kri,grav1,grav2}; matching points in pairs or other groups \cite{HP03, DM06, HPPS09, AGH09}; thinning and splitting of a Poisson process \cite{HLS09, AHS09}. Coloring extensions of i.i.d.\ processes on $\mathbb{Z}^d$ are considered in \cite{BHSW09}. We now proceed with formal definitions and statement of the main results. A non-empty, locally finite subset $S \subset \mathbb{R}^d$ defines a partition of $\mathbb{R}^d$, called the \emph{Voronoi tessellation}, as follows: The Voronoi cell $C(x)$ of a point $x\in S$ contains the points of $\mathbb{R}^d$ whose distance to $S$ is realized at $x$: \[ C(x) = \{z\in \mathbb{R}^d : d(z,x) = d(z,S)\}. \] Points in the intersection $C(x)\cap C(y)$ have equal distance to $x$ and $y$. It follows that the cells cover $\mathbb{R}^d$ and have disjoint interiors. For the purposes of coloring, we consider the adjacency graph $G$ of these cells, with vertices $S$ and edge $(x,y)$ if $C(x)\cap C(y)\neq\emptyset$. In the case $d=2$, this graph is called the {\em Delaunay triangulation}, and is a triangulation of the plane. (In general, this graph is the $1$-skeleton of a simplicial cover of $\mathbb{R}^d$.) A $k$-coloring of the Voronoi tessellation is a proper $k$-coloring of the Delaunay triangulation, i.e.\ an assignment of one of $k$ colors to each cell so that adjacent cells have distinct colors. Note that if $S$ does not contain four or more co-cyclic points, then no more than three cells meet at a single point. This is a.s.\ the case for the Poisson process. However, for greater generality one needs the more careful definition, where $(x,y)$ is an edge if $|C(x)\cap C(y)|>1$. This ensures that the graph is planar. \begin{figure}[t] \centering \includegraphics[width=.8\textwidth]{VM-4.eps} \caption{A proper $4$-coloring of a portion of the Poisson-Voronoi map.} \label{fig:4-color} \end{figure} Given a standard (unit intensity) Poisson process $\mathcal{P}\subset \mathbb{R}^2$, the Poisson-Voronoi map is the Voronoi map of its support. By the 4 color theorem, the Poisson-Voronoi map can always be properly colored with 4 colors. Our main question is whether it is possible to color the Poisson-Voronoi map in an isometry equivariant way and if so, how many colors are needed. To make this precise, let $\mathbb{M}$ be the space of locally finite sets in $\mathbb{R}^d$, endowed with the local topology and Borel $\sigma$-algebra.\footnote{It is also common to let $\mathbb{M}$ be the set of non-negative integer valued measures on $\mathbb{R}^d$ with $\mu(\{x\})\in\{0,1\}$. The distinction will not be important to us.} Let $\P$ be the probability on $\mathbb{M}$ which is the law of the Poisson process. Each realization $\mathcal{P} \in \mathbb{M}$ has the Delaunay graph associated with it. A (proper) $k$-coloring of $\mathcal{P}$ is a disjoint partition $\mathcal{P} = \cup_{i=1}^k \mathcal{P}_i$ such that if $x\sim y$ in the Delaunay graph of $\mathcal{P}$, then $x,y$ are not in the same $\mathcal{P}_i$. Thus the space of $k$-colored maps is a subset of $\mathbb{M}^k$. A {\em deterministic} $k$-coloring scheme of the Voronoi map is a measurable function $F:\mathbb{M} \to \mathbb{M}^k$ such that $F(\mathcal{P})$ is $\P$-a.s.\ a $k$-coloring of $\mathcal{P}$. Informally, given the point process, $F$ assigns a color to each point so that the result is a proper coloring. A {\em randomized} $k$-coloring scheme of the Voronoi map is a probability measure $\mu$ on $\mathbb{M}^k$, supported on proper $k$-colorings, such that the law under $\mu$ of $\mathcal{P} := \cup_{i=1}^k \mathcal{P}_i$ is $\P$. Given such a measure $\mu$, one may consider $\mu$ conditioned on $\mathcal{P}$. This conditional distribution is defined $\P$-a.s., and is supported on $k$-colorings of $\mathcal{P}$. Thus a randomized $k$-coloring can be interpreted as assigning to each $\mathcal{P}\in\mathbb{M}$ a probability measure on $k$ colorings of $\mathcal{P}$. Note that any deterministic coloring scheme is also a randomized one, with $\mu$ being the push-forward of $\P$ by $F$. A deterministic coloring scheme is said to be {\em isometry equivariant} if every isometry $\gamma$ of $\mathbb{R}^d$, acting naturally on $\mathbb{M}$ and $\mathbb{M}^k$, has $\gamma F(\mathcal{P}) = F(\gamma \mathcal{P})$. For randomized schemes equivariance is defined by $\mu \circ \gamma = \mu$. These definitions coincide for deterministic schemes. \begin{theorem}\label{thm:main} There exists a deterministic isometry equivariant 6-coloring scheme of the Poisson-Voronoi diagram in $\mathbb{R}^2$. \end{theorem} The requirement of determinism complicates things significantly. In contrast, we have the much simpler result \begin{prop}\label{prop:random} There exists a randomized isometry equivariant 4-coloring scheme of the Poisson-Voronoi diagram in $\mathbb{R}^2$. \end{prop} In dimensions other than 2 the problem is not as interesting. \begin{prop}\label{prop:other_dim} In $\mathbb{R}$, there is a randomized isometry equivariant coloring of the Poisson-Voronoi map with 2 colors and a deterministic one with 3 colors. In both cases this is the best possible. In $\mathbb{R}^d$ for $d>2$, the chromatic number of the Poisson-Voronoi map is a.s.\ $\infty$. \end{prop} The rest of the paper is organized as follows: In section~\ref{sec:outline} we outline the proof of \thmref{main}, and present our deterministic coloring algorithm and the two main propositions needed to prove its correctness. In Section~\ref{sec:gen} we discuss related questions: randomized colorings, dimensions other than $2$, and mention some open problems. Section~\ref{sec:proof} contains the proof of our main theorem. \section{Proof outline} \label{sec:outline} We outline the proof of Theorem~\ref{thm:main}. The idea is to find an isometry equivariant adaptation to the Voronoi map of a 6 coloring algorithm for finite planar graphs, originating in Kempe's attempted proof of the four color theorem. By Euler's formula it is known that any finite planar graph $G$ has a vertex of degree at most 5. The algorithm proceeds by iteratively removing such a vertex until the graph is empty, then putting back the vertices one by one in reverse order. As each vertex is put back into the graph, it is assigned a color distinct from those already assigned to any of its neighbors. Since a vertex has at most 5 neighbors when it is put back, this produces a proper 6 coloring. To adapt this algorithm to the Poisson-Voronoi isometry equivariant setting, one must deal with several issues. First, there exist infinitely many vertices of degree at most 5 and there is no way to pick just one of them in an isometry-equivariant way. Second, even if we iteratively remove {\em all} vertices of degree at most 5, the graph will not become empty after any finite number of steps. Finally, when returning the vertices, it is not clear in what order to do so (which may be important if some of them are neighbors). We need a way to order them which is isometry-equivariant. We overcome these issues by proving that for a Poisson-Voronoi map, the following two properties hold almost surely. Let $G=(V,E)$ be the Delaunay graph formed by the Poisson-Voronoi map. For a cell $v\in V$ write $A(v)$ for its area as a planar region. Inductively, define $G_0=G$ and $G_{i+1}$ as the graph formed from $G_i$ by removing all vertices of degree (in $G_i$) at most $5$. \begin{prop}\label{removal_prop} There exists an integer $M>0$ such that, almost surely, $G_M$ contains only finite connected components. \end{prop} \begin{prop}\label{area_prop} Almost surely, all cells have different areas and there is no infinite path in $G$ with decreasing areas. \end{prop} We now exhibit a deterministic algorithm which takes as input a graph $G=(V,E)$ with chromatic number at most $6$ and an area function $A:V\to\mathbb{R}^+$ satisfying the two propositions above and returns a proper 6-coloring of the graph. Since the algorithm only depends on the graph structure $G$ and areas $A$ which are preserved by isometries, it is clear that when applying it to the Delaunay graph of a Poisson-Voronoi map we will get a deterministic isometry-equivariant 6-coloring. The algorithm starts with all vertices uncolored. Once a vertex is colored, its color never changes. Consider first $G_M$. Each of its components is finite and hence may be colored with 4 colors in an isometry equivariant way (e.g.\ take the minimal coloring in lexicographic order, when the vertices of the component are ordered by their area). Next, having colored $G_k$, we color $G_{k-1}$ inductively. Once $G=G_0$ is colored, we are done. Consider the vertices of $G_{k-1}\setminus G_k$. Each has at most 5 neighbors in $G_{k-1}$. We order these vertices by increasing areas and wish to color them in order, i.e., coloring a vertex $v$ only after its neighbors of smaller area have been colored. The color of these neighbors is determined using the same method in an iterative manner. Proposition~\ref{area_prop} implies that there are just finitely many vertices that need to be considered before $v$ (see also Lemma~\ref{L:finite_predecessors}). Hence, going over these finitely many vertices in order of their areas, we color each one by a color which is unused by its neighbors (say, the minimal such color) until we finally color $v$. Proposition~\ref{removal_prop} is more difficult than Proposition~\ref{area_prop} and the main lemma required for its proof (Lemma~\ref{L:finite_radius_6core}) says that if we consider a square of side length $6R$ and iteratively remove vertices \emph{inside} this square having degree at most 5, then the square of side length $2R$ with the same center will eventually become empty with probability tending to $1$ as $R\to\infty$. This is shown using several probabilistic estimates and uses of Euler's formula. We then show that for well separated squares of side length $6R$, the events just described, applied to these squares, are nearly independent. A small variation on the above event (requiring that the boxes are also sealed; see below) makes separated boxes completely independent. Proposition~\ref{removal_prop} then follows by standard $k$-dependent percolation arguments. Proposition~\ref{area_prop} is proved using a similar but easier $k$-dependent percolation argument. As a corollary of the proofs of the above propositions we obtain that our coloring is finitary with exponential tails. That is, for any given point $p\in\mathbb{R}^2$, the probability that the color of the cell containing $p$ is not determined by the points of the Poisson process within a ball of radius $R$ around $p$ is at most $C e^{-c R}$ for some $C,c>0$. Note that instead of the area $A$, we could use any other parameter of the cell (e.g.\ diameter) which satisfies Proposition~\ref{area_prop} (in fact, one can relax the requirement that all cells have different areas to the requirement that adjacent cells have different areas). The sole purpose of $A$ is to induce a well founded order on cells which would ``break ties'' when putting back vertices. We chose to use the area because it is a very natural parameter to consider, but it is as easy to prove the required properties for other parameters (see Section~\ref{area_sec}). A related result is that there is no infinite path where each Poisson point is the closest to the previous one in the path \cite{KMN06}. \section{Generalizations, Extensions and Questions} \label{sec:gen} In this section we explain some variants and extensions of the question and settings discussed in our paper. \subsection{Randomized colorings} The fact that there is a randomized 4-coloring scheme of the Poisson-Voronoi map follows from the four color theorem by a soft argument. This involves an averaging consideration of ergodic theory and works for any amenable transitive space. \begin{proof}[Proof of Proposition~\ref{prop:random}] The $4$-color theorem implies existence of a measurable function $F$ (not necessarily equivariant) which assigns each Voronoi diagram a $4$-coloring. E.g. the lexicographically minimal proper coloring is easily seen to be a measurable function of the map. To get a randomized equivariant coloring, let $\tau_x$ be a translation by $x\in\mathbb{R}^2$, $\rho_\theta$ a rotation by $\theta$, and $\varepsilon$ the reflection about the $x$ axis. Let $\sigma = \tau_x \circ \rho_\theta \circ \varepsilon^u$ be a random isometry, where $u\in\{0,1\}$, $\theta\in[0,2\pi]$ and $x\in B(0,R)$ are uniform and independent. This defines a probability measure $F_R$ on $4$-colored maps by conjugating $F$ by $\sigma$. It is clear (due to compactness of the space of distributions over 4-colorings) that $\{F_R\}$ has a subsequential weak limit as $R\to\infty$, and any such limit is an isometry equivariant 4-coloring. \end{proof} \paragraph{Explicit Randomized Colorings} While the previous argument is clearly optimal with respect to the number of colors used, it is not constructive. It is instructive to consider an explicit construction with 7 colors. The construction below will be algorithmic, i.e. there is an algorithm, that determines the color of each cell by accessing a finite (but unbounded) number of cells along with a random independent bit for each cell. As a first stage, we explain how to get an 8-coloring. Start by assigning a fair coin toss to each cell independently. Consider the subgraph of $H\subset G$ where an edge is present if its endpoints have the same coin result. The connected components in this graph are components of site percolation on $G$ with $p=1/2$. By a result of Zvavitch \cite{Z96}, almost surely all connected components of both the heads and tails will be finite (in fact, Bollob{\'a}s and Riordan \cite{BR06} proved that the critical percolation threshold is indeed $p=\frac{1}{2}$). Color each ``head'' component independently with colors $\{0,1,2,3\}$ in some deterministic isometry-equivariant manner which is a function only of the cells of this component (e.g., again, a lexicographically minimal coloring with vertex order based on cell areas). Color the ``tail'' components with $\{4,5,6,7\}$. The result is a.s.\ a proper $8$-coloring of $G$. The randomness comes exclusively from the coin tosses. The color of a cell is determined by its connected component in $H$ (and the size of the corresponding cells). A trick suggested by Gady Kozma \cite{K07} reduces the number of colors required to $7$ as follows. A finite planar graph embedded in the plane has a unique unbounded face, called the{\em external} face. Attaching an additional vertex to the vertices of the external face preserves planarity. Thus a finite planar graph can be $4$-colored so vertices of the external face do not use one specified color. Now color the ``heads'' components using $\{0,1,2,3\}$ so that color $0$ does not appear at vertices of the external face of any component. Color the ``tails'' components using $\{0,4,5,6\}$ with the same constraint. Whenever two connected planar graphs are jointly embedded in the plane, one is contained in the external face of the other. Thus when a ``tails'' component is adjacent to a ``heads'' component, it is impossible for them to have adjacent vertices colored $0$, and the coloring is proper. As noted above, in order to determine the color of any cell, it is sufficient to know the map structure and the coin-tosses within a ball of a certain random radius around this cell. In addition, if one modifies the above algorithm by initially performing fair-independent rolls of a $3$-sided dice, instead of coin tosses (thus obtaining a proper 10-coloring in the final outcome, after applying Kozma's trick) then the distribution of the aforementioned radius will have exponential tails (see \cite{BR06}). The radius for our deterministic $6$-coloring also has exponential tails, as noted in the proof outline. \subsection{1-dimensional Poisson-Voronoi map} The deterministic isometry equivariant chromatic number of a graph may well be different from its usual chromatic number. For example, consider $Z^d$ translated by a uniform random variable in $[0,1]^d$ and rotated by a uniform random angle in $[0,2\pi]$. Clearly, the distribution of this random graph is isometry invariant and it is almost surely 2-colorable. Yet any deterministic isometry equivariant coloring must assign the same color to all vertices and hence cannot be proper. A different example is furnished by the 1-dimensional Poisson-Voronoi diagram, i.e., the ``Voronoi'' map composed of line segments around the points of a one-dimensional standard Poisson process. This map is 2-colorable, but we claim that its deterministic isometry equivariant coloring number is 3. First, it is seen to be at most 3 by considering the following algorithm: First color green all cells which are shorter than both their neighbors. Now, from each green cell, proceed to alternately color its neighbors to the right by red and blue, until the next green cell is reached. This produces a deterministic translation equivariant proper 3-coloring. To get an isometry equivariant coloring, instead of coloring red and blue from left to right, start from the shorter of the two green cells bounding the current stretch of uncolored cells. The following lemma states that at least 3 colors are needed. A similar argument appears in Holroyd, Pemantle, Peres and Schramm \cite{HPPS09}. \begin{lemma} \label{L:1-dim} There is no deterministic translation equivariant proper 2-coloring of the 1 dimensional Poisson-Voronoi map. \end{lemma} \begin{proof} In order to reach a contradiction, suppose $\mathcal{A}$ is such a coloring scheme. Since $\mathcal{A}$ is measurable there exists an integer $L$ and another scheme $\mathcal{B}$, such that the color $\mathcal{B}$ assigns to the cell at the origin depends only on the Poisson process in the interval $[-L,L]$ and the probability that $\mathcal{A}$ and $\mathcal{B}$ assign the same color to a given cell is at least $\frac78$. Consider also another point $x>2L$. By translation equivariance, the $\mathcal{B}$-color of the cell of $x$ is determined by the Poisson points in $[x-L,x+L]$. Hence, with probability at least $\frac34$ the $\mathcal{A}$-color of both these cells is the same as their $\mathcal{B}$-color. However, The $\mathcal{A}$-colors of these cells determine the parity of the number of cells (i.e.\ points) between them. But the parity of the number of points of the Poisson process in $[L,x-L]$ is independent of the $\mathcal{B}$-colors of the origin and of $x$, and tends to a uniform on $\{0,1\}$ as $x\to\infty$. Therefore, when $x$ is large enough there is a positive probability of a contradiction between this parity and the $\mathcal{A}$-colors of the origin and $x$, so this $\mathcal{A}$ coloring cannot exist. \end{proof} We remark that a variant of the $3$-coloring above can be used to color any invariant point process on $\mathbb{R}$ that is not an arithmetic progression (so that not all points are isomorphic). Furthermore, the proof of impossibility with 2 colors also applies to more general processes as we only use the fact that the parity of the number of points in $[L,x-L]$ is not (nearly) determined by the process in $[-L,L]$ and $[x-L,x+L]$ for $x$ large enough. \subsection{Higher dimensional Poisson-Voronoi maps} A natural generalization of our setting is to consider the 3-dimensional Poisson-Voronoi diagram. In this case it is not obvious whether one can properly color the diagram with finitely many colors even without the isometry equivariant condition. Dewdney and Vranch~\cite{DV77}, and Preparata~\cite{P77} discovered that $n$ Voronoi cells in $\mathbb{R}^3$ may be all pairwise adjacent. Indeed, \cite{DV77} shows that in $\mathbb{R}^3$, the Voronoi cells of $(x_i,x_i^2,x_i^3)_{i=1}^n$ satisfy this for any $\{x_1,\dots,x_n\}$. Since pairwise adjacency is preserved by sufficiently small perturbations, and since such configurations a.s.\ appear in the Poisson process, this implies that the chromatic number of the 3-dimensional Poisson-Voronoi diagram is almost surely infinite. Higher dimensional analogues also exist. Following Proposition~\ref{removal_prop}, one can still ask, as a weaker result than having an isometry equivariant coloring, what is the minimal $k$ such that if we iteratively remove all cells having degree at most $k$ we remain with finite components only? Such a $k$ necessarily exists by arguments similar to those of Proposition~\ref{removal_prop}. (Simulations indicate that $k=12$ may suffice in $\mathbb{R}^3$.) \subsection{Ramblings and open questions} \paragraph{Fewer colors.} Is there a deterministic $4$-coloring of the Poisson-Voronoi map? Theorem~\ref{thm:main} shows that $6$ colors suffice, while obviously at least $4$ are needed. Recent work by Adam Timar \cite{Timar} (in preparation) shows the existence of deterministic, equivariant 5-colorings using different methods. Our own methods are close to giving a $5$-coloring as well, in the following sense: Suppose we define $G_{k+1}$ by removing from $G_k$ all vertices of degree at most $4$. If Proposition~\ref{removal_prop} still holds then the same argument gives a $5$-coloring of $G_0$. To show this, it is enough to prove a statement similar to Lemma~\ref{L:finite_radius_6core} (roughly put, that the probability that a large component of the 5-core intersects the boundary of a box of size $R$ is small enough for some value of $R$). Simulations suggest that this is indeed the case. A small difficulty involved in the case of $5$ colors is that not every vertex is removed at some finite stage. Indeed, the 5-core of the Delaunay triangulation will not be empty, since it contains finite sub-graphs with minimal degree 5. The smallest such sub-graph is the dodecahedron, involving 12 vertices. Applying the same proof for 4 colors cannot work, since the 4-core of the Delaunay triangulation has an infinite component. Indeed, a vertex of degree 3 is necessarily in the interior of the triangle formed by its neighbors. It is straightforward to check that there are no infinite chains of triangles each one inside the next (since the probability of long edges decays exponentially; see also \lemref{rare_squares} below). Therefore, one can consider all the \emph{maximal} triangles in the Delaunay triangulation. This is also a triangulation of the plane, since every triangle is contained in a maximal one, and these are all disjoint. All the vertices of this triangulation also belong to $G_\infty$ (since none of them are in the interior of another triangle), and they are all in the same connected component, which is therefore infinite. Finally, while we only prove that some $G_M$ (again, deleting vertices of degree $<6$) has only finite connected components, simulations suggest that $M=2$ suffices while $M=1$ does not. In fact, it appears sufficient to delete in the second iteration roughly half the vertices of degree at most 5. Can one prove any of these assertions? \paragraph{Other properties of colorings.} If there is no deterministic $4$-coloring, one could consider intermediate properties between deterministic and unrestricted randomized colorings. For example, one may seek colorings that are ergodic, mixing, finitary, etc. Such properties were first brought to our attention by Russ Lyons \cite{L07}. \paragraph{Other planar processes.} It might be more interesting to consider other translation or isometry equivariant graph processes in the plane. These could be the Voronoi tessellation of some point process or more general planar graph processes. Except for some obvious counterexamples (see remarks before and after \lemref{1-dim}), is it true that every such process can be colored deterministically with $4$ colors? The aforementioned work of Timar \cite{Timar} shows the existence of deterministic 5-colorings. \paragraph{Hyperbolic geometry.} What can be done in the hyperbolic plane? Our argument can be adapted to give a deterministic coloring. However, the number of colors diverges as the density of the Poisson process tends to $0$, since the average degree diverges. For high enough density we can get a deterministic $6$-coloring. Is there a (deterministic or randomized) $k$-coloring with $k$ independent of the density? While the Poisson-Voronoi map is $4$-colorable by Proposition~\ref{prop:random}, our randomized constructions use amenability and fail for the hyperbolic plane. \paragraph{Prescribed color distribution.} What color distributions are achievable (with deterministic or randomized colorings)? We only show that coloring schemes exist such that the color of (say) the cell of 0 is supported on a finite set. If one asks for a particular distribution the question is interesting also in $\mathbb{R}^d$ for $d>2$. For example, in $\mathbb{R}^d$, it is possible to get a coloring so that color $i$ appears with exponentially (in $i$) small probability. What is the minimal possible entropy of the color of a cell? \paragraph {Fire percolation.} Given a set $S_0$ of vertices in the Delaunay triangulation, let $S_k$ be all vertices at graph distance exactly $k$ from $S_0$. Is it possible to select a set $S_0$ in a deterministic equivariant manner, so that for all $k$, $S_k$ has only finite connected components? If the answer is yes, then coloring the components of $S_k$ for even $k$ with colors $\{0,1,2,3\}$ and the components for odd $k$ by $\{4,5,6,7\}$ results in a deterministic 8-coloring of the Poisson-Voronoi map. Kozma's aforementioned trick can be used to get a 7-coloring in this way. \section{Proof of the main result} \label{sec:proof} In this section we prove Theorem~\ref{thm:main}. As explained in the proof outline, the proof is based on Propositions~\ref{removal_prop} and \ref{area_prop}. These in turn will be proved by reduction to $k$-dependent percolation. Section \ref{dep_perc_sec} below gives the basic fact about $k$ dependent percolation we shall need and introduces sealed squares, the tool which allows us to deduce that events taking place in distant locations are almost independent. In Section~\ref{area_sec} we prove the simpler Proposition~\ref{area_prop} and in Section~\ref{removal_sec} the more difficult Proposition~\ref{removal_prop}. Section~\ref{coloring_sec} shows how to deduce the main result from the two propositions. \textbf{Notation:} Throughout we shall denote by $G=(V,E)$ the Delaunay graph embedded in the plane where $V$ is the set of points of the Poisson process and the edges are straight lines connecting these points (this can be seen to be a planar representation of $G$). We will sometimes call the vertices \emph{centers} and say that a Voronoi cell is \emph{centered} at its vertex. We also let $A:V \to \mathbb{R}_+$ be the function which assigns to each vertex the area of the corresponding Voronoi cell. For $x\in\mathbb{R}^2$ we denote $Q(x,R):=x+[-R,R]^2$, i.e., a square centered at $x$ of side length $2R$. We let $B_R(x)$ or $B(x,R)$ stand for a closed ball of radius $R$ around $x$ (in the Euclidean metric). We write $d(x,y)$ for the Euclidean distance between $x,y\in\mathbb{R}^2$. Similarly $d(x,U):=\inf\{d(x,y)\ |\ y\in U\}$ for sets $U\subseteq\mathbb{R}^2$. \subsection{Dependent percolation and sealed squares} \label{dep_perc_sec} A process $\{A_x\}_{x\in\mathbb{Z}^2}$ is said to be {\em $k$-dependent} if for any sets $S,T \subset \mathbb{Z}^2$ at $\ell^\infty$-distance at least $k$, the restrictions of $A$ to $S$ and to $T$ are independent. Our processes will always take values in $\{0,1\}$. Vertices $x\in\mathbb{Z}^2$ with $A_x=1$ are called {\em open} (and others are {\em closed}). An {\em open component} is a connected component in $\mathbb{Z}^2$ of open vertices. A well known result of Liggett, Schonmann and Stacey \cite{LSS96} states that $k$-dependent percolation with sufficiently small marginals ($\mathbb{E} A_x$) is dominated by sub-critical Bernoulli percolation. The following simple lemma is weaker, and is a standard argument in percolation theory. We include a proof for completeness: \begin{lemma}\label{lem:k_dependent_percolation} For any $k$ there is some $p_0=p_0(k)<1$ such that if $\{A_x\}_{x\in\mathbb{Z}^2}$ is $k$-dependent and for all $x$, $P(A_x=1)\leq p_0$, then \[ \P(\exists \text{ an infinite open component})=0. \] \end{lemma} \begin{proof} The number of simple paths of length $L$ starting at a given $x\in \mathbb{Z}^2$ is bounded by $4^L$. Any simple path of length $L$ contains at least $\frac{L}{k^2}$ coordinates which are pairwise $k$-separated. Thus, the probability that any given path of length $L$ is open is at most $p_0^{L/k^2}$. The expected number of open paths originating at $x$ is bounded by \[ 4^L \cdot p_0^{L/k^2} = \left(4 p_0^{1/k^2}\right)^L. \] If $p_0 < 4^{-k^2}$ this quantity tends to $0$ as $L$ tends to infinity. However, an infinite open component must contain an open path of any length. \end{proof} \begin{defn} A set $S\subset\mathbb{R}^2$ is called {\em $\alpha$-sealed} w.r.t.\ the Poisson process $V$ if $d(x,V)\le \alpha$ for every point $x\in\partial S$. \end{defn} Thus a set is sealed if the point process is not far from any point on the boundary of $S$. This implies that the Voronoi cells of $V$ which intersect the boundary of $S$ are centered near the boundary. The purpose of this notation is that it bounds the dependency between the Voronoi map inside and outside the set. For a set $S$ we denote \[ S^\alpha = \{x \in \mathbb{R}^2 : d(x,S)\leq \alpha\} \] i.e.\ the closed (Euclidean) $\alpha$-neighborhood of $S$ (so that $\alpha$-sealed is equivalent to $\partial S\subset V^\alpha$). Note that being $\alpha$-sealed is determined by $V\cap (\partial S)^\alpha$. We denote by $S^{-\alpha}$ the points at distance at least $\alpha$ from the complement $S^c$ (the idea is that if $S=B_R(x)$ then $S^\alpha = B_{R+\alpha}(x)$ for any $\alpha\ge -R$). \begin{lemma}\label{L:sealed_independent} Condition on the points of $V\cap (\partial S)^\alpha$. On the event that $S$ is $\alpha$-sealed, the Voronoi map in $S^{-\alpha}$ is determined by the process $V\cap S^\alpha$. Moreover, the cell as well as all neighbors of $x\in V\cap S^{-\alpha}$ are contained in $S^\alpha$. \end{lemma} \begin{proof} The lemma follows from the following simple geometrical fact: If $V\cap (\partial S)^\alpha$ is such that $S$ is $\alpha$-sealed, then the center of the cell of any $z\in\partial S$ is in $(\partial S)^\alpha$. Thus the cells of centers in $(\partial S)^\alpha$ separate $S^{-\alpha}$ from $\mathbb{R}^2 \setminus S^\alpha$. It follows that the cell of $x\in S^{-\alpha}$ is contained in $S$, and is adjacent only to cells centered in $S^\alpha$. \end{proof} Next we argue that squares are likely to be $\alpha$-sealed \begin{lemma}\label{L:sealed_high_prob} The probability that $Q(0,R)$ is not $\alpha$-sealed is at most \[ \lceil 8 R/\alpha \rceil e^{-\pi\alpha^2/4}. \] \end{lemma} \begin{proof} Take an $\alpha/2$ net in $\partial Q(0,R)$, of size $\lceil 8R/\alpha \rceil$. Each of these points fails to have a center within distance $\alpha/2$ from it with probability $e^{-\pi \alpha^2/4}$. If none fail to have such a nearby center then the square is $\alpha$-sealed. A union bound gives the claim. \end{proof} \subsection{Areas behave --- Proposition~\ref{area_prop}} \label{area_sec} Our present goal is to prove Proposition~\ref{area_prop}. To this end we need two properties of the areas of Poisson-Voronoi cells. \begin{lemma}\label{L:cont_areas} Let $\mu$ be the law of the area of the cell containing the origin, then $\mu$ is absolutely continuous w.r.t.\ the Lebesgue measure. \end{lemma} A partition of $\mathbb{R}_+$ is a finite union $\mathbb{R}_+ = \bigcup_{i<M} [x_i,x_{i+1})$, given by a sequence $0 = x_0 < x_1 < \cdots < X_M = \infty$. The following is an immediate corollary of \lemref{cont_areas}. \begin{cor} \label{cor:partition_areas} For any $R,\varepsilon>0$ there is some sufficiently refined partition $\mathcal{A}$ of $\mathbb{R}_+$ such that for every interval $I \in \mathcal{A}$ the probability that there exists $v \in V \cap B_R(0)$ with $A(v) \in I$ is at most $\varepsilon$. \end{cor} However, just knowing that the area distribution is continuous is not enough, since the areas of different cells are not independent. For this reason we also need. \begin{lemma} \label{L:distinct_areas} Almost surely, all cells have different areas. \end{lemma} These two lemmas are intuitively obvious, though writing a precise proof is delicate. It is possible to get a somewhat simpler proof by replacing the area of a cell by some other quantity. For example, the distance to the nearest neighbor does not work since some centers have the same distance. However, total distance to the neighbors in the Delaunay graph does work. \begin{proof}[Proof of \lemref{cont_areas}] The idea of the proof is this: let $x$ be the center of the cell of the origin and let $y$ be the center of an adjacent cell. Conditioned on the location of all centers other than $y$, and on the direction of the vector $y-x$, we get that the area of $x$ is a differentiable function of $r=\|y-x\|$, the distance between $x$ and $y$, with positive derivative. Thus, $\mu$ conditioned on this $\sigma$-algebra is absolutely continuous w.r.t.\ Lebesgue and so $\mu$ itself must also be so. To make this precise, we partition $\mathbb{R}^d$ into cubes of size $\varepsilon^d$ centered around $\varepsilon \mathbb{Z}^d$. We condition on the number of points of the Poisson process in each of these cubes. We then use finer and finer partitions (say, with $\varepsilon_i=2^{-i}$) until we reach a partition which already reveals in what cube lies the center of the cell of the origin (i.e.\ $x$) as well as its nearest neighbor (i.e.\ $y$). We then continue according to the previous paragraph: we condition on the exact location of all points of the Poisson process except $y$ and on the direction of $y-x$. After that we get that $A(x)$ is now a monotone function of $r=\|y-x\|$ and its derivative is equal to the length of the intersection of the cells of $x$ and $y$, which is strictly positive. Since under this conditioning, the distribution of $r$ is absolutely continuous w.r.t.\ Lebesgue measure on some interval we get that the conditioned $\mu$ is also absolutely continuous w.r.t.\ Lebesgue and so is $\mu$ itself. \end{proof} \begin{proof}[Proof of \lemref{distinct_areas}] The proof is similar to that of \lemref{cont_areas}. Fixing any two points, $a$ and $b$ we wish to show that the probability that they belong to different cells with equal areas is zero. To that end, we find the two centers of the cells, $x$ and $y$ and find a third cell, centered at $z$, which is adjacent to one of these cells, say, $x$, but not to the other. (Such $z$ exists for any $x,y$ in any planar triangulation with no unbounded face.) Now $A(x)$ depends on the exact location of $z$, as in the proof of \lemref{cont_areas}, but $A(y)$ does not. Of course, all this needs to be done using fine partitions, etc. The lemma now follows by considering all possible values for $a$ and $b$ with rational coordinates. \end{proof} Note that the proof of \lemref{distinct_areas} above does not apply as is to higher dimensions, since in such dimensions, there are configurations with two distinct cells having the same neighbors. Of course, \lemref{distinct_areas} itself remain valid. We now prove Proposition~\ref{area_prop}. The key idea is that cells with areas in any sufficiently small interval are dominated by sub-critical percolation. \begin{proof}[Proof of Proposition~\ref{area_prop}] We show that there is some sufficiently refined partition $\mathcal{A}$ of $\mathbb{R}_+$, such that a.s.\ for any $I \in \mathcal{A}$ there is no infinite path in $G$ with all areas in $I$. The proposition will follow since an infinite path with decreasing areas will have all areas in the same interval of $\mathcal{A}$ from some point on. For some $R$ to be determined later, consider the lattice $L=(2R\mathbb{Z})^2$. For an interval $I$, if there is an infinite path of cells with areas in $I$, then there is an infinite path $\{x_i\}$ in $L$ so that every $Q(x_i,R)$ intersects such a cell. The probability that a square intersects a cell with area in $I$ can be made arbitrarily small, but these events are not independent. To overcome this we use sealed boxes. For an interval $I$, call a point $x\in L$ open if either $Q(x,R)$ intersects a cell with area in $I$, or if either one of $Q(x,R+\alpha)$ or $Q(x,R+3\alpha)$ is not $\alpha$-sealed. If there is an infinite path in $G$ with areas in $I$ then there is also an infinite open path in $L$. The event that the squares are sealed depends only on the Poisson process within $Q(x,R+4\alpha)$. We claim that on the event that they are sealed, the areas of cells intersecting $Q(x,R)$ also depend only on the process in $Q(x,R+4\alpha)$. Taking $\alpha=R/8$ it follows that the process of open boxes is $2$-dependent. To see this claim, note that the center of any cell intersecting $Q(x,R)$ must be within $Q(x,R+2\alpha)$. The second seal implies that the cell of this center is contained in $Q(x,R+4\alpha)$ and determined by the process in this box. To complete the proof, take some $\varepsilon>0$ so that a $2$-dependent percolation with marginal $\varepsilon$ is sub-critical (using Lemma~\ref{lem:k_dependent_percolation}). Using Lemma~\ref{L:sealed_high_prob}, fix $R$ large enough so that with $\alpha=R/8$, \[ \P(Q(x,R+i\alpha) \text{ is not $\alpha$-sealed}) < \varepsilon/3 \qquad \text{for $i=1,3$.} \] Next, using Corollary~\ref{cor:partition_areas} take a partition $\mathcal{A}$ fine enough that for any $I\in\mathcal{A}$, the probability that there exists $v\in V\cap Q(x,R+2\alpha)$ with area in $I$ is at most $\varepsilon/3$. Then for each $I\in\mathcal{A}$, the probability that any fixed $x$ is open is at most $\varepsilon$ and so the process of open points does not contain an infinite open path. \end{proof} \subsection{Deleting low degree vertices --- Proposition~\ref{removal_prop}} \label{removal_sec} In this section we prove Proposition~\ref{removal_prop}. Throughout the section $R>0$ is a parameter, assumed large enough as needed for the calculations which follow. We also define the square annuli $A(x,r,R) := Q(x,R) \setminus Q(x,r)$. We now introduce our main object of study in this section: \begin{defn} Inductively, let $G^R_{0}:=G$ and let $G^R_{n+1}$ denote the graph obtained from $G^R_{n}$ by deleting all vertices in $Q(0,3R)$ with $G^R_{n}$-degree at most $5$. Let $G^R_\infty:=\cap_{n=0}^\infty G^R_n$. \end{defn} Thus we iteratively delete vertices of degree at most 5, but only those vertices contained in a fixed large square. We aim to prove the following \begin{lemma}\label{L:finite_radius_6core} We have $\P\big(G^R_\infty \cap Q(0,R) \neq \emptyset \big) \xrightarrow[R\to\infty]{} 0$. \end{lemma} \begin{cor}\label{C:finite_radius_6core} For any $\varepsilon>0$, there are $R,M$ so that $\P\big(G^R_M \cap Q(0,R) \neq \emptyset \big) < \varepsilon$. \end{cor} \begin{proof} Pick $R$ such that $\P\big(G^R_\infty \cap Q(0,R) \neq \emptyset \big) < \varepsilon$. Since \[ \Big\{G^R_\infty \cap Q(0,R) \neq \emptyset\Big\} = \bigcap_M \Big\{G^R_M \cap Q(0,R) \neq \emptyset\Big\}, \] the bound will hold for that $R$ and sufficiently large $M$. \end{proof} Before embarking on the proof of \lemref{finite_radius_6core}, let us explain how one can get a similar and simpler result when deleting vertices of degree at most 6 (thus yielding a deterministic $7$-coloring). Suppose that $G^R_\infty$ contains a vertex in $Q(0,R)$. By \lemref{no_long_edges} $G^R_\infty$ is unlikely to contain edges longer then $\log R$ within $Q(0,3R)$. All vertices of $G^R_\infty$ in $Q(0,3R)$ have degree at least $7$. It is an easy consequence of Euler's formula that a planar graph with minimal degree $7$ has positive expansion (the boundary of any set is proportional to its size). This implies (in the absence of long edges) that the number of vertices of $G^R_\infty$ in $Q(0,3R)$ is exponential in $R$. Of course, this too is unlikely. When deleting vertices of degree at most 5, the remaining graph has minimal degree 6, which is not as obviously unlikely. However, this can only happen if $G$ contains a large segment of a triangular lattice, which we rule out below. We begin with two combinatorial lemmas on planar maps. For any finite graph $H$, let $LD=LD(H)$ be the number of vertices of low-degree, namely at most 5. For a finite simple planar map $H$, let $ME=ME(H)$ be the number of ``missing edges'': the number of edges that can be added to the map while keeping it planar and simple. A face of size $k$ can be triangulated using $k-3$ edges, after which no further edges can be added, and so $ME = \sum_f (\deg[f]-3)$ (where the sum also includes the external face and where we assume $|H|\ge 3$ so that $\deg[f]\ge 3$ for all faces). \begin{lemma}\label{L:LD_large} For any finite, connected and simple planar map $H$ with $|V[H]|\ge 3$ we have $LD \ge \frac25 ME + \frac{12}{5}$. \end{lemma} \begin{proof} Add $ME$ edges to make the map into a triangulation. Let $d'_v$ be the resulting vertex degrees, than we have $\sum_v (6-d'_v) = 12$ (using Euler's formula combined with the triangulation property $3F=2E=\sum d'_v$), and therefore $\sum_v (6-d_v) = 12 + 2 ME$. The claim follows since low-degree vertices contribute at most 5 to this sum, and high degree vertices at most 0, so that $\sum_v (6-d_v) \le 5 L D$. \end{proof} \begin{lemma}\label{L:large_6core} Fix $\rho>\ell>0$. Let $H$ be a simple planar graph embedded in $\mathbb{R}^2$ satisfying the following: \begin{enumerate} \item All vertices in $Q(0,3\rho)$ have degree at least $6$. \item All edges of $H$ with an endpoint in $Q(0,3\rho)$ have length at most $\ell$. \item There exists a vertex of $H$ in $Q(0,\rho)$. \end{enumerate} Then $H$ has at least $\frac{8\rho^2}{5\ell^2}$ vertices in $Q(0,3\rho)$. \end{lemma} Note that the order of magnitude $(\rho/\ell)^2$ is achieved by a triangular lattice with edge length $\ell$. \begin{proof} We assume that $H$ has only finitely many vertices in $Q(0,3\rho)$ since otherwise the conclusion is trivial. Fix a vertex $v\in Q(0,\rho)$. For $t\in[\rho,3\rho]$, let $H'_t$ be the sub-graph induced by vertices inside $Q(0,t)$, and let $H_t$ be the connected component of $v$ in $H'_t$. Note that the connected component of $v$ in $H$ is not contained in $Q(0,3\rho)$ since otherwise it would be a finite, connected and simple planar map with all degrees at least 6 which is impossible by Lemma~\ref{L:LD_large}. By our assumptions, all vertices of $H_t$ with neighbors in $H\setminus H_t$ (which includes all vertices of degree at most 5 in $H_t$) must be in the annulus $A(0,t-\ell,t)$. It follows that the external face of $H_t$ surrounds $v$ and exits $Q(0,t-\ell)$ and so has degree at least $\frac{2(t-\rho-\ell)}{\ell}$. Thus \[ ME(H_t) \ge \frac{2(t-\rho-\ell)}{\ell} - 3 = \frac{2(t-\rho)}{\ell} - 5. \] By \lemref{LD_large}, the number of vertices in $A(0,t-\ell,t)$ is at least $\frac25 ME(H_t) + \frac{12}{5} \ge \frac45 \frac{t-\rho}{\ell} + \frac25$. Let $M = \lfloor 2\rho/\ell\rfloor$. Splitting $A(0,\rho,3\rho)$ into annuli $A(0,\rho+(k-1)\ell,\rho+k\ell)$ for $k=1,\dots,M$ one finds that the number of vertices of $H$ in $Q(0,3\rho)$ is at least \[ 1 + \sum_{k=1}^M \left(\frac45 k + \frac25\right) = \frac{2(M+1)^2+3}{5} \ge \frac{2(2\rho/\ell)^2}{5}. \qedhere \] \end{proof} Next, a simple lemma showing that long edges in $G$ are unlikely. \begin{lemma}[No long edges]\label{L:no_long_edges} The probability of having an edge of length at least $\ell$ in $E[G]$ which intersects the square $Q(0,\rho)$ is at most \[ \left( \frac{\sqrt{32}\rho}{\ell}+8 \right)^2 e^{-\ell^2/32}. \] \end{lemma} \begin{proof} Suppose $(x,y)$ were such an edge, then some disc with $x,y$ on its boundary has no points in its interior. Consequently at least one of the two semi-circles with diameter $(x,y)$ has no points in its interior. This implies that there is an empty disc $B_{\ell/4}(z)$ for some $z\in Q(0,\rho+\ell)$ ($z$ might be outside $Q(0,\rho)$ since one of $x,y$ may be outside the square). Cover $Q(0,\rho+\ell)$ by $\left\lceil \frac{\rho+\ell}{\ell/\sqrt{32}} \right\rceil^2$ squares of side length $\ell/\sqrt{32}$. It follows that if such a long edge exists than one of the squares (the one containing $z$) must be empty, and the claim follows. \end{proof} We continue by showing that after some low-degree vertices are deleted, many large holes remain in the graph. \begin{defn} Call a square $Q(x,\rho)$ a \emph{typical square} if there exists some vertex $v \in Q(x,\rho)$ such that: \begin{enumerate} \item $\deg[v] < 6$. \item $v$ is not in the interior of any triangle in the Delaunay Graph $G$. \end{enumerate} Otherwise we call the square \emph{rare}. \end{defn} To make this clear, the second condition states that there are no $v_1,v_2,v_3 \in V$ which are pairwise adjacent in $G$ such that $v$ is contained in the interior of the triangle $(v_1,v_2,v_3)$. \begin{lem}[Rare squares are rare] \label{L:rare_squares} For some $\alpha,\beta>0$ we have $\P(Q(x,\rho) \text{ is rare}) \le \alpha\exp(-\beta \rho)$. \end{lem} \begin{proof} We may assume without loss of generality that $\rho\ge C$ for some large $C>0$ (otherwise the claim is trivial). Let $\gamma=\sqrt{\rho}$. The square $Q(0,\rho)$ contains at least $c\rho$ disjoint squares $Q(x,4\gamma)$ for some $c>0$. Call each of these squares {\em good} if it satisfies the following: \begin{enumerate} \item $Q(x,3\gamma)$ is $\gamma$-sealed, \item $Q(x,\gamma)$ contains a vertex $v$ of degree at most 5 which is not in the interior of any triangle with vertices in $Q(x,2\gamma)$. \end{enumerate} Note that by \lemref{sealed_independent}, the event that $Q(x,4\gamma)$ is good is determined by the Poisson process within it, and so these events are all independent. Each square has some probability $p>0$ of being good (independent of $\rho$ since long edges are unlikely by Lemma~\ref{L:no_long_edges}), so the probability that no square within $Q(0,\rho)$ is good is at most $e^{-\beta\rho}$ for some $\beta>0$. If the low-degree vertex in a good square is contained in a triangle of $G$ then an edge of that triangle must have length at least $\gamma$. Either the triangle intersects $Q(0,\rho)$, which by Lemma~\ref{L:no_long_edges} has probability at most $C_1\rho e^{-\rho/32} \le C_2e^{-\rho/33}$ for some $C_1,C_2>0$. Or the triangle contains $Q(0,\rho)$ in its interior, in which case for some integer $m\ge 1$, its longest edge has length at least $m\rho$ and intersects $Q(0,m\rho)$. By a union bound, this has probability at most $\sum_{m=1}^\infty C_2e^{-m^2\rho^2/33}\le C_3e^{-\rho^2/33}$ for some $C_3>0$. \end{proof} In what follows, define \begin{align*} L &:= \log R, & r &:= R^{1/3}. \end{align*} $L$ will be a bound on length of edges that appear (and can be reduced to $C\sqrt{\log R}$ for large enough $C$). The role of $r$ is more involved, and there is much freedom in the choice of $r$. Primarily, we consider a partition of boxes of size of order $R$ into boxes of size $r$. For simplicity, we assume that $6R/r$ is an odd integer ($R$ can be arbitrarily large under this condition). Define for each $x \in r\mathbb{Z}^2$ the square $Q_x:=Q(x,\frac{r}{2})$. Note that $Q(0,3R)$ is precisely tiled by the boxes $\{Q_x, x\in r\mathbb{Z}^2\cap Q(0,3R)\}$. We now define several events which we will show to be unlikely. \begin{align*} \Omega_0 &:=\{\text{$G^R_\infty$ has a vertex in $Q(0,R)$}\}, \\ \Omega_1 &:=\{\text{There exists $e\in E[G]$ of length at least $L$ that intersects $Q(0,3R)$}\}, \\ \Omega_2 &:=\{\text{$Q_x$ is rare for some $x\in r\mathbb{Z}^2\cap Q(0,3R)$}\}, \\ \Omega_3 &:=\{\text{There exists $x\in r\mathbb{Z}^2 \cap Q(0,3R)$ and $|V\cap Q_x| \ge 2r^2$}\}, \\ \Omega_4 &:=\{|V\cap A| > 2\mbox{Area}(A)\}, \text{ where } A = A(0,3R,3R+L). \end{align*} Thus \lemref{finite_radius_6core} states that $\P(\Omega_0)$ is small. \begin{lemma} With $L,r$ as above, $\P(\Omega_i) \xrightarrow[R\to\infty]{} 0$ for $i=1,2,3,4$. \end{lemma} \begin{proof} \lemref{no_long_edges} implies that $\P(\Omega_1) = O(R^2 e^{-L^2/32})$ is small. \lemref{rare_squares} implies $\P(\Omega_2) = O(R^{4/3} e^{-\beta r})$ (since there are $(6R/r)^2$ squares to consider). $\P(\Omega_3)$ and $\P(\Omega_4)$ are bounded by the fact that $\P\big(\Poi(\lambda)>2\lambda\big) \leq e^{-c\lambda}$ for some constant $c$. This gives respective bounds $O(R^2 e^{-cr^2})$ and $O(e^{-cRL})$. \end{proof} Define the set \[ S := \left\{x \in r\mathbb{Z}^2 \cap Q\left(0,3R\right) ~:~ Q_x \text{ is typical and } G^R_{\infty} \cap Q_x \ne \emptyset \right\}. \] \begin{lemma}\label{many_boundary_verts_lem} There exists $C>0$ such that if $\Omega_1^c$ holds then \[ \big| S \big| \leq C \Big|V\cap A(0,3R,3R+L) \Big|. \] \end{lemma} \begin{proof} Let $H$ be the sub-graph of $G^R_\infty$ induced by vertices in $Q(0,3R+L)$. On the event $\Omega_1^c$, the vertices of $H \cap Q(0,3R)$ all have degree at least 6. Thus low-degree vertices are all in the annulus $A(0,3R,3R+L)$ and by Lemma~\ref{L:LD_large}, \[ \Big|V\cap A(0,3R,3R+L) \Big| \ge LD(H) > \frac25 ME(H). \] For each $x\in S$ the square $Q_x$ is typical. Hence there is a vertex $v_x\in Q_x$ of degree at most 5 that is not contained in any triangle in $G$. The vertex $v_x$ is deleted in the first round and so is not in $H$. Let $f_x$ be the face of $H$ surrounding $v_x$. Note that $f_x$ must have an edge $e_x$ that intersects $Q_x$, since otherwise $Q_x$ is completely in the interior of $f_x$ and there could be no vertex of $H$ in $Q_x$. Now, on $\Omega_1^c$, the edge $e_x$ has length at most $L<r$ and therefore can intersect at most 3 different squares $Q_x$ (it can intersect 3 if it passes near a corner of $Q_x$). Since the face $f_x$ cannot be a triangle by definition of $v_x$ we deduce that \begin{equation*} \sum_{\substack{f \text{ face of }H \\ \deg[f]>3}} \deg[f] \ge \frac{1}{3}|S|. \end{equation*} Hence $ME(H) = \sum_{f\in H} (\deg[f]-3) \ge \frac{1}{12}|S|$ proving the claim (with $C=30$). \end{proof} \begin{proof}[Proof of \lemref{finite_radius_6core}] We show that $\Omega_0 \subset \bigcup_{i=1}^4 \Omega_i$. Assume by negation that $\Omega_0$ and $\Omega_i^c$ hold for $i=1,2,3,4$. Let $H$ be the restriction of $G^R_\infty$ to $Q(0,3R+L)$, and apply \lemref{large_6core} with $\rho=R$, $\ell=L$. $\Omega_1^c$ and $\Omega_0$ show that the lemma's hypotheses hold, thus $H$ has at least $\frac{8R^2}{5L^2}$ vertices in $Q(0,3R)$. On the other hand we show that $H$ is small. Tile $Q(0,3R)$ by boxes $Q_x$ with $x\in r\mathbb{Z}^2 \cap Q(0,3R)$. On $\Omega_4^c$, Lemma~\ref{many_boundary_verts_lem} implies that $|S| \leq C R L$ for some $C$. On $\Omega_3^c$ each of these includes at most $2r^2$ vertices, so the number of vertices of $H$ in $Q(0,3R)$ that are in typical boxes is at most $2r^2|S| \leq C R L r^2$. On $\Omega_2^c$ there are no vertices in rare boxes. Thus $\frac{8R^2}{5L^2} \leq C R L r^2$ which is a contradiction for $R$ large enough and our choice of $r$ and $L$. \end{proof} \begin{proof}[Proof of Proposition~\ref{removal_prop}] Define $G^{R,x}_M$ similarly to $G^R_M$, except that low degree vertices are deleted in $Q(x,3R)$ instead of $Q(0,3R)$. Consider the following dependent percolation process on the lattice $\Lambda=R\mathbb{Z}^2$. A point $x$ is open in one of 3 cases: \begin{enumerate} \item The square $Q(x,4R)$ is not $R$-sealed, \item $G^{R,x}_M$ has a vertex in $Q(x,R)$, \item $G$ has an edge of length at least $R/2$ intersecting $Q(x,R/2)$. \end{enumerate} We first argue that the event $\{x \text{ is open}\}$ is determined by the Poisson process in $Q(x,5R)$, so that the process is 11-dependent. Indeed, whether $Q(x,4R)$ is $R$-sealed depends only on the process in $Q(x,5R)$. If it is $R$-sealed, the restriction of the Voronoi map to $Q(x,3R)$ is determined by the process in $Q(x,5R)$, which determine the state of $x$. By Lemmas~\ref{L:sealed_high_prob}, \ref{L:finite_radius_6core} and \ref{L:no_long_edges}, we can choose $M,R$ so that $\P(x \text{ is open})$ is arbitrarily small. In particular, for some $M,R$, using Lemma~\ref{lem:k_dependent_percolation}, this percolation is dominated by sub-critical percolation, and has no infinite open component. Finally, we argue that if there were an infinite component in $G_M$ then there would also be an infinite component in our process on $\Lambda$. Consider all squares $Q(x,R/2)$ which intersect the edges of some infinite open component in $G_M$. For each such $x$, either there is a vertex of $G_M$ in $Q(x,R)$, or else the edge that passes through $Q(x,R/2)$ has both endpoints outside $Q(x,R)$. Since $G_M\subset G^{R,x}_M$, either case implies $x$ is open. \end{proof} \subsection{Equivariant Coloring}\label{coloring_sec} We now use Propositions~\ref{removal_prop} and \ref{area_prop} to construct a deterministic $6$-coloring scheme. Recall $G_n$ is derived from $G_{n-1}$ by deleting low degree vertices. Define the level of a vertex by \[ \ell(v) = \max\{n : v\in G_n \}. \] Thus a vertex has level 0 iff its degree is at most 5. For neighboring $v,w$ we direct the edge from $v$ to $w$, and write $v \to w$, if either $\ell(w)>\ell(v)$ or ($\ell(v)=\ell(w)$ and $A(w) < A(v)$) (Proposition~\ref{area_prop} gives that no two areas are equal). Let $\prec$ to be the transitive closure of $\to$. That is, $w\prec v$ iff there is a finite sequence such that $v=u_0\to u_1 \to \ldots \to u_n=w$. \begin{lemma}\label{L:finite_predecessors} A.s.\ every $v \in V$ has finitely many $\prec$-predecessors (in particular, $\prec$ is well founded). \end{lemma} \begin{proof} We first argue that there is no infinite directed path in $G$. By Proposition~\ref{area_prop} there are no infinite $A$-monotone paths, so any infinite directed path must have $\ell(v)\to\infty$. However, by Proposition~\ref{removal_prop} there are no infinite paths with $\ell>M$. Our conclusion then follows from K\"{o}nig's lemma: A locally finite tree with no infinite paths is finite. \end{proof} From this we get: \begin{prop} There exists a unique function $f:V \to \{0,\dots,5\}$ determined by the recursive formula: \[ f(u) = \mex \{ f(v) : u\to v \} \] where $\mex S = \min(\mathbb{N} \setminus S)$ is the minimal excluded integer function. \end{prop} \begin{proof} The proof is by induction on $\prec$, which is a well founded order by \lemref{finite_predecessors} (see e.g.\ \cite[Chapter 3]{Kunen}). Any $u \in V$ has at most 5 neighbors $v$ with $\ell(v) \ge \ell(u)$, so $|\{ v : u\to v\}| \le 5$ and so $f(u) < 6$ is well defined. Uniqueness holds since $f(u)$ is determined by $\{f(v) : v\prec u\}$. \end{proof} \thmref{main} now follows, since $\mathcal{P}_i = f^{-1}(i)$ defines a deterministic, isometry equivariant $6$-coloring. Note that the resulting coloring is finitary, that is, for every $x \in \mathbb{R}^2$ there exists a finite (but random) $R>0$ such that the color of the cell containing $x$ is a function of the Poisson process restricted to $B_R(x)$. Indeed, to determine $f(v)$ for $v \in V$, it is sufficient to know the graph $G$ induced on the $\prec$-predecessors of $v$. Furthermore, there exist $C,c>0$ such that $\P(R>s) \leq C e^{-c s}$. This is the case because Propositions~\ref{removal_prop} and \ref{area_prop} are proved using domination by sub-critical percolation.

https://arxiv.org/abs/1903.08214

A tighter bound on the number of relevant variables in a bounded degree Boolean function

A classical theorem of Nisan and Szegedy says that a boolean function with degree $d$ as a real polynomial depends on at most $d2^{d-1}$ of its variables. In recent work by Chiarelli, Hatami and Saks, this upper bound was improved to $C \cdot 2^d$, where $C = 6.614$. Here we refine their argument to show that one may take $C = 4.416$.

\section{Introduction} Given a Boolean function $f: \{0,1\}^n \to \{0,1\}$, there is a unique multilinear polynomial in $\R[x_1, \dots, x_n]$ which agrees with $f$ on every input in $\{0,1\}^n$. One important feature of this polynomial is its degree, denoted $\deg(f)$, which is known to be polynomially related to many other complexity measures, such as block sensitivity $\text{bs}(f)$, certificate complexity $C(f)$, decision tree depth $D(f)$, and approximate degree $\widetilde{\deg(f)}$ (see \cite{NS} and \cite{BdW}). One can also bound the number of relevant variables of $f$ (i.e. the variables which actually show up in a term with non-zero coefficient in the polynomial for $f$, also called the \emph{junta size} of $f$) entirely in terms of the degree: \begin{thm}[Nisan-Szegedy \cite{NS}]\label{NS} A function $f: \{0,1\}^n \to \{0,1\}$ with degree $d$ has at most $\frac{d}{2}\cdot 2^d$ relevant variables. \end{thm} The idea of Nisan and Szegedy's original proof is to lower bound the \emph{influence} of a relevant variable: a polynomial of degree $d$ has total influence at most $d$, and yet the derivative in the direction of a relevant coordinate is a degree $d-1$ polynomial which is not identically zero, so it is non-zero on a random input with probability at least $1/2^{d-1}$. In other words, each relevant coordinate has influence at least $1/2^{d-1}$, so there can be at most $d\cdot 2^{d-1}$ of them. Theorem \ref{NS} is has the correct exponential dependence on $d$ -- indeed, consider the function $f$ given by the complete binary decision tree of depth $d$ which queries a distinct coordinate at each vertex. This function has degree $d$ and $2^d - 1$ relevant variables. However, it remained open whether the multiplicative factor of $\Theta(d)$ was necessary until a recent paper by Chiarelli, Hatami and Saks showed that $O(2^d)$ suffices.\footnote{In the same paper, the authors also give an improved lower bound construction, namely, for each $d$, a degree $d$ function with $\frac{3}{2}\cdot 2^d - 2$ relevant variables. } \begin{thm}[Chiarelli, Hatami, Saks, \cite{CHS}]\label{6.6} A function $f: \{0,1\}^n \to \{0,1\}$ with degree $d$ has at most $(6.614)\cdot 2^d$ relevant variables. \end{thm} The main idea in \cite{CHS} is to replace influence by a different measure -- one which behaves more stably under restrictions of variables. Specifically, they define $$W(f) = \sum_{i \in R(f)} 2^{-\deg_i(f)}$$ where $R(f)$ is the set of relevant variables for $f$, and $\deg_i(f)$ is the degree of $f$ in $x_i$. It is straightforward to check that $W(f)$ does not decrease by more than $|H|\cdot 2^{-d}$ in expectation when randomly restricting a set $H$ of coordinates with $\deg_i(f) = \deg(f) = d$. If $H$ is chosen well, these contributions are summable, and hence $W(f)$ is bounded above by some universal constant. Since $|R(f)| \leq 2^{\deg(f)}W(f)$, this implies Theorem \ref{6.6}. The heart of the proof is therefore in choosing the set $H$ which is both \emph{small} enough so that $W(f)$ does not incur a heavy loss, and yet \emph{significant} enough that the restricted functions are of reduced complexity. The idea used in \cite{CHS} (originating in unpublished work of Nisan and Smolensky, see \cite{BdW}) is to build $H$ from a maximal collection of disjoint monomials of full degree. The number of such disjoint monomials is limited by the block sensitivity $\text{bs}(f)$, which is always at most $d^2$, and by maximality, all of the resulting restricted functions have degree $\leq d - 1$. \subsection{Our improvements} The above idea certainly does the trick, but there are two somewhat substantial sources of slack in the analysis: one is the global use of the worst-case bound $\text{bs}(f) \leq d^2$, which can be improved for any fixed $d$ with a finite computation. The other is that restricting a large disjoint collection of degree $d$ monomials actually causes a large drop in block sensitivity, which can be exploited. By leveraging both of these ideas, we are able to improve the constant $6.614$ by about 33$\%$: \begin{thm}[Main result]\label{main} A function $f: \{0,1\}^n \to \{0,1\}$ with degree $d$ has at most $(4.416)\cdot 2^d$ relevant variables. \end{thm} \section{Preliminaries} \textbf{Restrictions:} For a function $f:\{0,1\}^n \to \{0,1\}$, a set $H \subset[n]$, and an assignment $\alpha: H \to \{0,1\}$, we denote by $f_{\alpha}$ the restricted function obtained by setting the variables $x_h$ to $\alpha(h)$ for $h \in H$. We will sometimes use $f(\alpha_H, x)$ for $f_\alpha(x)$ if we want to be explicit about the set of coordinates which have received the assignment by $\alpha$. \textbf{Influence:} The influence of coordinate $i$ on $f$, or $\text{Inf}_i[f]$, is the probability that, for a uniformly random input $x$, flipping the $i$th bit of $x$ causes the value of $f(x)$ to flip. The total influence $\text{Inf}[f] = \sum_{i \in [n]} \text{Inf}_i[f]$ can also be expressed in terms of the Fourier coefficients of $f$, namely $$ \text{Inf}[f] = \sum_{S \subseteq [n]} |S|\widehat{f}(S)^2.$$ Since the degree of $f$ remains unchanged when $f$ is expressed as a multilinear polynomial over $\{0,1\}^n$ (as we consider in this paper) or $\{1, -1\}^n$ (as in the Fourier expansion), the above formula makes it clear that a Boolean function of degree $d$ has $\text{Inf}[f] \leq d$. As mentioned in the introduction, the following useful fact is from \cite{NS}, and can be proved by induction: \be\label{NS_fact} \text{Inf}_i[f] \geq 2^{1-\deg_i(f)} . \ee \textbf{Block sensitivity:} For a set $B \subset [n]$ and a string $x \in \{0,1\}^n$, we denote by $x^B$ the string obtained from $x$ by flipping all the bits $x_b$ for $b \in B$. Recall that the block sensitivity of $f$ at an input $x$ (denoted $\text{bs}_x(f)$) is the maximum number $b$ of disjoint blocks $B_1, \dots, B_b \subset [n]$ such that $f(x) \neq f(x^{B_i})$ for all $1 \leq i \leq b$, and the block sensitivity of $f$ (denoted $\text{bs}(f)$) is the maximum of $\text{bs}_x(f)$ over all inputs $x$. It is well-known that block sensitivity and degree are polynomially related: \be\label{bs_vs_deg} \deg(f)^{1/3} \leq \text{bs}(f) \leq \deg(f)^2 \ee although neither bound is known to be sharp. The best known constructions have $\text{bs}(f) = \Theta(\deg(f)^{1/2})$ and $\text{bs}(f) = \Theta(\deg(f)^{\log_3(6)}) = \Theta(\deg(f)^{1.6309...}) $ respectively. See \cite{BdW} and \cite{HKP} for details and for relationships to many other complexity measures. \textbf{The measure $W(f)$:} Recall that $$W(f) := \sum_{i \in R(f)} 2^{-\deg_i(f)},$$ where $R(f)$ is the set of relevant coordinates (i.e. coordinates $i$ for which $\text{Inf}_i[f] > 0$) and $\deg_i(f)$ is the degree of largest degree monomial appearing in $f$ (with non-zero coefficient) that contains $x_i$. The behavior of $W$ under restrictions boils down to the following inequality, whose simple proof we reproduce below for completeness. \begin{fact}[\cite{CHS}]\label{fact} For any relevant coordinates $i \neq j$, let $f_0$ and $f_1$ be the restrictions obtained from $f$ by setting $x_j$ to 0 and 1 respectively. Then \be \label{key_ineq}2^{-\deg_i(f)} \leq 2^{-\deg_i(f_0) - 1} + 2^{-\deg_i(f_1) - 1} \ee \end{fact} \begin{proof} Write $f = x_jf_1 + (1-x_j)f_0 = x_j(f_1 - f_0) + f_0$, from which it is clear that $\deg_i(f) \geq \deg_i(f_0)$. If $\deg_i(f) \geq 1 + \deg_i(f_0)$ then the inequality is true independently of $\deg_i(f_1)$. Otherwise, it must be that $\deg_i(f) = \deg_i(f_0)$, in which case the leading degree monomials for $x_i$ must cancel in $f_1 - f_0$. But this implies $\deg_i(f) = \deg_i(f_1) = \deg_i(f_0)$, and so the inequality becomes an equality in this case. \end{proof} By summing (\ref{key_ineq}) over $i \in R(f)$ and iterating over restrictions of more variables, one obtains \be \label{summed_ineq} W(f) \leq |H|\cdot 2^{-d} + \frac{1}{2^{|H|}}\sum_{\alpha: H \to \{0,1\}} W(f_\alpha) \ee for any set $H \subset [n]$ with $\deg_i(f) = d$ for all $i \in H$. As in \cite{CHS}, we define $$W_d := \max_{\deg(f) = d} W(f).$$ If $H$ is chosen as a maximal collection of degree $d$ monomials in $f$, then each $f_\alpha$ has degree at most $d-1$. An unpublished argument of Nisan and Smolensky (which we essentially use in the proof of Lemma \ref{bd_recursive} below) implies that $|H| \leq \deg(f)\cdot \text{bs}(f) \leq d^3$, and so (\ref{summed_ineq}) yields the recursive inequality $$W_d \leq d^3 \cdot 2^{-d} + W_{d-1}.$$ This is already summable, but the bound $W(f) \leq \frac{1}{2}\text{Inf}[f] \leq d/2$ is preferable for small $d$, and optimizing over the choice of the two bounds yields $W_d \leq 6.614$ for all $d$. \section{Improving the constant} \subsection{Don't spend it all in one place} Our first new idea is simply to keep track of block sensitivity through the restriction process: the main observation is Proposition \ref{bs decr} below, which says that if $f$ has $\ell$ disjoint monomials of maximum degree, then by assigning any values to the variables in these monomials, the block sensitivity of the restricted function decreases by $\ell$. So, if we have to restrict many variables in order to drop the degree of $f$ (i.e. to hit all the maximum degree monomials), then we must ``spend" our limited supply of block sensitivity, and in the future it will become much easier to lower the degree again. \\ \begin{prop}\label{bs decr} If $f: \{0,1\}^n \to \{0,1\}$ has $\ell$ \textbf{disjoint} monomials $M_1, \dots, M_\ell$, each of degree $d = \deg(f)$, then for any assignment $\alpha: \cup M_i \to \{0,1\}$, the restricted function $f_\alpha$ has \emph{$$\text{bs}(f_\alpha) \leq \text{bs}(f) - \ell.$$ } In particular, \emph{$\ell \leq \text{bs}(f)$}. \end{prop} \begin{proof} Let $M = \cup_{i=1}^\ell M_i$ and $b = \text{bs}(f_{\alpha}) = \text{bs}_y(f_\alpha)$, for some $y \in \{0,1\}^{[n] \setminus M}$. Then there are $b$ disjoint blocks $B_1, \dots, B_b \subset [n] \setminus M$ with $f(\alpha_M, y) \neq f(\alpha_M, y^{B_j})$ for each $j$. Since $M_i$ is a maximum degree monomial in $f$, each of the functions $\{0,1\}^{M_i} \ni x \mapsto f(x, z)$ is non-constant for any $z$. Therefore, for each $i$, there is a block $C_i \subset M_i$ with $f(\alpha_M^{C_i}, y) \neq f(\alpha_M, y)$. Therefore $\{C_1, \dots, C_\ell, B_1, \dots, B_b\}$ is a collection of disjoint sensitive blocks for $f$ at the input $(\alpha_M, y) \in \{0,1\}^n$, and so $\text{bs}(f) \geq b + \ell$. \end{proof} To keep track of $W$, degree and block sensitivity simultaneously, we define $$W(b, d) := \max_{\substack{f \text{ with }\text{bs}(f) \leq b \\ \text{ and } \deg(f) = d }} W(f).$$ Note that $W(0, d) = 0$, $W(b, 0) = 0$, and $W(b, d) \leq W_d$ for any $b$. By (\ref{bs_vs_deg}), we have $W(d^2, d) = W_d$, and we make the convention that $W(b, d) = 0$ for $b > d^2$. \begin{lem}\label{bd_recursive} For each $b,d$ with $b \leq d^2$, we have $$W(b,d) \leq \max_{(\ell, k) \in \{1, \dots, b\} \times \{1, \dots, d \}}\left(\ell \cdot d \cdot2^{-d} + W(b - \ell, d - k)\right) $$ \end{lem} \begin{proof} Suppose $f$ has degree $d$ and $\text{bs}(f) \leq b$. Let $M_1, \dots, M_\ell$ be a maximal collection of disjoint degree $d$ monomials in $f$, and let $H = \cup_i M_i$. By inequality (\ref{summed_ineq}), $$W(f) \leq \underbrace{|H|\cdot 2^{-d}}_{= \, \ell \cdot d \cdot 2^{-d}} + \underset{\alpha: H \to \{0,1\}}{\E}[W(f_\alpha)]$$ Because the collection $\{M_1, \dots, M_\ell\}$ is maximal, $H$ hits every degree $d$ monomial and hence each $f_\alpha$ has degree $d_{\alpha} \leq d - 1$. By Proposition \ref{bs decr}, each $f_{\alpha}$ has $\text{bs}(f_\alpha) \leq b - \ell$. Since $W(\cdot ,d)$ is monotone (for feasible inputs), it follows that for each $\alpha$, $W(f_\alpha) \leq W(b - \ell, d - k')$, where $k' = \arg\max_{k \in \{1, \dots, d\}} W(b - \ell, d - k)$. Taking the maximum over all possible values of $\ell \in \{1, \dots, b\}$ yields the desired bound. \end{proof} Since $W_d$ is bounded and increasing\footnote{It is shown in \cite{CHS} that $W_d \geq 2^{-d} + W_{d-1}$, and in fact their lower bound construction can be turned into a proof that $W_d \geq 2\cdot 2^{-d} + W_{d-1}$.}, so $W^* := \lim_{d \to \infty} W_d$ exists. Since $W_d = W(d^2, d)$, the following corollary comes easily from Lemma \ref{bd_recursive}. \begin{cor}\label{bd_cor} For any $d$, $$W^* \leq W(d^2, d) + \sum_{r = d+ 1}^{\infty} r^3 2^{-r} $$ \end{cor} Lemma \ref{bd_recursive} yields explicit bounds on $W(b,d)$ for any finite $(b,d)$, which in turn yields an explicit bound on $W^*$ via Corollary \ref{bd_cor}. For small values ($d \leq 9$), the bound $$W(b,d) \leq W_d \leq \max _{\deg(f) = d}\sum_{i \in R(f)} \frac{\text{Inf}_i[f]}{2} = \frac{d}{2} $$ is better than the one from Lemma \ref{bd_recursive}. Extracting numerical bounds recursively yields $$W(50^2, 50) \leq 5.07812...$$ which implies the same bound (to around 10 decimal digits) on $W^*$. \subsection{Tighter bounds on block sensitivity for low degree functions} To further reduce our estimate of $W^*$, we focus on functions of low degree, which clearly have the most influence on the bounds. Specifically, we produce sharper upper bounds on the block sensitivity of such functions, by solving a small set of linear programs. We begin with a simple reduction to linear program feasibility, using ideas from the original proof of $\text{bs}(f) \leq 2\deg(f)$ from \cite{NS}.\footnote{The ``2" in this bound can be removed by using repeated function composition (or \emph{tensorization}), as shown in \cite{Tal}.} \begin{fact}\label{bs_reduction} If there exists a function $f : \{0,1\}^n \to \{0,1\}$ of degree $d$ with block sensitivity $b$, then there exists another function $g: \{0,1\}^b \to \{0,1\}$ of degree $\leq d$ with $g(0) = 0$ and $g(w) = 1$ for each vector $w$ of hamming weight 1. \end{fact} \begin{proof} If $f(x)$ attains maximal block sensitivity at $z$, then $f(x \oplus z)$ attains maximal block sensitivity at 0, so without loss of generality we may assume $z = 0$, and possibly replacing $f$ by $1-f$ we may also assume that $f(0) = 0$. If $B_1, \dots, B_b$ are sensitive blocks for $f$ at 0, then define $$g(y_1, \dots, y_b) = f(\underbrace{y_1, \dots, y_1}_{B_1}, \dots, \underbrace{y_b, \dots, y_b}_{B_b})$$ so that for each coordinate vector $e_i$, $g(e_i) = f(\textbf{1}_{B_i}) = f(0^{B_i}) = 1$. \end{proof} For any $d \geq 1$, define the moment map $m_d: \R \to \R^d$ by $m(t) = (t, t^2,\dots, t^d)$. \begin{prop} If there exists a degree $d$ function $f: \{0,1\}^n \to \{0,1\}$ with block sensitivity $b$, then there exists $\tau \in \{0,1\}$ such that the following set of linear inequalities has a solution $p \in \R^d$: \bea \nn \langle p, m_d(1) \rangle &=& 1 \\ \label{LP} 0 \leq \langle p, m_d(k) \rangle &\leq & 1 \,\, \text{ for each } k \in \{2, \dots, b-1\} \\ \nn \langle p, m_d(b) \rangle &=& \tau \eea \end{prop} \begin{proof} If such an $f$ exists, then let $q(x_1, \dots, x_b) = \frac{1}{b!}\sum_{\sigma \in S_b}g(x_{\sigma(1)}, \dots, x_{\sigma(b)})$, where $g$ comes from Fact \ref{bs_reduction}, and set $\tau = g(1, 1, \dots, 1)$. It is well known (see \cite{BdW}) that there is a univariate polynomial $p: \R \to \R$ of degree at most $d$ such that for any $x \in \{0,1\}^b$, $q(x_1, \dots, x_b) = p(x_1 + \dots + x_b)$. For each $k \in \{1, \dots, b\}$, $p(k)$ is therefore the average value of $g$ on boolean vectors with hamming weight $k$, so in particular $p(k) \in [0,1]$. We also know $p(0) = g(0) = 0$, $p(b) = g(1,\dots, 1) = \tau$, and $p(1) = \frac{1}{n}\sum_i g(e_i) = 1$, and hence the coefficients of $p$ provide a solution to the set of linear inequalities. \end{proof} Using the simplex method with exact (rational) arithmetic in Maple, we compute the largest $b=b(d)$ for which the LP (\ref{LP}) is feasible for $1 \leq d \leq 14$, which yields upper bounds on block sensitivity for low degree boolean functions. These bounds are summarized in Table \ref{bs table}. \begin{table}[] \resizebox{0.9\textwidth}{!}{% \begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|l|l|l|l|} \hline $\deg(f)$ &1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 \\ \hline $\text{bs}(f) \leq$ &1 & 3 & 6 & 10 & 15 & 21 & 29 & 38 & 47 & 58 & 71 & 84 & 99 & 114 \\ \hline \end{tabular}% } \caption{LP bounds on block sensitivity for low degree functions.} \label{bs table} \end{table} Setting $W(b, d) = 0$ for $b > b(d)$, for each $d = 1, \dots, 14$, we can recompute the recursive bounds from Lemma \ref{bd_recursive} and obtain $$W(30^2, 30) \leq 4.41571 \implies W^* \leq 4.4158.$$ \noindent\textbf{Remark:} The largest known separation between block sensitivity and degree is exhibited by a function on 6 variables with degree 3. By a tensorization lemma in \cite{Tal}, any degree $d$ function $f$ with block sensitivity $b$ yields an infinite family of boolean functions $f_k$ with $\deg(f_k) = d^k$ and $\text{bs}(f_k) \geq b^k$. Hence, if an entry $(d, b(d))$ in Table \ref{bs table} is tight for some $d \geq 4$, then by Fact \ref{bs_reduction} there is a function on $b(d)$ variables exhibiting a larger-than-currently-known separation between degree and block sensitivity. If $\bs(f) = \deg(f)^{\log_3(6)}$ is in fact the optimal separation, then our techniques would show $W^* < 3.96$. \begin{table}[] \resizebox{\textwidth}{!}{% \begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|l|l|l|l|} \hline $\deg(f)$ &1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 \\ \hline $W(f) \leq$ &0.5 & 1 & 1.5 & 2.0 & 2.5 & 3.0 & 3.5 & 3.9375 & 4.096 & 4.203 & 4.273 & 4.311 & 4.335 & 4.348 \\ \hline \end{tabular}% } \caption{Bounds on $W(f)$ for low degrees, obtained using Lemma 6 and Table \ref{bs table}. } \label{W table} \end{table} \section{Discussion and concluding remarks}\label{other} While our methods are unlikely to produce the optimal $W^*$, they do suggest a few interesting questions. \begin{itemize} \item The proof seems to suggest that block sensitivity limits junta size, and for small $d$, the values of $W(b,d)$ are much lower when $b \ll d^2$ than when $b \sim d^2$. A classical result of Simon \cite{Simon} says that $|R(f)| \leq s(f)4^{s(f)}$. Interestingly, we can obtain a proof of a weaker version of Simon's theorem using only the techniques in this paper and in \cite{CHS}. The idea is to define an analogue of $W$ for sensitivity instead of degree: $$S(f) := \sum_{i \in R(f)} 2^{-s_i(f)}, \, \, \, \text{ where } s_i(f) := \max_{\{x \,: \,f(x^i) \neq f(x)\}} (s_x(f)+s_{x^i}(f)). $$ \textbf{Claim:} \textit{Fact \ref{fact} holds with $\deg_i(f)$ replaced by $s_i(f)$.} \textit{Proof:} Since $s_i(f) \geq \max\{s_i(f_0), s_i(f_1)\}$ is clear from the definitions, we can assume that $f_0$ does not depend on $x_i$. Without loss of generality suppose $j = 1$ and that $y$ has $f(1, y) = 1 \neq f(1, y^i)$ and $s_i(f_1) = s_y(f_1) + s_{y^i}(f_1)$. First suppose $f_0(y) = 0$. Since $f(1, y) = 1$, this means $f$ is also sensitive to $j$ at input $(1, y)$, and so $s_i(f) \geq 1 + s_i(f_1)$. If $f_0(y) = 1$, then $f_0(y^i)$ is also 1 because $f_0$ does not depend on $x_i$. But then $f$ is sensitive to $j$ at input $(1,y^i)$, and so either way $s_i(f) \geq 1 + s_i(f_1)$, which implies the claim. \qed From here we arrive at the analogue of (\ref{summed_ineq}), and we can proceed in a number of ways. As in the proof of Lemma \ref{bd_recursive}, we can restrict maximum degree monomials until we run out of block sensitivity, yielding $|R(f)| \leq \deg(f)\cdot\text{bs}(f)\cdot 4^{s(f)}$. In any case, it seems reasonable to conjecture a Nisan-Szegedy theorem for block sensitivity, namely that any boolean function $f$ is a $\text{poly}(\text{bs}(f))\cdot2^{\text{bs}(f)}$-junta. \item More generally, it would be interesting to characterize certain \emph{ternary} relationships between complexity measures. Many of the examples we know which achieve optimal or best-known separations between two measures tend to have the property that a third measure is equal or very close to one of the other two. (For example, the best known gap of the form $\text{bs}(f) \ll \deg(f)$ is attained by a Tribes function on $n$ variables with $\deg(f) = n$ and $\text{bs}(f) = s(f) = C(f) = \sqrt{n}$.) Moreover, these examples almost always have $|R(f)| = \text{poly}(\deg(f))$. Meanwhile, the known examples of functions with nearly-optimal junta size do not exhibit any super-constant separation between the measures $\deg(f), \text{bs}(f), s(f)$ and $D(f)$. (It is possible to hybridize small, well-separated functions with large juntas, but the separations and the junta size both suffer some loss.) \item Finally, Table \ref{bs table} suggsts that $\text{bs}(f) \leq c_0\deg(f)^2$, for $c_0 \approx 0.59$. If you enjoyed reading this paper, perhaps you would enjoy trying to compute the optimal value of $c_0$. One consequence of showing that $\text{bs} < \deg^2$ is that any separation between $\text{bs}$ and degree proven by simply tensorizing a single example would necessarily (in the absence of more sophisticated arguments) look like $\text{bs}(f) \geq \deg(f)^{2 - \epsilon}$, for some $\epsilon > 0$. \end{itemize}

https://arxiv.org/abs/2008.08172

Curves on the torus intersecting at most k times

We show that any set of distinct homotopy classes of simple closed curves on the torus that pairwise intersect at most $k$ times has size $k + O(\sqrt{k} \log k)$. Prior to this work, a lemma of Agol, together with the state of the art bounds for the size of prime gaps, implied the error term $O(k^{21/40})$, and in fact the assumption of the Riemann hypothesis improved this error term to the one we obtain $O(\sqrt{k} \log k)$. By contrast, our methods are elementary, combinatorial, and geometric.

\section{Introduction} Let $T \approx \mathbb{R}^2/\mathbb{Z}^2$ be the closed oriented surface of genus one. We indicate the homotopy class of an embedding of $S^1$ briefly by `curve'. By pulling a curve tight and lifting it to the universal cover, the collection of curves on $T$ is in one-to-one correspondence with slopes $\mathbb{Q} \cup \{\infty\}$. From this vantage point, the \emph{intersection number} of a pair of curves on $T$ (that is, the minimum possible number of intersection points among representatives from the pair of homotopy classes) can be computed explicitly via \[ \iota\left( \frac pq , \frac ab\right) = \ | \ pb-qa \ | \ ~. \] A collection of curves is called a \emph{$k$-system} when any pair of curves has intersection number at most $k$. Let $\eta_S(k)$ equal the maximum size of a $k$-system on the closed surface $S$. It was first shown by \cite{JMM} that $\eta_S(k)$ goes to infinity with $k$. The determination of the growth rate of $\eta_S(k)$, as a function of both $k$ and the genus $g$ of $S$, is a subtle counting problem, about which much remains unknown \cite{Przytycki, Aougab, ABG, Greene1, Greene2}. Notably, Greene has used probabilistic methods, leveraging the hyperbolic geometric bounds of Przytycki, to obtain $\eta_S(k) =O( g^{k+1}\log g)$, when $k$ is fixed and $g$ grows \cite[Thm.~3]{Greene2}. For the study of $\eta_S(k)$ with $g$ fixed, the simplest nontrivial case is evidently $S=T$. While studying Dehn filling slopes of 3-manifolds, Agol observed that $\eta_T(k)$ is at most one more than the smallest prime greater than $k$ \cite{Agol}, and via the Prime Number Theorem this implies $\frac{\eta_T(k)}k \to 1$ as $k\to \infty$ \cite{Agol-unpub}. More can be said. The size of prime gaps, large and small, is a major field of study. The currently best upper bound is due to Baker-Harman-Pintz, which, together with Agol's observation, implies that $\eta_T(k) = k+O(k^{21/40})$ \cite{BHP}. Cram\'er showed that a positive resolution of the Riemann hypothesis would provide $\eta_T(k) = k + O(\sqrt k \log k)$ \cite{Cramer1}, and he formulated a stronger conjecture that would imply $\eta_T(k) = k+O((\log k)^2)$ \cite{Cramer2}; although there seems to be general suspicion in the analytic number theory community that Cram\'er's error term should be replaced by $O((\log k)^{2+\epsilon})$ \cite{Granville, OSS}. All of these estimates pass through Agol's remarkable prime number bound, but it is reasonable to be skeptical about whether estimation of $\eta_T(k)$ should depend on such notoriously subtle and difficult questions. The purpose of this note is to sharpen currently available estimates, without reference to fine data about the distribution of the primes. Our methods are elementary, combinatorial, and hyperbolic. \begin{theorem} \label{main thm} There is a constant $C>0$ so that $\eta_T(k)\le k+C\sqrt k \log k$. \end{theorem} As for sharpness, it deserves remarking that there is a dearth of nontrivial lower bounds for $\eta_T(k)$. In fact, we are unaware of any example of a $k$-system of size $k+7$ on the torus. The function $\eta_T(k)$ admits a dual formulation: let $\kappa_S(n)$ indicate the minimum, taken over collections of $n$ curves on $S$, of the maximum pairwise intersection. (For clarity, we will often use `$n$' to indicate the size of a set of curves and `$k$' for an intersection number.) It is not hard to see that \[ \eta_S(k) = \max \{ n: \kappa_S(n) \le k\} \ , \text{ and } \ \kappa_S(n) = \min \{k : \eta_S(k)\ge n\}~. \] Our path towards \autoref{main thm} will be to first estimate $\kappa_T(n)$ from below. There is a kind of convexity to exploit in the study of $\kappa_T$, originally observed by Agol. As remarked above, curves on $T$ are in correspondence with slopes $\mathbb{Q}\cup \{\infty\}$. The latter form the vertices of the \emph{Farey complex} $\mathcal{F}$, in which a set of slopes form a simplex when they pairwise intersect once. Any collection of $n$ curves that is maximal with respect to inclusion among $k$-systems determines a collection of vertices in $\mathcal{F}$ so that the induced simplicial complex is a triangulated $n$-gon (see \autoref{lem: convexity} for detail). Conversely, any triangulation of an $n$-gon can be realized as a subcomplex of $\mathcal{F}$, in a way that is unique up to the action of $\mathrm{PSL}(2,\mathbb{Z}) \curvearrowright \mathcal{F}$ by simplicial automorphism. As $\mathrm{PSL}(2,\mathbb{Z})$ preserves the intersection form, the (multi-)set of pairwise intersection numbers is a well-defined function on the set of triangulations of an $n$-gon. The set of triangulations of an $n$-gon forms the vertex set of a well-studied simplicial complex in combinatorics called the \emph{associahedron} $\mathcal{A}_n$ (there is a slight indexing issue; the object we refer to as $\mathcal{A}_n$ is the $(n-2)$-dimensional associahedron). One obtains a `max intersection' function $\kappa: \mathcal{A}_n \to \mathbb{N}$ induced by the intersection form on $\mathcal{F}$, and the above discussion leads to $\kappa_T(n)=\min \kappa$ (see \autoref{prop:kappa=kappa}). \autoref{main thm} follows from the following: \begin{theorem} \label{thm: kappa} There is a constant $C>0$ so that, for any $\tau\in \mathcal{A}_n$, we have $\kappa(\tau) \ge n - C \sqrt n \log n $. \end{theorem} We briefly describe the proof of this theorem. The Farey complex $\mathcal{F}$ admits a natural embedding into a compactification of the hyperbolic plane $\mathbb{H}^2 \,\cup\, \partial_\infty \mathbb{H}^2$, so that the vertices of $\mathcal{F}$ embed naturally as $\mathbb{Q} \cup\{\infty\} \hookrightarrow \mathbb{R}\cup\{\infty\} \approx \partial_\infty \mathbb{H}^2$, with edges between vertices mapping to geodesics. The hyperbolic plane $\mathbb{H}^2$ admits a maximal $\mathrm{PSL}(2,\mathbb{Z})$-invariant horospherical packing $\{H_{p/q}: \frac pq \in \mathbb{Q}\cup\{\infty\}\}$, where $H_{p/q}$ is centered at $p/q\in \mathbb{R}\cup\{\infty\}$, so that a set of slopes span a simplex in $\mathcal{F}$ precisely when the corresponding horospheres are pairwise tangent. (The horospheres $\{H_{p/q}\}$ are called \emph{Ford circles} in the literature \cite{Ford, ConwayGuy, BonahonVid}.) A sketch of our proof of \autoref{thm: kappa} is as follows: \begin{enumerate} \item Locate a `nice horoball' $H$ for $\tau$, so that $\mathrm{ht}(\tau,H)$, the \emph{height of $\tau$ relative to $H$}, is $O\left(\sqrt{\kappa(\tau)}\right)$. See \autoref{def:horoball height width} and \autoref{prop: exists horoball}. \item Use $H$ to construct a convex combination of pairwise intersection numbers for $\tau$ whose sum is at least $n - O(h \log h)$, where $h= \mathrm{ht}(\tau,H)$. It follows that there is a pair of horoballs of $\tau$ with intersection number at least $n-O(h \log h)$. See \autoref{prop: control with height}. \end{enumerate} The proof of \autoref{thm: kappa} is now one line: if $\kappa(\tau)\le n$, then $\kappa(\tau) \ge n-C \sqrt n \log n$. The first step above uses the hyperbolic geometry of $\mathbb{H}^2$ in an essential way, in which we exploit a simple relationship between intersection numbers, hyperbolic geometry, and Ford circles (see \autoref{lem:i vs d}). \subsection*{Organization} We describe the reduction from $\kappa_T(n)$ to $\kappa:\mathcal{A}_n\to\mathbb{N}$ in \S\ref{sec:prelim}, analyze several examples in \S\ref{sec:examples}, bound $\kappa(\tau)$ from below in \S\ref{sec:estimating kappa}, locate a good horoball for $\tau$ in \S\ref{sec:controlling height}, and prove \autoref{main thm} in \S\ref{sec:pf of main thm}. \subsection*{Acknowledgements} The authors thank Josh Greene and Ian Agol for valuable feedback on an early draft of this paper. We are especially grateful to Ian Agol for sharing with us an unpublished note that contained \autoref{lem: convexity}, and for suggesting an alternative to our proof of \autoref{prop: exists horoball} (see \autoref{rem: Agol proof}). \section{Preliminaries} \label{sec:prelim} We collect here some useful facts about intersection numbers and the Farey graph $\mathcal{F}$. For more of the beautiful connections between hyperbolic geometry, the Farey graph, continued fractions, and Diophantine approximation, we suggest the reader consult \cite{Hatcher, Series, Series2, Springborn}. \subsection{Horoballs in trees, Farey labellings, and intersection numbers} Dual to a triangulation of an $n$-gon $\tau\in\mathcal{A}_n$ there is a trivalent tree with $n$ leaves embedded in the plane, which we refer to as $\tau^*$. Because $\tau^*$ is embedded in the plane, the three edges incident to vertices of $\tau^*$ are cyclically ordered. Hence any non-backtracking path in $\tau^*$ induces a sequence of left-right turns. The vertices of $\tau$ (that is, the slopes of the $k$-system) correspond to `horoball' regions in $\tau^*$: \begin{definition}[`Horoballs in trees'] \label{def:horoball} A \emph{horoball} of $\tau$ is a union of edges in a path of the dual tree $\tau^*$ that is composed of uni-directional turns (that is, only left or only right turns), which is moreover maximal with respect to inclusion among all such uni-directional subsets of $\tau^*$. \end{definition} For any triangulation $\tau$, the dual tree $\tau^*$ admits an orientation-preserving embedding to the regular trivalent tree dual to $\mathcal{F}$. Any such choice of an embedding determines a map from horoballs of $\tau^*$ to $\mathbb{Q}\cup \{\infty\}$, by recording the center of the corresponding horoball in $\mathbb{H}^2$. \begin{definition}[`Farey labellings and intersection numbers'] A \emph{Farey labelling} of $\tau$ is the map from horoballs to $\mathbb{Q}\cup\{\infty\}$ obtained from an orientation-preserving embedding from $\tau^*$ to the tree dual to $\mathcal{F}$. The \emph{intersection number} $\iota(H_1,H_2)$ of a pair of horoballs $H_1$ and $H_2$ is given by the intersection number of the slopes corresponding to $H_1$ and $H_2$ in a Farey labelling of $\tau$. \end{definition} \noindent We leave it as an exercise for the reader to show that intersection numbers in $\tau$ are well-defined. Farey labellings are especially pleasant because the vertices spanning a simplex of $\mathcal{F}$ satisfy a remarkably simple relationship. Namely, if $\frac pq$ and $\frac ab$ span an edge of $\mathcal{F}$, then the two other vertices of $\mathcal{F}$ that span a triangle with $\frac pq$ and $\frac ab$ are $\frac{p+ a}{q+ b}$ and $\frac{p-a}{q- b}$; this is the `Farey addition' rule. Farey addition can be used to construct a Farey labelling of $\tau$: Choose labels $1/0$, $0/1$, and $1/1$ for the three horoballs incident to some vertex of $\tau^*$, and use Farey addition to successively add labels to neighboring horoballs. \subsection{Monotonicity of intersection numbers and left-right sequences} \label{subsec:monotonicity} The intersection number $\iota(H_1,H_2)$ admits a description more intrinsic to the structure of $\tau^*$, which we now describe. There is a unique (possibly degenerate) non-backtracking path $\sigma$ between the pair of horoballs $H_1$ and $H_2$, and this path determines a sequence of left-right turns $(\ell_1,\ell_2,\ldots,\ell_s)$, where $\sigma$ makes $\ell_1$ turns in the same direction, followed by $\ell_2$ turns in the opposite direction, etc. The quantity $\iota(H_1,H_2)$ is given by the numerator of the continued fraction with coefficients $(\ell_1,\ell_2,\ldots,\ell_s)$ \cite[Thm.~5.3]{GLRRX}. \begin{remark} \label{rem:ambiguity} Observe that there is ambiguity in this computation of $\iota(H_1,H_2)$. For one, the non-backtracking path $\sigma$ may go either from $H_1$ to $H_2$, or from $H_2$ to $H_1$. Moreover, one must declare that $\sigma$ is starting with either `left' or `right' at its origin vertex, so that it can be observed whether $\sigma$ is switching directions or not at later vertices. These choices may be made arbitrarily and independently, and this ambiguity has no affect on the calculation of $\iota(H_1,H_2)$. See \cite[Fig.~3, Ex.~1]{GLRRX}. \end{remark} This viewpoint suggests a certain monotonicity. \begin{lemma}[`Monotonicity of intersection numbers'] \label{lem:monotonicity} Suppose that $\sigma$ and $\sigma'$ are non-backtracking paths with respective left-right sequences $(\ell_1,\ldots,\ell_s)$ and $(\ell_1',\ldots,\ell_{s'}')$. If $s'\ge s$ and $\ell_i'\ge \ell_i$ for each $i=1,\ldots,s$, then the intersection number determined by $\sigma'$ is at least that determined by $\sigma$. \end{lemma} \begin{proof} This lemma is almost exactly \cite[Lem.~5.5]{GLRRX}, with the sole difference that we may have $s'>s$. Therefore to prove the claim we may assume that $\sigma'$ contains $\sigma$ as an initial subpath. Choose a Farey labelling with label $1/0$ at the horoball forming the origin of $\sigma$ (and $\sigma'$), and labels $0/1$ and $1/1$ at the two neighboring horoballs that intersect $\sigma$. Compare the denominators of the Farey labels of the horoballs at the terminuses of $\sigma$ and $\sigma'$; because these labels are computed using Farey addition, it is evident that the denominator of the horoball for $\sigma'$ is at least that of $\sigma$. The intersection number of any horoball with $1/0$ is given by the denominator of its Farey label, so the claim follows. \end{proof} \subsection{Intersection numbers and hyperbolic distance} The quotient of $\mathbb{H}^2$ by $\mathrm{PSL}(2,\mathbb{Z})$ is a hyperbolic orbifold with one cusp and two orbifold points, one of order $2$ and one of order $3$. The preimage of the maximal horoball neighborhood of the cusp under the covering projection $\mathbb{H}^2 \to \mathbb{H}^2/\mathrm{PSL}(2,\mathbb{Z})$ is $\mathcal{H}=\{H_{p/q}:p/q\in\mathbb{Q}\cup\{\infty\}\}$, a $\mathrm{PSL}(2,\mathbb{Z})$-invariant collection of horoballs centered at the completed rationals. The following lemma is an exercise in hyperbolic geometry. \begin{lemma} \label{lem:distance from horoball} Every point in $\mathbb{H}^2$ is within $\log\frac2{\sqrt 3}$ of a horoball in $\mathcal{H}$. \end{lemma} There is a simple fundamental relationship between intersection numbers of curves on the torus and hyperbolic distance between the corresponding horoballs. \begin{lemma} \label{lem:i vs d} We have $\ \displaystyle d_{\mathbb{H}^2}\left(H_{p/q},H_{a/b} \right) = 2 \log \iota\left( \frac pq , \frac ab \right) ~$ for any $\frac pq, \frac ab \in \mathcal{F}$. \end{lemma} \begin{proof} Applying an element of $\mathrm{PSL}(2,\mathbb{Z})$, we may assume that $p/q=\infty$ in the upper half-plane model for $\mathbb{H}^2 \, \cup \, \partial_\infty \mathbb{H}^2$. The horosphere $H_{a/b}$ is given by $\{z: \left| z-(\frac1b + \frac i{2b^2}) \right| = \frac1{2b^2}\}$ (see e.g.~\cite{Athreya}), so \[ d_{\mathbb{H}^2}\left(H_{p/q},H_{a/b} \right) = d_{\mathbb{H}^2}\left( a/b + \frac i{b^2} , a/b + i\right) = 2\log b = 2 \log \iota\left( \frac pq , \frac ab \right) ~. \qedhere \] \end{proof} \subsection{Width and height for horoballs}\label{sec:width/height} The interior of each edge of $\tau^*$ is incident to exactly two horoball regions. Thus, for any choice of horoball $H$ in $\tau^*$, there are exactly two other horoball regions, distinct from $H$, that are incident to the interiors of the extreme edges of $H$. Call these $H_1$ and $H_2$. \begin{definition} \label{def:horoball height width} The \emph{width of $\tau$ relative to $H$} is $w = \iota(H_1,H_2)$. The \emph{height of $\tau$ relative to $H$} is \[ \mathrm{ht}(\tau,H) = \max\, \{ \ \iota(H, H') \ : H' \text{ is a horoball of }\tau \ \}~. \] \end{definition} For the remainder of this article, we will suppress the difference between the triangulation $\tau\in\mathcal{A}_n$ and its dual tree $\tau^*$. The translation between them is quite natural, and the difference can henceforth be understood from context. \subsection{The two kappas} Recall from the introduction the quantity $\kappa_T(n)$, which is the minimum, taken over collections $\Gamma$ of $n$ curves on $S$, of the maximum pairwise intersection number of curves in $\Gamma$. \begin{definition}[`Max Intersection Function'] \label{def:kappa} The function $\kappa:\mathcal{A}_n\to \mathbb{N}$ is defined by \[ \kappa(\tau) = \max\{\ \iota(H_1,H_2) : H_1 \text{ and }H_2 \text{ are horoballs of }\tau \ \}~. \] \end{definition} \noindent As noted in the introduction, we claim that $\kappa_T(n)= \min \{ \kappa(\tau): \tau\in \mathcal{A}_n \}$. That $\kappa_T(n)\le \min \kappa$ is easy: for any $\tau\in\mathcal{A}_n$, choose a Farey labelling. The quantity $\kappa(\tau)$ is equal to the maximum pairwise intersection number of the $n$ slopes obtained in this Farey labelling, and $\kappa_T(n)$ is the minimum of the maximum pairwise intersection of any $n$ slopes, so $\kappa_T(n) \le \kappa(\tau)$ for each $\tau\in\mathcal{A}_n$. The reverse inequality is slightly less obvious, and relies on a certain convexity of maximal $k$-systems in $\partial_\infty \mathbb{H}^2$. The following lemma makes this precise.\footnote{We are grateful to Ian Agol for sharing an unpublished note with us which contained \autoref{lem: convexity} \cite{Agol-unpub}.} \begin{lemma} \label{lem: convexity} If $\Gamma$ is a $k$-system on $T$ which is maximal with respect to inclusion among $k$-systems, then $\mathcal{F}$ induces a triangulation of the $n$-gon which forms the convex hull of $\Gamma\subset \partial_\infty \mathbb{H}^2$. \end{lemma} \begin{proof} Let $g_{ab}$ indicate the geodesic in $\mathbb{H}^2$ with endpoints $a,b\in\partial_\infty \mathbb{H}^2$. Suppose that $\alpha,\beta\in\Gamma\subset \mathcal{F}$, that $\Delta$ is a Farey triangle intersecting $g_{\alpha\beta}$, and that $\delta$ is a vertex of $\Delta$ (and, hence, of $\mathcal{F}$). \autoref{lem:monotonicity} implies that both $\iota(\delta,\alpha)$ and $\iota(\delta,\beta)$ are at most $\iota(\alpha,\beta)$, which is at most $k$ by assumption. For any $\gamma\in\Gamma$, either $g_{\alpha\gamma}$ or $g_{\beta\gamma}$ intersect $\Delta$, so it follows that $\iota(\gamma,\delta)\le k$ as well. Maximality of $\Gamma$ implies that $\delta\in\Gamma$, so the convex hull of $\Gamma$ is equal to the union of Farey triangles spanning elements of $\Gamma$. \end{proof} This demonstrates that $\kappa_T(n) \ge \min \kappa$, so we may conclude: \begin{proposition} \label{prop:kappa=kappa} We have $\displaystyle \kappa_T(n) = \min_{\tau\in\mathcal{A}_n} \kappa$. \end{proposition} \section{Illustrative examples} \label{sec:examples} There are several natural elements of $\mathcal{A}_n$ that we can use to observe $\kappa:\mathcal{A}_n\to \mathbb{N}$. \begin{remark} \label{rem:unlabelled} Technically, the vertices of the associahedron correspond to triangulations of a \emph{labelled} convex polygon. Notice however that the max intersection function $\kappa:\mathcal{A}_n\to\mathbb{N}$ is invariant under permutation of labels, so we can safely refer to $\kappa(\tau)$ for elements $\tau$ of $\mathcal{A}_n$ without reference to a particular ordering of the horoballs of $\tau$. \end{remark} \begin{itemize} \item The element $\mathrm{ch}(n)\in\mathcal{A}_n$ (for `chain') contains a horoball of width $n$. We have $\kappa(\mathrm{ch}(n))=n$, and the height relative to the horoball of width $n$ is $1$. \\ \item The element $\mathrm{ach}(n)\in\mathcal{A}_n$ (for `alternating chain') contains a path of length $n-3$ that switches direction $n-4$ times. Here we have $\kappa(\mathrm{ach}(n))=F_n$, the $n$th Fibonacci number, the largest width horoball of $\mathrm{ach}(n)$ is $3$, and the height relative to any horoball is at least $F_{\lfloor \frac n2\rfloor }$. \\ \item The element $\mathrm{reg}(r)\in\mathcal{A}_n$ (for `regular'), with $n=3\cdot 2^{r-1}$, is formed by choosing the subtree of the homogeneous (infinite) trivalent tree that is induced on all vertices at combinatorial distance at most $r$ from a fixed vertex. The tree $\mathrm{reg}(r)$ contains the alternating chain $\mathrm{ach}(2r+1)$ as a subtree, and in fact we have $\kappa(\mathrm{reg}(r))=\kappa(\mathrm{ach}(2r+1))=F_{2r+1}$. The largest width of a horoball of $\mathrm{reg}(r)$ is given by $2r-1$, and the height of $\mathrm{reg}(r)$ relative to this horoball is $F_{r+1}$. \\ \item The element $\mathrm{Far}(h)\in\mathcal{A}_n$ (for the `Farey series'), with $n=2+\sum_{k\le h} \phi(k)$, is the subgraph of $\mathcal{F}$ induced on fractions in $\mathbb{Q}\cap [0,1]$ that can be written with denominator $\le h$, together with $1/0$. Observe that $\kappa(\mathrm{Far}(h)) =h^2-2h$, the largest width of a horoball of $\mathrm{Far}(h)$ is given by $h$, while the height relative to this horoball is given by $h-1$. \\ \end{itemize} We collect this information in \autoref{pic:elements} and \autoref{table}. \begin{figure} \centering \vspace{1.5cm} \begin{minipage}[t][3cm][t]{.2\textwidth} \centering \vspace{-.7cm} \includegraphics[width=4cm]{pic-chain.png} \vspace{1.5cm} \subcaption{$\mathrm{ch}(n)$} \label{pic:chain} \end{minipage}\hfill \begin{minipage}[t][3cm][t]{.2\textwidth} \centering \includegraphics[width=4cm]{pic-achain.png} \vspace{1.5cm} \subcaption{$\mathrm{ach}(n)$} \label{pic:achain} \end{minipage}\hfill \begin{minipage}[t][3cm][t]{.2\textwidth} \centering \vspace{-2cm} \includegraphics[width=4cm]{pic-reg.png} \subcaption{$\mathrm{reg}(r)$} \label{pic:chain} \end{minipage}\hfill \begin{minipage}[t][3cm][t]{.2\textwidth} \vspace{-1.4cm} \centering \includegraphics[width=4cm]{pic-Farey.png} \vspace{.6cm} \subcaption{$\mathrm{Far}(h)$} \label{pic:chain} \end{minipage}\hfill \caption{Several dual trees of elements in $\mathcal{A}_n$.} \label{pic:elements} \end{figure} \renewcommand{\arraystretch}{1.5} \begin{table}[] \begin{tabular} {| >{\centering}p{.1\textwidth} | >{\centering}p{0.2\textwidth} | >{\centering}p{0.3\textwidth} | >{\centering\arraybackslash}p{0.2\textwidth} | } \hline $\tau$ & $\kappa(\tau)$ & Largest width horoball $H$ & $\mathrm{ht}(\tau,H)$ \\ \hline \hline $\mathrm{ch}(n)$ & $n$ & $n-2$ & $1$ \\ \hline $\mathrm{ach}(n)$ & $ \Phi^n$ & $3$ & $\ge \Phi^{n/2}$ \\ \hline $\mathrm{reg}(r)$ & $ n^{2\log_2\Phi}$ & $2\log_2n $ & $ n^{\log_2\Phi}$ \\ \hline $\mathrm{Far}(h)$ & $ \frac{\pi^2}3 n $ & $\sqrt n$ & $\sqrt n$ \\ \hline \end{tabular} \vspace{.5cm} \caption{Some data for attractive elements of $\mathcal{A}_n$. Some entries include only leading-order terms, ignoring multiplicative constants. Note that $\Phi = \frac{1+\sqrt 5}2$. } \label{table} \end{table} \begin{remark} \label{rem:An diameter} Observe the large difference between $\kappa(\mathrm{ach}(n))\approx (\frac{1+\sqrt5}2)^n$ and $\kappa(\mathrm{ch}(n))=n$. Though we have not discussed it, there is a natural simplicial structure on $\mathcal{A}_n$, with edges between triangulations of an $n$-gon that differ by a single diagonal flip. It is not hard to see that the diameter of $\mathcal{A}_n$ is at most $2n$ (in fact, this quantity can be determined precisely \cite{STT, Pournin}), so it follows that the change in $\kappa$ across an edge of $\mathcal{A}_n$ can be arbitrarily large. \end{remark} \section{Estimating kappa using heights} \label{sec:estimating kappa} Let $\tau\in\mathcal{A}_n$. The strategy to obtain a lower bound for $\kappa(\tau)$ is to find a set of pairwise intersection functions $\{I_\alpha\}$ for $\tau$, and estimate the convex combination \begin{equation} \label{eq:goal} \sum_\alpha r_\alpha I_\alpha \ge n - \epsilon(n)~, \end{equation} for some set of non-negative weights $\{r_\alpha\}$ with $\sum r_\alpha =1$, and error term $\epsilon(n)$. Of course, we have $\kappa(\tau)\ge I_\alpha$ for all $\alpha$, and by convexity we must have $I_\alpha \ge n-\epsilon(n)$ for some $\alpha$. We will make use of three facts from classical analytic number theory, which we group together in a single lemma for convenience. Below we indicate the interval $\{1,\ldots,m\}$ by $[m]$ and the subset of $\{1,\ldots,m\}$ relatively prime to $s$ by $[m]_s$, e.g.~Euler's totient function is $\phi(m) = \#[m]_m.$ The number of divisors of $n$ is indicated by $d(n)$. \begin{lemma} \label{lem:number theory} We have the following estimates: \begin{align} \label{eq:euler count} \#[m]_n & = \frac mn \phi(n) + O\left(d(n)\right)~, \\ \label{eq:dirichlet} \sum_{k\le h} d(k)& = h \log h + O(h)~, \text{ and }\\ \label{eq:walfisz} \sum_{k\le h} \frac {\phi(k)}k &= O(h)~. \end{align} \end{lemma} The estimate \eqref{eq:euler count} is a standard application of M\"obius inversion \cite[Lem.~3.4]{Cohen}. The second estimate \eqref{eq:dirichlet} is a weaker version of a famous theorem of Dirichlet \cite[Ch.~3]{Apostol}, and a more precise form of \eqref{eq:walfisz} can be found in \cite{Walfisz}. (Note that the error term in \eqref{eq:euler count} is in fact $O(\vartheta(n))$, where $\vartheta(n)$ is the number of square-free divisors of $n$. However, in the sum $\sum_{j\le h} \vartheta(j)$, one finds the same order of growth as $\sum_{j\le h} d(j)$ \cite{Mertens}, so in our application, \autoref{lem:count n}, this improvement is immaterial.) In this section we will show: \begin{proposition} \label{prop: control with height} Let $H$ be a horoball of $\tau\in\mathcal{A}_n$, and let $h = \mathrm{ht}(\tau,H)$. There is a constant $C>0$ so that we have $\kappa(\tau) \ge n - C h \log h$. \end{proposition} As in \S\ref{sec:width/height}, let $H_1$ and $H_2$ be the two horoballs that are incident to $H$ along its two extreme edges. By construction, the non-backtracking path $\sigma$ from $H_1$ to $H_2$ is contained in $H$; we indicate the vertices $\sigma$ passes through in order by $p_1,\ldots,p_w$. See \autoref{fig:branches}. \begin{figure}[h] \centering \includegraphics[width=10.5cm]{pic-branches2.png} \caption{The horoball $H$ determines branches for $\tau$, and extreme horoballs $H_1$ and $H_2$.} \label{fig:branches} \end{figure} For each $j$, the complement $\tau \setminus p_j$ consists of three components, and we indicate (the closure of) the unique such component that doesn't intersect $H$ as the $j$th \emph{branch} $B(j)$. \begin{figure} \centering \includegraphics[width=10cm]{pic-numerators.png} \caption{Each branch has minimal and maximal numerators at height $k$.} \label{pic:numerators} \end{figure} Label $H$ with $1/0$, label the horoball neighboring $H$ along the edge $\overline{p_j p_{j+1}}$ with $0/1$, and label the horoball neighbor of $H$ along $\overline{p_{j-1}p_j}$ with $1/1$. Now Farey addition determines how to fill in labels for the remaining horoballs, and let $\mathrm{ht}_H \, B(j)$ indicate the maximum denominator among Farey labels for horoballs intersecting $B(j)$. The vertices of $B(j)$ at height $k$ are given by $\frac {a_1}k, \frac{a_2}k,\ldots,\frac{a_{n_k}}k$. The \emph{minimum (resp.~max) numerator at height $k$} of $B(j)$, relative to $H$, is $\min a_i$ (resp.~$\max a_i$). See \autoref{pic:numerators}. \begin{figure} \centering \includegraphics[width=10cm]{pic-heights.png} \caption{The horoballs of $\tau$ may be filtered according to heights from $H$.} \label{fig:heights} \end{figure} Observe that we may count the $n$ horoball regions of $\tau$ by filtering them according to their heights on the branches. That is, for each $i\in \mathbb{N}$, let $X(k) = \{ j : \mathrm{ht}_H \, B(j) \ge k\}$ (that is, the set of indices $j$ where the $j$th branch has height at least $k$), and let $x_k = \#X(k)$. See Figure~\ref{fig:heights}. \begin{figure} \centering \includegraphics[width=10cm]{pic-function.png} \caption{Extremal horoballs of $\tau$ at height $k$ from $H$.} \label{fig:function} \end{figure} Consider $j_k^+$ (resp.~$j_k^-$), the maximal (resp.~minimal) index of $X(k)$. Let $\ell_k$ be the maximal numerator of $B(j_k^-)$ and let $r_k$ be the minimal numerator of $B(j_k^+)$. See \autoref{fig:function}. \begin{lemma} \label{lem:count n} We have \[ \sum_{k=1}^h \phi(k)x_k \ge n + \sum_{k=1}^h \phi(k) + \sum_{k=1}^h \frac{\phi(k)}k (r_k - \ell_k) - O\left( h \log h\right)~. \] \end{lemma} \begin{proof} Observe that the sum of the number of horoballs of $\tau$ at height $k$ relative to $H$, as $k$ goes from $1$ to $h$, is exactly $n-2$. The horoballs of $B(j)$ at height $k$ relative to $H$ have size at most $\phi(k)$, so the total number of horoballs of $\tau$ at height $k$ from $H$ is at most $\phi(k) x_k$. Observe that we may count the vertices of $B(j_k^-)$ and $B(j_k^+)$ with slightly more care: When the maximal and minimal numerators of $B(j)$ at height $k$ are $\ell$ and $r$, the horoballs of $B(j)$ at height $k$ relative to $H$ are at most $\phi(k)-\#[r-1]_k$, and at most $\#[\ell]_k$. Therefore the total number of horoballs of $\tau$ at height $k$ relative to $H$ are at most \[ \phi(k)x_k - (\underbrace{\phi(k)-\#[\ell_k]_k}_{\text{overcount in }B(j_k^-)} ) - \underbrace{\#[r_k-1]_k}_{\text{overcount in }B(j_k^+)} ~. \] By \eqref{eq:euler count}, the latter is at most \[ \phi(k)x_k - \phi(k) + \frac{\ell_k}k \phi(k) - \frac{r_k-1}k \phi(k) + O(d(k))~. \] The sum of this expression as $k$ goes from $1$ to $h$ is at least $n-2$, so rearranging we find that \[ \sum_{k=1}^h \phi(k)x_k \ge n +\sum_{k=1}^h \phi(k) + \sum_{k=1}^h \frac{\phi(k)}k (r_k-\ell_k) - 2- \sum_{k=1}^h \frac{\phi(k)}k - C\sum_{k=1}^h d(k)~. \] By \eqref{eq:dirichlet} and \eqref{eq:walfisz}, the last three terms can be replaced by $O(h\log h)$. \end{proof} The reader may observe how \autoref{lem:count n} is somewhat suggestive of \eqref{eq:goal}. Given heights $k$ and $k'$, consider the intersection $I_{kk'}$ between the horoball of $B(j_k^-)$ with maximal numerator $\ell_k$ and the horoball of $B(j_{k'}^+)$ of minimal numerator $r_{k'}$. We may compute: \begin{equation*} I_{kk'} = kk' |j_{k'}^+ - j_k^-| + k'\ell_k - kr_{k'}~ \end{equation*} Therefore, we have \begin{equation*} I_{kk'}+I_{k'k} = kk'\left( \big| j_{k'}^+-j_k^-\big| + \big| j_{k}^+-j_{k'}^-\big| \right) + (k'\ell_k - kr_{k'} + k\ell_{k'}-k'r_k)~. \end{equation*} Notice that $\big| j_{k'}^+-j_k^-\big| + \big| j_{k}^+-j_{k'}^-\big| \ge x_k+x_{k'}-2$, so dividing by $kk'$ we find \begin{equation} \label{eq:sum of Is} \frac1{kk'} \left(I_{kk'} + I_{k'k}\right) \ge x_k+x_{k'} - 2 + \left(\frac{\ell_k}k- \frac{r_k}k\right) + \left(\frac{\ell_{k'}}{k'} - \frac{r_{k'}}{k'}\right)~. \end{equation} With \autoref{lem:count n} in mind, we would like to choose pairs $\{k,k'\}\subset [h]$ so that the sum over the choices made of the terms `$x_k+x_{k'}$' on the righthand side of \eqref{eq:sum of Is} is equal to $\sum \phi(k)x_k$. The following proposition makes this idea feasible. \begin{proposition} \label{prop:Gammas} For each $h\in \mathbb{N}$, there is a graph $\Gamma_h$ satisfying: \begin{enumerate} \item The vertex set of $\Gamma_h$ is given by $\{1,\ldots, h\}$. \item \label{item:valence} The valence of vertex $k$ is $\phi(k)$. \item \label{item:weights} The sum $\displaystyle \sum_{k\sim k'} \frac2{kk'}$ over the edges of $\Gamma_h$ is equal to $1$. \end{enumerate} \end{proposition} See \autoref{pic:graph} for a picture of $\Gamma_6$. \begin{proof} Declare $k\sim k'$ when $\gcd(k,k')=1$ and $k+k'>h$. For property~\eqref{item:valence}, choose a vertex $k$. Each integer $i$ relatively prime to $k$ can be shifted by $k$ to $i+k$, another integer relatively prime to $k$. For $1\le i\le k$, we may choose the maximum $n$ so that $i+nk\le h$. The result is a bijection of $[k]_k$, the integers in $[k]$ relatively prime to $k$, with the set of integers $k'$, relatively prime to $k$, less than $h$, and so that $k+k'>h$. Therefore the valence of $k$ is $\#[k]_k=\phi(k)$. For property~\eqref{item:weights}, observe that, to transform $\Gamma_h$ into $\Gamma_{h+1}$, the edges $k\sim k'$ with $k+k'=h+1$ are deleted and replaced by edges $k\sim(h+1)$ and $k'\sim(h+1)$. Because $k+k'=h+1$, the edge weight $\frac2{kk'}$ in $\Gamma_h$ is equal to the sum of edge weights $\frac2{k(h+1)}+\frac2{k'(h+1)}$ in $\Gamma_{h+1}$, so the sum $\sum_{k\sim k'}\frac2{kk'}$ is independent of $h$. For the base case $h=2$, observe that $\frac2{1\cdot2}=1$. \end{proof} \begin{figure} \centering \includegraphics[height=7cm]{pic-graph.png} \caption{The graph $\Gamma_6$ as described in \autoref{prop:Gammas}. The sum of the edge weights is $1$.} \label{pic:graph} \end{figure} Let $E$ indicate the set of edges of $\Gamma_h$. Now \autoref{prop:Gammas} and \eqref{eq:sum of Is} imply: \begin{align*} \sum_E \frac1{kk'} \left(I_{kk'} + I_{k'k}\right) &\ge \ \sum_E \ \left( x_k+x_{k'} - 2 + \left(\frac{\ell_k}k- \frac{r_k}k\right) + \left(\frac{\ell_{k'}}{k'} - \frac{r_{k'}}{k'}\right) \right) \\ &= \sum_k \phi(k) x_k \ -\ \sum_k \phi(k) \ + \ \sum_k \frac{\phi(k)}k \left( \ell_k - r_k \right) \end{align*} Applying \autoref{lem:count n}, we find that \begin{equation} \label{eq:convex estimate} \sum_E \frac1{kk'} \left(I_{kk'} + I_{k'k}\right) \ge n - O\left( h \log h\right)~. \end{equation} Because $\sum_E \frac2{kk'} =1$ by \autoref{prop:Gammas}, inequality \eqref{eq:convex estimate} proves \autoref{prop: control with height}. \section{Finding a horoball of controlled relative height} \label{sec:controlling height} For many $\tau\in\mathcal{A}_n$, there exist $H$ so that the $\mathrm{ht}(\tau,H)$ is $O(1)$, so \autoref{prop: control with height} demonstrates that $\kappa(\tau)=n-O(1)$. However, such a horoball need not exist, e.g.~every horoball of $\mathrm{ach}(n)$ has height at least $\approx (\frac{1+\sqrt 5}2)^{n/2}$. Nonetheless, $\kappa(\mathrm{ach}(n))$ is quite large (on the order $(\frac{1+\sqrt 5}2)^n$), so one might hope that it is always possible to find horoballs of small height relative to $\kappa(\tau)$. We show: \begin{proposition} \label{prop: exists horoball} There exists a constant $C>0$ so that, for any $\tau\in \mathcal{A}_n$, there exists a horoball $H$ of $\tau$ so that the height of $\tau$ relative to $H$ is controlled as $\mathrm{ht}(\tau,H) \le C \sqrt{\kappa(\tau)}$. \end{proposition} \begin{remark} \label{rem: sharpness heights} The reader can observe that the conclusion above fits the data in \autoref{table}. It is tempting to hope for an improvement of \autoref{prop: exists horoball} along the following lines: as $\kappa(\tau)$ gets closer to $\min \kappa$ (e.g.~if $\kappa(\tau)\le n$), one should be able to find horoballs of $\tau$ with relative heights $\ll \sqrt n$. On the other hand, the row containing $\tau=\mathrm{Far}(h)$, with $n\approx \frac3{\pi^2}h^2$, makes this hope seem quite remote. Indeed, $\kappa(\mathrm{Far}(h))$ is greater than $n$ only by the innocuous looking linear factor $\frac {\pi^2}3 \approx 3.3$, and yet every horoball has relative height $\ge h \approx \sqrt n$. \end{remark} \begin{proof} Let $K_1,K_2$ be horoballs of $\tau$ so that $\iota(K_1,K_2)=\kappa(\tau)$, and let $r=\log \kappa(\tau)$. By \autoref{lem:i vs d}, we have $d_{\mathbb{H}^2}(K_1,K_2) = 2r$. Let $x\in \mathbb{H}^2$ be the midpoint of the geodesic from $K_1$ to $K_2$. By \autoref{lem:distance from horoball} there is some Farey horoball $H$ so that $d_{\mathbb{H}^2}(H,x)\le \log \frac2{\sqrt3}$. The Farey horoball $H$ is incident to the geodesic between $K_1$ and $K_2$, so by convexity it must be a horoball in $\tau$. \begin{figure} \centering \includegraphics[height=10cm]{pic-thin-triangle2.png} \caption{Hyperbolicity guarantees that a horoball $K$ far from $x$ is far from some $K_i$ as well.} \label{fig:pic-thin-triangle} \end{figure} We claim that $H$ satisfies the requisite bound. Let $K$ be any other horoball of $\tau$. Because $\mathbb{H}^2$ is $\delta$-hyperbolic, the point $x$ is within $\delta$ of the geodesic segment between $K$ and $K_i$ for $i$ equal to either $1$ or $2$. A standard application of the triangle inequality (see \autoref{fig:pic-thin-triangle}) then yields \[ d_{\mathbb{H}^2}(K,K_i) \ge d_{\mathbb{H}^2}(K,x) + d_{\mathbb{H}^2}(x,K_i) - 2\delta = d_{\mathbb{H}^2}(K,x) + r -2\delta~. \] Because $d_{\mathbb{H}^2}(K,K_i)\le 2r$, we conclude that $d_{\mathbb{H}^2}(K,x)\le r+2\delta$, and \[ d_{\mathbb{H}^2}(H,K)\le d_{\mathbb{H}^2}(H,x) + d_{\mathbb{H}^2}(x,K) \le \log \frac2{\sqrt 3}+ r +2\delta~. \] By \autoref{lem:i vs d} we conclude that \[ \iota(H,K) = e^{\frac12 d_{\mathbb{H}^2}(H,K) } \le \frac2{\sqrt 3}e^{2\delta} \sqrt{\kappa(\tau)}~. \qedhere \] \end{proof} \begin{remark} \label{rem: Agol proof} Ian Agol has suggested a slightly different version of the above proof: choose Farey labels for the horoballs in $\tau$, and enlarge the horoballs by $\log \kappa(\tau)+\log\frac2{\sqrt3}$. A variation on \autoref{lem:distance from horoball} together with \autoref{lem:i vs d} implies that every trio of these horoballs mutually intersect, so by Helly's theorem there is a point $x$ in their common intersection \cite{Helly}, and one may finish as above. \end{remark} \section{From kappa to eta} \label{sec:pf of main thm} As stated in the introduction, it is an exercise to show that \begin{equation*} \label{eq:eta kappa} \eta_T(k) = \max \{ n: \kappa_T(n)\le k\}~. \end{equation*} By \autoref{thm: kappa}, we may conclude $\eta_T(k)\le \max\{n: n-C\sqrt n\log n \le k\}$. \autoref{main thm} now follows from the following lemma: \begin{lemma} \label{lem:functions} Suppose that $C>0$ is a constant, and that $f:\mathbb{R}\to\mathbb{R}$ is an increasing, sublinear function, with $f(x)=o(x)$. There is a $D>0$ so that for any $k\ge 1$ we have $\max\{ x: x- Cf(x)\le k\} \le k+ D f(k)$. \end{lemma} \begin{proof} Because $f(x)=o(x)$, there is a $C_1>0$ large enough so that $C_1-1>Cf(C_1)$. Because $f$ is sublinear, for any $k\ge 1$ we have \[ (C_1-1)k > Ck \, f(C_1) \ge C \, f(C_1k)~. \] Adding $k$ to both sides and rearranging we find \begin{equation} \label{eq:too big} C_1k-C\, f(C_1k) > k~. \end{equation} Let $F(k)=\max\{ x: x-Cf(x)\le k\}$. By \eqref{eq:too big} we have $F(k) < C_1k$. Of course, by definition of $F(k)$ we have $F(k) - Cf(F(k)) \le k$. Because $f$ is increasing and sublinear, we find \[ F(k) \le k+C f(F(k)) \le k+C f(C_1k) \le k+CC_1 f(k)~, \] as claimed. \end{proof} Because $\sqrt x\log x$ is increasing and sublinear, this completes the proof of \autoref{main thm}. \begin{remark} \label{rem:sloppy lemma} The conclusion of \autoref{lem:functions} holds under much weaker assumptions. For instance, sublinearity of $f$ can be replaced by the assumption that there is some $C_2>0$ so that $f(x+y) $ is at most $C_2 f(x) + C_2 f(y)$. \end{remark} \bigskip \bibliographystyle{alpha} \newcommand{\etalchar}[1]{$^{#1}$}

https://arxiv.org/abs/1710.06002

Covering compact metric spaces greedily

A general greedy approach to construct coverings of compact metric spaces by metric balls is given and analyzed. The analysis is a continuous version of Chvatal's analysis of the greedy algorithm for the weighted set cover problem. The approach is demonstrated in an exemplary manner to construct efficient coverings of the n-dimensional sphere and n-dimensional Euclidean space to give short and transparent proofs of several best known bounds obtained from deterministic constructions in the literature on sphere coverings.

\section{Introduction} Let $X$ be a compact metric space having metric $d$. Given a scalar $r\in\mathbb{R}_{\geq 0}$ we define the \emph{closed ball} of radius $r$ around center $x \in X$ by \[ B(x,r) = \{ y \in X : d(x,y) \leq r\}. \] The \emph{covering number} of the space $X$ and a positive number $r$ is \[ \mathcal{N}(X,r) = \min \left\{ |Y| : Y \subseteq X,\, \bigcup_{y \in Y} B(y,r) = X \right\}, \] i.e.\ it is the smallest number of balls with radius $r$ one needs to cover~$X$. Determining the covering number is a fundamental problem in metric geometry (see for example the classical book by Rogers \cite{Rogers1964a}) with many applications: compressive sensing \cite{Foucart2013a}, approximation theory and machine learning \cite{Cucker2002a} --- to name a few. \smallskip In this paper we are concerned with compact metric spaces which carry a probability measure $\omega$; a Borel measure normalized by $\omega(X) = 1$. We will assume that this probability measure behaves homogeneously on balls and is non degenerate, i.e.\ it satisfies the following two conditions: \begin{enumerate} \item[(a)] $\omega(B(x,s))=\omega(B(y,s))$ for all $x,y\in X$, and for all $s \geq 0$, \item[(b)] $\omega(B(x,\varepsilon))>0$ for all $x\in X$, and for all $\varepsilon>0$. \end{enumerate} By (a) the measure of a ball does only depend on the radius $s$ and not on the center $x$, so we simply denote $\omega(B(x,s))$ by $\omega_s$ throughout the paper. \begin{theorem} \label{thm:main} Let $(X,d)$ be a compact metric space with probability measure $\omega$ satisfying conditions~(a) and~(b). Then for every $\varepsilon$ with $r/2 > \varepsilon > 0$ the covering number satisfies \[ \frac{1}{\omega_r} \leq \mathcal{N}(X,r) \leq \frac{1}{\omega_{r-\varepsilon}} \left(\ln\left(\frac{\omega_{r-\varepsilon}}{\omega_\varepsilon}\right)+1\right). \] \end{theorem} The lower bound is obvious (using the $\sigma$-subadditivity of $\omega$). We give a proof for the upper bound in Section~\ref{sec:proof}. Our proof is based on a greedy approach to covering. We iteratively choose balls which cover the maximum measure of yet uncovered space. This greedy algorithm has been analyzed in the finite setting of the \textsc{set cover} problem which is a fundamental problem in combinatorial optimization. The \textsc{set cover} problem is defined as follows. Given a collection $S_1, \ldots, S_m$ of the ground set $\{1, \ldots, n\}$ and given costs $c_1, \ldots, c_m$ the task is find a set of indices $I \subseteq \{1, \ldots, m\}$ such that $\bigcup_{i \in I} S_i = \{1, \ldots, n\}$ and $\sum_{i \in I} c_i$ is as small as possible. Computationally, the \textsc{set cover} problem is difficult; Dinur and Steurer \cite{Dinur2014a} showed that for every $\varepsilon > 0$ it is $\mathrm{NP}$-hard to find an approximation to the \textsc{set cover} problem within a factor of $(1-\varepsilon) \ln n$. On the other hand, Chv\'atal \cite{Chvatal1979a} (previously, Johnson \cite{Johnson1974a}, Stein \cite{Stein1974a} and Lov\'asz \cite{Lovasz1975a} proved similar results for the case of uniform costs $c_1 = \ldots = c_m = 1$) showed that the greedy algorithm gives an $(\ln n + 1)$-approximation for the \textsc{set cover} problem. More specifically, Chv\'atal showed that the natural linear programming relaxation of \textsc{set cover} \[ \begin{split} \text{minimize } & \;\sum_{i=1}^m c_i x_i\\ \text{subject to } & \; x_1, \ldots, x_m \geq 0\\ & \; \sum_{i : j \in S_i} x_i \geq 1 \text{ for all } j = 1, \ldots, n \end{split} \] is at most a factor of $H_k = \sum_{n=1}^k \frac{1}{n} \leq \ln k + 1$, with $k = \max_i |S_i|$, away from an optimal solution of \textsc{set cover}. He proved this bound by exhibiting an appropriate feasible solution of the dual of the linear programming relaxation. The greedy algorithm is used to construct this feasible solution. In Section~\ref{sec:proof} we transfer Chv\'atal's argument from the finite \textsc{set cover} setting to the setting of compact metric spaces. Function $g$ appearing there features the feasible solution of the dual linear program. This will provide a proof of Theorem~\ref{thm:main}. In Section~\ref{sec:applications} we apply Theorem~\ref{thm:main} to three concrete geometric settings and we retrieve some of the best known asymptotic results, unifying many results on sphere coverings. We think that the $\mathrm{NP}$-hardness of getting $(1-\varepsilon)\ln n$-approximations for the \textsc{set cover} problem is a natural barrier for getting better asymptotic results for geometric covering problems. This might serve as an explanation why progress for example on the sphere covering problem has been very slow since the initial work of Rogers \cite{Rogers1964a}. We are not the first observing the strong relation between geometric covering problems and \textsc{set cover}\footnote{In fact, we realized this only after we, in an attempt to understand geometric covering problems from an optimization point of view, wrote down the main body of this paper.}. In recent papers, Artstein-Avidan and Raz \cite{Artstein2011a}, Artstein-Avidan and Slomka \cite{Artstein2015a} and especially Nasz\'odi \cite{Naszodi2016a} used the results of Lov\'asz \cite{Lovasz1975a} to unify old results and prove new results on geometric coverings. However, they apply the results from \textsc{set cover} directly after choosing a finite $\varepsilon$-net. Since we consider an infinite analogue of \textsc{set cover} we do not need to use an $\varepsilon$-net and by this we sometimes get slightly better constants and more importantly we think that the analysis becomes rather beautiful. Using the relation between geometric covering problems and \textsc{set cover} has already turned out to be fruitful: Prosanov \cite{Prosanov2017a} found new upper bounds for the chromatic number of distance graphs on the unit sphere, Nasz\'odi and Polyanskii \cite{Naszodi2017a} studied multi covers by this approach. \section{Proof of Theorem~\ref{thm:main}} \label{sec:proof} We shall prove that the following greedy algorithm (Algorithm~\ref{algo:Chvcover}) will provide a covering of $X$ with at most \[ \frac{1}{\omega_{r-\varepsilon}} \left(\ln\left(\frac{\omega_{r-\varepsilon}}{\omega_\varepsilon}\right)+1\right) \] many balls of radius $r$. \begin{algorithm}[h] \caption{Greedy algorithm}\label{algo:Chvcover} \begin{algorithmic}[1] \State $i \gets 0$ \State $S_x^i = B(x,r-\varepsilon)$ for all $x\in X$ \While {$\bigcup_{j=1}^i B(y^j,r) \neq X$} \State $i \gets i+1$ \State Choose $y\in X$ with $\omega(S_y^{i-1})\geq \omega(S_x^{i-1})$ for all $x\in X$ \State $y^i = y$ \State $S_x^i = S_x^{i-1}\setminus S_y^{i-1}$ for all $x\in X$ \EndWhile \end{algorithmic} \end{algorithm} We split the proof into three lemmas where the following identity will become important: \begin{equation} \label{eq:fundamental} S_x^{i-1} = B(x,r-\varepsilon) \setminus \bigcup_{j=1}^{i-1} B(y^j, r-\varepsilon). \end{equation} \smallskip The first lemma states that the step of the algorithm when we want to choose $y \in X$, with $\omega(S^{i-1}_y) \geq \omega(S^{i-1}_x)$ for all $x \in X$, is indeed well-defined. \begin{lemma} \label{lem:welldefined} In every iteration $i$ the supremum $\sup\{\omega(S_x^{i-1}) : x \in X\}$ is attained. \end{lemma} \begin{proof} We shall show that the function $f_i \colon X \to \mathbb{R}$, $f_i(x)=\omega(S_x^{i-1})$ is continuous for every iteration $i$. This implies that $f_i$ attains its maximum since $X$ is compact. For $x, y \in X$ we have \[ \begin{split} |f_i(x)-f_i(y)| & = |\omega(S_x^{i-1}) - \omega(S_y^{i-1})|\\ & = |\omega(S_x^{i-1} \setminus S_y^{i-1}) + \omega(S_x^{i-1} \cap S_y^{i-1})\\ & \qquad - (\omega(S_y^{i-1} \setminus S_x^{i-1}) + \omega(S_y^{i-1} \cap S_x^{i-1}))|\\ & = |\omega(S_x^{i-1} \setminus S_y^{i-1}) - \omega(S_y^{i-1}\setminus S_x^{i-1})|\\ & \leq \max\{\omega(S_x^{i-1}\setminus S_y^{i-1}), \omega(S_y^{i-1}\setminus S_x^{i-1})\}. \end{split} \] Without loss of generality, the maximum is attained at $\omega(S_x^{i-1}\setminus S_y^{i-1})$. Then by \eqref{eq:fundamental} we see \[ S_x^{i-1}\setminus S_y^{i-1} \subseteq B(x,r - \varepsilon) \setminus B(y,r - \varepsilon). \] By the triangle inequality \[ B(x,r - \varepsilon) \setminus B(y,r - \varepsilon) \subseteq B(y,r - \varepsilon +d(x,y))\setminus B(y,r - \varepsilon). \] Now consider the indicator function $\mathbbm{1}_{B(y,r - \varepsilon +d(x,y))\setminus B(y,r - \varepsilon)}$. When $y$ tends to $x$, then we have a monotonously decreasing sequence of measurable functions tending to $0$. By applying the theorem of monotone convergence we obtain that the integral \[ \int \mathbbm{1}_{B(y,r - \varepsilon +d(x,y))\setminus B(y,r - \varepsilon)}(z)\, d\omega(z) \] tends to $0$ as well. Hence, $f_i(y)$ tends to $f_i(x)$. \end{proof} The second lemma states that the algorithm terminates after finitely many iterations. \begin{lemma}\label{lem:termination} Algorithm \ref{algo:Chvcover} terminates after at most $\omega_\varepsilon^{-1}$ iterations and returns a covering. \end{lemma} \begin{proof} Consider the $i$-th iteration of the algorithm and suppose there exists $z\in X$ with $z\notin \bigcup_{j=1}^{i-1} B(y^j,r)$. From the triangle inequality it follows that \[ B(z,\varepsilon)\cap B(y^j,r-\varepsilon) = \emptyset. \] Together with \eqref{eq:fundamental} it implies that $B(z,\varepsilon)\subseteq S_z^{i-1}$. Choose $y \in X$ with $\omega (S^{i-1}_y) \geq \omega (S^{i-1}_x)$ for every $x \in X$. Hence we have \[ \omega(S^{i-1}_y) \geq \omega (S^{i-1}_z) \geq \omega(B(z,\varepsilon)) = \omega_\varepsilon >0, \] where $\omega_\varepsilon$ is positive by assumption (b) and thus \[ 1 = \omega(X) \geq \sum_{j=1}^i\omega(S_{y^j}^{j-1})\geq i\cdot\omega_\varepsilon, \] where the first inequality follows because the sets $S_{y^j}^{j-1}$, with $j = 1, \ldots, i$, are pairwise disjoint. So after at most $\omega_\varepsilon^{-1}$ iterations, the algorithm terminates with a covering. \end{proof} The third lemma gives the desired upper bound for the covering number. \begin{lemma} Algorithm~\ref{algo:Chvcover} terminates after at most \[ \frac{1}{\omega_{r-\varepsilon}} \left(\ln\left(\frac{\omega_{r-\varepsilon}}{\omega_{\varepsilon}}\right) + 1\right) \] iterations. In particular, this number gives an upper bound for the covering number $\mathcal{N}(X,r)$. \end{lemma} \begin{proof} Let $Y \subseteq X$ denote the covering produced by Algorithm \ref{algo:Chvcover} after $|Y|$ iterations. We shall prove \begin{equation} \label{densitystatementm} \ln\left(\frac{\omega_{r - \varepsilon}}{\omega_\varepsilon}\right)+ 1 \geq |Y|\cdot \omega_{r - \varepsilon}. \end{equation} For this we define the symmetric kernel $K \colon X\times X \to \mathbb{R}$ by \[ K(x,y)= \begin{cases} 1, & \text{if } y \in B(x,r - \varepsilon)\\ 0, & \text{otherwise.} \end{cases} \] For every $x\in X$ the following equality \[ \int K(x,y) \, d\omega(y) = \omega_{r - \varepsilon} \] holds because for every fixed $x \in X$ we have $K(x,y)=\mathbbm{1}_{B(x,r-\varepsilon)}(y)$ for all $y \in X$. We will exhibit an integrable function $g \colon X \rightarrow \mathbb{R}$ satisfying \begin{equation} \label{help1m} \int K(x,y) g(x) \, d\omega(x) \leq \ln\left(\frac{\omega_{r - \varepsilon}}{\omega_\varepsilon}\right)+1 \end{equation} for all $y\in X$ and satisfying \begin{equation} \label{help2m} \int g(x)\, d\omega(x) = |Y|. \end{equation} Combining \eqref{help1m} and \eqref{help2m}, we get \[ \begin{split} \ln\left(\frac{\omega_{r-\varepsilon}}{\omega_{\varepsilon}}\right) + 1 & \geq \int \int K(x,y) g(x) \, d\omega(x) d\omega(y)\\ & = \int g(x) \int K(x,y) \, d\omega(y) d\omega(x) \\ & = \int g(x) \omega_{r - \varepsilon} \, d\omega(x) \\ & = |Y| \cdot \omega_{r - \varepsilon} \end{split} \] and we have proven~\eqref{densitystatementm}. \bigskip Now we only have to exhibit the function $g$. \smallskip For brevity, we denote $\omega^{i-1}_y=\omega(S_y^{i-1})$. We define $g$ as follows: \[ g(x) = \begin{cases} (\omega_{y^i}^{i-1})^{-1}, & \text{ if } x\in S_{y^i}^{i-1},\\ 0, & \text{otherwise,} \end{cases} \] which is a valid definition since the sets $S_{y^i}^{i-1}$ are pairwise disjoint. Also observe that $g$ is an integrable function on the compact set $X$. From this definition of~$g$ we immediately get \eqref{help2m}: \[ \int g(x)\, d\omega(x) = \sum_{i=1}^{|Y|} \omega_{y^i}^{i-1} (\omega_{y^i}^{i-1})^{-1} = |Y|, \] To prove \eqref{help1m} we fix $y \in X$. We observe the equality \[ B(y,r - \varepsilon)\cap S_{y^i}^{i-1} = S_y^{i-1}\setminus S_y^i, \] which describes which part of $B(y,r - \varepsilon)$ is cut away in iteration $i$. Then, \[ \begin{split} \int K(x,y) g(x) \, d\omega(x) & = \sum_{i=1}^{|Y|} \int K(x,y) \mathbbm{1}_{S_{y^i}^{i-1} }(x) (\omega^{i-1}_{y^i})^{-1} \, d\omega(x)\\ & = \sum_{i=1}^{|Y|} \int \mathbbm{1}_{S_y^{i-1}\setminus S_y^i}(x) (\omega^{i-1}_{y^i})^{-1} \, d\omega(x)\\ & = \sum_{i=1}^{|Y|} (\omega^{i-1}_y - \omega^i_y) (\omega^{i-1}_{y^i})^{-1} \end{split} \] For $y\in X$ consider the last iteration $b$ such that \begin{equation} \label{help4m} \omega_{r-\varepsilon} = \omega(B(y,r - \varepsilon)) = \omega_y^0\geq \omega_y^1 \geq \ldots \geq \omega_y^b \geq \omega(B(y, \varepsilon)) = \omega_{\varepsilon} \end{equation} holds (here we used $r/2 > \varepsilon$). Note that $b < |Y|$. Note also that $\omega_y^{i-1} \leq \omega^{i-1}_{y^i}$ holds. We split the sum above into two parts: \[ \begin{split} \sum_{i=1}^{|Y|} (\omega^{i-1}_y - \omega^i_y) (\omega^{i-1}_{y^i})^{-1} & = \sum_{i=1}^{b} (\omega^{i-1}_y - \omega^i_y) (\omega^{i-1}_{y^i})^{-1} + \sum_{i=b+1}^{|Y|} (\omega^{i-1}_y - \omega^i_y) (\omega^{i-1}_{y^i})^{-1}\\ & \leq \sum_{i=1}^{b} (\omega^{i-1}_y - \omega^i_y) (\omega^{i-1}_{y})^{-1} + (\omega^b_y - \omega^{b+1}_y) (\omega^b_y)^{-1} \\ & \qquad + \sum_{i = b+2}^{|Y|} (\omega^{i-1}_y - \omega^i_y) \omega_{\varepsilon}^{-1}\\ & \leq \left(\sum_{i=1}^{b} (\omega^{i-1}_y - \omega^i_y) (\omega^{i-1}_{y})^{-1} + \frac{\omega^b_y - \omega_\varepsilon}{\omega^b_y} \right) \\ & \qquad + \left(\frac{\omega_{\varepsilon} - \omega^{b+1}_y}{\omega_{\varepsilon}} + \frac{\omega^{b+1}_y - \omega^{|Y|}_y}{\omega_{\varepsilon}}\right). \end{split} \] The first sum is a lower Riemann sum of the function $x \mapsto \frac{1}{x}$ in the interval $[\omega_\varepsilon, \omega_{r-\varepsilon}]$ and thus we have $\ln\left(\frac{\omega_{r - \varepsilon}}{\omega_{\varepsilon}}\right)$ as an upper bound. The second sum is clearly bounded above by $1$. Hence, \eqref{help1m} holds. \end{proof} \section{Applications of Theorem~\ref{thm:main}} \label{sec:applications} \subsection{Covering the $n$-dimensional sphere} \label{ssec:sphere} As a first application of Theorem~\ref{thm:main} we consider the problem of covering the $n$-dimensional sphere \[ X = S^n = \{x \in \mathbb{R}^{n+1} : x \cdot x = 1\}, \] equipped with spherical distance \[ d(x,y) = \arccos x \cdot y \in [0,\pi] \] and with the rotationally invariant probability measure $\omega$, by spherical caps / metric balls $B(x,r)$. Clearly, properties (a) and (b) are satisfied in this setting. Again we set $\omega_r = \omega(B(x,r))$. We are especially interested in the covering number $\mathcal{N}(S^n, r)$ when $0 < r < \pi/2$ or equivalently in the covering density defined by $\omega_r \cdot \mathcal{N}(S^n, r)$. Theorem~\ref{thm:main} says that the covering density is at most \begin{equation} \label{eq:density-bound} \frac{\omega_r}{\omega_{r-\varepsilon}} \left( \ln\left(\frac{\omega_{r-\varepsilon}}{\omega_{\varepsilon}}\right) + 1 \right). \end{equation} This upper bounds holds for every $\varepsilon$ with $0 < \varepsilon < r$. By choosing $\varepsilon$ depending on the dimension $n$ and on the spherical distance $r$ we can find an upper bound for the covering density which only depends on $n$. For this we recall a useful estimate of fractions of the form $\omega_{tr}/\omega_r$ due to B\"or\"ozky~Jr.\ and Wintsche \cite{Böröczky2003a}: \begin{equation} \label{eq:bw-bound} \frac{\omega_{tr}}{\omega_r} \leq t^n \quad \text{whenever } r < tr < \frac{\pi}{2}. \end{equation} We set $\varepsilon = r/(\mu n +1)$ with parameter $\mu > 1$ which we are going to adjust later. Furthermore, we set \[ t = \frac{r}{r-\varepsilon} = 1 + \frac{1}{\mu n} \] and \[ t' =\frac{r-\varepsilon}{\varepsilon} = \mu n. \] By using \eqref{eq:density-bound} and \eqref{eq:bw-bound} we have the following upper bound for the covering density \[ \begin{split} \frac{\omega_r}{\omega_{r-\varepsilon}} \left( \ln\left(\frac{\omega_{r-\varepsilon}}{\omega_{\varepsilon}}\right) + 1 \right) & \leq \left( 1 + \frac{1}{\mu n}\right)^n (n \ln \mu n + 1)\\ & \leq e^{1/\mu} (n \ln \mu n + 1)\\ & \leq \left(1 + \frac{1}{\mu - 1}\right) (n \ln \mu n + 1) \end{split} \] Thus we have proven: \begin{corollary} The covering density of the $n$-dimensional sphere by spherical balls is at most \[ \left(1 + \frac{1}{\mu - 1}\right) (n \ln \mu n + 1) \text{ for all } \mu > 1. \] In particular, for $\mu = \ln n$, the covering density is at most \[ n \ln n + n \ln\ln n + n + o(n). \] \end{corollary} In the asymptotic case the best known bound is $(1/2 + o(1)) n \ln n$ due to Dumer \cite{Dumer2007a} which comes from a randomized construction. Our corollary slightly improves the previously best known non-asymptotic bound $n \ln n + n \ln\ln n + 2n + o(n)$ by B\"or\"ozky~Jr.\ and Wintsche \cite{Böröczky2003a} also coming from a randomized construction. \subsection{Covering $n$-dimensional Euclidean space} \label{ssec:space} As a second application we consider coverings of $n$-dimensional Euclidean space~$\mathbb{R}^n$ by congruent balls. We get a covering of $\mathbb{R}^n$ by applying Theorem~\ref{thm:main} to the torus $\mathbb{T}^n = \mathbb{R}^n / \mathbb{Z}^n$ which is a compact metric space satisfying properties (a) and (b). Then we periodically extend the obtained covering of $\mathbb{T}^n$ to a covering of the entire $\mathbb{R}^n$ having the same covering density. We repeat the choices and calculations as in the previous section (which are slightly simpler here because clearly $\omega_{tr}/\omega_r = t^n$ holds where here $\omega$ denotes the Lebesgues measure) and get: \begin{corollary} The covering density of the $n$-dimensional Euclidean space by congruent balls is at most \[ \left(1 + \frac{1}{\mu - 1}\right) (n \ln \mu n + 1) \text{ for all } \mu > 1. \] In particular, for $\mu = \ln n$, the covering density is at most \[ n \ln n + n \ln\ln n + n + o(n). \] \end{corollary} We remark that this bound coincides with the currently best known bound by G. F\'ejes Toth \cite{FejesToth2009a} coming from a deterministic construction. The best known bound coming from a randomized construction is $(1/2 + o(1)) n \ln n$ due to Dumer \cite{Dumer2007a} \subsection{More general coverings} \label{ssec:general} At last we want to demonstrate that the greedy approach to geometric covering problems is quite flexible. It is not restricted to finding coverings of compact metric spaces by balls but can be extended to finding coverings of compact metric spaces by finite unions of balls \[ \bigcup_{i=1}^N B(y_i,r), \] where we choose the initial points $y_1,\ldots , y_N\in X$ arbitrarily. We make this statement precise in the general setting of a compact metric space $(X,d)$. Consider the group of continuous isometries of $(X,d)$, these are all continuous bijective maps $\tau \colon X \to X$ which preserve the distance between every two points $x,y \in X$. We assume that the group acts transitively on $X$ and that $\omega(\tau A) = \omega(A)$ holds for all continuous isometries $\tau$ and all measurable sets $A$. Then by the theorem of Arzel\`a-Ascoli (see for example \cite[Chapter 4.6]{Folland1999a}) the group of continuous isometries is relatively compact in the compact space of continuous maps mapping $X$ to itself equipped with the supremum norm. We need this compactness for Lemma \ref{lem:welldefined}. So we can transfer the analysis of the greedy algorithm given in Section~\ref{sec:proof} to this setting. With small modifications this extension can for example be applied to prove the following theorem due to Nasz\'odi \cite[Theorem 1.3]{Naszodi2016a}: \begin{theorem} \label{Naszth} Let $K\subseteq \mathbb{R}^n$ be a bounded measurable set. Then there is a covering of $\mathbb{R}^n$ by translated copies of $K$ of density at most \[ \inf\left\{ \frac{\omega(K)}{\omega(K_{-\delta})}\left( \ln\left(\frac{\omega\left(K_{-\delta/2} \right)}{\omega(B(0,\delta/2))}\right) +1\right) : \delta > 0, K_{-\delta} \neq \emptyset\right\}, \] where $K_{-\delta} = \{x\in K :\ B(x,\delta)\subseteq K\}$ is the $\delta$-inner parallel body of $K$. \end{theorem} Here, we only sketch the proof, though filling in the details is easy. As in Section~\ref{ssec:space} we can work on the torus $\mathbb{T}^n$. We approximate the body $K$ and its inner parallel bodies by a finite union of balls for which \[ \bigcup_{i=1}^N B(y_i, \delta) \subseteq K \] and \[ K_{-\delta} \subseteq \bigcup_{i=1}^N B(y_i, \delta/2) \subseteq K_{-\delta/2} \] holds. In the end going back from the torus $\mathbb{T}^n$ to $\mathbb{R}^n$ we get a covering of $\mathbb{R}^n$ by translated copies of $K$ with density at most \[ \inf\left\{ \frac{\omega(K)}{\omega(K_{-\delta/2})}\left( \ln\left(\frac{\omega\left(K_{-\delta/2} \right)}{\omega(B(0,\delta/2))}\right) +1\right) : \delta > 0, K_{-\delta} \neq \emptyset\right\}, \] improving the result of Nasz\'odi slightly. Another alternative of proving this bound is to verify that the proof of Theorem \ref{thm:main} also holds if we consider translates of $K_{-\delta/2}\subseteq \mathbb{R}^n$ instead of balls $B(x,r-\varepsilon)$. This further requires that $K_{-\delta}$ is nonempty and to consider translates of Minkowski sums $x + K_{-\delta/2} + B(0, \zeta)$ instead of $B(x, r - \varepsilon + \zeta)$ in the parts of the proofs of Lemmas \ref{lem:welldefined} and \ref{lem:termination} where we apply the triangle inequality. \section*{Acknowledgements} We thank Markus Schweighofer, Cordian Riener, and the anonymous referee for helpful remarks.

https://arxiv.org/abs/1201.0425

Spectral gaps of random graphs and applications

We study the spectral gap of the Erdős--Rényi random graph through the connectivity threshold. In particular, we show that for any fixed $\delta > 0$ if $$p \ge \frac{(1/2 + \delta) \log n}{n},$$ then the normalized graph Laplacian of an Erdős--Rényi graph has all of its nonzero eigenvalues tightly concentrated around $1$. We estimate both the decay rate of the spectral gap to $1$ and the failure probability, up to a constant factor. We also show that the $1/2$ in the above is optimal, and that if $p = \frac{c \log n}{n}$ for $c < 1/2,$ then there are eigenvalues of the Laplacian restricted to the giant component that are separated from $1.$We then describe several applications of our spectral gap results to stochastic topology and geometric group theory. These all depend on Garland's "p-adic curvature" method, a kind of spectral geometry for simplicial complexes. These can all be considered to be high-dimensional expander properties.

\section{Introduction} Studying the spectral properties of random matrices has played a central role in probability theory ever since Wigner's paper establishing the semi-circular law for symmetric matrices with independent centered entries above the diagonal~\cite{wigner}. The theory of these matrices is rich and well-developed, and its techniques and theorems provide great insight into the adjacency matrix of an {Erd\H{o}s--R\'{e}nyi } random graph. In this paper we study the normalized Laplacian matrix of the {Erd\H{o}s--R\'{e}nyi } random graph. In particular, we work with the binomial random graph $G(n,p),$ which has $n$ vertices and whose every edge is included independently with probability $p.$ For a connected graph $G,$ the normalized Laplacian has smallest eigenvalue $\lambda_1=0,$ and the remainder of its eigenvalues $\{\lambda_i\}_{i=2}^n$ lie in the interval $0 < \lambda_i \le 2$. The spectral gap, $\lambda_2,$ is the principal quantity of interest in many applications, and it has received much attention in the literature~\cite{Chung,ChungLuVu,ChungRadcliffe,CojaOghlan}. For the {Erd\H{o}s--R\'{e}nyi } graph, the eigenvalues $\{\lambda_i\}_{i=2}^n$ tend to cluster around $1,$ and hence we define $\lambda(G)=\max_{i \neq 1}|1-\lambda_i|.$ The quantity $1-\lambda(G)$ is sometimes referred to as the absolute gap. The methods in the previous papers are successful in establishing the correct order for $\lambda(G)$ of $C(np)^{-1/2}$ when the density of edges is sufficiently large, but they do not extend to $p$ very near the connectivity threshold $\log n / n$. Our main result is as follows. \begin{theorem} \label{thm:ergap_giant} Fix $\delta > 0$ and let \(p \geq (\frac12+\delta) \log n / n\). Let $d=p(n-1)$ denote the expected degree of a vertex. Let $\tilde G$ be the giant component of the {Erd\H{o}s--R\'{e}nyi } graph. For every fixed $\epsilon > 0,$ there is a constant $C=C(\delta, \epsilon),$ so that \[ \lambda(\tilde G)< \frac{C}{\sqrt{d}}. \] with probability at least $1-Cn\exp(-(2-\epsilon)d)-C\exp(-d^{1/4}\log n).$ \end{theorem} This result improves on a number of previous results. These earlier results are discussed in more detail in Section \ref{sec:rmt_background}. In brief, the state of the art is due to Coja--Oghlan \cite{CojaOghlan} who obtains gap $1-O(d^{-1/2})$ for $p \ge C \log n / n$, where $C> 0$ is a sufficiently large constant. We are able to extend this to $C=1,$ and appropriately modifying the statement for the giant component, we extend this to $C = \tfrac12$. We note that Theorem \ref{thm:ergap_giant} is vacuous for $p \leq \frac{1}{2} \log n / n.$ Indeed, the next result shows that for smaller values of $p,$ the gap is no longer $1-o(1).$ \begin{theorem} \label{thm:ergap_giant2} For $p$ satisfying $p = \omega(\sqrt{\log n}/n)$ and $p \leq \tfrac 12 \log n / n$ \[ \lambda(\tilde G) \geq \tfrac 12, \] with high probability. \end{theorem} Here, and throughout the paper, we use ``with high probability'' (w.h.p.) to mean that the probability approaches one as the number of vertices $n \to \infty$. We additionally use ``with overwhelming probability'' to mean that the probability approaches $1$ faster than any power of $n.$ For $p = \Omega( \sqrt{\log n}/n),$ Fountoulakis and Reed~\cite{FR} show that the mixing time is large, and hence provide a lower bound for $\lambda(\tilde G)$ in this regime. So $G(n,p)$ has $\lambda(\tilde G)$ bounded away from $0$, but at $ \frac{1}{2} \log n / n$ there is a phase transition, and at this point $\lambda(\tilde G) = o(1)$. The proof in fact shows that both $\lambda_2(\tilde G)$ and $\lambda_{|\tilde G|}$ are separated from $1$ by at least $\tfrac 12 - o(1).$ We also consider an {Erd\H{o}s--R\'{e}nyi } process version of the spectral gap theorem. In particular, we show that if random edges are added one at a time, at the moment of connectivity the random graph already has spectral gap $1-o(1)$. More precisely, we have the following. \begin{theorem} \label{thm:erstop} Let $\tau_c$ be the connection time for the {Erd\H{o}s--R\'{e}nyi } graph process $G(n,m).$ Then there is a constant $C$ so that with high probability \[ \lambda(G(n,\tau_c)) \leq C/\sqrt{\log n}. \] \end{theorem} This theorem follows immediately from Theorem~\ref{thm:stopping_time}, which contains a stronger result about the {Erd\H{o}s--R\'{e}nyi } process valid for the spectral gap of the giant component. \section*{Applications to random topology} As we will see, Theorem \ref{thm:ergap_giant} is extremely useful in the study of random topological spaces and random groups. We now provide several examples where this theorem yields sharp results. All of these new results depend on the combination of the spectral gap theorem with ``Garland's method'' and its refinements by Ballman and \'Swi\k{a}tkowski \cite{BS}, and by \.Zuk \cite{Zuk,zuk2}. \vspace{.05in}\noindent{\bf $\bullet$ Kazhdan's property (T).} Linial and Meshulam~\cite{LM06} introduce an analogous measure $Y_2(n,p)$ to the binomial random graph for random $2$-dimensional simplicial complexes. This is the probability distribution on all simplicial complexes with vertex set $[n]=\{1,2,\dots,n\}$, with complete $1$-skeleton (i.e.\ with all possible $n \choose 2$ edges), and such that each of the ${n \choose 3}$ possible $2$-dimensional faces are included independently with probability $p$. We use the notation $Y \sim Y_2(n,p)$ to indicate a complex drawn from this distribution. We first prove a structure theorem for the random fundamental group, for a certain range of $p$. \begin{theorem} \label{thm:structure} Suppose $\delta > 0$ is fixed, $$ p \ge \frac{ ( 1 + \delta) \log n}{n},$$ and $Y \sim Y_2(n,p)$. Then w.h.p.\ $\pi_1(Y)$ is isomorphic to the free product of a (T) group $G$, and a free group $F$, where the free group $F$ has one generator for every isolated edge in $Y$. \end{theorem} As a corollary, we also show that the threshold for $\pi_1(Y)$ to have property (T) agrees precisely with the homology-vanishing threshold found by Linial and Meshulam \cite{LM06}. This is true in the strong ``hitting time'' sense. \vspace{.05in}\noindent{\bf $\bullet$ Random $d$-dimensional simplicial complexes.} Meshulam and Wallach further generalize the $2$-dimensional model to random $d$-dimensional complexes $Y_d(n,p)$ \cite{MW09}. Their main result is that $ p = d \log n / n$ is a sharp threshold for vanishing of cohomology $H^{d-1}(Y, {\bf k})$ where ${\bf k}$ is a finite field of field of characteristic $0$. The proof requires delicate cocycle counting arguments. The new spectral gap results give a completely different proof of the Meshulam--Wallach theorem, in the case that $k$ is a field of characteristic $0$. The Meshulam--Wallach theorem is stronger topologically, since it can handle positive characteristic. But our new proof is very short (given the spectral gap theorem), and the result is actually sharper probabilistically. For example, we obtain ``hitting time'' results in an accompanying stochastic growth process, and also we recover a simple proof of the Poisson distribution of Betti numbers in the critical window. \vspace{.05in}\noindent{\bf $\bullet$ Density model of random groups.} Antoniuk et.\ al.\ study the phase transitions that occur in the density model of random graphs \cite{densitymodel}. Using our spectral bounds instead of those in \cite{CojaOghlan} their results can be strengthened to show a hitting time result like in the case of the Meshulam--Wallach model described above. \vspace{.05in}\noindent{\bf $\bullet$ Random flag complexes.} Using the spectral gap theorem and Garland's method, similar cohomology vanishing results were recently obtained for a different model of random simplicial complex by the second author in \cite{flag}. As a corollary, for every $d \ge 3$, there is a wide range of $p$ for which $X(n,p)$ is rationally homotopy equivalent to a wedge of $d$-spheres. \vspace{.05in}\noindent{\bf $\bullet$ Right-angled Coxeter groups.} Group cohomology of random right-angled Coxeter groups were studied in \cite{DK12}. Applying the same techniques as in the random flag complex paper \cite{flag}, it is shown that for a certain measure and range of parameter, random right-angled Coxeter groups are rational duality groups with high probability. This is actually a special case of a more general statement that shows that the same holds for random graph products of finite groups. \section*{Organization} Section \ref{sec:rmt_background} contains the background about the spectrum of the normalized Laplacian of {Erd\H{o}s--R\'{e}nyi } random graphs. Section \ref{sec:rt} does the same for our applications of our spectral results to random topology. In Section \ref{sec:est} we show how to transfer adjacency matrix estimates to the normalized Laplacian under some assumptions on the structure of the graph. In Section~\ref{sec:pbnds} we show that an {Erd\H{o}s--R\'{e}nyi } graph satisfies these structural conditions with high probability. In Section~\ref{sec:proc} we show that the Linial-Meshulam process has large gap in all its codimension-2 links. In Section~\ref{sec:cohom} we show how to apply the {Ballman--\'Swi\k{a}tkowski } criterion to prove the structure theorem for rational cohomology, and in Section~\ref{sec:propT} we show how to apply \.Zuk's criterion to prove the structure theorem for the fundamental group. In Section~\ref{sec:KSz} we apply the Kahn-Szemer\'erdi machinery to show that the adjacency matrix of the {Erd\H{o}s--R\'{e}nyi } graph has a gap of the correct order for any $p$ with $p = \Omega( \log n / n).$ Finally, we include one appendix which proves the precise versions of the tail bounds for binomial variables that we use. \section{Background: random matrices and spectral gaps} \label{sec:rmt_background} One of the central problems in modern probability theory is to understand the spectrum of a Wigner matrix, a random symmetric matrix with independent, centered, variance $1$ entries. From Wigner's celebrated semicircle law, it can be inferred that the largest eigenvalue of such a matrix is around $2\sqrt{n}.$ In fact a much stronger result is known for a large class of Wigner matrices, for which it is seen that \[ \lambda_1 \sim 2\sqrt{n} + Xn^{-1/6} \] where $X$ follows the GOE Tracy-Widom law. When the entry distributions are $\operatorname{Bernoulli}(p)$ -- i.e. when this is the adjacency matrix of an {Erd\H{o}s--R\'{e}nyi } graph -- it was recently shown by Knowles, L. Erd\H{o}s, Yau and Yin~\cite{KEYY} that for $p \gg n^{-1/3},$ the analogous results hold for the second largest eigenvalue. A related question, and what we need here, is to determine the ``gap'' between the first and second eigenvalues of the spectrum. In the Wigner case these two eigenvalues will be vanishingly close. But the entries of our matrices have nonzero, identical expectations, and this has the effect of boosting the largest eigenvalue beyond the remainder of the spectrum, producing a large gap. Estimating this gap typically arises in the study of adjacency and Laplacian matrices of graphs, where its applications include determining the mixing time of Markov chains and quantifying the expansion properties of a graph, to name a few (see~\cite{Chung} for further applications). We study the normalized Laplacian matrix of a graph. A good introduction to the properties of the normalized Laplacian are available in~\cite{Chung}. Let $\pi_+$ be the projection map onto the vertices with positive degree, let $T$ be the diagonal matrix of degrees, and let $A$ be the adjacency matrix. The normalized Laplacian is defined as \[ { L} = \pi_{+} - T^{-1/2}AT^{-1/2}, \] where $T^{-1/2}$ is taken to be $0$ in coordinates where the degree is $0.$ For the rest of the paper we let $0 =\lambda_1 \leq \lambda_2 \leq \ldots \leq \lambda_n \leq 2$ be the eigenvalues of ${ L}.$ (Note that some authors use an alternate definition of normalized Laplacian, with a $\pi_{+}$ replaced by $\operatorname{Id}$.) The principal nontrivial property we will employ about ${ L}$ is that the dimension of the kernel is equal to the number of components of $G$. An immediate consequence is that for a graph with multiple nontrivial components, $\lambda_2=0.$ In particular, for $p \ll \log n / n,$ the normalized Laplacian has no spectral gap. That said, it still makes sense to consider the spectral gap of ${ L}$ restricted to the giant component. Many of the tools which were developed for estimation of the spectral gap were designed with Wigner matrices in mind, and thus their usefulness for matrices of sparse graphs is limited. In particular, we recall the bound of F\"uredi and Koml\'os~\cite{FurediKomlos} which can be extended to show that when $p \gg \log^6 n/n,$ the spectral gap of the adjacency matrix of an {Erd\H{o}s--R\'{e}nyi } graph is of smaller order than the largest eigenvalue. Improvements along this line of reasoning bring the range of feasible $p$ to as low as $p \gg \log^2 n/n$ \cite{Vu07, ChungLuVu}. The global dependence alone is not enough to disturb some of the Wigner-like features of ${ L};$ especially, for $p \gg \log^2 n / n,$ the smallest positive eigenvalue of ${ L}$ is $1-o(1).$ More generally, Chung, Lu and Vu~\cite{ChungLuVu} show that \[ \max_{ i \neq 1} |1 - \lambda_i| \leq \frac{C}{\sqrt{np}} + \text{error}. \] As the entries of $T^{-1/2}AT^{-1/2}$ have variance roughly on the order of $1/n^2p$, the order of this bound agrees with what is seen in the Wigner case. The method used to prove these bounds is the trace method, i.e. estimating the expectation of high powers of the trace of a matrix. The state-of-the-art for this method is precisely what produces the spectral gap bound in the $p \gg \log^2 n/n$ regime. This method starts to lose effectiveness as the density of edges decreases to the connectivity threshold, where it becomes difficult to analyze the powers necessary to get the desired results. Other methods of attack include matrix martingales, which have been used to analyze the spectra of {Erd\H{o}s--R\'{e}nyi } graphs and other related models. See for example Oliveria~\cite{Oliveira} or Chung and Radcliffe~\cite{ChungRadcliffe}. However, these methods have limited success in the $p = O(\log n/n)$ regime. Finally, there is the method of Kahn and Szemer\'edi~\cite{FKSz}, first developed for bounding the spectral gap of $d$-regular graphs, which has been adapted quite successfully for estimating the spectral gap in the $p = O(\log n / n)$ regime. Coja-Oghlan~\cite{CojaOghlan} shows that with $p \geq c \log n/n,$ the gap is $1-O(1/\sqrt{np})$ with probability $1-o(1),$ and Feige and Ofek~\cite{FeigeOfek} show an analogous claim for the adjacency matrix. This method is one component of the argument in this paper. The spectral gap of the normalized Laplacian is strongly related to other probabilistic quantities of the graph, in particular to properties of simple random walk (see~\cite{Chung} for more details) and to the Cheeger constant. Direct analysis of these quantities is also possible, which then implicitly give bounds on the spectral gap. Benjamani et.\ al. take a combinatorial approach and study the Cheeger constant (also called isoperimetric constant, or conductance) throughout the evolution of the random graph process \cite{cheegerconstant}. Likewise Fountoulakis and Reed study the mixing time of simple random walk on the giant component through the conductance~\cite{FR} in the strictly supercritical regime $\tfrac{1+\epsilon}{n} < p < \tfrac{\sqrt{\log n}}{n}.$ Ding et.\ al.\ studied probabilistic aspects of the graph including the mixing time of simple random walk on the giant component as the graph emerges from the critical window \cite{peresrandomwalk}. All these works show that the giant component can be written as a well connected expanding core together with small (logarithmic size) graphs attached to the core. We also employ a version of this decomposition to analyze the spectral properties of the graph. \subsection{Gap theorem proof approach} \label{sec:gap_proof} To bound $\max_{i > 1} \left| 1 - \lambda_i\right|$ it suffices instead to bound the spectrum of what is essentially $I - { L}.$ Given the graph $G$ with vertices $\{1,2,\dots,n\}$ we define the matrix \[ M_{u,v} = \begin{cases} \frac{1}{\sqrt{\deg(u)}\sqrt{\deg(v)}}& \text{if $u$ is adjacent to $v$,} \\ 0 &\text{otherwise.} \end{cases} \] Thus if all degrees are positive we have \[M = T^{-1/2}AT^{-1/2},\] and it is easily checked that for any vertex set $W$ of a connected component of $V$, $T^{1/2}\one_{W}$ is an eigenvector with eigenvalue one. Like in~\cite{CojaOghlan}, we prove this theorem in part by comparison with the spectrum of the adjacency matrix, which exhibits no transition in the $p = \Omega(\log n/ n)$ regime. The spectra of the adjacency matrix is handled by the technique of Kahn and Szemer\'erdi, and is a slight improvement over an earlier estimate of Feige and Ofek~\cite{FeigeOfek}. Letting $S=\{x~\vert~x^t\one = 0\},$ this states that \[ |x^tAy| \leq C\sqrt{d} \|x\|\|y\| \] for all $x\in S$ and all $y \in \mathbb{R}^n,$ provided $p = \Omega(\log n / n).$ When $p > (1+\epsilon) \log n / n,$ the comparison is relatively straightforward, by virtue of the fact that with high probability all the degrees in the graph are larger than $d/M$ for some sufficiently large $M$. In particular, this means that $\|T^{-1/2}\| = O(M/\sqrt{d}).$ One must additionally show that $T^{-1/2}\one$ is nearly parallel to $\one,$ i.e. $T^{-1/2}$ nearly maps the space $S$ to itself. In sum, these two facts show that for $x\in S,$ $T^{-1/2}x$ is still nearly in $S$ and has norm $\|T^{-1/2}x\| \leq \sqrt{M}\|x\|/\sqrt{d}.$ Thus, \[ |x^t M x| = |(T^{-1/2}x) A (T^{-1/2}x)| \approx C\sqrt{d} \|T^{-1/2}x\|^2 = O(\|x\|^2d^{-1/2}), \] giving the desired result. Likewise, when $p > \frac{\log n + (\log n)^{1/2+\delta}\log\log n}{n},$ the degree of the graph is still at least $d^{1/2+\delta}$ w.h.p. In this case, the $T^{-1/2}$ still nearly maps $S$ to $S$, but now $\|T^{-1/2}x\| = O(d^{-1/4-\delta/2}).$ This allows one to show that \[ \max_{i > 1} \left| 1 - \lambda_i\right| < d^{-\delta}, \] which is essentially the approach taken by an earlier version of this paper. To get theorems all the way down to the connectivity threshold and below, where the minimum degree drops to $0,$ an additional argument is needed. This is because it is no longer the case that $\|T^{-1/2}\| = O(1/\sqrt{d}).$ The key structure theorem that allows the comparison to go through is an analysis of the graph structure surrounding low-degree vertices. Precisely, we show that near the connectivity threshold, there are no edges between low-degree vertices, and low-degree vertices and do not even have shared neighbors (see Proposition~\ref{prop:key_structure}). Thus, they are only connected through the large, high-degree core. This is enough to ensure that the desired spectral properties persist all the way down to around $p \sim 1/2 \log n / n.$ On the other hand, below $p \sim 1/2 \log n / n,$ low-degree vertices in the giant component begin to connect with high probability. Indeed, it is possible to show that there are even two degree $2$ vertices that connect to each other and the high-degree core. This is enough to ensure that $\lambda_2$ of the giant component is at most a little above $\tfrac12$ and $\lambda_n$ is at least $\tfrac32.$ \section{Random topology} \label{sec:rt} In~\cite{LM06}, Linial and Meshulam introduce an analogous measure $Y_2(n,p)$ to the binomial random graph for random $2$-dimensional simplicial complexes. This is a probability distribution over all simplicial complexes with vertex set $[n]=\{1,2,\dots,n\}$ with complete $1$-skeleton (i.e.\ with all possible $n \choose 2$ edges). Each of the ${n \choose 3}$ possible $2$-dimensional faces are included independently with probability $p$. We use the notation $Y \sim Y_2(n,p)$ to indicate a complex drawn from this distribution. Meshulam and Wallach~\cite{MW09} extend this definition to a $d$-dimensional complex, formed by taking the complete $(d-1)$-skeleton of the $n$-vertex simplex, and including $d$-dimensional faces independently with probability $p.$ The distributions can be made into stochastic growth processes in a natural way. Let $Y_2(n,m)$ be the random $2$-complex that has the uniform distribution over all simplicial complexes with $n$ vertices, $n \choose 2$ edges, and exactly $m$ two-dimensional faces. In the random complex process $\{ Y_2(n,m) \}$, faces are added one at a time, uniformly randomly from all faces which have not already been chosen. In the same way, we can define the process $\{Y_d(n,m)\}$ by including $d$-faces one at a time. We also define a time-changed version of this process $\LMP,$ more suitable to working with the binomial complex. Instead of including the faces one at a time, create independent $\operatorname{Exp}(1)$ clocks for every $d$-face. When one of the clocks rings, include the corresponding face. If we let $p(t) = 1 - e^{-t},$ then $\LMP$ has the distribution $Y_d(n,p(t)).$ \subsection{Cohomology vanishing} The foundational work on the Linial-Meshulam complexes is a cohomological analogue of the~{Erd\H{o}s--R\'{e}nyi } connectivity theorem. \begin{LMW} \label{thm:LMW} Let ${\bf k}$ be any finite field, $d \ge 2$ fixed, $f(n) \to \infty$ be any slowly growing function, and $Y \sim Y_d(n,p)$. If $$ p \ge \frac{d\log n + f(n)}{n},$$ then w.h.p.\ $H^{d-1}(Y, k) = 0$, and if $$ p \le \frac{d\log n-f(n)}{n},$$ then w.h.p.\ $H^{d-1}(Y, k) \neq 0$. \end{LMW} For the case that $d=2$ and ${\bf k}={\mathbb{Z}}_2,$ this is due to Linial and Meshulam~\cite{LM06}, while for the version stated, this is due to Meshulam and Wallach~\cite{MW09}. By the universal coefficient theorem, these results imply the corresponding theorem for the cohomology with $\mathbb{Q}$ coefficients. It remains an open problem whether or not the same theorem holds with $\mathbb{Z}$ coefficients. It is shown by the authors in~\cite{HKP} that for $p \geq 80d\log n / n$, $H^{d-1}(Y, \mathbb{Z}) = 0$ by other techniques. The threshold $p \sim d\log n / n$ is also the threshold for the existence of isolated $(d-1)$-faces in the complex, i.e. faces that are not included in any $d$-face. Indeed, the presence of isolated faces is precisely the reason that the cohomology is nonzero below this threshold. In fact, a finer statement can be made about the number of isolated $(d-1)$-faces. \begin{lemma} \label{lem:poisson_faces} Let $I$ denote the number of isolated $(d-1)$ faces in $Y_d(n,p).$ Suppose that for fixed $c,$ \[ p = \frac{d \log n + c+o(1)}{n}. \] Then $I$ converges in law to $\operatorname{Poisson}( e^{-c}/d!).$ \end{lemma} The proof of this lemma is standard and can be proved in the same manner as the Poisson convergence of the number of isolated vertices in $G(n,p).$ See Proposition 4.13 of~\cite{Ross}. Using spectral techniques, we give a new proof of the Linial-Meshulam-Wallach theorem, although only with $\mathbb{Q}$ coefficients. However, for $\mathbb{Q}$ coefficients, we also sharpen the theorem by proving a process version. More strikingly, this theorem shows that long before the last isolated $(d-1)$-faces disappear, the only obstruction to vanishing cohomology are those isolated $(d-1)$-faces. Its proof follows almost immediately from spectral arguments and Garland's method (see Section~\ref{sec:cohom}). \begin{theorem} \label{thm:Hhit} Consider the random complex process $\{ \LMP \}$. Let $I_t$ denote the number of isolated $(d-1)$-faces in the complex at time $t.$ Fix any $\delta >0$ and define $t_0$ so $p(t_0) = (d-1+\delta)\log n / n.$ Then w.h.p.\ for all time $t \geq t_0,$ \[ H_{d-1}(\LMP, \mathbb{Q}) \cong \mathbb{Q}^{I_t}. \] \end{theorem} As w.h.p. $I_{t_0} > 0$ we immediately get the following hitting time corollary. \begin{corollary} \label{cor:Hhit} Consider the random complex process $\{ Y_d(n,m) \}$. Let $$M_1 = \min \{ m \mid Y(n,m) \mbox{ has no isolated } (d-1)-\mbox{dimensional faces}\},$$ and let $$M_2 = \min \{ m \mid H^{d-1}(Y(n,m), \mathbb{Q}) = 0 \}.$$ Then w.h.p.\ $M_1 = M_2$. \end{corollary} Further, it is standard to show at this point that the Betti numbers are asymptotically Poisson. \begin{corollary} \label{cor:HPoisson} Suppose that for fixed $c,$ \[ p = \frac{d \log n + c+o(1)}{n}. \] Then $b_{d-1}(Y_d(n,p))$ converges in law to $\operatorname{Poisson}( e^{-c}/d!).$ \end{corollary} Note that this follows immediately from Lemma~\ref{lem:poisson_faces} and Theorem~\ref{thm:Hhit}. \subsection{The fundamental group} For the $2$-dimensional complex, a fair bit is known about the fundamental group $\pi_1(Y).$ Babson and the first two authors find the threshold for the fundamental group to be trivial~\cite{bhk11}. \begin{theorem} [Babson--Hoffman--Kahle] If $p = n^{-\alpha}$ where $\alpha < 1/2$ then w.h.p. $\pi_1(Y)$ is a nontrivial word hyperbolic group. If $p\geq n^{-1/2}\log(n)$ then $\pi_1(Y)$ is trivial. \end{theorem} Cohen et al. \cite{CCFK12} show that if $p = o ( 1/n),$ then w.h.p.\ $\pi_1(Y)$ is free. Finally, Costa and Farber describe the cohomological dimension $\operatorname{cd} \pi_1(Y)$ in various regimes \cite{CF12,CF13}. \begin{theorem} [Costa--Farber] Let $Y \sim Y(n,p)$, and set $p = n^{-\alpha}$. \begin{enumerate} \item If $\alpha > 1$ then w.h.p.\ $\operatorname{cd} \pi_1(Y) = 1$, \item if $1 > \alpha > 3/5$ then w.h.p.\ $\operatorname{cd} \pi_1(Y)=2$, and \item if $3/5 > \alpha > 1/2$ then w.h.p.\ $\operatorname{cd} \pi_1(Y) = \infty$. \end{enumerate} \end{theorem} For the $2$-dimensional complex, we combine the new spectral results with Garland's method to show a threshold theorem for $\pi_1(Y)$ to have property~(T). A group $G$ is said to have property~(T) if every unitary action of $G$ on a Hilbert space that has almost invariant vectors also has a nonzero invariant vector. The first explicit examples of expanders, due to Margulis, were constructed using Cayley graphs on quotients of (T) groups such as $SL(3,\mathbb{Z})$ \cite{Margulis}. Conversely, expansion properties of some graphs associated to the generating set of a group can imply property~(T) (see~\cite{Zuk}). Property~(T) has found use in many different areas of mathematics. For example, groups with property~(T) lead to good mixing properties in ergodic theory --- a process which mixes slowly must leave some subsets almost invariant. In particular, if a group $\Gamma$ has property~(T), then every ergodic $\Gamma$ system is also strongly ergodic \cite{glasner}. See the monograph \cite{Bekka} for a comprehensive overview of property~(T). We show the following: \begin{theorem} \label{thm:structure} Suppose $\delta > 0$ is fixed, $$ p \ge \frac{ ( 1 + \delta) \log n}{n},$$ and $Y \sim Y_2(n,p)$. Then w.h.p.\ $\pi_1(Y)$ is isomorphic to the free product of a (T) group $G$, and a free group $F$, where the free group $F$ has one generator for every isolated edge in $Y$. \end{theorem} Theorem \ref{thm:structure} might be viewed as a group-theoretic analogue of the fact that for $p \ge (1/2 + \delta) \log n / n$, the random graph $G \sim G(n,p)$ is w.h.p.\ a giant component, which is an expander, and isolated vertices. We have the following corollary of Theorem \ref{thm:structure}, which shows that the threshold for property~(T) is the same as the Linial--Meshulam theorem for vanishing of $\mathbb{Z} / 2$-homology. \begin{corollary} \label{cor:Twin} Let $\omega \to \infty$ as $n \to \infty$, and $Y \sim Y_2(n,p)$. If $$p \ge \frac{2 \log{n} + \omega}{n}$$ then $\mathbb{P}[\pi_1(Y) \mbox{ has property~(T)}] \to 1$. \label{T} \end{corollary} We also describe a process version of this structure theorem that holds below the connectivity threshold. \begin{theorem} \label{thm:Thit} Consider the random complex process $\{ \LMP \}$. Let $\tilde F_t$ be a free group with the number of generators equal to the number of isolated edges in the complex $\LMP$. Fix any $\delta >0$ and define $t_0$ so $p(t_0) = (1+\delta)\log n / n.$ Then w.h.p.\ for all $t \geq t_0,$ \[ \pi_1(Y_2(n,p(t))) \cong G_{t} * \tilde F_{t} \] where $G_t$ has property~(T). \end{theorem} Note that Theorem~\ref{thm:structure} follows immediately from this. As the number of isolated edges at time $t_0$ is positive w.h.p, we get the following hitting time corollary. \begin{corollary} \label{cor:Thit} Consider the random complex process $\{ Y_2(n,m) \}$. Let $$M_1 = \min \{ m \mid Y(n,m) \mbox{ has no isolated edges} \},$$ and let $$M_2 = \min \{ m \mid \pi_1( Y(n,m) ) \mbox{ is (T)} \}.$$ Then w.h.p.\ $M_1 = M_2$. \end{corollary} \begin{remark} We can additionally give an explicit Kazhdan pair for the (T) group. Setting $S$ to be the canonical generating set based at vertex $1,$ i.e. all loops cycles of the form $1 \to x \to y \to 1$ for distinct vertices $x$ and $y,$ then $(S,\sqrt{2}(1-o(1)))$ is a Kazhdan pair (see Remark 5.5.3 of~\cite{Bekka}). \end{remark} \section{Spectral estimates} \label{sec:est} In this section we give some with conditions on an arbitrary graph $G$ on $n$ vertices which facilitate a large spectral gap. Fix positive constants $C_1, C_2, C_3$ and $M$. In this section $d$ can be any function of $n$ with $d=d(n) \geq 1,$ and this is always satisfied by $d=(n-1)p,$ the convention taken in other sections. Recall that $T$ is the diagonal matrix of degrees. Let $W$ denote the set of vertices $x$ for which $\deg x > 0$ and $I$ be the number of isolated vertices in the graph. For any set of vertices $S,$ let $\one_{S}$ denote the vector that is one in every coordinate corresponding to $S$ and $0$ elsewhere. Let $0= \lambda_1 \leq \lambda_2 \leq \cdots \leq \lambda_{n}$ be the eigenvalues of the normalized Laplacian ${ L}[G],$ so that $\lambda_1 = \lambda_2 = \cdots = \lambda_{I+1} = 0.$ We also define a set of vertices of small degree. Let \begin{equation} \label{the fuzz} \ensuremath{\aleph_M} = \{ v \in V ~:~ \deg(v) \leq d/M \}. \end{equation} We now define four conditions that will ensure a spectral gap. \begin{enumerate} \item {\bf Bounded degree (b.d.c)} Every vertex has degree at most $C_1 d.$ \item {\bf Adjacency matrix} \[ \sup_{ \substack{ \|x\| = 1, x^t \one = 0 \\ \|y\| = 1 } } | x^tAy| \leq C_2\sqrt{d}. \] \item {\bf Fuzz} There are no edges between vertices of $\ensuremath{\aleph_M},$ $|\ensuremath{\aleph_M}| \leq \frac{n}{2}$ and \[ \max_{u \in \ensuremath{\aleph_M}^c} \dset{u}{\ensuremath{\aleph_M}} \leq 1.\] \item {\bf Parallel eigenspaces} \[ \sup_{ \substack{ \|x\| = 1, \\ x^t T^{1/2}\one_{W} = 0 } } | x^t T^{-1/2} \one_{\ensuremath{\aleph_M}^c}| \leq C_3\frac{\sqrt{n}}{d}. \] \end{enumerate} With these definitions we can now state our main result on spectral gaps. \begin{lemma} \label{lem:deterministic} Let $G$ be a graph on $n$ vertices and let $C_1,C_2,C_3$ and $M$ be constants. If $G$ satisfies the four conditions above then there is a constant $C=C(C_1,C_2,C_3,M)$ so that \[ \max_{i > I + 1} \left| 1 - \lambda_i\right| < \frac{C}{\sqrt{d}}. \] \end{lemma} \begin{proof} Let $W$ be the set of vertices $x$ for which $\deg x > 0.$ By the spectral theorem, ${ L}$ admits a basis of orthogonal eigenvectors. Let $v$ be a normalized eigenvector of ${ L}$ corresponding to an eigenvalue $\lambda_i$ with $i > I+1.$ Setting $l_1,l_2,\ldots, l_I$ to be the isolated vertices, a basis for the kernel of ${ L}$ is given by $\{ T^{1/2}\one, \delta_{l_1},\delta_{l_2}, \ldots, \delta_{l_I}\},$ where $\delta_a$ is $1$ in the $a^{th}$ coordinate and $0$ elsewhere. As $v$ is orthogonal to all of these, it is orthogonal to $T^{1/2}\one_{W}.$ Hence, \[ \left| 1 - \lambda_i\right| = \left| v^t T^{-1/2} A T^{-1/2} v\right| \leq \sup_{\substack{ \|x\| = 1, \\ x^tT^{1/2}\one_{W}=0}} \left| x^t T^{-1/2} A T^{-1/2} x \right|. \] As this holds for all such $i > I+1,$ it suffices to bound the right hand side. Orthogonally decompose $T^{-1/2}x = u + v,$ where $u$ is supported on vertices in $\ensuremath{\aleph_M}^c$ and $v$ is supported on vertices in $\ensuremath{\aleph_M}.$ Further decompose $u = u_0 + u_1$ by letting $u_1$ be the projection of $u$ along $\one_{\ensuremath{\aleph_M}^c}.$ Expanding the quadratic form, we may write \begin{equation} \label{decomposition} \left| x^t T^{-1/2} A T^{-1/2} x \right| \leq 2|u_0^tAu| + |u_1^tAu_1| + |v^tAv| + 2|v^tAu|. \end{equation} Each of these terms will be seen to have the right order bound, completing the proof. As $u_0 \perp \one_{\ensuremath{\aleph_M}^c}$ and is supported only on $\ensuremath{\aleph_M}^c,$ we have that $u_0 \perp \one.$ By the definitions of $\ensuremath{\aleph_M}$ and $x,$ we have that \[ \|u_0\|^2 \leq \|u\|^2 = \sum_{i \in \ensuremath{\aleph_M}^c} \frac{|x_i|^2}{\deg i} \leq \frac{M}{d}. \] Hence by the adjacency matrix condition and the above equation we have that \begin{equation} \label{partone} |u_0^tAu| \leq C_2\sqrt{d} \|u_0\| \|u\| = \frac{C_2M}{\sqrt{d}}. \end{equation} As $u_1$ is the projection of $u$ along $\one_{\ensuremath{\aleph_M}^c},$ we have \[ u_1 = (u^t \one_{\ensuremath{\aleph_M}^c}) \frac{ \one_{\ensuremath{\aleph_M}^c}}{|\ensuremath{\aleph_M}^c|} = (x^t T^{-1/2} \one_{\ensuremath{\aleph_M}^c}) \frac{ \one_{\ensuremath{\aleph_M}^c}}{|\ensuremath{\aleph_M}^c|}. \] Because $\ensuremath{\aleph_M}^c \geq \frac{n}{2},$ the parallel eigenspaces condition implies that we have \( \|u_1\| \leq \frac{\sqrt{2}C_3}{d}. \) The norm of $A$ is at most the maximum degree of the graph, and by the bounded degree condition this is at most $C_1d.$ Hence, we get that \begin{equation} \label{parttwo} |u_1^tAu_1| \leq \frac{2C_1C_3^2}{d} \end{equation} For the third term, we note that by the $\ensuremath{\aleph_M}$ condition there are no edges between vertices of $\ensuremath{\aleph_M},$ and hence \begin{equation} \label{partthree} v^tAv = 0. \end{equation} Finally, we may expand $v^tAu$ as \[ v^tAu = \sum_{i \in \ensuremath{\aleph_M}} \frac{x_i}{\sqrt{\deg i}} \sum_{\substack{ j\in \ensuremath{\aleph_M}^c,\\ j \sim i}} u_j. \] By Cauchy-Schwarz, this is bounded by \[ |v^tAu|^2 \leq \sum_{i \in \ensuremath{\aleph_M}} \frac{1}{{\deg i}} \biggl(\sum_{\substack{ j\in \ensuremath{\aleph_M}^c,\\ j \sim i}} u_j\biggr)^2 \leq \sum_{i \in \ensuremath{\aleph_M}} \sum_{\substack{ j\in \ensuremath{\aleph_M}^c,\\ j \sim i}} \left(u_j\right)^2. \] Now each $j \in \ensuremath{\aleph_M}^c$ has at most one neighbor in $\ensuremath{\aleph_M}$, and hence we have \begin{equation} \label{partfour} \left|v^tAu\right| \leq \|u\| = \frac{\sqrt{M}}{\sqrt{d}}. \end{equation} Plugging (\ref{partone}), (\ref{parttwo}), (\ref{partthree}) and (\ref{partfour}) into (\ref{decomposition}) completes the proof. \end{proof} In the remainder of this section we prove a condition on a graph that will imply an upper bound on the spectral gap. This lemma shows that our previous argument breaks down when the set $\ensuremath{\aleph_M}$ fails to be isolated. \begin{lemma} \label{lem:deterministic2} Suppose that $H$ is connected graph and that there are vertices $u,v,w,x$ for which the induced graph on $u,v,w,x$ is a path with endpoints $u$ and $x$. Suppose further that $\deg v = \deg w = 2$ and $\deg u, \deg x \geq m.$ Let $0= \lambda_1 \leq \lambda_2 \leq \cdots \leq \lambda_{|H|}$ be the eigenvalues of the normalized Laplacian ${ L}[H],$ then \[ \lambda_{|H|} \geq \tfrac{3}{2} \] and \[ \lambda_{2} \leq \tfrac{1}{2} + O(1/\sqrt{m}) \] \end{lemma} \begin{proof} For each case, we construct an appropriate approximate eigenvector. For the first, consider vector $f$ with $f(v) = 1,$ $f(w)=-1$ and $f(y)=0$ for all other $y$. This vector is orthogonal to $T^{1/2}\one$, the first eigenvector of ${ L}.$ Now $T^{-1/2}f$ is just $f/\sqrt{2}$ while $f^tAf = -2.$ Thus, \[ \frac{f^tT^{-1/2}AT^{-1/2}f}{\|f\|^2} = -\frac{1}{2}, \] and so $\lambda_{|S|} \geq 1 - \tfrac{-1}{2} = \tfrac{3}{2}.$ For the lower bound let $f$ be given by $f(v) = f(w) = 1/\sqrt{2}$ while $f(x) = -1/\sqrt{\deg x}$ and $f(u) = -1/\sqrt{\deg u}.$ Then we have $f \perp T^{1/2}\one.$ By direct computation, \[ {f^tT^{-1/2}AT^{-1/2}f} = \frac{1}{2} - \frac{1}{\deg x} - \frac{1}{\deg u}, \] while \[ \|f\|^2 \leq 1 + \frac{1}{\deg x} + \frac{1}{\deg u}. \] Thus, combining everything, we have that \[ \lambda_{2} \leq 1- \frac{\tfrac12 - \tfrac{2}{m}}{\sqrt{1+\frac{2}{m}}} = \tfrac{1}{2} + O(1/\sqrt{m}). \] \end{proof} \section{Probability bounds} \label{sec:pbnds} \begin{lemma} \label{lem:regularity} For each $\delta >0$ and $m \geq 0,$ there is a constant $C=C(\delta, m)$ so that the following conditions hold with probability at least $1-C\exp(-md)$ and $1-C\exp(-md^{1/4}\log n)$ respectively. \begin{enumerate} \item {\bf Bounded degree condition (b.d.c)} Every vertex has degree at most $C d.$ \item {\bf Discrepancy} For every pair of vertex sets $A$ and $B$, letting $e(A,B)$ denote the number of edges between the sets and $\mu(A,B) = \tfrac{|A||B|d}{n}$, one of \begin{enumerate} \item $ \tfrac{e(A,B)}{\mu(A,B)} \leq C $ \item $ e(A,B)\log \tfrac{e(A,B)}{\mu(A,B)} \leq C( |A| \vee |B|)\log \tfrac{n}{|A| \vee |B|} $ \item $ |A| \leq d^{1/4}/100, |B| \leq d^{1/4}/100 $ \end{enumerate} occurs. \end{enumerate} \end{lemma} Both of these bounds are consequences of tail bounds of binomial variables, and they are relatively standard in the literature (see, e.g. \cite{FKSz},\cite{FeigeOfek},\cite{CojaOghlan}). This one differs in that we look for more control over the order of decay of the failure probability. \begin{proposition} \label{prop:adj} For each $\delta >0$ and $m \geq 0,$ there is a constant $C=C(\delta, m)$ sufficiently large so that if \(p \geq \delta \log n / n\) then \[ \sup_{ \substack{ \|x\| = 1, x^t \one = 0 \\ \|y\| = 1 } } | x^tAy| \leq C\sqrt{d} \] with probability at least $1-C\exp(-md^{1/4}\log n) - C\exp(-md).$ \end{proposition} This follows from the standard Kahn-Szemer\'erdi argument, and it is essentially proven in both Feige and Ofek~\cite{FeigeOfek} and the original Friedman, Kahn and Szemer\'erdi paper~\cite{FKSz}. This version has a sharper estimate on the failure probability than~\cite{FeigeOfek}, which in turn follows from Lemma~\ref{lem:regularity}. We will delay the proof of both this and the previous lemma to Section~\ref{sec:KSz}. Additionally, the bounded degree condition is needed to make estimates about low degree vertices. Recall the definition of $\ensuremath{\aleph_M}$ from (\ref{the fuzz}). We show that this set is both small and structurally very simple for sufficiently large $M.$ \begin{proposition} \label{prop:key_structure} For each $\delta >0$ and each $\epsilon >0,$ if \(p \geq \delta \log n / n\) there is an $M=M(\delta,\epsilon) > 1$ \begin{enumerate} \item $|\ensuremath{\aleph_M}| < n/(100d)$ \item $\ensuremath{\aleph_M}$ is an independent set, \item and \( \max_{u \in \ensuremath{\aleph_M}^c} \dset{u}{\ensuremath{\aleph_M}} \leq 1\) \end{enumerate} with probability at least $1 - Cn\exp(-(2-\epsilon)d)- C\exp(-cn)$ for some absolute constant $c>0.$ \end{proposition} \begin{proof} \noindent{\emph{ (i)}} We start by estimating the size of $\ensuremath{\aleph_M},$ which we do by a simple union bound. Namely by symmetry we have \begin{align*} \Pr\left[ |\ensuremath{\aleph_M}| \geq k \right] \leq {n \choose k} \Pr\left[ \deg u_i \leq d/M, 1 \leq i \leq k \right]. \end{align*} Let $S$ be the set of vertices $u_{k+1},\ldots,u_n,$ then we have \[ \Pr \left[ \deg u_i \leq d/M, 1 \leq i \leq k \right] \leq \Pr \left[ \dset{u_i}{S} \leq d/M, 1 \leq i \leq k \right], \] which are now independent $\operatorname{Binom}(n-k,p)$ variables. Applying Lemma~\ref{magic}, we get \begin{align*} \log \Pr\left[ |\ensuremath{\aleph_M}| \geq k \right] \leq k\left[ (1+ \log\frac{n}{k})-(d - kp) + \frac{d}{M}( 1 + \log(M)) \right]. \end{align*} Setting $k = [n/(100d)],$ we may make $M$ sufficiently large that \[ (1+ \log\frac{n}{k})-(d - kp) + \frac{d}{M}( 1 + \log(M)) \leq -\frac{d}{2} \] for all $n \geq n_0(\delta).$ Hence we have that $|\ensuremath{\aleph_M}| < n/(100d)$ with probability at least $1 - O\exp(-cn)$ for some absolute constant $c>0.$ \noindent{\emph{ (ii)}} We begin by bounding the probability that there is an edge between any two vertices of $\ensuremath{\aleph_M}.$ Note that we may assume that $d < n/100,$ lest $\ensuremath{\aleph_M} = \emptyset$ by the previous bound. From the union bound and symmetry, we have that \[ \Pr \left[ \ensuremath{\aleph_M} \text{ is not an independent set} \right] \leq n^2 \Pr \left[ v \in S, w \in S, \ebetween{v}{w} \right]. \] Thus it suffices to compute this probability, which we do by conditioning $\deg v = d_1$ and $\deg w = d_2.$ Note that the law of the neighborhood $N$ of $\{v,w\}$ under this conditioning is \emph{not} uniform over all such neighborhoods. For a possible neighborhood $H$ of $\{v,w\},$ let $E(H)$ denote the number of edges in this neighborhood. Then we have that \[ \Pr \left[ N = H \middle\vert \deg v = d_1, \deg w = d_2 \right] = \frac{1}{Z} \left(\frac{p}{1-p}\right)^{E(H)}, \] for a suitable normalization constant $Z.$ Thus, we have that \begin{align*} \Pr \left[ \ebetween{v}{w} \middle\vert \deg v = d_1, \deg w = d_2 \right] &\leq \frac{ \Pr \left[ \ebetween{v}{w} \middle\vert \deg v = d_1, \deg w = d_2 \right] }{ \Pr \left[ \noebetween{v}{w} \middle\vert \deg v = d_1, \deg w = d_2 \right] } \\ &= \frac{1-p}{p} \frac{ { n - 2 \choose d_1 - 1} { n - 2 \choose d_2 - 1} } { {n - 2 \choose d_1} {n - 2 \choose d_2} }. \end{align*} As we consider only $d_1$ and $d_2$ that are less than $d / M,$ and as $d < n/100,$ we may bound this as $Cd/n$ for some absolute constant $C.$ It remains to estimate the probability that both $v$ and $w$ are in $\ensuremath{\aleph_M}.$ Hence we have \[ \Pr\left[ \deg v \leq d/M, \deg w \leq d/M \right] \leq \Pr\left[ X \leq d/M \right]^2, \] where $X \sim \Binomial(n-2,p).$ Applying Lemma~\ref{magic}, we have that \begin{equation} \label{eq:cofuzz} \Pr\left[ \deg v \leq d/M, \deg w \leq d/M \right] \leq \exp\left[ -2d + \frac{2d}{M}( 1+ \log M + O(1) ) \right] \end{equation} Thus by adjusting $M$ to be sufficiently large, we have \[ \Pr \left[ \ensuremath{\aleph_M} \text{ is not an independent set} \right] = O(nd\exp(-(2-\epsilon/2)d)) = O(n\exp(-(2-\epsilon)d)). \] \noindent{\emph{ (iii)}} This follows in much the same way as the proof of (ii). Here though, we require that the degrees of $\ensuremath{\aleph_M}^c$ are not too large. By Lemma~\ref{lem:regularity}, these degrees can be bounded by some $Cd$ with probability at least $1-O(\exp(-2d)),$ and so it suffices to assume it. From the union bound and symmetry, we have that \begin{multline*} \Pr \left[ \exists u \in \ensuremath{\aleph_M}^c~:~\dset{u}{\ensuremath{\aleph_M}} \geq 2 ~\cap~\ensuremath{ \bf{b.d.c.}} \right] \\ \leq n^3 \Pr \left[ u \in \ensuremath{\aleph_M}^c, v \in \ensuremath{\aleph_M}, w \in \ensuremath{\aleph_M}, \ebetween{u}{v}, \ebetween{u}{w}~\cap~\ensuremath{ \bf{b.d.c.}} \right]. \end{multline*} Again we condition on the degrees $\deg u = d_1, \deg v = d_2,$ and $\deg w = d_3,$ and bound \begin{align*} \Pr \left[ \ebetween{u}{v}, \ebetween{u}{w} \middle\vert \deg u = d_1, \deg v = d_2, \deg w = d_3 \right] \hspace{-2in}&\hspace{2in} \\ &\leq \frac{ \Pr \left[ \ebetween{u}{v}, \ebetween{u}{w} \middle\vert \deg u = d_1, \deg v = d_2, \deg w = d_3 \right] }{ \Pr \left[ \noebetween{u}{v}, \noebetween{u}{w}, \noebetween{v}{w} \middle\vert \deg u = d_1, \deg v = d_2, \deg w = d_3 \right] } \\ &= \left(\frac{1-p}{p}\right)^2 \frac{ { n - 3 \choose d_1 - 2} { n - 3 \choose d_2 - 1} { n - 3 \choose d_3 - 1} + \frac{p}{1-p} { n - 3 \choose d_1 - 2} { n - 3 \choose d_2 - 2} { n - 3 \choose d_3 - 2} } { {n - 3 \choose d_1} {n - 3 \choose d_2} {n - 3 \choose d_3} }. \end{align*} As before, we have $d_1$ and $d_2$ are less than $d / M.$ As we also require the $\ensuremath{ \bf{b.d.c.}}$ to hold, we may take $d_1 \leq c_{1} d$ and as $d < n/100,$ we may bound this as $C c_{1} d^2/n^2$ for some absolute constant $C.$ From~\eqref{eq:cofuzz}, we have that \[ \Pr\left[ v \in \ensuremath{\aleph_M}, w \in \ensuremath{\aleph_M} \right] = O(\exp(-(2-\epsilon/2)d)), \] and so we conclude that \[ \Pr \left[ \max_{u \in \ensuremath{\aleph_M}^c} \dset{u}{\ensuremath{\aleph_M}} > 1 \right] = O(n\exp(-(2-\epsilon)d)). \] \end{proof} Our next lemma shows that the variance of the degree distribution is not too much larger than its expectation. \begin{lemma} \label{lem:T_distortion} For each fixed $\delta >0$ and $m \geq 0,$ there is a constant $C=C(\delta, m)$ sufficiently large so that if \(p \geq \delta \log n / n\) then \[ \sum_{v \in V} \left( \deg v - d \right)^2 \leq Cnd. \] with probability at least $1-C\exp(-md).$ \end{lemma} \begin{proof} Note that this sum is the square Euclidean norm of the vector $(A-dI)\one.$ Further, it is possible to write the norm as \[ \| (A-dI)\one \| = \sup_{\|x\|=1} |x^t(A-dI)\one|. \] For any fixed vector $x$, we orthogonally decompose it as $x = v + c\one,$ where $|c| \leq 1/\sqrt{n}.$ We have that $v^t(A-dI)\one = v^tA\one,$ and so by Proposition~\ref{prop:adj}, for any $m$ there is a constant $C$ so that \[ \sup_{\substack{\|v\|=1 \\ v^t\one = 0}} |v^tA\one| \leq C\sqrt{nd} \] with probability at least $1-O(\exp(-md)).$ It remains to bound $\one^t(A-dI)\one,$ which is \[ \one^t(A-dI)\one = \left(\sum_{v \in V} \deg v\right) - nd. \] Note that $\sum_{v \in V} \deg v \sim 2 \Binomial( {n\choose 2}, p),$ and so by standard Chernoff bounds, we have that \[ \Pr \left[ \left|\one^t(A-dI)\one \right| \geq t \right] \leq C\exp( - \frac{t^2}{Cnd} ) \] for some absolute constant $C$ and all $t \leq nd.$ By taking $t = mn\sqrt{d},$ we have that $\left|\one^t(A-dI)\one \right| \leq mn\sqrt{d}$ with probability at least $1-O(\exp(-mn))$ for sufficiently large $n.$ Recalling that $|c| \leq 1/\sqrt{n},$ we have that \[ \left|c\one^t(A-dI)\one\right| = O(\sqrt{nd}). \] which completes the proof. \end{proof} Using the previous lemma, we show that $T^{-1/2}$ tends to map the orthogonal complement of the first eigenvector of $M$ to the approximate orthogonal complement of the first eigenvector of $A$. \begin{lemma} \label{lem:T_flattens} Let $W$ be the set of vertices $x$ for which $\deg x > 0,$ and let $\ensuremath{\aleph_M}$ be as in Proposition~\ref{prop:key_structure}. For each $\delta >0$ and $m \geq 0,$ there is a constant $C=C(\delta, m)$ sufficiently large so that if \(p \geq \delta \log n / n\) then \[ \sup_{ \substack{ \|x\| = 1, \\ x^t T^{1/2}\one_{W} = 0 } } | x^t T^{-1/2} \one_{\ensuremath{\aleph_M}^c} | \leq C\frac{\sqrt{n}}{d} \] with probability at least $1-C\exp(-md).$ \end{lemma} \begin{proof} As we have that $|\ensuremath{\aleph_M}| < n/(100d)$ by Proposition~\ref{prop:key_structure}, it follows that \[ |x^tT^{1/2}\one_{\ensuremath{\aleph_M}}| \leq \| T^{1/2}\one_{\ensuremath{\aleph_M}}\| \leq \sqrt{d|\ensuremath{\aleph_M}|} = O\left(\sqrt{n}\right). \] Further, we have that \( x^tT^{1/2}\one_{\ensuremath{\aleph_M}} = -x^tT^{1/2}\one_{\ensuremath{\aleph_M}^c}, \) and hence it suffices to show that \[ \sup_{ \substack{ \|x\| = 1, \\ x^t T^{1/2}\one = 0 } } \left|x^t( T^{-1/2} - T^{1/2}/d)\one_{\ensuremath{\aleph_M}^c}\right| \leq C\frac{\sqrt{n}}{d}. \] Taking norms, \[ \left|x^t( T^{-1/2} - T^{1/2}/d)\one_{\ensuremath{\aleph_M}^c}\right| \leq \left\|( T^{-1/2} - T^{1/2}/d)\one_{\ensuremath{\aleph_M}^c}\right\|. \] Squaring this norm, we get \[ \left\|( T^{-1/2} - T^{1/2}/d)\one_{\ensuremath{\aleph_M}^c}\right\|^2 =\sum_{v \in \ensuremath{\aleph_M}^c} \left( \frac{1}{\sqrt{\deg v}} - \frac{\sqrt{\deg v}}{d} \right)^2 \leq \frac{M}{d^3} \sum_{v \in \ensuremath{\aleph_M}^c} \left( \deg v - d \right)^2. \] Lemma~\ref{lem:T_distortion} completes the proof. \end{proof} \subsection*{Proofs of the main theorems} \begin{pfofthm}{Theorem~\ref{thm:ergap_giant}} By Lemma~\ref{lem:deterministic}, the theorem follows from Lemma~\ref{lem:regularity}, Proposition~\ref{prop:adj}, Proposition~\ref{prop:key_structure}, and Lemma~\ref{lem:T_flattens}. \end{pfofthm} We wish to now show the lower bounds for $\lambda(\tilde{G}).$ We will use Lemma~\ref{lem:deterministic2}, and this requires that we show: \begin{proposition} \label{prop:handles} If $p = \omega(\sqrt{\log n}/n)$ and $p \leq \tfrac12\log n / n$ then with high probability, there are four distinct vertices $a,b,c,d$ in the giant component for which the degrees of $a$ and $d$ are at least $np/2,$ the degrees of $b$ and $c$ are $2,$ and the induced subgraph on $(a,b,c,d)$ is a path. \end{proposition} We first show by the second moment method that such four-tuples $(a,b,c,d)$ exist in the graph with high probability. We then show that with high probability, the small components have maximal degree $o(np),$ and hence these four-tuples must have been part of the giant component. \begin{lemma} \label{lem:handles_exist} Suppose that $p = \omega( 1 / n)$ and that $p \leq \tfrac12\log n / n.$ Then, with high probability, there are four-tuples $(a,b,c,d)$ for which the degrees of $a$ and $d$ are at least $np/2,$ the degrees of $b$ and $c$ are $2,$ and the induced subgraph on $(a,b,c,d)$ is a path. \end{lemma} \begin{proof} Define the pair of events \begin{align*} A(a,b,c,d) &= \{ a \leftrightarrow b \leftrightarrow c \leftrightarrow d, \deg b = \deg c = 2 \}~\text{ and } \\ B(a,b,c,d) &= A(a,b,c,d) \cap \{ \deg a \geq np/2, \deg d \geq np/2 \}. \end{align*} Set $S$ to be the number of occurrences of $B,$ i.e. \[ S = \sum_{ [n] \choose 4 } \one[B(a,b,c,d)]. \] We need to show that $S > 0$ with high probability. The probability of $A$ can be explicitly calculated as \[ \Pr \left[ A(a,b,c,d) \right] = p^3(1-p)^{2(n-3)}. \] Meanwhile, conditional on $A(a,b,c,d),$ the probability of $B(a,b,c,d)$ is exactly the probability of having two vertices of degree at least $np/2 - 1$ in $G(n-2,p).$ Set $Q = \Pr\left[ X \geq np/2 \right]$ where $X\sim \Binomial(n,p).$ Note that as $np \to \infty,$ we have that $Q=1-o(1).$ Furthermore, as $np \to \infty$ we have that \[ \Pr \left[ B(a,b,c,d) ~\middle\vert~ A(a,b,c,d) \right] = Q^2(1-o(1)), \] simply by conditioning on the edge between $a$ and $d.$ By summing over all possible tuples, it follows that \( \mathbb{E} S = \Theta( nQ^2(np)^3e^{-2np} ) = \omega(1). \) \newcommand{\bvec}{ (b_i)_{i=1}^4 } \newcommand{\avec}{ (a_i)_{i=1}^4 } For the variance of $S$, we need to compute probabilities of the pairs $B( \avec ) \cap B( \bvec).$ Note that if $a_2 = b_2$ then the only way both can happen is if $a_i = b_i$ for all $i \in [4].$ Analogous conclusions hold if $a_2 = b_3$ or if $a_3 \in \{b_2,b_3\}.$ Thus, the only nontrivial way for the events $B( \avec)$ and $B( \bvec)$ to intersect is if \begin{enumerate} \item all $a_i$ and $b_i$ are distinct, \item $a_1 = b_1$ and the rest are distinct, \item $a_1 = b_1,$ $a_4=b_4,$ and the rest are distinct, or \item $a_i = b_i$ for all $i.$ \end{enumerate} Note that there's no need to consider $a_1 = b_4,$ as the event $B( \bvec )$ is preserved under reversing the $a_i.$ Set $T_i$ to be the pairs of tuples satisfying each of the $4$ cases. If the pair is in $T_1,$ then \[ \Pr \left[ B( \avec) \cap B( \bvec) \middle\vert A( \avec) \cap A( \bvec) \right] =Q^4(1-o(1)) \] as once more, this is the statement that four vertices in $G(n-4,p)$ have degree at least $(np/2-1)$. We also have that \[ \Pr \left[ A( \avec) \cap A( \bvec) \right] = p^6(1-p)^{4n-16}, \] so that \[ \Pr \left[ B( \avec) \cap B( \bvec) \right] = \Pr \left[ B( \avec) \right]^2(1-o(1)). \] Thus the contribution of the pairs in $T_1$ to the variance of $S$ is $o( (\mathbb{E} S)^2).$ For terms from $T_2,$ the same reasoning as above shows that \[ \Pr \left[ B( \avec) \cap B( \bvec) \right] = Q^3p^6(1-p)^{4n}(1-o(1)) \] For such pairs, however, we have that $|T_2| = \Theta(n^7),$ and hence the contribution to the variance of $S$ is $o( (\mathbb{E} S)^2).$ In the same way, the contributions of $T_3$ and $T_4$ are smaller still. As each is individually of order $o( (\mathbb{E} S)^2 )$, we have that $S>0$ with high probability. \end{proof} \begin{lemma} \label{lem:giant_component_size} Suppose that $p = \omega(1/n),$ then for any $\epsilon >0,$ the number of vertices not in the giant component is at most $n e^{-(1-\epsilon)np}$ with high probability. \end{lemma} \begin{proof} Set $R$ to be the number of vertices not in the giant component. Then $R < r$ if and only if there is no collection $W$ of at least $r$ vertices for which $W$ is disconnected from $W^c.$ The expected number $\mathbb{E} N_r$ of such collections $W$ is given by \[ \mathbb{E} N_r = (1-p)^{r(n-r)} { n \choose r }. \] Set $r_0 = n e^{-(1-\epsilon)np}.$ We will show that \( \sum_{r=r_0}^{n/2} \mathbb{E} N_r \to 0, \) which implies the lemma. Subdivide the sum into two pieces $S_1$ and $S_2,$ given by $S_1 = \sum_{r_0}^{\lfloor \epsilon n/4 \rfloor} \mathbb{E} N_r$ and $S_2 = \sum_{\lfloor \epsilon n/4 \rfloor}^{n/2} \mathbb{E} N_r.$ For $\lfloor \epsilon n/4 \rfloor \leq r \leq n/2,$ \[ \mathbb{E} N_r = (1-p)^{r(n-r)} { n \choose r } \leq e^{-c_\epsilon n^2p} 2^n, \] for some $c_\epsilon > 0,$ which decays exponentially in $n$ as $np \to \infty.$ Hence $S_2 \to 0.$ As for $S_1,$ we claim that for any $\alpha > 0$ there is an $n \geq n_0(\alpha,\epsilon)$ sufficiently large so that for all $r_0 < r < \epsilon n/4,$ $\mathbb{E} N_{r+1} \leq \alpha \mathbb{E} N_r$ for all $n \geq n_0(\alpha,\epsilon).$ Estimating for these $r,$ \begin{align*} \frac{\mathbb{E} N_{r+1}}{\mathbb{E} N_r} &= (1-p)^{n-2r+1} \frac{n-r-1}{r+1} \\ &\leq \frac{ne^{-np+2rp} }{r}.\\ &\leq \frac{ne^{-(1-\epsilon/2)np} }{r}.\\ &\leq e^{-\epsilon np/2}. \end{align*} Hence, as $np \to \infty,$ this is eventually less than any positive $\alpha.$ As $S_1$ is dominated by a geometric series, and $S_1 = O( \mathbb{E} N_{r_0}).$ For this leading term, we get that \[ \mathbb{E} N_{r_0} \leq e^{-pr_0(n-r_0)} \left( \frac{en}{r_0}\right)^{r_0} \leq \exp\left( -\epsilon n^2pe^{-(1-\epsilon)np}(1-o(1)) \right) \to 0, \] completing the proof. \end{proof} \begin{lemma} \label{lem:dust_degree_bound} If $p = \omega( \sqrt{\log n}/ n),$ then with high probability, the maximum degree of the vertices not in the giant component is at most $np/100$ with high probability. \end{lemma} \begin{proof} Set $R$ to be the number of vertices not in the giant component. By Lemma~\ref{lem:giant_component_size}, we have that $R \leq ne^{-np/2}$ with high probability. Suppose that $W$ is a fixed collection of vertices of size $r.$ Conditional on there being no edges between $W$ and $W^c,$ the law of the induced graph on $W$ is simply that of $G(r,p).$ Let $X \sim \operatorname{Binom}(r-1,p).$ Then by Lemma~\ref{maybe} there are absolute constants $c>0$ and $M > 0$ so that \[ \Pr \left[ X > np/100 \right] \leq \exp(-cnp\log(n/r)) \] provided $r<n/M.$ Setting $E_W$ to be the event that $W$ and $W^c$ are not connected \[ \Pr \left[ \max_{w \in W} \deg w > np/100 ~\middle\vert~ E_W \right] \leq r\exp(-cnp\log(n/r)). \] Let $Y$ be the max degree of all vertices not in the largest component As the previous bound holds for all $W$ in consideration, we get that \[ \Pr \left[ Y > np/100 ~\middle\vert~ R = r \right] \leq r\exp(-cnp\log(n/r)). \] This bound is monotone increasing in $r,$ and so we get that \[ \Pr \left[ Y > np/100 ~\middle\vert~ R \leq ne^{-np/2} \right] \leq n\exp(-c(np)^2(1-o(1))) \] for some absolute constant $c$. Thus by the assumption on $np,$ the desired claim holds. \end{proof} \begin{pfofthm}{ Theorem~\ref{thm:ergap_giant} and Proposition~\ref{prop:handles} } For Proposition~\ref{prop:handles}, the previous three Lemmas~\ref{lem:handles_exist},~\ref{lem:giant_component_size}, and~\ref{lem:dust_degree_bound} show the desired claim that there are tuples $(a,b,c,d)$ of vertices in the giant component for which $\deg a$ and $\deg d$ are at least $np/2,$ vertices $b$ and $c$ have degree $2,$ and the induced graph on these vertices is a path. Letting $H$ be the giant component of the graph, then there is a constant $C$ so that the eigenvalues of the Laplacian of $H$ satisfy \[ \lambda_{|H|} \geq \tfrac{3}{2} \] and \[ \lambda_{2} \leq \tfrac{1}{2} + C/\sqrt{np}, \] by Lemma~\ref{lem:deterministic2}. \end{pfofthm} \section{Gap process theorem} \label{sec:proc} In this section we prove a general process-version theorem for the spectral gap below the connectivity threshold. We recall the definition of $\LMP[t][k],$ the continuous time Linial-Meshulam process. Let $F_k$ denote the collection of all $k$-faces of the $n$-simplex, and let $\{T_\sigma, \sigma \in F_k\}$ be an i.i.d.\ family of $\operatorname{Exp}(1)$ variables. Define $\{\LMP[t][k], t \geq 0\}$ to be the continuous time Markov process where $\LMP[0][k]$ is the complete $(k-1)$-skeleton of the $n$-simplex and its $k$-faces are given by \[ F_k(\LMP[t][k]) = \{ \sigma \in F_k~:~T_\sigma \leq t\}. \] Thus $\LMP[t][k]$ is the complex whose $k$-faces have been born up to time $t,$ and $\LMP[\infty][k]$ is the complete $k$-skeleton of the $n$-simplex. For $k=1,$ this recovers the standard continuous time {Erd\H{o}s--R\'{e}nyi } process. For fixed $t,$ $\LMP[t][k]$ is the Bernoulli complex $Y_k(n,p(t))$ with $p(t) = 1 - e^{-t}.$ Let $d(t) = (n-1)p(t).$ Fix a $\delta$ with $0 < \delta < \frac12$ and define $t_0$ to be that time so that \[ p(t_0) = \begin{cases} (\tfrac 12 + \delta) \log n / n & k = 1, \\ (k - 1 + \delta) \log n / n & k > 1. \end{cases} \] \begin{theorem} \label{thm:stopping_time} Let $\tLMP[t][k]$ denote the process that has every isolated $(k-1)$-face removed. There is a constant $C=C(k,\delta)$ so that with high probability the Laplacian of every $\mbox{lk}(f)$ of \( {\tLMP[t][k]} \) has \[ \max_{i > 1} \left| 1 - \lambda_i\right| < \frac{C}{\sqrt{d(t)}}. \] for all $t \geq t_0.$ \end{theorem} Note that any isolated $(k-1)$-face creates an isolated vertex in some link and a vertex in a codimension-2 link is isolated if and only if the corresponding $(k-1)$-face is isolated. Thus, an equivalent formulation is that every link of $\LMP[t][k]$ consists of isolated vertices and a giant component whose gap is $1-C/\sqrt{d(t)}$ for all time $t \geq t_0.$ In the higher-dimensional setting, the proof is more complicated than simply studying each link individually and taking the union bound. The key is to study the ``low-degree'' simplices globally. To this end, for each $\mbox{lk}(f)$ and for any $M \geq 1,$ let \begin{equation} \label{eq:fuzz_t} \ensuremath{\aleph_M}[f][t] =\{ w \in \mathcal{V}(\mbox{lk}(f))~:~ \deg_{\mbox{lk}(f)}(w) \leq d(t_0)/M \}. \end{equation} Note that this makes each $\ensuremath{\aleph_M}[f][t]$ monotone decreasing in $t.$ \begin{lemma} \label{lem:supersmall} There is an $M=M(k,\delta)$ and an $\epsilon=\epsilon(k,\delta)$ so that \[ \sum_{f \in F_{k-2}}\left| \ensuremath{\aleph_M}[f][t_0] \right|^2 \leq n^{1 - \epsilon} \] with overwhelming probability. \end{lemma} \begin{proof} For $k=1,$ there is only one link to consider, and so it suffices to show that \( \left|\ensuremath{\aleph_M}[\emptyset][t_0]\right| \leq n^{1/2 - \epsilon}. \) For $k>1,$ we proceed by showing that for any $\epsilon$ there is an $M$ so that both \begin{enumerate} \item \( \max_{f \in F_{k-2}} \left| \ensuremath{\aleph_M}[f][t_0] \right| \leq n^{\epsilon} \) \item \( \sum_{f \in F_{k-2}}\left| \ensuremath{\aleph_M}[f][t_0] \right| \leq n^{1 - 2\epsilon} \) \end{enumerate} hold with overwhelming probability. The first condition follows from an identical argument to the first part of Proposition~\ref{prop:key_structure}; the $k=1$ case follows from an identical argument, and we just sketch the $k>1$ case. As before, for any $1 > \eta > 0,$ there is an $M(\delta,\eta)$ sufficiently large so that for a fixed set of vertices $w_1,w_2,\ldots,w_{\lceil {n}^{\epsilon} \rceil},$ \[ \Pr \left[ \deg_{\mbox{lk}(f)}(w_i) \leq d(t_0)/M,~\forall~1 \leq i \leq\lceil n^{\epsilon} \rceil \right] = O(\exp(-{n}^{\epsilon}d(t_0)(1-\eta))). \] This overwhelms the $O(\exp( (1-\epsilon)n^{\epsilon}\log n))$ possible choices of vertices as \[ d(t_0)/\log n > (1+\delta)(1+o(1)) \] and $\eta$ may be chosen sufficiently small. As there are only $O(n^{k-1})$ many links to consider, this may be taken to hold for all links simultaneously with overwhelming probability. We now turn to the second condition. For a fixed $(k-1)$-dimensional face $f$, let $X_f$ denote the number of $k$-faces in $\LMP[t_0][k]$ containing $f.$ Note that every degree of every link of a $(k-2)$-dimensional face of $\LMP[t_0][k]$ can be identified with some $X_f,$ and in fact we have \[ \frac{1}{k} \sum_{f \in F_{k-2}}\left| \ensuremath{\aleph_M}[f][t_0] \right| = \sum_{f \in F_{k-1}} \one[X_f \leq d(t_0)/M]. \] Thus by adjusting $\epsilon,$ it suffices to show the claim for the right hand side. Call a collection $S$ of $(k-1)$-faces \emph{balanced} if \[ \max_{w \in F_{k-2}} \left| \{ \sigma \in S~:~ w \subset \sigma \}\right| \leq n^{\epsilon}. \] Note that for each $w \in F_{k-2},$ the collection $\{ X_{\sigma} ~:~ \sigma \in S, w \subset \sigma \}$ can be identified with the degrees of a collection of vertices in the link $\mbox{lk}(w).$ Therefore, we may assume that the collection is balanced. By symmetry we have \begin{multline*} \Pr\left[ \exists~f_1, f_2,\ldots,f_r~:~ X_{f_i} \leq d(t_0)/M,~1\leq i \leq r,~\{f_i\}~\text{balanced} \right] \\ \leq { {n \choose k} \choose r} \Pr\left[ X_{f_i} \leq d(t_0)/M, 1 \leq i \leq r,~\{f_i\}~\text{balanced} \right]. \end{multline*} Let $X$ denote the number of $k$-faces that contain some $f_i.$ If every $X_{f_i} \leq d(t_0)/M,$ it follows that $X \leq rd(t_0)/M.$ Each $f_i$ is contained in $n-k$ possible $k$-faces, but it may be possible that some $f_i$ and $f_j$ are both contained in a single $k$-face. If this occurs, however, it must be that $|f_i \cap f_j|=k-1.$ In other words, each contains a common $(k-2)$-face. Furthermore, there is at most one $k$-face that contains both $f_i$ and $f_j.$ A fixed face $f_j$ contains $k$ distinct $(k-2)$-faces $q_1,q_2,\ldots,q_k.$ As $\{f_i\}$ is balanced, each $q_l$ is contained in at most $n^{\epsilon}$ distinct $f_i.$ Thus there are at most $n^{\epsilon}k$ many $k$-faces that contain $f_j$ and some other $f_i,$ and this implies there are at least $r(n-k-n^{\epsilon}k)$ distinct possible $k$-faces that contain some $f_i.$ It follows that $X$ stochastically dominates a \( \operatorname{Binom}\left( \bigl\lceil r(n-k-n^{\epsilon}k)\bigr\rceil, p(t_0)\right) \) variable. Applying Lemma~\ref{magic}, we get \begin{multline*} \Pr\left[ X_{f_i} \leq d(t_0)/M, 1 \leq i \leq r,~\{f_i\}~\text{balanced} \right] \leq \Pr\left[ X \leq rd(t_0)/M \right] \\ \leq \exp\left( -r(n-k-n^{\epsilon}k)p(t_0) + \tfrac{rd(t_0)}{M}(1 + \log\tfrac{M(1+r(n-k-n^{\epsilon}k))p(t_0)}{rd(t_0)}) \right). \end{multline*} Thus, we get \begin{multline*} \log \Pr\left[ \exists~f_1, f_2,\ldots,f_r~:~ X_{f_i} \leq d(t_0)/M,~1\leq i \leq r \right] \\ \leq r\left[ (k\log n - \log r)-d(t_0) + \frac{d(t_0)}{M}( 1 + \log(M)) \right](1+o(1)). \end{multline*} Since $d(t_0) \geq (k - 1 + \delta) \log n - o(1),$ we can set $r = [n^{1 - \delta/2}]$ and make $M$ sufficiently large that \[ (k\log n - \log r)-d(t_0) + \frac{d(t_0)}{M}( 1 + \log(M)) \to -\infty. \] Taking $\epsilon = \delta/4,$ we have shown the desired claim. \end{proof} With global control on the number of exceptional vertices, the proof now reduces to essentially a union bound over all later times and links. \begin{lemma} \label{lem:proc_easy} There is a constant $C=C(k)$ so that with high probability, every $\mbox{lk}(f)$ of $\LMP[t][k]$ satisfies \begin{enumerate} \item {\bf Bounded degree (b.d.c)} Every vertex has degree at most $C d(t).$ \item {\bf Adjacency matrix} The adjacency matrix of the link satisfies \[ \sup_{ \substack{ \|x\| = 1, x^t \one = 0 \\ \|y\| = 1 } } | x^tAy| \leq C\sqrt{d(t)}. \] \item {\bf Parallel eigenspaces} Setting $\ensuremath{\aleph_M} = \ensuremath{\aleph_M}[f][t]$ and $T$ to be the diagonal matrix of degrees of the link, \[ \sup_{ \substack{ \|x\| = 1, \\ x^t T^{1/2}\one_{W} = 0 } } | x^t T^{-1/2} \one_{\ensuremath{\aleph_M}^c}| \leq C\frac{\sqrt{n}}{d(t)}. \] \end{enumerate} for all $t \geq t_0.$ \end{lemma} \begin{proof} Let $I$ be the interval $[t_1,t_2],$ where $t_0 \leq t_1 \leq t_2.$ The probability that there are two faces that appear in this interval can be bounded by \[ \Pr \left[ \exists~\sigma_1, \sigma_2~:~T_{\sigma_1} \in I \text{ and } T_{\sigma_2} \in I \right] \leq { n \choose k}^2 \left( p(t_2) - p(t_1) \right)^2. \] Let $r$ be the smallest integer so that $p(t_0) + rn^{-2k-1} \geq 1.$ Set $p_i = p(t_0) + i n^{-2k-1}$ for all $0 \leq i < r,$ and set $p_r = 1.$ Let $t_i$ be such that $p(t_i) = p_i,$ and set $t_r = \infty.$ Note that for $t \in [t_i, t_{i+1}),$ $\LMP[t][k] \neq \LMP[t_i][k]$ and $\LMP[t][k] \neq \LMP[t_{i+1}][k]$ implies there must be two faces $\sigma_1$ and $\sigma_2$ for which $T_{\sigma_1}, T_{\sigma_2} \in [t_i,t_{i+1}).$ Hence, \begin{align*} \Pr \left[ \exists ~t \geq t_0 ~:~ \LMP[t][k] \neq \LMP[t_i][k]~\forall~ 0 \leq i \leq r \right] &\leq \sum_{i=0}^{r-1} \Pr \left[ \exists~\sigma_1, \sigma_2~:~T_{\sigma_1}, T_{\sigma_2} \in I \right] \\ &\leq \sum_{i=0}^{r-1} n^{-2k-2} \leq n^{-2}. \end{align*} By applying Lemma~\ref{lem:regularity},Proposition~\ref{prop:adj}, and Lemma~\ref{lem:T_flattens} with $m$ sufficiently large, we may thus assure that there is a constant sufficiently large that these properties occur for all links of all $\LMP[t_i][k],$ for $0 \leq i \leq {r-1}.$ \end{proof} \begin{lemma} \label{lem:fuzz_process} There is an $M=M(k,\delta)$ and a constant $C=C(M,k)$ so that with $t_1$ satisfying $p(t_1) = C \log n / n,$ all \( \ensuremath{\aleph_M}[f][t] = \emptyset \) for $t \geq t_1$ with high probability. Further, for all $t_1 \geq t \geq t_0$ every $\mbox{lk}(f)$ of $\LMP[t][k]$ satisfies \begin{enumerate} \item $|\ensuremath{\aleph_M}| \leq \frac{n}{2},$ \item $\ensuremath{\aleph_M}$ is an independent set, \item and \( \max_{u \in \ensuremath{\aleph_M}^c} \dset{u}{\ensuremath{\aleph_M}} \leq 1\) \end{enumerate} with $\ensuremath{\aleph_M}=\ensuremath{\aleph_M}[f][t].$ \end{lemma} \begin{proof} There is an $M_1$ so that this holds for $\LMP[t_0][k]$ by Proposition~\ref{prop:key_structure} and by taking the union bound over all links. Likewise, there is an $M_2$ so that the conclusions of Lemma~\ref{lem:supersmall} holds. Take $M$ to be the maximum of these, and note that from monotonicity, the conclusions of both the proposition and lemma hold. As $\ensuremath{\aleph_M}[f][t]$ is monotone in $t$ also, we have that \[ |\ensuremath{\aleph_M}[f][t]| \leq |\ensuremath{\aleph_M}[f][t_0]| \leq n/2 \] is satisfied for all $n$ sufficiently large. From a union bound and Lemma~\ref{magic}, we may choose $C=C(M,k)$ sufficiently large so that with probability going to $1,$ \[ \ensuremath{\aleph_M}[f][t_1] = \emptyset \] for all $f \in F_{k-2}.$ Let $\tau_i$ be the times at which the $i^{th}$ face is added to $\LMP[t][k]$ after time $t_0,$ and let $\tau_0 = t_0.$ Likewise, let $\Delta_i$ denote the $i^{th}$ face, and let $\mathscr{F}(\tau_i) = \sigma(\LMP[\tau_i][k]).$ Let $N$ denote the largest $i$ so that $\tau_i \leq C\log n / n.$ From Chernoff bounds, there are at most $100 C (\log n) n^{k}$ many $k$-dimensional faces in $\LMP[t_1][k]$ with overwhelming probability, and hence $N \leq 100 (\log n) n^k$ with overwhelming probability. We begin by bounding the probability that a newly added face creates an edge between two vertices of $\ensuremath{\aleph_M}[f][t]$ for some $f \in F_{k-2}.$ \begin{align} \label{eq:indset} \Pr\left[ \exists~ u,v \in \ensuremath{\aleph_M}[f][\tau_i]~:~u,v \in \Delta_{i+1} \middle \vert \mathscr{F}(\tau_i) \right] &\leq \frac{ | \ensuremath{\aleph_M}[f][\tau_i] |^2 }{ |F_k| - |\LMP[\tau_i][k]| } \\ \notag &\leq \frac{ | \ensuremath{\aleph_M}[f][\tau_0] |^2 }{ |F_k| - |\LMP[t_1][k]| }. \end{align} Let $E_{i,f}$ denote the event that \begin{enumerate} \item \( |\LMP[t_1][k] | \leq 100 C n^k\log n, \) \item \( \sum_{f \in F_{k-2}} | \ensuremath{\aleph_M}[f][t_0] |^2 \leq n^{1 - \epsilon}, \) \item there exist $u$ and $v$ in $\ensuremath{\aleph_M}[f][\tau_i]$ so that $u \in \Delta_{i+1}$ and $v \in \Delta_{i+1}.$ \end{enumerate} By conditioning, we have that \begin{align*} \Pr \left[ \cup_{i,f} E_{i,f} \right] &\leq \mathbb{E} \sum_{i=0}^N \sum_{f \in F_{k-2}} \frac{ | \ensuremath{\aleph_M}[f][\tau_0] |^2 \one[E_{i,f}] }{ |F_k| - |\LMP[t_1][k]| } \\ &\leq \mathbb{E} \sum_{i=0}^N \frac{ \sum_{f \in F_{k-2}} | \ensuremath{\aleph_M}[f][\tau_0] |^2 \one[E_{i,f}] }{ |F_k| - 100C n^k \log n } \\ &\leq \mathbb{E} \sum_{i=0}^N \frac{ n^{1-\epsilon}\one[\LMP[t_1][k] \leq 100Cn^k\log n] }{ |F_k| - 100C n^k \log n } \\ &\leq \frac{ {(100Cn^k\log n)} n^{1-\epsilon} }{ |F_k| - 100C n^k \log n } = O(n^{-\epsilon} \log n). \end{align*} Thus with high probability, no face added between $t_0$ and $t_1$ creates an edge between two elements of any $\ensuremath{\aleph_M}[f][t].$ We now turn to bounding the probability that a newly added face connects an element of $\ensuremath{\aleph_M}[f][t]$ to a neighbor of $\ensuremath{\aleph_M}[f][t].$ Let $\mathcal{N}_M^f(t)$ be the set of neighbors of $\ensuremath{\aleph_M}[f][t],$ and let $D(t)$ be an upper bound for the degree of a vertex of any link of $\LMP[t][k].$ Note that \( |\mathcal{N}_M^f(t)| \leq D(t) | \ensuremath{\aleph_M}[f][t] |. \) Then \begin{align*} \Pr\left[ \exists~ u\in \ensuremath{\aleph_M}[f][\tau_i], v\in \mathcal{N}_M^f(t) ~:~u,v \in \Delta_{i+1} \middle \vert \mathscr{F}(\tau_i) \right] \leq \frac{ D(\tau_i)| \ensuremath{\aleph_M}[f][\tau_i] |^2 }{ |F_k| - |\LMP[\tau_i][k]| }. \end{align*} With high probability, there is a constant $K$ so that all the degrees can be bounded by $K \log n$ for all $t \leq t_1.$ This failure probability is at most a logarithmic factor more than the failure probability in~\eqref{eq:indset}. Hence the same proof shows that with high probability, no added face increases \[ \max_{u \in \mathcal{V}(\mbox{lk}(f)) \setminus \ensuremath{\aleph_M}[f][t]} e(u,\ensuremath{\aleph_M}[f][t]). \] \end{proof} \begin{pfofthm}{Theorem~\ref{thm:stopping_time}} We are essentially ready to apply Lemma~\ref{lem:deterministic}. The only concern is that in~\eqref{eq:fuzz_t}, the set $\ensuremath{\aleph_M}[f][t]$ is defined in terms of $d(t_0)$ and not $d(t).$ However, as noted in Lemma~\ref{lem:fuzz_process}, all these sets disappear once $p(t_1)=C\log n / n,$ at which point $d(t)$ has only risen by a factor of $K=\frac{p(t_1)}{p(t_0)}.$ Thus, \[ Q^f(t) = \{ w \in \mathcal{V}(\mbox{lk}(f))~:~ \deg_{\mbox{lk}(f)}(w) \leq d(t)/KM \} \subseteq \ensuremath{\aleph_M}[f][t], \] for all $t \leq t_1,$ and by monotonicity, all the desired properties of $\ensuremath{\aleph_M}[f][t]$ transfer to $Q^f(t).$ Thus Lemmas~\ref{lem:proc_easy} and~\ref{lem:fuzz_process} show all the needed properties of Lemma~\ref{lem:deterministic} apply, completing the proof. \end{pfofthm} \section{Cohomology structure theorem} \label{sec:cohom} The structure theorem for cohomology relies on the following theorem of {Ballman--\'Swi\k{a}tkowski }~\cite{BS}. \begin{BSC} \label{thm:BStool} If \( \Delta \) is a finite, pure \(d\)-dimensional simplicial complex, so that for every $(d-2)$-dimensional face $\sigma,$ the normalized Laplacian $L= L [\mbox{lk}(\sigma) ]$ satisfies \( \lambda_2 > 1- \frac{1}{d} \) then \(H^{d-1}(\Delta, \mathbb{Q}) =0.\) \end{BSC} \begin{proof}[Proof of Theorem~\ref{thm:Hhit}] Recall that we define $t_0$ so that $p(t_0) = (d-1+\delta)\log n / n.$ Let $\tilde{Y_t}$ denote the simplicial complex $\LMP$ with all its isolated $(d-1)$-faces deleted. By Theorem~\ref{thm:stopping_time}, w.h.p. for all $t \geq t_0$, all links of $\tilde{Y_t}$ have $\lambda_2(L) = 1-o(1).$ We need to check that $\tilde{Y_t}$ is pure $d$-dimensional, i.e. that every face is contained in some $d$-dimensional face. Note that this can only fail if there is some $(d-2)$-dimensional face of $\LMP[t]$ that is not contained in any $d$-dimensional face. As this is a monotone property, it suffices to check that $\LMP[t_0]$ has no such $(d-2)$-faces. Put $I$ to be the number of isolated $(d-2)$-faces in $\LMP[t_0].$ Then \[ \mathbb{E} I = {n \choose d-1} (1-p(t_0))^{ n^2/2}(1-o(1)), \] which decays exponentially in $n.$ Hence, $\tilde{Y_t}$ is pure $d$-dimensional for all $t \geq t_0$, and so Theorem~\ref{thm:BStool} applies. It follows that $H^{d-1}( \tilde{Y_t}, \mathbb{Q}) = 0,$ and it remains to compare $H^{d-1}( \tilde{Y_t}, \mathbb{Q})$ and $H^{d-1}( \LMP, \mathbb{Q}).$ For what remains, fix $t \geq t_0.$ It will follow from induction that each additional $(d-1)$-face we glue to $\tilde{Y}$ increases the dimension of the $(d-1)$ cohomology by $1.$ Let $Z$ be the complex formed by including one of the isolated $(d-1)$-faces of $Y$ back into $\tilde{Y}.$ Let $B$ be a neighborhood of the included $(d-1)$ face that is homotopic to a single $(d-1)$-simplex. Then the Mayer-Vietoris sequence (see Chapter 3 of~\cite{Hatcher}) for the $(d-1)$-dimensional cohomology is \[ \cdots \rightarrow H^{d-1}(Z,\mathbb{Q}) \rightarrow H^{d-1}(\tilde{Y},\mathbb{Q}) \oplus H^{d-1}(B,\mathbb{Q}) \rightarrow H^{d-1}(\tilde{Y} \cap B,\mathbb{Q}) \rightarrow H^{d}(Z,\mathbb{Q}). \] As \(\tilde{Y} \cap B\) is homotopic to a $(d-2)$-dimensional sphere, \( H^{d-1}(\tilde{Y} \cap B,\mathbb{Q}) = 0 \). Also, \(H^{d-1}(B,\mathbb{Q}) = \mathbb{Q} \), and so this sequence becomes \[ 0 \rightarrow H^{d-1}(Z,\mathbb{Q}) \rightarrow H^{d-1}(\tilde{Y},\mathbb{Q}) \oplus \mathbb{Q} \rightarrow 0, \] or otherwise stated, \( H^{d-1}(Z,\mathbb{Q}) \cong H^{d-1}(\tilde{Y},\mathbb{Q}) \oplus \mathbb{Q} \). Each additional isolated $(d-1)$-faces increases the dimension by one by the very same argument, which completes the proof. \end{proof} \section{Property (T)} \label{sec:propT} The proof here is nearly identical to the proof of the cohomology vanishing structure theorem. To establish our results concerning property (T) of random fundamental groups, we will use the following theorem of \.Zuk. \begin{Zuk} \label{thm:tool} If $X$ is a pure $2$-dimensional locally-finite simplicial complex so that for every vertex $v$, the vertex link $\mbox{lk}(v)$ is connected and the normalized Laplacian $L= L [\mbox{lk}(v) ]$ satisfies $\lambda_2(L) > 1/2$, then $\pi_1(X)$ has property~(T). \end{Zuk} \begin{proof}[Proof of Theorem \ref{thm:structure}] Recall that we define $t_0$ so that $p(t_0) = (d-1+\delta)\log n / n.$ Let $\tilde{Y_t}$ denote the simplicial complex $\LMP$ with all its isolated edges deleted. By Theorem~\ref{thm:stopping_time}, w.h.p. for all $t \geq t_0$, all links of $\tilde{Y_t}$ have $\lambda_2(L) = 1-o(1).$ Then by \.Zuk's criterion, $\pi_1(\tilde{Y_t})$ has property~(T) for all $t \geq t_0.$ Fix $t \geq t_0.$ It only remains to compare the fundamental groups $\pi_1(\tilde{Y})$ and $\pi_1(Y)$. But attaching a $1$-cell to a connected CW complex $W$ adds a free $\mathbb{Z}$-factor to the fundamental group $\pi_1(W)$, by the Seifert--van Kampen theorem (see Theorem 1.20 of~\cite{Hatcher}). So we only need to check that deleting all the isolated edges in $Y$ does not result in a disconnected complex $\tilde{Y}$. Removing less than $n-1$ edges from the complete graph $K_n$ can not disconnect it; indeed, to separate a component of order $k$ form the rest of the graph requires removing at least $k(n-k)$ edges, which is minimized when $k=1$. Thus we need only check that the number of isolated edges is fewer than $n-1.$ From monotonicity, it suffices to show that at time $t_0$ the number of isolated edges is w.h.p. $o(n).$ By linearity of expectation, the expected number of edges deleted $\mathbb{E}[D]$ is given by \begin{align*} \mathbb{E}[D] &= { n \choose 2} (1 - p(t_0) )^{n-2}\\ & \le \frac{1}{2} n^2 \exp ( -p(t_0) (n-2) )\\ & \le O \left( n^{1-c} \right) \end{align*} for some constant $c > 0$. By the second moment method, for example, $D$ is tightly concentrated around its mean, so w.h.p.\ $\tilde{Y}$ is connected. The claim follows. \end{proof} Corollary \ref{cor:Twin} quickly follows. \begin{proof}[Proof of Corollary \ref{cor:Twin}] Let $I$ denote the number of isolated edges. The expected number of isolated edges $\mathbb{E}[I]$ is \begin{align*} \mathbb{E} [i] &= {n \choose 2} (1 - p)^{n-2} \leq n^2 e^{-np} \end{align*} Taking $p = (2\log n + f(n))/n,$ where $f(n) \to \infty,$ this is seen to go to $0,$ completing the proof. \end{proof} \section{Kahn-Szemer\'erdi argument} \label{sec:KSz} We begin with a proof of the regularity conditions. \begin{pfofthm}{Lemma \ref{lem:regularity}} For any vertex $v,$ $\deg(v)$ is a binomial random variable with mean $d>\delta \log(n)$. By Lemma \ref{maybe}, \( \mathbb{P}(\deg(v)>c_{0} d) \leq \exp\left(-\tfrac{d c_{0} \log c_{0} }{3}\right) \) provided $c_{0} > 4.$ Thus taking the union bound over all vertices, we get that \[ \Pr\left[ \ensuremath{ \bf{b.d.c.}} \text{ fails } \right] \leq \exp( d (\tfrac{1}{\delta}-\tfrac{c_{0} \log c_{0} }{3})). \] By taking $c_{0}$ sufficiently large, we may take \[ \frac{1}{\delta}-\frac{c_{0} \log c_{0} }{3} \leq -m, \] completing the proof of the first claim. We will now turn to showing the discrepancy property, for which we need to show there are constants $c_i=c_i(\delta,m)$ so that \begin{enumerate} \item $ \tfrac{e(A,B)}{\mu(A,B)} \leq c_{1} $ \item $ e(A,B)\log \tfrac{e(A,B)}{\mu(A,B)} \leq c_{2}( |A| \vee |B|)\log \tfrac{n}{|A| \vee |B|} $ \item $ |A| \vee |B| \leq d^{1/4}/100 $ \end{enumerate} Note that these properties are monotone in $c_i$, and so we are free to increase the constants as need be throughout the proof. Let $D$ be the event that the discrepancy condition fails and let $D(A,B)$ be the event that the discrepancy condition fails for sets $A$ and $B$. Then by the union bound \begin{eqnarray*} \mathbb{P}(D) &\leq& \mathbb{P}(\exists A,B \text{ with $|A| \wedge |B|\geq n/e$}): D(A,B) \text{ occurs})\\ &+& \mathbb{P}(\exists A,B \text{ with $|A| \vee |B| \geq n/e \geq |A| \wedge |B|$}): D(A,B) \text{ occurs})\\ &+&\sum_{A,B: \ |A| \vee |B|< n/e}\mathbb{P}(D(A,B)) \end{eqnarray*} Taking $c_{1} > e^2$, then when $|A| \wedge |B| \geq \tfrac{n}{e},$ \[ e(A,B)>c_{1} \mu(A,B)>c_{1} (n/e)^2d/n>nd. \] Thus, there are at least $nd$ edges in the graph. The distribution of the number of edges is binomial with mean $n(n-1)p/2=nd/2$, and so the probability of this is going to zero exponentially in $nd$, i.e. \begin{equation} \label{verystupid} \mathbb{P}(\exists A,B \text{ with $|A| \wedge |B|\geq n/e$}): D(A,B) \text{ occurs})= O(\exp(-cnd)) \end{equation} for some absolute constant $c >0.$ If $|A| \vee |B| \geq \tfrac{n}{e}> |A| \wedge |B|,$ and if the bounded degree condition holds, then $e(A,B) \leq (|A| \vee |B|)c_{0} d$ and \[ \frac{e(A,B)}{\mu(A,B,n)} \leq \frac{c_{0} nd(|A| \vee |B|)}{|A||B|d} = \frac{c_{0} n}{|A| \wedge |B|} \leq c_{0} e \] Thus taking $c_{1} > c_{0} e$, we have that \begin{eqnarray} \mathbb{P}(\exists A,B \text{ with $|A| \vee |B| \geq n/e \geq |A| \wedge |B|$}): D(A,B) \text{occurs}) &\leq& \mathbb{P}(\ensuremath{ \bf{b.d.c.}} \text{fails}) \nonumber\\ &=&O(\exp(-md)). \label{twoam} \end{eqnarray} Now we need to deal with the case that both $A$ and $B$ are less than $\tfrac{n}{e},$ but at least one is greater than $d^{1/4}/100.$ Take $c_{2} > 18 + 1200m.$ For emphasis, we will write $\mu(A,B,n)=\mu(A,B) = \frac{|A||B|d}{n}.$ Choose $r=r(A,B,n)=c_{1} \vee r_1$ where $r_1$ is the solution to \[ \mu(A,B,n) r_1\log(r_1) = c_{2}(|A| \vee |B|)\log \tfrac{n}{|A| \vee |B|}. \] For any $A$, $B$ and $n$ we must have either \begin{itemize} \item $e(A,B) \leq r \mu (A,B,n)$ and $r =c_{1} $ \item $e(A,B) \leq r \mu (A,B,n)$ and $r =r_1$ or \item $e(A,B)>r \mu(A,B,n)$ \end{itemize} Thus if $D(A,B)$ occurs then at least one of the following three events occur. \begin{itemize} \item $D_1=D_1(A,B)=\bigg\{e(A,B)\leq r \mu(A,B,n),\, r=c_{1} \text{ and }\\ {} \hspace{1.9in} e(A,B)> c_{1} \mu(|A|,|B|,n) \bigg\}$ \item $D_2=D_2(A,B)=\bigg\{e(A,B)\leq r{\mu(A,B,n)}, r=r_1 \text{ and } \\ {} \hspace{1.9in} e(A,B)\log \tfrac{e(A,B)}{\mu(A,B,n)}> c_{2} (|A| \vee |B|)\log \tfrac{n}{|A| \vee |B|}\bigg\}$ \item $D_3=D_3(A,B)=\{e(A,B)>r \mu(A,B,n)\}$ \end{itemize} For $D_1$ the conditions are mutually exclusive as $e(A,B)$ can not be simultaneously greater than and less than or equal to $c_{1} \mu(A,B,n)$. Thus $D_1(A,B)$ is empty. For $D_2$ we get similar contradiction after a little work. \begin{eqnarray*} e(A,B) \log\tfrac{e(A,B)}{\mu(A,B,n)} &>&c_{2} (|A|\vee |B|)\log \tfrac{n}{|A| \vee |B|}\\ e(A,B) \log\tfrac{e(A,B)}{\mu(A,B,n)} &>& \mu(A,B,n)r_1\log r_1 \\ \tfrac{e(A,B)}{\mu(A,B,n)} \log\tfrac{e(A,B)}{\mu(A,B,n)} &>& r_1 \log r_1\\ \tfrac{e(A,B)}{\mu(A,B,n)} &>& r_1 \\ e(A,B) &>& r_1 \mu(A,B,n) \\ e(A,B) &>& r \mu(A,B,n). \end{eqnarray*} This is a contradiction so $D_2(A,B)$ is also empty. Now we bound $\mathbb{P}(D_3(A,B)).$ As $e(A,B)$ is binomial with mean at most $\mu(A,B,n)$, Lemma \ref{maybe} implies \[ \mathbb{P}(D_3(A,B)) \leq \exp\left(-\tfrac{\mu(|A|,|B|,n) r \log r}{3}\right) \] for any $r\geq 4.$ For all $A,B$ we have $D \subset D_1 \cup D_2 \cup D_3$ and $\mathbb{P}(D_1(A,B))=\mathbb{P}(D_2(A,B))=0$. Combining this with (\ref{verystupid}) and (\ref{twoam}) we get \begin{eqnarray*} \mathbb{P}(D) & \leq & \mathbb{P}(\exists A,B:\ D(A,B) \text{ occurs})\\ & \leq & \mathbb{P}(\exists A,B:\ |A|,|B|<n/e \text{ and } D(A,B) \text{ occurs})+O(\exp(-md))\\ & \leq & \mathbb{P}(\exists A,B:\ |A|,|B|<n/e \text{ and } D_3(A,B) \text{ occurs})+O(\exp(-md))\\ & \leq & \sum_{|A|,|B|}\mathbb{P}(D_3(A,B)) +O(\exp(-md)) \\ & \leq & \sum_{a,b} \sum_{|A|=a,|B|=b} \exp\left(-\tfrac{\mu r \log r}{3}\right) +O(\exp(-md))\\ & \leq & \sum_{a,b} {n \choose a} {n\choose b}\exp\left(-\tfrac{\mu(a,b,n) r \log r}{3}\right) +O(\exp(-md)), \end{eqnarray*} where the sums are over all pairs $(a,b)$ with $d^{1/4}/100 \leq a \vee b \leq n/e.$ To evaluate the last term we get \begin{eqnarray*} \tfrac{\mu r \log r}{3} & \geq &\bigg(6 + 400m\bigg)\bigg((|A| \vee |B|) \log\tfrac{n}{|A| \vee |B|}\bigg)\\ & > & \bigg(2+2+2+(400m))\bigg)\bigg((|A|\vee |B|) \log\tfrac{n}{|A|\vee|B|}\bigg)\\ & >& 2|A|( \log\tfrac{n}{|A|}) +2|B|( \log\tfrac{n}{|B|}) +2\log n +4md^{1/4}\log\tfrac{100n}{d^{1/4}} \\ & >& |A|(1 + \log\tfrac{n}{|A|}) +|B|(1 + \log\tfrac{n}{|B|}) +2\log n +3md^{1/4}\log n. \end{eqnarray*} The first line is due to the definitions of $r$ and $c_{2}$. In the third line we use the monotonicity of $x \log \tfrac{n}{x}$ on $[1,n/e]$ by substituting in $|A|$, $|B|,$ $1$ and $d^{1/4}/100$ for $x$. In the fourth line we use that $|A| \vee |B| \leq \tfrac{n}{e}$ so $\log \tfrac{n}{|A|},\log \tfrac{n}{|B|}>1$ Exponentiating we get \[ \exp\left[ \tfrac{\mu r \log r}{3} \right] \geq \left(\tfrac{en}{|A|}\right)^n \left(\tfrac{en}{|B|}\right)^n n^2 \exp(3md^{1/4}(\log n)) \] It follows that \begin{eqnarray*} {n \choose a} {n\choose b}\exp\left(-\tfrac{\mu(a,b,n) r \log r}{3}\right) &\leq & {n \choose a}{n \choose b}\left(\tfrac{en}{a}\right)^{-n} \left(\tfrac{en}{b}\right)^{-n} n^{-2}\exp(-3md^{1/4}\log n)\\ &\leq & n^{-2}\exp(-3md^{1/4}\log n). \end{eqnarray* Putting this together we get \begin{eqnarray*} \mathbb{P}(D) & \leq & \sum_{d/100 \leq a \vee b \leq n/e }{n \choose a} {n\choose b}\exp\left(-\tfrac{\mu(a,b,n) r \log r}{3}\right) +O(\exp(-md))\\ & \leq & n^2n^{-2}\exp(-3md^{1/4}\log n)+O(\exp(-md)). \end{eqnarray*} Thus the lemma is satisfied.\end{pfofthm} We finally give a quick sketch of how Proposition~\ref{prop:adj} follows from Lemma~\ref{lem:regularity}. This is nearly the same as Theorem 2.5 of~\cite{FeigeOfek}, and so we will cite them heavily. \begin{pfofthm}{ Proposition \ref{prop:adj}} We recall that we wish to bound \[ \sup_{ \substack{ \|x\| = 1, x^t \one = 0 \\ \|y\| = 1 } } | x^tAy| \leq C\sqrt{d}. \] For this we will relax the supremum to a finite, discrete space. Define \[ \mathcal{U}=\left\{ \frac{ z}{2\sqrt{n}} ~:~ z \in \mathbb{Z}^n, \|z\|^2 \leq 4n\right\} ~~~~~\text{ and }~~~~~ \mathcal{T}=\left\{ z \in \mathcal{U} ~:~ z \perp \one \right\}. \] As $\mathcal{U}$ is $\tfrac{1}{2}$-net of the sphere, and $S=\{x~:\|x\| = 1, x^t \one = 0\}$ is in the convex hull of $\mathcal{T}$ (by Lemma 2.3 of~\cite{FeigeOfek}), we have that \[ \sup_{ \substack{ \|x\| = 1, x^t \one = 0 \\ \|y\| = 1 } } | x^tAy| \leq 4 \sup_{ \substack{ x \in \mathcal{T} \\ y \in \mathcal{U}} } | x^tAy|. \] Further, we have that $|\mathcal{T}| \leq |\mathcal{U}| \leq C^n$ for some absolute constant $C.$ For a fixed pair of vectors $(x,y) \in \mathcal{T} \times \mathcal{U},$ define the {\bf light couples} $\mathcal{L} = \mathcal{L}(x,y)$ to be all those ordered pairs $(u,v) \in \{1,2,\dots,n\}^2$ so that $|x_uy_v| \leq \tfrac{\sqrt{d}}{n},$ and let the {\bf heavy couples} $\mathcal{H}=\mathcal{H}(x,y)$ be all those pairs that are not light. We will use the notation \[ \ensuremath{ \operatorname{light}(x,y)} = \sum_{ (u,v) \in \mathcal{L}} x_uA_{uv}y_v, \] and the notation \[ \ensuremath{ \operatorname{heavy}(x,y)} = \sum_{ (u,v) \in \mathcal{H}} x_uA_{uv}y_v, \] Now the contribution of the light couples can be bounded by a martingale argument to show that for every $c > 0,$ there is a $K$ so that \[ \Pr \left[ |\ensuremath{ \operatorname{light}(x,y)}| > K\sqrt{d} \right] \leq e^{-cn}. \] (c.f. Claim 2.7 of~\cite{FeigeOfek}--note we are asserting a stronger claim than what Feige and Ofek claim, as $y \in \mathcal{U},$ but the same proof goes through without modification). And thus, for each $m$ there is a constant $C=C(m)$ so that \[ \Pr \left[ \sup_{(x,y) \in \mathcal{T} \times \mathcal{U}} |\ensuremath{ \operatorname{light}(x,y)}| > C\sqrt{d} \right] \leq Ce^{-mn}. \] To control the heavy couples, the discrepancy property suffices (c.f. Corollary 2.11 of~\cite{FeigeOfek} or Section 2.3 of~\cite{FKSz}). Again, the proof is identical to either of those two claims, although it is not exactly either one, and it shows that there is a constant $C=C(\delta,m)$ sufficiently large so that \begin{multline*} \Pr \left[ \sup_{(x,y) \in \mathcal{T} \times \mathcal{U}} |\ensuremath{ \operatorname{heavy}(x,y)}| > C\sqrt{d} \right] \\ \leq \Pr \left[ \ensuremath{ \bf{b.d.c.}} \text{ fails } \right] + \Pr \left[ \text{ discrepancy fails } \right]. \end{multline*} \end{pfofthm} \begin{appendices} \section{Estimates of Binomial Random Variables} \begin{lemma} \label{magic} Let $X$ be a binomial random variable with mean $\mu$. Then for any $t \leq \mu$ \[ \mathbb{P} \left[ X \leq t \right] \leq \exp\left[ -\mu + t( 1+ \log \tfrac{\mu}{t}) \right], \] \end{lemma} \begin{pfofthm}{Lemma~\ref{magic}} The proof follows from a standard estimate on the Laplace transform combined with Markov's inequality. For any $\lambda \in \mathbb{R},$ the Laplace transform of $X \sim \operatorname{Binomial}(n,p)$ can be bounded by \begin{align*} \mathbb{E} e^{\lambda X} &= \left(pe^{\lambda} + (1-p)\right)^n \\ &= \left(1 + p(e^{\lambda}-1)\right)^n \\ &\leq \exp\left[\mu(e^{\lambda}-1)\right]. \end{align*} Provided that $\lambda < 0,$ the tail bound now can be bounded by Markov's inequality by \begin{align*} \mathbb{P} \left[ X \leq t \right] &= \mathbb{P} \left[ e^{\lambda X} \geq e^{\lambda t} \right] \\ &\leq \left[\mathbb{E} e^{\lambda X}\right] e^{-\lambda t} \\ &\leq \exp\left[\mu(e^{\lambda}-1) - \lambda t\right]. \end{align*} Assuming that $t < \mu,$ this bound holds with $\lambda=\log (t/\mu),$ which upon evaluation gives \[ \mathbb{P} \left[ X \leq t \right] \leq \exp\left[\mu(e^{\log (t/\mu)}-1) - \log (t/\mu) t\right] = \exp \left[ -\mu + t(1 + \log \tfrac{\mu}{t}) \right]. \] \end{pfofthm} \begin{lemma} \label{maybe} Let $X$ be a binomial random variable with mean $\mu$. Then for any $t>4$ \[ \mathbb{P} \left[ X \geq t \mu \right] \leq \exp\left[ -\frac{t \mu \log(t)}{3} \right], \] \end{lemma} \begin{pfofthm}{Lemma \ref{maybe}} The proof here is identical in approach to the proof of Lemma~\ref{magic}. As there, it is possible to bound the Laplace transform of $X$ as \[ \mathbb{E} e^{\lambda X} \leq \exp\left[\mu(e^{\lambda}-1)\right], \] for any real $\lambda.$ For $\lambda > 0,$ the tail bound follows from Markov's inequality by \begin{align*} \mathbb{P} \left[ X \geq t\mu \right] &= \mathbb{P} \left[ e^{\lambda X} \geq e^{\lambda t\mu} \right] \\ &\leq \left[\mathbb{E} e^{\lambda X}\right] e^{-\lambda t\mu} \\ &\leq \exp\left[\mu(e^{\lambda}-1) - \lambda t\mu \right]. \end{align*} For $t > 1,$ it is possible to take $\lambda = \log t.$ This gives the bound on the tail probability \[ \mathbb{P} \left[ X \geq t\mu \right] \leq\exp\left[\mu\left(t - 1 - t \log t\right)\right]. \] To complete the proof, it remains to show that $t-1 \leq \tfrac{2}{3}t\log t$ when $t \geq 4.$ The function $\frac{t}{t-1}\log t$ is monotonically increasing for $t > 1,$ and thus it suffices to show that $\frac{4}{3}\log 4 \geq \tfrac{3}{2},$ or equivalently that $\log 4 \geq \tfrac{9}{8}.$ This follows from $\log 4 = \int_1^4 \tfrac 1 x dx$ and bounding the integral from below by a right Riemann sum. \end{pfofthm} \end{appendices} \bibliographystyle{plain}

https://arxiv.org/abs/1605.07832

Antistrong digraphs

An antidirected trail in a digraph is a trail (a walk with no arc repeated) in which the arcs alternate between forward and backward arcs. An antidirected path is an antidirected trail where no vertex is repeated. We show that it is NP-complete to decide whether two vertices $x,y$ in a digraph are connected by an antidirected path, while one can decide in linear time whether they are connected by an antidirected trail. A digraph $D$ is antistrong if it contains an antidirected $(x,y)$-trail starting and ending with a forward arc for every choice of $x,y\in V(D)$. We show that antistrong connectivity can be decided in linear time. We discuss relations between antistrong connectivity and other properties of a digraph and show that the arc-minimal antistrong spanning subgraphs of a digraph are the bases of a matroid on its arc-set. We show that one can determine in polynomial time the minimum number of new arcs whose addition to $D$ makes the resulting digraph the arc-disjoint union of $k$ antistrong digraphs. In particular, we determine the minimum number of new arcs which need to be added to a digraph to make it antistrong. We use results from matroid theory to characterize graphs which have an antistrong orientation and give a polynomial time algorithm for constructing such an orientation when it exists. This immediately gives analogous results for graphs which have a connected bipartite 2-detachment. Finally, we study arc-decompositions of antistrong digraphs and pose several problems and conjectures.

\section{Introduction} We refer the reader to~\cite{bang2009} for notation and terminology not explicitly defined in this paper. An {\bf antidirected path} in a digraph $D$ is a path in which the arcs alternate between forward and backward arcs. The digraph $D$ is said to be {\bf anticonnected} if it contains an antidirected path between $x$ and $y$ for every pair of distinct vertices $x,y$ of $D$. Anticonnected digraphs were studied in~\cite{chartrandUM51}, where several properties such as antihamiltonian connectivity have been considered. We will show in Theorem~\ref{antipathcheck} below that it is NP-complete to decide whether a given digraph contains an antidirected path between given vertices. Our main purpose is to introduce a related connectivity property based on the concept of a {\bf forward antidirected trail}, i.~e.~a walk with no arc repeated which begins and ends with a forward arc and in which the arcs alternate between forward and backward arcs. A digraph $D$ is {\bf antistrong} if it has at least three vertices and contains a forward antidirected $(x,y)$-trail for every pair of distinct vertices $x,y$ of $D$. We say that $D$ is {\bf $k$-arc-antistrong} if it has at least three vertices and contains $k$ arc-disjoint forward antidirected $(x,y)$-trails for all distinct $x,y \in V(D)$. \medskip The paper is organized as follows. First we show that, from an algorithmic point of view, anticonnectivity is not an easy concept to work with, since deciding whether a digraph contains an antidirected path between a given pair of vertices is NP-complete. Then we move to the main topic of the paper, antistrong connectivity, and show that this relaxed version of anticonnectivity is easy to check algorithmically. In fact, we show in Section~\ref{ASsec} that there is a close relation between antistrong connectivity of a digraph $D$ and its so called bipartite representation $B(D)$, namely $D$ is antistrong if and only if $B(D)$ is connected. This allows us in Section~\ref{ASaugsec} to find the minimum number of new arcs we need to add to a digraph which is not antistrong so that the resulting digraph is antistrong. Furthermore, using the bipartite representation we show in Section~\ref{matroidsec} that the arc-minimal antistrong spanning subdigraphs of a digraph $D$ form the bases of a matroid on the arc-set of $D$. More generally, we show that the subsets of $A$ which contain no closed antidirected trails are the independent sets of a matroid on $A$. In Section~\ref{ASorsec} we study the problem of deciding whether a given undirected graph has an antistrong orientation. We show how to reduce this problem to a matroid problem and give a characterization of those graphs that have an antistrong orientation. For the convenience of readers who are not familiar with matroids, we also provide, as an appendix, a purely graph theoretical proof of the key step which is Lemma \ref{lem:mat3}. Both proofs can be converted to polynomial time algorithms which either finds an antistrong orientation of the given input graph $G$ or produce a certificate which shows that $G$ has no such orientation. In Section~\ref{detachsec} we show that being orientable as an antistrong digraph can be expressed in terms of connected 2-detachments of graphs (every vertex $v$ is replaced by two copies $v',v''$ and every original edge $uv$ becomes an edge between precisely one of the 4 possible pairs $u'v',u'v'',u''v',u''v''$) with the extra requirement that the 2-detachment is bipartite and contains no edge of the form $u'v'$ or $u''v''$. This imediately leads to a characterization of graphs having such a 2-detachment. Finally, in Section~\ref{nosepsec} we show that one can decide in polynomial time whether a given digraph $D$ has a spanning antistrong subdigraph $D'$ so that $D-A(D')$ is connected in the underlying sense (while it is NP-hard to decide whether a given digraph contains a non-separating strong spanning subdigraph).\\ We conclude the paper with some remarks and open problems. \iffalse \section{Notation and Terminology} We follow the notation from~\cite{bang2009}. We denote the vertex set and arc set of a digraph $D$ by $V(D)$ and $A(D)$, respectively and write $D=(V,A)$ where $V=V(D)$ and $A=A(D)$. Unless otherwise specified, the numbers $n$ and $m$ will always be used to denote the number of vertices and arcs/edges, respectively, in the digraph or graph in question. Digraphs and graphs will not have parallel arcs or edges unless explicitly stated. We will use the notation $[k]$ for the set of integers $\{1,2,\ldots{},k\}$. The {\bf underlying graph} of a digraph $D$, denoted $U\!G(D)$, is obtained from $D$ by suppressing the orientation of each arc and deleting one copy from parallel edges arising from 2-cycles in $D$. A digraph $D$ is {\bf connected} if $UG(D)$ is a connected graph. When $xy$ is an arc of $D$ we say that $x$ is the {\bf tail} and $y$ the {\bf head} of the arc. For a digraph $D=(V,A)$ the {\bf out-degree} $d^+_D(x)$ and {\bf in-degree} $d^-_D(x)$ of a vertex $x\in V$ is the number of arcs in $A$ of the form $xy$ or $yx$, respectively. A vertex of in-degree zero or out-degree zero is a {\bf source} or a {\bf sink}, respectively. An arc $vw$ is {\bf forward} in a path/trail/walk $v_1v_2\ldots{}v_k$ if we have $v=v_i,w=v_{i+1}$ for some $i\in \{1,\dots,k-1\}$. It is {\bf backward} if we have $w=v_i,v=v_{i+1}$ for some $i\in \{1,\dots,k-1\}$. An {\bf antidirected trail} in a digraph is a trail (a walk with no arc repeated) in which the arcs alternate between forward and backward arcs. An {\bf antidirected path} is an antidirected trail where no vertex is repeated. An {\bf antidirected cycle} is the digraph we obtain by identifying the end vertices of an antidirected path with an even number of arcs. An antidirected trail $T$ is a {\bf forward} antidirected trail if it starts {\em and} ends with a forward arc. Two vertices $x,y$ in a digraph $D$ are {\bf anticonnected} if $D$ contains an antidirected path with end vertices $x$ and $y$. A digraph $D$ is {\bf anticonnected} if it contains an antidirected path between $x$ and $y$ for every pair of distinct vertices $x,y$ of $D$. Anticonnected digraphs were studied in~\cite{chartrandUM51}, where several properties such as antihamiltonian connectivity have been considered. It is easy to see that being anticonnected is not an equivalence relation on the vertex set of a digraph as we may have an antidirected path between $x$ and $y$ and between $y$ and $z$ but none between $x$ and $z$. In fact, as we shall show below in Theorem~\ref{antipathcheck}, it is NP-complete to decide whether there exists an antidirected $(x,y)$-path in a digraph $D$. The purpose of this paper is to introduce a new notion, antistrong connectivity, which turns out to take polynomial time to check and show that this has connections to several important topics in combinatorial optimization. A digraph $D$ is {\bf antistrong} if it contains a forward antidirected $(x,y)$-trail for every choice of $x,y\in V(D)$. Let us call $D$ {\bf $k$-arc-antistrong} if it contains $k$ arc-disjoint forward antidirected $(x,y)$-trails for all $x,y \in V(D)$.\\ {\bf More to be written here} \fi \section{Anticonnectivity}\label{anticonsec} It was shown in~\cite{chartrandUM51} that every connected graph $G$ has an anticonnected orientation. This can be seen by considering a breath first search tree rooted at some vertex $r$. Let $\{r\}=L_0,L_1,L_2,\ldots{},L_k$ be the distance classes of $G$. Orient all edges between $r$ and $L_1$ from $r$ to these vertices, orient all edges between $L_1\cup L_3$ and $L_2$ from $L_2$ to $L_1\cup L_3$, orient all edges from $L_4$ to $L_3\cup L_5$ etc. Finally, orient all the remaining, not yet oriented edges arbitrarily. \iffalse There are various results in the literature on antidirected cycles \begin{itemize} \item It is NP-complete to decide whether $D$ has an antidirected 2-factor~\cite{diwanDM311}. \item Every digraph with minimum semidegree (min of in and out-degree) at least $\frac{24}{46}n$ has an antidirected 2-factor~\cite{diwanDM311}. \item In~\cite{bogdanowiszAMS6} there are given necessary and sufficient conditions under which a circulant digraph $D$ can be decomposed into $k$ pairwise arc-disjoint antidirected Hamilton cycles, each induced by two jumps.\mmk{What is a jump?} \end{itemize} \fi We will show that it is NP-complete to decide if a digraph has an antidirected path between two given vertices. We need the following result which is not new, as it follows from a result in~\cite{gabowITSE2} on the vertex analogue, but we include a new and short proof for completeness. \begin{theorem} \label{avoidpairs} It is NP-complete to decide for a given graph $G=(V,E)$, two specified vertices $x,y\in V$ and pairs of distinct edges ${\cal P}=\{(e_1,f_1),(e_2,f_2),\ldots{},(e_p,f_p)\}$, all from $E$, whether $G$ has an $(x,y)$-path which avoids at least one edge from each pair in $\cal P$. \end{theorem} {\bf Proof: } We first slightly modify a very useful polynomial reduction, used in many papers such as~\cite{bangTCS438}, from 3-SAT to a simple path problem and then show how to extend this to a reduction from 3-SAT to the problem above. For simplicity our proof uses multigraphs but it is easy to change to graphs. Let $W[u,v,p,q]$ be the graph (the variable gadget) with vertices $\{u,v,y_1,y_2,\dots{}y_p,z_1,z_2,\ldots z_q\}$ and the edges of the two $(u,v)$-paths $uy_1y_2\ldots{}y_pv, uz_1z_2\ldots{}z_qv$. \\ Let ${\cal F}$ be an instance of 3-SAT with variables $x_1,x_2,\ldots{},x_n$ and clauses $C_1,C_2,\ldots{},C_m$. The ordering of the clauses $C_1,C_2,\ldots{},C_m$ induces an ordering of the occurrences of a variable $x$ and its negation $\overline{x}$ in these. With each variable $x_i$ we associate a copy of $W[u_i,v_i,p_i+1,q_i+1]$ where $x_i$ occurs $p_i$ times and $\overline{x_i}$ occurs $q_i$ times in the clauses of $\cal F$. Identify end vertices of these graphs by setting $v_i=u_{i+1}$ for $i=1,2,\ldots{},n-1$. Let $s=u_1$ and $t=v_n$ and denote by $G'$ the resulting graph. In $G'$ we respectively denote by $y_{i,j}$ and $z_{i,j}$ the vertices $y_j$ and $z_j$ in the copy of $W$ associated with the variable $x_i$. \\ Next, for each clause $C_i$ we associate this with 3 edges from $G'$ as follows: assume $C_i$ contains variables $x_j,x_k,x_l$ (negated or not). If $x_j$ is not negated in $C_i$ and this is the $r$th copy of $x_j$ (in the order of the clauses that use $x_j$), then we associate $C_i$ with the edge $y_{j,r}y_{j,r+1}$ and if $C_i$ contains $\overline{x_j}$ and this is the $k$th occurrence of $\overline{x_j}$, then we associate $C_i$ with the edge $z_{j,k}z_{j,k+1}$. We make similar associations for the other two literals of $C_i$. Thus for each clause $C_i$ we now have a set $E_i$ of three distinct edges $e_{i,1},e_{i,2},e_{i,3}$ from $G'$ and $E_i\cap E_j=\emptyset$ for $i\neq j$. \\ Now it is easy to check that $G'$ has an $(s,t)$-path which avoids at least one edge from each of the sets $E_1,E_2,\ldots{},E_m$ if and only if $\cal F$ is satisfiable. Indeed, the $(s,t)$-path goes through the `$z$-vertices' of the copy of $W$ associated with $x_i$ if and only if $x_i$ is set to {\sc true} to satisfy $\cal F$. Let us go back to the original problem. Let $H$ be the multigraph consisting of vertices $c_0,c_1,\ldots{},c_m$ and three edges (denoted $f_{i,1},f_{i,2},f_{i,3}$) from $c_{i-1}$ to $c_i$ for $i\in \{1,\dots,m\}$. Let $G$ denote the multigraph we obtain from $G'$ and $H$ by identifying $t$ and $c_0$. Let $x=s$ and $y=c_m$. Finally, form three disjoint pairs of arcs $(e_{i,1},f_{i,1}), (e_{i,2},f_{i,2}),(e_{i,3},f_{i,3})$ between $E_i$ and $\{f_{i,1},f_{i,2},f_{i,3}\}$ for every $i \in \{1 \dots m\}$.\\ By the observations above it is easy to check that $G$ has an $(x,y)$-path which avoids at least one arc from each of the forbidden pairs if and only if $\cal F$ is satisfiable. \hfill$\diamond$\\\vspace{2mm} \begin{theorem} \label{antipathcheck} It is NP-complete to decide whether a given digraph contains an antidirected path between given vertices $x,y$. \end{theorem} {\bf Proof: } The following proof is due to Anders Yeo (private communication, April 2014). Let $G=(V,E)$ be a graph with two specified vertices $x,y\in V$ and pairs of distinct edges ${\cal P}=\{(e_1,f_1),(e_2,f_2),\ldots{},(e_p,f_p)\}$, all from $E$. We will show how to construct a digraph $D_G$ with specified vertices $s,t$ such that $D_G$ contains an antidirected $(s,t)$-path if and only if $G$ has an $(x,y)$-path which avoids at least one edge from each pair in $\cal P$. Since the construction can be done in polynomial time this and Theorem~\ref{avoidpairs} will imply the result. Let $k$ be the maximum number of pairs in ${\cal P}$ involving the same edge from $E$. Let $P$ be an antidirected path of length $2k+2$ which starts with a forward arc (and hence ends with a backward arc). Now construct $D_G$ as follows: start from $G$ and first replace every edge $uv$ with a private copy $P_{uv}$ of $P$ (no internal vertices are common to two such paths). Then for each pair $(e_i,f_i)\in {\cal P}$ we identify one sink of $P_{e_i}$ with one source of $P_{f_i}$ so that the resulting vertex has in- and out-degree $2$. By the choice of the length of $P$ we can identify in pairs, i.~e.~no three vertices will be identified. Note that all the original vertices of $G$ will be sources in $D_G$. The remaining (new vertices) will be called {\it internal} vertices. Finally let $s=x$ and $t=y$. We claim that $D_G$ has an antidirected $(s,t)$-path if and only if $G$ has an $(x,y)$-path which uses at most one edge from each of the pairs in ${\cal P}$. Suppose first that $xx_1x_2\ldots{}x_{r-1}x_ry$ is a path in $G$ which uses at most one edge from each of the pairs in ${\cal P}$. Then $P_{xx_1}P_{x_1x_2}\ldots{}P_{x_{r-1}x_r}P_{x_ry}$ is an antidirected $(s,t)$-path in $D_G$ (no vertex is repeated since the identifications above where only done for paths corresponding to pairs in ${\cal P}$). Conversely, suppose $D_G$ contains an antidirected $(s,t)$-path $Q$. By the way we identified vertex pairs according to ${\cal P}$, the internal vertices have in- and out-degree at most $2$, and if an internal vertex is on two paths $P_{e_i},P_{f_i}$ then it has both its in-neighbours on $P_{e_i}$ and both its out-neighbours on $P_{f_i}$. This implies that $Q$ will either completely traverse a path $P_{e_i}$ or not touch any internal vertex of that path. Hence it cannot traverse both $P_{e_i}$ and $P_{f_i}$ if $(e_i,f_i)\in {\cal P}$, and it follows that if we delete all internal vertices of $Q$ and add back the edges of $G$ corresponding to each of the traversed paths, we obtain an $(x,y)$-path in $G$ that uses at most one edge from each pair in ${\cal P}$. \hfill$\diamond$\\\vspace{2mm} \section{Properties of antistrong digraphs}\label{ASsec} It follows from our definition that every pair of vertices of an antistrong digraph is joined by a trail of odd length. This immediately gives \begin{lemma} \label{nobipAS} No bipartite digraph is antistrong. \end{lemma} For every digraph $D$ we can associate an undirected bipartite graph which contains all the information we need to study antistrong connectivity. The {\bf bipartite representation}~\cite[Page 19]{bang2009} of a digraph $D=(V,A)$ is the bipartite graph $B(D)=(V'\cup V'',E)$, where $V'=\{v'|v\in V\}$, $V''=\{v''|v\in V\}$ and $E=\{v'w''|vw\in A\}$. \begin{proposition} \label{ASchar} Let $D=(V,A)$ be a digraph with $|V|\ge 3$. The following are equivalent. \begin{enumerate} \item $D$ is antistrong \item $B(D)$ is connected. \item For every choice of distinct vertices $x,y$, the digraph $D$ contains both an antidirected $(x,y)$-trail $T_{x,y}$ of even length starting on a forward arc and an antidirected $(x,y)$-trail $\bar{T}_{x,y}$ of even length starting on a backward arc. \end{enumerate} \end{proposition} {\bf Proof: } Suppose (a) holds. Then, following the edges corresponding to the arcs of a forward antidirected $(x,y)$-trail, $B(D)$ contains an $(x',y'')$-path for every pair of distinct vertices $x,y\in V$. Now, if $x$ and $y$ are distinct vertices of $D$, we choose a third vertex $z$ in $D$ ($z\neq x$ and $z\neq y$), and the union of an $(x',z'')$-path and a $(z'',y')$-path contains an $(x',y')$-path in $B(D)$. Similarly we obtain an $(x'',y'')$-path in $B(D)$ for every pair of distinct vertices $x,y\in V$. Finally, for any $x\in V$, an $(x',x'')$-path in $B(D)$ can be found in the union of an $(x',y')$-path and a $(y',x'')$-path, where $y$ is a vertex of $D$ distinct from $x$. Hence $(a)\Rightarrow{}(b)$ holds. Conversely, $(b)\Rightarrow{}(a)$ holds, since any $(x',y'')$-path in $B(D)$ corresponds to a forward antidirected $(x,y)$-path in $D$ which starts and ends with a forward arc.\\ Now to prove $(b)\Rightarrow{}(c)$, it suffices to remark that $T_{x,y}$ and $\bar{T}_{x,y}$ correspond to an $(x',y')$-path and an $(x'',y'')$-path in $B(D)$, respectively. Finally, to see that $(c)\Rightarrow{}(b)$ holds, it suffices to show that if (c) holds, then $B(D)$ contains an $(x',y'')$-path for all $x,y\in V$ (possibly equal). This follows by considering a neighbour $z''$ of $x'$ and a $(z'',y'')$-path in $B(D)$. \hfill$\diamond$\\\vspace{2mm} Proposition~\ref{ASchar} immediately implies the next result. \begin{corollary} \label{checkAS} One can check in linear time whether a digraph is antistrong. \end{corollary} Recall that a digraph is {\bf $\mathbf{k}$-strong} if it has at least $k+1$ vertices and it remains strong after deletion of any set of at most $k-1$ vertices. The digraph obtained from three disjoint independent sets $X,Y,Z$ each of size $k$ by adding all arcs from $X$ to $Y$, from $Y$ to $Z$, and from $Z$ to $X$ is $k$-strong. However, $B(D)$ has three connected components. This shows that no condition on the strong connectivity will guarantee that a digraph is antistrong.\\ Recall that $D$ is $k$-arc-antistrong if it contains $k$ arc-disjoint forward antidirected $(x,y)$-trails for every ordered pair of distinct vertices $x,y$. We can check in time $O(mk)$ whether a digraph has $k$ arc-disjoint forward antidirected $(x,y)$-trails for given vertices $x,y$, because they correspond to edge-disjoint $(x',y'')$-paths in $B(D)$ whose existence can be checked by using flows, see e.g.~\cite[Section 5.5]{bang2009}. So we can check in polynomial time if a digraph is $k$-arc-antistrong. \begin{theorem} If $D$ is $2k$-arc-antistrong, then it contains $k$ arc-disjoint antistrong spanning subdigraphs. \end{theorem} {\bf Proof: } Since $D$ is $2k$-arc-antistrong, $B(D)$ is $2k$-edge-connected. We can now use Nash-Williams'~theorem (see~\cite[Theorem~9.4.2]{bang2009} for instance) to deduce that $B(D)$ has $k$ edge-disjoint spanning trees. Proposition~\ref{ASchar} now gives the required set of $k$ arc-disjoint antistrong spanning subdigraphs of $D$. \hfill$\diamond$\\\vspace{2mm} \begin{theorem} There exists a polynomial time algorithm which for a given digraph $D$ and a natural number $k$ either returns $k$ arc-disjoint spanning antistrong subdigraphs of $D$ or correctly answers that no such set exists. \end{theorem} {\bf Proof: } This follows from the fact that such subdigraphs exist if and only if $B(D)$ has $k$ edge-disjoint spanning trees, and the existence of such trees can be checked via Edmonds' algorithm for matroid partition~\cite{edmondsJRNBS69}. \hfill$\diamond$\\\vspace{2mm} The corresponding problem for containing two arc-disjoint {\em strong} spanning subdigraphs is NP-complete (see e.g.~\cite[Theorem~13.10.1]{bang2009}). \begin{theorem} \label{twospstrongD} It is NP-complete to decide whether a digraph $D$ contains two spanning strong subdigraphs $D_1,D_2$ which are arc-disjoint. \end{theorem} \section{Antistrong connectivity augmentation}\label{ASaugsec} Note that every complete digraph on at least 3 vertices is antistrong. Hence it is natural to ask for the minimum number of new arcs one has to add to a digraph in order to make it antistrong. \begin{theorem} \label{augmenttoAS} There exists a polynomial time algorithm for finding, for a given digraph $D=(V,A)$ on at least 3 vertices, a minimum cardinality set of new arcs $F$ such that the digraph $D'=(V,A \dot{\cup} F)$ is antistrong. \end{theorem} {\bf Proof: } Let $D$ be a digraph on $n\geq 3$ vertices which is not antistrong. By Proposition~\ref{ASchar}, its bipartite representation $B(D)$ is not connected. First observe that in the bipartite representation each new arc added to $D$ will correspond to an arc from a vertex $u'$ of $V'$ to a vertex $v''\in V''$ such that $u\neq v$ back in $V$. So we are looking for the minimum number of new edges of type $u'v''$ with $u\neq v$ whose addition to $B(D)$ makes it connected while preserving the bipartition $V',V''$. Note that, as long as $n\geq 3$, in which case $B(D)$ has at least 6 vertices, we can always obtain a connected graph by adding edges that are legal according to the definition above. So the number of edges we need is exactly the number of connected components of $B(D)$ minus one.\footnote{This number is also equal to $(2n-1)-r(A)$ where $r$ is the rank function of the matroid $M(D)$ which we define in Section~\ref{matroidsec}.} To find an optimal augmentation we add all missing edges between $V'$ and $V''$ to $B(D)$, except for those of the form $v'v''$ and give the new edges cost 1, while all original edges get cost 0. Now find a minimum weight spanning tree in the resulting weighted complete bipartite graph. The edges of cost 1 correspond to an optimal augmenting set back in $D$. \hfill$\diamond$\\\vspace{2mm} The complexity of the analogous question for $k$-arc-antistrong connectivity is open.\\ \begin{problem} \label{addtogetkAS} Given a digraph $D$ and a natural number $k$, can we find in polynomial time a minimum cardinality set of new arcs whose addition to $D$ results in a digraph $D'$ which is $k$-arc-antistrong? \end{problem} Problem~\ref{addtogetkAS} is easily seen to be equivalent to the following problem on edge-connectivity augmentation of bipartite graphs. \begin{problem} \label{bipaug} Given a natural number $k$ and a bipartite graph $B=(X,Y,E)$ with $|X|=|Y|=p$ which admits a perfect matching $M$ in its bipartite complement, find a minimum cardinality set of new edges $F$ such that $F\cap M=\emptyset$ and $B+F$ is $k$-edge-connected and bipartite with the same bipartition as $B$. \end{problem} Theorem~\ref{augmenttoAS} can be extended to find the minimum number of new arcs whose addition to $D$ gives a digraph with $k$ arc-disjoint antistrong spanning subdigraphs $D_1,\ldots{},D_k$, provided that $V(D)$ is large enough to allow the existence of $k$ such subdigraphs. Note that since each $D_i$ needs at least $2n-1$ arcs and we do not allow parallel arcs, we need $n$ to be large enough, in particular we must have $n\geq 2k+1$. \begin{theorem} \label{addtogetkdisjointAS} There exists a polynomial time algorithm for determining, for a given digraph $D$ on at least 3 vertices, whether one can add some edges to $D$ such that the resulting digraph is simple (no parallel arcs) and has $k$ arc-disjoint antistrong spanning subdigraphs. In the case when such a set exists, the algorithm will return a minimum cardinality set of arcs $A'$ such that $D'=(V,A\cup A')$ contains $k$ arc-disjoint antistrong spanning subdigraphs. \end{theorem} {\bf Proof: } This follows from the fact that the minimum set of new arcs is exactly the minimum number of new edges, not of the form $v'v''$ that we have to add to $B(D)$ such that the resulting bipartite graph is simple and has $k$ edge-disjoint spanning trees. This number can be found using matroid techniques as follows. Add all missing edges from $V'$ to $V''$ and give those of the form $v'v''$ very large cost (larger than $2nk$) and the other new edges cost 1. Now, if the resulting complete bipartite digraph $K_{n,n}$ has $k$-edge-disjoint spanning trees of total cost less than $2nk$, then the set of new edges added will form a minimum augmenting set and otherwise no solution exists. Recall from matroid theory that $k$ edge-disjoint spanning trees in $K_{n,n}$ correspond to $k$ edge-disjoint bases in the cycle matroid $M(K_{n,n})$ of $K_{n,n}$ which again corresponds to an independent set of size $k(2n-1)$ in the union $M=\bigvee_{i=1}^kM(K_{n,n})$. This means that we can solve the problem by finding a minimum cost base $B$ of $M$ and then either return the arcs which correspond to edges of cost 1 in $B$ or decide that no solution exists when the cost of $B$ is more than $2kn$. We leave the details to the reader.\hfill$\diamond$\\\vspace{2mm} \section{A matroid for antistrong connectivity}\label{matroidsec} Having seen the equivalence between antistrong connectivity of digraph $D$ on $n$ vertices and connectivity of its bipartite representation $B(D)$ (see Proposition~\ref{ASchar}), and recalling from matroid theory that $B(D)$ is connected if and only if the cycle matroid $M(B(D))$ has rank $|V(B(D))|-1$, it is natural to ask how antistrong connectivity can be expressed as a matroid property on $D$ itself. For $F\subseteq A$, we denote by $h(F)$ and $t(F)$ the numbers of vertices that are heads, respectively tails, of one or more arcs in $F$. Recall that the independent sets of the cycle matroid $M(G)$ of a graph $G=(V,E)$ are those subsets $I\subseteq E$ for which we have $|I'| \leq \nu(I')-1$ for all $\emptyset\neq I'\subseteq I$, where $\nu(I')$ is the number of end vertices of the edges in $I'$. Inspired by this we define set $I$ of arcs in a digraph $D=(V,A)$ to be {\bf independent} if \begin{equation} \label{goodset} |I'| \leq h(I')+t(I')-1 \hspace{3mm}\mbox{for all }\emptyset \neq I'\subseteq I, \end{equation} A set $S\subseteq A$ is {\bf dependent} if it is not independent. \begin{proposition} \label{goodisforest} Let $D=(V,A)$ be a digraph. A subset $I\subseteq A$ is independent if and only if the corresponding edge set $I$ in $B(D)$ forms a forest. Every inclusion-minimal dependent set $S\subseteq A$ corresponds to a cycle in $B(D)$ and conversely. \end{proposition} {\bf Proof: } Suppose $I\subseteq A$ is independent and consider the corresponding edge set $\tilde I$ in $B(D)$. If $\tilde I$ is not a forest, then some subset $\tilde I'\subseteq \tilde I$ will be a cycle $C$ in $B(D)$ with $p$ vertices in each of $V',V''$ for some $p\geq 2$. The set $\tilde I'$ corresponds to a set $I'\subseteq I$ with $h(I')+t(I')-1=p+p-1<2p=|I'|$, contradicting that $I$ is independent. The other direction follows from the fact that every forest $F$ in $B(D)$ spans at least $|E(F)|+1$ vertices in $B(D)$ and every subset of a forest is again a forest. The last claim follows from the fact that every minimal set of edges which does not form a forest in $B(D)$ forms a cycle in $B(D)$. \hfill$\diamond$\\\vspace{2mm} The previous proposition implies that a set of arcs of a digraph is dependent if and only if it contains a closed trail of even length consisting of alternating forward and backward arcs. We will refer to such a trail as a {\bf closed antidirected trail}, or {\bf CAT} for short. \begin{theorem}\label{antimatroid} Let $D=(V,A)$ be a digraph and $\cal I$ be the family of all independent sets of arcs in $D$. Then $M(D)=(A,{\cal I})$ is a graphic matroid with rank equal to the size of a largest collection of arcs containing no closed alternating trail. \end{theorem} {\bf Proof: } It follows immediately from Proposition~\ref{goodisforest} that a set $I$ belongs to ${\cal I}$ if and only if the corresponding edge set $\tilde I$ is independent in the cycle matroid on $B(D)$. \hfill$\diamond$\\\vspace{2mm} \begin{theorem} A digraph $D$ is antistrong if and only if $M(B(D))$ has rank $2|V|-1$. \end{theorem} {\bf Proof: } The rank of $M(B(D))$ equals the size of a largest acyclic set of edges in $B(D)$. This has size $2|V|-1$ precisely when $B(D)$ has a spanning tree $H$. Back in $D$, the arcs corresponding to $E(H)$ contain antidirected forward trails between any pair of distinct vertices. \hfill$\diamond$\\\vspace{2mm} \section{Antistrong orientations of graphs}\label{ASorsec} Recall that, by Robbins' theorem (see e.g.~\cite[Theorem 1.6.1]{bang2009}) a graph $G$ has a strongly connected orientation if and only if $G$ is 2-edge-connected. For antistrong orientations we have the following consequence of Proposition~\ref{ASchar} which implies that there is no lower bound to the (edge-) connectivity which guarantees an antistrong orientation of a graph. \begin{proposition} \label{nobipOK} No bipartite graph can be oriented as an antistrong digraph. \end{proposition} The purpose of this section is to characterize graphs which can be oriented as antistrong digraphs. \begin{theorem}\label{orient} Suppose $G=(V,E)$ and $|E|=2|V|-1$. Then $G$ has an antistrong orientation if and only if \begin{eqnarray} |E(H)|&\leq& 2|V(H)|-1 \mbox{ for all nonempty subgraphs $H$ of $G$, and}\label{condition}\\ |E(H)|&\leq& 2|V(H)|-2 \mbox{ for all nonempty bipartite subgraphs $H$ of $G$.}\label{bipcond}\\ \nonumber \end{eqnarray} \end{theorem} We derive Theorem~\ref{orient} from the following characterization of graphs which can be oriented as digraphs with no closed antidirected trail (CAT). \begin{theorem} \label{CATfree} A graph $G=(V,E)$ has an orientation with no CAT if and only if $G$ satisfies (\ref{condition}) and (\ref{bipcond}). In particular no $n$ vertex graph with at least $2n$ edges and no $n$ vertex bipartite graph with at least $2n-1$ edges admits a CAT-free orientation. \end{theorem} It is not hard to see that Theorem~\ref{CATfree} implies Theorem~\ref{orient}. Assume that Theorem~\ref{CATfree} holds and consider a graph $G=(V,E)$ with $|E|=2|V|-1$. Suppose that $G$ has an antistrong orientation $D$. Then $B(D)$ is connected by Proposition~\ref{ASchar}. As $B(D)$ has $2|V|-1=|V(B(D))|-1$ edges it is a tree. So $D$ is a CAT-free orientation of $G$ and, by Theorem~\ref{CATfree}, conditions (\ref{condition}) and (\ref{bipcond}) hold for $G$. Conversely, if (\ref{condition}) and (\ref{bipcond}) hold for $G$, then $G$ has a CAT-free orientation by Theorem~\ref{CATfree}, and we can deduce as above that this orientation is also an antistrong orientation of $G$.\\ We next show that (\ref{condition}) and (\ref{bipcond}) are necessary conditions for a CAT-free orientation. For the necessity of (\ref{condition}), suppose that some nonempty subgraph $H$ has $|E(H)|\geq 2|V(H)|$ and that $D$ is any orientation of $G$. Then $B(D)$ has at least $2|V(H)|$ edges between $V(H)'$ and $V(H)''$, implying that it contains a cycle. Hence $D$ is not CAT-free. The necessity of (\ref{bipcond}) can be seen as follows. Suppose $H$ is a bipartite subgraph on $2|V(H)|-1$ edges and let $\vec{H}$ be an arbitrary orientation of $H$. Since no bipartite graph has an antistrong orientation it follows that $B(\vec{H})$ is not connected, and, as it has $2|V(H)|-1=|V(B(\vec{H}))|-1$ edges, it contains a cycle. This corresponds to a CAT in $\vec{H}$.\\ Most of the remainder of this section is devoted to a proof of sufficiency in Theorem~\ref{CATfree}. We first show that, for an arbitrary graph $G'=(V',E')$, the edge sets of all subgraphs $G$ of $G'$ which satisfy (\ref{condition}) and (\ref{bipcond}) are the independent sets of a matroid on $E'$. We then show that this matroid is the matroid union of the cycle matroid and the `even bicircular matroid' of $G'$ (defined below). This allows us to partition the edge-set of a graph $G$ which satisfies (\ref{condition}) and (\ref{bipcond}) into a forest and an `odd pseudoforest'. We then use this partition to define a CAT-free orientation of $G$. We first recall some results from matroid theory. We refer a reader unfamiliar with submodular functions and matroids to~\cite{frank2011}. Suppose $E$ is a set and $f:2^E\to {\mathbb{Z}}$ is a submodular, nondecreasing set function which is nonnegative on $2^E \setminus \{\emptyset\}$. Edmonds~\cite{E1970}, see~\cite[Theorem 13.4.2]{frank2011}, showed that $f$ induces a matroid $M_f$ on $E$ in which $S\subseteq E$ is independent if $|S'|\leq f(S')$ for all $\emptyset\neq S'\subseteq S$. The rank of a subset $S\subseteq E$ in $M_f$ is given by the min-max formula \begin{equation}\label{eq:minmax} r_f(S)=\min_{\mathcal{P}}\left\{\left|S\setminus\bigcup_{T\in {\mathcal{P}}}T\right|+\sum_{T\in{\mathcal{P}}}f(T)\right\}, \end{equation} where the minimum is taken over all subpartitions ${\mathcal{P}}$ of $S$ (where a {\bf subpartition} of $S$ is a collection of pairwise disjoint nonempty subsets of $S$). Note that the matroid $M(D)$ defined in the previous section is induced on the arc-set of the digraph $D$ by the set function $h+t-1$. Given a graph $G=(V,E)$ and $S\subseteq E$ let $G[S]$ be the {\bf subgraph induced} by $S$ i.e. the subgraph of $G$ with edge-set $S$ and vertex-set all vertices incident to $S$. Let $\nu,\beta:2^E\to {\mathbb{Z}}$ by putting $\nu(S)$ equal to the number of vertices incident to $S$, and $\beta(S)$ equal to the number of bipartite components of $G[S]$. It is well known that $\nu$ is submodular, nondecreasing, and nonnegative on $2^E$ and that $M_{\nu-1}(G)$ is the cycle matroid of $G$. The function $\nu-\beta$ is also submodular, nondecreasing, and nonnegative on $2^E $ since it is the rank function of the matroid on $E$ whose independent sets are the edge sets of the {\bf odd pseudoforests} of $G$, i.~e.~subgraphs in which each connected component contains at most one cycle, and if such a cycle exists then it is odd, see \cite[Corollary 7D.3]{Z}. We will refer to this matroid as the {\bf even bicircular matroid} of $G$. The above mentioned properties of $\nu$ and $\nu-\beta$ imply that $2\nu-1-\beta$ is submodular, nondecreasing, and nonnegative on $2^E \setminus \{\emptyset\}$. We will show that the independent sets in $M_{2\nu-1-\beta}(G)$ are the edge sets of the subgraphs which satisfy (\ref{condition}) and (\ref{bipcond}). \begin{lemma}\label{lem:mat1} Let $G=(V,E)$ be a graph and ${\mathcal{I}}=\{I\subseteq E\,:\mbox{ $G[I]$ satisfies (\ref{condition}) and (\ref{bipcond})}\}$. Then ${\mathcal{I}}$ is the family of independent sets of the matroid $M_{2\nu-1-\beta}(G)$. In addition, the rank of a subset $S\subseteq E$ in this matroid is $r_{2\nu-1-\beta}(S)=\min_{\mathcal{P}}\left\{|S\setminus\bigcup_{T\in {\mathcal{P}}}T|+\sum_{T\in {\mathcal{P}}}\left(2\nu(T)-1-\beta(T)\right)\right\}$ where the minimum is taken over all subpartitions ${\mathcal{P}}$ of $S$. \end{lemma} \noindent {\bf Proof. } We first suppose that some $S\subseteq E$ is not independent in $M_{2\nu-1-\beta}(G)$. Then we may choose a nonempty $S'\subseteq S$ with $|S'|>2\nu(S')-1-\beta(S')$, and subject to this condition, such that $|S'|$ is as small as possible. The minimality of $S'$ implies that $H=G[S']$ is connected. So $\beta(S')=1$ if and only if $H$ is bipartite (and 0 otherwise) and we may now deduce that that $H\subseteq G[S]$ fails to satisfy (\ref{condition}) or (\ref{bipcond}). We next suppose that $G[S]$ fails to satisfy (\ref{condition}) or (\ref{bipcond}) for some $S\subseteq E$. Then there exists a nonempty subgraph $H$ of $G[S]$ such that either $|E(H)|>2|V(H)|-1$, or $H$ is bipartite and $|E(H)|>2|V(H)|-2$. Then $S'=E(H)$ satisfies $|S'|>2\nu(S')-1-\beta(S')$ so $S$ is not independent in $M_{2\nu-1-\beta}(G)$. The expression for the rank function of $M_{2\nu-1-\beta}(G)$ follows immediately from (\ref{eq:minmax}). \hfill$\diamond$\\\vspace{2mm} The {\bf matroid union} of two matroids $M_1=(E,{\mathcal{I}}_1)$ and $M_2=(E,{\mathcal{I}}_2)$ on the same ground set $E$ is the matroid $M_1\vee M_2=(E,{\mathcal{I}})$ where ${\mathcal{I}}=\{I_1\cup I_2\,:\,I_1\in {\mathcal{I}}_1\mbox{ and } I_1\in {\mathcal{I}}_1\}$. Suppose $f_1,f_2:E\to {\mathbb{Z}}$ are submodular, nondecreasing, and nonnegative on $2^E \setminus \{\emptyset\}$. Then $f_1+f_2$ will also be submodular, nondecreasing, and nonnegative on $2^E \setminus \{\emptyset\}$ and hence will induce the matroid $M_{f_1+f_2}$. Every independent set in $M_{f_1}\vee M_{f_2}$ is independent in $M_{f_1+f_2}$, but the converse does not hold in general. Katoh and Tanigawa~\cite[Lemma 2.2]{KT} have shown that the equality $M_{f_1+f_2}=M_{f_1}\vee M_{f_2}$ does hold whenever the minimum in formula (\ref{eq:minmax}) for the ranks $r_{f_1}(S)$ and $r_{f_2}(S)$ is attained for the same subpartition of $S$, for all $S\subset E$. This allows us to deduce \begin{lemma} \label{lem:mat2} For any graph $G=(V,E)$, we have $M_{2\nu-1-\beta}(G)=M_{\nu-1}(G)\vee M_{\nu-\beta}(G)$. \end{lemma} \noindent {\bf Proof. } This follows from the above mentioned result of Katoh and Tanigawa, and the facts that $r_{\nu-1}(S)=\sum_{T\in{\mathcal{P}}}(\nu(T)-1)$ and $r_{\nu-\beta}(S)=\sum_{T\in{\mathcal{P}}}(\nu(T)-\beta(T))$ where ${\mathcal{P}}$ is the partition of $S$ given by the connected components of $G[S]$ (since $r_{\nu-1}(S)$ and $r_{\nu-\beta}(S)$ are equal to the number of edges in a maximum forest and a maximum odd pseudoforest, respectively, in $G[S]$). \hfill$\diamond$\\\vspace{2mm} Lemma~\ref{lem:mat1} and Lemma~\ref{lem:mat2} immediately give the following. \begin{lemma}\label{lem:mat3} Let $G=(V,E)$ be a graph. Then $G$ satisfies (\ref{condition}) and (\ref{bipcond}) if and only if $E$ can be partitioned into a forest and an odd pseudoforest. \end{lemma} \noindent We provide an alternative graph theoretic proof of this lemma in the appendix. We next show that every graph whose edge set can can be partitioned into a spanning tree and an odd pseudoforest has a CAT-free orientation.\\ \begin{figure}[h] \begin{center} \includegraphics[height=85mm]{asillu2.pdf}% \end{center} \caption{A CAT-free orientation of the union of a spanning tree $T$ and a spanning odd pseudoforest $P$; $T$ governs the bipartition $X,Y$ (white/grey), its edges are drawn outside (or on) the disk spanned by the vertices. The edges of $P$ are embedded in the interior of that disk, the root vertex is the encircled topmost one, the precious edge is the dashed one.} \label{twotreesfig} \end{figure} \begin{theorem} \label{tree+forest} Let $G$ be the edge-disjoint union of a spanning tree $T$ and an odd pseudoforest $P$.\footnote{Note that $G$ may have parallel edges, but no more than two copies of any edge, in which case one copy is in $T$ and the other in $P$} Then $G$ has a CAT-free orientation. In addition, such an orientation can be constructed in linear time given $T$ and $P$. \end{theorem} {\bf Proof: } Let $X,Y$ be the unique (up to renaming the two sets) bipartition of $T$ and orient all edges of $T$ from $Y$ to $X$. If $P$ has no edges we are done since there are no cycles in $G$. Let $P_1,\ldots{},P_k$ be the connected components of $P$. We shall show that we can orient the edges of $P_1,\ldots{},P_k$ in such a way that none of the resulting arcs of these (now oriented) pseudoforests $\vec{P_1},\ldots{},\vec{P_k}$ can belong to a closed antidirected trail. Clearly this will imply the lemma. For each $P_i$ we choose a root vertex $r_i$ of $P_i$ as follows. If $P_i$ is a tree then we choose $r_i$ to be an arbitrary vertex of $P_i$. If $P_i$ contains an odd cycle $C_i$ then we choose $r_i$ to be a vertex of $C_i$ such that $r_i$ has as many neighbours on $C_i$ as possible in the same set of the bipartition $(X,Y)$ as $r_i$. Since $C_i$ is odd we may choose one such neighbour $s_i$ of $r_i$. We will refer to the edge $r_is_i$ as a {\bf precious edge} of $P_i$. Put $T_i=P_i-r_is_i$ if $P_i$ contains a cycle and otherwise put $T_i=P_i$. We orient the edges of $T_i$ as follows. Every edge of $P_i$ with one end in $X$ and the other in $Y$ is oriented from $X$ to $Y$. Every edge $uv$ of $T_i$ with $u,v\in X$ is oriented towards $r_i$ in $T_i$ (so if $v$ is closer to $r_i$ than $u$ in $T_i$ we orient the edge from $u$ to $v$ and otherwise we orient it from $v$ to $u$, see Figure~\ref{twotreesfig}). Every edge $pq$ of $T_i$ with $p,q\in Y$ is oriented away from $r_i$ in $T_i$. Finally, if $P_i$ contains a precious edge $r_is_i$, then we orient $r_is_i$ from $r_i$ to $s_i$ if $r_i,s_i\in X$, and from $s_i$ to $r_i$ if $r_i,s_i\in Y$. Let $D$ denote the resulting orientation of $G$. The digraph $D$ can be constructed in linear time if we traverse each tree $T_i$ by a breath first search rooted at $r_i$. We use induction on $|E(P)|$ to show that the above construction results in a CAT-free digraph. As noted above, this is true for the base case when $E(P)=\emptyset$. Suppose that $E(P)\neq \emptyset$ and choose an edge $uv$ in some $P_i$ according to the following criteria. If $P_i$ is not a cycle then choose $v$ to be a vertex of degree one in $P_i$ distinct from $r_i$ and $u$ to be the neighbour of $v$ in $P_i$. If $P_i$ is an odd cycle then choose $v=r_i$ and $u=s_i$. We will show that $uv$ belongs to no CAT in $D$. By symmetry, we may suppose that $v\in X$. We first consider the case when $v$ is a vertex of degree one in $P_i$. Below $d^+(v),d^-(v)$ denote the out-degree, respectively, the in-degree of the vertex $v$ in $D$. We have two possible subcases: \begin{itemize} \item $u\in Y$. Since $v\in X$, we oriented $uv$ from $v$ to $u$. All the other edges incident to $v$ belong to $T$ and were oriented towards $v$. Then $d^+(v)=1$ and the arc $vu$ cannot be part of a CAT. \item $u\in X$. Since $v\in X$, we oriented $uv$ from $v$ to $u$ (as $u$ is closer to $r_i$ than $v$ in $T_i$). As previously we have $d^+(v)=1$ and the arc $vu$ cannot be part of a CAT. \end{itemize} Since $D-uv$ is CAT-free by induction, $D$ is also CAT-free. We next consider the case when $P_i$ is an odd circuit. Let $t_i$ be the neighbour of $r_i$ in $P_i$ distinct from $s_i$. We again have two possible subcases: \begin{itemize} \item $t_i\in X$. Since $r_i\in X$, we oriented the edge $t_ir_i$ from $t_i$ to $r_i$. Then $d^+(r_i)=1$, and the arc $r_is_i$ cannot be part of a CAT. Since $D-r_is_i$ is CAT-free by induction, $D$ is also CAT-free. \item $t_i\in Y$. Let $q_i$ be the neighbour of $s_i$ in $P_i$ which is distinct from $r_i$. The choice of $r_i$ implies that $q_i\in Y$, and hence that $s_iq_i$ is oriented from $s_i$ to $q_i$. Then $d^+(s_i)=1$, and the arc $s_iq_i$ cannot be part of a CAT. Since $D-s_iq_i$ can be obtained by applying our construction to $T\cup (P-s_iq_i)$ by choosing $s_i$ as the root vertex in $P_i-s_iq_i$ rather than $r_i$, it is CAT-free by induction. Hence $D$ is also CAT-free. \end{itemize} \hfill$\diamond$\\\vspace{2mm} \noindent{}{\bf Proof of Theorem~\ref{CATfree} (sufficiency):} Let $G=(V,E)$ be a graph satisfying (\ref{condition}) and (\ref{bipcond}). By Lemma~\ref{lem:mat3}, $E$ can be partitioned into a forest $F$ and an odd pseudoforest $P$. By adding a suitable set of edges to $G$, we may assume that $|E|=2|V|-1$. (This follows by considering the matroid $M_{2\nu-1-\beta}(2K_n)$ on the edge set of the graph $2K_n$ with vertex set $V$ in which all pairs of vertices are joined by two parallel edges. It is easy to check that $2K_n$ has an edge-disjoint forest and odd pseudoforest with a total of $2|V|-1$ edges. Thus the rank of $M_{2\nu-1-\beta}(2K_n)$ is $2|V|-1$. Since $E$ is an independent set in $M_{2\nu-1-\beta}(2K_n)$, it can be extended to an independent set with $2|V|-1$ edges.) The fact that $|E|=2|V|-1$ implies that $F$ is a spanning tree of $G$. We can now apply Theorem~\ref{tree+forest} to deduce that $G$ has a CAT-free orientation. \hfill$\diamond$\\\vspace{2mm} We have seen that Theorem~\ref{CATfree} implies Theorem~\ref{orient}, and hence that a graph $G=(V,E)$ has an antistrong orientation if and only if the rank of $M_{2\nu-1-\beta}(G)$ is equal to $2|V|-1$. We can now apply the rank formula (\ref{eq:minmax}) to characterize graphs which admit an antistrong orientation. \begin{theorem}\label{thm:char} A graph $G=(V,E)$ has an antistrong orientation if and only if \begin{equation}\label{ASorchar} e({\mathcal{Q}})\geq |{\mathcal{Q}}|-1+b({\mathcal{Q}}) \end{equation} for all partitions ${\mathcal{Q}}$ of $V$, where $e({\mathcal{Q}})$ denotes the number of edges of $G$ between the different parts of ${\mathcal{Q}}$ and $b({\mathcal{Q}})$ the number of parts of ${\mathcal{Q}}$ which induce bipartite subgraphs of $G$. \end{theorem} \noindent {\bf Proof: } Suppose that $G$ has no antistrong orientation. Then the rank of $M_{2\nu-1-\beta}(G)$ is less than $2|V|-1$ so there exists a subpartition ${\mathcal{P}}$ of $E$ such that \begin{equation}\label{ASorchar1} \alpha({\mathcal{P}}):=\left|E\setminus\bigcup_{T\in {\mathcal{P}}}T\right|+\sum_{T\in{\mathcal{P}}}\left(2\nu(T)-1-\beta(T)\right)< 2|V|-1 \end{equation} by Lemma \ref{lem:mat1}. We may assume that ${\mathcal{P}}$ has been chosen such that: \begin{enumerate} \item[(i)] $\alpha({\mathcal{P}})$ is as small as possible; \item[(ii)] subject to (i), $|{\mathcal{P}}|$ is as small as possible; \item[(iii)] subject to (i) and (ii), $|\bigcup_{T\in {\mathcal{P}}}T|$ is as large as possible. \end{enumerate} Let ${\mathcal{P}}=\{E_1,E_2,\ldots,E_t\}$ and let $H_i=(V_i,E_i)$ be the subgraph of $G$ induced by $E_i$ for all $1\leq i\leq t$. We will show that $H_i$ is a (vertex-)induced connected subgraph of $G$ and that $V_i\cap V_j=\emptyset$ for all $i\neq j$. First, suppose that $H_i$ is disconnected for some $1\leq i\leq t$. Then we have $H_i=H_i'\cup H_i''$ for some subgraphs $H_i'=(V_i',E_i')$ and $H_i''=(V_i'',E_i'')$ with $V_i'\cap V_i''=\emptyset$. Let ${\mathcal{P}}'=({\mathcal{P}}\setminus\{E_i\})\cup\{E_i',E_i''\}$. We have $$2\nu(E_i)-1-\beta(E_i)>2\nu(E_i')-1-\beta(E_i')+2\nu(E_i'')-1-\beta(E_i'')$$ since $\nu(E_i)=\nu(E_i')+\nu(E_i'')$ and $\beta(E_i)= \beta(E_i')+\beta(E_i'')$. This implies that $\alpha({\mathcal{P}}')<\alpha({\mathcal{P}})$ and contradicts (i). Hence $H_i$ is connected and $\beta(E_i)\in \{0,1\}$ for all $1\leq i \leq t$. Next, suppose that $V_i\cap V_j\neq \emptyset$ for some $1\leq i<j\leq t$. Let ${\mathcal{P}}'=({\mathcal{P}}\setminus\{E_i,E_j\})\cup\{E_i\cup E_j\}$. We have $$2\nu(E_i)-1-\beta(E_i)+2\nu(E_j)-1-\beta(E_j)\geq 2\nu(E_i\cup E_j)-1-\beta(E_i\cup E_j)$$ since, if $|V_i\cap V_j|=1$, then $\nu(E_i)+\nu(E_j)= \nu(E_i\cup E_j)+1$ and $\beta(E_i)+\beta(E_j)\leq \beta(E_i\cup E_j)+1$, and, if $|V_i\cap V_j|\ge 2$, then $\nu(E_i)+\nu(E_j)\geq \nu(E_i\cup E_j)+2$ and $\beta(E_i)+\beta(E_j)\leq 2$. This implies that $\alpha({\mathcal{P}}')\leq\alpha({\mathcal{P}})$. Since $|{\mathcal{P}}'|<|{\mathcal{P}}|$ this contradicts (i) or (ii). Hence $V_i\cap V_j=\emptyset$ for all $i\neq j$. Finally, suppose that $H_i\neq G[V_i]$. Then some $e\in E\setminus\bigcup_{T\in {\mathcal{P}}}T$ has both end vertices in $E_i$. Let $E_i'=E_i+e$ and ${\mathcal{P}}'={\mathcal{P}} - E_i +E_i'$. This implies that $\alpha({\mathcal{P}}')\leq\alpha({\mathcal{P}})$. Since $|{\mathcal{P}}'|=|{\mathcal{P}}|$ and $|\bigcup_{T\in {\mathcal{P}}'}T|>|\bigcup_{T\in {\mathcal{P}}}T|$, this contradicts (i) or (iii). Hence $H_i=G[V_i]$. Let ${\mathcal{Q}}$ be the partition of $V$ obtained from $\{V_1,V_2,\ldots,V_t\}$ by adding the remaining vertices of $G$ as singletons. Then $\left|E\setminus\bigcup_{T\in {\mathcal{P}}}T\right|=e({\mathcal{Q}})$ and $\sum_{T\in{\mathcal{P}}}\left(2\nu(T)-1-\beta(T)\right)=2|V|-|{\mathcal{Q}}|-b({\mathcal{Q}})$. We can now use (\ref{ASorchar1}) to deduce that $e({\mathcal{Q}})< |{\mathcal{Q}}|-1+b({\mathcal{Q}})$. \smallskip For the converse, suppose that $e({\mathcal{Q}})< |{\mathcal{Q}}|-1+b({\mathcal{Q}})$ for some partition ${\mathcal{Q}}=\{V_1,V_2,\ldots,V_s\}$ of $V$. We may assume that $G[V_i]$ is connected for all $1\leq i \leq s$ since we can replace $V_i$ by the vertex sets of the components of $G[V_i]$ in ${\mathcal{Q}}$ and maintain this inequality. Let $G[V_i]=(V_i,E_i)$ for $1\leq i\leq s$ and ${\mathcal{P}}=\{E_i\,:\,E_i\neq \emptyset,\,1\leq i\leq s\}$. Then $\left|E\setminus\bigcup_{T\in {\mathcal{P}}}T\right|=e({\mathcal{Q}})$ and $\sum_{T\in{\mathcal{P}}}\left(2\nu(T)-1-\beta(T)\right)=2|V|-|{\mathcal{Q}}|-b({\mathcal{Q}})$. (Note that a set $V_i\in {\mathcal{Q}}$ with $|V_i|=1$ has no corresponding edge set in ${\mathcal{P}}$ and contributes $2-1-1$ to the right hand side of the last equation.) This implies that $$\left|E\setminus\bigcup_{T\in {\mathcal{P}}}T\right|+\sum_{T\in{\mathcal{P}}}\left(2\nu(T)-1-\beta(T)\right)= e({\mathcal{Q}})+\sum_{T\in{\mathcal{P}}}\left(2\nu(T)-1-\beta(T)\right)< 2|V|-1$$ and hence $G$ has no antistrong orientation. \hfill$\diamond$\\\vspace{2mm} \\[2mm] \iffalse OR \\[2mm] Suppose, on the other hand, that $G$ has an antistrong orientation. Then $G$ contains an edge-disjoint spanning tree $T$ and odd pseudoforest $P$ with $|E(P)|=|V(G)$|. Let ${\mathcal{Q}}=\{V_1,V_2,\ldots,V_s\}$ be a partition of $V$ and let $H$ be the graph obtained by contracting each set $V_i$ to a single vertex $v_i$ and then adding a loop at $v_i$ whenever $G[V_i]$ is not bipartite. Then there must be at least $|{\mathcal{Q}}|-1$ edges of $T$ and at least $b({\mathcal{Q}})$ edges of $P$ joining the vertices of $H$. Hence $e({\mathcal{Q}})\geq |{\mathcal{Q}}|-1+b({\mathcal{Q}})$. \fi \begin{corollary} Every $4$-edge-connected nonbipartite graph has an antistrong orientation. \end{corollary} {\bf Proof: } Suppose $G=(V,E)$ is $4$-edge-connected and not bipartite and let ${\mathcal{Q}}$ be a partition of $V$. If ${\mathcal{Q}}=\{V\}$ then $e({\mathcal{Q}})=0=|{\mathcal{Q}}|-1+b({\mathcal{Q}})$ since $G$ is not bipartite, and if ${\mathcal{Q}}\neq \{V\}$ then $e({\mathcal{Q}})\geq 2|{\mathcal{Q}}|\geq |{\mathcal{Q}}|-1+b({\mathcal{Q}})$ since $G$ is $4$-edge-connected. Hence $G$ has an antistrong orientation by Theorem~\ref{thm:char}. \hfill$\diamond$\\\vspace{2mm} \begin{corollary} \label{cor:3trees-antistrong} Every nonbipartite graph with three edge-disjoint spanning trees has an antistrong orientation. \end{corollary} {\bf Proof: } We give two proofs of this corollary. Suppose $G=(V,E)$ is a nonbipartite graph with three edge-disjoint spanning trees and let ${\mathcal{Q}}$ be a partition of $V$. If ${\mathcal{Q}}=\{V\}$ then $e({\mathcal{Q}})=0=|{\mathcal{Q}}|-1+b({\mathcal{Q}})$ since $G$ is not bipartite, and if ${\mathcal{Q}}\neq \{V\}$ then $e({\mathcal{Q}})\geq 3(|{\mathcal{Q}}|-1)$ since $G$ has three edge-disjoint spanning trees. Since $|{\mathcal{Q}}|\ge 2$, $2(|{\mathcal{Q}}|-1)\ge |{\mathcal{Q}}| \ge b({\mathcal{Q}})$ and $e({\mathcal{Q}})\geq |{\mathcal{Q}}|-1+b({\mathcal{Q}})$. Hence $G$ has an antistrong orientation by Theorem~\ref{thm:char}. We could also remark that if $T_1$, $T_2$ and $T_3$ denote three edge-disjoint spanning trees of $G$, then there exists $e\in G$ such that $T_1+e$ is not bipartite. Then depending if $e\in T_2$ or not, $\{T_1+e,T_3\}$ or $\{T_1+e,T_2\}$ is an edge-disjoint pair of a spanning odd pseudo-tree and a spanning tree of $G$. Let $H$ denote this subgraph of $G$. Then using Theorem~\ref{tree+forest}, $H$ has a CAT-free orientation which is also an antistrong orientation of $H$ since $|E(H)|=2|V(H)|-1$. So $G$ has also an antistrong orientation. \hfill$\diamond$\\\vspace{2mm} Corollary \ref{cor:3trees-antistrong} is tight in the sense that there exist graphs with many edge-disjoint trees, two spanning and the others missing just three vertices, which have no antistrong orientation. Consider the graph $G$ obtained by identifying one vertex of a complete bipartite graph $K_{k,k}$ and a complete graph $K_4$. Then $G$ has no antistrong orientation. Indeed, consider the partition ${\mathcal{Q}}$ of $V(G)$ into four parts: the copy of $K_{k,k}$, and one part for each remaining vertex of $K_4$. We have $e({\mathcal{Q}})=6 < |{\mathcal{Q}}|-1+b({\mathcal{Q}})= 4-1+4$ and then $G$ has no antistrong orientation by Theorem~\ref{thm:char}. \medskip Since $M_{2\nu-1-\beta}(G)=M_{\nu-1}(G)\vee M_{\nu-\beta}(G)$, we can use Edmonds' matroid partition algorithm~\cite{edmondsJRNBS69} to determine the rank of $M_{2\nu-1-\beta}(G)$ in polynomial time, and hence determine whether $G$ has an antistrong orientation. Moreover, when such an orientation exists, we can use Edmonds' algorithm to construct an edge-disjoint spanning tree and pseudoforest with a total of $2|V|-1$ edges, and then use the construction from the proof of Theorem~\ref{tree+forest} to obtain the desired antistrong orientation in polynomial time. This gives \begin{corollary} There exists a polynomial time algorithm which finds, for a given input graph $G$, either an antistrong orientation $D$ of $G$, or a certificate, in terms of a subpartition $\cal P$ which violates (\ref{ASorchar}), that $G$ has no such orientation. \end{corollary} \section{Connected bipartite 2-detachments of graphs}\label{detachsec} We now show a connection between antistrong orientations of a graph $G$ and so-called detachments of $G$. We need only the special case of 2-detachments (see e.g.~\cite{nashwilliamsJLMS31} for results on detachments). A {\bf 2-detachment} of a graph $G=(V,E)$ is any graph $H=(V'\cup V'',E')$ which can be obtained from $G$ by replacing every vertex $v\in V$ with two new vertices $v',v''$ and then for each original edge $uv$ adding precisely one of the four edges $u'v',u'v'',u''v',u''v''$ to $E'$. \begin{lemma}\label{detach} A graph $G=(V,E)$ has an antistrong orientation if and only if $G$ has a 2-detachment $H=(V'\cup V'',E')$ which is connected and bipartite with bipartition $V',V''$ (we call such a 2-detachment {\bf good}). \end{lemma} {\bf Proof: } Suppose $G$ has a good 2-detachment $H=(V'\cup V'',E')$. Then there are no edges of the form $u'v'$ and no edges of the form $u''v''$. Hence the orientation $D$ that we get by orienting the edges of the form $u'v''$ from $u$ to $v$ will be an antistrong orientation of $G$ by Proposition~\ref{ASchar}. Conversely, if $D$ is an antistrong orientation of $G$, then $B(D)$ is a good 2-detachment of $G$. \hfill$\diamond$\\\vspace{2mm} We can now use Theorem~\ref{thm:char} and the subsequent remark to deduce the following. \begin{theorem} \label{good2detach} A graph $G=(V,E)$ has a good 2-detachment if and only if \begin{equation} \label{detachcond} e({\mathcal{Q}})\geq |{\mathcal{Q}}|-1+b({\mathcal{Q}}) \end{equation} for all partitions ${\mathcal{Q}}$ of $V$. Furthermore, there exists a polynomial time algorithm which returns such a $2$-detachment when it exists and otherwise returns a certificate, in terms of a partition violating (\ref{detachcond}), that no such detachment exists. \end{theorem} \section{Non-separating antistrong spanning subdigraphs}\label{nosepsec} While there are polynomial time algorithms for checking the existence of two edge-disjoint spanning trees~\cite{edmondsJRNBS69}, or two arc-disjoint branchings (spanning out-trees) in a digraph (see e.g.~\cite[Corollary 9.3.2]{bang2009}), checking whether we can delete a strong spanning subdigraph and still have a connected digraph is difficult. Let $U\!G(D)$ denote the underlying undirected graph of a digraph $D$. \begin{theorem}\cite{bangTCS438} \label{nondisconstrong} It is NP-complete to decide whether a given digraph $D$ contains a spanning strong subdigraph $H$ such that $U\!G(D-A(H))$ is connected. \end{theorem} If we replace ``strong'' by ``antistrong'' above, the problem becomes solvable in polynomial time. \begin{theorem} \label{deleteanticon} We can decide in polynomial time for a given digraph $D=(V,A)$ on $n$ vertices whether $D$ contains a spanning antistrong subdigraph $H=(V,A')$ such that $U\!G(D-A')$ is connected. \end{theorem} {\bf Proof: } We may assume that $D$ is antistrong, since this can be checked in linear time by verifying that $B(D)$ is connected. Let $M_1=(A,{\cal I})$ be the cycle matroid of of the underlying graph $U\!G(D)$ of $D$ and let $M_2=M(D)=(A,{\cal I}(D))$ be the matroid from Section~\ref{matroidsec} whose bases are the antistrong sets consisting of $2n-1$ arcs. Let $M=M_1\vee{}M_2$ be the union of the matroids $M_1,M_2$, that is, a set $X$ of arcs is independent in $M$ if and only we can partition $X$ into $X_1, X_2$ such that $X_i$ is independent in $M_i$. For each of the matroids $M_1, M_2$ we can check in polynomial time whether a given subset $X$ of arcs is independent in $M_1$ and $M_2$ (for $M_1$ we need to check that there is no cycle in $U\!G(D)[X]$ and for $M_2$ we need to check that there is no cycle in the subgraph of $B(D)[E_X]$ induced by the edges $E_X$ corresponding to $X$ in $B(D)$% ). Thus it follows from Edmonds' algorithm for matroid partitioning~\cite{edmondsJRNBS69} that we can find a base of $M$ in polynomial time using the independence oracles of $M_1,M_2$. The desired digraph $H$ exists if and only if the size of a base in $M$ is $(2n-1)+(n-1)=3n-2$. \hfill$\diamond$\\\vspace{2mm} A similar proof gives the following. \begin{theorem} We can decide in polynomial time whether a digraph $D$ contains $k+\ell$ arc-disjoint spanning subdigraphs $D_1,\ldots{},D_{k+\ell}$ such that $D_1,\ldots{},D_k$ are antistrong and $U\!G(D_{k+1}),\ldots{},U\!G(D_{k+\ell})$ are connected. \end{theorem} \section{Remarks and open problems} We saw in Theorem~\ref{twospstrongD} that it is NP-complete to decide whether a given digraph contains two arc-disjoint spanning strong subdigraphs. We would be interested to know what happens if we modify the problem as follows. \begin{question} Can we decide in polynomial time whether a digraph $D$ contains arc-disjoint spanning subdigraphs $D_1,D_2$ such that $D_1$ is antistrong and $D_2$ is strongly connected? \end{question} Inspired by Theorem~\ref{deleteanticon} it is natural to ask the following intermediate question. \begin{question} \label{anti+2con} Can we decide in polynomial time whether a digraph $D$ contains arc-disjoint spanning subdigraphs $D_1,D_2$ such that $D_1$ is antistrong and $U\!G(D_2)$ is 2-edge-connected? \end{question} The following conjecture was raised in~\cite{bangC24}. \begin{conjecture}\cite{bangC24} There exists a natural number $k$ such that every $k$-arc-strong digraph has arc-disjoint strong spanning subdigraphs $D_1,D_2$. \end{conjecture} \noindent Perhaps the following special case may be easier to study. \begin{conjecture} There exists a natural number $k$ such that every digraph $D$ which is both $k$-arc-strong and $k$-arc-antistrong has arc-disjoint strong spanning subdigraphs $D_1,D_2$. \end{conjecture} \begin{problem} Does there exist a polynomial algorithm for deciding whether a given undirected graph $G$ has an orientation $D$ which is both strong and antistrong? \end{problem} \bigskip \centerline{*} \bigskip {\bf Acknowledgement.} Bang-Jensen and Jackson wish to thank Jan van den Heuvel for stimulating discussions about antistrong connectivity. They also thank the Mittag-Leffler Institute for providing an excellent working environment.

https://arxiv.org/abs/1301.4590

Computing Hypermatrix Spectra with the Poisson Product Formula

We compute the spectrum of the "all ones" hypermatrix using the Poisson product formula. This computation includes a complete description of the eigenvalues' multiplicities, a seemingly elusive aspect of the spectral theory of tensors. We also give a general distributional picture of the spectrum as a point-set in the complex plane, and use our techniques to analyze the spectrum of "sunflower hypergraphs", a class that has played a prominent role in extremal hypergraph theory.

\section{Introduction} There are few extant techniques that allow one to determine the characteristic polynomial of a symmetric hypermatrix in the sense of Qi (\cite{Qi_2005}). Indeed, the lack of a simple formula for the symmetric hyperdeterminant makes many questions about hypermatrices much more difficult than the same questions for matrices. In particular, we know of no simple way of determining the algebraic multiplicities of a hypermatrix's eigenvalues. The characteristic polynomial and the symmetric hyperdeterminant are both defined using polynomial resultants, and so finding better methods of computing the resultant is one way to approach this problem. There is a tool for inductively computing resultants, the so-called ``Poisson product formula'', which is given in \cite{Cox_Little_OShea_2005}. Presently, we show that it can be used to compute the characteristic polynomial for some types of hypermatrices, thereby completely describing their spectra. The statement of the product formula (from \cite{Cox_Little_OShea_2005}) is as follows. \begin{theorem} Let \( F_0, F_1, \ldots, F_n \) be homogeneous polynomials of respective degrees \( d_0, \ldots, d_n \) in \( K[x_0, \ldots, x_n] \) where \( K \) is an algebraically closed field. For \( 0\leq i\leq n, \) let \( \overline{F_i} \) be the homogeneous polynomial in \( K[x_1, \ldots, x_n] \) obtained by substituting \( x_0=0 \) in \( F_i \), and let \( f_i \) be the polynomial in \( K[x_1, \ldots, x_n] \) obtained by substituting \( x_0=1 \) in \( F_i. \) Let \( V \) be the set of simultaneous zeros of the system of polynomials \( \{f_1, f_2, \ldots, f_n\}, \) that is, \( V \) is the affine variety defined by the polynomials. If \( \Res(\overline{F_1}, \overline{F_2}, \ldots, \overline{F_n}) \neq 0 \), then \( V \) is a zero-dimensional variety (a finite set of points), and \[\Res(F_0, F_1, \ldots, F_n) = \Res(\overline{F_1}, \overline{F_2}, \ldots, \overline{F_n})^{d_0}\prod_{p\in V} f_n(p)^{m(p)}\] where \( m(p) \) is the \emph{multiplicity} of a point \( p\in V. \) \end{theorem} A few notes are in order. In our case, our field will be \( K = \overline{\mathbb{C}[\lambda]} \), the algebraic closure of the (fraction field of) polynomials in an indeterminant \( \lambda \) with complex coefficients. All computations in the sequel will concern characteristic polynomials of hypermatrices with complex entries, which are always monic (more importantly, nonzero), so the condition \( \Res(\overline{F_1}, \overline{F_2}, \ldots, \overline{F_n}) \neq 0 \) always holds. The \emph{multiplicity} referred to above is defined in \cite{Cox_Little_OShea_2005}, but will be inconsequential in our uses. In each case, we show that every point of the variety has multiplicity at most one, and so identically one. Indeed, the variety \( V \) referred to above has at most \( \prod_{i=1}^n \deg(f_i) \) points in it by Bezout's theorem, and in our applications we exhibit exactly this many distinct points. \section{The All-Ones Hypermatrix} Fix \( m \geq 2 \). Let \( \mathbb{J}_n^m \) denote the order \( m \), dimension \( n \) hypermatrix consisting of all ones. We write \( \hat{0} \) and \( \hat{1} \) for the all-zeros and all-ones vectors of dimension \( n \), respectively. For \( i \in [n] \), define \[ F_i = \lambda x_i^{m-1} - \left ( \sum_{j=1}^n x_j \right )^{m-1}. \] Evidently, the characteristic polynomial of \( \mathbb{J}_n^m \) is \( \phi_{\mathbb{J}^m_n}(\lambda) = \phi(\lambda) = \Res(F_1,\ldots,F_n) \), a polynomial in \(\lambda\) of degree \(n (m-1)^{n-1}\). First, we consider the spectrum of \( \mathbb{J}_n^m \) as a set, i.e., \( S = \{\lambda : \phi(\lambda)=0\} \), or alternatively, \[ S = \left \{ \lambda : \exists \mathbf{x} \forall i \in [n] (F_i(\mathbf{x}) = 0) \right \}. \] Note that \( \mathbf{x} = (1,-1,0,0,\ldots,0) \) is a solution of the system \( \mathcal{F} = \{F_i = 0\}_{i=1}^n \) for \( \lambda = 0 \), so \( 0 \in S \). Now, assume \( \lambda \neq 0 \), and let \( \mathbf{x} \) be an eigenvector of \( \lambda \). Note that \( \mathbf{x} \) has all nonzero entries, since otherwise \( x_k = 0 \) would imply \[ \lambda x_k^{m-1} = 0 = \left ( \sum_{j=1}^n x_j \right )^{m-1}, \] so that \( \lambda x_i^{m-1} = 0 \), i.e., \( x_i = 0 \), for all \( i \in [n] \), whence \( \mathbf{x} = \hat{0} \), a contradiction. Next, note that \begin{equation} \label{eq5} x_i^{m-1} = \frac{\left ( \sum_{j=1}^n x_j \right )^{m-1}}{\lambda} \end{equation} for all \( i \in [n] \). Since the right-hand side of (\ref{eq5}) is independent of \( i \), we have that \( x_i^{m-1} = x_j^{m-1} \) for all \( i,j \in [n] \). Since the \( F_i \) are homogeneous (in \( \mathbf{x} \)), we may assume that \( x_1 = 1 \) and therefore \( x_i = \zeta_{m-1}^{a_i} \) for all \( i \in [n] \), where \( \zeta_q \) denotes a primitive \( q \)-th root of unity. Therefore, each eigenvector \( \mathbf{x} \) corresponding to a nonzero \( \lambda \) consists of \( r_i \) coordinates \( \zeta_{m-1}^i \) for some \( r_0,\ldots,r_{m-2} \in \mathbb{N} \) with \( \sum_{i=0}^{m-2} r_i = n \). It follows that \( \lambda = \sum_{i=0}^{m-2} r_i \zeta_{m-1}^i \), so we may conclude that \[ S = \left \{ \left ( \mathbf{r} \cdot \hat{\zeta} \right )^{m-1} : \mathbf{r} \in \mathbb{N}^{m-1}, \mathbf{r} \cdot \hat{1} = n \right \} \cup \{0\}. \] where \( \hat{\zeta} \) denotes the vector \( (1,\zeta_{m-1},\ldots,\zeta_{m-1}^{m-2}) \). Although the above argument, which fully describes the {\em set} spectrum, gives no information about the multiplicities of the eigenvalues, it does yield a natural conjecture: the multiplicity of \( \xi = ( \mathbf{r} \cdot \hat{\zeta} )^{m-1} \) should be \( 1/(m-1) \) times the number of ways to choose \( s_0,\ldots,s_{m-2} \) so that \[ \mathbf{s} \cdot \hat{\zeta} = \xi^{1/(m-1)} \zeta_{m-1}^j \] for some \( j \in \{0,\ldots,m-2\} \). Write \( \mu(\mathbf{r}) \) for this quantity. In order to affirm our intuition, we apply the Poisson product formula. To wit, let \begin{eqnarray*} \overline{F}_i = & F_i(x_1,x_2,\ldots,x_{n-1},0) \qquad & \text{for } i \in [n] \text{ and}\\ f_i = & F_i(x_1,x_2,\ldots,x_{n-1},1) \qquad & \text{for } i \in [n]. \end{eqnarray*} Then, we define \begin{equation} \label{eq6} V = \mathbb{V}(f_1,\ldots,f_{n-1}) \subseteq \mathbb{K}^{n-1} \end{equation} the zero locus of the \( f_i \)'s, where \( \mathbb{K} = \overline{\mathbb{C}[\lambda]} \), the algebraic closure of \( \mathbb{C}[\lambda] \). Then the product formula states that \begin{equation} \label{eq7} \Res(F_1,\ldots,F_n) = \Res(\overline{F}_1,\ldots,\overline{F}_{n-1})^{\deg(F_n)} \prod_{\mathbf{x} \in V} f_n(\mathbf{x})^{m(\mathbf{x})}, \end{equation} where \( m(\mathbf{x}) \) denotes the multiplicity of the point \( \mathbf{x} \in V \). As we shall see below, there are \( (m-1)^{n-1} \) distinct \( \mathbf{x} \) in the product above -- the maximum possible number of such points by B\'{e}zout's Theorem applied to (\ref{eq6}), since \( \deg(f_i) = m-1 \) for all \( i \in [n-1] \). Therefore, \( m(\mathbf{x}) = 1 \) for all \( \mathbf{x} \in V \). Note also that the hypothesis of the product formula, that \( \psi = \Res(\overline{F}_1,\ldots,\overline{F}_{n-1}) \neq 0 \), holds here because \( \psi \) is a nonvanishing polynomial in \( \lambda \). \begin{lemma} \label{lem:rootsofunity} Suppose \( p(x) = \alpha (x - \beta_1) \cdots (x - \beta_r) \) is a polynomial of degree \( r \) with lead coefficient \( \alpha \). Then \[ \prod_{j=0}^{r-1} p(\zeta_r^j x) = \alpha^r \prod_{j=0}^{r-1} (x^r - \beta_j^r) \] \end{lemma} \begin{proof} We can write \begin{align*} \prod_{j=0}^{r-1} p(\zeta_r^j x) &= \prod_{j=0}^{r-1} [\alpha (\zeta_r^j x - \beta_1) \cdots (\zeta_r^j x - \beta_r)] \\ &= \alpha^r \prod_{j=0}^{r-1} \zeta_r^{jr} (x - \zeta_r^{-j} \beta_1) \cdots (x - \zeta_r^{-j} \beta_r)] \\ &= \alpha^r (x^r - \beta_1^r) \cdots (x^r - \beta_r^r). \end{align*} \end{proof} \begin{theorem} \label{thm:charpolyJJ} Let \( \phi_n(\lambda) \) denote the characteristic polynomial of \( \mathbb{J}_n^m \), \( n \geq 2 \). Then \[ \phi_n(\lambda) = \lambda^{(n-1) (m-1)^{n-1}} \prod_{\substack{\mathbf{r} \in \mathbb{N}^{m-1} \\ \mathbf{r} \cdot \hat{1} = n}} \left ( \lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1} \right )^{\mu(\mathbf{r})/(m-1)}. \] \end{theorem} \begin{proof} We proceed by induction on \( n \) using (\ref{eq7}). The base case \( n=1 \) is trivial. For the inductive step, we write \begin{align*} \phi_n(\lambda) &= \Res(F_1,\ldots,F_n) \\ &= \Res(\overline{F}_1,\ldots,\overline{F}_n)^{m-1} \prod_{\mathbf{x} \in V} f_n(\mathbf{x})^{m(\mathbf{x})} \\ &= \phi_{n-1}(\lambda)^{m-1} \prod_{\mathbf{x} \in V} f_n(\mathbf{x})^{m(\mathbf{x})}. \end{align*} By the inductive hypothesis, \[ \phi_{n-1}(\lambda) = \lambda^{(n-2) (m-1)^{n-2}} \prod_{\mathbf{r} \cdot \hat{1} = n-1} \left ( \lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1} \right )^{\mu(\mathbf{r})/(m-1)}. \] Hence, \begin{equation} \label{eq8} \phi_{n}(\lambda) = \lambda^{(n-2) (m-1)^{n-1}} \prod_{\mathbf{r} \cdot \hat{1} = n-1} \left ( \lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1} \right )^{\mu(\mathbf{r})} \prod_{\mathbf{x} \in V} f_n(\mathbf{x})^{m(\mathbf{x})}. \end{equation} Note that \[ V = \left \{ \mathbf{x} \in \mathbb{K}^{n-1} : \forall i \in n \left (\lambda x_i^{m-1} - \left (1 + \sum_{j=1}^{n-1} x_j \right )^{m-1} = 0 \right ) \right \}. \] We claim that \( x_i \neq 0 \) for any point \( \mathbf{x} \in V \). Indeed, suppose \( x_k = 0 \). Then \[ 0 = \lambda x_k^{m-1} = \left ( 1 + \sum_{j=1}^{n-1} x_j \right )^{m-1}. \] Since this implies that, for any \( j \in [n-1] \), \[ \lambda x_j^{m-1} = 0, \] it must be the case that \( x_j = 0 \) for all \( j \in [n-1] \), since \( \lambda \) is a nonzero element of \( \mathbb{K} \). Thus, \[ 0 = \left ( 1 + \sum_{j=1}^{n-1} x_j \right )^{m-1} = (1 + 0)^{m-1} = 1, \] we have a contradiction. Therefore, \( x_i \neq 0 \) for any \( \mathbf{x} \in V \). For any \( i \in [n-1] \), we may write \[ x_i^{m-1} = \frac{\left (1 + \sum_{j=1}^{n-1} x_j \right )^{m-1}}{\lambda}. \] Since \( x_i \) is nonzero, it must be the case that \( x_i = \zeta_{m-1}^{j_i} a \) for some \( j_i \in \{0,\ldots,m-2\} \) and each \( i \in [n-1] \), where \[ a = \frac{1 + \sum_{j=1}^{n-1} x_j}{\lambda^{1/{m-1}}} \] (for any choice of the root). Suppose that \( r_0 \) of the \( x_i \) are \( a \), \( r_1 \) of the \( x_i \) are \( \zeta_{m-1} a \), \( r_2 \) of the \( x_i \) are \( \zeta_{m-1}^2 a \), etc., where \( \mathbf{r} \cdot \hat{1} = n-1 \). Then \[ a = \frac{1 + a (\mathbf{r} \cdot \hat{\zeta})}{\lambda^{1/(m-1)}}, \] so that \( a = (\lambda^{1/(m-1)} - (\mathbf{r} \cdot \hat{\zeta}))^{-1} \). Since then \( a \) is uniquely determined by the choice of \( \mathbf{r} \), this yields exactly \( (m-1)^{n-1} \) distinct points \( \mathbf{x} \in V \), one for each function \( \nu : [n-1] \rightarrow \{0,\ldots,m-2\} \), where \( \nu(j) \) represents the exponent \( j_i \) of \( \zeta_{m-1} \) chosen for the \( j \)-th coordinate of \( \mathbf{x} \). (Thus, the statement above concerning the multiplicities of the \( \mathbf{x} \in V \) is verified.) If \( \mathbf{x} \in V \), then \( \zeta_{m-1}^j \mathbf{x} \) is also a point of \( V \) for each \( j \in [m-2] \). Note that \begin{align*} f_n(\mathbf{x}) &= \lambda - \left ( 1 + \frac{\mathbf{r} \cdot \hat{\zeta}}{\lambda^{1/(m-1)} - (\mathbf{r} \cdot \hat{\zeta})} \right )^{m-1} \\ &= \lambda - \left ( \frac{\lambda^{1/(m-1)}}{\lambda^{1/(m-1)} - (\mathbf{r} \cdot \hat{\zeta})} \right )^{m-1} \\ &= \lambda \left [ 1 - \left ( \lambda^{1/(m-1)} - (\mathbf{r} \cdot \hat{\zeta}) \right )^{1-m} \right ] \\ &= \lambda \frac{ \left ( \lambda^{1/(m-1)} - (\mathbf{r} \cdot \hat{\zeta}) \right )^{m-1} - 1 }{\left ( \lambda^{1/(m-1)} - (\mathbf{r} \cdot \hat{\zeta}) \right )^{m-1}}. \end{align*} Then, \begin{align*} \prod_{j=0}^{m-2} f_n(\zeta_{m-1}^j \mathbf{x}) &= \lambda^{m-1} \frac{ \prod_{j=0}^{m-2} \left [ \left ( \lambda^{1/(m-1)} - \zeta_{m-1}^j (\mathbf{r} \cdot \hat{\zeta}) \right )^{m-1} - 1 \right ]}{\prod_{j=0}^{m-2} \left ( \lambda^{1/(m-1)} - \zeta_{m-1}^j (\mathbf{r} \cdot \hat{\zeta}) \right )^{m-1}} \\ &= \lambda^{m-1} \frac{ \prod_{j=0}^{m-2} g(\zeta_{m-1}^j (\mathbf{r} \cdot \hat{\zeta}))}{ \prod_{j=0}^{m-2} h(\zeta_{m-1}^j (\mathbf{r} \cdot \hat{\zeta}))}, \end{align*} where \( g(x) = \left ( \lambda^{1/(m-1)} - x \right )^{m-1} - 1 \) and \( h(x) = \left ( \lambda^{1/(m-1)} - x \right )^{m-1} \). All \( m-1 \) roots of \( h(x) \) are \( \lambda^{1/(m-1)} \), and its lead coefficient is \( (-1)^{m-1} \). Therefore, by Lemma \ref{lem:rootsofunity}, \begin{align*} \prod_{j=0}^{m-1} h(\zeta_{m-1}^j (\mathbf{r} \cdot \hat{\zeta})) &= (-1)^{(m-1)^2} \prod_{j=0}^{m-1} ((\mathbf{r} \cdot \hat{\zeta})^{m-1} - \lambda) \\ &= (-1)^{(m-1)} ((\mathbf{r} \cdot \hat{\zeta})^{m-1} - \lambda)^{m-1} \\ &= (\lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1})^{m-1}. \end{align*} Define \( k(y) = \lambda - (x+y)^{m-1} \). Then \( k(y) = 0 \) implies \( y = \lambda^{1/(m-1)} \zeta_{m-1}^j - x \) for some \( 0 \leq j \leq m-2 \), so Lemma \ref{lem:rootsofunity} yields \begin{align*} \prod_{j=0}^{m-2} k(\zeta_{m-1}^j y) &= (-1)^{m-1} \prod_{j=0}^{m-2} (y^{m-1} - (\lambda^{1/(m-1)} \zeta_{m-1}^j - x)^{m-1}) \\ &= \prod_{j=0}^{m-2} ((\lambda^{1/(m-1)} - \zeta_{m-1}^{-j} x)^{m-1} \zeta_{m-1}^{j(m-1)} - y^{m-1}) \\ &= \prod_{j=0}^{m-2} ((\lambda^{1/(m-1)} - \zeta_{m-1}^{-j} x)^{m-1} - y^{m-1}). \end{align*} Hence, \begin{align*} \prod_{j=0}^{m-2} g(\zeta_{m-1}^j x) &= \prod_{j=0}^{m-2} ((\lambda^{1/(m-1)} - \zeta_{m-1}^{j} x)^{m-1} - 1) \\ &= \prod_{j=0}^{m-2} ((\lambda^{1/(m-1)} - \zeta_{m-1}^{-j} x)^{m-1} - 1) \\ &= \prod_{j=0}^{m-2} k(\zeta_{m-1}^j \cdot 1) \\ &= \prod_{j=0}^{m-2} (\lambda - (x+\zeta_{m-1}^j)^{m-1}). \end{align*} and we may conclude that \[ \prod_{j=0}^{m-2} f_n(\zeta_{m-1}^j \mathbf{x}) = \lambda^{m-1} \frac{ \prod_{j=0}^{m-2} (\lambda - ((\mathbf{r} \cdot \hat{\zeta})+\zeta_{m-1}^j)^{m-1})}{(\lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1})^{m-1}}. \] Let us define, then, the function \[ \tilde{f}_n(\mathbf{x}) = \frac{\lambda \left ( \prod_{j=0}^{m-2} (\lambda - ((\mathbf{r} \cdot \hat{\zeta})+\zeta_{m-1}^j)^{m-1}) \right )^{1/(m-1)}}{\lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1}}. \] By the above computation, we may write \begin{align*} \prod_{\mathbf{x} \in V} f_n(\mathbf{x}) &= \prod_{\mathbf{x} \in V} \tilde{f}_n(\mathbf{x}) \\ &= \prod_{\mathbf{r} \cdot \hat{1} = n-1} \left ( \frac{\lambda \left ( \prod_{j=0}^{m-2} (\lambda - ((\mathbf{r} \cdot \hat{\zeta})+\zeta_{m-1}^j)^{m-1}) \right )^{1/(m-1)}}{\lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1}} \right )^{\mu(\mathbf{r})} \\ &= \frac{ \lambda^{(m-1)^{n-1}} }{\prod_{\mathbf{r} \cdot \hat{1} = n-1} ( \lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1})^{\mu(\mathbf{r})}} \\ & \qquad \cdot \prod_{\mathbf{r} \cdot \hat{1} = n-1} \left ( \prod_{j=0}^{m-2} (\lambda - ((\mathbf{r} \cdot \hat{\zeta})+\zeta_{m-1}^j)^{m-1}) \right )^{\mu(\mathbf{r})/(m-1)}. \end{align*} The total multiplicity of the term \( \lambda - ((\mathbf{r} \cdot \hat{\zeta})+\zeta_{m-1}^j) \) in the right-hand product is \( \mu(\mathbf{r}^\prime)/(m-1) \), where \( \mathbf{r}^\prime = \mathbf{r} + \mathbf{e}_j \), \( \mathbf{e}_j \) the elementary basis vector with nonzero coordinate \( j \). Therefore, \begin{align*} \prod_{\mathbf{x} \in V} f_n(\mathbf{x}) &= \frac{ \lambda^{(m-1)^{n-1}} }{\prod_{\mathbf{r} \cdot \hat{1} = n-1} ( \lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1})^{\mu(\mathbf{r})}} \\ & \qquad \cdot \prod_{\mathbf{r}^\prime \cdot \hat{1} = n} (\lambda - (\mathbf{r}^\prime \cdot \hat{\zeta}))^{m-1})^{\mu(\mathbf{r}^\prime)/(m-1)}. \end{align*} Substituting this expression into (\ref{eq7}) yields \[ \phi_{n}(\lambda) = \lambda^{(n-1) (m-1)^{n-1}} \prod_{\mathbf{r} \cdot \hat{1} = n} (\lambda - (\mathbf{r} \cdot \hat{\zeta})^{m-1})^{\mu(\mathbf{r})/(m-1)}, \] which is the desired conclusion. \end{proof} Our next result illustrates the geometry of the eigenvalues of \(\mathbb{J}_n^m\) as a point-set in the complex plane. Let \( \nu_n \) denote the probability measure on \( \mathbb{C} \) defined by choosing an element of the spectrum of \( \mathbb{J}_n^m \) (with multiplicity) uniformly at random. This distribution is straightforward to describe exactly in several cases, and to describe up to total variation in all other cases. \begin{theorem} Let \( \binom{n}{a_1,\ldots,a_k} \) denote the multinomial coefficient, which we define to be zero unless \( n = \sum_{j=1}^k a_j \) and \( a_j \in \mathbb{N} \) for all \( j \in [k] \), in which case it equals \( n! / (a_1! \cdots a_k!) \). The distribution \( \nu_n \) is everywhere zero except as follows: \begin{enumerate} \item When \( m=2 \), \( \nu_n(0) = (n-1)/n \) and \( \nu_n(n) = 1/n \). \item When \( m=3 \), \(\nu_n(k) = 2^{-n+1} \binom{n}{n + \sqrt{k}}/n\) for \(k \neq 0\) and \(\nu_n(0) = (n-1)/n + 2^{-n} \binom{n}{n/2}/n\). \item When \( m=4 \), \[ \nu_n\left ( \frac{x + y \sqrt{-3}}{2} \right ) = \frac{3^{-n+1}}{n} \binom{n}{\frac{n+a}{3}, \frac{2n-a+3b}{6}, \frac{2n-a-3b}{6}} \] for \((x,y) \in \mathbb{Z}^2 \setminus \{(0,0)\}\) if there exist \(a,b \in \mathbb{Z}\) so that \(4x = a(a^2 - 9b^2)\) and \(4y = 3b(a^2-b^2)\), and \[ \nu_n(0) = \frac{n-1}{n} + \frac{3^{-n}}{n} \binom{n}{n/3, n/3, n/3}. \] \item When \( m=5 \), \[ \nu_n(x+yi) = \frac{4^{-n+1}}{n} \binom{n}{(n+a+b)/2} \binom{n}{(n-a+b)/2} \] for \((x,y) \in \mathbb{Z}^2 \setminus \{(0,0)\}\), if there exist \(a,b \in \mathbb{Z}\) so that \(x = a^4 - 6a^2b^2 + b^4\) and \(y = 4 ab (a^2 - b^2)\), and \[ \nu_n(0) = \frac{n-1}{n} + \frac{4^{-n}}{n} \binom{n}{n/2}^2. \] \item For any \(m \geq 4\), \[ \sqrt{n} \nu_n \circ \psi_n - \frac{n-1}{\sqrt{n}} \delta_0 \overset{D}{\longrightarrow} \mathcal{N}(0,1) \circ \rho_{m-1}, \] where \( \mathcal{N}(t,\sigma) \) is a standard normal distribution on \( \mathbb{C} \) with mean \( t \) and standard deviation \( \sigma \), \(\psi_n(\xi) = \xi/\sqrt{n}\) for \(\xi \in \mathbb{C}\), \(\delta_0\) is the probability distribution consisting of a single atom at \(0\), and \(\rho_r\) is the multivalued map which sends \(\xi \in \mathbb{C}\) to all of its \(r\)-th roots, \(r \in \mathbb{N}^+\). \end{enumerate} \end{theorem} \begin{proof} Note that there are \(n (m-1)^{n-1}\) eigenvalues of \(\mathbb{J}_n^m\), counted with (algebraic) multiplicity. In the statement of Theorem \ref{thm:charpolyJJ}, the exponent of \(\lambda\) to the left of the product is \((n-1)(m-1)^{n-1}\); this accounts for a \( (n-1)(m-1)^{n-1} / (n (m-1)^{n-1}) = (n-1)/n\) fraction of the roots of the characteristic polynomial. It follows that for, \(\xi \neq 0\), \(\nu(\xi)\) is \(1/n\) times the probability that a simple random walk whose steps are the \((m-1)\)-st roots of unity ends at some \(m-1\)-st root of \(\xi\) after \(n\) steps, and, for \(\xi = 0\), \(\nu(\xi)\) is this same quantity, plus \((n-1)/n\). \begin{enumerate} \item For \(m=2\), there is only one \((m-1)\)-st root of unity: \(1\). Therefore, a simple random walk of length \(n\) with this as its only choice of step will end at \(n\) with probability \(1\). It follows that \(\nu_n(n) = 1/n\) and \(\nu_n(0) = (n-1)/n\). \item For \(m=3\), there are two \((m-1)\)-st roots of unity: \(1\) and \(-1\). Therefore, a walk of length \(n\) with \(\pm 1\) as its steps will end at \(k\) if and only if it consists of \(x\) steps in the direction of \(1\) and \(y\) steps in the direction of \(-1\), where \(x-y = k\) and \(x+y = n\). Solving for \(x\) and \(y\), one obtains \(x = (n+k)/2\) and \(y = (n-k)/2\). The number of such walks is \(\binom{n}{(n+k)/2}\), so that the probability of ending at \(k\) is \(2^{-n} \binom{n}{(n+k)/2}\). Since the only elements of \(\mathbb{C}\) both of whose square roots are elements of \(\mathbb{Z}\) are themselves elements of \(\mathbb{Z}\), the probability that a random walk with steps \(\pm 1\) ends at an \((m-1)\)-st root of \(k \in \mathbb{Z} \setminus \{0\}\) is \[ 2^{-n} \left ( \binom{n}{(n+\sqrt{k})/2} + \binom{n}{(n-\sqrt{k})/2} \right ) = 2^{-n+1} \binom{n}{(n + \sqrt{k})/2}. \] and \(2^{-n} \binom{n}{n/2}\) if \(k = 0\). Therefore, \(\nu_n(k) = 2^{-n+1} \binom{n}{n + \sqrt{k}}/n\) for \(k \neq 0\) and \(\nu_n(0) = (n-1)/n + 2^{-n} \binom{n}{n/2}/n\). \item For \(m=4\), there are three \((m-1)\)-st roots of unity: \(S = \{1, \zeta_3, \zeta_3^2\}\), where \(\zeta = \exp(2 \pi i / 3) = (-1+i\sqrt{3})/2 \). Therefore, a walk of length \(n\) with \(S\) as its steps will end at \(a/2 + b \sqrt{-3}/2\), \(a,b \in \mathbb{Z}\) with \(a \equiv n \pmod{2}\), if and only if it consists of \(\alpha\) steps in the direction of \(1\), \(\beta\) steps in the direction of \(\zeta\), and \(\gamma\) steps in the direction of \(\zeta^2\), where \(\alpha,\beta,\gamma \in \mathbb{N}\) and \begin{eqnarray*} \alpha + \beta + \gamma & = & n \\ \alpha - \beta/2 - \gamma/2 & = & a/2 \\ \beta \sqrt{3}/2 - \gamma \sqrt{3}/2 & = & b \sqrt{3}/2. \end{eqnarray*} Solving for \(\alpha\), \(\beta\), and \(\gamma\) yields: \begin{eqnarray*} \alpha & = & \frac{1}{3} a + \frac{1}{3} n \\ \beta & = & -\frac{1}{6} a + \frac{1}{2} b + \frac{1}{3} n \\ \gamma & = & -\frac{1}{6} a - \frac{1}{2} b + \frac{1}{3} n. \end{eqnarray*} Therefore, the number of walks on the lattice generated by \(S\) that end at \(a/2 + b \sqrt{-3}/2\) is given by \[ F(n,a,b) = \binom{n}{\frac{n+a}{3}, \frac{2n-a+3b}{6}, \frac{2n-a-3b}{6}}. \] Note that \[ \left ( \frac{a}{2} + \frac{b \sqrt{-3}}{2} \right )^3 = \frac{a(a^2 - 9 b^2)}{8} + \frac{3\sqrt{-3}b(a^2 - b^2)}{8}. \] Since \(a \equiv b \pmod{2}\) implies that \(4 | a^2 - b^2\) and \(4 | a^2 - 9b^2\), the right-hand side is an element of \(\frac{1}{2} \mathbb{Z}[\sqrt{-3}]\). We may conclude that the probability that a simple random walk whose steps are the cube roots of unity ends after \(n\) steps at a cube root of \(x/2 + y\sqrt{-3}/2 \neq 0\), \(x,y \in \mathbb{Z}\), is \(3 \cdot 3^{-n} F(n,a,b)\) if there exist \(a,b \in \mathbb{Z}\) so that \(4x = a(a^2 - 9b^2)\) and \(4y = 3b(a^2-b^2)\), and \(0\) otherwise. Therefore, for \((x, y) \neq (0,0)\), \[ \nu_n\left ( \frac{x + y \sqrt{-3}}{2} \right ) = \frac{3^{-n+1}}{n} \binom{n}{\frac{n+a}{3}, \frac{2n-a+3b}{6}, \frac{2n-a-3b}{6}} \] if there exist \(a,b \in \mathbb{Z}\) so that \(4x = a(a^2 - 9b^2)\) and \(4y = 3b(a^2-b^2)\), and \[ \nu_n(0) = \frac{n-1}{n} + \frac{3^{-n}}{n} \binom{n}{n/3, n/3, n/3}. \] \item For \(m=5\), there are four \((m-1)\)-st roots of unity: \(S = \{\pm 1, \pm i\}\). A simple random walk of length \(n\) with \(S\) as its steps can be recast as a walk on \(\mathbb{Z}^2\) using the standard identification \(\mathbb{C} \equiv \mathbb{R}^2\); it is well-known (see, for example, \cite{DoyleSnell84}), that such a walk will end at \(a + bi\), \(a,b \in \mathbb{Z}\), with probability \[ 4^{-n} \binom{n}{(n+a+b)/2} \binom{n}{(n-a+b)/2}. \] Since \((a+bi)^4 = (a^4 - 6a^2b^2 + b^4) + 4 ab (a^2 - b^2) i \in \mathbb{Z}[i]\), we may conclude that, for \((x,y) \in \mathbb{Z}^2 \setminus \{(0,0)\}\), \[ \nu_n(x+yi) = \frac{4^{-n+1}}{n} \binom{n}{(n+a+b)/2} \binom{n}{(n-a+b)/2} \] if there exist \(a,b \in \mathbb{Z}\) so that \(x = a^4 - 6a^2b^2 + b^4\) and \(y = 4 ab (a^2 - b^2)\), and \[ \nu_n(0) = \frac{n-1}{n} + \frac{4^{-n}}{n} \binom{n}{n/2}^2. \] \item Let \(X_r\) denote a length \(n\) simple random walk on \(\mathbb{C}\) whose steps are the \(r\)-th roots of unity, \(r \geq 1\), and let \(Y_r = (\Re(X_r),\Im(X_r))\). The covariance matrix of \(Y_r\) is given by \begin{align*} \cov(Y_r) &= \mathbb{E} [\|Y_r\|_2^2] \\ &= \left [ \begin{array}{cc} \mathbb{E} [\Re(X_r)^2] & \mathbb{E}[\Re(X_r)\Im(X_r)] \\ \mathbb{E}[\Re(X_r)\Im(X_r)] & \mathbb{E}[\Im(X_r)^2] \end{array} \right ]. \end{align*} Let \(\zeta_r = \exp(2 \pi i/r)\). Note that \[ \sum_{j=0}^{r-1} \zeta_r^{2j} = \frac{\zeta_r^{2r}-1}{\zeta_r^2 - 1} = 0 \] as long as the denominator is not zero, i.e., \(r (= m-1) \geq 3\). Taking the real and imaginary parts of this equation yields \[ \sum_{j=0}^{r-1} \cos(4 \pi i j / r) = \sum_{j=0}^{r-1} \sin(4 \pi i j / r) = 0. \] Therefore, \begin{align*} \mathbb{E} [\Re(X_r)^2] &= \frac{1}{r} \sum_{j=0}^{r-1} \cos^2(2 \pi ij/r) \\ &= \frac{1}{2r} \sum_{j=0}^{r-1} \cos(4 \pi ij/r) + 1\\ &= 1/2. \end{align*} It also follows that \begin{align*} \mathbb{E} [\Im(X_r)^2] &= \frac{1}{r} \sum_{j=0}^{r-1} \sin^2(2 \pi ij/r) \\ &= \frac{1}{2r} \sum_{j=0}^{r-1} 1 - \cos(4 \pi ij/r) \\ &= 1/2. \end{align*} Furthermore, \begin{align*} \mathbb{E} [\Re(X_r)\Im(X_r)] &= \frac{1}{r} \sum_{j=0}^{r-1} \sin(2 \pi ij/r) \cos(2 \pi ij/r) \\ &= \frac{1}{2r} \sum_{j=0}^{r-1} 2 \sin(4 \pi ij/r) \\ &= 0. \end{align*} Therefore, \(\cov(Y_r) = \frac{1}{2} I\), and, by the Multidimensional Central Limit Theorem (see, e.g., \cite{vanderVaart98}), \[ \frac{X_r}{\sqrt{n}} \overset{D}{\longrightarrow} \mathcal{N}(0,1). \] It follows that \[ \sqrt{n} \nu_n \circ \psi_n - \frac{n-1}{\sqrt{n}} \delta_0 \overset{D}{\longrightarrow} \mathcal{N}(0,1) \circ \rho_{m-1}. \] \end{enumerate} \end{proof} When \(m = 7\), it does not appear possible to write down a closed-form expression for \(\nu_n(\xi)\) (\cite{Zeilberger13}). For other \(m\), however, we conjecture the following statement that strengthens the conclusion of part (5) above. \begin{conj} For \( m = 6 \) or \( m > 7 \), \(\epsilon > 0\), and any measurable set \(E \subset \mathbb{C}\), \[ \lim_{n \rightarrow \infty} n^{-\epsilon} \left | \int_E n \nu_n \circ \psi_n - (n-1) \delta_0 - \sqrt{n} \mathcal{N}(0,1) \circ \rho_{m-1} \right | = 0 \] where \( \mathcal{N}(t,\sigma) \) is a standard normal distribution on \( \mathbb{C} \) with mean \( t \) and standard deviation \( \sigma \), \(\psi_n(\xi) = \xi/\sqrt{n}\) for \(\xi \in \mathbb{C}\), \(\delta_0\) is the probability distribution consisting of a single atom at \(0\), and \(\rho_r\) is the multivalued map which sends \(\xi \in \mathbb{C}\) to all of its \(r\)-th roots, \(r \in \mathbb{N}^+\). \end{conj} \section{Sunflower Hypergraphs} \label{sect:sunflowers} In previous work \cite{Cooper_Dutle_2012}, the authors computed the (normalized) spectrum of the adjacency hypermatrix of some classes of hypergraphs. Other than (possibly partial) matchings, the techniques used therein disallowed computation of the multiplicities of eigenvalues for any infinite class. Here, however, we determine the characteristic polynomial for an infinite class of \emph{sunflower} hypergraphs. A \emph{sunflower} \( \mathcal{S}(m,q,k) \) hypergraph is defined as follows. Let \( m>0 \), and \( q, k \) satisfy \( 0<q<k. \) Let \( S \) be a set of \( q \) vertices (``seeds'') and define \( m \) disjoint sets \( \{E_i\}_{i=1}^m \) of \( k-q \) vertices each (``petals''). The edges of the hypergraph are the sets \( S\cup E_i \) for \( 1\leq i\leq m. \) When \( k=2 \), sunflower graphs are normally referred to as \emph{stars.} In general, when \( q=k-1 \), \( \mathcal{S}(m,k-1, k) \) is a complete \( k \)-cylinder (i.e., \( k \)-partite, \( k \)-uniform hypergraph) with block sizes \( m \) and \( 1^{k-1} \), which the authors considered in \cite{Cooper_Dutle_2012}. By using the product formula, we determine the spectrum, including multiplicities, for single-seed sunflowers of uniformity \( 3 \). \begin{theorem} \label{sunflower_spec_mult} The characteristic polynomial for \( \mathcal{S}(n,1,3) \) is \[\phi_{\mathcal{S}(n,1,3)}(\lambda) = \lambda^{(2n-2)4^n} \prod_{r=0}^n (\lambda^3-r)^{\binom{n}{r}3^r}. \] \end{theorem} \begin{proof} Label the \( 2n+1 \) vertices of \( \mathcal{S}(n,1,3) \) as \( v_0 \) for the seed, and pairs of vertices \( \{v_{i,1}, v_{i,2}\} \) for \( i \in [n]. \) Then the equations defining the eigenpairs for our sunflower are the following: \[F_0 = \lambda x_0 - \sum_{i=1}^n x_{i,1}x_{i,2}\] and for each \( i\in [n], \) the pair of equations \[\left\{ \begin{array}{c} F_{i,1} = \lambda x_{i,1}^2- x_0x_{i,2} \\ F_{i,2} = \lambda x_{i,2}^2- x_0x_{i,1} \end{array} \right\}. \] Consider \( \Res( \overline{F_{1,1}}, \overline{F_{1,2}}, \ldots, \overline{F_{n,1}}, \overline{F_{n,1}}) \). Note that \( \overline{F_{i,j}} = \lambda x_{i,j}^2 \) for every \( i\in [n] \) and all \( j\in [2], \) which is the set of equations used to define the characteristic polynomial of the hypermatrix with all zero entries. This clearly has 0 as its only eigenvalue, and so \[ \Res( \overline{F_{1,1}}, \overline{F_{1,2}}, \ldots, \overline{F_{n,1}}, \overline{F_{n,1}}) = \lambda^{2n2^{2n-1}}.\] As mentioned in the introduction, this is a non-zero polynomial in \( \lambda \), and so the product formula applies. This formula gives that \begin{align} \phi_{\mathcal{S}(n,1,3)}(\lambda) & = \Res(F_0, F_{1,1}, F_{1,2}, \ldots, F_{n,1}, F_{n,1}) \nonumber \\ & = \Res( \overline{F_{1,1}}, \overline{F_{1,2}}, \ldots, \overline{F_{n,1}}, \overline{F_{n,1}})^2 \prod_{p\in V} f_0(p)^{m(p)} \nonumber \\ \label{sunfcalc} & = \lambda^{2n2^{2n}}\prod_{p\in V} f_0(p)^{m(p)}. \end{align} We next determine the points of the variety \( V \). Fix \( i\in [n] \), and consider the pair of polynomials \[\left\{ \begin{array}{c} f_{i,1} = x_{i,2} -\lambda x_{i,1}^2 \\ f_{i,2} = x_{i,1} -\lambda x_{i,2}^2. \end{array} \right\} \] Let \( \zeta_3 \) denote a primitive third root of unity, and choose \( x_{i,1}\in \{ 0, \frac{1}{\lambda}, \frac{\zeta_3}{\lambda},\frac{\zeta_3^2}{\lambda}\} \). It's easy to verify that this value for \( x_{i,1} \) can be uniquely extended to a zero for our two equations \( f_{i,1}, f_{i,2} \) by setting \( x_{i,2} = \lambda x_{i,1}^2 \). It's also easy to check that \[x_{i,1}x_{i,2} = \begin{cases} 0 &\text{ if } x_{i,1}=0 \\ \frac{1}{\lambda^2} & \text{ otherwise. } \end{cases}\] A point of our variety must be a zero for each of the \( n \) pairs of polynomials \( \{f_{i,1}, f_{i,2}\} \). Because the variables appearing in any particular pair don't appear in any of the other polynomials describing the variety, a solution can be found by combining solutions for the pairs. We saw above that we can find a unique solution for a pair of these polynomials by choosing \( x_{i,1} \) from one of the four values \( 0, \frac{1}{\lambda}, \frac{\zeta_3}{\lambda},\frac{\zeta_3^2}{\lambda} \). Hence we can build \( 4^n = 2^{2n} \) distinct zeros, which is the most we can have in a zero dimensional variety defined by \( 2n \) quadratics. Hence all points of the variety are given by the above construction, and the mulitplicity of each point is exactly one. Next, we note that \( f_0(\mathbf{x}) = \lambda - \sum_{i=1}^n x_{i,1}x_{i,2}. \) As we saw above, for any point \( p\in V \), each term in this sum is either 0 or \( 1/\lambda^2 \). If we let \( r = |\{ i \mid p_{i,1} \neq 0\}| \) denote the number of nonzero pairs of coordinates, we see that \[f_0(p) = \lambda-\frac{r}{\lambda^2} = \frac{\lambda^3-r}{\lambda^2}.\] For a fixed \( r \), there are \( \binom{n}{r}3^r \) such points \( p\in V. \) Thus we can compute the second factor for our resultant as \begin{align*} \prod_{p\in V} f_0(p)^{m(p)} & = \prod_{r=0}^n\left(\frac{\lambda^3-r}{\lambda^2}\right)^{\binom{n}{r}3^r} \\ & = \lambda^{-2\cdot 4^n} \prod_{r=0}^n \left(\lambda^3-r\right)^{\binom{n}{r}3^r}.\end{align*} Substituting this expression into \eqref{sunfcalc} gives the claimed result. \end{proof} Unfortunately, we were unable to extend the result to \( \mathcal{S}(n,q,k) \) sunflowers. Following the technique above first requires that \( q=1. \) If we attempt to use an arbitrary \( k, \) we find that there are \( k^{k-2}+1 \) distinct solutions for the \( k-1 \) equations arising from a particular petal. Hence there are \( (k^{k-2}+1)^n \) solutions, whereas we need \( \left((k-1)^{k-1}\right)^n \) distinct solutions to use the product formula. These two values agree precisely when \( k=3 \), making the computation above possible.

https://arxiv.org/abs/2111.04339

Sharp Sobolev regularity of restricted X-ray transforms

We study $L^p$-Sobolev regularity estimate for the restricted X-ray transforms generated by nondegenerate curves. Making use of the inductive strategy in the recent work by the authors, we establish the sharp $L^p$-regularity estimates for the restricted X-ray transforms in $\mathbb R^{d+1}$, $d\ge 3$. This extends the result due to Pramanik and Seeger in $\mathbb R^3$ to every dimension.

\section{Introduction} Let $\gamma$ be a smooth curve from $I=[-1,1]$ to $\mathbb R^{d}$. We consider \[ \mathfrak R f(x,s)=\psi(s)\int f(x+t\gamma(s),t)\chi(t)dt, \quad f\in \mathcal S(\mathbb{R}^{d+1}), \] where $\psi$ and $\chi$ are smooth functions supported in the interior of the intervals $I$ and $[1,2]$, respectively. The operator $\mathfrak Rf$ is referred to as the restriction of X-ray transform to the line complex generated in the direction $(\gamma,1)$. We say $\gamma$ is nondegenerate if \begin{align}\label{nonv} \det( \gamma'(s), \dots, \gamma^{(d)}(s) )\neq0, \quad s\in I. \end{align} The operator $\mathfrak Rf$ is a model case of general class of restricted X-ray transforms (see \cite{GG, GS00, GU1, GU2, GU3}). Especially in $\mathbb R^3$, under the nondegeneracy assumption \eqref{nonv}, $\mathfrak Rf$ is a typical example of Fourier integral operators with one sided fold singularities (\cite{GS94}). Regularity properties of $\mathfrak Rf$ have been studied in terms of $L^p$ improving and $L^p$ Sobolev regularity estimates. $L^p$ improving property of $\mathfrak R$ is well understood by now (see \cite{GS98, GSW, Ob, La}). In fact, the problem was considered in a more general framework: $L^p$--$L_s^q(L_x^r)$ mixed norm estimates for $\mathfrak R$ were studied by various authors (see \cite{W, E, CE1}), so the optimal estimates were established except for some endpoint cases. (See also \cite{CE2, TW, G, DS1, DS2} for related results.) The $L^2$--$L_{1/(2d)}^2$ estimate is easy to obtain via $TT^*$ argument and van der Corput's lemma (\cite{GS00}) (see also \cite{GU3, GS94} for the sharp $L^2$ Sobolev estimates for general class of operators). Interpolation between this estimate and the trivial $L^\infty$ estimate gives the sharp $L^p$--$L_{1/(pd)}^p$ estimate for $p\ge 2$, which is optimal in that $L^p$--$L_{\alpha}^p$ fails if $\alpha < 1/(pd)$ (see Proposition \ref{nece} below). However, when $p< 2$, the sharp $L^p$ regularity estimate is less straightforward. Such estimate was not known until recently. When $d=2$, the optimal $L^p-L_{1-1/p}^p$ estimate was established for $1< p<4/3$ by Pramanik and Seeger's conditional result \cite{PS06} and the sharp decoupling inequality for the cone (in $\mathbb R^3$) due to Bourgain and Demeter \cite{BD}. Via interpolation those estimates give the sharp $L^p$--$L_{1/(2d+)}^p$ estimate for $4/3\le p<2$ but it remains open whether the endpoint $L^p$--$L_{1/(2d)}^p$ estimate holds for $4/3\le p<2$. (See Conjecture \ref{sobolev} below.) In $\mathbb R^3$, the result has been extended to more general operators. Pramanik and Seeger \cite{PS20} obtained sharp $L^p$ regularity estimates for Fourier integral operator with folding canonical relations. Bentsen \cite{Bentsen2} (also see \cite{Bentsen}) extended their result to a class of radon transforms with fold and blowdown singularities. \medskip However, in higher dimensions ($d\ge 3$) the sharp $L^p$ regularity estimate for $\mathfrak R$ has been left open for $1<p<2$. We set $p_d=\frac{2d}{2d-1}$ and \[ \alpha(p)= \begin{cases} 1-\frac 1p, & \ 1\le p< p_d, \\ \ \frac 1{2d}, & \ p_d \le p\le 2. \end{cases} \] It is natural to conjecture the following. \begin{conj} \label{sobolev} Let $d\ge3$ and $1< p < 2$. Suppose $\gamma$ is a smooth nondegenerate curve. Then, $\mathfrak R$ boundedly maps $L^p$ to $L_\alpha^p$ for $\alpha \le \alpha(p)$. \end{conj} \noindent Failure of $L^p$--$L_\alpha^p$ boundedness for $\alpha> \alpha(p)$ can be shown by a slight modification of the examples in \cite{PS06}. See Proposition \ref{nece}. The following is our main result which verifies the conjecture except some endpoint cases. \begin{thm} \label{thm:improving} Let $d\ge3$. Suppose $\gamma$ is nondegenerate. Then, for $1\le p<p_d$, \begin{equation} \label{A1} \|\mathfrak R f\|_{L^p_{1-1/p}(\mathbb R^{d+1})}\le C\|f\|_{p}. \end{equation} \end{thm} Interpolation with $L^2-L_{1/(2d)}^2$ estimate yields \eqref{A1} for $\alpha <\alpha(p)$ when $p_d \in [p_d,2)$. By a standard scaling argument (\cite{PS06, PS07}) the result in Theorem \ref{thm:improving} can be extended to the curves of finite type. We say that a curve $\gamma: I \mapsto \mathbb R^{d}$ is of finite type if there is an $L$ such that $\sspan \{ \gamma^{(1)}(s), \dots, \gamma^{(L)}(s)\}=\mathbb R^d$ for each $s\in I$. For the finite type curve, the smallest of such $L$ is defined to be the type of $\gamma$ at $s$. The supremum of the type over all $s \in I$ is called the maximal type of $\gamma$ (see, e.g., \cite{PS07, HL}). \begin{cor}\label{type} Let $d\ge3$, $1<p<2$, and $L>d$. Suppose $\gamma$ is a curve of maximal type $L$. Then, $\mathfrak Rf$ is bounded from $L^p(\mathbb R^{d+1})$ to $L_\alpha^p(\mathbb R^{d+1})$ if $\alpha \le \min( \alpha(p), 1/(Lp))$. \end{cor} For $p\in [2,\infty]$ it is easy to obtain the sharp $L^p$ to $L_{1/(Lp)}^p$ estimate, which can be shown by using $L^2$ to $L_{1/(2L)}^2$ boundedness and interpolation in a similar manner as above. Corollary \ref{type} and Proposition \ref{nece} settle the problem of the optimal Sobolev regularity estimate for $\mathfrak R$ if $L\ge 2d$. However, some endpoint cases remain left open when $d<L\le 2d-1$, not to mention such estimates for the nondegenerate curve. \medskip In order to prove Theorem \ref{thm:improving} we adapt the inductive strategy developed by the authors in \cite{KLO} to obtain the sharp $L^p$ regularity estimates for the convolution operator \[ \mathcal A_t g(x) =\int g(x-t\gamma(s)) \psi(s)\,ds, \quad t\neq 0, \] for $ g\in \mathcal S(\mathbb{R}^{d})$. It is known (see \cite{PS07, BGHS}) that $\mathcal A_t$ maps $L^p(\mathbb R^{d})$ boundedly into $L^p_{\alpha}(\mathbb R^{d})$ only if $\alpha\le \widetilde\alpha (p)$ for $2\le p\le \infty$ where \[ \widetilde\alpha (p)= \begin{cases} \frac 1d(\frac 12+\frac 1p), & \ 2 \le p \le 2(d-1), \\[3pt] \ \ \frac 1p, & \ 2(d-1)< p \le \infty. \end{cases} \] When $d=3$, the sharp $L^p$--$L_{1/p}^p$ estimate of $\mathcal A_t$ for $p>4$ is due to Pramanik and Seeger \cite{PS07} and the aforementioned decoupling inequality \cite{BD} (see also \cite{OS, OSS, TV} for earlier results). The result was extended to $d=4$ by the recent work of Beltran, Guo, Hickman, and Seeger \cite{BGHS}. When $d\ge5$, the sharp $L^p-L_{\widetilde \alpha(p)-} ^p$ boundedness was recently established in \cite{KLO} (see also \cite{PS07, KLO2, BGHS2} for recent progress in related problems). A modification of the argument in \cite{KLO} yields the optimal $L^p$--$L_{1/p}^p$ estimate for $p >2(d-1)$. See Remark \ref{optimal}. Since its proof is almost identical to that of Theorem \ref{thm:improving}, we state the result without proof. \begin{thm}\label{averaging} Let $d\ge5$ and $p>2d-2$. Suppose $\gamma$ is a smooth nondegenerate curve. Then, $\mathcal A_t$ maps $L^p(\mathbb R^{d})$ boundedly into $L^p_{1/p}(\mathbb R^{d})$. \end{thm} For nonnegative quantities $A$ and $ D$, we denote $A\lesssim D$ if there exists a (independent) constant $C$ such that $A\le CD$, however the constant $C$ may differ at each occurrence depending on the context. \section{Estimates with localized frequency} In this section, we reduce the proof of Theorem \ref{thm:improving} to obtaining Proposition \ref{microlocal} which we prove at the end of this section. In order to prove Theorem \ref{thm:improving}, we consider the dual operator $\mathcal R:=\mathfrak R^*$, which is given by \[\mathcal R f(x,t)=\chi(t)\int f(x-t\gamma(s),s)\psi(s)ds.\] By duality it suffices to show \begin{align}\label{dual2} \| \mathcal Rf\|_{L^p(\mathbb R^{d+1})} \lesssim \|f\|_{L_{-1/p}^p}, \quad 2d<p<\infty. \end{align} For the purpose we closely follow the lines of argument in \cite{KLO}. For a smooth function $a(s,t,\xi)$ defined on $I \times [1,2] \times \mathbb R^d$, we define an integral operator by \[ \mathcal R[a]f(x,t)=(2\pi)^{-d}\iint e^{i(x-t\gamma(s))\cdot\xi}a(s,t,\xi) \,\mathcal F_x f (\xi,s)ds d\xi. \] Here $\mathcal F_x$ denotes the Fourier transform in $x$. Note that $\mathcal Rf =\mathcal R[a]f$ if $a(s,t,\xi)=\psi(s)\chi(t)$. We assume that \begin{equation} \label{curveB} \sum_{j=0}^{3d+1}|\gamma^{(j)}(s)|\,\le\, 10^{-2} B, \quad s\in I \end{equation} for some constant $B\ge1$. We obtain the estimate \eqref{dual2} in an inductive manner under the localized nondegeneracy assumption: \begin{align} \label{sumN} \sum_{\ell=1}^L & |\langle \gamma^{(\ell)}(s), \xi\rangle| \ge B^{-1}|\xi|. \end{align} When $L<d$, \eqref{sumN} can not be true in general even if $\gamma$ is nondegenerate. However, an appropriate decomposition in frequency domain makes it possible that \eqref{sumN} holds if $(s,t,\xi) \in \supp a$ for some $t$. For an integer $1 \le L \le d$, by $ \Vol (v_1,\dots,v_L)$ we denote the $L$-dimensional volume of the parallelepiped generated by vectors $v_1,\dots, v_L \in \mathbb R^d$. We denote \[ \mathbb A_k=\{ \xi \in \mathbb R^d: 2^{k-1} \le |\xi| \le 2^{k+1}\}, \quad k\ge 0. \] \begin{thm}\label{lclb} Let $1\le L \le d$, $B\ge1$, and $\gamma\in \mathrm C^{3d+1}(I)$ satisfy \eqref{curveB}. Suppose $a(s,t,\xi)$ is a smooth function supported in $I \times [1,2]\times \mathbb A_k$ such that \[ |\partial_t^{j}\partial_\xi^\alpha a(s,t,\xi)|\le B |\xi|^{-|\alpha|} \] for $0\le j\le 2L$ and $|\alpha|\le d+L+2$. Suppose \begin{equation} \label{lindepN} \Vol \big( \gamma^{(1)}(s), \dots, \gamma^{(L)}(s) \big) \ge B^{-1}, \quad s\in I, \end{equation} and suppose \eqref{sumN} holds whenever $(s,t,\xi) \in \supp a$ for some $t$. Then, for $p> 2L$ \begin{align}\label{smoothing} \| \mathcal R[a]f\|_{L^p(\mathbb R^{d+1})}\le C 2^{-\frac{k}p}\|f\|_p. \end{align} \end{thm} Once we have Theorem \ref{lclb}, the estimate \eqref{dual2} is rather easy to show. Let $\beta \in \mathrm C_0^\infty((1/2,2))$ such that $\sum_{k}\beta(2^{-k} s)=1$ for $s>0$ and $\beta_0\in \mathrm C_0^\infty([-1,1])$ such that $\beta_0= 1$ on $[-1/2,1/2]$. We define $f_k$ by \[\widehat{f_k}(\xi,\tau)=\beta(2^{-k}|\xi|) \beta_0(r_1^{-1}2^{-k}|\tau|)\widehat f(\xi,\tau),\] where $r_1=1+\sup_{s \in \supp \psi}|\gamma'(s)|$. Then \[\mathcal F_x(\mathcal R(f_k-f))(\xi,t)=(2\pi)^{-1}\chi(t)\iint e^{i(s\tau-t\gamma(s)\cdot \xi)}\psi(s) ds (\widehat{f_k}- \widehat{f}\,)(\xi,\tau)d\tau.\] One can easily see that $\| \mathcal Rf-\mathcal Rf_k \|_{L^2(\mathbb R^{d+1})}\lesssim 2^{-kN}\|f\|_2$ for any $N>0$ if $\widehat f$ is supported on $\mathbb A_k$.\footnote{By integration by parts, we have $\int e^{i(s \tau-\gamma(s)\cdot \xi)}\psi(s)ds=O((1+|\tau|)^{-N})$ for $|\tau| \ge r_1|\xi|$.} Since $\gamma$ is nondegenerate, using Theorem \ref{lclb} with $L=d$ and $a=\psi(s)\chi(t)\beta(2^{-k}|\xi|)$, we get \begin{align}\label{one} \| \mathcal R f_k\|_{L^p(\mathbb R^{d+1})} \lesssim 2^{-\frac kp} \|f\|_p \end{align} for $p>2d$. Then, making use of the vector-valued version of the Fefferman-Stein inequality for the $\#$-function, one can obtain \eqref{dual2} from \eqref{one}. This can be done in the same manner as in \cite{PS06} (also see \cite{PS07, BGHS}). So, we omit the details. \medskip We prove Theorem \ref{lclb} by induction on $L$. When $L=1$, the estimate \eqref{smoothing} is easy to prove by $TT^*$ argument. Indeed, let us set $\widetilde{\mathcal R}f=\mathcal F_x(\mathcal R[a] \mathcal F_x^{-1}\!f)$ so that \[ \widetilde{\mathcal R}f(\xi,t) = \int e^{-it\gamma(s)\cdot \xi} a(s,t,\xi) f(\xi,s)\,ds. \] Then, we have $\widetilde{\mathcal R}^*\widetilde{\mathcal R}f(\xi,s)=\int \mathcal K(s,s',\xi)f(\xi,s')\,ds'$ where $\mathcal K$ is given by $\mathcal K(s,s',\xi)=\int e^{-it(\gamma(s)-\gamma(s'))\cdot \xi} \overline a(s,t,\xi) a(s',t,\xi)\,dt$. Since \eqref{sumN} holds with $L=1$ on the support of $a$, by integration by parts we have \[ |\mathcal K(s,s',\xi)| \le C (1+2^k|s-s'|)^{-2}. \] Thus, $|\widetilde{\mathcal R}^* \widetilde{\mathcal R}f(\xi,s)|\le C \int (1+2^k|s-s'|)^{-2} |f(\xi,s')|\,ds'$. By Young's convolution inequality we get $\|\widetilde{\mathcal R}^* \widetilde{\mathcal R}f \|_2 \lesssim 2^{-k}\|f\|_2$. Via Plancherel's theorem this gives $\| \mathcal R[a]f\|_2 \lesssim 2^{-k/2}\|f\|_2$. Interpolating this with the trivial $L^\infty$ estimate, we obtain Theorem \ref{lclb} for $L=1$. Theorem \ref{lclb} with $L\ge2$ follows from the next proposition. \begin{prop}\label{newprop} Let $2 \le N \le d$. Suppose Theorem \ref{lclb} holds with $L=N-1$. Then Theorem \ref{lclb} holds with $L=N$. \end{prop} We now fix $2\le N\le d$. For the rest of the paper we assume \begin{align*} 2^{-k /N}\le \delta \le 2^{-3dN}B^{-N}. \end{align*} \subsection{A class of symbols adapted to $\gamma$} We define a class of symbols adapted to the curve $\gamma$. For $\gamma$ satisfying \eqref{lindepN} with $L=N-1$ and $s\in I$, we define a linear map $\widetilde {\mathcal L}_s^\delta :\mathbb R^d \mapsto \mathbb R^d$ by \begin{align*} (\widetilde{\mathcal L}_{s}^\delta)^\intercal \gamma^{(j)}(s) &=\delta^{N-j}\gamma^{(j)}(s), \qquad j=1,\dots,N-1, \\ (\widetilde{\mathcal L}_{s}^\delta)^\intercal v &=v. \qquad \qquad \quad~ v\in \big(\text{span} \big\{ \gamma^{(j)}(s) : j=1,\dots,N-1 \big\}\big)^{\perp}. \end{align*} Let $(\tau,\xi)\in \mathbb R \times \mathbb R^d$. We also set a linear map $\mathcal L_s^\delta: \mathbb R^{d+1} \mapsto \mathbb R^{d+1}$ by \[ \mathcal L_{s}^\delta(\tau, \xi)= \big(\delta^N \tau - \gamma(s)\cdot \widetilde{\mathcal L}_s^\delta \xi,\, \widetilde{\mathcal L}_s^\delta \xi\big). \] We set \[ G(s)=(1,\gamma(s))\] and define a class of symbols which are adapted to subcurves of $G(s)$ with length about $\delta$. \begin{defn} Let $\gamma$ be a smooth curve satisfying \eqref{curveB} and \eqref{lindepN} with $L=N-1$. Let $s_{\!\,\mathsmaller{0}} \in (-1,1)$ and $0<\delta \le1$ such that $[s_{\!\,\mathsmaller{0}}-\delta, s_{\!\,\mathsmaller{0}}+\delta]\subset I$. Then, we set \begin{align*} \Lambda_k(\delta,s_{\!\,\mathsmaller{0}})=\big\{(\tau,\xi) \in \mathbb R \times \mathbb A_k : &\,|\langle G^{(j)}(s_{\!\,\mathsmaller{0}}),(\tau,\xi)\rangle| \le B2^{k+5} \delta^{N-j}, \, j=0,\dots N-1\big\}. \end{align*} We denote by $\mathfrak A_k(\delta,s_{\!\,\mathsmaller{0}})=\mathfrak A_k(\delta, s_{\!\,\mathsmaller{0}},d,N,B,\gamma)$ the set of smooth functions $\mathfrak a$ which satisfy the following\,$:$ \begin{align}\label{symsupp} \supp \mathfrak a &(s,t,\tau,\xi) \subset [s_{\!\,\mathsmaller{0}}-\delta,s_{\!\,\mathsmaller{0}}+\delta]\times [1,2]\times \Lambda_k(\delta,s_{\!\,\mathsmaller{0}}), \end{align} and \begin{align} \label{symineq2} &\big|\partial^{j}_t\partial^{\alpha}_{\tau, \xi} \mathfrak a\big( s,t, \mathcal L_{s_{\!\,\mathsmaller{0}}}^\delta(\tau,\xi) \big) \big| \le B|(\tau,\xi)|^{-|\alpha|} \end{align} for $0\le j\le 2N$ and $|\alpha|\le d+N+2$. \end{defn} We set $\supp_{\xi} \mathfrak a=\cup_{s,t,\tau} \supp \mathfrak a(s,t,\tau,\cdot)$ and $\supp_{s,\xi} \mathfrak a=\cup_{t,\tau} \supp \mathfrak a(\cdot,t,\tau,\cdot)$. Also $\supp_{\tau,\xi} \mathfrak a$ and $\supp_{s} \mathfrak a$ are defined in the same manner. We say \emph{a statement $S(s, \xi)$, depending on $(s, \xi)$, holds on $\supp \mathfrak a$ $($the support of $\mathfrak a$$)$ if $S(s,\xi)$ holds for $(s,\xi)\in \supp_{s,\xi} \mathfrak a$}, i.e., $(s,t, \tau,\xi) \in \supp \mathfrak a$ for some $t$ and $\tau$. We also adopt the same convention with other variables. We define \begin{align}\label{T} \mathcal F(\mathcal T [\mathfrak a] f)(\xi,\tau) =\iint e^{-it'(\tau+ \gamma(s)\cdot \xi)}\mathfrak a(s,t',\tau,\xi)dt' \, \mathcal F_x f (\xi,s)ds. \end{align} Clearly, we have $\mathcal R[a]=\mathcal T[a]$ for $a=a(s,t,\xi)$ by the Fourier inversion formula. The following is an analogue of \cite[Lemma 2.6]{KLO} which is obtained by estimating the kernel of $\mathcal T[\mathfrak a]$. \begin{lem}\label{kernel} Let $\widetilde \chi \in \mathrm C_0^\infty((2^{-2},2^2))$ such that $\widetilde \chi=1$ on $[3^{-1},3]$. Let $\mathfrak a$ be a smooth function which satisfies \eqref{symsupp} and \eqref{symineq2} with $j=0$ and $|\alpha| \le d+3$. Then, we have \begin{equation}\label{ker-est} \| \mathcal T[\mathfrak a]f\|_{L^\infty(\mathbb R^{d+1})} \le C\delta \|f\|_\infty, \end{equation} and \begin{align}\label{ker-est2} \|(1-\widetilde\chi(t)) \mathcal T[\mathfrak a]f\|_{L^p(\mathbb R^{d+1})} \le C2^{-k}\delta^{1-\frac 1p-N} \|f\|_p \end{align} for $p\ge2$. \end{lem} \begin{proof} Note that $\mathcal T[\mathfrak a]f (x,t) = \int K[\mathfrak a](s,t, \cdot)\ast f(\cdot,s)(x) \,ds$ where \[ K[\mathfrak a](s,t,x)=C_d \iiint e^{i(t-t')\tau+i(x-t\gamma(s))\cdot \xi}\mathfrak a(s,t',\tau,\xi) \, d\xi d\tau dt' \] for $C_d=(2\pi)^{-d-1}$. Since \eqref{symineq2} holds with $j=0$ and $|\alpha| \le d+3$, by changing variables $(\tau,\xi) \rightarrow 2^k\mathcal L_{s_{\!\,\mathsmaller{0}}}^\delta(\tau,\xi)$ followed by repeated integration by parts, we have \[|K[\mathfrak a](s,t,x)|\lesssim \delta^{\frac {N(N+1)}2}2^{k(d+1)}\int\big(1+2^k|(\delta^N(t-t'),(\widetilde{\mathcal L}_{s_{\!\,\mathsmaller{0}}}^\delta)^\intercal(x-t\gamma(s)))|\big)^{-d-3}dt'. \] This yields $\|K[\mathfrak a](s,t,\cdot)\|_{L^1_x}\lesssim 1$. From \eqref{symsupp} we have $\mathcal T[\mathfrak a]f (x,t) = \int_{s_{\!\,\mathsmaller{0}}-\delta}^{s_{\!\,\mathsmaller{0}}+\delta} K[\mathfrak a](s,t, \cdot)\ast f(\cdot,s)(x) \,ds$. Thus, \eqref{ker-est} follows. By \eqref{symsupp} and H\"older's inequality we see \begin{align*} \|(1-\widetilde\chi)\mathcal T[\mathfrak a]f\|_p^p\lesssim \delta^{p-1} \int_1^2 \int_{s_{\!\,\mathsmaller{0}}-\delta}^{s_{\!\,\mathsmaller{0}}+\delta} \|(1-\widetilde\chi(t))K[\mathfrak a](s,t, \cdot)\ast f(\cdot,s)\|_{L^p_x}^pdsdt. \end{align*} Using the previous estimate for $K[\mathfrak a](s,t,x)$, we have $\|(1-\widetilde\chi(t))K[\mathfrak a](s,t,\cdot)\|_{L^1_x}\lesssim 2^{-k} \delta^{-N} |t-1|^{-1}(1-\widetilde\chi(t))$. This and Young's convolution inequality give \eqref{ker-est2}. \end{proof} \subsection{Frequency localization} To prove Proposition \ref{newprop} we make some preliminary decompositions and reductions. Recall that $\beta_0\in \mathrm C_0^\infty([-1,1])$ such that $\beta_0= 1$ on $[-1/2,1/2]$. For a fixed constant $\delta_{\!\,\mathsmaller{0}}$ which is to be chosen later, we set \[ a_N(s,t,\xi)=a(s,t,\xi)\prod_{j=1}^{N-1} \beta_0 \Big(100dB\delta_{\!\,\mathsmaller{0}}^{-N}\frac{\gamma^{(j)}(s)\cdot\xi}{|\xi|} \Big). \] Suppose \eqref{sumN} holds with $L=N$. Then, we have \begin{align}\label{lowerN} |\gamma^{(N)}(s)\cdot\xi|\ge (2B)^{-1}|\xi| \end{align} on the support of $a_N$ since $|\gamma^{(j)}(s)\cdot\xi|\le (100dB)^{-1}\delta_{\!\,\mathsmaller{0}}^N|\xi|$ for $1\le j\le N-1$. Clearly, \eqref{sumN} holds with $L=N-1$ and $B$ replaced with $(100dB)^{-1}\delta_{\!\,\mathsmaller{0}}^N$ on $\supp (a-a_N)$. Thus, by the induction hypothesis (Theorem \ref{lclb} with $L=N-1$ and $B$ replaced by $(100dB)^{-1}\delta_{\!\,\mathsmaller{0}}^N$) we have \[ \|\mathcal R[a-a_N]f\|_p\lesssim 2^{-\frac kp}\|f\|_p \] for $p> 2(N-1)$. So, we need only to consider $\mathcal R[a_N]$ instead of $\mathcal R[a]$. We set \[ \Gamma_k(a_N,h)=\big\{ \xi \in \mathbb A_k: \dist( |\xi|^{-1}\xi, |\xi'|^{-1} \xi') \le h, \quad~ \text{for some } \xi' \in \supp_\xi a_N \big\}. \] By finite decomposition of the support of $a_N$, we may assume that \eqref{lowerN} holds for $s \in [-1-\delta_{\!\,\mathsmaller{0}},1+\delta_{\!\,\mathsmaller{0}}]$ and $\xi \in \Gamma_k(a,\delta_{\!\,\mathsmaller{0}})$ for a small constant $\delta_{\!\,\mathsmaller{0}}>0$, which we specify later. Thus, by the implicit function theorem, there is a $\mathrm C^{2d+2}$ function $\sigma(\xi)$ such that \[ \langle\gamma^{(N-1)}(\sigma(\xi)),\xi\rangle=0. \] In what follows, any symbol which subsequently appears is given by decomposing the symbol $a_N$. Consequently, we may assume the symbols are supported in $\Gamma_k(a_N,\delta_{\!\,\mathsmaller{0}})$\footnote{Thus, $\sigma$ is a well defined on the $\xi$-supports of symbols.} and \eqref{lowerN} holds on their supports. Furthermore, by finite decomposition we may assume \[\supp_s a_N \subset [s_{\!\,\mathsmaller{0}}-\delta_{\!\,\mathsmaller{0}}, s_{\!\,\mathsmaller{0}}+\delta_{\!\,\mathsmaller{0}}]\] for some $s_{\!\,\mathsmaller{0}} \in (-1,1)$. It is not difficult to see that the contribution of the frequency part $\{(\tau,\xi): |\tau+\gamma(s)\cdot \xi| \gtrsim 2^{k+1} \delta_{\!\,\mathsmaller{0}}^N, \, \forall s \in I\}$ is not significant. To see this, let us set \[ {\mathfrak a}_0(s,t,\tau,\xi)= a_N(s,t,\xi) \beta_0 \big( \delta_{\!\,\mathsmaller{0}}^{-2N} 2^{-2k}|\tau+\gamma(s)\cdot \xi |^2\big) \] and $\mathfrak a_1(s,t,\tau,\xi)=\mathfrak a_0(s,t,\tau,\xi)-a_N(s,t,\xi)$. By Fourier inversion, we have \[ \mathcal R[a_N]f = \mathcal T[\mathfrak a_0]f +\mathcal T[\mathfrak a_1]f. \] For the operator $\mathcal T[\mathfrak a_1]$, we have a better estimate. Note that $|\tau+\gamma(s)\cdot\xi| \gtrsim 2^{k} \delta_{\!\,\mathsmaller{0}}^N$ on $\supp \mathfrak a_1$. Integration by parts in $t$ gives \begin{align*} \mathcal F(\mathcal T [\mathfrak a_1] f)(\xi,\tau) =\iint e^{-it'(\tau+ \gamma(s)\cdot \xi)} \frac{ \partial_{t'} \mathfrak a_1(s,t',\tau,\xi)}{i(\tau+ \langle \gamma(s), \xi \rangle)}dt' \, \mathcal F_x f (\xi,s)ds. \end{align*} Since $\mathfrak a=2^{k}\delta_{\!\,\mathsmaller{0}}^{N}(\tau+\langle \gamma(s), \xi \rangle )^{-1} \partial_{t'}\mathfrak a_1$ satisfies \eqref{symsupp} and \eqref{symineq2} with $\delta=\delta_{\!\,\mathsmaller{0}}$ and $B=C_1\delta_{\!\,\mathsmaller{0}}^{-C}$ for some $C,C_1$ when $j=0$ and $|\alpha| \le d+3$, Lemma \ref{kernel} gives \[ \|\mathcal T [ \mathfrak a_1]f\|_p\lesssim 2^{-k}\|f\|_p, \quad p\ge 2. \] Thus, disregarding $\mathcal T [ \mathfrak a_1]$, we need only to show $\| \mathcal T [ \mathfrak a_0]f\|_{L^p(\mathbb R^{d+1})}\le C 2^{-\frac{k}p}\|f\|_p$ instead of \eqref{smoothing}. Let us fix a constant \begin{equation}\label{delta-min} \delta_{\!\,\mathsmaller{0}}=\min\{ \delta_*, 2^{-3dN}B^{-N} \} \end{equation} where $\delta_*$ is a small number to be chosen later (see \eqref{delta-def} below). Since $\delta_{\!\,\mathsmaller{0}}$ is the fixed constant, it is clear that $C^{-1} \mathfrak a_\nu \in \mathfrak A_k(\delta_{\!\,\mathsmaller{0}},s_{\!\,\mathsmaller{0}})$ for a large constant $C>0$. Therefore, Proposition \ref{newprop} follows if we show the next proposition. \begin{prop}\label{microlocal} Let $ \mathfrak a\in \mathfrak A_k(\delta_{\!\,\mathsmaller{0}},{s_{\!\,\mathsmaller{0}}})$. Suppose $\gamma$ satisfies \eqref{lindepN} with $L=N$ and suppose Theorem \ref{lclb} holds with $L=N-1$. Then, for $p> 2N$ we have \begin{align}\label{overline-a0} \big\| \mathcal T[\mathfrak a]f \big\|_{p} \le C 2^{-\frac{k} p}\|f\|_{p}. \end{align} \end{prop} \subsection{Decoupling of $\mathcal T[\mathfrak a^\mu]$} Let $\delta_0$ and $\delta_1$ be positive numbers such that \begin{align}\label{orderB} 2^{3d}B\delta_0^{(N+1)/N} \le \delta_1\le \delta_0 \le \delta_{\!\,\mathsmaller{0}}. \end{align} For $n\ge 0$, we set \begin{equation}\label{Jnmu} \mathfrak J_n^\mu=\{ \nu \in \mathbb Z: |2^n\delta_1 \nu - \delta_0 \mu| \le \delta_0\}. \end{equation} We prove Proposition \ref{microlocal} by using the following proposition iteratively (cf. \cite[Proposition 4.1]{KLO}), which we show in Section \ref{sec3}. \begin{prop}\label{iterative} Let $\delta_0$ and $\delta_1$ be positive numbers satisfying \eqref{orderB} and $2^{-k/N}\le \delta_1$. Let $\mathfrak a^\mu \in \mathfrak A_k(\delta_0,\delta_0\mu)$ for $\delta_0\mu \in [-1,1]\cap \delta_0\mathbb Z$ such that $\supp_{\xi} \mathfrak a^\mu\subset \Gamma_k(a, \delta_{\!\,\mathsmaller{0}})$ for each $\mu$. Suppose \eqref{lindepN} holds with $L=N$ and suppose Theorem \ref{lclb} holds with $L=N-1$. Then, if $2N<p\le N(N+1)$, there exist $\epsilon=\epsilon(p,N)>0$, a constant $C=C(p,d,N,B)$ and symbols $\mathfrak a_{\nu} \in \mathfrak A_k(\delta_1,\delta_1\nu)$, $\nu \in \cup_\mu\mathfrak J_0^\mu$ such that \begin{align}\label{eq-iter} \Big( \sum_\mu \| \mathcal T[ \mathfrak a^\mu]f\|_p^p\Big)^{\frac 1p}\! \! \le \! C \Big( \frac{\delta_0}{\delta_1} \Big)^{1-\frac {N+1}{p}-\epsilon}\! \Big( \sum_\nu \| \mathcal T[\mathfrak a_{\nu}]f\|_p^p\Big)^{\frac 1p} \!\!\!+\!C \delta_0^{1-\frac {N+1}p } 2^{-\frac{k}p}\|f\|_p. \end{align} \end{prop} In \cite{KLO} a similar inequality was used to show the sharp smoothing estimate for $\mathcal A_t$ with $\epsilon$-loss on the regularity. Compared with \cite{KLO}, we have the optimal bound $C2^{-k/p}$ for the error part in the right hand side of \eqref{eq-iter}. This is possible since we have the optimal regularity estimate when $L=1$, i.e., Theorem \ref{lclb} with $L=1$. Such bound allows us to impose an induction assumption with the optimal regularity in Proposition \ref{newprop}. Even if the decoupling inequality has $\epsilon$-loss in its bound, as to be seen below, the loss is harmless. Consequently, we can obtain the optimal estimate \eqref{one}. \begin{rmk} \label{optimal} The same argument continues to work for the convolution operator $\mathcal A_t$ since we also have the optimal regularity estimate for $\mathcal A_t$, that is to say, \[\| \mathcal A _t[a]f \|_{L^p(\mathbb R^d)} \le C 2^{-k/p} \|f\|_{L^p(\mathbb R^d)}, \quad 2 \le p \le \infty\] when \eqref{sumN} holds with $L=2$. {\rm(}See \cite[Section 5]{KLO}{\rm)}. Therefore, we get Theorem \ref{averaging}. However, a similar approach breaks down if one attempts to obtain the optimal local smoothing estimate for $\mathcal A_t$ $($see \cite[Theorem 1.3]{KLO}$)$. To remove epsilon-loss in the local smoothing estimates, we need the decoupling inequality \eqref{c-decoupling2} for $4N-2<p \le N(N+1)$. However, there is no such estimate when $N=2$. Since the proof is based on the induction, the approach does not work for the local smoothing estimate to give the optimal estimate. \end{rmk} The proof of Proposition \ref{iterative} is based on a conic extension of the decoupling inequality for the nondegenerate curve, see Theorem \ref{rcdecoupling}. We first prove Proposition \ref{microlocal} using Proposition \ref{iterative}, whose proof is given in Section \ref{sec3}. \subsection{Proof of Proposition \ref{microlocal}} Let $\epsilon$ and $C$ be the numbers in Proposition \ref{iterative}, which depend only on $d,p,N, B$. Let us set \begin{equation}\label{delta-def} \delta_*= (2^{3d}B)^{-N} C^{-2N/\epsilon}. \end{equation} We take $\delta_0=\delta_{\!\,\mathsmaller{0}}$ (\eqref{delta-min}). Then, for $j=1,\dots, J-1$, we set \begin{align}\label{defn-j} \delta_j=(2^{3d}B)^{N(\frac{N+1}N)^{j}-N} \delta_0^{(\frac{N+1}N)^{j}}, \end{align} where $J$ is the largest integer such that $\delta_{\!J-1}>2^{-k/N}$. We also set \[ \delta_{\!J}=2^{-\frac kN}. \] Then, $J\le C_1 \log k$ for some $C_1$. It is easy to check \begin{align} \label{del-con} 2^{3d}B\delta_j^{\frac{N+1}N} \le \delta_{j+1}< \delta_j, \quad \, j=0, \dots, J-1. \end{align} Since $\mathfrak a \in \mathfrak A_k(\delta_0,\delta_0\mu)$ and \eqref{del-con} holds, by applying Proposition \ref{iterative} there are symbols $\mathfrak a_{\nu_1} \in \mathfrak A_k(\delta_{1}, \delta_{1}\nu)$ such that \eqref{eq-iter} holds with $\mathfrak a_{\nu}$ replaced by $\mathfrak a_{\nu_1}$. Applying Proposition \ref{iterative} repeatedly, we have symbols $\mathfrak a_\nu \in \mathfrak A_k(\delta_{\!J}, \delta_{\!J}\nu)$, $\delta_{\!J}\nu\in [s_{\!\,\mathsmaller{0}}-\delta_{\!\,\mathsmaller{0}}, s_{\!\,\mathsmaller{0}}+\delta_{\!\,\mathsmaller{0}}]$, such that \begin{align}\label{sum1} \big\| \mathcal T[\mathfrak a]f \big\|_{p} \le C^J\delta_{\!J}^{\frac {N+1}{p}-1+\epsilon} \Big( \sum_\nu \| \mathcal T[\mathfrak a_{\nu}]f\|_p^p\Big)^{\frac 1p} \!+\sum_{j=0}^{J-1}C^{j+1} \delta_j^{\epsilon } 2^{-\frac{k}p}\|f\|_p. \end{align} Since $\delta_0 \le (2^{3d}B)^{-N} C^{-2N/\epsilon}$, it follows from \eqref{defn-j} that $\delta_j \le (2^{3d}B)^{-N} C^{-\frac {2N}\epsilon (\frac{N+1}N)^j}$ for $0 \le j \le J-1$. Thus, it follows that $ \sum_{j=0}^{J-1}C^{j+1} \delta_j^{\epsilon }\le (2^{3d}B)^{-\epsilon N} \sum_{j=0}^{J-1}2^{-j} \le C_2 $ for some constant $C_2=C_2(d,B,p,N)$. For $2\le p \le\infty$, we claim \begin{align}\label{center} \Big( \sum_\nu \| \mathcal T[\mathfrak a_{\nu}]f\|_p^p\Big)^{\frac 1p}\lesssim \delta_J^{1-\frac 1p}\|f\|_p \end{align} where $\mathfrak a_\nu \in \mathfrak A_k(\delta_{\!J}, \delta_{\!J}\nu)$ and $\delta_{\!J}\nu\in [s_{\!\,\mathsmaller{0}}-\delta_{\!\,\mathsmaller{0}}, s_{\!\,\mathsmaller{0}}+\delta_{\!\,\mathsmaller{0}}]$. Indeed, by \eqref{ker-est} we have \eqref{center} for $p=\infty$. Since $\supp_s \mathfrak a_\nu \subset I_\nu:=[\delta_{\!J}\nu-\delta_{\!J}, \delta_{\!J}\nu+\delta_{\!J}]$, using \eqref{T}, Plancherel's Theorem, and Holder's inequality in $s$, we obtain \[ \sum_{\nu} \| \mathcal T [\mathfrak a_\nu] f\|_2^2 \lesssim \sum_\nu \Big( \int_{I_\nu} \, \| \mathcal F_x f(\cdot, s) \|_2 \,ds \Big)^2 \lesssim \delta_J \|f\|_2^2. \] This gives \eqref{center} with $p=2$ and the desired bound \eqref{center} follows by interpolation. Combining \eqref{center} with \eqref{sum1}, we have \[ \big\| \mathcal T[\mathfrak a]f \big\|_{p} \le (C^{J}2^{ -\frac kp}2^{- \frac{k} N\epsilon} + C_2 2^{-\frac kp})\|f\|_p. \] Since $J\le C_1\log k$, taking $k$ sufficiently large we get \eqref{overline-a0} for $2N<p\le N(N+1)$. Interpolating this and the trivial estimate with $p=\infty$, we obtain \eqref{overline-a0} for $p>2N$. This completes the proof of Proposition \ref{microlocal}. \section{Proof of Proposition \ref{iterative}} \label{sec3} In this section, we prove Proposition \ref{iterative} by making use of the decoupling inequality for the curve. To do this, we need to decompose carefully a given symbol $\mathfrak a$ near the set $\{(\xi,s)\in \supp_{\xi,s} \mathfrak a: \gamma^{(N-1)}(s)\cdot \xi=0\}$. Before proceeding to prove Proposition \ref{iterative}, we first consider the operator $\mathcal R[a]$ given by a symbol $a$ adapted to a short curve. \subsection{Rescaling} Let $s_{\!\,\mathsmaller{0}} \in (-1,1)$ and $0<\delta \le \delta_{\!\,\mathsmaller{0}}$ satisfy $[s_{\!\,\mathsmaller{0}}-\delta, s_{\!\,\mathsmaller{0}}+\delta]\subset I$. Suppose $\gamma$ is a smooth curve satisfying \eqref{lindepN} with $L=N$. Then, \eqref{lindepN} with $L=N-1$ holds with $B$ replaced by $B^{2}$. We set \[ \mathcal M_s^\delta= \delta^{-N} \widetilde{\mathcal L}_s^\delta \] and define $\gamma_{s_{\!\,\mathsmaller{0}}}^\delta=\gamma_{s_{\!\,\mathsmaller{0}}}^\delta(s;N)$ by \[ \gamma_{s_{\!\,\mathsmaller{0}}}^\delta(s)= (\mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta)^\intercal \big(\gamma(\delta s+s_{\!\,\mathsmaller{0}})-\gamma(s_{\!\,\mathsmaller{0}})\big). \] By Taylor expansion of $\gamma^{(j)}(\delta s+s_{\!\,\mathsmaller{0}})$ at $s=0$, we see $\gamma=\gamma_{s_{\!\,\mathsmaller{0}}}^\delta$ satisfies \eqref{curveB} with $B$ replaced by $2B$ if $\delta$ is sufficiently small. We define \[ \mathcal R^{\gamma_{s_{\!\,\mathsmaller{0}}}^{\delta}}[a]f(x,t)=(2\pi)^{-d}\iint e^{i(x-t\gamma_{s_{\!\,\mathsmaller{0}}}^{\delta}(s))\cdot\xi}a(s,t,\xi) \mathcal F_x f (\xi,s)ds d\xi. \] The next lemma which is an analogue of \cite[Lemma 2.7]{KLO} is important for carrying out our inductive argument. \begin{lem}\label{lem:res} Let $2\le N\le d$, $s_{\!\,\mathsmaller{0}} \in (-1,1)$, and $\mathfrak a \in \mathfrak A_k(\delta,s_{\!\,\mathsmaller{0}})$. Let $\gamma \in \mathrm C^{3d+1}(I)$ satisfy \eqref{curveB} and \eqref{lindepN} with $L=N$. Suppose \[ \sum_{j=1}^{N-1} \delta^{j} |\langle \gamma^{(j)}(s), \xi \rangle |\ge B^{-1}2^k \delta^N \] holds on $[s_{\!\,\mathsmaller{0}}-\delta, s_{\!\,\mathsmaller{0}}+\delta]\times \supp_\xi \mathfrak a$. Then, there exist $\widetilde a $, $\widetilde f$, and constants $C$, $\widetilde B$, and $\tilde\delta_{*}=\tilde\delta_{*}(B,N,d)$ such that \begin{align}\label{res:TA} \big\| \widetilde\chi \mathcal T [\mathfrak a] f \big\|_p = \delta^{1-\frac 1p} \big\|\mathcal R^{\gamma_{s_{\!\,\mathsmaller{0}}}^{\delta}} [\, \raisebox{-.2ex}{$\widetilde {a}$} \,] \widetilde f \big\|_p, \end{align} $ \|\widetilde f\|_p=\|f\|_p$ and the following hold for $0<\delta<\tilde\delta_{*}$$:$ \begin{align} \label{Vol1} &\Vol \big( (\gamma_{s_{\!\,\mathsmaller{0}}}^\delta)^{(1)}(s), \dots, (\gamma_{s_{\!\,\mathsmaller{0}}}^\delta)^{(N-1)}(s) \big) \ge \widetilde B, \quad s\in I {\,\rm ;} \\ \label{lowerg} \qquad \sum_{j=1}^{N-1} &\big| \big\langle(\gamma_{s_{\!\,\mathsmaller{0}}}^\delta)^{(j)}(s),\xi \big\rangle \big|\ge \widetilde B^{-1}2^k\delta^N, \quad (s,\xi)\in I\times \supp_\xi \widetilde{a} {\,\rm ;} \\ \label{fourierF} \supp \widetilde a(s,t,\xi) &\,\subset\, I \times [2^{-2},2^2] \times \big\{ \xi \in \mathbb R^d: C^{-1}\delta^N 2^k \le |\xi | \le C \delta^N 2^{k} \big\}{\,\rm ;} \\ \label{diff:a} &\qquad\quad | \partial_t^{j}\partial_\xi^\alpha \widetilde a(s,t,\xi) | \le \widetilde B |\xi|^{-|\alpha|} \end{align} for $0\le j \le 2N-2$, and $|\alpha|\le d+N+1$. \end{lem} \begin{proof} Using \eqref{T}, by Fourier inversion and changing variables $s\rightarrow \delta s+s_{\!\,\mathsmaller{0}}$ and $(\tau,\xi)\rightarrow (\tau-\gamma(s_{\!\,\mathsmaller{0}})\cdot \xi, \xi)$, we have \[ \mathcal T[\mathfrak a]f(x,t)=C_d\delta \iint e^{i\langle x-t\gamma(s_{\!\,\mathsmaller{0}}),\xi \rangle} b(s,t,\xi) \mathcal F_xf(\xi, \delta s+s_{\!\,\mathsmaller{0}})\,ds d\xi \] where $b(s,t,\xi)$ is given by \[ b(s,t,\xi)= \iint e^{it\tau} e^{-it' ( \tau+ \langle \gamma(\delta s+s_{\!\,\mathsmaller{0}})-\gamma(s_{\!\,\mathsmaller{0}}), \xi \rangle)} \mathfrak a(\delta s+s_{\!\,\mathsmaller{0}}, t', \tau-\gamma(s_{\!\,\mathsmaller{0}})\cdot \xi, \xi)\,dt' d\tau. \] Let $\widetilde f$ be given by $\mathcal F_x \widetilde f (\xi,s)= \delta^{\frac 1p} |\det \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta|^{1-\frac 1p} \mathcal F_x f(\mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta \xi, \delta s+s_{\!\,\mathsmaller{0}})$. Then we have $\|\widetilde f\|_p=\|f\|_p$. Changing variable $\xi \rightarrow \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta \xi$, we have \[ \mathcal T[\mathfrak a]f(x,t) = C_d \delta^{1-\frac 1p} |\det \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta |^{\frac 1p} \iint e^{i \langle x-t\gamma(s_{\!\,\mathsmaller{0}}), \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta \xi \rangle} b(s,t, \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta\xi ) \mathcal F_x \widetilde f(\xi,s)\,ds d\xi. \] Note that $(\tau-\gamma (s_{\!\,\mathsmaller{0}}) \cdot \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta \xi, \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta \xi)=\delta^{-N} \mathcal L_{s_{\!\,\mathsmaller{0}}}^\delta (\tau,\xi)$. Thus, if we set \begin{align*} \widetilde{ a}(s,t,\xi)&=\frac 1{2\pi}\iint e^{-it'(\tau+\gamma_{s_{\!\,\mathsmaller{0}}}^\delta(s)\cdot \xi)}\widetilde{\chi}(t)\mathfrak a(\delta s+s_{\!\,\mathsmaller{0}},t+t',\delta^{-N}\mathcal L_{s_{\!\,\mathsmaller{0}}}^\delta(\tau,\xi))dt'd\tau, \end{align*} then $\widetilde \chi(t) b(s,t, \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta \xi) = 2\pi e^{-it\gamma_{s_{\!\,\mathsmaller{0}}}^\delta (s)\cdot \xi} \widetilde a(s,t,\xi)$ and \begin{align*} \widetilde \chi(t)\mathcal T [\mathfrak a]f(x,t) =\delta^{1-\frac 1p} |\det \mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta|^{\frac 1p} \mathcal R^{\gamma_{s_{\!\,\mathsmaller{0}}}^\delta} [ \,\raisebox{-.3ex}{$\widetilde a$} \,] \widetilde f \big( (\mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta)^{\intercal} (x-t\gamma(s_{\!\,\mathsmaller{0}})),t\big). \end{align*} Therefore, changing of variable $x \rightarrow (\mathcal M_{s_{\!\,\mathsmaller{0}}}^\delta)^{-\intercal}x+t\gamma(s_{\!\,\mathsmaller{0}})$, we obtain \eqref{res:TA}. It now remains to prove \eqref{Vol1}, \eqref{lowerg}, \eqref{fourierF} and \eqref{diff:a}. One can easily show those estimates following the argument in \cite{KLO}. So, we omit the details. \end{proof} \subsection{Decomposition} We apply the decoupling inequality after projecting the support of $\mathfrak a^\mu$ to the subspace $\mathrm V_\mu$ generated by $\{G^{(j)}(\delta_0\mu) , j=0,\dots,N\}$ for each $\mu$. In order to appropriately decompose the symbols $\mathfrak a^\mu$, we need to approximate $\langle G(s), (\tau,\xi) \rangle $ in a local coordinate. For the purpose we follow the argument \cite[Section 4.2]{KLO} to which we refer the reader for the details. For each $\mu$, we define \[ y_\mu^j(\tau,\xi)= \langle G^{(j)}(\delta_0\mu ),(\tau,\xi)\rangle, \quad j=0,1,\dots,N. \] Then, by \eqref{lowerN} we have \begin{equation}\label{lowYN} |y_\mu^N(\tau,\xi)|\ge(2B)^{-1}|\xi|. \end{equation} We also set \[ \omega_\mu(\tau,\xi) = \frac{y_\mu^{N-1}(\tau,\xi)}{ y_\mu^N(\tau,\xi)}. \] Now we approximate $y_\mu^j(\tau,\xi)$ via Taylor series expansion of $G^{(j)}(s)$ at $s=\sigma(\xi)$. To do this, we set \[ \mathfrak g_\mu^N(\tau,\xi)=y_\mu^N(\tau,\xi), \] and then define $\mathfrak g_\mu^{j}$, $j=N-1,\dots,1,0$, recursively by \begin{align}\label{u-w} \mathfrak g_\mu^{j}(\tau,\xi)=y^{j}_\mu(\tau,\xi)-\sum_{\ell=j+1}^{N}\frac{\mathfrak g^{\ell}_\mu(\tau,\xi)}{(\ell-j)!} (\omega_\mu(\tau,\xi))^{\ell-j}. \end{align} In particular, $\mathfrak g_\mu^{N-1}(\tau,\xi)=0$. By \eqref{u-w}, we can write $y_\mu^m$ in terms of $\mathfrak g_\mu^j$ and $\omega_\mu$: \begin{equation}\label{uw2} y_\mu^m(\tau,\xi) = \sum_{\ell=m}^N \frac{\mathfrak g_\mu^\ell(\tau,\xi)}{(\ell-m)!} (\omega_\mu(\tau,\xi))^{\ell-m}, \quad m=0,\dots,N. \end{equation} Here $y_\mu^j$, $\mathfrak g_\mu^j$ are independent of $\tau$ except for $j=0$ but we include $\tau$ as a variable for notational convenience. We can now approximate $\delta_0\mu-\sigma(\xi)$ and $\langle G^{(j)}(\sigma(\xi)),(\tau,\xi)\rangle$ in terms of $\omega_\mu^j$ and $\mathfrak g^{j}_\mu$. \begin{lem}[{\cite[Lemma 4.3]{KLO}}] We have \begin{align}\label{angle} \delta_0\mu-\sigma(\xi) = \omega_\mu(\tau,\xi)-\mathcal E_1(\xi) \end{align} and $\mathcal E_1(\xi)=O(2^n\delta_1)$. Also, for $j=0,\dots,N-1$, we have \[ \mathfrak g_\mu^j(\tau,\xi)= \langle G^{(j)}(\sigma(\xi)), (\tau,\xi) \rangle -\sum_{\ell=j+1}^N \frac{\mathfrak g_\mu^\ell(\tau,\xi)}{(\ell-j)!} O(\delta_0^{2(\ell-j)}) +y_\mu^N(\tau,\xi) O(\delta_0^{N-j+1}). \] \end{lem} To decompose $\mathfrak a^\mu \in \mathfrak A_k(\delta_0,\delta_0\mu)$, we denote \[ \mathfrak G_N(s,\tau,\xi)=\sum_{j=0}^{N-2}(2^{-k}|\mathfrak g^{j}_\mu(\tau,\xi)|)^{\frac{2N!}{N-j}}+(s-\sigma(\xi))^{2N!}. \] Setting $\beta_N=\beta_0-\beta_0(2^{2N!}\cdot)$, we define \begin{align*} \mathfrak a_{0,\nu}^\mu(s,t,\tau,\xi) &=\mathfrak a^\mu (s,t,\tau,\xi) \, \beta_0 \big( \delta_1^{-2N!} \, \mathfrak G_N(s,\tau,\xi) \big) \, \zeta(\delta_1^{-1}s-\nu), \quad\, \nu \in \mathfrak J_0^\mu; \\[1ex] \mathfrak a_{n,\nu}^\mu(s,t,\tau,\xi) &=\mathfrak a^\mu(s,t,\tau,\xi)\, \beta_N \big( (2^{n}\delta_1)^{-2N!} \, \mathfrak G_N(s,\tau,\xi) \big) \, \zeta(2^{-n}\delta_1^{-1}s-\nu) \end{align*} for $n\ge1$ and $\nu \in \mathfrak J_n^\mu$ (see \eqref{Jnmu}). Then, it follows that \begin{align}\label{split} \sum_\mu \mathfrak a^\mu(s,t,\tau,\xi)= \sum_{n\ge 0}\sum_\mu \sum_{\nu \in \mathfrak J_n^\mu} \mathfrak a^\mu_{n,\nu}(s,t,\tau,\xi). \end{align} The following tells that $\mathfrak a_{n,\nu}^\mu$ is contained in a proper symbol class. \begin{lem} [{\cite[Lemma 4.4]{KLO}}]\label{symcheck} There exists a constant $C$ such that $C^{-1}\mathfrak a_{n,\nu}^\mu\in \mathfrak A_k(2^n\delta_1,2^n\delta_1\nu)$. \end{lem} \subsection{Decoupling inequality} Let us set \[ \gamma_{\circ, N+1}(s)=(s,s^2/2!,\dots, s^{N+1}/(N+1)!). \] We consider an anisotropic neighborhood of the curve. \begin{defn} Let $N\ge2$ and $\gamma_\circ=\gamma_{\circ, N+1}$. For $s \in I$, let $\mathbf b(s,\delta,N; \gamma_\circ)$ be the set of $(\tau,\xi) \in \mathbb R \times \mathbb R^N$ such that \begin{align*} 2^{-1} \le &| \langle \gamma_{\circ \phantom{1}}^{(N+1)}(s), (\tau,\xi) \rangle| \le1, \\ &| \langle \gamma_{\circ \phantom{1}}^{(j)}(s),(\tau,\xi) \rangle| \le \delta^{N+1-j}, \quad j=1,\dots,N. \end{align*} Let $\{s_k\}$ be $2^{-1}\delta$-separated points contained in $I$ such that $\cup_j [s_j-\delta, s_j+\delta]$ covers $I$. Then, we say $\mathcal B(\delta) =\{ \mathbf b(s_1,\delta,N; \gamma_\circ), \dots, \mathbf b(s_k,\delta, N;\gamma_\circ)\}$ is a reverse $\delta$-adapted cover generated by $\gamma_\circ$. \end{defn} We recall the decoupling inequality for a reverse $\delta$-adapted cover, which was shown in \cite{BGHS} (see also \cite[Corollary 3.4]{KLO}). \begin{thm}\label{rcdecoupling} Let $0<\delta\le1$. Let $\mathcal B(\delta)=\{\mathbf b(s_j,\delta, N;\gamma_\circ): 1\le j\le k\}$ be a reverse $\delta$-adapted cover generated by $\gamma_{\circ, N+1}$. Suppose $F=\sum_{\mathbf b \in \mathcal B(\delta)} F_{\mathbf b}$ and $\supp \widehat {F}_{\mathbf b}\subset \mathbf b$. Then, if $2 \le p\le N(N+1)$, for $\epsilon>0$ we have \begin{align*} \big\| F \big\|_{L^p(\mathbb R^{N+1})}\le C_\epsilon \delta^{-\epsilon} \Big(\sum_{\mathbf b \in \mathcal B(\delta)} \|F_{\mathbf b}\|_{L^p(\mathbb R^{N+1})}^2 \Big)^{\frac 12}. \end{align*} \end{thm} By H\"older's inequality, we have \begin{align}\label{c-decoupling2} \big\| F \big\|_{L^p(\mathbb R^{N+1})}\le C_\epsilon \delta^{-\frac12+\frac 1p-\epsilon} \Big(\sum_{\mathbf b \in \mathcal B(\delta)} \|F_{\mathbf b}\|_{L^p(\mathbb R^{N+1})}^p \Big)^{\frac 1p} \end{align} for $2 \le p \le N(N+1)$. We prove Proposition \ref{iterative} using \eqref{c-decoupling2}. \subsection{Proof of Proposition \ref{iterative}} The proof of Proposition \ref{iterative} is similar with that of \cite[Proposition 4.1]{KLO}. So, we shall be brief. Some details can be found in \cite{KLO}. Recalling \eqref{split} and applying the Minkowski inequality, we have \begin{equation}\label{Minkowski} \Big( \sum_\mu \| \mathcal T[\mathfrak a^\mu] f\|_p^p \Big)^{\frac 1p} \le \sum_{n\ge0} \Big( \sum_\mu \big\| \sum_{\nu \in \mathfrak J_n^\mu} \mathcal T[\mathfrak a_{n,\nu}^\mu]f \|_p^p \Big)^{\frac 1p}. \end{equation} In order to apply Theorem \ref{rcdecoupling} to the right hand side, we need to show that for each fixed $\mu,n$, the projections of the Fourier supports of $\mathcal T[\mathfrak a_{n,\nu}^\mu]f$, $\nu \in \mathfrak J_n^\mu$ are contained in a reverse adapted cover generated by $\gamma_\circ=\gamma_{\circ, N+1}$. Let us set $V=\sspan \{ \gamma'(\delta_0\mu),\dots, \gamma^{(N)}(\delta_0\mu)\}$. Since \eqref{lindepN} holds with $L=N$, we may write $\xi=\overline \xi+\sum_{j=N+1}^d y_j(\xi)v_j$ for $\overline \xi \in V$ and an orthonormal basis $\{v_{N+1},\dots,v_d\}$ of $V^\perp$. We denote \[ \overline y_\mu(\tau,\xi)=(y_\mu^0(\tau,\xi),\cdots, y_\mu^N(\tau,\xi)). \] Making change of variable $(\tau,\xi) \rightarrow \mathrm Y_\mu(\tau,\xi):= (\overline y_\mu(\tau,\xi), y_{N+1}(\xi),\dots,y_d(\xi))$, it suffices to consider the coordinate system generated by $\{ \overline y_\mu,y_{N+1},\dots,y_d\}$ instead of $(\tau,\xi)$. As before, denote $\gamma_\circ=\gamma_{\circ, N+1}$ for simplicity. We now claim that \begin{align} |\langle \overline y_\mu(\tau,\xi),\gamma_{\vphantom{1}\circ}^{(N+1)} (2^n\delta_1\nu-\delta_0&\mu) \rangle | \sim 2^k, \label{claim2} \\ |\langle \overline y_\mu(\tau,\xi),\gamma_{\vphantom{1}\circ}^{(j)}(2^n\delta_1\nu-\delta_0\mu) \rangle| &\lesssim 2^k (2^n\delta_1)^{N+1-j}, \quad 1 \le j \le N, \label{claim1} \end{align} whenever $(\tau,\xi) \in \supp_{\tau,\xi} \mathfrak a_{n,\nu}^\mu$. Denoting $\gamma_\circ=((\gamma_{\circ})_1,\dots, (\gamma_{\circ})_{N+1})$ and using \eqref{uw2}, we have \[ y_\mu^{m-1} (\tau,\xi)(\gamma_{\circ})_{m}^{(j)}(2^n\delta_1\nu-\delta_0\mu)= \sum_{\ell=m-1}^N \frac{\mathfrak g_\mu^\ell(\tau,\xi) (2^n\delta_1\nu-\delta_0\mu)^{m-j}}{(\ell+1-m)!(m-j)!} (\omega_\mu)^{\ell+1-m} \] for $m\ge j$. Thus, summation over $j \le m \le N+1$ gives \[ \langle \overline y_\mu, \gamma_{\vphantom{1}\circ}^{(j)}(2^n\delta_1\nu-\delta_0\mu) \rangle =\sum_{\ell=j-1}^N \frac{ \mathfrak g_\mu^\ell(\tau,\xi) }{(\ell+1-j)!} (2^n\delta_1 \nu- \delta_0\mu+\omega_\mu)^{\ell+1-j}. \] Since $|\sigma(\xi)-2^n\delta_1\nu| \le 2^{n+1}\delta_1$ on $\supp \mathfrak a_{n,\nu}^\mu$, by \eqref{angle} we have $|2^n\delta_1\nu-\delta_0\mu+\omega_\mu| \lesssim 2^n\delta_1$. Note $\mathfrak G_N \le (2^n\delta_1)^{2N!}$ on $\supp \mathfrak a_{n,\nu}^\mu$. So, we have $|\mathfrak g_\mu^j(\tau,\xi)| \le 2^k(2^n\delta_1)^{N-j}$ for $0\le j \le N-2$ on $\supp \mathfrak a_{n,\nu}^\mu$. Since $\mathfrak g_\mu^{N-1}=0$, we see \eqref{claim1} holds. The bound \eqref{claim2} follows from \eqref{lowYN} and this proves the claim. Let us consider the $(N+1)\times (N+1)$ matrix $\mathcal D_{\!s}^{\delta}=(\delta \gamma'(s), \dots, \delta^{N+1}\gamma^{(N+1)}(s))$. Since $\gamma_\circ(\delta_0 s-\delta_0\mu)-\gamma_\circ(-\delta_0\mu)=\mathcal D_{\!-\delta_0 \mu}^{\delta_0} \gamma_\circ(s)$, we note that $\gamma_{\vphantom{1}\circ}^{(j)}(\delta_0 s - \delta_0 \mu)=\delta_0^{-j} \mathcal D_{\!-\delta_0\mu}^{\delta_0} \gamma_{\vphantom{1}\circ}^{(j)}(s)$ for $j\ge 1$. If we change variables $\overline y_\mu \rightarrow 2^{k}\delta_0^{N+1} (\mathcal D_{\!-\delta_0\mu}^{\delta_0})^{-\intercal} \overline y_\mu$, \eqref{claim2} and \eqref{claim1} are, respectively, equivalent to \begin{align*} &\big| \big\langle \overline y_\mu(\tau,\xi),\gamma_{\vphantom{1}\circ}^{(N+1)} \Big( \frac{2^n\delta_1}{\delta_0}\nu \Big) \big\rangle \big| \sim 1, \\ &\big| \big\langle \overline y_\mu(\tau,\xi),\gamma_{\vphantom{1}\circ}^{(j)}\Big( \frac{2^n\delta_1}{\delta_0}\nu \Big) \big\rangle \big| \lesssim \Big(\frac{2^n\delta_1}{\delta_0}\Big)^{N+1-j}, \quad 1 \le j \le N. \end{align*} We define a linear map $\mathrm Y_\mu^{\delta_0}$ by setting \[ \mathrm Y_\mu^{\delta_0}(\tau,\xi)= \Big( 2^k \delta_0^{N+1} (\mathcal D_{\!-\delta_0\mu}^{\delta_0})^{-\intercal} \overline y_\mu(\tau,\xi),\, y_{N+1}(\xi),\dots,\,y_d(\xi)\Big). \] Then, it follows that \begin{equation} \label{supp-supp} \mathrm Y_\mu^{\delta_0} ( \supp_{\tau,\xi} \mathfrak a_{n,\nu}^\mu ) \subset \mathbf b_\nu:=\mathbf b\Big(\frac{2^n\delta_1}{\delta_0}\nu, C\frac{2^n\delta_1}{\delta_0}, N; \gamma_\circ\Big) \times \mathbb R^{d-N} \end{equation} for some $C>0$. For $p>2N$, we take $\epsilon_*=1/4-N/(2p)>0$ so that $1/2-1/p+\epsilon_*=1-(N+1)/p-\epsilon_*$. We obtain the decoupling inequality \eqref{c-decoupling2} with $\delta=2^n\delta_1/\delta_0$, $\epsilon=\epsilon_*$ and $\mathcal B(\delta)=\{ \mathbf b_\nu :\nu\in \mathfrak J_{n}^\mu\} $. Thus, by a trivial extension via the Minkowski's inequality we have \[ \big\| \sum_{\nu\in \mathfrak J_{n}^\mu} F_\nu \big\|_p \le C \Big( \frac{\delta_0}{2^n\delta_1} \Big)^{1-\frac {N+1} {p}-\epsilon} \Big(\sum_{\nu\in \mathfrak J_{n}^\mu} \big\| F_\nu \big\|_p^p \Big)^{\frac 1p} \] for $2N<p \le N(N+1)$ provided that $\supp \widehat{F_\nu}\subset \mathbf b_\nu$. From \eqref{T} we note that $\mathcal F(\mathcal T [\mathfrak a_{n,\nu}^\mu]f)$ is contained in $ \supp_{\tau,\xi} \mathfrak a_{n,\nu}^\mu$. Using \eqref{supp-supp} and the harmless change of variables $(\tau, \xi)\rightarrow ( \mathrm Y_\mu^{\delta_0})^{-1}(\tau,\xi)$, we obtain \[ \big\| \sum_{\nu\in \mathfrak J_{n}^\mu} \mathcal T [\mathfrak a_{n,\nu}^\mu]f \big\|_p \le C \Big( \frac{\delta_0}{2^n\delta_1} \Big)^{1-\frac {N+1} {p}-\epsilon} \Big(\sum_{\nu\in \mathfrak J_{n}^\mu} \big\| \mathcal T [\mathfrak a_{n,\nu}^\mu]f \big\|_p^p \Big)^{\frac 1p} \] for $2N<p \le N(N+1)$. Combining this and \eqref{Minkowski}, we get \begin{align*} \Big( \sum_\mu \| \mathcal T[\mathfrak a^\mu] f\|_p^p \Big)^{\frac 1p} \le C\Big( \frac{\delta_0}{\delta_1} \Big)^{1-\frac {N+1} {p}-\epsilon} \Big( \sum_\mu \sum_{\nu \in \mathfrak J_0^\mu} \big\| \mathcal T[\mathfrak a_{0,\nu}^\mu]f \|_p^p \Big)^{\frac 1p} +\mathrm S \end{align*} where \begin{align*} \mathrm S= C\sum_{n\ge1}\Big( \frac{\delta_0}{2^n\delta_1} \Big)^{1-\frac {N+1} {p}-\epsilon} \Big( \sum_\mu \sum_{\nu \in \mathfrak J_n^\mu} \big\| \mathcal T[\mathfrak a_{n,\nu}^\mu]f \|_p^p \Big)^{\frac 1p}. \end{align*} By Lemma \ref{symcheck}, note that $C^{-1}\mathfrak a_{0,\nu}^{\mu} \in \mathfrak A_k(\delta_1,\delta_1\nu)$ for some $C>0$. We set $\mathfrak a_{\nu}=C^{-1}\mathfrak a_{0,\nu}^{\mu(\nu)}$ where $\mu(\nu)$ satisfies $|\nu-\mu(\nu)|\le|\nu-\mu|$ for any $\mu$. To complete the proof, it remains to show $\mathrm S \le C \delta_0^{1-\frac {N+1}p} 2^{-\frac kp}\|f \|_p$. For $n\ge1$, we further decompose $\mathfrak a_{n,\nu}^\mu$. Let \[ \mathfrak G_N^0(s,\tau,\xi)=\sum_{j=1}^{N-2}(2^{-k}|\mathfrak g^{j}_\mu(\tau,\xi)|)^{\frac{2N!}{N-j}}+(s-\sigma(\xi))^{2N!}. \] Note that $\mathfrak G_N^0=\mathfrak G_N-(2^{-k} \mathfrak g_\mu^0)^{2(N-1)!}$. For $C_0 \ge 2^{3d}100B$, we set \[ \mathfrak a_{n,\nu}^{\mu,1} (s,t,\tau,\xi) = \mathfrak a_{n,\nu}^{\mu} (s,t,\tau,\xi) \beta_0 \Big( \frac{2^{-k} \mathfrak g_\mu^0(\tau,\xi) )^{2(N-1)!}} {C_0^{2N!} \mathfrak G_N^0(s,\tau,\xi)} \Big) \] and $\mathfrak a_{n,\nu}^{\mu,2} = \mathfrak a_{n,\nu}^{\mu}-\mathfrak a_{n,\nu}^{\mu,1}$, so that we have $\mathfrak a_{n,\nu}^{\mu} = \mathfrak a_{n,\nu}^{\mu,1}+\mathfrak a_{n,\nu}^{\mu,2}. $ To prove \eqref{eq-iter}, it suffices to show \begin{equation}\label{S12} \mathrm S_i \le C \delta_0^{1-\frac {N+1}p } 2^{-\frac{k}p}\|f\|_p, \quad i=1,2, \end{equation} where \begin{align}\label{Si} \mathrm S_i= C\sum_{n\ge1}\Big( \frac{\delta_0}{2^n\delta_1} \Big)^{1-\frac {N+1} {p}-\epsilon} \Big( \sum_\mu \sum_{\nu \in \mathfrak J_n^\mu} \big\| \mathcal T[\mathfrak a_{n,\nu}^{\mu,i}]f \|_p^p \Big)^{\frac 1p}, \quad i=1,2. \end{align} Before proceeding further, we note that $C^{-1} \mathfrak a_{n,\nu}^{\mu,i} \in \mathfrak A_k(2^n\delta_1,2^n\delta_1\nu)$, $i=1,2$ for some $C>0$. This can be checked by following the proof of Lemma \ref{symcheck} (see \cite{KLO}). We also recall the following from \cite{KLO}. \begin{lem}[Lemma 4.7 in \cite{KLO}] For $n\ge 1$ and a sufficiently large $C_0>1$, \begin{align} \label{sumss} \sum_{j=1}^{N-1} (2^n\delta_1)^{-(N-j)} |\langle\gamma^{(j)}(s),\xi\rangle|\gtrsim 2^k \ \text{on } \supp \mathfrak a_{n,\nu}^{\mu,1}, \\ \label{sumss2} (2^n\delta_1)^{-N} |\tau+\langle\gamma(s),\xi\rangle|\gtrsim 2^{k} \ \text{on } \supp \mathfrak a_{n,\nu}^{\mu,2}. \end{align} \end{lem} We first show \eqref{S12} with $i=1$. Since $C^{-1}\mathfrak a_{n,\nu}^{\mu,1} \in \mathfrak A_k(2^n\delta_1,2^n\delta_1\nu)$ for some $C>0$ and \eqref{sumss} holds, by applying Lemma \ref{lem:res} with $\delta=2^n\delta_1$ to $\mathfrak a_{n,\nu}^{\mu,1}$, we have \begin{align*} \big\| \widetilde\chi \mathcal T [\mathfrak a_{n,\nu}^{\mu,1}] f \big\|_p = (2^n\delta_1)^{1-\frac 1p} \big\|\mathcal R^{\gamma_{s_{\!\,\mathsmaller{0}}}^{2^n\delta_1}} [\, \raisebox{-.2ex}{$\widetilde {a}$} \,] \widetilde f \big\|_p. \end{align*} By decomposing $\supp_t \widetilde{a}$ and rescaling, we may assume that $\supp_t \widetilde{a} \subset [1,2]$. Thus we apply the induction hypothesis (Theorem \ref{lclb} with $L=N-1$) and obtain $ \|\widetilde \chi\mathcal T [\mathfrak a_{n,\nu}^{\mu,1}]f \|_p\lesssim (2^n\delta_1)^{1-\frac{N+1}p}2^{-\frac kp}\|f \|_p. $ On the other hand, by Lemma \ref{kernel}, we have $ \|(1-\widetilde \chi)\mathcal T [\mathfrak a_{n,\nu}^{\mu,1}]f \|_p\lesssim 2^{-k} (2^n\delta_1)^{1-\frac 1p-N}\|f \|_p. $ Combining those two bounds gives \begin{equation}\label{T11} \|\mathcal T [\mathfrak a_{n,\nu}^{\mu,1}]f \|_p\lesssim (2^n\delta_1)^{1-\frac{N+1}p}2^{-\frac kp}\|f \|_p. \end{equation} For ${\tilde s}\in I$, we define a projection operator $\mathcal P_{\tilde s}^\delta$ by \[ \mathcal P_{\tilde s}^\delta f =\mathcal F_\xi^{-1} \big( \beta_0 \big( (C_02^k)^{-1}| (\widetilde{\mathcal L}_{\tilde s}^\delta)^{-1}\xi | \big) \mathcal F_x (f (\cdot,s))(\xi)\big). \] The projection operator $\mathcal P_{\tilde s}^\delta$ is slightly different from a similar projection operator used in \cite[Section 4.3]{KLO} since there is an extra variable $s$. However, we can handle $\mathcal P_{\tilde s}^\delta$ in the same way. Since $C^{-1}\mathfrak a_{n,\nu}^{\mu,1}\in \mathfrak A_k(2^n\delta_1, 2^n\delta_1\nu)$ and $\supp_\xi \mathfrak a^{\mu,1}_{n,\nu}\subset \Gamma_k(a_N, \delta_{\!\,\mathsmaller{0}} )$, we can choose $C_0>0$ large enough so that $ \beta_0 \big( (C_02^k)^{-1}| (\widetilde{\mathcal L}_{2^n\delta_1\nu}^{2^n\delta_1})^{-1}\xi | \big) =1$ on the support of $\mathfrak a_{n,\nu}^{\mu,1}$ (see \cite[Lemma 4.8]{KLO}). Thus, it follows $\mathcal T[\mathfrak a_{n,\nu}^{\mu,1}]f=\mathcal T[\mathfrak a_{n,\nu}^{\mu,1}] \mathcal P_{2^n\delta_1\nu}^{2^n\delta_1} f$. By disjointness of the supports of $ \beta_0 \big( (C_02^k)^{-1}| (\widetilde{\mathcal L}_{2^n\delta_1\nu}^{2^n\delta_1})^{-1}\cdot | \big)$, one can easily see \begin{equation}\label{Psum} \Big( \sum_\mu \sum_{\nu \in \mathfrak J_n^\mu} \big\| \mathcal P_{2^n\delta_1\nu}^{2^n\delta_1} f\big\|_p^p \Big)^{\frac 1p} \le C\|f\|_p \end{equation} for $2 \le p \le \infty$, which can be shown by interpolating \eqref{Psum} with $p=2$ and $p=\infty$. The estimate \eqref{Psum} with $p=2$ follows by Plancherel's theorem. Thus, combining \eqref{T11} and \eqref{Psum}, we get \eqref{S12} for $i=1$. Now we consider the estimate for $\mathrm S_2$, which is easier. By \eqref{T} and integration by parts we have \[ \mathcal F(\mathcal T [\mathfrak a_{n,\nu}^{\mu,2}]f)(\xi,\tau) =\iint e^{-it'(\tau+ \gamma(s)\cdot \xi)} \frac{\partial_{t'} \mathfrak a_{n,\nu}^{\mu,2} (s,t',\tau,\xi)}{i(\tau+ \langle \gamma(s),\xi \rangle)} dt' \, \mathcal F_x (f (\cdot,s))(\xi)ds. \] Since $C^{-1} \mathfrak a_{n,\nu}^{\mu,2} \in \mathfrak A_k(2^n\delta_1,2^n\delta_1\nu)$ for some $C>0$, by \eqref{sumss2} we see \eqref{symsupp} and \eqref{symineq2} hold with $\mathfrak a=2^k (2^n\delta_1)^{N} (\tau+\langle\gamma(s),\xi\rangle)^{-1}\partial_t \mathfrak a_{n,\nu}^{\mu,2}$ and $\delta=2^n\delta_1$ for $j=0$ and $|\alpha| \le d+3$. Thus, by Lemma \ref{kernel} we have \[ \|\mathcal T [\mathfrak a_{n,\nu}^{\mu,2}]f \|_\infty \lesssim (2^n\delta_1)^{1-N}2^{-k}\|f\|_\infty. \] As before, noting $\mathcal T[\mathfrak a_{n,\nu}^{\mu,2}]f=\mathcal T[\mathfrak a_{n,\nu}^{\mu,2}] \mathcal P_{2^n\delta_1\nu}^{2^n\delta_1} f$, by Plancherel's theorem and \eqref{sumss2} we have \[ \big\|\mathcal T [\mathfrak a_{n,\nu}^{\mu,2}] f \big\|_2 \lesssim 2^{-k} (2^n\delta_1)^{-N} \int_{\{s: |s-2^n\delta_1\nu| \lesssim 2^n\delta_1 \}}\, \big\| \mathcal F_x (\mathcal P_{2^n\delta_1\nu}^{2^n\delta_1}f) (\cdot, s) \big\|_2\,ds. \] Then, H\"older's inequality gives $ \big\|\mathcal T [\mathfrak a_{n,\nu}^{\mu,2}] f \big\|_2 \lesssim 2^{-k} (2^n\delta_1)^{-N+\frac 12} \big\| \mathcal P_{2^n\delta_1\nu}^{2^n\delta_1} f\big\|_2. $ We interpolation this and the $L^\infty$ estimate to get $\|\mathcal T [\mathfrak a_{n,\nu}^{\mu,2}]f \|_p \lesssim 2^{-k} (2^n\delta_1)^{-N+1-\frac 1p} \| \mathcal P_{2^n\delta_1\nu}^{2^n\delta_1}f\|_p$ for $2\le p\le \infty$. Then, combining this with \eqref{Psum}, we obtain \begin{equation*} \Big( \sum_{\nu \in \mathfrak J_n^\mu} \|\mathcal T [\mathfrak a_{n,\nu}^{\mu,2}]f \|_p^p \Big)^{\frac 1p} \lesssim 2^{-k} (2^n\delta_1)^{-N+1-\frac 1p}\|f\|_p. \end{equation*} Since $2^{-\frac kN} \le \delta_1$, by \eqref{Si}, we get \eqref{S12} for $i=2$. \section{Necessary conditions} In this section, we obtain necessary conditions for $L^p$--$L_\alpha^p$ boundedness of $\mathfrak Rf$ when $\gamma$ is of maximal type $L$. When $d=2$, the necessary conditions were obtained in \cite{PS06}. Modifying the examples in \cite{PS06}, we obtain the following. \begin{prop}\label{nece} Let $d\ge3$, $L\ge d$, and $1 \le p \le \infty$. Let $\psi$ and $\chi$ be nontrivial, nonnegative continuous functions supported in the interiors of $I$ and $[1,2]$, respectively. Suppose there is an $s_{\!\,\mathsmaller{0}}$ such that $\psi(s_{\!\,\mathsmaller{0}})\neq0$ and $\gamma$ is of type $L$ at $s_{\!\,\mathsmaller{0}}$. Then, $\mathfrak Rf$ maps $L^p(\mathbb R^{d+1})$ boundedly to $L_\alpha^p(\mathbb R^{d+1})$ only if \[\alpha\le \min\big\{ 1-p^{-1}, \ (2d)^{-1}, \ (Lp)^{-1} \big \}.\] \end{prop} \begin{proof}[Proof of $\alpha \le 1-p^{-1}$] Let $t_{\mathsmaller 0} \in (1,2)$ such that $\chi(t_{\mathsmaller 0})>0$. We choose $\zeta \in \mathcal S(\mathbb R^{d})$ such that $\zeta \ge1$ on $[-1,1]^{d}$, $\supp \widehat \zeta \subset [1/2,4]^d$, and $\widehat\zeta=1$ on $[1,2]^d$. Let $\psi_0\in \mathrm C_c^\infty((-1,1))$ satisfy $\psi_0=1$ on $[-1/2,1/2]$. We take \[ f(x,t)=\zeta(\lambda x)\psi_0(\lambda r_0 |t-t_{\mathsmaller 0}|) \] where $r_0=1+\sup_{s \in I}|\gamma(s)|$. Then, we have $\mathfrak Rf(x,s) \gtrsim \lambda^{-1}$ if $|x+t_{\mathsmaller 0}\gamma(s)|\le c \lambda^{-1}$ and $|s-s_{\!\,\mathsmaller{0}}|<c$. Thus, $\| \mathfrak Rf \|_{L^p(\mathbb R^{d+1})} \gtrsim \lambda^{-1-d/p}$. We note \[ \mathcal F_x(\mathfrak Rf(\cdot, s)) = \lambda^{-d} \psi(s)\int \widehat \zeta(\lambda^{-1}\xi) e^{it\gamma(s)\cdot \xi}\psi_0(\lambda|t-t_{\mathsmaller 0}|)\chi(t)\,dt. \] Since $\supp_\xi \mathcal F_x(\mathfrak Rf)$ is contained in $\{ \xi : |\xi| \sim \lambda \}$, $\| \mathfrak Rf(\cdot,s) \|_{L_\alpha^p(\mathbb R^{d};dx)} \gtrsim \lambda^{\alpha-1-d/p}$. So, we have $\| \mathfrak Rf\|_{L_\alpha^p(\mathbb R^{d+1})} \gtrsim \lambda^{\alpha-1-d/p}$. Since $\|f\|_p \lesssim \lambda^{-(d+1)/p}$, we get $\alpha \le 1-1/p$. \end{proof} \begin{proof}[Proof of $\alpha \le 1/(2d)$] Let $\widetilde I \subset (-1,1)$ be a nonempty compact interval such that \eqref{nonv} holds for $s \in \widetilde I$. Also, we fix a constant $\rho \gg1$ which we choose later. Let $\{s_\ell\}\subset \widetilde I$ be a collection of $\rho\lambda^{-1/d}$-separated points which are as many as $C\rho^{-1}\lambda^{1/d}$. Since $G(s_\ell), G'(s_\ell), \dots, G^{(d-1)}(s_\ell)$ are linearly independent in $\mathbb R^{d+1}$, there is a unit vector $\Xi_\ell \in \big( \{ G^{(j)}(s_\ell): j=0,1,\dots,d-1\} \big)^\perp$. Let $\phi \in \mathcal S(\mathbb R^{d+1})$ such that $\phi \ge1$ on $[-3r_0,3r_0]^{d+1}$ and $\widehat \phi$ is supported in $[-1,1]^{d+1}$ where $r_0=1+\sup_{s \in I} |\gamma(s)|$. Let $\varepsilon_\ell \in \{\pm 1\}$ be independent random variables. Then, we consider \[ f(x,t)=\sum_\ell \varepsilon_\ell f_\ell(x,t):= \sum_\ell \varepsilon_\ell \phi(x,t) e^{i\lambda \Xi_\ell \cdot (t,x)}. \] Since $\langle \Xi_\ell, G^{(j)}(s_\ell) \rangle=0$ for $j=0,\dots,d-1$, by Taylor's theorem we have \begin{align}\label{est2} \langle \Xi_\ell, G(s) \rangle =\frac1{d!}\langle \Xi_\ell, G^{(d)}(s_\ell) \rangle(s-s_\ell)^{d}+O(|s-s_\ell|^{d+1}). \end{align} Thus $|t\Xi_\ell \cdot G(s)| \le 2^{-2}\lambda^{-1}$ whenever $s \in I_\ell:=\{ s \in \widetilde I : |s-s_\ell| \le c \lambda^{-1/d}\}$ for a $c>0$ small enough. Noting that \begin{align}\label{est1} \mathfrak Rf_\ell(x,s)= \psi(s) e^{i\lambda \Xi_\ell \cdot (0,x)}\int \phi (x+t\gamma(s),t ) e^{i\lambda t\Xi_\ell \cdot G(s)} \chi(t)\,dt, \end{align} we see $|\mathfrak Rf_\ell(x,s)| \gtrsim1$ if $ (x,s) \in B_\ell:=B(0, c ) \times I_\ell.$ Hence, $ \sum_\ell \big\| \mathfrak Rf_\ell \big\|_{L^p(B_\ell)}^p \gtrsim \rho^{-1}. $ On the other hand, by \eqref{est1}, \eqref{est2}, and integration by parts in $t$ we have $ |\mathfrak Rf_m(x,s)| \lesssim (1+\lambda |s_\ell-s_m|^{d})^{-N} $ for any $N\ge1$ if $m\neq \ell$ and $s \in I_\ell$. Since $\{s_\ell\}$ are $\rho \lambda^{-\frac{1}{d}}$-separated, we obtain \[ \sum_\ell \big\| \sum_{m\neq \ell} |\mathfrak Rf_m| \big\|_{L^p(B_\ell)}^p \lesssim \sum_\ell \sum_{m \neq \ell} (1+\lambda |s_\ell-s_m|^{d})^{-pN} \lambda^{- 1/d} \lesssim \rho^{-pdN-1}. \] Therefore, taking $\rho,N$ sufficiently large, we have $\|\mathfrak Rf \|_p \gtrsim 1$ for any choice of $\varepsilon_\ell$. By our choice of $\phi$ it follows that $\mathcal F_x (\mathfrak Rf)$ is supported on $\{\xi : C_1\lambda \le |\xi| \le C_2 \lambda \}$ for some positive constant $C_1, C_2$. Thus, we have $\| \mathfrak Rf \|_{L_\alpha^p} \gtrsim \lambda^\alpha\|\mathfrak Rf\|_p $. Combining this with the $L^p$--$L_\alpha^p$ estimate gives $\lambda^\alpha \le C \|f\|_p $ for any choice of $\varepsilon_\ell$. By Khinchine's inequality we have $ \mathbb E (\| f \|_p^p) \sim \int (\sum_\ell | f_\ell|^2 )^{\frac p2}\,dxdt \sim C_\rho \lambda^{\frac{p}{2d}}.$ Therefore, we see $\lambda^\alpha\lesssim \lambda^\frac1{2d}$ and then $\alpha \le 1/(2d)$ taking $\lambda\to \infty$. \end{proof} \begin{proof}[Proof of $\alpha \le 1/(Lp)$] Since $\gamma$ is of type $L$ at $s_{\!\,\mathsmaller{0}}$, by affine transformation and taking $\psi$ supported near $s_{\!\,\mathsmaller{0}}$, we may assume \[ \gamma(s+s_{\!\,\mathsmaller{0}})=\gamma(s_{\!\,\mathsmaller{0}})+(s^{a_1} \varphi_1(s),\dots, s^{a_d} \varphi_d(s)) \] for $1\le a_1 <\dots<a_d=L$ and smooth functions $\varphi_j$, $j=1,\dots,d$, where $\|\varphi_j-1/a_j!\|_{\mathrm C^{a_d+1}(I)}\le c$ for a small constant $c>0$. We may also assume $s_{\!\,\mathsmaller{0}}=0$ and furthermore $\gamma(0)=0$ by replacing $f(x,t)$ by $f(x-t\gamma(0),t)$. Let $\phi_1\in \mathcal S(\mathbb R)$ such that $\phi_1 \ge1$ on $[-1,1]$, and $\supp \widehat \phi_1\subset [1/2,4]$ with $\widehat\phi_1=1$ on $[1,2]$. Let $\psi_0\in \mathrm C_c^\infty((-1,1))$ satisfy $\psi_0=1$ on $[-1/2,1/2]$. We consider \[ f(x,t)=\prod_{j=1}^{d-1} \psi_0 ( \lambda^{a_j/L}x_j ) \phi_1 ( \lambda x_{d} ) \chi(t). \] Denoting $\|a\|=\sum_{j=1}^d a_j$, we have $\|f\|_p \lesssim \lambda^{-\|a\|/(Lp)}$. Set $ E_\lambda=\big\{ (x,s) \in \mathbb R^d \times I: |x_j| \le c \lambda^{-a_j/L}, j=1,\dots,d, \quad~ |s| \le c \lambda^{-1/L} \big\} $ for a sufficiently small $c>0$. Since $\gamma(s)=(s^{a_1} \varphi_1(s),\dots, s^{a_d} \varphi_d(s))$, $ | \langle x+t\gamma(s),e_j \rangle| \le 2^{-1}\lambda^{-a_j/L},$ $j=1,\dots,d$, for $(x,s) \in E_\lambda$ and $t\in [1,2]$. So, it follows $\mathfrak Rf (x,s)\gtrsim 1$ for $(x,s) \in E_\lambda$. Hence, $\| \mathfrak Rf\|_p\gtrsim \lambda^{-(\|a\|+1)/(Lp)}$. Since $\supp \mathcal F_{x_{d}} (\mathfrak Rf) \subset \{ \xi_{d} : |\xi_{d}| \sim \lambda\}$, $ \| \mathfrak Rf \|_{L_\alpha^p} \gtrsim \lambda^{\alpha-(\|a\|+1)/(Lp)}. $ Therefore, we obtain $\alpha \le 1/(Lp)$. \end{proof} \medskip \subsection*{Acknowledgement} H. Ko was supported by NRF2019R1A6A3A01092525. S. Lee and S. Oh were supported by NRF2021R1A2B5B02001786.

https://arxiv.org/abs/1904.01714

A LeVeque-Type Inequality on the ring of $p$-adic integers

We derive an inequality on the discrepancy of sequences on the ring of $p$-adic integers $\ZZ_p$ using techniques from Fourier analysis. The inequality is used to obtain an upper bound on the discrepancy of the sequence $\alpha_n = na +b$, where $a$ and $b$ are elements of $\ZZ_p$. This is a $p$-adic analogue of the classical LeVeque inequality on the circle group $\RR/\ZZ$.

\section{Introduction} \label{section: introduction} The theory of equidistribution of sequences modulo one was initiated by Hermann Weyl in $1916$. Since then, it has spurred a lot of interest in many areas of mathematics, including number theory, harmonic analysis, and ergodic theory. The standard reference in this subject is Kuipers and Niederreiter \cite{book:MR0419394}. Equidistribution of sequences on the ring of $p$-adic integers was previously studied in \cite{article:MR0237441, article:MR0245528, article:MR0147467}. In particular, Cugiani in \cite{article:MR0147467} defines equidistribution and shows that the sequence $na + b$ is equidistributed if $a$ is a unit. Beer does a quantitative analysis in \cite{article:MR0237441} and \cite{article:MR0245528}. Our aim is to derive a LeVeque-type inequality on the discrepancy of a finite sequence using Fourier analysis. Let $| \cdot |_p$ denote the $p$-adic absolute value on ${\mathbb Q}_p$, and let \[ {\mathbb Z}_p = \{ x \,| \,\, |x|_p \leq 1 \}, \] be the ring of $p$-adic integers. Any element of ${\mathbb Z}_p$ can be given a unique canonical expansion of the form $x = a_0 + a_1p + a_2p^2 + ....$, where the $a_i$ are elements of $\{0,1,2,...,p-1\}$ (see for example \cite{book:MR1488696,book:MR2298943}). For $k \geq 0$, and $a \in {\mathbb Z}_p$, we denote by \begin{eqnarray*} D(a,1/p^k) &=& \{ x \,| \,\, |x-a|_p \leq 1/p^k\} \cr &=& a + p^k{\mathbb Z}_p, \end{eqnarray*} a disc of radius $1/p^k$ centered at $a$. Note that ${\mathbb Z}_p$ can be written as the union of $p^k$ disjoint discs of the form \[ {\mathbb Z}_p = \bigcup_{j=0}^{p^k-1} D(j,1/p^k). \] Hence, it is natural to define a notion of equidistribution using such sets. \begin{dfn} \label{dfn: equidist} A sequence $\{ \alpha_n \}$ is said to be equidistributed in ${\mathbb Z}_p$ if for every $a$ in ${\mathbb Z}_p$ and every $k \in {\mathbb N} $, we have \[ \lim_{N \to \infty} \left| \frac{\left|D(a,1/p^k) \cap \{\alpha_1,...,\alpha_N \} \right|}{N} - \frac{1}{p^k} \right| = 0. \] \end{dfn} That is, the proportion of the first $N$ elements of $\{ \alpha_n\}$ lying in a disc $D(a,1/p^k)$ is equal to its measure in the limit of large $N$, and this holds true for all such discs. This definition of equidistribution in ${\mathbb Z}_p$ was first given by Cugiani in \cite{article:MR0147467}, where propositions \ref{prop: Weyl} and \ref{prop: linear-sequence} were also proved. The details are also given in Kuipers and Niederreiter \cite{book:MR0419394}. One also wants to measure how well a sequence distributes itself. To this end, we define the notion of discrepancy to quantify the idea that some sequences are better equidistributed than others. \begin{dfn} \label{dfn: discrepancy} The discrepancy of a finite sequence $\{\alpha_1,\alpha_2,...,\alpha_N \}$ in ${\mathbb Z}_p$ is \[ D_N = \mathrm{sup}_{a \in {\mathbb Z}_p, \, k \in {\mathbb N}} \left| \frac{\left|D(a,1/p^k) \cap \{\alpha_1,...,\alpha_N \} \right|}{N} - \frac{1}{p^k} \right|. \] \end{dfn} Some elementary arguments show that \[ \frac{1}{N} \leq D_N \leq 1. \] The main aim of this paper is to prove a Fourier analytic upper bound on the discrepancy of a set of $N$ elements $\{\alpha_1,\alpha_2,...,\alpha_N \}$ in ${\mathbb Z}_p$. Let ${\mathbb Z}(p^\infty)$ denote the Pr\"{u}fer $p$-group, the group of all $p$-th power roots of unity in ${\mathbb C}$. Suppose that $\zeta \in {\mathbb Z}(p^\infty)$ has order $p^n$, and let $x \in {\mathbb Z}_p$ have the canonical expansion $x = a_0 + a_1p + a_2p^2 +.... + a_{n}p^{n} + .....$. Then we interpret the notation $\zeta^x$ as \[ \zeta^x = \zeta^{a_0 + a_1p + a_2p^2 +.... + a_{n-1}p^{n-1}}. \] Every element of ${\mathbb Z}(p^\infty)$ has finite order, and we denote the order of $\zeta \in {\mathbb Z}(p^\infty)$ by $\|\zeta\|$. \begin{thm}[Main Theorem] \label{thm: Main} The discrepancy of a finite sequence $\{\alpha_1,...,\alpha_N\}$ in ${\mathbb Z}_p$ is bounded by \begin{eqnarray*} D_N &\leq & C(p) \left(\sum_{\zeta \in {\mathbb Z}(p^\infty)\backslash \{1\} } \frac{1}{\|\zeta \|^3} \left|\frac{1}{N} \sum_{n=1}^{N} \zeta^{\alpha_n} \right|^2 \right)^{\frac{1}{4}}, \end{eqnarray*} where $C(p)$ is a constant dependent on $p$. \end{thm} As an example application of Theorem \ref{thm: Main}, we have the following corollary \begin{cor} \label{cor: Linear-sequence} The sequence $na + b$ where $a$ is a unit in ${\mathbb Z}_p$ has discrepancy \[ D_N = O\left( N^{-1/2} \right). \] \end{cor} Some quantitative results on the discrepancy of $p$-adic sequences were done by Beer in \cite{article:MR0237441} and \cite{article:MR0245528}. In particular, the author proves in \cite{article:MR0237441} that the discrepancy of the sequence $na + b$ with $a$ a unit is exactly equal to $D_N = N^{-1}$, the best possible. It is not surprising that the LeVeque type inequality gives us a weaker bound, as this is the case in the classical setting on ${\mathbb R}/{\mathbb Z}$. Montogomery in \cite{book:MR1297543} provides a detailed discussion and considers some examples. In particular, the sequence $n\theta$ where $\theta = \frac{1 +\sqrt{5} }{ 2}$ has discrepancy $D_N \ll \log(N)/N$, where as the use of the LeVeque inequality gives only $D_N \ll N^{-2/3}$. Our paper is structured as follows. In section \ref{section: fanalysis-padic} we set up the relevant Fourier analysis that is required for our calculations. We prove the main theorem in section \ref{section: Main-theorem}. We analyze the quantitative behavior of the linear sequence $\alpha_n= na + b$ in section \ref{section: Linear-sequence}, and prove Corollary \ref{cor: Linear-sequence}. \section{Fourier analysis on ${\mathbb Z}_p$} \label{section: fanalysis-padic} If $G$ is a compact abelian group, then the set of all continuous group homomorphisms (or characters) from $G$ to the multiplicative unit circle ${\mathbb T} = \{z \in {\mathbb C} \, | \, |z| = 1 \}$ forms a discrete group under multiplication, the Pontyagrin dual group $\widehat{G}$ (see for example \cite{book:MR1038803}). Note that ${\mathbb Z}_p$ is a compact abelian group. The next lemma states that the dual group $\widehat{{\mathbb Z}}_p$ of ${\mathbb Z}_p$ is isomorphic to the Pr\"{u}fer $p$-group ${\mathbb Z}(p^\infty)$. The result is known, but we include a proof due to the lack of a suitable reference. \begin{lem} \label{lem: dualgroup} For each $\zeta \in {\mathbb Z}(p^\infty)$, the map $x \mapsto \zeta^x $ is a character of ${\mathbb Z}_p$. Moreover, the map \begin{eqnarray*} \Psi: \,\, {\mathbb Z}(p^\infty) &\longrightarrow & \widehat{{\mathbb Z}}_p \cr \zeta &\longmapsto & (x \mapsto \zeta^x ) \end{eqnarray*} is an isomorphism from the Pr\"{u}fer $p$-group ${\mathbb Z}(p^\infty)$ to the Pontyagrin dual group of ${\mathbb Z}_p$. \end{lem} \begin{proof} It is easily shown that the map $x \mapsto \zeta^x$ is a character of ${\mathbb Z}_p$. To show the injectivity of $\Psi$, suppose that $\zeta_1^x = \zeta_2^x$ for all $x$ in ${\mathbb Z}_p$. Then picking $x=1$, we get $\zeta_1 = \zeta_2$. We need argue that $\Psi$ is surjective. Let $\gamma$ be in the dual group of ${\mathbb Z}_p$. Since $\gamma(0) = 1$ and $\gamma$ is continuous, there exists a disc of radius $1/p^n$ centered at zero $D = p^n{\mathbb Z}_p$, such that $|\gamma(x) - \gamma(0)|< 1$ for all $x$ in $D$, and we can pick a smallest $n$ such that this is true. Moreover, since $D$ is a subgroup of ${\mathbb Z}_p$ we must have that the image $\gamma(D)$ is a subgroup of ${\mathbb T}$. Note that there does not exist any non-trivial subgroup of ${\mathbb T}$ satisfying the condition $|x -y| < 1$ for all elements $x$ and $y$ in the subgroup. Hence, we conclude that $\gamma(D) = \{ 1 \}$. Now suppose that $\gamma(1) = \zeta = e^{2\pi i \theta}$ for some $\theta$ in $[0, 1)$. Then, $\gamma(p^n) = \gamma(1)^{p^n} = e^{2\pi i p^n \theta} = 1$ or $p^n \theta \in {\mathbb Z}$. We conclude that $\theta = \frac{m}{p^n}$, where $p \ndiv m$ by the minimality of $n$. For any integer value $k$, we have $\gamma(k) = \gamma(1)^k = \zeta^k$. This completely determines $\gamma$, since we can write ${\mathbb Z}_p$ as the union of $p^n$ disjoint balls ${\mathbb Z}_p = \cup_{k=0}^{p^n-1} \left( k + p^n{\mathbb Z}_p \right)$ and for any $x \in {\mathbb Z}_p$ we have $x = k + p^ny$, and $\gamma(x) = \gamma(k)$. We conclude that \[ \gamma(x) = \zeta^x, \] for all $x$ in ${\mathbb Z}_p$ where $\zeta = e^{\frac{2\pi i m }{p^n}}$. \end{proof} Using Lemma ~\ref{lem: dualgroup}, we shall express the Fourier series of any $f \in L^1({\mathbb Z}_p)$ in terms of the elements of ${\mathbb Z}(p^\infty)$. As a compact group there exists a normalized Haar measure $\mu$ on ${\mathbb Z}_p$ (see for example \cite{book:MR1681462}). Let $f \in L^1({\mathbb Z}_p)$. The Fourier coefficients of $f$ are given by \begin{eqnarray*} \hat{f}(\zeta) = \int_{{\mathbb Z}^p} f(x) \zeta^{-x} d\mu, \end{eqnarray*} and the Fourier inversion formula gives \[ f(x) = \sum_{\zeta \in {\mathbb Z}(p^\infty)} \hat{f}(\zeta) \zeta^x, \] whenever $\sum_{\zeta \in {\mathbb Z}(p^\infty)} |\hat{f}(\zeta)| < \infty$. A Weyl type criterion holds for equidistribution in ${\mathbb Z}_p$. We state it here in terms of the elements of ${\mathbb Z}(p^\infty)$, although it holds for a more general class of Riemann integrable functions on ${\mathbb Z}_p$ (see \cite{book:MR0419394}). \begin{prop}[Weyl's Criterion] \label{prop: Weyl} A sequence $\{ \alpha_n \}$ is equidistributed in ${\mathbb Z}_p$ if and only if for every non-trivial $\zeta$ in $ {\mathbb Z}(p^\infty) $ we have \[ \lim_{N \to \infty} \frac{1}{N} \sum_{n=1}^N \zeta^{\alpha_n} = 0. \] \end{prop} We denote by ${\mathcal X}_{D(a,R)}(x)$ the characteristic function of the disc $D(a,R)$ centered at $a$ of radius $R$. We have the following change of variables formula, the proof of which is elementary and we omit. \begin{prop} \label{prop: subformula} Let $f: {\mathbb Z}_p \to {\mathbb C}$ be an integrable function. Then \[ \int_{{\mathbb Z}_p} {\mathcal X}_{D(a,1/p^k)}(x)f(x) \, d\mu(x) = \frac{1}{p^k} \int_{{\mathbb Z}_p} f(a + p^kx) \, d\mu(x). \] \end{prop} We use Proposition \ref{prop: subformula} to calculate the Fourier coefficients of the characteristic function of a disc. \begin{lem} \label{lemma: charfun} The Fourier coefficients of the characteristic function ${\mathcal X}_{D(a,1/p^k)}(x)$ are \[ \widehat{{\mathcal X}}_{D(a,1/p^k)}(\zeta) = \left\{ \begin{array}{cc} \zeta^{-a} p^{-k} & \text{if \,\,} \|\zeta\| \leq p^k, \\ & \\ 0 & \text{if \,\,} \|\zeta\| > p^k. \end{array} \right. \] \end{lem} \begin{proof}[Proof of Lemma ~\ref{lemma: charfun}] Suppose that $ \|\zeta\| \leq p^k $, then $\zeta^{p^kx} = 1$ for all $x$ in ${\mathbb Z}_p$. Therefore, we have \begin{eqnarray*} \int_{{\mathbb Z}_p} {\mathcal X}_{D(a,1/p^k)}(x) \zeta^{-x} \, d\mu(x) &=& p^{-k} \int_{{\mathbb Z}_p} \zeta^{-(a + p^kx)} \, d\mu(x) \cr &=& \zeta^{-a} p^{-k} \int_{{\mathbb Z}_p} \zeta^{-p^kx} \, d\mu(x) \cr &=& \zeta^{-a} p^{-k}. \end{eqnarray*} On the other hand suppose $ \|\zeta\| > p^k$, and let $\omega = \zeta^{p^k}$. Then $\|\omega\| = \|\zeta\| / p^k > 1$ and hence \begin{eqnarray*} \int_{{\mathbb Z}_p} {\mathcal X}_{D(a,1/p^{k})}(x) \zeta^{-x} \, d\mu(x) &=& \zeta^{-a}p^{-k} \int_{{\mathbb Z}_p} \zeta^{-p^kx} \, d\mu(x) \cr &=& \zeta^{-a} p^{-k} \int_{{\mathbb Z}_p} \omega^{-x} \, d\mu(x) \cr &=& 0. \end{eqnarray*} \end{proof} \section{Proof of the main theorem} \label{section: Main-theorem} Let $\{\alpha_1,\alpha_2,...,\alpha_N \}$ be a finite sequence in ${\mathbb Z}_p$. Define the function $f: {\mathbb Z}_p \times {\mathbb Z}_p \longrightarrow {\mathbb R} $ \[ f(x,y) = \frac{ \left| \{\alpha_1, \alpha_2,...,\alpha_N \} \, \cap D(x,|y|_p) \right| }{N} - |y|_p, \] where $D(x,|y|_p)$ is a disc of radius $|y|_p$ centered at $x$. The discrepancy of the points $\{\alpha_1,...,\alpha_N \}$ is then \[ D_N = \mathrm{sup}_{x, y \in {\mathbb Z}_p} \, \left| f(x,y) \right|. \] We suppress the $p$ in $|\cdot|_p$ as it would be clear from the context. We can also write \begin{eqnarray*} f(x,y) &=& \frac{1}{N} \left( \sum_{n=1}^{N} {\mathcal X}_{D(x,|y|)} (\alpha_n) \right) - |y| \cr & & \cr &=& \frac{1}{N} \left( \sum_{n=1}^{N} {\mathcal X}_{D(\alpha_n, |y|)} (x) \right) - |y|. \end{eqnarray*} Our proof of Theorem \ref{thm: Main} proceeds as follows. We shall bound the $L^2$ norm $\displaystyle \|f\|_2^2 = \iint_{{\mathbb Z}_p^2} |f(x,y)|^2 \, d\mu(x) d\mu(y)$ from below by $D_N^4$ using geometrical arguments, and from above by using Parseval's theorem. The two steps are given below as lemmas \begin{lem} \label{lemma: bound1} The discrepancy $D_N$ is bounded by \[ D_N^4 \leq C_1(p) \|f\|_2^2, \] where $C_1(p)$ is a constant dependent on $p$. \end{lem} \begin{lem} \label{lemma: bound2} The $L^2$ norm of the function $f$ is bounded by \[ \|f\|_2^2 \leq C_2(p) \sum_{\zeta \in {\mathbb Z}(p^\infty)\backslash \{1 \} } \frac{1}{\|\zeta\|^3} \left|\frac{1}{N} \sum_{n=1}^{N} \zeta^{x_n} \right|^2. \] where $C_2(p)$ is a constant dependent on $p$. \end{lem} The proof of Theorem \ref{thm: Main} then follows by combining Lemmas \ref{lemma: bound1} and \ref{lemma: bound2}. \begin{rem} For $x >0$, we use the notation $\lfloor x \rfloor $ and $\lceil x \rceil$ to denote \[ \lfloor x \rfloor = \mathrm{max}\{p^k \, | \, k \in {\mathbb Z},\, p^k \leq x \} \] \[ \lceil x \rceil = \min\{p^k \, | \, k \in {\mathbb Z},\, x \leq p^k \}. \] \\ Note that $ \lfloor x \rfloor \leq x < p \lfloor x \rfloor $ and $\frac{1}{p}\lceil x \rceil < x \leq \lceil x \rceil $. \end{rem} \begin{proof}[Proof of Lemma ~\ref{lemma: bound1}] Pick a point $(x_0,y_0)$ for which $f(x_0,y_0)$ is not zero. We consider each of the two possibilities $f(x_0,y_0) > 0$ and $f(x_0,y_0) <0$ separately. Our strategy in each case is to find a small neighborhood around the point $(x_0,y_0)$ where $|f(x,y)|$ is bounded away from zero. Using this fact and integrating over this neighborhood, we produce a bound of the form $\|f\|_2^2 \geq C(p) |f(x_0,y_0)|^4$, where $C(p)$ is a constant depending only on $p$. \subsection*{Case 1} Suppose that $\Delta = f(x_0,y_0) > 0.$ Let $R = \lfloor \Delta +|y_0| \rfloor$. Since, $|y_0| < |y_0| + \Delta$ and $|y_0|$ is in the value group of ${\mathbb Q}_p$, we have $|y_0| \leq R$. We consider the two cases $|y_0| < R$ and $|y_0| = R$. \subsection*{Case 1.1:} Suppose that $|y_0| < R$. We must then have $|y_0| \leq \frac{1}{p}R$. If we fix $|y| = \frac{1}{p}R$ and $|x-x_0| \leq \frac{1}{p}R$, then $D(x_0,|y_0|) \subseteq D(x,|y|)$. We get a nonnegative lower bound on $f(x,y)$ as follows \begin{eqnarray*} f(x,y) &=& \frac{1}{N} \sum_{n=1}^N {\mathcal X}_{D(x,|y|) }(\alpha_n) - |y| \cr &\geq& \frac{1}{N} \sum_{n=1}^N {\mathcal X}_{D(x_0,|y_0|) }(\alpha_n) - |y| \cr &=& |y_0| + f(x_0,y_0) - |y| \cr &=& |y_0| + \Delta - |y| \cr &\geq& \left(1 - \frac{1}{p} \right)R. \end{eqnarray*} We can bound the $L^2$ norm of $f$ from below by evaluating the required integral only on the set $|y| = \frac{1}{p} R, |x-x_0| \leq \frac{1}{p}R $ \begin{eqnarray*} \|f\|_2^2 &=& \iint_{{\mathbb Z}_p^2} |f(x,y)|^2 \, d\mu(x)d\mu(y) \cr &\geq& \iint_{|y|=\frac{1}{p}R, |x-x_0| \leq \frac{1}{p}R} |f(x,y)|^2 \, d\mu(x)d\mu(y) \cr &\geq & \iint_{|y|=\frac{1}{p}R, |x-x_0| \leq \frac{1}{p}R} \left( 1 - \frac{1}{p}\right)^2 R^2 \, d\mu(x)d\mu(y) \cr &=& \left( 1 - \frac{1}{p}\right)^3 \frac{1}{p^2} R^4 \cr &\geq& \left( 1 - \frac{1}{p}\right)^3 \frac{1}{p^6}\Delta^4 \cr &=& \frac{(p-1)^3 }{p^9} \Delta^4, \end{eqnarray*} using $R = \lfloor \Delta +|y_0| \rfloor \geq \frac{1}{p}(|y_0| + \Delta) \geq \frac{1}{p} \Delta$. \subsection*{Case 1.2: } Suppose that $|y_0|=R$. If we let $|y| = R$ and $|x-x_0| \leq R$, then $D(x_0,|y_0|) = D(x,|y|)$. From this, we get \begin{eqnarray*} f(x,y) &=& \frac{1}{N} \sum_{n=1}^N {\mathcal X}_{D(x,|y|) }(\alpha_n) - |y| \cr &=& \frac{1}{N} \sum_{n=1}^N {\mathcal X}_{D(x_0,|y_0|) }(\alpha_n) - |y_0| \cr &=& |y_0| + f(x_0,y_0) - |y| \cr &=& f(x_0,y_0) \cr &=& \Delta. \end{eqnarray*} Therefore, \begin{eqnarray*} \|f\|_2^2 &=& \iint_{{\mathbb Z}_p^2} |f(x,y)|^2 \, d\mu(x)d\mu(y) \cr &\geq& \iint_{|y|= R, |x-x_0| \leq R} \Delta^2 \, d\mu(x)d\mu(y) \cr &=& \left(1- \frac{1}{p} \right) R^2 \Delta^2 \cr &\geq& \left(1- \frac{1}{p} \right) \frac{1}{(p-1)^2} \Delta^4 \cr &=& \frac{1}{p(p-1)} \Delta^4, \end{eqnarray*} using $R + \Delta = |y_0| + \Delta < p\lfloor |y_0| + \Delta \rfloor = pR$ and therefore $\Delta < (p-1)R$. Finally, since $\frac{(p-1)^3}{p^9} < \frac{1}{p(p-1)}$ we conclude \[ \| f\|^2 \geq \frac{(p-1)^3}{p^9} \Delta^4 \] holds in both cases 1.1 and 1.2, so it holds in general for case 1. \subsection*{Case 2:} Suppose that $f(x_0,y_0) < 0$ and $ \displaystyle \Delta = |f(x_0,y_0)| = -f(x_0,y_0). $ In other words, the disc $D(x_0,|y_0|)$ contains fewer than the expected number of points $\alpha_n$. Now let $R = |y_0|$. Then if $|y| = R$ and $|x-x_0| \leq R$, by the strong triangle inequality $D(x,|y|) = D(x_0,|y_0|) $ and we have \begin{eqnarray*} f(x,y) &=& \frac{1}{N} \sum_{n=1}^N {\mathcal X}_{D(x,|y|) }(\alpha_n) - |y| \cr &=& \frac{1}{N} \sum_{n=1}^N {\mathcal X}_{D(x_0,|y_0|) }(\alpha_n) - |y_0| \cr &=& f(x_0,y_0). \end{eqnarray*} Therefore, \begin{eqnarray*} \|f\|_2^2 &=& \iint_{{\mathbb Z}_p^2} |f(x,y)|^2 \, d\mu(x)d\mu(y) \cr &\geq& \iint_{|y|=R, |x-x_0| \leq R} |f(x,y)|^2 \, d\mu(x)d\mu(y) \cr &=& \iint_{|y|=R, |x-x_0| \leq R} \Delta^2 \, d\mu(x)d\mu(y) \cr &=& \left(1- \frac{1}{p}\right) R^2 \Delta^2 \cr &\geq& \left(1- \frac{1}{p}\right)\Delta^4, \end{eqnarray*} where the last line follows because $\Delta \leq R$. To see this, note that \begin{eqnarray*} \Delta &=& - f(x_0,y_0) \cr &=& |y_0| - \frac{1}{N} \sum_{n=1}^N {\mathcal X}_{D(x_0,|y_0|) }(\alpha_n) \cr &\leq& |y_0| \cr &=& R. \end{eqnarray*} \end{proof} Next, we need to prove Lemma \ref{lemma: bound2}. Our goal is to find an upper bound on the $L^2$-norm of $f(x,y)$ using Parseval's theorem. Suppose $f(x,y)$ has a Fourier series \[ f(x,y) = \sum_{\zeta, \omega \, \in {\mathbb Z}(p^\infty) } \hat{f} (\zeta,\omega) \zeta^x \, \omega^y. \] Then by Parseval's theorem we would get \[ \|f \|_2^2 = \sum_{\zeta, \omega \, \in {\mathbb Z}(p^\infty) } |\hat{f}(\zeta,\omega)|^2. \] Therefore, we need to bound the Fourier coefficients of $f(x,y)$. The Fourier coefficients are \begin{eqnarray} \label{eqn: FourierCoefficients} \hat{f}(\zeta,\omega) & = & \iint_{{\mathbb Z}_p^2} f(x,y) \, \zeta^{-x}\omega^{-y} \,\, d\mu(x) d\mu(y) \cr & = & \frac{1}{N} \sum_{n=1}^N \iint_{{\mathbb Z}_p^2} {\mathcal X}_{D(\alpha_n, |y|)}(x) \, \zeta^{-x}\omega^{-y} \,\, d\mu(x) d\mu(y) \,\, \cr & & - \, \iint_{{\mathbb Z}_p^2} |y| \, \zeta^{-x}\omega^{-y} \,\, d\mu(x) d\mu(y). \end{eqnarray} Note that if $\zeta = 1$ we get \begin{eqnarray} \label{eqn: zerocoefficients} \hat{f}(1,\omega) &=& \frac{1}{N} \sum_{n=1}^N \iint_{{\mathbb Z}_p^2} {\mathcal X}_{D(\alpha_n, |y|)}(x) \, \omega^{-y} \,\, d\mu(x) d\mu(y) \,\, - \, \int_{{\mathbb Z}_p} |y| \,\omega^{-y} \,d\mu(y) \cr &=& \frac{1}{N} \sum_{n=1}^N \int_{{\mathbb Z}_p} |y| \omega^{-y} \, d\mu(y) \, - \int_{{\mathbb Z}_p} |y| \, \omega^{-y} \,d\mu(y) \cr &=& 0. \end{eqnarray} When $\zeta \neq 1$, the second integral in line 2 of Equation (\ref{eqn: FourierCoefficients}) is zero \begin{eqnarray*} \iint_{{\mathbb Z}_p^2} |y| \, \zeta^{-x}\omega^{-y} \,\, d\mu(x) d\mu(y) &=& \int_{{\mathbb Z}_p} |y| \omega^{-y} \left(\int_{{\mathbb Z}_p} \zeta^{-x} \, d\mu(x) \right) \, d\mu(y) \cr &=& 0. \end{eqnarray*} Therefore, \[ \hat{f}(\zeta, \omega) = \frac{1}{N} \sum_{n=1}^N \iint_{{\mathbb Z}_p^2} {\mathcal X}_{D(\alpha_n, |y|)}(x) \, \zeta^{-x}\omega^{-y}\,\, d\mu(x) d\mu(y). \] Using Lemma \ref{lemma: charfun}, we have \[ \int_{{\mathbb Z}_p} {\mathcal X}_{D(\alpha_n,|y|)}(x) \zeta^{-x} \, d\mu(x) = \left\{ \begin{array}{cc} \zeta^{-\alpha_n}\,|y| & \text{if \,} \|\zeta \| \leq 1/|y|, \\ & \\ 0 & \text{else}. \end{array} \right. \] Hence, for $\zeta \neq 1$, \[ \hat{f}(\zeta, \omega) = \frac{1}{N} \sum_{n=1}^N \zeta^{-\alpha_n} \, \int_{|y| \leq 1/\|\zeta\| } |y| \, \omega^{-y} \, d\mu(y). \] The following lemma makes some estimates that are useful in our succeeding calculations \begin{lem} \label{lemma: IntegralEst} Let $R = p^k, k \in {\mathbb Z}$ satisfy $0 < R < 1$, and let $\omega \in {\mathbb Z}(p^\infty).$ Then, \begin{equation} \label{eqn: IntegralEst1} \left|\int_{|y| \leq R} |y| \, \omega^{-y} \, d\mu(y) \right| \leq \frac{p}{\mathrm{max}(1/R, \|\omega \|)^2}. \end{equation} Moreover, \begin{equation} \label{eqn: IntegralEst2} \sum_{\omega \in {\mathbb Z}(p^\infty)} \left|\int_{|y| \leq R} |y| \, \omega^{-y} \, d\mu(y) \right|^2 \leq 2p^2R^3. \end{equation} \end{lem} \begin{proof}[Proof of Lemma \ref{lemma: IntegralEst}] Let $R = 1/p^k$ and let $\|\omega\| = p^l$. We have \begin{eqnarray} \label{eqn: IntegralEst3} \int_{|y| \leq R} |y| \, \omega^{-y} \, d\mu(y) &=& \sum_{j \geq k} \frac{1}{p^j} \int_{|y| = 1/p^j } \omega^{-y} \, d\mu(y) \cr &=& \sum_{j \geq k} \frac{1}{p^j} \int_{{\mathbb Z}_p} \left({\mathcal X}_{D(0,1/p^j) }(y) - {\mathcal X}_{D(0,1/p^{j+1}) }(y)\right) \omega^{-y} \, d\mu(y). \end{eqnarray} When $\|\omega\| \leq 1/R$, that is when $l \leq k$, using Lemma \ref{lemma: charfun} and (\ref{eqn: IntegralEst3}) we have \begin{eqnarray*} \int_{|y| \leq R} |y| \, \omega^{-y} \, d\mu(y) &=& \sum_{j \geq k} \frac{1}{p^j}\left(\frac{1}{p^j} - \frac{1}{p^{j+1}} \right) \cr &=& \frac{p}{(p+1)(1/R)^2}. \end{eqnarray*} Thus (\ref{eqn: IntegralEst1}) holds in this case. If $\|\omega\| > 1/R$, that is when $l \geq k+1$, again using Lemma \ref{lemma: charfun} and (\ref{eqn: IntegralEst3}) we have \begin{eqnarray*} \int_{|y| \leq R} |y| \, \omega^{-y} \, d\mu(y) &=& \sum_{j \geq l} \frac{1}{p^j}\left(\frac{1}{p^j} - \frac{1}{p^{j+1}} \right) - \frac{1}{p^{l-1}p^l} \cr &=& -\frac{p^2}{(p+1)\|\omega\|^2}, \end{eqnarray*} and thus (\ref{eqn: IntegralEst1}) holds in this case as well. To check (\ref{eqn: IntegralEst2}) , we use the fact that for each $j \geq 1$, the group ${\mathbb Z}(p^\infty)$ contains $p^j$ elements of order at most $p^j$, and $p^j - p^{j-1}$ elements of order exactly $p^j$. We then have \begin{eqnarray*} \sum_{\omega \in {\mathbb Z}(p^\infty)} \left|\int_{|y| \leq R} |y| \, \omega^{-y} \, d\mu(y) \right|^2 &\leq& p^2 \sum_{\omega \in {\mathbb Z}(p^\infty)} \frac{1}{\mathrm{max}(1/R, \|\omega \|)^4} \cr &=& p^2\left(\sum_{\|\omega\| \leq 1/R} R^4 + \sum_{\|\omega\| > 1/R} \frac{1}{\|\omega\|^4}\right) \cr &=& p^2 \left(R^3 + \frac{p-1}{p(p^3-1)}R^3 \right) \cr &<& 2p^2R^3. \end{eqnarray*} \end{proof} Finally, we prove Lemma \ref{lemma: bound2}. \begin{proof}[Proof of Lemma \ref{lemma: bound2}] Applying Parseval's theorem to $f(x,y)$ and using Equations (\ref{eqn: zerocoefficients}) and (\ref{eqn: IntegralEst2}), we conclude \begin{eqnarray*} \|f\|^2 &=& \sum_{\zeta, \omega \in {\mathbb Z}(p^\infty) \atop \zeta \neq 1} |\hat{f}(\zeta,\omega)|^2 \cr &=& \sum_{\zeta \in {\mathbb Z}(p^\infty) \atop \zeta \neq 1} \left( \sum_{\omega \in {\mathbb Z}(p^\infty)} \left| \int_{|y| \leq 1/\|\zeta\|} |y|\omega^{-y} \, d\mu(y) \right|^2 \right) \left|\frac{1}{N} \sum_{n= 1}^N \zeta^{\alpha_n} \right|^2 \cr &\leq& 2p^2 \sum_{\zeta \in {\mathbb Z}(p^\infty) \atop \zeta \neq 1} \frac{1}{\| \zeta\|^3} \left|\frac{1}{N} \sum_{n= 1}^N \zeta^{\alpha_n} \right|^2. \end{eqnarray*} \end{proof} \section{The linear sequence $na+b$ in ${\mathbb Z}_p$} \label{section: Linear-sequence} Consider the sequence $\alpha_n = na + b $. We have the following proposition, a proof of which is given in \cite{book:MR0419394} using elementary number theory. We present an alternate proof using Fourier analysis. \begin{prop} \label{prop: linear-sequence} The sequence $\alpha_n = n a + b$ is equidistributed in ${\mathbb Z}_p$ if and only if $a$ is a unit in ${\mathbb Z}_p$. \end{prop} \begin{proof} The forward implication follows from Weyls criterion (Proposition \ref{prop: Weyl}). For suppose, $a$ was not a unit. Then $a = p^kc$, where $k>0$ and $c$ is a unit. Now let $\zeta = e^{2\pi i/p^k}$. Then $\zeta^a =1$, and Weyl's criterion will not hold. For the reverse implication, let $\zeta \in {\mathbb Z}(p^\infty)$ with $\| \zeta \| = p^k$ for $k \geq 1$. There exists an $m$ such that $1 \leq m < p^k $, with $p \ndiv m$ and $\displaystyle \zeta = e^{2\pi im/p^k}$. Suppose that $a $ is a unit in ${\mathbb Z}_p$. Let $a = t_0 + t_1p + t_2p^2 +...$ be the canonical expansion of $a$, with $t_0 \neq 0$. Then we let $a_k = t_0 + t_1p + ... + t_{k-1}p^{k-1}$ be the truncation of this expansion to the first $k$ terms. We have \begin{eqnarray} \frac{1}{N}\left| \sum_{n=1}^{N} \zeta^{na + b} \right| &=& \frac{1}{N}\left| \sum_{n=1}^{N} \zeta^{na} \right| \cr & & \cr &=& \frac{1}{N} \left| \frac{ 1- \zeta^{(N+1)a_k} }{1 - \zeta^{a_k}} \right| \cr & & \cr \label{eqn: lin-sequence} &\leq& \frac{1}{N} \frac{2}{\left|1 - \zeta^{a_k} \right| } \cr &\leq& \frac{1}{N} \left| \frac{1}{\sin(\pi m a_k/p^k)} \right|. \end{eqnarray} Since $p \ndiv m$ and $p \ndiv a_k$, $\displaystyle \sin(\pi m a_k/p^k) \neq 0$ and hence $\frac{1}{N} \sum_{n=1}^{N} \zeta^{na + b} \to 0$ as $N \to \infty$; the proof of equidistribution now follows from Weyl's criterion. \end{proof} \begin{proof}[Proof of Corollary~\ref{cor: Linear-sequence}] Applying the bound given by Theorem \ref{thm: Main} we get \begin{eqnarray} \label{eqn: CorEqn1} D_N^4 &\ll& \sum_{\zeta \in {\mathbb Z}(p^\infty)\backslash \{1\} } \frac{1}{\|\zeta \|^3} \left|\frac{1}{N} \sum_{n=1}^{N} \zeta^{ na + b} \right|^2 \cr & & \cr &\leq& \frac{1}{N^2} \sum_{k=1}^\infty \frac{1}{p^{3k}} \sum_{1 \leq m < p^k \atop p \nmid m} \frac{1}{ \left|\sin \left(\pi \, m a_k/ p^k \right) \right|^2} \cr & & \cr &\leq& \frac{1}{N^2} \sum_{k=1}^\infty \frac{1}{p^{3k}} \sum_{1 \leq m < p^k} \frac{1}{ \left|\sin \left(\pi \, m a_k/ p^k \right) \right|^2} \cr & & \cr &\leq& \frac{1}{N^2} \sum_{k=1}^\infty \frac{1}{p^{3k}} \sum_{1 \leq l < p^k} \frac{1}{ \left|\sin \left(\pi \, l/ p^k \right) \right|^2} \cr & & \cr &\leq& \frac{2}{N^2} \sum_{k=1}^\infty \frac{1}{p^{3k}} \sum_{1 \leq l \leq p^k/2} \frac{1}{ \left|\sin \left(\pi \, l/ p^k \right) \right|^2}. \end{eqnarray} Note that the second inequality in Equation (\ref{eqn: CorEqn1}) comes from the last inequality in (\ref{eqn: lin-sequence}). For the fourth inequality, note that since $a$ is a unit we have $p \ndiv a_k$. Hence, $\gcd(a_k,p^k) =1$ and so $a_k$ generates ${\mathbb Z}/p^k{\mathbb Z}$. That is, ${\mathbb Z}/p^k{\mathbb Z} = \left\{ma_k \, | \, m = 0,..,p^k-1 \right\}$. The final inequality follows from the identities $|\sin(\theta)| = |\sin(-\theta)| = |\sin(\pi - \theta)|$, so that for $p^k/2 \leq l <p^k$ we have $|\sin(\pi l /p^k)| = |\sin(\pi(p^k-l)/p^k) |$. This allows us to double the sum over the first half of the interval. Note that in the interval $[0,\, \pi/2]$, $\sin(\theta)$ is bounded from below by $2\theta/\pi$, so that \[ \frac{1}{|\sin(\theta)|} \leq \frac{\pi}{2\theta}. \] This gives us \begin{eqnarray} \label{eqn: CorEqn2} \sum_{1 \leq l \leq p^k/2} \frac{1}{ \left|\sin \left(\pi \, l/ p^k \right) \right|^2} &\leq& \sum_{1 \leq l \leq p^k/2} \frac{p^{2k} }{4 l^2 } \cr & & \cr &\leq& \frac{p^{2k}}{4} \sum_{1 \leq l <\infty} \frac{1}{l^2} \cr & & \cr &\leq& \frac{p^{2k} \pi^2}{24}. \end{eqnarray} Finally, applying the bound from (\ref{eqn: CorEqn2}) to (\ref{eqn: CorEqn1}) we get \begin{eqnarray*} D_N^4 &\ll& \frac{\pi^2}{12N^2} \sum_{k=1}^{\infty} \frac{1}{p^{k}}. \end{eqnarray*} We conclude that $D_N = O\left(\frac{1}{\sqrt{N}}\right)$. \end{proof} \bibliographystyle{siam}

https://arxiv.org/abs/2103.11330

Expected Extinction Times of Epidemics with State-Dependent Infectiousness

We model an epidemic where the per-person infectiousness in a network of geographic localities changes with the total number of active cases. This would happen as people adopt more stringent non-pharmaceutical precautions when the population has a larger number of active cases. We show that there exists a sharp threshold such that when the curing rate for the infection is above this threshold, the mean time for the epidemic to die out is logarithmic in the initial infection size, whereas when the curing rate is below this threshold, the mean time for epidemic extinction is infinite. We also show that when the per-person infectiousness goes to zero asymptotically as a function of the number of active cases, the mean extinction times all have the same asymptote independent of network structure. Simulations bear out these results, while also demonstrating that if the per-person infectiousness is large when the epidemic size is small (i.e., the precautions are lax when the epidemic is small and only get stringent after the epidemic has become large), it might take a very long time for the epidemic to die out. We also provide some analytical insight into these observations.

\section{Conclusion} \label{sec:conclusion} We have developed a model for epidemic spread within and across population centers with state-dependent infectiousness. In this model, we directly prove (without mean-field assumptions) that there exists a sharp threshold for the curing rate \(\delta\) such that when \(\delta\) is more than a threshold, the epidemic dies out quickly (the mean lifetime is of logarithmic order in the initial infection size), and when \(\delta\) is less than the threshold, the mean lifetime of the epidemic is infinite. Although \(\delta\) is not typically something we can control, especially in the initial stages of a pandemic without vaccines or other medication, it is possible to lower the threshold by following more stringent precautions. While we do not provide prescriptive solutions for managing pandemics, we hope that this work would offer useful insights to policymakers. { While our model makes no mean-field assumptions to characterize the extinction time, we provide theoretical results only on its expected value. It is of interest to establish high-probability bounds on extinction time and characterize how strongly extinction time concentrates. Combining techniques in Claim~\ref{claim:ctmc-countable} with literature on (discrete-time) Markov concentration \cite{Kotzing2016, Paulin2015} might be pursued. There is also scope for developing broader and more realistic models of state-dependent infectiousness. Empirical work suggests that people take precautions against contagions not only in response to the actual number of infections, but also to other factors like the media attention on infection prevalence \cite{FenichelKC2013, SpringbornCMF2015}. These models should capture infectiousness as a function of both the actual infection prevalence and the spread of awareness through (social) media. Finally, it is important to accurately infer parameters of our model using historical and current epidemiological data so as to inform practical applications. } \section{Proof of Claim~\ref{claim:constant-infection-spectral-lower}} \label{appendix:constant-infection-spectral-lower} The following proof that we provide here closely resembles the proof of \cite[Theorem 3.1]{GaneshMT2005}. However, since we are not interested in the exact constant \(C\) like \cite{GaneshMT2005}, we avoid the use of matrix exponentials seen there. The rates of \eqref{eq:node-rates} (with constant \(\beta(\cdot)\) and \(\beta^\textsc{int}(\cdot)\)) give us the following differential equation: \begin{align*} \frac{d\ensuremath{\mathbb{E}}\left[\mathbf{X}(t)\right]}{dt}= \Big( \beta G + \beta^\textsc{int}I - \delta I \Big)\ensuremath{\mathbb{E}}\left[\mathbf{X}(t)\right], \end{align*} where \(I\) is the identity matrix (of correct size). Multiply each side of the equation with \(\mathbf{q}^\top\) (\(\mathbf{q}\) is the eigenvector of \(G\) corresponding to \(\lambda_r\)). This gives us \begin{align} \frac{d\ensuremath{\mathbb{E}}\left[\mathbf{q}^\top\mathbf{X}(t)\right]}{dt}= \mathbf{q}^\top\Big(\beta G + \beta^\textsc{int}I - \delta I\Big)\ensuremath{\mathbb{E}}[\mathbf{X}(t)]. \label{eq:preceding-equation} \end{align} Since \(\mathbf{q}\) is an eigenvector of \(G\) with eigenvalue \(\lambda_{r}\), and an eigenvector of \(I\) with eigenvalue \(1\) (every vector is an eigenvector of \(I\) with eigenvalue \(1\)), \eqref{eq:preceding-equation} gives us \begin{align*} \frac{d\ensuremath{\mathbb{E}}\left[\mathbf{q}^\top\mathbf{X}(t)\right]}{dt}= \Big(\beta \lambda_{r} + \beta^\textsc{int} - \delta \Big) \ensuremath{\mathbb{E}}\left[\mathbf{q}^\top\mathbf{X}(t)\right]. \end{align*} This is a differential equation in terms of \(\ensuremath{\mathbb{E}}\left[\mathbf{q}^\top\mathbf{X}(t)\right]\), and solving it gives us \begin{align*} \ensuremath{\mathbb{E}}\left[\mathbf{q}^\top\mathbf{X}(t)\right] = e^{t\left(\beta\lambda_{r}+\beta^\textsc{int}-\delta\right)}\mathbf{q}^\top\mathbf{X}(0). \end{align*} Let \(q_{\max}\) and \(q_{\min}\) denote the maximum and minimum elements of \(\mathbf{q}\), i.e., \(q_{\max}=\max_{i}{q}_i\) and \(q_{\min}=\min_{i}{q}_i\). Since \(\mathbf{q}\succ0\), \(q_{\min}\) is strictly positive. This gives us \begin{align} \ensuremath{\mathbb{E}}[X(t)] = \ensuremath{\mathbb{E}}\left[\mathbf{1}^\top\mathbf{X}(t)\right] \le e^{t\left(\beta\lambda_{r}+\beta^\textsc{int}-\delta\right)}\frac{q_{\max}n}{q_{\min}}. \label{eq:xt-exponential-bound} \end{align} The mean hitting time can be written as \begin{align*} \ensuremath{\mathbb{E}}\left[T_{\mathbf{X}(0)}\right] &= \int_{0}^\infty \ensuremath{\mathbb{P}}(X(t) > 1) dt \\ &= \int_0^\tau \ensuremath{\mathbb{P}}(X(t)>1) dt + \int_\tau^\infty \ensuremath{\mathbb{P}}(X(t)>1) dt \\ &\le \tau + \int_\tau^\infty \ensuremath{\mathbb{E}}[X(t)] dt \end{align*} for any \(\tau>0\). The last inequality follows from the fact that \(\ensuremath{\mathbb{P}}(X(t)>1)\le1\) since it is a probability (which gives the first term), and the Markov inequality which gives us \(\ensuremath{\mathbb{P}}(X(t)>1)\le\ensuremath{\mathbb{E}}[X(t)]\) (for the second term). Using \eqref{eq:xt-exponential-bound}, we get \begin{align*} \ensuremath{\mathbb{E}}[T_{\mathbf{X}(0)}] \le \tau+kne^{-\tau\Delta} \quad \text{for all}\quad \tau > 0, \end{align*} where \(k=\frac{q_{\max}}{q_{\min}(\delta-\beta\lambda_{r}-\beta^\textsc{int})}>0\) and \(\Delta=\delta-\beta\lambda_{r}-\beta^\textsc{int}>0\). Setting \(\tau=\frac{\ln n}{\Delta}\) gives us \(\ensuremath{\mathbb{E}}[T_{\mathbf{X}(0)}] \le C \ln n\). \section{Proof of Claim~\ref{claim:ctmc-countable}} \label{appendix:ctmc-countable} Divide the time axis into intervals of unit length. Given any (finite) \(t\in\mathbb{T}\), if the number of transitions in all intervals preceding and including \(t\) is finite, then the cardinality of the set \(\{s \mid s\in\mathbb{T}\ \text{and}\ s<t\}\) is finite. Further, this cardinality is unique for each \(t\), allowing us to map \(t\) to this unique natural number plus one. Thus we get an injective mapping (if the number of transitions in each interval is finite). At the start of the interval, assume that the Markov chain starts in state \(\mathbf{X}\) with \(\mathbf{1}^\top\mathbf{X}=n\). The probability that there are at least \(k\) transitions in the interval satisfies \begin{align} \ensuremath{\mathbb{P}}(\text{at least \(k\) transitions in interval}) \le \ensuremath{\mathbb{P}}\left(\sum_{j=0}^{k-1}X_j \le 1\right), \label{eq:ktransitionsle1} \end{align} where \(\{X_j\}\) are the amounts of time it takes to transition out of the first \(k\) states starting from \(\mathbf{X}\) at the beginning of the interval. Since the total rate of transition rate out of \(\mathbf{X}\) is given by \(\mathbf{1}^\top\left(\beta(n)\mathbf{G}+\beta^\textsc{int}(n)\mathbf{I}+\delta\mathbf{I}\right)\mathbf{X}\), the total transition rate out of any state with at most \(n\) infections is less than or equal to \((\beta_{\max}d_{\max}+\beta^\textsc{int}_{\max}+\delta)n\). Recall that \(\beta_{\max}=\sup_{i\in\ensuremath{\mathbb{N}}}\beta(i)\), \(\beta^\textsc{int}_{\max}=\sup_{i\in\ensuremath{\mathbb{N}}}\beta^\textsc{int}(i)\), and \(d_{\max}\) is the maximum degree among nodes of \(\mathcal{G}\). Define \(\tau=\beta_{\max}d_{\max}+\beta^\textsc{int}_{\max}+\delta\). So in the worst case, which gives the greatest probability on the right side of \eqref{eq:ktransitionsle1}, we have \(X_j\sim\exp(\tau(n+j))\). This gives us \begin{align*} \ensuremath{\mathbb{P}}\left(\sum_{j=0}^{k-1}X_j \le 1\right) &\le \ensuremath{\mathbb{P}}\left(e^{-\sum_{j=0}^{k-1}X_j}\ge e^{-1}\right) \\ &\le e\prod_{j=0}^{k-1}\ensuremath{\mathbb{E}}\left[e^{-X_j}\right] \\ &= e\prod_{j=0}^{k-1}\frac{\tau n+\tau j}{1+\tau n+\tau j} \\ &=\frac{e}{\prod_{j=0}^{k-1}\left(1+\frac{1}{\tau n+\tau j}\right)}. \end{align*} If \(\prod_{j=0}^{k-1}\left(1+\frac{1}{\tau n+\tau j}\right)\to\infty\) as \(k\to\infty\), then the probability that there are infinite transitions in the interval goes to \(0\). But this is equivalent to \(\sum_{j=0}^{k-1}\ln\left(1+\frac{1}{\tau n+\tau j}\right)\to \infty\) as \(k\to\infty\). This gives us \begin{align*} \sum_{j=0}^{k-1}\ln\left(1+\frac{1}{\tau n+\tau j}\right) &= \sum_{j=0}^{k-1}\frac{\ln \left(1+\frac{1}{\tau n+\tau j}\right)}{\frac{1}{\tau n+\tau j}}\frac{1}{\tau n+\tau j}. \end{align*} For a large enough \(j\), we can make \(\frac{\ln\left(1+\frac{1}{\tau n+\tau j}\right)}{\frac{1}{\tau n+\tau j}}\) arbitrarily close to \(1\). This implies \begin{align*} \sum_{j=0}^{k-1}\frac{\ln\left(1+\frac{1}{\tau n+\tau j}\right)}{\frac{1}{\tau n+\tau j}}\!\frac{1}{\tau n+\tau j}\!\! >\!\! \left(1-\epsilon\right)\!\!\sum_{j=l}^{k-1}\frac{1}{\tau n+\tau j}\to\infty, \end{align*} where \(l\) is chosen to be large enough so that \(\frac{\ln\left(1+\frac{1}{\tau n+\tau j}\right)}{\frac{1}{\tau n+\tau j}}\) is at most \(\epsilon\) away from \(1\) for all \(j\ge l\). The sum goes to infinity because the sum of the harmonic series goes to infinity. Since this ensures that the Markov chain only has a finite number of transitions in any interval, it concludes the proof. \section{Proof of Claim~\ref{claim:epsilon-bound}} \label{appendix:epsilon-bound} Since \(\lim_{n\to\infty}\gamma(n)=0\), for any \(\epsilon > 0\), we can find an \(m_\epsilon\) such that for all \(n > m_\epsilon\), \(\gamma(n) < \epsilon\). Let \(T_{i,j}\) denote the time it takes to go from \(i\) infections to \(j\) infections (for the first time). Then we have \begin{align*} \ensuremath{\mathbb{E}}[T_n] = \ensuremath{\mathbb{E}}[T_{n,m_\epsilon}] + \ensuremath{\mathbb{E}}[T_{m_\epsilon}]. \end{align*} But the birth rate of the Markov chain between \(n\) and \(m_\epsilon\) is less than \(\epsilon\) (from the definition of \(m_\epsilon\)). So \(\ensuremath{\mathbb{E}}[T_{n,m_\epsilon}]\) should be less than the expected time to go from \(n\) to \(0\) in a Markov chain where all the birth rates are \(\epsilon\). This gives us (using Claim~\ref{claim:tn-constant}): \begin{align*} \ensuremath{\mathbb{E}}[T_n] \le \frac{\ln n}{\delta - \epsilon} + \ensuremath{\mathbb{E}}[T_{m_\epsilon}] + \frac{1}{\delta-\epsilon}. \end{align*} Since \(\ensuremath{\mathbb{E}}[T_{m_\epsilon}]\) depends only on \(\epsilon\) given a \(\gamma(\cdot)\), this concludes the proof for the second inequality. The first inequality is relatively straightforward since \(\frac{\ln (n+1)}{\delta}\) is the lower bound in Claim~\ref{claim:tn-constant} if the birth rate was \(0\) throughout. \section{Extinction time exponential in equilibrium point} \label{appendix:exponential-equal} For simplicity, we just consider the upper- and lower-bound Markov chains using the rates from \eqref{eq:rates} defined using the \(\gamma(\cdot)\) function here. We expect similar arguments to hold for the network-wide epidemic as well. Let \(\gamma(n)=\delta+\epsilon\) for all \(n \le N\) and \(\gamma(n)=0\) for all \(n > N\). Since this satisfies the condition of Theorem~\ref{thm:log-hitting}, we are guaranteed that the mean epidemic extinction time is logarithmic in the initial infection size. However, the mean extinction time also turns out to be exponential in \(N\), the ``equilibrium point,'' or the size of the epidemic where the rate of infectiousness \(\gamma(\cdot)\) goes below the curing rate \(\delta\). To see this, substitute these values into the expression for \(\ensuremath{\mathbb{E}}[T_1]\) from Claim~\ref{claim:t1-gamma}. We get { \begin{align*} \ensuremath{\mathbb{E}}[T_1] &= \frac{1}{\delta}\sum_{i=1}^{N+1}\frac{1}{i}\left(\frac{\delta+\epsilon}{\delta}\right)^{i-1} \\ &\ge\frac{1}{(N+1)\delta}\sum_{i=1}^{N+1}\left(1+\frac{\epsilon}{\delta}\right)^{i-1} \end{align*} } \vspace{-0.3em} \begin{align*} =\frac{\left(1+\frac{\epsilon}{\delta}\right)^{N+1}-1}{(N+1)\epsilon}.\quad\ \end{align*} If \(N\) or \(\epsilon\) are large enough, \(\ensuremath{\mathbb{E}}[T_1]\) is greater than an exponential of the form \(a^N\) for some \(a>1\). This implies that the mean die-out time is exponential in the equilibrium point \(N\). \section{Extension to General Networks} \label{sec:extension} So far, we have assumed that the connection graph among the population centers is symmetric (\(G_{uv}=G_{vu}\)) and unweighted (\(G_{uv}\in\{0,1\}\)). However, this is not true for many real-world networks: the rate of infection spread between any two connected centers need not be identical, and the rate of infection spread from \(u\) to \(v\) need not be equal to the rate of infection spread from \(v\) to \(u\) for a connected pair \((u,v)\). Thus, it is important to study the behavior of the epidemic under a general connection network given by a general asymmetric, (nonnegative) real-valued adjacency matrix \(G\).% \footnote{ Rather than defining the graph \(\mathcal{G}\) as a set \(\{(u,v)\}\) of node pairs, we now define it as a set of triples \(\{(u,v,e_{uv})\}\), where \(e_{uv}\) is the weight of the edge from \(u\) to \(v\). The adjacency matrix \(G\) concisely captures all this information. } However, it is still reasonable to assume that the graph is strongly connected, i.e., there exists a path with nonzero edges from any center \(u\) to any other center \(v\). This is because it is rarely the case that there exist no paths from one population center to another. Further, the intra-locality growth rate of infections need not be identical for all the population centers, as this rate typically depends on local factors like population density \cite{WongL2020} and social capital \cite{VarshneyS2020}. Let us use the parameter \(D_u>0\) to modulate the growth rate of the infection at location \(u\). Let \(D\) be a diagonal matrix with \(D_u\) as the \(u\)th element of its diagonal. These considerations give us the following expressions for the rates of epidemic spread. \begin{align} \mathbf{X}(t) &\to \mathbf{X}(t) + \mathbf{e}_u \nonumber \\ &\quad\quad\quad \text{at rate}\ \left[\Big(\beta(X(t))G+\beta^\textsc{int}(X(t))D\Big)\mathbf{X}(t)\right]_u,\nonumber \\ \mathbf{X}(t) &\to \mathbf{X}(t) - \mathbf{e}_u \ \text{at rate}\ \delta\left[\mathbf{X}(t)\right]_u, \label{eq:general-rates} \end{align} for all \(u\in\mathcal{L}\), where \(\left[\cdot\right]_u\) indicates the \(u\)th element of a vector. Let \(\rho(\cdot)\) denote the spectral radius of a matrix. We generalize Theorem~\ref{thm:spectral-bound} as Theorem~\ref{thm:general-bound}. \begin{theorem} Let the system start in some state \(\mathbf{X}\) that has \(n\) infections cumulatively, i.e., \(\mathbf{1}^\top\mathbf{X}=n\). For the epidemic described by \eqref{eq:general-rates}, the following hold. \begin{itemize} \item[(i)] If \(\rho\left(\beta_\infty G+\beta^\textsc{int}_\infty D\right)<\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right] \le C \ln n\) for some constant \(C>0\). \item[(ii)] If \(\rho\left(\beta_\infty G+\beta^\textsc{int}_\infty D\right)>\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right]=\infty\). \end{itemize} \label{thm:general-bound} \end{theorem} \begin{IEEEproof} Note that the Perron-Frobenius theorem holds for the matrix \(\beta_\infty G + \beta^\textsc{int}_\infty D\), and we can find a strictly positive eigenvector \(\mathbf{q}'\succ0\) of \(\beta_\infty G + \beta^\textsc{int}_\infty D\), which has the (positive, real) eigenvalue \(\rho\left(\beta_\infty G + \beta^\textsc{int}_\infty D\right)\) (see \cite{GraphSpectraBook}). The proof follows directly by replacing the \(\mathbf{q}\) used in Claim~\ref{claim:constant-infection-spectral-lower} and Theorem~\ref{thm:spectral-bound} with the Perron-Frobenius eigenvector of \(\beta_\infty G + \beta^\textsc{int}_\infty D\). \end{IEEEproof} While Theorem~\ref{thm:general-bound} provides a sharp threshold in terms of \(\rho\left(\beta_\infty G + \beta^\textsc{int}_\infty D\right)\), it is difficult to separate the contributions of the between-locality spreading term \(\beta_\infty G\) and the intra-locality spreading term \(\beta^\textsc{int}_\infty D\). It would be nice to have sufficient conditions for fast die-out and long-lasting epidemic in terms of expressions where these two contributions are decoupled. Towards this end, we provide two corollaries. \begin{corollary} Let the system start in some state \(\mathbf{X}\) that has \(n\) infections cumulatively, i.e., \(\mathbf{1}^\top\mathbf{X}=n\). If \(D\) is a scalar matrix \(\eta I\), i.e., if the intra-locality rate-modulating factor \(D_u=\eta\) for every locality \(u\), then the following hold. \begin{itemize} \item[(i)] If \(\beta_\infty\rho(G)+\beta^\textsc{int}_\infty\eta<\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right] \le C \ln n\) for some constant \(C>0\). \item[(ii)] If \(\beta_\infty\rho(G)+\beta^\textsc{int}_\infty\eta>\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right]=\infty\). \end{itemize} \label{cor:intra-constant} \end{corollary} \begin{IEEEproof} Please see Appendix~\ref{appendix:intra-constant}. \end{IEEEproof} For the next corollary, we need a theorem from \cite{Schwenk1986} which relates the spectral radius of nonnegative asymmetric matrices to the spectral radius of certain symmetric matrices. We state this as Claim~\ref{claim:schwenk} (in a form useful for us). \begin{claim}[from \cite{Schwenk1986}] For any nonnegative (square) matrix \(A\), \begin{align*} \rho\left(\sqrt{A \odot A^\top}\right) \le \rho(A) \le \rho\left(\frac{A+A^\top}{2}\right), \end{align*} where \(\odot\) is the element-wise product of matrices and \(\sqrt{\cdot}\) is the element-wise square root. \label{claim:schwenk} \end{claim} Note that the \(ij\)th element of \(\sqrt{A \odot A^\top}\) is \(\sqrt{A_{ij}A_{ji}}\) and the \(ij\)th element of \(\frac{A+A^\top}{2}\) is \(\frac{A_{ij}+A_{ji}}{2}\). Both \(\sqrt{A \odot A^\top}\) and \(\frac{A+A^\top}{2}\) are symmetric matrices. This reduction to symmetric matrices allows us to apply Weyl's inequalities on the conditions in Theorem~\ref{thm:general-bound}. We state this formally as Corollary~\ref{cor:weyl-derived}. See the textbook \cite{Bhatia1997} for the details regarding Weyl's inequalities. We also provide short proofs of the inequalities used here in Appendix~\ref{appendix:weyl-proof}. \begin{corollary} Let the system start in some state \(\mathbf{X}\) that has \(n\) infections cumulatively, i.e., \(\mathbf{1}^\top\mathbf{X}=n\). Then the following hold. \begin{itemize} \item[(i)] If \(\beta_\infty\rho\left(\frac{G+G^\top}{2}\right)+\beta^\textsc{int}_\infty\max_uD_u<\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right] \le C \ln n\) for some constant \(C>0\). \item[(ii)] If \(\beta_\infty\rho\left(\sqrt{G \odot G^\top}\right)+\beta^\textsc{int}_\infty\min_uD_u>\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right]=\infty\). \end{itemize} \label{cor:weyl-derived} \end{corollary} \begin{IEEEproof} Applying the upper bound in Claim~\ref{claim:schwenk} to the spectral-radius expression in part (i) of Theorem~\ref{thm:general-bound}, we get \begin{align*} \rho\left(\beta_\infty G + \beta^\textsc{int}_\infty D\right) \le \rho\left(\beta_\infty\tfrac{G+G^\top}{2}+\beta^\textsc{int}_\infty D\right). \end{align*} Since \(\beta_\infty\frac{G+G^\top}{2}\) and \(\beta^\textsc{int}_\infty D\) are both symmetric matrices, we can apply one of Weyl's inequalities (see \cite{Bhatia1997} or Appendix~\ref{appendix:weyl-proof}) to get \begin{align} \rho\left(\beta_\infty G + \beta^\textsc{int}_\infty D\right) \le \beta_\infty \rho\left(\tfrac{G+G^\top}{2}\right) + \beta^\textsc{int}_\infty \max_u D_u. \label{eq:weyl-upper-in-proof} \end{align} Equation~\eqref{eq:weyl-upper-in-proof} ensures that whenever the condition in part (i) of Corollary~\ref{cor:weyl-derived} is satisfied, the condition in part (i) of Theorem~\ref{thm:general-bound} is satisfied as well. This proves part (i) of Corollary~\ref{cor:weyl-derived}. For part (ii) of Corollary~\ref{cor:weyl-derived}, observe that \begin{align*} (\beta_\infty G + \beta^\textsc{int}_\infty D)&\odot (\beta_\infty G + \beta^\textsc{int}_\infty D)^\top \\ &= \beta_\infty^2 G \odot G^\top + (\beta^\textsc{int}_\infty)^2 D \odot D. \end{align*} This is because there is no position \(ij\) that has a nonzero element in both the matrices \(G\) and \(D\). Further, the matrix \(D\) is diagonal (and hence symmetric), and so we have \begin{align*} &\sqrt{ (\beta_\infty G + \beta^\textsc{int}_\infty D)\odot (\beta_\infty G + \beta^\textsc{int}_\infty D)^\top } \\ &\qquad\qquad\qquad\qquad = \beta_\infty\sqrt{G \odot G^\top} + \beta^\textsc{int}_\infty D. \end{align*} Using the lower bound in Claim~\ref{claim:schwenk}, we get \begin{align*} \rho\left(\beta_\infty G + \beta^\textsc{int}_\infty D\right) \ge \rho\left(\beta_\infty\sqrt{G \odot G^\top}+\beta^\textsc{int}_\infty D\right). \end{align*} Since \(\beta_\infty\sqrt{G \odot G^\top}\) and \(\beta^\textsc{int}_\infty D\) are both symmetric matrices, we can apply another one of Weyl's inequalities (see \cite{Bhatia1997} or Appendix~\ref{appendix:weyl-proof}) to get \begin{align*} \rho\left(\beta_\infty G + \beta^\textsc{int}_\infty D\right) \ge \beta_\infty\rho\left(\sqrt{G \odot G^\top}\right) + \beta^\textsc{int}_\infty \min_u D_u. \end{align*} Thus, whenever the condition in part (ii) of Corollary~\ref{cor:weyl-derived} is true, the condition in part (ii) of Theorem~\ref{thm:general-bound} is true as well. This concludes the proof of part (ii) of Corollary~\ref{cor:weyl-derived}. \end{IEEEproof} Unlike Theorem~\ref{thm:spectral-bound}, Theorem~\ref{thm:general-bound}, and Corollary~\ref{cor:intra-constant} where the thresholds are sharp, there is a gap between the thresholds for a quick die-out and long-lasting epidemic in Corollary~\ref{cor:weyl-derived}. However, Corollary~\ref{cor:weyl-derived} decouples the contributions of the graph structure \(G\) and the variation in intra-locality spreading \(D\) in the thresholds. \section{Proof of Corollary~\ref{cor:intra-constant}} \label{appendix:intra-constant} We need to show that \begin{align*} \rho\left(\beta_\infty G + \beta^\textsc{int}_\infty \eta I\right) = \beta_\infty \rho(G) + \beta^\textsc{int}_\infty\eta. \end{align*} Recall that the spectral radius of a matrix is defined as the maximum absolute value of the eigenvalues of the matrix. Let \(\lambda\) be an eigenvalue of \(\beta_\infty G + \beta^\textsc{int}_\infty\eta I\). This yields \begin{align*} \text{det}(\beta_\infty G + \beta^\textsc{int}_\infty \eta I - \lambda I)=0, \end{align*} or \begin{align*} \text{det}(\beta_\infty G - (\lambda-\beta^\textsc{int}_\infty\eta)I)=0. \end{align*} This implies \(\lambda-\beta^\textsc{int}_\infty\eta\) is an eigenvalue of \(\beta_\infty G\) for every eigenvalue \(\lambda\) of \(\beta_\infty G + \beta^\textsc{int}_\infty\eta I\). The Perron-Frobenius theorem (see \cite{GraphSpectraBook}) guarantees that there exists a positive eigenvalue of \(\beta_\infty G\) which has the maximum absolute value. Thus the maximum absolute value of \(\lambda\) is \(\beta_\infty\rho(G)+\beta^\textsc{int}_\infty\eta\). \section{Introduction} \label{sec:intro} Newly emerging infectious diseases that quickly spread across population centers in an increasingly interconnected world form a large portion of human infections \cite{MorensF2013}. These epidemics spread over contact networks and the characteristics of this spread have been widely studied \cite{ShirleyR2005, Newman2002, GiordanoBBCDDC2020, ColizzaBBV2006, KuchlerRS2020, ZhouWZSYHZOPS2020, GomezABMM2010}. In this work, we develop a state-dependent infectiousness model for the spread of epidemics over a network of population centers and analytically prove that the epidemic dynamics follow certain properties. Specifically, we characterize the expected time of epidemic extinction and show that it exhibits a threshold behavior where it is either logarithmic in the initial infection size or infinite depending on whether the curing rate is higher or lower than a threshold. We make no mean-field assumption while deriving this threshold. We believe our model captures important features of epidemic spreading not captured in prior literature, and our results advance the understanding of epidemic spread. We model the epidemic as a Markov spreading process over a network whose nodes represent population centers such as cities or large communities, and the connections between them indicate the amount of contact between the population centers. New infections could either be due to interactions with people from neighboring population centers, or due to community spread within the population center. We model these two components of epidemic spread separately. In a typical epidemic, especially in the early stages of newly emerging infections, vaccines and other pharmaceutical means to combat the disease are unlikely to be available. Further, in the early stages of the epidemic, the number of susceptible people in a typical population center is very large, and effectively infinite, until a large majority of the population has developed herd immunity. We capture these properties in a model where the number of infections in each population center can potentially grow without bound. {In cases where the infected population is a significant fraction of the total population, the epidemic would spread more slowly than what is predicted by our model. This is because for a given number of infected individuals, our model assumes that the susceptible population is larger than it actually is. So our model would over-estimate the effective rate at which the contagion spreads, and the number of (new) infections in our model stochastically dominates the actual number of infections.} Thus, in those settings, the threshold obtained from our model would still hold for the quick-extinction case. Whereas models at the person level \cite{GaneshMT2005, FagnaniZ2017, VanMieghemOK2008, SahnehVMS2017, SahnehCS2012} capture interactions between individual people and might help us predict the probability of a particular person getting infected, it is prohibitively expensive to collect information about all individuals in a city and compute over a network that treats each person as a distinct node. Population~center-level models allow us to predict the epidemic trajectory over a much larger number of people at the level of countries or even the world. Related work on \emph{metapopulation} \cite{ColizzaV2007, ColizzaPV2007, ColizzaV2008, WangL2014} also develops population~center-level models, but uses a mean~field-type approximation, which \emph{assumes} the existence of a (sharp) threshold and finds it. In contrast, sharp thresholds \emph{emerge} in our work. Like \cite{GaneshMT2005, FagnaniZ2017}, we directly characterize the time it takes for epidemic extinction. But unlike \cite{GaneshMT2005, FagnaniZ2017}, where there is a gap between the conditions for a short- and long-lasting epidemic, we prove there is a sharp threshold for the curing rate which separates the conditions for short- and long-lasting epidemics.% \footnote{Note that the ``mean~field'' described in \cite{FagnaniZ2017} is over the network, not the infection probabilities.} Note that our model is at the population~center-level (compared to the person-level model in \cite{GaneshMT2005, FagnaniZ2017}). Besides the work on metapopulation, other prior work which claim a sharp threshold between the two regimes \cite{VanMieghemOK2008, ChakrabartiWWLF2008, SahnehVMS2017, SahnehCS2012} have assumed it and employed a mean~field-type approximation. {Our analysis is significantly different from the analysis of the extinction time of the mean-field dynamics. The advantage of this stochastic-analysis framework is that it allows the possibility of obtaining tail bounds for the extinction time, whereas the existing mean-field models, in their current form, do not offer that scope. While we do not present tail bounds on extinction time in this work, in Sec.~\ref{sec:numerical-computations}, we plot the confidence bounds on the extinction times obtained from simulations of the stochastic dynamics.} Another key aspect of our model is that the per-person infectiousness of the epidemic is a function of the number of active cases in the system. { State-dependent infectiousness influences the epidemic trajectory as people tend to take more precautions \cite{YanMBFCO2021, FenichelKC2013, SpringbornCMF2015} and governments tend to impose more restrictions on travel, gatherings, etc. \cite{Bourassa2021, YanMBFCO2021} as the number of active cases increases. Moreover, these changes in contact can be well-described using changes in the parameters of standard epidemiological compartment models \cite{ChangPKGRGL2020}; models that incorporate these considerations may yield predictions that are significantly different from models that do not \cite{FenichelCCCPHHHMPSVV2011}. } As explained in \cite{Sattenspiel1990}, modeling the effects of human behavior on epidemic spread is necessary for realistic models. Although time-dependent infectiousness has been studied empirically in \cite{ChenLCL2020}, we analytically model infectiousness as a function of the number of active cases in the system, which provides a (tractable) theoretical basis to time-varying infectiousness. A person-level model for state-dependent infectiousness has been developed in \cite{FagnaniZ2017}, but as explained earlier, modeling the epidemic at the population~center-level allows us to predict the epidemic trajectory over a much larger number of people. We prove the population~center-level model has a sharp epidemic threshold for the extinction times, in contrast to the gap between the conditions for short- and long-lasting epidemics in \cite{FagnaniZ2017}. Related to this are \cite{SahnehVMS2017} and \cite{SahnehCS2012}, which develop person-level models where individual people get \emph{alerted} in the presence of infected neighbors and take more precautions or change their contacts. Other related work on epidemic extinction time include \cite{Holme2013} which estimates extinction time in SIR networks using simulations; \cite{Khatri2020} which calculates the extinction-time distribution in an aggregate non-network model; \cite{HolmeT2018} which computes the mean extinction times for all possible configurations of small networks; \cite{HindesS2016, ChenHZL2017} which use the Wentzel-Kramers-Brillouin approximation; and \cite{BallH2017} which characterizes the epidemic extinction times over a ``mean'' network formed from a given degree distribution. To summarize, our main contribution is a sharp, analytical, and direct (not mean-field) characterization of the extinction time in a population~center-level model with state-dependent infectiousness. This, to the best of our knowledge, is new. The remainder of this paper is organized as follows. Sec.~\ref{sec:model} describes our model. Under this model, Sec.~\ref{sec:sharp-threshold} proves the existence of a sharp threshold: if the curing rate \(\delta\) is greater than this threshold, the mean time for epidemic die-out starting from a state with a cumulative of \(n\) infections is of order \(\ln n\), and if the curing rate is below this threshold, the mean die-out time is infinite. Sec.~\ref{sec:extension} generalizes the results to settings with asymmetric and weighted graphs. Then Sec.~\ref{sec:vanishing-infectiousness} proves that the asymptotic mean extinction time is (exactly) equal to \(\frac{\ln n}{\delta}\) independent of graph structure if the per-person infectiousness functions go to zero asymptotically. This would happen if the level of precautions people take to combat the epidemic keep getting more stringent with increasing numbers of active cases. Sec.~\ref{sec:numerical-computations} provides simulation and computation results, and Sec.~\ref{sec:conclusion} concludes. \section{Model} \label{sec:model} Let there be a set of \emph{localities}% \footnote{Note that ``localities'' can refer to population centers at various levels of demographic aggregation. They could represent countries, states, cities, or even neighborhoods within a city. Indeed, there can be marked differences in how people react to a contagion even within a single large urban area \cite{KimCCJL2017}. } \(\mathcal{L}\), and at each locality \(u \in \mathcal{L}\), the number of infected people at time \(t\) is given by \(X_u(t)\). We assume each locality has a large enough population that for our purposes, for all \(u\), the range of \(X_u(t)\) is the set of all non-negative integers. There is a graph \(\mathcal{G}\) across the localities, and \((u, v) \in \mathcal{G}\) when the localities \(u\) and \(v\) are \emph{connected}. The adjacency matrix \(G\) of \(\mathcal{G}\) is the matrix having \(G_{uv}=1\) if \((u,v)\in\mathcal{G}\) and \(G_{uv}=0\) otherwise. For ease of presentation, we first assume that the graph is symmetric: \((u,v)\in\mathcal{G}\) implies \((v,u)\in\mathcal{G}\). We relax this assumption in Sec.~\ref{sec:extension}. A connection between two localities means that infected people in one locality can infect susceptible people in the other locality. Further, we assume the graph \(\mathcal{G}\) is connected, i.e., for every \(u,v\in\mathcal{L}\), there exists a path between \(u\) and \(v\) in \(\mathcal{G}\). Let the total number of people infected at time \(t\) be \(X(t)\), i.e., \(\sum_{u\in\mathcal{L}}X_u(t) = X(t)\). The rate of growth of the infection at locality \(u\) at time \(t\) consists of two components: \begin{enumerate} \item the intra-locality growth rate due to interactions within the locality given by \(\beta^\textsc{int}(X(t))X_u(t)\), and \item the between-locality growth rate, where the rate of growth due to \(v\) for each \((u,v)\in\mathcal{G}\) is given by \(\beta(X(t))X_v(t)\). \end{enumerate} Here, \(\beta(\cdot)\) and \(\beta^\textsc{int}(\cdot)\) are positive real-valued functions of the total number of infections in the system, which give the rate of growth of the infection per infecting agent. We assume that these per-person infectiousness functions are bounded. Let their suprema be given by \(\sup_{n\in\ensuremath{\mathbb{N}}}\beta(n)=\beta_{\max}\) and \(\sup_{n\in\ensuremath{\mathbb{N}}}\beta^\textsc{int}(n)=\beta^\textsc{int}_{\max}\). We also assume that the asymptotic limits for these functions exist as the total number of infections grows without bound: \(\lim_{n \to \infty}\beta(n)=\beta_\infty\) and \(\lim_{n \to \infty}\beta^\textsc{int}(n)=\beta^\textsc{int}_\infty\). Let the curing rate for every infected agent be \(\delta\). This is independent of the graph \(\mathcal{G}\), the level of precautions taken (\(\beta\) and \(\beta^\textsc{int}\)), or the number of infections at any node \(\{X_u(t)\}\), and just depends on the nature of the infection. We show the model pictorially in Fig.~\ref{fig:model}. \begin{figure}[h] \begin{center} \resizebox{0.8\columnwidth}{!}{ \begin{tikzpicture} \filldraw [black] (-1,0) circle (1.5pt); \filldraw [black] (1,0) circle (1.5pt); \filldraw [black] (0.3,0.6) circle (1.5pt); \filldraw [black] (-0.3,-0.6) circle (1.5pt); \begin{scope}[>=stealth, decoration={markings,mark=at position 0.25 with {\arrow[red]{<}}, mark=at position 0.75 with {\arrow[red]{>}}}] \draw[postaction={decorate}] (-1,0) -- (1,0); \draw[postaction={decorate}] (-1,0) -- node[above, red, midway, sloped] {\tiny \(\beta(X(t))\)} (0.3,0.6); \draw[postaction={decorate}] (1,0) -- (0.3,0.6); \draw[postaction={decorate}] (1,0) -- (-0.3,-0.6); \end{scope} \begin{scope}[black!30!brown, >=stealth, decoration={markings, mark=at position 1 with {\arrow{>}}}] \draw[postaction={decorate}] (-1,0.1) -- (-1,0.3); \draw[postaction={decorate}] (-0.3,-0.5) -- (-0.3,-0.3); \draw[postaction={decorate}] (0.3,0.7) -- (0.3,0.9); \draw[postaction={decorate}] (1,0.1) -- (1,0.3); \end{scope} \node at (0.2, 0.9) [anchor=west, black!30!brown, yshift=0.07cm] {\tiny \(\beta^\textsc{int}(X(t))\)}; \draw[->, thick, black!30!green, >=stealth] (1.5, 0.5) -- node[anchor=west] {\scriptsize \(\delta\)} (1.5, -0.5); \end{tikzpicture} } \end{center} \caption{The epidemic model, where nodes represent population centers, edges represent the connections between the centers, \(\beta(\cdot)\) and \(\beta^\textsc{int}(\cdot)\) are the between-locality and intra-locality infectiousness functions, and \(\delta\) is the curing rate. } \label{fig:model} \end{figure} For each \(u\in\mathcal{L}\), the above discussion implies the following rates for the infection: \begin{align} X_u(t) &\rightarrow X_u(t)+1\ \nonumber\\ \qquad\qquad&\text{at rate}\ \sum_{v:(u,v)\in\mathcal{G}}\!\!\!\beta(X(t))X_v(t) + \beta^{\textsc{int}}(X(t))X_u(t), \nonumber\\ X_u(t) &\rightarrow X_u(t)-1\ \text{at rate}\ \delta X_u(t). \label{eq:node-rates} \end{align} Let us use the vector \(\mathbf{X}(t)\) to denote the state of the system at time \(t\). The \(u\)th element of \(\mathbf{X}(t)\) is \(X_u(t)\), the number of infections at node \(u\) at time \(t\). Let \(T_\mathbf{X}\) denote the time it takes to go from a state \(\mathbf{X}\) to the all-zero state \(\mathbf{0}\). Once the epidemic reaches the all-zero state, it is \emph{extinct,} since one can only contract the infection from someone else,% \footnote{Note that this is true for many viral infections since the only host for these viruses are humans. However, this may not be the case for other infections.} and if there are no infected individuals, the epidemic can never rebound later. The mean extinction time (also called the mean \emph{hitting time}) starting from the state \(\mathbf{X}\) is given by \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right]\). \section{Simulations \& Numerical Computations} \label{sec:numerical-computations} In this section, we present some simulations and numerical computations to demonstrate the theoretical results of the preceding sections. \subsection{Network-wide simulations} For simulations, we use the network from \cite{ColizzaPV2007} which is a graph where the nodes represent the top \(500\) US airports and the edge weights are the number of seats scheduled on flights between the airports in the year 2002. We consider the top \(100\) of these \(500\) nodes and normalize the adjacency matrix with the mean column weight (this normalization just scales the values of \(\beta(\cdot)\)). We simulate the model described in Sec.~\ref{sec:model} using Gillespie's algorithm \cite{Gillespie1977}. \begin{center} \begin{figure} \subfloat[Curing rate below the threshold.]{\label{fig:graph-constant:below} \includegraphics[width=\columnwidth]{{r0.90}}} \subfloat[Curing rate above the threshold.]{\label{fig:graph-constant:above} \includegraphics[width=\columnwidth]{{r1.10}}} \caption{Epidemic trajectories using constant values for \(\beta(\cdot)\) and \(\beta^\textsc{int}(\cdot)\).} \label{fig:graph-constant} \end{figure} \end{center} First, in Fig.~\ref{fig:graph-constant}, we simulate using constant values for \(\beta(\cdot)\) and \(\beta^\textsc{int}(\cdot)\). Specifically, we set \(\beta(n)=\beta=2\) and \(\beta^\textsc{int}(n)=\beta^\textsc{int}=2\) for all \(n\), and choose \(\delta\) to get the value of \(\frac{\delta}{\beta\lambda_r+\beta^\textsc{int}}\) shown on the plot. For both the values of \(\delta\), we simulate the system {\(1000\)} times and show the trajectories of \(X(t)\) over time in the plot, {and the interval that contains \(95\%\) of the simulated states at each time instant. We obtain this \(95\%\) interval by finding the maximum and minimum state values after ignoring the top and bottom \(2.5\%\) of the simulations. }% We also show the plot of \(\ensuremath{\mathbb{E}}[X(t)]\) computed theoretically by solving the differential equation for \(\frac{d\ensuremath{\mathbb{E}}[X(t)]}{dt}\) (see Appendix~\ref{appendix:constant-infection-spectral-lower}). { As we can see in Fig.~\ref{fig:graph-constant:below}, when \(\beta\lambda_r+\beta^\textsc{int}>\delta\), most of the simulated trajectories of the system show an epidemic that is not dying out. Even though more than \(2.5\%\) of the simulations die out (as the \(95\%\) interval shows), since most of the simulations show an epidemic that becomes increasingly larger with time, the expected extinction time would be infinite, in line with what we have theoretically proven in Theorem~\ref{thm:spectral-bound}. On the other hand, in Fig.~\ref{fig:graph-constant:above}, when \(\beta\lambda_r+\beta^\textsc{int}<\delta\), all the trajectories of the system result in the epidemic dying out relatively quickly. Further, in this case, the confidence bounds on the extinction time are meaningfully defined, and we show the \(95\%\) confidence interval of the extinction time \(T_\mathbf{X}\) in Fig.~\ref{fig:graph-constant:above}. This interval is calculated in the same way as the \(95\%\) interval for the state trajectory. }% For all the simulations, we start with an initial epidemic size of \(100\), placed uniformly at random at one of the nodes. \begin{center} \begin{figure} \subfloat[Curing rate is greater than threshold when epidemic is small.]{ \includegraphics[width=\columnwidth]{{early-drop}} } \subfloat[Curing rate is greater than threshold only after epidemic gets very large.]{ \includegraphics[width=\columnwidth]{{late-drop}} } \caption{Epidemic trajectories when \(\beta(n)\) and \(\beta^\textsc{int}(n)\) change with \(n\).} \label{fig:graph-variable-drop} \end{figure} \end{center} When the values of \(\beta(n)\) and \(\beta^\textsc{int}(n)\) change with \(n\), if \(\beta(n)\lambda_r+\beta^\textsc{int}(n)<\delta\) or \(\beta(n)\lambda_r+\beta^\textsc{int}(n)>\delta\) for all \(n\), then the results are very similar to the ones in Fig.~\ref{fig:graph-constant}, and hence we omit these plots. In Fig.~\ref{fig:graph-variable-drop}, we show the results of simulations where \(\beta(n)\lambda_r+\beta^\textsc{int}(n)\) starts from a value greater than \(\delta\) for small \(n\), but eventually falls to a value smaller than \(\delta\) for larger \(n\). The value of \(n\) where this transition happens is shown on the plots in Fig.~\ref{fig:graph-variable-drop}. We can see in Fig.~\ref{fig:graph-variable-drop} that there seems to be a ``metastable'' state at the point where the infectiousness is equal to the curing rate. Note that since the value of \(\beta(n)\lambda_r+\beta^\textsc{int}(n)\) eventually falls below \(\delta\) for large enough \(n\), the condition in part (i) of Theorem~\ref{thm:spectral-bound} is true, and so the mean hitting time should be logarithmic in the initial infection size. However, these simulations suggest that the epidemic takes a very long time to die out in this case. It seems that the die-out times are in fact exponential in the infection size where the infectiousness and curing rate are equal. Please see Appendix~\ref{appendix:exponential-equal} for some insight into this behavior. This means that even though Theorem~\ref{thm:spectral-bound} guarantees that the mean die-out time would be logarithmic in the initial infection size if the asymptotic rate of infectiousness is less than the curing rate, it is still very important that measures such as lockdowns and other non-pharmaceutical precautions are implemented in the early stages of an epidemic. \subsection{Numerical computations for vanishing \(\gamma(\cdot)\)} Here, we provide some numerical computations to support Theorem~\ref{thm:log-hitting}. Note that in contrast to the network-wide simulations in Fig.~\ref{fig:graph-constant} and \ref{fig:graph-variable-drop} where we have used the infectiousness functions \(\beta(\cdot)\) and \(\beta^\textsc{int}(\cdot)\), we use \(\gamma(\cdot)\) here which captures the infectiousness for both the upper- and lower-bound Markov chains together in a single expression using \eqref{eq:rates}. We consider three different \(\gamma(\cdot)\) functions and plot the values of \(\ensuremath{\mathbb{E}}[T_n]\) computed using the recursion from \eqref{eq:sn-recursion} (with the base case from Claim~\ref{claim:t1-gamma}). We plot this in Fig.~\ref{fig:multiplicative-factor}. \begin{figure} \begin{center} \includegraphics[width=0.9\columnwidth]{multiplicative-factor} \caption{\(\ensuremath{\mathbb{E}}[T_n]\) for slightly different \(\gamma(\cdot)\) functions.} \label{fig:multiplicative-factor} \end{center} \end{figure} Fig.~\ref{fig:multiplicative-factor} shows that even small changes in \(\gamma(\cdot)\) can cause large changes in the values of \(\ensuremath{\mathbb{E}}[T_n]\). Further, Fig.~\ref{fig:multiplicative-factor} may seem to indicate that even these small changes cause the mean hitting times to not converge to the same asymptote. This would be contrary to what we expect from Theorem~\ref{thm:log-hitting}. However, the reason we do not see all the three curves in Fig.~\ref{fig:multiplicative-factor} converge to the same asymptote is that the convergence happens extremely slowly. This is not very surprising, given that the asymptote is the function \(\frac{\ln n}{\delta}\). Since the logarithmic function increases very slowly, differences between \(\ensuremath{\mathbb{E}}[T_n]\) for different \(\gamma(\cdot)\) functions at small values of \(n\) take a very long time to become insignificant, and the \(\ensuremath{\mathbb{E}}[T_n]\) values become close to each other only at very large values of \(n\). To demonstrate this, consider \(\gamma(n)=\frac{k}{n}\). We choose this function because it leads to easier analysis. Similar arguments hold for any other function as well. Substituting this into Claim~\ref{claim:t1-gamma} gives us \begin{align} \ensuremath{\mathbb{E}}[T_1] &= \frac{1}{\delta} + \frac{k}{1\cdot2\cdot\delta^2} + \frac{k^2}{1\cdot2\cdot3\cdot\delta^3} + \cdots \nonumber \\ &= \frac{1}{k}\left(\frac{k}{\delta}+\frac{k^2}{2!\delta^2}+\frac{k^3}{3!\delta^3} + \cdots\right) \nonumber \\ &= \frac{e^{k/\delta}-1}{k}. \label{eq:t1-kbyn} \end{align} Equation~\eqref{eq:t1-kbyn} is quite sensitive to the value of \(k\). For example, with \(\delta=1\), we get a derivative of \(\frac{4e^5+1}{25}\approx23.79\) at \(k=5\). Small changes in the value of \(k\) can significantly change the value of \(\ensuremath{\mathbb{E}}[T_1]\). We can use the recursion from \eqref{eq:sn-recursion} to analytically find the value of \(\ensuremath{\mathbb{E}}[T_2]\) to find that \(\ensuremath{\mathbb{E}}[T_2]\) is even more sensitive to the value of \(k\). Since \(\ensuremath{\mathbb{E}}[T_n]\) is of the form \(\ensuremath{\mathbb{E}}[T_2]+\sum_{i=3}^nS_i\), and \(S_n\) asymptotically reaches \(\frac{1}{n}\), these differences in \(\ensuremath{\mathbb{E}}[T_2]\) become negligible only for a very large value of \(n\). We can verify this using Fig.~\ref{fig:difference} where we plot the values of \(S_n\) for different \(\gamma(\cdot)\) functions. We see that all of them eventually reach the asymptote \(\frac{1}{n}\). This means that for large enough \(n\), the mean hitting times will all be indistinguishable from \(\ln n\). However, we need an extremely large value of \(n\) for the differences to become negligible. \begin{figure} \begin{center} \includegraphics[width=0.9\columnwidth]{difference} \caption{\(S_n\) for large \(n\). Note that for these large values of \(n\), \(\gamma(n)\) is too small for accurately computing the recursion in \eqref{eq:sn-recursion} using even 128-bit floating point arithmetic. To compute \(S_n\), we need to divide a very small value, \(S_{n-1}\delta-\frac{1}{n-1}\), by another very small value, \(\gamma(n-1)\). Insufficient numerical precision can lead to garbage values for \(S_n\). Arbitrary-precision arithmetic (such as the one provided by \texttt{mpmath} \cite{mpmath}) is needed.} \label{fig:difference} \end{center} \end{figure} \section{Sharp Threshold} \label{sec:sharp-threshold} In this section, we state our main result as Theorem~\ref{thm:spectral-bound}. \begin{theorem} Let \(\lim_{n\to\infty}\beta(n)=\beta_\infty\) and \(\lim_{n\to\infty}\beta^\textsc{int}(n)=\beta^\textsc{int}_\infty\). Let \(\lambda_r\) denote the spectral radius of the adjacency matrix of the (symmetric) undirected graph \(\mathcal{G}\). Let the system start in some state \(\mathbf{X}\) that has \(n\) infections cumulatively, i.e., \(\mathbf{1}^\top\mathbf{X}=n\). If the graph \(\mathcal{G}\) is connected, then the following hold. \begin{itemize} \item[(i)] If \(\beta_\infty\lambda_r+\beta^\textsc{int}_\infty<\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right] \le C \ln n\) for some constant \(C>0\). \item[(ii)] If \(\beta_\infty\lambda_r+\beta^\textsc{int}_\infty>\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right]=\infty\). \end{itemize} \label{thm:spectral-bound} \end{theorem} Before we prove Theorem~\ref{thm:spectral-bound}, let us first observe a property of the spectral radius, \(\lambda_r\), of \(G\). Since we have assumed that \(\mathcal{G}\) is connected, the Perron-Frobenius theorem (see \cite{GraphSpectraBook}) implies that every element of the eigenvector \(\mathbf{q}\) corresponding to \(\lambda_{r}\) is strictly positive, i.e., \(\mathbf{q}\succ0\). We prove Theorem~\ref{thm:spectral-bound} in two parts: (i) \(\beta_\infty\lambda_r+\beta^\textsc{int}_\infty<\delta\), and (ii) \(\beta_\infty\lambda_r+\beta^\textsc{int}_\infty>\delta\). \subsection{Curing rate above the threshold} For proving part (i) of Theorem~\ref{thm:spectral-bound}, the following claim, which follows from analyzing the time evolution of \(\mathbf{X}\) along the eigen-direction of \(G\), is useful. \begin{claim} When \(\beta(n)\) and \(\beta^\textsc{int}(n)\) are constant, i.e., \(\beta(n)=\beta\) and \(\beta^\textsc{int}(n)=\beta^\textsc{int}\) for all \(n\), and \(\beta\lambda_r+\beta^\textsc{int}<\delta\), then \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right] \le C \ln n\) for some \(C>0\) where \(\mathbf{1}^\top\mathbf{X}=n\). \label{claim:constant-infection-spectral-lower} \end{claim} \begin{IEEEproof} Please see Appendix~\ref{appendix:constant-infection-spectral-lower}. \end{IEEEproof} We are now ready to prove part (i) of Theorem~\ref{thm:spectral-bound}. \begin{IEEEproof}[Proof of part {\normalfont (i)} of Theorem~\ref{thm:spectral-bound}] Since we have \(\lim_{n\to\infty}\beta(n)=\beta_\infty\), \(\lim_{n\to\infty}\beta^\textsc{int}(n)=\beta^\textsc{int}_\infty\), and \(\beta_\infty\lambda_r+\beta^\textsc{int}_\infty<\delta\), it follows from the definition of limit \cite{ThomasF1996} that there is an \(m\) such that \begin{align} \beta(n)\lambda_r + \beta^\textsc{int}(n) < \delta,\ \text{for all}\ n \ge m. \label{eq:ngem-lower} \end{align} When the system starts in any state with a cumulative number of infections \(n\), which is greater than \(m\), it must go through a state where the cumulative number of infections is \(m\) to reach the all-zero state. However, if the system starts in a state that has less than \(m\) infections in total, then it may or may not reach a state with \(m\) infections. This gives us \begin{align*} \ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right] \le \ensuremath{\mathbb{E}}\left[T_{\mathbf{X},m}\right] + \max_{\mathbf{Y}:\mathbf{1}^{\!\!\top}\!\mathbf{Y}=m}\ensuremath{\mathbb{E}}\left[T_\mathbf{Y}\right], \end{align*} where \(T_{\mathbf{X},m}\) is the amount of time it takes to reach a state with a total of \(m\) infections starting from state \(\mathbf{X}\). Using \eqref{eq:ngem-lower}, we make the following observations. \begin{enumerate} \item[(a)] Between \(\mathbf{X}\) and any state with a total of \(m\) infections, \emph{every} state satisfies \(\beta(\cdot)\lambda_r+\beta^\textsc{int}(\cdot)<\delta\), and thus a system with a constant infectiousness equal to \(\max_{n\ge m}\left(\beta(n)\lambda_r+\beta^\textsc{int}(n)\right)\) satisfies Claim~\ref{claim:constant-infection-spectral-lower}. \item[(b)] Since our system has an infectiousness less than the system with constant infectiousness in point~(a) for \emph{every} state with \(n\ge m\), using a stochastic-dominance argument, the time it takes for epidemic extinction in the constant-infectiousness system should be greater (in expectation) than the time our system takes to go from \(\mathbf{X}\) to a state with \(m\) infections. \end{enumerate} Using these observations and Claim~\ref{claim:constant-infection-spectral-lower}, it follows that \(\ensuremath{\mathbb{E}}\left[T_{\mathbf{X},m}\right] \le C \ln n\) whenever \(n \ge m\). Since \(\max_{\mathbf{Y}:\mathbf{1}^{\!\!\top}\!\mathbf{Y}=m}\ensuremath{\mathbb{E}}\left[T_\mathbf{Y}\right]\) is a constant independent of \(n\), we have \begin{align*} \ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right] \le C' \ln n, \end{align*} when \(\mathbf{1}^\top\mathbf{X}=n\) and \(n \ge m\). For \(n<m\), \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right]\) is less than the constant \(\max_{\mathbf{Y}:\mathbf{1}^{\!\!\top}\!\mathbf{Y}=m}\ensuremath{\mathbb{E}}\left[T_\mathbf{Y}\right]\), and hence \(\ensuremath{\mathbb{E}}\left[T_\mathbf{X}\right] \le C' \ln n\) follows directly. \end{IEEEproof} \subsection{Curing rate below the threshold} We now move to the case where \(\beta_\infty\lambda_r+\beta^\textsc{int}_\infty>\delta\). For this case, we use the discrete-time Markov chain (DTMC) embedded in the continuous-time Markov chain (CTMC) \(\mathbf{X}(t)\). Let \(\mathbf{X}_0=\mathbf{X}(0)\), and let \(\mathbf{X}_k\) be the state of our system after \(k\) transitions. Then \(\mathbf{X}_0, \mathbf{X}_1, \ldots\) form a DTMC. However, the number of transitions in the CTMC must be countable for every sample path of the CTMC if the embedded DTMC is to include every transition in the CTMC. If the transitions in the CTMC are otherwise uncountably infinite, we cannot map all the transitions in the CTMC to transitions in the DTMC. If the CTMC's transitions are countable, and if the embedded DTMC is transient, there is a nonzero probability that the sequence \(\mathbf{X}_0, \mathbf{X}_1, \mathbf{X}_2, \ldots\) does not contain the all-zero state \(\mathbf{0}\), with zero infections at all nodes. This in turn implies there is a nonzero probability that our system does not reach the zero state starting from \(n\) infections initially (because the transitions are countable). This gives us an infinite mean hitting time \(\ensuremath{\mathbb{E}}[T_{\mathbf{X}_0}]\). We first state as Claim~\ref{claim:ctmc-countable} that our system has a countable number of transitions. We use this together with a theorem from \cite{MalyshevM1979} (stated as Theorem~\ref{thm:potential} here) to prove the transience of our system when \(\beta_\infty\lambda_r+\beta^\textsc{int}_\infty>\delta\). \begin{claim} Let \(\mathbb{T}\) be the set of all transition times for the continuous-time Markov chain given by \eqref{eq:node-rates}. There exists an injection from \(\mathbb{T}\) to \ensuremath{\mathbb{N}}\ with probability \(1\), i.e., the set \(\mathbb{T}\) is countable. \label{claim:ctmc-countable} \end{claim} \begin{IEEEproof} Please see Appendix~\ref{appendix:ctmc-countable} for the proof. Claim~\ref{claim:ctmc-countable} is similar to the results in \cite[Section~5.1]{ResnickAdventures}. \end{IEEEproof} To show that the embedded discrete-time Markov chain is transient, the following theorem from \cite{MalyshevM1979} (paraphrased in our notation) is useful. \begin{theorem}[from \cite{MalyshevM1979}] Let the state space of the Markov chain \(\mathbf{X}_0, \mathbf{X}_1, \ldots\) be given by \(\mathcal{S}\). If there exists a function \(V:\mathcal{S}\mapsto\ensuremath{\mathbb{R}}_+\cup\{0\}\) that satisfies the following properties: \begin{enumerate} \item[(a)] for some \(d>0\), \(\ensuremath{\mathbb{P}}\Big(\lvert V(\mathbf{X}_{k+1})-V(\mathbf{X}_k)\rvert>d\Big) = 0\ \ \text{for all}\ \ \mathbf{X}_{k}\ \text{and}\ \mathbf{X}_{k+1}\), \item[(b)] for some \(\epsilon>0\), and \(c>0\), \(\ensuremath{\mathbb{E}}\Big[V(\mathbf{X}_{k+1})-V(\mathbf{X}_k)\mid\mathbf{X}_k=\mathbf{X}\Big]>\epsilon\ \text{for all}\ \mathbf{X}\in\{\mathbf{Y} \mid V(\mathbf{Y})\ge c\}\), \end{enumerate} then the Markov chain \(\mathbf{X}_0,\mathbf{X}_1,\ldots\) is transient. \label{thm:potential} \end{theorem} Note that the conditions for transience in Theorem~\ref{thm:potential} are similar to Foster's well-known work~\cite{Foster1953}. While the conditions for positive recurrence from \cite{Foster1953} are still widely used, the conditions for transience require the potential function \(V\) to be bounded. The conditions for transience given in Theorem~\ref{thm:potential} from~\cite{MalyshevM1979} are easier to use. See \cite{SrikantY2013} for other variants. \begin{IEEEproof}[Proof of part {\normalfont (ii)} of Theorem~\ref{thm:spectral-bound}] We first prove that the DTMC embedded in our CTMC satisfies the conditions of Theorem~\ref{thm:potential}, which implies that the embedded DTMC is transient. Claim~\ref{claim:ctmc-countable} then ensures that the transience of the embedded DTMC implies transience of the CTMC. For the embedded DTMC, let us define the potential function \begin{align*} V(\mathbf{X})=\mathbf{q}^\top\mathbf{X}. \end{align*} Recall that \(\mathbf{q}\) is the Perron-Frobenius eigenvector of \(G\), which ensures that \(\mathbf{q}\succ0\) and so \(V(\mathbf{X})\) is a valid potential function. This gives us \(V(\mathbf{X}_{k+1})-V(\mathbf{X}_k)=\mathbf{q}^\top\left(\mathbf{X}_{k+1}-\mathbf{X}_k\right)\). Condition~(a) of Theorem~\ref{thm:potential} is straightforward to verify since \(\mathbf{X}_{k+1}-\mathbf{X}_k=\pm \mathbf{e}_i\) for some \(i\), where \(\mathbf{e}_i\) is the vector whose \(i\)th element is \(1\) and the rest are \(0\). So \(\ensuremath{\mathbb{P}}\big(\lvert V(\mathbf{X}_{k+1})-V(\mathbf{X}_k)\rvert>d) = 0\) for all \(d>q_{\max}\), where \(q_{\max}\) is the maximum element of \(\mathbf{q}\). We now define \(c\), and thus the set \(\{\mathbf{Y}\mid V(\mathbf{Y})\ge c\}\) in condition~(b) of Theorem~\ref{thm:potential}. We set \(c=q_{\max}m\), where \(m\) shall be determined later. This means that a sufficient condition for the transience of the embedded DTMC is that condition~(b) of Theorem~\ref{thm:potential} should hold in the set \(\mathcal{U}=\{\mathbf{Y}\mid V(\mathbf{Y})\ge q_{\max}m\}\). Note that since we have defined \(V(\mathbf{Y})=\mathbf{q}^\top\mathbf{Y}\), \(V(\mathbf{Y})\ge q_{\max}m\) implies \(\mathbf{1}^\top\mathbf{Y} \ge m\). Let \(\mathbf{X}_k=\mathbf{X}\in\mathcal{U}\), and let the sum of all rates in \eqref{eq:node-rates} when the system is in this state be \(R\). Let \(\mathbf{1}^\top\mathbf{X}=n\). Using \eqref{eq:node-rates}, we get \begin{align*} \ensuremath{\mathbb{E}}[V(\mathbf{X}_{k+1})-&V(\mathbf{X}_k)\mid\mathbf{X}_k=\mathbf{X}] \\ &= \mathbf{q}^\top\ensuremath{\mathbb{E}}\left[\mathbf{X}_{k+1}-\mathbf{X}_k\mid\mathbf{X}_k=\mathbf{X}\right] \\ &= \mathbf{q}^\top\times\frac{1}{R}\Big(\beta(n) G \mathbf{X} + \beta^\textsc{int}(n)\mathbf{X} - \delta \mathbf{X}\Big) \\ &= \frac{1}{R}(\beta(n)\lambda_{r}+\beta^\textsc{int}(n)-\delta)\mathbf{q}^\top\mathbf{X}. \end{align*} Observe that \begin{align} R &= \mathbf{1}^\top\left(\beta(n) G \mathbf{X}+\beta^\textsc{int}(n)\mathbf{X} +\delta\mathbf{X}\right) \nonumber \\ &\le (\beta(n) d_{\max}+\beta^\textsc{int}(n)+\delta)\mathbf{1}^\top\mathbf{X}, \label{eq:total-transition-rate} \end{align} where \(d_{\max}\) is the maximum degree in the graph. Since \(\mathbf{q}^\top\mathbf{X} \ge q_{\min}\mathbf{1}^\top\mathbf{X}\), where \(q_{\min}\) is the minimum element of \(\mathbf{q}\), we get \begin{align*} \ensuremath{\mathbb{E}}[V(\mathbf{X}_{k+1})-V(\mathbf{X}_k)&\mid\mathbf{X}_k=\mathbf{X}] \\ &\ge \frac{(\beta(n)\lambda_{r}+\beta^\textsc{int}(n)-\delta)q_{\min}}{ \beta(n) d_{\max}+\beta^\textsc{int}(n)+\delta} \end{align*} for all \(\mathbf{X}\in\mathcal{U}\). Since \(\beta(n)\to\beta_\infty\) and \(\beta^\textsc{int}(n)\to\beta^\textsc{int}_\infty\), the definition of limit ensures that for a sufficiently large \(m\), \(\frac{(\beta(n)\lambda_{r}+\beta^\textsc{int}(n)-\delta)q_{\min}}{\beta(n) d_{\max}+\beta^\textsc{int}(n)+\delta}\) is arbitrarily close to \(\frac{(\beta_\infty\lambda_{r}+\beta^\textsc{int}_\infty-\delta)q_{\min}}{ \beta_\infty d_{\max}+\beta^\textsc{int}_\infty+\delta}\) for all \(n \ge m\). Since \(\frac{(\beta_\infty\lambda_{r}+\beta^\textsc{int}_\infty-\delta)q_{\min}}{ \beta_\infty d_{\max}+\beta^\textsc{int}_\infty+\delta}>0\), we have \begin{align*} \ensuremath{\mathbb{E}}[V(\mathbf{X}_{k+1})-V(\mathbf{X}_k)&\mid\mathbf{X}_k=\mathbf{X}] \ge \epsilon > 0 \end{align*} for all \(\mathbf{X}\in\mathcal{U}\) for a sufficiently large \(m\). (Recall that \(\mathcal{U}=\{\mathbf{Y}\mid V(\mathbf{Y})\ge q_{\max}m\}\) which implies \(\mathbf{1}^\top\mathbf{X}\ge m\) for all \(\mathbf{X}\in\mathcal{U}\).) This proves the transience of the embedded DTMC. Transience of the embedded DTMC means that starting in state \(\mathbf{X}_0=\mathbf{X}\,(\neq\mathbf{0})\), there is a nonzero probability that the sequence of states \(\mathbf{X}_1,\mathbf{X}_2, \mathbf{X}_3, \ldots\) does not contain the all-zero state \(\mathbf{0}\) with nonzero probability (directly from the definition of transience used in \cite{MalyshevM1979} in their proof of Theorem~\ref{thm:potential}). Using Claim~\ref{claim:ctmc-countable}, this means that the CTMC defined in \eqref{eq:node-rates} has a nonzero probability of never reaching the all-zero state. Hence the average hitting time is infinite. \end{IEEEproof} \section{Vanishing Infectiousness} \label{sec:vanishing-infectiousness} In this section, we consider the special case where the per-person infectiousness functions decrease to zero as the number of active cases in the system increases: \(\beta_\infty=\beta^\textsc{int}_\infty=0\). For this, we define upper-bound and lower-bound Markov chains using the maximum and minimum node degrees. We then show that both these Markov chains have the same asymptotic mean hitting times if the per-person infectiousness functions go to zero asymptotically. Let the maximum node in-degree in \(\mathcal{G}\) be \(d_{\max}\) and the minimum node in-degree be \(d_{\min}\).% \footnote{ For weighted graphs, use the definitions \(d_{\max}=\max_u\sum_vG_{uv}\) and \(d_{\min}=\min_u\sum_vG_{uv}\). } Adding up \eqref{eq:node-rates} over all the localities \(u\in\mathcal{L}\) gives us the following \emph{upper-} and \emph{lower-bound} Markov chains for the system-wide epidemic. \begin{align} &\textit{Upper-bound Markov chain:} \nonumber\\ &X(t) \rightarrow X(t)+1\ \text{at rate}\ \Big(d_{\max}\beta\big(X(t)\big)\!\!+\!\! \beta^\textsc{int}\big(X(t)\big)\Big)X(t), \nonumber \\ &X(t) \rightarrow X(t)-1\ \text{at rate}\ \delta X(t), \label{eq:ub-rates} \end{align} and \begin{align} &\textit{lower-bound Markov chain:} \nonumber\\ &X(t) \rightarrow X(t)+1\ \text{at rate}\ \Big(d_{\min}\beta\big(X(t)\big)\!\!+\!\! \beta^\textsc{int}\big(X(t)\big)\Big)X(t), \nonumber \\ &X(t) \rightarrow X(t)-1\ \text{at rate}\ \delta X(t). \label{eq:lb-rates} \end{align} The mean hitting times of these upper- and lower-bound Markov chains are, respectively, higher and lower than the mean hitting times of the original epidemic. Proofs that they are in fact bounds are straightforward. We can see that the form of both \eqref{eq:ub-rates} for the upper-bound Markov chain and \eqref{eq:lb-rates} for the lower-bound Markov chain can be captured using a rate coefficient \(\gamma(\cdot)\) as follows. \begin{align} &X(t) \to X(t)+1\ \text{at rate}\ \gamma(X(t))X(t), \nonumber\\ &X(t) \to X(t)-1\ \text{at rate}\ \delta X(t). \label{eq:rates} \end{align} Any results we derive for a general \(\gamma(\cdot)\) apply for both the upper-bound and lower-bound Markov chains. So we now derive bounds for the hitting times of a general Markov chain satisfying \eqref{eq:rates}. Let \(T_n\) be the time it takes for the infection to go to \(0\) infections starting from \(n\) infections. Starting from \(n\) infections, the probability that the system given by \eqref{eq:rates} goes to \(n+1\) infections next (instead of \(n-1\) infections) is given by \(\frac{\gamma(n)}{\gamma(n)+\delta}\). Similarly, the probability that the system goes to \(n-1\) infections next after \(n\) infections is given by \(\frac{\delta}{\gamma(n)+\delta}\). This gives us \begin{align*} \ensuremath{\mathbb{E}}[T_n] &= \ensuremath{\mathbb{E}}[T_{n+1}]\frac{\gamma(n)}{\gamma(n)+\delta} + \ensuremath{\mathbb{E}}[T_{n-1}]\frac{\delta}{\gamma(n)+\delta} + \ensuremath{\mathbb{E}}[\tau_n], \end{align*} where \(\tau_n\) is the time it takes to make the next transition from \(n\) infections. Using \(\ensuremath{\mathbb{E}}[\tau_n]=\frac{1}{n(\gamma(n)+\delta)}\), rearranging the terms, and replacing \(n\) with \(n-1\) throughout, we get \begin{align*} \ensuremath{\mathbb{E}}[T_{n}] &= \ensuremath{\mathbb{E}}[T_{n-1}]\tfrac{\gamma(n-1)+\delta}{\gamma(n-1)} - \ensuremath{\mathbb{E}}[T_{n-2}]\tfrac{\delta}{\gamma(n-1)} - \tfrac{1}{(n-1)\gamma(n-1)} \end{align*} for \(n \ge 2\). Defining \(S_n = \ensuremath{\mathbb{E}}[T_n]-\ensuremath{\mathbb{E}}[T_{n-1}]\) yields \begin{align} S_{n+1}\gamma(n) - S_{n}\delta = -\frac{1}{n} \label{eq:sn-recursion} \end{align} for \(n \ge 1\) with \(S_1=\ensuremath{\mathbb{E}}[T_1]\). So if we can find \(\ensuremath{\mathbb{E}}[T_1]\), we will be able to compute all the mean hitting times (not necessarily in closed form). To compute \(\ensuremath{\mathbb{E}}[T_1]\), we compute the steady-state probability in state \(0\) of the transformed Markov chain in Fig.~\ref{fig:modified-markov}, whose hitting times are the same as the required Markov chain in \eqref{eq:rates}. The modification we have done to the Markov chain in \eqref{eq:rates} is the addition of the extra transition out of the zero state with a rate \(\theta\). This does not change the hitting time from any nonzero state since the time it takes to reach the zero state for the first time is independent of the rate of transition out of the zero state. However, the transformation gives us a positive-recurrent Markov chain, for which the steady-state probabilities are well-defined and non-trivial. Further, the mean hitting times are independent of the birth rate from \(0\), \(\theta\). \begin{figure*}[t!] \begin{center} \resizebox{0.8\textwidth}{!}{ \begin{tikzpicture}[>=stealth,thick] \tikzset{minimum size=3em} \node (0) [circle, draw] {0}; \node (1) [right=3em of 0, circle, draw] {1}; \node (2) [right=3em of 1, circle, draw] {2}; \node (dot) [right=3em of 2, circle] {\(\cdots\)}; \path (0) edge[bend left, ->, magenta] node [above] {\textcolor{magenta}{\(\theta\)}} (1)(1) edge[bend left, ->] node [above] {\(\gamma(1)\)} (2)(2) edge[bend left, ->] node [above] {\(2\gamma(2)\)} (dot); \path (dot) edge[bend left, ->] node [pos=0.5, below] {\(3\delta\)} (2)(2) edge[bend left, ->] node [below] {\(2\delta\)} (1)(1) edge[bend left, ->] node [below] {\(\delta\)} (0); \node (n) [right=3em of dot, circle, draw] {\(n-1\)}; \node (n1) [right=3em of n, circle, draw] {\(n\)}; \node (lastdot) [right=3em of n1, circle] {\(\cdots\)}; \path (dot) edge[bend left, ->] node [above] {} (n)(n) edge[bend left, ->] node [above] {\((n-1)\gamma(n-1)\)} (n1)(n1) edge[bend left, ->] node [above] {} (lastdot); \path (lastdot) edge[bend left, ->] node [below] {} (n1)(n1) edge[bend left, ->] node [below] {\(n\delta\)} (n)(n) edge[bend left, ->] node [below] {} (dot); \end{tikzpicture} } \caption{Modified Markov chain with same mean hitting times as the Markov chain in \eqref{eq:rates}. Adding \(\theta\) does not change the hitting times, but makes the chain positive-recurrent.} \label{fig:modified-markov} \end{center} \end{figure*} Let \(\pi_n\) be the steady-state probability of finding the chain in node \(n\). Local balance between node \(n-1\) and node \(n\) gives \begin{align*} \pi_{n-1}(n-1)\gamma(n-1) = \pi_nn\delta, \end{align*} which on expanding out yields \begin{align*} \pi_n &= \theta\pi_0\frac{\gamma(1)\gamma(2)\cdots\gamma(n-1)}{n\delta^n}, \end{align*} for \(n \ge 1\). Using \(\sum_{n=0}^\infty\pi_n=1\), we get \begin{align} \pi_0\left( 1 + \theta\left( \frac{1}{\delta} + \frac{\gamma(1)}{2\delta^2} + \frac{\gamma(1)\gamma(2)}{3\delta^3} + \cdots \right) \right) =1. \label{eq:pi0} \end{align} From renewal theory (see \cite[Chapter~7]{Ross2019}), we have \begin{align*} \pi_0 = \frac{\ensuremath{\mathbb{E}}[\tau_0]}{\ensuremath{\mathbb{E}}[\tau_0] + \ensuremath{\mathbb{E}}[T_1]}. \end{align*} Since the rate of transition out of the zero state (in the modified Markov chain) is \(\theta\), \(\ensuremath{\mathbb{E}}[\tau_0]=\frac{1}{\theta}\), and this gives \begin{align*} \ensuremath{\mathbb{E}}[T_1] = \frac{1}{\theta}\left(\frac{1}{\pi_0}-1\right). \end{align*} Substituting the expression for \(\pi_0\) from \eqref{eq:pi0} implies the following claim. \begin{claim} \label{claim:t1-gamma} The mean hitting time from one infected agent to zero infected agents is given by \begin{align*} { \ensuremath{\mathbb{E}}[T_1] = \frac{1}{\delta}\sum_{i=1}^\infty \frac{1}{i}\frac{\prod_{j=1}^{i-1}\gamma(j)}{\delta^{i-1}} } \end{align*} whenever the Markov chain in Fig.~\ref{fig:modified-markov} is positive recurrent. \end{claim} Our goal in this section has been to compute the asymptotic mean hitting times when \(\beta_\infty\) and \(\beta^\textsc{int}_\infty\) are \(0\). These conditions translate to \(\lim_{n\to\infty}\gamma(n)=0\) for both the upper-bound Markov chain \eqref{eq:ub-rates} and the lower-bound Markov chain \eqref{eq:lb-rates}. We get there by first computing the (asymptotic) mean hitting times when \(\gamma(n)=\alpha\), which we do in the next subsection. \subsection{Hitting time bounds when \(\gamma(\cdot)\) is a constant} Substituting \(\gamma(n)=\alpha\) in the expression for \(\ensuremath{\mathbb{E}}[T_1]\) in Claim~\ref{claim:t1-gamma}, we get \begin{align} { \ensuremath{\mathbb{E}}[T_1] = \frac{1}{\delta}\sum_{i=1}^\infty \frac{1}{i}\left(\frac{\alpha}{\delta}\right)^{i-1} } \label{eq:t1-alpha} \end{align} and expanding out \eqref{eq:sn-recursion} for \(\gamma(n)=\alpha\) gives us \begin{align*} S_n = S_{n-1}\frac{\delta}{\alpha} - \frac{1}{\alpha(n-1)}\qquad\quad \\ \end{align*} \vspace{-3em} { \begin{align*} \qquad\qquad&= \frac{\delta^2}{\alpha^2}S_{n-2} - \frac{\delta}{\alpha^2(n-2)} - \frac{1}{\alpha(n-1)} \\ &\ \ \vdots \\ &= \frac{\delta^{n-1}}{\alpha^{n-1}}\left( S_1 - \frac{1}{\delta}\sum_{i=1}^{n-1}\frac{1}{i}\left(\frac{\alpha}{\delta}\right)^{i-1} \right) \end{align*} } Since \(S_1=\ensuremath{\mathbb{E}}[T_1]\) by definition, substituting the expression from \eqref{eq:t1-alpha} gives us { \begin{align*} S_n &= \frac{\delta^{n-1}}{\alpha^{n-1}}\cdot\frac{1}{\delta}\sum_{i=n}^\infty\frac{1}{i}\left(\frac{\alpha}{\delta}\right)^{i-1} \\ &= \frac{1}{\delta n}\sum_{i=0}^\infty\frac{n}{n+i}\left(\frac{\alpha}{\delta}\right)^{i}, \end{align*} } and since \(\frac{n}{n+r}<1\) for all positive integers \(r\), we get \begin{align*} \frac{1}{\delta n}\ \le\ &S_n\ \le\ \frac{1}{(\delta-\alpha)n}, \end{align*} using the geometric series \(1+\frac{\alpha}{\delta}+\frac{\alpha^2}{\delta^2}+\cdots=\frac{\delta}{\delta-\alpha}\), which implies \begin{align*} \frac{1}{\delta}\sum_{i=1}^n\frac{1}{i}\ \le\ &\ensuremath{\mathbb{E}}[T_n]\ \le\ \frac{1}{\delta-\alpha}\sum_{i=1}^n\frac{1}{i}. \end{align*} This directly leads us to the following claim. \begin{claim} When the per-person infectiousness is given by \(\gamma(n)=\alpha\) for all \(n\) for some \(\alpha\in(0,\delta)\), the mean hitting time to go to zero infections starting from \(n\) infections satisfies \begin{align*} \frac{\ln (n+1)}{\delta} \ \ \le\ \ \ensuremath{\mathbb{E}}[T_n]\ \ \le\ \ \frac{1 + \ln n}{\delta-\alpha}. \end{align*} \label{claim:tn-constant} \end{claim} \subsection{When \(\lim_{n\to\infty}\gamma(n)=0\)} When the infectiousness functions \(\beta(\cdot)\) and \(\beta^\textsc{int}(\cdot)\) go to zero, i.e., \(\beta_\infty=0\) and \(\beta^\textsc{int}_\infty=0\), the \(\gamma(\cdot)\) for both the upper-bound Markov chain in \eqref{eq:ub-rates} and the lower-bound Markov chain in \eqref{eq:lb-rates} go to zero. Hence, if we can derive the asymptotic mean hitting time for \(\lim_{n\to\infty}\gamma(n)=0\), it will give us matching asymptotes for the upper and lower bounds, which means we have the exact asymptote. We will show that for any arbitrarily small \(\alpha\), we can use Claim~\ref{claim:tn-constant} to show that the asymptote for \(\ensuremath{\mathbb{E}}[T_n]\) is arbitrarily close to \(\frac{\ln n}{\delta}\). We state this formally as Theorem~\ref{thm:log-hitting}. \begin{theorem} If \(\lim_{n\to\infty}\gamma(n)=0\), then the mean hitting times of the Markov chain in Fig.~\ref{fig:modified-markov} satisfy \begin{align*} \lim_{n\to\infty}\frac{\delta\ensuremath{\mathbb{E}}[T_n]}{\ln n} = 1. \end{align*} \label{thm:log-hitting} \end{theorem} Before proving Theorem~\ref{thm:log-hitting}, let us first state a claim which will be useful. \begin{claim} If \(\lim_{n\to\infty}\gamma(n)=0\), then for any \(\epsilon > 0\), the Markov chain in Fig.~\ref{fig:modified-markov} satisfies \begin{align*} \frac{\ln (n+1)}{\delta} \le \ensuremath{\mathbb{E}}[T_n] \le \frac{\ln n}{\delta - \epsilon} + h(\epsilon)\quad \text{for all}\ n, \end{align*} for some function \(h(\epsilon)\) that is independent of \(n\). \label{claim:epsilon-bound} \end{claim} \begin{IEEEproof} Please see Appendix~\ref{appendix:epsilon-bound}. \end{IEEEproof} We are now ready to prove Theorem~\ref{thm:log-hitting}. \begin{IEEEproof}[Proof of Theorem~\ref{thm:log-hitting}] Proving \(\lim_{n\to\infty}\frac{\delta\ensuremath{\mathbb{E}}[T_n]}{\ln n}=1\) is equivalent to proving that for any \(\epsilon > 0\), we can find an \(n_\epsilon\) such that \(\left|\frac{\delta\ensuremath{\mathbb{E}}[T_n]}{\ln n}-1\right| < \epsilon\) for all \(n > n_\epsilon\) (from the definition of limit \cite{ThomasF1996}). For any \(\epsilon\), substitute \(\min\left(\frac{\epsilon\delta}{4}, \frac{\delta}{2}\right)\) for \(\epsilon\) in Claim~\ref{claim:epsilon-bound}. This gives us \begin{align*} \frac{\delta\ensuremath{\mathbb{E}}[T_n]}{\ln n} - 1 \le \frac{\epsilon}{2} + \frac{\delta\max\left(h\left(\frac{\epsilon\delta}{4}\right),h\left(\frac{\delta}{2}\right)\right)}{\ln n}. \end{align*} For sufficiently large \(n\), we get \begin{align*} \frac{\delta\ensuremath{\mathbb{E}}[T_n]}{\ln n} - 1 < \epsilon. \end{align*} Further, from the lower bound in Claim~\ref{claim:epsilon-bound}, we get \begin{align*} \frac{\delta\ensuremath{\mathbb{E}}[T_n]}{\ln n}-1 \ge \frac{\ln (n+1)}{\ln n}-1. \end{align*} For a sufficiently large \(n\), \(\frac{\ln(n+1)}{\ln n}-1\) can be made arbitrarily close to \(0\). Thus we get \begin{align*} \left|\frac{\delta\ensuremath{\mathbb{E}}[T_n]}{\ln n}-1\right| < \epsilon \end{align*} for all sufficiently large \(n\), which concludes the proof. \end{IEEEproof} \subsection{Putting it together for the original epidemic on \(\mathcal{G}\)} For both the upper-bound Markov chain in \eqref{eq:ub-rates} and the lower-bound Markov chain in \eqref{eq:lb-rates}, the infectiousness per person goes to zero if both \(\beta(\cdot)\) and \(\beta^\textsc{int}(\cdot)\) go to zero as \(n \to \infty\). Since Theorem~\ref{thm:log-hitting} applies for any chain with \(\lim_{n\to\infty}\gamma(n)=0\), both these upper- and lower-bound Markov chains satisfy Theorem~\ref{thm:log-hitting}. Since both these chains have the same asymptote, by sandwiching, even the original epidemic on \(\mathcal{G}\) must have the same asymptote. This gives us the following corollary. \begin{corollary} If \(\lim_{n\to\infty}\beta(n)=0\) and \(\lim_{n\to\infty}\beta^\textsc{int}(n)=0\), then for any locality graph \(\mathcal{G}\), we have \begin{align*} \lim_{n \to \infty}\frac{\delta\ensuremath{\mathbb{E}}[T_n]}{\ln n} = 1, \end{align*} where \(T_n\) is the time taken by the epidemic to go from a cumulative of \(n\) infections in the system to \(0\). \label{cor:graph-hitting} \end{corollary} Corollary~\ref{cor:graph-hitting} implies that if the per-person infectiousness functions go to zero asymptotically, i.e., if the (non-pharmaceutical) precautions get arbitrarily more stringent as the number of cases increases, then the mean hitting times have the asymptote \(\frac{\ln n}{\delta}\) independent of the locality graph. \section{Spectral Radius of Sum of Symmetric and Diagonal Matrices} \label{appendix:weyl-proof} In this appendix, we prove a special case of Weyl's inequality which suffices for the purposes of this paper. We state this formally in Claim~\ref{claim:weyl-special}. \begin{claim} Let \(P\) be any nonnegative symmetric matrix and \(Q\) be any nonnegative diagonal matrix. Then \begin{align*} \rho(P) + \min_i Q_{ii} \le \rho(P+Q) \le \rho(P) + \max_i Q_{ii}, \end{align*} where \(\rho(\cdot)\) denotes the spectral radius. \label{claim:weyl-special} \end{claim} \begin{IEEEproof} Recall that the spectral radius of a matrix is the maximum absolute value of the eigenvalues of the matrix. For symmetric matrices, the eigenvalues are all real, and since \(P\) and \(Q\) are nonnegative, the Perron-Frobenius theorem ensures that there is a positive eigenvalue which has the maximum absolute value. Thus we have \begin{align*} \rho(P+Q) &= \max_{\lVert \mathbf{x} \rVert=1}{\mathbf{x}^\top (P+Q) \mathbf{x}} = \max_{\lVert \mathbf{x} \rVert=1}\left(\mathbf{x}^\top P \mathbf{x} + \mathbf{x}^\top Q \mathbf{x}\right). \end{align*} Let \(\tilde{\mathbf{x}}\) be the unit vector \(\mathbf{x}\) which maximizes \(\mathbf{x}^\top P \mathbf{x}\), i.e., \(\rho(P)=\max_{\lVert \mathbf{x} \rVert=1}\mathbf{x}^\top P \mathbf{x} =\tilde{\mathbf{x}}^\top P \tilde{\mathbf{x}}\). This gives \begin{align*} \max_{\lVert \mathbf{x} \rVert=1}\left(\mathbf{x}^\top P \mathbf{x} + \mathbf{x}^\top Q \mathbf{x}\right) &\ge \tilde{\mathbf{x}}^\top P \tilde{\mathbf{x}} + \tilde{\mathbf{x}}^\top Q \tilde{\mathbf{x}} \\ &\ge \rho(P) + \min_i Q_{ii}, \end{align*} where the second inequality follows since \(\min_iQ_{ii}\) is the least value of \(\mathbf{x}^\top Q \mathbf{x}\) subject to \(\lVert \mathbf{x} \rVert=1\) since \(Q\) is a diagonal matrix. This proves the lower bound of Claim~\ref{claim:weyl-special}. For the upper bound, we have \begin{align*} \max_{\lVert \mathbf{x} \rVert=1}\left(\mathbf{x}^\top P \mathbf{x} + \mathbf{x}^\top Q \mathbf{x}\right) &\le \max_{\lVert \mathbf{x} \rVert=1} \mathbf{x}^\top P \mathbf{x} +\max_{\lVert \mathbf{x} \rVert=1} \mathbf{x}^\top Q \mathbf{x} \\ &\le \rho(P) + \max_i Q_{ii}, \end{align*} which concludes the proof. \end{IEEEproof}

https://arxiv.org/abs/1310.1570

Local Maxima of Quadratic Boolean Functions

How many strict local maxima can a real quadratic function on $\{0,1\}^n$ have? Holzman conjectured a maximum of $n \choose \lfloor n/2 \rfloor$. The aim of this paper is to prove this conjecture. Our approach is via a generalization of Sperner's theorem that may be of independent interest.

\section{Introduction} Let $\Theta$ be a real quadratic function (polynomial of total degree $\le 2$) in $n$ variables $x_1,\ldots x_n$. A \textit{strict local maximum} (or just \textit{local maximum}) of $\Theta$ on the discrete cube $Q_n=\{0,1\}^n$ is a point whose value is strictly larger than all of its neighbors. As we are only concerned with the value of $\Theta$ on $Q_n$, we may assume when convenient that all terms are of degree 2, as the constant term is irrelevant, and we can replace $x_i$ with $x_i^2$ if necessary. In this paper, we prove the following conjecture attributed to Ron Holzman (see \cite[p.~3]{problem}): \begin{thm1} Let $\Theta$ be a quadratic function on $Q_n$. Then $\Theta$ has at most $n \choose \lfloor n/2 \rfloor$ local maxima. \end{thm1} This bound is attained for example when $\Theta=-(x_1+\ldots +x_n-\lfloor n/2 \rfloor)^2$. From this, we can deduce the following: \begin{cor2} A (possibly degenerate) parallelepiped in $\mathbb{R}^n$ can have at most $n \choose \lfloor n/2 \rfloor$ vertices that are strictly closer to the origin than all of their neighbors. \end{cor2} \begin{proof} Let $\Theta$ be the form $-\sum x_i^2$; we are counting strict local maxima of $\Theta$ on the parallelepiped. As being a strict local maximum is clearly an open condition, we may perturb the sides so that the parallelepiped is not degenerate. There is an affine transformation $\tau$ taking this parallelepiped to $Q_n$, so composing $\Theta$ with $\tau^{-1}$, we are done by Theorem 1. \end{proof} [We remark in passing that Corollary 2 implies the Littlewood-Offord theorem of Kleitman \cite{kleitman}. Indeed, the theorem is equivalent to showing that the number of vertices of a parallelepiped with all side lengths at least $2$ that can land in the interior of a disc of radius $1$ is at most $n \choose n/2$. The result follows by noting that a vertex landing inside the disc must have all neighbors outside of the disc, hence farther from the disc's center.] Let $[n]=\{1,2,\ldots n\}$, and let $\Delta$ denote symmetric set difference. We define the upper $i$th level of a family of subsets $\mathcal{F}$ of $[n]$ as $$\mathcal{F}^{i+}=\{A \subseteq [n] : i \not\in A, A \cup \{i\} \in \mathcal{F}\},$$ and the lower $i$th level as $$\mathcal{F}^{i-}=\{A \subseteq [n] : i \not\in A, A \in \mathcal{F}\}.$$ Finally, for a subset $S \subseteq [n]$, we define $\mathcal{F}\Delta S$ be the family of sets $A \Delta S$ for all $A \in \mathcal{F}$. We will find that a key step is to prove the following combinatorial result: \begin{thm3} Given $S_i \subseteq [n]$ for $i=1,2,\ldots, n$, and a family of subsets $\mathcal{F}$ such that, for each $i$, no element of $(\mathcal{F}\Delta S_i)^{i+}$ contains an element of $(\mathcal{F} \Delta S_i)^{i-}$, then $|\mathcal{F}| \le {n \choose\lfloor n/2\rfloor}$. \end{thm3} Theorem 3 is perhaps interesting in its own right, because it gives a generalization of Sperner's theorem (which is the case when all $S_i=\emptyset$). At the end of the paper, we analyze the cases of when the form attains the maximum number of local maxima. It turns out our method allows us not only to deduce the structure of the quadratic function when we attain equality, but also when we are within $\frac{1}{n}{n \choose \lfloor n/2 \rfloor}$ of the optimal solution $n \choose \lfloor n/2 \rfloor$. We do not know how close this bound of $\frac{1}{n}{n \choose \lfloor n/2 \rfloor}$ is to being optimal. This paper is self-contained. See \cite{bela} for general background on set systems and Sperner's theorem. \section{Proof} We frequently view the discrete cube as the family of all subsets of $[n]$ in the obvious way (using the $x_i$ as indicator functions). Assuming all terms of $\Theta$ are of degree 2 (so that it is a quadratic form), we have its associated symmetric matrix $(q_{ij})$, where $\Theta(x_1,\ldots, x_n)=\sum_{i,j} q_{ij}x_ix_j$. There are $2^n$ graph automorphisms one gets by taking some subset $S$ of $[n]$, and replacing each $A \subseteq [n]$ by $A\Delta S$: we preserve the quadratic form by replacing $x_i$ by $1-x_i$ for all $i \in S$. When we apply such an automorphism we will say we are ``changing the origin", or ``changing the base", since this is how one can view such an action geometrically (after suitable changes of signs). \subsection{Proof that Theorem 3 implies Theorem 1} By perturbing the form slightly, we may assume all $q_{ij}$ are nonzero. We note first that changing the origin to $B=(b_1,\ldots, b_n)$ has the effect on the off-diagonal entries of $(q_{ij})$ of flipping the signs in the $i$th row and column for all $i$ with $b_i=1$ (so leaving unchanged any off-diagonal $q_{ij}$ with $b_i=b_j=1$). The behaviour of the diagonal entries is more complicated, and does not need to be considered. The difference between the value of $\Theta$ on the $x_i=1$ plane and the $x_i=0$ plane as a function of the remaining coordinates is given by the linear function $2\sum_{j \ne i} q_{ij}x_i+q_{ii}$. Changing the origin to $S_i$, where $S_i$ is the set of all $j \ne i$ with $q_{ij}$ positive, we can assume $q_{ij}<0$ for all $j \ne i$. In this new coordinate system associated to $S_i$, we cannot have $B$ in the $x_i=1$ plane containing $C$ in the $x_i=0$ plane, with both $B,C$ local maxima. This is because $B$ being a local maximum means we must have $2\sum_{j \ne i} q_{ij}b_j+q_{ii}>0$ and $C$ being a strict local maximum means $2\sum_{j \ne i} q_{ij}c_j+q_{ii}<0$, so we get $2\sum_{j \ne i} q_{ij} (b_j-c_j)>0$, which is clearly false. Also, $B$ containing $C$ in the $S_i$-coordinate system is equivalent in the original coordinate system to the statement $B \Delta S_i \supseteq C \Delta S_i$. Hence we have reduced to Theorem 3. \subsection{Proof of Theorem 3} Inspired by Kleitman's proof of the Littlewood-Offord problem \cite{kleitman} (or see \cite[Ch.4]{bela} for general background), we seek a ``symmetric quasichain decomposition" of $[n]$ (described below), where a ``quasichain" will be a family with some property $P$ which implies at most one member of $\mathcal{F}$ lies inside it. The heart of this proof is the definition of a ``quasichain", which allows this method to go through. The definition is rather surprising and seems contrived, however as we will see, once we have this definition the proof is straightforward. A ``symmetric quasichain decomposition" in this case will then be a decomposition of the $n$-cube inductively built up from the $1$-cube by taking the quasichain decomposition of the $k$-cube, duplicating each quasichain with $k+1$ added to each set, then removing precisely one set from each duplicate and adding it to the original in such a way that the new families formed remain quasichains. It is not hard to prove (or see \cite[p.~17-20]{bela}) that this process will result in exactly $n \choose \lfloor n/2 \rfloor$ quasichains, as required. For sets $B,C \subseteq [n]$, we write $B \supseteq_{S_i} C$ to mean $B \Delta S_i \supseteq C \Delta S_i$. We define a \textit{quasichain} to be a colored tournament (with colors in $[n]$) on a family of subsets $\mathcal{G}=\{G_1,\ldots G_k\}$ of $[n]$, such that \begin{enumerate}[i)] \item Whenever there is a directed edge from $G_x$ to $G_y$ of color $i$, then $i \in G_x\Delta S_i$, $i \not \in G_y\Delta S_i$, and $ G_x \supseteq_{S_i} G_y$. \item For any subset of the colors (including the empty set), if we swap the direction of edges associated to those colors, then the resulting tournament is acyclic (or equivalently transitive). \end{enumerate} It is easy to check that the acyclicity condition is equivalent to saying that no triangle has 3 distinct colors, any monochromatic triangle is acyclic, and any triangle with 2 distinct colors has the 2 edges with the same color either both leaving, or both entering, the same vertex. Note that a quasichain does \textit{not} contain all possible information about $\supseteq_{S_i}$ contaiment between its various members, but rather remembers only one such containment for every pair (just enough information to guarantee that at most one element from the pair can be a local maximum from the condition in the theorem). We write $G_x \to^i G_y$ when there is an edge from $G_x$ to $G_y$ of color $i$. Sometimes we will reduce to the case $i \not \in S_i$, in which case $G_x \to^i G_y$ implies $i \in G_x$, $i \not\in G_y$, and $G_x \supseteq_{S_i} G_y$. \begin{proof} Note that for a fixed family $\mathcal{F}$, the hypothesis of the theorem is easily checked to be invariant under base-change by an arbitrary subset $A$ (i.e. taking $F \to \mathcal{F}\Delta A$ and $S_i \to S_i \Delta A$). It is also invariant replacing $S_i$ by $S_i^c$ for any $i$, so in particular we may assume that $i \not\in S_i$ for all $i$. Also note that given a quasichain $\mathcal{G}$, it remains a quasichain after replacing $S_i$ by $S_i^c$ if one swaps the directions of all $i$-colored edges (this is the reason we need the acyclicity condition in the induction hypothesis, so that complementing the $S_i$ preserves the property of being a quasichain). If we have a quasichain $\mathcal{G}$, then base-changing to $A$ followed by complementing every $S_i$ for which $i \in S_i$ has the following effect: each $S_i$ is changed to $S_i\Delta A$ if $i \not\in A$ or $(S_i \Delta A)^c$ otherwise, $G_i$ turns into $G_i\Delta S_i$, and the direction of all $i$-colored arrows are swapped if $i \in A$. (We remark for motivation that this net effect is equivalent to what happens when we base-change by $A$ in the quadratic form case, where $S_j=\{i \ne j : q_{ij}>0\}$, as the complementation is ``built in" to the definition of $S_j$). Given $S_1,\ldots, S_n$, assume by induction we have a symmetric quasichain decomposition for $Q_{n-1}$ (using colors in $[n-1]$) associated to the sets $S_i \cap \{1,\ldots, n-1\}$ for $i=1,\ldots, n-1$. We will produce from this a symmetric quasichain decomposition for $Q_n$ (using colors in $[n]$) associated to the sets $S_i$ for $i=1,\ldots, n$. By base-changing with resect to $\{i : n \in S_i\}$, and complementing as described above, we can assume without loss of generality that $i \not \in S_i$ and $n \not\in S_i$, for $i=1,\ldots,n$. Note that if we treat a quasichain $\mathcal{C}$ in $Q_{n-1}$ as a quasichain in $Q_n$, then it remains a quasichain (obviously), and if we add $n$ to each set in $\mathcal{C}$ (denoted $\mathcal{C}+n$), then this is also a quasichain. By the standard symmetric quasichain construction discussed previously, it suffices to show that there exists an element of $\mathcal{C}+n$ which we can remove from it and add to $\mathcal{C}$ such that the two newly constructed directed graphs are quasichains. As $\mathcal{C}$ is acyclic, it has a maximal element $A$. A subgraph of a quasichain is clearly a quasichain, so $\mathcal{C}+n-\{A+n\}$ is a quasichain (here minus means we remove $A+n$ from the quasichain). Thus it suffices to show that $\mathcal{C}+\{A+n\}$ is in fact a quasichain once we appropriately color edges containing $A+n$. If $A \supseteq_{S_i} B$, then since $n \not\in S_i$, we have $A+n \supseteq_{S_i} A \supseteq_{S_i} B$, so by transitivity of $\supseteq_{S_i}$, we have $A+n \supseteq_{S_i} B$. Thus if $A \to^i B$, then $i\not\in B$, and $i \in A$, so clearly $i \in A+n$, and we may set $A+n \to^i B$. Also, as both $n \not\in S_n$ and $n \not\in A$, we may also set $A+n \to^n A$. Thus it suffices to show that the newly constructed graph satisfies the acyclicity condition. After swapping some directions associated to a subset of the colors, we have a tournament $H$, with vertices $x$, $y$ corresponding to $A$, $A+n$ respectively. For all $z$ then, we have either $x \to z$, $y\to z$, or $z\to x$, $z \to y$. If we have a cycle, then identifying $x$ and $y$ yields a cycle in the original graph (since $x$ and $y$ have the same incoming/outgoing edges, this identification is well-defined), which is a contradiction. \end{proof} \tikzset { e1/.style= { red }, e2/.style= { blue }, e3/.style= { black }, e4/.style= { purple }, e5/.style= { orange }, } \section{Analysis of the quadratic form close to equality} From now on, we are working with the $S_j$ associated to a quadratic form $\Theta$ (recall we defined $S_j=\{i \ne j : q_{ij}>0\}$). Then $i \not \in S_i$, $i \in S_j$ if and only if $j \in S_i$, and these properties are invariant under base-change. Base-change by $S_1$ so that $S_1=\emptyset$. We will show that if all $S_i$ are not the empty set (i.e. in this coordinate system there is an off-diagonal entry which is positive), then we are bounded away from the optimal solution by a factor of $\frac{1}{n}$. Indeed, suppose without loss of generality $S_2$ contains 3. Then $S_3$ contains 2, so the intersections with $\{1,2,3\}$ of $S_1,S_2,S_3$ are $\emptyset, \{3\},\{2\}$ respectively. The first three stages yield the quasichain decomposition below (in bold lines): \begin{tikzpicture}[->,>=stealth',node distance=3cm, thick,main node/.style={circle,draw,font=\sffamily\Large\bfseries}] \node[main node] (0) {$\emptyset$}; \node[main node] (1) [right of=0] {1}; \node[main node] (12) [below of=0] {12}; \node[main node] (13) [right of=12] {13}; \node[main node] (3) [right of=1] {3}; \node[main node] (123) [below of=3] {123}; \node[main node] (2) [right of=3] {2}; \node[main node] (23) [below of=2] {23}; \path[every node/.style={rectangle,draw, fill=white, font=\sffamily\small, pos=0.2}] (1) edge[e1] node {1} (0) (12) edge[e1] node {1} (0) edge[e2] node {2} (1) edge[e2] node {2} (13) (13) edge[e3] node {3} (1) edge[e1] node {1} (0); \path[every node/.style={rectangle,draw, fill=white, font=\sffamily\small, pos=0.2}] (3) edge[e3,dashed] node {3} (2) (123) edge[e1] node {1} (3) edge[e1,dashed] node {1} (2) edge[e1,dashed] node {1} (23) (23) edge[e2,dashed] node {2} (3) edge[e3] node {3} (2); \end{tikzpicture} Using the dotted lines, we see that we can ``glue" together two of these quasichains to make a single quasichain. We want to understand the ``evolution" of this quasichain as it goes through the symmetric chain algorithm. We know that a 2-quasichain after $k$ steps becomes $k+1 \choose \lfloor (k+1)/2 \rfloor$ quasichains. As it evolves into a 4-quasichain plus two 2-quasichains after two steps, if $g(k)$ is the number of quasichains the 4-quasichain evolves into after $k$ steps, it satisfies $${k+3 \choose \lfloor (k+3)/2 \rfloor}=2{k+1 \choose \lfloor (k+1)/2 \rfloor}+g(k).$$ After we reach $n$-dimensions therefore, these two 4-quasichains will have evolved to $$2g(n-3)=2{n \choose \lfloor n/2 \rfloor}-4{n-2 \choose \lfloor (n-2)/2 \rfloor}.$$ The difference between the actual bound and this is $$4{n-2 \choose \lfloor (n-2)/2 \rfloor}-{n \choose \lfloor n/2 \rfloor},$$ which is equal to $$\frac{1}{n-c}{n \choose \lfloor n/2 \rfloor}$$ where $c$ is 0 or 1 depending on whether $n$ is odd or even respectively. If every $S_i$ is in fact empty, then the condition the $S_i$ create is precisely the normal antichain condition, so the family creates an actual antichain. Thus we have the following: \begin{thm4} If the number of local maxima of a quadratic function on $Q_n$ is greater than $(1-\frac{1}{n-c}){n\choose n/2}$, where $c=0,1$ if $n$ is odd/even respectively, then there is an automorphism of $Q_n$ such that after applying it, the local maxima form an antichain. \end{thm4} \section{Acknowledgements} I would like to thank Imre Leader for helpful conversations. I would also like to thank Trinity College, Cambridge for providing me with financial support.

https://arxiv.org/abs/2204.01947

Tournaments and Even Graphs are Equinumerous

A graph is called odd if there is an orientation of its edges and an automorphism that reverses the sense of an odd number of its edges, and even otherwise. Pontus von Brömssen (né Andersson) showed that the existence of such an automorphism is independent of the orientation, and considered the question of counting pairwise non-isomorphic even graphs. Based on computational evidence, he made the rather surprising conjecture that the number of pairwise non-isomorphic even graphs on $n$ vertices is equal to the number of pairwise non-isomorphic tournaments on $n$ vertices. We prove this conjecture using a counting argument with several applications of the Cauchy-Frobenius Theorem.

\section{Introduction} In a paper on the asymptotics of random tournaments, Pontus Andersson \cite[page 252]{MR1662785} introduced the concept of an \emph{even graph} in the following fashion: Given a graph $X$, assign an arbitrary orientation to its edges. Then an automorphism $g \in \mathrm{Aut}(X)$ \emph{reverses the sense} of an edge $e = \{u,v\}$ if $e$ is oriented from $u$ to $v$, but $e^g$ is oriented from $v^g$ to $u^g$. Changing the orientation of the graph might alter the \emph{number} of edges whose sense is reversed by $g$, but Andersson showed that it does not alter the \emph{parity} of this number. He defined a graph to be \emph{odd} if it has an automorphism reversing the sense of an odd number of edges and \emph{even} otherwise. (This is somewhat overloading the adjectives ``even'' and ``odd'', which are already used for several different concepts relating to both permutations and graphs but, to avoid confusion, we shall be explicit when using any of these other meanings.) Figure~\ref{fig:4vert} shows the pairwise non-isomorphic graphs on $4$ vertices, along with an odd automorphism, i.e., an automorphism reversing the sense of an odd number of edges, if one exists. The four graphs with no odd automorphism listed are the four even graphs on 4 vertices. A \emph{tournament} is a directed graph $D$ with arc set $A(D)$ such that for any distinct vertices $\{v,w\}$, either $(v,w) \in A(D)$ or $(w,v) \in A(D)$, but not both. Equivalently, a tournament is an \emph{oriented complete graph}. Figure~\ref{fig:tourn4} illustrates the four pairwise non-isomorphic tournaments on four vertices. As we shall see, it is no coincidence that the numbers of even graphs and tournaments on $4$ vertices are the same. Pontus von Br\"omssen computed the number of pairwise non-isomorphic even graphs on up to $10$ vertices in order to enter the sequence into the On-Line Encyclopedia of Integer Sequences \cite{A000568}. He was surprised to see that his numbers coincided perfectly with the leading entries of the sequence A000568 of \cite{A000568}, which counts the number of pairwise non-isomorphic tournaments. In a comment on the sequence, he asked whether the numbers of even graphs and tournaments actually coincide for all $n$. In this note, we answer this question positively. The most obvious and satisfying way to prove that two sequences counting combinatorial structures are the same is to find a reasonably natural bijection between the sets enumerated by the two sequences. This not only proves that the sequences are equal, but also gives a convincing explanation for \emph{why} they are the same. The second-best option is to find a \emph{counting argument} showing that the two sequences are given by the same formula or expression (or satisfy the same recurrence, etc.). By using the Cauchy-Frobenius Theorem and double counting, we find expressions for the numbers of pairwise non-isomorphic graphs, tournaments and odd graphs, showing that for any fixed number of vertices,\begin{center} Number of graphs = Number of tournaments + Number of odd graphs. \end{center} As a graph is either even or odd, this immediately implies our main result: \begin{theorem} \label{thm:evenandtourn} For all $n \geqslant 2$, the number of pairwise non-isomorphic even graphs on $n$ vertices is equal to the number of pairwise non-isomorphic tournaments on $n$ vertices. \end{theorem} \begin{figure} \begin{center} \begin{tikzpicture} \tikzstyle{vertex}=[circle, draw=black, fill = blue!15, inner sep = 0.25mm] \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \node at (0.5,-0.5) {\small \phantom{$\sigma=(0,3)$}}; \pgftransformxshift{3cm} \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v3); \node at (0.5,-0.5) {\small $\sigma=(0,3)$}; \pgftransformxshift{3cm} \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v3); \draw (v1)--(v3); \pgftransformxshift{3cm} \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v3); \draw (v1)--(v3); \draw (v2)--(v3); \end{tikzpicture} \vspace{1cm} \begin{tikzpicture} \tikzstyle{vertex}=[circle, draw=black, fill = blue!15, inner sep = 0.25mm] \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v2); \draw (v1)--(v3); \node at (0.5,-0.5) {\small $\sigma=(0,2)$}; \pgftransformxshift{3cm} \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v2); \draw (v0)--(v3); \draw (v1)--(v3); \node at (0.5,-0.5) {\small $\sigma=(0,3)(1,2)$}; \pgftransformxshift{3cm} \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v2); \draw (v0)--(v3); \draw (v2)--(v3); \node at (0.5,-0.5) {\small $\sigma=(0,2)$}; \pgftransformxshift{3cm} \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v2); \draw (v0)--(v3); \draw (v1)--(v3); \draw (v2)--(v3); \node at (0.5,-0.5) {\small $\sigma=(0,2)$}; \end{tikzpicture} \vspace{1cm} \begin{tikzpicture} \tikzstyle{vertex}=[circle, draw=black, fill = blue!15, inner sep = 0.25mm] \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v2); \draw (v0)--(v3); \draw (v1)--(v2); \draw (v1)--(v3); \node at (0.5,-0.5) {\small \phantom{$\sigma=(0,3)$}}; \pgftransformxshift{3cm} \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v2); \draw (v0)--(v3); \draw (v1)--(v2); \draw (v1)--(v3); \draw (v2)--(v3); \node at (0.5,-0.5) {\small $\sigma=(2,3)$}; \pgftransformxshift{3cm} \node [vertex] (v0) at (0,0) {\footnotesize $0$}; \node [vertex] (v1) at (1,0) {\footnotesize $1$}; \node [vertex] (v2) at (1,1) {\footnotesize $2$}; \node [vertex] (v3) at (0,1) {\footnotesize $3$}; \draw (v0)--(v1); \draw (v0)--(v2); \draw (v0)--(v3); \draw (v1)--(v2); \draw (v1)--(v3); \draw (v2)--(v3); \node at (0.5,-0.5) {\small $\sigma=(0,1)$}; \end{tikzpicture} \caption{$4$-vertex graphs with an odd automorphism if one exists} \label{fig:4vert} \end{center} \end{figure} \begin{figure} \begin{center} \begin{tikzpicture}[scale=1.2] \tikzstyle{vertex}=[circle, draw=black, fill = teal, inner sep = 0.9mm] \node [vertex] (v0) at (0,0) {}; \node [vertex] (v1) at (1,0) {}; \node [vertex] (v2) at (1,1) {}; \node [vertex] (v3) at (0,1) {}; \draw [-latex] (v0)--(v1); \draw [-latex] (v0)--(v2); \draw [-latex] (v0)--(v3); \draw [-latex] (v1)--(v2); \draw [-latex] (v1)--(v3); \draw [-latex] (v2)--(v3); \end{tikzpicture} \qquad \begin{tikzpicture}[scale=1.2] \tikzstyle{vertex}=[circle, draw=black, fill = teal, inner sep = 0.9mm] \node [vertex] (v0) at (0,0) {}; \node [vertex] (v1) at (1,0) {}; \node [vertex] (v2) at (1,1) {}; \node [vertex] (v3) at (0,1) {}; \draw [-latex] (v0)--(v1); \draw [-latex] (v0)--(v2); \draw [-latex] (v0)--(v3); \draw [-latex] (v1)--(v2); \draw [-latex] (v3)--(v1); \draw [-latex] (v2)--(v3); \end{tikzpicture} \qquad \begin{tikzpicture}[scale=1.2] \tikzstyle{vertex}=[circle, draw=black, fill = teal, inner sep = 0.9mm] \node [vertex] (v0) at (0,0) {}; \node [vertex] (v1) at (1,0) {}; \node [vertex] (v2) at (1,1) {}; \node [vertex] (v3) at (0,1) {}; \draw [-latex] (v0)--(v1); \draw [-latex] (v2)--(v0); \draw [-latex] (v0)--(v3); \draw [-latex] (v1)--(v2); \draw [-latex] (v1)--(v3); \draw [-latex] (v2)--(v3); \end{tikzpicture} \qquad \begin{tikzpicture}[scale=1.2] \tikzstyle{vertex}=[circle, draw=black, fill = teal, inner sep = 0.9mm] \node [vertex] (v0) at (0,0) {}; \node [vertex] (v1) at (1,0) {}; \node [vertex] (v2) at (1,1) {}; \node [vertex] (v3) at (0,1) {}; \draw [-latex] (v0)--(v1); \draw [-latex] (v2)--(v0); \draw [-latex] (v3)--(v0); \draw [-latex] (v1)--(v2); \draw [-latex] (v1)--(v3); \draw [-latex] (v2)--(v3); \end{tikzpicture} \end{center} \caption{Tournaments on four vertices} \label{fig:tourn4} \end{figure} \section{Counting Graphs} \label{sec:graphcount} When counting graphs (or tournaments, even graphs etc.) on $n$ vertices we normally assume there is some fixed vertex set of size $n$ (often $\{1,2,\ldots,n\}$) and then distinguish between counting \emph{pairwise distinct} graphs and counting \emph{pairwise non-isomorphic} graphs. For brevity the former is normally referred to as counting \emph{labelled} graphs, and the latter as counting \emph{unlabelled} graphs. In general, counting labelled graphs is easier than counting unlabelled graphs. The Cauchy-Frobenius Theorem is a fundamental tool that can be used to express the number of unlabelled graphs as a sum, each of whose terms is the size of a specific set of labelled graphs. Before applying this theorem however, we need to establish some notation. Let $[n] = \{1,2,\ldots,n\}$ denote the vertex set of all of our graphs and tournaments on $n$ vertices. Then define \[ E_n = \{ \{u,v\} : u, v \in [n], u \ne v \}, \] so that $E_n$ is the set of \emph{unordered pairs} of distinct vertices. We think of $E_n$ as being the set of \emph{all possible edges} in a graph with vertex set $[n]$. Suppose that $X$ and $Y$ are graphs with vertex set $[n]$ and edge-sets $E(X)$ and $E(Y)$ respectively. Then $X$ and $Y$ are \emph{isomorphic} if and only if there is some permutation $g \in \mathrm{Sym}(n)$ such that $E(Y) = E(X)^g$, where \[ E(X)^g = \{ \{u^g,v^g\} \mid \{u,v\} \in E(X)\}. \] So the number of unlabelled graphs is the \emph{number of orbits} of the symmetric group $\mathrm{Sym}(n)$ acting on $\Omega = \mathcal{P}(E_n)$ (the powerset of $E_n$). In order to count orbits, we turn to the following well-known theorem (see Neumann \cite{MR562002} for the fascinating history of this result). \begin{theorem}[Cauchy-Frobenius Theorem]\label{thm:cauchyfrob} Let $G$ be a permutation group acting on a set $\Omega$. The number of orbits of $G$ on $\Omega$ is given by the expression \[ \frac{1}{|G|} \sum_{g \in G} |\mathrm{fix}(g)|, \] where $\mathrm{fix}(g)$ is the set of elements of $\Omega$ fixed by $g$. \end{theorem} In other words, the number of orbits of $G$ on $\Omega$ is equal to the \emph{average number of fixed points} of the elements of $G$. In order to apply this theorem, we need to know how many \emph{subsets of $E_n$} are fixed by a permutation $g \in \mathrm{Sym}(n)$ in its induced action on $E_n$. Equivalently, we need to know how many \emph{labelled graphs} are fixed by $g$. So for $g \in \mathrm{Sym}(n)$, let $g_E$ denote the permutation induced by $g$ on $E_n$. For example, if $n = 4$ and $g = (1,2,3,4)$, then \[ g_E = \left( \{1,2\}, \{2,3\}, \{3,4\}, \{1,4\} \right)\ \left( \{1,3\}, \{2,4\} \right). \] Now any subset of $E_4$ that is fixed by $g_E$ must either contain \emph{all} of the pairs $\{ \{1,2\}, \{2,3\}, \{3,4\}, \{1,4\}\}$ or \emph{none} of them, and similarly for $\{\{1,3\}, \{2,4\}\}$. So there are $2^2 = 4$ subsets of $E_n$ (or labelled graphs) fixed by $g_E$. In general, a subset of $E_n$ is fixed by $g_E$ if and only if it is the union of the cycles of $g_E$ (here, and later, we identify a cycle of $g_E$ with its support, i.e., the set of edges that it moves). So letting $c(g_E)$ denote the number of cycles of $g_E$, there are exactly $2^{c(g_E)}$ labelled graphs fixed by $g_E$. By the Cauchy-Frobenius Theorem, we conclude that the number of isomorphism classes of graphs on $n$ vertices is given by the expression \begin{equation}\label{exp:graphs} \frac{1}{n!} \sum_{g \in \mathrm{Sym}(n)} 2^{c(g_E)}. \end{equation} \section{Counting Tournaments} In this section, we will derive an analogous expression for the number of unlabelled tournaments on $n$ vertices, essentially by considering \emph{arcs} rather than \emph{edges}. We start by defining \[ A_n = \{ (u,v) : u, v \in [n], u \ne v\}, \] so that where $E_n$ was the set of all possible \emph{edges}, $A_n$ is the set of all possible \emph{arcs}. So, for $g \in \mathrm{Sym}(n)$, let $g_A$ denote the permutation induced by $g$ on $A_n$. For example, if $n = 4$ and $g=(1,2,3,4)$, then {\small \[ g_A = \left((1,2),(2,3),(3,4),(4,1)\right)\ \left((2,1),(3,2),(4,3),(1,4)\right)\ \left((1,3),(2,4),(3,1),(4,2)\right). \]}\noindent In this case $g_A$ has three cycles, say $C_1$, $C_2$ and $C_3$. The first two cycles $C_1$ and $C_2$ are closely related, in the sense that $C_2$ can be obtained from $C_1$ by reversing every arc, and vice versa. In contrast, reversing every arc in $C_3$ simply yields $C_3$ again. We use $c(g_A)$ to denote the number of cycles of $g_A$. We need a little bit more notation and terminology to discuss the general case. If $a$ refers to an ordered pair, say $(u,v)$, then $\overline{a}$ refers to its \emph{reverse} $(v,u)$ and $e(a)$ refers to the unordered pair $\{u,v\}$. Suppose that $C = (a_1, a_2, \ldots, a_k)$ is a cycle of $g_A$ and let $ \overline{C} = (\overline{a_1}, \overline{a_2}, \ldots, \overline{a_k}) $ be the cycle obtained from $C$ by reversing each arc. If $C = \overline{C}$, then we say that $C$ is \emph{self-paired}, otherwise \emph{non self-paired}. So in general, $g_A$ has some number (maybe zero) of self-paired cycles, the remaining non-self-paired cycles occur in pairs of the form $\{C, \overline{C}\}$. In order to use the Cauchy-Frobenius Theorem, we need to understand the relationship between the cycle structures of $g$, $g_A$ and $g_E$. If $C$ is a non self-paired cycle $(a_1, a_2, \ldots, a_k)$ of $g_A$, then its \emph{undirected image} \[ \left( e(a_1), e(a_2), \ldots, e(a_k) \right) \] is a cycle of $g_E$, and clearly $\overline{C}$ has the same undirected image. If instead $C$ is a self-paired cycle of $g_A$, then \[ C = \left(a_1, a_2, \ldots, a_k, \overline{a_1}, \overline{a_2}, \ldots, \overline{a_k} \right), \] for some $a_1,\ldots,a_k \in A_n$. As $e(a) = e(\overline{a})$, if each arc $a$ is replaced by $e(a)$ we obtain the cycle \[ \left(e(a_1), e(a_2), \ldots, e(a_k), e(a_1), e(a_2), \ldots, e(a_k) \right), \] which is a cycle of $g_E$ ``wrapped around twice''. In this case, we define the \emph{undirected image} of $C$ to be the cycle $\left(e(a_1), e(a_2), \ldots, e(a_k)\right)$. As we shall shortly see, it is the presence or absence of self-paired cycles that is crucial when counting both tournaments and odd graphs. This, in turn, depends on the \emph{order} $|g|$ of $g$. \begin{lemma}\label{le:selfpairedeven} The permutation $g_A$ has a self-paired cycle if and only if $|g|$ is even. \end{lemma} \begin{proof} If $g$ has even order, then it has a cycle of even length, say $(1,2, \ldots, 2k)$. Then the cycle \[ \left( (1,k+1), (2,k+2), \ldots, (k,2k), (k+1,1), \ldots, (2k,k) \right) \] is a self-paired cycle of $g_A$. Conversely, if $g_A$ has a self-paired cycle, then this cycle has even length, and so $|g_A|$ is even, implying that $|g|$ is even. \end{proof} \begin{corollary}\label{cor:tournaut} The automorphism group of a tournament has odd order. \end{corollary} \begin{proof} Let $T$ be a tournament, and suppose that $g \in \mathrm{Aut}(T)$. If $g$ has even order, then by \cref{le:selfpairedeven}, $g_A$ has a self-paired cycle $C$, which contains at least one pair of arcs of the form $\{a,\overline{a}\}$. The tournament is fixed by $g$ if and only if its arc set is a union of cycles of $g_A$, and so $C$ must either be a subset of $A(T)$ or disjoint from $A(T)$. However, neither is possible because exactly one of $a$, $\overline{a}$ is an arc of $T$. \end{proof} \begin{theorem}\label{t:tournaments} The number of isomorphism classes of tournaments on $n$ vertices is given by the expression \begin{equation}\label{exp:tourn} \frac{1}{n!} \sum_{\substack{g \in \mathrm{Sym}(n)\\ |g|\ \mathrm{odd}}} 2^{c(g_E)}. \end{equation} \end{theorem} \begin{proof} We apply the Cauchy-Frobenius Theorem with $G=\mathrm{Sym}(n)$ and $\Omega$ the set of (labelled) tournaments. By \cref{cor:tournaut}, permutations of even order do not contribute to the sum. If $g$ has odd order, then $g_A$ has no self-paired cycles, and a tournament is fixed by $g$ if and only if its arc set is a union of cycles of $g_A$ containing \emph{exactly one} cycle from each pair $\{C, \overline{C}\}$ pair. Therefore $g$ fixes exactly $2^{c(g_A)/2}$ tournaments. As $g_A$ has no self-paired cycles, $c(g_A)/2 = c(g_E)$ and so the number of isomorphism classes of tournaments on $n$ vertices is given by \eqref{exp:tourn}. \end{proof} An explicit expression for the number of isomorphism classes of tournaments was first given by Davis \cite{MR55294}, although expressed as a sum over conjugacy classes of permutations rather than a sum over individual permutations. \section{Counting Odd Graphs} We now derive a similar expression for counting pairwise non-isomorphic \emph{odd graphs} on $n$ vertices. Let $X$ be a graph on the vertex set $[n]$. Whether or not an automorphism $g \in \mathrm{Aut}(X)$ is an odd automorphism for $X$ is independent of the orientation of $X$, and it is convenient henceforth to assume $V(X) = [n]$ and each edge is oriented from the lower- to the higher-numbered vertex. An \emph{inversion} of a permutation $h \in \mathrm{Sym}(n)$ is a pair $\{u,v\}$ of distinct elements of $[n]$ such that $u<v$ and $u^h > v^h$. An edge of $X$ has its sense reversed by a permutation $g \in \mathrm{Aut}(X)$ if and only if it is an inversion of $g$. So the set of edges of a graph $X$ whose sense is reversed by a permutation $g \in \mathrm{Aut}(X)$ is the set of inversions of $g$ that are also edges of $X$. Now, given $g \in \mathrm{Aut}(X)$, the \emph{sign} $\mathrm{sgn}_X(g)$ of $g$ with respect to $X$ is the value \[ \mathrm{sgn}_X(g) = \prod_{\{u,v\} \in E(X)} \frac{u^g - v^g}{u - v}. \] Each term in the product is well-defined because $\frac{u^g - v^g}{u - v} = \frac{v^g - u^g}{v - u}$. Because $g \in \mathrm{Aut}(X)$, the numerator and denominator of this expression are each a product of the same multiset of values, but permuted, and possibly with changes of sign. It follows therefore that $\mathrm{sgn}_X(g) \in \{-1,1\}$. An edge $\{u,v\}$ contributes $\pm 1$ to the product, with the contribution being $-1$ if and only if its sense is reversed by $g$. Therefore $g$ is an odd automorphism for $X$ if and only if $\mathrm{sgn}_X(g) = -1$. \begin{lemma} The function $\mathrm{sgn}_X$ is a group homomorphism from $\mathrm{Aut}(X)$ to the multiplicative group $(\{-1,1\}, \times)$. \end{lemma} \begin{proof} Suppose that $g$, $h \in \mathrm{Aut}(X)$. Then \begin{align*} \mathrm{sgn}_X(gh) &= \prod_{\{u,v\} \in E(X)} \frac{u^{gh}-v^{gh}}{u - v}\\ &= \prod_{\{u,v\} \in E(X)} \frac{u^g-v^g}{u - v} \cdot \frac{(u^g)^h - (v^g)^h} {u^g-v^g}\\ &= \prod_{\{u,v\} \in E(X)} \frac{u^g-v^g}{u - v} \cdot \prod_{\{u,v\} \in E(X)}\frac{(u^g)^h - (v^g)^h} {u^g-v^g}\\ &= \mathrm{sgn}_X(g) \ \mathrm{sgn}_X(h), \end{align*} where the second product in the penultimate line is equal to $\mathrm{sgn}_X(h)$ because $g \in \mathrm{Aut}(X)$. \end{proof} \begin{corollary}\label{lem:halfodd} Exactly half of the permutations in $\mathrm{Aut}(X)$ are odd automorphisms for $X$ if and only if $X$ is an odd graph. \end{corollary} \begin{proof} The following are equivalent: (a)~exactly half of the permutations in $\mathrm{Aut}(X)$ are odd automorphisms for $X$, (b) $\mathrm{sgn}_X$ maps onto $\{-1,1\}$, (c)~there exists $g\in\mathrm{Aut}(X)$ with $\mathrm{sgn}_X(g)=-1$, and (d)~$X$ is an odd graph. \end{proof} \begin{lemma}\label{le:cycle} Suppose that $g \in \mathrm{Sym}(n)$. Then a cycle of $g_E$ contains an odd number of inversions of $g$ if and only if it is the undirected image of a self-paired cycle of $g_A$. \end{lemma} \begin{proof} Let $C$ be a cycle of $g_A$ that is not self-paired, and let $C'$ be its undirected image. Label each arc $(u,v)$ of $C$ with $0$ if $u < v$ or $1$ otherwise, and count the number of times the label \emph{changes} (from $0$ to $1$ or vice-versa) as $C$ is traversed exactly once in cyclic order starting and ending at the same arc. This number is precisely the number of inversions of $g$ that lie in $C'$. As the label at the starting and ending point of this traversal is the same, the number of changes is even. Therefore $C'$ contains an \emph{even number} of inversions of $g$. Now suppose that $C'$ is the undirected image of a self-paired cycle \[ C = (a_1, a_2, \ldots, a_k, \overline{a_1},\overline{a_2},\ldots,\overline{a_k}) \] of $g_A$. As previously, label each arc $(u,v)$ in $C$ with $0$ if $u<v$ and $1$ if $v < u$, and again count the total number of label changes as the cycle $C$ is traversed from $a_1$ to $\overline{a_1}$. (This corresponds to traversing $C'$ exactly once in cyclic order, but takes into account the fact that $g^k$ reverses the sense of $a_1$.) As the start and end of this traversal have \emph{different} labels, there are an \emph{odd number} of label changes. Therefore the undirected image of any self-paired cycle of $g_A$ contains an \emph{odd number} of inversions of $g$. \end{proof} Returning to the example of $g=(1,2,3,4)$ from Section \ref{sec:graphcount}, we see that the first cycle of $g_E$ contains two inversions of $g$, namely $\{3,4\}$ and $\{1,4\}$, while the second cycle of $g_E$ contains just one. Finally we have enough to count odd graphs. \begin{theorem}\label{t:odd} The number of isomorphism classes of odd graphs on $n$ vertices is given by the expression \begin{equation}\label{eqn:oddgraph} \frac{1}{n!} \sum_{\substack{g \in \mathrm{Sym}(n)\\ |g|\ \mathrm{even}}} 2^{c(g_E)}. \end{equation} \end{theorem} \begin{proof} We will double-count the elements of the following set \[ S = \{(X, g) : g \text{ is an odd automorphism for the $n$-vertex odd graph } X\}. \] If $X$ is an odd graph, then there are $n!/|\mathrm{Aut}(X)|$ labelled graphs isomorphic to $X$, and by \cref{lem:halfodd}, each of these has $|\mathrm{Aut}(X)|/2$ odd automorphisms, meaning that each isomorphism class of odd graphs contributes $n!/2$ pairs to $S$. So if there are $K$ pairwise non-isomorphic odd graphs, then $|S| = K n!/2$. Now, a permutation $g \in \mathrm{Sym}(n)$ is an odd automorphism for a graph $X$ if and only if $E(X)$ is a union of cycles of $g_E$ that (collectively) contain an \emph{odd number} of inversions of $g$. If $g$ has odd order, then by \cref{le:selfpairedeven} every cycle of $g_A$ is non self-paired, and hence by \cref{le:cycle}, every cycle of $g_E$ contains an even number of inversions of $g$. Therefore $S$ contains no pairs $(X,g)$ for which $g$ has odd order. If $g$ has even order, then $g_E$ has at least one cycle, say $C$, that is the undirected image of a self-paired cycle and therefore contains an odd number of inversions of $g$ (again using Lemmas \ref{le:selfpairedeven} and \ref{le:cycle}). There are $2^{c(g_E)-1}$ subsets of $E_n$ obtained by taking the union of a subset of the cycles of $g_E$ other than $C$. If $\widehat{E}$ is one of these subsets, then \emph{exactly one} of $\widehat{E}$ and $\widehat{E} \cup C$ contains an odd number of inversions of $g$. (If $\widehat{E}$ contains an even number of inversions of $g$, then $\widehat{E} \cup C$ contains an odd number of inversions of~$g$.) Therefore when $g$ has even order, it contributes $2^{c(g_E)-1}$ pairs $(X,g)$ to $S$. It follows that \[ |S| = K \frac{n!}{2} = \sum_{\substack{g \in \mathrm{Sym}(n)\\ |g|\ \mathrm{even}}} 2^{c(g_E)-1} \] and so $K$ is given by the expression~\eqref{eqn:oddgraph}. \end{proof} Every permutation of $\mathrm{Sym}(n)$ has even order or odd order, and so we obtain, from Theorems~\ref{t:tournaments} and~\ref{t:odd}, the final expression \[ \underbrace{\frac{1}{n!} \sum_{g \in \mathrm{Sym}(n)} 2^{c(g_E)}}_{\text{\# Graphs}} = \underbrace{\frac{1}{n!} \sum_{\substack{g \in \mathrm{Sym}(n)\\ |g|\ \mathrm{odd}}} 2^{c(g_E)}}_{\text{\# Tournaments}} + \underbrace{\frac{1}{n!} \sum_{\substack{g \in \mathrm{Sym}(n)\\ |g|\ \mathrm{even}}} 2^{c(g_E)}}_{\text{\# Odd Graphs}}. \] Therefore, the number of isomorphism classes of $n$-vertex tournaments is equal to the number of isomorphism classes of $n$-vertex even graphs, and we have proved Theorem \ref{thm:evenandtourn}. \section{Acknowledgements} The authors thank the Centre for the Mathematics of Symmetry and Computation at the University of Western Australia for suppporting the 2022 CMSC Research Retreat where this problem was solved. SDF was supported by a St Leonard's International Doctoral Fees Scholarship and a School of Mathematics \& Statistics PhD Funding Scholarship at the University of St Andrews. SPG was supported by the Australian Research Council Discovery Project DP190100450. \bibliographystyle{plain}

https://arxiv.org/abs/0709.4046

Equivalence Theorems in Numerical Analysis : Integration, Differentiation and Interpolation

We show that if a numerical method is posed as a sequence of operators acting on data and depending on a parameter, typically a measure of the size of discretization, then consistency, convergence and stability can be related by a Lax-Richtmyer type equivalence theorem -- a consistent method is convergent if and only if it is stable. We define consistency as convergence on a dense subspace and stability as discrete well-posedness. In some applications convergence is harder to prove than consistency or stability since convergence requires knowledge of the solution. An equivalence theorem can be useful in such settings. We give concrete instances of equivalence theorems for polynomial interpolation, numerical differentiation, numerical integration using quadrature rules and Monte Carlo integration.

\section{Introduction} For a numerical method the three most important aspects are its consistency, convergence and stability. These three were related in the well known equivalence theorem of Lax and Richtmyer for finite difference methods for certain partial differential equations \cite{LaRi1956}. We show that in a very general setting of numerical methods, in which a numerical method is posed as a family of operators acting on data, there is a Lax-Richtmyer type equivalence theorem : a consistent method is convergent if and only if it is stable. After proving the theorem in two general settings (Theorem~\ref{thm:equiv_linear} and Theorem~\ref{thm:equiv_nonlinear}), we prove it in specific instances of numerical integration using quadrature, numerical integration using Monte Carlo methods, numerical differentiation and polynomial interpolation. Consistency is a measure of how good the discretization is. Roughly, it says that the discretization is close to the smooth operator in some sense. If the discrete solution converges to the smooth solution then the numerical method is said to be convergent. Note that for discussing consistency and convergence one needs some information about the smooth problem and the smooth solution. However, numerical stability is purely a property of the discrete scheme. Roughly, stability means that the propagated error is controlled by the error in the data. Hence there is a similarity between numerical stability and well-posedness. In practice, convergence can be the hardest to prove among consistency, convergence and stability, since the actual solution is usually not known. Hence equivalence theorems can be useful in such situations. Equivalence theorems essentially say that we need not worry about the convergence while solving a problem numerically as long as its discretization is consistent with the smooth problem and the discrete scheme is stable. In addition, such theorems also show that unstable schemes will not converge for some data. These are the same advantages that are often appreciated in the setting of the classical Lax-Richtmyer equivalence theorem for finite difference schemes (see \cite{Strikwerda2004}, page 32). After the preliminaries in the next section, in Section~\ref{sec:ccs} we define consistency, convergence and stability in a general context of operators acting on data and we then prove equivalence theorems in this setting. The three sections that follow specialize these notions to specific classes of basic numerical methods. The convergence and stability theory for these example areas are well understood, but equivalence theorems have not been discussed in these areas in the literature. As given here, the equivalence theorems in these example areas serve only as concrete instances for illustration of the main ideas. We make no claims that these examples have direct practical importance in numerical analysis. However, with proper generalizations, the ideas might be of use in practical situations. For example, convergence of multidimensional interpolation can be related to its stability and consistency, as we sketch in Section~\ref{sec:interp}. Until now the advantages of equivalence theorem have been limited to finite difference methods. We suggest that similar benefits may be possible in many areas of numerical analysis. \section{ Preliminaries} \label{sec:prelim} For the convenience of the reader, we state some definitions and theorems used later on in the paper. Uniform boundedness principle is one of the fundamental building blocks of functional analysis and it is useful for proving equivalence theorems in the linear operator setting. It says that a sequence of pointwise bounded continuous linear operators defined on a complete normed linear space are uniformly bounded. \begin{theorem}[\bf Uniform Boundedness Principle] Let $\{F_{i \in I}\}$ be a set of bounded linear operators from a Banach space $V$ to a normed linear space $W$, where $I $ is an arbitrary set. Assume for every $v \in V$, the set $\{F_{i }(v)\}$ is bounded. Then $\sup_{i \in I} \left \Vert F_i \right \Vert < \infty$. \end{theorem} \begin{proof} See \cite{AbMaRa1988} or \cite{Conway1990}. The main ingredient of the proof is Baire category theorem. \end{proof} The next lemma will be used in Section~\ref{sec:interp} for proving that polynomial interpolation operators are bounded. \begin{lemma}\label{lem:closed_kernel} Let $V$ and $W$ be normed linear spaces over $\mathbb{R}$, where $W$ is finite dimensional. Let $T : V \rightarrow W$ be a surjective linear operator with a closed kernel. Then $T$ is continuous. \end{lemma} \begin{proof} Since $K = T^{-1}(0)$ is closed, $V/K$ is a normed linear space (see for instance Theorem 4.2 on page 70 in \cite{Conway1990}). For $v \in V$ the norm in $V/K$ is defined as usual to be $\Vert v + K \Vert_{V/K} := \inf\{\Vert v + k \Vert\,\text{s.t.}\, k \in K\}$, which is the distance of $v$ from $K$. Let $\widehat{T}: V/K \rightarrow W$ be the unique linear map such that $T = \widehat{T} \circ \pi$, where $\pi: V \rightarrow V/K$ is the quotient map. Recall that the quotient map is continuous. Note also that $\widehat{T}$ is a bijection. Then since $\widehat{T}$ is a linear bijection between $V/K$ and $W$, and $W$ is finite dimensional, we have that $V/K$ is finite dimensional. Thus $\widehat{T}$ is continuous (see page 56 of \cite{AbMaRa1988}). Hence $T$ is continuous since it is the composition of two continuous functions. \end{proof} \begin{remark} Note that all that one needs above is that the image of $T$ in $W$ be finite dimensional and the kernel be closed. The surjectivity of $T$ is not required in that case. \end{remark} The remaining part of this section deals with some basic facts about probability theory which are needed when we discuss Monte Carlo integration. These are not needed for the general equivalence theorems of Section~\ref{sec:ccs} or other sections with the exception of Section~\ref{subsec:mc}. Let $(\Omega, \Sigma, P)$ denote a probability space \cite{Varadhan2001} where $\Omega$ is the space of outcomes, $\Sigma$ is a $\sigma$-algebra of subsets of $\Omega$, and $P$ is a countably additive probability measure on $\Sigma$. \begin{definition} A random variable $X: \Omega \rightarrow \mathbb{R}$ is a measurable function, i.e., for every Borel set $B \subset \mathbb{R}$, $X^{-1}(B) \in \Sigma$. \end{definition} The probability measure $\alpha = PX^{-1}$ defined on $\mathbb{R}$ is called the distribution of $X$. We will assume that the random variable $X$ is continuous, i.e., there exists a nonnegative function $f(x)$, called the probability density function (pdf) of the random variable $X$, defined on $\mathbb{R}$ such that for all Borel sets $A \subset \mathbb{R}$ \[ P\left[\omega : X(\omega) \in A \right]=\alpha(A) = \int_{A}f\, . \] The mean or the expectation of the random variable $X$ if it exists is $E(X) = \int_{\mathbb{R}} x f(x) dx$. \begin{definition} Two random variables $X, Y$ are said to be independent if \[ P[\omega : X(\omega) \in A, Y(\omega) \in B] =P[\omega : X(\omega) \in A]P[\omega : Y(\omega) \in B]\, , \] for all Borel sets $A,B$. \end{definition} \begin{theorem}[\bf Strong Law of Large Numbers]\label{thm:lawof} Let $ X_1, X_2, \ldots, X_n, \ldots$ be a sequence of independent and identically distributed random variables with finite mean $\mu$. Then \[ P\left[\omega : \lim_{n \rightarrow \infty}\dfrac{1}{n}\sum _{i=1}^{n}X_i(\omega) =\mu \right]=1\, . \] \end{theorem} \begin{proof} See \cite{Chung2001}. \end{proof} \begin{theorem} \label{thm:independent} Let $X$,$Y$ be two independently and identically distributed random variables. If $f$ is a Borel measurable function \cite{Rudin1987} on $\mathbb{R}$, then the random variables $f(X)$,$f(Y)$ are independent and identically distributed. \end{theorem} \begin{proof} Let $A, B$ be any Borel sets in $\mathbb{R}$. \begin{align*} P\left[\omega : f(X(\omega)) \in A, f(Y(\omega)) \in B \right] &=P\left[\omega : X(\omega) \in f^{-1}(A), Y(\omega) \in f^{-1}(B) \right]\\ &=P\left[\omega : X(\omega) \in f^{-1}(A) \right] P\left[\omega : Y(\omega) \in f^{-1}(B) \right]\\ &=P\left[\omega : f(X(\omega)) \in A \right] P\left[\omega : f(Y(\omega)) \in B \right]\, . \end{align*} Hence $f(X), f(Y)$ are independent. Let $\alpha $ be the distribution measure of both $X$ and $Y$, i.e., \[\alpha(A) = P \left[\omega: X(\omega) \in A \right] = P \left[\omega: Y(\omega) \in A \right]\, .\]Let $\beta_{1}$ $( \beta_{2})$ be the distribution of $f(X)$ $(f(Y)$ respectively). Therefore, \[\beta_{1}(A)=P\left[\omega : f(X(\omega)) \in A \right]= P\left[\omega : X(\omega) \in f^{-1}(A) \right]=\alpha(f^{-1}(A))\, ,\] \[\beta_{2}(A)=P\left[\omega : f(Y(\omega)) \in A \right]= P\left[\omega : Y(\omega) \in f^{-1}(A) \right]=\alpha(f^{-1}(A))\, .\] Hence $\beta_1 = \beta_2 (= \alpha f^{-1})$. \end{proof} \begin{remark} We need the condition of Borel measurability on $f$ because for every Borel set $A$, we want $f^{-1}A$ to be a Borel set so that $X^{-1}(f^{-1}A) \subseteq \Sigma$. Otherwise, if $f$ is not a Borel measurable function, then $f^{-1}A$ need not be a Borel measurable set and then $X^{-1}(f^{-1}A)$ would not be in $\Sigma$. \end{remark} \section{Consistency, Convergence and Stability} \label{sec:ccs} If a smooth problem can be formulated as an operator applied to data, then the discretization can usually be formulated as a family of operators depending on some discretization parameter, typically a measure of the mesh size. For example, the parameter might be a measure of the distance between nodes in quadrature, or between interpolation points in polynomial interpolation, etc. The smooth and discrete operators can be made to act on the same space. In some cases this may be done for example, by considering continuous functions instead of discrete data. We then define the discrete scheme to be convergent if the discrete operators converge to the smooth operator pointwise on the entire space. We define consistency to be convergence on a dense subspace. If the discrete operators are bounded linear then the definition of stability is uniform boundedness of the family of discrete operators. However, when the discrete operators are general nonlinear operators, we define stability as asymptotic pointwise boundedness of the family of operators. The precise definitions are given below in Definition~\ref{defn:ccs_linear} and~\ref{defn:ccs_nonlinear}. The two definitions of stability lead to two different proofs for the equivalence theorem, both of which appear in this section. Theorem~\ref{thm:equiv_linear} is a Lax-Richtmyer type equivalence theorem applicable to general numerical analysis problems when the discrete operators involved are bounded linear. When this condition is dropped, we get Theorem~\ref{thm:equiv_nonlinear}. Specific incarnations of the linear case theorem are proved later in the context of numerical integration (Theorems~\ref{thm:quadrature}), numerical differentiation (Theorem~\ref{thm:differentiation}), and polynomial interpolation (Theorem~\ref{thm:interp}). We treat the general problem of Monte Carlo integration without the assumption of linearity. This leads to a proof of an equivalence theorem (Theorem~\ref{thm:montecarlo}) analogous to the nonlinear case but with a probabilistic flavor. In the classical Lax-Richtmyer equivalence theorem for partial differential equations, the main ingredient in the proof of stability implying convergence is essentially triangle inequality and a density argument \cite{LaRi1956}. Similarly, to prove the equivalence theorems in this paper, the main tools we use are uniform boundedness principle, triangle inequality and some density arguments. Let $V$ be a Banach space and $W$ a normed linear space. Let $h \in (0,1)$. The upper limit of $h$ is not relevant because we will be considering limits as $h \rightarrow 0$. Let $T,T_{h} : V \rightarrow W$ be set of bounded linear operators. \begin{definition}[Linear Discrete Operator Case]\label{defn:ccs_linear} If $\lim_{h \rightarrow 0} \Vert(T_h-T)v\Vert =0$ for every $v \in V$, then $T_h$ is said to {\bf converge} to $T$, and if $\lim_{h \rightarrow 0} \Vert(T_h-T)v\Vert =0$ for every $v \in V_0$, where $V_0$ is a dense subspace of $V$, then $T_h$ is said to be {\bf consistent} with $T$. If $\sup_{h} \Vert T_h \Vert < \infty$ then $T_h$ is called {\bf stable}. \end{definition} Our definition of consistency is motivated by the definition of consistency for finite difference schemes for certain PDEs. For a partial differential equation $P u = f$, where $u$ is the unknown, and a finite difference scheme $P_{k,h} v = f$, the finite difference scheme is called consistent if for any smooth function $\phi(t,x)$, $P\phi - P_{k,h}\phi \rightarrow 0$ as $k,h \rightarrow 0$. Here $k$ and $h$ are a measure of the space and time meshes respectively. The convergence is pointwise convergence at every $(t,x)$. This definition is from \cite{Strikwerda2004}. Although density is not explicitly mentioned in this definition, note that smooth functions are dense in the typical function spaces in which the solutions live. \begin{remark}\label{rem:consistency} Observe that by our definition of consistency, convergence implies consistency. However, there are other definitions of consistency under which there exist inconsistent schemes which converge. See for instance, \cite{Yamamoto2002} and Example 1.4.3 in \cite{Strikwerda2004} in the context of finite difference schemes for PDEs. \end{remark} The stability definition above is equivalent to discrete well-posedness, i.e., well-posedness of the discrete problems which is that for any $h$ and for any $v_1,v_2$ in $V$, $\Vert T_h(v_1 - v_2) \Vert \le K \Vert v_1 - v_2 \Vert$ where $K \ge 0$ is some constant. \begin{theorem}[\bf Equivalence Theorem for Linear Discrete Operators] \label{thm:equiv_linear} A consistent family of operators $T_h$ is convergent if and only if it is stable. \end{theorem} \begin{proof} Suppose the family is convergent, i.e., $\lim_{h \rightarrow 0} T_h v =Tv$ for every $v \in V$. Hence $\Vert T_h v\Vert \leq K( v)$ where $K(v)$ is a constant possibly depending on $v$. Since $V$ is complete, by uniform boundedness principle we have uniform boundedness of $T_h$, i.e., $\sup_{h } \Vert T_h \Vert\leq K $, for some $K \geq 0$, and hence stability. Conversely, suppose that $T_h$ is consistent and stable. Since $V_0$ is a dense subspace of $V$, for a given $v \in V$ choose $v_0 \in V_0$ such that \[ \Vert v- v_0 \Vert \leq \dfrac{\epsilon}{3 \max \{\left\Vert T \right \Vert, \sup_{h} \left\Vert T_h \right\Vert\}}\, . \] Because of consistency, there exists $ h_0 \in (0,1)$ such that, for all $h \leq h_0$ we have that $\Vert T_h v_0 - T v_0 \Vert \leq \epsilon /3$. Hence for all $h \leq h_0$ \begin{align*} \Vert T_h v - T v \Vert &\leq \Vert T_h v - T_h v_0\Vert + \Vert T_h v_0 - T v_0 \Vert+\Vert T v_0 - T v \Vert \\ & \leq \Vert T_h \Vert \Vert v-v_0\Vert + \Vert T_h v_0 - T v_0 \Vert+ \Vert T \Vert \Vert v_0 - v\Vert \\ & \leq \dfrac{\epsilon}{3}+\dfrac{\epsilon}{3}+\dfrac{\epsilon}{3} = \epsilon \, . \end{align*} Therefore we have convergence. \end{proof} For the case when the discrete operators are nonlinear we give a different definition of stability since uniform boundedness is unlikely. This allows us to prove an equivalence theorem in this case as well. Most of the examples we give in later sections are linear ones. The proof of the equivalence theorem in the case of nonlinear discrete operators (Theorem~\ref{thm:equiv_nonlinear}) is different from the linear case. Let $T_n$ be a sequence of operators (not necessarily linear or continuous) from $V$ to $W$, where $V$ and $W$ are normed linear. Let $T: V\rightarrow W$ be a bounded linear operator. \begin{definition}[Nonlinear Discrete Operator Case]\label{defn:ccs_nonlinear} The discrete operator $T_n$ is said to {\bf converge} to $T$ if, for any given $v \in V$, the sequence $T_n v$ converges to $Tv$ in $W$, i.e., for each $v \in V$, given $\epsilon > 0$ there exists $n_0 \in \mathbb{N}$, such that for all $n \geq n_0$, ${\left \Vert T_n v - T v \right \Vert } \leq \epsilon$. If there is a dense subspace $V_0$ of $V$ such that for any $v$ in $V_0$, the sequence $T_n v$ converges to $Tv$ in $W$ then $T_n$ is said to be {\bf consistent} with $T$. The operator $T_n$ is {\bf stable at} $v_0 \in V$ if for any $\epsilon > 0$ there exists a $\delta > 0$, such that for each $v \in V$ with ${\left \Vert v - v_0 \right \Vert } \leq \delta$, there exists $ n_0 \in \mathbb{N} $ such that for all $n \geq n_0$, ${\left \Vert T_n v - T_n v_0 \right \Vert } \leq \epsilon$. It is {\bf stable} if it is stable at every $v_0 \in V$. \end{definition} Remark~\ref{rem:consistency} about convergence implying consistency that was made earlier about the linear case is also valid for the nonlinear case covered in Definition~\ref{defn:ccs_nonlinear} since the consistency and convergence definitions are the same in both cases. \begin{theorem}[\bf Equivalence Theorem for Nonlinear Discrete Operators] \label{thm:equiv_nonlinear} A consistent family of operators $T_n$ is convergent if and only if it is stable. \end{theorem} \begin{proof} Suppose $T_n$ is convergent. This implies that given $v, v_0$ in $V$ and $\epsilon > 0$ there exists $n_v, n_{v_{0}} \in \mathbb{N}$ such that ${\left\Vert T_n v -T v \right \Vert} \leq \epsilon/3 $ for all $n \geq n_v$ and $\left\Vert T_n v_0 -T v_0 \right \Vert \leq \epsilon/3$ for all $n \geq n_{v_0}$. We need to find an $n_0 \in \mathbb{N}$ and a $\delta > 0$ such that for a given $v \in V$ with $\left\Vert v -v_0\right\Vert < \delta$, $\left\Vert T_n v -T_n v_0 \right\Vert \leq \epsilon$ for all $ n \geq n_0$. But, $\left\Vert T_n v -T_n v_0 \right\Vert = \left\Vert T_n v - T v + T v - T v_0 + T v_0- T_n v_0 \right\Vert \leq \left\Vert T_n v - T v \right\Vert + \left\Vert T v-T v_0 \right\Vert + \left\Vert T v_0- T_n v_0 \right \Vert$. Since $T$ is bounded linear operator, if we choose $\delta$ appropriately, then $\Vert T v -T v_0\Vert \leq \Vert T\Vert \Vert v - v_0 \Vert \leq \epsilon/3$. Letting $n_0 = \max (n_v, n_{v_0})$ we get $\left\Vert T_n v -T_n v_0 \right\Vert \leq \epsilon$ for all $ n \geq n_0$. Since $ v_0$ was arbitrary, $T_n$ is stable if it is convergent. Conversely, suppose we assume stability and consistency. We'll show convergence at $v_0$. Thus we need to find an $n_0 \in \mathbb{N}$ such that ${\left\Vert T_n v_0 - Tv_0 \right\Vert} \leq \epsilon$ for all $n \geq n_0$. Stability at $v_0$ means that there exists $ \delta >0$ such that for each $v_0' \in V_0$ with ${\left\Vert v_0'- v_0 \right\Vert} \leq \delta$ there exists $n_1 \in \mathbb{N}$ with ${\left\Vert T_n v_0 - T_n v_0' \right\Vert} \leq \epsilon /3$ for all $n \geq n_1$. Since $T$ is bounded linear operator, if we choose $\delta$ appropriately we can also make $\left\Vert T v_0' - T v_0 \right\Vert \leq \epsilon/3$. Choose such a $\delta$ and $v_0' \in V_0$. Note that that $\left\Vert T_n v_0 - Tv_0 \right\Vert \leq \left\Vert T_n v_0 - T_nv_0' \right\Vert + \left\Vert T_n v_0' - T v_0' \right\Vert + \left\Vert T v_0'-T v_0 \right\Vert$. By the choice of $v_0'$ the first and last terms on the right hand side of the above inequality are already at most $\epsilon/3$. By consistency, there exists $n_2 \in \mathbb{N}$ such that ${\left\Vert T_n v_0'- T v_0' \right\Vert}\leq \epsilon/3$ whenever $n \geq n_2$. Choose $n_0 = \max (n_1, n_2)$. Then for all $n \ge n_0$ we have $\left\Vert T_n v_0 - T v_0 \right\Vert \leq \epsilon$. Hence we have convergence at $v_0$. Since $v_0 $ was arbitrary, we have that stability implies convergence. \end{proof} In the linear case above (Definition~\ref{defn:ccs_linear} and Theorem~\ref{thm:equiv_linear}), we used a real parameter $h$. Typically this will be a measure of size of discretization, such as the maximum distance between adjacent nodes in the partition of an interval. In the nonlinear case (Definition~\ref{defn:ccs_nonlinear} and Theorem~\ref{thm:equiv_nonlinear}) we chose to use a natural number $n$ as the parameter. This might stand, for example, for the number of times sampling is done in Monte Carlo integration. This change from real $h$ to natural number $n$ was done to give both flavors of the definitions and proofs. Each can be written using either $h$ or $n$. \begin{remark}\label{rem:conv_stable} Note that in the proofs of both the equivalence theorems above, we did not assume consistency to show that convergence implies stability. It would have been redundant anyway, to assume consistency when we already have convergence (see Remark~\ref{rem:consistency}). \end{remark} \begin{remark} In \cite{AtHa2005} (page 67) consistency, convergence and stability are defined in a general setting of linear operators. The problem setting is the solution of equation $L v = w$, where $L$ is a bounded linear operator. This is discretized as $L_n v_n = w$. Here the unknown is $v$ and $v_n$ and $w$ is known. Under their definitions they show one side of the equivalence theorem, that a consistent method is convergent if it is stable. Our setting however is that of ``direct'' problems. In our case the object being approximated discretely is $T f$ which is approximated by $T_h f$ where $f$ is known data. In \cite{AtHa2005} an inverse of the operator $L$ is required. In our case $T$ may go from the space of continuous functions to reals (as in Section~\ref{sec:integration}) so that an inverse may not exist. \end{remark} \section{Numerical Integration} \label{sec:integration} As the first application of the ideas of Section~\ref{sec:ccs} we now discuss the notions of consistency, convergence and stability of numerical integration. We only address definite integrals of continuous functions on the real line. The two most successful methods for numerical integration are quadrature rules and Monte Carlo integration and these are covered below in Sections~\ref{subsec:quadrature} and ~\ref{subsec:mc}. The definitions and proof of theorem for quadrature are identical to the linear case in Section~\ref{sec:ccs}. For Monte Carlo integration however, the notions of consistency, convergence and stability need to be put into a probabilistic setting and the equivalence theorem proof uses some probabilistic reasoning. Otherwise, the pattern of the proof follows that of Theorem~\ref{thm:equiv_nonlinear}. In the quadrature case we apply the theorem to infer convergence of Gaussian quadrature from a simple proof of its stability and we also discuss composite trapezoidal rule and the instability of Newton-Cotes quadrature. In the case of Monte Carlo integration we discuss the Sample Mean method as an example. \subsection{Quadrature}\label{subsec:quadrature} The numerical approximation of definite integrals is often done using quadrature rules \cite{Heath2002}. Let $V = (C[a,b], {\Vert \cdot \Vert}_{\infty})$. For $f \in V$ define $I(f) = \int_{a}^{b}f(x) dx$, which can be approximated by a sequence of quadratures $I_{n}(f) = \sum_{i=0}^{n} {w_{i}^{(n)}f(x_{i}^{(n)})}$, where $a \leqslant x_{0}^{(n)} < x_{1}^{(n)} < \cdots < x_{n}^{(n)} \leqslant b$ is a partition of $[a,b]$. The points $x_i$ are called nodes. These nodes are not necessarily equally spaced or progressive. (In progressive quadrature if the number of nodes is increased from $n_1$ to $n_2$ then only $n_2-n_1$ nodes are new.) The real linear functional $I: V \rightarrow \mathbb{R}$ is a bounded linear operator and $\left\Vert I \right\Vert = b-a$. Each $I_n: V \rightarrow \mathbb{R}$ is also a linear functional on $V$ and \[ \left\vert I_n(f) \right\vert = \left\vert \sum_{i=0}^{n} {w_{i}^{(n)} f\left(x_{i}^{(n)}\right)} \right\vert \leq \sum_{i=0}^{n} \left\vert w_{i}^{(n)} \right\vert \left\vert f\left(x_{i}^{(n)}\right)\right\vert \leq {\left\Vert f \right\Vert}_{\infty} \sum_{i=0}^{n} \left\vert w_{i}^{(n)} \right\vert \, .\] Thus $\left\Vert I_n \right\Vert = \sum_{i=0}^{n} \left\vert w_{i}^{(n)} \right\vert$, and so each $I_n$ is also a bounded linear functional. With these preliminaries, we can define the stability, consistency and convergence of quadrature rules in exactly the same way as was done in Definition~\ref{defn:ccs_linear}. \begin{definition} A quadrature rule $I_n$ is said to {\bf converge} to $I$ if $I_n(f) \rightarrow I(f)$ for every $f$ in $V$. It is {\bf consistent} if it converges on a dense subspace of $V$ and {\bf stable} if $\sup_{n} \left \Vert I_n \right\Vert < \infty$. \end{definition} \begin{remark} The motivation for the above definition of stability is the following. \begin{align*} \left\vert I_n(f_1) - I_n(f_2)\right\vert &= \left \vert \sum_{i=0}^{n} {w_{i}^{(n)} f_{1}\left(x_{i}^{(n)}\right)} - \sum_{i=0}^{n} {w_{i}^{(n)} f_{2}\left(x_{i}^{(n)}\right)} \right\vert\\ &= \left\vert \sum_{i=0}^{n} w_{i}^{(n)}\left(f_{1}\left(x_{i}^{(n)} \right) - f_{2}\left(x_{i}^ {(n)}\right)\right) \right\vert\\ &\leq \sum_{i=0}^{n} \left\vert w_{i}^{(n)} \right\vert {\left\Vert f_1 -f_2 \right\Vert}_ {\infty}\, . \end{align*} Thus $\sup_{n} \sum_{i=0}^{n} \left\vert w_{i}^{(n)} \right\vert < \infty $ gives discrete well-posedness, i.e., stability. \end{remark} Now we prove an equivalence theorem for quadrature whose statement and proof is exactly the same same as the general linear equivalence theorem proved earlier (Theorem~\ref{thm:equiv_linear}). We have decided to use the natural number $n$ as a parameter here instead of the real parameter $h$ of Theorem~\ref{thm:equiv_linear}, because this is the natural setting for quadrature ($n+1$ is the number of nodes). \begin{theorem}[\bf Equivalence Theorem for Quadrature] \label{thm:quadrature} A consistent quadrature rule is convergent if and only if it is stable. \end{theorem} \begin{proof} Suppose $I_n$ converges to $I$, i.e., $I_n(f) \rightarrow I(f)$ for every $f$ in $V$. This implies that for any given $f$ in $V$, the sequence $\left\{I_n(f)\right\}$ is bounded. Since each $I_n$ is a bounded linear functional, we can apply the uniform boundedness principle which gives us that $\sup_{n} \left\Vert I_n \right\Vert < \infty$ which is the definition of stability. Conversely assume stability and consistency. By definition of consistency then $I_n(f) \rightarrow I(f)$ for all $f$ in $V_0$ where $V_0$ is a dense subspace of $V$. Stability means that $\sup_{n} \left\Vert I_n \right\Vert < \infty$. By the density of $V_0$ in $V$, given $f \in V$ choose $f_{0} \in V_0 $ such that \[{\Vert f-f_0 \Vert}_{\infty} \leq \dfrac{\epsilon}{3 \max \{\Vert I \Vert, \sup_{n} \Vert I_n \Vert \}}\, .\]Hence \begin{align*}\Vert I (f) - I_n (f) \Vert &\leq \Vert I (f) - I (f_0) \Vert + \Vert I (f_0) - I_n (f_0) \Vert + \Vert I_n (f_0) - I_n (f) \Vert \\ & \leq \Vert I \Vert {\Vert f-f_0 \Vert}_{\infty} + \Vert I (f_0) - I_n (f_0) \Vert + \Vert I_n \Vert {\Vert f_0 -f \Vert} _{\infty}\\ &\leq \Vert I \Vert {\Vert f-f_0 \Vert}_{\infty} + \Vert I (f_0) - I_n (f_0) \Vert + \sup_{n} \Vert I_n \Vert {\Vert f-f_0 \Vert}_{\infty}\\ &\leq \dfrac{\epsilon}{3}+ \Vert I (f_0) - I_n (f_0) \Vert +\dfrac{\epsilon}{3} \, . \end{align*} By consistency, there exists $n_0 \in \mathbb{N}$ such that $\Vert I (f_0) - I_n (f_0) \Vert \leq \epsilon/3$ for all $n \geq n_0$. Hence $\Vert I (f) - I_n (f) \Vert \leq \epsilon$ for all $n \geq n_0$. Therefore $I_n(f) \rightarrow I(f)$ for every $f$ in~$V$. \end{proof} Now we use the equivalence theorem to show convergence of Gaussian quadrature rules and of the composite trapezoidal rule, followed by the non convergence and instability of Newton-Cotes rules. \begin{example}[Gaussian Quadrature] In Gaussian quadrature all the weights $w_{i}^{(n)} \geq 0$. Hence $\left\Vert I_n \right\Vert = \sum_{i=0}^{n} w_{i}^{(n)}$. Gaussian quadrature rule $I_n$ is exact for all polynomials of degree less or equal to $2n-1$, i.e., $I_n(p) = \int_{a}^{b} p(x) dx$ for all polynomials $p$ of degree less than or equal to $2n-1$. Since the space of such polynomials is dense in $V$, Gaussian quadrature is consistent. Moreover, $I_n(1) = \sum_{i=0}^{n} w_{i}^{(n)} = \int_{a}^{b}1 dx = b-a$. Thus $\sup_{n} \left\Vert I_n \right\Vert = \sup_{n} \sum_{i=0}^{n} w_{i}^{(n)} = b-a$. Therefore Gaussian quadrature is stable. Then by the above theorem, it is convergent. \end{example} \begin{example}[Composite Trapezoidal Rule] The composite trapezoidal rule has non negative weights. Moreover, it is exact for all piecewise linear polynomials which is a dense subspace of $V$. Hence by the argument as in the above example, we have stability and consistency and therefore convergence of the composite trapezoidal rule. \end{example} \begin{example}[Newton-Cotes] Define $I_n$ to be the Newton-Cotes quadrature rule. The nodes are are equally spaced in this quadrature rule and it integrates polynomials of a certain degree exactly. Thus $I_n$ is a consistent family by our definition. However it is not convergent. An example continuous function for which $I_n (f)$ does not converge to $I(f)$ is the Runge's function $f(x) = (1/1+25x^2)$ in the interval $[-1,1]$ (see \cite{SuMa2003}, page 208). This function also appears in the polynomial interpolation section in Example~\ref{eg:runge_interp}. By Theorem~\ref{thm:quadrature} Newton-Cotes should also be unstable. Indeed it is known that in Newton-Cotes rules some of the weights have negative sign and that this leads to instability, i.e., it is known that $\sup_{n} \left\Vert I_n \right \Vert = \sup_{n} \sum_{i=0}^{n} \vert w_{i}^{(n)} \vert = \infty$ (see page 350 of \cite{Heath2002}). \end{example} \subsection{Monte Carlo Integration}\label{subsec:mc} For well behaved functions, i.e., functions with continuous derivatives, the deterministic quadrature rule is very efficient at least in one dimension. However, if the function fails to be well behaved or in the case of multidimensional integrals, other techniques can be competitive. In this section we will define convergence, consistency and stability for Monte Carlo integration. For simplicity, this is done for functions in $C[a,b]$. The notation and results from probability theory that are needed were reviewed in Section~\ref{sec:prelim}. Let $V = (C[a,b], {\Vert \cdot \Vert}_{\infty})$ and for an $f \in V$, let $I: V\rightarrow \mathbb{R}$ be defined as $I(f)= \int_{a}^{b}f$. Let $ (W, {\Vert \cdot \Vert}_{{\infty}'})$ denote the space all of bounded random variables defined on a probability space $(\Omega, \Sigma, P)$, where ${\Vert X \Vert}_{{\infty}'} = \operatorname{ess. sup_{\omega \in \Omega}} \vert X(\omega)\vert$. Let $M_n : V \rightarrow W$ be a sequence of maps, not necessarily bounded linear. For a specific Monte Carlo integration method see Example~\ref{eg:sample_mean} below. In that example the discrete operators $M_n$ are bounded linear. We have chosen to state and prove the equivalence theorem for Monte Carlo integration (Theorem~\ref{thm:montecarlo}) without this assumption. This is done to illustrate how the nonlinear case proof of an equivalence theorem works in a probabilistic setting such as this. We can look upon $I$ as a map from $V$ to $W$ by defining $I(f)$ to be the constant random variable, i.e., $I(f): \Omega \rightarrow \mathbb{R}$ is defined as $I(f)(\omega) = I(f)$ for all $\omega \in \Omega$. \begin{definition} A Monte Carlo integration is said to be {\bf convergent} if for any given $f \in V$, \[ P \left[ \omega \in \Omega : \lim_{n \rightarrow \infty} M_n(f)(\omega) = I(f)(\omega) \right] =1 \, . \] It is {\bf consistent} if there is a dense subspace $V_0 $ of $V$ such that for any $f \in V_0$, it is convergent. It is said to be {\bf stable at} $f_0 \in V$ if for any $\epsilon >0$ there exists a $\delta > 0$ such that for each $f \in V$ with ${\Vert f - f_0 \Vert}_{\infty} \leq \delta$, there exists $n_0 \in \mathbb{N}$ such that for all $n \geq n_0$, \[ P \left[ \omega \in \Omega : \left \vert M_n(f)(\omega) - M_n(f_0)(\omega) \right \vert \leq \epsilon \right] =1\, . \] It {\bf stable} if it is stable at every $f_0 \in V$. \end{definition} Now we state and prove an equivalence theorem for which the proof is similar to Theorem~\ref{thm:equiv_nonlinear}, but with a probabilistic flavor. \begin{theorem}[\bf Equivalence Theorem for Monte Carlo Integration] \label{thm:montecarlo} A consistent Monte Carlo integration is convergent if and only if it is stable. \end{theorem} \begin{proof} Suppose it is convergent. To show stability, we need to find an $n_0 \in \mathbb{N}$ and $\delta > 0$ such that for a given $f \in V$ with ${\Vert f - f_0 \Vert}_{\infty} \leq \delta$, and for all $n \geq n_0$, we have $ P \left[ \omega : \left \vert M_n(f)(\omega) - M_n(f_0)(\omega) \right \vert \leq \epsilon \right] =1$. Outside a set of $P$-measure 0 in $\Omega$ and for all $n \in \mathbb{N}$, we have the following inequality \begin{align*} \left\vert M_n(f)(\omega)-M_n(f_0)(\omega) \right\vert &\leq \left\vert M_n(f)(\omega)- I(f)(\omega) \right\vert \\ &+\left\vert I(f)(\omega) - I(f_0)(\omega) \right\vert +\left \vert I(f_0)(\omega)-M_n(f_0)(\omega) \right \vert \, . \end{align*} By convergence, there are $n_1$ and $n_2$ such that for all $n \geq n_1$, \[ P \left [ \omega : \left \vert M_n(f)(\omega) - I(f)(\omega) \right \vert \leq \epsilon/3 \right ] = 1 \, , \] and for all $n \geq n_2$, \[ P \left [ \omega : \left \vert M_n(f_0)(\omega) - I(f_0)(\omega) \right \vert \leq \epsilon/3 \right ] = 1 \, . \] The integral operator is bounded, because if ${\Vert f-f_0\Vert}_{\infty} \leq \delta$, then $\vert I(f)-I(f_0) \vert \leq \delta(b-a)$. Since $I(f)(\omega) = I(f)$ for all $\omega \in \Omega$ and by the boundedness of the integral operator, for an appropriately chosen $\delta > 0$, we have $\vert I(f)-I(f_0) \vert \leq \epsilon/3$. Hence \[ P \left [ \omega : \left \vert I(f)(\omega) - I(f_0)(\omega) \right \vert = \left \vert I(f) - I(f_0) \right \vert \leq \epsilon/3 \right ] =1\, . \] Define the three sets \begin{align*} \Omega_1 &= \{ \omega \in \Omega : P \left [ \omega : \left \vert M_n(f)(\omega) - I(f)(\omega) \right \vert \leq \epsilon/3 \right ] =1 \}\\ \Omega_2 &= \{ \omega \in \Omega : P \left [ \omega : \left \vert I(f)(\omega) - I(f_0)(\omega) \right \vert \leq \epsilon/3 \right ] =1\}\\ \Omega_3 &= \{ \omega \in \Omega : P \left [ \omega : \left \vert I(f_0)(\omega) - M_n(f_0)(\omega) \right \vert \leq \epsilon/3 \right ] =1 \}\, . \end{align*} It is easy to see that $\Omega_2=\Omega$. Let $\Omega' = \bigcap_{i=1}^3 \Omega_i$. Let $n_0 = \max(n_1, n_2)$. For all $n \geq n_0$, $P(\Omega_i) =1$ for $1 \leq i \leq 3$. Since $P$ is countably additive, measure of a countable union of sets of measure zero is zero. Therefore $P(\Omega')=1$. Thus for all $\omega \in \Omega'$, except for set of $P$-measure zero we have $\left \vert M_n(f)(\omega)-M_n(f_0)(\omega) \right\vert \leq \epsilon$ for all $n \geq n_0$. Hence $P \left[ \omega : \left \vert M_n(f)(\omega) - M_n(f_0)(\omega) \right \vert \leq \epsilon \right] =1$ for all $n \geq n_0$. Since $f_0$ was arbitrary the Monte Carlo integration is stable if it is convergent. Conversely, suppose we assume stability and consistency. Therefore, given $f_0$ and $\epsilon > 0$, there exists $\delta > 0$, such that for each $f \in V$ with ${\Vert f - f_0 \Vert}_{\infty} \leq \delta$, there exists an $n_0 \in \mathbb{N}$ such that for all $n \geq n_0$, \[ P \left[ \omega \in \Omega : \left \vert M_n(f)(\omega) - M_n(f_0)(\omega) \right \vert \leq \epsilon \right] =1\, . \] By the density of $V_0$ in $V$, we can choose $g \in V_0$ such that ${\Vert g - f_0 \Vert}_{\infty} \leq \delta$. By the boundedness of $I$, for an appropriately chosen $\delta$, we have ${\left \vert I(g)-I(f_0) \right \vert} \leq \epsilon/3$. Again, outside a set of $P$-measure 0 in $\Omega$ and for all $n \in \mathbb{N}$, we have the following inequality \begin{align*} \left\vert M_n(f_0)(\omega)-I(f_0)(\omega) \right\vert &\leq \left\vert M_n(f_0)(\omega)- M_n(g)(\omega) \right\vert \\ &+\left\vert M_n(g)(\omega) - I(g)(\omega) \right\vert +\left \vert I(g)(\omega)-I(f_0)(\omega) \right \vert \, . \end{align*} Stability at $f_0 $ means that there exists $\delta > 0 $ such that for each $g \in V_0$ with ${\Vert g - f_0 \Vert}_{\infty} \leq \delta$ there exists $n_1 \in \mathbb{N}$ with \[ P \left [ \omega : \left \vert M_n(f_0)(\omega) - M_n(g)(\omega) \right \vert \leq \epsilon/3 \right ] =1 \}\, , \] for all $n \geq n_1$. By consistency, there is an $n_2$ such that for all $n \geq n_2$, \[ P \left [ \omega : \left \vert M_n(g)(\omega) - I(g)(\omega) \right \vert \leq \epsilon/3 \right ] =1 \}\, . \] Moreover, \[ P \left [ \omega : \left \vert I(g)(\omega) - I(f_0)(\omega) \right \vert \leq \epsilon/3 \right ] =1\}\, . \] By an identical argument as in the previous half of the proof, on a subset $\Omega'$ of $\Omega$ with $P$-measure one, we have \[ P \left [ \omega \in \Omega' : \left \vert M_n(f_0)(\omega) - I(f_0)(\omega) \right \vert \leq \epsilon \right ] =1 \}\, , \] for all $n \geq n_0 = \max(n_1, n_2)$. Hence we have convergence at $f_0$. Since $f_0$ was arbitrary we have that stability implies convergence. \end{proof} As an example of numerical integration using Monte Carlo methods we will examine the Sample-Mean Monte Carlo method and discuss its convergence and stability. We prove both convergence and stability separately. An alternative would have been to prove consistency and one of the other two properties and infer the third from Theorem~\ref{thm:montecarlo}. \begin{example}[Sample-Mean Monte Carlo Method] \label{eg:sample_mean} Suppose we want to compute the approximate value of the integral $I= \int_{a}^{b} f(x) dx$. First, we choose any function $g \in C[a,b]$ with the property that $g > 0$ and $\int_{a}^{b}g =1$. Since $g \in C[a,b]$ and is positive, there exists $m \in \mathbb{R} $ such that $ 0 < m \leq g(x) $ in $[a,b]$. Then there exists some random variable $X$ with range in $[a,b]$, such that $g(x)$ is the pdf of $X$\cite{Tucker1967}. Then, consider the random variable $Y= f(X)/g(X)$. Therefore, \[ E(Y)= \int_{a}^{b}\dfrac{f(x)}{g(x)}g(x)dx = \int_{a}^{b}f =I\, . \] Now, choose a sequence of independent and identically distributed random variables $X=X_1$,$ X_2$, $\ldots$,$X_n$, $\ldots$ Since $f,g$ are continuous, they are Borel measurable. Since $g \neq 0$, $f/g$ is Borel measurable. Hence the sequence of random variables $Y = f(X)/g(X)= f(X_1)/g(X_1)= Y_1, Y_2 = f(X_2)/g(X_2), \ldots, Y_n=f(X_n)/g(X_n), \ldots$ are independently and identically distributed by Theorem \ref{thm:independent}. Hence for the chosen pdf $g$, we can define $M_{n}: V\rightarrow W$ as \[M_{n}(f) = \dfrac{1}{n}\sum _{i=1}^{n}\dfrac{f(X_i)}{g(X_i)}\, .\]By Theorem \ref{thm:lawof}, \[ P\left[\omega : \lim_{n \rightarrow \infty}\dfrac{1}{n}\sum _{i=1}^{n}Y_i(\omega) = I \right]=1\, . \] Thus we have the convergence of $M_{n}$. Since $M_{n}$ is linear in $f$, we can check for the boundedness of this linear operator \[ \left \Vert M_{n} \right \Vert = \sup_{{\Vert f \Vert}_{\infty}=1} {\left \Vert \dfrac{1}{n}\sum _{i=1}^{n}\dfrac{f(X_i)}{g(X_i)} \right \Vert}_{{\infty}^{'}} \leq \dfrac{1}{n} \sum _{i=1}^{n} {\left \Vert \dfrac{f}{g} \right \Vert}_{\infty} \leq \dfrac{1}{m} \, . \] To get the estimate on the norm of $M_n$ we have used the fact that ${\Vert h(X)\Vert}_{\infty'} \leq {\Vert h \Vert}_{\infty}$, where $h = f/g$. This is true because \[ {\Vert h(X) \Vert}_{\infty'}= \operatorname{ess. sup}_{\omega \in \Omega} \vert h(X(\omega)) \vert = \inf \{M : P[\omega \in \Omega : \vert h(X(\omega))\vert \geq M ] =0 \} \, , \] and hence the bound on $h(X)$ is controlled by the bound on $h$. The bound on $M_{n}$ is independent of $n$ and depends only on $g$. To exhibit stability we have to show that given $\epsilon > 0$ there exists $\delta > 0, n_0 \in \mathbb{N}$ such that for all $n \geq n_0$ \[ P \left[ \omega : \left \vert M_{n}(f)(\omega) - M_{n}(f_0)(\omega) \right \vert \leq \epsilon \right] =1\, , \] whenever ${\Vert f - f_0 \Vert}_{\infty} \leq \delta$. But outside a set of $P$-measure zero and for all $n \in \mathbb{N}$, we have \[ \left \vert M_{n}(f)(\omega) - M_{n}(f_0)(\omega) \right \vert \leq {\left\Vert M_{n}(f) - M_{n}(f_0) \right\Vert}_{\infty'} \leq \left\Vert M_{n} \right\Vert {\Vert f- f_0 \Vert}_{\infty}\, . \] Therefore, for an appropriately chosen $\delta > 0 $ and for all $n \in \mathbb{N}$, we have \[ P \left[ \omega : \left \vert M_{n}(f)(\omega) - M_{n}(f_0)(\omega) \right \vert \leq \epsilon \right] =1\, , \] hence stability. The random variables $X_n$ are not necessarily unique for a given $g$. However, it does not matter as the Sample Mean Monte Carlo method is stable and convergent if we choose a positive pdf $g \in C[a,b]$. \end{example} \section{Numerical Differentiation} \label{sec:differentiation} If sufficiently differentiable functions are considered and a sum of sup norms on the function and its derivatives is used as the norm then smooth differentiation is bounded linear. Numerical differentiation can be posed as a parameterized collection of linear operators. If they are assumed to be bounded as well then the definitions and equivalence theorem from Section~\ref{sec:ccs} apply here. We show in this section, and it is no surprise, that the equivalence theorem is actually not needed for proving convergence or stability for the usual finite difference formulas for the first derivative such as forward, backward and central difference formulas. We also give the example of the lowest order, 3 points finite difference formula for second derivative, for which equivalence theorem is also not required. For these and similar simple formulas, the proofs of stability and convergence can be done directly and independently of each other and are easy. Thus there is no practical benefit of an equivalence theorem in the context of such simple finite difference formulas for numerical differentiation in one dimension. However the equivalence theorem might be of benefit in proving stability or convergence for formulas of high order accuracy for arbitrary derivatives on non equally spaced grids such as the finite difference formulas in \cite{Fornberg1988,Fornberg1996}. The mesh size parameter $h$ appears as $h^k$ in the denominator of these and similar formulas and stability proofs might be tedious. Here $k$ is the order of the derivative. We do not discuss these formulas further in this paper. For $k \geq 1$, define ${\Vert f \Vert}_{C^k} := \sum_{i=0}^{k} {\Vert f^{(i)} \Vert}_{\infty}$ where $f^{(i)}$ denotes the $i$-th derivative of $f$. Let $V^k = (C^k [a,b], \Vert \cdot \Vert_{C^k})$ and let $W = (C[a,b], \Vert \cdot \Vert_{\infty})$. Consider $D_{h}^{(k)}, D^{(k)} : V^k \rightarrow W$, where $D_{h}^{(k)}$ and $D^{(k)}$ are the discrete and smooth differentiation operators respectively. Here $h \in (0,1)$, is a parameter of the discrete operators and is a measure of how far apart the points used in, say, a finite difference formula are. In the chosen norms, $D^{(k)}$ is a bounded linear operator. In addition, assume also that $D_{h}^{(k)}$ are bounded linear operators. The definitions of consistency, convergence and stability in Definition~\ref{defn:ccs_linear} can now be repeated in the context of numerical differentiation. \begin{definition} Numerical differentiation is said to be {\bf convergent} if \[ \lim_{h\rightarrow 0}{\left\Vert \left(D_h^{(k)}-D^{(k)}\right)f\right \Vert}_{\infty} =0\, , \] for every $f \in V^k$ and it is {\bf consistent} if it converges on a dense subspace of $V^k$. It is said to be {\bf stable} if $\Vert D_h^{(k)} \Vert \leq C$ where $C$ is independent of the parameter~$h$. \end{definition} \begin{theorem}[\bf Equivalence Theorem for Numerical Derivatives] \label{thm:differentiation} A consistent finite difference scheme is convergent if and only if stable. \end{theorem} \begin{proof} The proof is identical to the proof of Theorem~\ref{thm:equiv_linear}, with $T_h$ replaced by $D_h^{(k)}$ and $T$ by $D^{(k)}$. \end{proof} \begin{example}[Forward Difference] As an example we now consider the basic forward difference approximation to the first derivative which we will show to be stable and convergent. The equivalence theorem is not required in this case although using it would reduce the work required in proving stability and convergence. It is easy enough to prove convergence and stability separately in an elementary way. Consider a point $x \in [a,b-h]$. Let $h \in (0,1)$ such that $x+h < b$. We will show that the forward difference approximation to the first derivative \[ D_h^{(1)}f(x) := \dfrac{f(x+h)-f(x)}{h}\, , \] is both convergent and stable. First note that \begin{align*} \left\Vert D_h^{(1)}\right\Vert=\sup_{{\left\Vert f \right\Vert}_{C^1}=1} {\left\Vert D_h^{(1)}f \right\Vert}_{\infty} &=\sup_{{\left\Vert f \right\Vert}_{C^1}=1} \sup_{x \in [a,b-h]}\left\vert \dfrac{f(x+h)-f(x)}{h} \right\vert \\ & = \sup_{{\left\Vert f \right\Vert}_{C^1}=1} \sup_{x \in [a,b-h]} \left\vert f'(x + \theta h) \right\vert \leq 1 \, , \end{align*} for some $0 < \theta <1$. Thus $D_h^{(1)}$ is stable. Moreover, by Mean Value Theorem we have \begin{align*} \lim_{h\rightarrow 0}{\left\Vert D_h^{(1)}f-D^{(1)}f\right\Vert}_{\infty} &= \lim_{h\rightarrow 0} \sup_{x \in [a,b-h]} \left\vert \dfrac{f(x+h)-f(x)}{h}-f'(x) \right\vert\\ &=\lim_{h\rightarrow 0} \sup_{x \in [a,b-h]} \left\vert f'(x + \theta h)-f'(x) \right\vert =0 \, , \end{align*} which means that $D_h^{(1)}$ converges to $D^{(1)}$. The convergence and stability of the other commonly used finite difference schemes like backward difference and central difference schemes can be similarly proved. \end{example} The stability proved above means that a slight perturbation of $f \in V^1$ does not drastically change the computed numerical derivative. Note however that if the $C^{1}$ norm is replaced by the sup norm, i.e., if the space $V^1$ is replaced by $(C^1[a,b], {\Vert \cdot \Vert}_{\infty})$, then the finite difference schemes above are highly sensitive to small perturbations, leading to instability. This can be seen in the following example. \begin{example}[Unboundedness in sup Norm]\label{eg:unbounded_deriv} Define a sequence $g_h(x)= \sin (2\pi x/h)$ in $\left(C^{1}[0,1], \Vert \cdot \Vert_{\infty}\right)$. Note that $\left\Vert g_h \right\Vert_{\infty}=1$. However, \[ {\left\Vert D_h^{(1)}(g_h) \right\Vert}_{\infty} = \sup_{x \in [0,1-h]}\left\vert\dfrac{\sin (2\pi(x+h)/h)- \sin (2\pi x/h)} {h}\right\vert\, . \] By Mean Value Theorem, \[ \dfrac{\sin (2\pi (x+h)/h)- \sin (2\pi x/h)}{h} = \dfrac{2\pi}{h}\cos (2\pi c/h) \, , \] for some $c \in (0,1-h)$ and so \[ {\left\Vert D_h^{(1)}(g_h) \right\Vert}_{\infty} = \sup_{c \in (0,1-h)}\left\vert \dfrac{2\pi}{h}\cos (2\pi c / h) \right\vert =\dfrac{2\pi}{h} \, . \] This implies that ${\Vert D_h^{(1)} \Vert}_{\infty} > 1/h$ for all $ h \in (0,1)$. \end{example} \begin{example}[Second Derivative] We now discuss the convergence and stability of the most basic finite difference operator for second derivative. We will show that the finite difference approximation for second derivative \[ D_h^{(2)}f(x) := \dfrac{f(x+h)-2f(x)+f(x-h)}{h^2}\, , \] is both convergent and stable. The operators $D_h^{(2)}$ are consistent and so it is actually enough to prove just stability or convergence due to Theorem~\ref{thm:differentiation}. But as in the first derivative case, we will prove stability and convergence separately, since both are easy to prove. To prove stability, note that \[ \left\Vert D_h^{(2)} \right\Vert =\sup_{{\left\Vert f \right\Vert}_{C^2}=1} {\left\Vert D_h^{(2)}f \right\Vert}_{\infty} =\sup_{{\left\Vert f \right\Vert}_{C^2}=1}\sup_{x \in [a+h,b-h]} \left\vert \dfrac{f(x+h)-2f(x)+f(x-h)}{h^2} \right\vert \, . \] By two applications of the Mean Value Theorem, we have that for some $0 < \alpha, \beta <1$, the above is equal to \[ \sup_{{\left\Vert f \right\Vert}_{C^2}=1}\sup_{x \in [a+h,b-h]} \left\vert \dfrac{f'(x+\alpha h)-f'(x- \beta h)}{h} \right\vert \leq \sup_{{\left\Vert f \right\Vert}_{C^2}=1} {\left\Vert f'' \right\Vert}_{\infty} \leq 1\, . \] Hence the $D_h^{(2)}$ is stable. Moreover, once again by the Mean Value Theorem, for some $0<\alpha, \beta, \gamma <1$ we have \begin{align*} \lim_{h\rightarrow 0}{\left\Vert D_h^{(2)}f-D^{(2)}f \right\Vert}_{\infty}&= \lim_{h\rightarrow 0} \sup_{x\in [a+h,b-h]} \left\vert\dfrac{f(x+h)-2f(x)+f(x-h)}{h^2}-f''(x) \right\vert\\ &=\lim_{h\rightarrow 0} \sup_{x\in [a+h,b-h]}\left\vert \dfrac{f'(x+ \alpha h)-f'(x- \beta h)}{h}-f''(x) \right\vert\\ &=\lim_{h\rightarrow 0} \sup_{x\in [a+h,b-h]} \left\vert f''(x-\beta h + \gamma h (\alpha + \beta))- f''(x)\right\vert=0 \, , \end{align*} which means convergence. Again, if the $C^2$ norm is replaced by the $C^1$ norm or the sup norm, the operator $D_h^{(2)}$ fails to be stable. Examples similar to Example~\ref{eg:unbounded_deriv} above can be constructed to exhibit the unboundedness of the operator $D_h^{(2)}$. \end{example} \section{Polynomial Interpolation} \label{sec:interp} The interpolation problem is to find a function that takes on prescribed values at specified points. In one dimension the data is given as $(x_i, y_i)$ for $i=0, \ldots,n$, with $x_0 < x_1 < \cdots < x_n$, and we look for a function $f : \mathbb{R} \rightarrow \mathbb{R}$ called interpolating function such that $f(x_i)= y_i$ for all $i$. In order to define the notions of consistency, convergence and stability it is convenient to pose the interpolation problem as interpolation of continuous functions. For any given data which is to be interpolated, there is the obvious unique piecewise linear continuous function that interpolates the data. Stability with respect to changes in the given values or the locations of the values are both captured by stability with respect to changes in the piecewise linear continuous function. We will only address interpolation of continuous functions by polynomials. It is easy to show the existence and uniqueness of the interpolant polynomial in the case of one dimensional interpolation, i.e., when the dimension of the domain of the function is one \cite{Heath2002, Walsh1965}. Let $\{p_n\}$ be a sequence of polynomials of degree at most $n$, interpolating a function $f$ in $V = \left(C[a,b], {\left\Vert \cdot \right \Vert}_{\infty}\right)$ on a set of interpolation points $x_i^{(n)}$, where $a \leq x_0^{(n)} < x_1^{(n)} < \cdots < x_n^{(n)} \leq b$. We can define polynomial interpolation as a map $P_n : V \rightarrow \mathbb{P}_n \subset V$, where $\mathbb{P}_n$ is the space of polynomials of degree at most $n$ and $P_n f $ is defined as the unique polynomial $p \in \mathbb{P}_n$ which interpolates $f$ at the given nodes. A basic fact about error in polynomial interpolation is given by the following lemma. \begin{lemma}\label{lem:interp_error} Let $f \in C^{(n+1)}[a,b]$, and let $a \leq x_0 < x_1 < \cdots < x_n \leq b$ be a partition of $[a,b]$. Then for all $x \in (a,b)$, there exists $\xi \in (\min\{x_0,x\}, \max\{x_n,x\})$ such that \[ f(x) - (P_n f)(x) = \frac{f^{(n+1)}(\xi)}{(n+1) !} \prod_{i=0}^{n}(x-x_i)\, . \] \end{lemma} \begin{proof} The proof requires repeated use of Rolle's Theorem. See page 119 of \cite{AtHa2005} for details. \end{proof} The following Lemma~\ref{lem:interp_is_linear} shows that the $P_n$ operators are linear, and Lemma~\ref{lem:interp_is_bounded} shows that they are bounded. The corresponding smooth operator is the identity map which is bounded linear. Thus the linear case definitions of consistency, convergence and stability given in Definition~\ref{defn:ccs_linear} apply here as well. \begin{lemma}[Linearity]\label{lem:interp_is_linear} The interpolation operator $P_n : V \rightarrow \mathbb{P}_n \subset V$ is linear. \end{lemma} \begin{proof} Let $P_n(f) = p_1$ and $P_n(g)=p_2$, where $f,g \in V$ and $p_1, p_2$ are the unique polynomials in $\mathbb{P}_n$ such that $f(x_i) = p_1(x_i)$ and $g(x_i) = p_2(x_i)$ for $0 \leq i \leq n$. Hence $(f+g)(x_i) = (p_1+p_2)(x_i)$. By uniqueness of polynomial interpolation $P_n(f+g) = p_1+p_2$. Hence $P_n(f +g) = P_nf+P_ng$. Similarly using uniqueness we can show that $P_n(cf) = cP_n(f)$, for any $c \in \mathbb{R}$. \end{proof} \begin{lemma}[Boundedness]\label{lem:interp_is_bounded} Each interpolation operator $P_n$ is bounded, i.e., $ \Vert P_n \Vert = \sup_{{\Vert f \Vert}_{\infty}=1} {\Vert P_n(f) \Vert}_{\infty}\leq K(n)$ where $K(n) > 0$. \end{lemma} \begin{proof} By Lemma~\ref{lem:interp_error}, $P_n (f) = f$, where $f$ is any polynomial of degree $\leq n$. Hence $P_n$ maps onto $\mathbb{P}_n$. Let $f_m$ be a sequence in $ P_n^{-1}(0)$ converging to $f$ in $ V$. Since $f_m(x_i)=0$ for all $0 \leq i \leq n$ and $\lim_{m \rightarrow \infty}f_m(x) =f(x)$ for all $x \in [a,b]$, we have $f(x_i)= 0$ for all $0 \leq i \leq n$. Therefore $f \in P_n^{-1}(0)$. Hence the kernel of $P_n$ is closed. Therefore, by Lemma~\ref{lem:closed_kernel}, $P_n$ is continuous and hence bounded. \end{proof} \begin{definition} Interpolation is {\bf convergent} if for any given $f \in V$ the sequence of interpolant polynomials $P_n f$ converges to $f$ in $V$. It is {\bf consistent} if there is a dense subspace $V_0$ of $V$ such that for any $f$ in $V_0$, the sequence of interpolant polynomials $P_n f$ converges to $f$ in $V$. Interpolation is said to be {\bf stable} if $\sup_n\Vert P_n\Vert < \infty$. \end{definition} If the function being interpolated is a polynomial then the interpolant is exact. This follows from Lemma~\ref{lem:interp_error}. Since polynomials are dense in $V$ (Weierstrass Approximation Theorem) this implies that interpolation of continuous functions by polynomials is consistent. We now prove an equivalence theorem for polynomial interpolation. In the notation of Theorem~\ref{thm:equiv_linear} the discrete operators $T_n$ of the theorem correspond to the polynomial interpolation operator $P_n$ and the smooth operator $T$ corresponds to the identity operator. \begin{theorem}[\bf Equivalence Theorem for Interpolation]\label{thm:interp} A consistent polynomial interpolation is convergent if and only if it is stable. \end{theorem} \begin{proof} The proof is identical to the one given for the general linear equivalence theorem (Theorem~\ref{thm:equiv_linear}) with discretization parameter $h$ replaced by $n$. Thus the proof of the equivalence theorem for quadrature (Theorem~\ref{thm:quadrature}) can be used here with appropriate modifications. \end{proof} \begin{example}[Runge's Function] \label{eg:runge_interp} It is well known that for the Runge's function $f(x)= 1/(1+25x^2)$ in the interval $[-1,1]$ the polynomial interpolants do not converge if the interpolation points are uniformly spaced. See for example \cite{IsKe1966}. As we noted earlier, interpolation of continuous functions by polynomials is consistent. However, the method is not stable if equispaced points are used for interpolation. In light of the equivalence theorem above (Theorem~\ref{thm:interp}) this agrees with the lack of convergence. To see the lack of stability, let $p(x)$ be any polynomial which is arbitrarily close to the Runge's function. Existence of $p(x)$ is guaranteed by the Weierstrass approximation Theorem. Now, the interpolating polynomials for $p(x)$ are exact as we increase the number of interpolating points. However, the interpolating polynomial for the Runge's function is not close to $p(x)$. This shows that polynomial interpolation of continuous functions using equispaced points is not discretely well-posedness, i.e., it is not stable. \end{example} Interpolating real valued functions whose domain has dimension greater than one is a much harder problem, unlike the one dimensional case where polynomial interpolant always exists and is unique. For example, in 2 variable interpolation, if data is specified at 3 points that are collinear, then there are infinitely many linear interpolants. The geometry of the point locations becomes an important factor in existence and uniqueness of the interpolant. For a result of this type in the 2 variable case see, for example, \cite{Phillips2003}. We need the existence and uniqueness of interpolant so that there is a well defined operator $P_n$. There does exist a generalization, called Ciarlet's error formula \cite{CiWa1971,CiRa1972,GaSa2000}, of Lemma \ref{lem:interp_error} to the multivariable case. Moreover, multivariate polynomials are dense in the sup norm, in the space of multivariate continuous functions on compact subsets of $\mathbb{R}^n$ (Stone-Weierstrass Theorem) \cite{Davis1975}. Thus when multivariate polynomial interpolation does exist and is unique, we get consistency as before. Consistency, stability and convergence can be defined as described in the univariate case and thus an equivalence theorem exists for the multivariate polynomial interpolation of continuous functions. Of course the conditions for stability and convergence are more complicated in the multivariate case and we do not address those here. In contrast with interpolation, approximation methods do not require agreement with data at specific points. For example one might seek a polynomial $p$ such that $\min_{p \in \mathbb{P}_n} {\Vert f- p\Vert}_{\infty}$ is attained. In this particular case one can show existence and uniqueness. See for instance Theorem 7.5.6 in \cite{Davis1975}. One can define the notion of convergence, consistency and stability for approximation problems in an identical fashion as for interpolation and prove a theorem like Theorem~\ref{thm:interp}. In the particular case considered above, the approximation always converges (Theorem 7.6.1 of \cite{Davis1975}) and hence is stable. \section{Conclusions and Future Work} We have shown that equivalence theorems can be proved in a general setting in numerical analysis. As in the classical Lax-Richtmyer equivalence theorem in PDEs, these theorems state that a consistent method is convergent if and only if it is stable. The notion of stability we used was that of discrete well posedness and we defined consistency to be convergence on a dense subspace. The discretizations of the smooth problems were considered to be linear or nonlinear operators on normed linear spaces and depending on some parameter which measures the discretization size. We showed that our general equivalence theorems require basic tools like uniform boundedness principle, triangle inequality, and density arguments. Convergence implying stability was always obtained independent of consistency. In this paper, we have studied stability with respect to perturbation of input data of the discrete operators. However, one could also study stability with respect to locations of points, shape of domain etc. One can also investigate equivalence theorems in other numerical analysis contexts. Some examples are, multidimensional quadrature rules, multidimensional Monte Carlo integration, optimization, eigenvalue problems in PDEs, formulas for numerical differentiation of any order and accuracy, such as those given in \cite{Fornberg1996}. We defined consistency as convergence on a dense subspace. In our examples, many times the dense subspace was polynomials in continuous functions. It may be worthwhile to study if and how the choice of particular dense subspace affects the theory and applications of equivalence theorems. \bibliographystyle{amsplain}

https://arxiv.org/abs/2009.09351

Counteracting Inequality in Markets via Convex Pricing

We study market mechanisms for allocating divisible goods to competing agents with quasilinear utilities. For \emph{linear} pricing (i.e., the cost of a good is proportional to the quantity purchased), the First Welfare Theorem states that Walrasian equilibria maximize the sum of agent valuations. This ensures efficiency, but can lead to extreme inequality across individuals. Many real-world markets -- especially for water -- use \emph{convex} pricing instead, often known as increasing block tariffs (IBTs). IBTs are thought to promote equality, but there is a dearth of theoretical support for this claim.In this paper, we study a simple convex pricing rule and show that the resulting equilibria are guaranteed to maximize a CES welfare function. Furthermore, a parameter of the pricing rule directly determines which CES welfare function is implemented; by tweaking this parameter, the social planner can precisely control the tradeoff between equality and efficiency. Our result holds for any valuations that are homogeneous, differentiable, and concave. We also give an iterative algorithm for computing these pricing rules, derive a truthful mechanism for the case of a single good, and discuss Sybil attacks.

\section{Towards an implementation}\label{sec:comp} Theorem~\ref{thm:main} guarantees the existence of Walrasian equilibria maximizing CES welfare, but says nothing about how to find these equilibria. As discussed in Section~\ref{sec:results}, we could always explicitly ask each agent for her valuation, and directly solve Program~\ref{pro:ces}. However, agents are generally not able to articulate their entire valuations, and even if they are, doing so could be extremely tedious. In this section, we give an iterative algorithm for computing the WE given by Theorem~\ref{thm:main}. The algorithm will just compute the optimal allocation; Lemma~\ref{lem:x-to-q} shows how the equilibrium pricing rule can easily be obtained once the optimal allocation is in hand. Our algorithm is computationally equivalent to running the general-purpose ellipsoid method on Program~\ref{pro:ces}, i.e., it explores the exact same sequence of allocations. The key is that we are able to implement the ellipsoid method only using valuation gradient queries, i.e., ``tell me the gradient of your valuation at this point". We immediately inherit the correctness and polynomial-time convergence properties of the ellipsoid algorithm. Throughout this section, we make the same assumptions as in Theorem~\ref{thm:main}: each $v_i$ is concave, homogeneous of degree $r$, and differentiable. First, recall that the pricing rule from Theorem~\ref{thm:main} takes the form $p(x_i) = \rho r^{\frac{\rho - 1}{\rho}} (\sum_{j \in M} q_j x_{ij})^{1/\rho}$. Since $\rho$ and $r$ are constants, it suffices to compute $\q = q_1,\dots,q_m$. Helpfully, Theorem~\ref{thm:main} tells us that if $\q$ are optimal Lagrange multipliers for Program~\ref{pro:ces}, then $(\mathbf{x}, p)$ is a WE for any $\mathbf{x} \in \Psi(\rho)$. The next lemma states if we know an $\mathbf{x} \in \Psi(\rho)$, and have access to the gradients of the agents' valuations at $\mathbf{x}$, we can determine optimal Lagrange multipliers. \begin{restatable}{lemma}{lemXToQ}\label{lem:x-to-q} Let $\mathbf{x} \in \Psi(\rho)$. Then we can determine optimal Lagrange multipliers $\q$ using only\\ $\nabla v_1(x_1),\dots,\nabla v_n(x_n)$. \end{restatable} \begin{proof} First, using Euler's Theorem for homogeneous functions (Theorem~\ref{thm:euler}), we can obtain $v_1(x_1),\dots,v_n(x_n)$ using only $\nabla v_1(x_1),\dots,\nabla v_n(x_n)$. Next, since $\mathbf{x} \in \Psi(\rho)$, the KKT conditions for Program~\ref{pro:ces} imply that whenever $x_{ij} > 0$, $q_j = v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$. Fix a $j \in M$. If $x_{ij} > 0$ for some agent $i$, then $q_j = v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$. We know all the values on the right hand side, so we can compute $q_j$. if $x_{ij} = 0$ for all $i \in N$, then complementary slackness implies that $q_j = 0$. \end{proof} Thus it suffices to find an allocation $\mathbf{x} \in \Psi(\rho)$, which is equivalent to finding an $\mathbf{x}$ that is optimal for Program~\ref{pro:ces}. There are many iterative algorithms for solving convex programs of this form. Furthermore, many only require (1) oracle access to the objective function and its gradient, and (2) \emph{a separation oracle}\footnote{A separation oracle is an algorithm which, given a point $x$ and a convex set $\mathcal{X}$, determines whether $x \in \mathcal{X}$. If $x\not\in\mathcal{X}$, it must return a separating hyperplane (if $\mathcal{X}$ is specified by a set of constraints, returning a violated constraint is sufficient).} for the constraint set (and no additional assumptions of strong convexity or other properties). For the sake of specificity, we focus on the \emph{ellipsoid method}~\cite{Bubeck2015}, but any algorithm with these properties is sufficient for our purposes. \begin{lemma}[\cite{Bubeck2015}]\label{lem:converge} Let $f$ be a convex function and let $\mathcal{X}$ be a convex set. Consider the program $\min_{x \in \mathcal{X}} f(x)$. Let $\E$ be a ball containing the minimum of $f$, and suppose there exists a polynomial-time separation oracle for $\mathcal{X}$. Then the ellipsoid method starting from $\E$ requires only oracle access to $f$ and $\nabla f$, and converges to the minimum of $f$ in polynomial time. \end{lemma} In our case, we have a trivial polynomial-time separation oracle: simply check each constraint to see if it is violated. For the gradient of our objective function, we have $\mfrac{\partial}{\partial x_{ij}} \Big(\frac{1}{\rho} \sum_{i \in N} v_i(x_i)^\rho\Big) = \mfrac{\partial v_i(x_i)}{\partial x_{ij}} v_i(x_i)^{\rho-1}$. By Euler's Theorem for homogeneous functions (Theorem~\ref{thm:euler}), we have \begin{align} \frac{\partial v_i(x_i)}{\partial x_{ij}} v_i(x_i)^{\rho-1} = \frac{\partial v_i(x_i)}{x_{ij}} \Big(r^{-1} \sum_{\ell \in M} x_{i\ell} \frac{\partial v_i(x_i)}{x_{i\ell}}\Big)^{\rho-1} \label{eq:euler} \end{align} Similarly, \begin{align} \frac{1}{\rho} \sum_{i \in N} v_i(x_i)^\rho = \frac{1}{\rho} \sum_{i \in N} \Big(r^{-1} \sum_{j \in M} x_{ij} \frac{\partial v_i(x_i)}{\partial x_{ij}}\Big)^\rho\label{eq:euler2} \end{align} Therefore for any allocation $\mathbf{x}$, we can compute both the objective function value and the gradient of the objective function using only the gradients of $v_1,\dots,v_n$. The final ingredient we need is an initial ball guaranteed to contain the optimum; we can simply enclose the entire feasible region in a ball of constant radius. Thus we get the following iterative algorithm for computing the equilibrium pricing rule: \begin{enumerate} \item Run the ellipsoid method (or any other suitable convex optimization algorithm) to solve Program~\ref{pro:ces}. \item At the start of each iteration, ask each agent $i$ for the gradient of $v_i$ at the current point $\mathbf{x}$. \item Whenever the algorithm requires the gradient of the objective function at $\mathbf{x}$, compute it via Equation~\ref{eq:euler}. \item Whenever the algorithm requires the objective function value at $\mathbf{x}$, compute it via Equation~\ref{eq:euler2}. \end{enumerate} Lemma~\ref{lem:converge} immediately implies correctness and polynomial-time convergence. \subsection{Eliciting the gradients of valuations} The above algorithm (as well as Lemma~\ref{lem:x-to-q}) requires us to have access the gradients of agents' valuations. We could simply ask each agent for this information explicitly; depending on the application domain, this may or may not be reasonable. An alternative approach is to relate $\nabla v_i(x_i)$ to agent $i$'s willingness to pay. For example, consider the following query to agent $i$: ``Suppose you have already bought the bundle $x_i$. What is the smallest marginal price for good $j$ such that you would not buy more of good $j$?" The KKT conditions for agent $i$'s demand set imply that the answer to this question is exactly $\mfrac{\partial v_i(x_i)}{\partial x_{ij}}$ (assuming that the agent would not buy more if she is indifferent). Such a query could be implemented in a variety of ways. One possibility would be gradually increasing the hypothetical marginal price of good $j$ in a continuous fashion, and asking agent $i$ to say ``stop" when she would no longer buy more of good $j$ (in a ``moving-knife"-like fashion). Also, rather asking agents about hypothetical marginal prices, one could build the necessary marginal prices into an actual pricing rule, e.g., even one as simple as $p(x_i) = \sum_{j \in M} c_j x_{ij}$. The choice of implementation would depend heavily on the specific problem setting; our point here is that there are a variety of ways to elicit $\nabla v_i(x_i)$ via queries about what agent $i$ would purchase in different (hypothetical) situations. \section{Conclusion} In this paper, we studied a simple family of convex pricing rules, motivated by the widespread use of convex pricing in the real world, especially for water. We proved that these pricing rules implement CES welfare maximization in Walrasian equilibrium, providing a formal quantitative interpretation of the frequent informal claim that convex pricing promotes equality. Furthermore, by tweaking the exponent of the pricing rule, the social planner can precisely control the tradeoff between equality and efficiency. This result also shows that convex pricing is not necessarily economically inefficient, as often claimed; it simply maximizes a different welfare function than the traditional utilitarian one. Improved implementation is perhaps the most important of the future directions we propose. One concrete possibility is a \emph{\tat}: an iterative algorithm where on each step, each agent reports her demand for the current pricing rule, and the pricing rule is adjusted accordingly. Demand queries are arguably easier for agents to answer than valuation gradient queries. Some implementation questions -- in particular, how to deal with Sybil attacks -- would likely need to be handled on a case-by-case basis. Aside from the implementation itself, there is the additional challenge of convincing market designers to consider using this type of convex pricing. Equality is generally thought to be desirable, but sellers may be concerned that this will decrease their revenue. In future work, we hope to show that our pricing rule guarantees a good approximation of the optimal revenue for sellers. Another possible direction would be CES welfare maximization for indivisible goods. The analogous pricing rule would be $p(S) = (\sum_{j \in S} q_j)^{1/\rho}$, where $S$ is a set of indivisible goods. It seems like very different theoretical techniques would be needed in this setting (along with perhaps a gross substitutes assumption), but we suspect that the same intuition of convex pricing improving equality would hold. \section{Negative results}\label{sec:counter} Even when Sybil attacks are not possible, there are limitations to implementation in WE. This section presents several relevant counterexamples. \subsection{Linear pricing poorly approximates CES welfare for $\rho \ne 1$} Recall that for an allocation $\mathbf{x}$, $\Phi(\rho, \mathbf{x})$ denotes the CES welfare of $\mathbf{x}$. In contrast, $\Psi(\rho)$ denotes the set of allocations with optimal CES welfare with respect to $\rho$. Our first negative result relates to linear pricing. In particular, can linear pricing guarantee a reasonable approximation of CES welfare? We show that the answer is no, justifying the need for nonlinear pricing. In particular, for any $\rho \in (0,1)$, the gap between the CES welfare of any linear pricing equilibrium and the optimal CES welfare can be arbitrarily large. Note that as $\rho$ goes to zero, $\frac{1}{\rho} - 1$ goes to infinity, so the denominator of the bound (and thus the gap in CES welfare) in the following theorem can indeed be arbitrarily large. \begin{theorem}\label{thm:rho=1-bad} Let $m=1$, $\rho \in (0,1]$, $v_1(x) = (1+\ep) x$ for some $\ep > 0$, and $v_i(x) = x$ for all $i \ne 1$. Suppose $(\mathbf{x}, p)$ is a WE where $p$ is linear. Then \[ \frac{\Phi(\rho, \mathbf{x})}{\max_{\mathbf{y}} \Phi(\rho, \mathbf{y})} \le \frac{1+\ep}{n^{\frac{1}{\rho}-1}} \] \end{theorem} \begin{proof} By the First Welfare Theorem, $\mathbf{x}$ must maximize utilitarian (i.e., $\rho = 1$) welfare. Thus by Lemma~\ref{lem:m=1-x}, $\mathbf{x}$ must give the entire good to agent 1: $x_1 = 1$ and $x_i = 0$ for $i \ne 1$. Thus the CES welfare of $\mathbf{x}$ with respect to $\rho$ is \[ \Phi(\rho, \mathbf{x}) = \Big(\sum_{i \in N} v_i(x_i)^\rho\Big)^{1/\rho} = \big( (1+\ep)^\rho \big)^{1/\rho} \] In contrast, consider the allocation $\mathbf{y}$ such that $y_i = 1/n$ for all $i \in N$: \[ \Phi(\rho, \mathbf{y}) = \Big( \sum_{i \in N} v_i(1/n)^{\rho} \Big)^{1/\rho} \ge \Big(\sum_{i \in N} (1/n)^\rho \Big)^{1/\rho} = \Big(n (1/n)^{1/\rho} \Big)^{1/\rho} = n^{\frac{1}{\rho} - 1} \] Thus $\max_{\mathbf{y}} \Phi(\rho, \mathbf{y}) \ge n^{\frac{1}{\rho} - 1}$, as required. \end{proof} \subsection{Theorem~\ref{thm:main} does not extend to nonuniform homogeneity degrees} In this section, we show that for all $\rho \in (0,1)$, Theorem~\ref{thm:main} does not extend to the case where different $v_i$'s have different homogeneity degrees. This shows that our result is tight in the sense that it is necessary to require the same homogeneity degree. We begin with the following lemma, which is a standard property of strictly concave and differentiable functions: it essentially states that any such function is bounded above by any tangent line. This lemma is sometimes called the ``Rooftop Theorem". \begin{lemma}\label{lem:rooftop} Let $f: \bbr \to \bbr$ be strictly concave and differentiable. Then for all $a,b \in \bbr$ where $a\ne b$, $f(a) < f(b) + f'(b)(a-b)$, where $f'$ denotes the derivative of $f$. \end{lemma} The next lemma is also quite standard; we provide a proof for completeness. \begin{lemma}\label{lem:concave-bound} Let $f: \bbr \to \bbr$ be strictly concave and differentiable, and let $x, a_1, \dots, a_k$ be nonnegative reals such that $\sum_{i=1}^k a_i = 0$. Then $\sum_{i =1}^k f(x + a_i) < kf(x)$. \end{lemma} \begin{proof} The lemma follows from Lemma~\ref{lem:rooftop} and arithmetic: \begin{align*} \sum_{i =1}^k f(x + a_i) <&\ \sum_{i=1}^k (f(x) + f'(x)(x+a_i - x))\\ =&\ \sum_{i=1}^k f(x) + f'(x) \sum_{i=1}^k a_i\\ =&\ \sum_{i=1}^k f(x) + f'(x) \cdot 0\\ =&\ kf(x) \end{align*} \end{proof} We are now ready to present our counterexample. \begin{theorem}\label{thm:different-r} Let $n =2$ and $m=1$, and for $x \in \bbrpos$, let $v_1(x) = x$ and $v_2(x) = \sqrt{2x}$. Then for all $\rho \in (0,1)$, there exists no allocation $\mathbf{x} \in \Psi(\rho)$ and pricing rule $p: \bbrpos \to \bbrpos$ such that $(\mathbf{x}, p)$ is a WE. \end{theorem} \begin{proof} Suppose for sake of contradiction that such $\mathbf{x}, p$ do exist. We first claim that $x_1 > x_2$. Suppose the opposite: then $x_2 \ge 1/2 \ge x_1$. Thus $v_2(x_2) \ge 1 > 1/2 \ge v_1(x_1)$. We also have $\mfrac{\partial v_2(x_2)}{\partial x_2} = \mfrac{1}{\sqrt{2x_2}} \le 1 = \mfrac{\partial v_1(x_1)}{\partial x_1}$. Thus $v_2(x_2) > v_1(x_2)$ and $\mfrac{\partial v_2(x_2)}{\partial x_2} \le \mfrac{\partial v_1(x_1)}{\partial x_1}$. Since $\rho < 1$, $\rho - 1 < 0$, so we have $v_2(x_2)^{\rho - 1}\mfrac{\partial v_2(x_2)}{\partial x_2} < v_1(x_1)^{\rho - 1}\mfrac{\partial v_1(x_1)}{\partial x_2}$. But this contradicts $\mathbf{x} \in \Psi(\rho)$, so we have $x_1 > x_2$ as claimed.\footnote{This immediately implies $\frac{\partial v_2(x_2)}{\partial x_2} < 1 =\frac{\partial v_1(x_1)}{\partial x_1}$, which, in combination with $x_1 > x_2$, rules out convex $p$. However, we still need to rule out non-convex $p$.} Since $(\mathbf{x}, p)$ is a WE, we must have $x_i \in D_i(p)$ for both agents $i$. Thus for any $x \ne x_i$, $v_i(x_i) - p(x_i) \ge v_i(x) - p(x)$. Therefore \begin{align*} v_1(x_1) - p(x_1) \ge&\ v_1(x_2) - p(x_2) \quad \text{and} \quad v_2(x_2) - p(x_2) \ge v_2(x_1) - p(x_1)\\ v_1(x_1) + v_2(x_2) - p(x_1) - p(x_2) \ge&\ v_1(x_2) + v_2(x_1) - p(x_1) - p(x_2)\\ v_1(x_1) + v_2(x_2) \ge&\ v_1(x_2) + v_2(x_1)\\ v_1(x_1) - v_1(x_2) \ge&\ v_2(x_1) - v_2(x_2) \end{align*} Since $x_1 > 1/2 > x_2$ and $x_1 + x_2 = 1$, let $x_1 = 1/2 + \ep$ and $x_2 = 1/2 - \ep$. Then we have $v_1(x_1) - v_1(x_2) = 2\ep$. For $v_2(x_1) - v_2(x_2)$, we have \begin{align*} v_2(x_1) - v_2(x_2) =&\ \sqrt{1+2\ep} - \sqrt{1-2\ep}\\ =&\ \frac{(\sqrt{1+2\ep} - \sqrt{1-2\ep})(\sqrt{1+2\ep} - \sqrt{1-2\ep})} {\sqrt{1+2\ep} + \sqrt{1-2\ep}}\\ =&\ \frac{(1+2\ep) - (1-2\ep)}{\sqrt{1+2\ep} + \sqrt{1-2\ep}}\\ =&\ \frac{4\ep}{\sqrt{1+2\ep} + \sqrt{1-2\ep}} \end{align*} Applying Lemma~\ref{lem:concave-bound} with $f(x) = \sqrt{x}$, $x=1$, $k=2$, and $(a_1, a_2) = (2\ep, -2\ep)$, we get $\sqrt{1+2\ep} + \sqrt{1-2\ep} < 2$. Thus $v_2(x_1) - v_2(x_2) > 4\ep/2 = 2\ep = v_1(x_1) - v_1(x_2)$. However, this contradicts $v_1(x_1) - v_1(x_2) \ge v_2(x_1) - v_2(x_2)$, as we showed above. We conclude that there is no $\mathbf{x} \in \Psi(\rho)$ and pricing rule $p$ such that $(\mathbf{x}, p)$ is a WE. \end{proof} \subsection{CES welfare maximization for $\rho \le 0$} In this section, we show that there is no pricing rule supporting CES welfare maximization for any $\rho < 0$. For $\rho = 0$ (i.e., Nash welfare), the situation is slightly different. We do show, however, that Nash welfare maximization cannot be supported by a differentiable pricing rule. \begin{theorem}\label{thm:neg-rho-counter} Consider the instance with $n=2$, $m=1$, $v_1(x) = x$ and $v_2(x) = 2x$. Then for every $\rho < 0$, there is no pricing rule $p$ and allocation $\mathbf{x} \in \Psi(\rho)$ such that $(\mathbf{x}, p)$ is a WE. \end{theorem} \begin{proof} For any $\rho < 0$ and any $\mathbf{x} \in \Psi(\rho)$, we must have $x_1 > x_2$. Assume $(\mathbf{x}, p)$ is a WE for some pricing rule $p$: then $x_1 \in D_1(p)$, so $v_1(x_1) - p(x_1) \ge v_1(x_2) - p(x_2)$. Thus $p(x_1) \le p(x_2) + v_1(x_1) - v_2(x_2) = p(x_2) + x_2 - x_1$. Therefore \begin{align*} v_2(x_1) - p(x_1) \ge&\ 2x_1 - (p(x_2) + x_2 - x_1)\\ =&\ 3x_1 - x_2 - p(x_2)\\ >&\ 2x_1 - p(x_2)\\ >&\ 2x_2 - p(x_2)\\ =&\ v_2(x_2) - p(x_2) \end{align*} Thus agent 2 would rather purchase $x_1$ than $x_2$, so $x_2 \not\in D_2(p)$. Therefore $(\mathbf{x}, p)$ is not a WE. \end{proof} For $\rho = 0$, the situation is different. Recall that Fisher market equilibrium always maximizes Nash welfare, and we can simulate Fisher market budgets by setting \[ p(x_i) = \begin{cases} 0 &\ \text{ if } \sum_{j \in M} q_j x_{ij} \le 1\\ \infty &\ \text{ otherwise} \end{cases} \] where $q_1\dots q_m$ are the optimal Lagrange multipliers in the convex program for maximizing Nash welfare. Gale and Eisenberg's famous result implies that for such a pricing rule, a WE always exists, and all WE maximize Nash welfare~\cite{Eisenberg1961,Eisenberg1959}. Note that for $\sum_{j \in M} q_j x_{ij} > 1$, $p(x_i) = \infty$ can be implemented by setting $\mfrac{\partial p(x_i)}{\partial x_{ij}}$ to be at least $\max_{i \in N} \max_{x_i \in [0,1]^m} \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$. This ensures that no agent purchases a bundle $x_i$ such that $\sum_{j \in M} q_j x_{ij} > 1$. The above pricing rule is somewhat artificial, however. One natural question is whether Nash welfare maximization can be implemented with a differentiable pricing rule. We next show that the answer is no. \begin{theorem}\label{thm:nash-counter} Consider the instance with $n=2$, $m=1$, $v_1(x) = x$ and $v_2(x) = 2x$. Then there is no allocation $\mathbf{x}$ maximizing Nash welfare and differentiable pricing rule $p$ such that $(\mathbf{x}, p)$ is a WE. \end{theorem} \begin{proof} Suppose the opposite: that such $\mathbf{x}, p$ exist. The unique $\mathbf{x}$ maximizing Nash welfare must have $x_1 = x_2 = 1/2$. Since $p, v_1,$ and $v_2$ are all differentiable, we have $x_i \in D_i(p)$ if and only if $\mfrac{\dif p(x_i)}{\dif x_i} = \mfrac{\dif v_i(x_i)}{\dif x_i}$. Since $x_1 = x_2$, we have $\mfrac{\dif p(x_1)}{\dif x_1} = \mfrac{\dif p(x_2)}{\dif x_2}$. Thus implies $\mfrac{\dif v_1(x_1)}{\dif x_1} = \mfrac{\dif v_2(x_2)}{\dif x_2}$, which is a contradiction. We conclude that no such $\mathbf{x}, p$ exist. \end{proof} \section{Omitted proofs}\label{sec:proofs} \thmEuler* \begin{proof} Fix an arbitrary $\B \in \bbrpos^m$ and let $g(\lambda) = f(\lambda\B)$. Since $f$ is differentiable, so is $g$, and its derivative is given by the multidimensional chain rule: $\frac{\dif g(\lambda)}{\dif \lambda} = \sum_{j = 1}^m b_j \frac{\partial f(\lambda\B)}{\partial b_j}$. Since $f$ is homogeneous of degree $r$, we have $f(\lambda\B) = \lambda^r f(\B)$ for all $\lambda \ge 0$. Thus $g(\lambda) = \lambda^r f(\B)$ for all $\lambda \ge 0$, so we can differentiable both sides of this equation to get $\sum_{j = 1}^m b_j \frac{\partial f(\lambda\B)}{\partial b_j} = r\lambda^{r-1} f(\B)$. This holds for all $\lambda \ge 0$, so setting $\lambda = 1$ completes the proof. \end{proof} \lemHomoOneGood* \begin{proof} By Euler's Theorem (Theorem~\ref{thm:euler}), we have $x \mfrac{\dif f(x)}{\dif x} = r f(x)$ for all $x \in \bbrpos$. Let $y = f(x)$. We can solve this differential equation explicitly: \begin{align*} \frac{1}{y}\cdot \frac{\dif y}{\dif x} =&\ \frac{r}{x}\\ \int \frac{1}{y}\cdot \frac{\dif y}{\dif x} \dif x =&\ \int \frac{r}{x} \dif x\\ \int \frac{1}{y} \dif y =&\ r \int \frac{1}{x} \dif x\\ \ln y =&\ r \ln x + \ln c \end{align*} where $c$ (and thus $\ln c$) is some constant. Therefore \begin{align*} e^y =&\ e^{r \ln x + \ln c}\\ y =&\ c x^r \end{align*} Thus $f(x) = c x^r$, as required. \end{proof} \lemOptOneGood* \begin{proof} As in Section~\ref{sec:main}, strong duality for Program~\ref{pro:ces} implies that any optimal $\mathbf{x}$ must satisfy the KKT conditions. Thus $\mathbf{x} \in \Psi(\rho)$ if and only if there exists $q \in \bbrpos$ such that (1) stationarity holds: $\mfrac{\partial v_i(x_i)}{\partial x_i} v_i(x_i)^{\rho -1} \le q$ for all $i \in N$, and if $x_i > 0$, the inequality holds with equality, and (2) complementary slackness holds: either $\sum_{i \in N} x_i = 1$, or $q = 0$. Since we assume that $w_i > 0$ for all $i \in N$, any allocation with $\sum_{i \in N} x_i < 1$ is not Pareto optimal, and thus cannot be optimal for Program~\ref{pro:ces}. In other words, we must have $q > 0$. Thus complementary slackness simply requires that $\sum_{i \in N} x_i = 1$, and we can focus on stationarity. We first handle $\rho = 1$. In this case, $\mfrac{\partial v_i(x_i)}{\partial x_i} v_i(x_i)^{\rho -1} = \mfrac{\partial v_i(x_i)}{\partial x_i} = w_i$. Thus if $\mathbf{x} \in \Psi(\rho)$ we must have $w_i \le q$, and if $x_i > 0$, then $w_i = q$. This implies that $q = \max_{k \in N} w_k$. Thus if $x_i > 0$, then $w_i = \max_{k \in N} w_k$, as required. For the rest of the proof, we assume $r\rho \ne 1$. Since $r,\rho \in (0,1]$, we have $0 < r\rho < 1$. By the definition of $v_i$, for an arbitrary allocation $\mathbf{x}$ and $i \in N$ we have \[ \frac{\partial v_i(x_i)}{\partial x_i} v_i(x_i)^{\rho -1} = (w_i r x_i^{r-1})(w_i^{\rho-1} x_i^{r(\rho-1)}) = r w_i^\rho x_i^{r\rho - 1} \] Thus $q \ge rw_i^\rho x_i^{r\rho - 1}$. Since $r\rho < 1$, if $x_i = 0$, then $x_i^{r\rho-1}$ is undefined. Therefore stationarity is satisfied if and only if $q = r w_i^\rho x_i^{r\rho-1}$ for all $i \in N$, which is equivalent to \begin{align} x_i = (q/r)^{\frac{1}{r\rho-1}} w_i^{\frac{\rho}{1-r\rho}} \label{eq:x-expr} \end{align} Furthermore, if $\mathbf{x}$ satisfies Equation~\ref{eq:x-expr} for all $i \in N$, then $\sum_{i \in N} x_i = 1$ is equivalent to \begin{align*} \sum_{i \in N} (q/r)^{\frac{1}{r\rho-1}} w_i^{\frac{\rho}{1-r\rho}} =&\ 1\\ (q/r)^{\frac{1}{r\rho-1}} =&\ \Big(\sum_{i \in N} {w_i}^{\frac{\rho}{1-r\rho}}\Big)^{-1} \end{align*} Therefore $\mathbf{x}$ satisfies stationarity and complementary slackness (and thus satisfies $\mathbf{x} \in \Psi(\rho)$) if and only if \[ x_i = \frac{{w_i}^{\frac{\rho}{1-r\rho}}}{\sum_{k \in N} {w_k}^{\frac{\rho}{1-r\rho}}} \] as required. \end{proof} \section{Connections to Fisher markets}\label{sec:fisher} The focus of this paper is on markets for quasilinear utilities, where agents can spend as much money as they want, and the amount spent is incorporated into their resulting utility. The other predominant market model assumes each agent $i$ has a finite budget $B_i$ of money to spend, and has no value for leftover money (in general, this implies that each agent $i$ spends exactly $B_i$). This is called the \emph{Fisher market} model.\footnote{There are also more general versions of this model that allow each agent's initial endowment to be goods instead of money (``exchange economies") and/or allow production (``Arrow-Debreu markets").} In this section, we explore connections between our results and the Fisher market model. In the Fisher market model, each agent's utility $u_i(x_i)$ is simply $v_i(x_i)$. For pricing rule $p$, the Fisher market demand set is given by \[ D_i^F(p) = \argmax_{x_i \in \bbrpos^m:\ p(x_i) \le B_i} v_i(x_i) \] We will reserve the notation $D_i(p)$ for the demand set in the quasilinear case, i.e., $D_i(p) = \argmax_{x_i \in \bbrpos^m} (v_i(x_i) - p(x_i))$. For an allocation $\mathbf{x}$, agent budgets $\bigB = (B_1,\dots,B_n)$, and a pricing rule $p$, $(\mathbf{x}, \bigB, p)$ is a \emph{Fisher market Walrasian equilibrium} if (1) $x_i \in D_i^F(p)$ for all $i \in N$, and (2) $\sum_{i \in N} x_{ij} \le 1$ for all $j \in M$, and if good $j$ has nonzero cost, $\sum_{i \in N} x_{ij} = 1$.\footnote{Recall that good $j$ has nonzero cost for $j$ if there is a bundle $x_i$ such that $x_{i\ell} = 0$ for all $\ell \ne j$, but $p(x_i) > 0$.} These are the same two conditions for Walrasian equilibrium in quasilinear markets: the only change is the definition of the demand set. To distinguish, we will use the terms ``Fisher WE" and ``quasilinear WE". \subsection{CES welfare maximization in Fisher markets} In the quasilinear model, agents can express not only their relative values between goods, but also the absolute scale of their valuation (i.e., the ``intensity" of their preferences) by choosing how much money to spend. In contrast, agents in the Fisher market model spend exactly their budget, and so have no way to express the absolute scale of their valuation. This should make us pessimistic about the possibility of CES welfare maximization in the Fisher market model in general. Indeed, consider a single good and two agents with valuations $v_1(x) = x$, $v_2(x) = 2x$. For any $\rho > 0$, any optimal allocation $\mathbf{x} \in \Psi(\rho)$ has $x_2 > x_1$. But if $B_1 = B_2$, any Fisher market Walrasian equilibrium will always have $x_1 = x_2$, since both agents simply spend their entire budget on the single good. However, in general we can convert a quasilinear WE to a Fisher WE if the agents' budgets are sized appropriately. Specifically, we need agent $i$'s budget to be exactly the amount she pays in the quasilinear WE: \begin{theorem}\label{thm:fisher-we} Suppose $(\mathbf{x}, p)$ is a quasilinear WE, and let $B_i = p(x_i)$. Then $(\mathbf{x}, \bigB, p)$ is a Fisher WE. \end{theorem} \begin{proof} For all $i \in N$, $x_i$ is affordable to agent $i$ under $p$ by definition of $B_i$. Suppose there were another bundle $y_i$ such that $p(y_i) \le B_i$ but $v_i(y_i) > v_i(x_i)$. That would contradict $x_i \in D_i(p)$ for the quasilinear case, since $u_i(y_i) = v_i(y_i) - p(y_i) > v_i(x_i) - p(y_i) \ge v_i(x_i) - p(x_i) = u_i(x_i)$. Therefore $x_i \in D_i^F(p)$ for all $i \in N$. Furthermore, the market clearing conditions for Fisher WE and quasilinear WE are identical. We conclude that $(\mathbf{x}, \bigB, p)$ is a Fisher WE. \end{proof} Combining the above result with Theorem~\ref{thm:main} gives us the following corollary for CES welfare maximization: \begin{corollary}\label{cor:fisher-we} Assume each $v_i$ is homogeneous of degree $r$, concave, and differentiable. Let $\rho \in (0, 1]$, and $p(x_i) = \rho r^{\frac{\rho -1}{\rho}} (\sum_{j \in M} q_j x_{ij})^{1/\rho}$, where $q_1,\dots, q_m$ are optimal Lagrange multipliers for Program~\ref{pro:ces}. For $\mathbf{x} \in \Psi(\rho)$, let $B_i = p(x_i)$ for all $i \in N$, Then $(\mathbf{x}, \bigB, p)$ is a Fisher WE. \end{corollary} Perhaps the more interesting connection relates to the welfare function being optimized. In the case of linear pricing, the Fisher market Walrasian equilibria are exactly the budget-weighted maximum Nash welfare allocations.\footnote{Recall that Nash welfare corresponds to $\rho = 0$, and the budget-weighted Nash welfare of an allocation $\mathbf{x}$ is $\prod_{i \in N} v_i(x_i)^{B_i}$.} One natural question is whether the Fisher market equilibria from Theorem~\ref{thm:fisher-we} also optimize a budget-weighted CES welfare function. We answer this in the affirmative. Recall that we define $\Phi_\bigB(\rho, \mathbf{x}) = \big(\sum_{i \in N} B_i v_i(x_i)^{\rho}\big)^{1/\rho}$, and $\Psi_\bigB(\rho) = \argmax_\mathbf{x} \Phi_\bigB(\rho, \mathbf{x})$. \begin{lemma}\label{lem:fisher-double-opt} Assume each $v_i$ is concave and differentiable. Let $\xprime$ be any allocation, let $a_i = v_i(x_i')$ for each $i \in N$, and let $\rho \in (0, 1]$. Then $\xprime \in \Psi(\rho)$ if and only if $\xprime \in \Psi_\mathbf{a}(\rho - 1)$. \end{lemma} \begin{proof} When $\rho = 1$, $\rho - 1 = 0$, so Program~\ref{pro:ces} does not apply, and we must handle this case separately. We first consider $\rho \ne 1$. The Lagrangian for Program~\ref{pro:ces} for $\Psi_\mathbf{a}(\rho-1)$ is $L(\mathbf{x}, \q) = \frac{1}{\rho-1} \sum_{i \in N} a_i v_i(x_i)^{\rho-1} - \sum_{j \in M} q_j (\sum_{i \in N} x_{ij} - 1)$. The KKT conditions imply that $\mathbf{x} \in \Psi_\mathbf{a}(\rho-1)$ if and only if there exist Lagrange multipliers $q_1,\dots, q_m$ such that: \begin{enumerate} \item Stationarity: $\mfrac{\partial L(\mathbf{x}, \q)}{\partial x_{ij}} = a_i v_i(x_i)^{\rho - 2} \mfrac{\partial v_i(x_i)}{\partial x_{ij}} \le 0$ for all $i,j$.\footnote{Note that since $\xprime$ is a fixed allocation, $a_i$ is just some constant, so the differentiation does not affect it.} Furthermore, if $x_{ij} > 0$, the inequality holds with equality. \item Complementary slackness: for all $j \in M$, either $\sum_{i \in N} x_{ij} = 1$, or $q_j = 0$. \end{enumerate} For Program~\ref{pro:ces} for $\Psi(\rho)$, as before we have $L'(\mathbf{x}, \q) = \frac{1}{\rho} \sum_{i \in N} v_i(x_i)^\rho - \sum_{j \in M} q_j (\sum_{i \in N} x_{ij} - 1)$. Thus the KKT conditions imply that $\mathbf{x} \in \Psi(\rho)$ if and only if there exist $q_1',\dots,q_m' \in \bbrpos$ such that (1) $v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}} \le q_j$ for all $i,j$, and when $x_{ij} > 0$, the inequality holds with equality, and (2) for all $j \in M$, either $\sum_{i \in N} x_{ij} = 1$, or $q_j = 0$. Note that if $q_j = q_j'$ for all $j \in M$, the complementary slackness conditions become equivalent. Next, for $\mathbf{x} = \xprime$ we have \[ v_i(x_i')^{\rho - 1} \frac{\partial v_i(x_i')}{\partial x'_{ij}} = v_i(x'_i) v_i(x'_i)^{\rho - 2} \frac{\partial v_i(x'_i)}{\partial x'_{ij}} = a_i v_i(x'_i)^{\rho - 2} \frac{\partial v_i(x'_i)}{\partial x'_{ij}} \] Therefore for given $q_j$, we have $q_j \ge v_i(x'_i)^{\rho - 1} \mfrac{\partial v_i(x'_i)}{\partial x'_{ij}}$ if and only if $q_j \ge a_i v_i(x'_i)^{\rho - 2} \mfrac{\partial v_i(x'_i)}{\partial x'_{ij}}$, and $q_j = v_i(x'_i)^{\rho - 1} \mfrac{\partial v_i(x'_i)}{\partial x'_{ij}}$ if and only if $q_j \ge a_i v_i(x'_i)^{\rho - 2} \mfrac{\partial v_i(x'_i)}{\partial x'_{ij}}$. Now suppose $\xprime \in \Psi(\rho)$. Then there exist $q_1,\dots, q_m \in \bbrpos$ that satisfy both stationarity and complementary slackness. Then as we showed above, $\xprime$ and $q_1,\dots, q_m$ satisfy stationarity for $\Psi_\mathbf{a}(\rho-1)$. Furthermore, the complementary slackness conditions are equivalent, so we have $\xprime \in \Psi_\mathbf{a}(\rho-1)$. Similarly, suppose $\xprime \in \Psi_\mathbf{a}(\rho-1)$. Then there exist $q_1,\dots, q_m$ satisfying stationarity and complementary slackness, so the same $q_1,\dots, q_m$ along with $\xprime$ satisfy the KKT conditions for $\Psi(\rho)$. Therefore $\Psi(\rho)$, and we conclude that $\xprime \in \Psi(\rho)$ if and only if $\xprime \in \Psi_\mathbf{a}(\rho-1)$ for $\rho \ne 1$. All of the above was for $\rho \ne 1$; it remains to handle the case of $\rho = 1$. In this case, we can use the same KKT conditions for $\Psi(\rho)$, but must use a different convex program for $\Psi_\mathbf{a}(\rho-1)$. Consider the following convex program for maximizing Nash welfare (i.e., CES welfare for $\rho = 0$): \begin{alignat}{2} \max\limits_{\mathbf{x} \in \bbrpos^{n\times m}} &\ \sum_{i \in N} a_i \log v_i(x_i) \label{pro:eg} \\ s.t.\ &\ \sum\limits_{i \in N} x_{ij}\leq 1\quad &&\ \forall j \in M \nonumber \end{alignat} This is known as the Eisenberg-Gale program~\cite{Eisenberg1961,Eisenberg1959}. In this case, the stationarity condition requires that $\mfrac{\partial}{\partial x_{ij}} a_i \log v_i(x_i) = a_i v_i(x_i)^{-1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}} \le q_j$ for all $i,j$, and when $x_{ij} > 0$, the inequality holds with equality. Since $\rho = 1$ here, we have $\rho - 2 = - 1$. Thus the stationarity condition for $\Psi_\mathbf{a}(\rho - 1)$ requires that $a v_i(x_i)^{\rho-2} \mfrac{\partial v_i(x_i)}{\partial x_{ij}} \le q_j$ for all $i,j$ (and if $x_{ij} > 0$, this holds with equality). This is exactly what we had above, and since we are using the same KKT conditions for $\Psi(\rho)$, this case reduces to the case for $\rho \ne 1$. Therefore for $\rho = 1$, $\xprime \in \Psi(\rho)$ if and only if $\xprime \in \Psi_\mathbf{a}(\rho-1)$. \end{proof} Combining Theorem~\ref{thm:fisher-we} and Lemma~\ref{lem:fisher-double-opt}, we get: \begin{theorem}\label{thm:full} Assume each $v_i$ is homogeneous of degree $r$, concave, and differentiable, let $\rho \in (0, 1]$, let $q_1,\dots,q_m \in \bbrpos$, and let $p(x_i) = \rho \big(\sum_{j \in M} q_j x_{ij}\big)^{1/\rho}$. Given $\mathbf{x} \in \Psi(\rho)$, let $B_i = p(x_i)$. Then all of the following hold: \begin{enumerate} \item $(\mathbf{x}, p)$ is a quasilinear WE. \item $(\mathbf{x}, \bigB, p)$ is a Fisher WE. \item $\mathbf{x} \in \Psi_\bigB(\rho - 1)$ \end{enumerate} \end{theorem} \begin{proof} The first and second conditions hold by Theorems~\ref{thm:main} and \ref{thm:fisher-we}, respectively. Then Corollary~\ref{cor:utility-price} implies that $p(x_i) = r\rho v_i(x_i)$. Let $a_i = v_i(x_i) = \mfrac{B_i}{r\rho}$. Thus by Lemma~\ref{lem:fisher-double-opt}, we have $\mathbf{x} \in \Psi_\mathbf{a}(\rho-1)$. Since scaling all agents' multipliers by the same factor does not affect $\Psi_\mathbf{a}(\rho-1)$, we have $\mathbf{x} \in \Psi_{r\rho \mathbf{a}}(\rho-1) = \Psi_\bigB(\rho-1)$, as required. \end{proof} It is worth noting that the special case of Theorem~\ref{thm:fisher-we} for $\rho = 1$ and Leontief utilities with $w_{ij} \in \{0,1\}$\footnote{This is also known as the bandwidth allocation setting, where each good represents a link in a network, and agent $i$ has $w_{ij} = 1$ for every link $j$ on a fixed path (and $w_{ij} = 0$ otherwise).} is implied by the work of Kelly et al.~\cite{Kelly1998}. \section{Introduction}\label{sec:intro} Markets are one of the oldest mechanisms for distributing resources; indeed, commodity prices were meticulously recorded in ancient Babylon for over 300 years~\cite{Spek2005,Spek2003}. In a market, buyers and sellers exchange goods according to some sort of pricing system, and \emph{Walrasian equilibrium}\footnote{This is also known as market equilibrium, competitive equilibrium, and general equilibrium, depending on the context.} occurs when the demand of the buyers exactly equals the supply of the sellers. This concept was first studied by Walras in the 1870's~\cite{Walras1874}. In 1954, Arrow and Debreu showed that under some conditions, a Walrasian equilibrium is guaranteed to exist~\cite{Arrow1954}. Most of the literature on Walrasian equilibrium only considers \emph{linear} pricing, meaning the cost of a good is proportional to the quantity purchased. In this paper, we consider the problem of allocating divisible goods to competing agents via a market mechanism. We assume each agent has quasilinear utility: an agent's utility is her value for the resources she obtains (her \emph{valuation}), minus the money she spends (her \emph{payment}). The First Welfare Theorem states that in this setting, the linear-pricing Walrasian equilibria are exactly the allocations maximizing utilitarian welfare, i.e., the sum of agent valuations. Thus linear pricing \emph{implements} utilitarian welfare in Walrasian equilibrium (sometimes abbreviated ``WE"). The result is powerful, but also limiting. Maximizing utilitarian welfare yields the most efficient outcome, but may also cause maximal inequality (see Figure~\ref{fig:linear}). One common alternative is \emph{convex} pricing. In this paper, we study convex pricing rules $p$ of the form \begin{align*} p(x_i) = \Big(\sum_{j} q_j x_{ij}\Big)^{1/\rho} \end{align*} where $x_i$ is \emph{bundle} agent $i$ receives, $x_{ij} \in \bbrpos$ is the fraction of good $j$ she receives, $q_1,\dots,q_m$ are constants, and $\rho \in (0,1]$ determines the curvature of the pricing rule. Like linear pricing, $p$ is still \emph{anonymous}, meaning that agents' payments depend only on their purchases (and not on their preferences, for example). When $\rho =1$, $p$ reduces to linear pricing. When $\rho < 1$, $p$ is strictly convex, meaning that doubling one's consumption will more than double the price. This will make it easy to buy a small amount, but hard to buy a large amount, which intuitively should lead to a more equal distribution of resources. As the curvature of the pricing rule grows, this effect should be amplified, leading to a different equality/efficiency tradeoff. Our work seeks to formalize that claim. We will show that the Walrasian equilibria of these convex pricing rules are guaranteed to maximize a \emph{constant elasticity of substitution} (CES) welfare function, where the choice of $\rho$ determines the specific welfare function and thus the precise equality/efficiency tradeoff (Theorem~\ref{thm:main}). Our result holds for a wide range of agent valuations. \input{water_fig} \paragraph{Convex pricing in the real world.} Convex pricing is especially pervasive in the water sector, where such pricing rules are known as \emph{increasing block tariffs} (IBTs)~\cite{Whittington1992}, typically implemented with discrete blocks of water (hence the name). IBTs have been implemented and empirically studied in Israel~\cite{Becker2015}, South Africa~\cite{Burger2014}, Spain~\cite{Garcia-Rubio2015}, Jordan~\cite{Klassert2018}, and the United States~\cite{Renwick2000}, among many other countries. IBTs are often claimed to promote equality in water access~\cite{Whittington1992}, but there has been limited theoretical evidence supporting this (see~\cite{Monteiro2011} for one of the only examples). On the other hand, a common concern is that IBTs may lead to poor ``economic efficiency"~\cite{Boland2000,Monteiro2011}. Our work shows that at least on a theoretical level, convexity of pricing does not necessarily lead to inefficiency: it simply maximizes a different welfare function than the traditional utilitarian one. In particular, it maximizes a CES welfare function. \paragraph{The Second Welfare Theorem and personalized pricing.} The Second Welfare Theorem is perhaps the most famous theoretical result regarding implementation in Walrasian equilibrium. It states any Pareto optimum can be a WE when an arbitrary redistribution of initial wealth is allowed.\footnote{Specifically, for any Pareto optimal allocation, there exists a redistribution of initial wealth which makes that allocation a WE. However, our quasilinear utility model does not have a concept of initial wealth (alternatively, initial wealth is simply an additive constant in agents' utilities which does not affect their behavior), so this result is not as mathematically relevant. See Section~\ref{sec:prior} for additional discussion.} Another method that achieves the same goal is \emph{personalized pricing}, where different agents can be charged different (linear) prices. In contrast, convex pricing is anonymous: agents purchasing the same bundle always pay the same price. Each of these approaches certainly has its own pros and cons. In this paper, our goal is not to claim that convex pricing is ``better" than other approaches (or vice versa). Regardless of which is ``better" in any given situation, convex pricing \emph{is} widely used in practice, and is often claimed to promote equality. Our goal in this paper is to formally quantify that claim. \subsection{CES welfare functions}\label{sec:ces-intro} A welfare function~\cite{Bergson1938,Samuelson1947} assigns a real number to each possible outcome, with higher numbers (i.e, higher welfare) indicating outcomes that are more desirable to the social planner. Different welfare functions represent different priorities; our focus will be the tradeoff between overall efficiency and individual equality. For a fixed constant $\rho \in (-\infty, 0) \cup (0,1]$, the constant elasticity of substitution (CES) welfare of outcome $\mathbf{x}$ is \[ \Phi(\rho, \mathbf{x}) = \bigg(\sum_{\text{agents } i} v_i(\mathbf{x})^{\rho}\bigg)^{1/\rho} \] where $v_i(\mathbf{x})$ is agent $i$'s value for $\mathbf{x}$. In general, different values of $\rho$ will lead to different optimal allocations, so whenever we say ``maximum CES welfare allocation", we mean with respect to a fixed value of $\rho$. See Section~\ref{sec:model} for an axiomatic characterization of CES welfare functions. For $\rho = 1$, this is utilitarian welfare, i.e., the sum of valuations. The limit as $\rho \to -\infty$ yields max-min welfare (the minimum valuation)~\cite{Rawls2009,Sen1977,Sen1976}, whereas $\rho \to 0$ yields Nash welfare (the product of valuations)~\cite{Kaneko1979,Nash1950}. The closer $\rho$ gets to $-\infty$, the more the social planner cares about individual equality (max-min welfare being the extreme case of this), and the closer $\rho$ gets to 1, the more the social planner cares about overall societal good (utilitarian welfare being the extreme case of this). For this reason, $\rho$ is called the \emph{inequality aversion} parameter, and this family of welfare functions is thought to exhibit an \emph{equality/efficiency} tradeoff. These welfare functions were originally proposed by Atkinson~\cite{Atkinson70}; indeed, his motivation was to measure the level of inequality in a society. Despite being extremely influential in the traditional economics literature (see~\cite{Cowell2011} for a survey), the CES welfare function has received almost no attention in the computational economics community.\footnote{To our knowledge, only three other computational economics papers have studied CES welfare in any context:~\cite{Arunachaleswaran2019,Goel2019nash,Plaut2019}.} Finally, note that $\Phi(\rho, \mathbf{x})$ is defined with respect to the each agent's valuation $v_i$ and \emph{not} her overall quasilinear utility $u_i$. We acknowledge that it is standard to define welfare with respect to the overall utility $u_i$, and we have two reasons for not doing so. First, in the case of scarce resources, a social planner may be interested in equality in \emph{consumption} (e.g., equality in water access), not just equality in utility derived. Second, it turns out mathematically that this is the welfare function maximized by convex pricing WE in our model; the version where $\Phi(\rho, \mathbf{x})$ considers $u_i$ may be not be maximized by the resulting WE. This may yield valuable qualitative insights about convex pricing; for example, does convex pricing lead to equality with respect to consumption but not necessarily with respect to underlying utility? \paragraph{CES welfare in healthcare.} These welfare functions have also seen substantial use in healthcare under the name of \emph{isoelastic welfare functions}. This began with~\cite{Wagstaff1991}, largely motivated by concerns abeout purely utilitarian approaches to healthcare (i.e., allocating resources to maximize total health in a community, without concern for equality). Since these decisions can affect who lives and who dies, significant effort has been invested into understanding the equality/efficiency tradeoff, with this class of welfare functions serving as a theoretical tool~\cite{Dolan1998,Ortun1996,Wagstaff1991}; see Section~\ref{sec:prior} for additional discussion. Despite the ongoing interest in this tradeoff, the healthcare literature has not (to our knowledge) considered convex pricing as a mechanism for balancing equality and efficiency. More broadly, our work can be thought of as weaving together the previously disjoint threads of CES welfare and convex pricing to provide theoretical support for the oft-cited but rarely quantified claim that IBTs promote equality. \section{CES welfare maximization for Leontief valuations}\label{sec:leontief} We say that $v_i$ is \emph{Leontief} if there exist weights $w_1,\dots,w_m \in \bbrpos$ such that \[ v_i(x_i) = \min_{j:\ w_{ij} \ne 0} \frac{x_{ij}}{w_{ij}} \] Leontief valuations are not differentiable, and so Theorem~\ref{thm:main} does not apply. In this section, we handle Leontief valuations as a special case. Although there are many non-differentiable valuations we could consider, there is substantial related work on Leontief valuations (\cite{Goel2019nash,Plaut2019}), so we find it worthwhile to show that our result does indeed extend to this case. Recall Program~\ref{pro:ces}: \begin{alignat*}{2} \max\limits_{\mathbf{x} \in \bbrpos^{n\times m}} &\ \frac{1}{\rho}\sum_{i \in N} v_i(x_i)^\rho \\ s.t.\ &\ \sum\limits_{i \in N} x_{ij}\leq 1\quad &&\ \forall j \in M \nonumber \end{alignat*} We will work with a specialized version of this for Leontief utilities: \begin{alignat}{2} \max\limits_{\mathbf{x} \in \bbrpos^{n\times m}, \bfa \in \bbrpos^m} \frac{1}{\rho}\sum_{i \in N} \alpha_i^\rho &\ \label{pro:leontief}\\ s.t.\ w_{ij} \alpha_i \le&\ x_{ij}\quad &&\ \forall i \in N, j \in M \nonumber\\ \sum\limits_{i \in N} x_{ij}\le&\ 1\quad &&\ \forall j \in M \nonumber \end{alignat} where we use $\bfa$ to denote the vector $(\alpha_1,\dots,\alpha_n) \in \bbrpos^n$. Also recall each agent's demand set $D_i(p) = \argmax_{x_i \in \bbrpos^m}\ \big( v_i(x_i) - p(x_i)\big)$. Similarly to Program~\ref{pro:leontief}, we consider the following equivalent (specialized) convex program for agent $i$'s demand set: \begin{alignat}{2} \max\limits_{x_i \in \bbrpos^m, \alpha_i \in \bbrpos} \big(\alpha_i - p(x_i)\big) &\ \label{pro:leontief-demand}\\ s.t.\ w_{ij} \alpha_i \le&\ x_{ij}\quad &&\ \forall j \in M \nonumber \end{alignat} \begin{theorem}\label{thm:leontief} Assume each $v_i$ is Leontief with weights $w_{i1},\dots,w_{im}$. Then for any $\rho \in (0, 1]$ and any allocation $\mathbf{x}$, we have $\mathbf{x} \in \Psi(\rho)$ if and only if there exist $q_1,\dots, q_m\in \bbrpos$ such that for the pricing rule $p(x_i) = \rho (\sum_{j \in M} q_j x_{ij})^{1/\rho}$, $(\mathbf{x}, p)$ is a WE. Furthermore, $q_1,\dots, q_m$ are optimal Lagrange multipliers (for the $\sum_{i \in N} x_{ij} \le 1$ constraints) for Program~\ref{pro:leontief}. \end{theorem} \begin{proof} We first claim that in an optimal solution $\mathbf{x},\bfa$ to either Program~\ref{pro:leontief} or Program~\ref{pro:leontief-demand}, we have $v_i(x_i) = \alpha_i$ for all $i \in N$: that is, that these programs are doing what we want them to. To see this, note that $\alpha_i \le x_{ij} / w_{ij}$ for all $j$ with $w_{ij} \ne 0$, so $\alpha_i \le v_i(x_i)$. Furthermore, at least one constraint involving $\alpha_i$ must be tight: otherwise, we could increase $\alpha_i$ and thus the objective value. In particular, we must have $\alpha_i = \min_{j:\ w_{ij} \ne 0} \frac{x_{ij}}{w_{ij}} = v_i(x_i)$. Thus Program~\ref{pro:leontief-demand} is indeed maximizing $v_i(x_i) - p(x_i)$, so $x_i \in D_i(p)$ if and only if $(x_i, \alpha_i)$ is optimal for Program~\ref{pro:leontief-demand} (for some $\alpha_i$). Similarly, Program~\ref{pro:leontief} is indeed maximizing $\frac{1}{\rho}\sum_{i \in N} v_i(x_i)^\rho$ subject to $\sum_{i \in N} x_{ij} \le 1$ for all $j \in M$, so $(\mathbf{x}, \bfa)$ is optimal for Program~\ref{pro:leontief} if and only if $\mathbf{x}$ is optimal for Program~\ref{pro:ces}. Therefore $\mathbf{x} \in \Psi(\rho)$ if and only if $(\mathbf{x}, \bfa)$ is optimal for Program~\ref{pro:leontief} (for some $\bfa$). Next, we write the Lagrangian of Program~\ref{pro:leontief}:\footnote{As in the proof of Theorem~\ref{thm:main}, we omit the $\mathbf{x} \in \bbrpos^{m\times n}$ constraint from the Lagrangian incorporate it into the KKT conditions instead.} \[ L(\mathbf{x}, \bfa, \q, \bflam) = \frac{1}{\rho} \sum_{i \in N} \alpha_i^\rho - \sum_{i \in N} \sum_{j \in M} \lambda_{ij} (w_{ij} \alpha_i - x_{ij}) - \sum_{j \in M} q_j \Big(\sum_{i \in N} x_{ij} - 1\Big) \] We have strong duality by Slater's condition, so the KKT conditions are both necessary and sufficient for optimality. That is, $(\mathbf{x}, \bfa)$ is optimal if and only if there exist $\q \in \bbrpos^m$, $\bflam \in \bbrpos^{m\times n}$ such that all of the following hold:\footnote{As in the proof of Theorem~\ref{thm:main}, primal and dual feasibility are trivially satisfied.} \begin{enumerate} \item Stationarity for $\mathbf{x}$: $\mfrac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial x_{ij}} \le 0$ for all $i,j$. Furthermore, if $x_{ij} > 0$, the inequality holds with equality. \item Stationarity for $\bfa$: $\mfrac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial \alpha_i} \le 0$ for all $i \in N$. Furthermore, if $\alpha_i > 0$, the inequality holds with equality. \item Complementary slackness for $\q$: for all $j \in M$, either $\sum_{i \in N} x_{ij} = 1$, or $q_j = 0$. \item Complementary slackness for $\bflam$: for all $i \in N$, $j\in M$, either $w_{ij} \alpha_i = x_{ij}$ or $\lambda_{ij} = 0$. \end{enumerate} Similarly, let $L_i'$ denote the Lagrangian of Program~\ref{pro:leontief-demand} for agent $i$: \[ L_i'(x_i, \alpha_i, \lambda_i) = \alpha_i - p(x_i) - \sum_{j \in M} \lambda_{ij} (w_{ij} \alpha_i - x_{ij}) \] where $\lambda_i = (\lambda_{i1},\dots,\lambda_{im}) \in \bbrpos^m$. We again have strong duality, so the KKT conditions are again necessary and sufficient. Let $L_i'(x_i, \alpha_i, \lambda_i)$ denote the Lagrangian of this program; then $(x_i, \alpha_i)$ is optimal for Program~\ref{pro:leontief-demand} if and only if all of the following hold: \begin{enumerate} \item Stationarity for $x_i$: $\mfrac{\partial L_i'(x_i, \alpha_i, \lambda_i)}{\partial x_{ij}} \le 0$ for all $j \in M$. If $x_{ij} > 0$, the inequality holds with equality. \item Stationarity for $\alpha_i$: $\mfrac{\partial L_i'(x_i, \alpha_i, \lambda_i)}{\partial \alpha_i} \le 0$. If $\alpha_i > 0$, the inequality holds with equality. \item Complementary slackness for $\lambda_i$: for all $i \in N$, $j\in M$, either $w_{ij} \alpha_i = x_{ij}$ or $\lambda_{ij} = 0$. \end{enumerate} We will claim that $(\mathbf{x}, \bfa, \q, \bflam)$ is optimal for Program~\ref{pro:leontief} if and only if for all $i \in N$, $(x_i, \alpha_i, \alpha_i^{1-\rho} \lambda_i)$ is optimal for Program~\ref{pro:leontief-demand}. Essentially, we show that if complementary slackness holds (for either program), the stationarity conditions are equivalent. To begin, we can explicitly compute the relevant partial derivatives for given $\mathbf{x}, \bfa, \q, \bflam$, with $p(x_i) = \rho (\sum_{j \in M} q_j x_{ij})^{1/\rho}$: \begin{align*} \frac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial x_{ij}} =&\ \lambda_{ij} - q_j \\ \frac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial \alpha_i} =&\ \alpha_i^{\rho - 1} - \sum_{j \in M} \lambda_{ij} w_{ij}\\ \frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial x_{ij}} =&\ \lambda'_{ij} - q_j \Big(\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{\frac{1-\rho}{\rho}}\\ \frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial \alpha_i} =&\ 1 - \sum_{j \in M} \lambda'_{ij} w_{ij} \end{align*} \textbf{Part 1:} $(\implies)$ Suppose that $\mathbf{x} \in \Psi(\rho)$. Then there exist $\bfa, \q, \bflam$ such that the KKT conditions for Program~\ref{pro:leontief} are satisfied for $(\mathbf{x}, \bfa, \q, \bflam)$. We first claim that $\alpha_i > 0$ for all $i \in N$. Suppose not: stationarity implies that $\alpha_i^{\rho-1} \le \sum_{j \in M} \lambda_{ij} w_{ij}$, but since $\rho - 1< 0$, the left hand side is not defined for $\alpha_i = 0$. Thus $\alpha_i > 0$. Therefore by stationarity for $\alpha_i$, we have $\alpha_i^{\rho-1} = \sum_{j \in M} \lambda_{ij} w_{ij}$. Let $\lambda'_{ij} = \alpha_i^{1-\rho} \lambda_{ij}$ for all $i,j$\footnote{Note that this is \emph{not} defining $\lambda'_{ij}$ to be a function of $\alpha_i$. This is defining $\lambda'_{ij}$ based on a fixed value of $\alpha_i$: in particular, the value from $(\mathbf{x}, \bfa, \q, \bflam)$, which we assumed to be optimal for Program~\ref{pro:leontief}. Consequently, the derivatives in the KKT conditions treat $\lambda'_{ij}$ as a constant.}. Then $\alpha_i^{\rho-1} = \sum_{j \in M} \lambda_{ij} w_{ij}$ is equivalent to $1 = \sum_{j \in M} \lambda'_{ij} w_{ij}$, and thus $\mfrac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial \alpha_i} = 0$ for all $i \in N$. Thus for all $i \in N$, $(x_i, \alpha_i, \lambda'_i)$ satisfies stationarity for $\alpha_i$ for Program~\ref{pro:leontief-demand}. We now turn to the $x_{ij}$ variables. Stationarity for $x_{ij}$ in Program~\ref{pro:leontief} implies that $\lambda_{ij} = q_j$ whenever $x_{ij} > 0$. Furthermore, complementary slackness for $\lambda_{ij}$ implies that if $\lambda_{ij} > 0$, $w_{ij} \alpha_i = x_{ij}$. Thus whenever $q_j > 0$ and $x_{ij} > 0$, $w_{ij} \alpha_i = x_{ij}$ and $\lambda_{ij} = q_j$. Therefore for all $i,j$, \begin{align*} \frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial x_{ij}} =&\ \lambda'_{ij} - q_j \Big(\sum_{\ell: q_\ell, x_{i\ell} > 0} q_\ell x_{i\ell} \Big)^{\frac{1-\rho}{\rho}}\\ =&\ \lambda'_{ij} - q_j \Big(\sum_{\ell \in M} \lambda_{i\ell} w_{i\ell} \alpha_i\Big)^{\frac{1-\rho}{\rho}}\\ =&\ \lambda'_{ij} - q_j \Big(\alpha_i \sum_{\ell \in M} \lambda_{i\ell} w_{i\ell} \Big)^{\frac{1-\rho}{\rho}}\\ =&\ \lambda'_{ij} - q_j (\alpha_i \alpha_i^{\rho-1})^{\frac{1-\rho}{\rho}}\\ =&\ \alpha_i^{1-\rho} \lambda_{ij} - q_j \alpha_i^{1-\rho}\\ =&\ \alpha_i^{1-\rho} \frac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial x_{ij}} \end{align*} We have $\mfrac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial x_{ij}} \le 0$ for all $i,j$ by stationarity (and the inequality holds with equality when $x_{ij} > 0$), so $\frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial x_{ij}} \le 0$ for all $j \in M$ (and the inequality holds with equality when $x_{ij} > 0$). Thus for each $i \in N$, $(x_i, \alpha_i, \lambda'_i)$ satisfies stationarity for Program~\ref{pro:leontief-demand} for $x_{ij}$ for all $j \in M$. As mentioned above, we have $w_{ij} \alpha_i = x_{ij}$ whenever $\lambda_{ij} > 0$. Since $\lambda'_{ij} > 0$ if and only if $\lambda_{ij} > 0$, we have $w_{ij} \alpha_i = x_{ij}$ whenever $\lambda'_{ij} > 0$. Thus for each $i \in N$, $(x_i, \alpha_i, \lambda'_i)$ satisfies complementary slackness for Program~\ref{pro:leontief-demand}. Therefore $(x_i, \alpha_i, \lambda'_i)$ satisfies the KKT conditions, and thus is optimal for Program~\ref{pro:leontief-demand}. Therefore $x_i \in D_i(p)$ for all $i \in N$. The complementary slackness condition for $\q$ is identical to the market clearing condition, so we conclude that $(\mathbf{x}, p)$ is a WE. \textbf{Part 2:} $(\impliedby)$ Suppose that $(\mathbf{x}, p)$ is a WE, where $p(x_i) = \rho (\sum_{j \in M} q_j x_{ij})^{1/\rho}$ for constants $q_1,\dots,q_m \in \bbrpos$. Then $x_i \in D_i(p)$ for all $i \in N$, so there exists $\bfa, \bflamp$ such that $(x_i, \alpha_i, \lambda'_i)$ is optimal for Program~\ref{pro:leontief-demand} for all $i \in N$. Thus by stationarity, we have $\frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial \alpha_i} \le 0$ and $\frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial x_{ij}} \le 0$ for all $i,j$ (and if $\alpha_i > 0$ and $x_{ij} > 0$, the inequalities hold with equality). Using the definition of $p$, we have $\frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial x_{ij}} = \lambda'_{ij} - q_j \big(p(x_i)/\rho\big)^{1-\rho}$. Thus $1 \le \sum_{j \in M} \lambda'_{ij} w_{ij}$ and $\lambda'_{ij} \le q_j \big(p(x_i)/\rho\big)^{1-\rho}$. We first claim that $\alpha_i > 0$ for all $i \in N$. For each agent $i$, there must exist $j \in M$ such that $\lambda'_{ij} > 0$ and $w_{ij} > 0$: otherwise $1 \le \sum_{j \in M} \lambda'_{ij} w_{ij}$ would be impossible. Consider any such $j$: then $0 < \lambda'_{ij} \le q_j \big(p(x_i)/\rho\big)^{1-\rho}$, so we must have $p(x_i) > 0$. Suppose $\alpha_i = 0$: then the optimal objective value of Program~\ref{pro:leontief-demand} is $\alpha_i - p(x_i) < 0$. But setting $x_{ij} = 0$ for all $j \in M$ achieves an objective value of 0, so $\alpha_i - p(x_i) < 0$ cannot be optimal. This is a contradiction, and so $\alpha_i > 0$ for all $i \in N$. Returning to the stationarity conditions, we then have $1 = \sum_{j \in M}\lambda'_{ij} w_{ij}$. Complementary slackness implies that $w_{ij} \alpha_i = x_{ij}$ whenever $\lambda'_{ij} > 0$, so we get \begin{align*} \alpha_i =&\ \sum_{j \in M} \lambda'_{ij} w_{ij} \alpha_i \\ \alpha_i =&\ \sum_{j: \lambda'_{ij} > 0} \lambda'_{ij} w_{ij} \alpha_i \\ \alpha_i =&\ \sum_{j \in M} \lambda'_{ij} x_{ij} \end{align*} Combining this with $\lambda'_{ij} = q_j \big(p(x_i)/\rho\big)^{1-\rho}$ whenever $x_{ij} > 0$ gives us \begin{align*} \alpha_i =&\ \sum_{j: x_{ij} > 0} \lambda'_{ij} x_{ij}\\ =&\ \sum_{j \in M} q_j x_{ij} \big(p(x_i)/\rho\big)^{1-\rho}\\ =&\ \big(p(x_i)/\rho\big)^{1-\rho} \sum_{j \in M} q_j x_{ij} \\ =&\ \big(p(x_i)/\rho\big)^{1-\rho} \big(p(x_i)/\rho\big)^\rho \\ =&\ p(x_i)/\rho \end{align*} Now let $\lambda_{ij} = \alpha_i^{\rho - 1} \lambda'_{ij}$ for all $i,j$. We claim that $(\mathbf{x}, \bfa, \q, \bflam)$ satisfies the KKT conditions for Program~\ref{pro:leontief}. For each $(i,j)$ pair, we have \[ \frac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial \alpha_i} = \alpha_i^{\rho - 1} - \sum_{j \in M} \lambda_{ij} w_{ij} = \alpha_i^{\rho - 1}\Big(1 - \sum_{j \in M} \lambda'_{ij} w_{ij}\Big) = \alpha_i^{\rho - 1} \frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial \alpha_i} \] Since $\alpha_i > 0$, stationarity for Program~\ref{pro:leontief-demand} implies that $\frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial \alpha_i} = 0$, so we have $\frac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial \alpha_i} = 0$. Next, we have \begin{align*} \frac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial x_{ij}} =&\ \lambda_{ij} - q_j\\ =&\ \alpha_i^{\rho - 1} \lambda'_{ij} - q_j\\ =&\ \alpha_i^{\rho - 1} ( \lambda'_{ij} - q_j \alpha_i^{1-\rho})\\ =&\ \alpha_i^{\rho - 1} \Big(\lambda'_{ij} - q_j \big(p(x_i)/\rho\big)^{1-\rho}\Big)\\ =&\ \alpha_i^{\rho - 1} \frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial x_{ij}} \end{align*} Stationarity for Program~\ref{pro:leontief-demand} implies that $\frac{\partial L_i'(x_i, \alpha_i, \lambda'_i)}{\partial x_{ij}} \le 0$ for all $i,j$ (and when $x_{ij} > 0$, this holds with equality), so we have $\frac{\partial L(\mathbf{x}, \bfa, \q, \bflam)}{\partial x_{ij}}$ for all $i,j$ (and when $x_{ij} > 0$, this holds with equality). Thus we have shown that $(\mathbf{x}, \bfa, \q, \bflam)$ satisfies stationarity for Program~\ref{pro:leontief}. As before, the market clearing condition is equivalent to complementary slackness for $\q$. By complementary slackness for $\bflamp$ (for Program~\ref{pro:leontief-demand}), we have $w_{ij} \alpha_i = x_{ij}$ whenever $\lambda'_{ij} > 0$. By definition, $\lambda'_{ij} > 0$ if and only if $\lambda_{ij} > 0$, so this implies the required complementary slackness for $\bflam$ (for Program~\ref{pro:leontief}). Therefore $(\mathbf{x}, \bfa, \q, \bflam)$ satisfies the KKT conditions for Program~\ref{pro:leontief}, and thus is optimal for that program. We conclude that $\mathbf{x} \in \Psi(\rho)$. \end{proof} \section{The First Welfare Theorem and linear pricing}\label{sec:linear} Recall our main result: \mainThm* For this class of valuations, Theorem~\ref{thm:main} for $\rho = 1$ implies the First Welfare Theorem: $p$ becomes a linear pricing rule, and CES welfare for $\rho = 1$ is just utilitarian welfare. In particular, Theorem~\ref{thm:main} implies both the existence of a WE, and that every WE maximizes utilitarian welfare. Typically, the ``the First Welfare Theorem" refers to just half of this: that every WE maximizes utilitarian welfare. The reason is that WE are not always guaranteed to exist: for divisible goods, generally at least concavity or quasi-concavity of valuations is necessary. On the other hand, very few assumptions are needed to show that linear pricing equilibria always maximize utilitarian welfare; for example, divisibility of goods is not needed. We provide a proof of this below. \begin{theorem}[The First Welfare Theorem]\label{thm:first-welfare} Let $\X_i \subset \bbr^m$ denote the set of feasible bundles for agent $i$ (not necessarily convex, and not necessarily the same for all agents). Let $D_i(p) = \argmax_{x_i \in \X_i} (v_i(x_i) - p(x_i))$ and assume $p$ is linear. Then if $(\mathbf{x}, p)$ is a WE, $\mathbf{x}$ maximizes utilitarian welfare. \end{theorem} \begin{proof} Since $p$ is linear, there exist $q_1,\dots,q_m$ such that $p(y_i) = \sum_{j \in M} q_j y_{ij}$ for any bundle $y_i$. Consider an arbitrary feasible allocation $\mathbf{y}$. Since $(\mathbf{x}, p)$ is a WE, we have $x_i \in D_i(p)$, so $v_i(x_i) - p(x_i) \ge v_i(y_i) - p(y_i)$. Therefore \begin{align*} v_i(x_i) - \sum_{j \in M} q_j x_{ij} \ge&\ v_i(y_i) - \sum_{j \in M} q_j y_{ij}\\ \sum_{i \in N} v_i(x_i) - \sum_{i \in N} \sum_{j \in M} q_j x_{ij} \ge&\ \sum_{i \in N} v_i(y_i) - \sum_{i \in N} \sum_{j \in M} q_j x_{ij}\\ \sum_{i \in N} v_i(x_i) - \sum_{j \in M} q_j \sum_{i \in N} x_{ij} \ge&\ \sum_{i \in N} v_i(y_i) - \sum_{j \in M} q_j \sum_{i \in N} x_{ij} \end{align*} Furthermore, $\sum_{i \in N} x_{ij} = 1$ for all $j \in M$ with $q_j > 0$. Also, since $\mathbf{y}$ is a valid allocation, $\sum_{i \in N} y_{ij} \le 1$ for all $j \in M$. Therefore \begin{align*} \sum_{i \in N} v_i(x_i) - \sum_{j \in M} q_j \sum_{i \in N} x_{ij} \ge&\ \sum_{i \in N} v_i(y_i) - \sum_{j \in M} q_j \sum_{i \in N} x_{ij}\\ \sum_{i \in N} v_i(x_i) - \sum_{j \in M} q_j \ge&\ \sum_{i \in N} v_i(y_i) - \sum_{j \in M} q_j\\ \sum_{i \in N} v_i(x_i) \ge&\ \sum_{i \in N} v_i(y_i) \end{align*} Thus the utilitarian welfare of $\mathbf{x}$ is at least as high as that of any other allocation. We conclude that $\mathbf{x}$ maximizes utilitarian welfare. \end{proof} Note that no assumptions at all were made on the nature of the valuations: all we needed was $x_i \in \argmax_{y_i \in \X_i} (v_i(y_i) - p(y_i))$, and $\sum_{j \in M} x_{ij} = 1$ whenever $q_j > 0$. The most natural cases are $\X_i = \bbrpos^m$ (divisible goods) and $\X_i = \{0,1\}^m$ (indivisible goods), but the result does hold more broadly. \section{Main result}\label{sec:main} We begin with our main result: for a wide range of valuations and any $\rho \in (0,1]$, a simple convex pricing rule leads to CES welfare maximization in Walrasian equilibrium. Our pricing rule has many additional interesting properties; to avoid redundancy, we refer the reader back to our discussion in Section~\ref{sec:results}. On a high level, the proof relies on the KKT conditions for CES welfare maximization and the KKT conditions for each agent's demand set, and uses Euler's Theorem for homogeneous functions to conjoin the two. This will result in the following theorem: \begin{restatable}{theorem}{mainThm}\label{thm:main} Assume each $v_i$ is homogeneous of degree $r$, concave, and differentiable. For any $\rho \in (0, 1]$ and any allocation $\mathbf{x}$, we have $\mathbf{x} \in \Psi(\rho)$ if and only if there exist $q_1,\dots, q_m\in \bbrpos$ such that for the pricing rule \[ p(x_i) = \rho r^{\frac{\rho -1}{\rho}} \Big(\sum_{j \in M} q_j x_{ij}\Big)^{1/\rho}, \] $(\mathbf{x}, p)$ is a WE. Furthermore, $q_1,\dots, q_m$ are optimal Lagrange multipliers for Program~\ref{pro:ces}. \end{restatable} \subsection{Proof setup} We begin by setting up the two relevant convex programs and proving several lemmas. For valuations $v_1\dots v_n$, nonnegative multipliers $\mathbf{a} = a_1\dots a_n$, and $\rho \in (-\infty, 0) \cup(0,1]$, consider the following nonlinear program for maximizing CES welfare: \begin{alignat}{2} \max\limits_{\mathbf{x} \in \bbrpos^{n\times m}} &\ \frac{1}{\rho}\sum_{i \in N} a_i v_i(x_i)^\rho \label{pro:ces} \\ s.t.\ &\ \sum\limits_{i \in N} x_{ij}\leq 1\quad &&\ \forall j \in M \nonumber \end{alignat} Since the constraints are linear and the objective function is concave (since $\rho \le 1$), Program~\ref{pro:ces} is a convex program. Program~\ref{pro:ces} depends on $\rho$, but we will leave this implicit when clear from context: we will simply say ``$\mathbf{x}$ is optimal for Program~\ref{pro:ces}" as opposed to ``$\mathbf{x}$ is optimal for Program~\ref{pro:ces} with respect to $\rho$". Note also that we are maximizing $\frac{1}{\rho}\sum_{i \in N} a_i v_i(x_i)^\rho$ instead of the true CES welfare $\Phi_\mathbf{a}(\rho, \mathbf{x}) = (\sum_{i \in N} a_i v_i(x_i)^\rho)^{1/\rho}$; this will lead to the same optimal allocation $\mathbf{x}$ and will simplify the analysis. When $\mathbf{a}$ is not specified, we assume that $\mathbf{a} = \one$. Nonuniform multipliers will only be used in Appendix~\ref{sec:fisher} when we consider connections to Fisher markets, but we include them here for completeness. Next, consider each agent's demand set given a pricing rule $p$: \begin{align} D_i(p) = \argmax_{x_i \in \bbrpos^m}\ \big(v_i(x_i) - p(x_i)\big) \label{pro:demand} \end{align} When $p$ is convex (as in Theorem~\ref{thm:main}), $-p$ is concave. Since $v_i$ is also concave, $v_i(x_i) - p(x_i)$ is concave, so each agent's demand set defines a convex program (Program~\ref{pro:demand}). Program~\ref{pro:demand} depends on $i$, the agent in question, but again we leave this implicit when it is clear from context. We will also use the following theorem, due to Euler. We include a short proof in Appendix~\ref{sec:proofs}.\footnote{The reason we provide a proof is that this theorem is often stated with the requirement of continuous differentiability, but in fact only requires differentiability; to avoid any confusion, we provide a proof only using differentiability.} \begin{restatable}[Euler's Theorem for homogeneous functions]{theorem}{thmEuler}\label{thm:euler} Let $f: \bbrpos^m \to \bbr$ be differentiable and homogenous of degree $r$. Then for any $\B = (b_1,\dots b_m) \in \bbrpos^m$, $\sum_{j = 1}^m b_j \mfrac{\partial f(\B)}{\partial b_j} = r f(\B)$. \end{restatable} Before we state and prove Theorem~\ref{thm:main}, we note one other property: for a pricing rule of the form $p(x_i) = c(\sum_{j \in M} q_j x_{ij})^{1/\rho}$ where $c>0$, good $j$ has nonzero cost (for the purposes of Walrasian equilibrium) if and only if $q_j = 0$. \subsection{Proof of Theorem~\ref{thm:main}} The proof of Theorem~\ref{thm:main} is divided into three parts. The first part involves setting up the KKT conditions for Programs~\ref{pro:ces} and \ref{pro:demand}. The second assumes that $\mathbf{x} \in \Psi(\rho)$ and proves that $(\mathbf{x}, p)$ is a WE, and the third assumes that $(\mathbf{x}, p)$ is a WE and proves that $\mathbf{x} \in \Psi(\rho)$. \begin{proof}[Proof of Theorem~\ref{thm:main}] \textbf{Part 1: Setup.} Let $\q$ denote the vector $(q_1,\dots,q_m) \in \bbrpos^m$; then the Lagrangian of Program~\ref{pro:ces} is $L(\mathbf{x}, \q) = \frac{1}{\rho} \sum_{i \in N} v_i(x_i)^\rho - \sum_{j \in M} q_j (\sum_{i \in N} x_{ij} - 1)$.\footnote{The expert reader may notice that we have omitted the $\mathbf{x} \in \bbrpos^{m\times n}$ constraint from the Lagrangian. We do this to slightly simplify the analysis. The effect on the KKT conditions is that stationarity changes from ``For all $i,j$, $\frac{\partial L(\mathbf{x}, \q)}{\partial x_{ij}} = 0$" to ``For all $i,j$, $\frac{\partial L(\mathbf{x}, \q)}{\partial x_{ij}} \le 0$, and the inequality holds with equality when $x_{ij} > 0$".} Since Program~\ref{pro:ces} is convex and satisfies strong duality by Slater's condition, the KKT conditions are both necessary and sufficient for optimality. That is, $\mathbf{x}$ is optimal for Program~\ref{pro:ces} (which is equivalent to $\mathbf{x} \in \Psi(\rho)$) if and only if there exist Lagrange multipliers $\q \in \bbrpos^m$ such that both of the following hold:\footnote{The KKT conditions also include primal feasibility and dual feasibility. Since we will only work with valid allocations $\mathbf{x}$ and nonnegative $q_1,\dots,q_m$, these two conditions are trivially satisfied.} \begin{enumerate} \item Stationarity: $\mfrac{\partial L(\mathbf{x}, \q)}{\partial x_{ij}} \le 0$ for all $i,j$. Furthermore, if $x_{ij} > 0$, the inequality holds with equality. \item Complementary slackness: for all $j \in M$, either $\sum_{i \in N} x_{ij} = 1$, or $q_j = 0$. \end{enumerate} For a given $(i,j)$ pair, $\mfrac{\partial L(\mathbf{x}, \q)}{\partial x_{ij}}$ is equal to $v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}} - q_j$, so stationarity for Program~\ref{pro:ces} is equivalent to: $q_j \ge v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$ for all $i,j$, and when $x_{ij} > 0$, the inequality holds with equality. Next consider Program~\ref{pro:demand}, which defines each agent's demand set. This program has no constraints (other than $x_i \in \bbrpos^m$), so we can ignore complementary slackness. Thus by the KKT conditions, $x_i \in D_i(p)$ if and only if for every $j \in M$, $\mfrac{\partial v_i(x_i)}{\partial x_{ij}} \le \mfrac{\partial p(x_i)}{\partial x_{ij}}$, and if $x_{ij} > 0$, the inequality holds with equality (stationarity). We can explicitly compute the partial derivatives of $p$: $\mfrac{\partial p(x_i)}{\partial x_{ij}} = r^{\frac{\rho -1}{\rho}} q_j \big(\sum_{\ell \in M} q_\ell x_{i\ell}\big)^{\frac{1-\rho}{\rho}}$. \textbf{Part 2: Optimal CES welfare implies WE.} Suppose that $\mathbf{x} \in \Psi(\rho)$. Then there exists $\q \in \bbrpos^m$ such that $q_j \ge v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$ for all $j$, and $q_j = v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$ whenever $x_{ij} > 0$. Using the latter in combination with Euler's Theorem for homogeneous functions, for each $(i,j)$ pair we have \begin{align*} \frac{\partial p(x_i)}{\partial x_{ij}} =&\ r^{\frac{\rho -1}{\rho}} q_j \Big(\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{\frac{1-\rho}{\rho}}\\ =&\ q_j \Big(r^{-1} \sum_{\ell: x_{i\ell} > 0} q_\ell x_{i\ell}\Big)^{\frac{1-\rho}{\rho}}\\ =&\ q_j \Big(r^{-1} \sum_{\ell: x_{i\ell} > 0} v_i(x_i)^{\rho - 1} \frac{\partial v_i(x_i)}{\partial x_{i\ell}} x_{i\ell}\Big)^{\frac{1-\rho}{\rho}} &&\ \text{(stationarity for $x_{i\ell}$ when $x_{i\ell} > 0$)}\\ =&\ q_j \Big(r^{-1} v_i(x_i)^{\rho - 1} \sum_{\ell \in M} \frac{\partial v_i(x_i)}{\partial x_{i\ell}} x_{i\ell}\Big)^{\frac{1-\rho}{\rho}}\\ =&\ q_j \big(r^{-1} v_i(x_i)^{\rho - 1} r v_i(x_i)\big)^{\frac{1-\rho}{\rho}} \quad\quad &&\ \text{(Euler's Theorem)}\\ =&\ q_j v_i(x_i)^{1-\rho} \end{align*} Thus $\mfrac{\partial p(x_i)}{\partial x_{ij}} = q_j v_i(x_i)^{1-\rho}$. Next, we claim that $x_i \in D_i(p)$ for all $i \in N$. Fix an agent $i$; we show by case analysis that $x_i$ satisfies stationarity (for Program~\ref{pro:demand}) for each $j \in M$. Case 1: $q_j = v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$. Then $\mfrac{\partial p(x_i)}{\partial x_{ij}} = q_j v_i(x_i)^{1-\rho} = v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}} v_i(x_i)^{1-\rho} = \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$, and we are done. Case 2: $x_{ij} = 0$ and $q_j \ge v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$. Then similarly, $\mfrac{\partial p(x_i)}{\partial x_{ij}} = q_j v_i(x_i)^{1-\rho} \ge v_i(x_i)^{\rho - 1} \mfrac{\partial v_i(x_i)}{\partial x_{ij}} v_i(x_i)^{1-\rho} = \mfrac{\partial v_i(x_i)}{\partial x_{ij}}$, and again we are done. Therefore $x_i \in D_i(p)$ for all $i \in N$. Since $\mathbf{x}$ is a valid allocation, $\sum_{i \in N} x_{ij} \le 1$ for all $j \in M$. This, combined with complementary slackness for Program~\ref{pro:ces}, is identical to the market clearing condition for Walrasian equilibrium. Thus we have shown that if $\mathbf{x} \in \Psi(\rho)$, there exist $q_1,\dots, q_m$ (which are optimal Lagrange multipliers for Program~\ref{pro:ces}) such that for pricing rule $p$ as defined, $(\mathbf{x}, p)$ is a WE. \textbf{Part 3: WE implies optimal CES welfare.} This is similar to Part 2. Suppose there exists $\q \in \bbrpos^m$ such that for pricing rule $p(x_i) = \rho r^{\frac{\rho -1}{\rho}} (\sum_{j \in M} q_j x_{ij})^{1/\rho}$, $(\mathbf{x}, p)$ is a WE. Recall the partial derivatives of $p$: $\mfrac{\partial p(x_i)}{\partial x_{ij}} = q_j \big(r^{-1} \sum_{\ell \in M} q_\ell x_{i\ell}\big)^{\frac{1-\rho}{\rho}}$. We multiply each side by $x_{ij} r^{-1}$, and sum both sides over $j$: \begin{align*} r^{-1} \sum_{j \in M} x_{ij} \frac{\partial p(x_i)}{\partial x_{ij}} =&\ \Big(r^{-1}\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{\frac{1-\rho}{\rho}} r^{-1} \sum_{j \in M} q_j x_{ij}\\ r^{-1}\sum_{j: x_{ij} > 0} x_{ij} \frac{\partial p(x_i)}{\partial x_{ij}} =&\ \Big(r^{-1}\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{\frac{1-\rho}{\rho} + 1} \end{align*} Since $(\mathbf{x}, p)$ is a WE, we have $x_i \in D_i(p)$ for all $i \in N$. Thus $\mfrac{\partial v_i(x_i)}{\partial x_{ij}} = \mfrac{\partial p(x_i)}{\partial x_{ij}}$ whenever $x_{ij} > 0$, so \begin{align*} r^{-1}\sum_{j: x_{ij} > 0} x_{ij} \frac{\partial v_i(x_i)}{\partial x_{ij}} =&\ \Big(r^{-1}\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{1/\rho}\\ r^{-1}\sum_{j \in M} x_{ij} \frac{\partial v_i(x_i)}{\partial x_{ij}} =&\ \Big(r^{-1}\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{1/\rho}\\ v_i(x_i) =&\ \Big(r^{-1}\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{1/\rho} \quad\quad \text{(Euler's Theorem)}\\ v_i(x_i)^{\rho-1} =&\ \Big(r^{-1}\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{\frac{\rho-1}{\rho}} \end{align*} Using this in combination with $\mfrac{\partial p(x_i)}{\partial x_{ij}} = q_j \big(r^{-1} \sum_{\ell \in M} q_\ell x_{i\ell}\big)^{\frac{1-\rho}{\rho}}$, we get \[ q_j = \frac{\partial p(x_i)}{\partial x_{ij}}\Big(r^{-1}\sum_{\ell \in M} q_\ell x_{i\ell}\Big)^{\frac{\rho-1}{\rho}} = \frac{\partial p(x_i)}{\partial x_{ij}} v_i(x_i)^{\rho - 1} \] Next, we claim that $(\mathbf{x}, \q)$ satisfies stationarity for Program~\ref{pro:ces}. We proceed by case analysis for each $(i,j)$ pair. Stationarity for Program~\ref{pro:demand} implies that these are the only two possible cases. Case 1: $\mfrac{\partial v_i(x_i)}{\partial x_{ij}} = \mfrac{\partial p(x_i)}{\partial x_{ij}}$. In this case, $q_j = \mfrac{\partial v_i(x_i)}{\partial x_{ij}} v_i(x_i)^{\rho - 1}$, and we are done. Case 2: $x_{ij} = 0$ and $\mfrac{\partial v_i(x_i)}{\partial x_{ij}} \le \mfrac{\partial p(x_i)}{\partial x_{ij}}$. In this case we have $q_j \ge \mfrac{\partial v_i(x_i)}{\partial x_{ij}} v_i(x_i)^{\rho - 1}$, and we are again done. Thus $(\mathbf{x}, \q)$ satisfies stationarity for Program~\ref{pro:ces}. Furthermore, the second condition of Walrasian equilibrium is again identical to the complementary slackness condition. We conclude that $\mathbf{x} \in \Psi(\rho)$, and that $q_1,\dots, q_m$ are optimal Lagrange multipliers for Program~\ref{pro:ces}. This completes the proof. \end{proof} The following corollary states that under this pricing rule, each agent's resulting value will be proportional to the her payment. This property will be helpful in future sections, and may also be interesting independently. \begin{restatable}{corollary}{corUtilityPrice}\label{cor:utility-price} Assume each $v_i$ is homogeneous of degree $r$, concave, and differentiable, and let $p(x_i) = (\sum_{j \in M} q_j x_{ij})^{1/\rho}$ for some $\q\in\bbrpos^m$. Then if $x_i \in D_i(p)$, $p(x_i) = \rho r v_i(x_i)$. \end{restatable} \begin{proof} As before, stationarity for Program~\ref{pro:demand} gives us $\mfrac{\partial v_i(x_i)}{\partial x_{ij}} = \mfrac{\partial p(x_i)}{\partial x_{ij}}$ whenever $x_{ij} > 0$. Also note that by definition, $p$ is homogeneous of degree $1/\rho$. Using these two properties in combination with Euler's Theorem, we get \begin{align*} \frac{\partial v_i(x_i)}{\partial x_{ij}} =&\ \frac{\partial p(x_i)}{\partial x_{ij}} \quad \text{for all $j \in M$ s.t. $x_{ij} > 0$} \\ \sum_{j \in M} x_{ij} \frac{\partial v_i(x_i)}{\partial x_{ij}} =&\ \sum_{j \in M} x_{ij} \frac{\partial p(x_i)}{\partial x_{ij}}\\ r v_i(x_i) =&\ \frac{1}{\rho} p(x_i) \end{align*} Multiplying both sides by $1/\rho$ completes the proof. \end{proof} \section*{Acknowledgements} This research was supported in part by NSF grant CCF-1637418, ONR grant N00014-15-1-2786, and the NSF Graduate Research Fellowship under grant DGE-1656518. \end{document} \section{Model}\label{sec:model} Let $N = \{1,2,\ldots n\}$ be a set of agents, and let $M = \{1,2,\dots m\}$ be a set of divisible goods. Throughout the paper, we use $i$ and $k$ to refer to agents and $j$ and $\ell$ to refer to goods. We need to determine an \emph{allocation} $\mathbf{x} \in \bbr_{\ge 0}^{n\times m}$, where $x_i \in \bbr_{\ge 0}^m$ is the \emph{bundle} of agent $i$, and $x_{ij} \in [0,1]$\footnote{Without loss of generality, we can normalize the supply of each good to be 1.} is the quantity of good $j$ allocated to agent $i$. An allocation cannot allocate more than the available supply: $\mathbf{x}$ is a valid allocation if and only if $\sum_i x_{ij} \leq 1$ for all $j$. We will also determine \emph{payments} $p_1,\dots,p_n$, where $p_i$ is the payment charged to agent $i$. Thus an outcome is an allocation $\mathbf{x}$ and payments $p_1,\dots,p_n$. Agent $i$'s utility for a bundle $x_i$ is denoted by $u_i(x_i) \in \bbr$. We assume that agents have quasilinear utility: for an allocation $\mathbf{x}$ and payment $p_i$, agent $i$'s utility is $u_i(x_i) = v_i(x_i) - p_i$, where $v_i$ is agent $i$'s \emph{valuation}. When each agent's payment only depends on the bundle she receives, i.e., $p_i = p(x_i)$, we call $p$ a \emph{pricing rule}. With the exception of Section~\ref{sec:truthful}, we will focus on pricing rules. For $v_i$, we make the following standard assumptions throughout the paper: \begin{enumerate} \item Nonzero: There exists a bundle $x_i$ such that $v_i(x_i) > 0$. \item Montone: If $x_{ij} \ge y_{ij}$ for all $j \in M$, then $v_i(x_i) \ge v_i(y_i)$. \item Normalized: $v_i(0,\dots, 0) = 0$. \end{enumerate} Our positive results require the following three additional properties, which we will mention explicitly whenever used: \begin{enumerate} \setcounter{enumi}{3} \item Concave: For any bundles $x_i, y_i$ and constant $\lambda \in [0,1]$, we have $v_i(\lambda x_i + (1 - \lambda) y_i) \geq \lambda v_i( x_i) + (1-\lambda)u_i(y_i)$. \item Homogeneous of degree $r$: for any bundle $x_i$ and constant $\lambda \ge 0$, $v_i(\lambda x_i) = \lambda^r v_i(x_i)$. For $0 < r < 1$, this models diminishing returns. Note that homogeneity implies normalization, and for monotone and concave $v_i$, we must have $r \ge 0$ and $r \le 1$ respectively. \item Differentiable: for any bundle $x_i$ and all $j \in M$, $\mfrac{\partial v_i(x_i)}{\partial x_{ij}}$ is defined. \end{enumerate} \paragraph{CES welfare.} For multipliers $\mathbf{a} = (a_1, a_2\dots a_n) \in \bbrpos^n$ and $\rho \in (-\infty, 0)\cup(0,1]$, the (weighted) CES welfare of an allocation $\mathbf{x}$ is $\Phi_\mathbf{a}(\rho, \mathbf{x}) = \big(\sum_{i \in N} a_i v_i(x_i)^{\rho}\big)^{1/\rho}$. For $\rho \ne 1$, $\Phi$ is strictly concave in $v_i(x_i)$ for all $i \in N$, so every optimal allocation $\mathbf{x}$ has the same valuation vector $v_1(x_1),\dots, v_n(x_n)$. We will use $\Psi_\mathbf{a}(\rho)$ to denote CES welfare maximization, i.e., $\Psi_\mathbf{a}(\rho) = \argmax_{\mathbf{x} \in \bbrpos^m:\ \sum_i x_{ij} \le 1\ \forall j} \Phi_\mathbf{a}(\rho, \mathbf{x})$. There may be multiple optimal allocations (for example, if there is a good which no one values), so $\Psi_\mathbf{a}(\rho)$ denotes a set. Thus $\mathbf{x} \in \Psi_\mathbf{a}(\rho)$ denotes that $\mathbf{x}$ has maximum CES welfare. When each agent has the same multiplier (other than Appendix~\ref{sec:fisher}, this will always be the case), we simply write $\Phi(\rho, \mathbf{x})$ and $\Psi(\rho)$. As an illustrative example, consider a single good and valuations that are homogeneous of degree 1. Utilitarian welfare results in the good being entirely allocated to agents with $w_i = \max_k w_k$, with other agents receiving nothing (see Figure~\ref{fig:linear}). In contrast, for $\rho < 1$, the unique allocation maximum CES welfare welfare gives the following bundle $x_i \in \bbrspos$ to each agent $i$ (Lemma~\ref{lem:m=1-x}): $x_i = \mfrac{{w_i}^{\frac{\rho}{1-\rho}}}{\sum_{k} {w_k}^{\frac{\rho}{1-\rho}}}$. One natural case is $\rho = 1/2$, which results in a proportional allocation. \paragraph{Axiomatic characterization.} CES welfare functions also admit an axiomatic characterization. Consider the following axioms: (1) Monotonicity: if one agent's valuation increases while all others are unchanged, the welfare function should prefer the new allocation, (2) Anonymity: the welfare function should treat all agents the same, (3) Continuity: the welfare function should be continuous.\footnote{A slightly weaker version of continuity is often used: if an allocation $\mathbf{x}$ is strictly preferred to an allocation $\mathbf{y}$, there should be neighborhoods $N(\mathbf{x})$ and $N(\mathbf{y})$ such that every $\xprime \in N(\mathbf{x})$ is preferred to every $\yprime \in N(\mathbf{y})$. This weaker version only requires a welfare \emph{ordering} and does not require that this ordering be expressed by a function. However, any such ordering which also satisfies the rest of our axioms is indeed representable by a welfare function~\cite{Debreu1960}, and so both sets of axioms end up specifying the same set of welfare functions/orderings.}, (4) Independence of common scale: scaling all agent valuations by the same factor should not affect which allocations have better welfare than others, (5) Independence of unconcerned agents: when comparing the welfare of two allocations, the comparison should not depend on agents who have the same valuation in both allocations, and (6) The Pigou-Dalton principle: when choosing between equally efficient allocations, the welfare function should prefer more equitable allocations~\cite{Dalton1920,Pigou1912}. Disregarding monotonic transformations of the welfare function (which of course do not affect which allocations have better welfare than others), the set of welfare functions satisfying these axioms is exactly the set of CES welfare functions with $\rho \in (-\infty,0)\cup(0,1]$, including Nash welfare~\cite{Moulin2003}.\footnote{This actually does not include max-min welfare, which satisfies weak monotonicity but not strict monotonicity.} This axiomatic characterization shows that we are not just focusing on an arbitrary class of welfare functions: CES welfare functions are arguably the most reasonable welfare functions. \paragraph{Walrasian equilibrium.} Given a pricing rule $p$, agent $i$'s \emph{demand set} is defined by $D_i(p) = \argmax_{x_i \in \bbrpos^m} u_i(x_i)$, or equivalently, $D_i(p) = \argmax_{x_i \in \bbrpos^m} \big(v_i(x_i) - p(x_i)\big)$. Given an allocation $\mathbf{x}$ and payment rule $p$, $(\mathbf{x}, p)$ is a\emph{Walrasian equilibrium} (WE) if both of the following hold: \begin{enumerate} \item Each agent receives a bundle in her demand set: $x_i \in D_i(p)$ for all $i \in N$. \item The market clears: for all $j \in M$, $\sum_{i \in N} x_{ij} \le 1$, and for all $j \in M$ with nonzero cost, $\sum_{i \in N} x_{ij} = 1$.\footnote{We say that good $j$ has nonzero cost for $j$ if there is a bundle $x_i$ such that $x_{i\ell} = 0$ for all $\ell \ne j$, but $p(x_i) > 0$.} \end{enumerate} \subsection{Related work}\label{sec:prior} The study of markets has a long history in economics~\cite{Arrow1954,Brainard2005,Fisher1892,Varian1974,Walras1874}. Recently, this topic has received substantial attention in the computer science community as well (see~\cite{Vazirani2007} for an algorithmic introduction). We first provide some important background on different market models and the First and Second Welfare Theorems, and then move on to more recent related work. \paragraph{Quasilinear markets and Fisher markets.} There are two primary market models for divisible goods. This paper considers the quasilinear utility model, where each agent can spend as much as she wants, and the amount spent is incorporated into her utility function. The other predominant model is the \emph{Fisher market} model~\cite{Brainard2005,Fisher1892}, where each agent has a fixed budget constraint, and the amount spent does not affect her resulting utility (as a result, each agent always spends exactly her budget). Although these two models share many of the same conceptual messages, some of the technical results vary. See Appendix~\ref{sec:fisher} for the technical relationship between the two models with respect to WE and CES welfare maximization. Since agents in Fisher markets always spend exactly their budgets, there is no way to elicit the absolute scale of agent valuations. Nash welfare maximization is invariant to this type of scaling, but no other CES welfare function is~\cite{Moulin2003}. For this reason, the Fisher market model is not well suited to reason about other welfare functions. In contrast, the quasilinear model \emph{does} allow agents to express the absolute scale of their valuation: specifically, by choosing how much to spend. That is one reason that we focus on the quasilinear model for this paper. The other is that convex pricing is most easily applied to a small submarket of the broader economy (e.g., water pricing), and quasilinear utility captures the fact that agents may wish to spend money on other goods outside of this submarket. In contrast, Arrow and Debreu's model (see below) can arguably capture the entire economy, so there is nothing outside of the market to spend money on. \paragraph{The First and Second Welfare Theorems.} Conceptually, the First Welfare Theorem establishes an efficiency property that any WE must satisfy, and the Second Welfare Theorem deals with implementing a wide range of allocations as WE. The two welfare theorems originate in the context of \emph{Arrow-Debreu} markets~\cite{Arrow1954}, which generalize Fisher markets to allow for (1) agents to enter the market with goods (as opposed to just money)\footnote{These are known as ``exchange markets" or ``exchange economies".} and (2) production of goods. The statements of the First and Second Welfare Theorems in that model are, respectively, ``Any (linear pricing) WE is Pareto optimal" and ``Any Pareto optimal allocation can be a (linear pricing) WE with \emph{transfers}, i.e., under a suitable redistribution of initial wealth". In the Fisher market and quasilinear utility models, the First Welfare Theorem can be strengthened to ``Any (linear pricing) WE maximizes budget-weighted Nash welfare"~\cite{Eisenberg1961,Eisenberg1959,Vazirani2007} and ``Any (linear pricing) WE maximizes utilitarian welfare", respectively. The version of the Second Welfare Theorem stated above is appropriate for Fisher markets, since agents' budgets constitute the ``initial wealth". However, for quasilinear utilities, there is no notion of initial wealth (alternatively, initial wealth is an additive constant in agents' utilities which does not affect their behavior). Thus for quasilinear utilities, allowing transfers actually does not affect the set of WE. This may seem counterintuitive, since the Second Welfare Theorem (which still holds in this setting) states that any Pareto optimum can be a WE. However, Pareto optimality here is referring to agents' overall quasilinear utilities, \emph{not} the agents' valuations. It can be shown that the only allocations which are Pareto optimal with respect to the quasilinear utilities are allocations maximizing utilitarian welfare, which are already covered by the First Welfare Theorem (without transfers). Thus on a technical level, the Second Welfare Theorem is not helpful in the world of quasilinear utilities. However, even when the Second Welfare Theorem is mathematically relevant, a centrally mandated redistribution of wealth is often out of the question in practice. \paragraph{The equality/efficiency tradeoff in healthcare.} In Section~\ref{sec:ces-intro}, we discussed how CES welfare has been studied from a theoretical perspective in healthcare~\cite{Dolan1998,Ortun1996,Wagstaff1991}. There have also been several empirical studies aiming to understand the general population's view of the equality/efficiency tradeoff, with results generally indicating a disapproval of purely utilitarian approaches to healthcare~\cite{Dolan2000,Wittrup-Jensen2008}. For example, a survey of 449 Swedish politicians found widespread rejection of purely utilitarian decision-making in healthcare, and under some conditions, the respondents were willing to sacrifice up to 15 of 100 preventable deaths in order to ensure equality across subgroups~\cite{Lindholm1998}. \paragraph{CES welfare and $\alpha$-fairness in networking.} CES welfare functions have also enjoyed considerable attention from the field of networking, under the name of \emph{$\alpha$-fairness} (the parameter $\alpha$ corresponds to $1-\rho$ in our definition). The $\alpha$-fairness notion was proposed by~\cite{Mo2000}, motivated in part as a generalization of the prominent \emph{proportional fairness} objective (which is equivalent to Nash welfare)~\cite{Kelly1998}. See~\cite{Bertsimas2012} and references therein for further background on $\alpha$-fairness in networking. To our knowledge, a market-based understanding was developed only for proportional fairness, starting with the seminal work of Kelly et al.~\cite{Kelly1998}. \paragraph{Nonlinear market mechanisms and CES welfare maximization.} We are aware of just two papers studying market mechanisms for CES welfare functions: \cite{Goel2019nash} and \cite{Plaut2019}. Like our work, both of these papers explore nonlinear pricing rules, but unlike our work, only consider Leontief valuations. Furthermore, both of those papers are in the Fisher market model and only achieve CES welfare maximization under strong assumptions on the absolute scale of the agents' valuations.\footnote{In particular, that each agent's weight for each good is either 0 or 1. This subclass of Leontief valuations is known as ``bandwidth allocation" valuations, where each good is a link in a network, and agents transmit data over fixed paths.} In contrast, our main result holds for any valuations that are homogeneous of degree $r$, differentiable, and concave, a much larger range of valuations. (Leontief valuations are not differentiable, but we handle them as a special case in Appendix~\ref{sec:leontief} and show that our result still holds.) It is worth noting that~\cite{Goel2019nash} focuses on the WE model, whereas~\cite{Plaut2019} considers strategic agents and Nash equilibria. On a related note, we are not aware of any broader results regarding general nonlinear pricing, i.e., what set of allocations can be implemented if we allow $p(x_i)$ to be any nondecreasing function of $x_i$ (but still require anonymity)? This could be an interesting direction for future work. The rest of the paper is organized as follows. Section~\ref{sec:model} formally defines the model. In Section~\ref{sec:main}, we present our main result: a simple convex pricing rule implements CES welfare maximization in WE for $\rho \in (0,1]$ (Theorem~\ref{thm:main}). Section~\ref{sec:comp} presents an iterative algorithm for computing these WE. In Section~\ref{sec:truthful}, we consider truthful mechanisms for CES welfare maximization. Section~\ref{sec:sybil} discusses Sybil attacks, and Section~\ref{sec:counter} presents our negative results. At this point we conclude the main paper. Appendix~\ref{sec:fisher} discusses connections to Fisher markets, Appendix~\ref{sec:leontief} shows that our main result extends to Leontief valuations, Appendix~\ref{sec:linear} discusses the First Welfare Theorem in more detail, and Appendix~\ref{sec:proofs} provides some proofs omitted from the main paper. \section{Results and related work}\label{sec:results} \paragraph{Main result: convex pricing implements CES welfare maximization in Walrasian equilibrium.} Our main result is that for convex pricing of the form $p(x_i) = (\sum_j q_j x_{ij})^{1/\rho}$ for any $\rho \in (0,1]$\footnote{The case of $\rho < 0$ is slightly unintuitive, as it can result in agents who care more receiving \emph{less} of the good. Consequently, implementation in WE is impossible; see Theorem~\ref{thm:neg-rho-counter}.}, a Walrasian equilibrium is guaranteed to exist, and every WE maximizes CES welfare with respect to $\rho$. This holds for a wide range of agent valuations. \begin{theoremNumbered}[\ref{thm:main} \textnormal{(Simplified version)}] Assume each valuation is homogeneous of degree $r$,\footnote{A valuation is homogeneous of degree $r$ if scaling any bundle by a constant $c$ scales the resulting value by $c^r$.} differentiable, and concave, and fix $\rho \in (0,1]$. Then an allocation $\mathbf{x} = (x_1,\dots, x_n)$ maximizes CES welfare if and only if there exist constants $q_1,\dots,q_m \in \bbrpos$ such that for the pricing rule \[ p(x_i) = \Big(\sum_{j} q_j x_{ij}\Big)^{1/\rho}, \] $\mathbf{x}$ and $p$ form a WE. \end{theoremNumbered} \noindent Note that the $\rho$ in $p(x_i)$ is the same $\rho$ for which CES welfare is maximized. We call the reader's attention to two important aspects of this result. Perhaps most importantly, our result is not simply a reformulation of the First Welfare Theorem: although maximizing CES welfare for valuations $v_1,\dots,v_n$ is equivalent to maximizing utilitarian welfare for valuations $v_1^\rho,\dots,v_n^\rho$, the First Welfare Theorem does not say anything about the agent demands in response to this convex pricing rule. The First Welfare Theorem also does not help with identifying the exact conditions under which Theorem~\ref{thm:main} holds, e.g., homogeneity of valuations.\footnote{In fact, not only is homogeneity necessary, but homogeneity of the same degree is necessary: if we allow the degree of homogeneity to differ across agents, the result no longer holds (Theorem~\ref{thm:different-r}).} Secondly, the class of homogeneous, differentiable, and concave valuations is quite large: it generalizes most of the commonly studied valuations, e.g., linear, Cobb-Douglas, and CES (note that here we are referring to CES agent valuations, not CES welfare functions). Although Leontief valuations are not differentiable, we handle them as a special case and show that the same result holds (Theorem~\ref{thm:leontief}). The following additional properties are of note: \begin{enumerate} \item For this class of utilities, Theorem~\ref{thm:main} generalizes the First Welfare Theorem:\footnote{One direction of the First Welfare Theorem (if $(\mathbf{x}, p)$ is a linear pricing WE, then $\mathbf{x}$ maximizes utilitarian welfare) holds in a much more general setting; see Appendix~\ref{sec:linear}.} when $\rho = 1$, $p(x_i)$ yields linear pricing and CES welfare yields utilitarian welfare. \item The constants $q_1,\dots,q_m$ will be the optimal Lagrange multipliers for a convex program maximizing CES welfare. This connection to duality will be very helpful for computing these WE (see Section~\ref{sec:comp}). \item Our pricing rule is strictly convex for $\rho < 1$, with the curvature growing as $\rho$ goes to 0. The smaller $\rho$ gets, the easier it is to buy a small amount, but the harder it is to buy a large amount. Intuitively, this should prevent any single individual from dominating the market and lead to a more equitable outcome. Furthermore, the marginal price at $x_i = \zero$ is zero, which ensures that everyone ends up with a nonempty bundle (in contrast to linear pricing: see Figure~\ref{fig:linear}). Theorem~\ref{thm:main} provides a tight relationship between the curvature of the pricing rule and the exact equality/efficiency tradeoff. \end{enumerate} \paragraph{Towards an implementation.} We also prove several supporting results: in particular, regarding implementation. The WE from Theorem~\ref{thm:main} can always be computed by asking each agent for her entire utility function, and then solving a convex program for maximizing CES welfare maximization to obtain the optimal Lagrange multipliers $q_1,\dots,q_m$. However, this is not very practical: people are generally not able to articulate a full cardinal utility function, and even if they are, doing so could require transmitting an enormous amount of information. Section~\ref{sec:comp} presents our first supporting result: an iterative algorithm for computing the WE, where in each step, each agent only needs to report the gradient of her valuation at the current point. Our algorithm is based on the ellipsoid method, and inherits its polynomial-time convergence properties. We recognize that even valuation gradient queries may be difficult for agents to answer, and we leave the possibility of an improved implementation -- in particular, a \emph{\tat}\footnote{A \tat is an iterative algorithm which only asks \emph{demand queries}, i.e., what would each agent purchase given the current prices. Demand queries may be easier to answer than the valuation gradient queries in our algorithm.} -- as an open question. \paragraph{Truthfulness.} Our second supporting result considers a different approach to implementation: \emph{truthful} mechanisms. Walrasian equilibria are generally not truthful: agents can lie about their preferences to affect the equilibrium prices for their personal gain.\footnote{Another interpretation is that WE assumes agents are \emph{price-taking} (i.e., treat the prices are given and do not lie about their preferences to affect the equilibrium prices) and breaks down when agents are \emph{price-anticipating}.} For $\rho = 1$, the Vickrey-Clarke-Groves (VCG) mechanism is known to truthfully maximize utilitarian welfare~\cite{Nisan2007}. For the case of a single good and any $\rho \in (0,1)$, we give a mechanism which truthfully maximizes CES welfare (Theorem~\ref{thm:truthful}). We also show that our mechanism is the unique truthful mechanism up to an additive constant in the payment rule (Theorem~\ref{thm:truthful-unique}). The proof of Theorem~\ref{thm:truthful-unique} is quite involved, and requires techniques from real analysis such as Kirszbraun's Theorem for Lipschitz extensions and the Fundamental Theorem of Lebesgue Calculus. \paragraph{Negative results.} We prove the following negative results. Most importantly, we show that for any $\rho \ne 1$, linear-pricing WE can have arbitrarily poor CES welfare (Theorem~\ref{thm:rho=1-bad}); were this not the case, perhaps it would suffice to simply use linear pricing and accept an approximation of CES welfare. Next, note that Theorem~\ref{thm:main} requires each agent's valuation to be homogeneous with the same degree $r$. We show that when agents' valuations have different homogeneity degrees, there exist instances where no pricing rule can implement CES welfare maximization in WE (Theorem~\ref{thm:different-r}), and thus our assumption is necessary. We also show that CES welfare maximization cannot be implemented in WE for $\rho < 0$ (Theorem~\ref{thm:neg-rho-counter}), and discuss the special case of $\rho = 0$ (i.e., Nash welfare). There is an additional crucial issue which any practical implementation of Theorem~\ref{thm:main} would need to address: \emph{Sybil attacks}. A Sybil attack is when a selfish agent attempts to gain an advantage in a system by creating fake identities~\cite{Douceur2002}. Since the pricing rule from Theorem~\ref{thm:main} is strictly convex for $\rho < 1$, an agent can decrease her payment by masquerading as multiple individuals and splitting her purchase across those identities.\footnote{In contrast, for $\rho = 1$, there is nothing to be gained by creating fake identities.} In Appendix~\ref{sec:sybil}, we propose a model for analyzing Sybil attacks in markets, and show that if these attacks are possible, there exist instances where no pricing rule can implement CES welfare maximization in WE (Theorem~\ref{thm:sybil-other-p}).\footnote{There are combinations of parameters, however, where our pricing rule is naturally robust to Sybil attacks: in particular, when $v_i(\mathbf{x}) (1-\rho) \le \kappa$ (where $v_i(\mathbf{x})$ is agent $i$'s value for the maximum CES welfare allocation and $\kappa$ is the identity creation cost). This suggests a natural way for an equality-focused social planner to choose a specific value for $\rho$: estimate the identity creation cost and scale of valuations in the system of interest, and pick $\rho$ to be as small as possible without incentivizing Sybil attacks.} \paragraph{Additional results.} In Appendix~\ref{sec:fisher}, we explore connections between our results in the quasilinear utility model, and the Fisher market fixed-budget model. Appendix~\ref{sec:leontief} shows that Theorem~\ref{thm:main} extends to Leontief valuations, which are not differentiable (so the main proof does not apply). Leontief valuations have been a focus of prior work, so we find is worthwhile to handle this as a special case. \section{Sybil attacks}\label{sec:sybil} In Sections~\ref{sec:comp} and \ref{sec:truthful}, we discussed two alternative approaches to implementation: an iterative query-based algorithm, and a truthful mechanism. However, there is an additional crucial issue which any practical implementation must address: since our pricing rule $p(x_i) = (\sum_{j \in M} q_j x_{ij})^{1/\rho}$ is strictly convex for $\rho < 1$, agents have an incentive to create fake identities. In particular, an agent can decrease her payment while receiving the same bundle by splitting the payment over multiple fake identities.\footnote{Note that for linear prices there is no such incentive.} This is known as a \emph{Sybil attack}. The truthful payment rule from Section~\ref{sec:truthful} is not strictly convex everywhere, but it is strictly convex on some intervals, and thus has the same vulnerability. \paragraph{Model of Sybil attacks.} We model this as follows. Let $\kappa$ denote the cost of creating a new identity. The cost could reflect inconvenience, risk of getting caught, or other factors, and would depend on the nature of the system. Let $\eta_i$ be the \emph{multiplicity} of agent $i$, i.e., the number of identities agent $i$ controls in the system. This includes both fake identities and the agent's single real identity, so we assume that $\eta_i \in \bbnspos$. For convex $p$, multiplicity $\eta_i$, and a desired bundle for purchase, it is always optimal for agent $i$ to split the purchase evenly across her identities.\footnote{This is essentially a multidimensional version of Jensen's inequality; see, e.g.,~\cite{Neuman1990}.} Thus we can assume that each identity purchases the same bundle $x_i$, and we define agent $i$'s utility as \[ u_i(x_i, \eta_i) = v_i(\eta_i x_i) - \eta_i p(x_i) - \eta_i \kappa \] We do not claim that this perfectly models the reality of Sybil attacks; for example, the identity creation cost is arguably sublinear (one someone has created a single fake identity, creating more might become easier). Our goal here is simply to show formally that at least in some cases, CES welfare maximization cannot be robust to Sybil attacks in general. \paragraph{Walrasian equilibrium.} We focus on a Walrasian model of Sybil attacks; the analogous analysis for truthful mechanisms is left as an open question. We define each agent's \emph{Sybil demand set} by \[ D_i(p) = \argmax_{x_i \in \bbrpos^m, \eta_i \in \bbnspos}\ u_i(x_i, \eta_i) \] Note that we require $\eta_i \in \bbnspos$. We define a \emph{Sybil Walrasian equilibrium} (SWE) to be an allocation $\mathbf{x}$, payment rule $p$, and vector of multiplicities $\bfeta = \eta_1,\dots,\eta_n$ such that \begin{enumerate} \item Each agent receive a bundle in her demand set: $(x_i, \eta_i) \in D_i(p)$ for all $i \in N$. \item The market clears: for all $j \in M$, $\sum_{i \in N} x_{ij} \le 1$. Furthermore, for any $j\in M$ with nonzero cost\footnote{Recall that good $j$ has ``nonzero cost" in our pricing rule if $q_j > 0$.}, $\sum_{i \in N} x_{ij} = 1$. \end{enumerate} In this section, we will focus on the case of homogeneity degree $r=1$. The following lemma states that for any pricing rule, a rational agent either creates no fake identities (i.e, $\eta_i = 1$), or creates an unbounded number (and consequently the demand set is empty). \begin{lemma}\label{lem:sybil-demand} Assume each $v_i$ is concave, differentiable, and homogeneous of degree 1. Let $\rho \in (0,1]$, define $p$ as in Theorem~\ref{thm:main}, and let $\mathbf{x} \in \Psi(\rho)$. Then we have \[ D_i(p) = \begin{cases} (x_i, 1) &\ \text{if } v_i(x_i) (1-\rho) \le \kappa\\ \emptyset &\ \text{otherwise} \end{cases} \] where $x_i$ is agent $i$'s bundle in $\mathbf{x}$. \end{lemma} \begin{proof} When $v_i$ is homogeneous of degree 1, for any bundle $y_i$, we have $u_i(y_i, \eta_i) = \eta_i v_i(y_i) - \eta_i p(y_i) - \eta_i \kappa = \eta_i\big(v_i(y_i) - p(y_i) - \kappa\big)$. Thus given a choice of $\eta_i$, $y_i$ must be chosen to maximize $v_i(y_i) - p(y_i) - \kappa$. Let $\mathbf{x} \in \Psi(\rho)$: then by Theorem~\ref{thm:main} $y_i$ optimizes $v_i(y_i) - p(y_i)$ (and thus $v_i(y_i) - p(y_i) - \kappa$) if and only if $y_i = x_i$. Therefore the demand set is equal to \[ D_i(p) = \Big(x_i,\ \argmax_{\eta_i \in \bbnspos}\ \eta_i \big(v_i(x_i) - p(x_i) - \kappa\big) \Big) \] That is, the demanded bundle must always be $x_i$, and $\eta_i$ is optimized accordingly. By Corollary~\ref{cor:utility-price}, $p(x_i) = \rho v_i(x_i)$, so $v_i(x_i) - p(x_i) - \kappa = v_i(x_i)(1-\rho) - \kappa$. Thus if $v_i(x_i) (1-\rho) \le \kappa$, then $1$ is an optimal choice for $\eta_i$, so $(x_i, 1) \in D_i(p)$. However, if $v_i(x_i) (1-\rho) > \kappa$, there is no optimal choice for $\eta_i$: specifically, $\eta_i$ goes to infinity. Thus if $v_i(x_i) (1-\rho) > \kappa$, $D_i(p) = \emptyset$. \end{proof} This immediately implies that if $\mathbf{x} \in \Psi(\rho)$ satisfies $v_i(x_i) (1-\rho) \le \kappa$ for all $i \in N$, the convex pricing rule from Theorem~\ref{thm:main} is naturally robust to Sybil attacks. \begin{theorem}\label{thm:sybil-r=1-good} Assume each $v_i$ is concave, differentiable, and homogeneous of degree 1. Let $\mathbf{x} \in \Psi(\rho)$ for $\rho \in (0,1]$, and define $p$ as in Theorem~\ref{thm:main}. Then if $v_i(x_i) (1-\rho) \le \kappa$ for all $i \in N$, $(\mathbf{x}, p, \one)$ is a SWE. \end{theorem} \begin{proof} By Lemma~\ref{lem:sybil-demand}, we have $(x_i, 1) \in D_i(p)$ for all $i \in N$ in this case. Theorem~\ref{thm:main} implies that the market clearing condition is met, so $(\mathbf{x}, p, \one)$ is a SWE. \end{proof} In other words, if the identity creation cost is small, $\rho$ is close to 1, and/or agents valuations are not too large, we need not worry about Sybil attacks. As discussed in Section~\ref{sec:results}, this suggests one possible way for a social planner to choose a value of $\rho$: estimate $\kappa$ and the magnitude of valuations, and choose $\rho$ to be as small as possible without incentivizing Sybil attacks. On the other hand, if $v_i(x_i) (1-\rho) > \kappa$, how bad are the consequences? Theorem~\ref{thm:sybil-r=1-bad} states that an agent's valuation at equilibrium has a hard cap at $\mfrac{\kappa}{1-\rho}$. This provides a hard maximum on the CES welfare in any SWE with $p$ thus defined: in particular, the CES welfare is at most $\big(\sum_{i \in N} \big(\frac{\kappa}{1-\rho}\big)^\rho\big)^{1/\rho} = n^{1/\rho} \frac{\kappa}{1-\rho}$. In general, each $v_i(x_i)$ (and thus the CES welfare) can be arbitrarily large, so Theorem~\ref{thm:sybil-r=1-bad} implies an unbounded ratio between the optimal CES welfare and that of any SWE with this $p$. \begin{theorem}\label{thm:sybil-r=1-bad} Assume each $v_i$ is concave, differentiable, and homogeneous of degree 1. Let $\rho \in (0,1]$, and define $p$ as in Theorem~\ref{thm:main}. Then for any allocation $\mathbf{x}$ and multiplicities $\bfeta$ such that $(\mathbf{x}, p, \bfeta)$ is a SWE, we have \[ v_i(x_i) \le \frac{\kappa}{1-\rho} \] \end{theorem} \begin{proof} Suppose $(\mathbf{x}, p, \bfeta)$ is a SWE for some allocation $\mathbf{x}$ and multiplicities $\bfeta$: then each $(x_i, \eta_i) \in D_i(p)$ for all $i \in N$; Thus $D_i(\mathbf{p}) \ne \emptyset$, so Lemma~\ref{lem:sybil-demand} implies that $v_i(x_i) (1-\rho) \le \kappa$, and consequently, $v_i(x_i) \le \frac{\kappa}{1-\rho}$. \end{proof} The next natural question is, can we circumvent this by using a different pricing rule? Theorem~\ref{thm:sybil-other-p} answers this in the negative. The counterexample uses an instance with a single good; recall that $x_i$ denotes a scalar in this case. \begin{theorem}\label{thm:sybil-other-p} Let $m=1$, $v_1(x_1) = w x_1$, and $v_i(x_i) = x_i$ for all $i \ne 1$. Let $(\mathbf{x}, p, \bfeta)$ be any SWE. Then for all $i \ne 1$, \[ v_i(x_i) \le \frac{\kappa}{w-1} \] \end{theorem} \begin{proof} Let $(\mathbf{x}, p, \bfeta)$ be any SWE. Fix an arbitrary $i \in N$. As in Lemma~\ref{lem:sybil-demand}, we have $u_i(x_i, \eta_i) = \eta_i (v_i(x_i) - p(x_i) - \kappa)$. Since $(x_i, \eta_i) \in D_i(p)$, we must have $\eta_i \in \argmax_{\eta_i' \in \bbnspos} \eta_i (v_i(x_i) - p(x_i) - \kappa)$ (note that we are not assuming anything about the bundle $x_i$). Since $\argmax_{\eta_i' \in \bbnspos} \eta_i' (v_i(x_i) - p(x_i) - \kappa)$ cannot be the empty set, we must have $v_i(x_i) \le p(x_i) + \kappa$ and $\eta_i = 1$. Focusing on agent 1, we further claim that $v_1(x_i) \le p(x_i) + \kappa$ for any $i \ne 1$. Suppose not: then agent 1 could purchase $x_i$ and set $\eta_1 = \infty$ to increase her utility. Thus $v_1(x_i) \le p(x_i) + \kappa$ for each $i \ne 1$. Now looking at the optimization for $i \ne 1$, we have $v_i(x_i) \ge p(x_i)$. Combining this with $v_1(x_i) \le p(x_i) + \kappa$, we get $v_1(x_i) \le v_i(x_i) + \kappa$. Plugging in our definitions of $v_1$ and $v_{i \ne 1}$, we get $w x_i \le x_i + \kappa$, so $x_i (w - 1) \le \kappa$. Substituting back in the definition of $v_i$, we get $v_i(x_i) \le \mfrac{\kappa}{w-1}$ for all $i \ne 1$, as required. \end{proof} Although the bound in Theorem~\ref{thm:sybil-other-p} is different from that in Theorem~\ref{thm:sybil-r=1-bad}, the implication is the same: this is a hard maximum on the value obtained by any agent other than agent 1. As $\kappa$ goes to zero, the fraction of the good agent 1 receives approaches 1, so the outcome approaches the maximum utilitarian welfare outcome (where agent 1 receives the entirety of the good). Therefore by Theorem~\ref{thm:rho=1-bad}, the CES welfare at any Sybil Walrasian equilibrium (for any pricing rule) can be arbitrarily bad in comparison to the optimal CES welfare. Thus in general, when Sybil attacks are possible, it is impossible to implement any bounded approximation of CES welfare maximization in Walrasian equilibrium. \section{Truthfulness}\label{sec:truthful} An alternative approach to implementation is via \emph{truthful} mechanisms. Walrasian equilibria are generally not truthful: agents can sometimes create more favorable equilibrium prices by lying about their preferences. In this section, we present a truthful mechanism for optimizing CES welfare in the case of a single good (Theorem~\ref{thm:truthful}), and show that it is unique up to additive constants in the payment rule (Theorem~\ref{thm:truthful-unique}). Note that uniqueness beyond additive constants in the payment rule can never be achieved without additional assumptions (e.g., individual rationality), since such constants do not affect the behavior of agents. Before formally stating and proving these results, we mention an important distinction between this section and Section~\ref{sec:comp}. Section~\ref{sec:comp} is an implementation of the WE from Theorem~\ref{thm:main} (which we know maximizes CES welfare). In contrast, the truthful mechanism from this section is an implementation of CES welfare maximization directly, not an implementation of the WE from Theorem~\ref{thm:main}. Indeed, we know that the payment rule from Theorem~\ref{thm:main} is not truthful, so we must consider a different payment rule if we desire truthfulness. To define our truthful mechanism we need the following two lemmas, whose proofs appear in Appendix~\ref{sec:proofs}. The first states that for a single good, homogeneous and differentiable functions take a very simple form. The second states that for a single good, the maximum CES welfare allocations take a very simple form; also, for $\rho \ne 1$, the optimum is unique. \begin{restatable}{lemma}{lemHomoOneGood}\label{lem:homo-m=1} Let $f: \bbrpos \to \bbrpos$ be differentiable and homogeneous of degree $r$. Then there exists $c \in \bbrpos$ such that $f(x) = c x^r$. \end{restatable} \begin{restatable}{lemma}{lemOptOneGood}\label{lem:m=1-x} Let $m=1$ and $v_i(x_i) = w_i x_i^r$ for all $i \in N$ where $r \in (0,1]$. Then $\rho \in (0,1]$ and $r\rho \ne 1$, $\mathbf{x} \in \Psi(\rho)$ if and only if \[ x_i = \frac{{w_i}^{\frac{\rho}{1-r\rho}}}{\sum_{k \in N} {w_k}^{\frac{\rho}{1-r\rho}}} \] If $\mathbf{x} \in \Psi(\rho)$ and $r = \rho = 1$, then whenever $x_i > 0$, $w_i = \max_{k \in N} w_k$. \end{restatable} We now define our mechanism. For $\rho = 1$, the VCG mechanism truthfully maximizes utilitarian welfare~\cite{Nisan2007}, so assume $\rho \in (0,1)$. We ask each agent $i$ to report $w_i$ (where $v_i(x_i) = w_i \cdot x_i^r$), assume the $w_i$'s are truthful, and output the (unique) optimal allocation $\mathbf{x} \in \Psi(\rho)$ according to Lemma~\ref{lem:m=1-x}. Let $\B = b_1,\dots,b_n$ be the vector of reported $w_i$'s. We then charge each agent $i$ the following payment:\footnote{Although this integral does not have a simple closed form, it can be expressed via the hypergeometric function.} \begin{align} p_i(\B) = \frac{r\rho}{1-r\rho} \Big(\sum_{k \ne i} b_k^\frac{\rho}{1-r\rho}\Big) \int_{b = 0}^{b_i} \frac{b^\frac{r\rho}{1-r\rho}}{\big(b^\frac{\rho}{1-r\rho} + \sum_{k \ne i} b_k^\frac{\rho}{1-r\rho}\big)^{r+1}}\dif b \label{eq:payment} \end{align} This payment is chosen so that the derivative of agent $i$'s utility at $b_i = w_i$ is 0. In particular, let $x_i(\B)$ denote agent $i$'s bundle under reports $\B$. Then we will have $\mfrac{\partial v_i(x_i(\B))}{\partial b_i} = rw_i\mfrac{\partial x_i(\B)}{\partial b_i} x_i(\B)^{r-1}$, and $\mfrac{\partial p_i(\B)}{\partial b_i} = rb_i\mfrac{\partial x_i(\B)}{\partial b_i} x_i(\B)^{r-1}$, so the derivative of agent $i$'s overall utility will be $\mfrac{\partial u_i(\B)}{\partial b_i} = r (w_i - b_i) \mfrac{\partial x_i(\B)}{\partial b_i} x_i(\B)^{r-1}$. This will imply that it is optimal for agent $i$ to truthfully report $b_i = w_i$. \begin{restatable}{theorem}{thmTruthful}\label{thm:truthful} Assume $m=1$, and that each $v_i$ is homogenous of degree $r$ (with $r$ publicly known), concave, and differentiable. Then for all $\rho \in (0,1)$, there is a truthful mechanism which outputs an allocation $\mathbf{x} \in \Psi(\rho)$. \end{restatable} \begin{proof} Since VCG satisfies the claim for $\rho = 1$, assume $\rho \in (0,1)$. Let $\mathbf{x}(\B)$ denote the allocation outputted given reports $\B$, and let $x_i(\B)$ denote agent $i$'s bundle: formally, $x_i(\B) = \mfrac{{b_i}^{\frac{\rho}{1-r\rho}}}{\sum_{k \in N} {b_k}^{\frac{\rho}{1-r\rho}}}$. Since $m=1$, Lemma~\ref{lem:homo-m=1} implies that for all $i \in N$, there exists $w_i \in \bbrpos$ such that $v_i(x) = w_i\cdot x^r$ for all $x \in \bbrpos$. Then by Lemma~\ref{lem:m=1-x}, $x_i(\B) \in \Psi(\rho)$, so it remains to prove truthfulness. Since we assume that each agent's valuation is not identically zero, we have $w_i > 0$. Also, by concavity and monotonicity of $v_i$, we have $r \in (0,1]$. Thus $0 < r\rho < 1$. Since we also have $b_i > 0$, all denominators in $p_i(\B)$ are nonzero and thus $p_i(\B)$ is well-defined. Let $v_i(\B) = v_i(x_i(\B)) = w_i x_i(\B)^r$ for brevity, and let $u_i(\B) = v_i(\B) - p_i(\B)$ denote agent $i$'s resulting utility under bids $\B$. Note that $x_i(\B)$, $v_i(\B)$, $p_i(\B)$, and $u_i(\B)$ are all differentiable with respect to $b_i$. Also let $\alpha = \mfrac{\rho}{1-r\rho}$ for brevity; then $p_i(\B) = r\alpha (\sum_{k \ne i} b_k^\alpha) \int_{b = 0}^{b_i} \mfrac{b^{r\alpha}}{(b^\alpha + \sum_{k \ne i} b_k^\alpha)^{r+1}}\dif b$ and $x_i(\B) = \mfrac{b_i^\alpha}{\sum_{k \in N} b_k^\alpha}$. To prove truthfulness, we need to show that $w_i \in \argmax_{b_i \in \bbr_{> 0}} u_i(\B)$, i.e., truthfully reporting $w_i$ is an optimal strategy for agent $i$.\footnote{Note that $u_i(\B)$ is not concave in $b_i$, since $p_i(\B)$ is not convex in $b_i$. Thus the KKT conditions do not apply, so we will have to use a different approach.} Since $u_i(\B)$ is differentiable with respect to $b_i$, we have $\mfrac{\partial u_i(\B)}{\partial b_i} = \mfrac{\partial v_i(\B)}{\partial b_i} - \mfrac{\partial p_i(\B)}{\partial b_i}$. The first term on the right hand side is \[ \frac{\partial v_i(\B)}{\partial b_i} = rw_i\frac{\partial x_i(\B)}{\partial b_i} x_i(\B)^{r-1} \] The second term is \begin{align*} \frac{\partial p_i(\B)}{\partial b_i} =&\ r \alpha \Big(\sum_{k \ne i} b_k^\alpha\Big) \frac{b_i^{r\alpha}}{(b_i^\alpha + \sum_{k \ne i} b_k^\alpha)^{r+1}}\\ =&\ r \alpha \Big(\sum_{k \ne i} b_k^\alpha\Big) \frac{b_i^{r\alpha}}{(\sum_{k \in N} b_k^\alpha)^{r+1}}\\ =&\ r \alpha \Big(\sum_{k \ne i} b_k^\alpha\Big) \frac{b_i^\alpha}{(\sum_{k \in N} b_k^\alpha)^2} \Big(\frac{b_i^\alpha}{\sum_{k \in N} b_k^\alpha}\Big)^{r-1}\\ =&\ r \alpha \Big(\sum_{k \ne i} b_k^\alpha\Big) \frac{b_i^\alpha}{(\sum_{k \in N} b_k^\alpha)^2} x_i(\B)^{r-1} \end{align*} Conveniently, we have $\mfrac{\partial}{\partial b_i} \big(\mfrac{b_i^\alpha}{\sum_{k \in N} b_k^\alpha}\big) = \alpha \big(\sum_{k \ne i} b_k^\alpha\big) \mfrac{b_i^{\alpha-1}}{(\sum_{k \in N} b_k^\alpha)^2}$. Thus \begin{align*} \frac{\partial p_i(\B)}{\partial b_i} =&\ r b_i \frac{\partial}{\partial b_i} \Big(\frac{b_i^\alpha}{\sum_{k \in N} b_k^\alpha}\Big) x_i(\B)^{r-1}\\ =&\ r b_i \frac{\partial x_i(\B)}{\partial b_i} x_i(\B)^{r-1} \end{align*} Therefore \[ \frac{\partial u_i(\B)}{\partial b_i} = r (w_i - b_i) \frac{\partial x_i(\B)}{\partial b_i} x_i(\B)^{r-1} \] Since $\frac{\partial x_i(\B)}{\partial b_i} > 0$ and $x_i(\B)^{r-1} > 0$ for all $b_i$, this implies \begin{enumerate} \item For all $b_i < w_i$, $\mfrac{\partial u_i(\B)}{\partial b_i} > 0$. \item For all $b_i > w_i$, $\mfrac{\partial u_i(\B)}{\partial b_i} > 0$. \item For $b_i = w_i$, $\mfrac{\partial u_i(\B)}{\partial b_i} = 0$. \end{enumerate} Therefore $w_i \in \argmax_{b_i \in \bbr_{> 0}}$ (in fact, $w_i$ is the unique maximizer). We conclude that the mechanism is truthful. \end{proof} From a technical standpoint, the harder task is proving that this mechanism is unique (up to additive constants in the payment rule). We assume without loss of generality that the mechanism asks each agent $i$ to report $w_i$, and let $\B=b_1,\dots,b_n$ be the vector of reported $w_i$'s. We use the standard notation of $(\B_{-i}, b_i')$ to denote the vector where the $i$th entry is $b_i'$, and the $k$th entry for each $k\ne i$ is $b_k$. The proof takes a real analysis approach, with Kirszbraun's Theorem for Lipschitz extensions~\cite{kirszbraun1934} playing a central role. On a high level, the proof proceeds as follows: (1) we establish some basic properties of the payment rule, (2) we show that the payment rule must be Lipschitz continuous not including $b_i = 0$, (3) there exists a Lipschitz extension $\phat$ including $b_i = 0$ (Kirszbraun's Theorem), (4) since $\phat$ is Lipschitz, it is differentiable almost everywhere and is equal to the integral of its derivative, (5) since it has the same derivative (when defined) as the payment rule from Theorem~\ref{thm:truthful}, the payment rules are equal (up to the constant of integration). \begin{theorem}\label{thm:truthful-unique} Assume $m=1$, and that each $v_i$ is homogenous of degree $r$ (with $r$ publicly known), concave, and differentiable. Fix $\rho \in (0,1)$, and let $\Gamma$ be a truthful mechanism which outputs an allocation $\mathbf{x} \in \Psi(\rho)$. Then the allocation rule is the same as in Theorem~\ref{thm:truthful}, and the payment rule $p_i(\B)$ is the same up to an additive constant. \end{theorem} \begin{proof} \textbf{Part 1: Setup and basic properties.} Since there is a unique optimal allocation (Lemma~\ref{lem:m=1-x}), $\Gamma$ must take $\B = (b_1,\dots, b_n)$ as honest and output the same allocation $\mathbf{x}(\B)$. It remains to consider the payment rule. Let $p_i(\B)$ denote the payment rule from Theorem~\ref{thm:main}, and let $\tilp(\B)$ denote the payment rule for $\Gamma$. Given reports $\B$, define $v_i(\B)$ as before, and let $u_i(\B) = v_i(\B) - \tilp(\B)$ be agent $i$'s resulting utility under $\Gamma$. From the point of view of a given agent $i$, the other agents' reports $\B_{-i}$ can be treated as a constant. Thus for brevity, write $x_i(b) = x_i(\B_{-i}, b)$, $p_i(b) = p_i(\B_{-i}, b)$, and $\tilp(b) = \tilp(\B_{-i}, b)$ for each $i \in N$. Fix an $i \in N$. Since $\Gamma$ is truthful, we must have $w_i \in \argmax_{b_i \in \bbr_{>0}} u_i(\B)$. Then by definition of $u_i(\B)$, we have $w_i \in \argmax_{b_i \in \bbr_{>0}} \big(w_i x_i(b_i)^r - \tilp(b_i)\big)$. Since $w_i$ could be any element of $\bbrspos$, and $\Gamma$ must be agnostic to $w_i$, we must have $b \in \argmax_{b_i \in \bbr_{>0}} \big(b x_i(b_i)^r - \tilp(b_i)\big)$ for all $b \in \bbrspos$. We first claim that $\tilp(b)$ is nondecreasing. Suppose the opposite: then there exists exists $b > b'$ such that $\tilp(b) < \tilp(b')$. But this means that if $w_i = b'$, reporting $b_i = w_i$ is never an optimal strategy, because the payment can be decreased by reporting $b_i = b$, and $x_i(b) \ge x_i(b')$ (since $x_i(b)$ is nondecreasing). Thus $\tilp(b)$ is nondecreasing. \medskip \textbf{Part 2: $\tilp$ is Lipschitz continuous.} Fix an arbitrary $b_i > 0$. Since $b_k > 0$ for all $k \ne i$, it can be seen from the definition of $x_i(b)$ that $x_i(b)^r$ is continuously differentiable on $[0,b_i]$. Therefore the maximum of $\mfrac{\dif x_i(b)^r}{\dif b}$ is a Lipschitz constant for $x_i(b)^r$, so $x_i(b)^r$ is Lipschitz continuous on $[0,b_i]$. Let $\kappa$ be this Lipschitz constant: then for all $b, b' \in [0, b_i]$, $|x_i(b)^r - x_i(b')^r| \le \kappa |b - b'|$. We claim that $\tilp$ is Lipschitz continuous on $(0, b_i]$ with constant $b_i \kappa$. Suppose the opposite: then there exist $b, b' \in (0, b_i]$ such that $|\tilp(b) - \tilp(b')| > b_i \kappa |b - b'|$. Assume without loss of generality that $b > b'$. Since $\tilp$ and $x_i$ are both nondecreasing, we then have $\tilp(b) - \tilp(b') > b_i \kappa (b - b')$ and $x_i(b)^r - x_i(b')^r \le \kappa(b-b')$. Since $b \in \argmax_{b_i \in \bbr_{>0}} \big(b x_i(b_i)^r - \tilp(b_i)\big)$, we have $b x_i(b)^r - \tilp(b) \ge b x_i(b')^r - \tilp(b')$ and thus $b(x_i(b)^r - x_i(b')^r) \ge \tilp(b) - \tilp(b')$. Therefore \[ b\kappa(b-b') \ge b\big(x_i(b)^r - x_i(b')^r\big) \ge \tilp(b) - \tilp(b') > b_i \kappa (b - b') \] Therefore $\mfrac{b\kappa}{b_i \kappa} > 1$, which contradicts $b \le b_i$. Therefore $\tilp$ is Lipschitz continuous on $(0, b_i]$. \medskip \textbf{Part 3: Kirszbraun's Theorem.} Thus by Kirszbraun's Theorem~\cite{kirszbraun1934}, $\tilp$ has a Lipschitz extension to $[0,b_i]$: that is, there exists $\phat: [0,b_i] \to \bbrpos$ such that $\phat$ is Lipschitz continuous on $[0,b_i]$, and $\phat(b) = \tilp(b)$ for $b \in (0, b_i]$. \medskip \textbf{Part 4: $\phat$ is the integral of its derivative.} Lipschitz continuity implies absolute continuity~\cite{Royden1988}, so $\phat$ is absolutely continuous on $[0, b_i]$. Thus by the Fundamental Theorem of Lebesgue Calculus~\cite{Royden1988}, $\phat$ is differentiable almost everywhere on $[0, b_i]$, its derivative $\mfrac{\dif \phat(b)}{\dif b}$ is integrable over $[0, b_i]$, and \[ \phat(b_i) - \phat(0) = \int_{b = 0}^{b_i} \frac{\dif \phat(b)}{\dif b} \dif b \] \textbf{Part 5: The derivatives of $\phat$ and $p_i$ match, so $\phat = p_i + c$.} Consider a $b > 0$ at which $\phat$ is differentiable. Then $\tilp$ is also differentiable, so $b \in \argmax_{b_i \in \bbr_{>0}} \big(b x_i(b_i)^r - \tilp(b_i)\big)$ implies $\mfrac{\dif \tilp(b)}{\dif b} = b\mfrac{\dif}{\dif b} (x_i(b)^r) = rb\mfrac{\dif x_i(b)}{\dif b} x_i(b)^{r-1}$.\footnote{Note that the $b$ in $bx_i(b_i)^r$ is a constant from the point of view of the argmax, so it is treated as a constant by the derivative. To be technically precise, we have $(\frac{\dif}{\dif b_i} b x_i(b_i)^r)|_{b_i = b} = rb\frac{\dif x_i(b)}{\dif b} x_i(b)^{r-1}$.} We showed in the proof of Theorem~\ref{thm:truthful} that $\mfrac{\partial p_i(\B)}{\partial b_i} = r b_i \frac{\partial x_i(\B)}{\partial b_i} x_i(\B)^{r-1}$; equivalently, $\mfrac{\dif p_i(b)}{\dif b} = r b \frac{\dif x_i(b)}{\dif b} x_i(b)^{r-1}$. Therefore for all $b > 0$ at which $\phat$ is differentiable, we have $\mfrac{\dif \phat(b)}{\dif b} = \mfrac{\dif p_i(b)}{\dif b}$. Since $\phat$ is differentiable almost everywhere, we have $\mfrac{\dif \phat(b)}{\dif b} = \mfrac{\dif p_i(b)}{\dif b}$ almost everywhere. Thus $ \mfrac{\dif p_i(b)}{\dif b}$ is also integrable over $[0,b_i]$, and $\int_{b = 0}^{b_i} \mfrac{\dif p_i(b)}{\dif b} \dif b = \int_{b = 0}^{b_i} \mfrac{\dif \phat(b)}{\dif b} \dif b$~\cite{Royden1988}. Therefore \begin{align*} \phat(b_i) =&\ \phat(0) + \int_{b = 0}^{b_i} \frac{\dif p_i(b)}{\dif b} \dif b\\ =&\ \phat(0) + p_i(b_i) \end{align*} where the second equality is from the definition of $p_i(\B)$. Therefore for all $b_i > 0$, $\tilp(b_i) = \phat(0) + p_i(b_i)$, and so $\tilp(\B) = \phat(0) + p_i(\B)$ for all $\B$. Since this holds for all $i \in N$, $\tilp(\B)$ is exactly the payment rule from Theorem~\ref{thm:truthful}, up to the additive constant of $\phat(0)$. \end{proof} It is worth noting that this truthful payment rule is quite complex; in particular, it may be hard to convince agents that it is actually in their best interest to be truthful. In contrast, the Walrasian pricing rule from Theorem~\ref{thm:main} is much simpler and more intuitive. That pricing rule is not truthful, but perhaps formal truthfulness is not crucial if a practical iterative implementation is possible. We do not claim that our algorithm from Section~\ref{sec:comp} is truly practical, but it could be a step in the right direction.

https://arxiv.org/abs/1902.07137

Recovery of a mixture of Gaussians by sum-of-norms clustering

Sum-of-norms clustering is a method for assigning $n$ points in $\mathbb{R}^d$ to $K$ clusters, $1\le K\le n$, using convex optimization. Recently, Panahi et al.\ proved that sum-of-norms clustering is guaranteed to recover a mixture of Gaussians under the restriction that the number of samples is not too large. The purpose of this note is to lift this restriction, i.e., show that sum-of-norms clustering with equal weights can recover a mixture of Gaussians even as the number of samples tends to infinity. Our proof relies on an interesting characterization of clusters computed by sum-of-norms clustering that was developed inside a proof of the agglomeration conjecture by Chiquet et al. Because we believe this theorem has independent interest, we restate and reprove the Chiquet et al.\ result herein.

\section{Introduction} Clustering is perhaps the most central problem in unsupervised machine learning and has been studied for over 60 years \cite{shais}. The problem may be stated informally as follows. One is given $n$ points, $\a_1,\ldots,\a_n$ lying in $\R^d$. One seeks to partition $\{1,\ldots,n\}$ into $K$ sets $C_1,\ldots,C_K$ such that the $\a_i$'s for $i\in C_m$ are closer to each other than to the $\a_i$'s for $i\in C_{m'}$, $m'\ne m$. Clustering is usually posed as a nonconvex optimization problem, and therefore prone to nonoptimal local minimizers, but Pelckmans et al.\ \cite{pelckmans}, Hocking et al.\ \cite{hocking}, and Lindsten et al.\ \cite{lindsten} proposed the following convex formulation for the clustering problem: \begin{equation} \min_{\x_1,\ldots,\x_n\in\R^d} \frac{1}{2}\sum_{i=1}^n \norm{\x_i-\a_i}^2 +\lambda\sum_{i<j}\norm{\x_i-\x_j}. \label{eq:son-clustering} \end{equation} This formulation is known in the literature as sum-of-norms clustering, convex clustering, or clusterpath clustering. Let $\x_1^*,\ldots,\x_n^*$ be the optimizer. (Note: \eqref{eq:son-clustering} is strongly convex, hence the optimizer exists and is unique.) The assignment to clusters is given by the $\x_i^*$'s: for $i,i'$, if $\x_i^*=\x_{i'}^*$ then $i,i'$ are assigned to the same cluster, else they are assigned to different clusters. It is apparent that for $\lambda=0$, each $\a_i$ is assigned to a different cluster (unless $\a_i=\a_{i'}$ exactly), whereas for $\lambda$ sufficiently large, the second summation drives all the $\x_i$'s to be equal (and hence there is one big cluster). Thus, the parameter $\lambda$ controls the number of clusters produced by the formulation. Throughout this paper, we assume that all norms are Euclidean, although \eqref{eq:son-clustering} has also been considered for other norms. In addition, some authors insert nonnegative weights in front of the the terms in the above summations. Our results, however, require all weights identically 1. Panahi et al.\ \cite{Panahi} developed several recovery theorems as well as a first-order optimization method for solving \eqref{eq:son-clustering}. Other authors, e.g., Sun et al.\ \cite{dsun1} have since extended these results. One of Panahi et al.'s results pertains to a mixture of Gaussians, which is the following generative model for producing the data $\a_1,\ldots,\a_n$. The parameters of the model are $K$ means $\bmu_1,\ldots,\bmu_K\in\R^d$, $K$ variances $\sigma_1^2,\ldots,\sigma_k^2$, and $K$ probabilities $w_1,\ldots,w_k$, all positive and summing to 1. One draws $n$ i.i.d.\ samples as follows. First, an index $m\in\{1,\ldots,K\}$ is selected at random according to probabilities $w_1,\ldots,w_K$. Next, a point $\a$ is chosen according to the spherical Gaussian distribution $N(\bmu_m,\sigma_m^2I)$. Panahi et al.\ proved that for the appropriate choice of $\lambda$, sum-of-norms clustering formulation \eqref{eq:son-clustering} will exactly recover a mixture of Gaussians (i.e., each point will be labeled with $m$ if it was selected from $N(\bmu_m,\sigma_m^2I)$) provided that for all $m,m'$, $1\le m<m'\le K$, \begin{equation} \norm{\bmu_m-\bmu_{m'}}\ge \frac{CK\sigma_{\max}}{w_{\min}}\mathrm{polylog}(n). \label{eq:panahibound} \end{equation} One issue with this bound is that as the number of samples $n$ tends to infinity, the bound seems to indicate that distinguishing the clusters becomes increasingly difficult (i.e., the $\bmu_m$'s have to be more distantly separated as $n\rightarrow\infty$). The reason for this aspect of their bound is that their proof technique requires a gap of positive width (i.e., a region of $\R^d$ containing no sample points) between $\{\a_i:i\in C_m\}$ and $\{\a_i:i\in C_{m'}\}$ whenever $m\ne m'$. Clearly, such a gap cannot exist in the mixture-of-Gaussians distribution as the number of samples tends to infinity. The purpose of this note is to prove that \eqref{eq:son-clustering} can recover a mixture of Gaussians even as $n\rightarrow\infty$. This is the content of Theorem~\ref{thm:mainrecovery} in Section~\ref{sec:mainrecovery} below. Naturally, under this hypothesis we cannot hope to correctly label all samples since, as $n\rightarrow\infty$, some of the samples associated with one mean will be placed arbitrarily close to another mean. Therefore, we are content in showing that \eqref{eq:son-clustering} can correctly cluster the points lying within some fixed number of standard-deviations for each mean. Radchenko and Mukherjee \cite{Radchenko} have previously analyzed the special case of mixture of Gaussians with $K=2$, $d=1$ under slightly different hypotheses. Our proof technique requires a cluster characterization theorem for sum-of-norms clustering derived by Chiquet et al.\ \cite{chiquet}. This theorem is not stated by these authors as a theorem, but instead appears as a sequence of steps inside a larger proof in a ``supplementary material'' appendix to their paper. Because we believe that this theorem is of independent interest, we restate it below and for the sake of completeness provide the proof (which is the same as the proof appearing in Chiquet et al.'s supplementary material). This material appears in Section~\ref{sec:chiquetproof}. \section{Cluster characterization theorem} \label{sec:chiquetproof} The following theorem is due to Chiquet et al.\ \cite{chiquet} but is not stated as a theorem by these authors; instead it appears as a sequence of steps in a proof of the agglomeration conjecture. Refer to the next section for a discussion of the agglomeration conjecture. We restate the theorem here because it is needed for our analysis and because we believe it is of independent interest. \begin{theorem} Let $\x_1^*,\ldots,\x_n^*$ denote the optimizer of \eqref{eq:son-clustering}. For notational ease, let $\x^*$ denote the concatenation of these vectors into a single $nd$-vector. Suppose that $C$ is a nonempty subset of $\{1,\ldots,n\}$. (a) Necessary condition: If for some $\hat\x \in \R^d$, $\x_i^*=\hat\x$ for $i\in C$ and $\x_i^*\ne\hat\x$ for $i\notin C$ (i.e., $C$ is exactly one cluster determined by \eqref{eq:son-clustering}), then there exist $\z_{ij}^*$ for $i,j\in C$, $i\ne j$, which solve \begin{equation} \begin{aligned} \a_i-\frac{1}{|C|}\sum_{l\in C}\a_l&= \lambda\sum_{j\in C-\{i\}} \z_{ij}^* &&\forall i\in C,\\ \norm{\z_{ij}^*} &\leq 1 &&\forall i,j\in C, i\ne j, \\ \z_{ij}^* &= -\z_{ji}^*&& \forall i,j\in C, i\ne j. \end{aligned} \label{eq:zstar1} \end{equation} (b) Sufficient condition: Suppose there exists a solution $\z_{ij}^* $ for $j\in C-\{i\}$, $i\in C$ to the conditions \eqref{eq:zstar1}. Then there exists an $\hat\x \in \R^d$ such that the minimizer $\x^*$ of \eqref{eq:son-clustering} satisfies $\x_i^*=\hat\x$ for $i\in C$. \label{thm:clustchar} \end{theorem} \noindent {\bf Note}: This theorem is an almost exact characterization of clusters that are determined by formulation \eqref{eq:son-clustering}. The only gap between the necessary and sufficient conditions is that the necessary condition requires that $C$ be exactly all the points in a cluster, whereas the sufficient condition is sufficient for $C$ to be a subset of the points in a cluster. The sufficient condition is notable because it does not require any hypothesis about the other $n-|C|$ points occurring in the input. \begin{proof} (Chiquet et al.) \underline{Proof for Necessity (a)} \\ As $\x^*$ is the minimizer of the problem (\ref{eq:son-clustering}), and this objective function, call it $f(\x)$, is convex, it follows that $\bz\in\partial f(\x^*)$, where $\partial f(\x^*)$ denotes the subdifferential, that is, the set of subgradients of $f$ at $\x^*$. (See, e.g., \cite{hiriarturrutylemarechal} for background on convex analysis). Written explicitly in terms of the derivative of the squared-norm and subdifferential of the norm, this means that $\x^*$ satisfies the following condition: \begin{equation} \x_i^*-\a_i+\lambda\sum_{j\ne i}\w_{ij}^*=\bz \qquad\forall i=1,\ldots,n, \label{eq:KKTcond} \end{equation} where $\w_{ij}^*$, $i=1,\ldots,n$, $j=1,\ldots,n$, $i\ne j$, are subgradients of the Euclidean norm function satisfying $$\w_{ij}^* = \left\{ \begin{array}{ll} \frac{\x_i^*-\x_j^*}{\norm{\x_i^*-\x_j^*}}, & \mbox{for $\x_i^*\ne \x_j^*$}, \\ \mbox{arbitrary point in $B(\bz,1)$}, & \mbox{for $\x_i^*=\x_j^*$}, \end{array} \right. $$ with the requirement that $\w_{ij}^*=-\w_{ji}^*$ in the second case. Here, $B(\x,r)$ is notation for the closed Euclidean ball centered at $\x$ of radius $r$. Since $\x_i^*=\hat\x$ for $i\in C$, $\x_i^*\ne\hat\x$ for $i\notin C$, the KKT condition for $i\in C$ is rewritten as \begin{equation} \hat\x-\a_i+\lambda\sum_{j\notin C}\frac{\hat \x-\x_j^*}{\norm{\hat\x-\x_j^*}} +\lambda\sum_{j\in C-\{i\}} \w_{ij}^*=\bz, \label{eq:hatx1} \end{equation} Define $\z_{ij}^*=\w_{ij}^*$ for $i,j\in C$, $i\ne j$. Then $$\norm{\z_{ij}^*} \leq 1, \z_{ij}^* = -\z_{ji}^*, \forall i,j\in C, i\ne j.$$ Substitute $\w_{ij}^* = \z_{ij}^*$ into the equation (\ref{eq:hatx1}) to obtain \begin{equation} \hat\x-\a_i+\lambda\sum_{j\notin C}\frac{\hat \x-\x_j^*}{\norm{\hat\x-\x_j^*}} +\lambda\sum_{j\in C-\{i\}} \z_{ij}^*=\bz, \label{eq:hatx4} \end{equation} Sum the preceding equation over $i\in C$, noticing that the last term cancels out, leaving $$ |C|\hat\x-\sum_{i\in C}\a_i+\lambda|C|\sum_{j\notin C}\frac{\hat \x-\x_j^*}{\norm{\hat\x-\x_j^*}}=\bz, $$ which is rearranged to (renaming $i$ to $l$): \begin{equation} \lambda\sum_{j\notin C}\frac{\hat \x-\x_j^*}{\norm{\hat\x-\x_j^*}}= -\hat\x+\frac{1}{|C|}\sum_{l\in C}\a_l. \label{eq:hatx2} \end{equation} Subtract \eqref{eq:hatx2} from \eqref{eq:hatx4}, simplify and rearrange to obtain \begin{equation} \a_i-\frac{1}{|C|}\sum_{l\in C}\a_l= \lambda\sum_{j\in C-\{i\}} \z_{ij}^* \qquad\forall i\in C, \label{eq:zstar2} \end{equation} as desired. \underline{Proof for Sufficiency (b)}\\ We will show that at the solution of \eqref{eq:son-clustering}, all the $\x_i^*$'s for $i\in C$ have a common value under the hypothesis that $\z_{ij}^* $ is a solution to the equation \eqref{eq:zstar1} for $i,j\in C$, $i\ne j$. First, define the following intermediate problem. Let $\tilde \a$ denote the centroid of $\a_i$ for $i\in C$: $$\tilde \a = \frac{1}{|C|}\sum_{l\in C}\a_l.$$ Consider the weighted problem sum-of-norms clustering problem with unknowns as follows: one unknown $\x\in\R^d$ is associated with $C$, and one unknown $\x_j$ is associated with each $j\notin C$ (for a total of with $n-|C|+1$ unknown vectors): \begin{equation} \min_{\x;\x_j} \frac{|C|}{2}\cdot \norm{\x-\tilde\a}^2 + \frac{1}{2}\sum_{j\notin C}\norm{\x_j-\a_j}^2 +\lambda|C|\sum_{j\notin C} \norm{\x-\x_j} + \lambda\sum_{\substack {{i,j\notin C} \\{i<j}}}\norm{\x_i-\x_j}. \label{eq:sonclust2} \end{equation} This problem, being strongly convex, has a unique optimizer; denote the optimizing vectors $\tilde \x$ and $\tilde \x_j$ for $j\notin C$. First, let us consider the optimality conditions for \eqref{eq:sonclust2}, which are: \begin{align} |C|(\tilde\x-\tilde\a)+\lambda|C|\sum_{j\notin C}\g_j&=\bz, \label{eq:sonclust2_opt1} \\ \tilde\x_i-\a_i-\lambda|C|\g_i+\lambda\sum_{j\notin C\cup\{i\}}\y_{ij}&=\bz \qquad \forall i\notin C, \label{eq:sonclust2_opt2} \end{align} with subgradients defined as follows: $$\g_j=\left\{ \begin{array}{ll} \frac{\tilde\x-\tilde\x_j}{\norm{\tilde\x-\tilde\x_j}}, & \mbox{for $\tilde\x_j\ne\tilde\x$}, \\ \mbox{arbitrary in $B(\bz,1)$}, & \mbox{for $\tilde\x_j=\tilde\x$}, \end{array} \right. \qquad\forall j\notin C,$$ and $$\y_{ij} = \left\{ \begin{array}{ll} \frac{\tilde\x_i-\tilde\x_j}{\norm{\tilde\x_i-\tilde\x_j}}, & \mbox{for $\tilde\x_i\ne\tilde\x_j$}, \\ \mbox{arbitrary in $B(\bz,1)$}, & \mbox{for $\tilde\x_i=\tilde\x_j$}, \end{array} \right. \qquad\forall i,j\notin C, i\ne j,$$ with the proviso that in the second case, $\y_{ij}=-\y_{ji}$. We claim that the solution for \eqref{eq:son-clustering} given by defining $\x_i^*=\tilde\x$ for $i\in C$ while keeping the $\x_j^*=\tilde\x_j$ for $j\notin C$, where $\tilde\x$ and $\tilde\x_j$ are the optimizers for \eqref{eq:sonclust2} as in the last few paragraphs, is optimal for \eqref{eq:son-clustering}, which proves the main result. To show that this solution is optimal for \eqref{eq:son-clustering}, we need to provide subgradients to establish the necessary condition. Define $\w_{ij}$ to be the subgradients of $\x_i\mapsto \norm{\x_i- \tilde\x_j^*}$ evaluated at $\tilde\x_i^*$ as follows: \begin{align*} \w_{ij}&=\g_{j} && \mbox{for $i\in C, j\notin C$}, \\ \w_{ij}&=\y_{ij} &&\mbox{for $i,j\notin C$, $i\ne j$}, \\ \w_{ij}&=\z_{ij}^*&&\mbox{for $i,j\in C$, $i\ne j$,} \end{align*} Before confirming that the necessary condition is satisfied, we first need to confirm that these are all valid subgradients. In the case that $i\in C$, $j\notin C$, we have constructed $\g_{j}$ to be a valid subgradient of $\x\mapsto \norm{\x-\tilde \x_j}$ evaluated at $\tilde \x$, and we have taken $\x_i^*=\tilde\x$, $\x_j^*=\tilde\x_j$. In the case that $i,j\notin C$, we have construct $\y_{ij}$ to be a valid subgradient of $\x\mapsto \norm{\x-\tilde\x_j}$ evaluated at $\tilde\x_i$, and we have taken $\x_i^*=\tilde\x_i$, $\x_j^*=\tilde\x_j$. In the case that $i,j\in C$, by construction $\x_i^*=\x_j^*=\tilde\x$, so any vector in $B(\bz,1)$ is a valid subgradient of $\x\mapsto\norm{\x-\tilde\x_j}$ evaluated $\tilde\x_i$. Note that since $\z_{ij}^*\in B(\bz,1)$, then $\w_{ij}$ defined above also lies in $B(\bz,1)$. Now we check the necessary conditions for optimality in \eqref{eq:son-clustering}. First, consider an $i\in C$: \begin{align*} \tilde \x_i^*-\a_i+\lambda\sum_{j\ne i}\w_{ij} &= \tilde \x - \a_i+\lambda \sum_{j\in C-\{i\}}\w_{ij}+ \lambda\sum_{j\notin C}\w_{ij} \\ &= \tilde \x - \a_i+\lambda \sum_{j\in C-\{i\}}\z_{ij}^*+\lambda\sum_{j\notin C}\g_j \\ &= \tilde \x - \a_i+\a_i-\frac{1}{|C|}\sum_{l\in C}\a_l +\lambda\sum_{j\notin C}\g_j &&\mbox{(by \eqref{eq:zstar1})} \\ &= \tilde\x -\tilde\a +\lambda\sum_{j\notin C}\g_j \\ & = \bz &&\mbox{(by \eqref{eq:sonclust2_opt1}).} \end{align*} Then we check for $i\notin C$: \begin{align*} \tilde \x_i^*-\a_i+\lambda\sum_{j\ne i}\w_{ij} &= \tilde \x_i-\a_i+\lambda\sum_{j\in C}\w_{ij}+\lambda \sum_{j\notin C\cup\{i\}} \w_{ij} \\ &= \tilde \x_i-\a_i+\lambda\sum_{j\in C}(-\g_i)+\lambda \sum_{j\notin C\cup\{i\}} \y_{ij} \\ & = \tilde \x_i-\a_i-\lambda|C|\g_i+\lambda \sum_{j\notin C\cup\{i\}} \y_{ij} \\ &= \bz && \mbox{(by \eqref{eq:sonclust2_opt2}).} \end{align*} \end{proof} \section{Agglomeration Conjecture} Recall that when $\lambda=0$, each $\a_i$ is in its own cluster in the solution to \eqref{eq:son-clustering} (provided the $\a_i$'s are distinct), whereas for sufficiently large $\lambda$, all the points are in one cluster. Hocking et al.\ \cite{hocking} conjectured that sum-of-norms clustering with equal weights has the following agglomeration property: as $\lambda$ increases, clusters merge with each other but never break up. This means that the solutions to \eqref{eq:son-clustering} as $\lambda$ ranges over $[0,\infty)$ induce a tree of hierarchical clusters on the data. This conjecture was proved by Chiquet et al.\ \cite{chiquet} using Theorem~\ref{thm:clustchar}. Consider a $\bar \lambda \geq \lambda$ and its corresponding sum-of-norms cluster model: \begin{equation} \min_{\x_1,\ldots,\x_n}\frac{1}{2}\sum_{i=1}^n\norm{\x_i-\a_i}^2 +\bar \lambda \sum_{i<j}\norm{\x_i-\x_j}. \label{eq:sonclust3} \end{equation} \begin{corollary} (Chiquet et al.) If there is a $C$ such that minimizer $\x^*$ of \eqref{eq:son-clustering} satisfies $\x_i^*=\hat\x$ for $i\in C$, $\x_i^*\ne\hat\x$ for $i\notin C$ for some $\hat\x \in \R^d$, then there exists an $\hat\x' \in \R^d$ such that the minimizer of \eqref{eq:sonclust3}, $\bar\x^*$, satisfies $\bar\x_i^*=\hat\x'$ for $i\in C$. \end{corollary} The corollary follows from Theorem~\ref{thm:clustchar}. If $C$ is a cluster in the solution of \eqref{eq:son-clustering}, then by the necessary condition, there exist multipliers $\z_{ij}^*$ satisfying \eqref{eq:zstar1} for $\lambda$. If we scale each of these multipliers by $\lambda/\bar\lambda$, we now obtain a solution to \eqref{eq:zstar1} for with $\lambda$ replaced by $\bar\lambda$, and the theorem states that this is sufficient for the points in $C$ to be in the same cluster in the solution to \eqref{eq:sonclust3}. It should be noted that Hocking et al.\ construct an example of unequally-weighted sum-of-norms clustering in which the agglomeration property fails. It is still mostly an open question to characterize for which norms and for which families of unequal weights the agglomeration property holds. Refer to Chi and Steinerberger \cite{Chi} for some recent progress. \section{Mixture of Gaussians} \label{sec:mainrecovery} In this section, we present our main result about recovery of mixture of Gaussians. As noted in the introduction, a theorem stating that every point is labeled correctly is not possible in the setting of $n\rightarrow\infty$, so we settle for a theorem stating that points within a constant number of standard deviations from the means are correctly labeled. \begin{theorem} \label{thm:mainrecovery} Let the vertices $\a_1,\ldots,\a_n\in\R^d$ be generated from a mixture of $K$ Gaussian distributions with parameters $\bmu_1,\ldots,\bmu_K$, $\sigma_1^2,\ldots,\sigma_K^2$, and $w_1,\ldots,w_K$. Let $\theta>0$ be given, and let \[V_m = \{\a_i: \norm{\a_i - \bmu_m} \leq \theta\sigma_m\},\quad m=1,\dots,K.\] Let $\epsilon>0$ be arbitrary. Then for any $m=1,\ldots,K$, with probability exponentially close to $1$ (and depending on $\epsilon$) as $n\rightarrow \infty$, for the solution $\x^*$ computed by \eqref{eq:son-clustering}, the points in $V_m$ are in the same cluster provided \begin{equation} \lambda \ge \frac{2\theta\sigma_m}{(F(\theta,d)w_m-\epsilon)n}. \label{eq:mainrecoverylambdalb} \end{equation} Here, $F(\theta,d)$ denotes the cumulative density function of the chi-squared distribution with $d$ degrees of freedom (which tends to $1$ rapidly as $\theta$ increases). Furthermore, the cluster associated with $V_m$ is distinct from the cluster associated with $V_{m'}$, $1\le m<m'<k$, provided that \begin{equation} \lambda <\frac{\norm{\bmu_{m}-\bmu_{m'}}}{2(n-1)}. \label{eq:mainrecoverylambdaub} \end{equation} \label{thm:recovery} \end{theorem} \begin{proof} Let $\epsilon>0$ be fixed. Fix an $m\in\{1,\ldots,K\}$. First, we show that all the points in $V_m$ are in the same cluster. The usual technique for proving a recovery result is to find subgradients to satisfy the sufficient condition, which in this case is Theorem~\ref{thm:clustchar} taking $C$ in the theorem to be $V_m$. Observe that conditions \eqref{eq:zstar1} involve equalities and norm inequalities. A standard technique in the literature (see, e.g., Cand\`es and Recht \cite{Candes2009}) is to find the least-squares solution to the equalities and then prove that it satisfies the inequalities. This is the technique we adopt herein. The conditions \eqref{eq:zstar1} are in sufficiently simple form that we can write down the least-squares solution in closed form; it turns out to be: \[\z_{ij}^* = \frac{1}{\lambda |V_m|}(\a_i - \a_j) \quad \forall i,j\in V_m, i\ne j.\] It follows by construction (and is easy to check) that this formula satisfies the equalities in \eqref{eq:zstar1}, so the remaining task is to show that the norm bound $\norm{\z_{ij}^*}\le 1$ is satisfied. By definition of $V_m$, $\Vert\a_i-\a_j\Vert\le 2\theta\sigma_m$. The probability that an arbitrary sample $\a_i$ is associated with mean $\bmu_m$ is $w_m$. Furthermore, with probability $F(\theta,d)$, this sample satisfies $\norm{\a_i-\bmu_m}\le \theta\sigma_m$, i.e., lands in $V_m$. Since the second choice in the mixture of Gaussians is conditionally independent from the first, the overall probability that $\a_i$ lands in $V_m$ is $F(\theta,d)w_m$. Therefore, $E[|V_m|]=F(\theta,d)w_mn$. By the Chernoff bound for the tail of a binomial distribution, it follows that the probability that $|V_m|\ge (F(\theta,d)w_m-\epsilon)n$ is exponentially close to 1 for a fixed $\epsilon>0$. Thus, provided $\lambda \ge 2\theta\sigma_m/((F(\theta,d)w_m-\epsilon)n)$, we have constructed a solution to \eqref{eq:zstar1} with probability exponentially close to 1. For the second part of the theorem, suppose $1\le m<m'\le K$. For each sample $\a_i$ associated with $\bmu_m$ satisfying $\Vert\a_i-\bmu_m\Vert\le \theta\sigma_m^2$ (i.e., lying in $V_m$), the probability is $1/2$ that \[(\a_i-\bmu_m)^T(\bmu_{m'}-\bmu_m)\le 0\] by the fact that the spherical Gaussian distribution has mirror-image symmetry about any hyperplane through its mean. Therefore, with probability exponentially close to 1 as $n\rightarrow\infty$, we can assume that at least one $\a_i\in V_m$ satisfies the above inequality. Similarly, with probability exponentially close to 1, at least one sample $\a_{i'}\in V_{m'}$ satisfies \[(\a_{i'}-\bmu_{m'})^T(\bmu_{m}-\bmu_{m'})\le 0.\] Then \begin{align} \norm{\a_i-\a_{i'}}^2 &=\norm{\a_i-\bmu_m-\a_{i'}+\bmu_{m'}+\bmu_m-\bmu_{m'}}^2 \notag \\ &=\norm{\a_i-\bmu_m-\a_{i}+\bmu_{m'}}^2 + 2(\a_i-\bmu_m)^T(\bmu_{m}-\bmu_{m'}) \notag \\ &\hphantom{=}\quad\mbox{}-2(\a_{i'}-\bmu_{m'})^T(\bmu_m-\bmu_{m'}) + \norm{\bmu_{m}-\bmu_{m'}}^2 \notag\\ &\ge \norm{\bmu_{m}-\bmu_{m'}}^2, \label{eq:aimumineq} \end{align} where, in the final line, we used the two inequalities derived earlier in this paragraph. Consider the first-order optimality conditions for equation \eqref{eq:son-clustering}, which are given by \eqref{eq:KKTcond}. Apply the triangle inequality to the summation in \eqref{eq:KKTcond} to obtain, \begin{align} \norm{\x_i^* - \a_i} &\leq \lambda(n-1),\mbox{ and } \label{xisai} \\ \norm{\x_{i'}^* - \a_{i'}} &\leq \lambda(n-1).\label{xisai2} \end{align} Therefore, \begin{align*} \norm{\x_i^*-\x_{i'}^*} &=\norm{\a_{i}-\a_{i'}+\x_i^*-\a_i-\x_{i'}^*+\a_{i'}} \\ &\ge \norm{\a_{i}-\a_{i'}}-\norm{\x_i^*-\a_i}-\norm{\x_{i'}^*-\a_{i'}} && \mbox{(by the triangle inequality)} \\ &\ge \norm{\bmu_{m'}-\bmu_m}-2\lambda(n-1) &&\mbox{(by \eqref{eq:aimumineq}, \eqref{xisai}, and \eqref{xisai2}).} \end{align*} Therefore, we conclude that $\x_i^*\ne \x_{i'}^*$, i.e., that $V_m$ and $V_{m'}$ are not in the same cluster, provided that the right-hand side of the preceding inequality is positive, i.e., \[\lambda <\frac{\norm{\bmu_{m}-\bmu_{m'}}}{2(n-1)}. \] This concludes the proof of the second statement. \end{proof} Clearly, there exists a $\lambda$ so that the solution to \eqref{eq:son-clustering} can simultaneously place all points in $V_m$ into the same cluster for each $m=1,\ldots,K$ while distinguishing the clusters provided that the right-hand side of \eqref{eq:mainrecoverylambdaub} exceeds the right-hand side of \eqref{eq:mainrecoverylambdalb}. In order to obtain a compact inequality that guarantees this condition, let us fix some values. For example, let us take $\theta=2d$ and let $c_d=F(2d,d)$. The Chernoff bound implies that $c_d\rightarrow 1$ exponentially fast in $d$. Let $w_{\min}$ be the minimum weight in the mixture of Gaussians. Let $\sigma_{\max}$ denote the maximum standard deviation in the distribution. Finally, let us take $\epsilon=c_dw_{\min}/2$. Then the above theorem states there is a $\lambda$ such that with probability tending to $1$ exponentially fast in $n$, the points in $V_m$, for any $m=1,\ldots,K$ are each in the same cluster, and these clusters are distinct, provided that \begin{equation} \min_{1\le m<m'\le K}\norm{\bmu_m-\bmu_{m'}}>\frac{16d\sigma_{\max}}{c_dw_{\min}}. \label{eq:ourbound} \end{equation} Compared to the Panahi et al.\ bound \eqref{eq:panahibound}, we have removed the dependence of the right-hand side on $n$ as well as the factor of $K$. (The dependence of the Panahi et al.\ bound on $d$ is not made explicit so we cannot compare the two bounds' dependence on $d$. Note that there is still an implicit dependence on $K$ in \eqref{eq:ourbound} since necessarily $w_{\min}\le 1/K$.) \section{Discussion} The analysis of the mixture of Gaussians in the preceding section used only standard bounds and simple properties of the normal distribution, so it should be apparent to the reader that many extensions of this result (e.g., Gaussians with a more general covariance matrix, uniform distributions, many kinds of deterministic distributions) are possible. The key technique is Theorem~\ref{thm:clustchar}, which essentially decouples the clusters from each other so that each can be analyzed in isolation. Such a theorem does not apply to most other clustering algorithms, or even to sum-of-norm clustering in the case of unequal weights, so obtaining similar results for other algorithms remains a challenge.

https://arxiv.org/abs/1804.07575

Optimal Sorting with Persistent Comparison Errors

We consider the problem of sorting $n$ elements in the case of \emph{persistent} comparison errors. In this model (Braverman and Mossel, SODA'08), each comparison between two elements can be wrong with some fixed (small) probability $p$, and \emph{comparisons cannot be repeated}. Sorting perfectly in this model is impossible, and the objective is to minimize the \emph{dislocation} of each element in the output sequence, that is, the difference between its true rank and its position. Existing lower bounds for this problem show that no algorithm can guarantee, with high probability, \emph{maximum dislocation} and \emph{total dislocation} better than $\Omega(\log n)$ and $\Omega(n)$, respectively, regardless of its running time.In this paper, we present the first \emph{$O(n\log n)$-time} sorting algorithm that guarantees both \emph{$O(\log n)$ maximum dislocation} and \emph{$O(n)$ total dislocation} with high probability. Besides improving over the previous state-of-the art algorithms -- the best known algorithm had running time $\tilde{O}(n^{3/2})$ -- our result indicates that comparison errors do not make the problem computationally more difficult: a sequence with the best possible dislocation can be obtained in $O(n\log n)$ time and, even without comparison errors, $\Omega(n\log n)$ time is necessary to guarantee such dislocation bounds.In order to achieve this optimal result, we solve two sub-problems, and the respective methods have their own merits for further application. One is how to locate a position in which to insert an element in an almost-sorted sequence having $O(\log n)$ maximum dislocation in such a way that the dislocation of the resulting sequence will still be $O(\log n)$. The other is how to simultaneously insert $m$ elements into an almost sorted sequence of $m$ different elements, such that the resulting sequence of $2m$ elements remains almost sorted.

\section{Introduction}\label{sec-introduction} \newcommand{w.h.p.}{w.h.p.} \newcommand{exp.}{exp.} We study the problem of \emph{sorting} $n$ distinct elements under \emph{persistent} random comparison \emph{errors}. This problem arises naturally when sorting is applied to real life scenarios. For example, one could use experts to compare items, with each comparison being performed by one expert. As these operations are typically expensive, one cannot repeat them, and the result may sometimes be erroneous. Still, one would like to reconstruct from these information the correct (or a nearly correct) order of the elements. In this classical model, which \emph{does not allow resampling}, each comparison is wrong with some fixed (small) probability $p$, and correct with probability $1-p$.\footnote{As in previous works, we assume $p<1/32$ though the results holds for $p<1/16$.} The comparison errors are independent over all possible pairs of elements, but they are persistent: Repeating the same comparison several times is useless since the result is always the same, i.e., always wrong or always correct. Because of errors, it is impossible to sort correctly and therefore, one seeks to return a ``nearly sorted'' sequence, that is, a sequence where the elements are ``close'' to their correct positions. To measure the quality of an output sequence in terms of sortedness, a common way is to consider the \emph{dislocation} of an element, which is the difference between its position in the output and its position in the correctly sorted sequence. In particular, one can consider the \emph{maximum dislocation} of any element in the sequence or the \emph{total dislocation} of the sequence, i.e., the sum of the dislocations of all $n$ elements. Note that sorting with persistent errors as above is much more difficult than the case in which comparisons can be repeated, where a trivial $O(n\log^2 n)$ time solution is enough to sort perfectly with high probability (simply repeat each comparison $O(\log n)$ times and take the majority of the results). Instead, in the model with persistent errors, it is impossible to sort perfectly as no algorithm can achieve a maximum dislocation that is smaller than $\Omega(\log n)$ w.h.p., or total dislocation smaller than $\Omega(n)$ in expectation \cite{geissmann_et_al}. Such a problem has been extensively studied in the literature, and several algorithms have been devised with the goal of sorting \emph{quickly} with small dislocation (see Table~\ref{tb-recurrent}). Unfortunately, even though all the algorithms achieve the best possible maximum dislocation of $\Theta(\log n)$, they use a truly superlinear number of comparisons (specifically, $\Omega(n^{c})$ with $c\geq 1.5$), and/or require significant amount of time (namely, $O(n^{3+c})$ where $c$ is a big constant that depends on $p$). This naturally suggests the following question: \begin{quote} \emph{What is the time complexity of sorting optimally with persistent errors?} \end{quote} \noindent In this work, we answer this basic question by showing the following result: \begin{quote} \emph{There exists an algorithm with \textbf{optimal running time} $O(n\log n)$ which achieves simultaneously \textbf{optimal maximum dislocation} $O(\log n)$ and \textbf{optimal total dislocation} $O(n)$, both \textbf{with high probability}.} \end{quote} The dislocation guarantees of our algorithm are optimal, due to the lower bound of \cite{geissmann_et_al}, while the existence of an algorithm achieving a maximum dislocation of $d = O(\log n)$ in time $T(n) = o(n \log n)$ would immediately imply the existence of an algorithm that sorts $n$ elements in $T(n) + O(n \log \log n) = o(n \log n)$ time, even in the absence of comparison errors, thus contradicting the classical $\Omega(n \log n)$ lower bound for comparison-based algorithms.\footnote{Indeed, once the approximately sorted sequence $S$ is computed, it suffices to apply any $O(n \log n)$ sorting algorithm on the first $m=2\max\{d, \log n\}$ elements of $S$, in order to select the smallest $m/2$ elements. Removing those elements from $S$ and repeating this procedure $\frac{n}{m}$ times, allows to sort in $T(n) + O(\frac{n}{m} \cdot m \log m)$ time.} Along the way to our result, we consider the problem of \emph{searching with persistent errors}, defined as follows: \begin{quote}\emph{We are given an approximately sorted sequence $S$, and an additional element $x \not\in S$. The goal is to compute, under persistent comparison errors, an \emph{approximate rank} (position) of $x$ which differs from the true rank of $x$ in $S$ by a \emph{small} additive error.} \end{quote} For this problem, we show an algorithm that requires $O(\log n)$ time to compute, w.h.p., an approximate rank that differs from the true rank of $x$ by at most $O(\max\{d,\log n\})$, where $d$ is the maximum dislocation of $S$. For $d=\Omega(\log n)$ this allows to insert $x$ into $S$ without any asymptotic increase of the maximum (and total) dislocation in the resulting sequence. Notice that, if $d$ is also in $O(n^{1-\epsilon})$ for any constant $\epsilon > 0$, this is essentially the best we can hope for, as an easy decision-tree lower bound shows that any algorithm must require $\Omega(\log n)$ time. Finally, we remark that \cite{Klein2011} considered the variant in which the original sequence is \emph{sorted}, and the algorithm must compute the correct rank. For this problem, they present an algorithm that runs in $O(\log n \cdot \log \log n)$ time and succeeds with probability $1-f(p)$, with $f(p)$ vanishing as $p$ goes to $0$. As by-product of our result, we can obtain the optimal $O(\log n)$ running time with essentially the same success probability. \begin{table} \centering \begin{tabular}{|c|cc|c|} \multicolumn{4}{c}{\textbf{Upper bounds}} \\[\smallskipamount] \hline \textbf{Running Time} & \textbf{Max Dislocation} & \textbf{Tot Dislocation} & \textbf{Reference} \\ \hline $O(n^{3+c})$ & $O(\log n)$ w.h.p. & $O(n)$ w.h.p. & \cite{Braverman2008} \\ $O(n^2)$ & $O(\log n)$ w.h.p. & $O(n\log n)$ w.h.p. & \cite{Klein2011} \\ $O(n^2)$ & $O(\log n)$ w.h.p. & $O(n)$ exp. & \cite{geissmann_et_al} \\ $\tilde{O}(n^{3/2})$ & $O(\log n)$ w.h.p. & $O(n)$ exp. & \cite{newwindowsort} \\ \hline \hline $O(n\log n)$ & $O(\log n)$ w.h.p. & $O(n)$ w.h.p. & \textbf{this work} \\ \hline \multicolumn{4}{c}{} \\[0pt] \multicolumn{4}{c}{\textbf{Lower bounds}} \\[\smallskipamount] \hline Any & $\Omega(\log n)$ w.h.p. & $\Omega(n)$ exp. &\cite{geissmann_et_al} \\ \hline \end{tabular} \caption{The existing approximate sorting algorithms and our result. The constant $c$ in the exponent of the running time of \cite{Braverman2008} depends on the error probability $p$ and it is typically quite large. We write $\Omega(f(n))$ w.h.p. (resp. exp.) to mean that no algorithm can achieve dislocaiton $o(f(n))$ with high probability (resp. in expectation).} \label{tb-recurrent} \end{table} \subsection{Main Intuition and Techniques} \paragraph*{Approximate Sorting} In order to convey the main intuitions behind our $O(n \log n)$-time optimal-dislocation approximate sorting algorithm, we consider the following ideal scenario: we already have a perfectly sorted sequence $A$ containing a random half of the elements in our input sequence $S$ and we, somehow, also know the position in which each element $x \in S\setminus A$ should be inserted into $A$ so that the resulting sequence is also sorted (i.e, the \emph{rank} of $x$ in $A$). If these positions alternate with the elements of $A$, then, to obtain a sorted version of $S$, it suffices to \emph{merge} $S$ and $S \setminus A$, i.e., to simultaneously insert all the elements of $S \setminus A$ into their respective positions of $A$. Unfortunately, we are far from this ideal scenario for several reasons: first of all, multiple, say $\delta$, elements in $S \setminus A$ might have the same rank in $A$. Since we do not know the order in which those elements should appear, this will already increase the dislocation of the merged sequence to $\Omega(\delta)$. Moreover, due to the lower bound of \cite{geissmann_et_al}, we are not actually able to obtain a perfectly sorted version of $A$ and we are forced to work with a permutation of $A$ having dislocation $d = \Omega(\log n)$, implying that the natural bound on the resulting dislocation can be as large as $d \cdot \delta$. This is a bad news, as one can show that $\delta = \Omega(\log n)$. However, it turns out that the number of elements in $S \setminus A$ whose positions lie in a $O(\log n)$-wide interval of $A$ is still $O(\log n)$, w.h.p., implying that the final dislocation of $A$ is just $O(\log n)$. But how do we obtain the approximately sorted sequence $A$ in the first place? We could just recursively apply the above strategy on the (unsorted) elements of $A$, except that this would cause a blow-up in the resulting dislocation due to the constant hidden by the big-O notation. We therefore interleave merge steps with invocations of (a modified version of) the sorting algorithm of \cite{geissmann_et_al}, which essentially reduces the dislocation by a constant factor, so that the increase in the worst-case dislocation will be only an \emph{additive} constant per recursive step. An additional complication is due to the fact that we are not able to compute the exact ranks in $A$ of the elements in $S\setminus A$. We therefore have to deal, once again, with approximations that are computed using the other main contribution of this paper: \emph{noisy binary search trees}, whose key ideas are described in the following. \paragraph*{Noisy Binary Search} As a key ingredient of our approximate sorting algorithm, we need to \emph{merge} an almost-sorted sequence with a set of elements, without causing any substantial increase in the final dislocation. More precisely, if we are given a sequence $S$ with dislocation $d$ and an element $x$, we want to compute an \emph{approximate rank} of $x$ in $S$, i.e., a position that differs by at most $O(\max\{d, \log n\})$ from the position that $x$ would occupy if the elements $S \cup \{ x \}$ were perfectly sorted. As a comparison, this same problem has been solved optimally in $O(\log n)$ time in the easier case in which errors are not persistent and $S$ is already sorted \cite{Feige1994}. The idea of \cite{Feige1994} is to locate the correct position of $x$ using a binary decision tree: ideally each vertex $v$ of the tree tests whether $x$ appears to belong to a certain \emph{interval} of $S$ and, depending on the results, one of the children of $v$ is considered next. As these intervals become narrower as we move from the root towards the leaves, which are in a one-to-one correspondence with positions of $S$, we eventually discover the correct rank of $x$ in $S$. In order to cope with failures, this process is allowed to \emph{backtrack} when inconsistent comparisons are observed, thus repeating some of the comparisons involving ancestors of $v$. Moreover, to guarantee that the result will be correct with high probability, a logarithmic number of consistent comparisons with a leaf are needed before the algorithm terminates. Notice how this process heavily depends on the fact that it is possible to gather more information on the true relative position of $x$ by repeating a comparison multiple times (in fact, it is easy to design a simple $O(\log^2 n)$-time algorithm by exploiting this fact). Unfortunately, this is not the case anymore when errors are \emph{persistent}. To overcome this problem we design a \emph{noisy binary search tree} in which the intervals element $x$ is compared with are, in a sense, \emph{dynamic}, i.e., they grow every time the associated vertex is visited. This, in turn, is a source of other difficulties: first, the intervals of the descendants of $v$ also need to be updated. Moreover, we can obtain inconsistent answers not only due to the erroneous comparisons, but also due to the fact that an interval that initially did not contain $x$ might now become too large. Finally, since intervals overlap, we might end up repeating the same comparison even when two different vertices of the tree are involved. We overcome these problems by using two search trees that initially comprise of disjoint intervals in $S$, which are selected in a way that ensures that all the bad-behaving vertices are confined into only one of the two trees. \subsection{Related works}\label{sec:related} Sorting with \emph{persistent errors} has been studied in several works, starting from \cite{Braverman2008} who presented the first algorithm achieving optimal dislocation (matching lower bounds appeared only recently in \cite{geissmann_et_al}). The algorithm in \cite{Braverman2008} uses only $O(n\log n)$ comparisons, but unfortunately its running time $O(n^{3+c})$ is quite large. For example, for a success probability of $1-1/n$, the analysis in \cite{Braverman2008} yields $c=\frac{110525}{(1/2-p)^4}$. On the contrary, all subsequent faster algorithms \cite{Klein2011,geissmann_et_al,newwindowsort} -- see Table~\ref{tb-recurrent} -- use a number of comparisons which is asymptotically equal to their respective running time. Other works considered error models in which repeating comparisons is expensive. For example, \cite{braverman2016parallel} studied algorithms which use a \emph{bounded number of rounds} for some ``easier'' versions of sorting (e.g., distinguishing the top $k$ elements from the others). In each round, a fresh set of comparison results is generated, and each round consists of $\delta \cdot n$ comparisons. They evaluate the algorithm's performance by estimating the number of ``misclassified'' elements and also consider a variant in which errors now correspond to missing comparison results In general, sorting in presence of errors seems to be computationally more difficult than the error-free counterpart. For instance, \cite{ajtai2016sorting} provides algorithms using \emph{subquadratic} time (and number of comparisons) when errors occur only between elements whose difference is at most some fixed threshold. Also, \cite{Damaschke16} gives a \emph{subquadratic} time algorithm when the number $k$ of errors is known in advance. As mentioned above, an easier error model is the one with \emph{non-persistent} errors, meaning that the same comparison can be \emph{repeated} and the errors are independent, and happen with some probability $p<1/2$. In this model it is possible to sort $n$ elements in time $O(n\log(n/q))$, where $1-q$ is the success probability of the algorithm \cite{Feige1994} (see also \cite{alonso,hadji} for the analysis of the classical Quicksort and recursive Mergesort algorithms in this error model). More generally, computing with errors is often considered in the framework of a two-person game called \emph{R\'{e}nyi-Ulam Game} (see e.g. the survey \cite{Pelc02} and the monograph \cite{Cicalese13}). \subsection{Paper Organization} The paper is organized as follows: in Section~\ref{sec:preliminaries} we give some preliminary definitions; then, in Section~\ref{sec:noisy_binary_search}, we present our noisy binary search algorithm, which will be used in Section~\ref{sec-optimal-sorting} to design a optimal randomized sorting algorithm. The proof of correctness of this algorithm will make use of an improved analysis of the sorting algorithm of \cite{geissmann_et_al}, which we discuss in Section~\ref{sec:windowsort}. Finally, in Section~\ref{sec:derand}, we briefly argue on how our sorting algorithm can be adapted so that it does not require any external source of randomness. Some proofs that only use arguments that are not related to the details of our algorithms are moved to the appendix. \section{Preliminaries} \label{sec:preliminaries} According to our error model, elements possess a true total linear order, however this order can only be observed through noisy comparisons. In the following, given two distinct elements $x$ and $y$, we will write $x \prec y$ (resp. $x \succ y$) to mean that $x$ is smaller (resp. larger) than $y$ according to the true order, and $x<y$ (resp. $x>y$) to mean that $x$ appears to be smaller (resp. larger) than $y$ according to the observed comparison result. Given a sequence or a set of elements $A$ and an element $x$ (not necessarily in $A$), we define $\rank(x, A) = |\{ y \in A : y \prec x\}|$ be as the \emph{true rank} of element $x$ in $A$ (notice that ranks start from $0$). Moreover, if $A$ is a sequence and $x \in A$, we denote by $\pos(x, A) \in [0, |S'|-1]$ the \emph{position} of $x$ in $A$ (notice that positions are also indexed from $0$), so that the \emph{dislocation} of $x$ in $A$ is $\disl(x,S) = |\pos(x,S)-\rank(x,S)|$, and the \emph{maximum dislocation} of the sequence $A$ is $\disl(S) = \max_{x \in S} \disl(x,S)$. \noindent For $z \in \mathbb{R}$, we write $\ln z$ and $\log z$ to refer to the natural and the binary logarithm of $z$, respectively. \section{Noisy Binary Search} \label{sec:noisy_binary_search} Given a sequence $S = \langle s_0, \dots, s_{n-1} \rangle$ of $n$ elements with maximum dislocation $d \ge \log n$, and element $x$ not in the sequence, we want to compute in time $O(\log n)$ an \emph{approximate rank} of $x$ in $S$, that is, a position where to insert $x$ in $S$ while preserving a $O(d)$ upper bound on dislocation of the resulting sequence. More precisely, we want to compute index $r_x$ such that $|r_x - \rank(x,S)|=O(d)$, in presence of persistent comparison errors: Errors between $x$ and the elements in $S$ happen independently with probability $p$, and whether the comparison $x$ between x and an element $y \in S$ is correct or erroneous does not depend on the position of $y$ in $S$, nor on the actual permutation of the sorted elements induced by their order in $S$ (i.e., we are not allowed to pick the order of the elements in $S$ as a function of the errors). We do not impose any restriction on the errors for comparisons that do not involve $x$. In the following, we will show an algorithm that computes such a rank $r_x$ in time $O(\log n)$. This immediately implies that $O(\log n)$ time also suffices to insert $x$ into $S$ so that the resulting sequence $\langle s_0, \dots, s_{r_x-1}, x, s_{r_x}, s_{n-1}\rangle$ still has maximum dislocation $O(d)$. \begin{remark} Notice that the $O(\log n)$ running time is asymptotically optimal for all $d=n^{1-\epsilon}$, for constant $\epsilon<1$, since a $\Omega(\log n - \log d) = \Omega(\log n)$ decision-tree lower bound holds even in absence of comparison errors. \end{remark} In the following, for the sake of simplicity, we let $c = 10^3$ and we assume that $n = 2 c d \cdot 2^h - 1$ for some non-negative integer $h$. Moreover, we focus on $p \le \frac{1}{32}$ even though this restriction can be easily removed to handle all $p < \frac{1}{2}$, as we argue at the end of the section. We consider the set $\{0, \dots, n\}$ of the possible ranks of $x$ in $S$ and we subdivide them into $2 \cdot 2^h$ ordered \emph{groups} $g_0, g_1, \dots$ each containing $cd$ contiguous positions, namely, group $g_i$ contains positions $c i d$, \dots, $c (i+1) d -1$. Then, we further partition these $2 \cdot 2^h$ groups into two ordered sets $G_0$ and $G_1$, where $G_0$ contains the groups $g_i$ with even $i$ ($i \equiv 0 \pmod{2}$) and $G_1$ the groups $g_i$ with odd $i$ ($i \equiv 1 \pmod{2}$). Notice that $|G_0| =|G_1|= 2^h$. In the next section, for each $G_j$, we shall define a \emph{noisy binary search tree} $T_j$, which will be the main ingredient of our algorithm. \subsection{Constructing $T_0$ and $T_1$} Let us consider a fixed $j \in \{0,1\}$ and define $\eta = 2 \lceil \log n \rceil$. The tree $T_j$ comprises of a binary tree of height $h + \eta$ in which the first $h+1$ levels (i.e., those containing vertices at depths $0$ to $h$) are complete and the last $\eta $ levels consists of $2^h$ paths of $\eta$ vertices, each emanating from a distinct vertex on the $(h+1)$-th level. We index the leaves of the resulting tree from $0$ to $2^h-1$, we use $h(v)$ to denote the depth of vertex $v$ in $T_j$, and we refer to the vertices $v$ at depth $h(v) \ge h$ as \emph{path-vertices}. Each vertex $v$ of the tree is associated with one \emph{interval} $I(v)$, i.e., as a set of contiguous positions, as follows: for a leaf $v$ having index $i$, $I(v)$ consists of the positions in $g_{2i+j}$; for a non-leaf path-vertex $v$ having $u$ as its only child, we set $I(v)=I(u)$; finally, for an internal vertex $v$ having $u$ and $w$ as its left and right children, respectively, we define $I(v)$ as the interval containing all the positions between $\min I(u)$ and $\max I(w)$ (inclusive). \begin{figure} \centering \includegraphics[width=.9\textwidth]{tree.pdf} \caption{An example of the noisy tree $T_0$. On the left side the shared pointers $L(\cdot)$ and $R(\cdot)$ are shown. Notice how $L(r)$ (and, in general, all the $L(\cdot)$ pointers on the leftmost side of the tree) points to the special $-\infty$ element. Good vertices are shown in black while bad vertices are white. Notice that, since $i^* \in I(w)$, we have $T^* = T_0$ and hence all the depicted vertices are either good or bad.} \label{fig:noisy_tree} \end{figure} Moreover, each vertex $v$ of the tree has a reference to two \emph{shared pointers} $L(v)$ and $R(v)$ to positions in $\{0, \dots, n\} \setminus \bigcup_{g_i \in S_j} g_i$. Intuitively, $L(v)$ (resp. $R(v)$) will always point to positions of $S$ occupied by elements that are \emph{smaller} (resp. \emph{larger}) than all the elements $s_i$ with $i \in I(v)$. For each leaf $v$, let $L(v)$ initially point to $\min I(v) - d - 1$ and $R(v)$ initially point to $\max I(v) + d$. A non-leaf path-vertex $v$ shares both its pointers with the corresponding pointers of its only child, while a non-path vertex $v$ shares its left pointer $L(v)$ with the left pointer of its left child, and its right pointer $R(v)$ with the right pointer of its right child. See Figure~\ref{fig:noisy_tree} for an example. Notice that we sometimes allow $L(v)$ to point to negative positions and $R(v)$ to point to positions that are larger than $n-1$. In the following we consider all the elements $s_i$ with $i < 0$ (resp. $i \ge n)$ to be copies a special $-\infty$ (resp. $+\infty$) element such that $-\infty \prec x$ and $-\infty < x$ in every observed comparison (resp. $+\infty \succ x$ and $+\infty > x$). \subsection{Walking on $T_j$} The algorithm will perform a discrete-time random walk on each $T_j$. Before describing such a walk in more detail, it is useful to define the following operation: \begin{definition}[test operation]\label{def:test} A \emph{test} of an element $x$ with a vertex $v$ is performed by (i) comparing $x$ with the elements $s_{L(v)}$ and $s_{R(v)}$, (ii) decrementing $L(v)$ by $1$ and, (iii) incrementing $R(v)$ by $1$. The tests succeeds if the observed comparison results are $x > s_{L(v)}$ and $x < s_{R(v)}$, otherwise the test fails. \end{definition} The walk on $T_j$ proceeds as follows. At time $0$, i.e., before the first step, the \emph{current} vertex $v$ coincides with the root $r$ of $T_j$. Then, at each time step, we \emph{walk} from the current vertex $v$ to the next vertex as follows: \begin{enumerate} \item We test $x$ with all the children of $v$ and, if \emph{exactly one} of these tests succeeds, we\emph{ walk to the corresponding child}. \item Otherwise, if \emph{all the tests fail}, we\emph{ walk to the parent} of $v$, if it exists. \end{enumerate} In the remaining cases, we ``walk'' from $v$ to itself. \smallskip \noindent We fix an upper bound $\tau = 240 \lfloor \log n \rfloor$ on the total number of steps we perform. The walk stops as soon as one of the following two conditions is met: \begin{description} \item[Success:] The current vertex $v$ is a leaf of $T_j$, in which case we return $v$; \item[Timeout:] The $\tau$-th time step is completed and the success condition is not met. \end{description} \subsection{Analysis} Let $i^* = \rank(x, S)$, and let $T^*$ be the unique tree in $\{ T_0, T_1 \}$ such that $i^*$ belongs to the interval of a leaf in $T^*$, and let $T'$ be the other tree. \begin{definition}[good/bad vertex] We say that a vertex $v$ of $T^*$ is \emph{good} if $i^* \in I(v)$ and \emph{bad} if either $i^* < \min I(v) - cd$ or $i^* > \max I(v) + cd$. \end{definition} Notice that in $T^*$ all the vertices are either good or bad, that the intervals corresponding to vertices at the same depth in $T^*$ are pairwise disjoint, and that the set of good vertices is exactly a root-to-leaf path. Moreover, all path-vertices of $T'$ are either bad or neither good nor bad. In both $T'$ and $T^*$ all the children of a bad vertex are also bad. \begin{lemma} \label{lemma:test_good} A test on a good vertex succeeds with probability at least $1-2p$. \end{lemma} \begin{proof} Let $v$ be a good vertex. Since $i^* \in I(v)$ and the pointer $L(v)$ only gets decremented, we have $L(v) \le \min I(v) - d - 1 \le i^* -d - 1$. Since $S$ has dislocation at most $d$, $\rank(s_{L(v)}, S) \leq L(v) + d \le i^*-1$. This implies that $s_{L(v)} \prec x$ and hence the probability to observe $S_{L(v)} > x$ during the test is at most $p$. Similarly, $R(v)$ is only incremented during the walk and hence $R(v) \ge \max I(v) + d \ge i^* + d$. Since $S$ has dislocation at most $d$, this implies that $\rank(s_{R(v)}, S) \ge R(v)-d \ge i^*$ and, in turn, that $s_{R(v)} \succ x$. Therefore, $s_{R(v)} < x$ is observed with probability at most $p$. By the union bound, $s_L(v) < x < s_R(v)$ with probability at least $1-2p$. \end{proof} \begin{lemma} \label{lemma:test_bad} A test on a bad vertex succeeds with probability at most $p$. \end{lemma} \begin{proof} Let $v$ be a bad vertex and notice that at most $\tau - 1 \le 240 \log n - 1 \le 240 d - 1 < (c-2)d - 1$ tests have been performed before the current test. We have that either $i^* < \min I(v) - cd$ or $i^* > \max I(v) + cd$. In the former case, we have $L(v) \ge \min I(v) - d - 1 - (\tau -1) \ge \min I(v) - d - (c-2)d > (i^* + cd) - (c-1)d = i^* + d$. Since $S$ has dislocation at most $d$, this implies $\rank(s_{L(x)}, S) \geq L(x) -d \ge i^*$. Therefore, $s_{L(x)} \succ x$ and thus $s_{L(x)} > x$ is also observed with probability at least $1-p$, causing the test to fail. Similarly, in the latter case, we have $R(v) \le \max I(v) + d + (\tau-1) \le \max I(v) + d + (c-2)d - 1 < (i^* - cd) + (c-1)d - 1 = i^* - d - 1$. Therefore, $\rank(s_{R(v)}, S) \le i^* - 1$, implying $s_{R(v)} \prec x$, and hence $s_{R(v)} < x$ is observed with probability at least $1-p$, causing the test to fail. \end{proof} \noindent We say that a step on $T_j$ from vertex $v$ to vertex $u$ is \emph{improving} iff either: \begin{itemize} \item $u$ is a good vertex and $h(u) > h(v)$ (this implies that $v$ is also good); or \item $v$ is a bad vertex and $h(u) < h(v)$. \end{itemize} Intuitively, each improving step is making progress towards identifying the interval containing the true rank $i^*$ of $x$, while each non-improving step undoes the progress of at most one improving step. \begin{lemma} \label{lemma:improving_step} Each step performed during the walk on $T^*$ is improving with probability at least $1-3p$. \end{lemma} \begin{proof} Consider a generic step performed during the walk on $T^*$ from a vertex $v$. If $v$ is a good vertex, then exactly one child $u$ of $v$ in $T^*$ is good (notice that $v$ cannot be a leaf). By Lemma~\ref{lemma:test_good}, the test on $u$ succeeds with probability at least $1-2p$. Moreover, there can be at most one other child $w \neq u$ of $v$ in $T^*$. If $w$ exists, then it must be bad and, by Lemma~\ref{lemma:test_bad}, the test on $w$ fails with probability at least $1-p$. By using the union bound on the complementary probabilities, we have that process walks from $v$ to $u$ with probability at least $1-3p$. If $v$ is a bad vertex, then all of its children are also bad. Since $v$ has at most $2$ children and, by Lemma~\ref{lemma:test_bad}, a test on a bad vertex fails with probability at least $1-p$, we have that all the tests fail with probability at least $1-2p$. In this case the process walks from $v$ to the parent of $v$ (notice that, since $v$ is bad, it cannot be the root of $T^*$). \end{proof} Since the walk on $T_j$ takes at most $\tau$ steps, any two distinct shared pointers $L(v)$ and $R(u)$ initially differ by at least $cd - 2d - 1 > (c-3)d$ positions, each step increases/decreases at most $4$ pointers (as it performs at most $2$ tests), and $\tau = 240 \lfloor \log n \rfloor < \frac{c-3}{4} d$, we can conclude that no two distinct pointers will ever point to the same position. This, in turn, implies: \begin{observation} \label{obs:independent_comparisons} Element $x$ is compared to each element in $S$ at most once. \end{observation} The following lemmas show that the walk on $T^*$ is likely to return a vertex corresponding to a good interval, while the walk on $T' \neq T^*$ is likely to either timeout or to return a non-bad vertex, i.e., a vertex whose corresponding interval contains positions that are close to the true rank of $x$ in $S$. \begin{lemma} \label{lemma:timeout} The walk on $T^*$ timeouts with probability at most $e \cdot n^{-6}$. \end{lemma} \begin{proof} For $t=1,\dots, \tau$, let $X_t$ be an indicator random variable that is equal to $1$ iff the $t$-th step of the walk on $T^*$ is improving. If the $t$-th step is not performed then let $X_t = 1$. Notice that if, at any time $t'$ during the walk, the number $X^{(t')} = \sum_{t=1}^{t'} X_t$ of improving steps exceeds the number of non-improving steps by at least $h + \eta$, then the success condition is met. This means that a necessary condition for the walk to timeout is $X^{(\tau)} - (\tau - X^{(\tau)}) < h + \eta$, which is equivalent to $X^{(\tau)} < \frac{h + \eta + \tau}{2}$. By Observation~\ref{obs:independent_comparisons} and by Lemma~\ref{lemma:improving_step}, we know that each $X_t$s corresponding to a performed steps satisfies $P(X_t = 1) \ge 1-3p > 1 - \frac{1}{10}$ (since $p \le \frac{1}{32}$), regardless of whether the other steps are improving. We can therefore consider the following experiment: at every time step $t=1,\dots, \tau$ we flip a coin that is heads with probability $q = \frac{9}{10}$, we let $Y_t = 1$ if this happens and $Y_t = 0$ otherwise. Clearly, the probability that $X^{(\tau)} < \frac{h + \eta + \tau}{2}$ is at most the probability that $Y^{(\tau)} = \sum_{t=1}^{\tau} Y_t < \frac{h + \eta + \tau}{2}$. By noticing that $\tau > 3 (h + \eta )$, and by using the Chernoff bound: \begin{multline*} \Pr\left(X^{(\tau)} < \frac{h + \eta + \tau}{2}\right) \le \Pr\left(Y^{(\tau)} < \frac{h + \eta + \tau}{2}\right) \le \Pr\left(Y^{(\tau)} < \frac{2\tau}{3}\right) = \Pr\left(Y < \frac{2 \mathbb{E}[Y]}{3q} \right) \\ < \Pr\left(Y < \left(1- \frac{1}{4}\right) \mathbb{E}[Y] \right) \le e^{-\frac{\tau q}{32}} < e^{-\frac{\tau}{40}} \le e^{1-6 \log n} < e \cdot n^{-6}. \end{multline*} \end{proof} \begin{lemma} \label{lemma:bad_vertex_returned} A walk on $T_j$ returns a bad vertex with probability at most $240 n^{-7}$. \end{lemma} \begin{proof} Notice that, in order to return a bad vertex $v$, the walk must first reach a vertex $u$ at depth $h(u)=h$, and then traverse the $\eta$ vertices of the path rooted in $u$ having $v$ as its other endpoint. We now bound the probability that, once the walk reaches $u$, it will also reach $v$ before walking back to the parent of $u$. Notice that all the vertices in the path from $u$ to $v$ are associated to the same interval, and hence they are all bad. Since a test on a bad vertex succeeds with probability at most $p$ (see Lemma~\ref{lemma:test_bad}) and tests are independent (see Observation~\ref{obs:independent_comparisons}), the sought probability can be upper-bounded by considering a random walk on $\{0 ,\dots, \eta+1 \}$ that: (i) starts from $1$, (ii) has one absorbing barrier on $0$ and another on $\eta+1$, and (iii) for any state $i \in [1, \eta]$ has a probability of transitioning to state $i+1$ of $p$ and to state $i-1$ of $1-p$. Here state $0$ corresponds to the parent of $u$, and state $i$ for $i>0$ corresponds to the vertex of the $u$--$v$ path at depth $h+i-1$ (so that state $1$ corresponds to $u$ and state $\eta+1$ corresponds to $v$). The probability of reaching $v$ in $\tau$ steps is at most the probability of being absorbed in $\eta$ (in any number of steps), which is (see, e.g., \cite[pp.~344--346]{feller1957introduction}): \[ \frac{ \frac{1-p}{p} - 1 }{ (\frac{1-p}{p})^{\eta+1} - 1 } \le \left( \frac{p}{1-p} \right)^\eta \le \left( \frac{1}{31} \right)^{2\log n} < \left( \frac{1}{2^4} \right)^{2\log n} = 2^{-8 \log n} = n^{-8}, \] where we used the fact that $p \le \frac{1}{32}$. Since the walk on $T_j$ can reach a vertex at depth $h$ at most $\tau$ times, by the union bound we have that the probability of returning a bad interval is at most $\tau n^{-8} \le 240 n^{-8} \log n \le 240 n^{-7}$. \end{proof} \noindent We are now ready to prove the main result of this section. \begin{theorem} \label{thm:noisy_binary_search} Let $S$ be an sequence of $n$ elements having maximum dislocation at most $d \ge \log n$ and let $x \not\in S$. Under our error model, an index $r_x$ such that $r_x \in [ \rank(x,S) - \alpha d, \rank(x,S) + \alpha d]$ can be found in $O(\log n)$ time with probability at least $1- O(n^{-6})$, where $\alpha > 1$ is an absolute constant. \end{theorem} \begin{proof} We compute the index $r_x$ by performing two random walks on $T_0$ and on $T_1$, respectively. If any of the walks returns a vertex $v$, then we return any position in the interval $I(v)$ associated with $v$. If both walks timeout we return an arbitrary position. From Lemma~\ref{lemma:timeout} the probability that both walks timeout is at most $\frac{e}{n^6}$ (as the walk on $T^*$ timeouts with at most this probability). Moreover, by Lemma~\ref{lemma:bad_vertex_returned}, the probability that at least one of the two walks returns a bad vertex is at most $\frac{480}{n^7}$. By the union bound, we have that with probability at most $1-\frac{480}{n^7} - \frac{e}{n^6} = 1 - O(\frac{1}{n^6})$, vertex $v$ exists and it is not bad. In this case, using the fact that $r_x \in [\min I(v), \max I(v)]$ and that $\max I(v) - \min I(v) < cd$, we have: \[ i^* \ge \min I(v) - cd > \max I(v) - 2cd \ge r_x - 2cd, \] and \[ i^* \le \max I(v) + cd < \min I(v) + 2cd \le r_x + 2cd. \] To conclude the proof it suffices to notice that the random walk requires at most $\tau = O(\log n)$ steps, that each step requires constant time, and that it is not necessary to explicitly construct $T_0$ and $T_1$ beforehand. Instead, it suffices to maintain a partial tree consisting of all the vertices visited by the random walk so far: vertices (and the corresponding pointers) are \emph{lazily} created and appended to the existing tree whenever the walk visits them for the first time. \end{proof} To conclude this section, we remark that our assumption that $p \le \frac{1}{32}$ can be easily relaxed to handle any constant error probability $p < \frac{1}{2}$. This can be done by modifying the test operation so that, when $x$ is tested with a vertex $v$, the majority result of the comparisons between $x$ and the set $\{ s_{L(v)}, s_{L(v)-1}, \dots, s_{L(v)-k+1} \}$ (resp. $x$ and the set $\{ s_{R(v)}, s_{R(v)+1}, \dots, s_{R(v)+k-1} \}$) of $\eta$ elements is considered, where $k$ is a constant that only depends on $p$. Consistently, the pointers $L(v)$ and $R(v)$ are shifted by $k$ positions, and the group size is increased to $k \cdot c$ to ensure that Observation~\ref{obs:independent_comparisons} still holds. Notice how our description for $p \le \frac{1}{32}$ corresponds exactly to the case $k=1$. The only difference in the statement Theorem~\ref{thm:noisy_binary_search} is that $\alpha$ is no longer an absolute constant, rather, it depends (only) on the value of $p$. \input{Sorting.tex} \section{WindowSort}\label{sec:windowsort} To make our algorithm description self-contained, and in order to prove Theorem~\ref{thm:windowsort} (thus strengtening the result of\cite{geissmann_et_al}), Algorithm~\ref{alg:windowsort} reproduces (a slightly modified version of) the pseudocode of \texttt{WindowSort}\xspace. \texttt{WindowSort}\xspace receives in input a sequence $S$ of $n$ elements, and an additional upper bound $d$ on the initial dislocation of $S$. The original \texttt{WindowSort}\xspace algorithm corresponds to the case $d=n$. \begin{algorithm}[t] $S_{2d} \gets S$\; \ForEach{$w=2d, d, d/2, \dots, 2$} { Let $\langle x_0, \dots, x_{n-1} \rangle$ be the elements in $S_{w}$\; \ForEach{$x_i \in S_w$} { $\score_w(x_i) \gets \max\{0, i - w\} + \{ x_j < x_i \, : \, |j-i| \le w \}$ } $S_{ w/2 } \gets $ sort the element of $S_{w}$ by non-decreasing value of $\score_w( \cdot )$\; } \Return $S_1$ \caption{\texttt{WindowSort}\xspace\unskip($S,d$)} \label{alg:windowsort} \end{algorithm} \texttt{WindowSort}\xspace iteratively computes a collection $\{S_{2d},S_{d}, S_{d/2},\dots, S_1\}$ of permutations of $S$: at every step, \texttt{WindowSort}\xspace maintains a \emph{window size} $w$ and builds $S_{w}$ from $S_{2w}$. Intuitively, for the algorithm to be successful, we would like $S_w$ to be a permutation of $S$ having maximum dislocation at most $w/2$, w.h.p. Even though this is true in the beginning (since we initially set $w=2d$), this property only holds up to a certain window size $w^*=\Theta(\log n)$. Nevertheless, it is still possible to show that the maximum dislocation of the returned sequence is $\Theta(\log n)$ and that the expected dislocation of each element is constant. We summarize the above discussion in the following lemma, which follows from the same arguments used in the proofs of Theorems~9 and 14 in \cite{geissmann_et_al}: \begin{lemma} \label{lemma:windowsort} Let $S$ be a sequence of $n$ element having maximum dislocation at most $d$. Then, with probability $1-\frac{1}{n^5}$, the following properties hold: (i) there exists a window size $w^* = \Theta(\log n)$ such that $\disl(S_{w^*})=O(\log n)$; and (ii) $\mathbb{E}[\disl(x,S_1)] = O(1)$. All the hidden constants depend only on $p$. \end{lemma} In the following, we shall prove that the total dislocation of the sequence returned by \texttt{WindowSort}\xspace is linear with high probability. We start by providing an upper bound to the change in position of an element between different iterations of \texttt{WindowSort}\xspace. This will also immediately imply that an element can only move by at most $O(w)$ positions between $S_{w^*}$ and $S_1$. \begin{lemma}\label{lem:move} For every $x \in S$, $| pos(x, S_{w}) - pos(x, S_{w'}) | \le 4 |w-w'|$. \end{lemma} \begin{proof} Without loss of generality let $w' < w$ (the case $w'>w$ is symmetric, and the case $w'=w$ is trivial). Consider a generic iteration of \texttt{WindowSort}\xspace corresponding to window size $w'' < w$. Let $i=\pos(x, S_{w})$ and $\Delta_{w''} = |i - pos(x, S_{w''/2})|$. \noindent For every element $y$ such that $pos(y, S_{w''}) < i-2w''$ we have: \[ \score_{w''}(x) \ge i - w'' > \pos(y, S_{w''}) + w'' \ge \score_{w''}(y), \] implying that $pos(x, S_{w''/2}) \ge i-2w''$. Similarly, for every element $y \in S$ such that $pos(y, S_{w''}) > i + 2w''$: \[ \score_{w''}(x) \le i+w'' < \pos(y, S_{w''})-w'' \le \score_{w''}(y), \] showing that the position of $x$ in $S_{w''/2}$ is at most $i+2w''$. We conclude that $\Delta_{w''} \le 2w''$, which allows us to write: \begin{multline*} |i - pos(x_i, S_{w'}) | = \Delta_{w} + \Delta_{w/2} + \Delta_{w/4} + \dots + \Delta_{2w'} \\ \le 2w + w + w/2 + \dots + 4w' = 4w - (2w' + w' + w'/2 + \dots) = 4w - 4w'. \end{multline*} \end{proof} The previous lemma also allows us to show that, once the window size $w$ is sufficiently small, the final position of an element only depends on a small subset of nearby elements. This, in turn, will imply that, once $S_w$ is fixed, the final positions of distant elements in $S_1$ are conditionally independent. The above property is formally shown in the following: \begin{lemma}\label{lem:dependent} Let $x \in S$. Given $S_{w}$, $pos(x, S_1)$ only depends on comparisons involving elements in positions $pos(x, S_1)-6w, \dots, pos(x, S_1)+6w$ in $S_{w}$. \end{lemma} \begin{proof} Let $w' = w/2^t$, for $t=1, 2, \dots$, and let $S_w'=\langle x_0,\dots,x_{n-1}\rangle$. We prove by induction on $t$, that $i=pos(x_i,S_{w'})$ only depends on the comparison results with elements in positions $i-r_t, \dots, i+r_t$ in $S_w$, where $r_t = 6 w (1 - 2^{-t} )$. Base case $t=1$: By Lemma~\ref{lem:move} we have that $ i - 2w \le pos(s_i, S_{w}) \le i + 2w$. Therefore, $x_i$ is can be only compared to elements $x_j$ such that $i - 3w \le pos(x_j, S_{w}) \le i + 3w$ Suppose now that the claim is true for some $t \ge 1$. We show that the claim also holds for $t+1$. Indeed, by Lemma~\ref{lem:move}, we have that $ i - 2w/2^t \le pos(x_i, S_{w/2^t}) \le i + 2w/2^t$ and hence it is only compared to $x_j$s such that $i - 3w/2^t \le pos(x_j, S_{w/2^t}) \le i + 3 w/2^t$. By induction hypothesis, these elements depend only on elements in positions $3w/2^t + 6w - 6w 2^{-t} = 6w - 3w/2^t = 6w (1-2^{-(t+1)})$. \end{proof} \noindent We are finally ready to prove Theorem~\ref{thm:windowsort} used in section \ref{sec:rifflesort}, that we restate here for convenience. \windowsortthm* \begin{proof} First of all notice that Lemma~\ref{lem:move}, ensures that the final dislocation of each element in $S_1$ will be at most $d + 4w$ where $w=2d$ is the initial window size of \texttt{WindowSort}\xspace, thus implying the $c_p \cdot d$ bound on the final maximum dislocation. We will condition on the event that for some $w^* = \Theta( \log n)$, $S_{w^*}$ has maximum dislocation $\delta = O(\log n)$. Let $G$ be the indicator random variable that describes this event, i.e., $G=1$ if the event happens. Clearly, by Lemma~\ref{lemma:windowsort}, we have that $P(G=1) \ge 1 - \frac{1}{n^5}$, implying that the final maximum dislocation will be at most $\delta + 4 w^* = O(\log n)$ with the same probability. Therefore, we now only focus on bounding the total dislocation of $S_1$. Let $S_1= \langle x_0,\dots,x_{n-1}\rangle $. and observe that $\mathbb{E}[\disl(x_i,S_1)\mid G=1] = O(1)$. Indeed, by Lemma~\ref{lemma:windowsort}, $\mathbb{E}[\disl(x_i,S_1)] \le c$ for all $x_i \in S$ and some constant $c>0$ depending only on $p$. Therefore: \begin{align*} c & \ge \mathbb{E}[\disl(x_i, S_1)] = \mathbb{E}[\disl(x_i,S_1)\mid G=1]\cdot P(G=1) + \mathbb{E}[\disl(x_i,S_1)\mid G=0]\cdot P(G=0)\\ &\ge \mathbb{E}[\disl(x_i,S_1)\mid G=1]\cdot \left(1-\frac{1}{n^5}\right), \end{align*} implying that $\mathbb{E}[\disl(x_i,S_1)\mid G=1] \le \frac{32}{31}c$ for all $n \ge 2$. We partition the elements of $S$ into $k = 20w^* + 4\delta$ sets, $P^{(0)},\dots,P^{(k-1)}$, such that $P^{(j)} = \{x \in S \, : \, \rank(x, S) = j \pmod{k} \}$. We now show that the final positions of two elements belonging to the same set are conditionally independent on $G=1$. Indeed, assuming $G=1$ and using Lemma~\ref{lem:move}, we have: \[ |\pos(x, S_1) - \rank(x, S)| \le |\pos(x, S_1) - \pos(x, S_{w^*})| + |\pos(x, S_{w^*}) - \rank(x, S)| < 4{w^*} + \delta, \] and, using again that $G=1$ together with Lemma~\ref{lem:dependent}, we know that $\pos(x, S_1)$ only depends on comparisons between elements in $\{ y \in S \, : \, | \pos(x, S_1) - \pos(y, S_{w^*})| \le 6{w^*} \}$ in $S_{w^*}$ which, in turn, is a subset of $\{ y \in S \, : \, |pos(x, S_1) - \rank(y, S) | \le 6{w^*} + \delta \}$. Combining the previous properties, we have that $pos(x, S_1)$ only depends on $\{ y \in S \, : \, |\rank(x, S) - \rank(y, S) | < 10{w^*} + 2\delta \}$, implying that two elements belonging to the same set depend on different comparisons results. Observe now that each set has size at least $\lfloor \frac{n}{k} \rfloor $ and at most $\lceil \frac{n}{k} \rceil = O(\frac{n}{\log n})$, define $D^{(j)} = \sum_{x_i\in P^{(j)}} \disl(x_i,S_1)$ to be the total dislocation of all elements in $P^{(j)}$, and let $\mu^{(j)} = \mathbb{E}[{D^{(j)}\mid G=1}] = \sum_{x \in P^{(j)}} \mathbb{E}[\disl(x, S_1) \mid G=1] \le \frac{32 c n}{31 k}$. We will use Hoeffding's inequality to prove that $D^{(j)}\le \frac{2 c n}{k}$ with high probability. Hoeffding's inequality is as follows: for independent random variables $X_1,\dots,X_n$, such that $X_i$ is in $[a_i,b_i]$, and $X = \sum_i X_i$, \[ P(X - \mathbb{E}[X]\ge t) \le \exp\left(-\frac{2t^2}{\sum_{i=1}^{n}(b_i-a_i)^2}\right)\, . \] \noindent Hence, since each $\disl(x_i, S_1)$ is between $0$ and $\delta + 4w^* = O(\log n)$ when $G=1$, we have: \[ P\left(D^{(j)} - \mu^{(j)} \ge \frac{30 c n}{31 k} \, \Big| \, G=1 \right) \le \exp\left(-\frac{1800}{961} \cdot \frac{\frac{n^2}{c^2 \delta^2}}{ \lceil \frac{n}{k} \rceil (\delta+ 4w^*)^2}\right) = \exp \left(-\Omega\left(\frac{n}{\log^3 n}\right)\right). \] Finally, by using the union bound over all $k = \Theta(\log n)$ sets and on the event $G=0$, we get that $\disl(S_1) \ge 2 c n$ with probability at most $ O(\log n) \cdot \exp \left(-\Omega\left(\frac{n}{\log^3 n}\right)\right) + \frac{1}{n^5}$, which is at most $\frac{1}{n^4}$ for sufficiently large values of $n$. \end{proof} \section{Derandomization} \label{sec:derand} \subsection{Partitioning $S$} \label{sec:derand_partitioning} In order to run \texttt{RiffleSort}\xspace we need to partition the input sequence $S$ into a collection of random sets $T_0, T_1, \dots, T_k$ where $k= \frac{\log n}{2}$ and each $T_i$ contains $m = \sqrt{n} \cdot 2^{i-1}$ elements that are chosen uniformly at random from the $n - \sqrt{n} \sum_{j=i+1}^k 2^{i-1} = 2 m$ elements in $S \setminus \bigcup_{j=i+1}^k T_j$. Notice also that this is the only step in the algorithm that is randomized. To obtain a version of \texttt{RiffleSort}\xspace that does not require any external source of randomness, i.e., that depends only on the input sequence and on the comparison results, we will generate such a partition by exploiting the intrinsic random nature of the comparison results. We start by showing that, with probability at least $1-\frac{1}{n^3}$, the partition $T_0, \dots, T_k$ can be found in $O(n)$ time using only $O(n)$ random bits. To this aim it suffices to show that, given a set $A$ of $2N$ elements, a random set $B \subset A$ of $N$ elements can be found in $O(N)$ time using $O(N)$ random bits, with probability at least $1-\frac{1}{N^7}$. We construct $B$ as follows: \begin{itemize} \item For each element of $A$ perform a coin-flip. Let $C$ be the set of all the elements whose corresponding coin flip is ``heads''. \item If $|C|=N$, return $B = C$. Otherwise, if $|C|<N$, randomly select a set $D$ of $N-|C|$ elements from $A \setminus C$ and return $B = C \cup D$. Finally, if $|C|>N$, randomly select a set $D$ of $|C|-N$ elements from $C$ and return $B= C \setminus D$. \end{itemize} Standard techniques show that this method of selecting $B$ is unbiased, i.e., all the sets $B \subset A$ of $N$ elements are returned with equal probability. We therefore move the proof of the following lemma to the Appendix. \begin{lemma} \label{lemma:random_selection_unbiased} For any set $X \subset A$ of $N$ elements, $\Pr(B = X) =\binom{2N}{N}^{-1}$. \end{lemma} We now show an upper bound on the number of random bits required. \begin{lemma} For sufficiently large values of $N$, the number of random bits required to select $B$ u.a.r. is at most $2N$ with probability at least $1 - N^{-7}$. \end{lemma} \begin{proof} Clearly, $C$ can be built using $3N$ random bits. Let $m=|C|$, since $E[|C|]=N$, by Hoeffding's inequality we have: \[ \Pr(| m - N | \ge 2 \sqrt{N \ln N} ) \le 2 e^{-8 \ln N} \le 2N^{-8}. \] This implies that, with probability at least $1-2N^{-8}$, the set $C$ contains at most $2 \sqrt{N \ln N}$ elements. Hence, $O(\sqrt{N} \cdot \mathrm{polylog} N)$ random bits suffice to select $D$ (using, e.g., a simple rejection strategy), and therefore $B$, with probability at least $1-N^{-8}$. The claim follows by using the union bound. \end{proof} From the above lemmas, it immediately follows that all the sets $T_0, \dots, T_k$ can be constructed in time $O(\sum_{i=0}^k |T_i|) = O(n)$ using at most $4n$ random bits with probability at least $1- n^{-\frac{7}{2}} \cdot \log N \ge 1 - n^{-3}$, for sufficiently large values of $n$. \subsection{Derandomized RiffleSort} As shown in \cite{newwindowsort}, it is possible to simulate ``almost-fair'' coin flips by xor-ing together a sufficiently large number of comparison results. Indeed, we can associate the two possible results of a comparison with the values $0$ and $1$, so that each comparison behaves as a Bernoulli random variable whose parameter is either $p$ or $1-p$. We can then use the following fact: let $c_1, \dots, c_k$ be $k = \Theta(\log n)$ independent Bernoulli random variables such that $P(c_i=1) \in \{p, 1-p\} \; \forall i=1,\dots,k$, then $| P(c_1 \oplus c_2 \oplus \dots \oplus c_k = 0) - \frac{1}{2}| \le \frac{1}{n^4}$. Therefore, if we consider the set $A$ containing the first $9 k$ elements from $S$ and we compare each element in $A$ to all the elements in $S \setminus A$, we obtain a collection of $9 k (n - k) \ge 8 k n$ comparison results (for sufficiently large values of $n$) from which we can generate $8n$ almost-fair coin flips. With probability at least $1 - \frac{8 k n }{n^4} - n^{-3} > 1 - \frac{1}{n^2}$ these almost-fair coin flips behave exactly as unbiased random bits, and they suffice to select a partition $T_0, \dots, T_k$ of $S \setminus A$.\footnote{This is true even if up to $|S \setminus A| - 1$ additional $+\infty$ elements are added to $S \setminus A$, as described in Section~\ref{sec:rifflesort}}. It is now possible to use \texttt{RiffleSort}\xspace on $S \setminus A$ to obtain a sequence $S'$ having maximum dislocation $d=O(\log n)$ and linear dislocation $O(n)$ (from Lemma~\ref{lemma:rifflesort_runtime} and Lemma~\ref{lemma:rifflesort_dislocation} this requires time $O(n \log n)$ and succeeds with probability at least $1- |S \setminus A|^{-\frac{3}{2}} > 1 - 3 n^{-\frac{3}{2}}$ since $|S \setminus A| \ge \frac{n}{2}$). What is left to do is to reinsert all the elements of $A$ into $S'$ without causing any asymptotic increase in the total and in the maximum dislocation. While one might be tempted to use the result of Section~\ref{sec-introduction}, this is not actually possible since the errors between the elements in $A$ and the elements in $S'$ now depend on the permutation $S'$. However, a simple (but slower) strategy, which is similar to the one used in \cite{newwindowsort}, works even when the sequence $S'$ is adversarially chosen as a function of the errors, as long as its maximum dislocation is at most $d$. Suppose that we have a guess $\tilde{r}$ on $\rank(x, S')$, we can determine whether $\tilde{r}$ is a good estimate on $\rank(x, S')$ by comparing $x$ with all the elements in positions from $\tilde{r} - cd$ to $\tilde{r} + cd -1$ in $S'$ and counting the number $m$ of \emph{mismatches}: a mismatch is an element $y$ such that either (i) $\pos(y, S') < \tilde{r}$ and $y > x$, or (ii) $\pos(y, S') \ge \tilde{r}$ and $y < x$. Suppose that our guess of $\tilde{r}$ is much smaller than the true rank of $x$, say $\tilde{r} < \rank(x, S') - c d$ for a sufficiently large constant $c$, then all the elements $y$ such that $\tilde{r} + d \le \rank(y, S') < \tilde{r} + (c-1)d$ are in $\{ z \in S' \, : \, \tilde{r} \le \pos(z, S') < \tilde{r} + cd \}$. Since $x \succ y$, we have that the observed comparison result is $x > y$ with probability at least $1-p$, and hence the expected number of mismatches $m$ is at least $(c-2)d(1-p)$, and a Chernoff bound can be used to show that with probability $1-\frac{1}{n^4}$, $m$ will exceed $\frac{1}{2}(c-2)d(1-p) \ge \frac{1}{3} c d$. A symmetric argument holds for the case in which $\tilde{r} \ge \rank(x, S') + c d$. On the contrary, if $\rank(x, S') - d \le \tilde{r} < \rank(x, S') + d$, all the elements $y$ such that either $\tilde{r} - (c-2) d \le \rank(y, S') < \tilde{r} -2d$ or $\tilde{r} + 2 d \le \rank(y, S') < \tilde{r} - (c-2) d$ are in the correct relative order w.r.t. $x$ in $S'$. This implies that the expected number of mismatches $m'$ between $x$ and all the elements $y$ will be at most $2(c-4)d p$, that $m' \le 4(c-4)dp$ with probability at least $1-\frac{1}{n^4}$, which implies that $m \le m' + 4d \le 4(c-4)d p + 4d < \frac{1}{3}c d$ with at least the same probability. Therefore, to compute a $r_x$ satisfying $|r_x - \rank(x, S')| = O(d)$, it suffices to count the number of mismatches for $\tilde{r} = 0, 2d, 4d, \dots $ and to select the value of $\tilde{r}$ minimizing their number. The total time required to to compute all $r_x$ for $x \in A$ is therefore $|A| \cdot O(\frac{n}{d} \cdot d) = O(n \log n)$, and the success probability is at least $1 - O(\frac{n}{d} \cdot |A|) \cdot \frac{1}{n^4} \ge 1 - \frac{1}{n^2}$, for sufficiently large values of $n$. Combining this with the success probability of \texttt{RiffleSort}\xspace, we obtain an overall success probability of at least $1 - \frac{1}{n}$. Finally, since the set $A$ only contains $O(\log n)$ elements, simultaneously reinserting them in $S'$ affects the maximum dislocation of $S'$ by at most an $O(\log n)$ additive term. Moreover, their combined contribution to the total dislocation is at most $O(\log^2 n)$. \clearpage \section{Optimal Sorting Algorithm}\label{sec-optimal-sorting} \subsection{The algorithm} \label{sec:rifflesort} We will present an optimal approximate sorting algorithm that, given a sequence $S$ of $n$ elements, computes, in $O(n\log n)$ worst-case time, a permutation of $S$ having maximum dislocation $O(\log n)$ and total dislocation $O(n)$, w.h.p. In order to avoid being distracted by roundings, we assume that $n$ is a power of $2$ (this assumption can be easily removed by padding the sequence $S$ with dummy $+\infty$ elements). Our algorithm will make use of the noisy binary search of Section~\ref{sec:noisy_binary_search} and of algorithm \texttt{WindowSort}\xspace presented in \cite{geissmann_et_al}. For our purposes, we need the following \emph{stronger} version of the original analysis in \cite{geissmann_et_al}, in which the bound on the total dislocation was only given in expectation: \begin{restatable}{theorem}{windowsortthm} \label{thm:windowsort} Consider a set of $n$ elements that are subject to random persistent comparison errors. For any dislocation $d$, and for any (adversarially chosen) permutation $S$ of these elements whose dislocation is at most $d$, $\texttt{WindowSort}\xspace(S,d)$ requires $O(n d)$ wost-case time and computes, with probability at least $1-\frac{1}{n^4}$, a permutation of $S$ having maximum dislocation at most $c_p \cdot \min\{d, \log n \}$ and total dislocation at most $c_p \cdot n$, where $c_p$ is a constant depending only on the error probability $p < \frac{1}{16}$. \end{restatable} We prove this theorem in Section~\ref{sec:windowsort}. Notice that \texttt{WindowSort}\xspace also works in a stronger error model in which the input permutation $S$ can be chosen adversarially after the comparison errors between all pairs of elements have been randomly fixed, as long as the total dislocation of $S$ is at most $d$. In the remaining of this section, we assume $p<1/32$ in order to be consistent with Section~\ref{sec:noisy_binary_search}, though both the above theorem and the algorithm we are going to present will only require $p<1/16$. Based on the binary search in Section~\ref{sec:noisy_binary_search}, we define an operation that allows us to add a linear number of elements to an almost-sorted sequence without any asymptotic increase in the resulting dislocation, as we will formally prove in the sequel. More precisely, if $A$ and $B$ are two disjoint subsets of $S$, we denote by $\texttt{Merge}\xspace(A,B)$ the sequence obtained as follows: \begin{itemize} \item For each element $x \in B$ compute and index $r_x$ such that $|\rank(s, A) - r_x| \le \alpha d$. This can be done using the noisy binary search of Section~\ref{sec:noisy_binary_search}, which succeeds with probability at least $1-\frac{1}{|A|^6}$. \item Insert \emph{simultaneously} all the elements $x \in B$ into $A$ in their computed positions $r_x$, breaking ties arbitrarily. Return the resulting sequence. \end{itemize} Our sorting algorithm, which we call \texttt{RiffleSort}\xspace (see the pseudocode in Algorithm~\ref{alg:rifflesort}), works as follows. For $k = \frac{\log n}{2}$, we first partition $S$ into $k+1$ subsets $T_0, T_1, \dots, T_{k}$ as follows. Each $T_i$, with $1 \leq i \leq k$, contains $2^{i-1} \sqrt{n}$ elements chosen uniformly at random from $S \setminus \{T_{i+1},T_{i+2},\ldots,T_k\}$, and $T_0 = S \setminus \{T_{1},T_{2},\ldots,T_k\}$ contains the leftover $n - \sqrt{n} \sum_{i=1}^k 2^{i-1} = \sqrt{n}$ elements. As its first step, \texttt{RiffleSort}\xspace will approximately sort $T_0$ using \texttt{WindowSort}\xspace, and then it will alternate merge operations with calls to \texttt{WindowSort}\xspace. While, the merge operations allow us to iteratively grow the set of approximately sorted elements to ultimately include all the elements in $S$, each operation also worsens the dislocation by a constant factor. This is a problem since the rate at which the dislocation increases is faster than the rate at which new elements are inserted. The role of the sorting operations is exactly to circumvent this issue: each \texttt{WindowSort}\xspace call locally rearranges the elements, so that all newly inserted elements are now closer to their intended position, resulting in a dislocation increase that is only an additive constant. The corresponding pseudocode is shown in Algorithm~\ref{alg:rifflesort}, in which $\gamma \ge \max\{202 \alpha, 909 \}$ is an absolute constant. \begin{algorithm}[t] $T_0, T_1, \dots, T_{k} \gets$ partition of $S$ computed as explained in Section~\ref{sec:rifflesort}\; $S_0 \gets \texttt{WindowSort}\xspace(T_0, \sqrt{n})$\; \ForEach{$i=1,\dots,k+1$} { $S_i \gets \texttt{Merge}\xspace(S_{i-1}, T_{i-1})$\; $S_i \gets \texttt{WindowSort}\xspace(S_i, \gamma \cdot c_p \cdot \log n)$\; } \Return $S_{k+1}$\; \caption{\texttt{RiffleSort}\xspace\unskip(S)} \label{alg:rifflesort} \end{algorithm} \subsection{Analysis} \begin{lemma} \label{lemma:rifflesort_runtime} The worst-case running time of Algorithm~\ref{alg:rifflesort} is $O(n \log n)$. \end{lemma} \begin{proof} Clearly the random partition $T_0, \dots, T_k$ can be computed in time $O(n \log n)$,\footnote{The exact complexity of this steps depends on whether we are allowed to generate a uniformly random integer in a range in $O(1)$ time. If this is not the case, then integers can be generated bit-by-bit using rejection. It is possible to show that the total number of required random bits will be at most $O(n)$ with probability at least $1-n^{-2}$ (see Section~\ref{sec:derand_partitioning}). To maintain a worst-case upper bound on the running time also in the unlikely event that $O(n \log n)$ bits do not suffice, we can simply stop the algorithm and return any arbitrary permutation of $S$. This will not affect the high-probability bounds in presented in the sequel.} and the first call to \texttt{WindowSort}\xspace requires time $O(|T_0| \cdot \sqrt{n}) = O(n)$ (see Theorem~\ref{thm:windowsort}). We can therefore restrict our attention to the generic $i$-th iteration of the for loop. The call to $\texttt{Merge}\xspace(S_{i-1}, T_{i-1})$ can be performed in $O(|S_i| \log n)$ time since, for each $x \in T_{i-1}$, the required approximation of $\rank(x, S_{i-1})$ can be computed in time $O(\log |T_{i-1}|)$ and $|T_{i-1}|< |S_i| \le n$, while inserting the elements in $S_{i-1}$ their computed ranks requires linear time in $|S_{i-1}| + |T_{i-1}| = |S_i|$. The subsequent execution of \texttt{WindowSort}\xspace with $d = O(\log n)$ requires time $O(|S_i| \log n)$, where the hidden constant does not depend on $i$. Therefore, for a suitable constant $c$, the time spent in the $i$-th iteration is $c |S_i| \log n$ and total running time of Algorithm~\ref{alg:rifflesort} can be upper bounded by: \[ c \sum_{i=1}^{k+1} |S_i| \log n = c \sqrt{n} \log n \cdot \sum_{i=1}^{k+1} 2^i < 2^{k+2} c \sqrt{n} \log n = 2 c n \log n. \] This completes the proof. \end{proof} The following lemma, that concerns a thought experiment involving urns and randomly drawn balls, is instrumental to bounding the dislocation of the sequences returned by the $\texttt{Merge}\xspace$ operations. Since it can be proved using arguments that do not depend on the details of \texttt{RiffleSort}\xspace, we postpone its proof to the appendix. \begin{lemma} \label{lemma:draw_no_long_monocolor} Consider an urn containing $N=2M$ balls, $M$ of which are white and the remaining $N$ are black. Balls are iteratively drawn from the urn without replacement until the urn is empty. If $N$ is sufficiently large and $ 9 \log N \le k \le \frac{N}{16}$ holds, the probability that any contiguous subsequence of at most $100k$ drawn balls contains $k$ or fewer white balls is at most $N^{-6}$. \end{lemma} We can now show that, if $A$ and $B$ contain randomly selected elements, the sequence returned by $\texttt{Merge}\xspace(A,B)$ is likely to have a dislocation that is at most a constant factor larger than the dislocation of $A$: \begin{lemma} \label{lemma:merge_constant_disl_increase} Let $A$ be a sequence containing $m$ randomly chosen elements from $S$ and having maximum dislocation at most $d$, with $\log n \le d = o(m)$. Let $B$ be a set of $m$ randomly chosen elements from $S \setminus A$. Then, for a suitable constant $\gamma$, and for large enough values of $m$, $merge(A,B)$ has maximum dislocation at most $\gamma d$ with probability at least $1-m^{-4}$. \end{lemma} \begin{proof} Let $\beta = \max\{\alpha, 9/2 \}$, $S'=\texttt{Merge}\xspace(A,B)$, and $S^* = \langle s_0^*, s_1^*, \dots, s_{2n-1}^* \rangle$ be the sequence obtained by sorting $S'$ according to the true order of its elements. Assume that: \begin{itemize} \item all the approximate ranks $r_x$, for $x \in B$, are such that $|r_x - \rank(x,A)| \le \beta d$; and \item all the contiguous subsequences of $S^*$ containing up to $2 \beta d+2$ elements in $A$ have length at most $200 \beta d + 200$. \end{itemize} We will show in the sequel that the above assumptions are likely to hold. Pick any element $x \in S'$. We will show that our assumptions imply that the dislocation of $x$ in $S'$ is at most $201d$. An element $y \in B$ can affect the final dislocation of $x$ in $S'$ only if one of the following two (mutually exclusive) conditions holds: (i) $y \prec x$ and $r_y \ge r_x$, or (ii) $y \succ x$ and $r_y \le r_x$. All the remaining elements in $B$ will be placed in the correct relative order w.r.t.\ $x$ in $S'$, and hence they do not affect the final dislocation of $x$. If (i) holds, we have: \[ r_x - \beta d \le r_y - \beta d \le \rank(y, A) \le \rank(x, A) \le r_x + \beta d, \] while, if (ii) holds, we have: \[ r_x -\beta d \le \rank(x, A) \le \rank(y, A) \le r_y + \beta d \le r_x + \beta d, \] and hence, all the elements $y \in B$ that can affect the dislocation of $x$ in $S'$ are contained in the set $Y = \{ y \in B : r_x - \beta d \le \rank(y, A) \le r_x + \beta d\}$. We now upper bound the cardinality of $Y$. Let $y^-$ be the $(r_x - \beta d - 1)$-th element of $A$; if no such element exists, then let $y^- = s^*_0$. Similarly, let $y^+$ be the $(r_x + \beta d)$-th element of $A$; if no such element exists, then let $y^+ = s^*_{2m-1}$. Due to our choice of $y^-$ and $y^+$ we have that $\forall y \in Y, y^- \preceq y \preceq y^+$, implying that all the elements in $Y$ appear in the contiguous subsequence $\overline{S}$ of $S^*$ having $y^-$ and $y^+$ as its endpoints. Since no more than $2 \beta d+2$ elements of $A$ belong to $\overline{S}$ , our assumption guarantees that $\overline{S}$ contains at most $200 \beta d+200$ elements. This implies that the dislocation of $x$ in $S'$ is at most $\beta d + |Y| \le \beta d + |\overline{S}| \le 201 \beta d + 200 \le \gamma d$, where the last inequality holds for large enough $n$ once we choose $\gamma = 202 \beta$. To conclude the proof we need to show that our assumptions holds with probability at least $1-|S'|^{-6}$. Regarding the first assumption, for $x\in B$, a noisy binary search returns a rank $r_x$ such that $|r_x - \rank(x,A)| \le \alpha d \le \beta d$ with probability at least $1 - O(\frac{1}{m^6})$. Therefore the probability that the assumption holds is at least $1-O(\frac{1}{m^5})$. Regarding our second assumption, notice that, since the elements in $A$ and $B$ are randomly selected from $S$, we can relate their distribution in $S^*$ with the distribution of the drawn balls in the urn experiment of Lemma~\ref{lemma:draw_no_long_monocolor}: the urn contains $N=2m$ balls each corresponding to an elements in $A \cup B$, a ball is white if it corresponds to one of the $M=m$ elements of $A$, while a black ball corresponds one of the $M=m$ elements of $B$. If the assumption does not hold, then there exists a contiguous subsequence of $S^*$ of at least $200 \beta d+200$ elements that contains at most $2 \beta d+2$ elements from $A$. By Lemma~\ref{lemma:draw_no_long_monocolor} with $k=2 \beta d+2$, this happens with probability at most $(2m)^{-6}$ (for sufficiently large values of $n$). The claim follows by using the union bound. \end{proof} We can now use Lemma~\ref{lemma:merge_constant_disl_increase} and Theorem~\ref{thm:windowsort} together to derive an upper bound to the final dislocation of the sequence returned by Algorithm~\ref{alg:rifflesort}. \begin{lemma} \label{lemma:rifflesort_dislocation} The sequence returned by Algorithm~\ref{alg:rifflesort} has maximum dislocation $O(\log n)$ and total dislocation $O(n)$ with probability $1- \frac{1}{n\sqrt{n}}$. \end{lemma} \begin{proof} For $i = 1, \dots, k+1$, we say that the $i$-th iteration of Algorithm~\ref{alg:rifflesort} is \emph{good} if the sequence $S_i$ computed at its end has both (i) maximum dislocation at most $c_p \log n$, and (ii) total dislocation at most $c_p |S_i|$. As a corner case, we say that iteration $0$ is good if the sequence $S_0$ also satisfies conditions (i) and (ii) above. We now focus on a generic iteration $i\ge 1$ and show that, assuming that iteration $i-1$ is good, iteration $i$ is also good with probability at least $1-\frac{1}{n^2}$. Since iteration $0$ is good with probability at least $1-\frac{1}{|S_0|^4}$ = $1- \frac{1}{n^2}$ and there are $k + 1 = O(\log n)$ other iterations, the claim will follow by using the union bound. Since iteration $i-1$ was good, the sequence $S_{i-1}$ has maximum dislocation $c_p \log n$ and hence the sequence resulting from call to $\texttt{Merge}\xspace(S_{i-1}, T_{i-1})$ returns a sequence with dislocation at most $\gamma c_p \log n$ with probability at least $1-\frac{1}{|T_{i-1}|^4} \ge 1- \frac{1}{n^2}$. If this is indeed the case, we have that the sequence $S_i$ returned by the subsequent call to \texttt{WindowSort}\xspace satisfies (i) and (ii) with probability at least $1 - \frac{1}{|S_{i+1}|^4} \ge 1 - \frac{1}{n^2}$. The claim follows by using the union bound and by noticing that the returned sequence is exactly $S_{k+1}$. \end{proof} We have therefore proved the following result, which follows directly from Lemma~\ref{lemma:rifflesort_dislocation} and Lemma~\ref{lemma:rifflesort_runtime}: \begin{theorem} \texttt{RiffleSort}\xspace is a randomized algorithm that approximately sorts, in $O(n \log n)$ worst-case time, $n$ elements subject to random persistent comparison errors so that the maximum (resp. total) dislocation of the resulting sequence is $O(\log n)$ (resp. $O(n)$), w.h.p. \end{theorem}

https://arxiv.org/abs/1412.6639

A geometric Hall-type theorem

We introduce a geometric generalization of Hall's marriage theorem. For any family $F = \{X_1, \dots, X_m\}$ of finite sets in $\mathbb{R}^d$, we give conditions under which it is possible to choose a point $x_i\in X_i$ for every $1\leq i \leq m$ in such a way that the points $\{x_1,...,x_m\}\subset \mathbb{R}^d$ are in general position. We give two proofs, one elementary proof requiring slightly stronger conditions, and one proof using topological techniques in the spirit of Aharoni and Haxell's celebrated generalization of Hall's theorem.

\section{Introduction} \subsection{Background} Let $F = \{S_1, \dots, S_m\}$ be a family of finite subsets of a common ground set $E$. A {\em system of distinct representatives} is an $m$-element subset $\{x_1, \dots, x_m\}\subset E$ such that $x_i \in S_i$ for all $1\leq i \leq m$. A classical result in combinatorics is {\em Hall's marriage theorem} \cite{hall} which states that a family $F = \{S_1, \dots, S_m\}$ has a system of distinct representatives if and only if $\left| \bigcup_{i\in I} S_i \right| \geq | I |$ for every non-empty subset $I\subset \{1, \dots, m\}$. In 2000, Aharoni and Haxell \cite{aharoni1} presented a remarkable generalization of Hall's theorem. Let $F = \{H_1, \dots, H_m\}$ be a family of hypergraphs on a common vertex set $V$. A {\em system of disjoint representatives} is an $m$-element set $\{E_1, \dots, E_m\}$ of pairwise (vertex) disjoint edges such that $E_i\in H_i$ for all $1\leq i \leq m$. The Aharoni and Haxell result gives a sufficient condition for a family of hypergraphs to have a system of disjoint representatives, and their result reduces to the Hall's theorem in the case when the $H_i$ are 1-uniform hypergraphs. Their result was used to prove Ryser's conjecture for $3$-uniform hypergraphs \cite{aharoni}, but perhaps more importantly, their proof introduced topological techniques into this classical branch of combinatorics. The connections with topological combinatorics were further investigated and generalized in \cite{aharoni2}, \cite{ah-be}, \cite{a-b-m}, \cite{chudn}, \cite{haxell}, \cite{kahle}, \cite{meshulam}, \cite{meshulam2}. \subsection{Our result} The purpose of this paper is to introduce a discrete geometric generalization of Hall's marriage theorem. We say that a subset $X\subset \mb{R}^d$ is in {\em general position} if every subset of size at most $d+1$ is affinely independent. Let $F = \{X_1, \dots, X_m\}$ be a family of finite sets in $\mb{R}^d$. A {\em system of general position representatives} is a subset $\{x_1, \dots, x_m\}$ in general position such that $x_i \in X_i$ for all $1\leq i \leq m$. For a finite set $X\subset \mb{R}^d$ let $\varphi(X)$ denote the maximal size of a subset of $X$ in general position. We have the following. \begin{theorem} \label{sgpr} For every integer $d\geq 1$ there exists a function $f_d : \mb{N}\to \mb{N}$ such that the following holds. Let $F = \{X_1, \dots, X_m\}$ be a family of finite sets in $\mb{R}^d$. If \[\varphi \left(\bigcup_{i\in I} X_i\right ) \geq f_d(|I|)\] for every non-empty subset $I\subset \{1, \dots, m\}$, then $F$ has a system of general position representatives. \end{theorem} Notice that for $d=1$, a set is in general position if its elements are pairwise distinct. Therefore we can set $f_1(k) = k$, in which case Theorem \ref{sgpr} reduces to Hall's theorem. Once the existence of the functions $f_d(k)$ has been established, a natural goal is to obtain good general upper bounds on these functions. In general we are interested in asymptotic bounds, that is, when $d$ is fixed and the number of sets in the family $F$ grows. Let us illustrate how the the size of $F = \{X_1, \dots, X_m\}$ plays a role. Suppose $m \leq d+1$. We claim that if $\varphi \left(\bigcup_{i\in I} X_i\right ) \geq |I|$ for every non-empty subset $I\subset \{1, \dots, m\}$, then $F$ has a system of general position representatives (which is the same condition as in Hall's theorem). This follows from the {\em matroid intersection theorem} due to Edmonds \cite{edmonds}. To see this, let the ground set be the disjoint union $E = X_1 \dot\cup \cdots \dot\cup X_m$. (We allow for the same point to appear in several $X_i$, but we keep track of its multiplicity.) Let $M_1$ be the matroid on $E$ whose independent sets are the affinely independent subsets, and let $M_2$ be the partition matroid induced by $X_1, \dots, X_m$. Let $r_1$ and $r_2$ be the respective rank functions. Given a subset $S\subset E$, let $I\subset \{1, \dots, m\}$ be the maximal subset such that $\bigcup_{i\in I} X_i \subset S$. We then have $r_1(S)\geq r_1\left(\bigcup_{i\in I}X_i \right)$ and $r_2(E - S) \geq m - |I|$. The matroid intersection theorem implies that $F$ has a system of general position representatives if $r_1(S) \geq |I|$ for every non-empty subset $S\subset E$. This inequality holds by our hypothesis since $r_1(S) = \min \{d+1, \varphi(S)\} \geq \min \{d+1, \varphi \left(\bigcup_{i\in I}X_i \right) \} \geq |I|$. It is also easily seen that when $m>d+1$, the condition $\varphi\left(\bigcup_{i\in I}X_i \right) \geq | I |$ is not sufficient to guarantee a system of general position representatives. Suppose $|X_1|$ $=\cdots$ $ = |X_{m-1}| =1$ and that $\bigcup_{i=1}^{m-1}X_i$ is in general position in $\mb{R}^d$. From every hyperplane spanned by a $d$-tuple from $\bigcup_{i=1}^{m-1}X_i$ choose an additional point, at random, to be included in the set $X_m$. Thus $X_m$ consists of $\binom{m-1}{d}$ points in general position. For every non-empty subset $I\subset \{1, \dots, m\}$ we have $\varphi\left(\bigcup_{i\in I}X_i \right) \geq | I |$, but $F$ has no system of general position representatives. \subsection{Outline of paper} We will present two proofs for the existence of the functions $f_d(k)$. The first proof uses an elementary pigeon-hole argument and gives an upper bound in $O(k^{d+1})$. This is given in Section \ref{elementary}. Our second proof uses more sophisticated techniques and gives an upper bound in $O(k^d)$. This is given in Section \ref{non-elem}, while the main auxiliary result (Theorem \ref{gencom}) is proved in Section \ref{dusty}. We do not know if this bound is optimal, and it is an interesting problem to determine better bounds on $f_d(k)$. The reader familiar with matroids will notice that many of our arguments rely on properties of the underlying matroid of the point set. This leads to generalizations of our results which will be discussed further in Section \ref{remarks}. (All matroids arising in our setting are loopless, so this will be implicitly assumed throughout.) Just as the seminal result of Aharoni and Haxell, our second proof of Theorem \ref{sgpr} relies on topological methods, and we assume the reader is familiar with some basic notions of combinatorial topology. By using a result of Kalai and Meshulam \cite[Proposition 3.1]{kalai-mesh} (also appearing implicitly in \cite{aharoni2}, \cite{aharoni1}, and \cite{meshulam}), Theorem \ref{sgpr} can be reduced to the problem of showing that a certain simplicial complex is highly connected. We remind the reader that a topological space $X$ is {\em $k$-connected } if every map $f \colon \mb{S}^i\to X$ extends to a map $\hat{f} \colon \mb{B}^{i+1} \to X$ for $i = -1, 0, 1, \dots, k$. Here $\mb{B}^{i+1}$ denotes the $(i+1)$-dimensional ball whose boundary is the $i$-dimensional sphere $\mb{S}^i$, and $(-1)$-connected means non-empty. The following observation is sufficient for our application: {\em A simplicial complex is $k$-connected if and only if its $(k+1)$-skeleton is $k$-connected.} Before getting to the details, let us conclude with a few words about the simplicial complex arising in our second proof of Theorem \ref{sgpr}. It was made explicit in \cite{meshulam}, that the key idea in the Aharoni and Haxell result is to capture {\em pairwise disjointness} among the members in a family of sets. This can be encoded by the {\em disjointness graph} of the family, and the resulting simplicial complex is the {\em clique complex} of the disjointness graph. However, the general position property is not a pairwise condition (for $d\geq 2$), and to encode the subsets in general position requires a simplicial complex, the {\em independence complex} of the underlying matroid of the point configuration. This in turn requires a higher-dimensional version of a clique complex, which we call the {\em completion}. A crucial observation concerning the completion of a complex is Lemma \ref{q-star}, which gives a local combinatorial condition on a simplicial complex which guarantees that its completion is $k$-connected. \section{Proof of Theorem \ref{sgpr}} \label{elementary} For positive integers $d$ and $k$ let \[A_d(k) \coloneqq \begin{cases} k &\mbox{if } k\leq d+1 \\ d\binom{k-1}{d}+1 & \mbox{if } k > d+1.\end{cases} \] Notice that $A_d(k)$ is in $O(k^d)$. \begin{lemma} \label{lemUp} Let $k$ be a positive integer. If $S$ and $T$ are sets in general position in $\:\mb{R}^d$ where $|S| = k-1$ and $|T|\geq A_d(k)$, then there exists a point $p$ in $T$ such that $S\cup \{p\}$ is in general position. \end{lemma} \begin{proof} For $k\leq d+1$ the result is a consequence of the augmentation property of the underlying matroid of a set of points in $\mathbb{R}^d$ (the independent sets are the affinely independent sets). Suppose now that $k\geq d+2$ and that $T$ is a set of points in general position with $|T| \geq A_d(k)$. Notice that $S$ spans $\binom{k-1}{d}$ affine hyperplanes. In each of these hyperplanes there can be at most $d$ points from $T$ since $T$ is in general position. Therefore there exists a point $p$ in $T$ which does not lie in any of these hyperplanes, implying that $S\cup \{p\}$ is in general position. \end{proof} Now let $B_d(k)=k(A_d(k)-1)+1.$ Notice that $B_d(k)$ is in $O(k^{d+1})$. \begin{theorem} Let $F=\{X_1, \ldots, X_m\}$ be a family of finite sets in $\mathbb{R}^d$. If \[\varphi\left(\bigcup_{i\in I} X_i\right)\geq B_d(|I|)\] for every non-empty subset $I\subset \{1,\ldots,m\}$, then $F$ has a system of general position representatives. \end{theorem} \begin{proof} By the hypothesis, $\varphi\left(\bigcup_{i=1}^m X_i\right) \geq m(A_d(m)-1) + 1$. By the pigeon-hole principle, there are at least $A_d(m)$ points in general position belonging to one of the sets $X_1, \dots, X_m$, so we may assume $\varphi(X_m)\geq A_d(m)$. Using the hypothesis for $I=\{1,\ldots,m-1\}$, the same reasoning implies that there are $A_d(m-1)$ points in general position belonging to one of the sets $X_1$, $\ldots$, $X_{m-1}$, so we may assume $\varphi(X_{m-1}) \geq A_d(m-1)$. Proceeding downwards we may assume that $\varphi(X_i) \geq A_d(i)$ for each $i\in \{1,\ldots,m\}$. Now we use Lemma \ref{lemUp} upwards. We take a point $p_1 \in X_1$. Suppose we have selected points $p_i\in X_i$ for $i\in \{1,2,\ldots,k-1\}$ such that $\{p_1,\ldots,p_{k-1}\}$ is in general position. Then Lemma \ref{lemUp} allows us to select a point $p_k \in X_k$ such that $\{p_1,\ldots,p_k\}$ is in general position. We continue up to $k=m$ to get the desired system of general position representatives. \end{proof} \section{A better upper bound by topological methods}\label{non-elem} \subsection{The general position complex} Let $X\subset \mb{R}^d$ be a finite (multi)set. Let us define the {\em general position complex} of $X$, denoted by $G(X)$, to be the simplicial complex \[G(X) \coloneqq \{S \subset X \; : \; S \mbox{ is in general position in } \mb{R}^d \}.\] Note that we allow for $X$ to have repeated points. The number of vertices of $G(X)$ is the cardinality of $X$, counting multiplicities. A key observation is that the connectivity of $G(X)$ can be bounded below in terms of $d$ and $\varphi(X)$. \begin{theorem} \label{gencom} For all integers $d\geq 1$ and $k\geq -1$ there exists a minimal positive integer $g_d(k)$ such that the following holds. If $X\subset \mb{R}^d$ is a finite (multi)set with $\varphi(X) \geq g_d(k)$, then $G(X)$ is $k$-connected. \end{theorem} A closely related simplicial complex is the {\em independence complex} of $X$, denoted by $M(X)$, defined as \[M(X) \coloneqq \{S \subset X \: : \: S \mbox{ is affinely independent}\}.\] The simplices of $M(X)$ are the independent sets of a matroid, the underlying matroid of $X$, which has rank $r = \min \{\varphi(X) , d+1\} = \dim M(X) + 1$. Note that $M(X)$ is the $(r-1)$-skeleton of $G(X)$. \begin{remark} \label{2nd remark} We postpone the proof of Theorem \ref{gencom}, but here we note the following special cases. \begin{itemize} \item For $d = 1$, a multiset $X$ consists of $n = \varphi(X)$ distinct points with mutliplicities $m_1, \dots, m_n$. The corresponding general position complex is the join of $n$ discrete sets of points. That is, $G(X) = V_1 \ast \cdots \ast V_n$, where $|V_i| = m_i$. If $|V_i|=1$ for any $i$, then $G(X)$ is contractible. If $|V_i| > 1$ for all $i$ it is known that $G(X)$ is homotopic to a wedge of $(n-1)$-dimensional spheres which is $(n-2)$-connected. Therefore $g_1(k) = k+2$. \item If $k \leq d-1$, then $g_d(k) = k+2$. In this case $G(X) = M(X)$, and the claim follows from the well-known fact that the independence complex of a rank $r$ matroid is $(r-2)$-connected (see e.g. \cite{bjorn, BjKoLo}). \end{itemize} \end{remark} \subsection{Colorful simplices} Let $K$ be a simplicial complex on the vertex set $V$, and let $V = V_1 \cup \cdots \cup V_m$ be a partition. A simplex $S\in K$ is called {\em colorful} if $|S\cap V_i| = 1$ for all $1\leq i \leq m$. For a non-empty subset $I \subset \{1, \dots, m\}$ let $K(I)$ denote the {\em induced subcomplex} $K\left[\bigcup_{i\in I}V_i\right]$. The following sufficient condition for the existence of a colorful simplex in $K$ was given in \cite[Proposition 3.1]{kalai-mesh} (where it is stated in terms of rational homology rather than connectedness), and in a more general form in \cite[Theorem 4.5]{aharoni2}. \begin{prop}[Kalai and Meshulam] \label{colorful} Let $K$ be a simplicial complex on the vertex set $V$ with partition $V = V_1 \cup \cdots \cup V_m$. If the induced subcomplex $K(I)$ is $(|I|-2)$-connected for every non-empty subset $I\subset\{1, \dots, m\}$, then $K$ contains a colorful simplex. \end{prop} \begin{proof}[Second proof of Theorem \ref{sgpr}] Let $F = \{X_1, \dots, X_m\}$ be a family of finite sets in $\mb{R}^d$ and let $X = X_1 \dot\cup \cdots \dot\cup X_m$ (that is, counting multiplicities). The members of $F$ induce a partition of the vertex set of $G(X)$, and $F$ has a system of general position representatives if and only if the general position complex $G(X)$ contains a colorful simplex. If $\varphi\left( \bigcup_{i\in I} X_i\right) \geq g_d(|I|-2)$ for every non-empty subset $I\subset \{1, \dots, m\}$, then, by Theorem \ref{gencom}, $G(X)$ satisfies the conditions of Proposition \ref{colorful}. Therefore $f_d(k)$ can be bounded above by $g_d(k-2)$. \end{proof} \begin{remark} In the next section we give an upper bound on $g_d(k)$ which is in $O(k^d)$. \end{remark} \section{Proof of Theorem \ref{gencom}}\label{dusty} \subsection{The completion of a simplicial complex} Let $k$ be a positive integer and $S$ a finite set with $|S| \geq k$. The collection of all subsets of $S$ of size at most $k$ is denoted by \[[S]_{k} \coloneqq \{T\subset S \: : \: |T|\leq k\}.\] Let us also define $[S]_0 \coloneqq \emptyset$. Let $K$ be a simplicial complex of dimension $d$ on the vertex set $V$. For the proof of Theorem \ref{gencom} we need the following simplicial complexes associated with $K$. \bigskip For a vertex $v\in V$, let $\mbox{st}_K(v)$ denote the {\em star} of $v$, which is defined as \[\mbox{st}_K(v) \hspace{1ex} \coloneqq \hspace{1ex} \{ S\subset V \: : \: S\cup \{v\} \in K \}.\] Notice that $\mbox{st}_K(v)$ is always non-empty since $v\in \mbox{st}_K(v)$. Also, if $L$ is a subcomplex of $K$ and $v$ is a vertex of $L$, then $\mbox{st}_L(v) \subset \mbox{st}_K(v)$. \bigskip For a vertex $v\in V$, let $\Gamma_K(v)$ denote the {\em neighborhood complex} of $v$, which is defined as \[\Gamma_K(v) \hspace{1ex} \coloneqq \hspace{1ex} \mbox{st}_K(v) \hspace{1ex} \cup \hspace{1ex} \{S \subset V - \{v\} \: :\: S \in K, |S| = d+1, [S]_d \subset \mbox{st}_K(v)\}.\] We warn the reader about the subtle dependence on $d = \dim K$. For instance, if $K$ is $0$-dimensional, i.e. a set of isolated vertices, then $\Gamma_K(v) = K$. However, if $K$ has positive dimension and $v$ is an isolated vertex of $K$, then $\Gamma_K(v) = \{v\}$. This shows that if $L$ is a subcomplex of $K$, then it is not generally true that $\Gamma_L(v)$ is a subcomplex of $\Gamma_K(v)$. \bigskip For $j\geq d$, let $\Delta_j(K)$ denote the {\em $j$-completion} of $K$, which is defined as \[\Delta_j(K) \hspace{1ex} \coloneqq \hspace{1ex} K \hspace{1ex} \cup \hspace{1ex} \{ S \subset V \: : \: |S|\geq j+2 , [S]_{j+1}\subset K\}.\] Let us also define $\Delta_j(\emptyset) \coloneqq \emptyset$. Notice that if $K$ is $0$-dimensional then $\Delta_0(K)$ is the $(|V|-1)$-dimensional simplex. Also, if $j >\dim K$, then $\Delta_j(K) = K$. Consequently, if $L$ is a subcomplex of $K$, then $\Delta_d(L) \subset \Delta_d(K)$. \begin{prop} \label{iden1} Let $K$ be a simplicial complex of dimension $d$ and let $\{K_i\}_{i\in I}$ be a finite family of subcomplexes of $K$. Then \[\Delta _d\left(\bigcap_{i\in I} K_i \right) \; = \; \bigcap_{i\in I} \Delta_d(K_i).\] \end{prop} \begin{proof} Since $\bigcap_{i\in I} K_i \subset K_i$, we have $\Delta_d\left(\bigcap_{i\in I}K_i\right)\subset \Delta_d(K_i)$. Therefore $\Delta_d\left(\bigcap_{i\in I}K_i\right) \subset \bigcap_{i\in I}\Delta_d(K_i)$. For the other direction, suppose $S\in \bigcap_{i\in I}\Delta_d(K_i)$. If $|S| \leq d+1$, then $S \in \bigcap_{i\in I}K_i\subset \Delta_d\left(\bigcap_{i\in I}K_i\right)$. If $|S| \geq d+2$, then $[S]_{d+1} \subset K_i$ for every $i\in I$. That is, $[S]_{d+1} \subset \bigcap_{i\in I} K_i$, and therefore $S \in \Delta_d\left(\bigcap_{i\in I}K_i\right)$. \end{proof} \begin{prop}\label{iden2} Let $K$ be a simplicial complex of dimension $d$ on the vertex set $V$ and let $v\in V$. Then \[\mbox{\em st}_{\Delta_d(K)}(v) \; = \; \Delta_d\left(\Gamma_K(v)\right).\] \end{prop} \begin{proof} We first show that $\mbox{st}_{\Delta_d(K)}(v) \subset \Delta_d(\Gamma_K(v))$. Suppose $S \in \mbox{st}_{\Delta_d(K)}(v)$, which, by definition, means that \[S \cup \{v\} \; \in \; \Delta_d(K) \; = \; K \; \cup \; \{T \subset V \: : \: |S|\geq d+2 , [S]_{d+1}\subset K\}.\] If $|S\cup \{v\}|\leq d+1$, then $S \cup \{v\}\in K$. This implies that $S\in \mbox{st}_K(v)$, and since $\mbox{st}_K(v) \subset \Gamma_K(v) \subset \Delta_d(\Gamma_K(v))$ we have $S\in \Delta_d(\Gamma_K(v))$. If $|S \cup \{v\}| \geq d+2$, then $[S \cup \{v\}]_{d+1} \subset K$. This implies that for every $T\in [S - \{v\}]_{d}$ we have $T \in \mbox{st}_K(v)$, and consequently $[S]_{d+1} \subset \Gamma_K(v)$. Therefore $S\in \Delta_d(\Gamma_K(v))$. \bigskip It remains to show that $\Delta_d(\Gamma_K(v)) \subset \mbox{st}_{\Delta_d(K)}(v)$. Suppose $S \in \Delta_d(\Gamma_K(v))$. If $|S|\leq d+1$, then $S \in\Gamma_K(v)$. Furthermore, if $|S - \{v\}| \leq d$ it follows that $S \in \mbox{st}_K(v)$. Since $K \subset \Delta_d(K)$ we have $S \in \mbox{st}_{\Delta_d(K)}(v)$. On the other hand, if $v\notin S$ and $|S| = d+1$, then, by definition, we have $S\in K$ and $[S]_d \subset \mbox{st}_K(v)$. This implies that $[S \cup \{v\}]_{d+1} \subset K$, and it follows that $S \cup \{v\} \in \Delta_d(K)$, which shows that $S \in \mbox{st}_{\Delta_d(K)}(v)$. If $|S|\geq d+2$, then $[S]_{d+1} \subset \Gamma_K(v)$. This implies that for every $T\in [S - \{v\}]_d$ we have $T \in \mbox{st}_K(v)$, and for every $T \in [S - \{v\}]_{d+1}$ we have $T \in K$. It follows that $[S \cup \{v\}]_{d+1} \subset K$, and therefore $S \cup \{v\} \in \Delta_d(K)$, which shows that $S \in \mbox{st}_{\Delta_d(K)}(v)$. \end{proof} \subsection{The Nerve theorem}\label{nerves} Let $F$ be a finite family of sets. The {\em nerve} of $F$, denoted by $N(F)$, is the simplicial complex on the vertex set $F$ whose simplices are the intersecting subfamilies of $F$, that is \[N(F) \coloneqq \{ G\subset F \: : \: \bigcap_{S\in G}S \neq \emptyset\}.\] We will use the following version of the Nerve theorem which is a consequence of \cite[Theorem 6]{bjorner}. \begin{theorem}[Bj{\"o}rner]\label{nerve} Let $K$ be a simplicial complex and $F = \{K_i\}_{i \in I}$ a finite family of subcomplexes such that $K = \bigcup_{i\in I}K_i$. Suppose every non-empty intersection $\bigcap_{t\in T} K_t$ is $(k + 1 - |T|)$-connected, $T \subset I$. Then $K$ is $k$-connected if and only if $N(F)$ is $k$-connected. \end{theorem} \subsection{The $q$-star property} Let $K$ be a simplicial complex of dimension $d$ on the vertex set $V$. For any integer $q\geq 1$, we say that $K$ is {\em $q$-star} if $|V|>q$, and for every subset $Y \subset V$ of size $q$ there exists a vertex $v\in V - Y$ such that $S \cup \{v\} \in K$ for every simplex $S \in K[Y]$ with $|S|\leq d$. The following is an extension of a result on clique complexes appearing in \cite[Theorem 3.1]{kahle}, and in a more general form in \cite[Theorem 1.5]{meshulam}. \begin{lemma} \label{q-star} Let $K$ be a simplicial complex of dimension $d$ and let $k$ be a non-negative integer. If $K$ is $(2k+2)$-star, then its $d$-completion $\Delta_d(K)$ is $k$-connected. \end{lemma} \begin{proof} For $d=0$ the statement holds because $\Delta_0(K)$ is a simplex which is contractible. We may therefore assume $d\geq 1$. If a complex $K$ of dimension $d\geq 1$ is $2$-star, then $K$ is connected which implies that $\Delta_d(K)$ is also connected. So the statement is clearly true for $k=0$, and we proceed by induction on $k$. Suppose $K$ is $(2k+2)$-star for $k>0$ and let $V$ be the vertex set of $K$. For each vertex $v\in V$, let $K_v = \mbox{st}_{\Delta_d(K)}(v)$. Define the family of subcomplexes $F = \{K_v\}_{v\in V}$. Clearly we have \[\Delta_d(K) = \bigcup_{v\in V} K_{v}.\] For a non-empty subset $W\subset V$, let $K_W = \bigcap_{v\in W}K_v$. Theorem \ref{nerve} implies that $\Delta_d(K)$ is $k$-connected, if we can show the following. \begin{enumerate}[(i)] \item $K_W$ is $(k + 1 - t)$-connected for every non-empty subset $W\subset V$ with $|W| = t$. \item The nerve $N(F)$ is $k$-connected. \end{enumerate} \bigskip Part (i). For every $v\in V$, $K_v$ is a cone which is contractible, and hence $k$-connected. Now consider $W\subset V$ with $|W| = t \geq 2$, and let $L_W = \bigcap_{v\in W} \Gamma_K(v)$. By Propositions \ref{iden1} and \ref{iden2} we have \[K_W \; = \; \bigcap_{v\in W}\mbox{st}_{\Delta_d(K)}(v) \; =\; \Delta_d\left(\bigcap_{v\in W} \Gamma_K(v)\right) \; = \; \Delta_d(L_W).\] By induction, it therefore suffices to prove that $L_W$ is $(2(k + 1 - t) + 2)$-star. Also notice that for $t\geq 2$ we have $2k + 2 - t \geq 2(k + 1 - t) + 2$, so it suffices to show that $L_W$ is $(2k+2- t)$-star. Let $X$ be the vertex set of $L_W$. Clearly a vertex $v$ belongs to $X$ if and only if $\{v,w\} \in K$ for all $w\in W$. This implies that $|X| > 2k+2-t$, since $K$ is $(2k+2)$-star. Next, we observe that for every $Y \subset X$ with $|Y| = 2k + 2 - t$ we can find a set $Z\subset X \cup W$ with $|Z| = 2k+2$ such that $Y\cup W \subset Z$. Since $K$ is $(2k+2)$-star, there exists $v\in V - Z$ such that \[S\cup \{v\} \in K \; \mbox{ for every } \; S\in K[Z] \; \mbox{ with } \; |S|\leq d. \label{eq:st} \tag{$\ast$}\] It follows from our previous observation that $v\in X$. Let $S\in L_W[Y]$ with $|S| \leq \dim L_W \leq d$. We need to show that $S\cup \{v\} \in L_W$, that is, $S\cup \{v\} \in \Gamma_K(w)$ for every $w\in W$. Notice that $S \cup \{w\} \subset Z$, so \eqref{eq:st} may be applied provided $|S\cup \{w\}|\leq d$. If $|S \cup \{w\}|\leq d$, then \eqref{eq:st} implies that $S\cup \{w\} \cup \{v\} \in K$. This just means that $S\cup \{v\} \in \mbox{st}_K(w)$, and consequently $S \cup \{v\} \in \Gamma_K(w)$. If $|S\cup \{w\}| = d+1$, then $|S| = d$, $w \notin S$, and $S \cup \{w\} \in K$. For every $T\in [S \cup \{w\}]_d$, it follows from \eqref{eq:st} that $T \cup \{v\} \in K$. In particular $S \cup \{v\} \in K$ and $[S\cup \{v\}]_d\subset \mbox{st}_K(w)$, and since $|S \cup \{v\}| = d+1$, we conclude that $S\cup \{v\} \in \Gamma_K(w)$. This shows that $L_W$ is $(2k + 2 - t)$-star. \bigskip Part (ii). Clearly $K_W$ is non-empty for any subset $W\subset V$ with $|W| = 2k+2$. Therefore the $(2k+1)$-skeleton of the nerve $N(F)$ is complete, which implies that $N(F)$ is $2k$-connected.\end{proof} \begin{proof}[Proof of Theorem \ref{gencom}] Let $K = M(X)$, the independence complex of $X$. Clearly the general position complex of $X$ is the $d$-completion of $K$, that is, $G(X) = \Delta_d(K)$. We want to show that $G(X)$ is $k$-connected provided $\varphi(X)$ is sufficiently large. In view of Remark \ref{2nd remark}, we may assume that $k\geq d$. We will show that if $\varphi(X) > d\binom{2k+2}{d}$, then $K$ is $(2k+2)$-star. This implies that $G(X)$ is $k$-connected by Lemma \ref{q-star}. Let $S\subset X$ with $|S| = 2k+2$. Let $H$ denote the set of hyperplanes spanned by affinely independent $d$-tuples in $S$. Therefore, if $\varphi(X) > d\binom{2k+2}{d} \geq d|H|$, then there exists a point $x\in X$ which is not contained in the affine hull of any subset $T\subset S$ with $|T|\leq d$, and consequently $K$ is $(2k+2)$-star. This gives an upper bound on $g_d(k)$ which is in $O(k^d)$.\end{proof} \section{Concluding remarks} \label{remarks} The natural problem that arises is to try to determine better (or exact) bounds for the functions $g_d(k)$. We have shown that $g_d(k) = k+2$ for $k \leq d-1$ and $g_d(k) \leq d\binom{2k+2}{d}+1$, otherwise, but we hardly believe this to be optimal. In fact, the exact same proof (and bound) works in a more general setting, which we now describe. Let $M$ be a matroid of rank $r$ on the ground set $E$. We say that a subset $S\subset E$ is {\em uniform} if $S$ is independent or $|S| >r$ and every member of $[S]_r$ is independent. The set of all uniform subsets of a matroid $M$ form a simplicial complex, which call the {\em uniformity complex} of $M$. Obviously, the uniformity complex of a matroid is the $(r-1)$-completion of its independence complex. If we let $\mu(M)$ denote the maximum size of a uniform subset of $M$, then we have the following generalization of Theorem \ref{gencom}. \begin{theorem}\label{uniformity} For all integers $r\geq 2$ and $k\geq -1$ there exists a minimal positive integer $h_r(k)$ such that the following holds. If $M$ is matroid of rank $r$ and $\mu(M) \geq h_r(k)$, then the uniformity complex of $M$ is $k$-connected. \end{theorem} \begin{proof} The same argument (as in the proof of Theorem \ref{gencom}) shows that if $\mu(M) > (r-1) \binom{2k+2}{r-1}$, then the independence complex of $M$ is $(2k+2)$-star. The theorem then follows from Lemma \ref{q-star}. \end{proof} We find it likely that there should be a sharp distinction in the asymptotic behavior between the function $h_r(k)$ and the corresponding function $g_{r-1}(k)$. More generally, we find it reasonable to expect the orientability of the matroid $M$ to have a strong quantitative effect on the connectivity of the uniformity complex, but we lack any evidence to support this. In fact, the only exact value we know (apart from what is covered by Remark \ref{2nd remark}) is $g_2(2) = h_3(2) = 7$. In conclusion we mention that, in view of Theorem \ref{uniformity}, it is straightforward to apply Proposition \ref{colorful} to obtain an analogue of Theorem \ref{sgpr} for {\em uniform systems of representatives}. Further generalizations can also be obtained by using the more general version of Lemma \ref{colorful} appearing in \cite[Theorem 4.5]{aharoni2}. We leave the details to the reader. \section{Acknowledgments} We are grateful to the anonymous referee for pointing out a mistake in our original proof and for making several other valuable comments and suggestions.

https://arxiv.org/abs/1105.5572

Lagrange's Theorem for Hopf Monoids in Species

Following Radford's proof of Lagrange's theorem for pointed Hopf algebras, we prove Lagrange's theorem for Hopf monoids in the category of connected species. As a corollary, we obtain necessary conditions for a given subspecies K of a Hopf monoid H to be a Hopf submonoid: the quotient of any one of the generating series of H by the corresponding generating series of K must have nonnegative coefficients. Other corollaries include a necessary condition for a sequence of nonnegative integers to be the sequence of dimensions of a Hopf monoid in the form of certain polynomial inequalities, and of a set-theoretic Hopf monoid in the form of certain linear inequalities. The latter express that the binomial transform of the sequence must be nonnegative.

\section*{Introduction} Lagrange's theorem states that for any subgroup $K$ of a group $H$, $H\cong K\times Q$ as (left) $K$-sets, where $Q=H/K$. In particular, if $H$ is finite, then $|K|$ divides $|H|$. Passing to group algebras over a field $\Bbbk$, we have that $\Bbbk H \cong \Bbbk K \otimes \Bbbk Q$ as (left) $\Bbbk K$-modules, or that $\Bbbk H$ is free as a $\Bbbk K$-module. Kaplansky~\cite{Kap:1975} conjectured that the same statement holds for Hopf algebras---group algebras being principal examples. It turns out that the result does not hold in general, as shown by Oberst and Schneider~\cite[Proposition~10]{ObeSch:1974} and \cite[Example~3.5.2]{Mon:1993}. On the other hand, the result does hold for certain large classes of Hopf algebras, including the finite dimensional ones by a theorem of Nichols and Zoeller~\cite{NicZoe:1989}, and the pointed ones by a theorem of Radford~\cite{Rad:1977}. Further (and finer) results of this nature were developed by Schneider~\cite{Sch:1990,Sch:1992}. Additional work on the conjecture includes that of Masuoka~\cite{Mas:1992} and Takeuchi~\cite{Tak:1979}; more information can be found in Sommerh\"auser's survey~\cite{Som:2000}. The main result of this paper (Theorem~\ref{t:main}) is a version of Lagrange's theorem for Hopf monoids in the category of connected species: if $\mathbf h}%{\bm{\sf h}$ is a connected Hopf monoid and $\mathbf k}%{\bm{\sf k}$ is a Hopf submonoid, there exists a species $\mathbf q}%{\bm{\sf q}$ such that $\mathbf h}%{\bm{\sf h}=\mathbf k}%{\bm{\sf k}\bm\cdot\mathbf q}%{\bm{\sf q}$. An immediate application is a test for Hopf submonoids (Corollary~\ref{c:Sp series}): if any one of the generating series for a species $\mathbf k}%{\bm{\sf k}$ does not divide in $\mathbb{Q}_{\geq 0}{[\![} x{]\!]}$ the corresponding generating series for the Hopf monoid $\mathbf h}%{\bm{\sf h}$ (in the sense that the quotient has at least one negative coefficient), then $\mathbf k}%{\bm{\sf k}$ is not a Hopf submonoid of $\mathbf h}%{\bm{\sf h}$. A similar test also holds for connected graded Hopf algebras (Corollary~\ref{c:cgVec series}). The proof of Theorem~\ref{t:main} for Hopf monoids in species parallels Radford's proof for Hopf algebras. (Hopf algebras are Hopf monoids in the category of vector spaces). The paper is organized as follows. In Section~\ref{s:Hopf algebras}, we recall Lagrange's theorem for Hopf algebras, focusing on the case of connected graded Hopf algebras. In Section~\ref{s:Hopf monoids}, we recall the basics of Hopf monoids in species and prove Lagrange's theorem in this setting. Examples and applications are given in Section~\ref{s:applications}. Among these, we derive certain polynomial inequalities that a sequence of nonnegative integers must satisfy in order to be the dimension sequence of a connected Hopf monoid in species. In the case of a set-theoretic Hopf monoid structure, we obtain additional necessary conditions in the form of linear inequalities which express that the binomial transform of the enumerating sequence must be nonnegative. In Section~\ref{s:dimensions} we provide information on the growth and support of the dimension sequence of a connected Hopf monoid. The latter must be an additive submonoid of the natural numbers. We conclude in Section~\ref{s:kernels} with information on the species $\mathbf q}%{\bm{\sf q}$ entering in Lagrange's theorem. In the dual setting, $\mathbf q}%{\bm{\sf q}$ is the Hopf kernel of a morphism, and for cocommutative Hopf monoids it can be described in terms of Lie kernels and primitive elements via the Poincar\'e-Birkhoff-Witt theorem. All vector spaces are over a fixed field $\Bbbk$ of characteristic $0$, except in Section~\ref{s:dimensions}, where the characteristic is arbitrary. \section{Lagrange's theorem for Hopf algebras}\label{s:Hopf algebras} We begin by recalling a couple of versions of this theorem. \begin{theorem} Let $H$ be a finite dimensional Hopf algebra over a field $\Bbbk$. If $K\subseteq H$ is any Hopf subalgebra, then $H$ is a free left (and right) $K$-module. \end{theorem} This is the Nichols-Zoeller theorem \cite{NicZoe:1989}; see also \cite[Theorem~3.1.5]{Mon:1993}. We will not make direct use of this result, but instead of the related results discussed below. A Hopf algebra $H$ is \demph{pointed} if all its simple subcoalgebras are $1$-dimensional. Equivalently, the group-like elements of $H$ linearly span the coradical of $H$. Given a subspace $K$ of $H$, set \[ K_+:=K\cap\ker(\epsilon), \] where $\epsilon:H\to\Bbbk$ is the counit of $H$. Let $K_+H$ denote the right $H$-ideal generated by $K_+$. \begin{theorem}\label{t:pointed} Let $H$ be a pointed Hopf algebra. If $K\subseteq H$ is any Hopf subalgebra, then $H$ is a free left (and right) $K$-module. Moreover, \[ H\cong K\otimes (H/K_+H) \] as left $K$-modules. \end{theorem} The first statement is due to Radford \cite[Section~4]{Rad:1977} and the second (stronger) statement to Schneider \cite[Remark~4.14]{Sch:1990}, \cite[Corollary~4.3]{Sch:1992}. Various generalizations can be found in these references as well as in Masuoka~\cite{Mas:1992} and Takeuchi \cite{Tak:1979}; see also Sommerha\"user~\cite{Som:2000}. We are interested in the particular variant given in Theorem~\ref{t:connected} below. A Hopf algebra $H$ is \demph{graded} if there is given a decomposition \[ H=\bigoplus_{n\geq0} H_n \] into linear subspaces that is preserved by all operations. It is \demph{connected graded} if in addition $H_0$ is linearly spanned by the unit element. \begin{theorem}\label{t:connected} Let $H$ be a connected graded Hopf algebra. If $K\subseteq H$ is a graded Hopf subalgebra, then $H$ is a free left (and right) $K$-module. Moreover, \[ H\cong K\otimes (H/K_+H) \] as left $K$-modules and as graded vector spaces. \end{theorem} \begin{proof} Since $H$ is connected graded, its coradical is $H_0 = \Bbbk$, so $H$ is pointed and Theorem~\ref{t:pointed} applies. Radford's proof shows that there exists a graded vector space $Q$ such that \[ H \cong K \otimes Q \] as left $K$-modules and as graded vector spaces. (The argument we give in the parallel setting of Theorem~\ref{t:main} makes this clear.) Note that $K_{+\!}=\bigoplus_{n\geq1} K_n$, hence $K_+H$ and $H/K_+H$ inherit the grading of $H$. To complete the proof, it suffices to show that $Q\cong H/K_+H$ as graded vector spaces. Let $\varphi:K\otimes Q\to H$ be an isomorphism of left $K$-modules and of graded vector spaces. We claim that \[ \varphi(K_+\otimes Q)=K_+H. \] In fact, since $\varphi$ is a morphism of left $K$-modules, \[ \varphi(K_+\otimes Q) = K_+\varphi(1\otimes Q)\subseteq K_+H. \] Conversely, if $k\in K_+$ and $h\in H$, writing $h=\sum_i \varphi(k_i\otimes q_i)$ with $k_i\in K$ and $q_i\in Q$, we obtain \[ kh=\sum_i \varphi(kk_i\otimes q_i)\in \varphi(K_+\otimes Q), \] since $K_+$ is an ideal of $K$. Now, since $K=K_0\oplus K_+$, we have \[ K\otimes Q =(K_0\otimes Q)\oplus (K_+\otimes Q) \] and therefore \[ H/K_+H = \varphi(K\otimes Q)/\varphi(K_+\otimes Q) \cong \varphi (K_0\otimes Q)\cong Q \] as graded vector spaces. \end{proof} Given a graded Hopf algebra $H$, let $\poincare{H}{x}\in \mathbb{N}{[\![} x{]\!]}$ denote its \demph{Poincar\'e series}---the ordinary generating function for the sequence of dimensions of its graded components, \begin{gather* \poincare{H}{x} := \sum_{n\geq0} \dim H_n \, x^n . \end{gather*} Suppose $H$ is connected graded and $K$ is a graded Hopf subalgebra. In this case, their Poincar\'e series are of the form \[ 1+a_1x+a_2x^2+\cdots \] with $a_i\in\mathbb{N}$ and the quotient $\poincare{H}{x}/\poincare{K}{x}$ is a well-defined power series in $\mathbb{Z}{[\![} x{]\!]}$. \begin{corollary}\label{c:cgVec series} Let $H$ be a connected graded Hopf algebra. If $K\subseteq H$ is any graded Hopf subalgebra, then the quotient $\poincare{H}{x} / \poincare{K}{x}$ of Poincar\'e series is nonnegative, i.e., belongs to $\mathbb{N}{[\![} x{]\!]}$. \end{corollary} \begin{proof} By Theorem~\ref{t:connected}, $H \cong K \otimes Q$ as graded vector spaces, where $Q=H/K_+H$. Hence $\poincare{H}{x} = \poincare{K}{x} \, \poincare{Q}{x}$ and the result follows. \end{proof} \begin{example} Consider the Hopf algebra $\textsl{QSym}$ of quasisymmetric functions in countably many variables, and the Hopf subalgebra $\textsl{Sym}$ of symmetric functions. They are connected graded, so by Theorem~\ref{t:connected}, $\textsl{QSym}$ is a free module over $\textsl{Sym}$. Garsia and Wallach prove this same fact for the algebras $\textsl{QSym}_n$ and $\textsl{Sym}_n$ of (quasi) symmetric functions in $n$ variables (where $n$ is a finite number)~\cite{GarWal:2003}. While $\textsl{QSym}_n$ and $\textsl{Sym}_n$ are quotient algebras of $\textsl{QSym}$ and $\textsl{Sym}$, they are not quotient coalgebras. Since a Hopf algebra structure is needed in order to apply Theorem~\ref{t:connected}, we cannot derive the result of Garsia and Wallach in this manner. The papers~\cite{GarWal:2003} and~\cite{LauMas:2011} provide information on the space $Q_n$ entering in the decomposition $\textsl{QSym}_n\cong\textsl{Sym}_n\otimes Q_n$. \end{example} \section{Lagrange's theorem for Hopf monoids in species}\label{s:Hopf monoids} We first review the notion of Hopf monoid in the category of species, following \cite{AguMah:2010}, and then prove Lagrange's theorem in this setting. We restrict attention to the case of connected Hopf monoids. \subsection{Hopf monoids in species}\label{ss:species} The notion of species was introduced by Joyal \cite{Joy:1981}. It formalizes the notion of combinatorial structure and provides a framework for studying the generating functions which enumerate these structures. The book \cite{BerLabLer:1998} by Bergeron, Labelle and Leroux expounds the theory of set species. Joyal's work indicates that species may also be regarded as algebraic objects; this is the point of view adopted in \cite{AguMah:2010} and in this work. To this end, it is convenient to work with vector species. A \demph{(vector) species} is a functor $\mathbf q}%{\bm{\sf q}$ from {finite sets} and bijections to {vector spaces} and linear maps. Specifically, it is a family of vector spaces $\mathbf q}%{\bm{\sf q}[I]$, one for each finite set $I$, together with linear maps $\mathbf q}%{\bm{\sf q}[\sigma]: \mathbf q}%{\bm{\sf q}[I] \to \mathbf q}%{\bm{\sf q}[J]$, one for each bijection $\sigma:I\to J$, satisfying \[ \mathbf q}%{\bm{\sf q}[\mathrm{id}_I] = \mathrm{id}_{\mathbf q}%{\bm{\sf q}[I]} \quad\hbox{and}\quad \mathbf q}%{\bm{\sf q}[\sigma\circ\tau] = \mathbf q}%{\bm{\sf q}[\sigma]\circ \mathbf q}%{\bm{\sf q}[\tau] \] whenever $\sigma$ and $\tau$ are composable bijections. The notation $\mathbf q}%{\bm{\sf q}[a,b,c,\ldots]$ is shorthand for $\mathbf q}%{\bm{\sf q}[\{a,b,c,\ldots\}]$ and $\mathbf q}%{\bm{\sf q}[n]$ is shorthand for $\mathbf q}%{\bm{\sf q}[1,2,\ldots,n]$. The space $\mathbf q}%{\bm{\sf q}[n]$ is an $S_n$-module via $\sigma\cdot v = \mathbf q}%{\bm{\sf q}[\sigma](v)$ for $v\in\mathbf q}%{\bm{\sf q}[n]$ and $\sigma\in S_n$. A species $\mathbf q}%{\bm{\sf q}$ is \demph{finite dimensional} if each vector space $\mathbf q}%{\bm{\sf q}[I]$ is finite dimensional. In this paper, all species are finite dimensional. A morphism of species is a natural transformation of functors. Let ${\sf Sp}$ denote the category of (finite dimensional) species. We give two elementary examples that will be useful later. \begin{example}\label{ex: example species} Let $\mathbf E}%{\bm{\sf l}$ be the \demph{exponential species}, defined by $\mathbf E}%{\bm{\sf l}[I] = \Bbbk\{\ast_I\}$ for all $I$. The symbol $\ast_I$ denotes an element canonically associated to the set $I$ (for definiteness, we may take $\ast_I=I$). Thus, $\mathbf E}%{\bm{\sf l}[I]$ is a $1$-dimensional space with a distinguished basis element. A richer example is provided by the species $\mathbf L}%{\bm{\sf l}$ of \demph{linear orders}, defined by $\mathbf L}%{\bm{\sf l}[I] = \Bbbk\{\hbox{linear orders on }I\}$ for all $I$ (a space of dimension $n!$ when $|I|=n$). \end{example} We use $\bm\cdot$ to denote the \demph{Cauchy product} of two species. Specifically, \[ \bigl(\mathbf p}%{\bm{\sf p} \bm\cdot \mathbf q}%{\bm{\sf q}\bigr)[I] := \bigoplus_{S\sqcup T = I} \mathbf p}%{\bm{\sf p}[S] \otimes \mathbf q}%{\bm{\sf q}[T] \quad\hbox{for all finite sets }I. \] The notation $S\sqcup T = I$ indicates that $S\cup T = I$ and $S \cap T = \emptyset$. The sum runs over all such \demph{ordered decompositions} of $I$, or equivalently over all subsets $S$ of $I$: there is one term for $S\sqcup T$ and another for $T\sqcup S$. The Cauchy product turns ${\sf Sp}$ into a symmetric monoidal category. The braiding simply switches the tensor factors. The unit object is the species $\mathbf{1}$ defined by \[ \mathbf{1}[I] := \begin{cases} \Bbbk & \text{if $I$ is empty,} \\ 0 & \text{otherwise.} \end{cases} \] A \demph{monoid} in the category $({\sf Sp},\bm\cdot)$ is a species $\mathbf m}%{\bm{\sf m}$ together with a morphism of species $\mu: \mathbf m}%{\bm{\sf m}\bm\cdot\mathbf m}%{\bm{\sf m} \to \mathbf m}%{\bm{\sf m}$, i.e., a family of maps \[ \mu_{S,T} : \mathbf m}%{\bm{\sf m}[S] \otimes \mathbf m}%{\bm{\sf m}[T] \to \mathbf m}%{\bm{\sf m}[I], \] one for each ordered decomposition $I = S\sqcup T$, satisfying appropriate associativity and unital conditions, and naturality with respect to bijections. Briefly, to each $\mathbf m}%{\bm{\sf m}$-structure on $S$ and $\mathbf m}%{\bm{\sf m}$-structure on $T$, there is assigned an $\mathbf m}%{\bm{\sf m}$-structure on $S\sqcup T$. The analogous object in the category of graded vector spaces is a graded algebra. For the species $\mathbf E}%{\bm{\sf l}$, a monoid structure is defined by sending the basis element $\ast_S \otimes \ast_T$ to the basis element $\ast_I$. For $\mathbf L}%{\bm{\sf l}$, a monoid structure is provided by concatenation of linear orders: $\mu_{S,T}(\ell_1 \otimes \ell_2) = (\ell_1, \ell_2)$. A \demph{comonoid} in the category $({\sf Sp},\bm\cdot)$ is a species $\mathbf c}%{\bm{\sf c}$ together with a morphism of species $\Delta:\mathbf c}%{\bm{\sf c} \to \mathbf c}%{\bm{\sf c}\bm\cdot\mathbf c}%{\bm{\sf c}$, i.e., a family of maps \[ \Delta_{S,T} : \mathbf c}%{\bm{\sf c}[I] \to \mathbf c}%{\bm{\sf c}[S] \otimes \mathbf c}%{\bm{\sf c}[T], \] one for each ordered decomposition $I=S\sqcup T$, which are natural, coassociative and counital. For the species $\mathbf E}%{\bm{\sf l}$, a comonoid structure is defined by sending the basis vector $\ast_I$ to the basis vector $\ast_S \otimes \ast_T$. For $\mathbf L}%{\bm{\sf l}$, a comonoid structure is provided by restricting a total order $\ell$ on $I$ to total orders on $S$ and $T$: $\Delta_{S,T}(\ell) = \ell\vert_S\otimes \ell\vert_T$. We assume that our species $\mathbf q}%{\bm{\sf q}$ are \demph{connected}, i.e., $\mathbf q}%{\bm{\sf q}[\emptyset]=\Bbbk$. In this case, the (co)unital conditions for a (co)monoid force the maps $\mu_{S,T}$ ($\Delta_{S,T}$) to be the canonical identifications if either $S$ or $T$ is empty. Thus, in defining a connected (co)monoid structure one only needs to specify the maps $\mu_{S,T}$ ($\Delta_{S,T}$) when both $S$ and $T$ are nonempty. A \demph{Hopf monoid} in the category $({\sf Sp}, \bm\cdot)$ is a monoid and comonoid whose two structures are compatible in an appropriate sense, and which carries an antipode. In this paper we only consider connected Hopf monoids. For such Hopf monoids, the existence of the antipode is guaranteed. The species $\mathbf E}%{\bm{\sf l}$ and $\mathbf L}%{\bm{\sf l}$, with the structures outlined above, are two important examples of connected Hopf monoids. For further details on Hopf monoids in species, see Chapter 8 of \cite{AguMah:2010}. The theory of Hopf monoids in species is developed in Part II of this reference; several examples are discussed in Chapters 12 and 13. \subsection{Lagrange's theorem for connected Hopf monoids} Given a connected Hopf monoid $\mathbf k}%{\bm{\sf k}$ in species, we let $\mathbf k}%{\bm{\sf k}_+$ denote the species defined by \[ \mathbf k}%{\bm{\sf k}_+[I] = \begin{cases} \mathbf k}%{\bm{\sf k}[I] & \text{ if }I\neq\emptyset, \\ 0 & \text{ if }I=\emptyset. \end{cases} \] If $\mathbf k}%{\bm{\sf k}$ is a submonoid of a monoid $\mathbf h}%{\bm{\sf h}$, then $\mathbf k}%{\bm{\sf k}_+\mathbf h}%{\bm{\sf h}$ denotes the right ideal of $\mathbf h}%{\bm{\sf h}$ generated by $\mathbf k}%{\bm{\sf k}_+$. In other words, \[ (\mathbf k}%{\bm{\sf k}_+\mathbf h}%{\bm{\sf h})[I]=\sum_{\substack{S\sqcup T=I\\ S\neq\emptyset}}\mu_{S,T}\bigl(\mathbf k}%{\bm{\sf k}[S]\otimes\mathbf h}%{\bm{\sf h}[T]\bigr). \] \begin{theorem}\label{t:main} Let $\mathbf h}%{\bm{\sf h}$ be a connected Hopf monoid in the category of species. If $\mathbf k}%{\bm{\sf k}$ is a Hopf submonoid of $\mathbf h}%{\bm{\sf h}$, then $\mathbf h}%{\bm{\sf h}$ is a free left $\mathbf k}%{\bm{\sf k}$-module. Moreover, \[ \mathbf h}%{\bm{\sf h}\cong \mathbf k}%{\bm{\sf k}\bm\cdot(\mathbf h}%{\bm{\sf h}/\mathbf k}%{\bm{\sf k}_+\mathbf h}%{\bm{\sf h}) \] as left $\mathbf k}%{\bm{\sf k}$-modules (and as species). \end{theorem} The proof is given after a series of preparatory results. Our argument parallels Radford's proof of the first statement in Theorem~\ref{t:pointed} \cite[Section~4]{Rad:1977}. The main ingredient is a result of Larson and Sweedler \cite{LarSwe:1969} known as the fundamental theorem of Hopf modules \cite[Theorem~1.9.4]{Mon:1993}. It states that if $(M,\rho)$ is a left Hopf module over $K$, then $M$ is free as a left $K$-module and in fact is isomorphic to the Hopf module $K\otimes Q$, where $Q$ is the space of \demph{coinvariants} for the coaction $\rho \colon M\to K\otimes M$. Takeuchi extends this result to the context of Hopf monoids in a braided monoidal category with equalizers \cite[Theorem~3.4]{Tak:1999}; a similar result (in a more restrictive setting) is given by Lyubashenko \cite[Theorem~1.1]{Lyu:1995}. As a special case of Takeuchi's result, we have the following. \begin{proposition}\label{p:FTHM} Let $\mathbf m}%{\bm{\sf m}$ be a left Hopf module over a connected Hopf monoid $\mathbf k}%{\bm{\sf k}$ in species. There is an isomorphism $\mathbf m}%{\bm{\sf m} \cong \mathbf k}%{\bm{\sf k} \bm\cdot \mathbf q}%{\bm{\sf q}$ of left Hopf modules, where \[ \mathbf q}%{\bm{\sf q}[I] := \bigl\{m\in \mathbf m}%{\bm{\sf m}[I] \mid \text{$\rho_{S,T}(m) = 0$ for $S\sqcup T=I$, $T\neq I$} \bigr\}. \] In particular, $\mathbf m}%{\bm{\sf m}$ is free as a left $\mathbf k}%{\bm{\sf k}$-module. \end{proposition} Here $\rho:\mathbf m}%{\bm{\sf m}\to\mathbf k}%{\bm{\sf k}\bm\cdot\mathbf m}%{\bm{\sf m}$ denotes the comodule structure, which consists of maps \[ \rho_{S,T}: \mathbf m}%{\bm{\sf m}[I] \to \mathbf k}%{\bm{\sf k}[S]\otimes\mathbf m}%{\bm{\sf m}[T], \] one for each ordered decomposition $I=S\sqcup T$. \medskip Given a comonoid $\mathbf h}%{\bm{\sf h}$ and two subspecies $\mathbf u}%{\bm{\sf u},\mathbf v}%{\bm{\sf v} \subseteq \mathbf h}%{\bm{\sf h}$, the \demph{wedge} of $\mathbf u}%{\bm{\sf u}$ and $\mathbf v}%{\bm{\sf v}$ is the subspecies $\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}$ of $\mathbf h}%{\bm{\sf h}$ defined by \[ \mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v} := \Delta^{-1}(\mathbf u}%{\bm{\sf u}\bm\cdot\mathbf h}%{\bm{\sf h} + \mathbf h}%{\bm{\sf h}\bm\cdot\mathbf v}%{\bm{\sf v}). \] The remaining ingredients needed for the proof are supplied by the following lemmas. \begin{lemma}\label{l:lemma 1} Let $\mathbf h}%{\bm{\sf h}$ be a comonoid in species. If $\mathbf u}%{\bm{\sf u}$ and $\mathbf v}%{\bm{\sf v}$ are subcomonoids of $\mathbf h}%{\bm{\sf h}$, then: \begin{itemize} \item[(i)] $\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}$ is a subcomonoid of $\mathbf h}%{\bm{\sf h}$ and $\mathbf u}%{\bm{\sf u}+\mathbf v}%{\bm{\sf v}\subseteq \mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}$; \item[(ii)] $\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}= \Delta^{-1}\bigl(\mathbf u}%{\bm{\sf u}\bm\cdot(\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}) + (\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v})\bm\cdot\mathbf v}%{\bm{\sf v}\bigr)$. \end{itemize} \end{lemma} \begin{proof} (i) The proofs of the analogous statements for coalgebras given in \cite[Section~3.3]{Abe:1980} extend to this setting. (ii) {}From the definition, $\Delta^{-1}\bigl(\mathbf u}%{\bm{\sf u}\bm\cdot(\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}) + (\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v})\bm\cdot\mathbf v}%{\bm{\sf v}\bigr)\subseteq \mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}$. Now, since $\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}$ is a subcomonoid, \[ \Delta(\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v})\subseteq \bigr((\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v})\bm\cdot(\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v})\bigr)\cap (\mathbf u}%{\bm{\sf u}\bm\cdot\mathbf h}%{\bm{\sf h} + \mathbf h}%{\bm{\sf h}\bm\cdot\mathbf v}%{\bm{\sf v}) = \mathbf u}%{\bm{\sf u}\bm\cdot(\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}) + (\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v})\bm\cdot\mathbf v}%{\bm{\sf v}, \] since $\mathbf u}%{\bm{\sf u},\mathbf v}%{\bm{\sf v}\subseteq \mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}$. This proves the converse inclusion. \end{proof} \begin{lemma}\label{l:lemma 1.5} Let $\mathbf h}%{\bm{\sf h}$ be a Hopf monoid in species and $\mathbf k}%{\bm{\sf k}$ be a submonoid. Let $\mathbf u}%{\bm{\sf u},\mathbf v}%{\bm{\sf v}\subseteq\mathbf h}%{\bm{\sf h}$ be subspecies which are left $\mathbf k}%{\bm{\sf k}$-submodules of $\mathbf h}%{\bm{\sf h}$. Then $\mathbf u}%{\bm{\sf u} \wedge \mathbf v}%{\bm{\sf v}$ is a left $\mathbf k}%{\bm{\sf k}$-submodule of $\mathbf h}%{\bm{\sf h}$. \end{lemma} \begin{proof} Since $\mathbf h}%{\bm{\sf h}$ is a Hopf monoid, the coproduct $\Delta:\mathbf h}%{\bm{\sf h}\to\mathbf h}%{\bm{\sf h}\bm\cdot\mathbf h}%{\bm{\sf h}$ is a morphism of left $\mathbf h}%{\bm{\sf h}$-modules, where $\mathbf h}%{\bm{\sf h}$ acts on $\mathbf h}%{\bm{\sf h}\bm\cdot\mathbf h}%{\bm{\sf h}$ via $\Delta$. Hence it is also a morphism of left $\mathbf k}%{\bm{\sf k}$-modules. By hypothesis, $\mathbf u}%{\bm{\sf u}\bm\cdot\mathbf h}%{\bm{\sf h}+\mathbf h}%{\bm{\sf h}\bm\cdot\mathbf v}%{\bm{\sf v}$ is a left $\mathbf k}%{\bm{\sf k}$-submodule of $\mathbf h}%{\bm{\sf h}\bm\cdot\mathbf h}%{\bm{\sf h}$. Hence, $\mathbf u}%{\bm{\sf u}\wedge\mathbf v}%{\bm{\sf v}=\Delta^{-1}(\mathbf u}%{\bm{\sf u}\bm\cdot\mathbf h}%{\bm{\sf h} + \mathbf h}%{\bm{\sf h}\bm\cdot\mathbf v}%{\bm{\sf v})$ is a left $\mathbf k}%{\bm{\sf k}$-submodule of $\mathbf h}%{\bm{\sf h}$. \end{proof} \begin{lemma}\label{l:lemma 2} Let $\mathbf h}%{\bm{\sf h}$ be a Hopf monoid in species and $\mathbf k}%{\bm{\sf k}$ a Hopf submonoid. Let $\mathbf c}%{\bm{\sf c}$ be a subcomonoid of $\mathbf h}%{\bm{\sf h}$ and a left $\mathbf k}%{\bm{\sf k}$-submodule of $\mathbf h}%{\bm{\sf h}$. Then $(\mathbf k}%{\bm{\sf k} \wedge \mathbf c}%{\bm{\sf c})/\mathbf c}%{\bm{\sf c}$ is a left $\mathbf k}%{\bm{\sf k}$-Hopf module. \end{lemma} \begin{proof} By Lemma~\ref{l:lemma 1.5}, $\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c}$ is a left $\mathbf k}%{\bm{\sf k}$-submodule of $\mathbf h}%{\bm{\sf h}$. Therefore, the quotient $(\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c})/\mathbf c}%{\bm{\sf c}$ by the left $\mathbf k}%{\bm{\sf k}$-submodule $\mathbf c}%{\bm{\sf c}$ is a left $\mathbf k}%{\bm{\sf k}$-module. We next argue that $(\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c})/\mathbf c}%{\bm{\sf c}$ is a $\mathbf k}%{\bm{\sf k}$-comodule. Consider the composite \[ \mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c} \map{\Delta} \mathbf k}%{\bm{\sf k}\bm\cdot(\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c}) + (\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c})\bm\cdot\mathbf c}%{\bm{\sf c} \onto \mathbf k}%{\bm{\sf k}\bm\cdot\bigl(\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c})/\mathbf c}%{\bm{\sf c}, \] where the first map is granted by Lemma~\ref{l:lemma 1} and the second is the projection modulo $\mathbf c}%{\bm{\sf c}$ on the second coordinate. Since $\mathbf c}%{\bm{\sf c}$ is a subcomonoid, the composite factors through $\mathbf c}%{\bm{\sf c}$ and induces \[ (\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c})/\mathbf c}%{\bm{\sf c} \to \mathbf k}%{\bm{\sf k}\bm\cdot\bigl(\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c})/\mathbf c}%{\bm{\sf c}. \] This defines a left $\mathbf k}%{\bm{\sf k}$-comodule structure on $(\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c})/\mathbf c}%{\bm{\sf c}$. Finally, the compatibility between the module and comodule structures on $(\mathbf k}%{\bm{\sf k}\wedge\mathbf c}%{\bm{\sf c})/\mathbf c}%{\bm{\sf c}$ is inherited from the compatibility between the product and coproduct of $\mathbf h}%{\bm{\sf h}$. \end{proof} We are nearly ready for the proof of the main result. First, recall the \demph{coradical filtration} of a connected comonoid in species \cite[Section~8.10]{AguMah:2010}. Given a connected comonoid $\mathbf c}%{\bm{\sf c}$, define subspecies $\mathbf c}%{\bm{\sf c}_{(n)}$ by \[ \mathbf c}%{\bm{\sf c}_{(0)}=\mathbf{1} \quad\text{and}\quad \mathbf c}%{\bm{\sf c}_{(n)} = \mathbf c}%{\bm{\sf c}_{(0)} \wedge \mathbf c}%{\bm{\sf c}_{(n-1)} \text{ \ for all $n\geq 1$.} \] We then have \[ \mathbf c}%{\bm{\sf c}_{(0)} \subseteq \mathbf c}%{\bm{\sf c}_{(1)} \subseteq \dotsb\subseteq \mathbf c}%{\bm{\sf c}_{(n)}\subseteq \dotsb \mathbf c}%{\bm{\sf c} \quad\text{and}\quad \mathbf c}%{\bm{\sf c}= \bigcup_{n\geq 0} \mathbf c}%{\bm{\sf c}_{(n)}. \] \begin{proof}[Proof of Theorem~\ref{t:main}] We show that there is a species $\mathbf q}%{\bm{\sf q}$ such that $\mathbf h}%{\bm{\sf h} \cong \mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q}$ as left $\mathbf k}%{\bm{\sf k}$-modules. As in the proof of Theorem~\ref {t:connected}, it then follows that $\mathbf q}%{\bm{\sf q}\cong \mathbf h}%{\bm{\sf h}/\mathbf k}%{\bm{\sf k}_+\mathbf h}%{\bm{\sf h}$. Define a sequence $\mathbf k}%{\bm{\sf k}^{(n)}$ of subspecies of $\mathbf h}%{\bm{\sf h}$ by \[ \mathbf k}%{\bm{\sf k}^{(0)}=\mathbf k}%{\bm{\sf k} \quad\text{and}\quad \mathbf k}%{\bm{\sf k}^{(n)} = \mathbf k}%{\bm{\sf k} \wedge \mathbf k}%{\bm{\sf k}^{(n-1)} \text{ \ for all $n\geq 1$.} \] Each $\mathbf k}%{\bm{\sf k}^{(n)}$ is a subcomonoid and a left $\mathbf k}%{\bm{\sf k}$-submodule of $\mathbf h}%{\bm{\sf h}$. This follows from Lemmas~\ref{l:lemma 1} and~\ref{l:lemma 1.5} by induction on $n$. Then Lemma~\ref{l:lemma 2} provides a left $\mathbf k}%{\bm{\sf k}$-Hopf module structure on the quotient species $\mathbf k}%{\bm{\sf k}^{(n)}/\mathbf k}%{\bm{\sf k}^{(n-1)}$ for all $n\geq1$. Hence $\mathbf k}%{\bm{\sf k}^{(n)}/\mathbf k}%{\bm{\sf k}^{(n-1)}$ is a free left $\mathbf k}%{\bm{\sf k}$-module, by Proposition~\ref{p:FTHM}. We claim that there exists a sequence of species $\mathbf q}%{\bm{\sf q}_n$ such that \[ \mathbf k}%{\bm{\sf k}^{(n)} \cong \mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q}_n \] as left $\mathbf k}%{\bm{\sf k}$-modules for all $n\geq 0$; that is, each $\mathbf k}%{\bm{\sf k}^{(n)}$ is a free left $\mathbf k}%{\bm{\sf k}$-module. Moreover, we claim that the $\mathbf q}%{\bm{\sf q}_n$ can be chosen so that \[ \mathbf q}%{\bm{\sf q}_0\subseteq \mathbf q}%{\bm{\sf q}_1\ \subseteq \dotsb\subseteq \mathbf q}%{\bm{\sf q}_n \subseteq\dotsb \] and the above isomorphisms are compatible with the inclusions $\mathbf q}%{\bm{\sf q}_{n-1}\subseteq\mathbf q}%{\bm{\sf q}_n$ and $\mathbf k}%{\bm{\sf k}^{(n-1)}\subseteq\mathbf k}%{\bm{\sf k}^{(n)}$. We prove the claims by induction on $n\geq 0$. We start by letting $\mathbf q}%{\bm{\sf q}_0=\mathbf{1}$. For $n\geq 1$, we have \[ \mathbf k}%{\bm{\sf k}^{(n-1)}\cong \mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q}_{n-1} \quad\text{and}\quad \mathbf k}%{\bm{\sf k}^{(n)}/\mathbf k}%{\bm{\sf k}^{(n-1)}\cong \mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q}'_n \] for some species $ \mathbf q}%{\bm{\sf q}'_n$ (the former by induction hypothesis and the latter by the above argument). Let \[ \mathbf q}%{\bm{\sf q}_{n} = \mathbf q}%{\bm{\sf q}_{n-1}\oplus \mathbf q}%{\bm{\sf q}'_n. \] By choosing an arbitrary $\mathbf k}%{\bm{\sf k}$-module section of the map $\mathbf k}%{\bm{\sf k}^{(n)}\onto\mathbf k}%{\bm{\sf k}^{(n)}/\mathbf k}%{\bm{\sf k}^{(n-1)}\cong \mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q}'_n$ (possible by freeness), we obtain an isomorphism \[ \mathbf k}%{\bm{\sf k}^{(n)}\cong \mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q}_{n} \] extending the isomorphism $\mathbf k}%{\bm{\sf k}^{(n-1)}\cong \mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q}_{n-1}$. This proves the claims. We utilize the coradical filtration of $\mathbf h}%{\bm{\sf h}$ to finish the proof. Since $\mathbf h}%{\bm{\sf h}$ is connected, $\mathbf h}%{\bm{\sf h}_{(0)}=\mathbf{1} \subseteq \mathbf k}%{\bm{\sf k}= \mathbf k}%{\bm{\sf k}^{(0)}$, and by induction, \[ \mathbf h}%{\bm{\sf h}_{(n)} \subseteq \mathbf k}%{\bm{\sf k}^{(n)} \quad\hbox{for all }n\geq0. \] Hence, \[ \mathbf h}%{\bm{\sf h}= \bigcup_{n\geq 0} \mathbf h}%{\bm{\sf h}_{(n)} =\bigcup_{n\geq 0} \mathbf k}%{\bm{\sf k}^{(n)} \cong\bigcup_{n\geq 0} \mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q}_{n} \cong\mathbf k}%{\bm{\sf k}\bm\cdot \mathbf q}%{\bm{\sf q} \text{ \ where \ } \mathbf q}%{\bm{\sf q}=\bigcup_{n\geq 0} \mathbf q}%{\bm{\sf q}_n. \] Thus $\mathbf h}%{\bm{\sf h}$ is free as a left $\mathbf k}%{\bm{\sf k}$-module. \end{proof} Let $\pi:\mathbf h}%{\bm{\sf h}\to\mathbf k}%{\bm{\sf k}$ be a morphism of Hopf monoids. The \demph{right Hopf kernel} of $\pi$ is the species defined by \begin{equation}\label{e:hker} \mathrm{Hker}(\pi)=\ker\bigl(\mathbf h}%{\bm{\sf h} \map{\Delta} \mathbf h}%{\bm{\sf h}\bm\cdot\mathbf h}%{\bm{\sf h}\map{\pi_{+}\bm\cdot\,\mathrm{id}} \mathbf k}%{\bm{\sf k}_{+\!}\bm\cdot\mathbf h}%{\bm{\sf h}\bigr), \end{equation} where $\pi_+:\mathbf h}%{\bm{\sf h}\to\mathbf k}%{\bm{\sf k}_+$ is $\pi$ followed by the canonical projection $\mathbf k}%{\bm{\sf k}\onto\mathbf k}%{\bm{\sf k}_+$. For the following result, we employ duality for Hopf monoids~\cite[Section~8.6.2]{AguMah:2010}. (We assume all species are finite dimensional.) \begin{theorem}\label{t:maindual} Let $\mathbf h}%{\bm{\sf h}$ be a connected Hopf monoid in the category of species and $\mathbf k}%{\bm{\sf k}$ a quotient Hopf monoid via a morphism $\pi:\mathbf h}%{\bm{\sf h}\onto\mathbf k}%{\bm{\sf k}$. Then $\mathbf h}%{\bm{\sf h}$ is a cofree left $\mathbf k}%{\bm{\sf k}$-comodule. Moreover, \[ \mathbf h}%{\bm{\sf h}\cong \mathbf k}%{\bm{\sf k}\bm\cdot \mathrm{Hker}(\pi) \] as left $\mathbf k}%{\bm{\sf k}$-comodules (and as species). \end{theorem} \begin{proof} By duality, $\mathbf k}%{\bm{\sf k}^*$ is a Hopf submonoid of $\mathbf h}%{\bm{\sf h}^*$, so $\mathbf h}%{\bm{\sf h}^*\cong\mathbf k}%{\bm{\sf k}^*\bm\cdot(\mathbf h}%{\bm{\sf h}^*/\mathbf k}%{\bm{\sf k}^*_+\mathbf h}%{\bm{\sf h}^*)$ by Theorem~\ref{t:main}. Dualizing again we obtain the result, since \[ \mathbf h}%{\bm{\sf h}^*/\mathbf k}%{\bm{\sf k}^*_+\mathbf h}%{\bm{\sf h}^*=\mathrm{coker}\bigl(\mathbf k}%{\bm{\sf k}^*_{+\!}\bm\cdot\mathbf h}%{\bm{\sf h}^* \map{\pi^*_+\bm\cdot\,\mathrm{id}} \mathbf h}%{\bm{\sf h}^{*\!}\bm\cdot\mathbf h}%{\bm{\sf h}^* \map{\Delta^*} \mathbf h}%{\bm{\sf h}^* \bigr). \qedhere \] \end{proof} \section{Applications and examples}\label{s:applications} \subsection{A test for Hopf submonoids} Two invariants associated to a (finite dimensional) species $\mathbf q}%{\bm{\sf q}$ are the \demph{exponential generating series} $\exponential{\mathbf q}%{\bm{\sf q}}{x}$ and the \demph{type generating series} $\type{\mathbf q}%{\bm{\sf q}}{x}$. They are given by \[ \exponential{\mathbf q}%{\bm{\sf q}}{x} = \sum_{n\geq 0} \dim \mathbf q}%{\bm{\sf q}[n]\, \frac{x^n}{n!} \quad\text{and}\quad \type{\mathbf q}%{\bm{\sf q}}{x} = \sum_{n\geq 0} \dim \mathbf q}%{\bm{\sf q}[n]_{S_n}\, x^n, \] where \[ \mathbf q}%{\bm{\sf q}[n]_{S_n} =\mathbf q}%{\bm{\sf q}[n]/\Bbbk\{v-\sigma\cdot v \mid v\in \mathbf q}%{\bm{\sf q}[n],\ \sigma\in S_n\}. \] Both are specializations of the \demph{cycle index series} $\cycle{\mathbf q}%{\bm{\sf q}}{x_1,x_2,\dotsc}$; see \cite[Section~1.2]{BerLabLer:1998} for the definition. Specifically, \begin{gather*} \exponential{\mathbf q}%{\bm{\sf q}}{x} = \cycle{\mathbf q}%{\bm{\sf q}}{x,0,0,\dotsc} \quad\text{and}\quad \type{\mathbf q}%{\bm{\sf q}}{x} = \cycle{\mathbf q}%{\bm{\sf q}}{x,x^2,x^3,\dotsc} . \end{gather*} The cycle index series is multiplicative under Cauchy product \cite[Section~1.3]{BerLabLer:1998}: if $\mathbf h}%{\bm{\sf h} = \mathbf k}%{\bm{\sf k} \bm\cdot \mathbf q}%{\bm{\sf q}$, then $\cycle{\mathbf h}%{\bm{\sf h}}{x_1,x_2,\dotsc} = \cycle{\mathbf k}%{\bm{\sf k}}{x_1,x_2,\dotsc} \, \cycle{\mathbf q}%{\bm{\sf q}}{x_1,x_2,\dotsc}$. By specialization, the same is true for the exponential and type generating series. Let $\mathbb{Q}_{\geq 0}$ denote the nonnegative rational numbers. An immediate consequence of Theorems~\ref{t:main} and~\ref{t:maindual} is the following. \begin{corollary}\label{c:Sp series} Let $\mathbf h}%{\bm{\sf h}$ and $\mathbf k}%{\bm{\sf k}$ be connected Hopf monoids in species. Suppose $\mathbf k}%{\bm{\sf k}$ is either a Hopf submonoid or a quotient Hopf monoid of $\mathbf h}%{\bm{\sf h}$. Then the quotient of cycle index series $\cycle{\mathbf h}%{\bm{\sf h}}{x_1,x_2,\dotsc} / \cycle{\mathbf k}%{\bm{\sf k}}{x_1,x_2,\dotsc}$ is nonnegative, i.e., belongs to $\mathbb{Q}_{\geq 0}{[\![} x_1,x_2,\dotsc {]\!]}$. In particular, the quotients $\exponential{\mathbf h}%{\bm{\sf h}}{x}/\exponential{\mathbf k}%{\bm{\sf k}}{x}$ and $\type{\mathbf h}%{\bm{\sf h}}{x}/\type{\mathbf k}%{\bm{\sf k}}{x}$ are also nonnegative. \end{corollary} Given a connected Hopf monoid $\mathbf h}%{\bm{\sf h}$ in species, we may use Corollary~\ref{c:Sp series} to determine if a given species $\mathbf k}%{\bm{\sf k}$ may be a Hopf submonoid (or a quotient Hopf monoid). \begin{example}\label{ex: submonoid} A \demph{partition} of a set $I$ is an unordered collection of disjoint nonempty subsets of $I$ whose union is $I$. The notation $ab{\textcolor{mymagenta}{.}} c$ is shorthand for the partition $\bigl\{\{a,b\},\,\{c\}\bigr\}$ of $\{a,b,c\}$. Let $\bm\Pi$ be the species of set partitions, i.e., $\bm\Pi[I]$ is the vector space with basis the set of all partitions of $I$. Let $\bm\Pi'$ denote the subspecies linearly spanned by set partitions with distinct block sizes. For example, \[ \bm\Pi[a,b,c] = \Bbbk\bigl\{ abc, a{\textcolor{mymagenta}{.}} bc, ab{\textcolor{mymagenta}{.}} c, a{\textcolor{mymagenta}{.}} bc, a{\textcolor{mymagenta}{.}} b{\textcolor{mymagenta}{.}} c \bigr\} \quad\text{and}\quad \bm\Pi'[a,b,c] = \Bbbk\bigl\{ abc, a{\textcolor{mymagenta}{.}} bc, ab{\textcolor{mymagenta}{.}} c, a{\textcolor{mymagenta}{.}} bc\bigr\}. \] The sequences $(\dim\bm\Pi[n])_{n\geq 0}$ and $(\dim\bm\Pi'[n])_{n\geq 0}$ appear in \cite{Slo:oeis} as A000110 and A007837, respectively. We have \[ \exponential{\bm\Pi}{x} = \exp\bigl(\exp(x)-1\bigr) = 1+x+x^2+\frac{5}{6}x^3+\frac{5}{8}x^4+\dotsb \] and \[ \exponential{\bm\Pi'}{x} = \prod_{n\geq 1}\bigl(1+\frac{x^n}{n!}\bigr) = 1+x+\frac{1}{2}x^2+\frac{2}{3}x^3+\frac{5}{24}x^4+\dotsb \,. \] A Hopf monoid structure on $\bm\Pi$ is defined in \cite[Section~12.6]{AguMah:2010}. There are many linear bases of $\bm\Pi$ indexed by set partitions, and many ways to embed $\bm\Pi'$ as a subspecies of $\bm\Pi$. Is it possible to embed $\bm\Pi'$ as a Hopf submonoid of $\bm\Pi$? We have \[ \exponential{\bm\Pi}{x} \big/ \exponential{\bm\Pi'}{x} = 1+\frac12 x^2 - \frac1{3} x^3 + \frac{1}{2} x^4 - \frac{11}{30} x^5 + \dotsb \,, \] so the answer is negative by Corollary~\ref{c:Sp series}. In fact, it is not possible to embed $\bm\Pi'$ as a Hopf submonoid of $\bm\Pi$ for any potential Hopf monoid structure on $\bm\Pi$. We remark that the type generating series quotient for the pair of species in Example~\ref{ex: submonoid} is nonnegative: \begin{align*} \type{{\bm\Pi}}{x} \ &= \ 1+x+2x^2+3x^3+5x^4+7x^5+11x^6+15x^7 + \dotsb \,, \\ \type{{\bm\Pi'}}{x} \ &= \ 1 + x+x^2+2x^3+2x^4+3x^5+4x^6+5x^7 + \dotsb \,, \\ \type{{\bm\Pi}}{x} \big/ \type{{\bm\Pi'}}{x} \ &= \ 1+x^2+2x^4+3x^6+5x^8+7x^{10} + \dotsb \,. \end{align*} This can be understood by appealing to the Hopf algebra $\textsl{Sym}$ of symmetric functions. A basis for its homogenous component of degree $n$ is indexed by integer partitions of $n$, so $\poincare{\textsl{Sym}}{x} = \type{{\bm\Pi}}{x}$. Moreover, $\type{{\bm\Pi'}}{x}$ enumerates the integer partitions with odd part sizes and $\textsl{Sym}$ does indeed contain a Hopf subalgebra with this Poincar\'e series. It is the algebra of Schur $Q$-functions. See \cite[III.8]{Mac:1995}. Thus $\type{{\bm\Pi}}{x} \big/ \type{{\bm\Pi'}}{x}$ is nonnegative by Corollary~\ref{c:cgVec series}. \end{example} \subsection{Tests for Hopf monoids}\label{ss:ord-exp} Let $(a_n)_{n\geq 0}$ be a sequence of nonnegative integers with $a_0=1$. Does there exist a connected Hopf monoid $\mathbf h}%{\bm{\sf h}$ with $\dim \mathbf h}%{\bm{\sf h}[n]=a_n$ for all $n$? The next result provides conditions that the sequence $(a_n)_{n\geq 0}$ must satisfy in order for this to be the case. The proof makes use of the \demph{Hadamard product\/} of Hopf monoids~\cite[Sections~8.1 and~8.13]{AguMah:2010}. If $\mathbf h}%{\bm{\sf h}$ and $\mathbf k}%{\bm{\sf k}$ are Hopf monoids, so is $\mathbf h}%{\bm{\sf h}\times\mathbf k}%{\bm{\sf k}$, with $(\mathbf h}%{\bm{\sf h}\times\mathbf k}%{\bm{\sf k})[I] = \mathbf h}%{\bm{\sf h}[I]\otimes \mathbf k}%{\bm{\sf k}[I]$ for each finite set $I$. The exponential species $\mathbf E}%{\bm{\sf l}$ is the unit element for the Hadamard product. \begin{corollary}[The (ord/exp)-test]\label{c:ord-exp} For any connected Hopf monoid in species $\mathbf h}%{\bm{\sf h}$, \[ \Bigl(\sum_{n\geq 0} \dim \mathbf h}%{\bm{\sf h}[n]\, x^n\Bigr) \Big/ \Bigl(\sum_{n\geq 0} \frac{\dim \mathbf h}%{\bm{\sf h}[n]}{n!} \, x^n\Bigr) \in \mathbb{Q}_{\geq 0}{[\![} x{]\!]}. \] \end{corollary} \begin{proof} Consider the canonical morphism of Hopf monoids $\mathbf L}%{\bm{\sf l} \onto \mathbf E}%{\bm{\sf l}$~\cite[Section~8.5]{AguMah:2010}; it maps any linear order $\ell\in\mathbf L}%{\bm{\sf l}[I]$ to the basis element $\ast_I\in\mathbf E}%{\bm{\sf l}[I]$. The Hadamard product then yields a morphism of Hopf monoids \[ \mathbf L}%{\bm{\sf l}\times\mathbf h}%{\bm{\sf h} \onto \mathbf E}%{\bm{\sf l}\times\mathbf h}%{\bm{\sf h}\cong\mathbf h}%{\bm{\sf h}. \] By Corollary~\ref {c:Sp series}, $\exponential{\mathbf L}%{\bm{\sf l}\times\mathbf h}%{\bm{\sf h}}{x}/\exponential{\mathbf h}%{\bm{\sf h}}{x}\in \mathbb{Q}_{\geq 0}{[\![} x{]\!]}$. Since $\exponential{\mathbf L}%{\bm{\sf l}\times\mathbf h}%{\bm{\sf h}}{x}=\sum_{n\geq 0} \dim \mathbf h}%{\bm{\sf h}[n]\, x^n$, the result follows. \end{proof} Let $a_n=\dim \mathbf h}%{\bm{\sf h}[n]$. Corollary~\ref{c:ord-exp} states that the ratio of the ordinary to the exponential generating function of the sequence $(a_n)_{n\geq 0}$ must be nonnegative. This translates into a sequence of polynomial inequalities, the first of which are as follows: \begin{equation}\label{e:ord-exp} 5a_3\geq 3a_2a_1, \quad 23a_4+12a_2a_1^2 \geq 20a_3a_1+6a_2^2. \end{equation} In particular, not every nonnegative sequence arises as the dimension sequence of a Hopf monoid. The following test is of a similar nature, but involves the type instead of the exponential generating function. The conditions then depend not just on the dimension sequence of $\mathbf h}%{\bm{\sf h}$, but also on its species structure. \begin{corollary}[The (ord/type)-test]\label{c:full-bosonic} For any connected Hopf monoid in species $\mathbf h}%{\bm{\sf h}$, \[ \Bigl(\sum_{n\geq 0} \dim \mathbf h}%{\bm{\sf h}[n]\, x^n\Bigr) \Big/ \Bigl(\sum_{n\geq 0} \dim \mathbf h}%{\bm{\sf h}[n]_{S_n}\, x^n\Bigr) \in \mathbb{N}{[\![} x{]\!]}. \] \end{corollary} \begin{proof} We argue as in the proof of Corollary~\ref{c:ord-exp}, using type generating functions instead. Since we have $\type{\mathbf L}%{\bm{\sf l}\times\mathbf h}%{\bm{\sf h}}{x}= \sum_{n\geq 0} \dim \mathbf h}%{\bm{\sf h}[n]\, x^n$, the result follows. \end{proof} \begin{remark} The previous result may also be derived as follows. According to~\cite[Chapter 15]{AguMah:2010}, associated to the Hopf monoid $\mathbf h}%{\bm{\sf h}$ there are two graded Hopf algebras $\mathcal K(\mathbf h}%{\bm{\sf h})$ and $\overline{\K}(\mathbf h}%{\bm{\sf h})$, as well as a surjective morphism \[ \mathcal K(\mathbf h}%{\bm{\sf h}) \onto \overline{\K}(\mathbf h}%{\bm{\sf h}). \] Moreover, the Poincar\'e series for these Hopf algebras are \[ \poincare{\mathcal K(\mathbf h}%{\bm{\sf h})}{x} = \sum_{n\geq 0} \dim \mathbf h}%{\bm{\sf h}[n]\, x^n \quad\text{and}\quad \poincare{\overline{\K}(\mathbf h}%{\bm{\sf h})}{x} = \sum_{n\geq 0} \dim \mathbf h}%{\bm{\sf h}[n]_{S_n}\, x^n. \] Corollary~\ref{c:full-bosonic} now follows from (the dual form of) Corollary~\ref{c:cgVec series}. \end{remark} \subsection{Additional tests for Hopf monoids}\label{ss:addtests} The method of Section~\ref{ss:ord-exp} can be applied in multiple situations in order to deduce additional inequalities that the dimension sequence of a connected Hopf monoid must satisfy. We illustrate this next. Let $k$ be a fixed nonnegative integer. Let $\mathbf E}%{\bm{\sf l}^{\bm\cdot k}$ denote the $k$-th Cauchy power of the exponential species $\mathbf E}%{\bm{\sf l}$. The space $\mathbf E}%{\bm{\sf l}^{\bm\cdot k}[I]$ has a basis consisting of functions $f:I\to [k]$. The species $\mathbf E}%{\bm{\sf l}^{\bm\cdot k}$ carries a Hopf monoid structure~\cite[Examples~8.17 and~8.18]{AguMah:2010} and any fixed inclusion $[k]\hookrightarrow[k{+}1]$ gives rise to an injective morphism of Hopf monoids $\mathbf E}%{\bm{\sf l}^{\bm\cdot k}\hookrightarrow\mathbf E}%{\bm{\sf l}^{\bm\cdot (k+1)}$. Employing the Hadamard product as in Section~\ref{ss:ord-exp}, we obtain an injective morphism of Hopf monoids \[ \mathbf E}%{\bm{\sf l}^{\bm\cdot k}\times\mathbf h}%{\bm{\sf h} \hookrightarrow \mathbf E}%{\bm{\sf l}^{\bm\cdot (k+1)}\times\mathbf h}%{\bm{\sf h} \] where $\mathbf h}%{\bm{\sf h}$ is an arbitrary connected Hopf monoid. From the nonnegativity of the first coefficients of $\exponential{\mathbf E}%{\bm{\sf l}^{\bm\cdot (k+1)}\times\mathbf h}%{\bm{\sf h}}{x} \big/ \exponential{\mathbf E}%{\bm{\sf l}^{\bm\cdot k}\times\mathbf h}%{\bm{\sf h}}{x}$ we obtain \[ (2k+1)a_2 \geq 2k a_1^2 \quad\text{and}\quad (3k^2+3k+1)a_3 \geq 3(3k^2+k)a_2a_1 - 6k^2a_1^3. \] These inequalities hold for every $k\in\mathbb{N}$. Letting $k\to\infty$ we deduce \begin{equation}\label{e:addcond} a_2\geq a_1^2 \quad\text{and}\quad a_3\geq 3a_2a_1-2a_1^3. \end{equation} \begin{example} Consider the species $\bel$ \demph{of elements}. The set $I$ is a basis of the space $\bel[I]$, so the dimension sequence of $\bel$ is $a_n=n$. This sequence does not satisfy the second inequality in~\eqref{e:addcond}. Therefore, the species $\bel$ does not carry any Hopf monoid structure. \end{example} \subsection{A test for Hopf monoids over $\mathbf E}%{\bm{\sf l}$} Our next result is a necessary condition for a Hopf monoid in species to contain or surject onto the exponential species $\mathbf E}%{\bm{\sf l}$. Given a sequence $(a_n)_{n\geq0}$, its \demph{binomial transform} $(b_n)_{n\geq0}$ is defined by \[ b_n := \sum_{i=0}^n \binom{n}{i}(-1)^{i}\,a_{n-i}. \] \begin{corollary}[The $\mathbf E}%{\bm{\sf l}$-test] \label{c:e-test} Suppose $\mathbf h}%{\bm{\sf h}$ is a connected Hopf monoid that either contains the species $\mathbf E}%{\bm{\sf l}$ or surjects onto $\mathbf E}%{\bm{\sf l}$ (in both cases as a Hopf monoid). Let $a_n=\dim \mathbf h}%{\bm{\sf h}[n]$ and $\overline{a}_n=\dim \mathbf h}%{\bm{\sf h}[n]_{S_n}$. Then the binomial transform of $(a_n)_{n\geq0}$ must be nonnegative and $(\overline{a}_n)_{n\geq0}$ must be nondecreasing. \end{corollary} More plainly, in this setting, we must have the following inequalities: \[ a_1\geq a_0, \quad a_2\geq 2a_1-a_0, \quad a_3\geq 3a_2-3a_1+a_0, \ \ \dotsc \] and $\overline{a}_n\geq \overline{a}_{n-1}$ for all $n\geq 1$. \begin{proof} By Corollary~\ref{c:Sp series}, the quotient $\exponential{\mathbf h}%{\bm{\sf h}}{x} / \exponential{\mathbf E}%{\bm{\sf l}}{x}$ is nonnegative. But $\exponential{\mathbf E}%{\bm{\sf l}}{x} = \exp(x)$, so the quotient is given by \[ b_0 + b_1 x + b_2 \frac{x^2}{2} + b_3 \frac{x^3}{3!} + \dotsb \,, \] where $(b_n)_{n\geq0}$ is the binomial transform of $(a_n)_{n\geq0}$. Replacing exponential for type generating functions yields the result for $(\overline{a}_n)_{n\geq0}$, since $\type{\mathbf E}%{\bm{\sf l}}{x} = \frac{1}{1-x}$. \end{proof} \begin{remark} Myhill's theory of \demph{combinatorial functions}~\cite{Dek:1990,Myh:1958} provides necessary and sufficient conditions that a sequence $(a_n)_{n\geq0}$ must satisfy in order for its binomial transform to be nonnegative: the sequence must arise from a particular type of operator defined on finite sets. Work of Menni~\cite{Men:2009} expands on this from a categorical perspective. It would be interesting to relate these ideas to the ones of this paper. \end{remark} We make a further remark regarding connected \demph{linearized} Hopf monoids. These are Hopf monoids of a set theoretic nature. See \cite[Section~8.7]{AguMah:2010} for details. Briefly, there are set maps $\mu_{S,T}:\mathrm{H}[S]\times\mathrm{H}[T]\to\mathrm{H}[I]$ and $\Delta_{S,T}:\mathrm{H}[I]\to\mathrm{H}[S]\times\mathrm{H}[T]$ which produce a Hopf monoid in (vector) species when the set species $\mathrm{H}$ is linearized. It follows that if $\mathbf h}%{\bm{\sf h}$ is a linearized Hopf monoid other than the trivial Hopf monoid $\mathbf{1}$, then there is a morphism of Hopf monoids from $\mathbf h}%{\bm{\sf h}$ onto $\mathbf E}%{\bm{\sf l}$. Thus, Corollary~\ref{c:e-test} provides a test for existence of a linearized Hopf monoid structure on $\mathbf h}%{\bm{\sf h}$. \begin{example} We return to the species $\bm\Pi'$ of set partitions into distinct block sizes. We might ask if this can be made into a linearized Hopf monoid in some way (after Example~\ref{ex: submonoid}, this would \emph{not} be as a Hopf submonoid of $\bm\Pi$). With $a_n$ and $b_n$ as above, we have: \begin{align*} (a_n)_{n\geq0} \ &= \ 1, \ 1, \ 1, \ 4, \ \ 5, \ 16, \ \ 82, \ \ 169, \ 541, \dotsc \,, \\[.5ex] (b_n)_{n\geq0} \ &= \ 1, \ 0, \ 0, \ 3, \, -8,\ 25, \, -9, \, -119, \ 736, \dotsc \,. \end{align*} Thus $\bm\Pi'$ fails the $\mathbf E}%{\bm{\sf l}$-test and the answer to the above question is negative. \end{example} \subsection{A test for Hopf monoids over $\mathbf L}%{\bm{\sf l}$} Let $\mathbf h}%{\bm{\sf h}$ be a connected Hopf monoid in species. Let $a_n=\dim \mathbf h}%{\bm{\sf h}[n]$ and $\overline{a}_n=\dim \mathbf h}%{\bm{\sf h}[n]_{S_n}$. Note that the analogous integers for the species $\mathbf L}%{\bm{\sf l}$ of linear orders are $b_n=n!$ and $\overline{b}_n=1$. Now suppose that $\mathbf h}%{\bm{\sf h}$ contains $\mathbf L}%{\bm{\sf l}$ or surjects onto $\mathbf L}%{\bm{\sf l}$ as a Hopf monoid. An obvious necessary condition for this situation is that $a_n\geq n!$ and $\overline{a}_n\geq 1$. Our next result provides stronger conditions. \begin{corollary}[The $\mathbf L}%{\bm{\sf l}$-test] \label{c:l-test} Suppose $\mathbf h}%{\bm{\sf h}$ is a connected Hopf monoid that either contains the species $\mathbf L}%{\bm{\sf l}$ or surjects onto $\mathbf L}%{\bm{\sf l}$ (in both cases as a Hopf monoid). If $a_n=\dim \mathbf h}%{\bm{\sf h}[n]$ and $\overline{a}_n=\dim \mathbf h}%{\bm{\sf h}[n]_{S_n}$, then \[ a_n\geq n a_{n-1} \quad\text{and}\quad \overline{a}_n\geq \overline{a}_{n-1} \quad(\forall\,n\geq1). \] \end{corollary} \begin{proof} It follows from Corollary~\ref{c:Sp series} that both $\exponential{\mathbf h}%{\bm{\sf h}}{x} / \exponential{\mathbf L}%{\bm{\sf l}}{x}$ and $\type{\mathbf h}%{\bm{\sf h}}{x} / \type{\mathbf L}%{\bm{\sf l}}{x}$ are nonnegative. These yield the first and second set of inequalities, respectively. \end{proof} Before giving an application of the corollary, we introduce a new Hopf monoid in species. A \demph{composition} of a set $I$ is an ordered collection of disjoint nonempty subsets of $I$ whose union is $I$. The notation $ab{\textcolor{mymagenta}{|}} c$ is shorthand for the composition $\bigl(\{a,b\},\,\{c\}\bigr)$ of $\{a,b,c\}$. Let $\mathbf{Pal}$ denote the species of set compositions whose sequence of block sizes is palindromic. So, for instance, \begin{gather*} \mathbf{Pal}[a,b] = \Bbbk\bigl\{ ab, \ a{\textcolor{mymagenta}{|}} b, \ b{\textcolor{mymagenta}{|}} a\bigr\}\\ \intertext{and} \mathbf{Pal}[a,b,c,d,e] = \Bbbk\bigl\{abcde, \ a{\textcolor{mymagenta}{|}} bcd{\textcolor{mymagenta}{|}} e, \ ab{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} de, \ a{\textcolor{mymagenta}{|}} b{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} e, \ \dotsc \bigr\}. \end{gather*} The latter space has dimension $171=1 + 5\binom{4}{3} + \binom{5}{2}3 + 5!$ while $\dim \mathbf{Pal}[5]_{S_5}=4$ (accounting for the four types of palindromic set compositions shown above). Given a palindromic set composition $F=F_1{\textcolor{mymagenta}{|}} \dotsb{\textcolor{mymagenta}{|}} F_r$, we view it as a triple $F=(F^-,F^0,F^+)$, where $F^-$ is the initial sequence of blocks, $F^0$ is the central block if this exists (if the number of blocks is odd) and otherwise it is the empty set, and $F^+$ is the final sequence of blocks. That is, \[ F^-=F_1{\textcolor{mymagenta}{|}} \dotsb {\textcolor{mymagenta}{|}} F_{\lfloor r/2\rfloor}, \qquad\quad F^0 = \begin{cases} F_{\lfloor r/2\rfloor+1} & \hbox{if $r$ is odd,}\\ \emptyset & \hbox{if $r$ is even,} \end{cases} \qquad\quad F^+ = F_{\lceil r/2+1\rceil}{\textcolor{mymagenta}{|}} \dotsb {\textcolor{mymagenta}{|}} F_{r}. \] Given $S\subseteq I$, let \[ F\vert_S := F_1\cap S\,{\textcolor{mymagenta}{|}}\, F_2\cap S\,{\textcolor{mymagenta}{|}}\, \dotsb \,{\textcolor{mymagenta}{|}}\, F_r\cap S\,, \] where empty intersections are deleted. Then $F\vert_S$ is a composition of $S$. Let us say that $S$ is \demph{admissible} for $F$ if \[ \#\bigl( F_i\cap S\bigr) = \# \bigl(F_{r+1-i}\cap S\bigr) \ \text{ \ for all $i=1,\ldots,r$.} \] In this case, both $F\vert_S$ and $F\vert_{I\setminus S}$ are palindromic. We employ the above notation to define product and coproduct operations on $\mathbf{Pal}$. Fix a decomposition $I = S\sqcup T$. \\ {\it Product.\ } Given palindromic set compositions $F\in\mathbf{Pal}[S]$ and $G\in\mathbf{Pal}[T]$, we put \[ \mu_{S,T}(F \otimes G) := \bigl(F^-{\textcolor{mymagenta}{|}} G^-, F^0\cup G^0, G^+{\textcolor{mymagenta}{|}} F^+\bigr). \] In other words, we concatenate the initial sequences of blocks of $F$ and $G$ in that order, merge their central blocks, and concatenate their final sequences in the opposite order. The result is a palindromic composition of $I$. For example, with $S=\{a,b\}$ and $T=\{c,d,e,f\}$, \[ (a{\textcolor{mymagenta}{|}} b) \otimes (c{\textcolor{mymagenta}{|}} de{\textcolor{mymagenta}{|}} f) \mapsto a{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} de{\textcolor{mymagenta}{|}} f{\textcolor{mymagenta}{|}} b. \] {\it Coproduct.\ } Given a palindromic set composition $F\in\mathbf{Pal}[I]$, we put \[ \Delta_{S,T}(F) := \begin{cases} F\vert_S \otimes F\vert_T & \hbox{if $S$ is admissible for $F$,} \\ 0 & \hbox{otherwise.} \end{cases} \] For example, with $S$ and $T$ as above, \[ ad{\textcolor{mymagenta}{|}} b{\textcolor{mymagenta}{|}} e{\textcolor{mymagenta}{|}} cf \mapsto 0 \quad\text{and}\quad e{\textcolor{mymagenta}{|}} abcd{\textcolor{mymagenta}{|}} f \mapsto (ab) \otimes (e{\textcolor{mymagenta}{|}} cd{\textcolor{mymagenta}{|}} f). \] These operations endow $\mathbf{Pal}$ with the structure of Hopf monoid, as may be easily checked. \begin{example} A linear order may be seen as a palindromic set composition (with singleton blocks). Both Hopf monoids $\mathbf{Pal}$ and $\mathbf L}%{\bm{\sf l}$ are cocommutative and not commutative. We may then ask if $\mathbf{Pal}$ contains (or surjects onto) $\mathbf L}%{\bm{\sf l}$ as a Hopf monoid. Writing $a_n=\dim\mathbf{Pal}[n]$, we have: \begin{gather*} (a_n)_{n\geq0} \ = \ 1, \ 1, \ 3, \ 7, \ 43, \ 171, \ 1581, \ 8793, \ 108347, \dotsc \,. \end{gather*} However, \begin{gather*} (a_n - na_{n-1})_{n\geq1} \ = \ 0, \ 1, \ -2, \ 15, \ -44, \ 555, \ -2274, \ \ 38003, \dotsc \,, \end{gather*} so $\mathbf{Pal}$ fails the $\mathbf L}%{\bm{\sf l}$-test and the answer to the above question is negative. \end{example} \subsection{Examples of nonnegative quotients} We comment on a few examples where the quotient power series $\exponential{\mathbf h}%{\bm{\sf h}}{x}/\exponential{\mathbf k}%{\bm{\sf k}}{x}$ is not only nonnegative but is known to have a combinatorial interpretation as a generating function. \begin{example} Consider the Hopf monoid $\bm\Pi$ of set partitions. It contains $\mathbf E}%{\bm{\sf l}$ as a Hopf submonoid via the map that sends $\ast_I$ to the partition of $I$ into singletons. We have \[ \exponential{\bm\Pi}{x}/\exponential{\mathbf E}%{\bm{\sf l}}{x} = \exp\bigl(\exp(x)-x-1\bigr), \] which is the exponential generating function for the number of set partitions into blocks of size strictly bigger than $1$. This fact may also be understood with the aid of Theorem~\ref{t:main}, as follows. Given $I=S\sqcup T$, the product of a partition $\pi\in\bm\Pi[S]$ and a partition $\rho\in\bm\Pi[T]$ is the partition $\pi\cdot\rho\in\bm\Pi[I]$ each of whose blocks is either a block of $\pi$ or a block of $\rho$. (In the notation of \cite[Section~12.6]{AguMah:2010}, we are employing the $h$-basis of $\bm\Pi$.) Now, the $I$-component of the right ideal $\mathbf E}%{\bm{\sf l}_+\bm\Pi$ is linearly spanned by elements of the form $\ast_S\cdot\pi$ where $I=S\sqcup T$ and $\pi$ is a partition of $T$. Then, since $\ast_S=\ast_{\{i\}}\cdot\ast_{S\setminus\{i\}}$ (for any $i\in S$), we have that $\mathbf E}%{\bm{\sf l}_+\bm\Pi[I]$ is linearly spanned by elements of the form $\ast_{\{i\}}\cdot\pi$ where $i\in I$ and $\pi$ is a partition of $I\setminus\{i\}$. But the above description of the product shows that these are precisely the partitions with at least one singleton block. \end{example} \begin{example}\label{eg:derangement} Consider the Hopf monoids $\mathbf L}%{\bm{\sf l}$ and $\mathbf E}%{\bm{\sf l}$ and the surjective morphism $\pi:\mathbf L}%{\bm{\sf l}\onto\mathbf E}%{\bm{\sf l}$ (as in the proof of Corollary~\ref{c:ord-exp}). We have \[ \exponential{\mathbf L}%{\bm{\sf l}}{x}=\frac{1}{1-x} \quad\text{and}\quad \exponential{\mathbf E}%{\bm{\sf l}}{x}=\exp(x). \] It is well-known~\cite[Example~1.3.9]{BerLabLer:1998} that \[ \frac{\exp(-x)}{1-x} = \sum_{n\geq 0} \frac{d_n}{n!}\,x^n \] where $d_n$ is the number of \demph{derangements} of $[n]$. Together with Theorem~\ref{t:maindual}, this suggests the existence of a basis for the Hopf kernel of $\pi$ indexed by derangements. We construct such a basis and expand on this discussion in Section~\ref{ss:derangement}. \end{example} \begin{example} Let $\bm{\Sigma}$ be the Hopf monoid of set compositions defined in~\cite[Section~12.4]{AguMah:2010}. It contains $\mathbf L}%{\bm{\sf l}$ as a Hopf submonoid via the map that views a linear order as a composition into singletons. (In the notation of \cite[Section~12.4]{AguMah:2010}, we are employing the $H$-basis of $\bm{\Sigma}$.) This and other morphisms relating $\mathbf E}%{\bm{\sf l}$, $\mathbf L}%{\bm{\sf l}$, $\bm{\Pi}$ and $\bm{\Sigma}$, as well as other Hopf monoids, are discussed in~\cite[Section~12.8]{AguMah:2010}. The sequence $(\dim \bm{\Sigma}[n])_{n\geq 0}$ is A000670 in \cite{Slo:oeis}. We have \[ \exponential{\bm{\Sigma}}{x}=\frac{1}{2-\exp(x)}. \] Moreover, it is known from~\cite[Exercise~5.4.(a)]{Sta:1999} that \[ \frac{1-x}{2-\exp(x)} = \sum_{n\geq 0} \frac{s_n}{n!}\,x^n \] where $s_n$ is the number of \demph{threshold} graphs with vertex set $[n]$ and no isolated vertices. Together with Theorem~\ref{t:main}, this suggests the existence of a basis for $ \bm{\Sigma}/\mathbf L}%{\bm{\sf l}_+ \bm{\Sigma}$ indexed by such graphs. We do not pursue this possibility in this paper. \end{example} \section{The dimension sequence of a connected Hopf monoid}\label{s:dimensions} Let $\mathbf h}%{\bm{\sf h}$ be a connected Hopf monoid and $a_n=\dim\mathbf h}%{\bm{\sf h}[n]$ for $n\in\mathbb{N}$. The results of Sections~\ref{ss:ord-exp} and~\ref{ss:addtests}, derived from Theorem~\ref{t:main}, impose restrictions on the sequence $a_n$ in the form of polynomial inequalities. The results of this section are neither weaker nor stronger than those of Section~\ref{s:applications}, but provide supplementary information on the dimension sequence $a_n$. They do not make use of Theorem~\ref{t:main}. In this section, the base field $\Bbbk$ is of arbitrary characteristic. \begin{proposition}\label{p:aiaj} For any $n,i$ and $j$ such that $n=i+j$, \begin{equation}\label{e:aiaj} a_n\geq a_ia_j. \end{equation} \end{proposition} \begin{proof} Since $\mathbf h}%{\bm{\sf h}$ is connected, the compatibility axiom for Hopf monoids (diagram (8.18) in~\cite[Section~8.3.1]{AguMah:2010}) implies that the composite \[ \mathbf h}%{\bm{\sf h}[S]\otimes\mathbf h}%{\bm{\sf h}[T] \map{\mu_{S,T}} \mathbf h}%{\bm{\sf h}[I] \map{\Delta_{S,T}} \mathbf h}%{\bm{\sf h}[S]\otimes\mathbf h}%{\bm{\sf h}[T] \] is the identity. The result follows by choosing any decomposition $I=S\sqcup T$ with $\abs{I}=n$, $\abs{S}=i$, and $\abs{T}=j$. \end{proof} \begin{remark} The second inequality in~\eqref{e:addcond} may be combined with~\eqref{e:aiaj} to obtain \[ a_3-a_2a_1 \geq 2a_1(a_2-a_1^2)\geq 0. \] Considerations of this type show that neither set of inequalities~\eqref{e:ord-exp},~\eqref{e:addcond} or~\eqref{e:aiaj} follows from the others. \end{remark} As a first consequence of Proposition~\ref{p:aiaj}, we derive a result on the growth of the dimension sequence. \begin{corollary}\label{c:expgrowth} If $a_1\geq 1$, then the sequence $a_n$ is weakly increasing. If moreover there exists $k\geq 1$ such that \[ a_k\geq 2 \quad\text{and}\quad a_i\geq 1\ \forall\, i=0,\ldots,k-1, \] then $a_n=O(2^{n/k})$. \end{corollary} \begin{proof} The first statement follows from $a_n\geq a_1a_{n-1}$. Now fix $k$ as in the second statement. Given $n\geq k$, write $n=qk+r$ with $q\in\mathbb{N}$ and $0\leq r\leq k-1$. From~\eqref{e:aiaj} we obtain \[ a_n\geq a_k^q a_r\geq 2^q = 2^{-r/k}2^{n/k}. \] Thus $a_n=O(2^{n/k})$. \end{proof} Define the \demph{support} of $\mathbf h}%{\bm{\sf h}$ to be the support of its dimension sequence; namely, \[ \supp(\mathbf h}%{\bm{\sf h})=\{n\in\mathbb{N} \mid a_n\neq 0\}. \] We turn to consequences of Proposition~\ref{p:aiaj} on the support. \begin{corollary}\label{c:aiaj} The set $\supp(\mathbf h}%{\bm{\sf h})$ is a submonoid of $(\mathbb{N},+)$. \end{corollary} \begin{proof} By (co)unitality of $\mathbf h}%{\bm{\sf h}$, $0\in \supp(\mathbf h}%{\bm{\sf h})$. (In fact, $a_0=1$ by connectedness.) By Proposition~\ref{p:aiaj}, the set $\supp(\mathbf h}%{\bm{\sf h})$ is closed under addition. \end{proof} We mention that, conversely, given any submonoid $S$ of $(\mathbb{N},+)$, there exists a connected Hopf monoid $\mathbf h}%{\bm{\sf h}$ such that $\supp(\mathbf h}%{\bm{\sf h})=S$. Indeed, let $\bm\Pi_S[I]$ be the space spanned by the set of partitions of $I$ whose block sizes belong to $S\setminus\{0\}$. Then $\bm\Pi_S$ is a quotient Hopf monoid of $\bm\Pi$ and $\supp(\bm\Pi_S)=S$. (The former follows from the formulas in~\cite[Section~12.6.2]{AguMah:2010}; we employ the $h$-basis of $\bm\Pi$.) \begin{example} Consider the special case of the previous paragraph in which $S$ is the submonoid of even numbers. Then $\Pi_S$ is the species of set partitions into blocks of even size. In particular, $a_n=0$ for all odd $n$, so the dimension sequence is neither increasing nor of exponential growth. This example shows that the hypotheses of Corollary~\ref{c:expgrowth} cannot be removed. \end{example} \begin{corollary}\label{c:numerical} The set $\supp(\mathbf h}%{\bm{\sf h})$ is either $\{0\}$ or infinite. The set $\mathbb{N}\setminus\supp(\mathbf h}%{\bm{\sf h})$ is finite if and only if $\gcd\bigl(\supp(\mathbf h}%{\bm{\sf h})\bigr)=1$. \end{corollary} \begin{proof} These statements hold for all submonoids of $\mathbb{N}$, hence for $\supp(\mathbf h}%{\bm{\sf h})$ by Corollary~\ref{c:aiaj}. For the second statement, see~\cite[Lemma~2.1]{RosGar:2009}. \end{proof} \begin{remark} We comment on counterparts for connected graded Hopf algebras of the results of this section. Consider the polynomial Hopf algebra $H=\Bbbk[x_1,\ldots,x_k]$, in which the generators $x_i$ are primitive and of degree $1$. The dimension sequence is $a_n=\binom{n+k-1}{k-1}$. In contrast to Corollary~\ref{c:expgrowth}, this sequence is polynomial even if $a_1>1$. It follows that Proposition~\ref{p:aiaj} has no counterpart for connected graded Hopf algebras $H$, and that the multiplication map $H_i\otimes H_j \to H_{i+j}$ is not injective in general. Corollaries~\ref{c:aiaj} and~\ref{c:numerical} fail for connected Hopf algebras over a field of positive characteristic. In characteristic $p$, a counterexample is provided by $H=\Bbbk[x]/(x^p)$ with $x$ primitive and of degree $1$. On the other hand, if the field characteristic is $0$, then the set $S=\{n\in\mathbb{N} \mid H_n\neq 0\}$ is a submonoid of $(\mathbb{N},+)$. This follows from the fact that in this case any connected Hopf algebra is a domain. We expand on this point in the Appendix. \end{remark} \section{Hopf kernels for cocommutative Hopf monoids}\label{s:kernels} Hopf kernels enter in the decomposition of Theorem~\ref{t:maindual} (and in dual form, in Theorem~\ref{t:main}). For cocommutative Hopf monoids, Hopf kernels and Lie kernels are closely related, as discussed in this section. We provide a simple result that allows us to describe the Hopf kernel in certain situations and we illustrate it with the case of the canonical morphism $\mathbf L}%{\bm{\sf l}\onto\mathbf E}%{\bm{\sf l}$. \subsection{Hopf and Lie kernels} The species $\Pc(\mathbf h}%{\bm{\sf h})$ of \demph{primitive elements} of a connected Hopf monoid $\mathbf h}%{\bm{\sf h}$ is defined by \[ \Pc(\mathbf h}%{\bm{\sf h})[I] = \{ x \in \mathbf h}%{\bm{\sf h}[I] \mid \Delta(x) = 1 \otimes x + x \otimes 1\} \] for each nonempty finite set $I$, and $\Pc(\mathbf h}%{\bm{\sf h})[\emptyset]=0$. Equivalently, \[ \Pc(\mathbf h}%{\bm{\sf h})[I] = \bigcap_{\substack{S\sqcup T=I \\ S,T\neq\emptyset}} \ker\bigl(\Delta_{S,T}:\mathbf h}%{\bm{\sf h}[I]\to \mathbf h}%{\bm{\sf h}[S]\otimes\mathbf h}%{\bm{\sf h}[T]\bigr). \] It is a Lie submonoid of $\mathbf h}%{\bm{\sf h}$ under the commutator bracket. See~\cite[Sections~8.10 and~11.9]{AguMah:2010} for more information on primitive elements. Let $\pi:\mathbf h}%{\bm{\sf h}\to\mathbf k}%{\bm{\sf k}$ be a morphism of connected Hopf monoids. It restricts to a morphism of Lie monoids $\Pc(\mathbf h}%{\bm{\sf h})\to\Pc(\mathbf k}%{\bm{\sf k})$, which we still denote by $\pi$. We define the \demph{Lie kernel} of $\pi$ as the species \[ \mathrm{Lker}(\pi)=\ker\bigl(\pi:\Pc(\mathbf h}%{\bm{\sf h})\to\Pc(\mathbf k}%{\bm{\sf k})\bigr). \] It is a Lie ideal of $\Pc(\mathbf h}%{\bm{\sf h})$. The Hopf kernel $\mathrm{Hker}(\pi)$ is defined in~\eqref{e:hker}. \begin{lemma}\label{l:lker} Let $\pi:\mathbf h}%{\bm{\sf h}\to\mathbf k}%{\bm{\sf k}$ be a morphism of connected Hopf monoids. Then \[ \mathrm{Lker}(\pi)\subseteq\mathrm{Hker}(\pi). \] \end{lemma} \begin{proof} Let $x\in\mathrm{Lker}(\pi)$. Then \[ (\pi_{+\!}\bm\cdot\mathrm{id})\Delta(x)=(\pi_{+\!}\bm\cdot\mathrm{id})(1\otimes x + x\otimes 1)=0, \] since $\pi_+(1)=0$ and $\pi(x)=0$. Thus $x\in\mathrm{Hker}(\pi)$. \end{proof} \begin{lemma}\label{l:hker} Let $\pi:\mathbf h}%{\bm{\sf h}\to\mathbf k}%{\bm{\sf k}$ be a morphism of connected Hopf monoids. Then $\mathrm{Hker}(\pi)$ is a submonoid of $\mathbf h}%{\bm{\sf h}$. \end{lemma} \begin{proof} By definition, \[ \mathrm{Hker}(\pi)=\Delta^{-1}\bigl(\mathrm{Eq}(\pi\bm\cdot\mathrm{id},\, \iota\epsilon\bm\cdot\mathrm{id})\bigr), \] where $\iota:\mathbf{1}\to\mathbf k}%{\bm{\sf k}$ is the unit of $\mathbf k}%{\bm{\sf k}$, $\epsilon:\mathbf h}%{\bm{\sf h}\to\mathbf{1}$ is the counit of $\mathbf h}%{\bm{\sf h}$, and $\mathrm{Eq}$ denotes the equalizer of two maps. Since $\pi$ and $\iota\epsilon$ are morphisms of monoids $\mathbf h}%{\bm{\sf h}\to\mathbf k}%{\bm{\sf k}$, the above equalizer is a submonoid of $\mathbf h}%{\bm{\sf h}\bm\cdot\mathbf h}%{\bm{\sf h}$. Since $\Delta$ is a morphism of monoids, $\mathrm{Hker}(\pi)$ is a submonoid of $\mathbf h}%{\bm{\sf h}$. \end{proof} The following result provides the announced connection between Lie and Hopf kernels for cocommutative Hopf monoids. It makes use of the Poincar\'e-Birkhoff-Witt and Cartier-Milnor-Moore theorems for species, which are discussed in~\cite[Section~11.9.3]{AguMah:2010}. \begin{proposition}\label{p:kernels} Let $\pi:\mathbf h}%{\bm{\sf h}\to\mathbf k}%{\bm{\sf k}$ be a surjective morphism of connected cocommutative Hopf monoids. Then $\mathrm{Hker}(\pi)$ is the submonoid of $\mathbf h}%{\bm{\sf h}$ generated by $\mathrm{Lker}(\pi)$. \end{proposition} \begin{proof} Lemmas~\ref{l:lker} and~\ref{l:hker} imply one inclusion. To conclude the equality, it suffices to check that the dimensions agree, or equivalently, that the exponential generating series are the same. (We are assuming finite dimensionality throughout.) First of all, from Theorem~\ref{t:maindual}, we have \[ \exponential{\mathrm{Hker}(\pi)}{x}=\exponential{\mathbf h}%{\bm{\sf h}}{x}/\exponential{\mathbf k}%{\bm{\sf k}}{x}. \] Now, since $\mathbf h}%{\bm{\sf h}$ is cocommutative, we have \[ \mathbf h}%{\bm{\sf h}\cong \Uc\bigl(\Pc(\mathbf h}%{\bm{\sf h})\bigr) \cong \Sc\bigl(\Pc(\mathbf h}%{\bm{\sf h})\bigr) = \mathbf E}%{\bm{\sf l}\circ \Pc(\mathbf h}%{\bm{\sf h}). \] The first is an isomorphism of Hopf monoids (the Cartier-Milnor-Moore theorem), the second is an isomorphism of comonoids (the Poincar\'e-Birkhoff-Witt theorem), and the third is the definition of the species underlying $\Sc\bigl(\Pc(\mathbf h}%{\bm{\sf h})\bigr)$~\cite[Section~11.3]{AguMah:2010}. It follows that \[ \exponential{\mathbf h}%{\bm{\sf h}}{x} = \exp\bigl(\exponential{\Pc(\mathbf h}%{\bm{\sf h})}{x}\bigr). \] For the same reason, \[ \exponential{\mathbf k}%{\bm{\sf k}}{x} = \exp\bigl(\exponential{\Pc(\mathbf k}%{\bm{\sf k})}{x}\bigr), \] and therefore \[ \exponential{\mathrm{Hker}(\pi)}{x}=\exp\bigl(\exponential{\Pc(\mathbf h}%{\bm{\sf h})}{x} - \exponential{\Pc(\mathbf k}%{\bm{\sf k})}{x}\bigr). \] On the other hand, since the functors $\Uc$ and $\Pc$ define an adjoint equivalence, they preserve surjectivity of maps. Thus, the induced map $\pi:\Pc(\mathbf h}%{\bm{\sf h})\to\Pc(\mathbf k}%{\bm{\sf k})$ is surjective, and we have an exact sequence \[ 0\to \mathrm{Lker}(\pi) \to \Pc(\mathbf h}%{\bm{\sf h}) \to \Pc(\mathbf k}%{\bm{\sf k}) \to 0. \] Hence, \[ \exponential{\mathrm{Lker}(\pi)}{x}=\exponential{\Pc(\mathbf h}%{\bm{\sf h})}{x} - \exponential{\Pc(\mathbf k}%{\bm{\sf k})}{x}. \] Since $\mathrm{Lker}(\pi)$ is a Lie submonoid of $\Pc(\mathbf h}%{\bm{\sf h})$, the submonoid of $\mathbf h}%{\bm{\sf h}$ generated by $\mathrm{Lker}(\pi)$ identifies with $\Uc\bigl(\mathrm{Lker}(\pi)\bigr)$. Therefore, as above, the generating series for the latter submonoid is \[ \exp\bigl(\exponential{\mathrm{Lker}(\pi)}{x}\bigr)= \exp\bigl(\exponential{\Pc(\mathbf h}%{\bm{\sf h})}{x} - \exponential{\Pc(\mathbf k}%{\bm{\sf k})}{x}\bigr)=\exponential{\mathrm{Hker}(\pi)}{x}, \] which is the desired equality. \end{proof} \begin{remark} The results of this section hold also for connected (not necessarily graded or finite dimensional) Hopf algebras. See~\cite[Example 4.20]{BCM:1986} for a proof of Proposition~\ref{p:kernels} in this setting. The proof above used finite dimensionality of the (components of the) species, but this hypothesis is not necessary. The proof in~\cite{BCM:1986} may be adapted to yield the result for arbitrary species. \end{remark} \subsection{The Lie kernel of $\pi:\mathbf L}%{\bm{\sf l}\onto\mathbf E}%{\bm{\sf l}$}\label{ss:cycle} We return to the discussion in Example~\ref{eg:derangement}. The primitive elements of the Hopf monoids $\mathbf E}%{\bm{\sf l}$ and $\mathbf L}%{\bm{\sf l}$ are described in~\cite[Example~11.44]{AguMah:2010}. We have that \[ \Pc(\mathbf E}%{\bm{\sf l})=\mathbf{X} \quad\text{and}\quad \Pc(\mathbf L}%{\bm{\sf l})=\mathbf{Lie} \] where $\mathbf{X}$ is the species of \demph{singletons}, \[ \mathbf{X}[I] = \begin{cases} \Bbbk & \text{ if $\abs{I}=1$, } \\ 0 & \text{ otherwise,} \end{cases} \] and $\mathbf{Lie}$ is the species underlying the \demph{Lie operad}. It follows that the Lie kernel of the canonical morphism $\pi:\mathbf L}%{\bm{\sf l}\onto\mathbf E}%{\bm{\sf l}$ is given by \begin{equation}\label{e:lker} \mathrm{Lker}(\pi)[I]= \begin{cases} \mathbf{Lie}[I] & \text{ if $\abs{I}\geq 2$, } \\ 0 & \text{ otherwise.} \end{cases} \end{equation} Before moving on to the Hopf kernel of $\pi$, we provide some more information on the species $\mathbf{Lie}$. Let $I$ be a finite nonempty set and $n=\abs{I}$. It is known that the space $\mathbf{Lie}[I]$ is of dimension $(n-1)!$. We proceed to describe a linear basis indexed by \demph{cyclic orders} on $I$. A cyclic order on $I$ is an equivalence class of linear orders on $I$ modulo the action \[ i_1{\textcolor{mymagenta}{|}} \cdots {\textcolor{mymagenta}{|}} i_{n-1} {\textcolor{mymagenta}{|}} i_n \ \mapsto \ i_n{\textcolor{mymagenta}{|}} i_1{\textcolor{mymagenta}{|}} \cdots {\textcolor{mymagenta}{|}} i_{n-1} \] of the cyclic group of order $n$. Each class has $n$ elements so there are $(n-1)!$ cyclic orders on $I$. We use $(b,a,c)$ to denote the equivalence class of the linear order $b{\textcolor{mymagenta}{|}} a{\textcolor{mymagenta}{|}} c$. We fix a finite nonempty set $I$ and choose a linear order $\ell_0$ on $I$, say \[ \ell_0 = i_1{\textcolor{mymagenta}{|}} i_2 {\textcolor{mymagenta}{|}} \cdots {\textcolor{mymagenta}{|}} i_n. \] The basis of $\mathbf{Lie}[I]$ will depend on this choice. Given a cyclic order $\gamma$ on $I$, let $S$ be the subset of $I$ consisting of the elements encountered when traversing the cycle from $i_1$ to $i_2$ clockwise, including $i_1$ but excluding $i_2$ (these are the first and second elements in $\ell_0$, respectively). Let $T$ consist of the remaining elements (from $i_2$ to $i_1$). Note that $i_1\in S$ and $i_2\in T$, so both $S$ and $T$ are nonempty. The cyclic order $\gamma$ on $I$ induces cyclic orders on $S$ and $T$. We denote them by $\gamma|_S$ and $\gamma|_T$. An element $p_\gamma\in \mathbf L}%{\bm{\sf l}[I]$ is defined recursively by \[ p_\gamma := [p_{\gamma|_S}, p_{\gamma|_T}]= p_{\gamma|_S}\cdot p_{\gamma|_T}- p_{\gamma|_T}\cdot p_{\gamma|_S}. \] The elements $p_{\gamma|_S}\in\mathbf L}%{\bm{\sf l}[S]$ and $p_{\gamma|_T}\in\mathbf L}%{\bm{\sf l}[T]$ are themselves defined with respect to the induced linear orders $(\ell_0)|_S$ and $(\ell_0)|_T$. The recursion starts with the case when $I$ is a singleton $\{a\}$. In this case, we set \[ p_{(a)}:=a\in \mathbf L}%{\bm{\sf l}[a] \] (the unique linear order). Clearly $a\in\mathbf L}%{\bm{\sf l}[a]$ is a primitive element. Since the primitive elements are closed under commutators, we have $p_\gamma\in \mathbf{Lie}[I]$. Moreover, we have the following. \begin{proposition}\label{p:liebase} For fixed $I$ and $\ell_0$ as above, the set \[ \bigl\{p_\gamma \mid \text{$\gamma$ is a cyclic order on $I$}\bigr\} \] is a linear basis of $\mathbf{Lie}[I]$. \end{proposition} \begin{proof} The construction of the elements $p_\gamma$ is a reformulation of the familiar construction of the \emph{Lyndon} basis of a free Lie algebra~\cite{Lot:1997,Reu:1993,Reu:2003}. Reading the elements of the cyclic order $\gamma$ clockwise starting at the minimum of $\ell_0$ gives rise to a Lyndon word on $I$ (without repeated letters). The cyclic orders $\gamma|_S$ and $\gamma|_T$ give rise to the Lyndon words in the canonical factorization of this Lyndon word. \end{proof} For example, suppose that $I=\{a,b,c,d\}$, $\ell_0=a{\textcolor{mymagenta}{|}} b{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} d$ and $\gamma=(b,a,c,d)$. Then \begin{align*} p_{(b,a,c,d)} &= [p_{(a,c,d)}, p_{(b)}]=\bigl[ [p_{(a)}, p_{(c,d)}], p_{(b)}\bigr] =\Bigl[ \bigl[ p_{(a)}, [p_{(c)}, p_{(d)}] \bigr], p_{(b)} \Bigr] =\Bigl[ \bigl[ a, [c, d] \bigr], b \Bigr] \\ &= a{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} b - a{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} b - c{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} a{\textcolor{mymagenta}{|}} b +d{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} a{\textcolor{mymagenta}{|}} b -b{\textcolor{mymagenta}{|}} a{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} d +b{\textcolor{mymagenta}{|}} a{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} c +b{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} a -b{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} a. \end{align*} \begin{remark} The vector species $\mathbf{Lie}$ is \emph{not} the linearization of the set species of cycles. Note also that, for a general bijection $\sigma:I\to J$, the $p$-basis of $\mathbf{Lie}[I]$ will not map to the $p$-basis of $\mathbf{Lie}[J]$ under $\mathbf L}%{\bm{\sf l}[\sigma]$. \end{remark} \subsection{The Hopf kernel of $\pi:\mathbf L}%{\bm{\sf l}\onto\mathbf E}%{\bm{\sf l}$}\label{ss:derangement} The above description~\eqref{e:lker} of the Lie kernel of $\pi:\mathbf L}%{\bm{\sf l}\onto\mathbf E}%{\bm{\sf l}$ together with Proposition~\ref{p:kernels} imply that the Hopf kernel of $\pi$ is given by \[ \mathrm{Hker}(\pi)[I]=\sum_{k\geq 1}\sum_{\substack{S_1\sqcup\cdots\sqcup S_k=I \\ \abs{S_r}\geq 2\,\,\forall r}} \mathbf{Lie}[S_1]\cdots\mathbf{Lie}[S_k]. \] An element in $\mathbf{Lie}[S_1]\cdots\mathbf{Lie}[S_k]$ is a $k$-fold product of primitive elements $x_r\in\mathbf{Lie}[S_r]$; each $S_r$ must have at least $2$ elements. We proceed to describe a linear basis for $\mathrm{Hker}(\pi)[I]$. As in Section~\ref{ss:cycle}, we fix a linear order $\ell_0=i_1{\textcolor{mymagenta}{|}} i_2 {\textcolor{mymagenta}{|}} \cdots {\textcolor{mymagenta}{|}} i_n$ on $I$. The basis will be indexed by \demph{derangements} of $\ell_0$. A derangement of $\ell_0$ is a linear order $\ell=j_1{\textcolor{mymagenta}{|}} j_2 {\textcolor{mymagenta}{|}} \cdots {\textcolor{mymagenta}{|}} j_n$ on $I$ such that $i_r\neq j_r$ for all $r=1,\ldots,n$. View linear orders as bijections $[n]\to I$ and define $\sigma:=\ell\circ\ell_0^{-1}$. Then $\sigma$ is a permutation of $I$ and $\ell$ is a derangement of $\ell_0$ precisely when $\sigma$ has no fixed points. Let $\ell$ be a derangement of $\ell_0$ and $\sigma$ the associated permutation. Let $S_1,\ldots,S_k$ be the orbits of $\sigma$ on $I$ labeled so that \[ \min S_1<\cdots<\min S_k \text{ \ according to $\ell_0$,} \] and let $\gamma_r$ be the cyclic order on $S_r$ induced by $\sigma$. In other words, $\sigma=\gamma_1\cdots\gamma_k$ is the factorization of $\sigma$ into cycles, ordered in this specific manner. Employing the $p$-basis of $\mathbf{Lie}$ from Section~\ref{ss:cycle} (defined with respect to $\ell_0$ and the orders induced by $\ell_0$ on subsets of $I$), we define an element $p_{\ell}\in\mathbf L}%{\bm{\sf l}[I]$ by \[ p_{\ell} :=p_{\gamma_1}\cdots p_{\gamma_k}. \] By assumption, $\abs{S_r}\geq 2$ for all $r$. Hence $p_{\gamma_r}\in\mathrm{Lker}(\pi)[S_r]$ and $p_{\ell}\in\mathrm{Hker}(\pi)[I]$. For example, let $I=\{e,i,m,s,t\}$, $\ell_0 =s{\textcolor{mymagenta}{|}} m {\textcolor{mymagenta}{|}} i {\textcolor{mymagenta}{|}} t {\textcolor{mymagenta}{|}} e$ and $\ell = i {\textcolor{mymagenta}{|}} t {\textcolor{mymagenta}{|}} e {\textcolor{mymagenta}{|}} m {\textcolor{mymagenta}{|}} s$. Then \[ \sigma = (s,i,e)(m,t), \quad S_1 = \{i,e,s\},\quad S_2 = \{m,t\}, \] and \[ p_\ell = p_{(s,i,e)}p_{(m,t)} = \left[p_{(s)},p_{(i,e)}\right]p_{(m,t)} = \bigl[ s,[i,e]\bigr][m,t]. \] \begin{proposition}\label{p:hopfbase} For fixed $I$ and $\ell_0$ as above, the set \[ \bigl\{p_\ell \mid \text{$\ell$ is a derangement of $\ell_0$}\bigr\} \] is a linear basis of $\mathrm{Hker}(\pi)[I]$. \end{proposition} \begin{proof} This follows from Proposition~\ref{p:liebase} and the Poincar\'e-Birkhoff-Witt theorem. \end{proof} \begin{example} We describe the $p$-basis of $\mathrm{Hker}(\pi)[I]$ in low cardinalities. Throughout, we choose \[ \ell_0=a{\textcolor{mymagenta}{|}} b {\textcolor{mymagenta}{|}} c {\textcolor{mymagenta}{|}}\cdots\,. \] The space $\mathrm{Hker}(\pi)[a,b]$ is $1$-dimensional, linearly spanned by \[ p_{b{\textcolor{mymagenta}{|}} a} = p_{(a,b)} =[a,b]. \] The space $\mathrm{Hker}(\pi)[a,b,c]$ is $2$-dimensional, linearly spanned by \begin{align*} p_{b{\textcolor{mymagenta}{|}} c{\textcolor{mymagenta}{|}} a} = p_{(a,b,c)} = [ p_{(a)},p_{(b,c)}] = \bigl[ a,[b,c]\bigr],\\ p_{c{\textcolor{mymagenta}{|}} a{\textcolor{mymagenta}{|}} b} = p_{(a,c,b)} = [p_{(a,c)},p_{(b)}] = \bigl[ [a,c], b\bigr]. \end{align*} The space $\mathrm{Hker}(\pi)[a,b,c,d]$ is $9$-dimensional. There are $6$ basis elements corresponding to $4$-cycles, such as \[ p_{c{\textcolor{mymagenta}{|}} a{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} b} = p_{(a,c,d,b)} = \Bigl[ \bigl[ a, [c, d] \bigr], b \Bigr], \] and $3$ basis elements corresponding to products of two $2$-cycles, such as \[ p_{b{\textcolor{mymagenta}{|}} a{\textcolor{mymagenta}{|}} d{\textcolor{mymagenta}{|}} c} = p_{(a,b)}p_{(c,d)} = [a,b]\cdot [c,d]. \] \end{example} \section*{Appendix The following fact was referred to in the last remark in Section~\ref{s:dimensions}. \begin{proposition}\label{p:domain} Let $H$ be a connected (not necessarily graded) Hopf algebra over a field of characteristic $0$. Then $H$ is a domain. \end{proposition} This result is proven in~\cite[Lemma~1.8(a)]{WZZ:2011}, where it is attributed to Le Bruyn. We provide a different proof here. \begin{proof} Let $K$ denote the associated graded Hopf algebra with respect to the coradical filtration of $H$. Since $H$ is connected, $K$ is commutative~\cite[Remark~1.7]{AguSot:2005}. Now by~\cite[Proposition~1.2.3]{Rad:1979}, $K$ embeds in a shuffle Hopf algebra. The latter is a free commutative algebra~\cite[Corollary~3.1.2]{Rad:1979}, hence a domain. It follows that $K$ and hence also $H$ are domains. \end{proof} Over a field of positive characteristic, the restricted universal enveloping algebra $\mathfrak u(\mathfrak g)$ of a finite dimensional nonzero Lie algebra $\mathfrak g$ is a connected Hopf algebra that is not a domain. Indeed, in this case $u(\mathfrak g)$ is finite dimensional~\cite[Theorem~V.12]{Jac:1962} and so has a nontrivial idempotent, being a Hopf algebra with integrals~\cite[Theorem~2.1.3]{Mon:1993}. \bibliographystyle{plain}

https://arxiv.org/abs/math/0502249

A characterization of spaces with discrete topological fundamental group

We offer a counterexample to a theorem in the literature and then repair the theorem as follows: The fundamental group of a locally path connected metric space inherits the discrete topology in a natural way if and only if the underlying space is semilocally simply connected.

\section{Introduction} In general $\pi _{1}(X,p),$ the based fundamental group of a space $X,$ admits a canonical topology and becomes the \textit{topological fundamental group.} To date significant progress has been made in the investigation of the topological fundamental group of certain non semilocally simply connected spaces such as the Hawaiian earring and the harmonic archipelago. For example, the Hawaiian earring group is not a Baire space \cite{fab3} and also fails to embed topologically in the inverse limit of free groups \cite {fab25} despite injectivity of the canonical homomorphism \cite{mor}. The harmonic archipelago is a compact metric space whose fundamental group is uncountable but has only two open sets \cite{fab2}. In particular neither of these groups has the discrete topology. When does $\pi _{1}(X,p)$ have the discrete topology? The fundamental groups of locally contractible spaces such as $n-$manifolds have the discrete topology. Indeed, a published theorem of another author (Theorem 5.1 \cite {Biss}) indicates that $\pi _{1}(X,p)$ has the discrete topology precisely when $X$ is semilocally simply connected. Unfortunately Theorem 5.1 is false. This paper contains a counterexample, an alternate version of the theorem, and some remarks and examples that lend further perspective on the results in question. \section{Definitions} Suppose $X$ is a metrizable space and $p\in X.$ Let $C_{p}(X)=\{f:[0,1]% \rightarrow X$ such that $f$ is continuous and $f(0)=f(1)=p\}.$ Endow $% C_{p}(X)$ with the topology of uniform convergence. The \textbf{topological fundamental group} $\pi _{1}(X,p)$ is the set of path components of $% C_{p}(X) $ topologized with the quotient topology under the canonical surjection $q:C_{p}(X)\rightarrow \pi _{1}(X,p)$ satisfying $q(f)=q(g)$ if and only of $f $ and $g$ belong to the same path component of $C_{p}(X).$ Thus a set $U\subset \pi _{1}(X,p)$ is open in $\pi _{1}(X,p)$ if and only if $q^{-1}(U)$ is open in $\pi _{1}(X,p).$ A space $X$ is \textbf{semilocally simply connected} at $p$ if $p\in X$ and there exists an open set $U$ such that $p\in U$ and $j_{U}:U\rightarrow X$ induces the trivial homomorphism $j_{U}^{\ast }:\pi _{1}(U,p)\rightarrow \pi _{1}(X,p).$ Roughly speaking, this means small loops in $X$ are inessential, but might bound large disks. A topological space $X$ is \textbf{discrete} if every subset is both open and closed. The above definitions are consistent with those found in Munkres \cite{Munkres}. \section{A counterexample} Looking beyond the notational ambiguity in case $X$ is not path connected, it is claimed falsely in Theorem 5.1 \cite{Biss} that $\pi _{1}(X)$ is discrete if and only if the topological space $X$ is semilocally simply connected. Here is a counterexample to Theorem 5.1. \begin{example} \label{e1}Let $X$ denote the following subset of the plane: \[ X=(\{0,1,\frac{1}{2},\frac{1}{3},...\}\times \lbrack 0,1])\cup ([0,1]\times \{0\})\cup ([0,1]\times \{1\}). \] Let $p=(0,0).$ Consider the inessential map $f$ which starts at $(0,0),$ goes straight up to $(0,1)$ and then comes back down again. Notice we may construct maps $f_{n}\in C_{p}(X)$ such that $f_{n}$ is inessential but $% f_{n}\rightarrow f$ uniformly. (Let $f_{n}$ trace a rectangle with sides $% \{0\}\times \lbrack 0,1]$ and $\{\frac{1}{n}\}\times \lbrack 0,1]$). This shows that the path component of the constant map is not open in $C_{p}(X).$ Consequently $\pi _{1}(X,p)$ does not have the discrete topology. On the other hand $X$ is semilocally simply connected since small loops in $X$ are null homotopic in $X.$ \end{example} \section{A characterization of discrete topological fundamental groups} Theorem 5.1 \cite{Biss} can be repaired by making the added assumption that $% X$ is locally path connected. \begin{theorem} \label{main}Suppose $X$ is a locally path connected metrizable space and $% p\in X.$ Then $\pi _{1}(X,p)$ is discrete if and only if $X$ is semilocally simply connected at $p.$ \end{theorem} \begin{proof} Let $F:C_{p}(X)\rightarrow \pi _{1}(X,p)$ denote the canonical quotient map. Let $P\in C_{p}(X)$ denote the constant map. Let $[P]$ denote the path component of $P$ in $C_{p}(X).$ Let $d$ be any metric on $X.$ Suppose $\pi _{1}(X,p)$ is discrete. Then the trivial element of $\pi _{1}(X,p)$ is an open subset. Hence, since $F$ is continuous, $[P]$ is open in $C_{p}(X).$ Suppose, in order to obtain a contradiction, that $\pi _{1}(X,p)$ is not semilocally simply connected at $p.$ Then there exists a sequence of maps $\alpha _{n}\rightarrow P$ such that $\alpha _{n}\in C_{p}(X)$ and $\alpha _{n}$ is essential in $X$ for each $n.$ On the other hand, since $[P]$ is open in $C_{p}(X)$ we can be sure that eventually $% \alpha _{n}$ is inessential in $X.$ Conversely, suppose $X$ is semilocally simply connected at $p.$ To prove $% \pi _{1}(X,p)$ is discrete we must show that each one point subset is open. Since $\pi _{1}(X,p)$ is a topological group (Proposition 3.1 \cite{Biss}) $% \pi _{1}(X,p)$ is homogeneous. Thus it suffices to prove that the trivial element of $\pi _{1}(X,p)$ is an open subset. Since $F$ is a quotient map it suffices to prove $[P]$ is open in $C_{p}(X).$ Suppose $f\in \lbrack p]$ and suppose $f_{n}\in C_{p}(X)$ and suppose $f_{n}\rightarrow f$ uniformly. We must prove that for large $n$ $f$ and $f_{n}$ are path homotopic. Observation1. Since $X$ is locally path connected and since $im(f)$ is compact, for each $\varepsilon >0$ there exists $\delta >0$ such that if $% x\in im(f)$ and $d(x,y)<\delta $ then there exists a path $\alpha _{xy}$ from $x$ to $y$ such that diam$(\alpha _{xy})<\varepsilon .$ Observation 2. Since $X$ is semilocally simply connected and since $im(f)$ is compact, there exists $\varepsilon >0$ such that if $x\in im(f)$ and if $% \alpha _{xx}$ is any path from $x$ to $x$ such that diam$(\alpha _{xx})<\varepsilon $ and then $\alpha _{xx}$ is null homotopic. Observation 3. The set $f\cup \{f_{1},f_{2},..\}$ is an equicontinuous collection of maps. Combining observations 1,2, and 3, for sufficiently large $n$ we can manufacture a homotopy from $f_{n}$ to $f$ described roughly as follows. Observation 1 says we can connect $f(\frac{i}{2^{n}})$ and $f_{n}(\frac{i}{% 2^{n}})$ by a small path and we can connect $f(\frac{i+1}{2^{n}})$ and $% f_{n}(\frac{i+1}{2^{n}})$ by a small path. Observation 3 says the restriction of $f$ and $f_{n}$ to $[\frac{i}{2^{n}},\frac{i+1}{2^{n}}]$ is small. Observation 2 says the boundary of the rectangle with corners $f(% \frac{i}{2^{n}})$, $f_{n}(\frac{i}{2^{n}}),f(\frac{i+1}{2^{n}})$ and $f_{n}(% \frac{i+1}{2^{n}})$ determines an inessential loop. Thus we can fill in the homotopy across $[\frac{i}{2^{n}},\frac{i+1}{2^{n}}]\times \lbrack 0,1].$ \end{proof} \begin{remark} Local path connectivity is only used in the 2nd part of the proof of Theorem \ref{main}. Thus all path connected spaces with discrete topological fundamental group must be semilocally simply connected. \end{remark} \begin{remark} There exist spaces which are not locally path connected but which have discrete fundamental group. For example the cone over $X$ in example \ref{e1}% . (i.e. let $Y$ be quotient of the space $X\times \lbrack 0,1]$ with $% X\times \{1\}$ considered as a point. Note $Y$ is simply connected.) \end{remark} \begin{remark} Theorem 5.1 of \cite{Biss} cannot be repaired by replacing the notion of `discrete' with `totally disconnected'. The space $X=S^{1}\times S^{1}..$ has totally disconnected fundamental group but $X$ is not semilocally simply connected. \end{remark}

https://arxiv.org/abs/1411.0867

On the equality of Hausdorff measure and Hausdorff content

We are interested in situations where the Hausdorff measure and Hausdorff content of a set are equal in the critical dimension. Our main result shows that this equality holds for any subset of a self-similar set corresponding to a nontrivial cylinder of an irreducible subshift of finite type, and thus also for any self-similar or graph-directed self-similar set, regardless of separation conditions. The main tool in the proof is an exhaustion lemma for Hausdorff measure based on the Vitali Covering Theorem.We also give several examples showing that one cannot hope for the equality to hold in general if one moves in a number of the natural directions away from `self-similar'. For example, it fails in general for self-conformal sets, self-affine sets and Julia sets. We also give applications of our results concerning Ahlfors regularity. Finally we consider an analogous version of the problem for packing measure. In this case we need the strong separation condition and can only prove that the packing measure and $\delta$-approximate packing pre-measure coincide for sufficiently small $\delta>0$.

\section{Introduction} Hausdorff measure and dimension are among the most important notions in fractal geometry and geometric measure theory used to quantify the size of a set. The Hausdorff content is a concept closely related to the Hausdorff measure, but perhaps less popular in the context of classical measure theory. That being said the Hausdorff content enjoys greater regularity than the Hausdorff measure and still gives the Hausdorff dimension as the critical exponent. The goal of this article is to understand further the relationship between Hausdorff measure and Hausdorff content in the context of some well-known and popular classes of fractals sets. In particular we are interested in when the Hausdorff measure and Hausdorff content of a set are equal in the Hausdorff dimension. This study was motivated by a question of Michael Barnsley posed to one of the authors. \subsection{Hausdorff measure and Hausdorff content} Let $F \subseteq \mathbb{R}^n$. For $s \geq 0$ and $\delta>0$ the \emph{$\delta$-approximate $s$-dimensional Hausdorff measure} of $F$ is defined by \begin{eqnarray*} \mathcal{H}_\delta^s (F) = \inf \bigg\{ \sum_{k=1}^{\infty} \text{diam}( U_k)^s : \{ U_k\}_{k=1}^{\infty} \text{ is a countable cover of $F$ }\\ \text{ by sets with $\text{diam}(U_k) \leq \delta$ for all $k$} \bigg\} \end{eqnarray*} and the $s$-dimensional Hausdorff (outer) measure of $F$ by $\mathcal{H}^s (F) = \lim_{\delta \to 0} \mathcal{H}_\delta^s (F)$. If one does not put any restriction on the diameters of the covering sets, then one obtains the Hausdorff content of $F$, namely, \[ \mathcal{H}_\infty^s (F) = \inf \bigg\{ \sum_{k=1}^{\infty} \text{diam}( U_k)^s : \{ U_k\}_{k \in \mathcal{K}} \text{ is a countable cover of $F$ by arbitrary sets} \bigg\}. \] The following chain of inequalities is evident \[ \mathcal{H}_\infty^s (F) \leq \mathcal{H}_\delta^s (F) \leq \mathcal{H}^s (F) \] (for every $\delta>0$) and, moreover, the Hausdorff dimension of $F$ is equal to \begin{eqnarray*} \dim_\text{H} F \ = \ \inf \Big\{ s \geq 0: \mathcal{H}^s (F) =0 \Big\} \ = \ \inf \Big\{ s \geq 0: \mathcal{H}^s_\infty (F) =0 \Big\}. \end{eqnarray*} Thus, for every $s> \dim_\text{H} F$, we have $ \mathcal{H}^s_\infty (F) = \mathcal{H}^s_\delta (F) = \mathcal{H}^s (F) =0$ and for every $s< \dim_\text{H} F$, we have $ \mathcal{H}_\infty^s (F) \leq \mathcal{H}^s_\delta (F) \leq \mathcal{H}^s (F) = \infty$, again for every $\delta>0$, with the final inequality strict if $F$ is bounded ($\mathcal{H}^s_\delta(F)$ is finite for every $\delta$ for $F$ bounded). The case when $s =\dim_\text{H} F$ is more subtle, and the case of interest. Then $\mathcal{H}^s (F)$ may be zero, positive and finite, or infinite, but $\mathcal{H}^s_\infty (F)$ must be finite if $F$ is bounded. Moreover, if $\mathcal{H}^s_\infty (F) = 0$, then $\mathcal{H}^s (F) = 0$ also. \\ \\ The goal of this article is to study situations where $\mathcal{H}^s_\infty(F) = \mathcal{H}^s (F) $ with $s=\dim_\text{H} F$. Sets with this property were studied by Foran \cite{foran}, where they were called $s$-straight sets, and later studied by Delaware \cite{delaware0, delaware}. There are many advantages to having this equality as Hausdorff content is more easily analysed. For example, the expression $\sum_{k=1}^{\infty} \text{diam}( U_k)^s$ gives a genuine upper bound for $\mathcal{H}^s_\infty(F)$ for every cover $\{U_k\}_{k=1}^{\infty}$, and for every $s \geq 0$ the function $\mathcal{H}^s_\infty$ acting on the set of compact subsets of a compact metric space equipped with the Hausdorff metric is an upper semicontinuous function, and thus Baire 1, whereas $\mathcal{H}^s$ is only Baire 2, see \cite{mattilamauldin}. Another consequence is that $ \mathcal{H}_\delta^s (F) = \mathcal{H}^s (F)$ for all $\delta>0$. For more details on Hausdorff measure and dimension, see \cite[Chapter 3]{falconer} and \cite{rogers}. \\ \\ We conclude this section with a well-known observation and include the proof for completeness. \begin{lma} \label{haus11} Let $F \subseteq \mathbb{R}^n$ be such that $\mathcal{H}^s_\infty(F) = \mathcal{H}^s (F)<\infty $ where $s=\dim_\text{\emph{H}} F$. Then for every $\mathcal{H}^s$-measurable subset $E \subseteq F$ we also have $\mathcal{H}^s_\infty(E) = \mathcal{H}^s (E)$. \end{lma} \begin{proof} A routine calculation using $\mathcal{H}^s$-measurable hulls verifies that $\mathcal{H}^s(E) \ = \ \mathcal{H}^s(F) \, - \, \mathcal{H}^s(F \setminus E)$ even if $F$ is not $\mathcal{H}^s$-measurable. Therefore \[ \mathcal{H}^s_\infty(E) \ \leq \ \mathcal{H}^s(E) \ = \ \mathcal{H}^s(F) \, - \, \mathcal{H}^s(F \setminus E) \ \leq \ \mathcal{H}^s_\infty(F) \, - \, \mathcal{H}^s_\infty(F \setminus E) \ \leq \ \mathcal{H}^s_\infty(E) \] which completes the proof. \end{proof} Of course, this result is not necessarily true if we replace $s$ by $\dim_\H E$. \section{Main results: general situations where $ \mathcal{H}_\infty^s (F) = \mathcal{H}^s (F)$ } Let $\mathcal{I} = \{0, \dots, M-1\}$ be a finite alphabet, let $\Sigma = \mathcal{I}^{\mathbb{N}}$ and $\sigma : \Sigma \to \Sigma$ be the one-sided left shift. We will write $i \in \mathcal{I}$, $\textbf{\emph{i}} = (i_0 , \dots, i_{k-1}) \in \mathcal{I}^k$ and $\alpha = (\alpha_0, \alpha_1, \dots) \in \Sigma$. We will also write $\alpha\vert_k = (\alpha_0, \dots, \alpha_{k-1}) \in \mathcal{I}^k$ for the restriction of $\alpha$ to its first $k$ coordinates. We equip $\Sigma$ with the standard metric defined by \[ d(\alpha,\beta) = 2^{-n(\alpha,\beta)} \] for $\alpha \neq \beta$, where $n(\alpha,\beta) = \max\{ n \in \mathbb{N} : \alpha\vert_n = \beta\vert_n\}$. We write $\mathcal{I}^* = \cup_{k \in \mathbb{N}} \mathcal{I}^k$ for the set of all finite words. For $\textbf{\emph{i}} = (i_0 , \dots, i_{k-1}) \in \mathcal{I}^*$, we write \[ [\textbf{\emph{i}}] = \big\{ \alpha \in \Sigma : \alpha \vert_k = \textbf{\emph{i}} \big\} \] for the \emph{cylinder} corresponding to $\textbf{\emph{i}}$ and we let $|\textbf{\emph{i}}|=k$ be the length of $\textbf{\emph{i}}$. Also, even though the shift is only defined on $\Sigma$, it will be convenient also to define it for $\textbf{\emph{i}} = (i_0 , \dots, i_{k-1}) \in \mathcal{I}^*$ by \[ \sigma( \textbf{\emph{i}} ) = \sigma \big( (i_0 , \dots, i_{k-1}) \big) = (i_1 , \dots, i_{k-1}). \] Any closed $\sigma$-invariant set $\Lambda \subseteq \Sigma$ is called a \emph{subshift}. Among the most important subshifts are \emph{subshifts of finite type} which we define as follows. Let $A$ be an $M \times M$ \emph{transition matrix} indexed by $\mathcal{I} \times \mathcal{I}$ with entries in $\{0,1\}$. We define the subshift of finite type corresponding to $A$ as \[ \Sigma_A \ = \ \Big\{\alpha = (\alpha_0 \alpha_1 \dots ) \in \Sigma : A_{\alpha_i, \alpha_{i+1}} = 1 \text{ for all } i =0, 1, \dots \Big\}. \] If every entry of $A$ is 1 then we call $\Sigma_A = \Sigma$ the \emph{full shift}. We say $\Sigma_A$ is irreducible (or transitive) if the matrix $A$ is irreducible, which means that for all pairs $i,j \in \mathcal{I}$, there exists $n \in \mathbb{N}$ such that $(A^n)_{i,j} >0$. We say $\Sigma_A$ is aperiodic (or mixing) if the matrix $A$ is aperiodic, which means that there exists $n \in \mathbb{N}$ such that $(A^n)_{i,j} >0$ for all pairs $i,j \in \mathcal{I}$ simultaneously. \\ \\ To each $i \in \mathcal{I}$ associate a similarity map $S_i$ on $\mathbb{R}^n$ with contraction ratio $r_i \in (0,1)$ which we assume for convenience maps $[0,1]^n$ into itself. For $\textbf{\emph{i}} = (i_0 , \dots, i_{k-1}) \in \mathcal{I}^*$, write \[ S_{\textbf{\emph{i}}} = S_{i_0} \circ \cdots \circ S_{i_{k-1}} \] and \[ r_{\textbf{\emph{i}}} = r_{i_0} \cdots r_{i_{k-1}}. \] Let $\Pi: \Sigma \to [0,1]^n$ be the natural coding map given by \[ \Pi(\alpha) = \bigcap_{k=1}^\infty S_{\alpha\vert_k} \big([0,1]^n\big). \] For a given subshift of finite type $\Sigma_A$, we are interested in the set $F_A:=\Pi(\Sigma_A)$. The set $F:=\Pi(\Sigma)$ corresponding to the full shift is called a \emph{self-similar set} and is the unique non-empty compact set satisfying \[ F \ = \ \bigcup_{i \in \mathcal{I}} S_i (F). \] The collection of contracting similarities $\{S_i\}_{i \in \mathcal{I}}$ is called an \emph{iterated function system} (IFS), see \cite[Chapter 9]{falconer}. We will also be interested in subsets of $F_A$ corresponding to the cylinders of $\Sigma_A$. In particular, for $\textbf{\emph{i}} \in \mathcal{I}^*$, let \[ F_A^\textbf{\emph{i}} = \Pi(\Sigma_A \cap [\textbf{\emph{i}}]), \] which may be empty. It can be shown via the implicit theorems of Falconer \cite{implicit}, \cite[Section 3.1]{techniques} that if $A$ is irreducible, then $\mathcal{H}^s(F_A) < \infty$ where $s = \dim_\H F_A$. Moreover, if $\mathcal{H}^s(F_A) > 0$, then $\mathcal{H}^s(F_A^\textbf{\emph{i}}) > 0$ for each $\textbf{\emph{i}} \in \mathcal{I}^*$ for which $F_A^\textbf{\emph{i}} \neq \emptyset$. \begin{thm} \label{main} Let $A$ be irreducible and let $s = \dim_\text{\emph{H}} F_A$. For all $\textbf{i} \in \mathcal{I}^*$ we have \[ \mathcal{H}_\infty^s \big(F_A^\textbf{i}\big) = \mathcal{H}^s \big(F_A^\textbf{i}\big). \] Moreover, we can extend this to unions of 1-cylinders in the same `family'. For all $i \in \mathcal{I}$, \[ \mathcal{H}_\infty^s \Bigg( \bigcup_{j \in \mathcal{I} : A_{i,j} = 1} F_A^j \Bigg) \ = \ \mathcal{H}^s \Bigg( \bigcup_{j \in \mathcal{I} : A_{i,j} = 1} F_A^j \Bigg). \] \end{thm} We will prove Theorem \ref{main} in Section \ref{mainproof}. It is natural to wonder if the equality is still satisfied for the full set, and not just cylinders and unions of cylinders in the same family. We give an example in Section \ref{examples} which shows that this is not true. Delaware \cite{delaware} proved that any set with finite $\mathcal{H}^s$ measure is $\sigma s$-straight, in that it can be decomposed as a countable union of $s$-straight sets. This proved a conjecture of Foran \cite{foran}. Theorem \ref{main} can be viewed as a strengthening of this result in the very special case of subshifts of finite type for self-similar sets. In particular, we prove that for irreducible $A$ the set $F_A$ can be decomposed into a \emph{finite} union of $s$-straight sets \[ F_A = \bigcup_{i \in \mathcal{I}} F_A^i. \] In this paper we only consider subshifts of finite type, but the same questions are valid for general subshifts and we therefore ask the following natural question. \begin{ques} Does there exists a system of similarities and a transitive subshift $\Lambda \subseteq \Sigma$, such that \[ \mathcal{H}_\infty^s \big(\Pi(\Lambda) \cap [\textbf{i}]\big) < \mathcal{H}^s \big(\Pi(\Lambda) \cap [\textbf{i}]\big) \] for some $\textbf{i} \in \mathcal{I}^*$? \end{ques} Note that for general subshifts, being \emph{transitive} means that there exists one dense orbit under the left shift. Transitive subshifts of finite type are precisely those with irreducible $A$ and so Theorem \ref{main} answers this question in the negative for subshifts of finite type. \\ \\ Theorem \ref{main} was shown for self-similar sets rather than subshifts of finite type by Bandt and Graf \cite[Proposition 3]{bandt-graf} assuming the \emph{open set condition} is satisfied. See \cite[Section 9.2]{falconer} for the definition and further properties of the open set condition. This result was generalised by Farkas \cite[Proposition 1.11]{farkas} for self-similar sets without assuming any separation condition. We state this result as a corollary of Theorem \ref{main}. \begin{cor} \label{corselfsim} Let $F \subseteq [0,1]^n$ be a self-similar set and let $s = \dim_\text{\emph{H}} F$. Then, regardless of separation conditions, $\mathcal{H}_\infty^s (F) = \mathcal{H}^s (F)$. \end{cor} \begin{proof} Since $F$ is self-similar it is modelled by a full shift and thus for any $i \in \mathcal{I}$ \[ F = \bigcup_{j \in \mathcal{I} : A_{i,j} = 1} F_A^j \] and so the result follows from Theorem \ref{main}. \end{proof} We note here that if the Hausdorff measure of a set is zero in a particular dimension, then the Hausdorff content is also zero in that dimension and so the equality is trivial. One might initially wonder if $ \mathcal{H}^{\dim_\text{H}F} (F)=0$ always holds when $F$ is a self-similar set which cannot be defined via a system which satisfies the open set condition, but this is false, see for example \cite[Example 8.6]{farkas}. Thus this result provides nontrivial information even when the open set condition is not satisfied. Recall that Schief \cite{schief} proved that $ \mathcal{H}^{s} (F)=0$ if $F$ is a self-similar set defined via a system which does not satisfy the open set condition and $s$ is the \emph{similarity dimension} but, as the example of Farkas shows, one can obtain positive Hausdorff measure in the Hausdorff dimension if this is less than the similarity dimension, even if the open set condition cannot be satisfied. \\ \\ A natural and important generalisation of self-similar sets is \emph{graph-directed self-similar sets}, which we now define. Let $\Gamma = G(\mathcal{V}, \mathcal{E})$ be a finite strongly connected directed multigraph with vertices $\mathcal{V} = \{1, \dots, N\}$ and a finite multiset of edges $ \mathcal{E}$. Write $\mathcal{E}_{i,j}$ for the multiset of all edges joining the vertex $i$ to the vertex $j$. For each $e \in \mathcal{E}$ associate a contracting similarity mapping $S_e$ on $\mathbb{R}^n$ with contraction ratio $r_e \in (0,1)$ which we again assume for convenience maps $[0,1]^n$ into itself. It is standard that there exists a unique family of non-empty compact sets $\{F_{i}\}_{i \in \mathcal{V}}$ satisfying \begin{equation} F_i = \bigcup_{j = 1}^{N} \bigcup_{e \in \mathcal{E}_{i,j}} S_e(F_j). \label{ssc-gda} \end{equation} Each set in the family $\{F_{i}\}_{i \in \mathcal{V}}$ is called a graph-directed self-similar set. Even though all self-similar sets are graph-directed self-similar sets, it was proved by Boore and Falconer \cite{boore} that graph-directed self-similar sets are genuinely more general than just self-similar sets. We obtain the following generalisation of Corollary \ref{corselfsim}. \begin{cor} \label{corgdselfsim} Let $F \subseteq [0,1]^n$ be a graph-directed self-similar set and let $s = \dim_\text{\emph{H}} F$. Then, regardless of separation conditions, $\mathcal{H}_\infty^s (F) = \mathcal{H}^s (F)$. \end{cor} Corollary \ref{corgdselfsim} follows from Theorem \ref{main} and the following proposition. \begin{prop} \label{prop_gda2ssoft} Let $\{F_{i}\}_{i \in \mathcal{V}}$ be the solution of a graph-directed self-similar iterated function system with directed graph $\Gamma = G(\mathcal{V}, \mathcal{E})$. Then there exists a subshift of finite type associated to the alphabet $\mathcal{I}=\mathcal{E}$ such that every $F_{i}$ is the union of 1-cylinders in the same family in the sense of Theorem \ref{main}. If $\Gamma$ is strongly connected then the constructed subshift of finite type is irreducible. \end{prop} \begin{proof} Let the alphabet be indexed by the edge set $\mathcal{E}$. Now, for two edges $e,f \in \mathcal{E}$, let $A_{e,f} = 1$ if and only if $f$ begins from the vertex where $e$ ended, i.e., it is possible to walk along $e$ and then along $f$. If $\Gamma$ is strongly connected, the matrix $A$ is irreducible. It is now straightforward to see that for all $e \in \mathcal{E}_{i,j}$ \[ F_A^e = S_e(F_j) \] and so for all $i \in \mathcal{V}$ we have \[ F_i = \bigcup_{j = 1}^{N} \bigcup_{e \in \mathcal{E}_{i,j}} F_A^e \] and, moreover, for any edge $e$ which finishes at $i$ \[ \bigcup_{j = 1}^{N} \mathcal{E}_{i,j} = \{ f \in \mathcal{E} : A_{e,f} = 1\} \] as required. \end{proof} Proposition \ref{prop_gda2ssoft} says that the solution of every graph-directed self-similar iterated function system is a subshift of finite type in some sense. This is a folklore result and appears for example in \cite[Proposition 2.2.6]{lind}. The next proposition states that the converse is true which will be useful in Section \ref{packing_sec}. Again this is a folklore result and appears for example in \cite[Proposition 2.3.9]{lind}. We include both results and their simple proofs for completeness. \begin{prop} \label{prop_ssoft2gda} Let $\Sigma_{A}$ be a subshift of finite type for the alphabet $\mathcal{I}$ where $A$ has at least one non-zero entry in every row. Then there exits a graph-directed self-similar iterated function system with directed graph $\Gamma = G(\mathcal{I}, \mathcal{E})$ with solution $\{F_{A}^{i}\}_{i \in \mathcal{I}}$. If $A$ is irreducible then $\Gamma$ is strongly connected. \end{prop} \begin{proof} We draw a directed edge $e=e_{i,j}$ from $i$ to $j$ if $A_{i,j}=1$, let $S_{e}=S_{i}$ and let $\mathcal{E}=\{e_{i,j}:i,j \in \mathcal{I}, A_{i,j}=1\}$. If $A$ is irreducible then $\Gamma$ is strongly connected. We have that \[ F_{A}^{i} \ =\bigcup_{j \in \mathcal{I},A_{i,j}=1} S_{i}\big(F_A^{j}\big) \ = \ \bigcup_{j \in \mathcal{I}} \bigcup_{e \in \mathcal{E}_{i,j}} S_e\big(F_A^j\big) \] and since there is a unique set of compact attractors associated to this graph-directed system, the proposition follows. \end{proof} \subsection{Extension to $k$-block subshifts of finite type} We only consider 2-block subshifts of finite type in this paper, i.e. where the forbidden words are of length 2, but note that our results can be extended to the more general $k$-block case, where the forbidden words are of length $k$. This is a natural simplification to make, as one can always reformulate a $k$-block subshift of finite type as a 2-block analogue over a larger alphabet, see \cite[Theorem 2.3.2]{lind}. Moreover, this can be done so that the two systems are topologically conjugate which means that for irreducible $k$-block systems the associated $2$-block system remains irreducible. The reformulation is straightforward and standard. The new alphabet is the set of words of length $(k-1)$ such that there is an allowable word of length $k$ beginning with that word of length $(k-1)$. Then, the 2-word (over the new alphabet) consisting of $(i_0, i_1, \dots, i_{k-2})$ followed by $(i_1, i_2 ,\dots, i_{k-1})$ is allowed if and only if $(i_0, i_1, \dots, i_{k-1})$ was allowed in the original $k$-block system. There is a naturally induced homeomorphism which conjugates the $k$-block system to the new $2$-block system. \section{Ahlfors regularity and the weak separation property} Our results have applications in studying Ahlfors regularity of self-similar sets and related fractals. Recall that a bounded set $F \subseteq \mathbb{R}^n$ with Hausdorff dimension $s$ is called \emph{Ahlfors regular} if there exists a constant $c \geq 1$ such that for all $r \in (0, \text{diam}(F) ]$ and $x \in F$ \[ c^{-1} r^s \, \leq \, \mathcal{H}^s\big( F \cap B(x,r) \big) \, \leq \, c r^s. \] It is straightforward to show that for an Ahlfors regular set the Hausdorff measure and Hausdorff content are equivalent in the Hausdorff dimension (equal up to a constant bound). It is also well-known that a self-similar set satisfying the open set condition is Ahlfors regular. Our results yield the following corollary. \begin{cor} \label{ahlforscor} Let $A$ be irreducible and let $s = \dim_\text{\emph{H}} F_A$. Then $\mathcal{H}^s (F_A) >0 $ if and only if $F_A$ is Ahlfors regular. Moreover, this extends to any cylinder, i.e., for all $\textbf{i} \in \mathcal{I}^*$, $\mathcal{H}^s (F_A^\textbf{i}) >0 $ if and only if $F_A^\textbf{i}$ is Ahlfors regular. \end{cor} \begin{proof} We will prove the result for $F_A$; the result for cylinders is similar and omitted. Fix $r \in (0, \text{diam}(F_A) ]$ and $x \in F$. The lower bound is straightforward and follows by choosing a first level cylinder with positive measure and then finding a copy of this cylinder inside $F_A \cap B(x,r)$ with diameter comparable to $r$ and then applying the scaling property for Hausdorff measure. For the upper bound, \begin{eqnarray*} \mathcal{H}^s\big( F_A \cap B(x,r) \big) & \leq & \sum_{i \in \mathcal{I}} \mathcal{H}^s\big( F^i_A \cap B(x,r) \big) \\ \\ & = & \sum_{i \in \mathcal{I}} \mathcal{H}_\infty^s\big( F^i_A \cap B(x,r) \big) \qquad \text{by Theorem \ref{main} and Lemma \ref{haus11}}\\ \\ & \leq & \sum_{i \in \mathcal{I}} \text{diam}\big( F^i_A \cap B(x,r) \big)^s \\ \\ & \leq & M(2r)^s \end{eqnarray*} completing the proof. \end{proof} Observe that the above corollary also applies to any collection of cylinders in $F_A$ and in particular to graph-directed self-similar sets. Also, no separation conditions are assumed. \\ \\ Let $F \subseteq \mathbb{R}^n$ be a self-similar set, not contained in any affine hyperplane. Recall that the \emph{weak separation property} is satisfied if the identity map is \emph{not} an accumulation point of the set \[ \{ S_{\textbf{\textbf{\emph{i}}}}^{-1} \circ S_{\textbf{\emph{j}}} : \textbf{\textbf{\emph{i}}}, \textbf{\textbf{\emph{j}}} \in \mathcal{I}^* \} \] equipped with the uniform norm, see \cite{zerner}. It was shown in \cite[Theorem 2.1]{assouadoverlaps} that if $F$ satisfies the weak separation property (which is weaker than the open set condition) then it is Ahlfors regular. It was also shown \cite[Theorem 1.4]{assouadoverlaps} that if $F$ does not satisfy the weak separation condition then the Assouad dimension $\dim_\text{A} F$ of $F$ is greater than or equal to 1. In general the Assouad dimension is an upper bound for the Hausdorff dimension and we refer the reader to \cite{assouadoverlaps} for the definition. This allows us to prove the following corollary. \begin{cor} \label{ahlforscor2} Let $F\subseteq \mathbb{R}^n$ be a self-similar set with Hausdorff dimension $s<1$ not contained in any affine hyperplane. Then the following are equivalent: \begin{itemize} \item[(1)] $F$ satisfies the weak separation property \item[(2)] $\mathcal{H}^s (F) >0$ \item[(3)] $0<\mathcal{H}^s (F) < \infty$ \item[(4)] $F$ is Ahlfors regular \item[(5)] the Hausdorff and Assouad dimensions of $F$ coincide. \end{itemize} \end{cor} \begin{proof} Zerner \cite[Corollary after Proposition 2]{zerner} proved that (1) $\Rightarrow$ (2), (2) and (3) are equivalent since any self-similar set has finite Hausdorff measure in its Hausdorff dimension, see \cite[Corollary 3.3]{falconer}, our result, Corollary \ref{ahlforscor}, shows that (2) $\Leftrightarrow$ (4), the fact that (4) $\Rightarrow$ (5) is straightforward and folklore (see, for example, \cite[Proposition 2.1 (viii)]{tyson}), and since $\dim_\text{H} F < 1$ the result mentioned above \cite[Theorem 1.4]{assouadoverlaps} shows that (5) $\Rightarrow$ (1). \end{proof} The fact that (2) $\Rightarrow$ (1) provides a partial solution to a conjecture of Zerner, see the discussion following Proposition 2 in \cite{zerner}. We note that Corollary \ref{ahlforscor2} also shows that for self-similar sets with Hausdorff dimension strictly less than 1, the weak separation property can be formulated in a way which only depends on the set itself and not the defining iterated function system. The additional assumption $\dim_\text{H} F< 1$ required in the above corollary seems a little strange at first. However, it turns out that this condition is sharp. Firstly consider $F$ in the line. It is straightforward to construct a self-similar set $F \subseteq [0,1]$ which fails the weak separation property, but for which $\mathcal{H}^1 (F) >0$. For example, use the contractions $x \mapsto x/2$, $x \mapsto x/3$ and $x \mapsto x/2+1/2$ and apply the argument from \cite[Example 3.1]{fraser} using the fact that $\log 2/ \log 3 \notin \mathbb{Q}$. We use a variation of this example to prove the following proposition demonstrating the (almost) sharpness of Corollary \ref{ahlforscor2}. \begin{prop} \label{ahlforscor3} For all $n \in \mathbb{N}\setminus \{1\}$ and all $s \in (1,n]$, there exists a self-similar set $F \subseteq [0,1]^n$ not contained in any affine hyperplane such that \item[(1)] $F$ fails the weak separation property \item[(2)] $\dim_\text{\emph{H}} F = s$ \item[(3)] $\mathcal{H}^s (F) >0$ \end{prop} \begin{proof} Let $r \in (0,1/2]$ be chosen such that \[ \frac{\log 2}{-\log r} \ = \ \frac{s-1}{n-1} \ =: \ t \] and let $F_1 = [0,1]$ be viewed as a self-similar attractor of an iterated function system which fails the weak separation property and all of the maps have contraction ratio $r$. Such an iterated function system can be constructed by modifying \cite[Section 2 (v)]{bandt-graf}. Also, let $E \subseteq [0,1]$ be the self-similar set defined by the maps $x \mapsto rx$ and $x \mapsto rx+(1-r)$, and observe that $\dim_\text{H} E = t$ and $\mathcal{H}^t(E) >0$ since the open set condition is satisfied, see \cite[Corollary 3.3]{falconer}. Now let $F = F_1 \times E^{n-1} \subseteq [0,1]^n$ be the product of $F_1$ with $n-1$ copies of $E$. It is easy to see that $F$ is not contained in any affine hyperplane and that it is a self-similar set defined via the natural product iterated function system. It follows from \cite[Theorem 8.10]{mattila} that $\dim_\text{H} F = 1 + (n-1) t = s$ and that $\mathcal{H}^s (F) >0$. Note that to compute the dimension of $F$ here we used the fact that the Hausdorff and packing dimensions coincide for any self-similar set \cite[Corollary 3.3]{techniques}. Finally it is easy to see that the weak separation property fails by virtue of it failing in the first coordinate. \end{proof} For $s=n$ in the above proposition our set $F$ is just $[0,1]^n$, which is not very interesting. We point out that it is possible to construct a set with the desired properties but which has empty interior. For example, it was shown in \cite{jordanetal} that there exists a self-similar set in the plane with positive $\mathcal{H}^2$ measure, but empty interior, and by \cite[Theorem 3]{zerner} such a set must fail the weak separation property. We end this section by asking the natural question, an answer to which would complete the study. \begin{ques} Is it true that for all $n \in \mathbb{N}\setminus \{1\}$ there exists a self-similar set $F \subseteq [0,1]^n$ not contained in any affine hyperplane such that $F$ fails the weak separation property, $\dim_\text{\emph{H}} F = 1$ and $\mathcal{H}^1 (F) >0$? \end{ques} \section{Examples where $\mathcal{H}_\infty^s (F) < \mathcal{H}^s (F) < \infty$ and future work} \label{examples} In this section we give examples which show that equality of Hausdorff measure and Hausdorff content in the critical dimension is actually a rather special property. In particular, we give several examples falling into natural classes of set for which one might hope to be able to extend Theorem \ref{main}, but for which equality does not hold. A natural situation to consider is attractors of more general iterated function systems. In general an iterated function system (IFS) is a finite collection of contractions $\{S_i\}_{i \in \mathcal{I}}$ on a compact metric space. The \emph{attractor} of this system is the unique non-empty compact set $F$ satisfying \[ F = \bigcup_{i \in \mathcal{I}} S_i (F). \] See \cite[Chapter 9]{falconer} and \cite{hutchinson} for more details on iterated function systems. Two of the most standard and important generalisations of self-similar sets are \emph{self-affine sets}, where the defining maps are affine maps on some Euclidean space, and \emph{self-conformal sets}, where the defining maps are conformal. We note that similarities are both affine and conformal. It is evident that for any compact set $F \subset \mathbb{R}^n$ with Hausdorff dimension equal to 1, we have \[ \mathcal{H}_\infty^1(F) \leq \text{diam}( F ). \] However, if $F$ is connected and not contained in a straight line, then \[ \mathcal{H}^1(F) > \text{diam}(F ). \] This phenomenon provides us with several simple counter examples. \\ \\ \textbf{Self-affine sets:} It was shown in \cite{bandt} that there exist self-affine curves $C$ in the plane which are differentiable at all but countably many points. In particular, these curves can have finite length but not lie in a straight line (see \cite[Example 10]{bandt} and \cite[Example 6.2]{kaenmaki}). Such sets have Hausdorff dimension 1 and by the above argument satisfy \[ 0 < \mathcal{H}_\infty^1 (C) < \mathcal{H}^1 (C) < \infty. \] \textbf{Self-conformal sets:} The upper half $A$ of the unit circle in the complex plane is a self-conformal set and has \[ \mathcal{H}_\infty^1 (A) = 2 < \pi = \mathcal{H}^1 (A). \] The maps in the defining IFS for $A$ are $z \mapsto \sqrt{z}$ and $z \mapsto i \sqrt{z}$, defined on a suitable open domain containing $A$. \\ \\ \textbf{Julia sets:} the unit circle $S^1$ is the Julia set for the complex map $z \mapsto z^2$ and satisfies \[ \mathcal{H}_\infty^1 (S^1) = 2 < 2 \pi = \mathcal{H}^1 (S^1). \] \textbf{Sub-self-similar sets:} Sub-self-similar sets, introduced by Falconer in \cite{subselfsim}, are compact sets $F$ satisfying \[ F \ \subseteq \ \bigcup_{i \in \mathcal{I}} S_i(F) \] for some IFS of similarities. For any such IFS with the unit square as its attractor, the boundary of the unit square $Q = \partial [0,1]^2$ is a sub-self-similar set and satisfies \[ \mathcal{H}_\infty^1 (Q) = \sqrt{2} < 4 = \mathcal{H}^1 (Q). \] Finally we give two simple examples which show that Theorem \ref{main} is sharp, in some sense. \\ \\ \textbf{Non-irreducible subshift of finite type:} Consider the subshift of finite type on the alphabet $\{0,1,2\}$ given by the matrix \[ A \ = \ \left ( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 1 & 1 & 0 \\ \end{array} \right ) \] and associate any iterated function system consisting of three similarities on $[0,1]$ which map $[0,1]$ to three disjoint intervals. Here $A$ is not irreducible and so does not fall into the class considered by Theorem \ref{main}. The limit set $F = \Pi\big(\Sigma_A\big)$ consists of only four points and so $F$ and all of its children have Hausdorff dimension 0, but nevertheless \[ \mathcal{H}_\infty^0 \big(F_A^2\big) = 1 < 2 = \mathcal{H}^0 \big(F_A^2\big). \] \textbf{Full set for irreducible and aperiodic subshift of finite type:} Now we will show that one cannot hope to have $\mathcal{H}_\infty^s (F_A) = \mathcal{H}^s (F_A)$ for even an \emph{aperiodic} subshift of finite type (which we recall is a stronger condition than irreducible). Consider the alphabet $\{0,1,2,3\}$ and let \[ A \ = \ \left ( \begin{array}{cccc} 1 & 1 & 0 & 0 \\ 0 & 0 & 1 & 1 \\ 0 & 0 & 1 & 1 \\ 1 & 1 & 0 & 0 \\ \end{array} \right ) \] which is quickly seen to be aperiodic. Define similarities on the unit square by \[ S_0(x,y) = (x/2,y/2), \qquad \qquad S_1(x,y) = (-x/2,y/2) + (1/2, 1/2), \] \[ S_2(x,y) = (x/2,y/2) + (1/2,0), \quad \text{and} \quad S_3(x,y) = (-x/2,y/2) + (1, 1/2). \] It is easy to see that \[ F_A =( \{0\} \times [0,1]) \cup (\{1\} \times [0,1]) \] which satisfies \[ \mathcal{H}_\infty^1 (F_A) = \sqrt{2} < 2 = \mathcal{H}^1 (F_A). \] Of course Theorem \ref{main} still correctly states that \[ \mathcal{H}^1_\infty \big(F_A^0 \cup F_A^1\big) = \mathcal{H}^1 \big(F_A^0 \cup F_A^1\big) \quad \text{and} \quad \mathcal{H}_\infty^1 \big(F_A^2 \cup F_A^3\big) = \mathcal{H}^1\big(F_A^2 \cup F_A^3\big), \] noting that \[ F_A^0 \cup F_A^1 = \{0\} \times [0,1] \quad \text{and} \quad F_A^2 \cup F_A^3 = \{1\} \times [0,1]. \] A possible direction for further study on this topic would be to consider the classes of sets studied in this section, namely, self-conformal, self-affine, sub-self-similar, or Julia sets, and try to prove that the Hausdorff measure and Hausdorff content agree in some interesting subclass. Alternatively, one could look for negative results, which prove that the Hausdorff measure and Hausdorff content are always distinct in certain subclasses. Also, all of our counter examples in these classes were using sets with dimension 1. Could there be different phenomena at work for non-integral dimensions? We suspect not, but have not investigated this further. Note that we cannot give a simple condition guaranteeing $\mathcal{H}_\infty^s(F) < \mathcal{H}^s(F)$ apart from for connected sets $F$ not lying in a straight line with Hausdorff dimension $s=1$. This is because such sets may be $s$-straight by the result of Delaware mentioned previously \cite{delaware}. \section{The question of packing measure} \label{packing_sec} In this section we address the question of whether analogous results can be obtained for packing measure and a suitably defined `packing content'. First we recall the definition of the packing measure. Packing measure, defined in terms of \emph{packings}, is a natural dual to Hausdorff measure, which was defined in terms of \emph{covers}. For $s \geq 0$ and $\delta>0$ the $\delta$-approximate $s$-dimensional packing pre-measure of $F$ is defined by \begin{eqnarray*} \mathcal{P}_{\delta}^s (F) = \sup \bigg\{ \sum_{k=1}^{\infty} \text{diam}(U_k )^s : \{ U_k \}_{k=1}^{\infty} \text{ is a countable collection of pairwise disjoint } \\ \text{closed balls centered in $F$ with $\text{diam}(U_k) \leq \delta$ for all $k$} \bigg\} \end{eqnarray*} and the $s$-dimensional packing pre-measure of $F$ by $\mathcal{P}_0^s (F) = \lim_{\delta \to 0} \mathcal{P}_{\delta}^s (F)$. To ensure countable subadditivity, the packing (outer) measure of $F$ is defined by \[ \mathcal{P}^s (F) = \inf \bigg\{\sum_i \mathcal{P}_0^s (F_i) : F \subseteq \bigcup_i F_i \bigg\}. \] It follows from the definition that \begin{equation} \mathcal{P}^s (F)\leq \mathcal{P}_{0}^s (F)\leq \mathcal{P}_{\delta}^s (F).\label{packineq} \end{equation} Similar to the Hausdorff dimension, the packing dimension of $F$ is defined to be \[ \dim_\text{P} F = \inf \Big\{ s \geq 0: \mathcal{P}^s (F) =0 \Big\}. \] The extra step in the definition of packing measure makes it often more difficult to handle than the Hausdorff measure. However, in our setting there is a useful simplification due to Feng-Hua-Wen \cite{packingmeasure} and Haase \cite{haase}. \begin{prop} Let $F$ be a compact subset of $\mathbb{R}^n$ with the property that for every open ball $B$ centered in $F$, there exists a bi-Lipschitz map $S$ on $\mathbb{R}^n$ such that $S(F) \subseteq B \cap F$. Then for all $s \geq 0$ we have \[ \mathcal{P}^s (F) \ = \ \mathcal{P}_0^s (F). \] \end{prop} \begin{proof} For any compact set $F \subset \mathbb{R}^n$, if $\mathcal{P}^s_0 (F) < \infty$, then $\mathcal{P}^s (F) \ = \ \mathcal{P}_0^s (F)$, by the main result in \cite{packingmeasure}. In the case when $\mathcal{P}^s_0 (F) = \infty$, the additional assumption implies that $\mathcal{P}^s_0 (B \cap F) = \infty$ for all open balls intersecting $F$, which by \cite[Lemma 4]{haase}, implies that $\mathcal{P}^s(F) = \infty$. \end{proof} For this reason we can concern ourselves only with the packing pre-measure, which is easier to understand. The first question is, how should we define the packing (pre) content? If we naively define it by just removing the bounds on the diameters of the balls in the packing, then the answer is always infinity, as long as $s>0$ and $F\neq\emptyset$. This is because one can just take a packing by a single ball with unbounded diameter. Possible alternatives would be either to insist that there are at least two balls in every packing, or to bound the radii by something concrete, such as the diameter of $F$ itself. However, it might be more natural to try to prove that for sufficiently small $\delta$, the equality $\mathcal{P}_0^s (F) \ = \ \mathcal{P}_{\delta}^s (F)$ is satisfied. We adopt this third approach. The next question is, do we expect this to be true in the same setting as Theorem \ref{main}? An archetypal question being: \\ \\ ``\emph{If $F$ is self-similar, then does there exists a $\delta_0>0$ such that for all $\delta \in (0, \delta_0)$ we have} \[ \mathcal{P}^s (F) \ = \ \mathcal{P}_0^s (F) \ = \ \mathcal{P}_{\delta}^s (F)?\text{''} \] One strange consequence of this would be that for such sets the packing measure is always strictly positive. In the same way that $\mathcal{H}^s_\delta(F)$ is always finite for bounded sets, we have that $\mathcal{P}^s_{\delta}(F)$ is always positive for arbitrary non-empty sets. Interestingly enough it was an important question for about 15 years whether or not it was possible for a self-similar set to have zero packing measure in its dimension, see \cite{problems_update}, but this was recently resolved by Orponen \cite{orponen}, who provided a family of self-similar sets for whose elements $F$ (of course not satisfying the open set condition) $ \mathcal{P}^{\dim_\P F} (F) = 0$. Thus the answer to the above question is immediately `no'. We have managed to prove a weaker result, however, which we state after briefly recalling the \emph{strong separation condition}. This is a strictly stronger condition than the open set condition and is satisfied if the images of the attractor under the maps in the defining system are pairwise disjoint. We also recall that for any self-similar set, the packing measure must be finite in the packing dimension, see \cite[Exercise 3.2]{techniques}. \begin{thm} \label{packing} Let $F \subseteq \mathbb{R}^n$ be a self-similar set which satisfies the strong separation condition and let $s = \dim_\text{\emph{P}} F$. Then, there exists a $\delta_0>0$ such that for all $\delta \in (0, \delta_0)$ we have \[ 0< \mathcal{P}^s (F) \ = \ \mathcal{P}_0^s (F) \ = \ \mathcal{P}_{\delta}^s (F)< \infty. \] \end{thm} We will prove Theorem \ref{packing} in Section \ref{packingproof}. By the above discussion, this result does not extend to $F$ which do not satisfy the open set condition. It is also easy to see that it does not extend to the open set condition case either. For example, the unit interval $I$ is a self-similar set satisfying the open set condition but not the strong separation condition. Elementary calculations yield that $\mathcal{P}^1(I) = 1$, but that $\mathcal{P}_\delta^1 (I)= 1+\delta$ for all $\delta$. We pose the question of whether the appropriate converse of Theorem \ref{packing} is true. \begin{ques} Does there exists a self-similar set $F$ satisfying the open set condition, but for which there is no IFS of similarities satisfying the strong separation condition with $F$ as the attractor, for which there exists a $\delta_0>0$ such that for all $\delta \in (0, \delta_0)$ we have \[ 0< \mathcal{P}^s (F) \ = \ \mathcal{P}_0^s (F) \ = \ \mathcal{P}_{\delta}^s (F)< \infty? \] \end{ques} We generalise Theorem \ref{packing} for graph-directed self-similar sets and subshifts of finite type. A graph-directed self-similar iterated function system satisfies the \emph{strong separation condition} if (\ref{ssc-gda}) is a disjoint union for every $i$. \begin{thm} \label{packing-gda} Let $\{F_{i}\}_{i \in \mathcal{V}}$ be the solution of a graph-directed self-similar iterated function system which satisfies the strong separation condition and let $s$ be the common packing dimension of the sets $\{F_i\}_{i \in \mathcal{V}}$. Then, there exists a $\delta_0>0$ such that for all $\delta \in (0, \delta_0)$ and all $i\in \mathcal{V}$ we have \[ 0< \mathcal{P}^s (F_i) \ = \ \mathcal{P}_0^s (F_i) \ = \ \mathcal{P}_{\delta}^s (F_i)< \infty. \] \end{thm} We will prove Theorem \ref{packing-gda} in Section \ref{packingproof}. Due to Proposition \ref{prop_ssoft2gda} it follows that this result generalises to subshift of finite types. \begin{thm} \label{packing-ssoft} Let $\Sigma_{A}$ be an irreducible subshift of finite type on the alphabet $\mathcal{I}$ and let $s = \dim_\text{\emph{P}} F_A$. Assume that \begin{equation} \label{ssoft-ssc} \{F_A^j\}_{j \in \mathcal{I} : A_{i,j} = 1} \end{equation} are disjoint for every $i\in \mathcal{I}$. Then, there exists a $\delta_0>0$ such that for all $\delta \in (0, \delta_0)$ and all $i\in \mathcal{I}$ we have \[ 0< \mathcal{P}^s \big(F_A^i \big) \ = \ \mathcal{P}_0^s \big(F_A^i \big) \ = \ \mathcal{P}_{\delta}^s \big(F_A^i \big)< \infty. \] \end{thm} Theorem \ref{packing-ssoft} follows from Theorem \ref{packing-gda} and Proposition \ref{prop_ssoft2gda} since (\ref{ssoft-ssc}) ensures that the strong separation condition is satisfied for the graph-directed system in Proposition \ref{prop_ssoft2gda}. \section{Proofs} \subsection{A useful exhaustion lemma} In this section we prove an exhaustion lemma for Hausdorff measure, similar to \cite[Proposition 1.9]{farkas}, which may be of interest in its own right. It shows that we can exhaust the Hausdorff measure of a (potentially overlapping) subset of a self-similar set modelled by a subshift of finite type by infinitely many, disjoint, images of first level cylinders. First we state a version of Vitali's covering theorem. Let $H\subset\mathbb{R}^{n}$. A collection of sets $\mathcal{A}$ is called a \textsl{Vitali cover} of $H$ if for each $x\in H$ and $\delta>0$ there exist $A\in\mathcal{A}$ with $x\in A$ and $0<\mathrm{diam}(A)<\delta$. \begin{prop} \label{lem:VCT} Let $H\subset\mathbb{R}^{n}$ be a $\mathcal{H}^{s}$-measurable set with $\mathcal{H}^{s}(H)<\infty$ and $B_{1},\ldots,B_{m}\subset\mathbb{R}^{n}$ be closed sets with $0<\mathrm{diam}(B_{i})<\infty$ and $0<\mathcal{H}^{s}(B_{i})<\infty$ for all $i\in\left\{ 1,\ldots,m\right\} $. Let $\mathcal{A}$ be a Vitali cover of $H$ such that every element of $\mathcal{A}$ is similar to $B_{i}$ for some $i\in\left\{ 1,\ldots,m\right\} $ and every element of $\mathcal{A}$ is a subset of $H$. Then there exists a disjoint sequence of sets (finite or countable) $A_{1},A_{2},\ldots\in\mathcal{A}$ such that $\mathcal{H}^{s}\left(H\setminus\left(\bigcup_{i=1}^{\infty}A_{i}\right)\right)=0$. \end{prop} \begin{proof} Assume that $A_{1},A_{2},\ldots\in\mathcal{A}$ is a disjoint sequence of sets. Let $M=\max_{1\leq i\leq m}\frac{\mathrm{diam}(B_{i})^{s}}{\mathcal{H}^{s}(B_{i})}$. If $A_{i}$ is similar to $B_{j}$ then \[ \mathrm{diam}(A_{i})^{s} \ = \ \mathcal{H}^{s}(A_{i})\, \frac{\mathrm{diam}(B_{j})^{s}}{\mathcal{H}^{s}(B_{j})} \ \leq \ \mathcal{H}^{s}(A_{i})\, M. \] Hence \[ \sum_{i=1}^{\infty}\mathrm{diam}(A_{i})^{s} \ = \ \sum_{i=1}^{\infty}\mathcal{H}^{s}(A_{i})\, M \ \leq \ \mathcal{H}^{s}(H)\, M \ < \ \infty. \] Thus the proposition follows from a version of Vitali's covering theorem \cite[Theorem 1.10]{vitali}. \end{proof} Let \[ \mathcal{I}_{A}^{*}=\left\{ \textbf{\emph{i}}\in\mathcal{I}^{*}:\exists\alpha\in\Sigma_{A}\, \text{and} \, k\in\mathbb{N}\, \text{such that} \,\alpha\vert_{k}=\textbf{\emph{i}}\right\} \] and for $\textbf{\emph{i}}\in\mathcal{I}_{A}^{*}$ let \[ \mathcal{I}_{A}^{\textbf{\emph{i}}*}=\left\{ \textbf{\emph{j}}\in\mathcal{I}_{A}^{*}:\textbf{\emph{j}}\vert_{\left|\textbf{\emph{i}}\right|}=\textbf{\emph{i}}\right\}. \] For $\textbf{\emph{i}}=(i_{0},\ldots,i_{k-1})\in\mathcal{I}^{*}$ with $\left|\textbf{\emph{i}}\right|\geq1$ we define $(\textbf{\emph{i}})_{0}=i_{0}$ and $(\textbf{\emph{i}})_{last}=i_{k-1}$ and $\tau(\textbf{\emph{i}})=(i_{0},\ldots,i_{k-2})$. If $\textbf{\emph{i}}=(i_{0},\ldots,i_{k-1}),\textbf{\emph{j}}=(j_{0},\ldots,j_{l-1})\in\mathcal{I}_{A}^{*}$ are such that $A_{(\textbf{\emph{i}})_{last},(\textbf{\emph{j}})_{0}}=1$ then we write $\textbf{\emph{i}}*\textbf{\emph{j}}=(i_{0},\ldots,i_{k-1},j_{0},\ldots,j_{l-1})\in\mathcal{I}_{A}^{*}$. \begin{lma} \label{exhaust} Let $A$ be an irreducible subshift of finite type, let $s=\dim_{\emph{H}}F_{A}$ and assume that $\mathcal{H}^{s}(F_{A})>0$. Then for each $j\in\mathcal{I}$, there exists a collection $\mathcal{I}_{\infty}^{j}$ of finite words $\textbf{i}\in\mathcal{I}^{*}$ that satisfies the following properties: \noindent (i) the first symbol is $j$, i.e. $(\textbf{i})_0=j$, \noindent (ii) the last symbol is $j$, i.e. $(\textbf{i})_{last}=j$, \noindent (iii) there exists $\alpha\in\Sigma_{A}$ and $k\in\mathbb{N}$ such that $\alpha\vert_{k}=\textbf{i}$ or, in other words, $\textbf{i}\in\mathcal{I}^{*}_A$, \noindent (iv) for $\textbf{i},\textbf{j}\in\mathcal{I}_{\infty}^{j}$ with $\textbf{i}\neq\textbf{j}$ we have that \[ F_{A}^{\textbf{i}}\cap F_{A}^{\textbf{j}}=\emptyset, \] (v) \[ \mathcal{H}^{s}\left(F_{A}^{j}\setminus\left(\bigcup_{\textbf{i}\in\mathcal{I}_{\infty}^{j}}F_{A}^{\textbf{i}}\right)\right)=0, \] \noindent (vi) the contraction ratios satisfy a Hutchinson-Moran type expression for the Hausdorff dimension, i.e. \[ \sum_{\textbf{i}\in\mathcal{I}_{\infty}^{j}}r_{\tau(\textbf{i})}^{s}=1. \] \end{lma} \begin{proof} Since $A$ is irreducible, for every $i\in\mathcal{I}\setminus\left\{ j\right\} $ we can find $\textbf{\emph{i}}_{i}\in\mathcal{I}_{A}^{i*}$ such that $(\textbf{\emph{i}}_{i})_{0}=i$ and $(\textbf{\emph{i}}_{i})_{last}=j$. Of course there are infinitely many such choices for $\textbf{\emph{i}}_{i}$, but for definiteness choose one with minimal length. Thus if $\textbf{\emph{i}}\in\mathcal{I}_{A}^{j*}$ and $(\textbf{\emph{i}})_{last}=i$ then $\tau(\textbf{\emph{i}})*\textbf{\emph{i}}_{i}\in\mathcal{I}_{A}^{j*}$ and $(\tau(\textbf{\emph{i}})*\textbf{\emph{i}}_{i})_{last}=j$. \foreignlanguage{english}{Let \begin{equation} r_{\mathrm{\mathrm{min}}}=\min\left\{ \frac{\mathcal{H}^{s}\left(F_{A}^{\textbf{\emph{i}}_{i}}\right)}{\mathcal{H}^{s}\left(F_{A}^{i}\right)}:i\in\mathcal{I}\setminus\left\{ j\right\} \right\} \in(0,1).\label{eq:r_min} \end{equation} } We define a sequence $\mathcal{I}_{0}^{j},\mathcal{I}_{1}^{j},\ldots$ inductively where $\mathcal{I}_{n}^{j}$ satisfies properties \textit{(i), (iii), (iv)} and \textit{(v)}. The collection of sets $\left\{ F_{A}^{\textbf{\emph{i}}}:\textbf{\emph{i}}\in\mathcal{I}_{A}^{j*},\left|\textbf{\emph{i}}\right|\geq2\right\} $ is a Vitali cover of $F_{A}^{j}$ and hence by Proposition \ref{lem:VCT} there exists $\mathcal{I}_{0}^{j}\subseteq\left\{ \textbf{\emph{i}}\in\mathcal{I}_{A}^{j*}:\left|\textbf{\emph{i}}\right|\geq2\right\} $ such that $F_{A}^{\textbf{\emph{i}}}\cap F_{A}^{\textbf{\emph{j}}}=\emptyset$ for $\textbf{\emph{i}},\textbf{\emph{j}}\in\mathcal{I}_{0}^{j}$, $\textbf{\emph{i}}\neq\textbf{\emph{j}}$ and \[ \mathcal{H}^{s}\left(F_{A}^{j}\setminus\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{0}^{j}}F_{A}^{\textbf{\emph{i}}}\right)\right)=0. \] Once $\mathcal{I}_{n}^{j}$ is defined we define $\mathcal{I}_{n+1}^{j}$ as follows. First, for each $\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}$ we define a set $\mathcal{I}_{n+1,\textbf{\emph{i}}}^{j}$. If $(\textbf{\emph{i}})_{last}=j$ then $\mathcal{I}_{n+1,\textbf{\emph{i}}}^{j}=\left\{ \textbf{\emph{i}}\right\} $. If $(\textbf{\emph{i}})_{last}=i\neq j$ then \[ \left\{ F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}}:\textbf{\emph{j}}\in\mathcal{I}_{A}^{i*},F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}}\cap F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}_{i}}=\emptyset\right\} \] is a Vitali cover of $F_{A}^{\textbf{\emph{i}}}\setminus F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}_{i}}$ and hence by Proposition \ref{lem:VCT} there exists \[ \mathcal{J}_{n+1,\textbf{\emph{i}}} \ \subseteq \ \left\{ \textbf{\emph{j}}:\textbf{\emph{j}}\in\mathcal{I}_{A}^{i*},F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}}\cap F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}_{i}}=\emptyset\right\} \] such that $F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{i}}_{1}}\cap F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{i}}_{2}}=\emptyset$ for all $\textbf{\emph{i}}_{1},\textbf{\emph{i}}_{2}\in\mathcal{J}_{n+1,\textbf{\emph{i}}}^{j}$, with $\textbf{\emph{i}}_{1}\neq\textbf{\emph{i}}_{2}$, and \[ \mathcal{H}^{s}\left(\left(F_{A}^{\textbf{\emph{i}}}\setminus F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}_{i}}\right)\setminus\left(\bigcup_{\textbf{\emph{j}}\in\mathcal{J}_{n+1,\textbf{\emph{i}}}^{j}}F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}}\right)\right) \ = \ 0. \] Now let \[ \mathcal{I}_{n+1,\textbf{\emph{i}}}^{j} \ = \ \left\{ \tau(\textbf{\emph{i}})*\textbf{\emph{j}}_{i}\right\} \, \cup \, \left\{ \tau(\textbf{\emph{i}})*\textbf{\emph{j}}:\textbf{\emph{j}}\in\mathcal{J}_{n+1,\textbf{\emph{i}}}^{j}\right\} \] and \[ \mathcal{I}_{n+1}^{j} \ = \ \bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{n}}\mathcal{I}_{n+1,\textbf{\emph{i}}}^{j}. \] \selectlanguage{english Finally we define \[ \mathcal{I}_{\infty}^{j} \ = \ \bigcap_{n_{1}=1}^{\infty}\bigcup_{n_{2}=n_{1}}^{\infty}\mathcal{I}_{n_{2}}^{j}. \] Clearly $F_{A}^{\textbf{\emph{i}}}\cap F_{A}^{\textbf{\emph{j}}}=\emptyset$ for $\textbf{\emph{i}},\textbf{\emph{j}}\in\mathcal{I}_{\infty}^{j}$, $\textbf{\emph{i}}\neq\textbf{\emph{j}}$. If $\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}$ and $(\textbf{\emph{i}})_{last}\neq j$ then $\textbf{\emph{i}}\notin\mathcal{I}_{n+l}^{j}$ for every positive integer $l$, hence $\textbf{\emph{i}}\notin\mathcal{I}_{\infty}^{j}$. So $(\textbf{\emph{i}})_{last}=j$ for all $\textbf{\emph{i}}\in\mathcal{I}_{\infty}^{j}$. Clearly \begin{equation} \mathcal{H}^{s}\left(F_{A}^{j}\setminus\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}}F_{A}^{\textbf{\emph{i}}}\right)\right) \ = \ 0\label{eq:Egyenlito lem 1} \end{equation} for every positive integer $n$. For $\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}$ such that $(\textbf{\emph{i}})_{last}=i\neq j$ we have that \[ \left\{ \textbf{\emph{j}}:\tau(\textbf{\emph{i}})*\textbf{\emph{j}}\in\mathcal{I}_{n+1},(\tau(\textbf{\emph{i}})*\textbf{\emph{j}})_{last}\neq j\right\} \ \subseteq \ \mathcal{J}_{n+1,\textbf{\emph{i}}}^{j} \] and \begin{eqnarray} \mathcal{H}^{s} \left(F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}_{i}}\right) \ \ = \ \ \mathcal{H}^{s}\left(S_{\tau(\textbf{\emph{i}})*\tau(\textbf{\emph{j}}_{i})}(F_{A}^{j})\right)&=& r_{\tau(\textbf{\emph{i}})}^{s}r_{\tau(\textbf{\emph{j}}_{i})}^{s}\mathcal{H}^{s}\left(F_{A}^{j}\right)\frac{\mathcal{H}^{s}\left(F_{A}^{i}\right)}{\mathcal{H}^{s}\left(F_{A}^{i}\right)} \nonumber \\ \nonumber \\ &=&\mathcal{H}^{s}\left(F_{A}^{\textbf{\emph{j}}_{i}}\right)\frac{\mathcal{H}^{s}\left(F_{A}^{\textbf{\emph{i}}}\right)}{\mathcal{H}^{s}\left(F_{A}^{i}\right)} \nonumber \\ \nonumber \\ &\geq& r_{\mathrm{\min}}\mathcal{H}^{s}\left(F_{A}^{\textbf{\emph{i}}}\right) \label{lowerestimate1} \end{eqnarray} by (\ref{eq:r_min}). Also $(\tau(\textbf{\emph{i}})*\textbf{\emph{j}}_{i})_{last}=j$ by definition. Therefore \[ \mathcal{I}_{n+1}^{j}\setminus\mathcal{I}_{\infty}^{j} \ \subseteq \ \bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}\setminus\mathcal{I}_{\infty}^{j}}\left\{ \tau(\textbf{\emph{i}})*\textbf{\emph{j}}:\textbf{\emph{j}}\in\mathcal{J}_{n+1,\textbf{\emph{i}}}^{j}\right\} \] and \begin{eqnarray*} \mathcal{H}^{s}\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{n+1}^{j}\setminus\mathcal{I}_{\infty}^{j}}F_{A}^{\textbf{\emph{i}}}\right) &\leq& \sum_{\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}\setminus\mathcal{I}_{\infty}^{j}}\mathcal{H}^{s}\left(\bigcup_{\textbf{\emph{j}}\in\mathcal{J}_{n+1,\textbf{\emph{i}}}^{j}}F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}}\right)\\ \\ &\leq& \sum_{\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}\setminus\mathcal{I}_{\infty}^{j}}\mathcal{H}^{s}\left( F_{A}^{\textbf{\emph{i}}} \setminus F_{A}^{\tau(\textbf{\emph{i}})*\textbf{\emph{j}}_{i}} \right)\\ \\ &\leq& \sum_{\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}\setminus\mathcal{I}_{\infty}^{j}}\left(\mathcal{H}^{s}\left(F_{A}^{\textbf{\emph{i}}}\right)-r_{\mathrm{min}}\mathcal{H}^{s}\left(F_{A}^{\textbf{\emph{i}}}\right)\right) \qquad \text{ by (\ref{lowerestimate1})}\\ \\ &=&\sum_{\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}\setminus\mathcal{I}_{\infty}^{j}}(1-r_{\mathrm{min}})\, \mathcal{H}^{s}\left(F_{A}^{\textbf{\emph{i}}}\right) \\ \\ &=&(1-r_{\mathrm{min}})\, \mathcal{H}^{s}\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{n}^{j}\setminus\mathcal{I}_{\infty}^{j}}F_{A}^{\textbf{\emph{i}}}\right). \end{eqnarray*} Hence \[ \mathcal{H}^{s}\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{n+1}^{j}\setminus\mathcal{I}_{\infty}^{j}}F_{A}^{\textbf{\emph{i}}}\right) \ \leq \ (1-r_{\mathrm{min}})^{n}\, \mathcal{H}^{s}\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{0}^{j}\setminus\mathcal{I}_{\infty}^{j}}F_{A}^{\textbf{\emph{i}}}\right) \ = \ (1-r_{\mathrm{min}})^{n}\, \mathcal{H}^{s}\left(F_{A}^{j}\right) \] for all $n\in\mathbb{N}$ and combined with (\ref{eq:Egyenlito lem 1}) we get that \[ \mathcal{H}^{s}\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{n+1}^{j}\cap\mathcal{I}_{\infty}^{j}}F_{A}^{\textbf{\emph{i}}}\right) \ \geq \ \big(1-(1-r_{\mathrm{min}})^{n}\big)\, \mathcal{H}^{s}\left(F_{A}^{j}\right). \] Thus \[ \mathcal{H}^{s}\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{\infty}^{j}}F_{A}^{\textbf{\emph{i}}}\right) \ \geq \ \mathcal{H}^{s}\left(F_{A}^{j}\right) \] and so \[ \mathcal{H}^{s}\left(F_{A}^{j}\setminus\left(\bigcup_{\textbf{\emph{i}}\in\mathcal{I}_{\infty}^{j}}F_{A}^{\textbf{\emph{i}}}\right)\right) \ = \ 0. \] \selectlanguage{british Thus the collection $\mathcal{I}_{\infty}^{j}$ satisfies properties \textit{(i)-(v)}. Property \textit{(vi)} follows easily from \textit{(iv)} and \textit{(v)} since \[ \mathcal{H}^{s}\left(F_{A}^{j}\right) \ = \ \sum_{\textbf{\emph{i}}\in\mathcal{I}_{\infty}^{j}}\mathcal{H}^{s}\left(F_{A}^{\textbf{\emph{i}}}\right) \ = \ \sum_{\textbf{\emph{i}}\in\mathcal{I}_{\infty}^{j}}\mathcal{H}^{s}\left(S_{\tau(\textbf{\emph{i}})}\left(F_{A}^{j}\right)\right) \ = \ \sum_{\textbf{\emph{i}}\in\mathcal{I}_{\infty}^{j}}r_{\tau(\textbf{\emph{i}})}^{s}\mathcal{H}^{s}\left(F_{A}^{j}\right) \] and the fact that we can divide by $\mathcal{H}^{s}\left(F_{A}^{j}\right)$. \end{proof} \subsection{Proof of Theorem \ref{main}} \label{mainproof} In this section we will prove our main result. It is trivially true if $\mathcal{H}^s(F_A) = 0$, so we assume otherwise. Fix $i \in \mathcal{I}$ and $\varepsilon>0$. Choose a countable open cover $\{U_k\}_{k \in \mathcal{K}}$ of $F_A^i$ which satisfies \begin{equation} \label{111} \sum_{k \in \mathcal{K}} \text{diam}( U_k )^s \ \leq \ \mathcal{H}^s_\infty(F_A^i) \, + \, \varepsilon. \end{equation} Since $F_A^i$ is bounded we can assume that there is a uniform bound on the diameters of the $U_k$. Let $\mathcal{I}^{i}_\infty$ be the `exhausting set' from Lemma \ref{exhaust}. For $m \in \mathbb{N}$, let \begin{equation} \label{abab} \mathcal{I}^{i,m}_\infty \ = \ \Big\{ \textbf{\emph{i}}' \in \mathcal{I}^* : \textbf{\emph{i}}' = \tau(\textbf{\emph{i}}^0) \tau(\textbf{\emph{i}}^1) \dots \tau(\textbf{\emph{i}}^{m-1}) \text{ where } \textbf{\emph{i}}^l \in \mathcal{I}^{i}_\infty \text{ for } l=0, \dots, m-1 \Big\}. \end{equation} By properties \emph{(i)} and \emph{(ii)} in Lemma \ref{exhaust} the set $\mathcal{I}^{i,m}_\infty$ is a set of restricted words from $\Sigma^{i}_A$. Moreover, it follows form property \emph{(v)} in Lemma \ref{exhaust} that, for all $m \in \mathbb{N}$, \begin{equation} \label{222} \mathcal{H}^s \left(F_A^i \setminus \bigcup_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty} S_\textbf{\emph{i}}(F_A^i) \right) \ = \ 0. \end{equation} Observe that, for each $m \in \mathbb{N}$, \[ \{ S_\textbf{\emph{i}}(U_k) \}_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty, k \in \mathcal{K}} \] is a cover of $\bigcup_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty} S_\textbf{\emph{i}}(F_A^i)$. Let $\delta>0$ and choose $m \in \mathbb{N}$ sufficiently large to ensure that \[ \sup_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty, \, k \in \mathcal{K}} \text{diam}\big( S_\textbf{\emph{i}}(U_k) \big) \leq \delta \] and thus \[ \{ S_\textbf{\emph{i}}(U_k) \}_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty, \, k \in \mathcal{K}} \] is a countable open $\delta$-cover of $\bigcup_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty} S_\textbf{\emph{i}}(F_A^i)$. It follows that \begin{eqnarray*} \mathcal{H}^s_\delta(F_A^i) &\leq& \mathcal{H}^s_\delta \left(\bigcup_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty} S_\textbf{\emph{i}}(F_A^i) \right) \ + \ \mathcal{H}^s_\delta \left(F_A^i \setminus \bigcup_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty} S_\textbf{\emph{i}}(F_A^i) \right) \\ \\ &\leq& \sum_{k \in \mathcal{K}} \sum_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty} \text{diam}\big( S_\textbf{\emph{i}}(U_k) \big)^s \qquad \qquad \text{by (\ref{222})} \\ \\ &\leq& \sum_{k \in \mathcal{K}} \text{diam}( U_k )^s \sum_{\textbf{\emph{i}} \in \mathcal{I}^{i,m}_\infty} r_\textbf{\emph{i}}^s \\ \\ &\leq& \Big(\mathcal{H}^s_\infty(F_A^i) \, + \, \varepsilon \Big) \, \left(\sum_{\textbf{\emph{i}} \in \mathcal{I}^{i}_\infty} r_{\tau(\textbf{\emph{i}})}^s\right)^m \qquad \qquad \text{by (\ref{111}) and (\ref{abab})} \\ \\ &=& \mathcal{H}^s_\infty(F_A^i) \, + \, \varepsilon \end{eqnarray*} by property \emph{(vi)} from Lemma \ref{exhaust}. Taking the limit as $\delta \to 0$ and noting that $\varepsilon>0$ was arbitrary, yields $\mathcal{H}^s(F_A^i) \leq \mathcal{H}^s_\infty(F_A^i) $. The reverse inequality is always satisfied. \\ \\ The final part of Theorem \ref{main} follows by a simple trick. Let $i \in \mathcal{I}$ and observe that \[ F_A^i \ = \ S_i\left( \bigcup_{j \in \mathcal{I} : A_{i,j} = 1} F_A^j \right) \] and so \[ r_i^s \, \mathcal{H}_\infty^s \left( \bigcup_{j \in \mathcal{I} : A_{i,j} = 1} F_A^j \right) \ = \ \mathcal{H}_\infty^s (F_A^i) \ = \ \mathcal{H}^s (F_A^i) \ = \ r_i^s \, \mathcal{H}^s \left( \bigcup_{j \in \mathcal{I} : A_{i,j} = 1} F_A^j \right) \] where the middle equality is due to the first part of the theorem. Dividing by $r_i^s$ completes the proof. \hfill \qed \subsection{Proof of Theorem \ref{packing} and Theorem \ref{packing-gda}} \label{packingproof} \emph{Proof of Theorem \ref{packing}}. Let $F \subseteq \mathbb{R}^n$ be the self-similar attractor of the IFS $\{S_i\}_{i \in \mathcal{I}}$ and assume $F$ satisfies the strong separation condition. This implies that we can find a bounded open set $\mathcal{O} \subseteq \mathbb{R}^n$ such that $F \subset \mathcal{O}$ and $\bigcup_{i \in \mathcal{I}}S_i(\mathcal{O}) \subseteq \mathcal{O}$ is a disjoint union. Let \[ \delta_0 \, = \, \frac {1}{2} \, \inf_{x \in F} \inf_{y \in\mathbb{R}^n \setminus\mathcal{O}} \ \lvert x-y \rvert \] which is strictly positive since $F$ is closed. \\ \\ First assume that $\mathcal{P}^s_{\delta}(F)<\infty$ for every $\delta \in (0, \delta_0)$. Later we will see that $\mathcal{P}^s_{\delta}(F)=\infty$ is impossible for $\delta \in (0, \delta_0)$. Let $\varepsilon>0$, let $\delta \in (0, \delta_0)$ and let $\{ B_k\}_{k \in \mathcal{K}}$ be a countable collection of disjoint closed balls centered in $F$ with diameter less than or equal to $\delta$ which satisfies \begin{equation} \label{333} \sum_{k \in \mathcal{K}} \text{diam}( B_k )^s \ \geq \ \mathcal{P}^s_{\delta}(F) - \varepsilon. \end{equation} Since $B_k \subset \mathcal{O}$ for all $k \in \mathcal{K}$ and by the choice of $\mathcal{O}$, the collection \[ \{ S_\textbf{\emph{i}}(B_k) \}_{\textbf{\emph{i}} \in \mathcal{I}^m, k \in \mathcal{K}} \] is a countable collection of disjoint closed balls centered in $F$. Let $\eta \in (0, \delta)$ and choose $m \in \mathbb{N}$ so large so that \[ \sup_{\textbf{\emph{i}} \in \mathcal{I}^m, k \in \mathcal{K}}\text{diam}\big(S_\textbf{\emph{i}}(B_k)\big) \, \leq \, \eta. \] It follows that \begin{eqnarray*} \mathcal{P}^s_{\eta}(F) &\geq& \sum_{k \in \mathcal{K}} \sum_{\textbf{\emph{i}} \in \mathcal{I}^m} \text{diam}\big( S_\textbf{\emph{i}}(B_k)\big)^s \\ \\ &=& \sum_{k \in \mathcal{K}} \text{diam}( B_k )^s \sum_{\textbf{\emph{i}} \in \mathcal{I}^m} r_\textbf{\emph{i}}^s \\ \\ &\geq& \Big(\mathcal{P}^s_{\delta}(F) - \varepsilon \Big) \, \left(\sum_{i \in \mathcal{I}} r_i^s\right)^m \qquad \qquad \text{by (\ref{333})} \\ \\ &=& \mathcal{P}^s_{\delta}(F) - \varepsilon \end{eqnarray*} by the Hutchinson-Moran formula for (packing) dimension \cite{hutchinson}. Taking the limit as $\eta \to 0$ and noting that $\varepsilon>0$ was arbitrary, yields $\mathcal{P}^s(F)=\mathcal{P}^s_{0}(F) \geq \mathcal{P}^s_{\delta}(F) $. The reverse inequality is always satisfied by (\ref{packineq}), which completes the proof. \\ \\ Now assume that $\mathcal{P}^s_{\delta}(F)=\infty$ for some $\delta \in (0, \delta_0)$. Via a similar argument to the one above, this implies that $\mathcal{P}^s_{\eta}(F)>K$ for every $K>0$ and hence $\mathcal{P}^s_{0}(F)=\infty$ but this is a contradiction since every self-similar set has finite packing measure (and pre-measure) in the packing dimension, see \cite[Exercise 3.2]{techniques}. \hfill \qed \\ \\ The reason this proof cannot be extended to the open set condition case is because in that case the number $\delta_0$ may be zero and iterations of packings may no longer be packings. This is one of the reasons packings are sometimes more difficult to control than covers. The proof of Theorem \ref{packing-gda} is similar and we just provide a sketch. First we prove a simple lemma. We say $v \leq v'$ for vectors $v,v' \in \mathbb{R}^N$ if each entry in $v$ is less than or equal to the corresponding entry in $v'$. We say that $v$ is non-negative if $0\leq v$. Similar notations apply to matrices. \begin{lma} \label{matrixlem} Let $A$ be a non-negative irreducible matrix of spectral radius $1$ and $x$ be a non-negative vector such that $A^m x\leq x$ for large enough $m$. Then $Ax=x$. \end{lma} \begin{proof} Observe that $A^m$ is also an irreducible matrix with spectral radius $1$. Hence it follows from \cite[Theorem 1.3.28]{matrix} that $A^m x = x$ and therefore $Ax=x$ by the Perron-Frobenius theorem. \end{proof} \emph{Proof of Theorem \ref{packing-gda}}. Let $A^s$ be the matrix with $(i,j)$th entry given by \[ A^s_{i,j} \, = \, \sum_{e \in \mathcal{E}_{i,j}}r^s_e. \] Let $s$ be the unique value for which the spectral radius of the matrix $A^s$ is $1$. Let $u^{\intercal}=(\mathcal{P}^s(F_1),...,\mathcal{P}^s(F_N))$. If $\Gamma$ is strongly connected then $A^s$ is irreducible. If further the strong separation condition is satisfied then $0<\mathcal{P}^s(F_i)<\infty$ for every $i$ and $A^s u=u$ (see \cite[Corollary 3.5]{techniques}. Let $u_{\delta}^{\intercal}=(\mathcal{P}_{\delta}^s(F_1),...,\mathcal{P}_{\delta}^s(F_N))$. Since the strong separation condition is satisfied there exists a collection of open sets $\{\mathcal{O}_{i}\}_{i \in \mathcal{V}}$ such that $F_i\subseteq \mathcal{O}_i$ and \[ \bigcup_{j=1}^N \bigcup_{e \in \mathcal{E}_{i,j}} S_e(\mathcal{O}_j) \, \subseteq \, \mathcal{O}_i \] is a disjoint union for every $i$. Let \[ \delta_0 \, = \, \frac {1}{2} \, \min_{i\in \mathcal{V}}\inf_{x \in F_i} \inf_{y \in\mathbb{R}^n \setminus \mathcal{O}_i} \ \lvert x-y \rvert. \] A similar argument to the proof of Theorem \ref{packing} shows that for large enough $m$ depending on $\eta$ we have that \begin{equation} \label{Asdelta} (A^s)^m u_{\delta} \, \leq \, u_{\eta} \, \leq \, u_{\delta} \end{equation} for $\delta \in (0,\delta_0)$ and $0<\eta<\delta$. It follows by Lemma \ref{matrixlem} that equality holds in (\ref{Asdelta}). Hence $ u_{\delta}=u_{\eta}=u$ for $0<\eta<\delta<\delta_0$. \hfill \qed \section*{Acknowledgements} \'A.F. was financially supported by an EPSRC doctoral training grant. Most of this work took place whilst J.M.F. was a research fellow at the University of Warwick, where he was financially supported by the EPSRC grant EP/J013560/1. \'A.F. visited J.M.F. at the University of Warwick to work on this project and both authors thank the department for its hospitality. The authors thank Kenneth Falconer for helpful comments on the exposition of the paper. Finally the authors thank Thomas Jordan and Mike Todd for providing some helpful references.

https://arxiv.org/abs/2001.01792

Proof of the simplicity conjecture

In the 1970s, Fathi, having proven that the group of compactly supported volume-preserving homeomorphisms of the $n$-ball is simple for $n \ge 3$, asked if the same statement holds in dimension $2$. We show that the group of compactly supported area-preserving homeomorphisms of the two-disc is not simple. This settles what is known as the "simplicity conjecture" in the affirmative. In fact, we prove the a priori stronger statement that this group is not perfect.An important step in our proof involves verifying for certain smooth twist maps a conjecture of Hutchings concerning recovering the Calabi invariant from the asymptotics of spectral invariants defined using periodic Floer homology. Another key step, which builds on recent advances in continuous symplectic topology, involves proving that these spectral invariants extend continuously to area-preserving homeomorphisms of the disc. These two properties of PFH spectral invariants are potentially of independent interest.Our general strategy is partially inspired by suggestions of Fathi and the approach of Oh towards the simplicity question. In particular, we show that infinite twist maps, studied by Oh, are not finite energy homeomorphisms, which resolves the "infinite twist conjecture" in the affirmative; these twist maps are now the first examples of Hamiltonian homeomorphisms which can be said to have infinite energy. Another consequence of our work is that various forms of fragmentation for volume preserving homeomorphisms which hold for higher dimensional balls fail in dimension two.

\section{Introduction} \subsection{The main theorem} \label{sec:intro_main_theo} Let $(S,\omega)$ be a surface equipped with an area form. An {\bf area-preserving homeomorphism} is a homeomorphism which preserves the measure induced by $\omega$. Let $\Homeo_c(\D, \omega)$ denote the group of area-preserving homeomorphisms of the two-disc which are the identity near the boundary. Recall that a group is {\bf simple} if it does not have a non-trivial proper normal subgroup. The following fundamental question was raised\footnote{ The history of when this question was raised seems complicated. It is asked by Fathi in the paper \cite{fathi}. However, it has been suggested to us by Ghys that the question dates back to considerably earlier, possibly to Mather or Thurston. } in the 1970s: \begin{question}\label{que:disc} Is the group $\Homeo_c(\D, \omega)$ simple? \end{question} Indeed, the algebraic structure of the group of volume-preserving homeomorphisms has been well-understood in dimension at least three since the work of Fathi \cite{fathi} from the 70s; but, the case of surfaces, and in particular Question~\ref{que:disc}, has long remained mysterious. Question~\ref{que:disc} has been the subject of wide interest. For example, it is highlighted in the plenary ICM address of Ghys \cite[Sec.\ 2.2]{Ghys_ICM}; it appears on McDuff and Salamon's list of open problems \cite[Sec.\ 14.7]{mcduff-salamon}; it has been one of the main motivations behind the development of $C^0$-symplectic topology, which we will further discuss in Section \ref{sec:simplicity_high_dim}; for other examples, see \cite{Banyaga, fathi, Ghys_lecture, Bounemoura, LeRoux10, LeRoux-6Questions, EPP}. It has generally been believed since the early $2000s$ that the group $\Homeo_c(\D, \omega)$ is not simple: McDuff and Salamon refer to this as the {\bf simplicity conjecture}. Our main theorem resolves this conjecture in the affirmative. \begin{theo} \label{thm:main} The group $\Homeo_c(\D, \omega)$ is not simple. \end{theo} In fact, we can obtain an a priori stronger result. Recall that a group $G$ is called {\bf perfect} if its commutator subgroup $[G,G]$ satisfies $[G, G] = G$. The commutator subgroup $[G, G]$ is a normal subgroup of $G$. Thus, every non-abelian simple group is perfect. However, in the case of certain transformation groups, such as $\Homeo_c(\D, \omega)$, a general argument due to Epstein and Higman \cite{Epstein, Higman} implies that perfectness and simplicity are equivalent; see Proposition \ref{prop:commutators}. Hence, we obtain the following corollary. \begin{corol}\label{corol:not_perfect} The group $\Homeo_c(\D, \omega)$ is not perfect. \end{corol} We remark that in higher dimensions, the analogue of Theorem~\ref{thm:main} contrasts our main result: by \cite{fathi}, the group $\Homeo_c(\D^n, \mathrm{Vol})$ of compactly supported volume-preserving homeomorphisms of the $n$-ball is simple for $n \ge 3$. It also seems that the structure of $\Homeo_c(\D,\omega)$ is radically different from that of the group $\Diff_c(\D,\omega)$ of compactly supported area-preserving diffeomorphisms, as we will review below. \subsubsection{Background} % To place our main result in its appropriate context, we begin by reviewing the long and interesting history surrounding the question of simplicity for various subgroups of homeomorphism groups of manifolds. Our focus will be on compactly supported homeomorphisms/diffeomorphisms of manifolds without boundary in the component of the identity\footnote{The simplicity question is interesting only for compactly supported maps in the identity component, because this is a normal subgroup of the larger group. The group $\Homeo_c(\D, \omega)$ coincides with its identity component.}. In the 1930s, in the ``Scottish Book", Ulam asked if the identity component of the group of homeomorphisms of the $n$--dimensional sphere is simple. In 1947, Ulam and von Neumann announced in an abstract \cite{Ulam-vonNeumann} a solution to the question in the Scottish Book in the case $n=2$. % In the 50s, 60s, and 70s, there was a flurry of activity on this question and related ones. First, the works of Anderson \cite{Anderson}, Fisher \cite{Fisher}, Chernavski, Edwards and Kirby \cite{EK} led to the proof of simplicity of the identity component in the group of compactly supported homeomorphisms of any manifold. These developments led Smale to ask if the identity component in the group of compactly supported diffeomorphisms of any manifold is simple \cite{Epstein}. This question was answered affirmatively by Epstein \cite{Epstein}, Herman \cite{Herman}, Mather \cite{ MatherI, Mather74, MatherII}, and Thurston \cite{thurston}\footnote{More precisely, Epstein, Herman and Thurston settled the question in the case of smooth diffeomorphisms, while Mather resolved the case of $C^r$ diffeomorphisms for $r < \infty$ and $r \neq \mathrm{dim(M)}+1.$ The case of $r = \mathrm{dim(M)}+1$ remains open.}. The connected component of the identity in volume-preserving, and symplectic, diffeomorphisms admits a homomorphism, called {\bf flux}, to a certain abelian group. Hence, these groups are not simple when this homomorphism is non-trivial. Thurston proved, however, that the kernel of flux is simple in the volume-preserving setting for any manifold of dimension at least three; see \cite[Chapter 5]{Banyaga_book}. In the symplectic setting, Banyaga \cite{Banyaga} then proved that this group is simple when the symplectic manifold is closed; otherwise, it is not simple as it admits a non-trivial homomorphism, called {\bf Calabi}, and Banyaga showed that the kernel of Calabi is always simple. We will recall the definition of Calabi in the case of the disc in Section \ref{sec:prop-norm-subgr}. As mentioned above, the simplicity of the identity component in compactly supported volume-preserving homeomorphisms is well-understood in dimensions greater than two,\footnote{The group is trivial in dimension $1$.} thanks to the article \cite{fathi}, in which Fathi shows that, in all dimensions, the group admits a homomorphism, called ``mass-flow"; moreover, the kernel of mass-flow is simple in dimensions greater than two. On simply connected manifolds, the mass-flow homomorphism is trivial, and so the group is indeed simple in dimensions greater than two --- in particular, $\Homeo_c(\D^n, \mathrm{Vol})$, the higher dimensional analogue of the group in Question \ref{que:disc}, is simple for $n>2$, as stated above. Thus, the following rather simple picture emerges from the above % cases of the simplicity question. In the non-conservative setting, the connected component of the identity is simple. In the conservative setting, there always exists a natural homomorphism (flux, Calabi, mass-flow) which obstructs the simplicity of the group. However, the kernel of the homomorphism is always simple. \subsubsection{The case of surfaces and our case of the disc} Despite the clear picture above, established by the end of the 70s, the case of area-preserving homeomorphisms of surfaces has remained unsettled --- the simplicity question has remained open for the disc and more generally for the kernel of the mass-flow homomorphism\footnote{We review the mass-flow homomorphism and discuss possible ideas that build on the present work for establishing simplicity of its kernel in Section \ref{sec:questions_simplicity}.} --- underscoring the importance of answering Question~\ref{que:disc}. In fact, the case of area-preserving homeomorphisms of the disc does seem drastically different. For example, the natural homomorphisms flux, Calabi, and mass-flow mentioned above that obstruct simplicity are all continuous with respect to a natural topology on the group; however, we will show in Corollary \ref{corol:no_continuous_homomorphism} that there can not exist a continuous homomorphism out of $\Homeo_c(\D, \omega)$ with a proper non-trivial kernel, when $\Homeo_c(\D, \omega)$ is equipped with the $C^0$-topology; we will review the $C^0$-topology in \ref{sec:c0}. Of course, $\Homeo_c(\D, \omega)$ could still admit a discontinuous homomorphism. For example, Ghys asks, in his % ICM address \cite{Ghys_ICM}, whether the Calabi invariant extends to $\Homeo_c(\D, \omega)$; this, and related works of Fathi and Oh, are further discussed in Section \ref{sec:extending_Calabi}. However, as far as we know, even if one does not demand that the homomorphism is continuous, there is still no natural geometrically constructed homomorphism out of $\Homeo_c(\D, \omega)$ that would in any sense be an analogue of the flux, Calabi, or mass-flow; rather, as we will see, our proof of non-simplicity goes by explicitly constructing a non-trivial normal subgroup and showing that it is proper. It seems likely to us that $\Homeo_c(\D, \omega)$ indeed has a more complicated algebraic structure. We will now see in the discussion of Le Roux's work below another way in which the case of area-preserving homeomorphisms of the disc is quite different. \subsubsection{``Lots" of normal subgroups and the failure of fragmentation} \label{sec:lots} \sloppy Le Roux \cite{LeRoux10} has previously studied the simplicity question for $\Homeo_c(\D, \omega)$, and this provides useful context for our work; it is also valuable to combine his conclusions with our Theorem~\ref{thm:main}. As Le Roux explains, Fathi's proof of simplicity in higher-dimensions does not work in dimension $2$, because it relies on the following ``fragmentation'' result: any element of $\Homeo_c(\D^n, \mathrm{Vol})$ can be written as the product of two others, each of which are supported on topological balls of three-fourths of the total volume; this is not true on the disc. Building on this, Le Roux constructs a whole family $P_\rho$, for $0 < \rho \le 1$, of quantitative fragmentation properties for $\Homeo_c(\D, \omega)$: the property $P_\rho$ asserts that there is some $\rho' < \rho$ and a positive integer $N$, such that any group element supported on a topological disc of area $\rho$ can be written as a product of $N$ elements, each supported on discs of area at most $\rho'$; we refer the reader to \cite{LeRoux10} for more details. For example, if Fathi's fragmentation result held in dimension $2$, then $ P_1$ would hold. Le Roux then establishes the following alternative: if any one of these fragmentation properties holds, then $\Homeo_c(\D, \omega)$ is simple; otherwise, there is a huge number of proper normal subgroups, constructed in terms of ``fragmentation norms''. Thus, in view of our Theorem~\ref{thm:main}, we have not just one proper normal subgroup but ``lots" of them; for example, combining Le Roux's work \cite[Cor.\ 7.1]{LeRoux10} with our Theorem~\ref{thm:main} yields the following. \begin{corol} \label{cor:leroux} Every compact\footnote{As above, we are working in the $C^0$-topology.} subset of $\Homeo_c(\D,\omega)$ is contained in a proper normal subgroup. \end{corol} As Le Roux explains \cite[Section 7]{LeRoux10}, this is ``radically'' different from the situation for the group $\Diff_c(\mathbb{D},\omega)$ of compactly supported area-preserving diffeomorphisms of the disc with its usual topology, where the normal closure of many one-parameter subgroups is the entire group; a similar argument can be used to show that Corollary~\ref{cor:leroux} is false for the other cases of the simplicity question reviewed above. In view of our Theorem~\ref{thm:main}, we also have the following result about the failure\footnote{One should stress that a weaker version of fragmentation does hold, even in dimension $2$, for example if we remove the requirement in the definition of $P_\rho$ that the product has $N$ elements then the corresponding property holds for all $\rho$ by \cite{fathi}.} of fragmentation in dimension $2$, which follows by invoking \cite[Thm.\ 0.2]{LeRoux10}. \begin{corol} \label{cor:leroux2} None of the fragmentation properties $ P_\rho$ for $0 < \rho \le 1$ hold in $\Homeo_c(\D, \omega)$. \end{corol} This generalizes a result of Entov-Polterovich-Py \cite[Sec.\ 5.1]{EPP}, who established Corollary~\ref{cor:leroux2} for $1/2 \le \rho \le 1.$ \subsection{Idea of the proof} Our goal here is to give an outline of the proof of Theorem~\ref{thm:main}. Along the way, we state several other results of potentially independent interest, including a resolution of the ``Infinite Twist Conjecture". \subsubsection{A proper normal subgroup of $\Homeo_c(\D, \omega)$}\label{sec:prop-norm-subgr} To prove Theorem~\ref{thm:main}, we will define below a normal subgroup of $\Homeo_c(\D, \omega)$ which is a variation on the construction of Oh-M\"uller \cite{muller-oh}. We will show that this normal subgroup is proper, and in fact, contains the commutator subgroup of $\Homeo_c(\D, \omega)$. Denote by $C^{\infty}_c(\S^1 \times \D)$ the set of time-dependent functions, also referred to as {\bf Hamiltonians}, of the disc whose supports are contained in the interior of $\D$. As will be recalled in Section \ref{sec:area-pres-diffeos}, one can associate to every $H \in C^{\infty}_c(\S^1 \times \D)$ a {\bf Hamiltonian flow} $\varphi^t_H$. Furthermore, every area-preserving diffeomorphism of the disc $\phi \in \Diff_c(\D, \omega)$ is a Hamiltonian diffeomorphism in the sense that there exists $H \in C^{\infty}_c(\S^1 \times \D) $ such that $\phi = \varphi^1_H$. The energy, or the {\bf Hofer norm}, of a Hamiltonian $H \in C^{\infty}_c(\S^1 \times \D)$ is defined by the quantity \[ \| H \|_{(1, \infty)} = \int_0^1 \left( \max_{x \in \D} H(t,\cdot) - \min_{x \in \D} H(t, \cdot)\right) dt.\] \begin{definition}\label{def:finite_energy_homeo} An element $\phi \in \Homeo_c(\D, \omega)$ is a {\bf finite-energy homeomorphism} if there exists a sequence of smooth Hamiltonians $H_i \in C^{\infty}_c(\S^1 \times \D)$ such that the sequence $\| H_i \|_{(1, \infty)}$ is bounded, {\it i.e.}\ $\exists \, C \in \R$ such that $\| H_i \|_{(1, \infty)} \leq C$, and the Hamiltonian diffeomorphisms $\varphi^1_{H_i}$ converge uniformly to $\phi$. We will denote the set of all finite-energy homeomorphisms by $\FHomeo_c(\D, \omega)$. \end{definition} We clarify the topology for convergence in the definition, namely the $C^0$ topology, in Section \ref{sec:c0}. We will show in Proposition \ref{prop:FHomeo_subgp} and Corollary \ref{corol:fhomeo_commutators} that $\FHomeo_c(\D, \omega)$ is a normal subgroup of $\Homeo_c(\D, \omega)$ that contains the commutator subgroup of $\Homeo_c(\D, \omega)$. Thus, Theorem~\ref{thm:main} will follow from the following result that we show: \begin{theo}\label{theo:FHomeo_properness} $\FHomeo_c(\D, \omega)$ is a proper normal subgroup of $\Homeo_c(\D, \omega)$. \end{theo} \begin{remark}\label{rem:Hameo} In defining $\FHomeo_c(\D, \omega)$ as above, we were inspired by the article of Oh and M\"uller \cite{muller-oh}, who defined a normal subgroup of $\Homeo_c(\D, \omega),$ denoted by $\Hameo_c(\D, \omega)$, which is usually referred as the group of {\bf hameomorphisms}; see Section \ref{sec:extending_Calabi} for its definition. It has been conjectured that $\Hameo_c(\D, \omega)$ is a proper normal subgroup of $\Homeo_c(\D, \omega)$; see for example \cite[Question 4.3]{muller-oh}. It can easily be verified that $\Hameo_c(\D, \omega) \subset \FHomeo_c(\D, \omega)$. Hence, it follows from the above theorem that $\Hameo_c(\D, \omega)$ is a proper normal subgroup of $\Homeo_c(\D, \omega)$. \end{remark} \subsubsection{Infinite energy homeomorphisms} \label{sec:infinitedistance} % Theorem \ref{theo:FHomeo_properness} gives an affirmative answer to \cite[Question 1]{LeRoux-6Questions} where Le Roux asks if there exist area-preserving homeomorphisms of the disc which are ``infinitely far in Hofer's distance" from area-preserving diffeomorphisms. To elaborate, for any $\phi \in \Homeo_c(\D, \omega) \setminus \FHomeo_c(\D, \omega)$ and any sequence $H_i \in C^{\infty}_c(\S^1 \times \D)$ of smooth Hamiltonians such that the Hamiltonian diffeomorphisms $\varphi^1_{H_i} \to \phi$ uniformly, then, $\Vert H_i\Vert_{(1, \infty)} \to \infty$. This implies in particular that for any diffeomorphisms $\varphi_i \in \Diff_c(\D,\omega)$ converging to $\phi$, the $\varphi_i$ become arbitrarily far from the identity in the Hofer metric on $\Diff_c(\D,\omega)$, hence Le Roux's formulation above; we will review the Hofer metric in \ref{sec:C0_continuity}. We remark that we can also regard such a $\phi$ as having ``infinite Hofer energy". Prior to our work, it was not known whether or not such infinite energy maps existed. % In the next section, we will see explicit examples of $\phi$ that we will show are in $\Homeo_c(\D, \omega) \setminus \FHomeo_c(\D,\omega)$. \subsubsection{The Calabi invariant and the infinite twist}\label{sec:calabi_inf_twist} The hard part of Theorem~\ref{theo:FHomeo_properness} is to show properness. Here we describe the key example of an area-preserving homeomorphism that we later show is not in $\FHomeo_c(\D, \omega)$. We first summarize some background that will motivate what follows. As mentioned above, for smooth, area-preserving compactly supported two-disc diffeomorphisms, non-simplicity is known, via the Calabi invariant. More precisely, the {\bf Calabi invariant} of $\theta \in \Diff_c(\D, \omega)$ is defined as follows. Pick any Hamiltonian $H \in C^{\infty}_c(\S^1 \times \D)$ such that $\theta = \varphi^1_H$. Then, $$\Cal(\theta) := \int_{\S^1}\int_{\D} H \, \omega \; dt.$$ It is well-known that the above integral does not depend on the choice of $H$ and so $\Cal(\theta)$ is well-defined; it is also known that $\Cal: \Diff_c(\D, \omega) \rightarrow \R$ is a non-trivial group homomorphism, {\it i.e.} $ \Cal(\theta_1 \theta_2) = \Cal(\theta_1) + \Cal(\theta_2)$. For further details on the Calabi homomorphism see \cite{calabi,mcduff-salamon}. We will need to know the value of the Calabi invariant for the following class of area-preserving diffeomorphisms. Let $f :[0, 1] \to \R$ be a smooth function vanishing near $1$ and define $\phi_f \in \Diff_c(\D, \omega)$ by $\phi_f(0) := 0$ and $ \phi_f(r, \theta) := (r, \theta + 2 \pi f(r)).$ If the function $f$ is taken to be (positive/negative) monotone, then the map $\phi_f$ is referred to as a \textbf{ (positive/negative) monotone twist}. Since we will be working exclusively with positive monotone twists, we will assume monotone twists are all positive, unless otherwise stated. Now suppose that $\omega = \frac{1}{2\pi} rdr \wedge d\theta$. A simple computation (see our conventions in Section \ref{sec:prelim_symp}) shows that $\phi_f$ is the time--1 map of the flow of the Hamiltonian defined by \begin{equation}\label{eq:hamiltonian_radial_twist} F(r, \theta) = \int_r^1 s f(s) ds. \end{equation} From this we compute: \begin{equation}\label{eq:calabi_radial_twist} \Cal(\phi_f) = \int_0^1 \int_r^1 sf(s) ds \; rdr. \end{equation} We can now introduce the element that will not be in $\FHomeo_c(\D, \omega)$. Let $f:(0,1] \rightarrow \R$ be a smooth function which vanishes near $1$, is decreasing, and satisfies $\displaystyle \lim_{r \to 0} f(r) = \infty$. Define $\phi_f \in \Homeo_c(\D, \omega)$ by $\phi(0) := 0$ and \begin{equation}\label{eq:inf_twist} \phi_f(r, \theta) := (r, \theta + 2 \pi f(r)). \end{equation} It is not difficult to see that $\phi_f$ is indeed an element of $\Homeo_c(\D, \omega)$ which is in fact smooth away from the origin. We will refer to $\phi_f$ as an {\bf infinite twist.} We will show that if \begin{equation} \label{eq:infinite_calabi} \int_0^1 \int_r^1 sf(s) ds \; r dr = \infty, \end{equation} then \begin{equation}\label{eqn:inftwist_notin_FHomeo} \phi_f \notin \FHomeo_c(\D, \omega). \end{equation} \subsubsection{The infinite twist conjecture} The idea outlined in the above section is inspired by Fathi's suggestion that the Calabi homomorphism could extend to the subgroup $\Hameo_c(\D, \omega)$, the normal subgroup constructed by Oh and M\"uller mentioned in Remark~\ref{rem:Hameo}; see Conjecture 6.8 in Oh's article \cite{Oh10}; we discuss this further in Section \ref{sec:extending_Calabi}. Moreover, \cite[Thm.\ 7.2]{Oh10} shows that if the Calabi invariant extended to $\Hameo_c(\D, \omega)$, then any infinite twist $\phi_f$, satisfying \eqref{eq:infinite_calabi}, would not be in $ \Hameo_c(\D, \omega)$. Hence, it seemed reasonable to conjecture % that $\phi_f \notin \Hameo_c(\D, \omega)$ if \eqref{eq:infinite_calabi} holds. Indeed, McDuff and Salamon refer to this\footnote{The actual formulation in \cite{mcduff-salamon} is slightly different than this, because it does not include the condition \eqref{eq:infinite_calabi}. However, without this condition, one can produce infinite twists that lie in $\Hameo_c(\D, \omega)$, by the following argument. Pick a function $f$ with $\lim_{r\to 0}f(r)=\infty$ such that the function $F$ from (\ref{eq:hamiltonian_radial_twist}) is bounded. For such a function $f$, the map $\phi_f$ satisfies the assumptions given in \cite{mcduff-salamon}, but it can be approximated by smooth monotone twist maps in a standard way, showing that $\phi_f$ actually belongs to $\Hameo_c(\D,\omega)$. The authors of \cite{mcduff-salamon} have confirmed in private communication with us that imposing condition \eqref{eq:infinite_calabi} is consistent with what they intended.} as the {\bf Infinite Twist Conjecture} and it is Problem 43 on their list of open problems; see \cite[Section 14.7]{mcduff-salamon}. Since, as stated in Remark~\ref{rem:Hameo}, $\Hameo_c(\D, \omega) \subset \FHomeo_c(\D, \omega)$, we obtain the following Corollary from \eqref{eqn:inftwist_notin_FHomeo}. \begin{corol}[``Infinite Twist Conjecture"] Any infinite twist $\phi_f$ satisfying \eqref{eq:infinite_calabi} % is not in $\Hameo_c(\D, \omega)$. \end{corol} Infinite twists can be defined on any symplectic manifold, and we discuss them further in \ref{sec:fhomeoq} in the context of future open questions. \subsubsection{Spectral invariants from periodic Floer homology}\label{sec:PFH_spectralinvariants_intro} To show that the infinite twist is not in $\FHomeo_c(\D, \omega)$, we use the theory of {\bf periodic Floer homology} (PFH), reviewed in Section \ref{sec:pfhs2}. PFH is a version of Floer homology for area-preserving diffeomorphisms which was introduced by Hutchings \cite{Hutchings-index, Hutchings-Sullivan-Dehntwist}. As with ordinary Floer homology, PFH can be used to define a sequence of functions $c_d : \Diff_c(\D, \omega) \rightarrow \R$, where $d\in \N$, called {\bf spectral invariants}, which satisfy various useful properties. We give the definition of $c_d$ in Section \ref{sec:PFH_spec_initial_properties}, see in particular Remark \ref{rem:PFHspec-disc}. The definition of PFH spectral invariants is due to Michael Hutchings \cite{Hutchings_unpublished}, but very few properties have been established about these. We will prove in Theorem \ref{thm:PFHspec_initial_properties} that the PFH spectral invariants satisfy the following properties: \begin{enumerate} \item Normalization: $c_d(\id) = 0$, \item Monotonicity: Suppose that $H \leq G$ where $H, G \in C^{\infty}_c(\S^1 \times \D)$. Then, $c_d(\varphi^1_H) \leq c_d(\varphi^1_G)$ for all $d \in \N$, \item Hofer Continuity: $|c_d(\varphi^1_H) - c_d(\varphi^1_G) | \leq d \Vert H-G \Vert_{(1, \infty)}$, \item Spectrality: $c_{d}(\varphi^1_H) \in \Spec_d(H)$ for any $H \in \mathcal{H}$, where $\Spec_d(H)$ is the {\bf order $d$ spectrum} of $H$ and is defined in Section \ref{sec:action_spectra}. \end{enumerate} A key property, which allows us to use the PFH spectral invariants for studying homeomorphisms (as opposed to diffeomorphisms) is the following theorem, which we prove in Section \ref{sec:C0_continuity} via the methods of continuous symplectic topology. \begin{theo}\label{theo:C0_continuity} The spectral invariant $c_d : \Diff_c(\D, \omega) \rightarrow \R$ is continuous with respect to the $C^{0}$ topology on $\Diff_c(\D, \omega)$. Furthermore, it extends continuously to $\Homeo_c(\D, \omega)$. \end{theo} Another key property is the following which was originally conjectured in greater generality by Hutchings \cite{Hutchings_unpublished}: \begin{theo}\label{theo:PFHspec_Calabi_property} The PFH spectral invariants $c_d : \Diff_c(\D, \omega) \rightarrow \R$ satisfy the {\bf Calabi property} \begin{equation} \label{eqn:calabi} \lim_{d\to \infty} \frac{c_d(\varphi)}{d} = \Cal(\varphi) \end{equation} if $\varphi$ is a monotone twist map of the disc. \end{theo} The property \eqref{eqn:calabi} can be thought of as a kind of analogue of the ``Volume Property" for ECH spectral invariants proved in \cite{CGHR}. ECH has many similarities to PFH, which is part of the motivation for conjecturing that something like \eqref{eqn:calabi} might be possible. \begin{remark}\label{rem:hutchings_conjecture} In fact, Hutchings \cite{Hutchings_unpublished} has conjectured that the Calabi property in Theorem \ref{theo:PFHspec_Calabi_property} holds more generally for all $\varphi \in \Diff_c(\D, \omega)$. The point is that we verify this conjecture for monotone twists; and, this is sufficient for our purposes. Hutchings' conjecture, if true in full generality, would have further implications for our understanding of the algebraic structure of $\Homeo_c(\D,\omega)$, which we further discuss in \ref{sec:fhomeoq}. \end{remark} \subsubsection{Proof of Theorem \ref{theo:FHomeo_properness}} \label{sec:proof_properness} We will now give the proof of Theorem \ref{theo:FHomeo_properness}, assuming the results stated above --- we will prove these later in the paper --- and the forthcoming Proposition~\ref{prop:FHomeo_subgp}, which states that $\text{FHomeo}_c$ is a normal subgroup. We begin by explaining the basic idea. As was already explained above, the challenge with our approach is to show that the infinite twist is not in $\FHomeo_c(\mathbb{D},\omega)$. Here is how we do this. Theorem \ref{theo:C0_continuity} allows to define the PFH spectral invariants for any $\psi \in \Homeo_c(\D, \omega)$. We will show, by using the Hofer Continuity property, that if $\psi$ is a finite-energy homeomorphism then the sequence of PFH spectral invariants $\{c_d(\psi)\}_{d\in \N}$ grows at most linearly. On the other hand, in the case of an infinite twist $\phi_f$, satisfying the condition in Equation \eqref{eq:infinite_calabi}, the sequence $\{c_d(\phi_f)\}_{d\in \N}$ has super-linear growth, as a consequence of the Calabi property \eqref{eqn:calabi}. From this we can conclude that $\phi_f \notin \FHomeo_c(\D, \omega)$, as desired. The details are as follows. We begin with the following lemma which tells us that for a finite-energy homeomorphism $\psi$ the sequence of PFH spectral invariants $\{c_d(\psi)\}_{d\in \N}$ grows at most linearly. \begin{lemma}\label{lem:linear_growth} Let $\psi \in \FHomeo_c(\D, \omega)$ be a finite-energy homeomorphism. Then, there exists a constant $C$, depending on $\psi$, such that $$\frac{c_d(\psi)}{d} \leq C, \, \forall d \in \N.$$ \end{lemma} \begin{proof} By definition, $\psi$ being a finite-energy homeomorphism implies that there exist smooth Hamiltonians $H_i \in C^{\infty}_c(\S^1 \times \D)$ such that the sequence $\| H_i \|_{(1, \infty)}$ is bounded, {\it i.e.}\ $\exists \, C \in \R$ such that $\| H_i \|_{(1, \infty)} \leq C$, and the Hamiltonian diffeomorphisms $\varphi^1_{H_i}$ converge uniformly to $\psi$. The Hofer continuity property and the fact that $c_d(\id) = 0$ imply that $$c_d(\varphi^1_{H_i}) \leq d \| H_i \|_{(1, \infty)} \leq d C,$$ for each $ d \in \N$. On the other hand, by Theorem \ref{theo:C0_continuity} $c_d(\psi) = \lim_{i \to \infty} c_d(\varphi^1_{H_i})$. We conclude from the above inequality that $c_d(\psi) \leq dC$ for each $ d \in \N$. \end{proof} We now turn our attention to showing that the PFH spectral invariants of an infinite twist $\phi_f$, which satisfies Equation \eqref{eq:infinite_calabi}, violate the inequality from the above lemma. We will need the following. \begin{lemma}\label{lem:approx_monotone_twists} There exists a sequence of smooth monotone twists $\phi_{f_i}\in \Diff_c(\D, \omega)$ satisfying the following properties: \begin{enumerate} \item The sequence $\phi_{f_i}$ converges in the $C^0$ topology to $\phi_f$, \item There exist Hamiltonians $F_i$, compactly supported in the interior of the disc $\D$, such that $\varphi^1_{F_i}= \phi_{f_i}$ and $F_i \leq F_{i+1}$, \item $\displaystyle \lim_{i\to \infty} \Cal(\phi_{f_i}) = \infty$. \end{enumerate} \end{lemma} \begin{proof} Recall that $f$ is a decreasing function of $r$ which vanishes near $1$ and satisfies $\displaystyle \lim_{r \to 0} f(r) = \infty$. It is not difficult to see that we can pick smooth functions $f_i: [0,1] \rightarrow \R$ satisfying the following properties: \begin{enumerate} \item $f_i = f$ on $[\frac{1}{i}, 1]$, \item $f_i \leq f_{i+1}$. \end{enumerate} Let us check that the monotone twists $\phi_{f_i}$ satisfy the requirements of the lemma. To see that they converge to $\phi_f$, observe that $\phi_f$ and $\phi_{f_i}$ coincide outside the disc of radius $\frac{1}{i}$. Hence, $\phi_f^{-1} \phi_{f_i}$ converges uniformly to $\id$ because it is supported in the disc of radius $\frac{1}{i}$. Next, note that by Formula \eqref{eq:hamiltonian_radial_twist}, $\phi_{f_i}$ is the time--1 map of the Hamiltonian flow of $ F_i(r, \theta) = \int_r^1 s f_i(s) ds$. Clearly, $F_i \leq F_{i+1}$ because $f_i \leq f_{i+1}$. Finally, by Formula \eqref{eq:calabi_radial_twist} we have $$ \Cal(\phi_{f_i}) = \int_0^1 \int_r^1 sf_i(s) ds \; rdr \geq \int_{\frac{1}{i}}^1 \int_r^1 sf_i(s) ds \; rdr = \int_{\frac{1}{i}}^1 \int_r^1 sf(s) ds \; r dr.$$ Recall that $f$ has been picked such that $\int_0^1 \int_r^1 sf(s) ds \; rdr = \infty$; see Equation \eqref{eq:infinite_calabi}. We conclude that $\displaystyle \lim_{i \to \infty} \Cal(\phi_{f_i}) = \infty$. \end{proof} We will now use Lemma \ref{lem:approx_monotone_twists} to complete the proof of Theorem \ref{theo:FHomeo_properness}. By the upcoming Proposition~\ref{prop:FHomeo_subgp}, $\text{FHomeo}_c$ is a normal subgroup; it is certainly non-trivial, so it remains to show it is proper. By the Monotonicity property, we have $c_d(\phi_{f_i}) \leq c_d(\phi_{f_{i+1}})$ for each $d\in \N$. Since $\phi_{f_i}$ converges in $C^0$ topology to $\phi_f$, we conclude from Theorem \ref{theo:C0_continuity} that $c_d(\phi_{f}) = \lim_{i \to \infty} c_d(\phi_{f_i})$. Combining the previous two lines we obtain the following inequality: $$ c_d(\phi_{f_i}) \leq c_d(\phi_{f}), \forall d,i \in \N. $$ Now the Calabi property of Theorem \ref{theo:PFHspec_Calabi_property} tells us that $\lim_{d\to \infty} \frac{c_d(\phi_{f_i})}d = \Cal(\phi_{f_i})$. Combining this with the previous inequality we get $ \Cal(\phi_{f_i}) \leq \lim_{ d\to \infty} \frac{c_d(\phi_f)}{d}$ for all $i$. Hence, by the third item in Lemma~\ref{lem:approx_monotone_twists} $$\lim_{ d\to \infty} \frac{c_d(\phi_f)}{d} = \infty,$$ and so by Lemma~\ref{lem:linear_growth} $\phi_f$ is not in $\FHomeo_c(\D, \omega)$. \begin{remark}\label{rem:only_need_inequality} The proof outlined above does not use the full force of Theorem \ref{theo:PFHspec_Calabi_property}; it only uses the fact that $ \lim_{d\to \infty} \frac{c_d(\varphi)}{d} \geq \Cal(\varphi) $. \end{remark} \subsubsection*{Organization of the paper} In Section \ref{sec:prelim_symp}, we review some of the necessary background from symplectic geometry, especially the case of surfaces. In Section \ref{sec:pfhs2} of the paper we review the construction of periodic Floer homology and the associated spectral invariants. Some of the properties of PFH spectral invariants, such as Hofer continuity and spectrality, are proven in Section \ref{sec:PFH_spec_initial_properties}. We prove Theorem \ref{theo:C0_continuity}, on $C^0$ continuity of PFH spectral invariants, in Section \ref{sec:C0_continuity}. Section \ref{sec:PFH_monotone_twist} of the article begins with a more precise version of Theorem \ref{theo:PFHspec_Calabi_property}. The rest of Section \ref{sec:PFH_monotone_twist} is dedicated to the development of a combinatorial model of the periodic Floer homology of monotone twists. In Section \ref{sec:comuting_spec_invariant}, we first use the aforementioned combinatorial model of PFH to compute the PFH spectral invariants for monotone twists. We then use this computation to prove that the PFH spectral invariants for monotone twists satisfy the Calabi property of Theorem~\ref{theo:PFHspec_Calabi_property}; this will be carried out in Section \ref{sec:proof_Calabi_property}. \subsubsection*{Acknowledgments} We are grateful to Michael Hutchings for many useful conversations, in particular for explaining the theory of PFH spectral invariants to us and for explaining his conjecture on the Calabi property of PFH spectral invariants. We warmly thank Lev Buhovsky for helpful conversations which helped us realize that the Calabi property of PFH spectral invariants could not possibly hold for spectral invariants satisfying the ``standard" spectrality axiom on the disc. We are grateful to Fr\'ed\'eric Le Roux for many helpful discussions on fragmentations. We thank Albert Fathi for sharing a copy of his unpublished manuscript \cite{Fathi-Calabi} with us and for helpful comments. We also thank Abed Bounemoura, Igor Frenkel, \'Etienne Ghys, Helmut Hofer, Dusa McDuff, Yong-Geun Oh, Dietmar Salamon, Felix Schlenck, Takashi Tsuboi, and Chris Wendl for helpful communications. DCG is partially supported by NSF grant DMS 1711976 and the Minerva Research Foundation. DCG is extremely grateful to the Institute for Advanced Study for providing a wonderful atmosphere in which much of this research was completed. The research leading to this article began in June, 2018 when DCG was a ``FSMP Distinguished Professor" at the Institut Math\'ematiques de Jussieu-Paris Rive Gauche (IMJ-PRG). DCG is grateful to the Fondation Sciences Math\'ematiques de Paris (FSMP) and IMJ-PRG for their support and hospitality. DCG also thanks Kei Irie, Peter Kronheimer, and Dan Pomerleano for helpful discussions. VH thanks Fr\'ed\'eric Le Roux for hours of discussions related to the problem studied here, which started in 2006. VH is partially supported by the ANR project ``Microlocal'' ANR-15-CE40-0007. SS: I am grateful to Lev Buhovsky, Fr\'ed\'eric Le Roux, Yong-Geun Oh and Claude Viterbo from whom I have learned a great deal about the various aspects of the simplicity question. This material is partially based upon work supported by the NSF under Grant No. DMS-1440140 while I was in residence at the MSRI in Berkeley during the Fall, 2018 semester; this support at MSRI also allowed for valuable in-person discussions with DCG. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (grant agreement No. 851701). \section{Preliminaries about the symplectic geometry of surfaces}\label{sec:prelim_symp} Here we collect some basic facts, and fix notation, concerning two-dimensional symplectic geometry and diffeomorphism groups. \subsection{Symplectic form on the disc and sphere} \label{sec:sphere} Let $\S^2 := \{(x,y, z) \in \R^3 : x^2 + y^2 + z^2 = 1\} \subset \R^3$ and $\D:= \{(x,y)\in \R^2: x^2 + y^2 \leq 1\}$. We equip the sphere $\S^2$ with the symplectic form $\omega:= \frac{1}{4 \pi}d\theta \wedge dz$, where $(\theta,z)$ are cylindrical coordinates on $\mathbb{R}^3$. Note that with this form, $\S^2$ has area $1.$ Let \[ S^+ = \{ (x,y,z) \in \S^2 : z \geq 0 \},\] be the northern hemisphere in $\S^2$. In certain sections of the paper, we will need to identify the disc $\D$ with $S^+$. To do this, we will take the embedding $\iota: \D \to \S^2$ given by the formula \begin{equation}\label{eq:identify_disc_north_hemisphere} \iota(r, \theta) = (\theta,1-r^2), \end{equation} where $(r, \theta)$ denotes the standard polar coordinates on $\R^2$. We will equip the disc with the area form given by the pullback of $\omega$ under $\iota$; explicitly, this is given by the formula $\frac{1}{2\pi} rdr\wedge d\theta$. We will denote this form by $\omega$ as well. Note that this gives the disc a total area of $\frac{1}{2}$. Any area form on $\S^2$ or $\D$ is equivalent to the above differential forms, up to multiplication by a constant. \subsection{The $C^0$ topology} \label{sec:c0} Here we fix our conventions and notation concerning the $C^0$ topology. Denote by $\Homeo(\S^2)$ the group of homeomorphisms of the sphere and by $\Homeo_c(\D)$ the group of homeomorphisms of the disc whose support is contained in the interior of $\D$. Let $d$ be a Riemannian distance on $\S^2$. Given two maps $\phi, \psi :\S^2 \to \S^2,$ we denote $$d_{C^0}(\phi,\psi)= \max_{x\in \S^2}d(\phi(x),\psi(x)).$$ We will say that a sequence of maps $\phi_i : \S^2 \rightarrow \S^2$ {\bf converges uniformly}, or $C^0$--converges, to $\phi$, if $d_{C^0}(\phi_i, \phi) \to 0$ as $ i \to \infty$. As is well known, the notion of $C^0$--convergence does not depend on the choice of the Riemannian metric. The topology induced by $d_{C^0}$ on $\Homeo(\S^2)$ is referred to as the {\bf $C^0$ topology}. The $C^0$ topology on $\Homeo_c(\D)$ is defined analogously as the topology induced by the distance $$d_{C^0}(\phi,\psi)= \max_{x\in \D}d(\phi(x),\psi(x)).$$ \subsection{Hamiltonian diffeomorphisms} \label{sec:area-pres-diffeos} Let \[ \Diff(\S^2, \omega):=\{\theta \in \Diff(\S^2): \theta^*\omega =\omega\}\] denote the group of area-preserving, in other words symplectic, diffeomorphisms of the sphere. Let $C^{\infty}(\S^1 \times \S^2)$ denote the set of smooth time-dependent Hamiltonians on $\S^2$. As alluded to in the introduction, a smooth Hamiltonian $H \in C^{\infty} (\S^1 \times \S^2)$ gives rise to a time-dependent vector field $X_H$, called the {\bf Hamiltonian vector field}, defined via the equation \[ \omega(X_{H_t}, \cdot) = dH_t.\] The {\bf Hamiltonian flow} of $H$, denoted by $\varphi^t_H$, is by definition the flow of $X_H$. A {\bf Hamiltonian diffeomorphism} is a diffeomorphism which arises as the time-one map of a Hamiltonian flow. It is easy to verify that every Hamiltonian diffeomorphism of $\S^2$ is area-preserving. And, as is well-known, every area-preserving diffeomorphism of the sphere is in fact a Hamiltonian diffeomorphism. As for the disc, as mentioned in the introduction, every $\theta \in \Diff_c(\D, \omega)$ is Hamiltonian, in the sense that one can find $H \in C^{\infty}_c(\S^1 \times \D)$ such that $\theta = \varphi^1_H$, where the notation is as in the sphere case. Here, $C^{\infty}_c(\S^1 \times \D)$ denotes the set of Hamiltonians of $\D$ whose support is compactly contained in $\D$. Note that $\Diff(\S^2, \omega) \subset \Homeo_0(\S^2, \omega)$ and $\Diff_c(\D, \omega) \subset \Homeo_c(\D, \omega)$. It is well-known that $\Diff(\S^2, \omega)$ and $\Diff_c(\D, \omega)$ are dense, with respect to the $C^0$ topology, in $\Homeo_0(\S^2, \omega)$ and $\Homeo_c(\D, \omega)$, respectively. \paragraph{Hamiltonians for monotone twists.} Recall the definition of monotone twists from Section \ref{sec:calabi_inf_twist}. Every monotone twist $\varphi$, being an element of $\Diff_c(\D, \omega)$, is a Hamiltonian diffeomorphism of the disc and so can be written as the time--$1$ map $\varphi^1_H$ of the Hamiltonian flow of some $H \in C^\infty_c(\S^1 \times \D)$. The Hamiltonian $H$ can be picked to be of the form \[ H = \frac{1}{2} h(z),\] where $h: \S^2 \rightarrow \R$ is a function of $z$ satisfying \[ h' \geq 0, h'' \geq 0, h(-1) = 0, h'(-1) = 0.\] \subsection{Rotation numbers}\label{sec:rotation_number} Let $p$ be a fixed point of $\varphi \in \Diff(\S^2, \omega)$. One can find a Hamiltonian isotopy $\{\varphi^t\}_{t \in [0,1]}$ such that $\varphi^0 = \id, \varphi^1 = \varphi$ and $\varphi^t(p) = p$ for all $t \in [0, 1]$. In this section, we briefly review the definitions and some of the properties of $\rot(\{\varphi^t\}, p)$, the rotation number of the isotopy $\{\varphi^t\}_{t \in [0,1]}$ at $p$, and $\rot(\varphi, p),$ the rotation number of $\varphi$ at $p$. Our conventions are such that $\rot(\varphi, p)$ will be a real number in the interval $(-\frac12, \frac12]$. For further details on this subject, we refer the reader to \cite{katok-hasselblatt}. The derivative of the isotopy $D_p\varphi^t: T_p \S^2 \rightarrow T_p\S^2$, viewed as a linear isotopy of $\R^2$, induces an isotopy of the circle $\{f_t\}_{t\in[0,1]}$ with $f_0 = \id$ and $f_1 = D_p \varphi$. We define $\rot(\{\varphi^t \}, p)$, the {\bf rotation number of the isotopy $(\varphi^t)_{t \in [0,1]}$ at $p$}, to be the Poincar\'e rotation number of the circle isotopy $\{f_t\}$. The number $\rot(\{\varphi^t \}, p)$ depends only on $\{D_p \varphi^t\}_{t\in [0,1]}$ and it satisfies the following properties. Let $\{\psi^t\}_{t\in[0,1]}$ be another Hamiltonian isotopy such that $\psi^0 = \id, \psi^1 = \varphi^1$, and $\psi^t(p) = p$ for $t\in[0,1]$. Then, \begin{enumerate} \item $\rot(\{\varphi^t \}, p) - \rot(\{\psi^t \}, p) \in \Z$, \item $\rot(\{\varphi^t \}, p) = \rot(\{\psi^t \}, p)$ if the two isotopies are homotopic rel. endpoints among isotopies fixing the point $p$. \end{enumerate} The above facts may be deduced from the standard properties of the Poincar\'e rotation number; see, for example, \cite{katok-hasselblatt}. Lastly, we define $\rot(\varphi, p)$, the {\bf rotation number of $\varphi$ at $p$}, to be the unique number in $(-\frac12, \frac12]$ which coincides with $\rot(\{\varphi^t\}, p)$ mod $\Z$. It follows from the discussion in the previous paragraph that it depends only on $D_p\varphi$. \medskip \subsection{The action functional and its spectrum}\label{sec:action_spectra} Spectral invariants take values in the ``action spectrum". We now explain what this spectrum is. Denote by $\Omega:= \{z: \S^1 \rightarrow \S^2 \}$ the space of all loops in $\S^2$. By a {\bf capping} of a loop $z: \S^1 \rightarrow \S^2,$ we mean a map \[ u : D^2 \rightarrow \S^2,\] such that $u|_{\partial D^2} = z$. We say two cappings $u, u'$ for a loop $z$ are {\bf equivalent} if $u, u'$ are homotopic rel $z$. Henceforth, we will only consider cappings up to this equivalence relation. Note that given a capping $u$ of a loop $z$, all other cappings of $z$ are of the form $u\# A$ where $A \in \pi_2(\S^2)$ and $\# $ denotes the operation of connected sum. A {\bf capped loop} is a pair $(z,u)$ where $z$ is a loop and $u$ is a capping for $z$. We will denote by $\tilde{\Omega}$ the space of all capped loops in the sphere. Let $H \in C^{\infty} (\S^1 \times \S^2)$ denote a smooth Hamiltonian in $\S^2$. Recall that $\mathcal{A}_H: \tilde{\Omega} \rightarrow \mathbb{R}$, the {\bf action functional} associated to the Hamiltonian $H$ is defined by \begin{equation} \label{eqn:actiondefn} \mathcal{A}_H(z,u) = \int_{0}^{1} H(t,z(t))dt \text{ } + \int_{D^2} u^*\omega. \end{equation} Note that $\mathcal{A}_H(z,u \# A ) = \mathcal{A}_H(z,u ) + \omega(A),$ for every $A \in \pi_2(\S^2)$. The set of critical points of $\mathcal{A}_H$, denoted by $\Crit(\mathcal{A}_H)$, consists of capped loops $(z,u) \in \tilde{\Omega}$ such that $z$ is a $1$--periodic orbit of the Hamiltonian flow $\varphi^t_H$. We will often refer to such $(z,u)$ as a capped $1$--periodic orbit of $\varphi^t_H$. The {\bf action spectrum} of $H$, denoted by $\Spec(H)$, is the set of critical values of $\mathcal{A}_H$; it has Lebesgue measure zero. It turns out that the action spectrum $\Spec(H)$ is independent of $H$ in the following sense: If $H'$ is another Hamiltonian such that $\varphi^1_H = \varphi^1_{H'}$, then there exists a constant $C \in \R$ such that $$ \Spec(H) = \Spec(H') + C,$$ where $\Spec(H') + C$ is the set obtained from $\Spec(H')$ by adding the value $C$ to every element of $\Spec(H')$. \cite[Lemma 3.3]{Schwarz} proves this in the case where $\omega$ vanishes on $\pi_2(M)$ and the proof generalizes readily to general symplectic manifolds. Moreover, it follows from the proof of \cite[Lemma 3.3]{Schwarz} that if $H, H'$ are supported in the northern hemisphere $ S^+ \subset \S^2$, then the above constant $C$ is zero and hence \begin{equation}\label{eqn:lemma3.3_schwarz} \Spec(H) = \Spec(H'). \end{equation} The PFH spectral invariants will take values in a more general set, which we call the {\bf higher order action spectrum} To define it, let $H, G$ be two Hamiltonians. The {\bf composition} of $H$ and $G$ is the Hamiltonian $H\#G(t,x) : = H(t,x) + G(t, (\phi^t_H)^{-1}(x))$. It is known that $\phi^t_{H\#G} = \phi^t_H \circ \phi^t_G$; see \cite[Sec. 5.1, Prop. 1]{hofer-zehnder}, for example. Denote by $H^k$ the $k$-times composition of $H$ with itself. For any $d>0$, we now define the {\bf order $d$ spectrum} of $H$ by $$\displaystyle \Spec_d(H) := \cup_{k_1+ \ldots + k_j = d} \;\; \Spec(H^{k_1}) + \ldots + \Spec(H^{k_j}).$$ Note that $\displaystyle \Spec_d(H)$ may equivalently be described as follows: For every value $a \in \Spec_d(H)$ there exist capped periodic orbits $(z_1, u_1), \ldots, (z_k, u_k)$ of $H$ the sum of whose periods is $d$ and such that \[ a = \sum \mathcal{A}_{H^{k_i}}(u_i, z_i). \] We can use the above to define the action spectrum for compactly supported disc maps. Recall from Section \ref{sec:sphere} our convention to identify the northern hemisphere of $\mathbb{S}^2$ with the disc; we will use this to define the action spectrum in the case of the disc. More precisely, if $H, H'$ are supported in the northern hemisphere $ S^+ \subset \S^2$, and generate the same time-1 map $\phi$, then we in fact have $\Spec_d(H) = \Spec_d(H')$ for all $d>0$. Indeed, as an immediate consequence of Equation \eqref{eqn:lemma3.3_schwarz} we have $\Spec(H^k) = \Spec(H'^k)$ for all $k \in \N$ and so it follows from the definition that $\Spec_d(H) = \Spec_d(H')$ for all $d>0$. Hence, if $\phi \in \DiffS$, then we can define the action spectra of $\phi$ without any ambiguity by setting \begin{equation} \label{eqn:spec_time1} \Spec_d(\phi) = \Spec_d(H), \end{equation} where $H$ is any Hamiltonian in $C^{\infty}_c(\S^1 \times S^+)$ such that $\phi = \varphi^1_H$. \subsection{Finite energy homeomorphisms} Recall that we defined finite-energy homeomorphisms in Definition \ref{def:finite_energy_homeo}. It is not hard to show that $\FHomeo_c(\D, \omega)$ is a normal subgroup of $\Homeo_c(\D, \omega)$ and $\FHomeo_c(\D, \omega)$ contains the commutator subgroup of $\Homeo_c(\D, \omega)$. In this section, we show that $\FHomeo_c(\D, \omega)$ is a normal subgroup. \begin{prop}\label{prop:FHomeo_subgp} $\FHomeo_c(\D, \omega)$ is a normal subgroup of $\Homeo_c(\D, \omega)$. \end{prop} \begin{proof} Consider smooth Hamiltonians $H, G \in C^{\infty}_c(\S^1 \times \D)$. As was partly mentioned in Section \ref{sec:action_spectra}, it is well-known (and proved for example in \cite[Sec. 5.1, Prop. 1]{hofer-zehnder}) % that the Hamiltonians \begin{equation} \label{eqn:somehams} H\#G(t, x) := H(t,x) + G(t, (\varphi^t_H)^{-1}(x)), \quad \bar{H}(t,x) := - H(t, \varphi^t_H(x)), \end{equation} generate the Hamiltonian flows $ \varphi^t_H \phi^t_G$ and $(\varphi^t_H)^{-1}$ respectively. Furthermore, given $\psi \in \Diff_c(\D, \omega)$, the Hamiltonian \[ H\circ \psi(t,x):= H(t, \psi(x))\] generates the flow $\psi^{-1} \varphi^t_H \psi$. We now show that $\FHomeo_c$ is closed under conjugation. Take $\phi \in \FHomeo_c(\D, \omega)$ and let $H_i$ and $C$ be as in Definition \ref{def:finite_energy_homeo}. % Let $\psi \in \Homeo_c(\D, \omega)$ and take a sequence $\psi_i \in \Diff_c(\D, \omega)$ which converges uniformly to $\psi$. Consider the Hamiltonians $H_i \circ \psi_i$. The corresponding Hamiltonian diffeomorphisms are the conjugations $\psi_i^{-1} \varphi^1_{H_i} \psi_i$ which converge uniformly to $\psi^{-1} \phi \psi$. Furthermore, \[\| H_i \circ \psi_i \|_{(1, \infty)} = \| H_i \|_{(1, \infty)} \leq C,\] where the inequality follows from the definition of $\FHomeo_c(\D, \omega)$. We will next check that $\FHomeo_c$ is a group. Take $\phi, \psi \in \FHomeo_c$ and let $H_i, G_i \in C^{\infty}_c(\S^1 \times \D)$ be two sequences of Hamiltonians such that $\varphi^1_{H_i}, \varphi^1_{G_i}$ converge uniformly to $\phi, \psi$, respectively, and $ \| H_i \|_{(1, \infty)}, \| G_i \|_{(1, \infty)} \leq C $ for some constant $C$. Then, the sequence $\varphi^{-1}_{H_i} \circ \varphi^1_{G_i}$ converges uniformly to $\phi^{-1} \circ \psi$. Moreover, by the above formulas, we have $\varphi^{-1}_{H_i} \circ \varphi^1_{G_i} = \varphi^1_{\overline{H}_i \#G_i}$. Since $ \| \overline H_i \# G_i \|_{(1, \infty)} \leq \| H_i \|_{(1, \infty)} + \| G_i \|_{(1, \infty)} \leq 2C,$ this proves that $\phi^{-1} \circ \psi \in \FHomeo_c$ which completes the proof that $\FHomeo_c$ is a group. \end{proof} \subsection{Equivalence of perfectness and simplicity} The goal of this section is to show that in the case of $\Homeo_c(\D, \omega)$ perfectness and simplicity are equivalent. \begin{prop}\label{prop:commutators} Any non-trivial normal subgroup $H$ of $\Homeo_c(\D, \omega)$ contains the commutator subgroup of $\Homeo_c(\D, \omega)$. Hence, $\Homeo_c(\D, \omega)$ is perfect if and only if it is simple. \end{prop} Before proving the above proposition, we state two of its corollaries. \begin{corol}\label{corol:fhomeo_commutators}The commutator subgroup of $\Homeo_c(\D, \omega)$ is contained in $\FHomeo_c(\D, \omega)$. \end{corol} \begin{remark}\label{rem:inf_twist_not_commutator} As a consequence, proving that the infinite twist $\phi_f$, introduced in Section \ref{sec:calabi_inf_twist}, is not in $\FHomeo_c(\D, \omega)$ proves also that $\phi_f$ cannot be written as a product of commutators. \end{remark} As promised in the introduction, we prove in the next corollary that $\Homeo_c(\D, \omega)$ admits no non-trivial continuous homomorphisms. This fact seems to be well-known to the experts, however, we do not know of a published reference for it. \begin{corol}\label{corol:no_continuous_homomorphism} The group $\Homeo_c(\D, \omega)$ admits no non-trivial homomorphism which is continuous with respect to the $C^0$ topology. \end{corol} \begin{proof} Let $H$ be a non-trivial normal subgroup of $\Homeo_c(\D, \omega)$. We will show that $H$ is dense with respect to the $C^0$ topology; this proves the corollary because the kernel of a continuous homomorphism is closed. By Proposition \ref{prop:commutators}, we know that $H$ contains the commutator subgroup of $\Diff_c(\D, \omega)$. Consequently, $H$ contains the kernel of the Calabi homomorphism as the commutator subgroup of $\Diff_c(\D, \omega)$ coincides with the kernel of the Calabi invariant \cite{Banyaga}. We claim that the kernel of the Calabi invariant is dense in $\Diff_c(\D, \omega)$; hence, it is dense in $\Homeo_c(\D, \omega)$. Indeed, take any $\psi \in \Diff_c(\D, \omega)$ and denote $a= \Cal(\psi)$. Pick Hamiltonians $H_n$ such that \begin{itemize} \item $H_n$ is supported in a disc of diameter $\frac{1}{n}$, \item $\int_\D H_n = -a$. Thus, $\Cal(\varphi^1_{H_n}) = -a.$ \end{itemize} Then, $\Cal(\varphi^1_{H_n} \circ \psi) = 0$ and $\varphi^1_{H_n} \circ \psi \xrightarrow{C^0} \psi$. \end{proof} The proof of Proposition \ref{prop:commutators} relies on a general argument, due to Epstein \cite{Epstein} and Higman \cite{Higman}, which essentially shows that perfectness implies simplicity for transformation groups satisfying certain assumptions. We will present a version of this argument, which we learned in \cite{fathi}, in our context. \begin{proof}[Proof of Proposition \ref{prop:commutators}] Pick $h \in H$ such that $ h \neq \mathrm{Id}$. We can find a closed {\it topological disc}, that is a set which is homeomorphic to a standard Euclidean disc $\D' \subset \D$ such that $h(\D') \cap \D' = \emptyset$. Denote by $\Homeo_c(\D', \omega)$ the subset of $ \Homeo_c(\D, \omega)$ consisting of area-preserving homeomorphisms whose supports are contained in the interior of $\D'$. We will first prove the following lemma. \begin{lemma} \label{lem:comh} The commutator subgroup of $\Homeo_c(\D', \omega)$ is contained in $H$. \end{lemma} \begin{proof} We must show that for any $f, g \in \Homeo_c(\D', \omega)$, the commutator $[ f,g] := fgf^{-1} g^{-1}$ is an element of $H$. First, observe that for any $f \in \Homeo_c(\D', \omega)$ we have \begin{equation}\label{eq:commutator1} [f, r] \in H \end{equation} for any $r \in H$. Indeed, by normality, $ frf^{-1} \in H$ and hence $ frf^{-1} r^{-1} \in H$. Next, we claim that for any $f,g \in \Homeo_c(\D', \omega)$ \begin{equation}\label{eq:commutator2} [ f,g] [g, hfh^{-1}] = f [g, [f^{-1}, h] \, ] f^{-1}. \end{equation} Postponing the proof of this identity for the moment, we will first show that it implies the lemma. Note that $g$ and $hfh^{-1}$ are, respectively, supported in $\D'$ and $h(\D')$ which are disjoint. Thus, $[g, hfh^{-1}] = \id$. Hence, Identity \eqref{eq:commutator2} yields $[ f,g] = f [g, [f^{-1}, h] \, ] f^{-1}$. Now, \eqref{eq:commutator1} implies that $ [g, [f^{-1}, h] \, ] \in H$ which, by normality of $H$, implies that $f [g, [f^{-1}, h] \, ] f^{-1} \in H$. This gives us the conclusion of the lemma. We complete the proof by proving Identity \eqref{eq:commutator2}: \begin{align*} & [ f,g] [g, hfh^{-1}] = ( fgf^{-1}g^{-1}) \, (ghfh^{-1}g^{-1}hf^{-1}h^{-1}) \\ = & fg ( f^{-1}hfh^{-1}) g^{-1}hf^{-1}h^{-1} = fg [f^{-1},h] g^{-1}hf^{-1}h^{-1} \\ = & fg [f^{-1},h] g^{-1}hf^{-1}h^{-1} f f^{-1} = fg [f^{-1},h] g^{-1} [h, f^{-1}] f^{-1} \\ = & f[g, [f^{-1}, h] \, ] f^{-1}. \end{align*} \end{proof} We continue with the proof of Proposition \ref{prop:commutators}. Fix a small $\varepsilon >0$ and let $\mathcal{E}$ be the set consisting of all $g \in \Homeo_c(\D, \omega)$ whose supports are contained in some topological disc of area $\varepsilon$. It is a well-known fact that the set $\mathcal{E}$ generates the group $\Homeo_c(\D , \omega)$. This is usually referred to as the \emph{fragmentation property} and it was proven by Fathi; see Theorems 6.6, A.6.2, and A.6.5 in \cite{fathi}. We claim that $[f, g] \in H$ for any $f,g \in \mathcal{E}$. Indeed, assuming $\varepsilon$ is small enough, we can find a topological disc $U$ which contains the supports of $f$ and $g$ and whose area is less than the area of $\D'$. There exists $r \in \Homeo_c(\D , \omega)$ such that $r(U) \subset \D'$. As a consequence, $rfr^{-1}, r g r^{-1}$ are both supported in $\D'$ and hence, by Lemma~\ref{lem:comh}, $[rfr^{-1}, r g r^{-1}] \in H$. Since $H$ is a normal subgroup of $ \Homeo_c(\D , \omega)$, and $[rfr^{-1}, r g r^{-1}] = r [f,g] r^{-1}$, we conclude that $[f, g] \in H$. Now, the set $\mathcal{E}$ generates $\Homeo_c(\D , \omega)$ and $[f,g] \in H$ for any $f,g \in \mathcal{E}$. Hence, the quotient group $\Homeo_c(\D , \omega)/ H$ is abelian. Thus, $H$ contains the commutator subgroup of $\Homeo_c(\D , \omega)$. \end{proof} \section{Periodic Floer Homology and basic properties of the PFH spectral invariants} \label{sec:pfhs2} In this section, we recall the definition of Periodic Floer Homology (PFH), due to Hutchings \cite{Hutchings-index, Hutchings-Sullivan-Dehntwist}, and the construction of the spectral invariants which arise from this theory, also due to Hutchings \cite{Hutchings_unpublished}. We will then prove that PFH spectral invariants satisfy the Monotonicity, Hofer Continuity, and Spectrality properties which we mentioned in the introduction. The spectral invariants appearing in Section \ref{sec:PFH_spectralinvariants_intro} are defined by identifying area-preserving maps of the disc, $\Diff_c(\D, \omega)$, with area-preserving maps of the sphere, which are supported in the northern hemisphere $S^+$, and using the PFH of $\S^2$. Thus, the three aforementioned properties will follow from related properties about PFH spectral invariants on $\S^2$; see Theorem \ref{thm:PFHspec_initial_properties} below. \subsection{Preliminaries on $J$--holomorphic curves and stable Hamiltonian structures}\label{sec:prelim_Jcurves} A stable Hamiltonian structure (SHS) on a closed three-manifold $Y$ is a pair $(\alpha, \Omega)$, consisting of a $1$--form $\alpha$ and a closed two-form $\Omega$, such that \begin{enumerate} \item $\alpha \wedge \Omega $ is a volume form on $Y$, \item $\ker(\Omega) \subset \ker(d\alpha)$. \end{enumerate} Observe that the first condition implies that $\Omega$ is non-vanishing, and as a consequence, the second condition is equivalent to $d\alpha = g\Omega$, where $g:Y \rightarrow \R$ is a smooth function. A stable Hamiltonian structure determines a plane field $\xi : = \ker(\alpha)$ and a {\bf Reeb} vector field $R$ on $Y$ given by $$R \in \ker(\Omega), \; \alpha(R) =1.$$ Closed integrals curves of $R$ are called {\bf Reeb orbits}; we regard Reeb orbits as equivalent if they are equivalent as currents. Stable Hamiltonian structures were introduced in \cite{BEHWZ, CiMo} as a setting in which one can obtain general Gromov-type compactness results, such as the SFT compactness theorem, for pseudo-holomorphic curves in $\R \times Y$. Here are two examples of stable Hamiltonian structures which are relevant to our story. \begin{example}\label{ex:contact_structure} A {\bf contact} form on $Y$ is a $1$-form $\lambda$ such that $\lambda \wedge d\lambda$ is a volume form. The pair $(\alpha, \Omega) := (\lambda, d \lambda)$ gives a stable Hamiltonian structure with $g \equiv 1$. The plane field $\xi$ is the associated contact structure and the Reeb vector field as defined above gives the usual Reeb vector field of a contact form. The {\bf contact symplectization} of $Y$ is \[ X := \mathbb{R} \times Y_\varphi , \] which has a standard symplectic form, defined by \begin{equation} \label{eqn:symp_form_symplectization} \Gamma = d(e^s \lambda), \end{equation} where $s$ denotes the coordinate on $\R$. \end{example} \begin{example}\label{ex:mapping_torus} Let $(S,\omega_S)$ be a closed surface and denote by $\varphi$ a smooth area-preserving diffeomorphism of $S$. Define the mapping torus $$Y_\varphi := \frac{S \times [0,1]}{(x,1) \sim (\varphi(x), 0) }.$$ Let $r$ be the coordinate on $[0,1]$. Now, $Y_\varphi$ carries a stable Hamiltonian structure $(\alpha, \Omega) := (dr, \omega_\varphi)$, where $\omega_\varphi$ is the canonical closed two form on $Y_\varphi$ induced by $\omega_S$. Note that the plane field $\xi$ is given by the vertical tangent space of the fibration $\pi: Y_\varphi \rightarrow \S^1$ and the Reeb vector field is given by $R =\partial_r$. Here, $g \equiv 0$. Observe that the Reeb orbits here are in correspondence with the periodic orbits % of $\varphi$. We define the {\bf symplectization} of $Y_\varphi$ \[ X := \mathbb{R} \times Y_\varphi ,\] which has a standard symplectic form, defined by \begin{equation} \label{eqn:standard2} \Gamma = ds \wedge dr + \omega_{\varphi}, \end{equation} where $s$ denotes the coordinate on $\R$. \end{example} We say an almost complex structure $J$ on $X = \R \times Y$ is {\bf admissible}, for a given SHS $(\alpha, \Omega)$, if the following conditions are satisfied: \begin{enumerate} \item $J$ is invariant under translation in the $\R$-direction of $\R\times Y$, \item $J \partial_s = R$, where $s$ denotes the coordinate on the $\R$-factor of $\R \times Y$, \item $ J \xi = \xi$, where $\xi := \ker(\alpha)$, and $\Omega(v, Jv) > 0$ for all nonzero $v \in \xi$. \end{enumerate} We will denote by $\mathcal{J}(\alpha, \Omega)$ the set of almost complex structures which are admissible for $(\alpha, \Omega)$. The space $\mathcal{J}(\alpha, \Omega)$ equipped with the $C^{\infty}$ topology is path connected, and even contractible. Define a {\bf $J$-holomorphic map} to be a smooth map \[ u: (\Sigma,j) \to (X,J),\] satisfying the equation \begin{equation} \label{eqn:jcurveequation} du \circ j = J \circ du, \end{equation} where $(\Sigma,j)$ is a closed Riemann surface (possibly disconnected), minus a finite number of punctures. As is common in the literature on ECH, we will sometimes have to consider $J$-holomorphic maps up to equivalence of currents, and we call such an equivalence class a {\bf $J$-holomorphic current}; see \cite{Hutchings-Notes} for the precise definition of this equivalence relation. An equivalence class of $J$-holomorphic maps under the relation of biholomorphisms of the domain will be called a {\bf $J$-holomorphic curve}; this relation might be more familiar to the reader, but is not sufficient for our needs. For future reference, we will call a $J$-holomorphic curve or current {\bf irreducible} when its domain is connected. A $J$-holomorphic map $u: (\Sigma,j) \to (X,J)$ is called {\bf somewhere injective} if there exists a point $ z \in \Sigma$ such that $u^{-1}(u(z))$ = \{z\} and $du: T_z\Sigma \rightarrow T_{u(z)} X$ is injective. We say that $u$ is a {\bf multiple cover} if there exists a branched cover $\phi : (\Sigma, j) \rightarrow (\Sigma', j')$ of degree greater than $1$ and a $J$--holomorphic map $u': (\Sigma',j') \to (X,J)$ such that $u = u' \circ \phi$. Every nonconstant irreducible $J$--holomorphic curve is either somewhere injective or multiply covered; and, two somewhere injective maps are equivalent as curves if and only if they are equivalent as currents. % In the lemma below we state a standard property of $J$-holomorphic curves which plays a key role in our arguments. % For a proof see the argument in \cite[Lem. 9.9]{Wendl-Notes}, for example. \begin{lemma} \label{lem:pointwisenonnegative} Suppose $J \in \mathcal{J}(\alpha, \Omega)$ where $(\alpha, \Omega)$ is a stable Hamiltonian structure on $Y$. If $C$ is a $J$-holomorphic curve in $\R \times Y$, then $\Omega$ is pointwise nonnegative on $C$. Furthermore, $\Omega$ vanishes at a point on $C$ only if $C$ is tangent to the span of $\partial_s$ and $R$. \end{lemma} \subsubsection{Weakly admissible almost complex structures on mapping cylinders} \label{sec:admissible_vs_weakly} In this article, we will be almost exclusively considering $J$--holomorphic curves and currents in the symplectization $X= \R \times Y_\varphi$, introduced above in Example \ref{ex:mapping_torus}. Now, the usual SHS on $\R \times Y_\varphi$ is $(dr, \omega_\varphi)$, defined above, and hence, we will be mostly considering almost complex structures $J$ which are admissible for this SHS, i.e.\ $J \in \mathcal{J}(dr, \omega_\varphi)$. However, in Section \ref{sec:PFH_monotone_twist} we will need the added flexibility of working with almost complex structures on $\R \times Y_\varphi$ which are admissible for some SHS of the form $(\alpha, \Omega)$ with $\Omega=\omega_\varphi$ and $(\alpha,\Omega)$ inducing the same orientation as $(dr,\omega_{\varphi})$ we will refer to such almost complex structures as {\bf weakly admissible}. Clearly, an admissible almost complex structure is weakly admissible. % We will now list several observations about weakly admissible almost complex structures which will be helpful in Section \ref{sec:PFH_monotone_twist}. Let $(\alpha_0, \omega_\varphi)$ and $(\alpha_1, \omega_\varphi)$ be SHSs inducing the same orientation and denote their Reeb vector fields by $R_0, R_1$. \begin{enumerate} \item There exists a positive function $\eta : Y_\varphi \rightarrow \R$ such that $R_0 = \eta R_1$. This is because $R_0, R_1 \in \ker(\omega_\varphi)$ and $(\alpha_0, \omega_\varphi)$ and $(\alpha_1, \omega_\varphi)$ induce the same orientation. In particular, $R_0, R_1$ have the same Reeb orbits. \item Define $\alpha_t = (1-t)\alpha_0 + t \alpha_1$. Then, $(\alpha_t, \omega_\varphi)$ is a SHS for all $t\in [0,1]$. In other words, the space of all SHS of the above form is convex, hence contractible. \item As a consequence of the previous observation, we see that the space of weakly admissible almost complex structures is path connected. Indeed, it is even contractible as it forms a fibration, over the space of SHS with $\Omega =\omega_\varphi$, whose base and fibres are contractible. \item Lastly, the conclusions of Lemma \ref{lem:pointwisenonnegative} hold for $J$-holomorphic curves, for weakly admissible $J$. \end{enumerate} \subsection{Definition of periodic Floer homology} \label{sec:pfhdefn} Periodic Floer homology (PFH) is a version of Floer homology, defined by Hutchings \cite{Hutchings-index, Hutchings-Sullivan-Dehntwist}, for area-preserving maps of surfaces. The cons\-truction of PFH is closely related to the better-known embedded contact homology (ECH) and, in fact, predates the construction of ECH. We now review the definition of PFH; for further details on the subject we refer the reader to \cite{Hutchings-index, Hutchings-Sullivan-Dehntwist}. Let $(S,\omega_S)$ be a closed\footnote{PFH can still be defined if $S$ is not closed, but we will not need this here.} surface with an area form, and $\varphi$ a {\bf nondegenerate} smooth area-preserving diffeomorphism. Non-degeneracy is defined as follows: A periodic point $p$ of $\phi$, with period $k$, is said to be non-degenerate if the derivative of $\varphi^k$ at the point $p$ does not have $1$ as an eigenvalue. We say $\varphi$ is {\bf $d$--nondegenerate} if all of its periodic points of period at most $d$ are nondegenerate; if $\varphi$ is $d$--nondegenerate for all $d$, then we say it is non-degenerate. A $C^\infty$-generic area-preserving diffeomorphism is nondegenerate. Recall the definition of $Y_\varphi$ from Example \ref{ex:mapping_torus} and take $ 0 \neq h \in H_1(Y_\varphi)$. If $\varphi$ is nondegenerate and satisfies a certain ``monotonicity" assumption,\footnote{If the monotonicity assumption does not hold, we can still define PFH, but we need a different choice of coefficients; this is beyond the scope of the present work. } which we do not need to discuss here as it automatically holds when $S= \S^2$, the {\bf periodic Floer homology} % $PFH(\varphi,h)$ is defined; it is the homology of a chain complex $PFC(\varphi, h)$ which we define below. \begin{remark}\label{rem:ECH} If we carry out the construction outlined below, nearly verbatim, for a contact SHS $(\lambda, d\lambda)$, rather than the SHS $(dr, \omega_\varphi)$, then we would obtain the {\bf embedded contact homology} ECH; see \cite{Hutchings-Notes} for further details. \end{remark} \subsubsection{PFH generators} The chain complex $PFC$ is freely generated over\footnote{We could also define PFH over $\mathbb{Z}$, but we do not need this here.} $\mathbb{Z}_2$, by certain finite orbit sets $\alpha = \lbrace (\alpha_i, m_i) \rbrace$ called {\bf PFH generators}. Specifically, we require that each $\alpha_i$ is an embedded Reeb orbit, the $\alpha_i$ are distinct, the $m_i$ are positive integers, $m_i = 1$ whenever $\alpha_i$ is \emph{hyperbolic},\footnote{Being hyperbolic means that the eigenvalues at the corresponding periodic point of $\varphi$ are real. Otherwise, the orbit is called {\it elliptic}. } and $\sum m_i [\alpha_i] = h$. \subsubsection{The ECH index} The $\mathbb{Z}_2$ vector space $PFC(\varphi, h)$ has a relative $\mathbb{Z}$ grading which we now explain. Let $\alpha =\{(\alpha_i, m_i) \rbrace, \beta= \lbrace (\beta_j, n_j) \rbrace$ be two $PFH$ generators in $PFC(\varphi, h)$. Define $H_2(Y_\varphi, \alpha, \beta)$ to be the set of equivalence classes of $2$--chains $Z$ in $Y_\varphi$ satisfying $\partial Z = \sum m_i \alpha_i - \sum n_i \beta_i$. Note that $H_2(Y_\varphi, \alpha, \beta)$ is an affine space over $H_2(Y_\varphi)$. We define the {\bf ECH index} \begin{equation} \label{eqn:ECH_index} I(\alpha,\beta, Z) = c_{\tau} (Z) + Q_{\tau}(Z) + \sum_i \sum^{m_i}_{k = 1} CZ_{\tau}(\alpha^k_i) - \sum_j \sum^{n_j}_{k = 1} CZ_{\tau}(\beta^k_j), \end{equation} where $\tau$ is (the homotopy class of a trivialization) of the plane field $\xi$ over all Reeb orbits, $c_{\tau}(Z)$ denotes the relative first Chern class of $\xi$ over $Z$, $Q_{\tau}(Z)$ denotes the relative intersection pairing, and $CZ_{\tau}(\gamma^k)$ denotes the Conley-Zehnder index of the $k^{th}$ iterate of $\gamma$; all of these quantities are computed using the trivialization $\tau$. We will review the definitions of $c_{\tau}$, $CZ_\tau$, and $Q_{\tau}$ in Section \ref{sec:index}. It is proven in \cite{Hutchings-index} that although the individual terms in the above definition do depend on the choice of $\tau$, the ECH index itself does not depend on $\tau$. According to \cite[Prop 1.6]{Hutchings-index}, the change in index caused by changing the relative homology class $Z$ to another $Z' \in H_2(Y_\varphi, \alpha, \beta)$ is given by the formula \begin{equation}\label{eqn:index_ambiguity} I(\alpha, \beta, Z) - I(\alpha, \beta, Z') = \langle c_1(\xi) + 2 PD(h), Z-Z' \rangle. \end{equation} \subsubsection{The differential} Let $J \in \mathcal{J}(dr, \omega_\varphi)$ be an almost complex structure on $X = \R \times Y_\varphi$ which is admissible for the SHS $(dr, \omega_\varphi)$ and define \[ \mathcal{M}^{I = 1}_{J}(\alpha,\beta)\] to be the space of $J$-holomorphic currents $C$ in X, modulo translation in the $\mathbb{R}$ direction, with ECH index $I(\alpha, \beta, [C]) =1$, which are asymptotic to $\alpha$ as $s \to + \infty$ and $\beta$ as $s \to - \infty$; we refer the reader to \cite{Hutchings-Sullivan-Dehntwist}, page 307, for the precise definition of asymptotic in this context. Assume now and below for simplicity that $S = \S^2$. (For other surfaces, a similar story holds, but we will not need this.) Then, for generic $J$, $\mathcal{M}^{I = 1}_{J}(\alpha,\beta)$ is a compact $0$-dimensional manifold and we can define the PFH differential by the rule \begin{equation} \label{eqn:diff} \langle \partial \alpha, \beta \rangle = \# \mathcal{M}^{I=1}_J(\alpha,\beta), \end{equation} where $\#$ denotes mod $2$ cardinality. It is shown in \cite{Hutchings-TaubesI,Hutchings-TaubesII} that\footnote{More precisely, \cite{Hutchings-TaubesII} proves that the differential in embedded contact homology squares to zero. As pointed out in \cite{Hutchings-TaubesI} and \cite{Lee-Taubes} this proof carries over, nearly verbatim, to our setting.} $\partial^2 = 0$, hence the homology $PFH$ is defined. Lee and Taubes \cite{Lee-Taubes} proved that the homology does not depend on the choice of $J$; in fact, \cite[Corollary 1.1]{Lee-Taubes} states that for any surface, the homology depends only on the Hamiltonian isotopy class of $\varphi$ and the choice of $h \in H_1(Y_\varphi)$. In our case, where $S = \S^2$, all orientation preserving area-preserving diffeomorphisms are Hamiltonian isotopic, and so we obtain a well-defined invariant which we denote by $PFH(Y_{\varphi},h).$ For future motivation, we note that the Lee-Taubes invariance results discussed here come from an isomorphism of PFH and a version of the Seiberg-Witten Floer theory from \cite{Kronheimer-Mrowka}. Importantly, for the applications to this paper, we can relax the assumption that $\varphi$ is nondegenerate to requiring only that $\varphi$ is $d$-nondegenerate, where $d$, called the {\bf degree}, is the positive integer determined by the intersection of $h$ with the fiber class of the map $\pi: Y_\varphi \rightarrow \S^1$ ; note that any orbit set $\alpha$ with $[\alpha] = h$ must correspond to periodic points with period no more than $d$. \subsubsection{The structure of PFH curves} \label{sec:structure} We now explain part of the motivation for PFH, and for the ECH index $I$; we will use some of the results below later in the paper as well. For additional details on the account here, we refer the reader to \cite{Hutchings-Notes}, for example. Let $C$ be a $J$--holomorphic curve which is asymptotic to orbit sets $\alpha, \beta$ and denote $I([C]):= I(\alpha, \beta, [C])$. A key fact which powers the definition of PFH is the {\bf index inequality} \begin{equation} \label{eqn:indexinequality} \text{ind}(C) \leq I([C]) - 2 \delta(C), \end{equation} valid for any somewhere injective curve $C$, where $\delta(C) \geq 0$ is a count of singularities of $C$; see \cite{Hutchings-Notes} for the precise definition of $\delta(C)$; here $\text{ind}(C)$ refers to the {\bf Fredholm index} \begin{equation} \label{eqn:indexformula_Fredholm} \text{ind}(C) = - \chi(C) + 2 c_{\tau}(C) + CZ^{ind}_{\tau}(C), \end{equation} where $\chi(C)$ denotes the Euler characteristic of $C$, and $CZ_\tau^{ind}$ is another combination of Conley-Zehnder terms that we will review in Section \ref{sec:index}. The Fredholm index is the formal dimension of the moduli space of curves near $C$, and \eqref{eqn:indexinequality} therefore says that the ECH index bounds this dimension from above. In the case relevant here, when $I([C])= 1$, we therefore get an important structure theorem for ECH index $1$ curves. To simplify the exposition, we continue to assume here and below that $S = \mathbb{S}^2.$ \begin{prop} [Lem.\ 9.5, \cite{Hutchings-index}, Cor.\ 2.2 \cite{Hutchings-Sullivan-Dehntwist}] \label{prop:structure_Jcurves} Assume that $J$ is admissible and generic and that $\varphi$ is $d$-nondegenerate\footnote{In \cite{Hutchings-index}, one also wants to assume a ``local linearity" condition around periodic points; however, this condition is not necessary --- as was stated in \cite{Hutchings-index}, this condition was added to simplify the analysis, and it can be dropped, as in \cite{Lee-Taubes}, using work of Siefring \cite{Siefring-relative,Siefring-intersection}.}. Let $C$ be a $J$-holomorphic current in $X$, asymptotic to orbit sets of degree at most $d$. Then $I([C]) \geq 0$, and if $I([C]) = 1$, then $C$ has exactly one embedded component $C'$ with $I(C') = 1$, and all other components, if they exist, are multiple covers of $\mathbb{R}$-invariant cylinders that do not intersect $C'$. \end{prop} By an $\R$--invariant cylinder, we mean a $J$--holomorphic curve, or current, $C$ given by a map $u : \R \times \S^1 \rightarrow X$ of the form $u(s,t) = (s, \alpha(t))$ where $\alpha(t)$ is an embedded Reeb orbit; these are also referred to as {\bf trivial cylinders}. \medskip We end this section by mentioning that for any SHS $(\alpha, \Omega)$ on $Y_\varphi$, the formula \eqref{eqn:indexformula_Fredholm} still gives the formal dimension of the moduli space of curves near $C$ when $J \in \mathcal{J}(\alpha, \Omega);$ see \cite[Sections 7 and 8]{Wendl-Notes}; later, we will make use of this. \subsection{Twisted PFH and the action filtration} \label{sec:specdefn} One can define a twisted version of PFH, where we keep track of the relative homology classes of $J$--holomorphic currents, that we will need to define spectral invariants. It has the same invariance properties of ordinary PFH, e.g. by \cite[Corollary 6.7]{Lee-Taubes}, where it is shown to agree with an appropriate version of Seiberg-Witten Floer cohomology. For the benefit of the reader, we provide a brief explanation of how the twisted version works in relationship with Seiberg-Witten theory in Remark~\ref{rmk:twisted} below. We refer the reader to \cite[Section 11.2.1]{Hutchings-Sullivan-T3}, \cite[Section 1.a.2]{Taubes-ECH=SW-V}. The main reason we want to use twisted PFH is because while PFH does not have a natural action filtration, the twisted PFH does. As above, we are continuing to assume $S = \S^2$. First, note that in this case, $Y_\varphi$ is diffeomorphic to $ \S^2 \times \S^1$ and so $H_1(Y_{\varphi}) = \mathbb{Z}$. A class $h \in H_1(Y_{\varphi})$ is then determined by its intersection with the homology class of a fiber of the map $\pi: Y_{\varphi} \to \S^1$, which we defined above to be the {\bf degree} and denote by the integer $d$; from now on we will write the integer $d$ in place of $h$, since these two quantities determine each other. Choose a reference cycle $\gamma_0$ in $Y_{\varphi}$ such that $\pi|_{\gamma_0}: \gamma_0 \to \S^1$ is an orientation preserving diffeomorphism and fix a trivialization $\tau_0$ of $\xi$ over $\gamma_0$. We can now define the $\widetilde{PFH}$ chain complex $\widetilde{PFC}(\varphi,d)$. A generator of $\widetilde{PFC}(\varphi,d)$ is a pair $(\alpha,Z)$, where $\alpha$ is a PFH generator of degree $d$, and $Z$ is a relative homology class in $H_2(Y_{\varphi},\alpha, d \gamma_0)$. The $\mathbb{Z}_2$ vector space $\widetilde{PFC}(\varphi,d)$ has a canonical $\mathbb{Z}$-grading $I$ given by \begin{equation} \label{eqn:twistedgrading} I(\alpha,Z) = c_{\tau} (Z) + Q_{\tau}(Z) + \sum_i \sum^{m_i}_{k = 1} CZ_{\tau}(\alpha^k_i). \end{equation} The terms in the above equation are defined as in the definition of the ECH index given by Equation \eqref{eqn:ECH_index}. Note that the above index depends on the choice of the reference cycle $\gamma_0$ and the trivialization $\tau_0$ of $\xi$ over $\gamma_0$. The index defined here is closely related to the ECH index of Equation \eqref{eqn:ECH_index}: Let $(\alpha, Z)$ and $(\beta, Z')$ be two generators of $\widetilde{PFC}(\varphi,d)$. Note that $ Z - Z'$ is a relative homology class in $H_2(Y_\varphi, \alpha, \beta)$. Then, it follows from\cite[Prop 1.6]{Hutchings-index} that $$I(\alpha, \beta, Z-Z') = I(\alpha, Z) - I(\beta, Z').$$ As a consequence, we see that the index difference $ I(\alpha, Z) - I(\beta, Z')$ does not depend on the choices involved in the definition of the index. We now define the differential on $\widetilde{PFC}(\varphi,d)$. We say that $C$ is a $J$--holomorphic curve, or current, from $(\alpha, Z)$ to $(\beta, Z')$ if it is asymptotic to $\alpha$ as $s \to + \infty$ and $\beta$ as $s \to -\infty$, and moreover satisfies \begin{equation} \label{eqn:importantequation10} Z' + [C] = Z, \end{equation} as elements of $H_2(Y_\varphi, \alpha, d\gamma_0).$ Suppose that $I(\alpha, Z) - I(\beta, Z') =1$ and let $J \in \mathcal{J}(dr, \omega_\varphi)$. We define \[ \mathcal{M}_J( (\alpha,Z), (\beta,Z') )\] to be the moduli space of $J$-holomorphic currents in $X = \R \times Y_\varphi$, modulo translation in the $\R$ direction, from $(\alpha,Z)$ to $(\beta,Z')$. As before, for generic $J \in \mathcal{J}(dr, \omega_\varphi)$, the above moduli space is a compact $0$--dimensional manifold and we define the differential by the rule $$\langle \partial (\alpha, Z), (\beta, Z') \rangle = \# \mathcal{M}_J( (\alpha,Z), (\beta,Z') ),$$ where $\#$ denotes mod $2$ cardinality. As before, $\partial^2 = 0$ by \cite{Hutchings-TaubesI, Hutchings-TaubesII}, and so the homology $\widetilde{PFH}$ is well-defined; by \cite{Lee-Taubes} it depends only on the degree $d$; we write more about this in Remark~\ref{rmk:twisted} below. We will write it as $\widetilde{PFH}(Y_{\varphi},d).$ By a direct computation in the case where $\varphi$ is an irrational rotation of the sphere, i.e. $\varphi(z, \theta) = (z, \theta + \alpha)$ with $\alpha$ being irrational, we obtain \begin{equation}\label{eq:PFH_sphere} \widetilde{PFH}_*(Y_\varphi,d) = \begin{cases} \mathbb{Z}_2, & \text{if } *=d \text{ mod } 2, \\ 0 & \text{otherwise.} \end{cases} \end{equation} Here is a brief outline of the computation leading to the above identity. The Reeb vector field in $Y_\varphi$ has two simple Reeb orbits $\gamma_+, \gamma_-$ corresponding to the north and the south poles. Both of these orbits are elliptic and so the orbit sets of $\widetilde{PFC}(\varphi, d)$ consist entirely of elliptic Reeb orbits. This implies that the difference in index between any two generators of $\widetilde{PFC}(\varphi, d)$ chain complex is an even integer; see \cite[Proposition 1.6.d]{Hutchings-index}. Thus, the PFH differential vanishes. Now, the above identity follows from the fact that for each index $k$, satisfying $k = d $ mod $2$, there exists a unique generator of index $k$ in $\widetilde{PFC}(\varphi, d)$; this fact may be deduced from the index computations carried out in Section \ref{sec:index}. \medskip The vector space $\widetilde{PFC}(\varphi,d)$ carries a filtration, called the {\bf action filtration}\footnote{The relation between the quantity $\cal A(\alpha, Z)$ and the Hamiltonian action functional discussed in Section \ref{sec:action_spectra} will be clarified in Lemma \ref{lemma:action-action}.}, defined by \[ \mathcal{A}(\alpha,Z) = \int_Z \omega_{\varphi} .\] We define $\widetilde{PFC}^L(\varphi,d)$ to be the $\Z_2$ vector space spanned by generators $(\alpha,Z)$ with $\mathcal{A}(\alpha, Z) \leq L$. By Lemma \ref{lem:pointwisenonnegative}, $\omega_{\varphi}$ is pointwise nonnegative along any $J$-holomorphic curve $C$, and so $\int_C \omega_{\varphi} \geq 0$. This implies that the differential does not increase the action filtration, i.e. $$\partial( \widetilde{PFC}^L(\varphi,d) ) \subset \widetilde{PFC}^L(\varphi,d). $$ Hence, it makes sense to define $\widetilde{PFH}^L(\varphi,d)$ to be the homology of the subcomplex $\widetilde{PFC}^L(\varphi,d)$. We are now in position to define the PFH spectral invariants. There is an inclusion induced map \begin{equation} \label{eqn:iind} \widetilde{PFH}^L(\varphi,d) \to \widetilde{PFH}(Y_\varphi,d). \end{equation} If $0 \ne \sigma \in \widetilde{PFH}(\varphi,d)$ is any nonzero class, then we define the {\bf PFH spectral invariant} \[ c_{\sigma}(\varphi)\] to be the infimum, over $L$, such that $\sigma$ is in the image of the inclusion induced map \eqref{eqn:iind} above. The number $c_{\sigma}(\varphi)$ is finite, because $\varphi$ is non-degenerate and so there are only finitely many Reeb orbit sets of degree $d$, and hence only finitely many pairs $(\alpha,Z)$ of a fixed grading. We remark that $c_{\sigma}(\varphi)$ is given by the action of some $(\alpha,Z)$. Indeed, this can be deduced from the following two observations: \begin{enumerate} \item If $L < L'$ are such that there exists no $(\alpha,Z)$ with $L \leq \mathcal{A}(\alpha, Z) \leq L'$, then the two vector spaces $\widetilde{PFC}^L(\varphi,d)$ and $\widetilde{PFC}^{L'}(\varphi,d)$ coincide and so $\widetilde{PFH}^L(\varphi,d) \to \widetilde{PFH}(Y_\varphi,d)$ and $\widetilde{PFH}^{L'}(\varphi,d) \to \widetilde{PFH}(Y_\varphi,d)$ have the same image. \item The set of action values $\{\mathcal{A}(\alpha, Z) : (\alpha, Z) \in \widetilde{PFC}^L(\varphi,d) \}$ forms a discrete subset of $\R$. This is a consequence of the fact that, as stated above, there are only finitely many Reeb orbit sets of degree $d$. \end{enumerate} In Remark~\ref{rmk:nojdependence} below we show that this does not depend on the choice of the admissible almost complex structure $J$. Note, however, that $c_{\sigma}(\varphi)$ does depend on the choice of the reference cycle $\gamma_0$. \subsection{Initial properties of PFH spectral invariants } \label{sec:PFH_spec_initial_properties} Let $p_ -= (0,0,-1) \in \S^2$. We denote $$\mathcal{S} := \{ \varphi \in \Diff(\S^2, \omega) : \varphi(p_-) = p_-, \, -\tfrac{1}{4} < \rot(\varphi, p_-) < \tfrac{1}{4} \},$$ where $\rot(\varphi, p_-)$ denotes the rotation number of $\varphi$ at $p_-$ as defined in Section \ref{sec:rotation_number}. We remark that our choice of the constant $\frac14$ is arbitrary; any other constant in $(0,\frac12)$ would be suitable for us; we just need to slightly enlarge the class of diffeomorphisms arising from $\Diff_c(\mathbb{D}^2,\omega)$, so as to facilitate computations. Recall from the previous section that the spectral invariant $c_{\sigma}$ depends on the choice of reference cycle $\gamma_0 \in Y_{\varphi}.$ For $\varphi \in S,$ there is a unique embedded Reeb orbit through $p_-$, and we set this to be the reference cycle $\gamma_0$. The grading on $\widetilde{PFH}$ depends on the choice of trivialization $\tau_0$ over $\gamma_0$; our convention in this paper is that we always choose $\tau_0$ such that the rotation number $\theta$ of the linearized Reeb\footnote{Following \cite[Section 3.2]{Hutchings-Notes}, we define the rotation number $\theta$ as follows: Let $\{\psi_t\}_{t\in \R}$ denote the $1$--parameter group of diffeomorphisms of $Y_\varphi$ given by the flow of the Reeb vector field. Then, $D\psi_t : T_{\gamma_0(0)}Y_\varphi \rightarrow T_{\gamma_0(t)}Y_\varphi$ induces a symplectic linear map $\phi_t :\xi_{\gamma_0(0)} \rightarrow \xi_{\gamma_0(t)} $, which using the trivialization $\tau_0$ we regard as a symplectic linear transformation of $\R^2$. We define $\theta$ to be the rotation number of the isotopy $\{\phi_t\}_{t\in [0,1]}$ as defined in Section \ref{sec:rotation_number}. } flow along $\gamma_0$ with respect to $\tau_0$ satisfies $- \frac14 < \theta < \frac14 $: this determines $\tau_0$ uniquely. We will want to single out some particular spectral invariants for $\varphi \in \mathcal{S}$, and show that they have various convenient properties; we will use these to define the spectral invariants for $\varphi \in \Diff_c(\D, \omega)$. Having set the above conventions, we do this as follows. Suppose that $\varphi \in \mathcal{S}$ is non-degenerate. According to Equation \eqref{eq:PFH_sphere}, for every pair $(d,k)$ with $k = d$ mod $2$, we have a distinguished nonzero class $\sigma_{d,k}$ with degree $d$ and grading $k$, and so we can define $$c_{d,k}(\varphi) := c_{\sigma_{d,k}}(\varphi).$$ Lastly, we also define\footnote{ Alternatively, one may define $c_d(\varphi) := c_{d,k}(\varphi)$ for any $-d \le k \le d$ satisfying $k=d$ mod $2$. These alternative definitions are all suitable for our purposes in this article.} $$c_d(\varphi) := c_{d, -d} (\varphi).$$ We will see in the proof of Theorem~\ref{thm:PFHspec_initial_properties} that the $c_{d,k}(\varphi)$ for nondegenerate $\varphi$ determine $c_{d,k}(\varphi)$ for all $\varphi$ by continuity. To prepare for what is coming, we identify a class of Hamiltonians $\mathcal{H}$ with the key property, among others, that $\mathcal{S} = \{\varphi^1_H: H \in \mathcal{H} \}$. Denote \begin{align*} \mathcal{H} := \{H \in C^{\infty}(\S^1 \times \S^2) : &\ \varphi^t_H(p_-) = p_- , H(t, p_-) = 0, \forall t \in [0,1], \\ & -\tfrac14 < \rot( \{\varphi^t_H\}, p_- ) < \tfrac14 \}, \end{align*} where $\rot( \{\varphi^t_H\} , p_-)$ is the rotation number of the isotopy $\{\varphi^t_H\}_{t \in [0,1] }$ at $p_-$; see Section \ref{sec:rotation_number}. Observe that $\mathcal{S} = \{\varphi^1_H: H \in \mathcal{H} \}.$ \medskip The theorem below, which is the main result of this section, establishes some of the key properties of the PFH spectral invariants and furthermore allows us to extend the definition of these invariants to all, possibly degenerate, $\varphi \in \mathcal{S}$. In the statement below $|| \cdot ||_{(1,\infty)}$ denotes the energy, or the {\bf Hofer norm}, on $C^{\infty}(\S^1 \times \S^2)$ which is defined as follows \[ \| H \|_{(1, \infty)} = \int_0^1 \left( \max_{x \in \S^2} H(t,x) - \min_{x \in \S^2} H(t, x)\right) dt.\] \begin{theo} \label{thm:PFHspec_initial_properties} The PFH spectral invariants $c_{d,k}(\varphi)$ admit a unique extension to all $\varphi \in \mathcal{S}$ satisfying the following properties: \begin{enumerate} \item Monotonicity: Suppose that $H \leq G$, where $H, G \in \mathcal{H}$. Then, \[ c_{d, k}(\varphi^1_H) \leq c_{d, k} (\varphi^1_G).\] \item Hofer Continuity: For any $H, G \in \mathcal H$, we have $$| c_{d,k} (\varphi^1_H) - c_{d, k} (\varphi^1_G) | \leq d || H - G ||_{(1,\infty)}.$$ \item Spectrality: $c_{d, k}(\varphi^1_H) \in \Spec_d(H)$ for any $H \in \mathcal{H}$. \item Normalization: $c_{d,-d}(\id) =0$. \end{enumerate} \end{theo} \begin{remark}\label{rem:PFHspec-disc} To define the PFH spectral invariant $c_{d, k}$ for $\varphi \in \Diff_c(\D, \omega) $, we use Equation \eqref{eq:identify_disc_north_hemisphere} to identify $ \Diff_c(\D, \omega)$ with area-preserving diffeomorphisms of the sphere which are supported in the interior of the northern hemisphere $S^+$. We similarly define $c_d: \Diff_c(\D, \omega) \rightarrow \R$ which was introduced in Section \ref{sec:PFH_spectralinvariants_intro}. It follows from Theorem~\ref{thm:PFHspec_initial_properties} that $c_d : \Diff_c(\D, \omega) \rightarrow \R$ satisfies the properties 1-4 in Section \ref{sec:PFH_spectralinvariants_intro}. \end{remark} The rest of this section is dedicated to the proof of the above theorem. The proof requires certain preliminaries. First, it will be convenient to explicitly identify $Y_{\varphi}$ with $\S^1 \times \S^2$. To do so pick $ H \in \mathcal{H}$ such that $\varphi = \varphi^1_H$.\footnote{We remark that the choice of $H\in \mathcal H$ such that $\varphi = \varphi^1_H$ is unique up to homotopy of Hamiltonian isotopies rel endpoints. This fact, which is not used in our arguments, may be deduced from properties of the rotation number.} We define \begin{equation} \label{eq:trivialization} \begin{split} & \S^1 \times \S^2 \rightarrow Y_{\varphi} \\ (t,x) & \mapsto \left( (\varphi^{t}_H)^{-1} (x), t \right), \end{split} \end{equation} where $t$ denotes the variable on $\S^1$. For future reference, note that this identifies the Reeb vector field on $Y_\varphi$ with the vector field \begin{equation} \label{eqn:trivializedreeb} \partial_t + X_H \end{equation} on $\S^1 \times \S^2$. The $2$-form $\omega_{\varphi}$ pulls back under this map to the form \[ \omega + dH \wedge dt \] where $\omega$ is the area form on $\S^2$. The Reeb orbit $\gamma_0$ maps under \ref{eq:trivialization} to the preimage of $p_-$ under the map $\S^1 \times \S^2 \to \S^2$; we will continue to denote it by $\gamma_0$. Moreover, the trivialization $\tau_0$ from above agrees (up to homotopy) under this identification with the trivialization over $\gamma_0$ given by pulling back a fixed frame of $T_{p_-} \S^2$ under the map $\S^1 \times \S^2 \to \S^2$. The map \ref{eq:trivialization} allows us to identify $\mathbb{R} \times \S^1 \times \S^2$ with the symplectization $X$ via \begin{align*} \mathbb{R} \times \S^1 \times \S^2 &\rightarrow X \\ (s,t,x) &\mapsto (s,(\varphi_{H}^t)^{-1}(x), t). \end{align*} The symplectic form $\Gamma$ on $X$ then pulls back to \begin{equation} \label{eqn:trivsymplectic} \omega_H = ds \wedge dt + \omega + dH \wedge dt. \end{equation} Let $H, K$ be two Hamiltonians in $\mathcal H$. As mentioned earlier, $\widetilde{PFH}(\varphi_H^1, d)$ is isomorphic to $\widetilde{PFH}(\varphi_K^1,d)$. The proof of this uses Seiberg-Witten theory, and is carried out in \cite[Corollary 6.1]{Lee-Taubes}; this isomorphism is canonical with a choice of reference cycle in $H_2(\S^1 \times \S^2,\gamma_0, \gamma_0)$; we say more about this in Remark~\ref{rmk:twisted} below. We take this reference cycle to be the constant cycle\footnote{This is the projection $\gamma_0 \times I \to \gamma_0$.} over $\gamma_0$. In this case, we will see below that the canonical isomorphism \begin{equation} \label{eqn:choice} \widetilde{PFH}(\varphi^1_H,d) \to \widetilde{PFH}(\varphi^1_K,d), \end{equation} preserves the $\mathbb{Z}$-grading. As is generally the case with related invariants, one might expect this isomorphism to be induced by a chain map counting certain $J$-holomorphic curves. In fact, it is not currently known how to define the map \eqref{eqn:choice} this way; the construction uses Seiberg-Witten theory. Nevertheless, the map in \eqref{eqn:choice} does satisfy a ``holomorphic curve" axiom which was proven by Chen \cite{Chen} using variants of Taubes' ``Seiberg-Witten to Gromov" arguments in \cite{Taubes-Gr=SW}. A similar ``holomorphic curve" axiom was proven in the context of embedded contact homology by Hutchings-Taubes; we compare the Chen proof to the Hutchings-Taubes one in Remark~\ref{rmk:chen} below. To state this holomorphic curve axiom in our context, take Hamiltonians $H, K \in \mathcal{H}$, and define for $s\in \R$ \[ G_s= K + \beta(s) \cdot (H-K)\] where $\beta: \R \rightarrow [0,1]$ is some non-decreasing function that is $0$ for $s$ sufficiently negative and $1$ for $s$ sufficiently positive. Now consider the form \[\omega_X = ds\wedge dt + \omega + dG \wedge dt,\] where, as throughout this article, $dG$ denotes the derivatives in the $\S^2$ directions. This is a symplectic form on $\mathbb{R} \times \S^1 \times \S^2$. Observe that, for $s > > 0$, the form $\omega_X$ agrees with the symplectization form $\omega_H$, and for $s < <0$, it agrees with the symplectization form $\omega_K$. Let $J_X$ be any $\omega_X$-compatible\footnote{Recall that an almost complex structure $J$ is {\bf compatible} with a symplectic form $\omega$ if $g(u,v) := \omega(u,Jv)$ defines a Riemannian metric.} almost complex structure that agrees with a generic $(dt,\omega_H)$ admissible almost complex structure $J_+$ for $s >> 0$ and with a generic $(dt,\omega_K)$ admissible almost complex structure $J_-$ for $s << 0$. Then, the holomorphic curve axiom says that \eqref{eqn:choice} is induced by a chain map \begin{equation} \label{eqn:chainchoice} \Psi_{H,K}: \widetilde{PFC}(\varphi^1_H,d,J_+) \to \widetilde{PFC}(\varphi^1_K,d,J_-), \end{equation} with the property that if $ \langle \Psi_{H,K} (\alpha,Z), (\beta,Z') \rangle \ne 0$, then there is an ECH index $0$ $J_X$-{\bf holomorphic building} $C$ from $\alpha$ to $\beta$ such that \begin{equation} \label{eqn:importantequation} Z' + [C] = Z, \end{equation} as elements of $H_2(\S^1 \times \S^2, \alpha, d \gamma_0);$ we say more about this in Remark~\ref{rmk:twisted} below. Here, by a $J_X$-holomorphic building from $\alpha$ to $\beta$, we mean a sequence of $J_i$-holomorphic curves \[ (C_0,\ldots,C_i, \ldots, C_k),\] such that the negative asymptotics of $C_i$ agree with the positive asymptotics of $C_{i+1}$, the curve $C_0$ is asymptotic to $\alpha$ at $+\infty$, and the curve $C_k$ is asymptotic to $\beta$ at $- \infty$; we refer the reader to \cite[Section 5.3]{Hutchings-Notes} for more details. We remark for future reference that the $C_i$ are called {\bf levels}, and each $J_i$ is either\footnote{More can be said, but we will not need this additional information} $J_X, J_+$ or $J_-$. The condition that the ECH index of the building is zero means that the sum of the ECH indices of the levels add up to zero. In particular, this index condition, together with \eqref{eqn:importantequation}, implies the earlier claim that the map \eqref{eqn:choice} preserves the $\mathbb{Z}$-grading by additivity of $c_{\tau}$ and $Q_{\tau}$, since the trivializations over $\gamma_0$ required to define the grading on $\widetilde{PFC}(\varphi^1_H)$ and $\widetilde{PFC}(\varphi^1_K)$ are the same. We will want to assume that $J_X$ is {\bf compatible with the fibration} $\R \times \S^1 \times \S^2 \rightarrow \mathbb{R} \times \S^1$ in the following sense: Let $\mathbb{V}$ be the vertical tangent bundle of this fibration and denote by $\mathbb{H}$ the $\omega_X$-orthogonal complement of $\mathbb V$; observe that $\mathbb{H}$ is spanned by the vector fields $ \partial_s$ and $ \partial_t + X_G $. Then, we will want $J_X$ to preserve $\mathbb{V}$ and $\mathbb{H}$. Given any admissible $J_{\pm}$ on the ends, we can achieve this as follows. On the horizontal tangent bundle $\mathbb{H}$, we always demand that $J_X$ sends $\partial_s$ to $\partial_t + X_G$. On the vertical tangent bundle, we observe that $\omega_X|_{\mathbb{V}} = \omega$, and in particular $\omega_X|_{\mathbb{V}}$ is independent of $s$; we can then connect $J_{+}|_{\mathbb{V}} $ to $J_{-}|_{\mathbb{V}}$ through a path of $\omega$-tamed almost complex structures on $\mathbb{V}$. We can now prove Theorem~\ref{thm:PFHspec_initial_properties}. \begin{proof}[Proof of Theorem~\ref{thm:PFHspec_initial_properties}] We begin by first supposing that the monotonicity, Hofer continuity, and spectrality properties hold when $\varphi^1_K, \varphi^1_H$ are nondegenerate and explain how this implies the rest of the theorem. To that end, let $H \in \mathcal H$, not necessarily nondegenerate, and take a sequence $H_i \in \mathcal H$ which $C^2$ converges to $H$ and such that $\varphi^1_{H_i}$ is nondegenerate. Then, we define $$ c_{d, k} (\varphi^1_H) = \lim_{i\to \infty} c_{d,k}(\varphi^1_{H_i}).$$ This limit exists thanks to the inequality $| c_{d,k} (\varphi^1_{H_i}) - c_{d, k} (\varphi^1_{H_j}) | \leq d || H_i - H_j ||_{(1,\infty)}.$ Moreover, the same inequality implies that the limit value does not depend on the choice of the sequence $H_i$ and so $c_{d, k} (\varphi^1_H)$ is well-defined for all $H \in \mathcal H$. Thus, we obtain a well-defined mapping $$ c_{d, k} : \mathcal{S} \rightarrow \R.$$ It can be seen that $c_{d, k}$ continues to satisfy the monotonicity and Hofer continuity properties for degenerate $\varphi^1_K, \varphi^1_H$. The spectrality property is also satisfied; this is a consequence of the Arzela-Ascoli theorem under the assumption that spectrality is satisfied in the nondegenerate case; we will not provide the details here. % Moreover, note that, by the Hofer continuity property, the mapping $ c_{d, k} : \mathcal{S} \rightarrow \R$ is uniquely determined by its restriction to the set of all non-degenerate $\varphi \in \mathcal S$. To prove that $c_{d, -d}(\id) =0$, it is sufficient to show that $c_{d, -d}(\varphi) = 0$ in the case where $\varphi$ is a positive irrational rotation of the sphere, i.e. $\varphi(z, \theta) = (z, \theta + \alpha)$ with $\alpha $ being a small and positive irrational number. As in the explanation for Equation \eqref{eq:PFH_sphere}, the chain complex $\widetilde{PFC}(\varphi, d)$ has a unique generator in indices $k$ such that $k =d$ mod $2$ and it is zero for other indices. The unique generator of index $-d$ is of the form $(\alpha, Z)$ where $\alpha = \{(\gamma_0, d)\}$ and $Z$ is the trivial class in $H_2(Y_\varphi, d\gamma_0, d\gamma_0)$; this fact can be deduced from the forthcoming index computations of Section \ref{sec:index}. The action $\mathcal{A}(\alpha, Z)$ is zero. This proves that $c_{d, -d} (\varphi) = 0 = c_{d, -d}(\id)$.\footnote{ With a similar argument one can prove that $c_{d, k}(\id) = 0$ for every $-d \le k \le d$ with $k=d$ mod $2$. } \medskip For the rest of the proof we will suppose that $\varphi^1_{H}, \varphi^1_K$ are nondegenerate. We will now prove the monotonicity and Hofer continuity properties. Let $(\alpha_1, Z_1) + \ldots + (\alpha_m, Z_m)$ be a cycle in $\widetilde{PFC}(\varphi^1_H,d)$ representing $\sigma_{d,k}$, with $$c_{\sigma_{d,k}}(\varphi^1_H) = \mathcal{A}(\alpha_1, Z_1) \geq \ldots \geq \mathcal{A}(\alpha_m, Z_m).$$ Let $(\beta, Z')$ be a generator in $ \widetilde{PFC}(\varphi^1_K,d)$ which has maximal action among generators which appear with a non-zero coefficient in $$\Psi_{H,K}\left( (\alpha_1, Z_1) + \\ \ldots + (\alpha_m, Z_m) \right).$$ Then, by the holomorphic curve axiom there is a $J_X$-holomorphic building from some $(\alpha_i, Z_i)$ to $(\beta, Z')$. For the rest of the proof we will write $(\alpha_i, Z_i) = (\alpha,Z)$ and will denote the $J_X$-holomorphic building by $C$. For the arguments below, which only involve energy and index arguments, we can assume that $C$ consists of a single $J_X$-holomorphic level -- in other words, is an actual $J_X$-holomorphic curve, rather than a building -- so to simplify the notation, we assume this. \medskip For the remainder of the proof we will need the following Lemma. \begin{lemma} \label{lemm:important_identity_inequality} The following identity holds: $$\mathcal{A}(\alpha, Z) - \mathcal{A}(\beta, Z') = \int_C \omega + dG\wedge dt + G' ds\wedge dt, $$ where $G'$ denotes $\frac{\partial G}{\partial s}$. Furthermore, we have $$\int_C \omega + dG\wedge dt \geq 0.$$ \end{lemma} \begin{proof}[Proof of Lemma \ref{lemm:important_identity_inequality}] We will begin by proving that \begin{equation} \label{eqn:actionequation} \mathcal{A}(\alpha,Z) - \mathcal{A}(\beta,Z') =\int_C \omega + d(Gdt), \end{equation} which establishes the first item because $\omega+d(Gdt) = \omega + dG \wedge dt + G' ds \wedge dt.$ Note that we can write \[ \mathcal{A}(\alpha,Z) = \int_Z \omega + d(Hdt), \quad \mathcal{A}(\beta,Z') = \int_{Z'} \omega + d(Kdt).\] Hence, Equation \eqref{eqn:actionequation} would follow if we show that \[ \int_C \omega = \int_Z \omega - \int_{Z'} \omega, \text{ and } \int_C d(Gdt) = \int_Z d(Hdt) - \int_{Z'} d(Kdt) .\] The first identity holds because all of these integrals are determined by the homology classes, and we have $[C] = Z - Z'.$ The second identity follows from the following chain of identities: \begin{align*} \int_C d(Gdt) &= \int_\alpha G dt - \int_\beta Gdt \\ &= \int_\alpha H dt - \int_\beta K dt \\ &= \int_Z d(Hdt) - \int_{Z'} d(Kdt), \end{align*} where the first equality holds by Stokes' theorem, the second follows from the definition of $G$, and the third is a consequence of Stokes' theorem combined with the fact that $H, K$ both belong to $\mathcal{H}$ and so vanish on $\gamma_0$. This completes the proof of the first item in the lemma. Now, we will show that $\int_C (\omega + dG \wedge dt) \geq 0$ by showing that the form $\omega + dG \wedge dt$ is pointwise non-negative along $C$. Indeed, at any point $p \in X$, we can write any vector as $v + h$, where $v \in \mathbb{V}$ and $h \in \mathbb H$ are vertical and horizontal tangent vectors as described in the paragraph before the proof of Theorem \ref{thm:PFHspec_initial_properties}. Since $C$ is $J_X$--holomorphic, it is sufficient to show that $(\omega + dG \wedge dt)(v + h, J_X v + J_X h) \geq 0.$ We will show that \begin{equation} \label{eq:1} (\omega + dG \wedge dt)(v + h, J_X v + J_X h) = \omega_X (v, J_Xv), \end{equation} which proves the inequality because $J_X$ is $\omega_X$-tame. Now, to simplify our notation we will denote $\Omega = \omega + dG \wedge dt$ for the rest of the proof. Expanding the left hand side of the above equation we get \[\Omega(v + h, J_Xv + J_Xh) = \Omega(v, J_Xv) + \Omega(h, J_Xh) + \Omega(v, J_Xh) + \Omega(h, J_Xv). \] We will now show that $\Omega(v, J_Xv) = \omega_X(v, J_Xv)$ and $\Omega(h, J_Xh)= \Omega(v, J_Xh) = \Omega(h, J_Xv) =0$ which clearly implies Equation \eqref{eq:1}. To see this, note that $v$ and $J_Xv$ are in the kernel of $ds \wedge dt$, hence \[ \Omega(v, J_Xv) = \omega_X(v, J_X v),\] \[ \Omega(v, J_X h) = \omega_X(v,J_X h) = 0, \quad \Omega(h, J_X v) =\omega_X(h,J_X v) = 0.\] It remains to show that $\Omega(h, J_X h) = 0$, that is $\Omega\vert_\mathbb{H} = 0$. This follows from the fact that $\bb H$ is spanned by $\lbrace \partial_s, \partial_t + X_G \rbrace$ and $\partial_s$ is in the kernel of $\Omega$. Indeed, a $2$--form on a $2$--dimensional vector space with non-trivial kernel is identically zero. \end{proof} \medskip Note that $c_{d,k}(\varphi^1_H) \geq \mathcal{A}(\alpha, Z)$ and $c_{d,k}(\varphi^1_K) \leq \mathcal{A}(\beta,Z').$ Hence, \begin{equation}\label{eqn:lower_bound_action} c_{d,k}(\varphi^1_H) - c_{d,k}(\varphi^1_K) \geq \mathcal{A}(\alpha,Z) - \mathcal{A}(\beta,Z'). \end{equation} As a consequence of this inequality, Monotonicity would follow from proving that if $H \geq K$, then $\mathcal{A}(\alpha,Z) - \mathcal{A}(\beta,Z') \geq 0.$ By the above lemma we have \begin{equation} \label{eqn:masterequation} \mathcal{A}(\alpha,Z) - \mathcal{A}(\beta,Z') \geq \int_C G'\,ds \wedge dt. \end{equation} If $H \geq K$, then $G' \geq 0$, and so $\int_C G' ds \wedge dt \geq 0$, which proves Monotonicity. As for Hofer Continuity, it is sufficient to show that \begin{equation}\label{eq:estimate_hofer} \left|\int_C G'\,ds \wedge dt \right| \leq d \Vert H-K \Vert_{(1, \infty)}. \end{equation} Indeed, this inequality combined with Inequalities \eqref{eqn:lower_bound_action} and \eqref{eqn:masterequation} implies that $c_{d,k}(\varphi^1_K) - c_{d,k}(\varphi^1_H) \leq d \Vert H-K \Vert_{(1, \infty)}.$ Similarly, by switching the role of $H$ and $K$, one gets $c_{d,k}(\varphi^1_H) - c_{d,k}(\varphi^1_K) \leq d \Vert K -H \Vert_{(1, \infty)}$ which then implies Hofer Continuity. It remains to prove Inequality \eqref{eq:estimate_hofer}. We know that the form $ds \wedge dt$ is pointwise nonnegative on $C$, because $ds \wedge dt|_{\bb V} = 0$, we saw in the proof of the previous lemma that $\Omega|_{\bb H} = 0$, and $J_X$ is $\omega_X$-tame. Hence \[ \left|\int_C G' ds \wedge dt \right| = \left|\int_C \beta'(s) (H-K) ds \wedge dt \right| \leq \int_C \beta'(s) |H-K| ds \wedge dt.\] Note that because $H, K$ both vanish at the point $p_-$, for all $t,x$ we have \[ |H(t,x) - K(t,x) | \leq \max_{\S^2} \left (H_t -K_t\right) -\min_{\S^2} \left (H_t -K_t \right) .\] Hence, we get \[ \left|\int_C G' ds \wedge dt \right| \leq \int_C \beta'(s) \left(\max_{\S^2} (H-K) - \min_{\S^2}(H - K)\right) ds \wedge dt.\] We can evaluate the second integral by projecting $C$ to the $(s,t)$ plane; this projection has degree $d$, and since $\int_{-\infty}^{+\infty}\beta'=1$, the second integral evaluates to \[ d ||H-K||_{(1,\infty)} .\] This completes the proof of Hofer Continuity. \medskip It remains to prove Spectrality. As stated in Section \ref{sec:specdefn}, the spectral invariant $c_{d,k}(\varphi^1_H)$ is the action of a PFH generator $(\alpha,Z)$ of degree $d$. Spectrality, hence Theorem \ref{thm:PFHspec_initial_properties}, is then a consequence of the following lemma. \end{proof} \begin{lemma}\label{lemma:action-action} Let $(\alpha, Z)$ be a PFH generator of degree $d$ for the mapping torus $Y_\varphi$ of $\varphi=\varphi_H^1$ with $H \in \mathcal{H}$. Then, $\cal A(\alpha, Z)$ belongs to $\Spec_d(H)$, as defined in Section \ref{sec:action_spectra}. \end{lemma} \begin{proof} For every orbit set $\alpha$ we will construct a specific relative class $Z_\alpha \in H_2(Y_\varphi, \alpha, \gamma_0)$ and will show that $\mathcal{A}(\alpha, Z_\alpha) \in \Spec_d(H)$. Any other $Z\in H_2(Y_\varphi, \alpha, \gamma_0)$ is of the form $Z_\alpha + k [\S^2]$ where $k \in \Z$. Hence, $\mathcal{A}(\alpha, Z) = \mathcal{A}(\alpha, Z_\alpha) + k$ and so we get that $\mathcal{A}(\alpha, Z) \in \Spec_d(H)$ for all $Z$. First, suppose that $d=1$. Let $q \in \Fix(\varphi)$ and suppose that $\alpha$ is the Reeb orbit in the mapping cylinder corresponding to $q$. The relative cycle $Z_{\alpha}$ will be of the form $Z_{\alpha}= Z_0 + Z_1 + Z_2$. We begin by choosing a path $\eta$ in $\S^2 \times\{0\} \subset Y_{\varphi^1_H}$ such that $\partial \eta = (q,0) - (p_-, 0)$. We parametrize this curve with a variable $x \in [0,1]$. We define $Z_0$ to be the chain induced by the map \begin{align*} \begin{split} &[0,1]^2 \to Y_{\varphi^1_H}, \;\; (x,t)\mapsto (\eta(x),t). \end{split} \end{align*} Its boundary is given by $\partial Z_0= \alpha - \gamma_0 + (\eta, 0) - (\varphi(\eta), 0)$. Note that $ \int_{Z_0} \omega_{\varphi} =0.$ We define $Z_1$ to be the chain induced by the map \begin{align*} \begin{split} &[0,1]^2 \to Y_{\varphi^1_H}, \;\; (t,x) \mapsto (\varphi^t_H( \eta(x)), 0). \end{split} \end{align*} Then, $\partial Z_1 = (\varphi(\eta), 0) - (\eta, 0) - (\varphi^t_H(q), 0)$. We can compute $\int_{Z_1} \omega_{\varphi}$ as follows: \begin{align*} \int_{Z_1} \omega_{\varphi} &= \int \int_{[0,1]^2} \omega \langle \partial_t \varphi^t_H(\eta(x) ) , \partial_x \varphi^t_H (\eta(x)) \rangle \\ &= \int \int_{[0,1]^2} \omega \langle X_{H_t} (\varphi^t_H(\eta(x ))) , \partial_x \varphi^t_H (\eta(x)) \rangle \\ &= \int \int _{[0,1]^2} dH_t ( \partial_x \varphi^t_H( \eta(x) ) ) = \int \int _{[0,1]^2} \partial_x H_t ( \varphi^t_H (\eta(x)) ) \\ &= \int^{1}_{0} H_t (\varphi^t_H(q)) - H_t(\varphi^t(p_-))dt = \int^1_0 H_t(\varphi^t_H(q)) dt. \end{align*} Next, we define $Z_2$ to be the chain induced by a map $(u_\alpha, 0)$ where $u_\alpha:D^2 \rightarrow \S^2$ is such that $u_\alpha|_{\partial D^2}$ is the Hamiltonian orbit $t\mapsto\varphi^t_H(q)$. Then, $\partial Z_2= (\varphi^t_H(q), 0)$ and $\int_{Z_2} \omega_{\varphi} = \int_{D^2} u_\alpha^* \omega$. We will furthermore require $u_\alpha$ to satisfy the following additional properties which will be used in Section \ref{sec:comb}: \begin{enumerate}[(i)] \item If $\alpha = \gamma_0$, the Reeb orbit corresponding to $p_-$, then we take $u_\alpha$ to be the constant disc with image $p_-$. \item If $\alpha \neq \gamma_0$, then we take $u_\alpha$ such that its image does not contain $p_-$. \end{enumerate} Finally, we set $Z_\alpha = Z_0 + Z_1 + Z_2$. Adding up the above quantities we obtain $\partial Z_\alpha=\partial Z_0 + \partial Z_1 + \partial Z_2 = \alpha - \gamma_0$ and \begin{equation}\label{eqn:action_Z_alpha} \mathcal{A}(\alpha,Z_\alpha) = \int_{Z_0} \omega_{\varphi} + \int_{Z_1} \omega_{\varphi} + \int_{Z_2} \omega_{\varphi} = \int_{D^2} u_\alpha^* \omega + \int^1_0 H_t(\varphi^t_H(q)) dt. \end{equation} Clearly, $\mathcal{A}(\alpha,Z_\alpha) \in \Spec(H)$. We remark that $Z_{\alpha}$ does not depend on the choice of $u_{\alpha}$ subject to the above conditions. Next, let $q$ be a periodic point of period $m \in \N$ and suppose that $\alpha$ is the Reeb orbit in the mapping cylinder corresponding to $q$. Then, $q$ is a fixed point of $\varphi_H^m$. Consider the mapping torus $Y_{\varphi_H^m}$. There is a map $c:Y_{\varphi_H^m} \to Y_{\varphi_H^1}$, pulling back $\omega_{\varphi^1_H}$ to $ \omega_{\varphi^m_H}$, given by mapping each interval $\S^2 \times [\frac{k}{m}, \frac{k+1}{m}]$ onto $Y_{\varphi}$ via the map \[ (x,t) \mapsto (\varphi_H^{k}(x) , m \cdot t - k).\] Now repeat the construction from above to produce a relative cycle $Z'$ in $Y_{\varphi^m}$. Define $Z_\alpha$ to be the pushforward of $Z'$ under the map $c$. Then \[ \int_{Z_\alpha} \omega_{\varphi_H^1} = \int_{Z'} \omega_{\varphi_H^m}.\] The map $\varphi_H^m$ is generated by the Hamiltonian $H^m$, so by the argument in the $d = 1$ case, $\int_{Z'} \omega_{\varphi_H^m} \in \Spec(H^m)$, hence the same holds for $\int_{Z_{\alpha}} \omega_{\varphi^1_H}.$ We note for future reference that as \eqref{eqn:action_Z_alpha} holds for $\mathcal{A}(\alpha,Z')$ with respect to $\varphi^m_H$, and $\varphi^m_H$ can be viewed as the time $1$-map of the Hamiltonian $F_t(x) = m H_{mt}(x)$, we have \begin{equation}\label{eqn:action_Z_alpha_2} \mathcal{A}(\alpha,Z_\alpha) = \int_{D^2} u_\alpha^* \omega + \int^m_0 H_t(\varphi^t_H(q)) dt, \end{equation} where $u_\alpha$ is a disc whose boundary is the Hamiltonian orbit $t \mapsto \varphi^t_H(q), t\in [0, m]$, which satisfies the analogues of properties (i) and (ii) above. Now let $\alpha = \lbrace (\alpha_i, m_i) \rbrace$, where the $\alpha_i$ are simple closed orbits of $\partial_t$. So, each $(\alpha_i,m_i) $ corresponds to a (not necessarily simple) orbit of a periodic point $q_i$ of $\varphi^1_H$. By using the construction in the previous paragraph, we can associate a relative cycle to each $(\alpha_i, m_i)$; the sum, over $i$, of all of these cycles gives a relative cycle from $\alpha$ to $d \gamma_0$, where $d$ is the sum of the periods of the periodic points $q_i$. The arguments in the previous paragraphs show that the action of this cycle $\mathcal{A}(\alpha, Z)$ is in $\Spec_d(H)$. \end{proof} \begin{remark} \label{rmk:nojdependence} In the special case where $H = K$, but the two $J_i$ are different, the Monotonicity argument above, applied first to $H \geq K$ and next to $K \geq H$, gives that the spectral invariant does not depend on $J$. \end{remark} \begin{remark} \label{rmk:chen} For the benefit of the reader, we provide a brief comparison between the Chen proof and the proof of the holomorphic curve axiom in the ECH context by Hutchings-Taubes; we also provide a sketch of the Chen proof. For brevity, we assume in the remark that the reader is familiar with some of the relevant background, referring to \cite{Hutchings-Taubes-ChordII} for relevant terminology. In the Hutchings-Taubes context, one starts with an exact symplectic cobordism between contact manifolds, and attaches symplectization-like ends, to get a non-compact symplectic manifold $(\overline{X},\omega).$ Then, Hutchings-Taubes prove a ``holomorphic curve axiom" for any $\omega$-compatible almost complex structure $J$ that agrees with symplectization-admissible almost complex structures $J_{\pm}$ on the symplectization ends. Their proof goes via adapting a ``Seiberg-Witten to Gromov" type degeneration that was needed by Taubes to show the equivalence of ECH and HM, to this context. Essentially, just as Taubes' proof that ECH = HM, they write down a family of perturbations to the four-dimensional Seiberg-Witten equations, so that when there is a solution counted by the Seiberg-Witten cobordism map, it degenerates after perturbation to a holomorphic building. Although in the Hutchings-Taubes case the symplectic form is assumed exact, it was observed by Hutchings (see e.g. \cite{Hutchings-blog}) that this is not essential. In the Chen context, one starts with a fibered symplectic cobordism between (fiberwise symplectic) mapping tori, and again attaches symplectization-like ends, analogously to the Hutchings-Taubes context, to get a non-compact symplectic manifold $(\overline{X},\omega)$; the main difference here is that the symplectic form on the ends in the Hutchings-Taubes context is the symplectization form for a contact form, whereas in the Chen context, the symplectic form is the symplectization form for the mapping torus. Just like in Hutchings-Taubes, Chen proves a ``holomorphic curve axiom" for any $\omega$-compatible almost complex structure $J$ that agrees with symplectization-admissible almost complex structures $J_{\pm}$ on the ends; this also essentially goes via a ``Seiberg-Witten to Gromov" type degeneration \end{remark} \begin{remark} \label{rmk:twisted} On the Seiberg-Witten side, the twisted theory corresponds to a version of the Floer homology, where, instead of taking the quotient of solutions by the full gauge group $\mathcal{G} = C^{\infty}(M,\S^1)$, one only takes the quotient by the subgroup $\mathcal{G}^0 \subset \mathcal{G}$ of gauge transformations in the connected component of the identity. This has an $H^1(Y)$ action, induced by the action via gauge transformations, which corresponds to the $H_2(Y)$ action on twisted PFH given by adding a homology class. As mentioned above, it was remarked by Taubes \cite{Taubes-ECH=SW-V}, Sec.\ 1, that the twisted invariant on the PFH/ECH side is canonical only up to a choice of element of $H_2(Y,\rho,\rho')$, where $\rho, \rho'$ are two reference cycles. Implicit in this assertion is that after a choice of reference cycle $R$, the isomorphism \eqref{eqn:chainchoice} satisfies a holomorphic curve axiom, for buildings $C$ satisfying \begin{equation} \label{eqn:cobordismequation} Z + R = [C] + Z'. \end{equation} This is the best way to think about \eqref{eqn:importantequation}: this corresponds to the case where our reference cycle is constant over $\rho$. Since \eqref{eqn:cobordismequation} is only implicit in Taubes, we also should emphasize that the full force of \eqref{eqn:cobordismequation} is not needed for our purposes. Indeed, all we need to know to prove Theorem~\ref{thm:PFHspec_initial_properties} is that the map \eqref{eqn:importantequation} satisfies the weaker axiom that if there is a nonzero coefficient of $\Psi_{H,K}(\alpha,Z)$ on $(\beta,Z')$, then there is an $I=0$ holomorphic building $C$ from $\alpha$ to $\beta$; this is proved in \cite{Chen}. With this weaker axiom, we can no longer assume \eqref{eqn:importantequation}. Still, however, most of the proof of Theorem~\ref{thm:PFHspec_initial_properties} goes through unchanged, except for two exceptions, which cancel each other out. The first difference is that \eqref{eqn:actionequation} in Lemma~\ref{lemm:important_identity_inequality} requires that \[ \int_C \omega = \int_{Z - Z'} \omega.\] The above equation does not hold unless we know that $[C] = Z - Z'$; and, without \eqref{eqn:importantequation}, we only know that $[C] = Z - Z' + x [\S^2]$ for some $x$. Thus, \eqref{eqn:actionequation} must be modified by subtracting $x$ from the right hand side, and so the identity asserted by Lemma~\ref{lemm:important_identity_inequality} must also be modified by subtracting $x$ from its right hand side. However, this change is balanced by the fact that as $I([C]) = 0$, we must also have $I(Z) - I(Z') = -x(2d+2)$, and so the left hand side of \eqref{eqn:lower_bound_action} must also be modified to $c_{d,k - x(2d+2)}(\varphi^1_H) - c_{d,k}(\varphi^1_K)$; since $c_{d,k - x(2d+2)}(\varphi^1_H) = c_{d,k}(\varphi^1_H) - x$, this precisely balances the required modification of the statement of Lemma~\ref{lemm:important_identity_inequality}. For more about the connection between the twisted theory and the relevant Seiberg-Witten Floer homology, we refer the reader to \cite{Taubes-ECH=SW-V} Sections 1 - 2. In \cite{Taubes-ECH=SW-V} Sections 1 - 2, Taubes is writing about twisted ECH; we have adapted what is written there to the PFH context, as suggested by \cite{Lee-Taubes}, Cor. 6.1. \end{remark} \section{$C^0$ continuity}\label{sec:C0_continuity} Here we prove Theorem~\ref{theo:C0_continuity}, using Theorem \ref{thm:PFHspec_initial_properties} from Section \ref{sec:pfhs2}. The central objects of Theorem ~\ref{theo:C0_continuity} are the maps $c_d:\Diff_c(\D^2, \omega)\to\R$. Remember from Section \ref{sec:PFH_spec_initial_properties} and Remark \ref{rem:PFHspec-disc} that these maps are defined from the spectral invariants $c_d:\cal S\to\R$, by identifying $\Diff_c(\D^2, \omega)$ with the group $\DiffS$ consisting of symplectic diffeomorphisms of $\S^2$, which are supported in the interior of the northern hemisphere $S^+$. In the present section, we directly work in the group $\DiffS$. More generally, given an open subset $U\subset \S^2$, we will denote by $\DiffU$ the set of all Hamiltonian diffeomorphisms compactly supported in an open subset $U$. Our proof is inspired by the proof of the $C^0$-continuity of barcodes (hence, of spectral invariants) arising from Hamiltonian Floer theory presented in \cite{LSV}. Other existing proofs of $C^0$-continuity of spectral invariants make use\footnote{The product is usually used to deduce continuity everywhere from continuity at $\id$. Without a product, we need another argument to prove continuity in the complement of the identity.} of the product structure on Hamiltonian Floer homology. It might be possible to define a ``quantum product" on PFH, see \cite{Hutchings-Sullivan-Dehntwist}, however, at the time of the writing of this article such structures do not exist. Let $d$ be a positive integer. As in \cite{LSV}, we treat separately the $C^0$-continuity of $c_d$ at the identity and elsewhere. Theorem \ref{theo:C0_continuity} will be a consequence of the following two propositions. \begin{prop}\label{prop:C0-continuity-at-id} The map $c_d:\DiffS\to \R$ is continuous at $\id$ with respect to the $C^0$-topology on $\DiffS$. \end{prop} \begin{prop}\label{prop:C0-continuity-elsewhere} Every area-preserving homeomorphism $\eta\in\HomeoS$ with $\eta\neq \id$ admits a $C^0$-neighborhood $\cal V$ such that the restriction of $c_d$ to $\cal{V}\cap\DiffS$ is uniformly continuous with respect to the $C^0$-distance. \end{prop} This last proposition readily implies that any $\eta\in \HomeoS\setminus\{\id\}$ admits a $C^0$-neighborhood $\cal V$ to which $c_d$ extends continuously. In particular, it extends continuously at $\eta$. Since this holds for any such homeomorphism $\eta$, this shows together with Proposition \ref{prop:C0-continuity-at-id} that $c_d$ extends to a map $\HomeoS\to\R$ continuous with respect to $C^0$-topology, hence Theorem \ref{theo:C0_continuity}. Proposition \ref{prop:C0-continuity-elsewhere} can be rephrased as follows. Any homeomorphism $\eta\in\HomeoS$, $\eta\neq \id$, admits a neighborhood $\cal{V}$ in $\HomeoS$ such that for all $\eps>0$, there exists $\delta>0$ satisfying: \begin{align} \forall \phi,\psi\in \cal{V} \cap{\DiffS}, \text{ if } d_{C^0}(\phi,\psi)<\delta \text{ then } |c_d(\phi)-c_d(\psi)|<\eps. \label{eq:uniform-continuity} \end{align} \paragraph{The Hofer norm.} Our proofs will make intensive use of the Hofer norm for Hamiltonian diffeomorphisms. We now recall its definition and basic properties. We refer the reader to \cite{Polterovich2001} and the references therein for a general introduction to the material presented here. We have seen earlier in the paper the definition of the Hofer norm of a Hamiltonian on the sphere and the disc. On a general symplectic manifold, the {\bf Hofer norm} of a compactly supported Hamiltonian diffeomorphism $\phi$ is defined as \[\|\phi\|=\inf\{\|H\|_{(1,\infty)}\},\] where the infimum runs over all compactly supported Hamiltonians $H$ whose time-$1$ map is $\phi$. It satisfies a triangle inequality \[ \|\phi\circ\psi\|\leq\|\phi\|+\|\psi\|,\] for all Hamiltonian diffeomorphisms $\phi, \psi$, it is conjugation invariant and moreover, we have $\|\phi^{-1}\|=\|\phi\|$ for all Hamiltonian diffeomorphisms $\phi$. We remark, in connection with Section \ref{sec:infinitedistance}, that we can also define a metric, called {\bf Hofer's metric}, by $d_{\mathrm{Hofer}}(\phi,\psi) = \| \phi \psi^{-1} \|$; this is a nondegenerate, bi-invariant metric on the group of compactly supported Hamiltonian diffeomorphisms, but we will not say more about it here because we do not need it for the results in this paper; for more details, we refer the reader to \cite{Polterovich2001}. The displacement energy of a subset $A$ of the ambient symplectic manifold is by definition the quantity \[e(A) :=\inf\{\|\phi\|: \phi(\overline{A})\cap \overline{A}=\emptyset\}.\] On a surface, it is known that for a disjoint union of closed discs, with each disc having area $a$, and whose union covers less than half the area of the surface, the displacement energy is $a$. \medskip \noindent\textbf{Important note:} We will use the Hofer norm on the symplectic manifold $\S^2\setminus\{p_-\}$. Thus, all the Hamiltonians considered in this section will be compactly supported in the complement of the south pole $p_-$; in particular, they belong to $\mathcal{H}$. \medskip Note that the second item of Theorem \ref{thm:PFHspec_initial_properties} can be reformulated as \begin{equation} \label{eq:hofer-bound} |c_d(\psi)-c_d(\phi)|\leq d\cdot\|\psi^{-1}\circ\phi\|, \end{equation} for all Hamiltonian diffeomorphisms $\phi, \psi\in\Diff_{\S^2\setminus \{p_-\}}(\S^2,\omega)$. To see this, let $K$ be a Hamiltonian such that $\psi=\varphi_K^1$. For any Hamiltonian $H$ such that $\psi^{-1}\circ\phi=\varphi_H^1$, then $\phi=\varphi_{K\# H}^1$ by (\ref{eqn:somehams}), thus by Theorem \ref{thm:PFHspec_initial_properties} we obtain \begin{align*} |c_d(\psi)-c_d(\phi)|&=|c_d(\varphi_K^1)-c_d(\varphi_{K\# H}^1)|\\ &\leq d\cdot \|K-K\# H\|_{(1,\infty)}=d\cdot\|H\|_{(1,\infty)}. \end{align*} Inequality (\ref{eq:hofer-bound}) follows. \subsection{Continuity at the identity} We first prove Proposition \ref{prop:C0-continuity-at-id}. The case $d=1$ can be proved with the same proof as \cite{Sey13}, using the so-called $\eps$-shift technique. We will generalize this idea to make the proof work for all $d\geq 1$. Let us start our proof with a lemma. \begin{lemma}\label{lemma:eps-shift} Let $d\geq 1$ and let $F$ be a time-independent Hamiltonian, compactly supported in $\S^2\setminus\{p_-\}$ and $f=\phi_F^1$ the time-one map it generates. Assume that the next two conditions are satisfied: \begin{enumerate}[(a)] \item for all $k\in\{1,\dots, d\}$, the $k$-periodic points of $f$ are precisely the critical points of $F$; \item none of the critical points of $F$ are in the closure of $S^+$. \end{enumerate} Then, there exists $\delta>0$ such that $c_d(\phi\circ f)=c_d(f)$, for any $\phi\in\DiffS$ with $d_{C^0}(\phi,\id)<\delta$. \end{lemma} Postponing the proof of this lemma, we now explain how it implies Proposition \ref{prop:C0-continuity-at-id}. \begin{proof}[Proof of Proposition \ref{prop:C0-continuity-at-id}] Let $\eps>0$. Let $F$ be a function on $\S^2$ satisfying the assumptions of Lemma \ref{lemma:eps-shift} and assume furthermore that \[ \max F-\min F\leq \frac\eps{2d}.\] For instance a $C^2$-small function supported in the complement of $p_-$ all of whose critical points are in the southern hemisphere is appropriate. Let then $\delta$ be as provided by Lemma \ref{lemma:eps-shift}. Let $\phi\in\DiffS$ be such that $d_{C^0}(\phi,\id)<\delta$. Then, we have \[c_d(\phi\circ f)=c_d(f).\] % Using Inequality (\ref{eq:hofer-bound}) and the fact that $c_d(\id)=0$, we obtain \[|c_d(\phi\circ f)|=|c_d(f)-c_d(\id)|\leq d\,\|f\|.\] We then deduce from Inequality (\ref{eq:hofer-bound}): \begin{align*} |c_d(\phi)|\leq |c_d(\phi\circ f)| + d\|f\|\leq 2d\|f\|. \end{align*} Now, by definition of the Hofer norm, $\|f\|\leq \max F-\min F$. Thus we get \[c_d(\phi)\leq \eps.\] This show the $C^0$-continuity of $c_d$ at $\id$. \end{proof} We will now prove the lemma. \begin{proof}[Proof of Lemma \ref{lemma:eps-shift}] Let $F$ be as in the statement of the lemma and $f=\phi_F^1$. We want to prove that $c_d(f)$ remains unchanged when we $C^0$-perturb $f$ with a Hamiltonian diffeomorphism supported in the northern hemisphere. To obtain this, we will first prove that the entire spectrum remains unchanged under such perturbations. Let $k\in\{1,\dots, d\}$; we begin by showing that the set of $k$-periodic points is unchanged by these perturbations. By assumption, there exists $c>0$ such that \[d(f^k(x),x)>c\] for all $x$ in the closure of $S^+$. Now note that the diffeomorphism $(\phi f)^k$ converges to $f^k$ uniformly when $\phi$ tends uniformly to $\id$. Thus, there exists $\delta>0$ such for $d_{C^0}(\phi, \id)<\delta$, the inequality $d((\phi f)^k(x),x)>0$ holds for all $x$ in the closure of $S^+$. In other words, $\phi\circ f$ has no $k$-periodic points in the closure of $S^+$. Since $\phi$ coincides with the identity outside $S^+$, this implies that $\phi\circ f$ and $f$ have the same $k$-periodic points, which are in turn the critical points of $F$. For the rest of the proof we pick $\delta$ such that the above holds for all $k\in\{1,\dots, d\}$, and $\phi$ such that $d_{C^0}(\phi,\id)<\delta$. We next show that the actions of these $k$-periodic points, i.e. the critical points of $F$, agree when computed with respect to $f$ and $\phi\circ f$. To compute these actions, let $H$ be a Hamiltonian supported in $S^+$ such that $\varphi_H^1=\phi$. By (\ref{eqn:somehams}), the isotopy $\varphi_H^t\varphi_F^t$ (whose time one map is $\phi\circ f$) is generated by the Hamiltonian $H\# F(t,x)=H(t,x)+F((\varphi_H^t)^{-1}(x))$. Let $y$ be a critical point of $F$. Then, $\varphi_F^t(y)=y$ for all $t\in[0,1]$, and since $y\notin S^+$, we also have $\varphi_H^t(y)=y$ for all $t\in[0,1]$. Thus $y$ remains fixed along the whole isotopy. A capping of such an orbit is a trivial capping to which is attached $\ell[\S^2]$ for some $\ell\in\Z$. Also note that since $H$ is supported in $S^+$, \[H\# F(t,\varphi_H^t\varphi_F^t(y))=H(t,y)+F(y)=F(y).\] Applying Formula (\ref{eqn:actiondefn}) we obtain \begin{align*} \mathcal{A}_{H\# F}(y,\ell[\S^2])&=\int_0^1 H\# F(t,\varphi_H^t\varphi_F^t(y))dt +\ell \mathrm{Area}(\S^2).\\ &= F(y)+\ell\\ &= \mathcal{A}_F(y,\ell[\S^2]). \end{align*} This shows that $\spec(H\#F)=\spec(F)$. A similar argument shows that $\spec((H\#F)^k)=\spec(F^k)$ for all $k\in\{1,\dots, d\}$, hence $\spec_d(H\# F)=\spec_d(F)$. By (\ref{eqn:spec_time1}), we have proved \begin{equation} \spec_d(\phi\circ f)=\spec_d(f)\label{eq:eps-shift} \end{equation} for all $\phi\in\DiffS$ such that $d_{C^0}(\phi, \id)<\delta$. \medskip There remains to deduce $c_d(\phi\circ f)=c_d(f)$ from this equality of spectrums. Given $\phi\in\DiffS$ such that $d_{C^0}(\phi, \id)<\delta$, one can construct, using the Alexander isotopy, a Hamiltonian isotopy $(\varphi_K^t)_{t\in[0,1]}$ in $\DiffS$, such that $d_{C^0}(\varphi_K^s,\id)<\delta$ for all $s\in[0,1]$ and $\varphi_K^1=\phi$; we refer the reader to \cite[Lemma 3.2]{Sey13} for the details. Equation (\ref{eq:eps-shift}) then implies $\spec_d(\varphi_K^s\circ f)=\spec_d(f)$, for all $s\in [0,1]$. Now, by Theorem~\ref{thm:PFHspec_initial_properties} the function $s\mapsto c_d(\varphi_K^s\circ f)$ is continuous and takes its values in the measure $0$ subset $\spec_d(f)\subset \R$. As a consequence, it is constant. This shows $c_d(\phi\circ f)=c_d(f)$ and concludes our proof. \end{proof} \subsection{Continuity away from the identity} We now turn our attention to Proposition \ref{prop:C0-continuity-elsewhere}. We want to prove \eqref{eq:uniform-continuity}, or equivalently, that any $\eta\in\HomeoS$, $\eta\neq \id$, admits an open neighborhood $\mathcal{V}$ such that for all $\eps>0$, there exists $\delta>0$ satisfying \begin{align} \forall f\in\mathcal{V}& \cap \DiffS, \forall g\in\DiffS, \label{eq:uniform-conitnuity2}\\ &\text{ if }d_{C^0}(g, \id)<\delta \text{ then } |c_d(gf)-c_d(f)|<\eps.\nonumber \end{align} Our proof will follow from three lemmas, which we now introduce. To state the first, let us introduce some terminology. We will say that a diffeomorphism $f$ \textbf{$d$-displaces} a subset $U$ if the subsets $U$, $f(U)$, $\dots$, $f^d(U)$ are pairwise disjoint. Our first lemma states that for $g$ supported in an open subset $d$-displaced by $f$, an even stronger version of \eqref{eq:uniform-conitnuity2} holds. It is inspired by \cite[Lemma 3.2]{usher10}, which can be regarded as the analogue for the $d = 1$ case. \begin{lemma}\label{lemme:cd-under-displacement} Let $f\in\DiffS$ and $B$ be an open topological disc % whose closure is included in $\S^2\setminus\{p_-\}$ and which is $d$-displaced by $f$. Then, for all $\phi\in\DiffB$, we have $c_d(\phi\circ f)=c_d(f)$. \end{lemma} We will prove this lemma in Section \ref{sec:proof-of-lemmas}. To apply the previous lemma, we need there to exist an appropriate open disc $B$. The next lemma is the key ingredient for this. % \begin{lemma}\label{lemma:d-disjoint-iterates} Let $\eta\in \Homeo_c(\D,\omega)$ with $\eta\neq\id$. Then, there exists a point $x\in \D$ such that $x,\eta(x), \eta^2(x), \dots, \eta^d(x)$ are pairwise distinct points. \end{lemma} In particular, by the lemma, there exists an open topological disc $B$ whose closure is $d$-displaced by $\eta$. If we then let $\cal V$ be the $C^0$ open neighborhood of $\eta$ given by the set of all $f\in\HomeoS$ which $d$-displace the closure of the disc $B$, then by Lemma \ref{lemme:cd-under-displacement}, we have $c_d(\phi\circ f)=c_d(f)$ for all $\phi\in\DiffB$ and $f\in \cal V\cap\DiffS$. Now it turns out that every map $g$ which is sufficiently $C^0$ close to $\id$ is close in Hofer distance to an element in $\DiffB$. This is the content of the next lemma. \begin{lemma}\label{lemma:small-hofer} Let $B$ be an open topological disc whose closure is included in $\S^2\setminus\{p_-\}$. For all $\eps>0$, there exists $\delta>0$, such that for all $g\in\DiffS$ with $d_{C^0}(g,\id)<\delta$, there is $\phi\in\DiffB$ such that $\|\phi^{-1}g\|\leq\eps$. \end{lemma} We will prove Lemma \ref{lemma:small-hofer} at the end of Section \ref{sec:proof-of-lemmas}. Assuming this, we can achieve the proof of \eqref{eq:uniform-conitnuity2} and hence of Proposition \ref{prop:C0-continuity-elsewhere}, as we now explain. \begin{proof}[Proof of Proposition \ref{prop:C0-continuity-elsewhere}.] Let $\eps>0$ and let $\delta>0$ be as provided by Lemma \ref{lemma:small-hofer}. Let also $f\in \cal V\cap\DiffS$. Then, for all $g$ satisfying $d_{C^0}(g,\id)<\delta$, there exists $\phi\in \DiffB$ such that $\|\phi^{-1}g\|<\eps$. Thus, using Lemma~\ref{lemme:cd-under-displacement}, Hofer continuity \eqref{eq:hofer-bound} and the conjugation invariance of the Hofer norm, we have: \[|c_d(gf)-c_d(f)|=|c_d(g f)-c_d(\phi f)|\leq d\|f^{-1}\phi^{-1}g f\|=d\|\phi^{-1}g\|<d\eps.\] Since $\eps$ is arbitrary, this concludes the proof of \eqref{eq:uniform-conitnuity2} and of Proposition \ref{prop:C0-continuity-elsewhere}, modulo the proofs of the lemmas \ref{lemme:cd-under-displacement}, \ref{lemma:d-disjoint-iterates} and \ref{lemma:small-hofer}. \end{proof} \subsection{Proofs of Lemmas \ref{lemme:cd-under-displacement}, \ref{lemma:d-disjoint-iterates} and \ref{lemma:small-hofer}.}\label{sec:proof-of-lemmas} We now provide the proofs of the lemmas stated in the preceding section. We start with the simplest one. \begin{proof}[Proof of Lemma \ref{lemma:d-disjoint-iterates}] Let $\eta\in \Homeo_c(\D,\omega)$ with $\eta\neq\id$ and let $d$ be a positive integer. It is known (see \cite{Constantin-Kolev}) that for any positive integer $N$, $\eta^N\neq \id$. Thus, there exists a point $x\in\D$ such that $\eta^{d!}(x)\neq x$. For such a point, $x,\eta(x), \eta^2(x),\dots, \eta^d(x)$ are pairwise distinct. Indeed, otherwise, there would be integers $0\leq k<\ell\leq d$ such that $\eta^k(x)=\eta^\ell(x)$, and we would get $\eta^{\ell-k}(x)=x$, in contradiction with $\eta^{d!}(x)\neq x$ since $d!$ is evenly divided by $\ell - k$. \end{proof} We now turn our attention to Lemma \ref{lemme:cd-under-displacement}. \begin{proof}[Proof of Lemma \ref{lemme:cd-under-displacement}] Let $F$ be a Hamiltonian supported in $S^+$ with $\varphi_F^1 = f$ and let $G$ be a Hamiltonian supported in $B$ with $\varphi_G^1=\phi$. We will prove that for all $s\in[0,1]$, $\spec_d(\varphi_G^s f)=\spec_d(f)$. This implies, as in the proof of Lemma \ref{lemma:eps-shift}, that the map $s\mapsto c_d(\varphi_G^s f)$ is constant, hence the lemma. Let $s\in [0,1]$. We will first verify that the diffeomorphism $\varphi_G^s f$ admits the same $k$-periodic points as $ f$, for all $k\in\{1,\dots, d\}$. For all $\ell\in\{1,\dots, d\}$, we have $B\cap f^\ell(B)=\emptyset$ and $B\cap f^{-\ell}(B)=\emptyset$. It follows that $\varphi_G^s( f^{\ell}(B)) = f^\ell(B),$ for all $\ell\in\{-d,\dots, d\}$, hence \[(\varphi_G^s f)^k( f^{-\ell}(B)) = f^{k-\ell}(B), \quad \forall k\in\{1,\dots, d\}, \forall\ell\in\{0,\dots, d\}. \] Since $ f^{-\ell}(B)\cap f^{k-\ell}(B)=\emptyset$ for such $k, \ell$, this implies that $\varphi_G^s f$ has no $k$-periodic points with $1\leq k\leq d$ in $\bigcup_{\ell=0}^d f^{-\ell}(B)$. We now fix $k\in\{1,\dots, d\}$. If $x\notin \bigcup_{\ell=0}^d f^{-\ell}(B)$, then $(\varphi_G^s\circ f)^k(x)= f^k(x)$. As a consequence, the $k$-periodic points of $\varphi_G^s\circ f$ are exactly those of $ f$. We will now prove that the corresponding action values coincide as well. For that purpose it is convenient to use an isotopy generating the $(\varphi_G^s\circ f)^k$ obtained by concatenation of isotopies rather than composition. Namely, the map $(\varphi_G^s\circ f)^k$ is the time-1 map of the isotopy $\psi^t$ which at time $t\in [\frac{\ell}{2k},\frac{\ell+1}{2k}]$ for $\ell\in\{0,\dots, 2k-1\}$ is given by % \[\psi^t= \begin{cases} \varphi_F^{\rho(2kt-\ell)}\circ(\varphi_G^s\circ f)^{\frac{\ell}{2}},&\quadif \ell\text{ is even,}\\ \varphi_G^{s\rho(2kt-\ell)}\circ f\circ(\varphi_G^s\circ f)^{\frac{\ell-1}{2}} ,&\quadif \ell\text{ is odd.} \end{cases} \] Here, $\rho:[0,1]\to[0,1]$ is a non-decreasing smooth function which is equal to $0$ near $0$ and equal to $1$ near $1$. The role of the time-reparametrization $\rho$ is simply to make the isotopy smooth at the gluing times. This isotopy is generated by the Hamiltonian $K$ given by the formula \[K_t(x)= \begin{cases} 2k\rho'(2kt-\ell)F_{\rho(2kt-\ell)}(x),&\quadif \ell\text{ is even,}\\ 2ks\rho'(2kt-\ell)G_{s\rho(2kt-\ell)}(x),&\quadif \ell\text{ is odd,} \end{cases} \] % for $\ell\in\{0,\dots, 2k-1\}$ and $t\in [\frac{\ell}{2k},\frac{\ell+1}{2k}]$. We will compute the spectrum of $\varphi_G^s\circ f$ with the help of this particular Hamiltonian. The action of a capped 1-periodic orbit $(z,u)$ of $K$, with $z=\varphi_K^t(x)$, is given by \begin{align*} \cal A_K(z,u) &= \int_{\D^2}u^*\omega + \int_{0}^{1}K_{t}(\psi^t(x))\,dt \\ & \int_{\D^2}u^*\omega +\sum_{\ell=0}^{2k-1}\int_{\frac{\ell}{2k}}^{\frac{\ell+1}{2k}}K_{t}(\psi^t(x))\,dt. \end{align*} After suitable change of variables, we obtain: \begin{align*} \cal A_K(z,u) = \int_{\D^2}u^*\omega &+\sum_{j=0}^{k-1}\left( \int_{0}^{1}F_{t}(\varphi_F^{t}\circ (\varphi_G^s\circ f)^j(x))\,dt \right.\\ & +\left.\int_{0}^{1} sG_{st}(\varphi_G^{st}\circ f\circ (\varphi_G^s\circ f)^j(x))\,dt \right). \end{align*} Since we showed above that $\varphi^s_G \circ f$ has no $k$-periodic points in $ f^{- 1}(B)$, we know that $ f \circ (\varphi_G^s \circ f)^j(x)$ does not belong to $B$, hence to the support of $G$. It follows that the integrand for the third integral above has to vanish and the integrand for the second integral above can be simplified, so that we get \begin{align*} \cal A_K(z,u)=\int_{\D^2}u^*\omega +\sum_{j=0}^{k-1}\int_{0}^{1} F_t(\varphi_F^t\circ f^j(x))\,dt. \end{align*} We see that this action does not depend on $s$. As a consequence, we get that $\spec_d(\varphi_G^s f)=\spec_d( f)$ for all $s\in[0,1]$. \end{proof} There remains to prove Lemma \ref{lemma:small-hofer}. Its proof will consist of two steps. First, by Lemma \ref{lemma:square-fragmentation} below, a diffeomorphism which is sufficiently $C^0$-close to identity can be appropriately fragmented into maps supported in balls of small area. Second, we observe that moving theses maps with small support into $B$ can be achieved with small Hofer norm; this is the content of Lemma \ref{lemma:moving-balls} below. Before starting the proof of Lemma \ref{lemma:small-hofer}, let us state the first of these two lemmas. \begin{lemma}[\cite{LSV}, Lemma 47]\label{lemma:square-fragmentation} Let $\omega_0$ denote the standard area form on $\R^2$. Let $m$ be a positive integer and $\rho$ a positive real number. For % $i=1,\dots, m$, denote by $U_i$ the open rectangle $(0,1)\times(\frac{i-1}{m},\frac {i}{m})$. Then, there exists $\delta>0$, such that for every $g\in\Diff_c((0,1)\times(0,1),\omega_0)$ with $d_{C^0}(g,\id)<\delta$, there exist $g_1\in\Diff_c(U_1,\omega_0)$, $\dots$, $g_m\in \Diff_c(U_m,\omega_0)$ and $\theta\in \Diff_c((0,1)\times(0,1),\omega_0)$ supported in a disjoint union of topological discs whose total area is less than $\rho$, such that $g=g_1\circ\dots\circ g_m\circ\theta$. \end{lemma} We can now give the promised proof. \begin{proof}[Proof of Lemma \ref{lemma:small-hofer}] Let $\eps>0$. We pick an integer $N$ satisfying $\frac1{2N}<\textrm{area}(B)$, $m$ a positive multiple of $N$ such that $2\frac{N+1}{m}<\eps$, and $\rho=\frac1{m}$. Let $\delta$ be as provided by Lemma \ref{lemma:square-fragmentation} and let $g\in\DiffS$ be such that $d_{C^0}(g,\id)<\delta$. The map $g$ admits a fragmentation into the form $g=g_1\circ\dots\circ g_m\circ\theta$, with all the $g_i$ supported in pairwise disjoint topological discs $U_i$ of area $\frac{1}{2m}$ and $\theta$ supported in a disjoint union of discs of total area less than $\frac1{2m}$ (here, the factor $\frac12$ comes from the fact that the northern hemisphere $S^+$ has area $\frac12$, while Lemma \ref{lemma:square-fragmentation} is stated on $(0,1)\times(0,1)$ which has area $1$). % For $j=1,\dots, N$, denote by $f_j$ the composition $f_j:=\prod_{i=0}^{\tfrac mN-1}g_{j+iN}$. Also write $f_{N+1}:=\theta$, so that, noting that the $g_i$ commute, we have the following formula \[g=\prod_{j=1}^{N+1}f_j.\] Each $f_j$ for $j=1,\dots, N$ is supported in $V_j=\bigcup_{i=0}^{\tfrac mN-1}U_{j+iN}$ whose area is $\frac{1}{2N}<\textrm{area}(B)$. Note that $V_j$ is a disjoint union of discs, each of area $\frac1{2m}$. By assumption, the support of $f_{N+1}=\theta$, which we denote by $V_{N+1}$, is also included in a disjoint union of discs of total area smaller than $\frac1{2m}$. Let us now state our second lemma, whose proof we postpone to the end of this section. \begin{lemma}\label{lemma:moving-balls} Let $a\in(0,\frac12)$, let $B_1, \dots, B_k\subset S^+$ be pairwise disjoint open topological discs each of area smaller than $a$, and $B\subset S^+$ a topological disc with $\mathrm{area}(B)> ka$. Then, there exists a Hamiltonian diffeomorphism $h\in \Diff_{\S^2\setminus\{p_-\}}(\S^2,\omega)$ which maps $B_1\cup\dots\cup B_k$ into $B$ and satisfies $\|h\|\leq 2a$. \end{lemma} In our situation, this lemma implies that for any $j=1, \dots, N+1$, there exists a Hamiltonian diffeomorphism $h_j\in \Diff_{\S^2\setminus\{p_-\}}(\S^2,\omega)$, such that $h_j(V_j)\subset B$ and $\|h_j\|\leq \frac1m$. Consider the diffeomorphism \[\phi=\prod_{j=1}^{N+1}h_j\circ f_j\circ h_j^{-1}.\] By construction, $\phi$ is supported in $B$. We will show that $\|\phi^{-1}g\|<\eps$, which will achieve the proof of Lemma \ref{lemma:small-hofer}. To prove that $\|\phi^{-1}g\|<\eps$, let us introduce $\tilde{g}_k=f_1\circ f_2\circ \dots \circ f_k$ and $\phi_k=(h_1\circ f_1\circ h_1^{-1})\circ (h_2\circ f_2\circ h_2^{-1})\circ\dots\circ (h_k\circ f_k\circ h_k^{-1})$, for $k=1,\dots, N+1$. In particular, $\tilde{g}_{N+1}=g$ and $\phi_{N+1}=\phi$. Then, for all $k=1,\dots, N$, we have \begin{align*} \phi_{k+1}^{-1}\circ \tilde{g}_{k+1}&= h_{k+1}\circ f_{k+1}^{-1}\circ h_{k+1}^{-1}\circ(\phi_{k}^{-1}\circ \tilde{g}_{k})\circ f_{k+1}\\ &= h_{k+1}\circ (f_{k+1}^{-1}\circ h_{k+1}^{-1}\circ f_{k+1})\circ (f_{k+1}^{-1}\circ(\phi_{k}^{-1}\circ \tilde{g}_{k})\circ f_{k+1}) \end{align*} Thus, by the triangle inequality and the conjugation invariance, \begin{align*} \|\phi_{k+1}^{-1}\circ \tilde{g}_{k+1}\|&\leq \|h_{k+1}\|+\|f_{k+1}^{-1}\circ h_{k+1}^{-1}\circ f_{k+1}\|+ \|f_{k+1}^{-1}\circ(\phi_{k}^{-1}\circ \tilde{g}_{k})\circ f_{k+1}\|\\ &\leq \|h_{k+1}\|+\|h_{k+1}^{-1}\|+ \|\phi_{k}^{-1}\circ \tilde{g}_{k}\|\\ &\leq \frac2m+ \|\phi_{k}^{-1}\circ \tilde{g}_{k}\|. \end{align*} Hence, by induction, $\| \phi_i^{-1} \circ \tilde{g}_i \| \le \frac{2i}{m},$ and so this yields, as wished: \[\|\phi^{-1}\circ g\|\leq 2\frac{N+1}m<\eps.\] \end{proof} The last remaining proof is now the following. \begin{proof}[Proof of Lemma \ref{lemma:moving-balls}] Denote $U:=B_1\cup\dots\cup B_k$. Since $U$ has smaller area than $B$, there exists a Hamiltonian diffeomorphism $\psi\in \Diff_{S^+}(\S^2,\omega)$ such that $\psi(U)\subset B$. The Hofer norm of $\psi$ may not be small, but we will replace $\psi$ with an appropriate commutator of $\psi$ whose Hofer norm will be controlled. Since the displacement energy of $U$ is smaller than $a<\frac12$, there exists a Hamiltonian diffeomorphism $\ell\in\Diff_{\S^2\setminus\{p_-\}}(\S^2,\omega)$ such that $\ell(\overline{U})\cap \overline{U}=\emptyset$ and $\|\ell\|\leq a$. Since $\ell(U)$ has area smaller than $\frac12$, there exists $\chi\in \Diff_{\S^2\setminus\{p_-\}}(\S^2,\omega)$ which fixes $U$ and such that $\chi(\ell(U))\cap S^+=\emptyset$, which in particular implies that $\chi(\ell(U))\cap \supp(\psi)=\emptyset$. Then, $\ell'=\chi\circ\ell\circ\chi^{-1}$ satisfies the following properties: \begin{enumerate}[(i)] \item $\ell'(U)\cap U=\emptyset$, \item $\|\ell'\|\leq a$ , \item $\ell'(U)\cap \mathrm{supp}(\psi)=\emptyset$. \end{enumerate} To prove Property (i), start from $\ell(U)\cap U=\emptyset$ and compose with $\chi$, to get $\chi\circ \ell (U)\cap \chi(U)=\emptyset$. Since $U=\chi^{-1}(U)$, we obtain $\chi\circ\ell\circ\chi^{-1}(U)\cap U=\emptyset$, hence Property (i). Property (ii) follows from the conjugation invariance of the Hofer norm. Property (iii) is a consequence of Property (i), since $\chi$ fixes $U.$ Now, set $h:=\psi\circ\ell'^{-1}\circ\psi^{-1}\circ\ell'$. By Property (iii), $\ell'^{-1}\circ\psi^{-1}\circ\ell'(U)=U$. Thus, $h(U)=\psi(U)\subset B$. Moreover, \[\|h\|\leq \|\psi\circ\ell'^{-1}\circ\psi^{-1}\|+\|\ell'\|=2\|\ell'\|\leq 2a.\] This concludes our proof. \end{proof} \section{The periodic Floer homology of monotone twists} \label{sec:PFH_monotone_twist} The goal of this section and the next is to prove Theorem~\ref{theo:PFHspec_Calabi_property} which establishes the Calabi property for monotone twist maps of the disc which were introduced in Section \ref{sec:calabi_inf_twist} . Recall from Remark \ref{rem:PFHspec-disc} that we define PFH spectral invariants for maps of the disc by identifying $\Diff_c(\D, \omega)$ with maps of the sphere supported in the northern hemisphere $S^+$. In particular, we will view any monotone twist as an element of the set $\mathcal{S}$ appearing in Theorem \ref{thm:PFHspec_initial_properties}. Theorem~\ref{theo:PFHspec_Calabi_property} will follow from the following theorem for the invariants $c_{d,k}$, which, as alluded to in the introduction, was originally conjectured in greater generality by Hutchings \cite{Hutchings_unpublished}. \begin{theo} \label{thm:s2case} Let $(d,k)$ be any sequence, with $k = d$ mod $2$, and $d$ tending to infinity. Then, for any positive monotone twist $\varphi$ we have: \begin{equation} \label{eqn:calabiconjecture} \Cal(\varphi) = \lim_{d \to \infty} \left( \frac{ c_{d,k}(\varphi)}{d} - \frac{k}{2(d^2 + d)} \right). \end{equation} \end{theo} A first observation, concerning Equation \eqref{eqn:calabiconjecture}, which is also due to Hutchings, is that it suffices to establish \eqref{eqn:calabiconjecture} for a single such sequence $(d,k_d)$ with $d = 1, 2, \ldots$ ranging over the positive integers. Indeed, % for $d$-nondegenerate $\varphi$, there is an automorphism of the twisted PFH chain complex given by \[ (\alpha,Z) \mapsto (\alpha, Z + [S^2]),\] where $[S^2]$ denotes the class of the sphere. This increases the grading by $2d + 2$, by Formula \eqref{eqn:index_ambiguity}. It also increases the action by $1$. So, we have \begin{equation} \label{eqn:auto} c_{d,k + 2d + 2} = c_{d,k} + 1 \end{equation} for all $\varphi$. Now, the right hand side of Equation \eqref{eqn:calabiconjecture} is invariant under increasing the numerator of the first fraction by one, and increasing the numerator of the second fraction by $2d + 2$. Moreover, as a corollary of Theorem \ref{theo:spec_computation}, we will obtain \begin{equation} \label{eqn:monotonicink} c_{d,k} \leq c_{d,k'}, \end{equation} when $k' \geq k$, with $k=k' = d$ mod $2$ and $\varphi$ a positive monotone twist\footnote{For more general $\varphi$, \eqref{eqn:monotonicink} can still be established, by using the PFH ``$U$-map", but we will not need this in the present work.}; see Remark \ref{rem:monotonicink}. Thus, given an arbitrary sequence $(\tilde{d},\tilde{k})$, we can assume by the above analysis that $\tilde{k}$ is within $2d+2$ of $k_{\tilde{d}}$, and $| c_{\tilde{d},\tilde{k}} - c_{\tilde{d},k_{\tilde{d}}}| \leq 1$; the limit on the right hand side of \eqref{eqn:calabiconjecture} is then the same for $c_{\tilde{d},k_{\tilde{d}}}$ and $c_{\tilde{d},\tilde{k}}$. \medskip The remainder of Section \ref{sec:PFH_monotone_twist} is dedicated to describing a combinatorial model of $\widetilde{PFH}(\varphi^1_H, d)$ where the Hamiltonian $H$ belongs to a certain class of Hamiltonians $\mathcal{D}$ with the following property: The set $\mathcal{D}$ is contained in $\mathcal{H}$, which we introduced in Section \ref{sec:PFH_spec_initial_properties}, and any Hamiltonian in $\mathcal{H}$ whose time-1 map is a monotone twist can be approximated, in the Hofer norm $\Vert \cdot \Vert_{(1, \infty)}$, arbitrarily well by Hamiltonians in $\mathcal D$. We will use this combinatorial model in Section \ref{sec:calabi_for_montone_twists} to prove Theorem \ref{thm:s2case}. The main result of Section \ref{sec:PFH_monotone_twist} is Proposition \ref{prop:model} which describes this combinatorial model. While the current section is the longest section of our paper, it might help the reader to note that our model is inspired by an extensive literature. In particular, our combinatorial model is inspired by similar combinatorial models that have been developed for computing the PFH of a Dehn twist \cite{Hutchings-Sullivan-Dehntwist}, the ECH of $T^3$ \cite{Hutchings-Sullivan-T3}, and the ECH of contact toric manifolds \cite{Choi}, and our methods in this section are inspired by the techniques used to establish these combinatorial models; we should also mention the appendix of \cite{Hutchings_beyond}, which was very influential to our thinking, and which we say more about later in this section. The index calculations in \cite{C-CG-F-H-R} were also useful. For the benefit of the reader, we will frequently point out the parallels to all this literature more explicitly below. \subsection{Perturbations of rotation invariant Hamiltonian flows} \label{sec:ham} Here, and through the end of Section \ref{sec:PFH_monotone_twist}, we consider Hamiltonian flows on $(\S^2, \omega=\frac1{4\pi}d\theta\wedge dz)$, for an autonomous Hamiltonian \[ H = \frac{1}{2} h(z),\] where $h$ is some function of $z$. We have \begin{equation} \label{eqn:Hvf} X_H = 2 \pi h'(z) \partial_{\theta}. \end{equation} Hence, \[ \varphi^1_H(\theta,z) = (\theta + 2 \pi h'(z), z) \] We further restrict $h$ to satisfy \[ h' > 0, h'' > 0, h(-1) = 0.\] Furthermore, we demand that $h'(-1), h'(1)$ are irrational numbers satisfying $h'(-1) \leq \frac{\varepsilon_0}{d}$ and $\lceil h'(1) \rceil - h'(1) \leq \frac{\varepsilon_0}{d}$, where $\varepsilon_0$ is a small positive number and $\lceil \cdot \rceil$ denotes the ceiling function. Let $\mathcal{D}$ denote the set of Hamiltonians $H$ that satisfy all of these conditions and observe that $\mathcal D \subset \mathcal H$ where $\mathcal H$ was defined in Section \ref{sec:PFH_spec_initial_properties}. As a consequence of Theorem \ref{thm:PFHspec_initial_properties}, we have well-defined PFH spectral invariants $c_{d,k}(\varphi^1_H)$ for all $H \in \mathcal{D}$. \medskip The periodic orbits of $\varphi^1_H$ are then as follows: \begin{enumerate} \item There are elliptic orbits $p_+$ and $p_-$, corresponding to the north and south poles, respectively. \item For each $p/q$ in lowest terms such that $h' = p/q$ is rational, there is a circle of periodic orbits, all of which have period $q$. \end{enumerate} These circles of periodic orbits are familiar from Morse-Bott theory, and are sometimes referred to as ``Morse-Bott circles". There is also a standard $\varphi^1_H$-admissible\footnote{This means that the almost complex structure is compatible with the standard SHS on the mapping torus for $\varphi^1_H$.} % almost complex structure $J_{std}$ respecting this symmetry; its action on $\xi = T(\S^2 \times \lbrace pt \rbrace) = T\S^2$ is given by the standard almost complex structure on $\S^2$. As is familiar in this context (see \cite[Section 3.1]{Hutchings-Sullivan-Dehntwist}), we can perform a $C^2$-small perturbation of $H$, to split such a circle corresponding to the locus where $h' = p/q$ into one elliptic and one hyperbolic periodic points, such that the elliptic one $e_{p,q}$ has slightly negative monodromy angle, and the eigenvalues for the hyperbolic one $h_{p,q}$ are positive. Furthermore, the $C^2$-small perturbation can be taken to be supported in an arbitrarily small neighborhood of the circle where $h' = p/q$. More precisely, given a $\varphi^1_H$ such as above, for any positive $d$ and arbitrarily small $\eps > 0$, we can find another area-preserving diffeomorphism $\varphi_0$ of $\S^2$, which we call a {\bf nice perturbation} of $\varphi^1_H$, such that: \begin{enumerate} \item The only periodic points of $\varphi_0$ which are of degree at most $d$ are $p_+, p_-$, and the orbits $e_{p,q}$ and $h_{p,q}$ from above such that $q \leq d.$ Furthermore, all of these orbits are non-degenerate. \item The eigenvalues of the linearized return map for $e_{p,q}$ are within $\eps$ of $1$. \item $\varphi_0(\theta,z) = \varphi^1_H(\theta,z)$ as long as $z$ is not within $\eps$ of a value such that $h' = p/q$ where $ q \leq d$. \end{enumerate} Observe that we can pick a time-dependent Hamiltonian $\tilde H $ such that $\varphi^1_{\tilde H} = \varphi_0$ and $\tilde H = H$ as long as $z$ is not within $\eps$ of a value such that $h' = p/q$ with $q \leq d$. It is also familiar from the work of Hutchings-Sullivan, see \cite{Hutchings-Sullivan-Dehntwist}, Lemma A.1, that we can choose our perturbation of $\varphi^1_H$ such that we can assume the following: \begin{enumerate} \item[4.] $\varphi_0$ is chosen so that ``Double Rounding", defined in Section \ref{sec:comb} below, can not occur for generic $J$ close to $J_{std}$. \end{enumerate} Later, it will be clear why it simplifies the analysis to rule out Double Rounding. \subsection{The combinatorial model} \label{sec:comb} We now aim to describe the promised combinatorial model of $\widetilde{PFH}$ for the Hamiltonians described in the previous section. Fix $d \in \N$ and $\varphi^1_H$, where $H \in \mathcal{D}$, for the remainder of Section \ref{sec:PFH_monotone_twist}. To begin, define a {\bf concave lattice path} to be a piecewise linear, continuous path $P$, in the $xy$-plane, such that $P$ starts and ends on integer lattice points, its starting point is on the $y$--axis, the nonsmooth points of $P$ are also at integer lattice points, and $P$ is concave, in the sense that it always lies above any of the tangent lines at its smooth points. Lastly, every edge of $P$ is labeled by either $e$ or $h$. Below, we will associate one such lattice path to every $\widetilde{PFH}$ generator $(\alpha, Z)$. Let $\alpha = \{(\alpha_i, m_i)\}$ be an orbit set of degree $d$, for a nice perturbation $\varphi_0$ of $\varphi^1_H$. First of all, recall that the simple Reeb orbits for $Y_{\varphi_0}$, with degree no more than $d$, are as follows : \begin{enumerate} \item The Reeb orbits $\gamma_{\pm}$ corresponding to $p_{\pm}$. \item For each $z$ such that $h'(z) = p/q$ in lowest terms, where $q \leq d$, there are Reeb orbits of degree $q$ corresponding to the periodic points $e_{p,q}$ and $h_{p,q}$, that we will also denote by $e_{p,q}$ and $h_{p,q}$. \end{enumerate} \begin{figure}[h!] \centering \def1\textwidth{1.0\textwidth} \input{lattice-path1.pdf_tex} \caption{The lattice path $P_{\alpha,Z}$ for $\alpha=\{(h_{1/3},1),(e_{2/3},1),(h_{2/3},1),(e_{4/3},1),(e_2,2)\}$, $Z=Z_\alpha-3[\S^2]$, and $P_{\alpha',Z'}$ for $\alpha'=\{(\gamma_-,3),(e_{2/9},1),(h_{2/3},1),(\gamma_+,1)\}$, $Z'=Z_\alpha+4[\S^2]$ (assuming $\lceil h'(1)\rceil=4$).}\label{fig:lattice-path1} \end{figure} We will now associate to the orbit set $\alpha = \{(\alpha_i, m_i)\}$ a concave lattice path $P_\alpha$ whose starting point we require to be $(0,0)$. If $(\gamma_-, m_-) \in \alpha$, we set $v_- = m_- (1,0)$ and label it by $e$. If $(\gamma_+, m_+) \in \alpha$ we set $v_+ = m_+ (1, \lceil h'(1) \rceil )$ and label it by $e$. Next, consider the orbits in $\alpha$ corresponding to $z=z_{p,q}$ such that $h'(z) = p/q$; note that there are at most two such entries in $\alpha$: one corresponding to $e_{p/q}$ and another corresponding to $h_{p/q}$. To these entries, we associate the labeled vector $v_{p,q} = m_{p,q} (q,p)$, where $m_{p,q}$ is the sum of multiplicities of $e_{p/q}$ and $h_{p/q}$; the vector is labeled $h$ if $(h_{p/q},1) \in \alpha$, and $e$ otherwise. (For motivation, note that by the conditions on the PFH chain complex, an $m_i$ corresponding to a hyperbolic orbit must equal $1$.) To build the concave lattice path $P_{\alpha}$ from all of the data in $\alpha$, we simply concatenate the vectors $v_-, v_{p,q}, v_+$ into a concave lattice path. Note that there is a unique way to do this: the path must begin with $v_-$, it must end with $v_+$, and the vectors $v_{p/q}$ must be concatenated in increasing order with respect to the ratios $p/q$. Now, given a chain complex generator $(\alpha, Z)$ for $\widetilde{PFC}$, we define an assignation \[ (\alpha,Z) \mapsto P_{\alpha,Z}\] which associates a concave lattice path $P_{\alpha,Z}$ to the generator $(\alpha,Z)$. More specifically, when $Z = Z_\alpha$, where $Z_{\alpha}$ is the class from Lemma~\ref{lemma:action-action}, we define $P_{\alpha, Z_{\alpha}}$ to be the concave lattice path $P_\alpha$. Since $H_2(Y_{\varphi}) = \mathbb{Z}$, generated by the class of $\S^2 \times \lbrace \text{pt} \rbrace$, for any other $(\alpha,Z)$, we have $Z = Z_{\alpha} + y [\S^2]$. We then define $P_{\alpha,Z}$ to be $P_{\alpha}$ shifted by the vector $(0,y)$. Figure \ref{fig:lattice-path1} shows two examples of such concave lattice paths. \bigskip We now state some of the key properties of the above assignation \[ (\alpha,Z) \mapsto P_{\alpha,Z}.\] \noindent {\bf Degree: } Note that since we are fixing the degree of $\alpha$ to be $d$, the horizontal displacement of $P_{\alpha, Z}$ must be $d$; we therefore call the horizontal displacement of a concave lattice path its {\bf degree}. Clearly, the degree of $P_{\alpha, Z}$ agrees with the degree of the $\widetilde{PFH}$ generator $(\alpha, Z)$. \medskip \noindent {\bf Action:} Define the {\bf action} $\mathcal{A}(P_{\alpha,Z})$ as follows. We first define the actions of the edges of $P_{\alpha,Z}$ by the formulae: \begin{gather*} \mathcal{A}(v_-) = 0, \; \mathcal{A}(v_+) = m_+\frac{h(1)}2, \\ \mathcal{A}(v_{p,q}) = \frac{m_{p,q}}{2} ( p (1-z_{p,q}) + q h(z_{p,q})), \end{gather*} where $v_-, v_+$, and $v_{p,q}$ are as above. % We then define the action of $P_{\alpha,z}$ to be \begin{equation} \label{eq:action_formula} \mathcal{A}(P_{\alpha,Z}) = y + \mathcal{A}(v_+) + \sum_{v_{p,q}} \mathcal{A}(v_{p,q}), \end{equation} where $y$ is such that $P(\alpha,Z)$ begins at $(0,y)$. We claim that by picking the nice perturbation $\varphi_0$ to be sufficiently close to $\varphi^1_H$ we can arrange for $\mathcal{A}(\alpha, Z)$ to be as close to $\mathcal{A}(P_{\alpha, Z})$ as we wish. To show this it is sufficient to prove it when $\alpha$ is a simple Reeb orbit and $Z=Z_\alpha$, where $Z_\alpha$ is the relative class constructed in the proof of Lemma \ref{lemma:action-action}. We have to consider the following three cases: \begin{itemize} \item If $\alpha = \gamma_-$, then $\mathcal{A}(\alpha, Z_\alpha) = 0$, by Equation \eqref{eqn:action_Z_alpha}, which coincides with $\mathcal{A}(1,0)$. \item If $\alpha = \gamma_+$, then $\mathcal{A}(\alpha, Z_\alpha) = \frac{h(1)}{2}$, by Equation \eqref{eqn:action_Z_alpha}, which coincides with $\mathcal{A}(1, \lceil h'(1) \rceil)$. Note that, in Equation \eqref{eqn:action_Z_alpha}, the term $\int_{\D^2} u_\alpha^* \omega$ is zero. \item The remaining case is when $\alpha = e_{p,q}$ or $h_{p,q}$; here, it is sufficient to show that the action of the Reeb orbits at $z_{p,q}$, for the unperturbed diffeomorphism $\varphi^1_H,$ is exactly the quantity $\frac{1}{2} \left( p (1-z_{p,q}) + q h(z_{p,q})\right)$. This follows from Equation \eqref{eqn:action_Z_alpha_2}: the term $\int_{\D^2} u_\alpha^* \omega$ is exactly $\frac{1}{2} p(1 - z_{p,q})$ and the term $\int_0^q H_t(\varphi^t_H(q))dt$ is exactly $\frac{1}{2} q h(z_{p,q})$. \end{itemize} \medskip \noindent {\bf Index:} Next, we associate an {\bf index} to a concave lattice path $P$ which begins at a point $(0, y)$, on the $y$--axis, and has degree $d$. First, we form (possibly empty) regions $R_{\pm}$, where $R_-$ is the closed region bounded by the $x$-axis, the $y$-axis, and the part of $P$ below the $x$-axis, while $R_+$ is the closed region bounded by the $x$-axis, the line $x = d$, and the part of $P$ above the $x$-axis. Let $j_+$ denote the number of lattice points in the region $R_+$, not including lattice points on $P$, and let $j_-$ denote the number of lattice points in the region $R_-$, not including the lattice points on the $x$--axis; see Figure \ref{fig:combinatorial-index} below. We now define \begin{equation}\label{eqn:index_path_unlabelled} j(P) := j_+(P) - j_-(P). \end{equation} This definition of $j$ is such that if one shifts $P$ vertically by $1$, then $j(P)$ increases by $d+1$ % Given a path $P_{\alpha,Z}$, associated to a $\widetilde{PFH}$ generator $(\alpha,Z)$, we define its index by \begin{equation}\label{eqn:index_path_labelled} I(P_{\alpha,Z}) := 2 j(P_{\alpha,Z}) -d + h, \end{equation} where $h$ denotes the number of edges in $P_{\alpha, Z}$ labeled by $h$. See Figure \ref{fig:combinatorial-index} for an example of computation of this combinatorial index. We will show in Section \ref{sec:index} that $I(P_{\alpha,Z})$ coincides with the PFH index of $I(\alpha,Z)$ as defined in Equation \eqref{eqn:twistedgrading}. \begin{figure}[h!] \centering \def1\textwidth{0.5\textwidth} \input{combinatorial-index.pdf_tex} \caption{The lattice points included in the count of $j(P)$ are circled. On this example, $j_+(P)=6$, $j_-(P)=5$, $d=6$, $h=1$, thus $j(P)=1$ and $I(P)=-3$.}\label{fig:combinatorial-index} \end{figure} \medskip \noindent {\bf Corner rounding and the differential:} Lastly, we define a combinatorial process which corresponds to the PFH differential. Let $P_{\beta}$ be a concave lattice path of degree $d$ which begins on the $y$--axis. Then, if we attach vertical rays to the beginning and end of $P_{\beta}$, in the positive $y$ direction, we obtain a closed convex subset $R_{\beta}$ of the plane; see Figure \ref{fig:rounding-corner1}. For any given corner of $P_{\beta}$, where we include the initial and final endpoints of $P_\beta$ as corners, we can define a {\bf corner rounding} operation by removing this corner, taking the convex hull of the remaining integer lattice points in $R_{\beta}$, and taking the lower boundary of this region, namely the part of the boundary that does not consist of vertical lines. Note that the newly obtained path is of degree $d$. \begin{figure}[h!] \centering \def1\textwidth{0.4\textwidth} \input{rounding-corner1.pdf_tex} \caption{The region $R_\beta$.}\label{fig:rounding-corner1} \end{figure} We now say that another concave lattice path $P_{\alpha}$ is obtained from $P_{\beta}$ by {\bf rounding a corner and locally losing one $h$}, if $P_{\alpha}$ is obtained from $P_{\beta}$ by a corner rounding such that the following conditions are satisfied; see the examples in Figure \ref{fig:rounding-corner2}: \begin{enumerate}[(i)] \item Let $k$ denote the number of edges in $P_\beta$, with an endpoint at the rounded corner, which are labeled $h$. We require that $k >0$; so $k=1$ or $k=2$. \item Of the new edges in $P_\alpha$, created by the corner rounding operation, exactly $k-1$ are labelled $h$. % \end{enumerate} \begin{figure}[h!] \centering \def1\textwidth{0.8\textwidth} \input{rounding-corner2.pdf_tex} \caption{Some examples for the ``rounding corner and locally losing one $h$'' operation. The path $P_\beta$ is in black, the new edges in $P_\alpha$ are in grey. The rounded corner is circled. We only label the relevant edges.}\label{fig:rounding-corner2} \end{figure} The notions introduced above and the proposition below give a complete combinatorial interpretation of the PFH chain complex: \begin{prop} \label{prop:model} Fix $d > 0$ and let $\varphi_0$ be a nice perturbation of $\varphi^1_H$, where $H \in \mathcal{D}$. Then, for generic $\varphi_0$-admissible almost complex structure $J$ close to $J_{std}$, there is an assignation \[ (\alpha,Z) \mapsto P_{\alpha,Z} \] with the following properties: \begin{enumerate} \item $\mathcal{A}(\alpha,Z) \sim \mathcal{A}(P_{\alpha,Z})$, \item $I(\alpha,Z) = I(P_{\alpha,Z})$, \item $\langle \partial (\alpha, Z) , (\beta, Z') \rangle \ne 0$ if and only if $P_{\alpha, Z}$ is obtained from $P_{\beta, Z'}$ by rounding a corner and locally losing one $h.$ \end{enumerate} Here, by $\mathcal{A}(\alpha,Z) \sim \mathcal{A}(P_{\alpha,Z})$, we mean that by choosing our nice perturbation $\varphi_0$ sufficiently close to $\varphi^1_H$, we can arrange for $\mathcal{A}(\alpha,Z)$ to be as close to $\mathcal{A}(P_{\alpha,Z})$ as we wish. \end{prop} We have already proven the first of the three listed properties in the above proposition. The second will be proven below in Section \ref{sec:index}. The proof of the third takes up the remainder of Section \ref{sec:index}. We have already mentioned that the proposition is inspired by similar combinatorial models, but it might help the reader to note that, more specifically, the second property most closely corresponds to the index calculations in \cite[Section 3.2]{C-CG-F-H-R}, while the third property is inspired\footnote{In particular, the idea of attaching vertical rays in our definition of corner rounding is inspired by \cite[Definition A.2]{Hutchings_beyond}.} by \cite[Conjecture A.3]{Hutchings_beyond}, which was later proven in \cite{Choi}, and which we return to below. % \bigskip Finally, we end this section by defining the {\bf Double Rounding} operation which will appear in the following sections and which has already been introduced in Section \ref{sec:ham}. Namely, if $P_{\beta,Z'}$ has three consecutive edges, all labeled by $h$, we say that $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ by {\bf double rounding} if we remove both interior lattice points for these three edges, take the convex hull of the remaining lattice points (in the region formed by adding vertical lines, as above), and label all new edges by $e$. \subsection{Computation of the index} \label{sec:index} In this section, we prove the second item in Proposition~\ref{prop:model}. Before giving the proof, we first summarize the definitions of the various terms of the PFH grading as defined in Equation \eqref{eqn:twistedgrading}, which we recall here: \begin{equation*} I(\alpha,Z) = c_{\tau} (Z) + Q_{\tau}(Z) + \sum_i \sum^{m_i}_{k = 1} CZ_{\tau}(\alpha^k_i). \end{equation*} These definitions can be found explained in detail, for example, in \cite[Sec. 2]{Hutchings-index}. To define the {\bf relative Chern class}, $c_{\tau}(Z)$, we first take a surface $S$, representing $Z$ in $[-1, 1] \times \S^2 \times \S^1$, assumed transverse to the boundary $\{-1,1\} \times \mathbb{S}^2 \times \mathbb{S}^1$ and embedded in $(-1,1) \times \mathbb{S}^2 \times \mathbb{S}^1$. We then define $c_{\tau}(Z)$ to be a signed count of zeros of a generic section $\Psi$ of $\xi|_S$, where recall that $\xi$ is the vertical tangent bundle of the fibration $[-1,1] \times \S^2 \times \S^1\to [-1,1]\times\S^1$, such that the restriction of $\Psi$ to $\partial S$ agrees with the trivialization $\tau$. We similarly define the {\bf relative intersection number} $Q_{\tau}(Z)$ by the formula \begin{equation} \label{eqn:intersection} Q_{\tau}(Z) : = c_1(N,\tau) - w_{\tau}(S), \end{equation} where $c_1(N,\tau)$, the {\bf relative Chern number of the normal bundle}, is a signed count of zeros of a generic section of $N|_S$, such that the restriction of this section to $\partial S$ % agrees with $\tau$; note that the normal bundle $N$ can be canonically identified with $\xi$ along $\partial S$. Meanwhile, the term $w_{\tau}(S)$, the {\bf asymptotic writhe}, is defined by using the trivialization $\tau$ to identify a neighborhood of each Reeb orbit with $\S^1 \times D^2 \subset \mathbb{R}^3$, and then computing the writhe\footnote{This is defined by identifying $\S^1 \times D^2$ with the product of an annulus and interval, projecting to the annulus, and counting crossings with signs.} of a constant $s$ slice of $S$ near the boundary using this identification. Finally, to define the {\bf Conley-Zehnder} index, we first clarify the definitions of elliptic and hyperbolic Reeb orbits, and define the {\bf rotation number} $\theta$ for a simple orbit, relative to the trivialization $\tau$. Specifically, the elliptic case is characterized by the property that the linearized return map has eigenvalues on the unit circle; in this case, one can homotope the trivialization so that the linearized flow at time $t$ with respect to the trivialization is always a rotation by angle $2 \pi \theta_t$, for a continuous function $\theta_t$, and then the rotation number is the change in $\theta_t$ as one goes around the orbit once. In the hyperbolic case, the linearized return map has real eigenvalues, and the linearized return map rotates by angle $2\pi k$ for some half-integer $k\in\tfrac12\Z$ as one goes around the orbit; the integer $k$ is the rotation number in this case. In either case, denoting by $\theta$ the rotation number, for any cover of $\gamma$, we have \begin{equation} \label{eqn:cz} CZ_{\tau}(\gamma^n) := \lfloor n \theta \rfloor + \lceil n \theta \rceil, \end{equation} \begin{proof} [Proof of the second item in Proposition \ref{prop:model}] By the index ambiguity formula, Equation (\ref{eqn:index_ambiguity}), we have \begin{equation}\label{eq:index-ambiguity-here} I(\alpha, Z+a[\S^2])=I(\alpha, Z)+a(2d+2). \end{equation} Therefore, we only have to compute the index $I(\alpha, Z'_{\alpha})$ for a given relative class $Z'_\alpha\in H_2(\S^2\times\S^1,\alpha,d\gamma_-)$. Let us now define the relative class $Z'_\alpha$ we will be using. Write $\alpha=\{(\gamma_-,m_-)\}\cup\{(\alpha_i,m_i)\}\cup\{(\gamma_+, m_+)\}$, where each $(\alpha_i, m_i)$ is either an $(h_{p_i/q_i}, 1)$ or an $(e_{p_i/q_i},m_{p_i/q_i})$. We define $Z'_\alpha=m_-Z'_-+m_+ Z'_++\sum_im_iZ'_{\alpha_i}$, where \begin{itemize} \item $Z_-'\in H_2(\S^2\times\S^1,\gamma_-,\gamma_-)$ is the trivial class, \item $Z_+'\in H_2(\S^2\times\S^1,\gamma_+,\gamma_-)$ is represented by the map \[S_+:[0,1]\times[0,q]\to \S^2\times\S^1, \quad (s,t)\mapsto (R_{t \lceil h'(1) \rceil}(\eta(s)),t),\] where $\eta$ is a meridian from the South pole $p_-$ to the north pole $p_+$, and $R_{t \kappa}$ denotes the rotation on $\S^2$ by the angle $2\pi t \kappa$. \item for $\alpha_i=e_{p,q}$ or $h_{p,q}$, the relative class $Z'_{\alpha_i}\in H_2(\S^2\times\S^1,\alpha_i,q\gamma_-)$ is represented by the map \[S_{\alpha_i}:[0,1]\times[0,q]\to \S^2\times\S^1, \quad (s,t)\mapsto (R_{t \frac{p}{q}}(\eta(s)),t),\] where $\eta$ is a portion of the great circle which begins at $p_-$ and ends at $z_{\frac{p}{q}}$. \end{itemize} The class $Z_\alpha$ from \ref{sec:comb} is related to the class $Z'_\alpha$ as follows. We have $Z_-=Z'_-$, $Z_+=Z'_++\lceil h'(1)\rceil\,[\S^2]$ and for $\alpha_i=e_{p,q}$ or $h_{p,q}$, then $Z_{\alpha_i}=Z'_{\alpha_i}+p[\S^2]$. If we denote by $(0,y_\alpha)$ and $(d,w_\alpha)$ the endpoints of $P_{\alpha, Z}$, we thus obtain $Z_\alpha=Z'_\alpha+ (w_\alpha-y_\alpha)[\S^2]$. Using (\ref{eq:index-ambiguity-here}), this yields \begin{equation} \label{eq:Zalpha-Zalpha_prime} I(\alpha, Z_{\alpha})=I(\alpha, Z'_{\alpha})+(w_\alpha-y_\alpha)(2d+2). \end{equation} We will now compute the index $I(\alpha, Z'_\alpha)$. For that purpose, we first need to make choices of trivializations along periodic orbits. Along the orbit $\gamma_-$, the trivialization is given by any frame of $T_{p_-}\S^2$ independent of $t$. Along $\gamma_+$, we take a frame which rotates positively with rotation number $\lceil h'(1)\rceil$. Along other orbits, we use as trivializing frame $(\partial_\theta, \partial_z)\in T\S^2$. We denote by $\tau$ these choices of trivialization. Recall that we are also assuming for simplicity that $h'(-1)$ is arbitrarily close to 0 and $h'(1)$ is arbitrarily close (but not equal) to its ceiling $\lceil h'(1)\rceil$. In order to compute the grading, we now have to compute the Conley-Zehnder index, the relative Chern class, and the relative self-intersection; we then have to put this all together to give the stated interpretation in terms of a count of lattice points. \medskip \noindent{\em Step 1: The Conley-Zehnder index.} We begin by computing the Conley-Zehnder index of each orbit, relative to the trivialization above. \begin{enumerate} \item The north pole orbit $\gamma_+$ is elliptic with rotation number $h'(1)-\lceil h'(1)\rceil$. Picking $h'(1)$ to be sufficiently close to its ceiling, we then find by \eqref{eqn:cz} that \[CZ(\gamma_+^k)=\lfloor k(h'(1) - \lceil h'(1) \rceil ) \rfloor + \lceil k(h'(1) - \lceil h'(1) \rceil ) \rceil = -1,\] for any $k=1,\dots, d$. \item The South pole orbit $\gamma_-$ is elliptic with rotation number $-h'(-1)$ with respect to the considered trivialization. Since $h'(-1)$ is positive but arbitrarily small, we then find by \eqref{eqn:cz} that: \[CZ(\gamma_-^k)=\lfloor -k\,h'(-1) \rfloor+\lceil -k\,h'(-1) \rceil=-1.\] \item For other orbits, we are in the same settings as \cite{Hutchings-Sullivan-Dehntwist}. Namely, for hyperbolic orbits, the rotation number is $0$, so from \eqref{eqn:cz} we have \[CZ(h_{p/q})=0,\] and for the elliptic orbits $e_{p/q}$, the rotation number is slightly negative, so that from \eqref{eqn:cz} we have \[CZ(e_{p/q}^k)= -1,\] for any $k =1, \ldots, d$. \end{enumerate} It follows from the above that the contribution of the Conley-Zehnder part to the index in \eqref{eqn:twistedgrading} is given by \begin{equation} \label{eqn:CZcomp} CZ_{\tau} (\alpha) = \sum_{i} \sum_{k=1}^{m_i} CZ(\alpha_i^k) + \sum_{k=1}^{m_-} CZ(\gamma_-^k) + \sum_{k=1}^{m_+} CZ(\gamma_+^k) = -M + h, \end{equation} where $M$ denotes the total multiplicity of all orbits, and $h$ denotes the total number of hyperbolic orbits. \medskip {\noindent \em Step 2: The relative Chern class.} The relative Chern class $ c_\tau(Z'_-)$ is obviously $0$. For $\alpha_i=e_{p,q}$ or $h_{p,q}$, we consider the representative $S_{\alpha_i}$ of $Z'_{\alpha_{i}}$ given above. We choose the section $\partial_\theta$ as a non-winding section of $\xi|_{S_i}$ along $\alpha_{i}$, and any constant non-zero vector along $q\gamma_-$. Then, the section $\partial_\theta$ over the orbit $\alpha_{p/q}$ has index $-p$ while the section over $q \gamma_-$ has index $0$. It follows that any extension of these sections over $S_{\alpha_{i}}$ must have $-p$ zeros. Hence, \[c_\tau(Z'_{\alpha_{i}})=-p.\] For $Z_+$, an argument analogous to that of the previous paragraph gives \[c_\tau(Z'_+)=-\lceil h'(1)\rceil.\] The Chern class is additive, so we conclude from the above that the $c_\tau$ term of the index is \begin{align} \label{eqn:chernclass} c_\tau(Z'_{\alpha}) &= \sum m_i c_\tau(Z'_{\alpha_{i}}) + m_- \; c_\tau(Z'_-) + m_+ \; c_\tau(Z'_+) \nonumber \\ &= -\sum m_i p_i - m_+ \lceil h'(1) \rceil =-w_\alpha+y_\alpha. \end{align} {\noindent \em Step 3: The relative self-intersection.} Inspired by an analogous construction performed in \cite[Lemma 3.7]{Hutchings-Sullivan-Dehntwist}, we construct a representing surface $S\subset [0,1]\times\S^2\times\S^1$ of $Z'_\alpha$ as a movie of curves as follows. Denote by $\sigma$ the variable in $[0,1]$, and $S_\sigma \subset \{\sigma\} \times \S^2\times\S^1$. We will describe $S_\sigma$ as $\sigma$ decreases from $1$ to $0$. \begin{itemize} \item For $\sigma=1$, $S_1$ is the union of the orbits appearing in $\alpha$ with non-zero multiplicity. \item For values of $\sigma$ close to 1, $S_\sigma$ consists of \begin{enumerate}[(a)] \item $m_i$ circles, parallel to the orbit $\alpha_i$, in the torus $\{z=z_{p_i/q_i}\}\times\S^1\subset\S^2\times\S^1$ (these circles have slope $\frac{q_i}{p_i}$ if we see this torus as $\R/\Z\times\R/\Z$), \item $m_+$ parallel circles with slope $\frac 1{\lceil h'(1) \rceil}$ in the torus $\{z=\sigma\}\times\S^1$, \item $m_-$ parallel ``vertical'' circles $\{pt\}\times\S^1$ in the torus $\{z=-\sigma\}\times\S^1$. \end{enumerate} \item As $\sigma$ decreases to $1/2$, we move all these circles to the same $\{z=\mathrm{constant}\}\times\S^1$ torus. As in \cite{Hutchings-Sullivan-Dehntwist}, we perform negative surgeries, around $\sigma = 1/2$ to obtain an embedded union of circles in a single $\{z=\mathrm{constant}\}\times\S^1$ torus; this union will consist of $k$ (straight) parallel embedded circles directed by a primitive integral vector $(a,b)$. The vector $(a,b)$ and the number $k$ of circles are determined by our data: for homological reasons we must have \begin{align*} kb=d,\qquad ka=\sum_i m_ip_i + m_+ \lceil h'(1) \rceil = w_\alpha-y_\alpha. \end{align*} \item As $\sigma$ decreases from $\frac12$ to $0$, we modify the torus in which the curves are located, without changing the curves themselves, so that $S_{0}$ is $\gamma_-$. \end{itemize} We will compute $Q_\tau (Z'_{\alpha})$ using the above surface $S$ and the formula \eqref{eqn:intersection}. To compute $c_1(N, \tau)$, we take $\psi \in \Gamma (N)$ as follows: we take $\psi = \pi_N \partial_\theta$ everywhere on $S$ except in a small neighborhood of the boundary components $\gamma_+, \gamma_-$ where the vector $\partial_\theta$ is not well-defined. The surface is constructed such that $\psi$ may be extended to a $\tau$-trivial section of $N$ over $\gamma_+$ without introducing any zeroes. However, extending $\psi$ to a $\tau$--trivial section of $N$ over $\gamma_-$ necessarily creates $-a$ zeroes for each of the $k$ embedded circles which gives a total of $-ka$ zeroes. As in \cite{Hutchings-Sullivan-Dehntwist}, the other zeroes of $\psi$ appear at the surgery points with negative signs and their number is given by \[ - \sum \det \begin{pmatrix} p & p' \\ q & q' \end{pmatrix},\] where the sum runs over all pairs of distinct edges $v_{p,q}$, $v_{p',q'}$ in $P_{\alpha,Z}$, with $\tfrac{p'}{q'}<\tfrac pq$. Geometrically, this sum can be interpreted as $- 2 \mathrm{Area}(\mathcal{R_\alpha}')$, where $\mathcal{R}_\alpha'$ is the region between $P_{\alpha,Z}$ and the straight line connecting $(0,y_\alpha)$ to $(d,w_\alpha)$. Thus, we obtain: \[ c_1(N, \tau) = -(w_\alpha-y_\alpha) - 2 \mathrm{Area}(\mathcal{R_\alpha}').\] We must now compute the writhe $w_\tau(S)$. By construction, there is no writhe near $\sigma = 1$. Near $\sigma = 0$ the writhe is given by the writhe of the braid $k(a,b)$ on the torus which is $(w_\alpha-y_\alpha)(1-d)$, so we get \[ w_{\tau}(S) = (w_\alpha-y_\alpha)(d-1).\] Summing the above, we therefore get \begin{align} Q_\tau(Z'_{\alpha}) = -(w_\alpha-y_\alpha) - 2 \mathrm{Area}(\mathcal{R_\alpha}')-(w_\alpha-y_\alpha)(d-1). \label{eqn:selfintersection} \end{align} {\noindent \em Step 4: The combinatorial interpretation.} We now put all of this together to prove the second item in Proposition~\ref{prop:model}. By combining \eqref{eqn:CZcomp}, \eqref{eqn:chernclass} and \eqref{eqn:selfintersection} and the definition of the grading \eqref{eqn:twistedgrading}, we have \begin{align*} I(\alpha,Z'_{\alpha}) = -M+h -(w_\alpha-y_\alpha)(d+1)-2 \mathrm{Area}(\mathcal{R}_\alpha'). \end{align*} Using Equation (\ref{eq:Zalpha-Zalpha_prime}), we obtain \[ I(\alpha,Z_{\alpha})=-M+h+2\mathrm{Area}(\mathcal{R}_\alpha)+(w_\alpha-y_\alpha),\] where $\mathcal{R}_\alpha$ denotes the region between $P_{\alpha,Z_{\alpha}}$ and the $x$-axis. By Pick's theorem, \[ 2 \mathrm{Area}(\mathcal{R}_\alpha) = 2 T - (M + d+ (w_\alpha-y_\alpha)) - 2,\] where $T$ denotes the total number of lattice points in the closed region $\mathcal{R}_{\alpha}$. So, by combining the previous two equations, we get \[ I(\alpha,Z_{\alpha}) = 2(T-M - 1) -d + h.\] Now $(T-M-1)$ is exactly the number of lattice points in the closed region $\mathcal{R}_{\alpha}$, not including the lattice points on the path, and so, $T-M-1 = j$, hence the second item of Proposition~\ref{prop:model} is proved for $Z=Z_\alpha$. Now remember that if we shift our path upwards by $(0,1)$, then $j$ increases by $d+1$. Thus, using (\ref{eq:index-ambiguity-here}), we deduce that the second item of Proposition~\ref{prop:model} is satisfied for all relative classes $Z$. \end{proof} \subsubsection{Fredholm index in the combinatorial model}\label{sec:fredholm_combinatorial} The goal of this section is to give a simple formula for the Fredholm index which relates it to our combinatorial model. Let $C$ be a $J$--holomorphic curve in $\mathcal{M}_{J}( (\alpha,Z), (\beta,Z'))$, where $J$ is weakly admissible. Recall from Equation \eqref{eqn:indexformula_Fredholm} that the Fredholm index of $C$ is given by the formula $$\text{ind}(C) = - \chi(C) + 2 c_{\tau}(C) + CZ^{\mathrm{ind}}_{\tau}(C),$$ where $\chi(C)$ denotes the Euler characteristic of the curve, $c_{\tau}(C)$ is the relative first Chern class which we discussed above, and $CZ^{\mathrm{ind}}_{\tau}$ is a term involving the Conley-Zehnder index defined as follows: Write $\alpha = \{(\alpha_i, m_i)\}$ and $\beta =\{(\beta_j, n_j)\}$. Suppose that $C$ has ends at $\alpha_i$ with multiplicities $q_{i,k}$ and ends at $\beta_j$ with multiplicities $q_{j,k}'$; note that we must have $\sum_k q_{i,k} = m_i$ and $\sum_k q_{j,k}' = n_j$. Then, $$ CZ^{\mathrm{ind}}_{\tau}(C) := \sum_i \sum_k CZ_\tau(\alpha_i^{q_{i,k}}) \; - \; \sum_j \sum_k CZ_\tau(\beta_j^{q_{j,k}'}).$$ The next lemma explains how to compute $\mathrm{ind}(C)$ from the combinatorial model. In this lemma, we denote the starting points of $P_{\alpha, Z}$ and $P_{\beta, Z'}$ by $(0, y_\alpha)$ and $(0, y_\beta)$, and their endpoints by $(d, w_\alpha)$ and $(d, w_\beta)$, respectively. \begin{lemma}\label{lem:fredholm_index} Let $\varphi_0$ be a nice perturbation of $\varphi^1_H$, where $H \in \mathcal{D}$, let $J$ be any weakly admissible almost complex structure, and let $C$ be any irreducible $J$-holomorphic curve from $(\alpha, Z)$ to $(\beta, Z')$. Then, \begin{equation} \label{eqn:Fredholmindex} \text{ind}(C) = -2 + 2g + 2e_- + h + 2v , \end{equation} where $g$ is the genus of $C$, $e_-$ is the number of negative ends of $C$ which are at elliptic orbits, $h$ is the number of ends of $C$ at hyperbolic orbits, and $v = (y_\alpha - y_\beta) + (w_\alpha - w_\beta) $. \end{lemma} \begin{proof} We will prove the lemma by describing each of the three terms $\chi(C), c_\tau(C)$, and $CZ_\tau^{\mathrm{ind}}(C)$, which appear in $\mathrm{ind}(C)$, in terms of our combinatorial model. The number of ends of the curve $C$ is given by the sum $e_-+ e_+ +h$, where $e_+$ denotes the number of positive ends of $C$ which are at elliptic orbits. The Euler characteristic of $C$ is given by the formula $$\chi(C) = 2 -2g - e_- - e_+ - h.$$ As for the Chern class, because $[C] = Z- Z'$, we can write $c_{\tau}(C) = c_{\tau}(Z) -c_\tau(Z')$. Now, by the index computations of the previous section, $c_{\tau}(Z) = w_\alpha + y_\alpha$. Similarly, $c_{\tau}(Z') = w_\beta + y_\beta$. It follows that $$c_{\tau}(C) = v.$$ To compute $CZ_\tau^{\mathrm{ind}}(C)$, note that by our computations from the previous section, we have $CZ_\tau(\alpha_i^{q_{i,k}}) = -1$ if $\alpha_i$ is elliptic and $0$, otherwise; a similar formula holds for $CZ_\tau(\beta_j^{q_{j,k}'})$. This implies that $$CZ_\tau^{\mathrm{ind}}(C) = -e_+ + e_-.$$ Combining the above, we get $\mathrm{ind}(C) = -2 + 2g + 2e_- + h + 2v.$ \end{proof} \begin{remark} \label{rmk:why!} As already mentioned in Section~\ref{sec:admissible_vs_weakly}, in some situations we want the added flexibility of being able to work with weakly admissible almost complex structures. Lemma~\ref{lem:fredholm_index} above is also stated for weakly admissible almost complex structures, as are the forthcoming Lemma \ref{lem:positivity} and Lemma~\ref{lem:above}. Ultimately, the reason we want to be able to work with weakly admissible almost complex structures is because of the very useful Lemma~\ref{lem:comp} below, which will allow us to connect an admissible $J$ to a weakly admissible one in order to facilitate computations; the point is that the proof of Lemma~\ref{lem:comp} requires Lemma~\ref{lem:fredholm_index}, Lemma \ref{lem:positivity}, and Lemma~\ref{lem:above} in the weakly admissible case. \end{remark} \subsection{Positivity} \label{sec:pos} We now begin the proof of the third item of Proposition~\ref{prop:model}; this will take several subsections and will require a close examination of those $J$--homolorphic curves in $\R \times \S^1 \times \S^2$ which appear in the definition of the PFH differential. (Recall from Section \ref{sec:specdefn} that $X$ is identified with $\R \times \S^1 \times \S^2$.) In this section, we prove a very useful lemma which puts major restrictions on what kind of $J$--holomorphic curves can appear. The lemma and its proof are inspired by \cite[Lemma 3.5]{Choi}, which is in turn inspired by arguments in \cite{Hutchings-Sullivan-Dehntwist} and \cite{Hutchings-Sullivan-T3}, see for example \cite[Lemma 3.11]{Hutchings-Sullivan-Dehntwist}. To prepare for the lemma of this section, we need to introduce some new terminology. For $-1 < z_0 < 1 $, define the {\bf slice} \[ S_{z_0} := \lbrace (s,t,\theta,z) \in \mathbb{R} \times \S^1 \times \S^2 : z = z_0\rbrace.\] This is homotopy equivalent to a two-torus, and in particular we have $H_1(S_{z_0}) = H_1(\S^1_t \times \S^1_{\theta})$; we identify $H_1(\S^1_t \times \S^1_{\theta})$ with $\mathbb{Z}^2$ so that the positively oriented circle factors $\S^1_t$ and $ \S^1_{\theta}$ correspond to the vectors $(1,0)$ and $(0,1)$, respectively. Let $C$ be any $J$-holomorphic curve. If $C$ is transverse to $S_{z_0}$ (which happens for generic $z_0$) and $C$ has no ends at $z_0$, then $C_{z_0}=C \cap S_{z_0}$ is a (possibly empty) compact $1$-dimensional submanifold of $C$. When non-empty, it is the boundary of the subdomain % given by $C \cap \{z\leq z_0\}$. Thus, the orientation of $C$ induces an orientation on $C_{z_0}$: our convention is that we take the opposite of the usual ``outer normal first''\footnote{ \label{footnote:orientation} Recall that in the ``outer normal first" orientation, a vector $v$ on $C_{z_0}$ is positive if for an inner normal vector $w$, the frame $(v,w)$ is positive.} convention. Therefore, we get a well-defined class $[C_{z_0}] \in H_1(S_{z_0})=\Z^2$, which we call the {\bf slice class}. \begin{lemma} \label{lem:positivity} Let $\varphi_0$ be a nice perturbation of $\varphi^1_H$, where $H \in \mathcal{D}$. Let $C$ be a $J$--holomorphic curve, where $J$ is weakly admissible, and let $z_0 \in (-1,1)$ be such that $C$ intersects $S_{z_0}$ transversally, and the nice perturbation vanishes in an open neighborhood of $z_0$. (In particular, $C$ has no ends at Reeb orbits near $z_0$.) Then, \begin{equation} \label{eqn:keyinequality} \left(1,h'(z_0)\right) \times [C_{z_0}] \geq 0, \end{equation} with equality only if $C$ does not intersect $S_{z_0}$. \end{lemma} Here, $(a,b) \times (c,d)$, where $(a,b), (c,d) \in \R^2$, is defined to be the quantity $ad -bc$. \begin{remark}\label{rem:positivity_irreducible} In the context of the above lemma, suppose that $C$ is irreducible and let $z_{\min} := \inf\{ z_0: [C_{z_0}] \neq 0\}$ and $z_{\max} := \sup\{ z_0: [C_{z_0}] \neq 0\}$; here, we consider the $z_0$ such that the above lemma is applicable. The curve $C$ is connected, because it is irreducible, and thus its projection to $\S^2$ is also connected. Therefore, $C$ is contained in $\{(s,t,\theta, z) \in \R \times \S^1 \times \S^2: z_{\min} \leq z \leq z_{\max} \}$ . \end{remark} \begin{remark}\label{rem:locality} In the context of the above lemma, suppose that $C$ is an irreducible $J$--holomorphic curve such that $[C_{z_0}] = 0$ for all $z_0$ satisfying the conditions of the lemma. Then, as a consequence of the above lemma, $C$ must be a {\it local} curve in the following sense: there exists $z_{p,q}$ such that all ends of $C$ are either at $e_{p,q}$ or $h_{p,q}$. \end{remark} \begin{proof} We use the fact that, by Lemma~\ref{lem:pointwisenonnegative} the canonical $2$-form $\omega_{\varphi}$ is pointwise nonnegative on $C$, with equality only if the tangent space is the span of $\partial_s$ and $R$. Namely, as a function of $z > z_0$, close to $z_0$, we have that the mapping \[ \rho:z \mapsto \int_{C \cap (\mathbb{R} \times \S^1_t \times \S^1_\theta \times [z_0,z]) } \omega_{\varphi} \] is non-decreasing, as $z$ increases. Hence, its derivative with respect to $z$ is nonnegative. We will prove Equation \eqref{eqn:keyinequality} by showing that \begin{equation}\label{eqn:derivative} \rho'(z) = \tfrac{1}{2}\left(1,h'(z)\right) \times [C_{z_0}], \end{equation} for $z$ close to $z_0$ and $z \geq z_0$. Recall from Section \ref{sec:specdefn} that $\omega_\varphi$ is identified with $\omega + d\tilde H \wedge dt$. Since $\tilde H$ coincides with $H$ near $z_0$ we can write $\omega_\varphi = \omega + dH \wedge dt. $ Now, we have \[\omega_{\varphi}=\omega + dH \wedge dt = \tfrac{1}{4 \pi} d(- z\, d \theta) + d(H\,dt) = d\left( -\tfrac{1}{4\pi}z\, d \theta + H(z)\, dt\right).\] Let $\alpha = H(z) dt - \frac{1}{4\pi}z\, d \theta$ and note that $\alpha$ restricts to a closed 1-form $\alpha_z$ on any slice $S_z$. % Our choice of orientation gives: \[\partial (C \cap (\mathbb{R} \times \S^1_t \times \S^1_\theta \times [z_0,z])) = C_{z_0} - C_z.\] Thus it follows from the above: \begin{align*} \rho(z)&=\int_{C \cap (\mathbb{R} \times \S^1_t \times \S^1_\theta \times [z_0,z])} d\alpha = \int_{C_{z_0} } \alpha - \int_{C_z} \alpha \\ &= \langle \alpha_{z_0} - \alpha_z , [C_{z_0}] \rangle. \end{align*} The rate of change of the expression above with respect to $z$ is given by \begin{equation} \label{eqn:todiff} \rho'(z)=- \left (\tfrac{1}{2}h'(z), -\tfrac{1}{2}\right) \cdot [C_{z_0}] = \tfrac{1}{2} \left(1, h'(z)\right) \times [C_{z_0}], \end{equation} which extends by continuity to $z_0$. This proves \eqref{eqn:keyinequality}. To prove that equality in Equation \eqref{eqn:keyinequality} forces the slice to be empty, assume that $\rho'(z_0)=0$. We have shown that $\rho'\geq 0$, and so $\rho'$ must have a local minimum at $z_0$. Hence, $\rho''$ and $\rho'$ vanish at $z_0$. We claim that this implies $[C_{z_0}] = 0$. To prove this, write $[C_{z_0}] = (a, b)$ and note that, by Equation \eqref{eqn:derivative}, we have \begin{equation*} \rho'(z) = \tfrac12(b - a\, h'(z))\quad\text{and}\quad \rho''(z) = -\tfrac12\,{ a\, h''(z)}. \end{equation*} Since we are assuming $h''(z_0) \ne 0$, the vanishing of both of the above quantities can only take place if $a = b =0$. We can therefore conclude that $[C_{z_0}] = 0$. By the assumption that $C$ intersects $S_{z_0}$ transversely, we conclude that $[C_{z}] = 0$ for $z$ sufficiently close to $z_0$. Hence, $\rho(z)=0$ for $z$ sufficiently close to $z_0$. But, by Lemma~\ref{lem:pointwisenonnegative} this can only occur if the tangent space to $C \cap (\mathbb{R} \times \S^1_t \times \S^1_\theta \times [z_0,z]) $ is always in the span of the Reeb vector field $R$ and $\partial_s$, which are tangent to ${S_z}$. But, since $C$ is transverse to $S_{z_0}$, this cannot happen for $z$ sufficiently close to $z_0$, unless the intersection $C_{z_0}=C\cap S_{z_0}$ is empty. \end{proof} The next lemma, which is inspired by \cite[Eq. 28]{Hutchings-Sullivan-Dehntwist} and \cite[Definition 1.4]{Choi}, allows us to compute the slice class $[C_{z_0}]$ from our combinatorial model. We suppose here that $(\alpha, Z), (\beta, Z')$ are two PFH generators for $\varphi_0$, the nice perturbation of $\varphi^1_H$, as described in Section \ref{sec:ham} and that $C$ is a $J$--holomorphic curve from $(\alpha,Z)$ to $(\beta,Z')$; recall that this means that $C$ is a $J$--holomorphic curve from $\alpha$ to $\beta$ such that $[C] = Z - Z'.$ Let $P_{\alpha,Z}$ and $P_{\beta,Z'}$ be the concave lattice paths associated to $(\alpha,Z)$ and $(\beta,Z')$, respectively, as described in Section \ref{sec:comb}. For any $z_0$, let $P^{z_0}_{\alpha}$ be the vector obtained by summing all of the vectors in the underlying path $P_{\alpha}$ which correspond to Reeb orbits that arise from $z < z_0$. Denote $$P^{z_0}_{\alpha, Z} = (0, y_\alpha) + P^{z_0}_{\alpha},$$ where $(0,y_{\alpha})$ denotes the starting point of $P_{\alpha,Z}$ on the $y$--axis. We define $P^{z_0}_{\beta}$ and $P^{z_0}_{\beta, Z'}$, analogously. \begin{lemma} \label{lem:slice} Let $C$ be a $J$-holomorphic curve from $(\alpha,Z)$ to $(\beta,Z')$ and let $z_0 \in (-1,1)$ be such that $\alpha, \beta$ have no Reeb orbits near $z_0$ and suppose that $C$ intersects $S_{z_0}$ transversally. Then, \begin{equation} \label{eqn:sliceformula} [C_{z_0}] = P^{z_0}_{\alpha, Z } - P^{z_0}_{\beta, Z'}. \end{equation} \end{lemma} \begin{proof} First, note that the first coordinate of $P_{\alpha}^{z_0}$ (hence that of $P_{\alpha, Z}^{z_0}$, too) corresponds to the class in $H_1(\S^1_t\times\S^2)\simeq\Z$ obtained by summing the contributions of the orbits in $\alpha$ that belong to the domain $\{z<z_0\}$. The second coordinate of $P_{\alpha}^{z_0}$ is given by the $\theta$ component of the class in $H_1(\S_t^1\times\S_\theta^1)$ obtained by summing the contributions of the orbits in $\alpha$, other than $\gamma_-$, which are included in $\{z<z_0\}$. Analogous statements hold for $P_{\beta}^{z_0}$. Pick any $z_-<z_+$ such that the assumptions of Lemma~\ref{lem:slice} hold for $z_-$ and $z_+$, % and denote the part of $C$ with $z_- \leq z \leq z_+$ by $C_{[z_-, z_+]}$. This is asymptotic to some orbit set $\alpha'$ at $+\infty$ and some orbit set $\beta'$ at $-\infty$. The boundary of $C_{[z_-, z_+]}$ has components corresponding to $\alpha', \beta', C_{z_+}$ and $C_{z_-}$. The positive ends $\alpha'$ have the orientation coming from the Reeb vector field and the negative ends $\beta'$ have the opposite orientation. We therefore have \begin{equation} \label{eq:sliceclass} [C_{z_+}] = [\alpha'] - [\beta'] + [C_{z_-}] \end{equation} in $H_1(\S^1_t \times \S^1_{\theta})\simeq \Z^2$. We can apply similar reasoning to the part $C_{[-1,z_0]}$ of $C$ with $-1 \leq z \leq z_0$ to learn that \[ \iota_*[C_{z_0}] = [\alpha'] - [\beta'], \] in $H_1(\S^1_t \times \S^2),$ where \[\iota_*: H_1(\S^1_t \times \S^1_{\theta}) \to H_1(\S^1_t \times \S^2)\] is the map induced by inclusion. In particular, the first component of $[C_{z_0}]$ is that of $[\alpha'] - [\beta']$, which is exactly the first coordinate of $P^{z_0}_{\alpha, Z } - P^{z_0}_{\beta, Z'}$. % We now turn our attention to the second coordinate of $[C_{z_0}]$. Note that $[C_{z_0}]$ is determined by $[C]=Z-Z'$ in the following way. Consider the Mayer-Vietoris sequence associated to the cover by the two open subsets $\R\times\S^1\times\{z\in\S^2: z<z_0+\delta\}$ and $\R\times\S^1\times\{z\in\S^2:z>z_0-\delta\}$. The class $-[C_{z_0}]$ % is the image of $[C]$ under the connecting map\footnote{Our convention is that the connecting map $\partial$ in Mayer-Vietoris is determined by the ordering of the cover; in particular, in the present situation, for any class $K$, $\partial K$ is given by taking the boundary of the component of $K$ in the region with $z < z_0 + \delta$.} \[H_2(\R\times\S^1\times\S^2, \alpha, \beta)\to H_1(\R\times\S^1\times\{z\in\S^2:z_0-\delta<z<z_0+\delta\})\simeq H_1(S_{z_0}),\] where $\delta>0$ is small enough. In particular, adding $y[\S^2]$ with $y\in\Z$ to the class $[C]$, adds $y$ to the second component of $[C_{z_0}]$, since recall that our convention for the orientation on $[C_{z_0}]$ is opposite the standard boundary orientation. To compute the second component of $[C_{z_0}]$, we apply Equation (\ref{eq:sliceclass}) in the case where $z_+=z_0$ and $z_-$ is such that $C_{[-1, z_-]}$ has no ends other than possibly $\gamma_{-}$. We obtain in this case that the second component of $[C_{z_0}]-[C_{z_-}]$ is the second coordinate of $P^{z_0}_{\alpha} - P^{z_0}_{\beta}$. There only remains to establish that the second component of $[C_{z_-}]$ is $y_\alpha-y_\beta$. We will use here the fact that $[C_{z_-}]$ is determined by $Z$ and $Z'$. In the case, where $Z=Z_\alpha$ and $Z'=Z_\beta$ (as defined in Section \ref{sec:comb}), the second coordinate of $[C_{z_-}]$ vanishes. In general, $[C]=Z_\alpha+y_\alpha[\S^2]-(Z_\beta+y_\beta[\S^2])$. Thus, the second component of $[C_{z_-}]$ is $y_\alpha-y_\beta$, which concludes the proof. \end{proof} \subsection{Paths can not cross} As a consequence of the results in Section \ref{sec:pos}, we can prove the following useful fact which will play an important role in our proof of Proposition \ref{prop:model}. It is inspired by \cite[Prop 3.12]{Hutchings-Sullivan-Dehntwist} and \cite[Prop 10.12]{Hutchings-Sullivan-T3}. \begin{lemma} \label{lem:above} Let $\varphi_0$ be a nice perturbation of $\varphi^1_H$, where $H \in \mathcal{D}$, and let $J$ be any weakly admissible almost complex structure. If there exists a $J$-holomorphic curve $C$ from $(\alpha,Z)$ to $(\beta,Z')$, then $P_{\beta,Z'}$ is never above $P_{\alpha,Z}.$ \end{lemma} \begin{proof} Let $(0, y_\alpha)$ and $(0, y_\beta)$ denote the starting points of $P_{\alpha,Z}$ and $P_{\beta,Z'}$, respectively. We will first show that $y_\alpha \geq y_\beta$. Denote by $m_\alpha, m_\beta$ the multiplicities of $\gamma_-$ in $\alpha$ and $\beta$, respectively. Let $z_0 = -1 + \varepsilon$ for some small $\varepsilon > 0$. We have $P^{z_0}_{\alpha,Z} = (m_\alpha, y_\alpha)$ and $P^{z_0}_{\beta,Z'} = (m_\beta, y_\beta) $. Hence, by Lemma \ref{lem:slice}, $[C_{z_0}] = (m_\alpha - m_\beta, y_\alpha - y_\beta)$. Applying Lemma \ref{lem:positivity}, we obtain $$(1, h'(z_0)) \times [C_{z_0}] = (y_\alpha - y_\beta) -h'(z_0) (m_\alpha - m_\beta) \geq 0.$$ By our conventions from Section \ref{sec:ham}, $ h'(z_0) \approx 0$ and thus $(1, h'(z_0)) \times [C_{z_0}] \approx y_\alpha - y_\beta$. Since $y_\alpha - y_\beta$ is integer valued, the above inequality yields $y_\alpha \geq y_\beta$. Next, let $(d, w_\alpha)$ and $(d, w_\beta)$ denote the endpoints of $P_{\alpha,Z}$ and $P_{\beta,Z'}$, respectively. We will now show that $w_\alpha \geq w_\beta$. Denote by $n_\alpha, n_\beta$ the multiplicities of $\gamma_+$ in $\alpha$ and $\beta$, respectively. Let $z_0 = 1 - \varepsilon$ for some small $\varepsilon > 0$. We have $P^{z_0}_{\alpha,Z} = (d, w_\alpha) - n_\alpha \left(1, \lceil h'(1) \rceil \right)$ and $P^{z_0}_{\beta,Z'} = (d, w_\beta) - n_\beta \left(1, \lceil h'(1) \rceil \right)$. Hence, by Lemma \ref{lem:slice}, $$[C_{z_0}] = (n_\beta - n_\alpha, w_\alpha - w_\beta + (n_\beta -n_\alpha) \lceil h'(1) \rceil ).$$ Now, applying Lemma \ref{lem:positivity}, we obtain $$(1, h'(z_0)) \times [C_{z_0}] = w_\alpha - w_\beta + (n_\beta - n_\alpha) \left( \lceil h'(1) \rceil - h'(z_0) \right) \geq 0.$$ By our conventions from Section \ref{sec:ham}, $ \lceil h'(1) \rceil - h'(z_0) \approx 0$ and thus $(1, h'(z_0)) \times [C_{z_0}] \approx w_\alpha - w_\beta$. Since $w_\alpha - w_\beta$ is integer valued, the above inequality yields $w_\alpha \geq w_\beta$. To complete the proof, suppose that the conclusion of the lemma does not hold. We have shown that $P_{\beta,Z'}$ cannot begin or end above $P_{\alpha,Z}$. Hence, we can find two intersection points $(a,b )$ and $(c,d)$, with $a< c$, between the two paths, such that the path $P_{\beta,Z'}$ is strictly above $P_{\alpha,Z}$ in the strip $\{(x,y)\in\R^2: a<x<c\}$. Let $U, L$ denote the parts of $P_{\beta,Z'}, P_{\alpha,Z}$, respectively, which are contained in $\{(x,y) \in \R^2: a \leq x \leq c \}$. Consider the line connecting $(a,b)$ and $(c,d)$. We can find a point $z_0$, with $-1 < z_0 < 1$, such that $h'(z_0) = \frac{d-b}{c-a}$. We will compute the slice class $[C_{z_0+ \varepsilon}]$, for sufficiently small $\varepsilon >0$ and will show that $$(1,h'(z_0 + \varepsilon)) \times [C_{z_0 + \varepsilon}] < 0,$$ which contradicts Lemma \ref{lem:positivity}. The reason for considering $[C_{z_0+ \varepsilon}]$ instead of $[C_{z_0}]$ itself is that there might be Reeb orbits in $\alpha, \beta$ corresponding to $z_0$ in which case we cannot apply Lemmas \ref{lem:positivity} \& \ref{lem:slice}. % To compute the slice class $[C_{z_0 + \varepsilon}]$, we will compute $P^{z_0 + \varepsilon}_{\alpha, Z}, P^{z_0 + \varepsilon}_{\beta, Z'}$ and use Lemma \ref{lem:slice}. We begin with $P^{z_0 + \varepsilon}_{\alpha, Z}$. Let $(p,q)$ be the corner of $P_{\alpha, Z}$, on $L$, with the following property: the edge in $P_{\alpha, Z}$ to the left of $(p,q)$ has slope at most $\frac{d-b}{c-a}$, and the edge to the right of $(p,q)$ has slope strictly larger than $\frac{d-b}{c-a}$. The corner $(p,q)$ exists because $L$ is strictly below the line passing through $(a,b)$ and $(c,d)$. Then, $P^{z_0 + \varepsilon}_{\alpha, Z} = (p,q)$. Now, denote $P^{z_0 + \varepsilon}_{\beta, Z'}= (p', q')$; this vector may be computed as follows: If the line passing through $(a,b)$ and $(c,d)$ is strictly above $U$, then $(p', q')$ is computed exactly as above. If not, the line passing through $(a,b)$ and $(c,d)$ must coincide with $U$; then $(p', q')$ is the endpoint of the edge in $P_{\beta, Z'}$ containing $U$. We obtain $$[C_{z_0 + \varepsilon}]= (p-p', q-q') $$ with $(p,q)$ and $(p', q')$ as described in the previous paragraph. Now, we have $$(1,h'(z_0 + \varepsilon)) \times [C_{z_0 + \varepsilon}] = \left(1, \tfrac{d-b}{c-a}\right) \times (p-p', q-q').$$ This quantity is negative because $L$ is strictly below $U$. Indeed, one can see this by applying a rotation, which does not change the determinant, so that $(1 , \frac{d-b}{c-a})$ is rotated to a positive multiple of $(1,0)$, and $(p-p',q-q')$ is rotated to a vector with a negative second component. Hence, $(1,h'(z_0 + \varepsilon)) \times [C_{z_0 + \varepsilon}] < 0$ which contradicts Lemma \ref{lem:positivity}. \end{proof} \subsection{Curves correspond to corner rounding}\label{sec:curves-corner-rounding} Using the results we have obtained thus far, we can now describe the configurations of concave paths which could give rise to a non-trivial term in the PFH differential. More precisely, we can now prove the ``only if'' part of Proposition~\ref{prop:model}, which we state as a lemma below: \begin{lemma} \label{lem:nice} Let $\varphi_0$ be a nice perturbation of $\varphi^1_H$, where $H \in \mathcal{D}$. Assume that $I(P_{\alpha,Z}) - I(P_{\beta,Z'}) = 1$. Then, for generic admissible $J$ close to $J_{std}$, \[ \langle \partial(\alpha,Z), (\beta,Z') \rangle \ne 0 \] only if $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$. \end{lemma} \begin{proof} Assume that \[ \langle \partial(\alpha,Z), (\beta,Z') \rangle \ne 0\] for some generically chosen $J$ and some generators $(\alpha, Z)$ and $(\beta, Z')$. We first choose $J$ generically to rule out double rounding, which we can do by the argument in \cite[Lemma A.1]{Hutchings-Sullivan-Dehntwist}.\footnote{In this argument, other than notational changes, we need to make one minor modification: the $2$-form $dt\wedge dy-ds\wedge dx$ in the proof of Lemma A.2 must be replaced with the $2$-form $dt\wedge d\theta - f(z) ds\wedge dz$, for a function $f$ determined by $h$. (We could give an explicit formula for $f$, but it is not necessary for what we write here.) The reason we need to add the function $f$ is because the almost complex structure $J_{std}$ does not map $\partial_z$ to $\partial_{\theta}$: in contrast, the almost complex structure $J_0$ from \cite{Hutchings-Sullivan-Dehntwist}, Lemma A.1 maps $\partial_x$ to $\partial_y$. However, the rest of the argument there can be repeated essentially verbatim, since $\int_C f(z) ds dz = \int_C d (s f(z) dz)$.} % By Lemma \ref{lem:above}, we know that $P_{\beta,Z'}$ is never above $P_{\alpha,Z}$. Consider the region between $P_{\alpha,Z}$ and $P_{\beta,Z'}$. We can take this region and decompose it into two kinds of subregions: {\bf non-trivial} subregions where $P_{\alpha,Z}$ is {\bf above} $P_{\beta,Z'}$ --- meaning that the parts of $P_{\alpha,Z}$ and $P_{\beta,Z'}$ intersect at most at two points in these regions; and, {\bf trivial} subregions where the concave paths (without the labels) coincide. See Figure \ref{fig:rounding-corner3}. \begin{figure}[h!] \centering \def1\textwidth{1.0\textwidth} \input{rounding-corner3.pdf_tex} \caption{Examples of trivial and non-trivial regions. The path $P_{\beta, Z'}$ is in black, the path $P_{\alpha, Z}$ is in grey were it does not coincide with $P_{\beta, Z'}$.}\label{fig:rounding-corner3} \end{figure} We will first show that there is at least one non-trivial region. Arguing by contradiction, assume there is no non-trivial region, hence that $P_{\alpha,Z}$ and $P_{\beta,Z'}$ coincide as {\em unlabeled concave paths.} Let $C$ be the unique embedded component of a given $J$-holomorphic curve from $(\alpha,Z)$ to $(\beta,Z')$; see Proposition \ref{prop:structure_Jcurves}. We claim that $C$ must be {\em local} in the following sense: It is a $J$--holomorphic cylinder from the hyperbolic orbit, near some $z= z_{p,q}$, that arises after the good perturbation, to the elliptic orbit near $z = z_{p,q}$; furthermore, it does not leave the neighborhood of $z_{p,q}$ where our good perturbation is non-trivial. Indeed, if $C$ were not local, then we could find $z_0 \in (-1, 1)$ such that both of Lemmas \ref{lem:positivity} \& \ref{lem:slice} would be applicable at $z_0$. Now, Lemma \ref{lem:positivity} would imply $[C_{z_0}] \neq 0$, while Lemma \ref{lem:slice} would imply $[C_{z_0}] = 0$ because the two (unlabeled) concave paths coincide. Hence, $C$ must be local. As explained in the proof of \cite[Lemma 3.14]{Hutchings-Sullivan-Dehntwist} local curves appear in pairs and so their mod $2$ count vanishes. We will prove that if there exists $C \in \mathcal{M}_J( (\alpha,Z), (\beta,Z'))$, then $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$. First, observe that it suffices to prove this under the assumption that $C$ is irreducible. Indeed, if $C$ is not irreducible, consider its embedded component $C'$. Then, as a consequence of Proposition \ref{prop:structure_Jcurves}, there exists PFH generators $(\alpha_1,Z_1)$ and $(\beta_1,Z_1')$ such that $C' \in \mathcal{M}_J( (\alpha_1,Z_1), (\beta_1,Z_1'))$ and $P_{\alpha_1,Z_1}$ is obtained from $P_{\beta_1,Z_1'}$ by rounding a corner and locally losing one $h$ if and only if $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ via the same operation. We will suppose for the rest of the proof that $C$ is irreducible. \begin{claim} Under the assumption that $C$ is irreducible, the region between $P_{\alpha, Z}$ and $P_{\beta, Z'}$ contains no trivial regions and one non-trivial region. \end{claim} \begin{proof}[Proof of Claim] Lemma \ref{lem:fredholm_index} implies that if the number of non-trivial regions between $P_{\alpha, Z}$ and $P_{\beta, Z'}$ is at least $2$, then the Fredholm index of $C$ is also at least $2$, because a non-trivial region makes a contribution of size at least $2$ to the sum $2e_- + h + 2 v$ from Equation \eqref{eqn:Fredholmindex}. Since $\mathrm{ind}(C) = 1$, we conclude that the number of non-trivial regions must be one. Next, we will prove that the number of trivial regions must be zero. First, we will show that an edge in $P_{\beta, Z'}$ cannot have lattice points in its interior. Indeed, if such an edge existed it would make a contribution of size at least $3$ to the term $2e_- + h$ in Equation \eqref{eqn:Fredholmindex}. This would then force $2e_- +h + 2v$ to be at least four which cannot happen because $\mathrm{ind}(C) =1$. Now, suppose that there exists a trivial region between $P_{\alpha, Z}$ and $P_{\beta, Z'}$. We will treat the case where there is a trivial region to the left of the non-trivial region, leaving the remaining case, which is very similar, to the reader. There can be at most one trivial region because each trivial region makes a contribution of size at least $1$ to $2e_- +h$, and so if there were two or more such regions $2e_- +h + 2v$ would be at least four. Let $v_{p, q}$ the vector/edge of the trivial region. Then, the edge in $P_{\beta, Z'}$ immediately to the right of $v_{p,q}$ corresponds to a vector $v_{p_1, q_1}$ with $\frac{p}{q} < \frac{p_1}{q_1}$; this inequality is strict because otherwise $P_{\beta, Z'}$ would have an edge with an interior lattice point. It follows that we can find $z_0$ such that $z_{p,q} < z_0< z_{p_1, q_1}$ and Lemma \ref{lem:slice} is applicable at $z_0$; moreover, $P^{z_0}_{\alpha, Z} - P^{z_0}_{\beta, Z'} = 0$. Hence, $[C_{z_0}] = 0$. This contradicts Remark \ref{rem:positivity_irreducible} because $C$ is asymptotic to orbits in both of $\{z \in \S^2: z \geq z_0\}$ and $\{z \in \S^2: z \leq z_0\}$. We therefore conclude that the region between $P_{\alpha, Z}$ and $P_{\beta, Z'}$ consists of a single non-trivial region. \end{proof} Continuing with the proof of Lemma \ref{lem:nice}, observe that by the second item of Proposition~\ref{prop:model}, we have \[ I(\alpha,Z)- I(\beta,Z') = 2j + h_{\alpha} - h_{\beta}, \] where $j$ is the number of lattice points in the region between $P_{\alpha,Z}$ and $P_{\beta,Z'}$, not including lattice points on $P_{\alpha,Z}$, and $h_\alpha, h_\beta$ denote the number of edges labeled $h$ in $P_{\alpha, Z}, P_{\beta, Z'}$, respectively. The number of edges in $P_{\beta, Z'}$, which we denote by $r_\beta$, satisfies the following inequality: $ h_\beta \leq r_\beta \leq j + 1$ . Hence, we have \[ I(\alpha,Z)- I(\beta,Z') \geq 2 (r_{\beta}-1) - r_{\beta} = r_{\beta} - 2,\] with equality if and only if $P_{\alpha,Z}$ and $P_{\beta,Z'}$ start at the same point, end at the same point, the region between $P_{\alpha,Z}$ and $P_{\beta,Z'}$ contains no interior lattice points, every edge of $P_{\beta,Z'}$ is labelled $h$, and no edge of $P_{\alpha,Z}$ is labeled $h$. Since $I(\alpha,Z)- I(\beta,Z') =1$, we can rewrite the above inequality as $r_{\beta} \leq 3$. If $r_\beta = 1$, then the equality $1 = 2j + h_{\alpha} - h_{\beta}$ can hold if and only if $j=1, h_{\alpha} = 0, h_\beta =1$. This implies that there are no interior lattice points in the region between $P_{\alpha, Z}$ and $P_{\beta, Z'}$; moreover, the two paths either begin at the same lattice point or end at the same lattice point. We see that in both cases $P_{\alpha, Z}$ is obtained from $P_{\beta, Z'}$ by rounding a corner and locally losing one $h$. Note that the corner rounding takes place at the extremity of $P_{\beta, Z'}$ which is not on $P_{\alpha, Z}$. If $r_{\beta} = 2$, and if the region between $P_{\alpha,Z}$ and $P_{\beta,Z'}$ has at least one interior lattice point, then $j \geq 2$, so as $h_{\beta} \leq 2$, the index difference must be at least $2$, which can not happen. Thus, the region between $P_{\alpha,Z}$ and $P_{\beta,Z'}$ must have no interior lattice points, and $h_\alpha = h_\beta - 1$. Thus, in this case $P_{\alpha,Z}$ must be obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$. If $r_{\beta} = 3$, then equality holds in the above inequality, so every edge of $P_{\beta,Z'}$ must be labelled $h$, no edge of $P_{\alpha,Z}$ can be, $P_{\alpha,Z}$ and $P_{\beta,Z'}$ must start and end at the same points and there are no interior lattice points between $P_{\alpha,Z}$ and $P_{\beta,Z'}$; thus $P_{\alpha,Z}$ must be obtained from $P_{\beta,Z'}$ by double rounding, which can not occur given our choice of $J$ from the beginning of the proof. % \end{proof} \subsection{Corner rounding corresponds to curves} \label{sec:corner_rounding_curves} In the previous section we proved that if $\langle \partial (\alpha, Z), (\beta, Z')\rangle \neq 0$, then $P_{\alpha, Z}$ is obtained from $P_{\beta, Z'}$ by rounding a corner and locally losing one $h$. To complete the proof of Proposition \ref{prop:model}, we must show the converse: \begin{lemma} \label{lem:converse} Let $\varphi_0$ be a nice perturbation of $\varphi^1_H$, where $H \in \mathcal{D}$. If $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$, then $\langle \partial (\alpha, Z), (\beta, Z')\rangle \neq 0$. In other words, counting mod $2$ we have $$\# \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z')) =1,$$ for generic admissible $J\in \mathcal{J}(dr,\omega_\varphi)$. \end{lemma} The proof of the above lemma takes up the rest of this section. As we will now explain, it is sufficient to prove the lemma under the assumption that every $C \in \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z') ) $ is irreducible: We can write $P_{\alpha, Z}$ and $P_{\beta, Z'}$ as concatenations \begin{align*} P_{\alpha, Z} = P_{\mathrm{in}} P_{\alpha_1, Z_1} P_{\mathrm{fin}},\\ P_{\beta, Z'} = P_{\mathrm{in}} P_{\beta_1, Z_1'} P_{\mathrm{fin}}, \end{align*} where $P_{\mathrm{in}}$ and $P_{\mathrm{fin}}$ correspond to the (possibly empty) trivial subregions between $P_{\alpha, Z}$ and $P_{\beta, Z'}$, and $P_{\alpha_1, Z_1}, P_{\beta_1, Z_1'} $ correspond to the non-trivial subregion; here we are using the terminology of Section \ref{sec:curves-corner-rounding}. The concave path $P_{\alpha_1, Z_1}$ is obtained from $P_{\beta_1, Z_1'} $ by rounding a corner and locally losing one $h$, and the region between $P_{\alpha_1, Z_1}$ and $P_{\beta_1, Z_1'} $ consists of a single non-trivial subregion. \begin{claim}\label{cl:every_curve_irreducible} Every current $C_1 \in \mathcal{M}_{J}( (\alpha_1,Z_1) , (\beta_1,Z'_1) )$ is irreducible. \end{claim} \begin{proof} By the structure of the corner rounding operation, $C_1$ has at most two negative ends. Thus, by degree considerations it must have at most two irreducible components (closed components can not exist for various reasons, for example because the curve has $I = 1$, see Proposition \ref{prop:structure_Jcurves}); and, if it has exactly two irreducible components, then it must have exactly two negative ends, with one component corresponding to each end; assume this for the sake of contradiction, and write the components as $D_0$ and $D_1$. By Lemma~\ref{lem:above}, each of the $D_i$ must correspond to a region between concave paths that do not go above $P_{\alpha_1,Z_1}$, and do not go below $P_{\beta_1,Z'_1}$. In fact, there are no such concave lattice paths, other than $P_{\alpha_1,Z_1}$ and $P_{\beta_1,Z'_1}$ because $P_{\alpha_1,Z_1}$ is obtained from $P_{\beta_1,Z_1'}$ by rounding a corner. Therefore, the region corresponding to each $D_i$ has a lower edge corresponding to one of the two edges of $P_{\beta_1, Z'_1}$ and the upper edges of each region must be on the edges of $P_{\alpha_1, Z_1}$. It follows from the concavity of the paths that each region must have $v \geq 1$, where $v$ is defined as in Lemma \ref{lem:fredholm_index}. Hence, by the Fredholm index formula \eqref{eqn:Fredholmindex}, each $D_i$ must have index at least one and so $C_1$ must have index $2$ which is not possible for generic $J$, by \eqref{eqn:indexinequality}. \end{proof} Next, note that as a consequence of Proposition \ref{prop:structure_Jcurves}, every curve $C \in \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z') )$ can be written as a disjoint union \begin{equation}\label{eqn:more_structure} C = C_{\mathrm{in}} \sqcup C_1 \sqcup C_{\mathrm{fin}}, \end{equation} where $C_1$ is the irreducible component of $C$ and $C_{\mathrm{in}}, C_{\mathrm{fin}}$ are unions of covers of trivial cylinders. It follows from Lemma \ref{lem:slice} and the equality case of Lemma \ref{lem:positivity} that $C_{\mathrm{in}}, C_{\mathrm{fin}}$ correspond to the orbits in $P_{\mathrm{in}}, P_{\mathrm{fin}}$ and that $C_1 \in \mathcal{M}_{J}( (\alpha_1,Z_1) , (\beta_1,Z'_1) ).$ Combining Claim \ref{cl:every_curve_irreducible}, and Equation \eqref{eqn:more_structure} in view of the discussion in the previous paragraph, we obtain a canonical bijection $$\mathcal{M}_{J}( (\alpha,Z) , (\beta,Z') ) \stackrel\sim\to \mathcal{M}_{J}( (\alpha_1,Z_1) , (\beta_1,Z'_1) )$$ given by removal of covers of trivial cylinders. We conclude from the above discussion that it is indeed sufficient to prove Lemma \ref{lem:converse} under the assumption that, for generic admissible $J$, every $C \in \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z') ) $ is irreducible. In terms of our combinatorial model, this assumption is equivalent to requiring that the region between $P_{\alpha,Z}$ and $P_{\beta, Z'}$ consists of one non-trivial subregion and no trivial subregions. This will be our standing assumption for the rest of this section. \subsubsection{Deformation of $J$} \label{sec:var} The next ingredient, which we will need to prove the Lemma \ref{lem:converse}, allows us to deform $J$ within the class of weakly admissible almost complex structures, while keeping the count of curves the same. The method of proof is inspired by the proofs of Lemmas 3.15 and 3.17 of \cite{Hutchings-Sullivan-Dehntwist}. Recall our standing assumption that the region between $P_{\alpha,Z}$ and $P_{\beta, Z'}$ consists of one non-trivial subregion and no trivial subregions. \begin{lemma} \label{lem:comp} Let $\varphi_0$ be a nice perturbation of $\varphi^1_H$, where $H \in \mathcal{D}$. Let $J_0$ be an admissible and $J_1$ a weakly admissible almost complex structure. Assume that $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$. Then, if $J_0$ and $J_1$ are generic, \[ \# \mathcal{M}_{J_0}( (\alpha,Z) , (\beta,Z') ) = \# \mathcal{M}_{J_1}( (\alpha,Z), (\beta,Z')).\] \end{lemma} \begin{remark}\label{rem:J_defined_on_X1} We will apply the above lemma in a setting where $J_0, J_1$ are only defined on $\R \times X_1$, where $X_1$ is a subset of $Y_\varphi$ with the following property: There exists a compact subset $K \subset X_1$ such that for any weakly admissible almost complex structure $J$ on $\R \times Y_\varphi$, every $C \in \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z') )$ is contained in $\R \times K$. Because all $J$--holomorphic curves are contained in $\R \times K$, with $K\subset X_1$ compact, the proof we give below for Lemma \ref{lem:comp} works verbatim in the setting of the previous paragraph as well. However, for clarity of exposition we give the proof in the setting were $J_0, J_1$ are globally defined. \end{remark} To prove the above lemma, we will need the following claim which will be used in the proof below and the next section. \begin{claim} \label{cl:curves_dont_go_over_both_poles} Suppose that $P_{\alpha, Z}$ and $P_{\beta, Z'}$ are as in Lemma \ref{lem:comp} and recall the definitions of $P^{z_0}_{\alpha,Z},P^{z_0}_{\beta,Z'}$ from Lemma \ref{lem:slice}. Let $z_{\min}$ and $z_{\max}$ be the minimum and maximum values of $z_0\in[-1, 1]$ such that $P^{z_0}_{\alpha,Z} - P^{z_0}_{\beta,Z'} \neq 0$. Then, either $-1 < z_{\min}$ or $z_{\max} < 1$. \end{claim} \begin{proof}[Proof of Claim] Denote by $(0, y_\alpha)$ and $(0, y_\beta)$ the starting points of $P_{\alpha, Z}$ and $P_{\beta, Z'}$, respectively, and by $(d, w_\alpha)$ and $(d, w_\beta)$ their endpoints. We begin by supposing $z_{\min} = -1$, and we will show this entails $z_{\max} < 1$. By Lemma \ref{lem:slice}, if $z_{\min} = -1$ then at least one of the following two scenarios must hold: First, the path $P_{\beta, Z'}$ begins with a horizontal edge $(1,0)$. Second, $y_\alpha > y_\beta$. In the first scenario, $P_{\beta, Z'}$ must have a second edge $(q,p)$ labelled $h$; this edge corresponds to some $z_{p/q} < 1$. Note that we must also have $w_\beta = w_\alpha$, since $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$. It then follows that $P^{z_0}_{\alpha,Z} - P^{z_0}_{\beta,Z'} = 0$ for $z_{p/q} < z_0$ and so $z_{\max} < 1$. Similarly, in the second scenario, $P_{\beta, Z'}$ has only one $(q,p)$ and this edge is labelled $h$. Since it is labelled by $h$, this edge must correspond to some $z_{p/q} < 1$. As in the first scenario, we must also have $w_\beta = w_\alpha$. It then follows that $P^{z_0}_{\alpha,Z} - P^{z_0}_{\beta,Z'} = 0$ for $z_{p/q} < z_0$ and so $z_{\max} < 1$. The case where $z_{\max} =1$ is similar to above and hence, we will not provide a proof. \end{proof} \begin{proof}[Proof of Lemma \ref{lem:comp}] Any curve $C$ in $\mathcal{M}_{J_0}$ or $\mathcal{M}_{J_1}$ has Fredholm index $1$, and so it follows from Equation \eqref{eqn:Fredholmindex} that the genus of $C$ must be zero; this is because $2e_- + h+ 2v \geq 2$ in our setting. We next argue as in the proof of Lemma 3.17 of \cite{Hutchings-Sullivan-Dehntwist}. As we explained in Section \ref{sec:admissible_vs_weakly}, we can connect $J_0$ and $J_1$ with a smooth family of weakly admissible almost complex structures $J_s, s\in [0,1]$. Consider the moduli space $\mathcal{M} := \cup_s \mathcal{M}_{J_s}((\alpha, Z), (\beta, Z'))$, for a generic choice of $J_s, s\in [0,1] $. There exists a global bound on the energy of curves in $\mathcal{M}$: indeed, any two $C, C'$ are homologous, as elements of $H_2(Y_\varphi, \alpha, \beta),$ and so have the same energy\footnote{ More precisely, this argument shows that, in the language of \cite{BEHWZ}, $C,C'$ have the same $\omega$-energy. It then follows from Proposition 5.13 of \cite{BEHWZ} that there exists a global bound on the $\lambda$-energy of the curves in $\mathcal{M}$, as well.} $\int_C \omega_\varphi = \int_{C'} \omega_\varphi$. Moreover, these curves all have genus zero as we explained in the previous paragraph. Hence, we can appeal to the SFT compactness theorem\footnote{The SFT compactness theorem holds in our setting where we are allowing the stable Hamiltonian structures to vary smoothly; see, for example, \cite[page 170]{Wendl-Notes}.} from \cite{BEHWZ} to conclude that if a degeneration of the moduli space $\mathcal{M}$ occurs at some $s$, then there is convergence to a broken $J_s$-holomorphic building $(C_0,\ldots,C_k)$; here, $C_i$ is the $i^{th}$ level of the building and it is a $J$-holomorphic curve between PFH generators $(\alpha_i, Z_i)$ and $(\alpha_{i+1}, Z_{i+1})$ with $(\alpha_0, Z_0) = (\alpha, Z)$ and $(\alpha_{k+1}, Z_{k+1}) = (\beta, Z')$. This building is in the homology class $Z-Z'$, has genus $0$, and the top and bottom levels must have at most one end at any Reeb orbit, by the partition conditions provided by Theorem 1.7 and Definition 4.7 in \cite{Hutchings-index} (here we are using the fact that the monodromy angles of our elliptic periodic orbits are close to zero and negative, and that since the curve is supposed irreducible, the orbits constituting $\beta$ have multiplicity $1$); the compactness theorem \cite{BEHWZ} guarantees that it is {\bf stable} and possibly {\bf nodal}; we will define these terms and elaborate on them below; each $C_i$ is also nontrivial, in the sense that no $C_i$ is a union of trivial cylinders Recall that a {\bf nodal} $J$-holomorphic curve $C$ is an equivalence class of tuples $(\Sigma,j, u, \Delta)$, where $(\Sigma,j,u)$ is a possibly disconnected $J$-holomorphic curve, and $\Delta \subset \Sigma$ is a finite set of points with even cardinality, called {\bf nodal points}; the set $\Delta$ has an involution $i$ with no fixed points, and $u(z) = u(i(z))$ for any $z \in \Delta$; see \cite[Sec.\ 2.1.6]{otherwendl} for the precise defintion. We next recall the stability conditions mentioned above. A nodal curve $(\Sigma,j,u,\Delta)$ is called {\bf stable} if every connected componet of $\Sigma - \Delta$ on which $u$ is constant has negative Euler characteristic. The compactness theorem of \cite{BEHWZ} guarantees that the curves $C_i$ in our building are all stable. To proceed, we will need the following claim. \begin{claim} \label{cl:no_nodes_no_ghosts} Each of the curves $C_i$, appearing in our building $(C_0, \ldots, C_k)$, has the following properties: \begin{enumerate} \item $C_i$ has no irreducible component whose domain is closed, no irreducible component which is constant, and no nodal points. \item Irreducible components of $C_i$ have Fredholm index $0$ or $1$. \end{enumerate} Moreover, we have \begin{equation} \label{eqn:sumind} \sum_i \text{ind}(C_i) = 1. \end{equation} \end{claim} We postpone proving the above claim and continue with our proof of Lemma \ref{lem:comp}. By Lemma~\ref{lem:positivity}, any $C_i$ must correspond to a region between paths that do not go above $P_{\alpha,Z}$, and do not go below $P_{\beta,Z'}$. In fact, there are no such concave lattice paths, other than $P_{\alpha,Z}$ and $P_{\beta,Z'}$ because $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ by rounding a corner. It follows that there exists $0 \leq j < k $ such that the paths $P_{\alpha_0, Z_0}, \ldots, P_{\alpha_j, Z_j}$ coincide with $P_{\alpha,Z}$ and the remaining paths $P_{\alpha_{j+1}, Z_{j+1}}, \ldots, P_{\alpha_k, Z_k}$ coincide with $P_{\beta,Z'}$, as {\it unlabeled} lattice paths. In other words, the curve $C_j$ is the unique curve in our building whose asymptotics correspond to two distinct unlabeled lattice paths. As a result of the last item of Claim \ref{cl:no_nodes_no_ghosts}, in combination with \eqref{eqn:sumind}, % we learn that one of the irreducible curves in our building must have index $1$, and all other irreducible curves must have index $0$. This conclusion about indices in turn implies that the curve $C_j$, whose asymptotics correspond to two distinct unlabeled lattice paths, must be irreducible. Indeed, we can repeat the argument given in the proof of Claim \ref{cl:every_curve_irreducible} to conclude that if $C_j$ were not irreducible it would then have index $2$. Next, we claim that every irreducible curve in the building must be a cylinder, other than $C_j$. To see this, let $P_{\alpha,Z}$ have $k$ edges and $P_{\beta,Z'}$ have $k'$ edges. Remember that since the curve is irreducible, $\beta$ admits either only $1$ edge or $2$ distinct edges. Then, by the partition condition considerations mentioned above, the Euler characteristic of the building must be $-k - k'+2$. Thus, the sum of the Euler characteristics of all curves in the building must also be $-k-k'+2$. Each curve must have at least one positive end and one negative end, so the Euler characteristic of each curve must be non-positive. And, since $C_j$ is irreducible, with ends at $k+k'$ distinct orbits, the Euler characteristic of $C_j$ must be at most $2-k-k'$. It follows that every component other than $C_j$ must have Euler characteristic $0$, hence must be a cylinder; and, $C_j$ has exactly one end at each of its Reeb orbits. For, $i\neq j$, the slice class is always zero, and therefore, $C_i$ consists of cylinders which are either trivial or local; see Remark \ref{rem:locality}. We proved above that each irreducible component of our building, and in particular these cylinders, can have Fredholm index $0$ or $1$. We will now show that the only non-trivial cylinders which could exist are local cylinders of Fredholm index $1$ with a positive end at a hyperbolic orbit and a negative end at an elliptic orbit. First, observe that any local cylinder with both ends at the same orbit must in fact be trivial; see for example \cite[Prop 9.1]{Hutchings-index}. Thus, a non-trivial cylinder must have one elliptic end and one hyperbolic end, and, by Formula \eqref{eqn:Fredholmindex}, for the Fredholm index to be non-negative, the positive and negative ends must be, respectively, hyperbolic and elliptic. The index will then be $1$. Since the sum of the indices of the curves $C_i$ is $1$, by \eqref{eqn:sumind}, % we therefore learn that if the building is non-trivial, then it must have a single non-trivial index $1$ local cylinder, with a positive end at a hyperbolic orbit, and a negative end at an elliptic orbit, and the curve $C_j$ must have index $0$, and so it must have two negative ends, both at hyperbolic orbits, and all positive ends at elliptic orbits; all other curves in the building must be trivial cylinders. Thus the building must have only two levels, and in the level structure, $C_j$ could be either on top or on bottom. In either case, as observed in \cite[Lemma 3.15]{Hutchings-Sullivan-Dehntwist}, there are two such canceling non-trivial local cylinders, so standard gluing arguments show that the mod $2$ count is unaffected by this degeneration. To complete our proof of Lemma \ref{lem:comp}, it remains to prove Claim \ref{cl:no_nodes_no_ghosts}. \begin{proof}[Proof of Claim \ref{cl:no_nodes_no_ghosts}] We begin by showing that $C_i$ has no non-constant irreducible component whose domain is closed. Let $z_{\min}$ and $z_{\max}$ be as in Claim \ref{cl:curves_dont_go_over_both_poles}. According to Lemma \ref{lem:slice} and Remark \ref{rem:positivity_irreducible}, the curves $C_i$ are all contained in $$\mathcal{Y}= \{(s,t,\theta, z) \in \R \times \S^1 \times \S^2: z_{\min} \leq z \leq z_{\max}\}.$$ If a non-constant closed curve was formed it would be contained in the above set $ \mathcal{Y}$ and therefore, it would be null homologous because, by Claim \ref{cl:curves_dont_go_over_both_poles}, either $-1 < z_{\min}$ or $z_{\max} <1$. However, closed non-trivial pseudo-holomorphic maps are never null homologous. We next prove that every non-constant irreducible curve in our building must have Fredholm index $0$ or $1$. To see why, first note that Equation \eqref{eqn:Fredholmindex} applies to any such curve $C$; since any such $C$ has both positive and negative ends, $C$ must have index at least $-1$, and the only way for $C$ to have index $-1$ is for $C$ to have exactly one negative end at a hyperbolic orbit, all positive ends at elliptic orbits, and $v = g = 0$. % If $C$ is such a curve, then by Equation \eqref{eqn:Fredholmindex} the index of any somewhere injective curve that it covers must be $-1$: such somewhere injective curves do not exist, even in (generic) one parameter families of $J$, as observed in \cite[Lemma 3.15]{Hutchings-Sullivan-Dehntwist}, because this is the index before modding out by translation. To finish the proof of the claim, it remains to show that the curves $C_i$ have no irreducible constant components and no nodal points, and then show that this implies \eqref{eqn:sumind}. % To do so it will be helpful to visualize a nodal $J$-holomorphic curve by associating each nodal point $z$ with $i(z)$, and we can think of the two associated points as a single node. Given a nodal $J$-holomorphic curve with domain $\Sigma$, we can form a new oriented topological surface $\widehat{\Sigma}$ by replacing each nodal point with a circle, and gluing these circles together; we refer the reader to \cite[Fig. 2.2]{otherwendl} for a helpful illustration, and we refer to \cite[Sec. 2]{otherwendl} for more details. An important identity which we will use below is that \begin{equation} \label{eqn:node} \chi(\widehat{\Sigma}) = \sum_i \chi(\Sigma_i) - N, \end{equation} where $\chi$ denotes the Euler characteristic, the sum is over all connected components $\Sigma_i$ of $\Sigma$, and $N$ denotes the number of nodal points, i.e.\ $N := \# \Delta$. % When we associate $\widehat{\Sigma}$ to a nodal curve $C$, which is an equivalence class, we write it as $\widehat{C}$; we should regard this as an equivalence class of topological surfaces, all of which are homeomorphic. The significance of the construction of the surface $\widehat{\Sigma}$ for us is as follows. Let $C \in \mathcal{M}$ be some curve that is close to breaking into the building $(C_0, \ldots, C_k)$ in the sense of \cite{BEHWZ}. The compactness result \cite{BEHWZ} then guarantees that \[ c_{\tau}(C) = \sum^k_{i=0} c_{\tau}(C_i), \quad \chi(C) = \sum^{k}_{i=0} \chi( \widehat{C_i} ).\] In particular, using this equation, the additivity of the terms in \eqref{eqn:indexformula_Fredholm}, and the relation \eqref{eqn:node}, we learn from \eqref{eqn:node} that \begin{equation} \label{eqn:keyindexresult} \text{ind}(C) = \sum^k_{i=0} \text{ind}(C_i) + N. \end{equation} For later use, it will be convenient to call the term $\chi(C)$ the {\bf Euler characteristic of the building} $(C_0,\ldots,C_k);$ we similarly call $\text{ind}(C)$ the {\bf index of the building}. These terms do not depend on the choice of $C \in \mathcal{M}$ that is close to breaking into the building $(C_0, \ldots, C_k)$ We now conclude from \eqref{eqn:keyindexresult} that $N = 0$, in other words nodal degenerations can not occur; note that \eqref{eqn:sumind} will also be an immediate consequence of this in view of \eqref{eqn:keyindexresult}. To see why $N = 0$, first observe that in the present situation, we have $\text{ind}(C) = 1$ in \eqref{eqn:keyindexresult}. We also showed that the index is nonnegative for nonconstant curves; and, we know that $N$ is even. Hence, we will be done with the proof of the claim if we rule out the possibility of $C_i$ having constant irreducible components. In the present situation, the space $\mathcal{M}$ consists of $g = 0$ curves, so each $\Sigma$ must have $g = 0$, and hence any constant component must have at least three nodal points, and index $-2$ by \eqref{eqn:indexformula_Fredholm}. Thus, any constant component must contribute at least $1$ to the right hand side of \eqref{eqn:keyindexresult}. Since we have now shown that every component must contribute nonnegatively to the right hand side of \eqref{eqn:keyindexresult}, then it follows that there must be exactly one component that contributes $1$, and the rest must contribute $0$. The component contributing $1$ to the sum can not be constant, since then the component would have to have exactly three nodes, and no other components could have any nodes at all; but the number of nodes is even. It therefore follows, again using the fact that $N$ is even, that in fact we must have $N = 0$, and no constant components. This completes the proof of the claim. \end{proof} \end{proof} \subsubsection{Existence of $J$--holomorphic curves}\label{sec:curves_exist} In this section, we complete the proof of Lemma \ref{lem:converse} by showing that $\# \mathcal{M}_{J_1}( (\alpha,Z) , (\beta,Z') ) =1$ for generic weakly admissible $J_1$. Below, we will introduce an open subset $X_1$ of $Y_\varphi$ satisfying the conditions of Remark \ref{rem:J_defined_on_X1}. More specifically, it will satisfy the following property: There exists a compact subset $K \subset X_1$ such that for any weakly admissible almost complex structure $J$ on $\R \times Y_\varphi$, every $C \in \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z') )$ is contained in $\R \times K$. We will then prove the following lemma. \begin{lemma}\label{lem:curves_exist} For generic weakly admissible almost complex structure $J_1$ on $\R \times X_1,$ $$\#\mathcal{M}_{J_1}( (\alpha,Z) , (\beta,Z') ) =1.$$ \end{lemma} We now explain why the above lemma completes the proof of Proposition \ref{prop:model}. Let $J_0$ be a generic and admissible almost complex structure on $Y_{\varphi}$, and connect its restriction to $K$ to $J_1$ via a generic smooth path $J_s$ of weakly admissible almost complex structures. Then, appeal to Lemma \ref{lem:comp}, Lemma \ref{lem:curves_exist} and Remark \ref{rem:J_defined_on_X1} to conclude that, counting mod 2, we have \[ \# \mathcal{M}_{J_0}( (\alpha,Z) , (\beta,Z') ) = \# \mathcal{M}_{J_1}( (\alpha,Z) , (\beta,Z') ) = 1,\] which implies that $\langle \partial (\alpha, Z), (\beta, Z')\rangle =1.$ \medskip It remains to prove Lemma \ref{lem:curves_exist}. This will occupy the remainder of this section. We begin by recalling certain preliminaries which will be used in the course of the proof. Let $\Omega \subset \mathbb{R}^2$ be a subset of the first quadrant, and consider the set \[ X_{\Omega} := \lbrace (z_1,z_2) | \pi (|z_1|^2, |z_2|^2) \in \Omega \rbrace \subset \mathbb{C}^2 = \mathbb{R}^4.\] When $\Omega$ is the region bounded by the axes and the graph of a function $f$ with $f'' \leq 0$, then we call $X_{\Omega}$ a {\bf convex toric domain}. When $\Omega$ is the region bounded by the axes and the graph of a function $f$ with $f'' \geq 0$, then we call $X_{\Omega}$ a {\bf concave toric domain}. Much work has been done about these domains, see \cite{C-CG-F-H-R, Hutchings_beyond, CG_concave_convex}. The boundary $\partial X_\Omega$ of a concave or convex toric domain is a contact manifold with the contact form $\lambda$ being the restriction to $\partial \Omega$ of the standard one-form on $\R^4$ \[ \frac{1}{2}(x_1 dy_1 - y_1 dx_1 + x_2 dy_2 - y_2 dx_2). \] Recall from Section \ref{sec:prelim_Jcurves} that $\mathbb{R} \times \partial X_\Omega$ is referred to as the {\it contact symplectization} and its symplectic form is given by $\omega = d(e^s \lambda)$, where $s$ denotes the coordinate on $\R$. In the argument below, we will be considering $J$--holomorphic curves in $\R \times \partial X_\Omega$ where $J$ is admissible for the SHS determined by the contact structure. Here, we will refer to such $J$ as {\bf contact admissible}. \medskip We now begin the proof of Lemma \ref{lem:curves_exist}. \begin{proof} Recall the definitions of $P^{z_0}_{\alpha,Z},P^{z_0}_{\beta,Z'}$ from Lemma \ref{lem:slice}. Let $z_{\min}$ and $z_{\max}$ be the minimum and maximum value of $z_0\in[-1, 1]$ such that $P^{z_0}_{\alpha,Z} - P^{z_0}_{\beta,Z'} \neq 0$. In view of Claim \ref{cl:curves_dont_go_over_both_poles}, either $-1 < z_{\min}$ or $z_{\max} < 1$. We will treat the two cases separately. \noindent \textbf{Case 1: $z_{\min}>-1$.} % Define $X_1:= \{(t, \theta, z) \in \S^1 \times \S^2: -1< z \}.$ Let $J$ be a weakly admissible almost complex structure on $\R \times Y_\varphi$ and denote by $M_J(\alpha, \beta; X_1)$ the set of % $J$-holomorphic curves from $\alpha$ to $\beta$ with image contained in $\R \times X_1$, modulo $\R$-translation. We claim that \begin{equation}\label{eqn:curves_in_X1} \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z')) = M_J(\alpha, \beta; X_1). \end{equation} To prove the above, first note that any $J$--holomorphic curve from $(\alpha, Z)$ to $(\beta, Z')$ is contained in $\R \times X_1$ because we have $[C_z] \neq 0$ only if $ z \geq z_{\min} > -1$; see Remark \ref{rem:positivity_irreducible}. This proves $\mathcal{M}_{J}( (\alpha,Z) , (\beta,Z')) \subset M_J(\alpha, \beta; X_1)$. As for the other inclusion, observe that, because $P_{\alpha, Z}$ is obtained from $P_{\beta, Z'}$ by corner rounding, $Z-Z'$ is an element of $H_2(Y_\varphi, \alpha, \beta)$ which can be represented by a chain in $X_1$. Moreover, it is the only such element of $H_2(Y_\varphi, \alpha, \beta)$ as other elements are of the form $[Z-Z'] + k [\S^2]$, where $k\in \Z$ is nonzero. This proves $M_J(\alpha, \beta; X_1) \subset \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z'))$. Observe that the previous paragraph implies, in particular, that $X_1$ has the property stated in Remark \ref{rem:J_defined_on_X1} with $K = \{(t, \theta, z) : z \geq z_{\min}\}$. We will next identify $X_1$ with a subset of the concave toric domain $X_\Omega$ defined below. Consider the concave toric domain $X_\Omega$, where $\Omega$ is the region bounded by the axes and the graph of $f(x)$, where $f(x) := \frac{h(1-2x)}{2}$ for $0 \leq x \leq 1$. The boundary $\partial X_\Omega$ is a contact manifold as described above. Consider the subset of $\partial X_\Omega$ given by $$ X_2 : = \lbrace (z_1,z_2) | \pi (|z_1|^2, |z_2|^2) \in \partial \Omega- \{(1, 0)\} \rbrace. $$ Note that this is $\partial X_\Omega$ with a Reeb orbit removed. Define the mappings \begin{align} \label{eqn:parametrization_concave} \begin{split} \psi:\ &X_1 \to X_2,\quad (t,\theta,z) \mapsto \left(\tfrac12(1-z),\theta,\tfrac12{h(z)},2 \pi t\right),\\ \Psi:\ &\R \times X_1 \to \R \times X_2,\quad (s,t,\theta,z) \mapsto \left(s,\tfrac12(1-z),\theta, \tfrac12{h(z)}, 2 \pi t\right). \end{split} \end{align} Here, we are regarding $\partial X_{\Omega} \subset \mathbb{C}^2$, and we are equipping $\mathbb{C}^2$ with coordinates $(\rho_1 := \pi |z_1|^2, \theta_1, \rho_2 := \pi |z_2|^2, \theta_2)$. Note that the standard-one form $\lambda$, defined above, is given in these coordinates by \begin{equation} \label{eqn:standardpolar} \lambda = \frac{1}{2\pi}(\rho_1 d \theta_1+\rho_2 d \theta_2). \end{equation} The above diffeomorphisms have the following properties: \begin{enumerate}[(i)] \item The Reeb vector field $R$ on $X_1$ pushes forward under $\psi$ to a positive multiple of the contact Reeb vector field $\hat R$ on $X_2$. \item The two-form $d\lambda$ on $X_2$ pulls back under $\psi$ to $\omega_{\varphi}$ on $X_1$. Thus, the SHS $(\lambda, d\lambda)$ on $X_2$ pulls back under $\psi$ to the SHS $(\psi^* \lambda, \omega_\varphi)$ on $X_1$. \end{enumerate} Item (i) above holds because at a point $(x, \theta_1, f(x), \theta_2)$, $R'$ is a positive multiple of \[ -f'(x) \partial_{\theta_1} + \partial_{\theta_2},\] see for example Eq. 4.14 in \cite{Hutchings-Notes}, while \[ R= \partial_t + 2\pi h'(z) \partial_\theta,\] by combining \eqref{eqn:trivializedreeb} and \eqref{eqn:Hvf}, and $f'(x) = - h'(z).$ Item (ii) holds because $d\lambda$ is the restriction of \[ \frac{1}{2\pi}(d\rho_1\wedge d \theta_1+d\rho_2\wedge d \theta_2),\] which pulls back to \[ \frac{1}{4 \pi} \left( d \theta \wedge dz + 2 \pi h'(z) dz \wedge dt \right),\] which is exactly $\omega_{\varphi}.$ \medskip Having established the above properties of the diffeomorphisms $\psi, \Psi$, we now proceed with the proof of Lemma \ref{lem:curves_exist}. By property (i), $\psi$ induces a bijection between the Reeb orbit sets of $R$ in $X_1$ and the Reeb orbit sets of $\hat R$ in $X_2$. We will denote the induced bijection by $$\alpha \mapsto \hat \alpha.$$ Now, suppose $P_{\alpha, Z}$ is obtained from $P_{\beta, Z'}$ via rounding a corner and locally losing one $h$. Let $\hat J$ be a contact admissible almost complex structure on the symplectization $\R \times \partial X_\Omega$ and consider $ \mathcal{M}_{\hat J}( \hat \alpha , \hat \beta)$ the space of $\hat J$-holomorphic currents $C$, modulo translation in the $\mathbb{R}$ direction, which are asymptotic to $\hat \alpha$ as $s \to + \infty$ and $\hat \beta$ as $s \to - \infty$. For a generic choice of contact admissible $\hat J$, this moduli space is finite and, as alluded to in Remark \ref{rem:ECH}, its mod $2$ cardinality determines\footnote{In general when defining the ECH differential, one would also demand the additional condition that the curves have ECH index $1$. However, it follows from the index calculations in \cite{Choi} that this is automatic given the asymptotics $\hat{\alpha}, \hat{\beta}$.} the ECH differential in the sense that $$\langle \partial_{ECH} \, \hat \alpha, \hat \beta \rangle = \# \mathcal{M}_{\hat J}( \hat \alpha , \hat \beta).$$ As we will explain below, it follows from results proven in \cite{Choi} that the following hold: \begin{enumerate} \item[A1.] The image of every curve in $ \mathcal{M}_{\hat J}( \hat \alpha , \hat \beta)$ is contained in $\R \times X_2$. \item[B1.] The mod 2 count of curves in $ \mathcal{M}_{\hat J}( \hat \alpha , \hat \beta)$ is $1$. \end{enumerate} We will now explain why A1 \& B1 imply Lemma \ref{lem:curves_exist}. Indeed, define $J_1$ to be the almost complex structure on $\R \times X_1$ given by the pull back under $\Psi$ of the restriction of a generic $\hat J$ as above. Then, $J_1$ is an almost complex structure on $\R \times X_1$ which is admissible for the SHS given by $(\psi^* \lambda, \psi^* d\lambda) = (\psi^* \lambda, \omega_\varphi)$; this SHS gives the same orientation as $(dr,\omega_{\varphi})$, by property (i) above, which means that $J_1$ is weakly admissible, see Section \ref{sec:admissible_vs_weakly}. Next, by property A1, $\Psi$ induces a bijection between $ \mathcal{M}_{\hat J}( \hat \alpha , \hat \beta)$ and $ \mathcal{M}_{J_1}( \alpha , \beta, X_1)$. Property B1 then implies that $\# \mathcal{M}_{J_1}( \alpha , \beta, X_1) = 1.$ Lastly, by Equation \eqref{eqn:curves_in_X1}, we have $$ \# \mathcal{M}_{J_1}( (\alpha,Z) , (\beta, Z')) = 1,$$ which proves Lemma \ref{lem:curves_exist} in the case $z_{\min} > -1$. Before elaborating on A1 \& B1, we will treat Case 2. \medskip \noindent \textbf{Case 2: $z_{\max}<1$.} The proof of Case 2 is similar, and in a sense dual, to that of Case 1. \medskip Define $X_1:= \{(t, \theta, z) \in \S^1 \times \S^2: z< 1 \}.$ Let $J$ be a weakly admissible almost complex structure on $\R \times Y_\varphi$ and denote by $M_J(\alpha, \beta; X_1)$ the set of $J$ holomorphic curves from $\alpha$ to $\beta$ with image contained in $\R \times X_1$, modulo translation. As in Case 1, we have \begin{equation}\label{eqn:curves_in_X1_case2} \mathcal{M}_{J}( (\alpha,Z) , (\beta,Z')) = M_J(\alpha, \beta; X_1). \end{equation} As before, $X_1$ has the property stated in Remark \ref{rem:J_defined_on_X1} with $K = \{(t, \theta, z) : z \leq z_{\max}\}$. We will next identify $X_1$ with a subset of the convex toric domain $X_\Omega$ defined below. Consider the convex toric domain $X_\Omega$, where $\Omega$ is the region bounded by the axes and the graph of $g(x) = \frac{h(1) - h(2x-1)}{2}$ for $0 \leq x \leq 1$. The boundary $\partial X_\Omega$ is a contact manifold as described above. Consider the subset of $\partial X_\Omega$ given by $$ X_2 : = \lbrace (z_1,z_2) | \pi (|z_1|^2, |z_2|^2) \in \partial \Omega- \{ (1,0) \} \rbrace. $$ Define the mappings \begin{align} \label{eqn:parametrization_convex} \begin{split} \psi:\ &X_1 \to X_2,\quad (t,\theta,z) \mapsto \left( \tfrac12(z+1), \theta, \tfrac12(h(1) - h(z)), 2 \pi t \right),\\ \Psi:\ &\R \times X_1 \to \R \times X_2, \quad (s,t,\theta,z) \mapsto \left( -s, \tfrac12(z+1), \theta, \tfrac12(h(1) - h(z)), 2 \pi t\right). \end{split} \end{align} Here, we are regarding $\partial X_{\Omega} \subset \mathbb{C}^2$, and we are equipping $\mathbb{C}^2$ with coordinates $(\rho_1 := \pi |z_1|^2, \theta_1, \rho_2 := \pi |z_2|^2, \theta_2)$. Similarly (but dual) to Case 1, these diffeomorphisms have the following properties: \begin{enumerate}[(i)] \item The Reeb vector field $R$ on $X_1$ pushes forward under $\psi$ to a positive multiple of the contact Reeb vector field $\hat R$ on $X_2$. \item The two-form $d\lambda$ on $X_2$ pulls back under $\psi$ to $-\omega_{\varphi}$ on $X_1$. Thus, the SHS $(\lambda, -d\lambda)$ on $X_2$ pulls back under $\psi$ to the SHS $(\psi^* \lambda, \omega_\varphi)$ on $X_1$. \end{enumerate} By property (i), $\psi$ induces a bijection $$\alpha \mapsto \hat \alpha,$$ between the Reeb orbit sets of $R$ in $X_1$ and the Reeb orbit sets of $\hat R$ in $X_2$. Now, suppose $P_{\alpha, Z}$ is obtained from $P_{\beta, Z'}$ via rounding a corner and locally losing one $h$. As in Case 1, let $\hat J$ be a contact admissible almost complex structure on $\R \times \partial X_\Omega$ and consider the moduli space $ \mathcal{M}_{\hat J}( \hat \beta , \hat \alpha)$ which for a generic choice of contact admissible $\hat J$ is finite and its mod $2$ cardinality determines\footnote{Just as in the previous case, the condition that the curves have ECH index $1$ is actually automatic here, by the asymptotics.} the ECH differential in the sense that $$\langle \partial_{ECH} \, \hat \beta, \hat \alpha \rangle = \# \mathcal{M}_{\hat J}( \hat \beta , \hat \alpha).$$ As we will explain below, it follows from results proven in \cite{Choi} that the following hold: \begin{enumerate} \item[A2.] The image of every curve in $ \mathcal{M}_{\hat J}( \hat \beta , \hat \alpha)$ is contained in $\R \times X_2$. \item[B2.] The mod 2 count of curves in $ \mathcal{M}_{\hat J}( \hat \beta, \hat \alpha)$ is $1$. \end{enumerate} We will now explain why A2 and B2 imply \ref{lem:curves_exist}. Define $J_1$ to be the almost complex structure on $\mathbb{R} \times X_1$ given by the pull back under $\Psi$ of the restriction % of $- \hat J$, for a generic $\hat J$ as above. Then, $J_1$ is an almost complex structure on $\mathbb{R} \times X_1$ which is admissible for the SHS given by $(\psi^* \lambda, \psi^* (- d \lambda)) = (\psi^* \lambda, \omega_{\varphi})$; this SHS has the same orientation % as $(dr,\omega_{\varphi})$, by property (i) above, which means that $J_1$ is weakly admissible; note that $-\hat{J}$ is admissible for the SHS given by $(-\lambda, - d \lambda)$. Next, by property A2, $\Psi$ induces a bijection between $\mathcal{M}_{- \hat{J} }( - \hat{\beta}, - \hat{\alpha} )$ and $\mathcal{M}_{J_1}(\alpha,\beta,X_1)$ (here, the notation $- \hat{\beta}$ means that the orientation is reversed, which we can regard as viewing $\hat{\beta}$ as an orbit set for $-\lambda$); and, there is a canonical bijection\footnote{For the convenience of the reader, we remark that this bijection is familiar from the ``charge conjugation invariance" property of ECH, see for example \cite{Hutchings-Notes}.} between $\mathcal{M}_{- \hat{J} } (- \hat \beta, - \hat \alpha)$ and $\mathcal{M}_{ \hat J} (\hat \beta, \hat \alpha)$ given by associating a $\hat J$-holomorphic curve \[ u: (\Sigma,j) \to (X_2,\hat J)\] to the $-\hat J$ holomorphic curve \[ u: (\Sigma,-j) \to (X_2, -\hat J).\] We can now repeat the rest of the argument given in Case 1 to see that A2 \& B2 imply Lemma \ref{lem:curves_exist}. As promised, we will next elaborate on properties A1, B1, from Case 1, and A2, B2, from Case 2. \subsubsection{Elaborations on properties A1, B1 and A2, B2}\label{sec:elaborations} For the benefit of the reader, we now elaborate on why A1, B1, A2, and B2 hold. We first explain items $A1$ and $A2$. We showed above, while proving Equations \eqref{eqn:curves_in_X1} and \eqref{eqn:curves_in_X1_case2}, that as a consequence of Lemma \ref{lem:positivity} any $J$--holomorphic curve from $(\alpha, Z)$ to $(\beta, Z')$ is contained in $\R \times X_1$. A reasoning similar to what we have given above, proves that items $A1, A2$ are consequences of \cite[Lemma 3.5]{Choi} which is analogous to our Lemma~\ref{lem:positivity}. We now explain items $ B1$ and $B2$, beginning with some context. The purpose of \cite{Choi} is to give a combinatorial realization of the ECH chain complex for any toric contact three-manifold. Here, we use Choi's results in the special case of concave and convex toric domains. It turns out that in these cases, the ECH chain complex has a particularly nice form, as was originally conjectured by Hutchings \cite[Conj. A.3]{Hutchings_beyond}. We first explain the ECH combinatorial model in the case where $X_\Omega$ is a convex toric domain, which appears in Case 2 above. Let $X_{\Omega}$ be such a domain, described by a smooth function $f: [0,a] \to [0,b]$. In this case, we can represent an ECH generator $\hat \alpha$ by an {\bf ECH convex lattice path} $P_{\hat \alpha}$: this is a piecewise linear lattice path that starts on the $y$-axis, ends on the $x$-axis, stays in the first quadrant, and is the graph of a concave function; we usually orient this convex lattice path ``to the right". Note that the region bounded by such a path and the coordinate axes is a convex region. Hutchings conjectured (\cite[Conj. A.3]{Hutchings_beyond}), and Choi proved (\cite[Cor. 4.15]{Choi}) that the ECH chain complex differential in this case has a particularly nice form: it is given by rounding a corner and locally losing one $h$. More precisely, we can augment any convex lattice path to a closed lattice path by adding edges along the axes, and then there is a nonzero differential coefficient $\langle \partial_{ECH} \hat \beta, \hat \alpha \rangle$ if and only if this augmentation of $P_{\hat \alpha}$ is obtained from the augmentation of $P_{\hat \beta}$ by rounding a corner and locally losing one $h$ The ECH combinatorial model in the case where $X_\Omega$ is concave, which appears in Case 1 above, is dual to this. We represent an ECH generator by an {\bf ECH concave lattice path}, which is defined analogously to the ECH convex lattice paths above, except that the path is the graph of a convex function instead; we augment by adding infinite rays along the axes. Then, by \cite[Prop. 4.14, Thm. 4.8]{Choi}, there is a nonzero differential coefficient $\langle \partial_{ECH} \hat \alpha, \hat \beta \rangle$ if and only if this augmentation of $P_{\hat \alpha}$ is obtained from the augmentation of $P_{\hat \beta}$ by rounding a corner and locally losing one $h$. We can now explain the proofs of $B1$ and $B2$. We begin with $B1$. In the assignation between an ECH orbit set $\hat \alpha$ and a concave lattice path $P_{\hat \alpha}$ in the combinatorial model of \cite{Choi}, an edge $(p,q)$ corresponds to a point $r$ on the graph of $f$ with slope $q/p$, or to the two points $(0,a)$ and $(b,0)$ where the graph of $f$ meets the axes. We also handle the elliptic orbits corresponding to $(0,a)$ and $(b,0)$ as in the PFH case: in the case of an $m$-fold cover of the elliptic orbit corresponding to $(0,a)$ (which is the only case relevant to the proof), we concatenate the line segment with slope $f'(0)$ and horizontal displacement $m$ to the beginning of the rest of the lattice path. Now, this does not yield a lattice path, because $f'(0)$ is irrational, and so we take the lowest lattice path above this concatenated path and then translate so that the path starts on the $y$-axis and ends on the $x$-axis. We now describe how the identification of the previous section between PFH and ECH orbit sets $\alpha \mapsto \hat \alpha,$ translates to an identification of the corresponding lattice paths $P_{\alpha, Z} \mapsto P_{\hat{\alpha}}:$ We take $P_{\alpha, Z}$ and translate it so that it starts on the $x$-axis and ends on the $y$-axis, and reflect across the $y$-axis, to get $P_{\hat \alpha}$, the ECH lattice path corresponding to $\hat{\alpha}$. Thus, since $P_{\alpha,Z}$ is obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$, the ECH lattice path $P_{\hat{\alpha}}$ is obtained from $P_{\hat{\beta}}$ by rounding a corner and locally losing one $h$, and so the above results apply to prove B1. The proof of $B2$ is similar. The identification $\alpha \mapsto \hat \alpha$ translates to an identification of the corresponding lattice paths $P_{\alpha, Z} \mapsto P_{\hat{\alpha}}:$ given a PFH lattice path $P_{\alpha, Z}$, we translate it so that it begins on the $y$--axis and ends on the $x$--axis, we then reflect it across the $x$-axis to get $P_{\hat{\alpha}}$. As with B1, this now implies B2. \end{proof} \begin{remark} \label{rmk:choi} To make our paper as self-contained as possible, while maintaining some amount of brevity, we sketch those parts of Choi's argument that are relevant for what we need. The first point to make is a historical one. Many of our arguments in Section \ref{sec:PFH_monotone_twist} are inspired by arguments in \cite{Choi}, which are themselves inspired by arguments in Hutchings-Sullivan; in particular, most of the ideas needed for Prop. 4.14 and Cor. 4.15 of \cite{Choi} have already been presented here, although we presented them in the PFH case rather than the ECH case. In particular, Choi proves analogues of all of the results in \ref{sec:comb} through \ref{sec:var}, and indeed his argument for reducing to considering regions given by locally rounding a corner and losing one $h$ is quite similar to the argument we give. So, the only differences really worth commenting on involve how Choi shows that such a region actually carries a nonzero count of curves --- here there are two main differences between our argument and Choi's that we should highlight. \begin{enumerate} \item Choi finds his curves by referencing a paper by Taubes \cite{Taubes-beasts}, which works out various moduli spaces of curves for a particular contact form on $\S^1 \times \S^2$. Taubes' contact form is Morse-Bott, so Choi does a perturbation as in [\cite{Hutchings-Sullivan-Dehntwist}, Lem. 3.17], [\cite{Hutchings-Sullivan-T3}, Thm. 11.11, Step 4] to break the Morse-Bott symmetry so as to obtain a nondegenerate contact form; Choi then does a deformation argument that is very analogous to what we do. We remark that this strategy was pioneered by Hutchings and Sullivan in the series of papers \cite{Hutchings-Sullivan-Dehntwist, Hutchings-Sullivan-T3}. In contrast, we find our curves by referencing Choi's paper rather than Taubes'. This means that we do not need any Morse-Bott argument. \item Choi uses an inductive argument to reduce to considering moduli spaces of twice and thrice punctured spheres. The reason he does this is to be able to use the above paper by Taubes, which does not directly address all the curves needed to analyze corner rounding operations, which could lead for example to curves with an arbitrary number of ends. This induction works by using the fact that the differential is already known to square to $0$, by \cite{Hutchings-TaubesI, Hutchings-TaubesII} --- once one shows the result for the curves one gets from Taubes, given an arbitrary region that one wishes to show corresponds to a nonzero count of curves, one proves that if it did not have a nonzero count, the differential could not possibly square to zero, essentially by concatenating with a region that can be analyzed through the Taubes curves. As with the above item, we again remark that this strategy was previously pioneered by Hutchings and Sullivan in \cite{Hutchings-Sullivan-Dehntwist, Hutchings-Sullivan-T3}. In contrast, we have no need to use this inductive argument, since we can reference Choi's work for our needed curves. \end{enumerate} As should be clear from the above, the main reason we have chosen to take a slightly different tactic from Choi concerning these two points is for brevity --- we could have also used a strategy like the above, if for some reason we had wanted to. Another point from the above is that the ideas in the part of Choi's paper relevant here already appear in the Hutchings-Sullivan papers \cite{Hutchings-Sullivan-Dehntwist, Hutchings-Sullivan-T3}: what is especially new and impressive in Choi's paper is the analysis of very general toric contact forms, (for example, the contact form on a toric domain that is neither concave nor convex) for which there could be curves corresponding to regions more general than those that come from rounding a corner and locally losing one $h$ --- however, we do not need to consider these particular kinds of contact forms in this paper. \end{remark} \section{The Calabi property for positive monotone twists}\label{sec:calabi_for_montone_twists} In this section we provide a proof of Theorem~\ref{thm:s2case}. We will first use the combinatorial model of PFH, from the previous section, to compute the PFH spectral invariants for monotone twists; this is the content of Theorem \ref{theo:spec_computation}. We will then use this computation to prove Theorem~\ref{thm:s2case}; this will be carried out in Section \ref{sec:proof_Calabi_property}. \subsection{Computation of the spectral invariants}\label{sec:comuting_spec_invariant} We will need to introduce some notations and conventions before stating, and proving, the main result of this section. Throughout this section, we fix $\varphi$ to be a (smooth) positive monotone twist map of the disc. Recall from Remark \ref{rem:PFHspec-disc} that we define PFH spectral invariants for maps of the disc by identifying $\Diff_c(\D, \omega)$ with maps of the sphere supported in the northern hemisphere $S^+ \subset \S^2$, where the sphere $\S^2$ is equipped with the symplectic form $\omega=\frac1{4\pi}d\theta\wedge dz$. Recall that every monotone twist map of the disc $\varphi$ can be written as the time--$1$ map of the flow of an autonomous Hamiltonian \[ H = \frac{1}{2} h(z),\] where $h: \S^2 \rightarrow \R$ is a function of $z$ satisfying \[ h' \geq 0, h'' \geq 0, h(-1) = 0, h'(-1) = 0.\] For the main result of this section, Theorem \ref{theo:spec_computation}, we will need to impose the additional assumption that \begin{equation}\label{eq:additional_assump} h'(1)\in \N. \end{equation} The reason for imposing the above assumption is that in our combinatorial model, Proposition \ref{prop:model}, the we assume that $h'(1)$ is assumed to be sufficiently close to $\lceil h'(1) \rceil$. Observe that every Hamiltonian $H$ as above can be $C^\infty$ approximated by the Hamiltonians considered in Section \ref{sec:ham}. Although $\varphi$ is degenerate, we can still define the notion of a {\bf concave lattice path for $\varphi$} as any lattice path obtained from a starting point $(0, y)$, with $y\in\Z$, and a finite sequence of consecutive edges $v_{p_i,q_i}$, $i=0, \dots, \ell$, such that: \begin{itemize} \item $v_{p_i,q_i}=m_{p_i, q_i}(q_i, p_i)$ with $q_i, p_i$ coprime, \item the slopes $p_i/q_i$ are in increasing order, \item we have $0\leq p_0/q_0$ and $p_\ell/q_\ell$ is at most $h'(1)$. % \end{itemize} If $p_0=0$, we will, as in Section \ref{sec:comb}, denote $v_-=m_-(1,0)= v_{p_0, q_0}$. If $p_\ell/q_\ell= h'(1)$, we will denote $v_+=m_+(1, h'(1) )= v_{p_l, q_l}.$ We also let $z_{p,q}$ be such that $h'(z_{p,q})=p/q$. We can also define the action of such a path just as in Section \ref{sec:comb}: We first define \begin{equation} \label{eqn:degenaction} \mathcal{A}(v_-) = 0, \quad \mathcal{A}(v_+) = m_+\frac{h(1)}2,\quad \mathcal{A}(v_{p,q}) = \frac{m_{p,q}}{2} ( p (1-z_{p,q}) + q h(z_{p,q})). \end{equation} We then define the action of a concave lattice path $P$ to be \begin{equation} \label{eq:action_formula_unlabelled} \mathcal{A}(P) = y + \mathcal{A}(v_+) % + \sum_{v_{p,q}} \mathcal{A}(v_{p,q}). \end{equation} The definition of $j(P)$ from Section \ref{sec:comb} (see Equation \eqref{eqn:index_path_unlabelled}) is still valid here. With this in mind, we have the following: \begin{theo}\label{theo:spec_computation} \label{thm:comb} Let $\varphi \in \Diff_c(\D, \omega)$ be a monotone twist satisfying the assumption \eqref{eq:additional_assump}. Then, for all integers $d>0$ and $k=d\mod 2$, \begin{equation} \label{eqn:spectralformula} c_{d,k}(\varphi) = \max \lbrace \mathcal{A}(P) : 2 j(P) - d = k \rbrace, \end{equation} where the max is over all concave lattice paths $P$ for $\varphi$ of horizontal displacement $d$. \end{theo} \begin{proof} We can take a $C^{\infty}$ small perturbation of $\varphi$ to a $d$--nondegenerate Hamiltonian diffeomorphism $\varphi_0$ which itself is a {\it nice perturbation} of some $\varphi_H^1$, where $H \in \mathcal{D}$ as in Section \ref{sec:ham}. Since $c_{d,k}(\varphi)$ is the limit of $c_{d,k}(\varphi_0)$, as we take smaller and smaller perturbations, it suffices to show that the analogous formula to \eqref{eqn:spectralformula} holds for $\varphi_0$. In other words, we wish to show \begin{equation} \label{eqn:perturbedformula} c_{d,k}(\varphi_0) = \max \lbrace \mathcal{A}(P_{\alpha,Z}) : I(P_{\alpha,Z}) = k \rbrace, \end{equation} where the max is over all concave lattice paths of horizontal displacement $d$. To prove \eqref{eqn:perturbedformula}, given $(d,k)$, consider the element $\sigma$ of the PFH chain complex for $\varphi_0$ given by \begin{equation*} \sigma := \sum (\alpha,Z) \end{equation*} where the sum is over all PFH generators $(\alpha,Z)$ where $\alpha$ consists of only elliptic orbits, is of degree $d$ and $I(\alpha, Z) = k$. Equivalently, the corresponding concave lattice path $P_{\alpha,Z}$ has edges which are all labelled $e$, it has degree $d$ and index $k$. We first claim that $\sigma$ is in the kernel of the PFH differential. Indeed, by Proposition \ref{prop:model}, the differential is the mod $2$ sum over every $(\beta, Z')$ such that $P_{\alpha,Z}$ can be obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$. Fix one such $P_{\beta,Z'}$. It has exactly one edge labelled $h$ and so there are exactly two concave paths, say $P_{\alpha,Z}$ and $P_{\tilde \alpha, \tilde Z}$, which are obtained from $P_{\beta,Z'}$ by rounding a corner and locally losing one $h$. The two paths $P_{\alpha,Z}$ and $P_{\tilde \alpha, \tilde Z}$ are different, because for example when you round the two corners for an edge, one rounding contains one of the corners, and the other contains the other corner. Now, $(\alpha, Z)$ and $(\tilde \alpha, \tilde Z)$ both contribute to $\sigma$ and thus, $(\beta,Z')$ appears exactly twice in the differential of $\sigma$; hence, its mod $2$ contribution to the differential is zero. Consequently, we see that $\sigma$ is in the kernel of the PFH differential. Now, by Proposition~\ref{prop:model}, no concave path with all edges labeled $e$ is ever in the image of the differential, because the concave path corresponding to the negative end of a holomorphic curve counted by the differential has more edges labeled $h$ than the concave path corresponding to the positive end, and in particular has at least one edge labeled $h$. So, $[\sigma] \ne 0$ in homology. In fact, $\sigma$ must carry the spectral invariant for similar reasons. Specifically, if there is some other chain complex element $\sigma'$ homologous to $\sigma$, then $\sigma + \sigma'$ must be in the image of the differential. Nothing in the image of the differential has a path with all edges labeled by $e$, so $\sigma'$ must contain all possible paths of degree $d$ and index $k$ with all edges labeled by $e$, and so its action must be at least as much as $\sigma$. In view of the above paragraph, to prove Equation \eqref{eqn:perturbedformula}, we must show that the supremum in the equation is attained by a path whose edges are labelled $e$. To see this, consider a path of degree $d$ and index $k$ with some edges labelled $h$. As a consequence of the combinatorial index formula in Proposition \ref{prop:model}, the number of edges labelled $h$ must be even; denote this number by $2r$. If we round $r$ corners and remove all $h$ labels, we obtain a path of the same grading all of whose edges are labelled $e$. Now, the newly obtained path has larger action because the corner rounding operation increases action. This completes the proof. \end{proof} \begin{remark}\label{rem:monotonicink} As a corollary to Theorem~\ref{thm:comb}, we find that \[ c_{d,k} \leq c_{d,k+2},\] as promised in Equation \eqref{eqn:monotonicink}. Indeed, if we take an action maximizing path of index $k$, with all edges labeled $e$ as above, and round a corner, then the grading increases by $2$, and the action increases as well. \end{remark} \subsection{Proof of Theorem \ref{thm:s2case} } \label{sec:proof_Calabi_property} We now give a proof of Theorem~\ref{thm:s2case}. It is sufficient to prove Theorem \ref{thm:s2case} for monotone twists $\varphi$ which can be written as $\varphi^1_H$ with the Hamiltonian $H$ satisfying \eqref{eq:additional_assump}. This is because the left and right hand sides of Equation \ref{eqn:calabiconjecture}, i.e.\ the Calabi invariant and the PFH spectral invariants, are (Lipschitz) continuous with respect to the Hofer norm and, moreover, every monotone twist can be approximated, in the Hofer norm, by monotone twists satisfying \eqref{eq:additional_assump}. Hence, we will suppose for the rest of this section that our monotone twists $\varphi$ satisfy \eqref{eq:additional_assump}. This allows us to apply Theorem \ref{theo:spec_computation}. Our proof relies on a version of the isoperimetric inequality for non-standard norms which we now recall; the idea of using this inequality is inspired by Hutchings' proof in \cite[Section 8]{QECH} of the ``Volume Property" for ECH spectral invariants for certain toric contact forms. Let $\Omega \subset \R^2$ be a convex compact connected subset. Using the standard Euclidean inner product, the dual norm associated to $\Omega$, denoted $|| \cdot ||^*_{\Omega}$, is defined for any $v\in\R^2$ by \begin{equation} \label{eqn:normmax} || v ||^*_{\Omega} = \max \{ v \cdot w: w \in \partial \Omega\}. \end{equation} Let $\Lambda \subset \R^2$ be an oriented, piecewise smooth curve and denote by $\ell_\Omega(\Lambda)$ its length measured with respect to $|| \cdot ||^*_{\Omega}$. When $\Lambda$ is closed, its length remains unchanged under translation of $\Omega$. For our proof, we will suppose that $\Omega$ is the region bounded by the graph of $h$, the horizontal line through $(1, h(1))$, and the vertical line through $(-1, 0)$. Denote by $\hat{\Omega}$ the region obtained by rotating $\Omega$ clockwise by ninety degrees; see Figure \ref{fig:isoperimetric1}. We orient the boundary $\partial \hat \Omega$ counterclockwise with respect to any point in its interior. \begin{figure}[h!] \centering \def1\textwidth{0.6\textwidth} \input{isoperimetric1.pdf_tex}\caption{The convex subsets $\Omega$, $\hat\Omega$.} \label{fig:isoperimetric1} \end{figure} \begin{proof}[Proof of Theorem \ref{thm:s2case}] Let $P$ be a concave lattice path of horizontal displacement $d$ for $\varphi$. Complete the path $P$ to a closed, convex polygon by adding a vertical edge at the beginning of $P$ and a horizontal edge at the end; orient this polygon counterclockwise, relative to any point in its interior; and, rotate it clockwise by ninety degrees. Call the resulting path $\Lambda$; see Figure \ref{fig:isoperimetric2}. We will need the following lemma. \begin{figure}[h!] \centering \def1\textwidth{0.6\textwidth} \input{isoperimetric2.pdf_tex}\caption{The path $P$ and the closed path $\Lambda$.} \label{fig:isoperimetric2} \end{figure} \begin{lemma}\label{lem:length_action} The following identities hold: \begin{enumerate} \item $\ell_{\Omega}(\partial\hat \Omega) = 2 (2h(1) - I)$, where $I : = \int_{-1}^1 h(z) dz$. \item $\ell_{\Omega}(\Lambda) = dh(1) + 2y + 2 V - 2 \mathcal{A}(P)$, where $V$ denotes the vertical displacement of the path $P$. % \end{enumerate} \end{lemma} \begin{proof}[Proof of Lemma \ref{lem:length_action}] According to the Isoperimetric Theorem \cite{Brothers-Morgan}, for any simple closed curve $\Gamma$, we have \begin{equation} \ell_{\Omega}^2(\Gamma) \geq 4 A(\Omega) A (\Gamma),\label{eq:isoperimetric} \end{equation} where $A(\Omega)$ and $A(\Gamma)$ denote the Euclidean areas of $\Omega$ and the region bounded by $\Gamma$, respectively. Moreover, equality holds when $\Gamma$ is a scaling of a ninety degree clockwise rotation of $\partial \Omega$; see \cite[Example 8.3]{QECH}. The first item follows immediately from the equality case of the theorem applied to $\Gamma = \partial \hat \Omega$ because $A(\Gamma) = A(\Omega) = 2h(1) - I$. Alternatively, Item 1 could be obtained via direct computation. We now prove the second item. The length of the polygon $\Lambda$ is given by the sum $ \sum_{e \in \Lambda} || e ||^*_{\Omega},$ where the sum is taken over the edge vectors $e$ of $\Lambda$. It follows from the method of Lagrange multipliers that $$|| e ||^*_{\Omega} = e \cdot p_e,$$ for some point $p_e \in \partial \Omega$ where $e$ points in the direction of the outward normal cone at $p_e$. Hence, we can write \begin{equation} \label{eq:length} \ell_{\Omega}(\Lambda) = \sum_{e \in \Lambda} e \cdot p_e = \sum_{e \in \Lambda} e \cdot ( p_e - m), \end{equation} where the second equality holds, for any $m \in \R^2$, because $\Lambda$ is closed. We will calculate $\ell_{\Omega}(\Lambda)$ using the choice $m=(1,0)$. Let $e$ denote one of the edges of $\Lambda$ corresponding to a vector $v_{p,q} = m_{p, q} (q,p) $ in $P$. Now, we have $e = m_{p,q} (p, -q),$ since we are taking a ninety degree clockwise rotation; moreover, $ p_e - m = (z_{p,q} -1, h(z_{p,q}))$. Thus, \begin{equation*} \begin{split} e \cdot (p_e -m) =& m_{p,q} (p, -q) \cdot (z_{p,q} -1, h(z_{p,q})) \\ =& m_{p,q} \left( p (z_{p,q} - 1 ) - q h(z_{p,q}) \right) \\ =& - 2 \mathcal A ( v_{p,q} ), \end{split} \end{equation*} where the final equation follows from \eqref{eqn:degenaction}. If $e$ is an edge of $\Lambda$ corresponding to either of the vectors $v = v_- = m_- (1,0)$ or $v = v_+ = m_+ (1, \lceil h'(1) \rceil)$, then a similar computation to the above yields $e \cdot (p_e -m) = - 2 \mathcal A ( v )$. Summing over all of the edges $e$ of $\Lambda$, corresponding to vectors in $P$, we obtain the quantity \begin{equation}\label{eq:action_1} 2y- 2 \mathcal{A} (P). \end{equation} The remaining two edges of $\Lambda$ are the vectors $e_1 = (0, d )$ and $e_2 = (-V,0)$ for which we have \begin{equation}\label{eq:action_2} \begin{split} e_1 \cdot (p_{e_1} -m) &= (0, d) \cdot (-1, h(1)) = dh(1), \\ e_2 \cdot (p_{e_2} -m) & =(- V, 0) \cdot (-2, 0) = 2V . \end{split} \end{equation} We obtain from Equations \eqref{eq:length}, \eqref{eq:action_1}, and \eqref{eq:action_2} that $\ell_{\Omega}(\Lambda) = dh(1) + 2y + 2V - 2 \mathcal{A}(P)$. \end{proof} \medskip \noindent {\em Step 1: Calabi gives the lower bound.} Here, we will prove the lower bound needed for establishing Equation \eqref{eqn:calabiconjecture}. In other words, we will show that for any sequence $(d,k)$, such that $d\to \infty$, we have \begin{equation}\label{eqn:lower_bound} \Cal(\varphi) \leq \liminf_{d \to \infty} \left( \frac{c_{d,k}(\varphi)}{d} - \frac12\frac{k}{d^2 + d} \right). \end{equation} To prove the above, fix $\eps > 0$. We will show that for all sufficiently large positive integers $d$, there exists a sequence of concave lattice paths $\{P_{\eps,d}\}$, % such that \begin{equation} \label{eqn:lowerbound_for_Pd} \left| \Cal(\varphi) - \left( \frac{ \mathcal{A}(P_{\eps,d})}{d} - \frac12\frac{k_d}{d^2 + d} \right) \right| \leq \eps, \end{equation} % where $k_d = 2j(P_{\eps,d}) - d $ denotes the combinatorial index of $P_{\eps,d}$. By Theorem \ref{theo:spec_computation}, we have $\mathcal{A}(P_{\eps,d}) \leq c_{d, k_d} (\varphi)$, and, by the argument we explained in Section \ref{sec:PFH_monotone_twist}, see the discussion after Theorem \ref{thm:s2case}, proving \eqref{eqn:lower_bound} for one sequence $\{(d,k)\}$ with $d$ ranging across all sufficiently large positive integers proves it for all such sequences, and so we conclude \eqref{eqn:lower_bound} from the above, since $\eps$ was arbitrary. \medskip We now turn to the description of the concave paths $P_{\eps,d}$. Let $P$ be a concave path approximating the graph of $h$ such that it begins at $(-1, 0)$, ends on the line $x = 1$, is piecewise linear, and its vertices are rationals with numerator an even integer and denominator $d$. Let $\Lambda$ be the convex polygon obtained as follows: Add a vertical edge at the beginning of $P$ and a horizontal edge at the end of it; orient this polygon counterclockwise, relative to any point in its interior; and, rotate it clockwise by ninety degrees. The convex polygon $\Lambda$ approximates $\partial \hat \Omega$. More precisely, given $\eps$, we pick, for all sufficiently large positive integers $d$, paths $P$, subject to the conditions above, and such that \begin{enumerate}[(A)] \item $P$ is within $\eps$ of the graph of $h$, \item $ \vert \ell_\Omega( \Lambda) - \ell_\Omega(\partial \hat \Omega) \vert \leq \eps$ which by Lemma \ref{lem:length_action} is equivalent to $$ \vert \ell_\Omega( \Lambda) - 2 (2h(1) - I) \vert \leq \eps,$$ \item The area of the region under the path $P$, and above the $x$--axis, is within $\eps$ of $I$. \end{enumerate} Let $P_{\eps,d}, \Lambda_{\eps,d}$ be the images of $P, \Lambda$, respectively, under the mapping $$(x, y) \mapsto \frac {d}{2}(x + 1, y).$$ The path $P_{\eps,d}$ is a concave lattice path of degree $d$. Recall that $\Cal(\varphi) = \frac{1}{4}I$. We will prove the two inequalities below, which will imply Equation \eqref{eqn:lowerbound_for_Pd}: \begin{equation}\label{eqn:toconsider} \begin{split} \left| \frac{ \mathcal{A}(P_{\eps,d} )}{d} - \frac I2 \right| \leq \frac{3\eps}4, \\ \left| \frac12\frac{k_d}{d^2 + d} - \frac{I}{4} \right|\leq \frac{\eps}{4}. \end{split} \end{equation} We first examine the term $ \frac{ \mathcal{A}(P_{\eps,d} )}{d} $. By Lemma \ref{lem:length_action}, and using the fact that $\ell_\Omega(\Lambda_{\eps,d}) = \frac{d}{2} \ell_\Omega(\Lambda)$, we obtain \begin{align*} \frac{ \mathcal{A}(P_{\eps,d} )}{d} =& \frac{dh(1) + 2 V - \ell_\Omega(\Lambda_{\eps,d})}{2d} \\ =& \frac{h(1)}2 + \frac{V}{d} - \frac{\ell_\Omega(\Lambda)}{4}. \end{align*} By item $(A)$ above, the term $\frac{V}{d}$ is within $\frac\eps2$ of $\frac{h(1)}2$. And, by item $(B)$ above, the term $\ell_{\Omega}(\Lambda)$ is within $\eps$ of $2(2h(1) - I),$ hence the first inequality in \eqref{eqn:toconsider}. As for the second inequality, we know from the index computations of Section \ref{sec:index} that, up to an error of $O(d)$, the index $k_d$ is twice the area between the $x$--axis and the path $P_{\eps,d}$. Because $P_{\eps,d}$ is a scaling of $P$ by a factor of $\frac{d}{2}$, item (C) above implies $$ - \frac{d^2}{2} \eps + O(d) \leq k_d - \frac{d^2}{2} I \leq \frac{d^2}{2} \eps + O(d), $$ which for sufficiently large $d$ yields the second inequality in \eqref{eqn:toconsider}. \end{proof} \noindent {\em Step 2: Calabi gives the upper bound.} We now complete the proof of Theorem~\ref{thm:s2case}. We emphasize again, for the convenience of the reader, that as mentioned in Remark \ref{rem:only_need_inequality}, we do not actually need this step of the proof for the proof of our main result Theorem~\ref{thm:main}. To complete the proof, we need to show that \begin{equation}\label{eqn:upper_bound} \Cal(\varphi) \geq \limsup_{d \to \infty} \left( \frac{ c_{d,k}(\varphi)}{d} - \frac12\frac{k}{d^2 + d} \right). \end{equation} To do this, we will show that \begin{equation}\label{eqn:upper_bound2} \Cal(\varphi) \geq \limsup_{d \to \infty} \left( \frac{ \mathcal{A}(P)}{d} - \frac12\frac{k}{d^2 + d} \right) \end{equation} for all degree $d$ concave lattice paths $P$ of combinatorial index $k$. Let $P$ be a concave lattice path of degree $d$ and combinatorial index $k$. Let $E$ be a real number with $E> h(1)$ and $E>\frac{2V}d$, and $\Omega_E$ be the compact convex subset of $\R^2$ bounded by the graph of $h$, the vertical segments $\{-1\}\times [0,E]$ and $\{1\}\times [h(1),E]$, and the horizontal segment $[0,d]\times\{E\}$. For example, with the notations of Lemma \ref{lem:length_action}, we have $\Omega=\Omega_{h(1)}$. The inequality (\ref{eqn:upper_bound2}) will follow from letting $E$ tend to $\infty$ after applying the isoperimetric inequality (\ref{eq:isoperimetric}) to the domain $\Omega_E$ and to the following curve $\Lambda_E$. To define $\Lambda_E$, consider the region delimited by our lattice path $P$, the vertical segments $\{0\}\times [y,y+\frac{dE}2]$ and $\{d\}\times [y+V,y+\frac{dE}2]$, and the horizontal segment $[0,d]\times\{y+\frac{dE}2\}$. Our curve $\Lambda_E$ is the boundary of this region, rotated clockwise by ninety degrees; see Figure \ref{fig:isoperimetric3}. \begin{figure}[h!] \centering \def1\textwidth{1\textwidth} \input{isoperimetric3.pdf_tex}\caption{The convex subset $\Omega_E$ and the path $\Lambda_E$.}\label{fig:isoperimetric3} \end{figure} The isoperimetric inequality (\ref{eq:isoperimetric}) gives: \begin{equation}\label{eq:isoperimetric2} \ell_{\Omega_E}^2(\Lambda_E)\geq 4 A(\Omega_E)A(\Lambda_E). \end{equation} The area factors are easily computed. We have: \[A(\Omega_E)=2E-I,\quad A(\Lambda_E)=\tfrac12d^2E-a(P),\] where $a(P)$ denote the area of the region between $P$, the horizontal segment $[0,d]\times \{y\}$ and the vertical segment $\{d\}\times[y,y+V]$ (in grey on Figure \ref{fig:isoperimetric3}). Moreover, a computation similar to that of item 2 in Lemma \ref{lem:length_action}, gives \[\ell_{\Omega_E}(\Lambda_E)=2dE +2y-2\mathcal{A}(P).\] Thus, (\ref{eq:isoperimetric2}) gives \[(2dE+2y-2\mathcal{A}(P))^2\geq 4(2E-I)(\tfrac12d^2E-a(P)).\] After simplification, we obtain \[-2dE\mathcal{A}(P)+\mathcal{A}(P)^2+ 2dEy-2y\mathcal{A}(P) + y^2 \geq -2a(P)E-\tfrac12d^2EI+Ia(P).\] Dividing by $2d^2E$ and letting $E$ go to $+\infty$ then yields: \[\frac14I\geq \frac{\mathcal{A}(P)}{d}-\frac{a(P)+dy}{d^2}.\] Now $2a(P)+2dy$ corresponds to $k$ up to an error $O(d)$. Thus, for all sequence of paths $P$ of degree $d$ and index $k$, \[\frac14I\geq \limsup_{d \to \infty} \left( \frac{ \mathcal{A}(P)}{d} - \frac12\frac{k}{d^2 + d} \right),\] from which (\ref{eqn:upper_bound2}) follows. \section{Discussion and open questions}\label{sec:questions} We discuss here some open questions relating to the main results of our article. \subsection{Simplicity on other surfaces}\label{sec:questions_simplicity} Let $M$ denote a closed manifold equipped with a volume form $\omega$ and denote by $\Homeo_0(M, \omega)$ the identity component in the group of volume-preserving homeomorphisms of $M$. In \cite{fathi}, Fathi constructs the {\bf mass-flow homomorphism} $$\mathcal{F}: \Homeo_0(M, \omega) \rightarrow H_1(M)/ \Gamma,$$ mentioned above, where $H_1(M)$ denotes the first homology group of $M$ with coefficients in $\R$ and $\Gamma$ is a discrete subgroup of $H_1(M)$ whose definition we will not need here. Clearly, $\Homeo_0(M, \omega)$ is not simple when the mass-flow homomorphism is non-trivial. This is indeed the case when $M$ is a closed surface other than the sphere. Fathi proved that $\mathrm{ker}(\mathcal{F})$ simple if the dimension of $M$ is at least three. The following question is posed in \cite[Appendix A.6]{fathi}. \begin{question}[Fathi]\label{que:massflow} Is $\mathrm{ker}(\mathcal{F})$ simple in the case of surfaces? In particular, is $\Homeo_0(\S^2, \omega)$, the group of area and orientation preserving homeomorphisms of the sphere, simple? \end{question} This question remains open. We remark that the two-sphere is the only closed manifold for which the simplicity question for volume-preserving homeomorphisms is unknown. There exists a circle of classical ideas and techniques, generally attributed to Epstein and Thurston (sometimes referred to as ``Thurston tricks''), which, under certain conditions, allow one to conclude that simplicity holds on a large class of manifolds once it is established for a single manifold. According to \cite{Bounemoura}, these classical methods can be used to prove that $\mathrm{ker}(\mathcal{F})$ is simple if and only if $\Homeo_c(\D, \omega)$ is perfect; this is the statement of \cite[Proposition 4.1.7]{Bounemoura}; % as a consequence of Corollary \ref{corol:not_perfect}, then, $\mathrm{ker}(\mathcal{F})$ is in principle non-simple for all closed surfaces. However, we have been unable to independently verify \cite[Proposition 4.1.7]{Bounemoura}, and indeed the author of \cite{Bounemoura} has withdrawn the stated proposition in a recent private communication with us. Hence, as far as we know, Question \ref{que:massflow} remains open. It is plausible that the aforementioned classical methods could be used to settle the above question. Another approach would be to adapt the methods of our paper; this requires further development of the theory of PFH spectral invariants. \subsection{$C^0$-symplectic topology and simplicity in higher dimensions}\label{sec:simplicity_high_dim} From a symplectic viewpoint, a natural generalization of area-preserving homeomorphisms to higher dimensions is given by {\bf symplectic homeomorphisms}. These are, by definition, those homeomorphisms which appear as $C^0$ limits of symplectic diffeomorphisms. By the celebrated rigidity theorem \cite{Eliashberg, Gromov} of Eliashberg and Gromov, a smooth symplectic homeomorphism is a symplectic diffeomorphism. These homeomorphisms form the cornerstone of the field of {\bf $C^0$-symplectic topology} which explores continuous analogues of smooth symplectic objects; see for example \cite{muller-oh, banyaga_homeo, HLS1, HLS2, BuOp}. The connection between $C^0$-symplectic topology and the simplicity conjecture is formed by the fact that, in dimension two, symplectic homeomorphisms are precisely the area and orientation preserving homeomorphisms of surfaces. Indeed, as we mentioned in Section \ref{sec:intro_main_theo}, the simplicity conjeture has been one of the driving forces behind the development of $C^0$-symplectic topology; for example, the articles \cite{muller-oh, Oh10, Viterbo-uniqueness, buhovsky-seyfaddini, humiliere, LeRoux-6Questions, LeRoux10, EPP, Sey13, Sey-displaced} were, at least partially, motivated by this conjecture. The connection to $C^0$-symplectic topology motivates the following generalization of Question \ref{que:disc}. \begin{question}\label{que:simplicity_ball} Is $\overline{\Symp}_c(\D^{2n}, \omega)$, the group of compactly supported symplectic homeomorphisms of the standard ball, simple?\footnote{An argument involving the Alexander isotopy shows that $\overline{\Symp}_c(\D^{2n}, \omega)$ is connected.} \end{question} As we will now explain Question \ref{que:massflow} admits a natural generalization to higher dimensions as well. To keep our discussion simple we will suppose that $(M, \omega)$ is a closed symplectic manifold. However, this assumption is not necessary and the question below can be reformulated for non-closed manifolds too. On a symplectic surface $(M, \omega)$, the group $\mathrm{ker}(\mathcal{F})$ discussed in the above section is often referred to as the group of {\bf Hamiltonian homeomorphisms} and is denoted by $\overline{\Ham}(M, \omega)$; see for example \cite{lecalvez}. The reason for this terminology is that it can be shown that $\mathrm{ker}(\mathcal{F})$ coincides with the $C^0$ closure of Hamiltonian diffeomorphisms. Hence, in this language, Question \ref{que:massflow} may be rephrased as the question of whether or not the group of Hamiltonian homeomorphisms is simple. On higher dimensional symplectic manifolds, the elements of the $C^0$ closure of $\Ham(M,\omega)$ are also called Hamiltonian homeomorphisms and have been studied extensively in $C^0$ symplectic topology; see, for example, \cite{muller-oh, BHS1, BHS2, kawamoto}. \begin{question}\label{question:ham-bar-simple} Is $\overline{\Ham}(M, \omega)$ a simple group? \end{question} In comparison, Banyaga's theorem states that the group of Hamiltonian diffeomorphisms is simple for closed $M$. \subsection{Finite energy homeomorphisms} \label{sec:fhomeoq} The group of finite energy homeomorphisms, $\FHomeo(M, \omega)$, can be defined on arbitrary symplectic manifolds; the construction is analogous to what is done in Section \ref{sec:prop-norm-subgr}. It forms a normal subgroup of $\overline{\Ham}(M, \omega)$. However, we do not know if it is a proper normal subgroup. Infinite twists can also be constructed on arbitrary symplectic manifolds: the construction of $\phi_f$, described in Section \ref{sec:calabi_inf_twist}, admits a generalization to $\D^{2n}$. And the analogue of Equation \eqref{eq:infinite_calabi}, the condition for having ``infinite" Calabi invariant, can also be formulated in higher dimensions. \begin{question}\label{que:inf_twist_higher_dimension} Is it true that infinite twists, which satisfy the higher dimensional analogue of Equation \eqref{eq:infinite_calabi}, are not finite energy homeomorphisms? \end{question} Clearly, a positive answer to this question would settle all of the above simplicity questions. In the case of surfaces, one could hope to apply the machinery of PFH to approach this question. However, a serious obstacle arises in higher dimensions where PFH, and the related Seiberg-Witten theory, have no known generalization. We now return to the case of the disc, where we know that $\FHomeo_c(\D, \omega)$ is a proper normal subgroup of $\Homeo_c(\D, \omega)$. This immediately gives rise to several interesting questions about $\FHomeo_c(\D, \omega)$. \begin{question} Is $\FHomeo_c(\D, \omega)$ simple? \end{question} As was mentioned in Remark \ref{rem:Hameo}, the Oh-M\"uller group $\Hameo_c(\D, \omega)$, which we introduce below, is a subgroup of $\FHomeo_c(\D, \omega)$, and it can easily be checked that it is a normal subgroup. \begin{question} Is $\Hameo_c(\D, \omega)$ a proper normal subgroup of $\FHomeo_c(\D, \omega)$? \end{question} Another interesting direction to explore is the algebraic structure of the quotient $\Homeo_c(\D, \omega) / \FHomeo_c(\D, \omega)$. At present we are not able to say much beyond the fact that this quotient is abelian; see Proposition \ref{prop:commutators}. Here are two sample questions. \begin{question} Is the quotient $\Homeo_c(\D, \omega) / \FHomeo_c(\D, \omega)$ torsion-free? Is it divisible? \end{question} It would also be very interesting to know if the quotient $\Homeo_c(\D, \omega) / \FHomeo_c(\D, \omega)$ admits a geometric interpretation. \subsection{Hutchings' conjecture and extension of the Calabi invariant}\label{sec:extending_Calabi} Ghys \cite{Ghys_ICM}, Fathi and Oh \cite[Conjecture 6.8]{muller-oh} have asked if the Calabi invariant extends to either of $\Hameo_c(\D, \omega)$ or $\Homeo_c(\D, \omega)$. It seems natural to add $\FHomeo_c(\D, \omega)$ to the list. \begin{question} Does the Calabi invariant admit an extension to any of $\Hameo_c(\D, \omega)$, $\FHomeo_c(\D, \omega)$, or $\Homeo_c(\D, \omega)$? \end{question} The reason for expecting that such extensions could exist is that the Calabi invariant exhibits certain topological properties: In an unpublished manuscript \cite{Fathi-Calabi}, Fathi shows that the Calabi invariant admits an interpretation as an average rotation number and uses this to extend the Calabi invariant to Lipschitz homeomorphisms; see also \cite{humiliere} for a different proof. Fathi's interpretation of the Calabi invariant enables a proof of topological invariance for Calabi's invariant \cite{Gambaudo-Ghys}, i.e.\ it is invariant under conjugation by area-preserving homeomorphisms. As was mentioned in Remark \ref{rem:hutchings_conjecture}, Hutchings has conjectured that the asymptotics of PFH spectral invariants recover the Calabi invariant for Hamiltonian diffeomorphisms. This conjecture, combined with Theorem \ref{theo:C0_continuity}, implies that the Calabi invariant admits an extension to $\Hameo_c(\D, \omega)$; we verify this below. We remark that extending the Calabi invariant to $\Hameo_c(\D, \omega)$ would imply that both $\Hameo_c(\D, \omega)$ and $\FHomeo_c(\D, \omega)$ are non-simple. Non-simplicity of $\Hameo_c(\D, \omega)$ is clear; that of $\FHomeo_c(\D, \omega)$ follows from the fact that if $\Hameo_c(\D, \omega)$ is not a proper normal subgroup then $\FHomeo_c(\D, \omega)$ would coincide with the non-simple $\Hameo_c(\D, \omega)$. We end our article by verifying that Hutchings' conjecture implies that the Calabi invariant extends to $\Hameo_c(\D, \omega)$. Let us begin with the definition\footnote{The definition we have given here is a slight variation of the one in \cite{muller-oh}; it can easily be checked that if $\phi$ is a hameomorphism in the sense of \cite{muller-oh}, then it is also a hameomorphisms in the sense described here.} of $\Hameo_c(\D, \omega)$. We say $\phi \in \Homeo_c(\D, \omega)$ is a {\bf hameomorphism} if there exists a sequence of smooth Hamiltonians $H_i \in C^{\infty}_c(\S^1 \times \D)$ and a continuous $H \in C^{0}_c(\S^1 \times \D)$ such that $$\varphi^1_{H_i} \xrightarrow{C^0} \phi, \text{ and } \| H- H_i \|_{(1, \infty)} \to 0.$$ The set of all hameomorphisms is denoted by $\Hameo_c(\D, \omega)$.\footnote{Oh and M\"uller use the terminology \emph{Hamiltonian homeomorphisms} for the elements of $\Hameo_c(\D, \omega)$. We have chosen to avoid this terminology because in the surface dynamics literature it is commonly used for referring to homeomorphisms which arise as $C^0$ limits of Hamiltonian diffeomorphisms.} It defines a normal subgroup of $\Homeo_c(\D, \omega)$ which is clearly contained in $\FHomeo_c(\D, \omega)$. Take $\phi \in \Hameo_c(\D, \omega)$ and $H \in C^{0}_c(\S^1 \times \D)$ as in the previous paragraph; we denote $ \varphi_H := \phi$. Define \begin{equation}\label{eqn:Calabi_Hameo} \Cal(\varphi_H) := \int_{\S^1} \int_\D H \, \omega \, dt. \end{equation} We first show that $\Cal$ is well-defined. First, note that because $\Cal$ is a homomorphism on $\Diff_c(\D,\omega)$, to show this it suffices to show that if $\varphi_H = \id$, then \begin{equation} \label{eqn:somanyequations} \int_{\S^1} \int_\D H \, \omega \, dt = 0. \end{equation} Suppose that $\varphi_H = \id$ and fix a sequence $(H_1, H_2, \ldots,)$ for $\varphi_H$ as in the definition of $\Hameo_c(\D, \omega)$. \begin{claim}\label{cl:last_claim} For all $i$, we have $\vert \frac{ c_d } { d } (\varphi^1_{H_i}) \vert \leq \| H- H_i \|_{(1, \infty)} .$ \end{claim} \begin{proof} By Hofer continuity of PFH spectral invariants, we have $\vert \frac{ c_d } { d } (\varphi^1_{H_j}) - \frac{ c_d } { d } (\varphi^1_{H_i}) \vert \leq \| H_j- H_i \|_{(1, \infty)} $, for all $i, j$. Fixing $i$ and taking the limit of this inequality, as $j \to \infty$, yields the claim, by % Theorem \ref{theo:C0_continuity} and item $4$ of Theorem \ref{thm:PFHspec_initial_properties}, since $\varphi^1_{H_j} \xrightarrow{C^0} \id$. \end{proof} We now establish \eqref{eqn:somanyequations}. For all $i, d \in \N$ we have \begin{align*} \left\vert \int_{\S^1} \int_\D H \, \omega \, dt \right\vert &\leq \left\vert \int_{\S^1} \int_\D H \, \omega \, dt \; - \Cal(\varphi^1_{H_i} ) \right\vert \\ &\qquad\qquad+ \left\vert \Cal(\varphi^1_{H_i}) - \frac{ c_d } { d } (\varphi^1_{H_i}) \right\vert + \left\vert \frac{ c_d } { d } (\varphi^1_{H_i}) \right\vert \\ &\leq \| H- H_i \|_{(1, \infty)} + \left\vert \Cal(\varphi^1_{H_i}) - \frac{ c_d } { d } (\varphi^1_{H_i}) \right\vert + \left\vert \frac{ c_d } { d } (\varphi^1_{H_i}) \right\vert \\ &\leq 2 \| H- H_i \|_{(1, \infty)} + \left\vert \Cal(\varphi^1_{H_i}) - \frac{ c_d } { d }(\varphi^1_{H_i}) \right\vert, \end{align*} where the last inequality follows from Claim \ref{cl:last_claim}. Now take any $\varepsilon > 0$. There exists $N \in \N$ such that if $i \geq N$, then $ \| H- H_i \|_{(1, \infty)} \leq \frac12 \varepsilon$. Hence, for $i \geq N$, we have \begin{align*} \left\vert \int_{\S^1} \int_\D H \, \omega \, dt \right\vert \leq \varepsilon + \left\vert \Cal(\varphi^1_{H_i}) - \frac{ c_d } { d }(\varphi^1_{H_i}) \right\vert. \end{align*} Assuming Hutchings's conjecture, we can now conclude that $ \vert \int_{\S^1} \int_\D H \, \omega \, dt \vert \leq \varepsilon,$ hence \eqref{eqn:somanyequations}, hence % the claimed extension of $\Cal$. It remains to show that \[ \Cal : \Hameo_c(\D, \omega) \rightarrow \R\] is indeed a homomorphism. Having shown that \eqref{eqn:Calabi_Hameo} is well-defined, this has in fact already been shown in \cite{Oh10} and so we will only provide a sketch of the argument. Take $\varphi_H, \varphi_G \in \Hameo_c(\D, \omega)$. Without loss of generality, we may suppose that $H(t,x)$ and $G(t,x)$ vanish for $t$ near $0 \in \S^1$; this can be achieved by replacing $H$ with the reparametrization $\rho'(t) H(\rho(t), x)$, where $\rho : [0,1] \to [0,1]$ coincides with $0$ near $0$ and with $1$ near $1$; see \cite[p.\ 31]{Polterovich2001} for more details on the reparametrization argument. It can be checked that $\varphi_H \circ \varphi_G = \varphi_{K}$ where $K$ is the concatenation of $H$ and $G$ given by the formula \[K(t,x) = \begin{cases} 2 H(2t,x), &\quadif t\in [0, \frac{1}{2}]\\ 2 G(2t-1,x),&\quadif t\in [\frac{1}{2}, 1] \end{cases}. \] It follows immediately from the above formula, and Equation \eqref{eqn:Calabi_Hameo}, that $\Cal(\varphi_H \circ \varphi_G) = \Cal(\varphi_H) + \Cal(\varphi_G)$. Verification of the fact that $\Cal(\varphi_H^{-1} )= - \Cal(\varphi_H)$ is similar and so we omit it. \bibliographystyle{alpha}

https://arxiv.org/abs/1303.4966

Automorphisms of groups and converse of Schur's theorem

An automorphism of a group G is called an IA-automorphism if it induces the identity automorphism on the abelianized group G/G'. Let IA(G) denote the group of all IA-automorphisms of G. We classify all finitely generated nilpotent groups G of class 2 for which IA(G) is isomorphic to Inn(G). In particular, we classify all finite nilpotent groups of class 2 for which each IA-automorphism is inner. As consequences, we give surprisingly very easy proofs of converse of Schur's theorem and also prove some other related results.

\section{\normalsize Introduction} Let $G$ be any group and let $G^{\prime}$ and $Z(G)$ respectively denote the derived group and the center of $G$. By $\mathrm{Inn}(G)$ we denote the inner automorphism group of $G$. Following Bachmuth [2], we call an automorphism of $G$ an IA-automorphism if it induces the identity automorphism on the abelianized group $G/G^{\prime}$. Let $\mathrm{IA}(G)$ denote the group of all IA-automorphisms of $G$ and let $\mathrm{IA}(G)^*$ denote its subgroup consisting of those IA-automorphisms which fix $Z(G)$ element-wise. IA-automorphisms have been well studied in past (see for example [2], [4], [5], [9], [11]). If $G$ is a free group of rank 2, then it is a classical result of Nielsen [9] that $\mathrm{IA}(G)= \mathrm{Inn}(G)$. Bachmuth [2] obtained the same result for free metabelian groups of rank 2. Recently Attar [1, Theorem 2.1] has proved that if $G$ is a finite $p$-group of class 2, then $\mathrm{IA}(G)= \mathrm{Inn}(G)$ if and only if $G^{\prime}$ is cyclic and $\mathrm{IA}(G)=\mathrm{IA}(G)^*$. In section 2, we classify all finitely generated nilpotent groups $G$ of class 2 for which (i) $\mathrm{IA}(G)\simeq \mathrm{Inn}(G)$ and (ii) $\mathrm{IA}(G)^*\simeq \mathrm{Inn}(G)$. In particular, we classify all finite nilpotent groups $G$ of class 2 for which (i) $\mathrm{IA}(G)= \mathrm{Inn}(G)$ and (ii) $\mathrm{IA}(G)^*=\mathrm{Inn}(G)$. In the year 1904 Schur [12] proved the result: If the index of the center of a group is finite, then its derived group is also finite. We call this result as Schur's theorem. Neumann [8, Corrolary 5.41] proved the converse of Schur's theorem for finitely generated groups (the converse of Schur's theorem is not true in general for example for an infinite extra-special $p$-group for an odd prime $p$). More precisely, he proved that if $G$ is finitely generated and if its derived group $G^{\prime}$ has finite order, then the index $[G:Z(G)]$ of the centre is also finite. We call this result as Neumann's theorem. Niroomand [10] generalized the Neumann's theorem to the following: If $G^{\prime}$ is finite and $G/Z(G)$ is finitely generated by $d$ elements, then \begin{equation} |G/Z(G)|\le |G^{\prime}|^d. \end{equation} Sury [13] generalized Niroomand's result to the following: Let $G$ be a group such that the set $K(G)$ of its commutators is finite. Then $G^{\prime}$ is finite and if $G/Z(G)$ can be generated by $d$ elements, then \begin{equation} |G/Z(G)| \le |K(G)|^d. \end{equation} Recently Yadav [15, Theorem A] has further generalized Niroomand's and Sury's results and has raised the question: what can one say about non-abelian groups for which equality holds in (1)? He himself classifies, upto isoclinism, all nilpotent groups $G$ of class 2 such that $G/Z(G)$ is finite and equality holds in (1). It can be easily seen that if $G$ is a finite $p$-group of maximal class, then equality holds in (1) if and only if $|G|=p^3$. In section 3, we prove that if $G$ is a finite $p$-group of co-class 2 for which equality holds in (1), then $|G|=p^4$ or $p^5$; and give necessary and sufficient conditions for a finitely generated nilpotent group $G$ of class 2 for which $G/Z(G)\simeq {G^{\prime}}^{r(G/Z(G))}$. As a consequence of the fact that $\mathrm{Inn}(G)\le\mathrm{IA}(G)^*$, we also obtain the converses of Schur's theorem proved by Niroomand [10], Sury [13] and Yadav [15, Theorem A]. Quite unexpectedly, our proofs are easy, in-fact very easy. Yadav [15, Question 2] has also raised the question that if $G$ is a finite nilpotent group for which the equality holds in (1), then is it necessary that $G^{\prime}$ is cyclic? We answer the question in negative and give a counter example of minimal possible order. By $\mathrm{Hom}(G,A)$ we denote the group of all homomorphisms of $G$ into an abelian group $A$. For $x\in G$, the set $\{[x,g]|g\in G\}$ of commutators is denoted as $[x,G]$. The frattini subgroup of $G$ is denoted as $\Phi(G)$ and $(G)_{ab}$ denotes the abelianized group $G/G^{\prime}$. By $C_{p}$ we denote a cyclic group of order $p$ and by $X^n$ we denote the direct product of $n$-copies of a group $X$. The rank of a group $G$ is the smallest cardinality of a generating set of $G$. The rank, torsion rank and torsion-free rank of a group $G$ are respectively denoted as $r(G)$, $d(G)$ and $\rho(G)$. For a finite group $G$, $G_p$, $\mathrm{exp}(G)$ and $\pi(G)$ respectively denote the Sylow $p$-subgroup, exponent and the set of primes dividing the order of $G$. All other unexplained notations, if any, are standard. We shall use the following two well known lemmas very frequently without any further referring. \begin{lm} Let $U,V$ and $W$ be abelian groups. Then\\ $(i)$ if $U$ is torsion-free of rank $m$, then $\mathrm{Hom}(U, V)\simeq V^{m}$,\\ $(ii)$ if $U$ is torsion and $V$ is torsion-free, then $\mathrm{Hom}(U, V)=1$, and\\ $(iii)$ if $U$ and $V$ are finite and $\mathrm{exp}(U)=\mathrm{exp}(V)$, then $Hom(U, V)\simeq U$ if and only if $V$ is a cyclic group. \end{lm} \begin{lm} Let $G$ be a finitely generated nilpotent group of class $2$. Then $\mathrm{exp}(T(G/Z(G)))=\mathrm{exp}(T(G^{\prime}))$, where $T(H)$ denotes the torsion part of a group $H$. \end{lm} Let $G$ be a group and $X$ be a central subgroup of $G$. If $\theta\in\mathrm{Hom}((G/X)_{ab},X)$, then $T_{\theta}:G\rightarrow G$ defined as $T_{\theta}(g)=g\theta(\overline{g})$ is an automorphism of $G$, where $\overline{g}$ is the coset determined by $g$ in $(G/X)_{ab}$. It is clear that $\theta \mapsto T_{\theta}$ is an isomorphism between $\mathrm{Hom}((G/X)_{ab},X)$ and the group of all automorphisms of $G$ that induce the identity on both $X$ and $G/X$. This observation is from [4, Sec. 2] and is a little specific one. We can, in-fact, have the following more general result: \begin{lm} Let $G$ be any group and $Y$ be a central subgroup of $G$ contained in a normal subgroup $X$ of $G$. Then the group of all automorphisms of $G$ that induce the identity on both $X$ and $G/Y$ is isomorphic onto $\mathrm{Hom}(G/X,Y)$. \end{lm} The following is an easy particular consequence: \begin{lm} Let $G$ be a nilpotent group of class $2$. Then $$\mathrm{IA}(G)\simeq \mathrm{Hom}(G/G^{\prime},G^{\prime})\;\;\mathrm{and} \;\;\mathrm{IA}(G)^*\simeq \mathrm{Hom}(G/Z(G),G^{\prime}).$$ \end{lm} \section{\normalsize IA-automorphisms} Let $G$ be a finitely generated nilpotent group of class 2. Let $G/Z(G)\simeq A\times \mathbb{Z}^a$, $G/G^{\prime}\simeq B\times \mathbb{Z}^b$ and $G^{\prime}\simeq C\times \mathbb{Z}^c$, where $A, B,C$ are respective torsion parts and $a,b,c$ are respective torsion-free ranks of $G/Z(G),G/G^{\prime}$ and $G^{\prime}$. Let $\pi(A)=\{p_1,p_2,\ldots,p_d\}$ and $\pi(B)=\{p_1,p_2,\ldots,p_e\}$, where $d\le e$ because $G/Z(G)$ is a quotient group of $G/G^{\prime}$. Suppose that $$A_{p_{i}}\simeq \displaystyle\prod_{j=1}^{m_{i}}C_{p_{i}^{\alpha_{ij}}}\;\; \mathrm{and}\;\;B_{p_{i}}\simeq \displaystyle\prod_{j=1}^{n_{i}}C_{p_{i}^{\beta_{ij}}}.$$ Then \[ A\simeq \displaystyle\prod_{i=1}^d \displaystyle\prod_{j=1}^{m_{i}}C_{p_{i}^{\alpha_{ij}}}\;\; \mathrm{and}\;\;B\simeq \displaystyle\prod_{i=1}^e \displaystyle\prod_{j=1}^{n_{i}}C_{p_{i}^{\beta_{ij}}}, \] where $ \alpha_{11}\geq \alpha_{12}\geq\ldots\ge\alpha_{1m_{1}},\alpha_{21}\geq \alpha_{22}\geq\ldots\geq \alpha_{2m_{2}},\ldots,\alpha_{d1}\geq \alpha_{d2}\geq\ldots\geq\alpha_{dm_{d}}$ and $ \beta_{11}\geq \beta_{12}\ge\ldots\ge\beta_{1n_{1}}, \beta_{21}\geq \beta_{22}\geq\ldots\beta_{2n_{2}},\ldots,\beta_{e1}\geq \beta_{e2}\geq\ldots\geq \beta_{en_{e}}$ are positive integers. Since $G/Z(G)$ is a quotient group of $G/G^{\prime}$, it follows that $m_i\le n_i$ and $\alpha_{ij}\le\beta_{ij}$ for all $i,1\le i\le d$ and for all $j,1\le j\le m_i$. We fix these notations, unless or otherwise stated, for the rest of this paper. \begin{thm} Let $G$ be a finitely generated nilpotent group of class $2$. Then \\ $(i)$ if $G$ is torsion-free, then $\mathrm{IA}(G)\simeq \mathrm{Inn}(G)$ if and only if $G^{\prime}$ is cyclic and $\rho (G/Z(G))=\rho(G/G^{\prime})$.\\ $(ii)$ if $G$ is not torsion-free, then $\mathrm{IA}(G)\simeq \mathrm{Inn}(G)$ if and only if one of the following four conditions holds:\\ $(a)$ $G$ is finite, $m_{i}=n_{i}$ for $1\leq i \leq d$, $G^{\prime}$ is cyclic, and for each $i,\;1\leq i \leq d$, either $(G/Z(G))_{p_{i}} $ is homocyclic or $\alpha_{i1}=\alpha_{it_{i}}$ for $ i1\leq it_{i}\leq i(r_{i}-1)$ and $\beta_{it_{i}}= \alpha_{it_{i}}$ for $ir_{i}\leq it_{i}\leq in_{i}$, where $ir_{i}$ is the smallest positive integer between $i1$ and $ in_{i}$ such that $\beta_{ir_{i}}< \alpha_{i1}$.\\ $(b)$ $G^{\prime}\simeq \mathbb{Z}$ and $\rho(G/Z(G))=\rho(G/G^{\prime})$.\\ $(c)$ $G^{\prime}$ is torsion, $G/G^{\prime}$ is torsion-free and $A\simeq C^b.$\\ $(d)$ $G/Z(G)$ and $G^{\prime}\simeq C\times \mathbb{Z}$ are mixed groups, $G/G^{\prime}$ is torsion-free, $A\simeq C^b$ and $\rho (G/Z(G))= \rho (G/G^{\prime})$.\\ $(iii)$ $\mathrm{IA}(G)^*\simeq \mathrm{Inn}(G)$ if and only if $G^{\prime}$ is cyclic. \end{thm} \noindent{\bf Proof.} (i) First Suppose that $\mathrm{IA}(G)\simeq \mathrm{Inn}(G)$. Then $$\mathrm{IA}(G)\simeq \mathrm{Hom}(G/G^{\prime},G^{\prime})\simeq G/Z(G)$$ by Lemma 1.4. If $G/G^{\prime}$ is finite, then $G/Z(G)$ is finite and hence by Schur's theorem $G^{\prime}$ is finite, which is not so. Thus $G/G^{\prime}\simeq B\times \mathbb{Z}^b$, where $b>0$. If $G/Z(G)$ has a torsion element, then so does $G^{\prime}$, which is not so. Thus $G/Z(G)$ is torsion-free. It then follows that $\mathbb{Z}^{bc}\simeq \mathbb{Z}^a$. Since $a\le b$, $c=1$ and $a=b$. Hence $G^{\prime}$ is cyclic and $\rho(G/Z(G))=\rho(G/G^{\prime})$. The converse follows easily.\\ (ii) It is not very hard to see that if $G$ satisfies any of the four conditions, then $\mathrm{IA}(G)\simeq \mathrm{Inn}(G)$. Conversely, if $\mathrm{IA}(G)\simeq \mathrm{Inn}(G)$, then \begin{equation} \mathrm{Hom}(B\times \mathbb{Z}^b,C\times \mathbb{Z}^c)\simeq A\times \mathbb{Z}^a. \end{equation} First suppose that $G$ is finite. Then $\mathrm{Hom}(B,C)\simeq A$. Since $d(A)\le d(B)$ and $\mathrm{exp}(A)=\mathrm{exp}(C)$, $G^{\prime}$ is cyclic and thus $G^{\prime}\simeq C_{{p_{1}}^{\alpha_{11}}}\times C_{{p_{2}}^{\alpha_{21}}}\times \ldots \times C_{{p_{d}}^{\alpha_{d1}}}$. Observe that \[\begin{array}{lcl} \mathrm{Hom}(B,C)&\simeq&\mathrm{Hom}(\displaystyle\prod_{i=1}^e\displaystyle\prod_{j=1}^{n_{i}} C_{p_{i}^{\beta_{ij}}}, \displaystyle\prod_{i=1}^{d}C_{p_{i}^{\alpha_{i1}}})\\ &\simeq&\mathrm{Hom}(\displaystyle\prod_{i=1}^d\displaystyle\prod_{j=1}^{n_{i}} C_{p_{i}^{\beta_{ij}}}, \displaystyle\prod_{i=1}^{d}C_{p_{i}^{\alpha_{i1}}})\\ &\simeq&\displaystyle\prod_{i=1}^d\mathrm{Hom}(\displaystyle\prod_{j=1}^{n_{i}} C_{p_{i}^{\beta_{ij}}}, \displaystyle\prod_{i=1}^{d}C_{p_{i}^{\alpha_{i1}}})\\ &\simeq&\displaystyle\prod_{i=1}^d\mathrm{Hom}(\displaystyle\prod_{j=1}^{n_{i}} C_{p_{i}^{\beta_{ij}}}, C_{p_{i}^{\alpha_{i1}}}),\\ \end{array}\] and $A\simeq\displaystyle\prod_{i=1}^d\displaystyle\prod_{j=1}^{m_{i}} C_{p_{i}^{\alpha_{ij}}}$. It thus follows that for each $i,1\le i\le d$, $$\mathrm{Hom}(\displaystyle\prod_{j=1}^{n_{i}} C_{p_{i}^{\beta_{ij}}}, C_{p_{i}^{\alpha_{i1}}})\simeq \displaystyle\prod_{j=1}^{m_{i}} C_{p_{i}^{\alpha_{ij}}}.$$ This immediately implies that $m_i=n_i$. Now either $\beta_{ij}\ge \alpha_{i1}$ for each $j,1\le j\le n_i$, or there exist smallest positive integer $ir_{i}$ between $i1$ and $in_i$ such that $\beta_{ir_{i}}< \alpha_{i1}$. In the first case $(G/Z(G))_{p_{i}}$ is homocyclic, and in the second case $\alpha_{i1}=\alpha_{it_{i}}$ for $i1\le it_{i}\le i(r_{i}-1)$ and $\beta_{it_{i}}=\alpha_{it_{i}}$ for $ir_i \le it_{i}\le in_i$. Next suppose that $G$ is infinite. The derived group $G^{\prime}$ can be torsion, mixed, or torsion-free. First assume that $G^{\prime}$ is torsion-free. Then $G/Z(G)$ is torsion-free and $G/G^{\prime}$ is not torsion by Schur's theorem. Equation (3) then implies that $\mathbb{Z}^{bc}\simeq \mathbb{Z}^a$. It follows that $a=b$ and $c=1$. Next assume that $G^{\prime}$ is torsion. Then $G/Z(G)$ is torsion by Neumann's theorem. If $G/G^{\prime}$ is torsion, then $G$ is finite. If $G/G^{\prime}$ is torsion free, then (3) implies that $A\simeq C^b$; and if $G/G^{\prime}$ is a mixed group, then (3) implies that $A\simeq \mathrm{Hom}(B,C)\times C^b$. Since $d(A)\le d(B)$, $b=0$ and hence $G$ is finite, which is not so. Finally assume that $G^{\prime}$ is a mixed group. Then $G/G^{\prime}$ is either mixed or torsion-free by Schur's theorem. If $G/G^{\prime}$ is torsion-free, then (3) implies that $A\simeq C^b$, $a=b$, $c=1$; and if $G/G^{\prime}$ is mixed, then (3) implies that $A\simeq \mathrm {Hom}(B,C)\times C^b$ and $a=bc$. It follows that $a=b=0$. Therefore both $G/Z(G)$ and $G/G^{\prime}$ are finite. By Schur's theorem $G^{\prime}$ is finite and hence $G$ is finite, a contradiction. \\ (iii) If $G^{\prime}$ is cyclic, finite or infinite, then, by using Lemma 1.4, it is easy to see that $\mathrm{IA}(G)^*\simeq \mathrm{Inn}(G)$. To prove the converse, first suppose that $G$ is torsion-free. Then $G^{\prime}$ is torsion-free and hence $G/Z(G)$ is torsion-free. It follows that $\mathrm{Hom}(\mathbb{Z}^a,\mathbb{Z}^c)\simeq \mathbb{Z}^a$, and hence $G^{\prime}\simeq \mathbb{Z}$ is cyclic. Next suppose that $G$ is torsion. Then both $G/Z(G)$ and $G^{\prime}$ are torsion. Thus $\mathrm{Hom}(A,C)\simeq A$ and hence $G^{\prime}$ is finite cyclic. Finally suppose that $G$ is a mixed group. Then either $G^{\prime}$ is a torsion, or torsion-free, or a mixed group. The case that $G^{\prime}$ is a torsion-free group can be handled as before. If $G^{\prime}$ is torsion, then $G/Z(G)$ is torsion by Neumann's theorem and hence as above $G^{\prime}$ is finite cyclic. Now suppose that $G^{\prime}$ is a mixed group. Then $G/Z(G)$ is either a mixed or a torsion-free group by Schur's theorem. If $G/Z(G)$ is mixed, then $\mathrm{Hom}(A,C)\times C^a\simeq A$. Since $\mathrm{exp}(A)=\mathrm{exp}(C)$, $a=0$ and $G^{\prime}$ is a finite cyclic group. Finally if $G/Z(G)$ is torsion-free, then $\mathrm{Hom}(\mathbb{Z}^a, C\times \mathbb{Z}^c)\simeq \mathbb{Z}^a$ implies that $G^{\prime}\simeq \mathbb{Z}$. \hfill $\Box$\\ Let $\mathrm{Aut}_c(G)$ denote the group of all conjugacy class preserving automorphisms of a group $G$. As a consequence of Theorem 2.1(iii), which generalizes corollary 2.3 of Attar [1], we have the following small but significant generalization of corollary 3.6 of Yadav [14]. \begin{cor} Let $G$ be a finitely generated nilpotent group of class $2$ such that $G^{\prime}$ is cyclic and $G/Z(G)$ is finite. Then $\mathrm{Aut}_c(G)=\mathrm{Inn}(G)$. \end{cor} {\bf Proof.} The proof follows from the fact that $\mathrm{Inn}(G)\le\mathrm{Aut}_c(G)\le\mathrm{IA}(G)^*.$ \hfill $\Box$\\ \begin{cor} Let $G$ be a $2$-generated finite nilpotent group of class $2$. Then any $\mathrm{IA}$-automorphism of $G$ is an inner automorphism. \end{cor} {\bf Proof.} Observe that $G^{\prime}$ is cyclic and $d(G/Z(G))=d(G/G^{\prime})=2$. Also $G/Z(G)$ is homocyclic by [7, Lemma 0.4]. The result now follows from Theorem 2.1. \hfill $\Box$ \begin{cor} Let $G$ be a $2$-generated torsion-free nilpotent group of class $2$. Then $\mathrm{IA}(G)\simeq \mathrm{Inn}(G)$. \end{cor} {\bf Proof.} Observe that $G^{\prime}\simeq \mathbb{Z}$ and $\rho(G/Z(G))=\rho(G/G^{\prime})=2$. Thus $\mathrm{IA}(G)\simeq \mathrm{Inn}(G)$ by Theorem 2.1.\hfill $\Box$ \\ Let $G$ be a finite $p$-group of nilpotency class 2 and let $G/Z(G)\simeq\prod_{i=1}^t C_{p^{\alpha_i}}$ and $G/G^{\prime}\simeq\prod_{j=1}^s C_{p^{\delta_j}}$, where $\alpha_1\geq\alpha_2\geq\ldots\geq\alpha_t$ and $\delta_1\geq\delta_2\geq\ldots\geq\delta_s$ are positive integers. Since $G/Z(G)$ is a quotient group of $G/G^{\prime}$, $t\le s$ and $\alpha_i\le \delta_i$ for $1\le i\le t$. Attar [1, Theorem 2.1] has proved that if $G$ is a finite $p$-group of class 2, then $\mathrm{IA}(G)= \mathrm{Inn}(G)$ if and only if $G^{\prime}$ is cyclic and $\mathrm{IA}(G)=\mathrm{IA}(G)^*$. Our next corollary, which is a particular case of Theorem 2.1, gives an explicit interpretation of this result. \begin{cor} Let $G$ be a finite $p$-group of class $2$. Then $\mathrm{IA}(G)= \mathrm{Inn}(G)$ if and only if $G^{\prime}$ is cyclic, $d(G/Z(G))=d(G/G^{\prime})$, and either $G/Z(G)$ is homocyclic or $\alpha_i=\alpha_1$ for $1\le i\le r-1$ and $\delta_i=\alpha_i$ for $r\le i\le s$, where $r$ is the smallest positive integer such that $\delta_r< \alpha_1.$ \end{cor} \section {\normalsize Converse of Schur's Theorem} Let $G$ be an arbitrary group such that $G^{\prime}$ is finite and $G/Z(G)$ is finitely generated. Let $G/Z(G)$ be minimally generated by $x_1Z(G),x_2Z(G),\ldots,x_dZ(G)$ and let $\alpha\in \mathrm{IA}(G)^*$. Then $\alpha$ fixes $Z(G)$ element-wise and $\alpha(x_i)=x_iy_i$, where $y_i\in G^{\prime}$ for each $i,1\le i\le d$. It thus follows that $\mathrm{IA}(G)^*$ is finite and $|\mathrm{IA}(G)^*|\le |G^{\prime}|^d.$ We have thus proved the following. \begin{thm} Let $G$ be an arbitrary group such that $r(G/Z(G))$ and $G^{\prime}$ are finite. Then $|\mathrm{IA}(G)^*|\le |G^{\prime}|^{r(G/Z(G))}$. \end{thm} In particular, since $\mathrm{Inn}(G)\le \mathrm{IA}(G)^*$, we obtain the following main theorem of Niroomand [10]. \begin{cor} Let $G$ be an arbitrary group such that $r(G/Z(G))$ and $G^{\prime}$ are finite. Then $|G/Z(G)|\le |G^{\prime}|^{r(G/Z(G))}$. \end{cor} Now, rather than $G^{\prime}$, suppose that $[x_i,G]$ is finite for all $i,1\le i\le d$. Let $a\in G$ and let $\alpha$ be the inner automorphism of $G$ defined by conjugation by $a$. Then $\alpha(x_i)=a^{-1}x_ia=x_i[x_i,a]$. Since $|[x_i,G]|=|x_{i}^{G}|\le |K(G)|$, we have thus generalized the main theorem of Sury [13] and proved Theorem A of Yadav [15] in the following theorem. \begin{thm} Let $G$ be an arbitrary group such that $G/Z(G)$ is finitely generated by $x_1Z(G),x_2Z(G),\ldots ,x_dZ(G)$ and $[x_i,G]$ is finite for all $i,1\le i\le d$. Then $|G/Z(G)|\le \Pi_{i=1}^{d}[x_i,G]$ and $G^{\prime}$ is finite. \end{thm} \begin{thm} Let $G$ be an arbitrary group. Then $\mathrm{Inn}(G)$ is finite if and only if $\mathrm{IA}(G)^*$ is finite. \end{thm} {\bf Proof.} If $\mathrm{Inn}(G)$ is finite, then $G'$ is finite by Schur's theorem. It then follows from Theorem 3.1 that $\mathrm{IA}(G)^*$ is finite. Converse is obvious. \hfill$\Box$\\ Niroomand [10] claims that inequality (1) is sharp and the bound he obtains is best possible. He gives example of the group $Q_8$, the quaternion group of order 8, to support his claim. It is clear from Theorem 3.3 that bound obtained by Niroomand is not the best possible. The misinterpretation was because of the group $G=Q_8$ in which $G^{\prime}=[x,G]$ for all $x\in G-\Phi(G)$. We observe from inequalities (1), (2) and Theorem 3.3 that $$|G/Z(G)|\le \Pi_{i=1}^{d}|[x_i,G]|\le |K(G)|^d\le |G^{\prime}|^d.$$ It follows that if $G$ is an arbitrary group such that $G/Z(G)$ is finitely generated, $G^{\prime}$ is finite and equality holds in (1), then $G^{\prime}=K(G)$. In particular, if $G$ is a finite group for which equality holds in (1), then $G'=K(G)$. This gives a new tool for finding groups $G$ for which $G'=K(G)$ (for more details on this problem one can see the excellent survey article of Kappe and Morse [6]). Also observe that if the equality is achieved in (1) by a group $G$, then for that group $\mathrm{Inn}(G)=\mathrm{Aut}_c(G)=\mathrm{IA}(G)^*$. It now becomes interested to classify the groups for which equality holds in (1). Of course it holds for all abelian groups. Yadav [15] also has asked the question that for what non-abelian groups equality holds in (1). He himself has classified, upto isoclinism, all such nilpotent groups $G$ of class 2 with $G/Z(G)$ finite. We here prove the following. \begin{thm} Let $G$ be a finitely generated nilpotent group of class $2$. Then $$G/Z(G)\simeq G^{\prime^{r(G/Z(G))}}$$ if and only if $G^{\prime}$ is cyclic and $G/Z(G)$ is homocyclic. \end{thm} {\bf Proof.} First suppose that $G^{\prime}$ is cyclic and $G/Z(G)$ is homocyclic. If $G^{\prime}$ is torsion-free, then $G/Z(G)$ is also torsion-free. It follows that $$G/Z(G)\simeq \mathbb{Z}^{r(G/Z(G))}\simeq G^{\prime^{r(G/Z(G))}}.$$ In case $G^{\prime}$ is finite, then $G/Z(G)$ is also finite by Neumann's theorem. Since exponents of $G/Z(G)$ and $G^{\prime}$ are same, $G/Z(G)\simeq G^{\prime^{r(G/Z(G))}}$. Conversely suppose that $G/Z(G)\simeq G^{\prime^{r(G/Z(G))}}$. It is sufficient to prove that $G^{\prime}$ is cyclic. If $G'$ is finite, then $G'$ is cyclic by Theorem 2.1(iii). If $G^{\prime}$ is torsion-free, then $G/Z(G)$ is torsion-free. Thus $G^{\prime}\simeq \mathbb{Z}^c$ and $G/Z(G)\simeq \mathbb{Z}^a$. Then $\mathbb{Z}^a\simeq \mathbb{Z}^{ac}$ implies that $G^{\prime}\simeq \mathbb{Z}$ is cyclic. Finally suppose that $G'=C\times \mathbb{Z}^c$ is mixed. Then by assumption $A\times \mathbb{Z}^a\simeq (C\times\mathbb{Z}^c)^{r(G/Z(G))}$. This is not possible because $a<r(G/Z(G))$ and $cr(G/Z(G))\ge r(G/Z(G))$. \hfill$\Box$ \begin{thm} Let $G$ be a finite non-abelian $p$-group of co-class $2$ for which equality holds in $(1)$. Then $|G|=p^4$ or $p^5$. \end{thm} {\bf Proof.} Let $|G|=p^n$. Then $p\le |Z(G)|\le p^2$ and $p^{n-3}\le |G^\prime| \le p^{n-2}$. Let $C^*$ denote the group of all automorphisms of $G$ which centralize both $G/Z(G)$ and $Z(G)$. It follows by assumption that $\mathrm{IA}(G)^*=\mathrm{Inn}(G)$ and by Lemma 1.3 that $C^*\approx \mathrm{Hom}(G/G^\prime, Z(G))$. Observe that if $Z(G)\le G^\prime$, then $C^*\le \mathrm{IA}(G)^*=\mathrm{Inn}(G)$ and thus $C^*=Z(\mathrm{Inn}(G))$. Since $d(G)\ge 2$, $Z(\mathrm{Inn}(G))$ cannot be cyclic. First suppose that $|Z(G)|=p$. Then $|Z_2(G)|=p^3$ and $|C^*|=|Z(\mathrm{Inn}(G))|=p^2$. Thus $d(G)=d(G/Z(G))=2$. If $|G^\prime|=p^{n-2}$, then by (1) $n=3$, which is not possible; and if $|G^\prime|=p^{n-3}$, then by (1) $n=5$. Next suppose that $|Z(G)|=p^2$. Then $Z(G)$ is not contained in $G^\prime$, because then $Z(\mathrm{Inn}(G))$ is cyclic. If $|G^\prime|=p^{n-2}$, then $G'=\Phi(G)$ and thus $d(G/Z(G))=1$, a contradiction. Therefore $G'=p^{n-3}$, $d(G/Z(G))=2$ and hence $n=4$ by (1). \hfill $\Box$ We end up the section by giving a counter example of minimal order to the following question posed by Yadav [15, Question 2]: Let $G$ be a finite nilpotent group for which equality holds in (1). Is it true that $G^{\prime}$ is cyclic? It follows from Theorem 2.1(iii) that if there is a finite $p$-group $G$ for which equality holds in (1) but $G^{\prime}$ is not cyclic, then it cannot be of class 2 or of maximal class and hence its order must be bigger than $p^4$. Our group is of order $32$ with elementary abelian derived group. \\ \noindent {\bf Example:} Consider the group $$G=\langle x,y| x^2y^{-4}=[x,y,x]=[x,y,y]y^{-4}=1\rangle$$ which is of nilpotency class 3. Since $d(G)=d(G/Z(G))=2$, $Z(G)\le \Phi(G)$. Take $u=[x,y]$. It is easily seen that every element of $G$ is of the form $x^iy^ju^k$, $1\le i,k \le 2$, $1\le j\le 8$; and $|x|=4, |y|=8$, $|u|=2$. Since $|\Phi(G)|=8$, $\Phi(G)=\langle y^2, u\rangle$. It is easy to see that $Z(G)=\langle y^4\rangle$ and $|Z(G)|=2$. If $|G'|=8$, then $|\gamma_3(G)|=4$ by [3, Theorem 1.5], a contradiction to $\gamma_3(G)\le Z(G)$. Thus $|G^{\prime}|=4$. Since $y^4,u\in G'$ are of order 2, $G^{\prime}$ is elementary abelian. It is obvious that $|G/Z(G)|=2^4=|G^{\prime}|^2$. \hfill$\Box$ \vspace{.2in} \noindent {\bf Acknowledgement.} Financial support of Council of Scientific and Industrial Research, Government of India for the research of second and third author is gratefully acknowledged.

https://arxiv.org/abs/1611.01655

Twenty (simple) questions

A basic combinatorial interpretation of Shannon's entropy function is via the "20 questions" game. This cooperative game is played by two players, Alice and Bob: Alice picks a distribution $\pi$ over the numbers $\{1,\ldots,n\}$, and announces it to Bob. She then chooses a number $x$ according to $\pi$, and Bob attempts to identify $x$ using as few Yes/No queries as possible, on average.An optimal strategy for the "20 questions" game is given by a Huffman code for $\pi$: Bob's questions reveal the codeword for $x$ bit by bit. This strategy finds $x$ using fewer than $H(\pi)+1$ questions on average. However, the questions asked by Bob could be arbitrary. In this paper, we investigate the following question: Are there restricted sets of questions that match the performance of Huffman codes, either exactly or approximately?Our first main result shows that for every distribution $\pi$, Bob has a strategy that uses only questions of the form "$x < c$?" and "$x = c$?", and uncovers $x$ using at most $H(\pi)+1$ questions on average, matching the performance of Huffman codes in this sense. We also give a natural set of $O(rn^{1/r})$ questions that achieve a performance of at most $H(\pi)+r$, and show that $\Omega(rn^{1/r})$ questions are required to achieve such a guarantee.Our second main result gives a set $\mathcal{Q}$ of $1.25^{n+o(n)}$ questions such that for every distribution $\pi$, Bob can implement an optimal strategy for $\pi$ using only questions from $\mathcal{Q}$. We also show that $1.25^{n-o(n)}$ questions are needed, for infinitely many $n$. If we allow a small slack of $r$ over the optimal strategy, then roughly $(rn)^{\Theta(1/r)}$ questions are necessary and sufficient.

\section{Introduction} \label{sec:introduction} A basic combinatorial and operational interpretation of Shannon's entropy function, which is often taught in introductory courses on information theory, is via the ``20 questions'' game (see for example the well-known textbook~\cite{DBLP:books/daglib/0016881}). This game is played between two players, Alice and Bob: Alice picks a distribution $\pi$ over a (finite) set of objects $X$, and announces it to Bob. Alice then chooses an object $x$ according to $\pi$, and Bob attempts to identify the object using as few Yes/No queries as possible, on average What questions should Bob ask? An optimal strategy for Bob is to compute a Huffman code for~$\pi$, and then follow the corresponding decision tree: his first query, for example, asks whether $x$ lies in the left subtree of the root. While this strategy minimizes the expected number of queries, the queries themselves could be arbitrarily complex; already for the first question, Huffman's algorithm draws from an exponentially large reservoir of potential queries (see Theorem~\ref{thm:minimum-redundancy-lb} for more details). Therefore, it is natural to consider variants of this game in which the set of queries used is restricted; for example, it is plausible that Alice and Bob would prefer to use queries that (i) can be communicated efficiently (using as few bits as possible), and (ii) can be tested efficiently (i.e.\ there is an efficient encoding scheme for elements of $X$ and a fast algorithm that given $x\in X$ and a query $q$ as input, determines whether $x$ satisfies the query $q$). We summarize this with the following meta-question, which guides this work: \emph{Are there ``nice'' sets of queries $\cQ$ such that for any distribution, there is a ``high quality'' strategy that uses only queries from $\cQ$?} Formalizing this question depends on how ``nice'' and ``high quality'' are quantified. We consider two different benchmarks for sets of queries: \begin{enumerate} \item \textbf{An information-theoretical benchmark}: A set of queries $\cQ$ has \emph{redundancy $r$} if for every distribution $\pi$ there is a strategy using only queries from $\cQ$ that finds $x$ with at most $H(\pi)+r$ queries on average when $x$ is drawn according to $\pi$. \item \textbf{A combinatorial benchmark}: A set of queries $\cQ$ is \emph{$r$-optimal} (or has \emph{prolixity $r$}) if for every distribution $\pi$ there is a strategy using queries from $\cQ$ that finds $x$ with at most $\opt{\pi}+r$ queries on average when $x$ is drawn according to $\pi$, where $\opt{\pi}$ is the expected number of queries asked by an optimal strategy for $\pi$ (e.g.\ a Huffman tree). \end{enumerate} Given a certain redundancy or prolixity, we will be interested in sets of questions achieving that performance that (i) are as small as possible, and (ii) allow efficient construction of high quality strategies which achieve the target performance. In some cases we will settle for only one of these properties, and leave the other as an open question. \paragraph{Information-theoretical benchmark.} Let $\pi$ be a distribution over $X$. A basic result in information theory is that every algorithm that reveals an unknown element $x$ drawn according to $\pi$ (in short, $x \sim \pi$) using Yes/No questions must make at least $H(\pi)$ queries on average. Moreover, there are algorithms that make at most $H(\pi) + 1$ queries on average, such as Huffman coding and Shannon--Fano coding. However, these algorithms may potentially use arbitrary queries. Are there restricted sets of queries that match the performance of $H(\pi)+1$ queries on average, for {\bf every} distribution $\pi$? Consider the setting in which $X$ is linearly ordered (say $X=[n]$, with its natural ordering: $1<\cdots<n$). Gilbert and Moore~\cite{GilbertMoore}, in a result that paved the way to arithmetic coding, showed that two-way comparison queries (``$x<c$?'') almost fit the bill: they achieve a performance of at most $H(\pi)+2$ queries on average. Our first main result shows that the optimal performance of $H(\pi)+1$ can be achieved by allowing also equality queries (``$x=c?$''): \begin{theorem*}[restatement of Theorem~\ref{thm:redundancy-1}] For every distribution $\pi$ there is a strategy that uses only comparison and equality queries which finds $x$ drawn from $\pi$ with at most $H(\pi)+1$ queries on average. Moreover, this strategy can be computed in time $O(n\log n)$. \end{theorem*} In a sense, this result gives an affirmative answer to our main question above. The set of comparison and equality queries (first suggested by Spuler~\cite{Spuler}) arguably qualifies as ``nice'': linearly ordered universes appear in many natural and practical settings (numbers, dates, names, IDs) in which comparison and equality queries can be implemented efficiently. Moreover, from a communication complexity perspective, Bob can communicate a comparison/equality query using just $\log_2 n + 1$ bits (since there are just $2n$ such queries). This is an exponential improvement over the $\Omega(n)$ bits he would need to communicate had he used Huffman coding. We extend this result to the case where $X$ is a set of vectors of length $r$, $\vec x = (x_1,\ldots,x_r)$, by showing that there is a strategy using entry-wise comparisons (``$x_i<c?$'') and entry-wise equalities (``$x_i=c?$'') that achieves redundancy $r$: \begin{theorem*}[restatement of Theorem~\ref{thm:redundancy-r-bounds}] Let $X$ be the set of vectors of length $r$ over a linearly ordered universe. For every distribution $\pi$ there is a strategy that uses only entry-wise comparison queries and entry-wise equality queries and finds $\vec x\sim\pi$ with at most $H(\pi) + r$ queries. \end{theorem*} \providecommand{\find}{\mathit{find}} The theorem is proved by applying the algorithm of the preceding theorem to uncover the vector $\vec{x}$ entry by entry. As a corollary, we are able to determine almost exactly the minimum size of a set of queries that achieves redundancy $r\geq 1$. In more detail, let $\uent(n,r)$ denote the minimum size of a set of queries $\cQ$ such that for every distribution $\pi$ on $[n]$ there is a strategy using only queries from $\cQ$ that finds $x$ with at most $H(\pi)+r$ queries on average, when $x$ is drawn according to $\pi$. \begin{corollary*}[Theorem~\ref{thm:redundancy-r-bounds}] For every $n,r\in\mathbb{N}$, \[ \frac{1}{e} r n^{1/ r } \leq \uent(n,r) \leq 2 r n^{1/ r }. \] \end{corollary*} Obtaining this tight estimate of $\uent(n,r) = \Theta(r n^{1/ r })$ hinges on adding equality queries; had we used only entry-wise comparison queries and the Gilbert--Moore algorithm instead, the resulting upper bound would have been $\uent(n,r) = O\bigl(rn^{2/r}\bigr)$, which is quadratically worse than the truth. \medskip \paragraph{Combinatorial benchmark.} The analytical properties of the entropy function make $H(\pi)$ a standard benchmark for the average number of bits needed to describe an element $x$ drawn from a known distribution $\pi$, and so it is natural to use it as a benchmark for the average number of queries needed to find $x$ when it is drawn according to $\pi$. However, there is a conceptually simpler, and arguably more natural, benchmark: $\opt{\pi}$ --- the average number of queries that are used by a best strategy for $\pi$ (several might exist), such as one generated by Huffman's algorithm. Can the optimal performance of Huffman codes be matched \emph{exactly}? Can it be achieved without using all possible queries? Our second main result answers this in the affirmative: \begin{theorem*}[restatement of Theorem~\ref{thm:minimum-redundancy-ub} and Theorem~\ref{thm:minimum-redundancy-lb}] For every $n$ there is a set $\cQ$ of $1.25^{n+o(n)}$ queries such that for every distribution over $[n]$, there is a strategy using only queries from $\cQ$ which matches the performance of the optimal (unrestricted) strategy \emph{exactly}. Furthermore, for infinitely many $n$, at least $1.25^{n-o(n)}$ queries are required to achieve this feat. \end{theorem*} One drawback of our construction is that it is randomized. Thus, we do not consider it particularly ``efficient'' nor ``natural''. It is interesting to find an explicit set $\cQ$ that achieves this bound. Our best explicit construction is: \begin{theorem*}[restatement of Theorem~\ref{thm:cone}] For every $n$ there is an explicit set $\cQ$ of $O(2^{n/2})$ queries such that for every distribution over $[n]$, there is a strategy using only queries from $\cQ$ which matches the performance of the optimal (unrestricted) strategy \emph{exactly}. Moreover, we can compute this strategy in time $O(n^2)$. \end{theorem*} It is natural to ask in this setting how small can a set of queries be if it is \emph{$r$-optimal}; that is, if it uses at most $\opt{\pi}+r$ questions on average when the secret element is drawn according to $\pi$, for small $r > 0$. Let $\uhuf(n,r)$ denote the minimum size of a set of queries $\cQ$ such that for every distribution $\pi$ on $[n]$ there is a strategy using only queries from $\cQ$ that finds $x$ with at most $\opt{\pi}+r$ queries on average when $x$ is drawn from $\pi$. We show that for any fixed $r>0$, significant savings can be achieved: \begin{theorem*}[restatement of Theorem~\ref{thm:min-redundancy-approx}] For all $r\in(0,1)$: \[(r\cdot n)^{\frac{1}{4r}}\lesssim \uhuf(n,r) \lesssim (r\cdot n)^{\frac{16}{r}}. \] \end{theorem*} Instead of the exponential number of questions needed to match Huffman's algorithm exactly, for fixed $r>0$ an $r$-optimal set of questions has polynomial size. In this case the upper bound is achieved by an explicit set of queries $\cQ_r$. We also present an efficient randomized algorithm for computing an $r$-optimal strategy that uses queries from $\cQ_r$. \paragraph{Related work} The ``20 questions'' game is the starting point of combinatorial search theory~\cite{Katona,AhlswedeWegener,ACD}. Combinatorial search theory considers many different variants of the game, such as several unknown elements, non-adaptive queries, non-binary queries, and a non-truthful Alice~\cite{RMKWS,SW,AD,DGW}. Both average-case and worst-case complexity measures are of interest. An important topic in combinatorial search theory is combinatorial group testing~\cite{Dorfman,DuHwang}. We are unaware of any prior work which has considered the quantities $\uent(n,r)$ or $\uhuf(n,r)$. However, several particular sets of questions have been analyzed in the literature from the perspective of redundancy and prolixity: Huffman codes~\cite{Gallager,Johnsen,CGT,MontgomeryAbrahams,Manstetten,MPK}, binary search trees~\cite{GilbertMoore,Nakatsu,Rissanen,Horibe}, and comparison-based sorting algorithms~\cite{F76,MY13}. \section{Preliminaries} \label{sec:definitions} \paragraph{Notation} We use $\log n$ for the base~$2$ logarithm of $n$ and $[n]$ to denote the set $\{1,\ldots,n\}$. Throughout the document, we will consider probability distributions over the set $X_n = \{x_1,\ldots,x_n\}$ of size $n$. In some cases, we will think of this set as ordered: $x_1 \prec \cdots \prec x_n$. If $\pi$ is a probability distribution over $X_n$, we will denote the probability of $x_i$ by $\pi_i$, and the probability of a set $S \subseteq X_n$ by $\pi(S)$. \paragraph{Information theory} We use $H(\pi)$ to denote the base~$2$ entropy of a distribution $\pi$ and $D(\pi\|\mu)$ to denote the Kullback--Leibler divergence. The \emph{binary entropy} function $h(p)$ is the entropy of a Bernoulli random variable with success probability $p$. When $Y$ is a Bernoulli random variable, the chain rule takes the following form: \[ H(X,Y) = h(\Pr[Y=1]) + \Pr[Y=0] H(X|Y=0) + \Pr[Y=1] H(X|Y=1). \] We call this the \emph{Bernoulli chain rule}. \paragraph{Decision trees} In this paper we consider the task of revealing \anunknown element $x$ from $X_n$ by using Yes/No questions. Such a strategy will be called a \emph{decision tree} or an \emph{algorithm}. A decision tree is a binary tree in which the internal nodes are labeled by subsets of $X_n$ (which we call \emph{questions} or \emph{queries}), each internal node has two outgoing edges labeled \emph{Yes} (belongs to the question set) and \emph{No} (doesn't belong to the question set), and the leaves are labeled by distinct elements of $X_n$. The depth of an element $x_i$ in a decision tree $T$, denoted $T(x_i)$, is the number of edges in the unique path from the root to the unique leaf labeled $x_i$ (if any). Decision trees can be thought of as annotated prefix codes: the code of an element $x_i$ is the concatenation of the labels of the edges leading to it. The mapping can also be used in the other direction: each binary prefix code of cardinality~$n$ corresponds to a unique decision tree over $X_n$. Given a set $\cQ \subseteq 2^{X_n}$ (a set of \emph{allowed questions}), a \emph{decision tree using $\cQ$} is one in which all questions belong to $\cQ$. A decision tree is \emph{valid} for a distribution $\mu$ if its leaves are labeled by elements of the support of $\mu$, and each element in the support appears as the label of some leaf. Given a distribution $\mu$ and a decision tree $T$ valid for $\mu$, the \emph{cost} (or \emph{query complexity}) of $T$ on $\mu$, labeled $T(\mu)$, is the average number of questions asked on a random element, $T(\mu) = \sum_{i=1}^n \mu_i T(x_i)$. Given a set $\cQ$ of allowed questions and a distribution $\mu$, the \emph{optimal cost} of $\mu$ with respect to $\cQ$, denoted $\costt{\cQ}{\mu}$, is the minimal cost of a valid decision tree for $\mu$ using~$\cQ$. \paragraph{Dyadic distributions and Huffman's algorithm} Huffman's algorithm~\cite{Huffman} finds the optimal cost of an unrestricted decision tree for a given distribution: \[ \opt{\mu} = \costt{2^{X_n}}{\mu}. \] We call a decision tree with this cost a \emph{Huffman tree} or an \emph{optimal decision tree} for $\mu$. More generally, a decision tree is \emph{$r$-optimal} for $\mu$ if its cost is at most $\opt{\mu} + r$. It will be useful to consider this definition from a different point of view. Say that a distribution is \emph{dyadic} if the probability of every element is either $0$ or of the form $2^{-d}$ for some integer~$d$. We can associate with each decision tree $T$ a distribution $\tau$ on the leaves of $T$ given by $\tau_i = 2^{-T(x_i)}$. This gives a correspondence between decision trees and dyadic distributions. In the language of dyadic distributions, Huffman's algorithm solves the following optimization problem: \[ \opt{\mu} = \min_{\substack{\tau \text{ dyadic} \\ \supp \tau = \supp \mu}} \sum_{i=1}^n \mu_i \log \frac{1}{\tau_i} = \min_{\substack{\tau \text{ dyadic} \\ \supp \tau = \supp \mu}} \left[ H(\mu) + D(\mu \| \tau) \right]. \] In other words, computing $\opt{\mu}$ amounts to minimizing $D(\mu\|\tau)$, and thus to ``rounding'' $\mu$ to a dyadic distribution. We call $\tau$ a \emph{Huffman distribution} for $\mu$. The following classical inequality shows that $\opt{\mu}$ is very close to the entropy of $\mu$: \[ H(\mu) \leq \opt{\mu} < H(\mu) + 1. \] The lower bound follows from the non-negativity of the Kullback--Leibler divergence; it is tight exactly when $\mu$ is dyadic. The upper bound from the Shannon--Fano code, which corresponds to the dyadic sub-distribution $\tau_i = 2^{-\lceil \log \mu_i \rceil}$ (in which the probabilities could sum to less than~$1$). \paragraph{Redundancy and prolixity} We measure the quality of sets of questions by comparing the cost of decision trees using them to the entropy (the difference is known as \emph{redundancy}) and to the cost of optimal decision trees (for the difference we coin the term \emph{prolixity}). In more detail, the \emph{redundancy} of a decision tree $T$ for a distribution $\mu$ is $T(\mu) - H(\mu)$, and its \emph{prolixity} is $T(\mu) - \opt{\mu}$. Given a set of questions $\cQ$, the redundancy $\rent(\cQ,\mu)$ and prolixity $\rhuf(\cQ,\mu)$ of a distribution $\mu$ are the best redundancy and prolixity achievable using questions from $Q$: \begin{align*} \rent(\cQ,\mu) &= \costt{\cQ}{\mu} - H(\mu), \\ \rhuf(\cQ,\mu) &= \costt{\cQ}{\mu} - \opt{\mu}. \end{align*} The redundancy of a set of questions $\cQ$, denoted $\rent(\cQ)$, is the supremum of $\rent(\cQ,\mu)$ over all distributions $\mu$ over $X_n$. The prolixity $\rhuf(\cQ)$ of $\cQ$ is defined similarly. These quantities are closely related, as the inequality $H(\mu) \leq \opt{\mu} < H(\mu)+1$ implies: \[ \rhuf(\cQ) \leq \rent(\cQ) \leq \rhuf(\cQ) + 1. \] A set of questions $\cQ$ is \emph{optimal} if $\rhuf(\cQ) = 0$, and $r$-optimal if $\rhuf(\cQ) \leq r$. \yuvalf{Not sure we use the latter.} \paragraph{The parameters $\uent(n,r)$ and $\uhuf(n,r)$} Our main object of study in this paper are the parameters $\uent(n,r)$ and $\uhuf(n,r)$. The parameter $\uent(n,r)$ is the cardinality of the minimal set of questions $\cQ \subseteq 2^{X_n}$ such that $\rent(\cQ) \leq r$. Similarly, the parameter $\uhuf(n,r)$ is the cardinality of the minimal set of questions $\cQ$ such that $\rhuf(\cQ) \leq r$. These quantities are closely related: \[ \uhuf(n,r) \leq \uent(n,r) \leq \uhuf(n,r-1). \] \paragraph{A useful lemma} The following simple lemma will be used several times in the rest of the paper. \begin{main-lemma} \label{lem:neat-sum} Let $p_1 \geq \ldots \geq p_n$ be a non-increasing list of numbers of the form $p_i = 2^{-a_i}$ (for integer $a_i$), and let $a \leq a_1$ be an integer. If $\sum_{i=1}^n p_i \geq 2^{-a}$ then for some $m$ we have $\sum_{i=1}^m p_i = 2^{-a}$. If furthermore $\sum_{i=1}^n p_i$ is a multiple of $2^{-a}$ then for some $\ell$ we have $\sum_{i=\ell}^n p_i = 2^{-a}$. \end{main-lemma} \begin{proof} Let $m$ be the maximal index such that $\sum_{i=1}^m p_i \leq 2^{-a}$. If $m = n$ then we are done, so suppose that $m < n$. Let $S = \sum_{i=1}^m p_i$. We would like to show that $S = 2^{-a}$. The condition $p_1 \leq \cdots \leq p_n$ implies that $a_{m+1} \geq \cdots \geq a_1$, and so $k \ensuremath{:=}\xspace 2^{a_{m+1}} S = \sum_{i=1}^m 2^{a_{m+1}-a_i}$ is an integer. By assumption $k \leq 2^{a_{m+1}-a}$ whereas $k + 1 = 2^{a_{m+1}} \sum_{i=1}^{m+1} p_i > 2^{a_{m+1}-a}$. Since $2^{a_{m+1}-a}$ is itself an integer (since $a_{m+1} \geq a_1 \geq a$), we conclude that $k = 2^{a_{m+1}-a}$, and so $S = 2^{-a}$. \smallskip To prove the furthermore part, notice that by repeated applications of the of the lemma we can partition $[n]$ into intervals whose probabilities are $2^{-a}$. The last such interval provides the required index~$\ell$. \end{proof} \section{Comparisons and equality tests} \label{sec:comparison-equality} Let $\pi$ be a distribution over $X_n = \{ x_1,\ldots,x_n \}$. A fundamental result in information theory is that the entropy of a distribution $\pi $ captures the average number of queries needed to identify a random $x\sim\pi$. More specifically, every algorithm asks at least $H(\pi)$ questions in expectation, and there are algorithms that ask at most $H(\pi) + 1$ questions on average (such as Huffman coding and Shannon--Fano coding). However, these algorithms may potentially use arbitrary questions. In this section we are interested in the setting where $X_n$ is linearly ordered: $x_1 \prec x_2 \prec \dots \prec x_n$. We wish to use questions that are compatible with the ordering. Perhaps the most natural question in this setting is a comparison query; namely a question of the form ``$x \prec x_i$?''. Gilbert and Moore~\cite{GilbertMoore} showed that there exists an algorithm that uses at most $H(\pi) + 2$ comparisons. Is this tight? Can comparison queries achieve the benchmark of $H(\pi) + 1$? A simple argument shows that their result is tight: let $n = 3$, and let $\pi$ be a distribution such that \[ \pi(x_1) = \epsilon/2,~~~~ \pi(x_2)=1-\epsilon,~~~~ \pi(x_3) = \epsilon/2, \] for some small $\epsilon$. Note that $H(\pi) = O(\epsilon\log\bigl(1/\epsilon\bigr))$, and therefore any algorithm with redundancy 1 must use the query ``$x=x_2$?'' as its first question. This is impossible if we only allow comparison queries (see Lemma~\ref{l52} for a more detailed and general argument). In fact, this shows that any set of questions that achieves redundancy $1$ must include all equality queries. So, we need to at least add all equality queries. Is it enough? Do comparison and equality queries achieve redundancy of at most 1? Our main result in this section gives an affirmative answer to this question: \begin{main-theorem} \label{thm:redundancy-1} Let $\qeq^{(n)} = \{ \{ x_i \} : 1 \leq i \leq n \}$ and $\qcomp^{(n)} = \{ \{x_1,\ldots,x_i\} : 1 \leq i \leq n-1 \}$. In other words, $\qeq^{(n)}$ consists of the questions ``$x = x_i$?'' for $i \in \{1,\ldots,n\}$, and $\qcomp^{(n)}$ consists of the questions ``$x \prec x_i$?'' for $i \in \{2,\ldots,n\}$. (Recall that $x$ is the \unknown element.) For all $n$, $\rent(\qcomp^{(n)} \cup \qeq^{(n)}) = 1$. \end{main-theorem} We prove the theorem by modifying the weight-balancing algorithm of Rissanen~\cite{Rissanen}, which uses only comparison queries and achieves redundancy~$2$ (as shown by Horibe~\cite{Horibe}). The original weight-balancing algorithm is perhaps the first algorithm that comes to mind: it asks the most balanced comparison query (the one that splits the distribution into two parts whose probability is as equal as possible), and recurses according to the answer. Our modified algorithm, Algorithm~$A_t$, first checks whether the probability of the most probable element $x_{\max}$ exceeds the threshold~$t$. If so, it asks the question ``$x=x_{\max}$?''. Otherwise, it proceeds as in the weight-balancing algorithm. The choice $t=0.3$ results in an algorithm whose redundancy is at most~$1$. While a naive implementation of Algorithm~$A_t$ takes time $O(n)$ to determine the next query, this can be improved to $O(\log n)$, given $O(n)$ preprocessing time. Moreover, the entire strategy can be computed in time $O(n\log n)$. The interested reader can find the complete details in~\cite{DaganThesis}. \subsection{Algorithm \texorpdfstring{$A_t$}{At}} Let $\pi$ be a distribution over $X_n$. Let $\xmax$ denote the most probable element, and let $\pimax$ denote its probability. Let $\xmid$ denote a point $x_i$ that minimizes $|\pi(\{x: x \prec x_i\})-1/2|$ over $i\in [n]$. We call $\xmid$ the middle\footnote{Note that one of $\{\xmid,x_{\Mid-1}\}$ is a median.} of $\pi$. Note that the query ``$x\prec \xmid$?'' is the most balanced query among all queries of the form ``$x \prec x_i$?''. Let $A\subseteq\{x_1,\ldots,x_n\}$. We use $\pi_A$ to denote $\pi(A)$; i.e.\ the probability of $A$. Specifically, we use $\pi_{\prec x_i}, \pi_{\succeq x_i}, \pi_{\neq x_i}$ to denote $\pi_A$ when $A$ is $\{x: x \prec x_i\}, \{x: x \succeq x_i\}, \{x : x\neq x_i\}$. We use $\pi|_A$ to denote the restriction of $\pi$ to $A$; i.e., the distribution derived by conditioning $\pi$ on $A$. Specifically, we use $\pi|_{\prec x_i}, \pi|_{\succeq x_i}, \pi|_{\neq x_i}$ to denote $\pi|_A$ when $A$ is $\{x: x \prec x_i\}, \{x: x \succeq x_i\}, \{x : x\neq x_i\}$. \begin{algorithm-description}{$A_t$} Given a threshold $t\in (0,1)$, Algorithm~$A_t$ takes as input a distribution $\pi$ over $X_n$ and \anunknown element $x$, and determines $x$ using only comparison and equality queries, in the following recursive fashion: \begin{enumerate} \item If $\pi(x_i) = 1$ for some element $x_i$, then output $x_i$. \item If $\pimax \geq{t}$ then ask whether $x = \xmax$, and either output $\xmax$, or continue with $\pi|_{\neq \xmax}$. \item If $\pimax < {t}$, ask whether $x \prec \xmid$, and continue with either $\pi|_{\prec \xmid}$ or $\pi|_{\succeq\xmid}$. \end{enumerate} (When recursing on a domain $D$, we identify $D$ with $X_{|D|}$.) \end{algorithm-description} The weight balancing algorithm of Rissanen~\cite{Rissanen} is the special case $t = 1$; in this case no equality queries are needed, and the resulting redundancy is~$2$, as shown by Horibe~\cite{Horibe}. We will show that for a certain range of values of ${t}$ (for example, ${t} = 0.3$), Algorithm~$A_t$ achieves redundancy at most~$1$, thus proving Theorem~\ref{thm:redundancy-1}. \smallskip Recall that $A_{{t}}(\pi)$ is the cost of $A_t$ on $\pi$, and let $R_t(\pi) \ensuremath{:=}\xspace A_t(\pi) - H(\pi)-1$. It is more convenient to present the proof in terms of $R_t(\pi)$ rather than in terms of the redundancy. Our goal, stated in these terms, is showing that there exists some $t$ for which $R_t(\pi)\leq 0$ for all distributions $\pi$. We next observe two simple properties of the algorithm $A_t$. The proof of Theorem~\ref{thm:redundancy-1} relies only on these properties. The first property is a recursive definition of $R_t$ that is convenient to induct on. See Figure~\ref{fig:recdef} for a pictorial illustration. \providecommand{\thet}{t} \providecommand{\theR}{R_t} \begin{figure} \begin{subfigure}{.5\linewidth} \begin{tikzpicture}[sibling distance=5cm] \node[circle,draw,fill=black,label={[label distance=0.5cm]below:\small $1-h(\pi_{\succeq \xmid})$}, label={[label distance=0.15cm]above:\small ``$x\prec\xmid$?''}] {} child{ node[circle, draw, fill=black] {} node[itria] { $\theR(\pi|_{ \prec \xmid})$} edge from parent node[near end, above=10pt] {\small $\pi_{\prec \xmid}$} } child{ node[circle, draw, fill=black] {} node[itria] { $\theR(\pi|_{\succeq \xmid})$} edge from parent node[near end, above=10pt] {\small $\pi_{\succeq \xmid}$} }; \end{tikzpicture} \caption{$\pimax\in (0,\thet)$} \label{fig-lem-b-t} \end{subfigure} \begin{subfigure}{.5\linewidth} \begin{tikzpicture}[sibling distance=5cm] \node[circle,draw,fill=black, label={[label distance=0.5cm]below:\small $1-h(\pi_{\neq \xmax})$},label={[label distance=0.15cm]above:\small ``$x\neq\xmax$?''}] {} child{ node[circle, draw, label={[label distance=0.125cm]below:\small $\theR(\pi|_{=\xmax})=-1$}] {\small $\xmax$} edge from parent node[near end, above=10pt] {\small $\pimax$} } child{ node[circle, draw, fill=black] {} node[itria] {\small $\theR(\pi|_{ \neq \xmax})$} edge from parent node[near end, above=10pt] {\small $1-\pimax$} } ; \end{tikzpicture} \caption{$\pimax\in [\thet,1)$} \label{fig-lem-b-m} \end{subfigure} \caption{Recursive definition of $\theR$}\label{fig:recdef} \end{figure} \begin{lemma} \label{lem:redundancy-1-formula} Let $\pi$ be a distribution over $X_n$. Then \[ R_t(\pi) = \begin{cases} -1 & \text{if } \pimax = 1, \\ 1 - h(\pimax) - \pimax + (1-\pimax) R_t(\pi|_{\neq \xmax}) & \text{if } \pimax\in[t,1), \\ 1 - h(\pi_{\prec\xmid}) + \pi_{\prec\xmid} R_t(\pi|_{\prec\xmid}) + \pi_{\succeq\xmid} R_t(\pi|_{\succeq\xmid}) & \text{if } \pimax\in(0,t). \end{cases} \] \end{lemma} \begin{proof} If $\pimax = 1$ then $A_t(\pi) = 0$ and $H(\pi) = 0$, so $R_t(\pi) = -1$. If $\pimax \in [t,1)$ then \begin{align*} R_t(\pi) &= A_t(\pi) - H(\pi) - 1 \\ &= \Bigl[1 + (1-\pimax)A_t(\pi|_{\neq \xmax})\Bigr] - \Bigl[h(\pimax) + (1-\pimax)H(\pi|_{\neq \xmax})\Bigr] - \Bigl[\pimax + (1-\pimax)\Bigr] \\ &= 1 - h(\pimax) - \pimax + (1-\pimax) \Bigl[A_t(\pi|_{\neq \xmax}) - H(\pi|_{\neq \xmax}) - 1\Bigr] \\ &= 1 - h(\pimax) - \pimax + (1-\pimax) R_t(\pi_{\neq \xmax}). \end{align*} If $\pimax \in (0,t)$ then \begin{align*} R_t(\pi) &= A_t(\pi) - H(\pi) - 1 \\ &= \Bigl[1 + \pi_{\prec\xmid} A_t(\pi|_{\prec\xmid}) + \pi_{\succeq\xmid} A_t(\pi|_{\succeq\xmid})\Bigr] \\&\qquad - \Bigl[h(\pi_{\prec \xmid}) + \pi_{\prec\xmid} H(\pi|_{\prec\xmid}) + \pi_{\succeq\xmid} H(\pi|_{\succeq\xmid})\Bigr] - \Bigl[\pi_{\prec\xmid} + \pi_{\succeq\xmid}\Bigr] \\ &= 1 - h(\pi_{\prec \xmid}) + \pi_{\prec\xmid} \Bigl[A_t(\pi|_{\prec\xmid}) - H(\pi|_{\prec\xmid}) - 1\Bigr] + \pi_{\succeq\xmid} \Bigl[A_t(\pi|_{\succeq\xmid}) - H(\pi|_{\succeq\xmid}) - 1\Bigr] \\ &= 1 - h(\pi_{\prec\xmid}) + \pi_{\prec\xmid} R_t(\pi|_{\prec\xmid}) + \pi_{\succeq\xmid} R_t(\pi|_{\succeq\xmid}). \qedhere \end{align*} \end{proof} The second property is that whenever $A_t$ uses a comparison query (i.e.\ when $\pimax < t$), then this question is balanced: \begin{lemma}\label{l412} Let $\pi$ be a distribution over $X_n$. Then $\pi_{\prec\xmid},\pi_{\succeq\xmid}\in[\frac{1-\pimax}{2},\frac{1+\pimax}{2}]$. \end{lemma} \begin{proof} By the definition of $\xmid$, it suffices to show that there exists some $j$ with $\pi_{\prec j}\in [\frac{1-\pimax}{2},\frac{1+\pimax}{2}]$. Indeed, for all $j$, $\pi_{\prec j+1} - \pi_{\prec j} = \pi_j\leq\pimax$, and therefore, if $j$ is the maximum element with $\pi_{\prec j}<\frac{1}{2}$ (possibly $j=0$), then either $\pi_{\prec j}$ or $\pi_{\prec j+1}$ are in $[\frac{1-\pimax}{2},\frac{1+\pimax}{2}]$. \end{proof} \subsection{Game \texorpdfstring{$G_t$}{Gt}} We will use the properties described in Lemma~\ref{l412} to bound $R_t(\pi)$. Since this involves quantifying over all distributions $\pi$, it is convenient to introduce a game that simulates $A_t$ on a distribution $\pi$. The game involves one player, Alice, who we think of as an adversary that chooses the input distribution $\pi$ (in fact, she only chooses $\pimax$), and wins a revenue of $R_t(\pi)$ (thus, her objective is to maximize the redundancy). This reduces our goal to showing that Alice's optimum revenue is nonpositive. The definition of the game is tailored to the properties stated in Lemma~\ref{lem:redundancy-1-formula} and Lemma~\ref{l412}. We first introduce $G_t$, and then relate it to the redundancy of $A_t$ (see Lemma~\ref{l416} below). \begin{algorithm-description}[Game]{$G_t$} Let $t\leq\frac{1}{3}$, and let $f,s\colon (0,1]\rightarrow \mathbb{R}$ be $f(x) \ensuremath{:=}\xspace 1-h(x) - x$ and $s(x) \ensuremath{:=}\xspace 1-h(x)$. The game $G_t$ consists of one player called Alice, whose objective is to maximize her revenue. The game $G_t$ begins at an initial state $p\in (0,1]$, and proceeds as follows. \begin{enumerate} \item If $p\in [t,1]$, the game ends with revenue $f(p)$. \item If $p\in (0,t)$, then Alice chooses a state\footnote{Note that $p'\in(0,1]$, since $p<t\leq\frac{1}{3}$ implies that $\bigl[\frac{2p}{1+p}, \frac{2p}{1-p}\bigr]\subseteq(0,1]$.} $p'\in\bigl[\frac{2p}{1+p}, \frac{2p}{1-p}\bigr]$ and recursively plays $G_t$ with initial state $p'$. Let $r'$ denote her revenue in the game that begins at $p'$. Alice's final revenue is \[ s\biggl(\frac{p}{p'}\biggr) + \frac{p}{p'}\cdot r'. \eofahere \] \end{enumerate} \end{algorithm-description} Note that given any initial state $p$ and a strategy for Alice, the game $G_t$ always terminates: indeed, if $p=p_0,p_1,p_2,\ldots$ is the sequence of states chosen by Alice, then as long as $p_i < t$, it holds that $p_{i+1} \geq \frac{2p_i}{1+p_i} > \frac{2p_i}{1+1/3}=\frac{3}{2}p_i$ (the second inequality is since $p_i < t\leq\frac{1}{3}$). So the sequence of states grows at an exponential rate, which means that for $\ell=O(\log(1/p_0))$, the state $p_\ell$ exceeds the threshold $t$ and the game terminates. For $p\in (0,1]$, let $r_t(p)$ denote the supremum of Alice's revenue in $G_t$ when the initial state is $p$, the supremum ranging over all possible strategies for Alice. Our next step is using the game $G_t$ to prove Theorem~\ref{thm:redundancy-1}: We will show that $t=0.3$ satisfies: \begin{enumerate}[label=(\roman*)] \item $r_t(p)\leq 0$ for all $p\in(0,1]$, and \item $R_t(\pi)\leq r_t(\pimax)$ for all $\pi$. \end{enumerate} Note that (i) and (ii) imply Theorem~\ref{thm:redundancy-1}. Before establishing (i) and (ii), we state and prove three simple lemmas regarding $G_t$ that are useful to this end. The first lemma will be used in the proof of Lemma~\ref{l416}, which shows that if $r_t(p)\leq 0$ for all $p\in (0,1]$, then $R_t(\pi)\leq r_t(\pimax)$. Its proof follows directly from the definition of $r_t$. \begin{lemma}\label{l413} For all $p < t$ and all $p'\in\big[\frac{2p}{1+p}, \frac{2p}{1-p}\big]$: \[r_t(p) \geq s\left(\frac{p}{p'}\right) + \frac{p}{p'}\cdot r_t(p').\] \end{lemma} The next two lemmas will be used in the proof of Lemma~\ref{l417}, which shows that $r_t(p)\leq 0$ for all $p\in (0,1]$ when $t=0.3$. The first one gives a tighter estimate on the growth of the sequence of states: \begin{lemma}\label{l414} Let $p_0,p_1,\ldots,p_k$ be the sequence of states chosen by Alice. For every $i\leq k-1$: \[ p_{k-1-i} < \frac{t}{2^i(1-t)+t}. \] \end{lemma} \begin{proof} We prove the bound by induction on $i$. The case $i=0$ follows from $p_{k-1}$ being a non-final state, and therefore $p_{k-1} < t$ as required. Assume now that $i > 0$. By the definition of $G_t$ it follows that $p_{k-1-(i-1)}\in\bigl[\frac{2p_{k-1-i}}{1+p_{k-1-i}}, \frac{2p_{k-1-i}}{1-p_{k-1-i}}\bigr]$, which implies that $p_{k-1-i} \in \bigl[\frac{p_{k-1-(i-1)}}{2+p_{k-1-(i-1)}}, \frac{p_{k-1-(i-1)}}{2-p_{k-1-(i-1)}}\bigr]$. Therefore, \begin{align*} p_{k-1-i} &\leq \frac{p_{k-1-(i-1)}}{2-p_{k-1-(i-1)}}\\ &< \frac{t/\bigl(2^{i-1}(1-t)+t\bigr)}{2-t/\bigl(2^{i-1}(1-t)+t\bigr)}\tag{by induction hypothesis on $i-1$}\\ &=\frac{t}{2^i(1-t)+t}. \qedhere \end{align*} \end{proof} The second lemma gives a somewhat more explicit form of the revenue of $G_t$: \begin{lemma}\label{l415} Let $p_0,p_1,\ldots, p_k$ be the sequence of states chosen by Alice. Let $r(p_0,\ldots,p_k)$ denote the revenue obtained by choosing these states. Then \[r(p_0,\ldots,p_k) = \sum_{i=0}^{k-1}\frac{p_0}{p_i} s\biggl(\frac{p_{i}}{p_{i+1}}\biggr) + \frac{p_{0}}{p_k}f(p_k).\] \end{lemma} \begin{proof} We prove the formula by induction on $k$. If $k=0$ then $p_k \geq t$ and $r(p_k) = f(p_k) = \frac{p_0}{p_k}f(p_k)$. When $k\geq 1$: \begin{align*} r(p_0,\ldots,p_k) &= s\biggl(\frac{p_0}{p_1}\biggr) + \frac{p_0}{p_1}\cdot r(p_1,\dots, p_k)\tag{by definition of $G_t$}\\ &= \frac{p_0}{p_0} s\biggl(\frac{p_0}{p_1}\biggr) + \frac{p_0}{p_1}\cdot \Bigl(\sum_{i=1}^{k-1}\frac{p_1}{p_i} s\biggl(\frac{p_{i}}{p_{i+1}}\biggr) + \frac{p_{1}}{p_k}f(p_k)\Bigr)\tag{by induction hypothesis}\\ &= \sum_{i=0}^{k-1}\frac{p_0}{p_i} s\biggl(\frac{p_{i}}{p_{i+1}}\biggr) + \frac{p_{0}}{p_k}f(p_k). \qedhere \end{align*} \end{proof} \subsection{Relating \texorpdfstring{$A_t$}{At} to \texorpdfstring{$G_t$}{Gt}} Next, we relate the revenue in $G_t$ to the redundancy of $A_t$ by linking $r_t$ and $R_t$. We first reduce the Theorem~\ref{thm:redundancy-1} to showing that there exists some $t\in (0,1]$ such that $r_t(p)\leq 0$ for all $p\in (0,1]$ (Lemma~\ref{l416}), and then show that $t = 0.3$ satisfies this condition (Lemma~\ref{l417}). \begin{lemma}\label{l416} Let $t$ be such that $r_t(p)\leq 0$ for all $p\in(0,1]$. For every distribution $\pi$, $$R_t(\pi)\leq r_t(\pimax).$$ In particular, such $t$ satisfies $R_t(\pi)\leq 0$ for all $\pi$. \end{lemma} \begin{proof} We proceed by induction on the size of $\supp{\pi}=\{x_i : \pi(x_i)\neq 0\}$. If $\bigl\lvert\supp{\pi}\bigr\rvert = 1$ then $\pimax=1$, and therefore $R_t(\pi)=-1$, $r_t(\pimax)=0$, and indeed $R_t(\pi)\leq r_t(\pimax)$. Assume now that $\bigl\lvert\supp{\pi}\bigr\rvert = k > 1$. Since $k>1$, it follows that $\pimax < 1$. There are two cases, according to whether $\pimax\in(0,t)$ or $\pimax\in [t,1)$. If $\pimax\in(0,t)$ then $A_t$ asks whether $x \prec \xmid$, and continues accordingly with $\pi|_{\prec\xmid}$ or $\pi|_{\succeq\xmid}$. Let $\sigma \ensuremath{:=}\xspace \pi|_{\prec\xmid}$ and $\tau \ensuremath{:=}\xspace \pi|_{\succeq\xmid}$. By Lemma~\ref{lem:redundancy-1-formula}: \begin{align*} R_t(\pi) &= 1 - h(\pi_{\prec\xmid}) + \pi_{\prec\xmid} R_t(\sigma) + \pi_{\succeq\xmid} R_t(\tau)\tag{since $\pimax\in(0,t)$}\\ &\leq 1 - h(\pi_{\prec\xmid}) + \pi_{\prec\xmid} r_t(\sigma_{\max}) + \pi_{\succeq\xmid} r_t(\tau_{\max})\tag{by induction hypothesis} \end{align*} Without loss of generality, assume that $\xmax \prec \xmid$. Therefore $\sigma_{\max} = {\pimax}/{\pi_{\prec\xmid}}$, and by Lemma ~\ref{l412}: \begin{equation}\label{eq1} \sigma_{\max}\in \left[\frac{2\pimax}{1+\pimax},\frac{2\pimax}{1-\pimax}\right]. \end{equation} Thus, \begin{align*} R_t(\pi) &\leq 1 - h(\pi_{\prec\xmid}) + \pi_{\prec\xmid} r_t(\sigma_{\max})\tag{since $r_t(\tau_{\max})\leq 0$}\\ &= 1 - h\left(\frac{\pimax}{\sigma_{\max}}\right) + \frac{\pimax}{\sigma_{\max}}r_t(\sigma_{\max})\tag{$\sigma_{\max} = \frac{\pimax}{\pi_{\prec\xmid}}$}\\ &=s\left(\frac{\pimax}{\sigma_{\max}}\right) + \frac{\pimax}{\sigma_{\max}}r_t(\sigma_{\max})\tag{by definition of $s$}\\ &\leq r_t(\pimax).\tag{by~\eqref{eq1} and Lemma~\ref{l413}} \end{align*} The analysis when $\pimax \in [t,1)$ is very similar. In this case $A_t$ asks whether $x = \xmax$, and continues with $\pi|_{\neq\xmax}$ if $x \neq \xmax$. Let $\sigma \ensuremath{:=}\xspace \pi|_{\leq \xmax}$. By Lemma~\ref{lem:redundancy-1-formula}, \begin{align*} R_t(\pi) &= 1 - h(\pimax) - \pimax + (1-\pimax) R_t(\sigma) \tag{since $\pimax \in (t,1)$} \\ &\leq 1 - h(\pimax) - \pimax + (1-\pimax) r_t(\sigma_{\max}) \tag{by induction hypothesis} \\ &\leq 1 - h(\pimax) - \pimax \tag{since $r_t(\sigma_{\max}) \leq 0$} \\ &= f(\pimax) \tag{by definition of $f$} \\ &= r_t(\pimax). \tag{by definition of $r_t$, since $\pimax \geq t$} \end{align*} \end{proof} The following lemma shows that $t=0.3$ satisfies $r_t(p)\leq 0$ for all $p\in (0,1]$, completing the proof of Theorem~\ref{thm:redundancy-1}. It uses some technical results, proved below in Lemma~\ref{la01}. \begin{lemma}\label{l417} Let $t=0.3$. Then $r_{t}(p)\leq 0$ for all $p\in(0,1]$. \end{lemma} \begin{proof} Let $p\in (0,1]$. We consider two cases: (i) $p\geq t$, and (ii) $p < t$. In each case we derive a constraint on $t$ that suffices for ensuring that $r_t(p) \leq 0$, and conclude the proof by showing that $t= 0.3$ satisfies both constraints. Consider the case $t\leq p$. Here $r_t(p) = f(p) = 1-h(p)-p$, and calculation shows that $f(p)$ is non-positive on $[0.23,1]$; therefore, $r_t(p)\leq 0$ for all $p\geq t$, as long as $t\geq 0.23$. Consider the case $p < t$. Here, we are not aware of an explicit formula for $r_t(p)$; instead, we derive the following upper bound, for all $p < t$: \begin{equation}\label{eq2} \frac{r_t(p)}{p} \leq \sum_{n=0}^\infty S\left(\frac{t}{2^n(1-t)+t}\right) + \max\left(F(t),F\Bigl(\frac{2t}{1-t}\Bigr)\right), \end{equation} where \[ S(x)=\frac{s\bigg(\displaystyle\frac{1+x}{2}\bigg)}{x}, ~~~~~ F(x)=\frac{f(x)}{x}. \] With~\eqref{eq2} in hand we are done: indeed, calculation shows that the right hand side of~\eqref{eq2} is non-positive in some neighborhood of $0.3$ (e.g.\ it is $-0.0312$ when $t=0.3$, it is $-0.0899$ when $t = 0.294$); thus, as these values of $t$ also satisfy the constraint from (i), this finishes the proof. It remains to prove~\eqref{eq2}. Let $p=p_0,p_1,p_2,\ldots,p_k$ be a sequence of states chosen by Alice. It suffices to show that $\frac{r(p_0,\ldots,p_k)}{p_0}\leq\sum_{n=0}^\infty S\left(\frac{t}{2^n(1-t)+t}\right) + \max\left(F(t),F\Bigl(\frac{2t}{1-t}\Bigr)\right).$ By Lemma~\ref{l415}: \begin{align*} r(p_0,\ldots,p_k) &= \sum_{i=0}^{k-1}\frac{p_0}{p_i} s\biggl(\frac{p_{i}}{p_{i+1}}\biggr) + \frac{p_{0}}{p_k}f(p_k)\\ &\leq \sum_{i=0}^{k-1}\frac{p_0}{p_i} s\biggl(\frac{1+p_{i}}{2}\biggr) + \frac{p_{0}}{p_k}f(p_k)\\ &=p_0\Bigl(\sum_{i=0}^{k-1}{S(p_i)} + F(p_k)\Bigr), \end{align*} where in the second line we used that $\frac{p_{i}}{p_{i+1}}\in[\frac{1-p_i}{2}, \frac{1+p_i}{2}]$, and the fact that $s(x)=1-h(x)$ is symmetric around $x=0.5$ and increases with $\lvert x-0.5 \rvert$. Therefore, \begin{align*} \frac{r(p_0,\ldots,p_k)}{p_0} &\leq \sum_{i=0}^{k-1}{S(p_i)} + F(p_k)\\ &\leq \sum_{i=0}^{k-1} S\left(\frac{t}{2^i(1-t)+t}\right) + F(p_k)\\ &\leq \sum_{i=0}^\infty S\left(\frac{t}{2^i(1-t)+t}\right) + \max\left(F(t),F\Bigl(\frac{2t}{1-t}\Bigr)\right), \end{align*} where in the second last inequality we used that $p_{k-1-i} < \frac{t}{2^i(1-t)+t}$ (Lemma~\ref{l414}) and that $S(x)$ is monotone (Lemma~\ref{la01} below), and in the last inequality we used that $p_k\in [t, \frac{2t}{1-t})$ and that $F(x)$ is convex (Lemma~\ref{la01} below). \end{proof} The following technical lemma completes the proof of Lemma~\ref{l417}. \begin{lemma} \label{la01} The function $S(x) = \frac{1-h(\frac{1+x}{2})}{x}$ is monotone, and the function $F(x) = \frac{1-h(x)-x}{x}$ is convex. \end{lemma} \begin{proof} The function $h(\frac{1-x}{2})$ is equal to its Maclaurin series for $x\in (-1,+1)$: \[ h\Bigl(\frac{1+x}{2}\Bigr) = 1 - \sum_{k=1}^{\infty}{\frac{\log_2{e}}{2k(2k-1)}\cdot x^{2k}}. \] Therefore, \[ S(x) = \sum_{k=1}^{\infty}{\frac{\log_2{e}}{2k(2k-1)}\cdot x^{2k-1}}, \] and \[ F(x) = \Bigl(\sum_{k=1}^{\infty}{\frac{\log_2{e}}{2k(2k-1)}}\cdot\frac{(1-2x)^{2k}}{x}\Bigr) - 1. \] Now, each of the functions $x^{2k-1}$ is monotone, and each of the functions $\frac{(1-2x)^{2k}}{x}$ is convex on $(0,\infty)$: its second derivative is \[ \frac{2(1-2x)^{2k-2}(1 + 4(k-1)x + 4(k-1)(2k-1)x^2)}{x^3} > 0. \] Therefore, $S(x)$ is monotone as a non-negative combination of monotone functions, and $F(x)$ is convex as a non-negative combination of convex functions. \end{proof} \section{Information theoretical benchmark --- Shannon's entropy} \label{sec:redundancy-r} In this section we study the minimum number of questions that achieve redundancy of at most $r$, for a fixed $r \geq 1$. Note that $r=1$ is the optimal redundancy: the distribution $\pi$ on $X_2$ given by $\pi_1 = 1-\epsilon,\allowbreak \pi_2=\epsilon$ has redundancy $1-\tilde{O}(\epsilon)$ (that is, $1-O(\epsilon \log(1/\epsilon))$ even without restricting the set of allowed questions. In the previous section we have shown that the optimal redundancy of $r=1$ can be achieved with just $2n$ comparison and equality queries (in fact, as we show below, there are only $2n-3$ of these queries). It is natural to ask how small the number of questions can be if we allow for a larger $r$. Note that at least $\log n$ questions are necessary to achieve any finite redundancy. Indeed, a smaller set of questions is not capable of specifying all elements even if all questions are being asked. The main result of this section is that the minimum number of questions that are sufficient for achieving redundancy $r$ is roughly $r\cdot n^{1/ \lfloor r \rfloor}$: \begin{main-theorem} \label{thm:redundancy-r-bounds} For every $r \geq 1$ and $n\in\mathbb{N}$, \[ \frac{1}{e}\lfloor r \rfloor n^{1/\lfloor r \rfloor} \leq \uent(n,r) \leq 2 \lfloor r \rfloor n^{1/\lfloor r \rfloor}. \] In particular, $\uent(n,r) = \Theta\bigl(\lfloor r\rfloor n^{1/\lfloor r\rfloor}\bigr)$. \end{main-theorem} \subsection{Upper bound} \label{sec:redundancy-r-ub} The upper bound in Theorem~\ref{thm:redundancy-r-bounds} is based on the following corollary of Theorem~\ref{thm:redundancy-1}: \begin{theorem}\label{thm:redundancy-r} Let $Y$ be a linearly ordered set, and let $Z=Y^k$ (we don't think of $Z$ as ordered). For any distribution $\pi$ on $Z$ there is an algorithm that uses only questions of the form (i) ``$\vec x_i \prec y$?'' and (ii) ``$\vec x_i = y$?'', where $i\in[k]$ and $y\in Y$, whose cost is at most $H(\pi) + k$. \end{theorem} \begin{proof} Let $Z_1Z_2\ldots Z_k \sim \pi$. Consider the algorithm which determines $Z_1,\ldots,Z_k$ in order, where $Z_i$ is determined by applying the algorithm from Theorem~\ref{thm:redundancy-1} on ``$Z_i \vert Z_1\ldots Z_{i-1}$", which is the conditional distribution of $Z_i$ given the known values of $Z_1,\ldots,Z_{i-1}$. The expected number of queries is at most \[ \bigl(H(Z_1) + 1\bigr) + \bigl(H(Z_2 \vert Z_1) + 1\bigr) + \cdots + \bigl(H(Z_k\vert Z_1\ldots Z_{k-1}) + 1\bigr) = H(Z_1\ldots Z_k) + k, \] using the chain rule. \end{proof} We use this theorem to construct a set of questions of size at most $2 \lfloor r \rfloor n^{1/\lfloor r \rfloor}$ that achieves redundancy $r$ for any distribution over $X_n$. Note that $n \leq \Bigl(\bigl\lceil n^{1/\lfloor r\rfloor}\bigr\rceil\Bigr)^{\lfloor r\rfloor}$. Therefore every element $x\in X_n$ can be represented by a vector $\vec x \in \allowbreak \bigl\{1,\dots,\lceil n^{1/\lfloor r\rfloor}\rceil \bigr\} ^ {\lfloor r\rfloor}$. Let $\cQ$ be the set containing all questions of the form (i) ``$\vec x_i = y$?'' and (ii) ``$\vec x_i \prec y$?''. By Theorem~\ref{thm:redundancy-r}, $r(\cQ)=\lfloor r\rfloor$. The following questions from $\cQ$ are redundant, and can be removed from $\cQ$ without increasing its redundancy: (i) ``$\vec x_i \prec 1$?'' (corresponds to the empty set and therefore provides no information), (ii) ``$\vec x_i \prec 2$?'' (equivalent to the question ``$\vec x_i = 1$?''), and (iii) ``$\vec x_i \prec \lceil n^{1/\lfloor r\rfloor}\rceil$?'' (equivalent to the question ``$\vec x_i = \lceil n^{1/\lfloor r\rfloor}\rceil$?''). The number of remaining questions is \[ \lfloor r \rfloor\cdot\Bigl(2\bigl\lceil n^{1/\lfloor r\rfloor}\bigr\rceil - 3\Bigr) \leq 2\lfloor r \rfloor\cdot\Bigl( n^{1/\lfloor r\rfloor} \Bigr). \] This proves the upper bound in Theorem~\ref{thm:redundancy-r-bounds}. \subsection{Lower bound} \label{sec:redundancy-r-lb} The crux of the proof of the lower bound in Theorem~\ref{thm:redundancy-r-bounds} is that if $\cQ$ is a set of questions whose redundancy is at most $r$ then every $x\in X_n$ can be identified by at most $\lfloor r\rfloor$ questions from $\cQ$. We say that the questions $q_1,\ldots,q_T$ \emph{identify} $x$ if for every $y\neq x$ there is some $i\leq T$ such that $q_i(x)\neq q_i(y)$. Define $t(n,r)$ to be the minimum cardinality of a set $\cQ$ of questions such that every $x\in X$ has at most $r$ questions in $\cQ$ that identify it. The quantity $t(n,r)$ can be thought of as a non-deterministic version of $u(n,r)$: it is the minimal size of a set of questions so that every element can be ``verified'' using at most $r$ questions. The lower bound on $u(n,r)$ follows from Lemma~\ref{l52} and Lemma~\ref{l53} below. \begin{lemma}\label{l52} For all $n,r$, $u(n,r)\geq t(n,\lfloor r \rfloor)$. \end{lemma} \begin{proof} It suffices to show that for every set of questions $\cQ$ with redundancy at most $r$, every $x\in X$ has at most $\lfloor r\rfloor$ questions in $\cQ$ that identify it. Consider the distribution $\pi$ given by $\pi(x) = 1-\epsilon$ and $\pi(y) = \epsilon/(n-1)$ for $y \neq x$. Thus $H(\pi) = \tilde{O}(\epsilon)$. Consider an algorithm for $\pi$ with redundancy $r$ that uses only questions from $\cQ$. Let $T$ be the number of questions it uses to find $x$. The cost of the algorithm is at least $(1-\epsilon)T$, and so $(1-\epsilon)T \leq H(\pi) + r = \tilde{O}(\epsilon) + r$, implying $T \leq \tilde{O}(\epsilon) + (1+\frac{\epsilon}{1-\epsilon})r$. For small enough $\epsilon > 0$, the right-hand side is smaller than $\lfloor r \rfloor + 1$, and so $T \leq \lfloor r \rfloor$. \end{proof} Lemma~\ref{l52} says that in order to lower bound $u(n,r)$, it suffices to lower bound $t(n,\lfloor r\rfloor)$, which is easier to handle. For example, the following straightforward argument shows that $t(n,R)\geq \frac{1}{2e}R n^{ 1/R}$, for every $R,n\in\mathbb{N}$. Assume $\cQ$ is a set of questions of size $u(n,R)$ so that every $x$ is identified by at most $R$ questions. This implies an encoding (i.e.\ a one-to-one mapping) of $x\in X_n$ by the $R$ questions identifying it, and by the bits indicating whether $x$ satisfies each of these questions. Therefore \begin{align*} n \leq \binom{|\cQ|}{\leq R }2^{R} \leq \Bigl(\frac{2e|\cQ|}{R}\Bigr)^{R}, \end{align*} where in the last inequality we used that $\binom{m}{\leq k}\leq \bigl(\frac{em}{k}\bigr)^k$ for all $m,k$. This implies that $t(n,\lfloor r\rfloor) \geq \frac{1}{2e}\lfloor r\rfloor n^{ 1/\lfloor r \rfloor}$. The constant $\frac{1}{2e}$ in front of $\lfloor r\rfloor n^{ 1/\lfloor r \rfloor}$ can be increased to $\frac{1}{e}$, using an algebraic argument: \begin{lemma}\label{l53} For all $n,R\in\mathbb{N}$: \[ t(n,R)\geq \frac{1}{e}R\cdot n^{1/R}. \] \end{lemma} \begin{proof} We use the so-called polynomial method. Let $\cQ$ be a set of questions such that each $x\in X$ can be identified by at most $R$ queries. For each $x\in X$, let $u_x$ be the $|\cQ|$-dimensional vector $u_x=\bigl(q_1(x),\ldots,q_{\lvert \cQ\rvert}(x)\bigr)$, and let $U=\{u_x : x\in X\}\subseteq\{0,1\}^{\lvert \cQ\rvert}$. We will show that every function $F\colon U \to \mathbb{F}_2$ can be represented as a multilinear polynomial of degree at most $R$ in $|\cQ|$ variables. Since the dimension over $\mathbb{F}_2$ of all such functions is $n$, whereas the dimension of the space of all multilinear polynomials of degree at most $R$ is $\binom{|\cQ|}{\leq R}$, the bound follows: \[ n \leq \binom{|\cQ|}{\leq R} \leq \Bigl(\frac{e|\cQ|}{R}\Bigr)^R \Longrightarrow n \geq \frac{1}{e} R \cdot n^{1/R}. \] It is enough to show that for any $u_x\in U$, the corresponding ``delta function'' $\delta_x\colon U\to\mathbb{F}_2$, defined as $\delta_x(u_{x})=1$ and $\delta_x(v) =0$ for $u_{x}\neq v\in U$, can be represented as a polynomial of degree at most $d$. Suppose that $q_{i_1},\ldots, q_{i_T}$ are $T\leq R$ questions that identify $x$. Consider the polynomial \[ P(y_1,\ldots,y_{\lvert \cQ\rvert}) = (y_{i_1} - q_{i_1}(x) + 1) \cdots (y_{i_r} - q_{i_r}(x) + 1). \] Clearly $P(u_x) = 1$. On the other hand, if $P(u_y) = 1$ then $q_{i_j}(y) = q_{i_j}(x)$ for all $j$, showing that $y = x$. So $P = \delta_x$, completing the proof. \end{proof} Our proof of the lower bound in Theorem~\ref{thm:redundancy-r-bounds} is based on $t(n,R)$, which is the minimum cardinality of a set of queries such that each element can be identified by at most $R$ questions. This quantity is closely related to witness codes~\cite{Meshulam,CRZ08}; see~\cite{DaganThesis} for more details. \section{Combinatorial benchmark --- Huffman codes} \label{sec:huffman} Section~\ref{sec:comparison-equality} shows that the optimal redundancy, namely $1$, can be achieved using only $O(n)$ questions. However, it is natural to ask for an \emph{instance}-optimal algorithm? That is, we are looking for a set of questions which matches the performance of minimum redundancy codes such as Huffman codes. Let us repeat the definition of an optimal set of questions that is central in this section. \begin{main-definition} \label{dfn:optimal-set-of-questions} A set \ensuremath{{\cal Q}}\xspace of subsets of $X_n$ is an \emph{optimal set of questions over \ensuremath{X_n}\xspace} if for all distributions $\mu$ on \ensuremath{X_n}\xspace, \[ \costt{\ensuremath{{\cal Q}}\xspace}{\mu} = \opt{\mu}. \] \end{main-definition} Using the above definition, $\uhuf(n,0)$ is equal to the minimal size of an optimal set of questions over $X_n$. Perhaps surprisingly, the trivial upper bound of $2^{n-1}$ on $\uhuf(n,0)$ can be exponentially improved: \begin{main-theorem} \label{thm:minimum-redundancy-ub} We have \[ \uhuf(n,0) \leq 1.25^{n+o(n)}. \] \end{main-theorem} We prove a similar lower bound, which is almost tight for infinitely many $n$: \begin{main-theorem} \label{thm:minimum-redundancy-lb} For $n$ of the form $n=5 \cdot 2^m$, \[ \uhuf(n,0) \geq 1.25^{n}/O(\sqrt{n}). \] For all $n$, \[ \uhuf(n,0) \geq 1.232^n/O(\sqrt{n}). \] \end{main-theorem} \begin{main-corollary} \label{cor:minimum-redundancy} We have \[ \limsup_{n\to\infty} \frac{\log \uhuf(n,0)}{n} = \log 1.25. \] \end{main-corollary} Unfortunately, the construction in Theorem~\ref{thm:minimum-redundancy-ub} is not explicit. A different construction, which uses $O(\sqrt{2}^n)$ questions, is not only explicit, but can also be implemented efficiently: \begin{main-theorem} \label{thm:cone} Consider the set of questions \[ \ensuremath{{\cal Q}}\xspace = \{ A \subseteq X_n : A \subseteq X_{\lceil n/2 \rceil} \text{ or } A \supseteq X_{\lceil n/2 \rceil} \}. \] The set $\ensuremath{{\cal Q}}\xspace$ consists of $2^{\lceil n/2 \rceil} + 2^{\lfloor n/2 \rfloor}$ questions and satisfies the following properties: \begin{enumerate} \item There is an indexing scheme $\ensuremath{{\cal Q}}\xspace = \{ Q_q : q \in \{0,1\}^{\lceil n/2 \rceil + 1} \}$ such that given an index $q$ and an element $x_i \in X_n$, we can decide whether $x_i \in Q_q$ in time $O(n)$. \item Given a distribution $\pi$, we can construct an optimal decision tree for $\pi$ using \ensuremath{{\cal Q}}\xspace in time $O(n^2)$. \item Given a distribution $\pi$, we can implement an optimal decision tree for $\pi$ in an online fashion in time $O(n)$ per question, after $O(n\log n)$ preprocessing. \end{enumerate} \end{main-theorem} \paragraph{Section organization.} Section~\ref{sec:dyadic} shows that a set of questions is optimal if and only if it is a \emph{dyadic hitter}, that is, contains a question splitting every non-constant dyadic distribution into two equal halves. Section~\ref{sec:antichains} discusses a relation to hitting sets for maximal antichains, and proves Theorem~\ref{thm:cone}. Section~\ref{sec:un0-density} shows that the optimal size of a dyadic hitter is controlled by the minimum value of another parameter, the \emph{maximum relative density}. We upper bound the minimum value in Section~\ref{sec:un0-ub}, thus proving Theorem~\ref{thm:minimum-redundancy-lb}, and lower bound it in Section~\ref{sec:un0-lb}, thus proving Theorem~\ref{thm:minimum-redundancy-ub}. \subsection{Reduction to dyadic hitters} \label{sec:dyadic} The purpose of this subsection is to give a convenient combinatorial characterization of optimal sets of questions. Before presenting this characterization, we show that in this context it suffices to look at dyadic distributions. \begin{lemma} \label{lem:optimal-dyadic} A set \ensuremath{{\cal Q}}\xspace of questions over \ensuremath{X_n}\xspace is optimal if and only if $\costt{\ensuremath{{\cal Q}}\xspace}{\mu} = \opt{\mu}$ for all \emph{dyadic} distributions $\mu$. \end{lemma} \begin{proof} Suppose that \ensuremath{{\cal Q}}\xspace is optimal for all dyadic distributions, and let $\pi$ be an arbitrary distribution over \ensuremath{X_n}\xspace. Let $\mu$ be a dyadic distribution such that \[ \opt{\pi} = \sum_{i=1}^n \pi_i \log \frac{1}{\mu_i}. \] By assumption, \ensuremath{{\cal Q}}\xspace is optimal for $\mu$. Let $T$ be an optimal decision tree for $\mu$ using questions from \ensuremath{{\cal Q}}\xspace only, and let $\tau$ be the corresponding dyadic distribution, given by $\tau_i = 2^{-T(x_i)}$ (recall that $T(x_i)$ is the depth of $x_i$). Since $\tau$ minimizes $T(\mu) = H(\mu) + D(\mu\|\tau)$ over dyadic distributions, necessarily $\tau = \mu$. Thus \[ T(\pi) = \sum_{i=1}^n \pi_i \log \frac{1}{\tau_i} = \sum_{i=1}^n \pi_i \log \frac{1}{\mu_i} = \opt{\pi}, \] showing that \ensuremath{{\cal Q}}\xspace is optimal for \ensuremath{\mu}\xspace. \end{proof} Given a dyadic distribution $\mu$ on \ensuremath{X_n}\xspace, we will be particularly interested in the collection of subsets of \ensuremath{X_n}\xspace that have probability exactly half under $\mu$. \begin{dfn}[Dyadic hitters]\label{dfn:dyadSetsandHitters} Let \ensuremath{\mu}\xspace be a non-constant dyadic distribution. A set $A \subseteq X_n$ \emph{splits} \ensuremath{\mu}\xspace if $\ensuremath{\mu}\xspace(A) = 1/2$. We denote the collection of all sets splitting \ensuremath{\mu}\xspace by \dyadof{\ensuremath{\mu}\xspace}. We call a set of the form \dyadof{\ensuremath{\mu}\xspace} a \emph{dyadic set}. We call a set of questions \ensuremath{{\cal Q}}\xspace a \emph{dyadic hitter} in $X_n$ if it intersects \dyadof{\ensuremath{\mu}\xspace} for all non-constant dyadic distributions \ensuremath{\mu}\xspace. (Lemma~\ref{lem:neat-sum} implies that \dyadof{\ensuremath{\mu}\xspace} is always non-empty.) \end{dfn} A dyadic hitter is precisely the object we are interested in: \begin{lemma}\label{lem:hitterIsOptimal} A set \ensuremath{{\cal Q}}\xspace of subsets of \ensuremath{X_n}\xspace is an optimal set of questions if and only if it is a dyadic hitter\xspace in \ensuremath{X_n}\xspace. \end{lemma} \begin{proof} Let \ensuremath{{\cal Q}}\xspace be a dyadic hitter\xspace in \ensuremath{X_n}\xspace. We prove by induction on $1\leq m \leq n$ that for a dyadic distribution \ensuremath{\mu}\xspace on \ensuremath{X_n}\xspace with support size $m$, $\costt{\ensuremath{{\cal Q}}\xspace}{\ensuremath{\mu}\xspace}=H(\ensuremath{\mu}\xspace)$. Since $\opt{\ensuremath{\mu}\xspace} = H(\ensuremath{\mu}\xspace)$, Lemma~\ref{lem:optimal-dyadic} implies that \ensuremath{{\cal Q}}\xspace is an optimal set of questions. The base case, $m = 1$, is trivial. Suppose therefore that \ensuremath{\mu}\xspace is a dyadic distribution whose support has size $m>1$. In particular, \ensuremath{\mu}\xspace is not constant, and so \ensuremath{{\cal Q}}\xspace contains some set $S \in \dyadof{\ensuremath{\mu}\xspace}$. Let $\alpha = \ensuremath{\mu}\xspace|_S$ and $\beta = \ensuremath{\mu}\xspace|_{\overline{S}}$, and note that $\alpha,\beta$ are both dyadic. The induction hypothesis shows that $\costt{\ensuremath{{\cal Q}}\xspace}{\alpha} = H(\alpha)$ and $\costt{\ensuremath{{\cal Q}}\xspace}{\beta} = H(\beta)$. A decision tree which first queries $S$ and then uses the implied algorithms for $\alpha$ and $\beta$ has cost \[ 1 + \frac{1}{2} H(\alpha) + \frac{1}{2} H(\beta) = h(\ensuremath{\mu}\xspace(S)) + \ensuremath{\mu}\xspace(S) H(\ensuremath{\mu}\xspace|_S) + \ensuremath{\mu}\xspace(\overline{S}) H(\ensuremath{\mu}\xspace|_{\overline{S}}) = H(\ensuremath{\mu}\xspace), \] using the Bernoulli chain rule; here $\ensuremath{\mu}\xspace|_S$ is the restriction of $\ensuremath{\mu}\xspace$ to the elements in $S$. Conversely, suppose that \ensuremath{{\cal Q}}\xspace is not a dyadic hitter, and let \ensuremath{\mu}\xspace be a non-constant dyadic distribution such that \dyadof{\ensuremath{\mu}\xspace} is disjoint from \ensuremath{{\cal Q}}\xspace. Let $T$ be any decision tree for \ensuremath{\mu}\xspace using \ensuremath{{\cal Q}}\xspace, and let $S$ be its first question. The cost of $T$ is at least \[ 1 + \ensuremath{\mu}\xspace(S) H(\ensuremath{\mu}\xspace|_S) + \ensuremath{\mu}\xspace(\overline{S}) H(\ensuremath{\mu}\xspace|_{\overline{S}}) > h(\ensuremath{\mu}\xspace(S)) + \ensuremath{\mu}\xspace(S) H(\ensuremath{\mu}\xspace|_S) + \ensuremath{\mu}\xspace(\overline{S}) H(\ensuremath{\mu}\xspace|_{\overline{S}}) = H(\ensuremath{\mu}\xspace), \] since $\ensuremath{\mu}\xspace(S) \neq \frac{1}{2}$. Thus $\costt{\ensuremath{{\cal Q}}\xspace}{\ensuremath{\mu}\xspace} > \opt{\ensuremath{\mu}\xspace}$, and so \ensuremath{{\cal Q}}\xspace is not an optimal set of questions. \end{proof} \subsection{Dyadic sets as antichains} \label{sec:antichains} There is a surprising connection between dyadic hitters and hitting sets for maximal antichains. We start by defining the latter: \begin{definition} A \emph{fibre} in $X_n$ is a subset of $2^{X_n}$ which intersects every maximal antichain in $X_n$. \end{definition} Fibres were defined by Lonc and Rival~\cite{LoncRival}, who also gave a simple construction, via cones: \begin{definition} The \emph{cone} $\cone{S}$ of a set $S$ consists of all subsets and all supersets of $S$. \end{definition} Any cone $\cone{S}$ intersects any maximal antichain $A$, since otherwise $A \cup \{S\}$ is also an antichain. By choosing $S$ of size $\lfloor n/2 \rfloor$, we obtain a fibre of size $2^{\lfloor n/2 \rfloor} + 2^{\lceil n/2 \rceil} - 1 = \Theta(2^{n/2})$. Our goal now is to show that every fibre is a dyadic hitter: \begin{theorem} \label{thm:fibre} every fibre is a dyadic hitter. \end{theorem} This shows that every cone is a dyadic hitter, and allows us to give a simple algorithm for constructing an optimal decision tree using any cone. We start with a simple technical lemma which will also be used in Section~\ref{sec:un0-ub}: \begin{definition} \label{def:tail} Let $\mu$ be a dyadic distribution over $X_n$. The \emph{tail} of $\mu$ is the largest set of elements $T \subseteq X_n$ such that for some $a \geq 1$, \begin{enumerate}[label=(\roman*)] \item The elements in $T$ have probabilities $2^{-a-1},2^{-a-2},\ldots,2^{-a-(|T|-1)},2^{-a-(|T|-1)}$. \item Every element not in $T$ has probability at least $2^{-a}$. \new{Shay: It could be good if we demonstrate the name ``tail'' by considering the tree representation of the dyadic distribution (I can do it sometimes later).} \end{enumerate} \end{definition} \begin{lemma} \label{lem:small-elements} Suppose that $\mu$ is a non-constant dyadic distribution with non-empty tail $T$. Every set in $\dyadof{\mu}$ either contains $T$ or is disjoint from $T$. \end{lemma} \begin{proof} The proof is by induction on $|T|$. If $|T|=2$ then there exist an integer $a \geq 1$ and two elements, without loss of generality $x_1,x_2$, of probability $2^{-a-1}$, such that all other elements have probability at least $2^{-a}$. Suppose that $S \in \dyadof{\mu}$ contains exactly one of $x_1,x_2$. Then \[ 2^{a-1} = \sum_{x_i \in S} 2^a \mu(x_i) = \sum_{x_i \in S \setminus \{x_1,x_2\}} 2^a \mu(x_i) + \frac{1}{2}. \] However, the left-hand side is an integer while the right-hand side is not. We conclude that $S$ must contain either both of $x_1,x_2$ or none of them. For the induction step, let the elements in the tail $T$ of $\mu$ have probabilities $2^{-a-1},2^{-a-2},\allowbreak\ldots,\allowbreak2^{-a-(|T|-1)},2^{-a-(|T|-1)}$. Without loss of generality, suppose that $x_{n-1},x_n$ are the elements whose probability is $2^{-a-(|T|-1)}$. The same argument as before shows that every dyadic set in $\dyadof{\mu}$ must contain either both of $x_{n-1},x_n$ or neither. Form a new dyadic distribution $\nu$ on $X_{n-1}$ by merging the elements $x_{n-1},x_n$ into $x_{n-1}$, and note that $\dyadof{\mu}$ can be obtained from $\dyadof{\nu}$ by replacing $x_{n-1}$ with $x_{n-1},x_n$. The distribution $\nu$ has tail $T' = T \setminus \{x_n\}$, and so by induction, every set in $\dyadof{\nu}$ either contains $T'$ or is disjoint from $T'$. This implies that every set in $\dyadof{\mu}$ either contains $T$ or is disjoint from $T$. \end{proof} The first step in proving Theorem~\ref{thm:fibre} is a reduction to dyadic distributions having full support: \begin{lemma} \label{lem:dyadic-full-support} A set of questions is a dyadic hitter in $X_n$ if and only if it intersects $\dyadof{\mu}$ for all non-constant full-support dyadic distributions $\mu$ on $X_n$. \end{lemma} \begin{proof} A dyadic hitter clearly intersects $\dyadof{\mu}$ for all non-constant full-support dyadic distributions on $X_n$. For the other direction, suppose that \ensuremath{{\cal Q}}\xspace is a set of questions that intersects $\dyadof{\mu}$ for every non-constant full-support dyadic distribution $\mu$. Let $\nu$ be a non-constant dyadic distribution on $X_n$ which doesn't have full support. Let $x_{\min}$ be an element in the support of $\nu$ with minimal probability, which we denote $\nu_{\min}$. Arrange the elements in $\overline{\supp\nu}$ in some arbitrary order $x_{i_1},\ldots,x_{i_m}$. Consider the distribution $\mu$ given by: \begin{itemize} \item $\mu(x_i) = \nu(x_i)$ if $x_i \in \supp \mu$ and $x_i \neq x_{\min}$. \item $\mu(x_{\min}) = \nu_{\min}/2$. \item $\mu(x_{i_j}) = \nu_{\min}/2^{j+1}$ for $j < m$. \item $\mu(x_{i_m}) = \nu_{\min}/2^m$. \end{itemize} In short, we have replaced $\nu(x_{\min}) = \nu_{\min}$ with a tail $x_{\min},x_{i_1},\ldots,x_{i_m}$ of the same total probability. It is not hard to check that $\mu$ is a non-constant dyadic distribution having full support on $X_n$. We complete the proof by showing that \ensuremath{{\cal Q}}\xspace intersects $\dyadof{\nu}$. By assumption, \ensuremath{{\cal Q}}\xspace intersects $\dyadof{\mu}$, say at a set $S$. Lemma~\ref{lem:small-elements} shows that $S$ either contains all of $\{x_{\min}\} \cup \overline{\supp\nu}$, or none of these elements. In both cases, $\nu(S) = \mu(S) = 1/2$, and so \ensuremath{{\cal Q}}\xspace intersects $\dyadof{\nu}$. \end{proof} We complete the proof of Theorem~\ref{thm:fibre} by showing that dyadic sets corresponding to full-support distributions are maximal antichains: \begin{lemma} \label{lem:antichain} Let $\mu$ be a non-constant dyadic distribution over $X_n$ with full support, and let $\ensuremath{D}\xspace=\dyadof{\mu}$. Then $\ensuremath{D}\xspace$ is a maximal antichain which is closed under complementation (i.e.\ $A\in \ensuremath{D}\xspace\implies X\setminus A\in \ensuremath{D}\xspace$). \end{lemma} \begin{proof} (i) That $\ensuremath{D}\xspace$ is closed under complementation follows since if $A\in \ensuremath{D}\xspace$ then $\mu(X\setminus A) = 1 - \mu(A) = 1/2$. (ii) That $\ensuremath{D}\xspace$ is an antichain follows since if $A$ strictly contains $B$ then $\mu(A) > \mu(B)$ (because $\mu$ has full support). (iii) It remains to show that $\ensuremath{D}\xspace$ is maximal. By~(i) it suffices to show that every $B$ with $\mu(B)>1/2$ contains some $A\in \ensuremath{D}\xspace$. This follows from applying Lemma~\ref{lem:neat-sum} on the probabilities of the elements in~$B$. \end{proof} Cones allow us to prove Theorem~\ref{thm:cone}: \begin{proof}[Proof of Theorem~\ref{thm:cone}] Let $S = \{x_1,\ldots,x_{\lfloor n/2 \rfloor}\}$. The set of questions \ensuremath{{\cal Q}}\xspace is the cone $\cone{S}$, whose size is $2^{\lfloor n/2 \rfloor} + 2^{\lceil n/2 \rceil} - 1 < 2^{\lceil n/2 \rceil + 1}$. An efficient indexing scheme for \ensuremath{{\cal Q}}\xspace divides the index into a bit $b$, signifying whether the set is a subset of $S$ or a superset of $S$, and $\lfloor n/2 \rfloor$ bits (in the first case) or $\lceil n/2 \rceil$ bits (in the second case) for specifying the subset or superset. \smallskip To prove the other two parts, we first solve an easier question. Suppose that $\mu$ is a non-constant dyadic distribution whose sorted order is known. We show how to find a set in $\dyadof{\mu} \cap \ensuremath{{\cal Q}}\xspace$ in time $O(n)$. If $\mu(S) = 1/2$ then $S \in \dyadof{\mu}$. If $\mu(S) > 1/2$, go over the elements in $S$ in non-decreasing order. According to Lemma~\ref{lem:neat-sum}, some prefix will sum to $1/2$ exactly. If $\mu(S) < 1/2$, we do the same with $\overline{S}$, and then complement the result. Suppose now that $\pi$ is a non-constant distribution. We can find a Huffman distribution $\mu$ for $\pi$ and compute the sorted order of $\pi$ in time $O(n\log n)$. The second and third part now follow as in the proof of Lemma~\ref{lem:hitterIsOptimal}. \end{proof} \subsection{Reduction to maximum relative density} \label{sec:un0-density} Our lower bound on the size of a dyadic hitter\xspace, proved in the following subsection, will be along the following lines. For appropriate values of $n$, we describe a dyadic distribution $\mu$, all of whose splitters have a certain size $i$ or $n-i$. Moreover, only a $\rho$ fraction of sets of size $i$ split $\mu$. We then consider all possible permutations of $\mu$. Each set of size $i$ splits a $\rho$ fraction of these, and so any dyadic hitter must contain at least $1/\rho$ sets. This lower bound argument prompts the definition of \emph{maximum relative density} (MRD), which corresponds to the parameter $\rho$ above; in the general case we will also need to optimize over~$i$. We think of the MRD as a property of dyadic sets rather than dyadic distributions; indeed, the concept of MRD makes sense for any collection of subsets of $X_n$. If a dyadic set has MRD $\rho$ then any dyadic hitter must contain at least $1/\rho$ questions, due to the argument outlined above. Conversely, using the probabilistic method we will show that roughly $1/\smallestdens{n}$ questions suffice, where $\smallestdens{n}$ is the minimum MRD of a dyadic set on \ensuremath{X_n}\xspace. \begin{definition}[Maximum relative density]\label{dfn:relativedensity} Let \ensuremath{D}\xspace be a collection of subsets of \ensuremath{X_n}\xspace. For $0\leq i\leq n$, let \[\reldensof{\ensuremath{D}\xspace}{i}\ensuremath{:=}\xspace \frac{\Bigl\lvert\bigl\{S\in \ensuremath{D}\xspace : \lvert S\rvert = i\bigr\}\Big\rvert}{\binom{n}{i}}.\] We define the \emph{maximum relative density} (MRD) of \ensuremath{D}\xspace, denoted \maxdensof{\ensuremath{D}\xspace}, as \[\maxdensof{\ensuremath{D}\xspace}\ensuremath{:=}\xspace \max_{i \in \{1,\ldots,n-1\}} \reldensof{\ensuremath{D}\xspace}{i}.\] We define \smallestdens{n} to be the minimum of \maxdensof{\ensuremath{D}\xspace} over all dyadic sets. That is, \smallestdens{n} is the smallest possible maximum relative density of a set of the form \dyadof{\ensuremath{\mu}\xspace}. \end{definition} The following theorem shows that $\uhuf(n,0)$ is controlled by \smallestdens{n}, up to polynomial factors. \begin{thm}\label{thm:hittersizeisdensity} Fix an integer $n$, and denote $M\ensuremath{:=}\xspace \frac{1}{\smallestdens{n}}$. Then \[ M\leq \uhuf(n,0)\leq n^2 \log n \cdot M. \] \end{thm} \begin{proof} Note first that according to Lemma \ref{lem:hitterIsOptimal}, $\uhuf(n,0)$ is equal to the minimal size of a dyadic hitter in \ensuremath{X_n}\xspace, and thus it suffices to lower- and upper-bound this size. Let \ensuremath{\sigma}\xspace be a uniformly random permutation on \ensuremath{X_n}\xspace. If $S$ is any set of size~$i$ then $\ensuremath{\sigma}\xspace^{-1}(S)$ is a uniformly random set of size~$i$, and so \[ \reldensof{\ensuremath{D}\xspace}{i} = \Pr_{\ensuremath{\sigma}\xspace\in\sym}[\ensuremath{\sigma}\xspace^{-1}(S) \in \ensuremath{D}\xspace] = \Pr_{\ensuremath{\sigma}\xspace\in\sym}[S \in \ensuremath{\sigma}\xspace(\ensuremath{D}\xspace)]. \] (Here $\sym$ is the group of permutations of $X_n$.) Fix a dyadic set \ensuremath{D}\xspace on \ensuremath{X_n}\xspace with $\maxdensof{\ensuremath{D}\xspace}=\smallestdens{n}$. The formula for $\reldensof{\ensuremath{D}\xspace}{i}$ implies that for any subset $S$ of \ensuremath{X_n}\xspace (of \emph{any} size), \[\Pr_{\ensuremath{\sigma}\xspace\in\sym} [S\in \ensuremath{\sigma}\xspace(\ensuremath{D}\xspace)]\leq \smallestdens{n}.\] Let \ensuremath{{\cal Q}}\xspace be a collection of subsets of \ensuremath{X_n}\xspace with $|\ensuremath{{\cal Q}}\xspace|<M$. A union bound shows that \[ \Pr_{\ensuremath{\sigma}\xspace\in\sym} [\ensuremath{{\cal Q}}\xspace \cap \ensuremath{\sigma}\xspace(\ensuremath{D}\xspace) \neq \emptyset] \leq |\ensuremath{{\cal Q}}\xspace| \smallestdens{n} < 1.\] Thus, there exists a permutation $\ensuremath{\sigma}\xspace$ such that $\ensuremath{{\cal Q}}\xspace\cap \ensuremath{\sigma}\xspace(\ensuremath{D}\xspace) =\emptyset$. Since $\ensuremath{\sigma}\xspace(\ensuremath{D}\xspace)$ is also a dyadic set, this shows that \ensuremath{{\cal Q}}\xspace is not a dyadic hitter. We deduce that any dyadic hitter must contain at least $M$ questions. \smallskip For the upper bound on $\uhuf(n,0)$, construct a set of subsets \ensuremath{{\cal Q}}\xspace containing, for each $i\in \{1,\ldots,n-1\}$, $M n \log n$ uniformly chosen sets $S\subseteq \ensuremath{X_n}\xspace$ of size $i$. We show that with positive probability, \ensuremath{{\cal Q}}\xspace is a dyadic hitter\xspace. Fix any dyadic set \ensuremath{D}\xspace, and let $i\in\{1,\ldots,n-1\}$ be such that $\reldensof{\ensuremath{D}\xspace}{i}=\maxdensof{\ensuremath{D}\xspace}\geq \smallestdens{n}$. The probability that a random set of size $i$ doesn't belong to \ensuremath{D}\xspace is at most $1-\maxdensof{\ensuremath{D}\xspace}\leq 1-\smallestdens{n}$. Therefore the probability that \ensuremath{{\cal Q}}\xspace is disjoint from \ensuremath{D}\xspace is at most \[(1-\smallestdens{n})^{M n \log n} \leq e^{-\smallestdens{n} M n \log n} = e^{-n\log n}<n^{-n}. \] As we show below in Claim~\ref{clm:numofdyadics}, there are at most $n^n$ non-constant dyadic distributions, and so a union bound implies that with positive probability, \ensuremath{{\cal Q}}\xspace is indeed a dyadic hitter\xspace. \end{proof} In order to complete the proof of Theorem~\ref{thm:hittersizeisdensity}, we bound the number of non-constant dyadic distributions: \begin{claim}\label{clm:numofdyadics} There are at most $n^n$ non-constant dyadic distributions on $X_n$. \end{claim} \begin{proof} Recall that dyadic distributions correspond to decision trees in which an element of probability $2^{-\ell}$ is a leaf at depth $\ell$. Clearly the maximal depth of a leaf is $n-1$, and so the probability of each element in a non-constant dyadic distribution is one of the $n$ values $0,2^{-1},\ldots,2^{-(n-1)}$. The claim immediately follows. \end{proof} Krenn and Wagner~\cite{KrennWagner} showed that the number of full-support dyadic distributions on $X_n$ is asymptotic to $\alpha \gamma^{n-1} n!$, where $\alpha \approx 0.296$ and $\gamma \approx 1.193$, implying that the number of dyadic distributions on $X_n$ is asymptotic to $\alpha e^{1/\gamma} \gamma^{n-1} n!$. Boyd~\cite{Boyd} showed that the number of monotone full-support dyadic distributions on $X_n$ is asymptotic to $\beta \lambda^n$, where $\beta \approx 0.142$ and $\lambda \approx 1.794$, implying that the number of monotone dyadic distributions on $X_n$ is asymptotic to $\beta (1+\lambda)^n$. \medskip The proof of Theorem~\ref{thm:hittersizeisdensity} made use of two properties of dyadic sets: \begin{enumerate} \item Any permutation of a dyadic set is a dyadic set. \item There are $e^{n^{O(1)}}$ dyadic sets. \end{enumerate} If $\mathcal{F}$ is any collection of subsets of $2^{X_n}$ satisfying the first property then the proof of Theorem~\ref{thm:hittersizeisdensity} generalizes to show that the minimal size $U$ of a hitting set for $\mathcal{F}$ satisfies \[ M \leq U \leq M n \log |\mathcal{F}|, \qquad \text{where } M = \frac{1}{\min_{D \in \mathcal{F}} \rho(D)}. \] \subsection{Upper bounding \smallestdenstitle{n}} \label{sec:un0-ub} Theorem~\ref{thm:minimum-redundancy-lb} will ultimately follow from the following lemma, by way of Theorem~\ref{thm:hittersizeisdensity}: \begin{lemma}\label{lem:densityUB} Fix $0 < \epsparam \leq 1/2$. There exists an infinite sequence of positive integers $n$ (namely, those of the form $\lfloor \frac{2^a}{2\epsparam} \rfloor$ for integer $a$) such that some dyadic set \ensuremath{D}\xspace in \ensuremath{X_n}\xspace satisfies $\maxdensof{\ensuremath{D}\xspace}\leq O(\sqrt{n}) 2^{-(h(\epsparam)-2\epsparam) n}$. \end{lemma} \begin{proof} We prove the lemma under the simplifying assumption that $1/\epsparam$ is an integer (our most important application of the lemma has $\epsparam \ensuremath{:=}\xspace 1/5$). Extending the argument for general $\epsparam$ is straightforward and left to the reader. Let $n$ be an integer of the form $\frac{2^{a}}{2\epsparam}$, for a positive integer $a$. Note that for $n$ of this form, $\epsparam n = 2^{a-1}$ is a power of two. Let $t = \epsparam n$, and construct a dyadic distribution \ensuremath{\mu}\xspace on \ensuremath{X_n}\xspace as follows: \begin{enumerate} \item For $i\in [2t-1]$, $\mu(x_i) = 2^{-a}=\frac{1}{2t}$. \item For $i\in [n-1] \setminus [2t-1]$, $\mu(x_i) = \mu(x_{i-1})/2 = 2^{-(a+i-2t+1)}$. \item $\mu(x_n) = \mu(x_{n-1})$. \end{enumerate} The corresponding decision tree is obtained by taking a complete binary tree of depth $a$ and replacing one of the leaves by a ``path'' of length $n-2^a$; see Figure~\ref{fig:hard-distribution}. Alternatively, in the terminology of Definition~\ref{def:tail} we form \ensuremath{\mu}\xspace by taking the uniform distribution on $X_{2t}$ and replacing $x_{2t}$ with a tail on $x_{2t},\ldots,x_n$. \begin{figure} \begin{center} \tikzset{internal/.style={circle,draw,inner sep=2pt}} \tikzset{leaf/.style={circle,fill,inner sep=2pt}} \begin{tikzpicture} \node [draw,regular polygon,regular polygon sides=3,dashed,text width=2cm] (triangle) {Complete binary tree on $2\epsparam n$ vertices}; \node (leaf) [internal] at (triangle.corner 3) {}; \node (v1) [leaf,below left=1cm of leaf] {}; \node (v2) [internal,below right=1cm of leaf] {}; \node (v3) [leaf,below left=1cm of v2] {}; \node (v4) [internal,below right=2cm of v2] {}; \node (v5) [leaf,below left=1cm of v4] {}; \node (v6) [leaf,below right=1cm of v4] {}; \path (leaf) edge (v1) (leaf) edge (v2) (v2) edge (v3) (v2) edge[dashed] node[right] {Path of length $(1-2\epsparam)n$} (v4) (v4) edge (v5) (v4) edge (v6) ; \end{tikzpicture} \end{center} \caption{The hard distribution used to prove Lemma~\ref{lem:densityUB}, in decision tree form} \label{fig:hard-distribution} \end{figure} We claim that $\ensuremath{D}\xspace\ensuremath{:=}\xspace \dyadof{\ensuremath{\mu}\xspace}$ contains only two types of sets: \begin{enumerate} \item Subsets of size $t$ of $X_{2t-1}$. \item Subsets of size $n-t$ containing $t-1$ elements of $X_{2t-1}$ and all the elements $x_{2t},\ldots,x_n$. \end{enumerate} It is immediate that any such set $S$ is in \ensuremath{D}\xspace. On the other hand, Lemma~\ref{lem:small-elements} shows that every set $S \in \ensuremath{D}\xspace$ either contains the tail $x_{2t},\ldots,x_n$ or is disjoint from it. If $S$ is disjoint from the tail then it must be of the first form, and if $S$ contains the tail then it must be of the second form. Using the estimate $\binom{n}{\epsparam n} \geq 2^{h(\epsparam)n}/O(\sqrt{n})$ (see for example~\cite{TCS14476}), we see that \[\reldensof{\ensuremath{D}\xspace}{t}= \reldensof{\ensuremath{D}\xspace}{n-t}=\frac{\binom{2t-1}{t}}{\binom{n}{t}} \leq \frac{2^{2t}}{\binom{n}{\epsparam n}}\leq O(\sqrt{n}) \frac{2^{2t}}{2^{h(\epsparam) n}} = O(\sqrt{n}) 2^{(2\epsparam - h(\epsparam)) n}.\] For $i\in \{1,\ldots,n-1\} \setminus \set{t,n-t}$ we have $\reldensof{\ensuremath{D}\xspace}{i}=0$. Thus indeed \[\maxdensof{\ensuremath{D}\xspace}\leq O(\sqrt{n}) 2^{(2\epsparam - h(\epsparam)) n}. \qedhere \] \end{proof} Theorem~\ref{thm:minimum-redundancy-lb} can now be easily derived. The first step is determining the optimal value of $\epsparam$: \begin{claim}\label{claim:maxOfFunc} We have \[ \max_{\epsparam \in [0,1]} 2^{h(\epsparam)-2 \epsparam} = 1.25, \] and the maximum is attained (uniquely) at $\epsparam=1/5$. \end{claim} \begin{proof} Let $f(\epsparam) = h(\epsparam)-2\epsparam$. Calculation shows that the derivative $f'(\epsparam)$ is equal to \[f'(\epsparam) = \log \left(\frac{1-\epsparam}{\epsparam}\right) -2,\] which is decreasing for $0<\epsparam<1$ and vanishes at $\epsparam =1/5$. Thus $f(\epsparam)$ achieves a unique maximum over $\epsparam\in (0,1)$ at $\epsparam=1/5$, where \[2^{f(1/5)} = 2^{h(1/5)-2\cdot 1/5} =1.25. \qedhere\] \end{proof} \paragraph{Proof of Theorem \ref{thm:minimum-redundancy-lb}:} \begin{proof} Let $\epsparam\ensuremath{:=}\xspace 1/5$. Claim~\ref{claim:maxOfFunc} shows that $2^{-(2\epsparam-h(\epsparam))}=1.25$. Fix any $n$ of the form $n=\frac{2^{a}}{2\epsparam}$ for a positive integer $a$. It follows from Lemma~\ref{lem:densityUB} together with the first inequality in Theorem~\ref{thm:hittersizeisdensity} that $\uhuf(n,0)\geq 1.25^n/O(\sqrt{n})$. A general $n$ can be written in the form $n = \frac{2^a}{2\beta}$ for a positive integer $a$ and $1/4 \leq \beta \leq 1/2$. Lemma~\ref{lem:densityUB} and Theorem~\ref{thm:hittersizeisdensity} show that for any integer $\ell \geq 0$, \[ \uhuf(n,0) \geq 2^{[h(\beta/2^\ell) - 2\beta/2^\ell]n}/O(\sqrt{n}). \] Calculation shows that when $\beta \leq \beta_0 \approx 0.27052059413118146$, this is maximized at $\ell = 0$, and otherwise this is maximized at $\ell = 1$. Denote the resulting lower bound by $L(\beta)^n/O(\sqrt{n})$, the minimum of $L(\beta)$ is attained at $\beta_0$, at which point its value is $L(\beta_0) \approx 1.23214280723432$. \end{proof} \subsection{Lower bounding \smallestdenstitle{n}} \label{sec:un0-lb} We will derive Theorem~\ref{thm:minimum-redundancy-ub} from the following lemma: \begin{lemma}\label{lem:min-redundancy-ubhelper} For every non-constant dyadic distribution \ensuremath{\mu}\xspace there exists $0<\epsparam<1$ such that \[\maxdensof{\dyadof{\ensuremath{\mu}\xspace}}\geq \frac{ 2^{(2\epsparam - h(\epsparam)) n}}{O(\sqrt n)^{O(\log n)}} = 2^{(2\epsparam - h(\epsparam)) n-o(n)}.\] \end{lemma} \begin{proof} Assume without loss of generality that the probabilities in \ensuremath{\mu}\xspace are non-increasing: \[ \ensuremath{\mu}\xspace_1 \geq \ensuremath{\mu}\xspace_2 \geq \cdots \geq \ensuremath{\mu}\xspace_n. \] The idea is to find a partition of $X_n$ of the form \[ X_n = \bigcup_{i=1}^\gamma (D_i \cup E_i) \] which satisfies the following properties: \begin{enumerate} \item $D_i$ consists of elements having the same probability $p_i$. \item If $D_i$ has an even number of elements then $E_i = \emptyset$. \item If $D_i$ has an odd number of elements then $\ensuremath{\mu}\xspace(E_i) = p_i$. \item $\gamma = O(\log n)$. (In fact, $\gamma = o(n/\log n)$ would suffice.) \end{enumerate} We will show later how to construct such a partition. The conditions imply that $\ensuremath{\mu}\xspace(D_i \cup E_i)$ is an even integer multiple of $p_i$, say $\ensuremath{\mu}\xspace(D_i \cup E_i) = 2c_ip_i$. It is not hard to check that $c_i = \lceil |D_i|/2 \rceil$. Given such a partition, we show how to lower bound the maximum relative density of $\dyadof{\ensuremath{\mu}\xspace}$. If $S_i \subseteq D_i$ is a set of size $c_i$ for each $i \in [\gamma]$ then the set $S = \bigcup_i S_i$ splits $\ensuremath{\mu}\xspace$: \[ \ensuremath{\mu}\xspace(S) = \sum_{i=1}^\gamma c_i p_i = \frac{1}{2} \sum_{i=1}^\gamma \ensuremath{\mu}\xspace(D_i \cup E_i) = \frac{1}{2}. \] Defining $c = \sum_{i=1}^\gamma c_i$, we see that each such set $S$ contains $c$ elements, and the number of such sets is \[ \prod_{i=1}^\gamma \binom{|D_i|}{c_i} \geq \prod_{i=1}^\gamma \frac{2^{2c_i}}{O(\sqrt{n})} = \frac{2^{2c}}{O(\sqrt{n})^{O(\log n)}}, \] using the estimate \[ \binom{m}{\lceil m/2 \rceil} = \Theta\left(\frac{2^{2\lceil m/2 \rceil}}{\sqrt{m}}\right), \] which follows from Stirling's approximation. In order to obtain an estimate on the maximum relative density of $\dyadof{\ensuremath{\mu}\xspace}$, we use the following folklore upper bound\footnote{Here is a quick proof: Let $Y$ be a uniformly random subset of $X_n$ of size $c$, and let $Y_i$ indicate the event $x_i \in Y$. Then $\log \binom{n}{c} = H(Y) \leq nH(Y_1) = nh(c/n)$.} on $\binom{n}{c}$: \[ \binom{n}{c} \leq 2^{h(c/n)n}. \] We conclude that the maximum relative density of $\dyadof{\ensuremath{\mu}\xspace}$ is at least \[ \maxdensof{\ensuremath{\mu}\xspace} \geq \reldensof{\ensuremath{\mu}\xspace}{c} \geq \frac{\prod_{i=1}^\gamma \binom{|D_i|}{c_i}}{\binom{n}{c}} \geq \frac{2^{2c-h(c/n)n}}{O(\sqrt{n})^{O(\log n)}}. \] To obtain the expression in the statement of the lemma, take $\epsparam \ensuremath{:=}\xspace c/n$. \medskip We now show how to construct the partition of $X_n$. We first explain the idea behind the construction, and then provide full details; the reader who is interested only in the construction itself can skip ahead. \paragraph{Proof idea} Let $q_1,\ldots,q_\gamma$ be the different probabilities of elements in \ensuremath{\mu}\xspace. We would like to put all elements of probability $q_i$ in the set $D_i$, but there are two difficulties: \begin{enumerate} \item There might be an odd number of elements whose probability is $q_i$. \item There might be too many distinct probabilities, that is, $\gamma$ could be too large. (We need $\gamma = o(n/\log n)$ for the argument to work.) \end{enumerate} The second difficulty is easy to solve: we let $D_1 = \{ x_1 \}$, and use Lemma~\ref{lem:neat-sum} to find an index $\ell$ such that $\ensuremath{\mu}\xspace(E_1) \ensuremath{:=}\xspace \ensuremath{\mu}\xspace(\{x_\ell,\ldots,x_n\}) = \ensuremath{\mu}\xspace_1$. A simple argument shows that all remaining elements have probability at least $\ensuremath{\mu}\xspace_1/n$, and so the number of remaining distinct probabilities is $O(\log n)$. (The reader should observe the resemblance between $E_1$ and the tail of the hard distribution constructed in Lemma~\ref{lem:densityUB}.) Lemma~\ref{lem:neat-sum} also allows us to resolve the first difficulty. The idea is as follows. Suppose that the current set under construction, $D_i$, has an odd number of elements, each of probability $q_i$. We use Lemma~\ref{lem:neat-sum} to find a set of elements whose total probability is $q_i$, and put them in $E_i$. \paragraph{Detailed proof} Let $N$ be the maximal index such that $\ensuremath{\mu}\xspace_N > 0$. Since $\ensuremath{\mu}\xspace$ is non-constant, $\ensuremath{\mu}\xspace_1 \leq 1/2$, and so Lemma~\ref{lem:neat-sum} proves the existence of an index $M$ such that $\ensuremath{\mu}\xspace(\{x_{M+1},\ldots,x_N\}) = \ensuremath{\mu}\xspace_1$ (we use the \emph{furthermore} part of the lemma, and $M = \ell-1$). We take \[ D_1 \ensuremath{:=}\xspace \{ x_1 \}, \quad E_1 \ensuremath{:=}\xspace \{ x_{M+1}, \ldots, x_n \}. \] Thus $\ensuremath{\mu}\xspace(D_1) = \ensuremath{\mu}\xspace(E_1) = \ensuremath{\mu}\xspace_1$, and so $\ensuremath{\mu}\xspace(\{x_2,\ldots,x_M\}) = 1-2\ensuremath{\mu}\xspace_1$ (possibly $M=1$, in which case the construction is complete). By construction $n\ensuremath{\mu}\xspace_M > \ensuremath{\mu}\xspace(E_1) = \ensuremath{\mu}\xspace_1$, and so $\ensuremath{\mu}\xspace_M < \ensuremath{\mu}\xspace_1/n$. In particular, the number of distinct probabilities among $\ensuremath{\mu}\xspace_2,\ldots,\ensuremath{\mu}\xspace_M$ is at most $\log n$. This will guarantee that $\gamma \leq \log n + 1$, as will be evident from the construction. The construction now proceeds in steps. At step $i$, we construct the sets $D_i$ and $E_i$, given the set of available elements $\{x_{\alpha_i},\ldots,x_M\}$, where possibly $\alpha_i = M+1$; in the latter case, we have completed the construction. We will maintain the invariant that $\ensuremath{\mu}\xspace(\{x_{\alpha_i},\ldots,x_M\})$ is an even multiple of $\ensuremath{\mu}\xspace_{\alpha_i}$; initially $\alpha_2 \ensuremath{:=}\xspace 2$, and $\ensuremath{\mu}\xspace(\{x_{\alpha_i},\ldots,x_M\}) = (1/\ensuremath{\mu}\xspace_1-2)\ensuremath{\mu}\xspace_1$ is indeed an even multiple of $\ensuremath{\mu}\xspace_2$. Let $\beta_i$ be the maximal index such that $\ensuremath{\mu}\xspace_{\beta_i} = \ensuremath{\mu}\xspace_{\alpha_i}$ (possibly $\beta_i = \alpha_i$). We define \[ D_i \ensuremath{:=}\xspace \{x_{\alpha_i},\ldots,x_{\beta_i}\}. \] Suppose first that $|D_i|$ is even. In this case we define $E_i \ensuremath{:=}\xspace \emptyset$, and $\alpha_{i+1} \ensuremath{:=}\xspace \beta_i+1$. Note that \[ \ensuremath{\mu}\xspace(\{x_{\alpha_{i+1}},\ldots,x_M\}) = \ensuremath{\mu}\xspace(\{x_{\alpha_i},\ldots,x_M\}) - |D_i| \ensuremath{\mu}\xspace_{\alpha_i}, \] and so the invariant is maintained. Suppose next that $|D_i|$ is odd. In this case $\ensuremath{\mu}\xspace(\{x_{\beta_i+1},\ldots,x_M\}) \geq x_{\alpha_i}$, since $\ensuremath{\mu}\xspace(\{x_{\beta_i+1},\ldots,x_M\})$ is an odd multiple of $\ensuremath{\mu}\xspace_{\alpha_i}$. Therefore we can use Lemma~\ref{lem:neat-sum} to find an index $\gamma_i$ such that $\ensuremath{\mu}\xspace(\{x_{\beta_i+1},\ldots,x_{\gamma_i}\}) = \ensuremath{\mu}\xspace_{\alpha_i}$. We take \[ E_i \ensuremath{:=}\xspace \{x_{\beta_i+1},\ldots,x_{\gamma_i}\} \] and $\alpha_{i+1} \ensuremath{:=}\xspace \gamma_i+1$. Note that \[ \ensuremath{\mu}\xspace(\{x_{\alpha_{i+1}},\ldots,x_M\}) = \ensuremath{\mu}\xspace(\{x_{\alpha_i},\ldots,x_M\}) - (|D_i|+1) \ensuremath{\mu}\xspace_{\alpha_i}, \] and so the invariant is maintained. The construction eventually terminates, say after step $\gamma$. The construction ensures that $\ensuremath{\mu}\xspace_{\alpha_2} > \ensuremath{\mu}\xspace_{\alpha_3} > \cdots > \ensuremath{\mu}\xspace_{\alpha_\gamma}$. Since there are at most $\log n$ distinct probabilities among the elements $\{x_{\alpha_2},\ldots,x_M\}$, $\gamma \leq \log n + 1$, completing the proof. \end{proof} Theorem \ref{thm:minimum-redundancy-ub} follows immediately from the second inequality in Theorem \ref{thm:hittersizeisdensity} together with the following lemma: \begin{lemma}\label{lem:densityLB} Fix an integer $n$ and let \ensuremath{D}\xspace be a dyadic set in \ensuremath{X_n}\xspace. Then \[\maxdensof{\ensuremath{D}\xspace}\geq 1.25^{-n-o(n)},\] and thus \[\smallestdens{n}\geq 1.25^{-n-o(n)}.\] \end{lemma} \begin{proof} Fix a dyadic set \ensuremath{D}\xspace in \ensuremath{X_n}\xspace. Lemma~\ref{lem:min-redundancy-ubhelper} implies that there exists $0<\epsparam<1$ such that $\maxdensof{\ensuremath{D}\xspace}\geq 2^{(2\epsparam - h(\epsparam)) n-o(n)}$. Using Claim \ref{claim:maxOfFunc} we have $2^{2\epsparam - h(\epsparam)}\geq \frac{4}{5}$, and so \[\maxdensof{\ensuremath{D}\xspace}\geq (4/5)^n\cdot 2^{-o(n)}= 1.25^{-n-o(n)}. \qedhere \] \end{proof} \section{Combinatorial benchmark with prolixity} \label{sec:prolixity} In the previous section we studied the minimum size of a set $\cQ$ of questions with the property that for every distribution, there is an optimal decision tree using only questions from $\cQ$. In this section we relax this requirement by allowing the cost to be slightly worse than the optimal cost. More formally, recall that $\uhuf(n,r)$ is the minimum size of a set of questions $\cQ$ such that for every distribution $\pi$ there exists a decision tree that uses only questions from $\cQ$ with cost at most $\opt{\pi} + r$. In a sense, $\uhuf(n,r)$ is an extension of $\uent(n,r)$ for $r\in(0,1)$: indeed, $\uent(n,r)$ is not defined for $r<1$ since for some distributions $\pi$ there is no decision tree with cost less than $H(\pi)+1$ (see Section~\ref{sec:comparison-equality}). Moreover, $\opt{\pi}$, which is the benchmark used by $\uhuf(n,r)$, is precisely the optimal cost, whereas $H(\pi)$, the benchmark used by $\uent(n,r)$ is a convex surrogate of $\opt{\pi}$. We focus here on the range $r\in(0,1)$. We prove the following bounds on $\uhuf(n,r)$, establishing that $\uhuf(n,r) \approx (r\cdot n)^{\Theta(1/r)}$. \begin{main-theorem}\label{thm:min-redundancy-approx} For all $r\in(0,1)$, and for all $n>1/r$: \[\frac{1}{n}(r\cdot n)^{\frac{1}{4r}}\leq \uhuf(n,r) \leq n^2(3r\cdot n)^{\frac{16}{r}}. \] \end{main-theorem} As a corollary, we get that the threshold of exponentiality is $1 / n$: \begin{main-corollary} \label{cor:minimum-redundancy-approx-exp} If $r = \omega(1/n)$ then $\uhuf(n,r) = 2^{o(n)}$. Conversely, if $r = O(1/n)$ then $\uhuf(n,r) = 2^{\Omega(n)}$. \end{main-corollary} For larger $r$, the following theorem is a simple corollary of Theorem~\ref{thm:redundancy-r-bounds} and the bound $\uhuf(n,r) \leq \uent(n,r) \leq \uhuf(n,r-1)$: \begin{main-theorem} \label{thm:prolixity-r-bounds} For every $r \geq 1$ and $n\in\mathbb{N}$, \[ \frac{1}{e}\lfloor r+1 \rfloor n^{1/\lfloor r+1 \rfloor} \leq \uhuf(n,r) \leq 2 \lfloor r \rfloor n^{1/\lfloor r \rfloor}. \] \end{main-theorem} \medskip Theorem~\ref{thm:min-redundancy-approx} is implied by the following lower and upper bounds, which provide better bounds when $r\in (0,1)$ is a negative power of~$2$. \begin{main-theorem}[Lower bound] \label{thm:minimum-redundancy-approx-lb} For every $r$ of the form $1/2^k$, where $k\geq 1$ is an integer, and $n>2^k$: \[ \uhuf(n,r) \geq (r\cdot n)^{\frac{1}{2r}-1}. \] \end{main-theorem} \begin{main-theorem}[Upper bound] \label{thm:minimum-redundancy-approx-ub} For every $r$ of the form $4/2^k$, where $k\geq 3$ is an integer, and $n>2^k$: \[ \uhuf(n,r+ r^2) \leq n^2 \bigl(\frac{3e}{4}r\cdot n\bigr)^{\frac{4}{r}}. \] \end{main-theorem} These results imply Theorem~\ref{thm:min-redundancy-approx}, due to the monotonicity of $\uhuf(n,r)$, as follows. Let $r\in(0,1)$. For the \emph{lower bound}, pick the smallest $t\geq r$ of the form $1/2^k$. Note that $t\leq 2r$, and thus: \[\uhuf(n,r) \geq \uhuf(n,t) \geq (t\cdot n)^{\frac{1}{2t}-1} \geq (r\cdot n)^{\frac{1}{4r}-1}\geq \frac{1}{n}(r\cdot n)^{\frac{1}{4r}}.\] For the \emph{upper bound}, pick the largest $t$ of the form $4/2^k$, $k\geq 3$ such that $t + t^2\leq r$. Note that $t\geq r/4$ (since $s=r/2$ satisfies $s+s^2 \leq 2s \leq r$), and thus \[\uhuf(n,r) \leq \uhuf(n, t+t^2) \leq n^2 \bigl(\frac{3e}{4}t\cdot n\bigr)^{\frac{4}{t}}\leq n^2(3r\cdot n)^{\frac{16}{r}}.\] \subsection{Lower bound} \label{sec:minimum-redundancy-approx-lb} Pick a sufficiently small $\delta>0$ (as we will soon see, $\delta < r^2$ suffices), and consider a distribution $\mu$ with $2^k-1$ ``heavy'' elements (this many elements exist since $n>1/r$), each of probability $\frac{1-\delta}{2^k-1}$, and $n-(2^k-1)$ ``light'' elements with total probability of $\delta$. Recall that a decision tree is \emph{$r$-optimal} if its cost is at most $\opt{\mu}+r$. The proof proceeds by showing that if $T$ is an $r$-optimal tree, then the first question in $T$ has the following properties: \begin{enumerate}[label=(\roman*)] \item it separates the heavy elements to two sets of almost equal sizes ($2^{k-1}$ and $2^{k-1}-1$), and \item it does not distinguish between the light elements. \end{enumerate} The result then follows since there are $\binom{n}{2^k-1}$ such distributions $\sigma$ (the number of ways to choose the light elements), and each question can serve as a first question to at most $\binom{n-(2^{k-1}-1)}{2^{k-1}}$ of them. To establish these properties, we first prove a more general result (cf.\ Lemma~\ref{lem:small-elements}): \begin{lemma}\label{l324} Let $\mu$ be a distribution over a finite set $X$, and let $A\subseteq X$ be such that for every $x\notin A$, $\mu\bigl(\{x\}\bigr) > \mu(A) + \epsilon$. Then every decision tree $T$ which is $\epsilon$-optimal with respect to $\mu$ has a subtree $T'$ whose set of leaves is $A$. \end{lemma} \begin{proof} By induction on $\lvert A\rvert$. The case $\lvert A\rvert =1$ follows since any leaf is a subtree. Assume $\lvert A\rvert > 1$. Let $T$ be a decision tree which is $\epsilon$-optimal with respect to $\mu$. Let $x,y$ be two siblings of maximal depth. Note that it suffices to show that $x,y\in A$, since then, merging $x,y$ to a new element $z$ with $\mu(\{z\}) = \mu(\{x\}) + \mu(\{y\})$ and applying the induction hypothesis yields that $A\cup\{z\}\setminus\{x,y\}$ is the set of leaves of a subtree of $T$ with $x,y$ removed. This finishes the proof since $x,y$ are the children of $z$. It remains to show that $x,y\in A$. Let $d$ denote the depth of $x$ and $y$. Assume toward contradiction that $x\notin A$. Pick $a',a''\in A$, with depths $d',d''$ (this is possible since $\lvert A\rvert > 1$). If $d'<d$ or $d''<d$ then replacing $a'$ with $x$ or $a''$ with $x$ improves the cost of $T$ by more than $\epsilon$, contradicting its optimality. Therefore, it must be that $d'=d''=d$, and we perform the following transformation (see Figure~\ref{fig:transformation}): the parent of $x$ and $y$ becomes a leaf with label $x$ (decreasing the depth of $x$ by 1), $y$ takes the place of $a'$ (the depth of $y$ does not change), and $a''$ becomes an internal node with two children labeled by $a',a''$ (increasing the depths of $a',a''$ by 1). Since $\mu\bigl(\{x\}\bigr) - \mu\bigl(\{a',a''\}\bigr) > \epsilon$, this transformation improves the cost of $T$ by more than $\epsilon$, contradicting its $\epsilon$-optimality. \end{proof} \begin{figure} \begin{subfigure}[t]{.5\linewidth} \centering \begin{tikzpicture}[baseline=(root.base), level 1/.style={sibling distance=4cm}, level 2/.style={sibling distance=2.5cm}] \node[circle, draw] (root) {} child{ node[circle, draw, fill=black] {} edge from parent[dashed] child{ node[circle, draw, solid] {$x$} edge from parent[solid] } child{ node[circle, draw, solid] {$y$} edge from parent[solid] } } child{ node[circle, draw] {} edge from parent[dashed] child{ node[circle, draw, solid] {$a'$} } child{ node[circle, draw, solid] {$a''$} } }; \end{tikzpicture} \caption{Original tree} \end{subfigure} \begin{subfigure}[t]{.5\linewidth} \centering \begin{tikzpicture}[baseline=(root.base), level 1/.style={sibling distance=2.5cm}, level 2/.style={sibling distance=2.5cm} level 2/.style={sibling distance=2.5cm}] \node[circle, draw] (root) {} child{node[circle, draw] {$x$} edge from parent[dashed] } child{ node[circle, draw] {} edge from parent[dashed] child{ node[circle, draw, solid] {$y$} } child{ node[circle, draw, fill=black, solid] {} child{ node[circle, draw, solid] {$a'$} edge from parent[solid] } child{ node[circle, draw, solid] {$a''$} edge from parent[solid] } } }; \end{tikzpicture} \caption{Transformed tree} \end{subfigure} \caption{The transformation in Lemma~\ref{l324}. The cost decreases by $\mu\bigl(\{x\}\bigr)-\mu\bigl(\{a',a''\}\bigr)>\epsilon$.}\label{fig:transformation} \end{figure} \begin{corollary}\label{c325} Let $\mu$ be a distribution over $X$, and let $A\subseteq X$ be such that for every $x\notin A$, $\mu\bigl(\{x\}\bigr) > \mu(A)$. Then every optimal tree $T$ with respect to $\mu$ has a subtree $T'$ whose set of leaves is $A$. \end{corollary} Property~(ii) follows from Lemma~\ref{l324}, which implies that if $\delta$ is sufficiently small then all light elements are clustered together as the leaves of some subtree. Indeed, by Lemma~\ref{l324}, this happens if the probability of a single heavy element (which is $\frac{1-\delta}{2^k-1}$) exceeds the total probability of all light elements (which is $\delta$) by at least $r$. A simple calculation shows that setting $\delta$ smaller than $r^2$ suffices. We summarize this in the following claim: \begin{claim}[light elements] Every $r$-optimal tree has a subtree whose set of leaves is the set of light elements. \end{claim} The next claim concerns the other property: \begin{claim}[heavy elements] In every $r$-optimal decision tree, the first question partitions the heavy elements into a set of size $2^{k-1}$ and a set of size $2^{k-1}-1$. \end{claim} \begin{proof} When $k = 2$, it suffices to prove that an $r$-optimal decision tree cannot have a first question which separates the heavy elements from the light elements. Indeed, the heavy elements in such a tree reside at depths $2,3,3$. Exchanging one of the heavy elements at depth~$2$ with the subtree consisting of all light elements (which is at depth~$1$) decreases the cost by $\frac{1-\delta}{2^k-1} - \delta > r$, showing that the tree wasn't $r$-optimal. Suppose that some $r$-optimal decision tree $T$ contradicts the statement of the claim, for some $k \geq 3$. The first question in $T$ leads to two subtrees $T_1,T_2$, one of which (say $T_1$) contains at least $2^{k-1}+1$ heavy elements, and the other (say $T_2$) contain at most $2^{k-1}-2$. One of the subtrees also contains a subtree $T'$ whose leaves are all the light elements. For the sake of the argument, we replace the subtree $T'$ with a new element $y$. We claim that $T_1$ contains an internal node $v$ at depth $D(v) \geq k-1$ which has at least two heavy descendants. To see this, first remove $y$ if it is present in $T_1$, by replacing its parent by its sibling. The possibly modified tree $T'_1$ contains at least $2^{k-1}+1$ leaves, and in particular some leaf at depth at least $k$. Its parent $v$ has depth at least $k-1$ and at least two heavy descendants, in both $T'_1$ and $T_1$. In contrast, $T_2$ contains at least two leaves (since $2^{k-1}-2 \geq 2$), and the two shallowest ones must have depth at most $k-2$. At least one of these is some heavy element $x_\ell$. Exchanging~$v$ and $x_\ell$ results in a tree $T^*$ whose cost $c(T^*)$ is at most \[ c(T^*) \leq c(T) + (D(v) - D(x_\ell))(2 - 1)\frac{1-\delta}{2^k-1} \leq c(T^*) - \frac{1-\delta}{2^k-1} < c(T) - r, \] contradicting the assumption that $T$ is $r$-optimal. (That $\frac{1-\delta}{2^k-1} > r$ follows from the earlier assumption $\frac{1-\delta}{2^k-1} > \delta + r$.) \end{proof} By the above claims, there are two types of first questions for $\mu$, depending on which of the two subtrees of the root contains the light elements: \begin{itemize} \item Type 1: questions that split the elements into a part with $2^{k-1}$ elements, and a part with $n-2^{k-1}$ elements. \item Type 2: questions that split the elements into a part with $2^{k-1}-1$ elements, and a part with $n-(2^{k-1}-1)$ elements. \end{itemize} If we identify a question with its smaller part (i.e.\ the part of size $2^{k-1}$ or the part of size $2^{k-1}-1$), we deduce that any set of questions with redundancy $r$ must contain a family $\cF$ such that (i) every set in $\cF$ has size $2^{k-1}$ or $2^{k-1}-1$, and (ii) for every set of size $n-(2^k-1)$, there exists some set in $\cF$ that is disjoint from it. It remains to show that any such family $\cF$ is large. Indeed, there are $\binom{n}{2^k-1}$ sets of size $n-(2^k-1)$, and since every set in $\cF$ has size at least $2^{k-1}-1$, it is disjoint from at most $\binom{n-(2^{k-1}-1)}{n-(2^k-1)}=\binom{n-(2^{k-1}-1)}{2^{k-1}}$ of them. Thus \[ |\cF| \geq \frac{\binom{n}{2^k-1}}{\binom{n-(2^{k-1}-1)}{2^{k-1}}} = \frac{n(n-1)\cdots(n-(2^{k-1}-1)+1)}{(2^k-1)(2^k-2)\cdots(2^{k-1}+1)} \geq \bigl(\frac{n}{2^k}\bigr)^{2^{k-1}-1} = (r\cdot n)^{\frac{1}{2r}-1}. \] \subsection{Upper bound} \label{sec:minimum-redundancy-approx-ub} \paragraph{The set of questions.} In order to describe the set of queries it is convenient to assign a cyclic order on $X_n$: $x_1 \prec x_2 \prec \cdots \prec x_n \prec x_1 \prec \cdots$. The set of questions $\cQ$ consists of all cyclic intervals, with up to $2^k$ elements added or removed. Since $r = 4 \cdot 2^{-k}$, the number of questions is plainly at most \[ n^2 \binom{n}{2^k} 3^{2^k} \leq n^2 \bigl(\frac{3e}{4}r\cdot n\bigr)^{\frac{4}{r}}, \] using the inequality $\binom{n}{d}\leq\bigl(\frac{en}{d}\bigr)^d$. \paragraph{High level of the proof.} Let $\pi$ be an arbitrary distribution on $X_n$, and let $r\in(0,1)$ be of the form $4\cdot 2^{-k}$, with $k\geq 3$. Let $\mu$ be a Huffman distribution for $\pi$; we remind the reader that $\mu$ is a dyadic distribution corresponding to some optimal decision tree for $\pi$. We construct a decision tree $T$ that uses only queries from $\cQ$, with cost \[ T(\pi)\leq \opt{\pi} + r + r^2 = \sum_{x \in X_n}{\pi(x)\log\frac{1}{\mu(x)}} + r + r^2. \] The construction is randomized: we describe a randomized decision tree $T_R$ (`$R$' denotes the randomness that determines the tree) which uses queries from $\cQ$ and has the property that for every $x\in X_n$, the expected number of queries $T_R$ uses to find $x$ satisfies the inequality \begin{equation}\label{eq4} \E_R[T_R(x)] \leq \log \frac{1}{\mu(x)} + r + r^2, \end{equation} where $T_R(x)$ is the depth of $x$. This implies the existence of a deterministic tree with cost $\opt{\mu} + r + r^2$: indeed, when $x\sim\mu$, the expected cost of $T_R$ is \[ \E_{\substack{x \sim \pi; R}}[T_R(x)]\leq \sum_{x \in X_n}{\pi(x)\Bigl(\frac{1}{\mu(x)} + r + r^2\Bigr)} = \opt{\pi} + r + r^2. \] Since the randomness of the tree is independent from the randomness of $\pi$, it follows that there is a choice of $R$ such that the cost of the (deterministic) decision tree $T_R$ is at most $\opt{\pi} + r + r^2$. \paragraph{The randomized decision tree.} The randomized decision tree maintains a dyadic \emph{sub-distribution} $\mu^{(i)}$ that is being updated after each query. A \emph{dyadic sub-distribution} is a measure on $X_n$ such that (i) $\mu^{(i)}(x)$ is either 0 or a power of 2, and (ii) $\mu^{(i)}(X_n)=\sum_{x\in X_n}\mu^{(i)}(x)\leq 1$. A natural interpretation of $\mu^{(i)}(x)$ is as a dyadic sub-estimate of the probability that $x$ is the \unknown element, conditioned on the answers to the first $i$ queries. The analysis hinges on the following properties: \begin{enumerate} \item $\mu^{(0)}=\mu$, \item $\mu^{(i)}(x)\in\bigl\{2\mu^{(i-1)}(x),\mu^{(i-1)}(x),0\bigr\}$ for all $x\in X_n$, \item if $x$ is the \unknown element then almost always $\mu^{(i)}(x)$ is doubled; that is, $\mu^{(i)}(x)>0$ for all $i$, and the expected number of $i$'s for which $\mu^{(i)}(x)=\mu^{(i-1)}(x)$ is at most $r+r^2$. \end{enumerate} These properties imply~\eqref{eq4}, which implies Theorem~\ref{thm:minimum-redundancy-approx-ub}. Next, we describe the randomized decision tree and establish these properties. The algorithm distinguishes between {light} and {heavy} elements. An element $x\in X_n$ is \emph{light} if $\mu^{(i)}(x) < 2^{-k}$. Otherwise it is \emph{heavy}. The algorithm is based on the following win-win-win situation: (i) If the total mass of the heavy elements is at least $1/2$ then by Lemma~\ref{lem:neat-sum}, there is a set $I$ of heavy elements whose mass is exactly $1/2$. Since the number of heavy elements is at most $2^k$, the algorithm can ask whether $x \in I$ and recurse by doubling the sub-probabilities of the elements that are consistent with the answer (and setting the others to zero). (ii) Otherwise, the mass of the heavy elements is less than $1/2$. If the mass of the light elements is also less than $1/2$ (this could happen since $\mu^{(i)}$ is a sub-distribution), then we ask whether $x$ is a heavy element or a light element, and accordingly recurse with either the heavy or the light elements, with their sub-probabilities doubled (in this case the ``true'' probabilities conditioned on the answers become larger than the sub-probabilities). (iii) The final case is when the mass of the light elements is larger than $1/2$. In this case we query a random cyclic interval of light elements of mass $\approx 1/2$, and recurse; there are two light elements in the recursion whose sub-probability is not doubled (the probabilities of the rest are doubled). Elements whose probability is not doubled occur only in case~(iii). \paragraph{The randomized decision tree: formal description.} The algorithm gets as input a subset $y_1,\ldots,y_m$ of $X_n$ whose order is induced by that of $X_n$, and a dyadic sub-distribution $q_1,\ldots,q_m$. Initially, the input is $x_1,\ldots,x_n$, and $q_i = \mu_i$. We say that an element is \emph{heavy} is $q_i \geq 2^{-k}$; otherwise it is \emph{light}. There are at most $2^k$ heavy elements. The questions asked by the algorithm are cyclic intervals in $y_1,\ldots,y_m$, with some heavy elements added or removed. Since each cyclic interval in $y_1,\ldots,y_m$ corresponds to a (not necessarily unique) cyclic interval in $X_n$ (possibly including elements outside of $y_1,\ldots,y_m$), these questions belong to $\cQ$. \begin{algorithm-description}{$T_R$} \begin{enumerate} \item If $m = 1$, return $y_1$. Otherwise, continue to Step~2. \item If the total mass of heavy elements is at least $1/2$ then find (using Lemma~\ref{lem:neat-sum}) a subset $I$ whose mass is exactly $1/2$, and ask whether $x \in I$. Recurse with either $\{ 2q_i : y_i \in I \}$ or $\{ 2q_i : y_i \notin I \}$, according to the answer. Otherwise, continue to Step~3. \item Let $S$ be the set of all light elements, and let $\sigma$ be their total mass. If $\sigma \leq 1/2$ then ask whether $x \in S$, and recurse with either $\{ 2q_i : y_i \in S \}$ or $\{ 2q_i : y_i \notin S \}$, according to the answer. Otherwise, continue to Step~4. \item Arrange all light elements according to their cyclic order on a circle of circumference $\sigma$, by assigning each light element $x_i$ an arc $A_i$ of length $q_i$ of the circle. Pick an arc of length $1/2$ uniformly at random (e.g.\ by picking uniformly a point on the circle and taking an arc of length $1/2$ directed clockwise from it), which we call the \emph{window\xspace}. Let $K \subseteq S$ consist of all light elements whose midpoints are contained in the window\xspace, and let $B$ consist of the light elements whose arcs are cut by the boundary of the window\xspace (so $|B| \leq 2$); we call these elements \emph{boundary elements}. Ask whether $x \in K$; note that $K$ is a cyclic interval in $y_1,\ldots,y_m$ with some heavy elements removed. If $x \in K$, recurse with $\{2q_i : y_i \in K \setminus B\} \cup \{q_i : y_i \in K \cap B\}$. The sum of these dyadic probabilities is at most~$1$ since the window\xspace contains at least $q_i/2$ of the arc $A_i$ for each $y_i \in K \cap B$. If $x \notin K$, recurse with $\{2q_i : y_i \in \overline{K} \setminus B \} \cup \{q_i : y_i \in \overline{K} \cap B\}$. As in the preceding case, the total mass of light elements in the recursion is at most $2(\sigma - 1/2)$ (since the complement of the window\xspace contains at least $q_i/2$ of the arc $A_i$ for each $y_i \in \overline{K} \cap B$), and the total mass of heavy elements is $2(1-\sigma)$, for a total of at most $(2\sigma-1)+(2-2\sigma) = 1$. \eofahere \end{enumerate} \end{algorithm-description} \paragraph{Analysis.} We now finish the proof by establishing the three properties of the randomized decision tree that are stated above. The first two properties follow immediately from the description of the algorithm, and it thus remains to establish the third property. Fix some $x\in X_n$, and let $d\in\mathbb{N}$ be such that $\mu(x)=2^{-d}$. We need to show that the expected number of questions that are asked when the \unknown element is $x$ is at most $d + r + r^2$. Let $q=q^{(i)}$ denote the sub-probability of $x$ after the $i$'th question; note that $q\in\{2^{-j} : j\leq d\}$. \begin{lemma}\label{l721} If $q \geq 2^{-k}$ then $q$ doubles (that is, $q^{(i+1)} = 2q^{(i)}$). Otherwise, the expected number of questions until $q$ doubles is at most $\frac{1}{1-4q}$. \end{lemma} \begin{proof} From the description of the algorithm, it is clear that the only case in which the sub-probability of $x$ is not doubled is when $x$ is one of the two boundary elements in Step~4. This only happens when $x$ is a light element (i.e.\ $q < 2^{-k}$). The probability that $x$ is one of the boundary elements is at most $2q/\sigma\leq 4q$, where $\sigma\geq1/2$ is the total mass of light elements: indeed, the probability that a given endpoint of the window\xspace lies inside the arc corresponding to $q$ is $q/\sigma$, since each endpoint is distributed uniformly on the circle of circumference $\sigma$. It follows that the distribution of the number of questions that pass until $q$ doubles is dominated by the geometric distribution with failure probability $4q$, and so the expected number of questions until $q$ doubles is at most $\frac{1}{1-4q}$. \end{proof} The desired bound on the expected number of questions needed to find $x$ follows from Lemma~\ref{l721}: as long as $q$, the sub-probability associated with $x$, is smaller than $2^{-k}$, it takes an expected number of $\frac{1}{1-4q}$ questions until it doubles. Once $q\geq 2^{-k}$, it doubles after every question. Thus, by linearity of expectation, the expected total number of questions is at most: \begin{align*} k + \sum_{j=k+1}^d \frac{1}{1-4\cdot2^{-j}} &< k + \sum_{j=k+1}^d [{1+ 4\cdot 2^{-j} + 2(4\cdot 2^{-j})^2}] \\ & = d + \sum_{j=k+1}^d [{4\cdot 2^{-j} + 2(4\cdot 2^{-j})^2}] \\ &< d + 4\cdot 2^{-k} + \frac{2}{3} (4\cdot 2^{-k})^2\\ &< \log \frac{1}{\mu(x)} + r + r^2. \end{align*} \section{Open questions} \label{sec:open-questions} Our work suggests many open questions, some of which are: \begin{enumerate} \item The main results of Section~\ref{sec:huffman} show that when $n = 5\cdot 2^m$, $\uhuf(n,0) = 1.25^{n \pm o(n)}$. We conjecture that there exists a function $G\colon [1,2] \to \mathbb{R}$ such that for $n = \alpha 2^m$, $\uhuf(n,0) = G(\alpha)^{n \pm o(n)}$. Our results show that $1.232 \leq G(\alpha) \leq 1.25$ and that $G(1.25) = 1.25$. What is the function $G$? \item Theorem~\ref{thm:minimum-redundancy-ub} constructs an optimal set of questions of size $1.25^{n + o(n)}$, but this set is not explicit. In contrast, Theorem~\ref{thm:cone} constructs explicitly an optimal set of questions of size $O(\sqrt{2}^n)$, which furthermore supports efficient indexing and efficient construction of optimal strategies. Can we construct such an explicit set of optimal size $1.25^{n + o(n)}$? \item The results of Section~\ref{sec:comparison-equality} show that $n \leq \uent(n,1) \leq 2n-3$. We conjecture that the limit $\beta = \lim_{n\to\infty} \frac{\uent(n,1)}{n}$ exists. What is the value of $\beta$? \end{enumerate} An interesting suggestion for future research is to generalize the entire theory to $d$-way questions. \bibliographystyle{plain}

https://arxiv.org/abs/2105.12806

A Universal Law of Robustness via Isoperimetry

Classically, data interpolation with a parametrized model class is possible as long as the number of parameters is larger than the number of equations to be satisfied. A puzzling phenomenon in deep learning is that models are trained with many more parameters than what this classical theory would suggest. We propose a partial theoretical explanation for this phenomenon. We prove that for a broad class of data distributions and model classes, overparametrization is necessary if one wants to interpolate the data smoothly. Namely we show that smooth interpolation requires $d$ times more parameters than mere interpolation, where $d$ is the ambient data dimension. We prove this universal law of robustness for any smoothly parametrized function class with polynomial size weights, and any covariate distribution verifying isoperimetry. In the case of two-layers neural networks and Gaussian covariates, this law was conjectured in prior work by Bubeck, Li and Nagaraj. We also give an interpretation of our result as an improved generalization bound for model classes consisting of smooth functions.

\section{Introduction} Solving $n$ equations generically requires only $n$ unknowns\footnote{As in, for instance, the inverse function theorem in analysis or B\'ezout's theorem in algebraic geometry. See also \cite{yun2019small, BELM20} for versions of this claim with neural networks.}. However, the revolutionary deep learning methodology revolves around highly overparametrized models, with many more than $n$ parameters to learn from $n$ training data points. We propose an explanation for this enigmatic phenomenon, showing in great generality that finding a \emph{smooth} function to fit $d$-dimensional data requires at least $nd$ parameters. In other words, overparametrization by a factor of $d$ is {\em necessary} for {\em smooth} interpolation, suggesting that perhaps the large size of the models used in deep learning is a {\em necessity} rather than a weakness of the framework. Another way to phrase the result is as a {\em tradeoff} between the size of a model (as measured by the number of parameters) and its ``robustness" (as measured by its Lipschitz constant): either one has a small model (with $n$ parameters) which must then be non-robust, or one has a robust model (constant Lipschitz) but then it must be very large (with $nd$ parameters). Such a tradeoff was conjectured for the specific case of two-layer neural networks and Gaussian data in \cite{bubeck2020law}. Our result shows that in fact it is a {\em universal} phenomenon, which applies to essentially any parametrized function class (including in particular deep neural networks) as well as a much broader class of data distributions. As in \cite{bubeck2020law} we obtain an entire tradeoff curve between size and robustness: our universal law of robustness states that, for any function class smoothly parametrized by $p$ parameters, and for any $d$-dimensional dataset satisfying mild regularity conditions, any function in this class that fits the data {\em below the noise level} must have its (Euclidean) Lipschitz constant larger than $\sqrt{\frac{nd}{p}}$. \begin{theorem}[Informal version of Theorem~\ref{thm:main}]\label{thm:inf} Let $\mathcal F$ be a class of functions from $\mathbb R^d\to \mathbb R$ and let $(x_i,y_i)_{i=1}^n$ be i.i.d. input-output pairs in $\mathbb R^d\times [-1,1]$. Assume that: \begin{enumerate} \item $\mathcal F$ admits a Lipschitz parametrization by $p$ real parameters, each of size at most $poly(n,d)$. \item The distribution $\mu$ of the covariates $x_i$ satisfies isoperimetry (or is a mixture theoreof). \item The expected conditional variance of the output (i.e., the ``noise level") is strictly positive, denoted $\sigma^2 := \mathbb E^{\mu}[Var[y|x]] > 0$. \end{enumerate} Then, with high probability over the sampling of the data, one has simultaneously for all $f \in \mathcal F$: \[ \frac{1}{n}\sum_{i=1}^n (f(x_i)-y_i)^2 \leq \sigma^2-\epsilon \,\, \Rightarrow \,\, \mathrm{Lip}(f) \geq \widetilde{\Omega}\left(\epsilon\sqrt{\frac{nd}{p}}\right) \,. \] \end{theorem} \begin{remark}\label{rem:exists} For the distributions $\mu$ we have in mind, for instance uniform on the unit sphere, there exists with high probability some $O(1)$-Lipschitz function $f:\mathbb R^d\to\mathbb R$ satisfying $f(x_i)=y_i$ for all $i$. Indeed, with probability $1-e^{-\Omega(d)}$ we have $||x_i-x_j||\geq 1$ for all $1\leq i\neq j\leq n$ so long as $n\leq poly(d)$. In this case we may apply the Kirszbraun extension theorem to find a suitable $f$ regardless of the labels $y_i$. More explicitly we may fix a smooth bump function $g:\mathbb R^+\to\mathbb R$ with $g(0)=1$ and $g(x)=0$ for $x\geq 1$, and then interpolate using the sum of radial basis functions \[ f(x)=\sum_{i=1}^n g(||x-x_i||)y_i. \] In fact this construction requires only $p=n(d+1)$ parameters to specify the values $(x_i,y_i)_{i\in [n]}$ and thus determine the function $f$. Hence $p=n(d+1)$ parameters suffice for robust interpolation, i.e. Theorem~\ref{thm:main} is essentially best possible for $L=O(1)$. A similar construction shows the same conclusion for any $p\in [\widetilde{\Omega}(n),nd]$, essentially tracing the entire tradeoff curve. This is because one can first project onto a fixed subspace of dimension $\tilde d=p/n$, and the projected inputs $x_i$ now have pairwise distances at least $\Omega\left(\sqrt{\frac{\tilde d}{d}}\right)$ with high probability. The analogous construction on the projected points now requires only $p=\tilde{d}n$ parameters and has Lipschitz constant $L=O\left(\sqrt{\frac{d}{\tilde d}}\right)=O\left(\sqrt{\frac{nd}{p}}\right)$. \end{remark} \subsection{Speculative implication for real data}\label{subsec:speculative} To put Theorem \ref{thm:inf} in context, we compare to the empirical results presented in \cite{madry2017towards}. In the latter work, they consider the MNIST dataset which consists of $n=6\times10^4$ images in dimension $28^2=784$. They trained robustly different architectures, and reported in Figure 4 the size of the architecture versus the obtained robust test accuracy (third plot from the left). One can see a sharp transition from roughly 10\% accuracy to roughly 90\% accuracy at around $2 \times 10^5$ parameters (capacity scale $4$ in their notation). Moreover the robust accuracy keeps climbing up with more parameters, to roughly 95\% accuracy at roughly $3 \times 10^6$ parameters. \newline How can we compare these numbers to the law of robustness? There are a number of difficulties that we discuss below, and we emphasize that this discussion is highly speculative in nature, though we find that, with a few leaps of faith, our universal law of robustness sheds light on the potential parameter regimes of interest for robust deep learning. \newline The first difficulty is to evaluate the ``correct" dimension of the problem. Certainly the number of pixels per image gives an upper bound, however one expects that the data lies on something like a lower dimensional sub-manifold. Optimistically, we hope that Theorem~\ref{thm:inf} will continue to apply for an appropriate \emph{effective dimension} which may be rather smaller than the literal number of pixels. This hope is partially justified by the fact that isoperimetry holds in many less-than-picturesque situations, some of which are stated in the next subsection. \newline The next difficulty is to estimate/interpret the noise value $\sigma^2$. From a theoretical point of view, this noise assumption is necessary for otherwise there could exist a smooth classifier with perfect accuracy in $\mathcal{F}$, defeating the point of any lower bound on the size of $\mathcal F$. We tentatively would like to think of $\sigma^2$ as capturing the contribution of the ``difficult" part of the learning problem, that is $\sigma^2$ could be thought of as the non-robust generalization error of reasonably good models, so a couple of $\%$ of error in the case of MNIST. With that interpretation, one gets ``below the noise level" in MNIST with a training error of a couple of $\%$. We believe that versions of the law of robustness might hold without noise; these would need to go beyond representational power and consider the dynamics of learning algorithms. \newline Finally another subtlety to interpret the empirical results of \cite{madry2017towards} is that there is a mismatch between what they measure and our quantities of interest. Namely the law of robustness talks about two things: the training error, and the worst-case robustness (i.e., the Lipschitz constant). On the other hand \cite{madry2017towards} measures the {\em robust generalization error}. Understanding the interplay between those three quantities is a fantastic open problem. Here we take the perspective that a small robust generalization error should imply a small training error and a small Lipschitz constant. Another important mismatch is that we stated our universal law of robustness for Lipschitzness in $\ell_2$, while the experiments in \cite{madry2017towards} are for robustness in $\ell_{\infty}$. We believe that a variant of the law of robustness remains true for $\ell_{\infty}$, a belief again partially justified by how broad isoperimetry is (see next subsection). \newline With all the caveats described above, we can now look at the numbers as follows: in the \cite{madry2017towards} experiments, smooth models with accuracy below the noise level are attained with a number of parameters somewhere in the range $2 \times 10^5 - 3 \times 10^6$ parameters (possibly even larger depending on the interpretation of the noise level), while the law of robustness would predict any such model must have at least $n d$ parameters, and this latter quantity should be somewhere in the range $10^6 - 10^7$ (corresponding to an effective dimension between $15$ and $150$). While far from perfect, the law of robustness prediction is far more accurate than the classical rule of thumb $\#$ parameters $\simeq$ $\#$ equations (which here would predict a number of parameters of the order $10^4$). \newline Perhaps more interestingly, one could apply a similar reasoning to the ImageNet dataset, which consists of $1.4 \times 10^7$ images of size roughly $2 \times 10^5$. Estimating that the effective dimension is a couple of order of magnitudes smaller than this size, the law of robustness predicts that to obtain good robust models on ImageNet one would need at least $10^{10}-10^{11}$ parameters. This number is larger than the size of current neural networks trained robustly for this task, which sports between $10^8-10^9$ parameters. Thus, we arrive at the tantalizing possibility that robust models for ImageNet do not exist yet simply because we are a couple orders of magnitude off in the current scale of neural networks trained for this task. \subsection{Related work} Theorem \ref{thm:inf} is a direct follow-up to the conjectured law of robustness in \cite{bubeck2020law} for (arbitrarily weighted) two-layer neural networks with Gaussian data. Our result does not actually prove their conjecture, because we assume here polynomially bounded weights. While this assumption is reasonable from a practical perspective, it remains mathematically interesting to prove the full conjecture for the two-layer case. We prove however in Section \ref{sec:polyweights} that the polynomial weights assumption is necessary as soon as one considers three-layer neural networks. Let us also mention the \cite[Theorem 6.1]{gao2019convergence} which showed a lower bound $\Omega(nd)$ on the VC dimension of any function class which can robustly interpolate \emph{arbitrary} labels on \emph{all} well-separated input sets $(x_1,\dots,x_n)$. \newline In addition to \cite{madry2017towards}, several recent works have experimentally studied the relationship between a neural network scale and its achieved robustness, see e.g., \cite{novak2018sensitivity,Xie2020Intriguing, gowal2021uncovering}. It has been consistently reported that larger networks help tremendously for robustness, beyond what is typically seen for classical non-robust accuracy. We view our universal law of robustness as putting this empirical observation on a more solid footing: scale is actually {\em necessary} to achieve robustness. \newline The law of robustness setting is closely related to the interpolation setting: in the former case one considers models optimizing ``beyond the noise level", while in the latter case one studies models with perfect fit on the training data. The study of generalization in this interpolation regime has been a central focus of learning theory in the last few years (see e.g., \cite{Belkin15849, mei2020generalization, Bartlett30063, Nakkiran2020Deep}), as it seemingly contradicts classical theory about regularization. More broadly though, generalization remains a mysterious phenomon in deep learning, and the exact interplay between the law of robustness' setting (interpolation regime/worst-case robustness) and (robust) generalization error is a fantastic open problem. Interestingly, we note that one could potentially avoid the conclusion of the law of robustness (that is, that large models are necessary for robustness), with early stopping methods that could stop the optimization once the noise level is reached. In fact, this theoretically motivated suggestion has already been empirically tested and confirmed in the recent work \cite{pmlr-v119-rice20a}, showing again a close tie between the conclusions one can draw from the law of robustness and actual practical settings. \newline Classical lower bounds on the gradient of a function include Poincar{\'e} type inequalities, but they are of a qualitately different nature compared to the law of robustness lower bound. We recall that a measure $\mu$ on $\mathbb R^d$ satisfies a Poincar{\'e} inequality if for any function $f$, one has $\mathbb E^{\mu}[ \|\nabla f\|^2 ] \geq C \cdot \mathrm{Var}(f)$ (for some constant $C>0$). In our context, such a lower bound has essentially no consequence since the variance of a function interpolating a finite dataset could be exponentially small. In fact this is tight, as one easily use similar constructions to those in \cite{bubeck2020law} to show that one can interpolate with an exponentially small expected norm squared of the gradient (in particular it is crucial in the law of robustness to consider the Lipschitz constant, i.e., the supremum of the norm of the gradient). On the other hand, our isoperimetry assumption is related to a certain strenghtening of the Poincar{\'e} inequality known as log-Sobolov inequality (see e.g., \cite{Ledoux}). If the covariate measure satisfies only a Poincar{\'e} inequality, then we could prove a weaker law of robustness of the form $\mathrm{Lip} \gtrsim \frac{n\sqrt{d}}{p}$ (using for example the concentration result obtained in \cite{bobkov1997poincare}). We also note that a relation between high-dimensional phenomenon such as concentration and adversarial examples has been hypothesized before, such as in \cite{gilmer2018adversarial}. \subsection{Isoperimetry} Concentration of measure and isoperimetry are perhaps the most ubiquitous features of high-dimensional geometry. In short, they assert in many cases that Lipschitz functions on high-dimensional space concentrate tightly around their mean. Our result assumes that the distribution $\mu$ of the covariates $x_i$ satisfies such an inequality in the following sense. \begin{definition}\label{defn:iso} A probability measure $\mu$ on $\mathbb R^d$ satisfies $c$-isoperimetry if for any bounded $L$-Lipschitz $f:\mathbb R^d\to\mathbb R$, and any $t\geq 0$, \begin{equation} \label{eq:iso} \mathbb P[|f(x)- \mathbb E[f]|\geq t]\leq 2e^{-\frac{dt^2}{2cL^2}}. \end{equation} \end{definition} In general, if a scalar random variable $X$ satisfies $\mathbb P[|X|\geq t]\leq 2e^{-t^2/C}$ then we say $X$ is $C$-subgaussian. Hence isoperimetry states that the output of any Lipschitz function is $O(1)$-subgaussian under suitable rescaling. Distributions satisfying $O(1)$-isoperimetry include high dimensional Gaussians $\mu=\mathcal N\left(0,\frac{I_d}{d}\right)$ and uniform distributions on spheres and hypercubes (normalized to have diameter $1$). Isoperimetry also holds for mild perturbations of these idealized scenarios, including\footnote{The first two examples satisfy a logarithmic Sobolev inequality, which implies isoperimetry \cite[Proposition 2.3]{ledoux1999concentration}.}: \begin{itemize} \item The sum of a Gaussian and an independent random vector of small norm \cite{chen2021dimension}. \item Strongly log-concave measures in any normed space \cite[Proposition 3.1]{bobkov2000brunn}. \item Manifolds with positive Ricci curvature \cite[Theorem 2.2]{gromov1986isoperimetric}. \end{itemize} Due to the last condition above, we believe our results are realistic even under the \emph{manifold hypothesis} that high-dimensional data tends to lie on a lower-dimensional submanifold. This viewpoint on learning has been studied for decades, see e.g. \cite{hastie1989principal,kambhatla1993fast,roweis2000nonlinear,tenenbaum2000global,narayanan2010sample,fefferman2016testing}. We also note that our formal theorem (Theorem \ref{thm:main}) actually applies to distributions that can be written as a mixture of distributions satisfying isoperimetry. Let us also point out that from a technical perspective, our proof is not tied to the Euclidean norm and applies essentially whenever Definition~\ref{defn:iso} holds. The main difficulty in extending the law of robustness to e.g. the earth-mover distance seems to be identifying realistic cases which satisfy isoperimetry. \newline Our proofs will repeatedly use the following simple fact: \begin{proposition}\cite[Proposition 2.6.1]{vershynin2018high},\cite[Exercise 3.1]{van2014probability} \label{prop:versh} If $X_1,\dots,X_n$ are independent, $C$-subgaussian, with mean $0$, then $\frac{1}{\sqrt{n}}\sum_{i=1}^n X_i$ is $18C$-subgaussian. \end{proposition} \section{A finite approach to the law of robustness} For the function class of two-layer neural networks, \cite{bubeck2020law} investigated several approaches to prove the law of robustness. At a high level, the proof strategies there relied on various ways to measure how ``large" the set of two-layer neural networks can be (specifically, they tried a geometric approach based on relating to multi-index models, a statistical approach based on the Rademacher complexity, and an algebraic approach for the case of polynomial activations). \newline In this work we take here a different route: we shift the focus from the function class $\mathcal F$ to an {\em individual} function $f \in \mathcal F$. Namely, our proof starts by asking the following question: for a fixed function $f$, what is the probability that it would give a good approximate fit on the (random) data? For simplicity, consider for a moment the case where we require $f$ to actually interpolate the data (i.e., perfect fit), and say that $y_i$ are random $\pm1$ labels. The key insight is that isoperimetry implies that {\em either} the $0$-level set of $f$ {\em or} the $1$-level set of $f$ must have probability smaller than $\exp\left(- \frac{d}{\mathrm{Lip}(f)^2} \right)$. Thus, the probability that $f$ fits all the $n$ points is at most $\exp\left(- \frac{n d}{\mathrm{Lip}(f)^2} \right)$ so long as both labels $y_i\in \{-1,1\}$ actually appear a constant fraction of the time. In particular, using an union bound\footnote{In this informal argument we ignore the possibility that the labels $y_i$ are not well-balanced. Note that the probability of this rare event is not amplified by a union bound over $f\in\mathcal F$.}, for a finite function class $\mathcal F$ of size $N$ with $L$-Lipschitz functions, the probability that there exists a function $f \in \mathcal F$ fitting the data is at most \[ N \exp\left(- \frac{n d}{L^2} \right) = \exp\left(\log(N) - \frac{n d}{L^2} \right) \,. \] Thus we see that, if $L \ll \sqrt{\frac{n d}{\log(N)}}$, then the probability of finding a fitting function in $\mathcal F$ is very small. This basically concludes the proof, since via a standard discretization argument, for a smoothly parametrized family with $p$ (bounded) parameters one expects $\log(N) = \tilde{O}(p)$. \newline We now give the formal proof, which applies in particular to approximate fit rather than exact fit in the argument above. The only difference is that we will identify a well-chosen subgaussian random variable in the problem. We start with the finite function class case: \begin{theorem} \label{thm:finite} Let $(x_i,y_i)$ be i.i.d. input-output pairs in $\mathbb R^d \times [-1,1]$ such that: \begin{enumerate} \item The distribution $\mu$ of the covariates $x_i$ can be written as $\mu = \sum_{\ell=1}^k \alpha_{\ell} \mu_{\ell}$, where each $\mu_{\ell}$ satisfies $c$-isoperimetry and $\alpha_{\ell} \geq 0, \sum_{\ell=1}^k \alpha_{\ell}=1$. \item The expected conditional variance of the output is strictly positive, denoted $\sigma^2 := \mathbb E^{\mu}[Var[y|x]] > 0$. \end{enumerate} Then one has: \begin{align*} & \mathbb{P} \left( \exists f \in \mathcal F : \frac{1}{n} \sum_{i=1}^n (y_i - f(x_i))^2 \leq \sigma^2 - \epsilon \right) \\ & \leq 4 k \exp\left(- \frac{n \epsilon^2}{8^3 k} \right) + 2 \exp\left(\log(|\mathcal F|) - \frac{\epsilon^2 n d}{9^4 c L^2} \right) \,. \end{align*} \end{theorem} We start with a lemma showing that, to optimize beyond the noise level one must necessarily correlate with the noise part of the labels. In what follows we denote $g(x) = \mathbb E[y | x]$ for the target function, and $z_i = y_i - g(x_i)$ for the noise part of the observed labels (namely $y_i$ is the sum of the target function $g(x_i)$ and the noise term $z_i$). \begin{lemma} \label{lem:corr} One has \[ \mathbb{P} \left( \exists f \in \mathcal F : \frac{1}{n} \sum_{i=1}^n (y_i - f(x_i))^2 \leq \sigma^2 - \epsilon \right) \leq 2 \exp\left(- \frac{n \epsilon^2}{8^3} \right) + \mathbb{P} \left( \exists f \in \mathcal F : \frac{1}{n} \sum_{i=1}^n f(x_i) z_i \geq \frac{\epsilon}{4} \right) \,. \] \end{lemma} \begin{proof} The sequence $(z_i^2)$ is i.i.d., with mean $\sigma^2$, and such that $|z_i|^2 \leq 4$. Thus Hoeffding's inequality yields: \begin{equation} \label{eq:conc1} \mathbb{P} \left( \frac{1}{n} \sum_{i=1}^n z_i^2 \leq \sigma^2 - \frac{\epsilon}{6} \right) \leq \exp\left(- \frac{n \epsilon^2}{8^3} \right) \,. \end{equation} On the other hand the sequence $(z_i g(x_i))$ is i.i.d., with mean $0$ (since $\mathbb E[z_i | x_i] = 0$), and such that $|z_i g(x_i)| \leq 2$. Thus Hoeffding's inequality yields: \begin{equation} \label{eq:conc2} \mathbb{P} \left( \frac{1}{n} \sum_{i=1}^n z_i g(x_i) \leq - \frac{\epsilon}{6} \right) \leq \exp\left(- \frac{n \epsilon^2}{8^3} \right) \,. \end{equation} Let us write $Z= \frac{1}{\sqrt{n}}(z_1,\hdots,z_n), G= \frac{1}{\sqrt{n}}(g(x_1),\hdots, g(x_n))$, and $F= \frac{1}{\sqrt{n}}(f(x_1),\hdots,f(x_n))$. We claim that if $\|Z\|^2 \geq \sigma^2 - \frac{\epsilon}{6}$ and $\langle Z, G\rangle \geq - \frac{\epsilon}{6}$, then for any $f \in \mathcal F$ one has \[ \|G + Z - F\|^2 \leq \sigma^2 - \epsilon \, \Rightarrow \, \langle F, Z \rangle \geq \frac{\epsilon}{4} \,. \] This claim together with \eqref{eq:conc1} and \eqref{eq:conc2} conclude the proof. On the other hand the claim itself directly follows from: \[ \sigma^2 - \epsilon \geq \|G + Z - F\|^2 = \|Z + G - F\|^2 = \|Z\|^2 + 2 \langle Z, G-F\rangle + \|G- F\|^2 \geq \sigma^2 - \frac{\epsilon}{2} - 2 \langle Z, F \rangle \,. \] \end{proof} We can now proceed to the proof of Theorem \ref{thm:finite}: \begin{proof} First note that without loss of generality we can assume that the range of any function in $\mathcal F$ is included in $[-1,1]$ (indeed clipping the values improves both the fit to any $y \in [-1,1]$ and the Lipschitz constant). We also assume wlog that all functions in $\mathcal F$ are $L$-Lipschitz. For clarity let us start with the case $k=1$. By the isoperimetry assumption we have that $\sqrt{\frac{d}{c}}\frac{f(x_i)-\mathbb{E}[f]}{L}$ is $1$-subgaussian. Since $|z_i| \leq 2$, we also have that $\sqrt{\frac{d}{c}}\frac{(f(x_i)-\mathbb E[f]) z_i}{L}$ is $4$-subgaussian. Moreover, the latter random variable has zero-mean since $\mathbb E[z|x] = 0$. Thus by Proposition \ref{prop:versh} we have: \[ \mathbb{P} \left( \sqrt{\frac{d}{c n L^2}} \sum_{i=1}^n (f(x_i)-\mathbb E[f]) z_i \geq t \right) \leq 2 \exp\left( - (t/9)^2 \right) \,. \] We rewrite the above as: \begin{equation} \label{eq:conc3} \mathbb{P} \left(\frac{1}{n} \sum_{i=1}^n (f(x_i)-\mathbb E[f]) z_i \geq \frac{\epsilon}{8} \right) \leq 2 \exp\left( - \frac{\epsilon^2 n d}{9^4 c L^2} \right) \,. \end{equation} Since we assumed that the range of the functions is in $[-1,1]$ we have $\mathbb E[f] \in [-1,1]$ and hence: \begin{equation} \label{eq:conc4} \mathbb{P} \left(\exists f \in \mathcal F : \frac{1}{n} \sum_{i=1}^n \mathbb E[f] z_i \geq \frac{\epsilon}{8} \right) \leq \mathbb{P} \left(\left| \frac{1}{n} \sum_{i =1}^n z_i \right| \geq \frac{\epsilon}{8} \right) \,. \end{equation} (This step is the analog of requiring the labels $y_i$ to be well-balanced in the example of perfect interpolation.) By Hoeffding's inequality, the above quantity is smaller than $2 \exp( - n \epsilon^2 / 8^3)$ (recall that $|z_i| \leq 2$). Thus we obtain with an union bound: \begin{eqnarray*} \mathbb{P} \left(\exists f \in \mathcal F : \frac{1}{n} \sum_{i=1}^n f(x_i) z_i \geq \frac{\epsilon}{4} \right) & \leq & |\mathcal F| \cdot \mathbb{P} \left(\frac{1}{n} \sum_{i=1}^n (f(x_i)-\mathbb E[f]) z_i \geq \frac{\epsilon}{8} \right) + \mathbb{P} \left(\left| \frac{1}{n} \sum_{i =1}^n z_i \right| \geq \frac{\epsilon}{8} \right) \\ & \leq & 2 |\mathcal F| \cdot \exp\left( - \frac{\epsilon^2 n d}{9^4 c L^2} \right) + 2 \exp\left( - \frac{n \epsilon^2}{8^3} \right) \,. \end{eqnarray*} Together with Lemma \ref{lem:corr} this concludes the proof for $k=1$. \newline We now turn to the case $k >1$. We first sample the mixture component $\ell_i \in [k]$ for each data point $i \in [n]$, and we now reason conditioned on these mixture components. Let $S_{\ell} \subset [n]$ be the set of data points sampled from mixture component $\ell \in [k]$, that is $x_i, i \in S_{\ell},$ is i.i.d. from $\mu_{\ell}$. We now have that $\sqrt{\frac{d}{c}}\frac{f(x_i)-\mathbb{E}^{\mu_{\ell_i}}[f]}{L}$ is $1$-subgaussian (notice that the only difference is that now we need to center by $\mathbb{E}^{\mu_{\ell_i}}[f]$, which depends on the mixture component). In particular using the same reasoning as for \eqref{eq:conc4} we obtain (crucially note that Proposition \ref{prop:versh} does not require the random variables to be identically distributed): \begin{equation} \label{eq:conc5} \mathbb{P} \left(\frac{1}{n} \sum_{i=1}^n (f(x_i)-\mathbb E^{\mu_{\ell_i}}[f]) z_i \geq \frac{\epsilon}{8} \right) \leq 2 \exp\left( - \frac{\epsilon^2 n d}{9^4 c L^2} \right) \,. \end{equation} Next we want to appropriately modify \eqref{eq:conc4}. To do so note that: \[ \max_{m_1, \hdots, m_k \in [-1,1]} \sum_{i=1}^n m_{\ell_i} z_i = \sum_{\ell=1}^k \left| \sum_{i \in S_{\ell}} z_i \right| \,, \] so that we can rewrite \eqref{eq:conc4} as: \[ \mathbb{P} \left(\exists f \in \mathcal F : \frac{1}{n} \sum_{i=1}^n \mathbb E^{\mu_{\ell_i}}[f] z_i \geq \frac{\epsilon}{8} \right) \leq \mathbb{P} \left(\frac{1}{n} \sum_{\ell=1}^k \left| \sum_{i \in S_{\ell}} z_i \right| \geq \frac{\epsilon}{8} \right) \,. \] Now note that $\sum_{\ell=1}^k \sqrt{|S_{\ell}|} \leq \sqrt{n k}$ and thus we have: \[ \mathbb{P} \left(\frac{1}{n} \sum_{\ell=1}^k \left| \sum_{i \in S_{\ell}} z_i \right| \geq \frac{\epsilon}{8} \right) \leq \mathbb{P} \left(\sum_{\ell=1}^k \left| \sum_{i \in S_{\ell}} z_i \right| \geq \frac{\epsilon}{8} \sqrt{\frac{n}{k}} \sum_{\ell=1}^k \sqrt{|S_{\ell}|} \right) \leq \sum_{\ell=1}^k \mathbb{P} \left( \left| \sum_{i \in S_{\ell}} z_i \right| \geq \frac{\epsilon}{8} \sqrt{\frac{n}{k}} \sqrt{|S_{\ell}|} \right) \,. \] Finally by Hoeffding's inequality, we have for any $\ell \in [k]$, $\mathbb{P} \left( \left| \sum_{i \in S_{\ell}} z_i \right| \geq t \sqrt{|S_{\ell}|} \right) \leq 2 \exp \left( - \frac{t^2}{8} \right)$, and thus the last display is bounded from above by $2 k \exp \left( - \frac{n \epsilon^2}{8^3 k} \right)$. The proof can now be concluded as in the case $k=1$. \end{proof} Finally we can now state and prove the formal version of the informal Theorem~\ref{thm:inf} from the introduction. \begin{theorem}\label{thm:main} Let $\mathcal F$ be a class of functions from $\mathbb R^d\to \mathbb R$ and let $(x_i,y_i)_{i=1}^n$ be i.i.d. input-output pairs in $\mathbb R^d\times [-1,1]$. Fix $\epsilon, \delta \in (0,1)$. Assume that: \begin{enumerate} \item The function class can be written as $\mathcal F = \{ f_{{\boldsymbol{w}}}, {\boldsymbol{w}} \in \mathcal W\}$ with $\mathcal{W} \subset \mathbb R^p$, $\mathrm{diam}(\mathcal W) \leq W$ and for any ${\boldsymbol{w}}_1, {\boldsymbol{w}}_2 \in \mathcal W$, \[||f_{{\boldsymbol{w}}_1}-f_{{\boldsymbol{w}}_2}||_{\infty}\leq J||{\boldsymbol{w}}_1-{\boldsymbol{w}}_2||.\] \item The distribution $\mu$ of the covariates $x_i$ can be written as $\mu = \sum_{\ell=1}^k \alpha_{\ell} \mu_{\ell}$, where each $\mu_{\ell}$ satisfies $c$-isoperimetry, $\alpha_{\ell} \geq 0, \sum_{\ell=1}^k \alpha_{\ell}=1$, and $k$ is such that $9^4 k \log(8k / \delta) \leq n \epsilon^2$. \item The expected conditional variance of the output is strictly positive, denoted $\sigma^2 := \mathbb E^{\mu}[Var[y|x]] > 0$. \end{enumerate} Then, with probability at least $1-\delta$ with respect to the sampling of the data, one has simultaneously for all $f \in \mathcal F$: \[ \frac{1}{n}\sum_{i=1}^n (f(x_i)-y_i)^2 \leq \sigma^2-\epsilon \,\, \Rightarrow \,\, \mathrm{Lip}(f) \geq \frac{\epsilon}{2^9 \sqrt{c}} \sqrt{\frac{nd}{p \log(60 W J \epsilon^{-1}) + \log(4/\delta)}} \,. \] \end{theorem} \begin{proof} Define $\mathcal{W}_L\subseteq\mathcal W$ by $\mathcal{W}_L=\{{\boldsymbol{w}}\in \mathcal W:\mathrm{Lip}(f_{{\boldsymbol{w}}})\leq L\}.$ Denote $\mathcal{W}_{L,\epsilon}$ for an $\frac{\epsilon}{6J}$-net of $\mathcal{W}_L$. We have in particular $|\mathcal W_{\epsilon}|\leq (60WJ\epsilon^{-1})^{p}$. We apply Theorem \ref{thm:finite} to $\mathcal F_{L,\epsilon} = \{f_{{\boldsymbol{w}}}, {\boldsymbol{w}} \in \mathcal{W}_{L,\epsilon}\}$: \begin{align*} & \mathbb{P} \left( \exists f \in \mathcal F_{L,\epsilon} : \frac{1}{n} \sum_{i=1}^n (y_i - f(x_i))^2 \leq \sigma^2 - \frac{\epsilon}{2} \text{ and } \mathrm{Lip}(f) \leq 2 L \right) \\ & \leq 4 k \exp\left(- \frac{n \epsilon^2}{9^4 k} \right) + 2 \exp\left(p \log(60WJ\epsilon^{-1}) - \frac{\epsilon^2 n d}{8^6 c L^2} \right) \,. \end{align*} Observe that if $\|f-g\|_{\infty} \leq \frac{\epsilon}{6}$ and $\|y\|_{\infty},\|f\|_{\infty},\|g\|_{\infty}\leq 1$, then $\frac{1}{n} \sum_{i=1}^n (y_i - f(x_i))^2 \leq \frac{\epsilon}{2} + \frac{1}{n} \sum_{i=1}^n (y_i - g(x_i))^2$. (We may again assume without loss of generality that all functions in $\mathcal F$ map to $[-1,1]$.) Thus we obtain for any $L>0$: \begin{align*} & \mathbb{P} \left( \exists f \in \mathcal F : \frac{1}{n} \sum_{i=1}^n (y_i - f(x_i))^2 \leq \sigma^2 - \epsilon \text{ and } \mathrm{Lip}(f) \leq L \right) \\ & \leq 4 k \exp\left(- \frac{n \epsilon^2}{9^4 k} \right) + 2 \exp\left(p \log(60WJ\epsilon^{-1}) - \frac{\epsilon^2 n d}{8^6 c L^2} \right) \,. \end{align*} The first assumption ensures that for any ${\boldsymbol{w}}\in \mathcal{W}_L$, there is ${\boldsymbol{w}}'\in\mathcal{W}_{L,\epsilon}$ with $\|f_{{\boldsymbol{w}}}-f_{{\boldsymbol{w}}'}\|_{\infty}\leq\frac{\epsilon}{6}$. The second assumption shows the probability just above is at most $\delta$ when $L=\frac{\epsilon}{2^9 \sqrt{c}} \sqrt{\frac{nd}{p \log(60 W J \epsilon^{-1}) + \log(4/\delta)}}$. This concludes the proof. \end{proof} \section{Deep neural networks} \label{sec:continuous} We now specialize the law of robustness (Theorem \ref{thm:main}) to multi-layer neural networks. We consider a rather general class of depth $D$ neural networks described as follows. First, we require that the neurons are partitioned into layers $\mathcal L_1,\dots,\mathcal L_D$, and that all connections are from $\mathcal L_i\to\mathcal L_j$ for some $i<j$. This includes the basic feed-forward case in which only connections $\mathcal L_i\to\mathcal L_{i+1}$ are used as well as more general skip connections. We specify (in the natural way) a neural network by matrices $W_j$ of shape $|\mathcal L_j|\times \sum_{i<j}|\mathcal L_i|$ for each $1\leq j \leq D$, as well as $1$-Lipschitz non-linearities $\sigma_{j,\ell}$ and scalar biases $b_{j,\ell}$ for each $(j,\ell)$ satisfying $\ell\in |\mathcal L_j|$. We use fixed non-linearities $\sigma_{j,\ell}$ as well as a fixed architecture, in the sense that each matrix entry $W_{j}[k,\ell]$ is either always $0$ or else it is variable (and similarly for the bias terms). \newline To match the notation of Theorem \ref{thm:main}, we identify the parametrization in terms of the matrices $(W_{j})$ and bias terms $(b_{j,\ell})$ to a single $p$-dimensional vector ${\boldsymbol{w}}$ as follows. A variable matrix entry $W_{j}[k,\ell]$ is set to $w_{a(j,k,\ell)}$ for some fixed index $a(j,k,\ell)\in [p]$, and a variable bias term $b_{j,\ell}$ is set to $w_{a(j,\ell)}$ for some $a(j,\ell)\in [p]$. Thus we now have a parametrization ${\boldsymbol{w}} \in \mathbb R^p \mapsto f_{{\boldsymbol{w}}}$ where $f_{{\boldsymbol{w}}}$ is the neural network represented by the parameter vector ${\boldsymbol{w}}$. Importantly, note that our formulation allows for weight sharing (in the sense that a shared weight is counted only as a single parameter). For example, this is important to obtain an accurate count of the number of parameters in convolutional architectures. \newline In order to apply Theorem \ref{thm:main} to this class of functions we need to estimate the Lipschitz constant of the parametrization ${\boldsymbol{w}} \mapsto f_{{\boldsymbol{w}}}$. To do this we introduce three more quantities. First, we shall assume that all the parameters are bounded in magnitude by $W$, that is we consider the set of neural networks parametrized by ${\boldsymbol{w}} \in [-W,W]^p$. Next, for the architecture under consideration, denote $Q$ for the maximum number of matrix entries/bias terms that are tied to a single parameter $w_a$ for some $a \in [p]$. Finally we define \[ B({\boldsymbol{w}}) = \prod_{j\in [D]} \max(\|W_{j}\|_{op},1). \] Observe that $B({\boldsymbol{w}})$ is an upper bound on the Lipschitz constant of the network itself, i.e., the map $x \mapsto f_{{\boldsymbol{w}}}(x)$. It turns out that a uniform control on it also controls the Lipschitz constant of the {\em parametrization} ${\boldsymbol{w}} \mapsto f_{{\boldsymbol{w}}}$. Namely we have the following lemma: \begin{lemma}\label{lem:NNlip} Let $x \in \mathbb R^d$ such that $\|x\| \leq R$, and ${\boldsymbol{w}}_1, {\boldsymbol{w}}_2 \in \mathbb R^p$ such that $B({{\boldsymbol{w}}_1}),B({{\boldsymbol{w}}_2})\leq\overline{B}$. Then one has \[ |f_{{\boldsymbol{w}}_1}(x)-f_{{\boldsymbol{w}}_2}(x)|\leq \overline{B}^2Q R\sqrt{p} \|{\boldsymbol{w}}_1-{\boldsymbol{w}}_2\| \,. \] Moreover for any ${\boldsymbol{w}} \in [-W,W]^p$ with $W\geq 1$, one has \[ B({\boldsymbol{w}}) \leq (W \sqrt{pQ})^D. \] \end{lemma} \begin{proof} Fix an input $x$ and define $g_x$ by $g_x({\boldsymbol{w}})= f_{{\boldsymbol{w}}}(x)$. A standard gradient calculation for multi-layer neural networks directly shows that $\|\nabla g_x({\boldsymbol{w}})\|_{\infty} \leq B({\boldsymbol{w}}) Q R$ so that $\|\nabla g_x({\boldsymbol{w}})\| \leq B({\boldsymbol{w}}) Q R \sqrt{p}$. Since the matrix operator norm is convex (and nonnegative) it follows that $B({\boldsymbol{w}}) \leq B({\boldsymbol{w}}_1)B({\boldsymbol{w}}_2)\leq \overline{B}^2$ on the entire segment $[{\boldsymbol{w}}_1,{\boldsymbol{w}}_2]$ by multiplying over layers. Thus $\|\nabla g_x({\boldsymbol{w}})\| \leq \overline{B}^2 Q R \sqrt{p}$ on that segment, which concludes the proof of the first claimed inequality. The second claimed inequality follows directly from $\|W_{j}\|_{op}\leq \|W_{j}\|_2 \leq W\sqrt{p Q}$. \end{proof} Lemma \ref{lem:NNlip} shows that when applying Theorem \ref{thm:main} to our class of neural networks one can always take $J= R (W Q p)^{D}$ (assuming that the covariate measure $\mu$ is supported on the ball of radius $R$). Thus in this case the law of robustness (under the assumptions of Theorem \ref{thm:main}) directly states that with high probability, any neural network in our class that fits the training data well below the noise level must also have: \begin{equation} \label{eq:withdepth} \mathrm{Lip}(f) \geq \tilde{\Omega} \left(\sqrt{\frac{n d}{D p}} \right) \,, \end{equation} where $\tilde{\Omega}$ hides logarithmic factors in $W, p, R, Q$, and the probability of error $\delta$. Thus we see that the law of robustness, namely that the number of parameters should be at least $n d$ for a smooth model with low training error, remains intact for constant depth neural networks. If taken at face value, the lower bound \eqref{eq:withdepth} suggests that it is better in practice to distribute the parameters towards {\em depth} rather than {\em width}, since the lower bound is decreasing with $D$. On the other hand, we note that \eqref{eq:withdepth} can be strengthened to: \begin{equation} \label{eq:withoutdepth} \mathrm{Lip}(f) \geq \tilde{\Omega} \left( \sqrt{\frac{n d}{p \log(\overline{B})}} \right) \,, \end{equation} for the class of neural networks such that $B({\boldsymbol{w}}) \leq \overline{B}$. In other words the dependency on the depth all but disappears by simply assuming that the quantity $B({\boldsymbol{w}})$ (a natural upper bound on the Lipschitz constant of the network) is polynomially controlled. Interestingly many works have suggested to keep $B({\boldsymbol{w}})$ under control, either for regularization purpose (for example \cite{bartlett2017spectrally} relates $B({\boldsymbol{w}})$ to the Rademacher complexity of multi-layer neural networks) or to simply control gradient explosion during training, see e.g., \cite{arjovsky2016unitary,cisse2017parseval,mhammedi2017efficient,miyato2018spectral, jiang2018computation,yoshida2017spectral}. Moreover, in addition to being well-motivated in practice, the assumption that $\overline{B}$ is polynomially controlled seems also somewhat unavoidable in theory, since $B({\boldsymbol{w}})$ is an {\em upper bound} on the Lipschitz constant $\mathrm{Lip}(f_{{\boldsymbol{w}}})$. Thus a theoretical construction showing that the lower bound in \eqref{eq:withdepth} is tight (at some large depth $D$) would necessarily need to have an exponential gap between $\mathrm{Lip}(f_{{\boldsymbol{w}}})$ and $B({\boldsymbol{w}})$. We are not aware of any such example, and it would be interesting to fully elucidate the role of depth in the law of robustness (particularly if it could give recommendation on how to best distribute parameters in a neural network). \section{Generalization Perspective} The law of robustness can be phrased in a slightly stronger way, as a generalization bound for classes of Lipschitz functions based on data-dependent Rademacher complexity. In particular, this perspective applies to any Lipschitz loss function, whereas our analysis in the main text was specific to the squared loss. We define the data-dependent Rademacher complexity $\text{Rad}_{n,\mu}(\mathcal F)$ by \begin{equation} \text{Rad}_{n,\mu}(\mathcal F)=\frac{1}{n}\mathbb E^{\sigma_i,x_i}\left[\sup_{f\in \mathcal F}\left|\sum_{i=1}^{n} \sigma_i f(x_i)\right| \right] \end{equation} where the values $(\sigma_i)_{i\in [n]}$ are i.i.d. symmetric Rademacher variables in $\{-1,1\}$ while the values $(x_i)_{i\in [n]}$ are i.i.d. samples from $\mu$. \begin{lemma}\label{lem:gen1} Suppose $\mu=\sum_{i=1}^k \alpha_i\mu_i$ is a mixture of $c$-isoperimetric distributions. For finite $\mathcal F$ consisting of $L$-Lipschitz $f$ with $|f(x)|\leq 1$ for all $(f,x)\in \mathcal F\times \mathbb R^d$, we have \begin{equation} \text{Rad}_{n,\mu}(\mathcal F)\leq O\left(\max\left(\sqrt{\frac{k}{n}},L\sqrt{\frac{c\log(|\mathcal F|)}{nd}}\right)\right) .\end{equation} \end{lemma} The proof is identical to that of Theorem~\ref{thm:finite}. Note that $\text{Rad}_{n,\mu}(\mathcal F)$ simply measures the ability of functions in $\mathcal F$ to correlate with random noise. Using standard machinery it implies the following generalization bound: \begin{corollary} For any loss function $\ell(t,y)$ which is bounded and $1$-Lipschitz in its first argument and any $\delta\in [0,1]$, in the setting of Lemma~\ref{lem:gen1} we have with probability at least $1-\delta$ the uniform convergence bound: \[\sup_{f\in \mathcal F}\left|\mathbb E^{(x,y)\sim \mu}[\ell(f(x),y)]-\frac{1}{n}\sum_{i=1}^n \ell(f(x_i),y_i)\right|\leq O\left(\max\left(\sqrt{\frac{k}{n}},L\sqrt{\frac{c\log(|\mathcal F|)}{nd}},\sqrt{\frac{\log(1/\delta)}{n}}\right)\right) \,.\] \end{corollary} \begin{proof} Using McDiarmid's concentration inequality it is enough to bound the left hand side in expectation over $(x_i, y_i)$. Using the symmetrization trick, one reduces this task to upper bound \[ \mathbb E^{x_i,y_i,\sigma_i} \sup_{f \in \mathcal{F}} \frac{1}{n}\sum_{i=1}^n \sigma_i \ell(f(x_i),y_i) \,. \] Fixing the pairs $(x_i,y_i)$ and using the contraction lemma (see e.g., \cite[Theorem 26.9]{shalev2014understanding}) the above quantity is upper bounded by $\text{Rad}_{n,\mu}(\mathcal{F})$ which concludes the proof. \end{proof} Of course, one can again use an $\epsilon$-net to obtain an analogous result for continuously parametrized function classes. The law of robustness, now for a general loss function, follows as a corollary (the argument is similar to [Proposition 1, \cite{BELM20}]). Let us point out that many papers have studied the Rademacher complexity of function classes such as neural networks (see e.g. \cite{bartlett2017spectrally}, or \cite{yin2019rademacher} in the context of adversarial examples). The new feature of our result is that isoperimetry of the covariates yields improved generalization guarantees. \section{Necessity of Polynomially Bounded Weights} \label{sec:polyweights} In \cite{bubeck2020law} it was conjectured that the law of robustness should hold for the class of {\em all} two-layer neural networks. In this paper we prove that in fact it holds for arbitrary smoothly parametrized function classes, as long as the parameters are of size at most polynomial. In this section we demonstrate that this polynomial size restriction is necessary for bounded depth neural networks. First we note that {\em some} restriction on the size of the parameters is certainly necessary in the most general case. Indeed one can build a single-parameter family, where the single real parameter is used to approximately encode all Lipschitz functions from a compact set in $\mathbb R^d$ to $[-1,1]$, simply by brute-force enumeration. In particular no tradeoff between number of parameters and attainable Lipschitz constant would exist for this function class. Showing a counter-example to the law of robustness with unbounded parameters and ``reasonable" function classes is slightly harder. Here we build a three-layer neural network, with a single fixed nonlinearity $\sigma : \mathbb R \to \mathbb R$, but the latter is rather complicated and we do not know how to describe it explicitly (it is based on the Kolmogorov-Arnold theorem). It would be interesting to give similar constructions using other function classes such as ReLU networks. \begin{theorem}\label{thm:KA} For each $d\in\mathbb Z^+$ there is a continuous function $\sigma:\mathbb R\to \mathbb R$ and a sequence $(b_{\ell})_{\ell\leq 2^{2^d}}$ such that the following holds. The function $\Phi_a$ defined by \begin{equation}\label{eq:thmka}\Phi_a(x)=\sum_{\ell=1}^{2^{2^d}}\sigma(a-\ell)\sum_{i=1}^{2d}\sigma\left(b_{\ell}+\sum_{j=1}^d\sigma(x_j+b_{\ell})\right),\quad\quad |a|\leq 2^{2^d}\end{equation} is always $O(d^{3/2})$-Lipschitz, and the parametrization $a\to\Phi_a$ is $1$-Lipschitz. Moreover for $n\leq \frac{2^d}{100}$, given i.i.d. uniform points $x_1,\dots,x_n\in\mathbb S^{d-1}$ and random labels $y_1,\dots,y_n\in\{-1,1\}$, with probability $1-e^{-\Omega(d)}$ there exists $\ell\in [2^{2^d}]$ such that $\Phi_{\ell}(x_i)=y_i$ for at least $\frac{3n}{4}$ of the values $i\in [n]$. \end{theorem} \begin{proof} For each coordinate $i\in [d]$, define the slab $\term{slab}_i=\{x\in \mathbb S^{d-1}:|x_i|\leq \frac{1}{100d^{3/2}}\}$ and set $\term{slab}=\bigcup_{i\in [d]}\term{slab}_i$. Then it is not difficult to see that $\mu(\term{slab})\leq \frac{1}{10}$. We partition $\mathbb S^{d-1}\backslash\term{slab}$ into its $2^d$ connected components, which are characterized by their sign patterns in $\{-1,1\}^d$; this defines a piece-wise constant function $\gamma:\mathbb S^{d-1}\backslash\term{slab}\to \{-1,1\}^d$. If we sample the points $x_1,\dots,x_n$ sequentially, each point has probability at least $\frac{4}{5}$ to be in a new cell - this implies that with probability $1-e^{-\Omega(n)}$, at least $\frac{3n}{4}$ are in a unique cell. It therefore suffices to give a construction that achieves $\Phi(x_i)=y_i$ for all $x_i\notin \term{slab}$ such that $\gamma(x_i)\neq \gamma(x_j)$ for all $j\in [n]\backslash \{i\}$. We do this now. For each of the $2^{2^d}$ functions $g_{\ell}:\{-1,1\}^d \to \{-1,1\}$, we now obtain the partial function $\tilde h_{\ell}=g_{\ell}\circ \gamma:\mathbb S^{d-1}\backslash\term{slab}\to \{-1,1\}.$ By the Kirszbraun extension theorem, $\tilde h_{\ell}$ extends to an $O(d^{3/2})$-Lipschitz function $h_{\ell}:\mathbb S^{d-1}\to [-1,1]$ on the whole sphere. The Kolmogorov-Arnold theorem guarantees the existence of an exact representation \begin{equation}\label{eq:KA0}\Phi_{\ell}(x)=\sum_{i=1}^{2d}\sigma_{\ell}\left(\sum_{j=1}^d\sigma_{\ell}(x_j)\right)\end{equation} of $h_{\ell}$ by a two-layer neural network for some continuous function $\sigma_{\ell}:\mathbb R\to\mathbb R$ depending on $\ell$. It suffices to give a single neural network capable of computing all functions $(\Phi_{\ell})_{\ell=1}^{2^{2^d}}$. We extend the definition of $\Phi_{a}$ to any $a\in\mathbb R$ via: \begin{equation}\label{eq:KA}\Phi_a(x)=\sum_{\ell=1}^{2^{2^d}}\sigma(a-\ell)\Phi_{\ell}(x)\end{equation} where $\sigma:\mathbb R\to\mathbb R$ satisfies $\sigma(x)=(1-|x|)_+$ for $|x|\leq 2^{2^d}$. This ensures that \eqref{eq:KA} extends \eqref{eq:KA0}. To express $\Phi_a$ using only a single non-linearity, we prescribe further values for $\sigma$. Let \[U=2^{2^d}+d\cdot\max_{x\in [-1,1],\ell\in [2^{2^d}]}|\sigma_{\ell}(x)|\] so that $\left|\sum_{j=1}^d\sigma_{\ell}(x_j)\right|\leq U$ for all $x\in\mathbb S^{d-1}$. Define real numbers $b_{\ell}=10\ell U+2^{2^d}$ for $\ell\in [2^{2^d}]$ and for all $|x|\leq U$ set \[\sigma(x+b_{\ell})=\sigma_{\ell}(x).\] Due to the separation of the values $b_{\ell}$ such a function $\sigma$ certainly exists. Then we have \[\Phi_{\ell}(x)=\sum_{i=1}^{2d}\sigma\left(b_{\ell}+\sum_{j=1}^d\sigma(x_j+b_{\ell})\right).\] Therefore with this choice of non-linearity $\sigma$ and (data-independent) constants $b_{\ell}$, some function $\Phi_{\ell}$ fits at least $\frac{3n}{4}$ of the $n$ data points with high probability, and the functions $\Phi_a$ are parametrized in a $1$-Lipschitz way by a single real number $a\leq 2^{2^d}$. \end{proof} \begin{remark}The representation~\eqref{eq:thmka} is a three-layer neural network because the $\sigma(a-\ell)$ terms are just matrix entries for the final layer. \end{remark} \begin{remark} The construction above can be made more efficient, using only $O(n\cdot 2^{n})$ uniformly random functions $g_{\ell}:\{-1,1\}^d\to \{-1,1\}$ instead of all $2^{2^{\ell}}$. Indeed by the coupon collector problem, this results in all functions from $\{\gamma(x_i):i\in [n]\}\to \{-1,1\}$ being expressable as the restriction of some $g_{\ell}$, with high probability. \end{remark} \bibliographystyle{alpha}

https://arxiv.org/abs/1311.2984

Edge-maximality of power graphs of finite cyclic groups

We show that among all finite groups of any given order, the cyclic group of that order has the maximum number of edges in its power graph. Contains corrections to published version.

\section{Introduction} In this paper, we resolve in the affirmative a conjecture of Mirzargar et al. \cite[Conjecture 2]{m} concerning the number of edges in the power graph of a finite group. Motivated by the work of Kelarev and Quinn \cite{KQ1,KQ2,k,KQS}, Chakrabarty, Ghosh, and Sen \cite{ch} introduced undirected power graphs to study semigroups and groups. Other relevant work includes \cite{c1,c2}. The reader is encouraged to see \cite{AbKeCh} which surveys the literature to date. \begin{defn} Let $G$ be a finite group. Let $\langle g \rangle$ denote the cyclic subgroup of $G$ generated by $g\in G$. \begin{enumerate} \item The {\em directed power graph} $\overrightarrow{{\mathcal{P}}}(G)$ of $G$ is the directed graph with vertex set $G$ and directed edge set $\overrightarrow{E}(G) = \{(g,h)\,|\, g, h\in G,\, h\in\langle g \rangle-\{g\}\}$. That is, there is an edge from one group element to a second whenever the second is a positive power of the first and distinct from the first. % \item \label{defn:first} The {\it undirected power graph} (or {\em power graph} ) ${{\mathcal{P}}}(G)$ of $G$ is the undirected graph with vertex set $G$ and edge set $E(G)=\{ \{g,h\} \, | \, (g,h)\in \overrightarrow{E}(G) \hbox{ or } (h,g)\in\overrightarrow{E}(G) \}$. That is, two distinct group elements are adjacent whenever one of them is a positive power of the other. \end{enumerate} \end{defn} We recall the following property of directed power graphs of cyclic groups. \begin{thm} \cite[Main Theorem]{a} \label{cor:maximume} Among all finite groups of a given order, the cyclic group of that order has the maximum number of edges in its directed power graph. \end{thm} Our main theorem, which resolves \cite[Conjecture 2]{m}, is the undirected analog of Theorem \ref{cor:maximume}. \begin{thm} \label{thm:main} Among all finite groups of a given order, the cyclic group of that order has the maximum number of edges in its power graph. \end{thm} Theorem \ref{thm:main} gives that (ii) implies (iii) in the following. The others are trivial. \begin{cor} \label{cor:isomisom} Let $G$ be a finite group. Then the following are equivalent for all $n\geq 1$.\begin{enumerate} \item ${\mathcal{P}}(G)\cong{\mathcal{P}}(\mathbb{Z}_n)$. \item $|E(G)|=|E(\mathbb{Z}_n)|$. \item $G\cong \mathbb{Z}_n$. \end{enumerate} \end{cor} One special case of Theorem \ref{thm:main} is already known. \begin{thm}\cite[Theorem 2.12]{ch} \label{lem:PGcyclicprimepower1} A finite group has a complete power graph if and only if it is cyclic and has prime power order. \end{thm} \section{Edges in power graphs} Let $G$ be a finite group. For $g \in G$, let ${\mathrm{o}}(g)$ denote the order of $g$ as a group element and let $\deg(g)$ denote the degree of $g$ as a vertex of ${\mathcal{P}}(G)$. Throughout $\phi(n)$ shall denote the Euler totient function of the natural number $n$. Pick $g\in G$. Observe that $g$ has out-degree ${\mathrm{o}}(g) -1$ since there is a directed edge from $g\in G$ to each element of $\langle g\rangle-\{g\}$. There is a directed edge from each $h\in G-\{g\}$ to $g$ for which $g\in\langle h \rangle$, so the in-degree of $g$ is $|\{ h\in G-\{g\}\,|\, g\in\langle h\rangle\}|$. To account for directed edges which give the same undirected edge in the power graph of $G$, we introduce the following set. \begin{defn} The set of {\em bidirectional edges} $\overleftrightarrow{E}(G)$ of $\overrightarrow{{\mathcal{P}}}(G)$ is set of unordered pairs $\{ \{g,h\} \,|\, (g,h)\in \overrightarrow{E}(G) \hbox{ and } (h,g)\in\overrightarrow{E}(G) \}$. That is, $\overleftrightarrow{E}(G)$ consists of pairs of distinct elements, each of which is a positive power of the other. \end{defn} \begin{lem} \label{lem:bidi=gensame} Let $G$ be a finite group, and let $g$, $h$ be distinct elements of $G$. Then $\{g,h\}\in\overleftrightarrow{E}(G)$ if and only if $\langle g\rangle=\langle h\rangle$. \end{lem} \begin{proof} Straightforward from the definition of adjacency in the directed power graph. \end{proof} \begin{lem}\cite[Theorem 4.2]{ch} \label{lem:edgesizes} Let $G$ be a finite group of order $n$. Then \begin{eqnarray} \label{eq:dEGsize} |\overrightarrow{E}(G)| &=& \sum_{g\in G} ({\mathrm{o}}(g) -1), \\ \label{eq:bdEGsize} |\overleftrightarrow{E}(G)| &=& \frac{1}{2} \sum_{g\in G}( \phi({\mathrm{o}}(g))-1 ), \\ \label{eq:EGsize} |E(G)|&=& |\overrightarrow{E}(G)|-|\overleftrightarrow{E}(G)| = \frac{1}{2}\sum_{g\in G} \left(2{\mathrm{o}}(g) - \phi({\mathrm{o}}(g)) - 1\right). \end{eqnarray} \end{lem} \begin{proof} The sum in (\ref{eq:dEGsize}) adds out-degrees of vertices, and thus counts each directed edge once. Now (\ref{eq:bdEGsize}) follows from Lemma \ref{lem:bidi=gensame} and the fact that a cyclic group of order ${\mathrm{o}}(g)$ has ($\phi({\mathrm{o}}(g))$)-many generators. Indeed, $\phi({\mathrm{o}}(g))-1$ such edges leave $g$, and summing over all $G$ double counts these edges. For the first equality in (\ref{eq:EGsize}), count the edges in the directed power graph, and subtract one for each pair of oppositely oriented directed edges to avoid double counting. The second equality in (\ref{eq:EGsize}) follows from (\ref{eq:dEGsize}) and (\ref{eq:bdEGsize}). \end{proof} We give the number of edges in the undirected power graph of $\mathbb{Z}_n$. We use the following notation. \begin{nta} \label{nta:nfactored} Let $n$ be a positive integer. Write $n = p_1^{\alpha_1}p_2^{\alpha_2}\cdots p_k^{\alpha_k}$ for primes $p_1 < p_2 < \cdots <p_k$. Let $q=p_1$ and $p=p_k$ be the least and greatest prime divisors of $n$, and abbreviate $\beta=\alpha_1$ and $\alpha=\alpha_k$. \end{nta} It is well-known (see \cite[p.~27]{a'}, for instance) that \begin{equation} \label{eq:phi(n)primes} \phi(n) = p_1^{\alpha_1-1}(p_1-1)p_2^{\alpha_2-1}(p_2-1) \cdots p_k^{\alpha_k-1}(p_k-1). \end{equation} As a consequence, we have \cite[p.~28]{a'} \begin{equation} \label{eq:phi(product)} \phi(nm) =\phi(n)\phi(m) \frac{\gcd(n,m)}{\phi(\gcd(n,m))}. \end{equation} \begin{lem} \label{lem:sumZn} (See also \cite{burton:ENT}, page 143, exercise 5) With Notation \ref{nta:nfactored}, \begin{eqnarray} \label{eq:sumordZn} \sum_{z\in\mathbb{Z}_n} {\mathrm{o}}(z) &=&\sum_{d|n} \phi(d)d \ = \ \prod_{h=1}^{k} \frac{p_h^{2\alpha_h+1}+1}{p_h+1},\\ % \label{eq:sumphiordZn} \sum_{z\in \mathbb{Z}_n} \phi({\mathrm{o}}(z)) &=& \sum_{d|n} \phi(d)^2 \ = \ \prod_{h=1}^{k} \frac{p_h^{2\alpha_h}(p_h-1)+2}{p_h+1}. \end{eqnarray} \end{lem} \begin{proof} For each $z\in \mathbb{Z}_n$, ${\mathrm{o}}(z)$ is a divisor $d$ of $n$. For each divisor $d$, there are $\phi(d)$-many other elements of $\mathbb{Z}_n$ with the same order. Thus the first equality in (\ref{eq:sumordZn}) holds. Similarly, for each of the $\phi(d)$-many elements of $\mathbb{Z}_n$ with the same order as $z$, $\phi({\mathrm{o}}(z)) = \phi(d)$. Thus the first equality in (\ref{eq:sumphiordZn}) holds. When $n=1$, its only divisor is 1, and $\phi(1)=\phi(1)^2=1$. Thus both sums on the left are 1. There are no prime divisors of 1, so the product on the right is empty, and hence 1. Thus both second equalities holds when $k=0$. Assume that $n$ has $k\geq 1$ distinct prime divisors and that both second equalities holds for all $n$ with at most $k-1$ distinct prime divisors. Partition the divisors of $n$ according to the highest power of $q$ which divides it. As $d$ runs over $d|n$ with $q^\ell|f$ and $q^{\ell+1}\not|f$, we have $d = f q^\ell$ as $f$ runs over the divisors of $n/q^{\beta}$. Since $q\not| n/q^{\beta}$, (\ref{eq:phi(n)primes}) gives that $\phi(f q^\ell)= q^{\ell-1}(q-1)\phi(f)$. Compute \[ \sum_{d|n} \phi(d)d =\sum_{\ell=0}^\beta \sum_{f|n/q^\beta} \phi(q^\ell f)f q^\ell = \sum_{\ell=0}^\beta \phi(q^\ell)q^\ell \sum_{f|n/q^\beta} \phi(f)f = \left(1+\sum_{\ell=1}^\beta q^{\ell-1}(q-1)q^\ell\right) \sum_{f|n/q^\beta} \phi(f)f, \] and \[ \sum_{\ell=1}^\beta q^{2\ell-1}(q-1) = \sum_{\ell=0}^{\beta-1} (q-1)q^{2\ell+1} = \frac{q-1}{q} \cdot \sum_{\ell=0}^{\beta-1}(q^2)^\ell = q(q-1) \cdot \frac{(q^{2\beta}-1)}{q^2-1} = q\cdot\frac{q^{2\beta}-1}{q+1} = \frac{q^{2\beta+1}+1}{q+1} - 1. \] Now the second equality in (\ref{eq:sumordZn}) follows by induction. Similarly, \[ \sum_{d|n} \phi(d)^2 = \sum_{\ell=0}^\beta \sum_{f|n/q^\beta} \phi(q^\ell f)^2 = \sum_{\ell=0}^\beta \phi(q^\ell)^2 \sum_{f|n/q^\beta} \phi(f)^2 = \left(1+\sum_{\ell=1}^\beta (q^{\ell-1}(q-1))^2\right) \sum_{f|n/q^\beta} \phi(f)^2, \] and \[ \sum_{\ell=1}^\beta (q^{\ell-1}(q-1))^2 = (q-1)^2 \sum_{\ell=0}^{\beta-1} q^{2\ell} = (q-1)^2 \frac{q^{2\beta}-1}{q^2-1} = \frac{(q-1)(q^{2\beta}-1)}{q+1}. \] Now the second equality in (\ref{eq:sumphiordZn}) follows by induction. \end{proof} \begin{cor} \label{cor:peelsumphiordZn} With Notation \ref{nta:nfactored},\footnote{This corollary was after the published version. It makes explicit a fact proved in Lemma \ref{lem:sumZn} for later use.} \[ \sum_{d|n} \phi(d)d = \frac{q^{2\beta+1}+1}{q+1}\cdot\sum_{f|n/q^\beta} \phi(f)f. \] \end{cor} \begin{cor} With Notation \ref{nta:nfactored}, \[ |E(\mathbb{Z}_n)| = \sum_{d|n} \phi(d)(d-\frac{\phi(d)}{2}) -\frac{n}{2} = \prod_{h=1}^{k} \frac{p_h^{2\alpha_h+1}+1}{p_h+1} - \frac{1}{2} \prod_{h=1}^{k} \frac{p_h^{2\alpha_h}(p_h-1)+2}{p_h+1} - \frac{n}{2}. \] \end{cor} \begin{lem} With Notation \ref{nta:nfactored}, pick $z \in\mathbb{Z}_{n}$, and write $e={\mathrm{o}}(z)$. Then \begin{equation} \label{eq:degz-Zn} \deg(z) = e -1 -\phi(e) + \sum_{{ d | n/e}} \phi(de) = e -\phi(e) -1 +\phi(e) \sum_{d | n/e} \phi(d)\frac{\gcd(d,e)}{\phi(\gcd(d,e))}. \end{equation} \end{lem} \begin{proof} The term ${\mathrm{o}}(z) -1=e-1$ is the out-degree of $z$. There is a directed edge from each element $x\in \mathbb{Z}_{n}$ to $z$ whenever ${\mathrm{o}}(z)|{\mathrm{o}}(x)$. For such an $x$, ${\mathrm{o}}(x)$ is ${\mathrm{o}}(z)$ times a divisor of $n/{\mathrm{o}}(z)$. There are $\phi(k{\mathrm{o}}(z))$-many generators of $\langle x\rangle$. Thus the in-degree of $z$ is $\sum_{d | n/e} \phi(d e)$. However, to avoid double counting when dropping the orientation, we must exclude those elements with the same order as $z$, i.e., the case $k=1$ where $\phi(ke)=\phi(e)$. Hence the first equality holds. The second equality follows from (\ref{eq:phi(product)}) since the summand for $d=1$ is 1. \end{proof} \begin{lem} \label{lem:Inducedsubgaph} Let $G$ be a group, and let $H$ be a subgroup of $G$. \begin{enumerate} \item $\overrightarrow{{\mathcal{P}}}(H)$ is an induced subgraph of $\overrightarrow{{\mathcal{P}}}(G)$. \item All out-edges from an element of $H$ terminate at an element of $H$. \item ${\mathcal{P}}({H})$ is an induced subgraph of ${\mathcal{P}}(G)$. \end{enumerate} In particular, the adjacencies and non-adjacencies between elements of $H$ are the same in $\overrightarrow{{\mathcal{P}}}(H)$ and $\overrightarrow{{\mathcal{P}}}(G)$, and similarly for ${\mathcal{P}}({H})$ and ${\mathcal{P}}(G)$. \end{lem} \begin{proof} Straightforward from the definitions. \end{proof} \section{Some inequalities}\label{sec:prelim} We develop some inequalities. \begin{lem} \label{lem:inequalities} Let $G$ be a finite group. Then the following hold. \begin{enumerate} \item $|G|\leq |E(G)|+1$ with equality if and only if $G$ is an elementary abelian $2$-group. \item $|E(G)|\leq |\overrightarrow{E}(G)|$ with equality if and only if $G$ is an elementary abelian $2$-group. \item \label{item:equality} \footnote{This statement of Lemma \ref{item:equality} went awry in the published version. The proof is essentially the same.} $2 |\overleftrightarrow{E}(G)|\leq|\overrightarrow{E}(G)|-|G|+1$ with equality if and only if every nonidentity element of $G$ has prime order, i.e., $G$ is an EPO-group. \end{enumerate} \end{lem} \begin{proof} (i): Every nonidentity element of $G$ is adjacent to the identity, so $|G|\leq |E(G)|+1$. Equality holds if and only if every nonidentity element of $G$ has order two. (ii): Every undirected edge arises from a directed edge, so the inequality holds. By (\ref{eq:EGsize}), $|E(G)| = |\overrightarrow{E}(G)|$ if and only if there are no bidirectional edges if and only if for all $g\in G$, $g$ is the only generator of $\langle g\rangle$ if and only if for all $g\in G$, one is the only number both less than and coprime to ${\mathrm{o}}(g)$ if and only if every element of $G$ has order two. (iii): The ($|G|-1$)-many edges to the identity are not bidirectional, and the bidirectional edges come from pairs of directed edges. Thus $2 |\overleftrightarrow{E}(G)|\leq |\overrightarrow{E}(G)|-|G|+1$. Now equality holds in (iii) if and only if every edge not incident to the identity is bidirectional. If some $g\in G$ has composite order, say ${\mathrm{o}}(g)=pm$ for a prime $p$ and $m>1$, then the edge between $g$ and $g^p$ is not bidirectional. Thus equality fails in this case. If every element of $G$ has prime order, then each element generate a cyclic group of prime order. These subgroups only have the identity in common. Thus every edge not incident to the identity is bidirectional, so equality holds. \end{proof} The remaining inequalities in this section pertain to $\mathbb{Z}_n$, and so are number theoretic in nature. \begin{lem}\cite[Main Theorem]{a} \label{lem:PGcyclicprimepower2} Let $G$ be a group of finite order $n$. Then $\sum_{g\in G} {\mathrm{o}}(g) \leq \sum_{z\in\mathbb{Z}_n} {\mathrm{o}}(z)$, with equality if and only if $G$ is isomorphic to $\mathbb{Z}_n$. \end{lem} Theorem \ref{cor:maximume} is a straightforward consequence of (\ref{eq:dEGsize}) and Lemma \ref{lem:PGcyclicprimepower2}. Our next goal is to improve the bound ${\sum_{d|n} \phi(d)d > {n^2}/{p}}$ given in \cite[Lemma D]{a}. \begin{thm} \label{thm:phiddineq} \footnote{The published version of Theorem \ref{thm:phiddineq}, did not note the forbidden cases. This was our mistake. The corrected proof is a careful reworking of the original. Only (\ref{eq:betterbound}) was required in the sequel, so this error had no impact upon the sequel.} % With Notation \ref{nta:nfactored}, if $n>1$ does not have one of the following prime factorizations $2^a$, $2^a 3^b$, $2^a 3^b 5^c$, or $2^a 3^b 5^c7^d$ for positive $a$, $b$, $c$, $d$, then \begin{equation} \label{eq:betterboundodd} \sum_{d|n} \phi(d)d \geq \left( \prod_{h=1}^{k} \frac{p_h+1}{p_h}\right)\cdot \frac{n^2}{p}. \end{equation} If $n$ is not a power of 2, then \begin{equation} \label{eq:betterbound} \sum_{d|n} \phi(d)d \geq \left( \frac{q+1}{q}\right)\cdot \frac{n^2}{p}. \end{equation} \end{thm} \begin{proof} We prove (\ref{eq:betterboundodd}) by induction on the number $k$ of distinct prime divisors of $n$. We launch the induction by showing that (\ref{eq:betterboundodd}) holds for the ``first'' sets of prime factors that are not forbidden. Abbreviate \[ S(n) = \sum_{d|n} \phi(d)d, \quad R(n) = \left( \prod_{h=1}^{k} \frac{p_h+1}{p_h}\right)\cdot \frac{n^2}{p}; \quad \hbox{ so}\qquad \frac{S(n)}{R(n)} = p\cdot \prod_{h=1}^{k}\frac{\displaystyle{p_h^2+\frac{1}{p_h^{2\alpha_h-1}}}} {(p_h+1)^2}. \] Note that (\ref{eq:betterboundodd}) holds when $S(n)/R(n)\geq 1$. The term $1/p_h^{2\alpha_h-1}$ is a decreasing function of $\alpha_h$ and $\alpha_h$ can be arbitrarily large. Thus to verify (\ref{eq:betterboundodd}), it suffices to show that \[ T(n):= p\cdot \prod_{h=1}^{k}\frac{\displaystyle{p_h^2}} {(p_h+1)^2} \geq 1. \] Note that $T(n)$ does not depend upon the exponents. We readily compute that $T(3^{\alpha_2}) = 27/16$, $T(2^{\alpha_1} 5^{\alpha_2})=5/4$, $T(2^{\alpha_1}3^{\alpha_2}7^{\alpha_3})=343/256$, $T(2^{\alpha_1}3^{\alpha_2}5^{\alpha_3}11^{\alpha_4})=33275 / 20736$, $T(2^{\alpha_1}3^{\alpha_2}5^{\alpha_3}7^{\alpha_4} 11^{\alpha_5})=1630475/1327104$. This establishes the initial step for the induction on $k$ for $n$ not of a forbidden form. We now proceed to the inductive step. Suppose $n=p_1^{\alpha_1} p_2^{\alpha_2} \cdots p_k^{\alpha_k}$ with $k\geq 2$, and let $n'=p_1^{\alpha_1} p_2^{\alpha_2} \cdots p_{k-1}^{\alpha_{k-1}}$. Assume $n'$ is not one of the forbidden forms, so $p_k\geq p_{k-1}+2$ and $S(n')\geq R(n')$ by induction. Observe that by Corollary \ref{cor:peelsumphiordZn} \[ S(n) = S(n')\cdot \frac{p_k^{2\alpha_{k}+1}+1}{p_k + 1}, \qquad R(n) = R(n') \cdot p_{k-1}\cdot \frac{p_k+1}{p_k}\cdot \frac{p_k^{2\alpha_k}}{p_k}.\] Since $p_{k-1}\leq p_k-2$, we compute \[\frac{S(n)}{R(n)} = \frac{S(n')}{R(n')}\left(\frac{p^2(p+p^{-2\alpha_k})}{(p+1)^2p_{k-1}}\right) \geq \frac{S(n')}{R(n')}\left(\frac{p^3+p^{-2\alpha_k+2}}{p^3-3p-2 }\right) \geq 1. \] Thus (\ref{eq:betterboundodd}) holds by induction. Except in the forbidden cases this implies (\ref{eq:betterbound}) since each factor $(p_h+1)/p_h>1$. Thus, since $n$ is not a power of 2, we need only consider three cases for (\ref{eq:betterbound}). Abbreviate \[ Q(n) = \frac{q+1}{q}\frac{n^2}{p}, \quad U(n) = \frac{p q}{q+1}\prod_{h=1}^{k}\frac{\displaystyle{p_h}} {p_h+1}; \quad\hbox{ so }\quad \frac{S(n)}{Q(n)} = \frac{p q}{q+1}\prod_{h=1}^{k}\frac{\displaystyle{p_h+\frac{1}{p_h^{2\alpha_h}}}} {p_h+1} \geq U(n). \] To prove (\ref{eq:betterbound}) , it suffice to show that $U(n)\geq 1$ for forbidden $n$ which are not powers of 2. One computes $U(2^{\alpha_1} 3^{\alpha_2})=1$, $U(2^{\alpha_1}3^{\alpha_2}5^{\alpha_3})=25/18$, $U(2^{\alpha_1}3^{\alpha_2}5^{\alpha_3}7^{\alpha_4})=245/144$. \end{proof} A hybrid of $T(n)$ and $S(n)/R(n)$ allows one to verify (\ref{eq:betterboundodd}) for families with arbitrary powers, yielding $\frac{S}{R}(2) = 1$, $\frac{S}{R}(2\cdot 3 \cdot 5^{\alpha_3}) \geq 875/864$, $\frac{S}{R}(2^ 2 \cdot3\cdot 5\cdot 7^{\alpha_4}) \geq 184877/184320$, $\frac{S}{R}(2\cdot 3^{\alpha_2} \cdot 5^{\alpha_3}\cdot 7^{\alpha_4}) \geq 1225/1152$. Thus (\ref{eq:betterboundodd}) holds for these arguments for all positive integers $\alpha_2$, $\alpha_3$, $\alpha_4$. Since $S(n)/R(n)$ is decreasing in the $\alpha_h$, (\ref{eq:betterboundodd}) can be shown to fail for some $n$ by showing $S(n')/R(n')<1$ for some $n'$ with the same distinct prime divisors but possibly smaller, positive exponents. We compute $\frac{S}{R}(2^2) = 11/12$, $\frac{S}{R}(2\cdot 3) = 7/8$, $\frac{S}{R}(2^2\cdot 3 \cdot 5) = 539/576$, $\frac{S}{R}( 2\cdot 3^2 \cdot 5) = 427/432$, $\frac{S}{R}(2^3\cdot 3 \cdot 5\cdot 7) = 90601/92160$, $\frac{S}{R}(2^ 2 \cdot3^2\cdot 5\cdot 7) = 201971/207360$, $\frac{S}{R}(2^ 2 \cdot3\cdot 5^2\cdot 7) = 1725031/1728000$. Thus (\ref{eq:betterboundodd}) fails for numbers with forbidden prime factorizations other than those noted above. \begin{lem} With Notation \ref{nta:nfactored}, pick $z \in\mathbb{Z}_{n}$, and write $e={\mathrm{o}}(z)$. Then \[ \deg(z) \geq e -1 +\phi(e) (\frac{n}{e}-1), \] with equality if and only if $e$ and $n/e$ are coprime. \end{lem} \begin{proof} Consider (\ref{eq:degz-Zn}). The term ${\gcd(d,e)}/{\phi(\gcd(d,e))}$ is at least 1 with equality if and only if $\gcd(d,e)=1$. If $n/e$ has any divisors not coprime to $e$, then the inequality must be strict. \end{proof} \begin{lem} With Notation \ref{nta:nfactored}, \begin{equation} \label{eq:phigeqn/p} {\phi(n)} \geq \frac{n}{p}, \end{equation} with equality if and only if $n=2^\alpha 3^\beta$ with $\alpha \geq 1$ and $\beta\geq 0$. \end{lem} \begin{proof} If $n=p^\alpha$, then $\phi(p^\alpha)=p^{\alpha-1}(p-1) \geq p^{\alpha-1} = p^\alpha/p$. The inequality holds in this case, with equality if and only if $p=2$. Now assume that $n$ has $k\geq 2$ distinct prime factors. By (\ref{eq:phi(n)primes}), $\phi(n)/n = (p_1 -1) (p_2-1) \cdots (p_k-1)/(p_1 p_2 \cdots p_k)$. Note that $p_{i+1} - 1\geq p_{i-1}$ $(1\leq i \leq k-1)$, with equality precisely when $k=2$, $p_1=2$, and $p_{2}=3$. Telescoping the middle terms gives $\phi(n)/n\geq (p_1-1)/p_k = (q-1)/p\geq 1/p$, with equality under the stated conditions, i.e., $n=2^\alpha 3^\beta$ for positive $\alpha$ and nonnegative $\beta$. \end{proof} \begin{lem} \label{lem:avez} With Notation \ref{nta:nfactored}, assume that $n$ is not equal to $2^\alpha$ with $\alpha\geq 1$. Then \footnote{Fixed a sign error in first equation--only used in Lemma \ref{lem:lemma1}, but doesn't change that result.} \begin{align*} \sum_{z\in \mathbb{Z}_n} \deg(z) &\geq 2 \frac{n^2}{p} - \frac{n}{p}-1& \hbox{if $\phi(n) \leq n/q$,} \\ \sum_{z\in \mathbb{Z}_n} \deg(z) &>(n-1)\left(\frac{n}{q}+1\right) & \hbox{if $\phi(n)> n/q$.} \end{align*} \end{lem} \begin{proof} For any prime $p$ (other than 2), both $\phi(p^\alpha)=p^{\alpha-1}(p-1)>p^\alpha/p$ by (\ref{eq:phi(n)primes}) and $|E(\mathbb{Z}_{p^\alpha})|=(p^\alpha-1)(p^\alpha-1)/2 >p^\alpha/p$ by Theorem \ref{lem:PGcyclicprimepower1}. Thus we may assume that $n$ is not a prime power. Each of the $\phi(n)$-many generators of $\mathbb{Z}_n$ has degree $n-1$, as does the identity. So summing over the generators and the identity gives \[ \sum_{z} \deg(z) = \phi(n) (n-1) + n -1 = (\phi(n)+1) (n-1).\] If $\phi(n) > n/q$, then \[ \sum_{z\in \mathbb{Z}_n} \deg(z) > (n-1)\left(\frac{n}{q}+1\right)>\frac{n^{2}}{q}.\] By (\ref{eq:degz-Zn}), $z \in \mathbb{Z}_{n}$ has degree $\deg(z) = {\mathrm{o}}(z) -1-\phi( {\mathrm{o}}(z) ) + \sum_{ d | n/{\mathrm{o}}(z)} \phi(d{\mathrm{o}}(z))$. For each of the ($n-\phi(n)-1$)-many nonidentity nongenerators $z$ of $\mathbb{Z}_n$, the summand corresponding to $d=1$ is $\phi( {\mathrm{o}}(z) )$ and to $d=n/{\mathrm{o}}(z)$ is $\phi(n)$. By (\ref{eq:phigeqn/p}), $\phi(n)\geq n/p$. Thus $\deg(z)\geq {\mathrm{o}}(z) -1 +n/p$. Equality holds for all nonidentity nongenerators if and only if $n=6$. Now\footnote{The sign error originated in the following line.} \[ \sum_{z\in \mathbb{Z}_n} \deg(z) \geq (\sum_{z\in \mathbb{Z}_n} {\mathrm{o}}(z)) - n +( n-\varphi(n)-1) \frac{n}{p} + n-1 \geq (\sum_{z\in \mathbb{Z}_n} {\mathrm{o}}(z)) + \frac{n^2}{p} -(\phi(n)+1)\frac{n }{p} - 1. \] If $\phi(n) \leq n/q$, then by (\ref{eq:betterbound}) \[ \sum_{z\in \mathbb{Z}_n} \deg(z) \geq \frac{q+1}{q}\cdot \frac{n^2}{p} + \frac{n^2}{p} -(\frac{n}{q}+1)\frac{n }{p} - 1 = 2 \frac{n^2}{p} -\frac{n}{p}-1.\] \end{proof} Recall that for any undirected graph $\Gamma$ \[ |E(\Gamma)| = \frac{1}{2}\sum_{g\in \Gamma} \deg(g).\] \begin{lem} \label{lem:phi(n)ineq} With Notation \ref{nta:nfactored}, suppose $\phi(n) > n/q$. Then $n$ is odd. \end{lem} \begin{proof} In light of (\ref{eq:phi(n)primes}), $\phi(n) > n/q$ if and only if $(p_{1} - 1) \cdots (p_{k} - 1) > p_{2} \cdots p_{k}$. Assume $q=p_1 = 2$, so $p_{1} - 1 = 1$. Then the above inequality gives $(p_{2} - 1) \cdots (p_{k} - 1) > p_{2}\cdots p_{k}$, which is impossible. Therefore $n$ is odd. \end{proof} \begin{lem} \label{lem:d(r)} Assume that $r$ and $s$ are natural numbers such that $s | r$. Then $2r - \phi(r) - 1 \geq 2s - \phi(s) - 1$, with equality if and only if $s=r$. \end{lem} \begin{proof} Write $r= st$, and observe that that $\phi(st) \leq \phi(s)t$. Then \[ 2 r - \phi(r) - 1 \geq 2st - \phi(s)t -1 = t(2s - \phi(s)) - 1 \geq 2s - \phi(s) - 1. \] Observe that if $t>1$, then this inequality is strict. \end{proof} \section{Direct products} Let $G$ be a finite group. Suppose $G=U\times V$ is the direct product of normal subgroups $U$ and $V$. We view the underlying set of $G$ as the cartesian product $U\times V =\{uv\,|\, u\in U, v\in V\}$, where we use juxtaposition (and suggestive notation) to avoid confusion with directed edges in a power graph. We use $\cdot$ for the usual the group product on $U\times V$, namely $uv\cdot u'v' = (uu')(vv')$. We write $g\rightarrow h$ when $(g,h)\in \overrightarrow{E}(G)$, that is, when $h=g^a$ but $h\not=g$ for some $a$. \begin{lem} \label{lem:gendirectprod} Let $G$ be a finite group, and let $U$ and $V$ be subgroups of $G$. Define \begin{eqnarray*} D &=& \begin{array}{l} \{(uv, u'v')\,|\, uu'\in \overrightarrow{E}(U), vv'\in \overrightarrow{E}(V)\}\\ \phantom{X}\cup \{(uv, u'v)\,|\, uu'\in \overrightarrow{E}(U), v\in V\} \cup \{(uv, uv')\,|\, u\in U, vv'\in \overrightarrow{E}(V)\}, \end{array}\\ B &=& \begin{array}{l} \{\{uv, u'v'\}\,|\, uu' \in \overleftrightarrow{E}(U), vv' \in \overleftrightarrow{E}(V)\}\\ \phantom{X}\cup \{\{uv, u'v\}\,|\, uu' \in \overleftrightarrow{E}(U), v\in V\} \cup \{\{uv, uv'\}\,|\, u\in U, vv' \in \overleftrightarrow{E}(V)\}. \end{array} \end{eqnarray*} Suppose $G$ is the direct product $U\times V$ of $U$ and $V$. Then \begin{eqnarray} \label{eq:dEuvin} \overrightarrow{E}(U\times V) &\subseteq& D, \\ \label{eq:bdEuvin} \overleftrightarrow{E}(U\times V) &\subseteq& B. \end{eqnarray} \end{lem} \begin{proof} Let $u$, $u'\in U$ and $v$, $v'\in V$. First suppose $uv\rightarrow u'v'$ in $\overrightarrow{{\mathcal{P}}}(U\times V)$, so for some positive integer $c$, $u'v'=(uv)^{\cdot c} = u^c v^c$. The product is direct, so $u'= u^c$ and $v'=v^c$. Thus one of the following holds: (1d) $u\rightarrow u'$ in $\overrightarrow{{\mathcal{P}}}(U)$ and $v\rightarrow v'$ in $\overrightarrow{{\mathcal{P}}}(V)$, (2d) $u'=u$ and $v\rightarrow v'$ in $\overrightarrow{{\mathcal{P}}}(V)$, (3d) $u\rightarrow u'$ in $\overrightarrow{{\mathcal{P}}}(U)$ and $v'=v$, or (4d) $u'=u$ and $v'=v$. Thus (\ref{eq:dEuvin}) holds. Now suppose $u'v'\leftrightarrow uv$ in $\overleftrightarrow{{\mathcal{P}}}(U\times V)$. As above, one of the following holds: (1u) $u'\leftrightarrow u$ in $\overleftrightarrow{{\mathcal{P}}}(U)$ and $v'\leftrightarrow v$ in $\overleftrightarrow{{\mathcal{P}}}(V)$, (2u) $u'=u$ and $v'\leftrightarrow v$ in $\overleftrightarrow{{\mathcal{P}}}(V)$, (3u) $u'\leftrightarrow u$ in $\overleftrightarrow{{\mathcal{P}}}(U)$ and $v'=v$, or (4u) $u'=u$ and $v'=v$. Thus (\ref{eq:bdEuvin}) holds. \end{proof} \begin{lem} \label{lem:directprodeq} With reference to Lemma \ref{lem:gendirectprod}, suppose $G=U\times V$. Then the following hold. \begin{enumerate} \item $\overrightarrow{E}(U\times V) = D$ if and only if $(|U|,|V|) = 1$. \item If $(|U|,|V|) = 1$, then $\overleftrightarrow{E}(U\times V) = B$. \end{enumerate} \end{lem} \begin{proof}(i): Suppose that $\gcd(|U|, |V|) = 1$. Let $u \rightarrow u'$ in $\overrightarrow{{\mathcal{P}}}(U)$ and $v \rightarrow v'$ in $\overrightarrow{{\mathcal{P}}}(V)$. Then $uv \rightarrow u'v'$ in $\overrightarrow{{\mathcal{P}}}(U\times V)$. Also for each $u \in U$ and $v \in V$ we have $uv \rightarrow uv'$ in $\overrightarrow{{\mathcal{P}}}(U\times V)$ and $uv \rightarrow u'v$ in $\overrightarrow{{\mathcal{P}}}(U\times V)$. Therefore $D \subseteq \overrightarrow{E}(U\times V)$, and thus equality holds in this case. We now show that there are no other instances of equality. Suppose some prime $p$ divides $\gcd(|U|, |V|)$. Then $U$ and $V$ contain respective elements $u$ and $v$ of order $p$. Observe that there is no edge $uv\rightarrow ue$ although it appears in the third subset in the definition of $D$. Thus the inclusion is strict unless no prime divides $\gcd(|U|, |V|)$, i.e., $\gcd(|U|, |V|)=1$. (ii): The proof is similar to the first part of the proof of (i). \end{proof} \begin{cor} \label{cor:directprodsize} With reference to Lemma \ref{lem:gendirectprod}, suppose that $G = U\times V$ and $(|U|,|V|) = 1$. Then \begin{enumerate} \item The edge set of the directed power graph of $U\times V$ has size \[ |\overrightarrow{E}(U\times V)| = |\overrightarrow{E}(U)|\cdot |\overrightarrow{E}(V)| + |\overrightarrow{E}(U)|\cdot |V| + |U| \cdot |\overrightarrow{E}(V)|. \] \item The set of bidirectional edges of the directed power graph of $U\times V$ has size \[ |\overleftrightarrow{E}(U \times V)| = 2|\overleftrightarrow{E}(U)|\cdot |\overleftrightarrow{E}(V)| + |\overleftrightarrow{E}(U)|\cdot |V| + |U|\cdot |\overleftrightarrow{E}(V)|. \] \end{enumerate} \end{cor} \begin{proof} Referring to the proof of (\ref{eq:dEuvin}), there are $|\overrightarrow{E}(U)|\cdot |\overrightarrow{E}(V)|$ pairs in case (1d), $|U| \cdot |\overrightarrow{E}(V)|$ pairs in case (2d), $|\overrightarrow{E}(U)|\cdot |V|$ pairs in case (3d), but no possible edges in case (4d). Therefore (i) holds. Referring to the proof of (\ref{eq:bdEuvin}), there are $2|\overleftrightarrow{E}(U)|\cdot |\overleftrightarrow{E}(V)|$ pairs in case (1u) since for all $\{u,u'\} \in \overleftrightarrow{{\mathcal{P}}}(U)$ and $\{v,v'\} \in \overleftrightarrow{{\mathcal{P}}}(V)$ we have that $u'v'\leftrightarrow uv$ and $u'v\leftrightarrow uv'$ are distinct bidirectional edges in $\overleftrightarrow{E}(G)$. There are $|U| \cdot |\overleftrightarrow{E}(V)|$ pairs in case (2u), and $|\overleftrightarrow{E}(U)|\cdot |V|$ pairs in case (3u), but no possible edges in case (4u). Thus (ii) holds. \end{proof} \begin{lem} \label{lem:prodedgedifference} Let $U$, $V$, and $V'$ be finite groups with $|V|=|V'|$ and $(|U|,|V|) = 1$. Then \[ |{E}(U\times V)| - |{E}(U\times V')| = (|\overrightarrow{E}(U)| -2|\overleftrightarrow{E}(U)| ) ( |\overrightarrow{E}(V)|- |\overrightarrow{E}(V')|) + (2|\overleftrightarrow{E}(U)| + |U|) (|{E}(V)| - |{E}(V')|). \] \end{lem} \begin{proof} Expand $|{E}(U\times V)| - |{E}(U\times V')|$ with (\ref{eq:EGsize}): \[ |{E}(U\times V)| - |{E}(U\times V')|=( |\overrightarrow{E}(U\times V)|-|\overrightarrow{E}(U\times V)| ) - (|\overleftrightarrow{E}(U\times V)| - |\overleftrightarrow{E}(U\times V')|). \] Since $|V|=|V'|$, Corollary \ref{cor:directprodsize} gives \begin{eqnarray*} \overrightarrow{E}(U\times V)| -|\overrightarrow{E}(U\times V')| &=& (|\overrightarrow{E}(U)| +|U|) ( |\overrightarrow{E}(V)|- |\overrightarrow{E}(V')|),\\ |\overleftrightarrow{E}(U\times V)| - |\overleftrightarrow{E}(U\times V')| &=& (2|\overleftrightarrow{E}(U)| + |U|) (|\overleftrightarrow{E}(V)|-|\overleftrightarrow{E}(V')|). \end{eqnarray*} In the second line, use (\ref{eq:EGsize}) to write $|\overleftrightarrow{E}(V)| = |\overrightarrow{E}(V)|-|{E}(V)|$, and similarly for $|\overleftrightarrow{E}(V')|$. Thus \[ |\overleftrightarrow{E}(U\times V)| - |\overleftrightarrow{E}(U\times V')| = (2|\overleftrightarrow{E}(U)| + |U|) ((|\overrightarrow{E}(V)|-|\overrightarrow{E}(V')|) - (|E(V)|-|E(V')|)). \] Combining the above and simplifying gives the desired result. \end{proof} \begin{cor} \label{cor:whenprodhasmoreedges} With reference to Lemma \ref{lem:prodedgedifference}, suppose $|E(V)|\geq |E(V')|$ and $|\overrightarrow{E}(V)|\geq |\overrightarrow{E}(V')|$. Then $|{E}(U\times V)| \geq |{E}(U\times V')|$, with equality if and only if $|E(V)|= |E(V')|$ and $|\overrightarrow{E}(V)|= |\overrightarrow{E}(V')|$. \end{cor} \begin{proof} Observe that $|\overrightarrow{E}(U)| > 2|\overleftrightarrow{E}(U)|$ since each bidirectional edge is counted twice in $\overrightarrow{E}(U)$ and since $\overrightarrow{E}(U)$ contains one-way edges from each element of $U$ to the identity. Clearly, $2|\overleftrightarrow{E}(U)| + |U|>0$. With the given assumptions, Lemma \ref{lem:prodedgedifference} gives that $|{E}(U\times V)| - |{E}(U\times V')|\geq 0$, so the result follows. Equality holds if and only if $|E(V)|= |E(V')|$ and $|\overrightarrow{E}(V)|= |\overrightarrow{E}(V')|$ since all terms on the right are positive, so the right side is zero if and only if these equalities hold. \end{proof} \section{Semidirect products} We recall semidirect products. Recall that a group $G$ is the (internal) semidirect product of a normal subgroup $U$ and a subgroup $V$ if and only if $G=UV$ and $U\cap V=\{e\}$ \cite{r}. To uniquely determine $G$ from $U$ and $V$, we specify a homomorphism $\varphi:V\rightarrow \mathrm{Aut}(U)$. As is the custom, write $G=U\rtimes_\varphi V$ in this situation. When there is no ambiguity, we write $\varphi v$ for $\varphi(v)$ and place such automorphisms as a superscript of the element of $U$ to which it is applied. The elements of $G=U\rtimes_\varphi V$ can be identified with the cartesian product of the underlying sets of $U$ and $V$, just as is the case for the direct product. We use $\ast$ for the group product on $U\rtimes_\varphi V$, namely $uv \ast u'v' = (u(u')^{\varphi v})(vv')$. We write $(uv)^{ \cdot c}$ and $(uv)^{\ast c}$ to denote the $c$th powers of $uv$ under the corresponding operations. We write ${\mathrm{o}}_{\cdot}(uv)$ and ${\mathrm{o}}_{\ast}(uv)$ to denote the order of $uv$ relative to the corresponding multiplication, and we write $(uv)^{\cdot{-}1}$ and $(uv)^{\ast{-}1}$ for the corresponding inverses. With this notation, \begin{eqnarray} \label{eq:semidirect} uv^{\ast c} &=& (uu^{\varphi v} u^{\varphi v^2} \cdots u^{\varphi v^{c-2}} u^{\varphi v^{c-1}}) (v^c), \\ (uv)^{\ast-1} &=& (u^{-1})^{\varphi v^{-1}}(v^{-1}) \nonumber. \end{eqnarray} \begin{lem} \label{lem:astpowercyclic} Suppose that $G$ is a finite group and that $G=U\rtimes_\varphi V$ is the semidirect product of a normal cyclic subgroup $U$ and a subgroup $V$. Pick $u\in U$ and $v\in V$. Then $u^{\varphi v} = u^r$ for some $r$, and $(uv)^{\ast c} = u^t v^c$, where $t = 1+r+r^2 +\cdots +r^{c-1}$. \end{lem} \begin{proof} Consider the subgroup $\langle u \rangle$ of $U$. Since $U$ is cyclic and since $\langle u \rangle$ is the unique subgroup of $U$ of its order, it must be the case that $\langle u \rangle$ is a characteristic subgroup of $U$. Thus $u^{\varphi v} = u^r$ for some $r$. Now $u^{\varphi v^b} = u^{r^b}$. With this we compute \[ (u v)^{ \ast c} = (uu^{\varphi v} u^{\varphi v^2} \cdots u^{\varphi v^{c-2}} u^{\varphi v^{c-1}}) (v^c) = (u^1 u^r u^{r^2} \cdots u^{r^{c-1}}) (v^c) = (u^{1+r+r^2 +\cdots +r^{c-1}}) (v^c) = u^t v^c. \] \end{proof} \begin{lem} \label{lem:Semidirprodcyc} Suppose that $G$ is a finite group and that $G=U\rtimes_\varphi V$ is the semidirect product of a normal cyclic subgroup $U$ and a subgroup $V$. Then \begin{eqnarray} \label{eq:sdEuvin} \overrightarrow{E}(U\rtimes_\varphi V) &\subseteq& D, \\ \label{eq:sbdEuvin} \overleftrightarrow{E}(U\rtimes_\varphi V) &\subseteq& B. \end{eqnarray} In particular, if $(|U|,|V|)=1$, then $E(U\rtimes_\varphi V)\subseteq E(U\times V)$ and $\overleftrightarrow{E}(\mathbb{Z}_{|U|}\rtimes_\varphi \mathbb{Z}_{|V|}) \subseteq \overleftrightarrow{E}(\mathbb{Z}_{|U|}\times \mathbb{Z}_{|V|})$. \end{lem} \begin{proof} Pick $u$, $u'\in U$ and $v$, $v'\in V$, and say $u^{\varphi v} = u^r$. Suppose $uv\rightarrow u'v'$ in $\overrightarrow{{\mathcal{P}}}(U\rtimes_\varphi V)$. Then $u'v' = (uv)^{\ast c} = u^t v^c$ for some $t$ as in Lemma \ref{lem:astpowercyclic}. Since the product is semidirect, $u'=u^{t}$ and $v'=v^c$. Arguing as in the proof of Lemma \ref{lem:gendirectprod}, we reach the same conclusion. \end{proof} \begin{lem} \label{lem:divideorder1} Suppose that $G$ is a finite group and that $G=U\rtimes_\varphi V$ is the semidirect product of a normal abelian subgroup $U$ and a subgroup $V$. Assume $U$ and $V$ have coprime orders. Then for all $u \in U$ and $v\in V$, ${\mathrm{o}}_{\ast}(uv)|{\mathrm{o}}_{\cdot}(uv)$. \end{lem} \begin{proof} To prove the result we show that $(u v)^{\ast {\mathrm{o}}_{\cdot }({uv})}$ is the identity of $U\rtimes_\varphi V$. Note that ${\mathrm{o}}_{\cdot}(u v) = {\mathrm{o}}(u){\mathrm{o}}(v)/\gcd({\mathrm{o}}(u),{\mathrm{o}}(v)) = {\mathrm{o}}(u){\mathrm{o}}(v)$. By (\ref{eq:semidirect}), \begin{eqnarray*} (u v)^{ \ast {\mathrm{o}}(v)} &=& (uu^{\varphi v} u^{\varphi v^2} \cdots u^{\varphi v^{{\mathrm{o}}(v)-2}} u^{\varphi v^{{\mathrm{o}}(v)-1}}) (v^{{\mathrm{o}}(v)})\\ &=& (uu^{\varphi v} u^{\varphi v^2} \cdots u^{\varphi v^{{\mathrm{o}}(v)-2}} u^{\varphi v^{{\mathrm{o}}(v)-1}})(e). \end{eqnarray*} Thus \begin{eqnarray*} (u v)^{ \ast {\mathrm{o}}_{\cdot}({u v})} &=&\left[ (uu^{\varphi v} u^{\varphi v^2} \cdots u^{\varphi v^{{\mathrm{o}}(v)-2}} u^{\varphi v^{{\mathrm{o}}(v)-1}})(e)\right]^{\ast {\mathrm{o}}(u)}\\ &=& (u^{{\mathrm{o}}(u)}(u^{\varphi v})^{{\mathrm{o}}(u)} (u^{\varphi v^2})^{{\mathrm{o}}(u)} \cdots (u^{\varphi v^{{\mathrm{o}}(v)-2}})^{{\mathrm{o}}(u)} (u^{\varphi v^{{\mathrm{o}}(v)-1}})^{{\mathrm{o}}(u)})(e)\\ &=& (e) (e), \end{eqnarray*} since $U$ is abelian, as required. \end{proof} For later use we recall a couple facts concerning involving semidirect products. \begin{lem}\cite[Theorem 10.30]{r} (The Schur--Zassenhaus theorem) \label{lem:schurzass} Let $G$ be a finite group, and let $K$ be a normal subgroup of $G$ with $(|K|, |G:K|) = 1$. Then $G$ is a semidirect product of $K$ and $G/K$. In particular, there exists a subgroup $H$ of $G$ with order $|G : K|$ satisfying $G = HK$ and $H\cap K=\{e\}$. \end{lem} \begin{lem}\cite[Theorem 1.2 (i)]{b1} \label{lem:normalindex} Let $p$ be an odd prime, and suppose that $P$ is a non-abelian $p$-group with a cyclic subgroup of index $p$. Then $P \cong \mathbb{Z}_{p^{\alpha-1}}\rtimes_\varphi \mathbb{Z}_{p}$, the center of $P$ has order $|Z(P)| = p^{\alpha - 2}$, and $P$ has presentation \[ P \cong M_{p^{\alpha}} = \langle a,b | a^{p^{\alpha - 1}} = b^{p} = 1 , b^{-1}ab = a^{1+p^{\alpha - 2}} \rangle. \] \end{lem} \section{Some group theory} We need a few more results from group theory. \begin{defn} \label{def:pcomp} Let $p$ be a prime. Let $G$ be a finite group, and let $P$ be a Sylow $p$-subgroup of $G$. A {\em $p$-complement} in $G$ is a subgroup with index equal to the order of a Sylow $p$-subgroup. \end{defn} \begin{thm} \cite[Theorem 10.21]{r} (Burnside's transfer theorem) With the notation of Definition \ref{def:pcomp}, if $P\subseteq Z(N_G(P))$, then $G$ has a normal $p$-complement. \end{thm} \begin{cor} \cite[Corollary 10.24]{r} \label{cor:qcomplementsmall} With the notation of Definition \ref{def:pcomp}, suppose $G$ contains a cyclic $q$-Sylow subgroup, where $q$ is the least prime divisor of $|G|$. Then $G$ has a normal $q$-complement. \end{cor} \begin{lem} \label{lem:normalS} A finite group containing a cyclic subgroup of prime index has a non-trivial normal Sylow subgroup. \end{lem} \begin{proof} Let $G$ be a finite group and suppose $C$ is a cyclic subgroup of prime index $p$ in $G$. Induct on the number of distinct prime factors of $|G|$. If $G$ is a $p$-group, then $G$ itself is a normal Sylow $p$-subgroup. Assume that there is a prime $r$ different from $p$ which divides $|G|$. Let $R$ be a Sylow $r$-subgroup of $C$. Now $|G:R| = p\cdot |C:R|$ is not divisible by $r$, so $R$ is a Sylow $r$-subgroup of $G$. We are done if $R$ is normal in $G$, so assume that this is not the case. Now $R$ is normal in the cyclic subgroup $C$, so $C\leq N_G(R)< G$. Since $C$ has prime index in $G$, it must be the case that $N_G(R)=C$. In particular, $R\leq Z(N_G(R))$, so $G$ has a normal $r$-complement $N$ by Burnside's transfer theorem. Thus $RN=G$, and hence $CN=G$ as well. Now $N\cap C$ is cyclic since $C$ is, and $|N:N\cap C|=|CN:C|=|G:C| = p$. Suppose $N\cap C$ is nontrivial. Then by induction $N$ has a nontrivial normal Sylow subgroup $S$. In fact, $S$ is characteristic in $N$, so $S$ is normal in $G$. Observe that $S$ is a Sylow subgroup of $G$ since $|G:N|$ and $|N|$ are coprime. Suppose $N \cap C = 1$. Then $|G|=|NC|=|N||C|/|N\cap C| = (|G|/|R|)(|G|/p)/1$, so $|G|=p r^\alpha$ for some positive power $\alpha$. Now $R$ and $P$ are cyclic subgroups of $G$, so Corollary \ref{cor:qcomplementsmall} implies that if $r < p$ then $G$ has a normal $r$-complement, and if $p < r$ then $G$ has a normal $p$-complement. In either case, there is a normal Sylow subgroup of $G$. \end{proof} \begin{lem} \label{lem:divideorder2} Let $G$ be a finite group of order $n$. Suppose $n$ is not a power of 2, and let $r>2$ be a prime divisor of $n$. Suppose that $G$ contains both a cyclic $r$-complement and a normal $r$-Sylow subgroup which itself contains a cyclic subgroup of index $r$. Then there is a bijection $\lambda : G \rightarrow \mathbb{Z}_{n}$ such that ${\mathrm{o}}(g)|{\mathrm{o}}(\lambda(g))$ for each $g \in G$. \end{lem} \begin{proof} Let $U$ be a cyclic $r$-complement of $G$. Let $R$ denote a normal Sylow $r$-subgroup of $G$ which contains a cyclic subgroup $C$ with index $r$. Say $|R|=r^\alpha$. By consideration of group orders, $G=RU$ and $R\cap U=\{e\}$. Since $R$ is normal in $G$, both $G = R\rtimes_\varphi U$ and $G/R\cong U$. Suppose $R$ is nonabelian. By Lemma \ref{lem:normalindex}, the center $Z(R)$ of $R$ has order $r^{\alpha-2}$. Now $Z(R)\subseteq C$ since otherwise by consideration of indices $R=Z(R)C$; this would imply that $R$ is abelian, contrary to our assumption. Hence $Z(R)$ is cyclic; moreover, it is the unique subgroup of $C$ with index $r$. Now $Z(R)$ is a characteristic subgroup of the normal (characteristic) subgroup $R$, so $Z(R)$ is normal in $G$. Since $Z(R)$ is a normal cyclic subgroup of $G$, Lemma \ref{lem:divideorder1} gives that ${\mathrm{o}}_{\ast}(ab) | {\mathrm{o}}_{\cdot}(ab)$ for all $ab\in Z(R)\rtimes_\varphi U$. Observe that $Z(R)\times U$ is cyclic. Thus we may take $\lambda$ restricted to $Z(R)\rtimes_\varphi U$ to be a bijection from $Z(R)\rtimes_\varphi U$ to $Z(R)\times U \cong\mathbb{Z}_{|Z(R)|}\times \mathbb{Z}_{|U|} \subseteq \mathbb{Z}_{n}$ for which ${\mathrm{o}}(x)|{\mathrm{o}}(\lambda(x))$ for each $x\in Z(R)\rtimes_\varphi U$. Now since $G = RU$, we can write $G = Z(R)U \cup (R - Z(R))U$. Suppose that $x \in (R - Z(R))U$. Then $x = ru$ where $r \in R - Z(R)$ and $u \in U$. Since $R$ is not cyclic it contains no element of order $r^\alpha$. Hence the order of $x$ is divisible by $r^{\alpha - 1}{\mathrm{o}}(u)$. We identify $\mathbb{Z}_{|U|}$ with $U$, so $\mathbb{Z}_n = \mathbb{Z}_{|Z(R)|}U \cup (\mathbb{Z}_{r^{\alpha}} - \mathbb{Z}_{|Z(R)|})U$. Since the elements of $\mathbb{Z}_{r^{\alpha}} - \mathbb{Z}_{|Z(R)|}$ have order $r^{\alpha - 1}$ or $r^{\alpha}$ it follows that $r^{\alpha -1}{\mathrm{o}}(u) | {\mathrm{o}}(z){\mathrm{o}}(u)$ for each $z \in \mathbb{Z}_{r^{\alpha}} - \mathbb{Z}_{|Z(R)|}$. Thus any bijection from $G-(Z(R)\rtimes_\varphi U)$ to $\mathbb{Z}_{n}-\mathbb{Z}_{|Z(R)|}\mathbb{Z}_{|U|}$ will extend $\lambda$ to all of $G$ with the desired properties. Now suppose $R$ is abelian. If $R$ is cyclic then by Lemma \ref{lem:divideorder1} we have nothing to prove. Suppose $R$ is not cyclic. Then Lemma \ref{lem:divideorder1} gives that ${\mathrm{o}}_{\ast}(ab) | {\mathrm{o}}_{\cdot}(ab)$ for all $ab\in R\rtimes_\varphi U$. Now it is enough to show that there is a bijection from $R\times U$ to $\mathbb{Z}_n$ with desired property. By the assumption $R$ has a cyclic subgroup of index $r$. Thus $R \cong \mathbb{Z}_{r^{\alpha - 1}} \times \mathbb{Z}_r$, and $R\times U \cong \mathbb{Z}_{r^{\alpha - 1}} \times \mathbb{Z}_r \times U$. Now if we write $R\times U = (R - \mathbb{Z}_{r^{\alpha - 1}})U \cup \mathbb{Z}_{r^{\alpha - 1}}U$ then arguing as in the above we reach the same conclusion. \end{proof} \begin{lem} \label{lem:orddivbij:E=E,cyc} Let $G$ be a finite group, and suppose that there is a bijection $\lambda : G \rightarrow \mathbb{Z}_{n}$ such that ${\mathrm{o}}(g)|{\mathrm{o}}(\lambda(g))$ for each $g \in G$. Then $|E(G)| =|E(\mathbb{Z}_n)|$ if and only if $G\cong \mathbb{Z}_n$. \end{lem} \begin{proof} Suppose $|E(G)| =|E(\mathbb{Z}_n)|$. By Lemma \ref{lem:d(r)}, for all $g\in G$, $2{\mathrm{o}}(g) -\phi({\mathrm{o}}(g)) - 1 \leq 2{\mathrm{o}}(\lambda(g)) - \phi({\mathrm{o}}(\lambda(g))) -1$. By (\ref{eq:EGsize}), and since $\lambda$ is a bijection 2$|E(G)| = \sum_{g \in G}2{\mathrm{o}}(g) - \phi({\mathrm{o}}(g)) - 1 = \sum_{g \in G}2{\mathrm{o}}(\lambda(g)) - \phi({\mathrm{o}}(\lambda(g))) - 1 = 2|E(\mathbb{Z}_{n})|$. This equality and the preceding inequality imply that for all $g\in G$, $2{\mathrm{o}}(g) - \phi({\mathrm{o}}(g)) = 2{\mathrm{o}}(\lambda(g)) - \phi({\mathrm{o}}(\lambda(g)))$. Pick a generator $z$ of $\mathbb{Z}_n$, and let $g= \lambda^{-1}(z)$. Then $2{\mathrm{o}}(g) -\phi({\mathrm{o}}(g)) = 2{\mathrm{o}}(\lambda(g)) - \phi({\mathrm{o}}(\lambda(g))) = 2{\mathrm{o}}(z) - \phi({\mathrm{o}}(z)) =2n - \phi(n)$. Suppose $G$ is not cyclic. Then ${\mathrm{o}}(g)<n$ and ${\mathrm{o}}(g)$ divides $n={\mathrm{o}}(z)$. Lemma \ref{lem:d(r)} implies that $2{\mathrm{o}}(g) -\phi({\mathrm{o}}(g)) <2n - \phi(n)$. This contradicts the above. Thus $G$ must be cyclic, and hence isomorphic to $\mathbb{Z}_n$. The converse is clear. \end{proof} \section{Proof of main theorem} The proof of Theorem \ref{thm:main} is developed in a series of technical lemmas. \begin{nta} \label{nta:groupnwfactors} Let $G$ be a finite group of order $n$, and adopt the conventions of Notation \ref{nta:nfactored} for the prime factorization of $n$. \end{nta} \begin{lem} \label{lem:cyc,primnotce} With Notation \ref{nta:groupnwfactors}, the following hold. \begin{enumerate} \item No cyclic group is a counterexample to Theorem \ref{thm:main}. \item No group of prime power order is a counterexample to Theorem \ref{thm:main}. \end{enumerate} \end{lem} \begin{proof} Part (i) is clear, and Part (ii) follows from Theorem \ref{lem:PGcyclicprimepower1}. \end{proof} \begin{lem} \label{lem:cynormsylnotce} With Notation \ref{nta:groupnwfactors}, suppose $G=P\rtimes_\phi T$ is the semidirect product of a normal cyclic Sylow subgroup $P$ and a subgroup $T$ with order coprime to that of $P$. Then $G$ is not a minimal counterexample to Theorem \ref{thm:main}. \end{lem} \begin{proof} Suppose $G$ is a minimal counterexample to Theorem \ref{thm:main}. By Lemma \ref{lem:Semidirprodcyc}, $|E(P\times T)|\geq |E(P\rtimes_\varphi T)| = |E(G)| \geq |E(\mathbb{Z}_{n})|$. Note that $P$ is isomorphic to $\mathbb{Z}_{|P|}$. By construction $\gcd(|P|,|T|)=1$. For the sake of comparison, let $T' = \mathbb{Z}_{|T|}$, and observe that $P\times T'$ is isomorphic to $\mathbb{Z}_{n}$. Identify $\mathbb{Z}_{n}$ and $P \times T'$. Suppose $T$ is not cyclic. Since $|T|<|G|$, and since $G$ is assumed to be a minimal counterexample, $|E(T)|< |E(T')|$. Also, by Theorem \ref{cor:maximume}, $|\overrightarrow{E}(T)| < |\overrightarrow{E}(T')|$. Now by Corollary \ref{cor:whenprodhasmoreedges}, $|E(P\times T)| < |E(P\times T')|= |E(\mathbb{Z}_{n})|$. This implies that $|E(\mathbb{Z}_{n})|>|E(\mathbb{Z}_{n})|$, which is absurd. This contradiction leads to the conclusion that in this case, $G$ is not a minimal counterexample to Theorem \ref{thm:main}. Suppose $T$ is cyclic. Then $G$ is the semidirect product of cyclic subgroups of coprime orders. Note that $G$ is not cyclic by Lemma \ref{lem:cyc,primnotce}. Now is $P\times T$; in fact, $P\times T$ is isomorphic to $\mathbb{Z}_n$ since $\gcd(|P|,|T|)=1$. Thus $|E(G)| \leq |E(\mathbb{Z}_{n})|= |E(P\times T)|$. It remains to show that equality does not hold. Suppose $|E(G)| = |E(\mathbb{Z}_{n})|$. Since $T$ is a cyclic $p$-complement and $P$ is a cyclic Sylow $p$-subgroup, Lemma \ref{lem:divideorder1} gives that there is a bijection $\theta:G\rightarrow\mathbb{Z}_{n}$ such that ${\mathrm{o}}(g) \ | \ {\mathrm{o}}(\theta(g))$ for all $g \in G$. Now Lemma \ref{lem:orddivbij:E=E,cyc}, leads to the conclusion that $G$ is cyclic. This contradiction implies that in this case, $G$ is not a minimal counterexample to Theorem \ref{thm:main} either. The result follows. \end{proof} \begin{lem} \label{lem:largeordntce} With Notation \ref{nta:groupnwfactors}, if there exists $g\in G$ with ${\mathrm{o}}(g) > n / p$, then $G$ is not a minimal counterexample to Theorem \ref{thm:main}. \end{lem} \begin{proof} Since ${\mathrm{o}}(g) > n / p$, we have $|G:\langle g\rangle| =|G|/{\mathrm{o}}(g) < p$. Thus $\langle g \rangle$ contains a Sylow $p$-subgroup $P$ of $G$. In particular, $P$ is cyclic and normalized by $\langle g\rangle$; hence, $P \ \lhd \ G$. By Lemma \ref{lem:schurzass}, $G = P\rtimes_\varphi T$ (semidirect product). The result follows from Lemma \ref{lem:cynormsylnotce}. \end{proof} \begin{lem} \label{lem:lemma1} With Notation \ref{nta:groupnwfactors}, if $\phi(n) \leq \frac{n}{q}$, then $G$ is not a minimal counterexample to Theorem \ref{thm:main}. \end{lem} \begin{proof} Suppose $G$ is a minimal counterexample to Theorem \ref{thm:main}. Then $|E(G)|\geq |E(\mathbb{Z}_n)|$. Note that $G$ is not cyclic and $n$ is not a prime power by Lemma \ref{lem:cyc,primnotce}. In particular, $n$ is not a power of 2. By assumption and Lemmas \ref{lem:edgesizes} and \ref{lem:avez}, \[ \sum_{g\in G} 2{\mathrm{o}}(g) - \phi({\mathrm{o}}(g))-1 = \sum_{g\in G} \deg(g) = 2|E(G)| \geq 2|E(\mathbb{Z}_n)| = \sum_{z\in \mathbb{Z}_n} \deg(z)\geq 2 \frac{n^2}{p} - \frac{n}{p}-1. \] Cancel -1 on the right and $-\phi({\mathrm{o}}(e))=-1$ on the left, and then drop the remaining $- \phi({\mathrm{o}}(g))$ on the left. Add $n= \sum_{g\in G} 1$ to both sides to find \[ \sum_{g\in G} 2{\mathrm{o}}(g) > 2 \frac{n^2}{p} - \frac{n}{p} +n. \] Thus there is at least one $g\in G$ for which ${\mathrm{o}}(g) > n / p$. The result follows from Lemma \ref{lem:largeordntce}. \end{proof} \begin{lem} \label{lem:cenotphi>n/q} With Notation \ref{nta:groupnwfactors}, if $\phi(n) > \frac{n}{q}$, then $G$ is not a minimal counterexample to Theorem \ref{thm:main}. \end{lem} \begin{proof} Suppose $G$ is a counterexample of minimal order. Then arguing as in Lemma \ref{lem:lemma1} we find \[ \sum_{g\in G} 2{\mathrm{o}}(g) - \phi({\mathrm{o}}(g))-1 \geq (n-1)\left(\frac{n}{q}+1\right). \] Each term $\phi({\mathrm{o}}(g))$ is at least one, so we may add $2n$ to both sides and then drop all remaining contributions of $-\phi({\mathrm{o}}(g))$. Also subtract $2$, the contribution of $g=e$, from each side: \[ \sum_{g\in G-\{e\}} 2{\mathrm{o}}(g) \geq (n-1)\left(\frac{n}{q}+1\right)+ 2n-2. \] Thus among the $n-1$ terms on the left, there is at least one nonidentity $g$ with ${\mathrm{o}}(g)\geq n/2q + 1/2 + 1$. In particular, ${\mathrm{o}}(g)> n/2q$. Suppose $p > 2q$. Then $n / p < n / 2q$, and thus $ {\mathrm{o}}(g) > n/2q > n/p$. In this case the result follows from Lemma \ref{lem:largeordntce}. Note $p \not= 2q$ since $p$ is prime. Now suppose $p < 2q$ (this case occurs when $n$ is the product of twin primes, for example). Then ${\mathrm{o}}(g) > n / 2q$, so $|G : \langle g \rangle| = |G|/{\mathrm{o}}(g) < n / (n/2q) = 2q$. Note that $n$ is odd by Lemma \ref{lem:phi(n)ineq}, and thus $|G : \langle g \rangle|$ is a prime number, say $s$. Now by Lemma \ref{lem:normalS}, $G$ has a normal $r$-Sylow subgroup $R$ for some prime $r$. By Lemma \ref{lem:schurzass}, $G$ contains an $r$-complement $U$. Thus $G=R\rtimes_\varphi U$. Note that $\langle g \rangle$ contains a Sylow $t$-subgroup of $G$ for $t\not=s$. If $r \neq s$, then $R\subseteq \langle g \rangle$, so it is cyclic. The hypotheses of Lemma \ref{lem:cynormsylnotce} are satisfied by $R$ and $U$, so we conclude that $G$ is not a minimal counterexample in this case. Suppose that $r = s$. Say $|R|=r^\beta$. Since $|G:\langle g\rangle|=r$, the subgroup $H=\langle g^{r^{\beta-1}}\rangle$ has order $|H|=|G|/r^\beta$. Thus $H$ is a cyclic $r$-complement. Note that $q>2$ by Lemma \ref{lem:phi(n)ineq}. Now Lemma \ref{lem:divideorder2} gives a bijection, say again $\lambda$, from $G$ to $\mathbb{Z}_{n}$ such that ${\mathrm{o}}(g) \ | \ {\mathrm{o}} (\lambda(g))$ for all $g\in G$. In this case Lemma \ref{lem:orddivbij:E=E,cyc} leads to the conclusion that $G$ is cyclic. This contradiction implies that $G$ is not a minimal counterexample to Theorem \ref{thm:main}. \end{proof} We are now ready to prove Theorem \ref{thm:main}. \begin{proof} Since Lemmas \ref{lem:lemma1} and \ref{lem:cenotphi>n/q} exhaust the possibilities, there is no counterexample of minimal order to Theorem \ref{thm:main}, and hence no counterexamples at all. Thus Theorem \ref{thm:main} holds for all groups of all orders. \end{proof} We may restate Theorem \ref{thm:main} without reference to power graphs as follows. \begin{cor} Let $G$ be a finite group of order $n$. Then $\displaystyle{ \sum_{d | n} (2d - \phi(d))\phi(d) \geq \sum_{g\in G} (2{\mathrm{o}}(g) - \phi({\mathrm{o}}(g)))}$, with equality if and only if $G$ is cyclic. \end{cor} \noindent{\bf Acknowledgment.} The second author thanks his advisors Professor Hassan Yousefi Azari and Professor Ali Reza Ashrafi for providing him with this question as a conjecture. We thank Professor I. Martin Isaacs, from the University of Wisconsin-Madison for his valuable comments. We thank Professor Andrei Kelarev from the University of Newcastle (Australia) for discussion on the background and uses of power graphs of groups.

https://arxiv.org/abs/2102.13193

Minimum Spanning Tree Cycle Intersection Problem

Consider a connected graph $G$ and let $T$ be a spanning tree of $G$. Every edge $e \in G-T$ induces a cycle in $T \cup \{e\}$. The intersection of two distinct such cycles is the set of edges of $T$ that belong to both cycles. We consider the problem of finding a spanning tree that has the least number of such non-empty intersections.

\section{Introduction} In this article we present what we believe is a novel problem in graph theory, namely the Minimum Spanning Tree Cycle Intersection (\emph{MSTCI}) problem. \bigskip The problem can be expressed as follows. Let $G$ be a graph and $T$ a spanning tree of $G$. Every edge $e \in G-T$ induces a cycle in $T \cup \{e\}$. The intersection of two distinct such cycles is the set of edges of $T$ that belong to both cycles. Consider the problem of finding a spanning tree that has the least number of such pairwise non-empty intersections. \bigskip The problem arose while investigating a (yet unpublished) method for \emph{mesh deformation} in the area of \emph{digital geometry processing}, see \cite{Botsch:2010}. The method requires to solve a linear system and the sparsity of the matrix is related to the solution of this problem. \bigskip The remaining of this section is dedicated to express the MSTCI problem in the context of well established theories to motivate other points of view. Section 2 sets some notation and convenient definitions. In Section 3 the complete graph case is analyzed. Section 4 presents a variety of interesting properties, and a conjecture in the slightly general case of a graph (not necessarily complete) that admits a star spanning tree. Section 5 explores programatically the space of spanning trees to provide evidence that the conjecture is well posed. Section 6 collects the conclusions of the article. \subsection{Intersection graph theory} \emph{Intersection graph theory} is a longstanding and central area of graph theory covered in important textbooks \cite{McKee:1999}. It is concerned with the study of \emph{intersection graphs}. \smallskip Let $\mathscr{F} = \{S_1,\dots,S_k\}$ be any family of sets. The \emph{intersection graph} denoted $\Omega(\mathscr{F})$ is the graph having $\mathscr{F}$ as vertex set with $S_i$ adjacent $S_j$ if and only if $i \neq j$ and $S_i \cap S_j \neq \varnothing$. \bigskip Let $G=(V,E)$ be a graph and $\mathscr{F} = \{S_1,\dots,S_k\}$ be the family of edge sets corresponding to the cycles of $G$. Then $\Omega(\mathscr{F})$ has as vertex set the cycles of $G$ and two vertices are adjacent if the corresponding cycles share at least one edge. In this setting, let $T$ be a spanning tree of $G$ and $G-T \subset E$ the set of complementary edges of $T$. An edge $e \in G-T$ induces a cycle in $T \cup \{e\}$ which obviously is a cycle of $G$. So $T$ has a canonical mapping to some subgraph of $\Omega(\mathscr{F})$. The MSTCI problem can be expressed in the following terms: find the spanning tree $T$ such that it maps to the sparsest possible subgraph in $\Omega(\mathscr{F})$. \subsection{Matroid theory} The classical \emph{matroid theory} as developed by Tutte in \cite{Tutte:1965} is a fundamental theoretical toolbox with very deep insights in graph theory. \bigskip The familiy of sets described in the previous subsection closely resembles the \emph{polygon matroid}. An interesting formulation of the MSTCI problem can be expressed in terms of its dual matroid $B(G)$, namely the \emph{bond matroid}. \bigskip The bond matroid can be defined as follows. Let $G=(V,E)$ be a connected graph. The \emph{atoms} of $B(G)$ are the edge subsets $A \subset E$ such that $G-A$ determine two connected components $G_1$ and $G_2$ and such that every edge of $A$ has one end in each component. \bigskip A \emph{dendroid} $D$ of a matroid $M$ is a set that intersects all the atoms and is minimal, meaning that if we delete an element of $D$ then there exists an atom $A \in M$ such that $A \cap D = \varnothing$. The dendroids of $B(G)$ are the sets of edges of the spanning trees of $G$. \bigskip Let $T$ be a spanning tree of $G$. So the edges of $T$ define a dendroid $D$ of $B(G)$. Note that for every edge $e \in T$, $T-\{e\}$ determine two subtrees that span two connected components $G_1$ and $G_2$ of $G$. So there is a natural injective map between the edges of $T$ and the atoms of $B(G)$: $$\phi_T: E(T) \rightarrow B(G)$$ $$e \mapsto A$$ where $A$ is the atom corresponding to the set of edges linking $G_1$ and $G_2$. As a remark, note that $e \in A$. In the language of matroid theory this set of $|V|-1$ atoms (ie. the image of $\phi_T$) of $B(G)$ is called the \emph{dendroid basis} determined by $D$. \bigskip To formulate the MSTCI problem in this framework we have to be precise about the pairwise intersection of cycles. In this sense let $T$ be a spanning tree of a connected graph $G=(V,E)$. Let $S$ be the dendroid basis determined by the edges of $T$ and $A \in S$ an atom. Clearly there exist a unique edge $e \in T$ such that $\phi_T(e) = A$. Note that each edge $e' \in A-\{e\}$ determine a cycle $c \in T \cup \{e'\}$ and that $e \in c$. So two such cycles have non-empty intersection. It is not difficult to realize that every non-empty pairwise intersection is of this form. If we manage to count the set of this pair of edges of all the atoms of $S$ we could express the MSTCI problem as an alternative minimization problem: find the spanning tree such that its corresponding dendroid basis has the least number of such pair of edges. \subsection{Homology theory} The importance of \emph{homological methods} \cite{Cartan:1956} in topology and geometry cannot be overemphasized. These methods constitute fundamental algebraic tools that enable the computation of invariant quantities of spaces. In its original form the main invariant was the number of ``holes'' of a space known as its \emph{Betti numbers}. \bigskip We introduce some elementary notions based these notes \cite{Dewan:2010}. Let $R$ be a commutative ring, and suppose we have a sequence of $R$-modules $M_i$ and homomorphisms $d_i$ $$\dots \xrightarrow{d_3} M_2 \xrightarrow{d_2} M_1 \xrightarrow{d_1} M_0 \xrightarrow{d_0} M_{-1} \xrightarrow{d_{-1}} \dots$$ such that $\forall i \ d_i \circ d_{i+1} = 0$. Such a sequence is called a \emph{chain complex} of $R$-modules, and denoted $M_\bullet$. Because the composition of adjacent homomorphisms is trivial, $im(d_{i+1}) \subseteq ker(d_i)$. Therefore, for each module in the chain there is a quotient $$H_i(M_\bullet) = \frac{ker(d_i)}{im(d_{i+1})}$$ called the \emph{i-th homology group} of $M_\bullet$. \smallskip In graph theory this definitions become very concrete. Let $R$ be a \emph{principal ideal domain} (ie. $\mathbb{Z}$) and let $G=(V,E)$ be a graph. Consider an arbitrary orientation of the edges that maps every edge $e \in E$ to a triple: $e \mapsto (e,s,t)$, where $s,t \in V$ are the source and target ends of the orientation of $e$. Let $M_0$ be the free $R$-module over $V$ (formal linear combinations of the vertices of $G$). And $M_1$ the $R$-module generated by the oriented edges subject to the relations $(e,s,t) + (e,t,s) = 0$ (where $(e,t,s)$ expresses de traversal of the edge $(e,s,t)$ in the opposite direction). Now consider the \emph{boundary homomorphism} $$\partial: M_1 \rightarrow M_0$$ defined on the generators as $\partial(e,s,t) = t - s$. Not surprisingly $ker(\partial)$ is denoted the \emph{cycle space} of $G$, $\mathscr{C}(G)$. Since if we consider in $M_1$ the linear combination representing a cycle $c$ of $G$ then $\partial(c) = 0$. It is not difficult to check that $rank(\mathscr{C}(G)) = |E| - |V| + c$ where $c$ is the number of connected components of $G$. Now we can define the following chain complex $G_\bullet$: $$\dots \xrightarrow{d_3 = 0} 0 \xrightarrow{d_2 = 0} M_1 \xrightarrow{d_1 = \partial} M_0 \xrightarrow{d_0 = 0} 0 \xrightarrow{d_{-1} = 0} \dots$$ As $im(d_2) = 0$ then $H_1(G_\bullet) = ker(\partial) = \mathscr{C}(G)$. \smallskip An interesting result is the following. Let $T$ be a spanning tree of $G$. Note that if $e \in G-T$ then $T \cup {e}$ has only one cycle: $c_e$. The set of those cycles generate $\mathscr{C}(G)$. Symbolically: $$<\{c_e \ \forall e \in G-T \}> = \mathscr{C}(G) = H_1(G_\bullet)$$ In other words: each spanning tree determines some basis of $H_1(G_\bullet)$. In particular if $T$ is a spanning tree that is a solution of the MSTCI problem then the \emph{tree intersection number} (defined in the next section) of $T$ could be a finer invariant of $G$. \section{Preliminaries} \subsection{Overview} In the first part of this section we present some of the terms that are used in the article. Then we define the notion of \emph{closest-point} and \emph{closest-point-set}. Finally we show a convenient cycle partition. \subsection{Notation} Let $G=(V,E)$ be a graph and $T$ a spanning tree of $G$, then we will refer to the edges $e\in T$ as \emph{tree-edges} and to the edges $e \in G-T$ as \emph{cycle-edges}. \smallskip Every cycle-edge $e$ induces a cycle in $T \cup \{e\}$, we will call such a cycle a \emph{tree-cycle}. And we shall call $C_T$ to the set of tree-cycles of $T$. \smallskip The intersection of two tree-cycles is the set of edges of $T$ that belong to both cycles. We will define three functions concerning the intersection of tree-cycles. \smallskip The first is $\cap_{T}(\cdot, \cdot): C_T \times C_T \rightarrow \{0,1\}$ $$\cap_{T}(c_i, c_j) := \begin{cases} 1 & c_i \cap c_j \neq \varnothing \land c_i \neq c_j \\ 0 & c_i \cap c_j = \varnothing \lor c_i = c_j \\ \end{cases} $$ Note that the trivial case $c_i = c_j$ is excluded. This arbitrary decision will simplify future computations. \smallskip The second is $\cap_{T}(\cdot): C_T \rightarrow \mathbb{N}$ $$\cap_{T}(c_i) := \sum_{c_j \in C_T} \cap_{T}(c_i, c_j)$$ We will call $\cap_{T}(c)$ the \emph{cycle intersection number} of $c$. Given a tree-cycle $c$ we will denote $\cap_{T,c}$ as the set of tree-cycles that have non-empty intersection with $c$. More precisely: $$\cap_{T,c} \equiv \{c' \in C_T: \cap_{T}(c, c') = 1 \}$$ Note that $|\cap_{T,c}| = \cap_{T}(c)$. \smallskip To define the third function consider $\mathscr{T}_G$ to be the set of spanning trees of $G$, so the definition will be as follows: $\cap_G: \mathscr{T}_G \rightarrow \mathbb{N}$ $$\cap_G(T) := \frac{1}{2}\sum_{c \in C_T} \cap_{T}(c)$$ We will call $\cap_G(T)$ the \emph{tree intersection number} of $T$. Clearly the set $\displaystyle \min_{T}\cap_G(T)$ is the set of solutions of the MSTCI problem. If the graph is clear from the context we could drop the subindex and simply write: $\cap(T)$. \bigskip We shall call \emph{star} spanning tree to a spanning tree that has one vertex that connects to all other vertices. And $K_n$ to the complete graph on $n$ nodes. If $G=(V,E)$ we will say that $|V| = n$ is the number of vertices of $G$, $|c| = k$ is the length of the cycle $c$ and $|p|$ is the length of the path $p$. Also $uTv$ will denote the unique path between $u,v \in V$ in the spanning tree $T$; $d_T(v)$ will be the degree of $v \in V$ relative to the spanning tree $T$. We will denote $N(v)$ to the set of neighbor nodes of $v \in V$. \smallskip Finally we use the terms ``node'' and ``vertex'' interchangeably. \subsection{Closest point} In this section we prove the following simple fact: if $G=(V,E)$ is a connected graph, $T$ a spanning tree of $G$ and $c \in C_T$ a tree-cycle, then for every node $v \in V$ there exists a unique node $w \in c$ that minimizes the distance to $v$ in $T$. We shall denote that node $closest-point(v,c)$. \bigskip \begin{lemma} Let $G=(V,E)$ be a connected graph, $T$ a spanning tree of $G$ and $c \in C_T$ a tree-cycle then for every node $v \in V$ there exists a unique node $w \in c$ such that $$|vTw| \leq |vTu| \ \forall u \in c$$ Proof. The proof proceeds by contradiction. If $v \in c$ it is obviously its own unique closest point. Suppose that $v \notin c$ and that there are two distinct nodes $w, w' \in c$ such that $|vTw| = |vTw'| \leq |vTu| \ \forall u \in c$. Obviously $w' \notin vTw$ and $w \notin vTw'$, we conclude that $vTw \cup wTw' \cup vTw'$ determine a cycle in $T$ which contradicts the fact that $T$ is a tree.$\square$ \end{lemma} The uniqueness of the $closest-point(v,c)$ leads to the following definition. \begin{definition} Let $G=(V,E)$ be a connected graph, $T$ a spanning tree of $G$ and $c \in C_T$ a tree-cycle, then the set of closest points to a node $w \in c$ is defined as follows $$closest-point-set(u,c) := \{v \in V-c: $$ $$closest-point(v,c)=u \}$$ \end{definition} \subsection{Tree cycle intersection partition} Now we define a partition of the set $\cap_{T,c}$. More precisely, let $G$ be a connected graph, $T$ a spanning tree of $G$ and $c \in C_T$ a tree-cycle. As defined above the set $\cap_{T,c}$ is the set of tree-cycles that have non-empty intersection with $c$. \bigskip Let us consider any tree-cycle $c' \in \cap_{T,c}$ induced by a cycle-edge $e=(v,w)$. In this setting we can define the following partition: \begin{itemize} \item \emph{Internal tree-cycles}: $c'$ is \emph{internal} if $v, w \in c$. \item \emph{External tree-cycles}: $c'$ is \emph{external} if $v \notin c$ and $w \in c$. \item \emph{Transit tree-cycles}: $c'$ is \emph{transit} if $v, w \notin c$. \end{itemize} Let us denote them $\cap^i_{T,c}$, $\cap^e_{T,c}$, $\cap^t_{T,c}$, respectively. This partition will be convenient to simplify the computation of the \emph{intersection number} of $c$. \section{Tree cycles of complete graphs} \subsection{Overview} In this section we analyze the complete graph case $G=K_n$. First we deduce a formula to compute the cycle intersection number. Then we prove that the tree-cycles of a star spanning tree achieve the minimum cycle intersection number. Finally we conclude that the star spanning trees are the unique solutions of the MSTCI problem. \subsection{Cycle intersection number formula} In this subsection we consider the problem of finding a formula to count tree-cycle intersections. More precisely, let $G=K_n$, $T$ a spanning tree of $G$ and $c$ a tree-cycle, we intend to derive a formula to calculate $\cap_{T}(c)$. \bigskip The idea behind the formula is to consider the partition of $\cap_{T,c}$, defined in the previous section. And then by combinatorial arguments compute the number of elements in each class. \smallskip We shall analyze in turn the three classes: $\cap^i_{T,c}$, $\cap^e_{T,c}$, $\cap^t_{T,c}$. In this section we will consider $c' \in \cap_{T}(c)$ to be a tree-cycle induced by a cycle-edge $e=(v,w)$. \bigskip The simplest case is the internal tree-cycles class: $\cap^i_{T,c}$. Let $c'$ be an internal tree-cycle. By definition the nodes $v$ and $w$ belong to $c$, so the following holds: $(c' \cap T) \subset c$ because there is a unique path from $v$ to $w$ in $T$. So basically counting the number of internal tree-cycles reduces to count the pairings of the nodes of $c$ excluding some obvious cases. The cases that should be excluded are: the pairing of a node with itself and with its neighbors in $c$. Then the number of internal tree-cycles is: $$|\cap^i_{T,c}| = \frac{(k-3) k}{2}, $$ where $k$ is, as before, equal to $|c|$. The quotient is obviously due to the fact that every cycle is counted twice. \smallskip Next we consider the class of external tree-cycles. Now let $c'$ be an external tree-cycle. In this case exactly one of the extremal nodes ($v$ or $w$) belong to $c$. Without loss of generality (as we are considering undirected edges), suppose that $v \notin c$ and $w \in c$. Clearly $w \neq closest-point(v,c)$ because in that case $c' \cap c = \varnothing$ and consequently $c' \notin \cap_{T,c}$ which contradicts our hypothesis. As that is the only particular case that should be excluded, the number of external tree-cycles is: $$|\cap^e_{T,c}| = (n-k) (k-1),$$ where $n=|V|$ is the number of vertices of $G$ and $k=|c|$ is the length of $c$. Last we consider the class of transit tree-cycles. In this case the key observation depends on the $closest-point-set$ definition of the previous section. Lets define two classes of cycle-edges: \begin{enumerate} \item A cycle-edge $e=(v,w)$ is called \emph{intraset cycle-edge} if both $v,w \in closest-point-set(u_i,c)$ for some $u_i \in c$ \item A cycle-edge $e=(v,w)$ is called \emph{interset cycle-edge} if $v \in closest-point-set(u_i,c)$ and $w \in closest-point-set(u_j,c)$ where $u_i, u_j \in c$ and $u_i \neq u_j$ \end{enumerate} Then: \begin{itemize} \item Every \emph{intraset cycle-edge} induce a tree-cycle $c'$ such that $c' \cap c = \varnothing$ \item Every \emph{interset cycle-edge} induce a tree-cycle $c'$ such that $c' \cap c \neq \varnothing$ \end{itemize} So we should consider \emph{interset cycle-edges} or equivalently, the pairing of the nodes that are in different sets. Let $q_i = |closest-point-set(w_i,c)|$ be defined for all $w_i \in c$, then the number of transit tree-cycles is: $$|\cap^t_{T,c}| = \sum_{i<j} q_i q_j = \frac{1}{2} \sum_{i=1}^k q_i (n-k-q_i)$$ Finally, the intersection number formula is the aggregation of the three classes: $$\cap_{T}(c) = |\cap_{T,c}| = |\cap^i_{T,c}| + |\cap^e_{T,c}| + |\cap^t_{T,c}| =$$ $$\frac{(k-3) k}{2} + (n-k) (k-1) + \frac{1}{2} \sum_{i=1}^k q_i (n-k-q_i),$$ where $n$ is the number of vertices of $G$, $k=|c|$ and $q_i = |closest-point-set(w_i,c)|$ for $w_i \in c$. \subsection{Main result} In this subsection we start by defining \emph{transiteless} tree-cycles. Then we prove two lemmas. The first shows that for every cycle $c \in G=K_n$ we can build a spanning tree $T$ such that $c$ is a tree-cycle of $T$ and the intersection number $\cap_{T}(c)$ is minimal. And the second calculates the intersection number of tree-cycles of star spanning trees. Finally we prove the main result of this section, namely that star spanning trees minimize $\cap(\cdot)$ in the case of complete graphs. \bigskip \begin{definition} Let $G=(V,E)$ be a connected graph, $T$ a spanning tree of $G$ and $c \in C_T$ a tree-cycle, we call $c$ a \emph{transitless} tree-cycle if $|\cap^t_{T,c}| = 0$. \end{definition} As an important remark, note that the number of elements in the internal and external classes of $c$ are independent of the spanning tree because they depend exclusively on the numbers $n=|V|$ and $k=|c|$. So two spanning trees $T_1$ and $T_2$ that have $c$ as a tree-cycle induce an intersection number (for $c$) that only differ in the number of elements in their transit classes. We conclude that transitless tree-cycles have minimal intersection number $\cap_{T}(c)$. \begin{lemma}\label{lemmatransitless} Let $G=K_n$ and let $c$ be a cycle of $G$ then the following procedure lead to a spanning tree $T$ that minimizes the intersection number of $c$: \begin{enumerate} \item Remove exactly one edge $e \in c$ \item Choose a vertex $v \in c$ \item Define $T$ as follows: \begin{itemize} \item The edges $c - e$ \item A star centered at $v$ of the vertices $(G-c) \cup \{v\}$ \end{itemize} \end{enumerate} \smallskip Proof. Note that $T$ is a spanning tree of $G$, and $c$ is a tree-cycle of $T$. So if we prove that $|\cap^t_{T,c}| = 0$ then the intersection number $\cap_{T}(c)$ is minimal. This is the case, because by construction: \begin{itemize} \item $|closest-point-set(u,c)| = 0$ $\forall u \in c, u \neq v$ \item $|closest-point-set(v,c)| = n-k$ \end{itemize} So $|\cap^t_{T,c}| = \sum_{i<j} q_i q_j = 0$.$\square$ \end{lemma} \begin{lemma} Let $G=K_n$ and let $T_s$ be a star spanning tree of $G$. Then the the following property holds $$\cap_{T_s}(c) = 2(n - 3)$$ for any tree-cycle $c$ of $T_s$. \smallskip Proof. Clearly the tree-cycles in $T_s$ have the same intersection number (by symmetry). Let $c$ be a tree-cycle of $T_s$. Note that $c$ is a triangle ($|c| = 3$), so the corresponding internal tree-cycles class is empty: $|\cap^i_{T,c}| = 0$. Also note that $c$ is a transitless tree-cycle because its nodes are: the central node and two leaf nodes of $T_s$. So the external tree-cycle class is the only non-empty class: $$\cap_{T_s}(c) = |\cap^e_{T_s,c}| = 2(n - 3)$$ \end{lemma} \begin{proposition}\label{propstar} Let $G=K_n$ and let $T_s$ be a star spanning tree of $G$. Then the following property holds $$\cap_{T_s}(\cdot) \leq \cap_{T}(\cdot)$$ where $T$ is any spanning tree of $G$. \smallskip Proof. We shall prove the proposition by contradiction. Suppose that a spanning tree $T$ and a tree-cycle $c$ of $T$ exist such that: $$\cap_{T}(c) < \cap_{T_s}(\cdot) = 2(n - 3)$$ We can assume that $c$ is transitless because, if it's not the case, by lemma \ref{lemmatransitless} we can build a spanning tree $T'$ such that $\cap_{T'}(c) < \cap_{T}(c)$. In this context the inequality can be expressed as $$\cap_{T}(c) = |\cap^i_{T,c}| + |\cap^e_{T,c}| = $$ $$\frac{(k-3) k}{2} + (n-k) (k-1) < 2(n - 3)$$ Expanding and simplifying the expression we have $$\frac{-1}{2} k^2 + (n-\frac{1}{2}) k - 3n + 6 < 0$$ The roots of this quadratic polynomial are: $r_1 = 3$ and $r_2 = 2(n-2)$. We should consider two cases depending on the relation of the roots: \begin{enumerate} \item $r_1 < r_2$ \item $r_1 > r_2$ \end{enumerate} The case $r_1 = r_2$ can be discarded because it leads to a fractional number of nodes ($n = \frac{7}{2}$). In the first case the inequality holds for $k < r_1 = 3$ or $k > r_2 = 2(n-2)$. The case $k < 3$ is an obvious contradiction since the size of the cycle must be $|c| = k \geq 3$. The case $k > 2(n-2)$ combined with the fact that $k \leq n$ induces the following inequality $$r_1 = 3 < r_2 = 2(n-2) < k \leq n$$ which implies a contradiction: $3 < n < 4$, since $n$ is a positive integer. The second case ($r_1 = 3 > r_2 = 2(n-2)$) imply $n < \frac{7}{2}$. So the only case that should be considered is $k = n = 3$ since $k \leq n$. But, this inequality is false because $k = 3$ is a root of the quadratic polynomial.$\square$ \begin{corollary}\label{coromain} Let $G=K_n$ and let $T_s$ be a star spanning tree of $G$. Then the the following property holds $$\cap(T_s) \leq \cap(T),$$ where $T$ is any spanning tree of $G$. \smallskip Proof. As expressed by proposition \ref{propstar}, a tree-cycle of a star spanning tree has the minimal intersection number among all tree-cycles. Since any tree-cycles of a star spanning tree has the same intersection number, we conclude that the tree intersection number of a star spanning tree $\cap(T_s)$ is minimal among all spanning trees. $\square$ \end{corollary} \smallskip This corollary can be further improved to a strict inequality. In other words: star spanning trees are the unique minimizers of $\cap(T)$. \smallskip \begin{corollary}\label{coromain2} Let $G=(V,E)=K_n$ where $|V| = n > 4$ and let $T_s$ be a star spanning tree of $G$. Then, the following property holds $$\cap(T_s) < \cap(T),$$ where $T$ is any non-star spanning tree of $G$. \smallskip Proof. A careful reading of proposition \ref{propstar} leads to the conclusion that the equality $\cap_{T_s}(c) = \cap_{T}(c)$ is achieved when $k$ is either $r_1=3$ or $r_2=2(n-2)$ (the roots of the quadratic polynomial). If $k = r_2 = 2(n-2)$, and taking into account that $3 \leq k \leq n$, we conclude that $\frac{7}{2} \leq n \leq 4$; this case is explicitly excluded from our hypotheses (in fact, it is not difficult to check that the three non-isomorphic spanning trees of $K_4$ all have the same tree intersection number). \smallskip The other possibility is $k = r_1 = 3$. As all the tree-cycles of $T_s$ fall into this category, it is enough to show that $T$ has a tree-cycle $c$ such that $|c| = k > 3$ to conclude our thesis. Let $w \in V$ be a node with maximum degree in $T$. And let $d_T(w)$ denote the degree of $w$ in $T$ and $N(w)$ to the set of neighbors of $w$ in $T$. As $T$ is a non-star spanning tree then $2 \leq d_{T}(w) < n-1$. So there is a node $u \in V$ such that $u \notin N(w)$ in $T$, and there is a node $k \in N(w)$ such that $k \notin wTu$. Notice that the edge $e=(u,k) \notin T$ (since it would induce a cycle). Hence, it is a cycle-edge. Note that the tree-cycle induced by $e$ has length at least 4.$\square$ \end{corollary} \end{proposition} This result can be summarized in the following way: star spanning trees are the unique solutions of the MSTCI problem for complete graphs. \section{Further generalization} \subsection{Overview} Now we explore some aspects of a slightly more general case, namely: the MSTCI problem in the context of a graph (not necessarily complete) $G=(V,E)$ that admits a star spanning tree $T_s$. In the first part we present a formula to calculate $\cap(T_s)$. In the second part we show that $\cap(T_s)$ is a local minimum in the domain of what we refer to as the ``spanning tree graph''. In the third part we prove a result that suggests a general observation: the fact that a spanning tree of a graph $G$ being a solution of the MSTCI problem doesn't depend on an intrinsic property of $T$ but on the particular embedding of $T$ in $G$. Finally we conjecture a generalization of Corollary \ref{coromain}: $\cap(T_s) \leq \cap(T)$ for every spanning tree $T$ of $G$. \subsection{Formulas for star spanning trees} In this subsection we present two formulas for graphs $G=(V,E)$ that admit a star spanning tree $T_s$. Let us denote $v \in V$ to the central node of $T_s$. \bigskip The first formula corresponds to the cycle intersection number of a tree-cycle $c = (u,v,w) \in C_{T_s}$, namely $\cap_{T_s}(c)$. Recall from the previous section that $c$ does not intersect neither transit nor internal tree-cycles: $|\cap^t_{T,c}| = 0$ and $|\cap^i_{T,c}| = 0$. So its non-empty intersections are the tree-cycles in the set $\cap^e_{T,c}$. Note that the remaining incident edges to $u$ and $w$, are the only source of tree-cycles that have non-empty intersection with $c$. So the formula is straightforward: $$\cap_{T_s}(c) = d(u) - 2 + d(w) - 2,$$ where $d(u)$ and $d(w)$ are the degrees of $u$ and $w$, resp. Now we shall deduce a formula for the tree intersection number $\cap(T_s)$. Based on the preceding observations and the fact that each node $u \in V - \{v\}$ is involved in $d(u)-1$ tree-cycles and for each of those tree-cycles it produces $d(u) - 2$ non-empty intersections (note that the pairwise intersections are counted twice). The formula is as follows: $$\cap(T_s) = \frac{1}{2} \sum_{u \in V - \{v\}} (d(u) - 1) (d(u) - 2) = $$ $$\frac{1}{2} \sum_{u \in V - \{v\}} d(u)^2 - 3 d(u) + 2$$ If we denote ${\bf d}$ as the degree vector of $G$, that is: a vector that has in the $i$-th component the degree of the $i$-th vertex. And taking into account that $\sum_{u \in V} d(u) = 2m$ where $m = |E|$, the formula can be expressed as: $$\cap(T_s) = \frac{1}{2} [||{\bf d}||_{2}^2 - 6m - (n-1)(n-6)]$$ \subsection{Star spanning tree as a local minimum} In this subsection we prove that a star spanning tree is a local minimum respect to the tree intersection number in the domain of the \emph{spanning tree graph}. We start by defining this second order graph of the original graph $G=(V,E)$. Then we analyze the structure of the neighbors of a star spanning tree $T_s$. Finally we demonstrate the result by a bijection between tree-cycles to conclude that $\cap(T_s)$ is a local minimum. \bigskip \begin{definition} Let $G=(V,E)$ be a graph, and $S$ a subgraph of $G$. We denote as $e \leftrightarrow e'$ to the operation of replacing the edge $e \in S$ with the edge $e' \in G-S$. We call this operation \emph{edge replacement on $S$}. \end{definition} \begin{definition} Let $G=(V,E)$ be a graph. We denote $SP_{_G}$ to the graph that has one node for every spanning tree of $G$ and an edge between two nodes if the corresponding spanning trees differ in exactly one edge replacement. We call this graph the \emph{spanning tree graph of $G$}. \end{definition} \subsubsection{Neighborhood of $T_s$} Let $G=(V,E)$ be a graph that admits a star spanning tree $T_s$ with $v \in V$ as its center. Let $\alpha_{_{T_s}}$ be the node corresponding to $T_s$ in $SP_{_G}$ and let $\alpha_{_T}$ (with corresponding spanning tree $T$) be any neighbor of $\alpha_{_{T_s}}$. By definition $T_s$ and $T$ differ in exactly one edge replacement $e \leftrightarrow e'$ where $e=(v,w) \in T_s$ and $e'=(u,w) \in T$. Note that $T$ is exactly the same as $T_s$ except that the node $w$ is no longer connected to the central node $v$ but is connected to the intermediate node $u$. This similar structure has direct consequences in the intersection numbers of both trees. \bigskip Now we prove the result of this section. \begin{theorem}\label{teo:local} Let $G=(V,E)$ be a graph that admits a star spanning tree $T_s$ with $v \in V$ as its center. Then, $T_s$ is a local minimum respect to the tree intersection number in the domain of $SP_{_G}$. \smallskip Proof. Let $T$ be a spanning tree corresponding to a neighbor of $T_s$ in $SP_{_G}$. Then we want to prove that $\cap(T_s) \leq \cap(T)$. We shall proceed by defining a bijection between the tree-cycles of both trees $\{c \leftrightarrow d: c \in C_{_{T_s}} \wedge d \in C_{_T} \}$ such that $\cap_{_{T_s}}(c) \leq \cap_{_{T}}(d)$, this strategy clearly implies the thesis since by definition: $$\cap(T_s) = \frac{1}{2}\sum_{c} \cap_{T_s}(c) \leq \frac{1}{2}\sum_{d} \cap_{T}(d) = \cap(T)$$ \smallskip Let $e_{_{T_s}} \leftrightarrow e_{_T}$ with $e_{_{T_s}}=(v,w) \in T_s$ and $e_{_{T}}=(u,w) \in T$ be the edge replacement in $SP_{_G}$. Consider the following simple facts: \begin{itemize} \item $e_{_{T}}$ is a cycle-edge in $T_s$, with corresponding tree-cycle $c$ \item $e_{_{T_s}}$ is a cycle-edge in $T$, with corresponding tree-cycle $d$ \item Except for $e_{_{T_s}}$ and $e_{_{T}}$, $T_s$ and $T$ have the same set of cycle-edges. For every $e \in E - T_s - T$ we denote $c_e$ and $d_e$ to the corresponding tree-cycles in $T_s$ and $T$, resp. \end{itemize} according to this naming convention, we can define the following ``natural'' bijection between tree-cyles: $$\{c \leftrightarrow d\} \cup \{c_e \leftrightarrow d_e: e \in E - T_s - T\}$$ In order to compare the intersection numbers of the bijected pairs it is convenient to distiguish the following partition: \begin{itemize} \item Case 1: the pair induced by the edge replacement, $\{c \leftrightarrow d\}$ \item Case 2: pairs induced by cycle-edges non-incident to $u$ nor to $w$, $\{c_e \leftrightarrow d_e: e \in E - T_s - T \wedge u \notin e \wedge w \notin e\}$ \item Case 3: pairs induced by cycle-edges incident to $u$ or $w$, $\{c_e \leftrightarrow d_e: e \in E - T_s - T \wedge (u \in e \lor w \in e) \}$ \end{itemize} Case 1 is the easiest: note that $c$ and $d$ are the same tree-cycle $(u,v,w)$, which is a transitless triangle, so its intersection number is determined by its external intersections: $$\cap_{_{T_s}}(c) = d(u) - 2 + d(w) - 2 = \cap_{_{T}}(d)$$ Case 2 is similar, let $e=(h,k)$ be a cycle-edge non-incident to $u$ or to $w$ and $c_e \leftrightarrow d_e$ its corresponding pair of bijected tree-cycles. Clearly $e$ determines the transitless triangle $(h,v,k)$ both in $T_s$ and $T$ and as $d_{_{T_s}}(h) = d_{_{T_s}}(k) = d_{_{T}}(h) = d_{_{T}}(k) = 1$, then every other edge incident to $h$ or $k$ induces a tree-cycle that intersects $(h,v,k)$. We conclude that: $$\cap_{_{T_s}}(c_e) = d(h) - 2 + d(k) - 2 = \cap_{_{T}}(d_e)$$ Case 3 is the one that should be analyzed more carefully. As we already know how to calculate intersection numbers of tree-cycles in $T_s$, we will focus on the tree-cycles of $T$. We will further divide this partition in two subpartitions: \begin{itemize} \item Case 3.1: pairs induced by cycle-edges incident to $u$, $\{c_e \leftrightarrow d_e: e \in E - T_s - T \wedge u \in e \}$ \item Case 3.2: pairs induced by cycle-edges incident to $w$, $\{c_e \leftrightarrow d_e: e \in E - T_s - T \wedge w \in e \}$ \end{itemize} In case 3.1 the situation is as follows: the cycle-edge $e=(u,k)$ defines the tree-cycle $c_e = d_e = (u,v,k)$ (both in $T$ and $T_s$). The important details are: \begin{itemize} \item $d_T(u) = 2$: $u$ induces $d(u) - 3$ intersections \item $d_T(k) = 1$: $k$ induces $d(k) - 2$ intersections \item $d_T(w) = 1$: $w$ induces $d(w) - 1$ intersections \item $d(w) \geq 2$ since it is connected at least to $u$ and $v$ in $G$ \item $w$ may have an incident cycle-edge connecting it to $k$, so we should avoid counting twice that intersection \end{itemize} Now we claim that $$\cap_{_{T}}(d_e) \geq d(u) - 3 + d(k) - 2 + d(w) - 1 - \epsilon(w,k)$$ $$\geq d(u) - 2 + d(k) - 2 = \cap_{_{T_s}}(c_e)$$ where $$\epsilon(w,k)= \begin{cases} 1 & (w,k) \in E \\ 0 & otherwise \\ \end{cases} $$ The inequality follows since $d(w) - 1 - \epsilon(w,k) \geq 1$. \bigskip In case 3.2 the situation is as follows: the cycle-edge $e=(w,h)$ defines the tree-cycle $d_e = (w,u,v,h)$ in $T$ and $c_e = (w,v,h)$ in $T_s$. The important details are: \begin{itemize} \item $d_T(u) = 2$: $u$ induces $d(u) - 2$ intersections \item $d_T(h) = 1$: $h$ induces $d(h) - 2$ intersections \item $d_T(w) = 1$: $w$ induces $d(w) - 2$ intersections \item $u$ may have an incident cycle-edge connecting it to $h$, so we should avoid counting twice that intersection \end{itemize} And we claim that $$\cap_{_{T}}(d_e) \geq d(w) - 2 + d(h) - 2 + d(u) - 2 - \epsilon(u,h)$$ $$\geq d(w) - 2 + d(h) - 2 = \cap_{_{T_s}}(c_e)$$ The inequality follows since $d(u) - 2 - \epsilon(u,h) \geq 0$. $\square$ \end{theorem} \subsection{Intrinsic tree invariants} In this subsection we consider the following question: is there any correlation between an \emph{intrinsic tree invariant} and the tree intersection number of the spanning trees for every graph? If so we could formulate an alternative characterization of the MSTCI problem expressed in terms of the invariant. \bigskip By \emph{intrinsic tree invariant} we denote a map $f: \mathscr{T} \rightarrow \mathbb{R}$ on the set of all trees. Of particular interest are the degree-based topological indices \cite{Gutman:2013}. The topological index that motivated our question is the \emph{atom-bond connectivity} (ABC) index \cite{Estrada:1998}. As shown by \cite{Furtula:2009} the star trees are maximal among all trees respect to the ABC index. In the previous section we proved that in the complete graph the star spanning trees are minimal respect to the tree intersection number. Consequently we can formulate a natural question: is there a negative correlation between the ABC index of the spanning trees and their corresponding intersection numbers? \bigskip We will prove that the answer to our question is negative. Without loss of generality we will consider positive correlation (negative correlation is analogous). The underlying idea of the proof is as follows: suppose that there exists an intrinsic tree invariant $f: \mathscr{T} \rightarrow \mathbb{R}$ such that for every graph $G$ the intersection number $\cap(\cdot)$ is positively correlated with $f$, this can be expressed as: $$f(T_1) \leq f(T_2) \iff \cap_G(T_1) \leq \cap_G(T_2), \forall G, T_1, T_2$$ According to this property if we consider two trees $T_1$ and $T_2$ and two graphs $G$ and $H$ such that $T_1, T_2 \in \mathscr{T}_{G}$ and $T_1, T_2 \in \mathscr{T}_{H}$, then this equivalence follows: $$\cap_{G}(T_1) \leq \cap_{G}(T_2) \iff \cap_{H}(T_1) \leq \cap_{H}(T_2)$$ So it suffices to show that there exist $T_1$, $T_2$, $G$ and $H$ such that the equivalence is not satisfied to answer the question negatively. \smallskip First we prove a simple lemma regarding the tree intersection number of a spanning tree $T$ under the removal of a cycle-edge. Namely if a cycle-edge $e$ is removed from $G$ then the tree intersection number of $T$ decreases exactly in the intersection number of its corresponding tree-cycle. \begin{lemma}\label{lemma:edge_removal} let $G=(V,E)$ be a graph, $T \in \mathscr{T}_G$ a spanning tree, $e \in G-T$ a cycle-edge, and $c$ the corresponding tree-cycle, then the following holds: $$\cap_{G-e}(T) = \cap_{G}(T)-\cap_{T}(c)$$ \smallskip Proof. As the spanning tree $T$ is the same in both graphs: $G$ and $G-e$, then the remaining cycle-edges define the same tree-cycles so their pairwise intersection relations are identical. As $c$ is not a cycle in $G-e$ then the equality follows. $\square$ \end{lemma} \begin{theorem}\label{intrinsic} There is no intrinsic tree invariant $f: \mathscr{T} \rightarrow \mathbb{R}$ positively correlated with the intersection number $\cap_G(\cdot)$ for every graph $G$. \smallskip Proof. We will proceed by contradiction: let $f$ be such an intrinsic tree invariant. Then by definition for arbitrary graphs $G$ and $H$ the following equivalences hold $$f(T_1) \leq f(T_2) \iff \cap_{G}(T_1) \leq \cap_{G}(T_2)$$ $$f(T_1) \leq f(T_2) \iff \cap_{H}(T_1) \leq \cap_{H}(T_2)$$ Where $T_1, T_2 \in \mathscr{T}_{G}$ and $T_1, T_2 \in \mathscr{T}_{H}$. This in turn imply that $$\cap_{G}(T_1) \leq \cap_{G}(T_2) \iff \cap_{H}(T_1) \leq \cap_{H}(T_2)$$ \smallskip The proof will be based on showing two graphs and two spanning trees such that the latter equivalence is not valid. \begin{itemize} \item Let $G$ be the complete graph $K_n$ \item Let $H$ be the graph $K_n - \{e_{i,1}, \dots, e_{i,n-3}\}$ where the edges $e_{i,1}, \dots, e_{i,n-3}$ are $n-3$ edges incident to some arbitrary node $v_i$. We will refer to $v_i$ as the \emph{almost disconnected} node of $H$. Note that $d(v_i) = 2$. \item Let $T_1$ be the star spanning tree $T_s$ \item Let $T_2$ be the spanning tree defined as $T_s - \{e_i\} \cup \{e_{i,j}\}$, where $e_i$ is the edge that connects some arbitrary node $v_i$ (in $H$ this role will be played by the almost disconnected node) to the center of the star and $e_{i,j}$ is an edge that connects $v_i$ to a different node $v_j$. \end{itemize} It is easy to check that $T_1$ and $T_2$ are spanning trees of both $G$ and $H$. If we also suppose that $|V| = n > 4$ then by corollary \ref{coromain2} $$\cap_G(T_1) < \cap_G(T_2)$$ By the previous equivalence it is expected that $\cap_H(T_1) < \cap_H(T_2)$ as well. But we will show that this is not the case. \smallskip By a suitable labelling of the nodes of $H$ we can refer to: the center of the star spanning tree as $v_1$, the almost disconnected node of $H$ as $v_2$ and the other neighbor of $v_2$ as $v_3$. By lemma \ref{lemma:edge_removal} arises that $$\cap_H(T_1) = \cap_{H-e_{_{2,3}}}(T_1) + \cap_{T_1}(c_{_{2,3}})$$ $$\cap_H(T_2) = \cap_{H-e_{_{1,2}}}(T_2) + \cap_{T_2}(c_{_{1,2}})$$ Where $c_{_{2,3}}$ and $c_{_{1,2}}$ are the tree-cycles induced by $e_{_{2,3}}$ and $e_{_{1,2}}$ in $T_1$ and $T_2$, resp. Since the remaining tree-cycles corresponding to both trees are the same then $$\cap_{H-e_{_{2,3}}}(T_1) = \cap_{H-e_{_{1,2}}}(T_2)$$ And this imply the following $$\cap_H(T_1) - \cap_H(T_2) = \cap_{T_1}(c_{_{2,3}}) - \cap_{T_2}(c_{_{1,2}})$$ It is an easy exercise to check that $$\cap_{T_1}(c_{_{2,3}}) = \cap_{T_2}(c_{_{1,2}}) = d(v_3) - 2 = n - 3$$ At this point we can conclude that $$\cap_H(T_1) = \cap_H(T_2)$$ Contradicting the fact that $f$ is positively correlated with the tree intersection number for every graph.$\square$ \end{theorem} The underlying key fact of this result is that a spanning tree $T$ that solves the the MSTCI problem for a graph $G$ does not depend on intrinsic properties of $T$ but on the embedding of $T$ in $G$. \bigskip Note that as an interesting side effect this demonstration shows that a star spanning tree is not necessarily a strict local minimum in the spanning tree graph (see previous subsection). \subsection{Intersection number conjecture} In this subsection we present the conjecture $\cap(T_s) \leq \cap(T)$ for every spanning tree $T$ generalizing theorem \ref{teo:local}. Then we explore two ideas to simplify a hypothetical counterexample of the conjecture. The first is based on the notion of \emph{interbranch} cycle-edge. We show that if a non-star spanning tree $T$ exists such that $\cap(T) < \cap(T_s)$, then the inequality must hold if we remove the interbranch cycle-edges. The second is based on the notion of \emph{principal subtree}. In this case we show that the inequality must hold for some principal subtree of $T$. This ideas will be of practical use in the next section. \subsubsection{The conjecture statement} We present below the conjecture that generalizes the case of complete graphs. \bigskip \begin{conjecture}\label{conj} Let $G=(V,E)$ be a graph that admits a star spanning tree $T_s$, then $$\cap(T_s) \leq \cap(T)$$ for every spanning tree $T \in \mathscr{T}_G$. \end{conjecture} \bigskip As an important remark, a demonstration of this result seems difficult if approached by a local-to-global strategy as in the complete graph case exposed previously. \subsubsection{Counterexample simplification} \bigskip In this part we consider some ideas to simplify a hypothetical counterexample of conjecture \ref{conj}. \bigskip Below we define the notion of \emph{interbranch} cycle-edge. \bigskip \begin{definition} Let $G=(V,E)$ be a graph that admits a star spanning tree $T_s$ and let $v \in V$ be the center of $T_s$. Let $T \in \mathscr{T}_G$ be a spanning tree. We call \emph{interbranch cycle-edge of $T$} to any cycle-edge of $T$, $e=(u,w)$, such that $closest-point(v,c) \neq u,w$, where $c$ is the induced tree-cycle of $e$ in $T$. \end{definition} The intuition behind this definition is that the paths $vTu$ and $vTw$ belong to different branches. Or equivalently, $u$ and $w$ are not collinear with respect to $v$ in $T$. The following lemma shows that if we can find a counterexample to the conjecture \ref{conj} (ie.: $\cap(T) < \cap(T_s)$) then we can build a simpler counterexample removing from $G$ the interbranch cycle-edges of $T$. \smallskip \begin{lemma}\label{lemma:reduction1} Let $G=(V,E)$ be a graph that admits a star spanning tree $T_s$ with $v \in V$ as its center. Let $T \in \mathscr{T}_G$ be a spanning tree such that $\cap_{G}(T) < \cap_{G}(T_s)$ and let $\Delta_T$ be the set of interbranch cycle-edges of $T$, then $$\cap_{_{G-\Delta_T}}(T) < \cap_{_{G-\Delta_T}}(T_s)$$ \smallskip Proof. Let $e=(u,w) \in \Delta_T$. Note that $e$ is also a cycle-edge in $T_s$ since $v \neq u, w$ by definition of interbranch cycle-edge. So $e$ determines the tree-cycle $c$ in $T_s$ and the tree-cycle $c'$ in $T$. By the intersection number formula it arises that $\cap_{_{T_s}}(c) = d(u) - 2 + d(w) - 2$. On the other hand, since the other neighbors of $u$ and $w$ are connected to $v$, they belong to distinct tree-cycles with non-trivial intersection with respect to $c'$ in $T$. We conclude that $$\cap_{_{T}}(c') \geq d(u) - 2 + d(w) - 2 = \cap_{T_s}(c).$$ Hence by lemma \ref{lemma:edge_removal}, \begin{align*} \cap_{G-e}(T) = & \cap_{G}(T) - \cap_{T}(c') < &\\ &\cap_{G}(T_s) - \cap_{T_s}(c) = \cap_{G-e}(T_s) \end{align*} Applying the same procedure for every edge in $\Delta_T$, the claimed inequality follows.$\square$ \end{lemma} \begin{definition} Let $T=(V,E)$ be a rooted tree graph with root $v \in V$. Let $w \in N(v)$ then we call \emph{principal subtree respect to w} to the subtree spanned by $v$ and the nodes $u \in V$ such that $w \in vTu$. \end{definition} The next lemma expresses the intersection number of a spanning tree (without interbranch cycle-edges) as the sum of the intersection number of its principal subtrees. \begin{lemma}\label{lemma:partition} Let $G=(V,E)$ be a graph that admits a star spanning tree $T_s$ with $v \in V$ as its center. Let $T$ be a spanning tree of $G$ without interbranch cycle-edges (ie: $\Delta_T = \varnothing$), then the following holds $$\cap_G(T) = \sum_{w \in N(v)} \cap_{G_w}(T_w),$$ where $T_w$ is the principal subtree of $w \in N(v)$ considering $T$ as a rooted tree with $v$ as its root. And $G_w$ is the subgraph spanned by $T_w$. \smallskip Proof. As $\Delta_T=\varnothing$ there are no cycle-edges connecting any two such principal subtrees. This implies that the nonempty intersections between tree-cycles of $T$ must occur inside each subtree. This determines a partition of $C_T$ and the claimed expression follows.$\square$ \end{lemma} The following corollary in line with lemma \ref{lemma:reduction1} further simplifies a hypothetical counterexample of conjecture \ref{conj}. \begin{corollary}\label{corollary:reduction2} Let $G=(V,E)$ be a graph that admits a star spanning tree $T_s$ with $v \in V$ as its center. Let $T$ be a spanning tree of $G$ without interbranch cycle-edges (ie: $\Delta_T = \varnothing$) such that $\cap(T) < \cap(T_s)$ then $$\cap(T_w) < \cap(G_w \wedge T_s)$$ for some $G_w$. Where $T_w$ is the principal subtree of $w \in N(v)$ considering $T$ as a rooted tree with $v$ as its root; $G_w$ is the subgraph of $G$ spanned by $T_w$; $G_w \wedge T_s$ is the subtree of $T_s$ restricted to $G_w$, namely the intersection between $G_w$ and $T_s$. \smallskip Proof. First note that the $G_w$'s are edge disjoint since $\Delta_T = \varnothing$. This partition of the edges of $G$ also determines a partition of $T_s$ such that $\cap(T_s) = \sum_{w \in N(v)} \cap(G_w \wedge T_s)$. As the parts are in a natural bijective relation since they are the subtrees of $T$ and $T_s$ restricted to each $G_w$, we can express the intersection number of $T$ and $T_s$ as follows $$\cap(T) = \sum_{w \in N(v)} \cap(T_w) < \sum_{w \in N(v)} \cap(G_w \wedge T_s) = \cap(T_s)$$ And from the bijection we can deduce that $\cap(T_w) < \cap(G_w \wedge T_s)$ for some $G_w$. $\square$ \end{corollary} \section{Programmatic exploration} \subsection{Overview} In this section we present some experimental results to reinforce conjecture \ref{conj}. We proceed by trying to find a counterexample based on our preceding observations. In the first part we focus on the complete analysis of small graphs, ie: graphs of at most 9 nodes. In the second part we analyze larger families of graphs by random sampling instances. \subsection{General remarks} In the previous section we showed that the space of candidate counterexamples of conjecture \ref{conj} can be reduced. The general picture is as follows: \begin{itemize} \item Let $G=(V,E)$ be a graph that admits a star spanning tree $T_s$ with $v \in V$ as its center \item In the case that we can find some non-star spanning tree $T$ of $G$ such that $\cap(T) < \cap(T_s)$ \item Then we can ``simplify'' the instance by removing the interbranch cycle-edges with respect to $T$ in $G$ without affecting the inequality (see lemma \ref{lemma:reduction1}) \item We can further reduce the instance by focusing on the case where $d_T(v) = 1$, that is: the degree of $v$ restricted to $T$ is 1 (see corollary \ref{corollary:reduction2}) \end{itemize} This considerations can be used to implement algorithms to explore the space of spanning trees more efficiently. Since the algorithms will generate instances in this `reduced' form instead of a brute force approach. \subsection{Complete analysis of small graphs} In this subsection we present an algorithm to explore the spanning tree space. The algorithm proceed by exhaustively analyzing all the reduced graphs of a given number of nodes. The size of the space increases exponentially with respect to the number of nodes, so it has a major limitation: it can be used to analyze only small graphs. The main part is sketched in Algorithm \ref{alg:alg1}. \bigskip The details of the algorithm are the following: \begin{itemize} \item The input parameter $n$ is the number of nodes of the graphs to explore \item $GenerateAllTrees(n-1)$ is a function that returns the list of all trees of $n-1$ nodes \item $GenerateGraph(w,T')$ is a function that builds a graph $G$. Based on the tree $T'$, it adds a new node ($v$) that will play the role of the central node of a star spanning tree, then adds the edge $(v,w)$ to define our candidate tree counterexample $T$. Finally adds all the other edges that link $v$ to the rest of the nodes to obtain $G$. It returns the graph $G$ and ($\bar{\Delta}$) the set of ``possible'' non-interbranch cycle edges. \item $IntersectionNumber(\phi, G)$ is a function that calculates the intersection number of $T$ in $G \cup \phi$, where $\phi \subset \bar{\Delta}$ is a subset of supplementary edges of $G$. \item $StarIntersectionFormula(\phi, G)$ is a function that calculates the intersection number of the star spanning tree in $G \cup \phi$ \item The algorithm finds a counterexample of the conjecture if: $IntersectionNumber(\phi, G) < StarIntersectionFormula(\phi, G)$ \end{itemize} Note that the analyzed graphs are reduced in the sense explained previously. The cycle-edges are non-interbranch by construction and $d_T(v) = 1$ since $v$ is only connected to $w$ in $T$ (ie. there is a single principal subtree). As the algorithm iterates over all possible spanning subtrees $T'$ and all the combinations of possible non-interbranch cycle-edges, every instance is guaranteed to be explored at least once. \begin{algorithm} \caption{CounterexampleSearch($n$)} \begin{algorithmic} \State $\mathscr{T} \gets GenerateAllTrees(n-1)$ \ForEach {tree $T' \in \mathscr{T}$} \ForEach {node $w \in T'$} \State $G, \bar{\Delta} \gets GenerateGraph(w,T')$ \ForEach {subset $\phi \subset \bar{\Delta}$} \State \textbf{check} $(IntersectionNumber(\phi, G) <$ \State $StarIntersectionFormula(\phi, G))$ \EndFor \EndFor \EndFor \end{algorithmic} \label{alg:alg1} \end{algorithm} \begin{table} \caption{Small instances results} \centering \begin{tabular}{cr} \toprule nodes & instances (approx.) \\ \midrule 4 & 5 \\ 5 & 33 \\ 6 & 251 \\ 7 & 4200 \\ 8 & 125000 \\ 9 & 7900000 \\ \bottomrule \label{tab:small-inst} \end{tabular} \end{table} \bigskip To generate all non-isomorphic trees of $|V|-1$ nodes we used the package \emph{nauty} \cite{McKay:2014}. \bigskip The proposed algorithm did not find a counterexample of the intersection conjecture. Table \ref{tab:small-inst} shows the size of the experiments. The column \emph{nodes} is the number of nodes of the graph family, ie: $|V|$. The column \emph{instances} is the number of instances processed. \subsection{Random sampling of large graphs} In this section we present another algorithm to explore the spanning tree space. The strategy in this case is to sample reduced graphs of a given number of nodes. The main part is sketched in Algorithm \ref{alg:alg2}. \bigskip The details of the algorithm are the following: \begin{itemize} \item The input parameters are: $n$ the number of nodes of the graphs and $k$ the size of the sample \item $GenerateRandomTree(n)$ is a function that returns a random tree $T$ of $n$ nodes, where the node $v$ that will play the role of center of the star has degree 1 restricted to $T$. \item $GenerateGraph(T)$ is a function that builds a reduced graph $G$. Based on the tree $T$, adds all the edges that link $v$ to the rest of the nodes to obtain $G$. It returns the graph $G$ and ($\bar{\Delta}$) a random set of non-interbranch cycle edges. \item $IntersectionNumber(\phi, G)$ same as algorithm \ref{alg:alg1} \item $StarIntersectionFormula(\phi, G)$ same as algorithm \ref{alg:alg1} \item The algorithm finds a counterexample of the conjecture if: $IntersectionNumber(\phi, G) < StarIntersectionFormula(\phi, G)$ \end{itemize} \begin{algorithm} \caption{CounterexampleRandomSearch($n, k$)} \begin{algorithmic} \For{\texttt{i := 1..k}} \State $T \gets GenerateRandomTree(n)$ \State $G, \bar{\Delta} \gets GenerateRandomGraph(T)$ \State \textbf{check} $(IntersectionNumber(\phi, G) <$ \State $StarIntersectionFormula(\phi, G))$ \EndFor \end{algorithmic} \label{alg:alg2} \end{algorithm} \begin{table} \caption{Random instances results} \centering \begin{tabular}{cr} \toprule nodes & instances \\ \midrule 25 & 3000000 \\ 50 & 300000 \\ 100 & 30000 \\ 200 & 15000 \\ 400 & 300 \\ \bottomrule \label{tab:random-inst} \end{tabular} \end{table} \bigskip We used a uniformly distributed random number generator. To generate trees we used a simple algorithm that randomly connects a new node to an already connected tree. The non-interbranch cycle-edge set is built by associating a \emph{Bernoulli} trial to each such possible edge. To achieve some diversity for each tree we built three different sets to obtain a sparse, medium and dense sets based on corresponding probabilities $0.1$, $0.5$, $0.9$. \bigskip The proposed algorithm did not find a counterexample of the intersection conjecture. Table \ref{tab:random-inst} shows the size of the experiments. The column \emph{nodes} is the number of nodes of the graph family, ie: $|V|$. The column \emph{instances} is the number of instances processed. \section{Conclusion} In this article we introduced the \emph{Minimum Spanning Tree Cycle Intersection} (MSTCI) problem. \bigskip We proved by enumerative arguments that the star spanning trees are the unique solutions of the problem in the context of complete graphs. \bigskip We conjectured a generalization to the case of graphs (not necessarily complete) that admit a star spanning tree. In this sense we showed that the star spanning tree is a local minimum in the domain of the \emph{spanning tree graph}. We deduced a closed formula for the tree intersection number of star spanning trees in this setting. We proposed two ideas to attempt to find a counterexample of the conjecture. Those ideas were the basis of two strategies to programatically explore the space of solutions in the pursue of a counterexample. The negative result of the experiments suggest that the conjecture is well posed. Unlike the complete graph context, in this slightly more general case, star spanning trees are not unique; there are other spanning trees $T$ such that $\cap(T_s) = \cap(T)$. \bigskip We proved a general result that shows that spanning trees that solve the MSTCI problem don't depend on some intrinsic property but on their particular embedding in the ambient graph. \bigskip An interesting direction of research is to consider the MSTCI problem for other families of graphs, ie.: graphs that do not admit a star spanning tree. Of particular interest is the class of triangular meshes, ie.: graphs that model the immersion of compact surfaces in the 3D euclidean space. \bigskip Another interesing direction of research is related to proving to which complexity class the MSTCI problem belongs to. In case of belonging to the NP-hard class, it will be important to find approximate, probabilistic and heuristic algorithms.

https://arxiv.org/abs/1507.00829

Anti-concentration for polynomials of independent random variables

We prove anti-concentration results for polynomials of independent random variables with arbitrary degree. Our results extend the classical Littlewood-Offord result for linear polynomials, and improve several earlier estimates.We discuss applications in two different areas. In complexity theory, we prove near optimal lower bounds for computing the Parity, addressing a challenge in complexity theory posed by Razborov and Viola, and also address a problem concerning OR functions. In random graph theory, we derive a general anti-concentration result on the number of copies of a fixed graph in a random graph.

\section{Introduction} Let $\xi$ be a Rademacher random variable (taking value $\pm 1$ with probability $1/2$) and $A=\{a_1,\dots,a_n\}$ be a multi-set in $\R$ (here $n \rightarrow \infty$). Consider the random sum $$S := a_1 \xi_1 + \dots + a_n \xi_n $$ where $\xi_i$ are iid copies of $\xi$. In 1943, Littlewood and Offord, in connection with their studies of random polynomials \cite{littlewood1943number}, raised the problem of estimating $\P(S \in I)$ for {\it arbitrary} coefficients $a_i$. They proved the following remarkable theorem: \begin{theorem} \label{theorem:LO} There is a constant $B$ such that the following holds for all $n$. If all coefficients $a_i$ have absolute value at least 1, then for any open interval $I$ of length 1, $$\P (S \in I) \le B n^{-1/2} \log n . $$ \end{theorem} Shortly after the Littlewood-Offord result, Erd\H{o}s \cite{erdos1945lemma} removed the $\log n$ term to obtain the optimal bound using an elegant combinatorial proof. Littlewood-Offord type results are commonly referred to as anti-concentration (or small-ball) inequalities. Anti-concentration results have been developed by many researchers through decades, and have recently found important applications in the theories of random matrices and random polynomials; see, for instance, \cite{nguyen2013small} for a survey. The goal of this paper is to extend Theorem \ref{theorem:LO} to higher degree polynomials. Consider \begin{equation} P (x_1, \dots, x_n) := \sum_{ S \subset \{1, \dots, n\}; |S| \le d } a_S \prod_{j \in S} x_j. \label{form} \end{equation} The first result in this direction, due to Costello, Tao, and the third author, \cite{costello2006random}, is \begin{theorem} \label{theorem:LOpoly0} There is a constant $B$ such that the following holds for all $d,n$. If there are $m n^{d-1} $ coefficients $a_S$ with absolute value at least 1, then for any open interval $I$ of length 1, $$\P (P(\xi_1, \dots, \xi_n) \in I) \le B m^{- \frac{1} {2^ {(d^2+d)/2 } } } . $$ \end{theorem} The exponent $ \frac{1} {2^ {(d^2+d)/2 } }$ tends very fast to zero with $d$, and it is desirable to improve this bound. For the case $d=2$, Costello \cite{costello2013bilinear} obtained the optimal bound $n^{-1/2+o(1) }$. In a more recent paper \cite{razborov2013real}, Razborov and Viola proved \begin{theorem} \label{theorem:LOpolyRV} There is a constant $B$ such that the following holds for all $d,n$. If there are pairwise disjoint subsets $S_1, \dots, S_r$ each of size $d$ such that $a_{S_i}$ have absolute value at least 1 for all $i$, then for any open interval $I$ of length 1, $$\P (P(\xi_1, \dots, \xi_n) \in I) \le B r^{- \frac{1}{ d 2 ^{d+1} }} . $$ \end{theorem} This theorem improves the bound in Theorem \ref{theorem:LOpoly0} to $ m^{- \frac{1}{ d 2 ^{d+1} }} $ via a simple counting argument. Researchers in analysis also considered anti-concentration of polynomials, for entirely different reasons. Carbery and Wright \cite {carbery2001distributional} consider polynomials with $\xi_i$ being iid Gaussian and showed \begin{theorem} \label{thm:CW} There is a constant $B$ such that $$ \P ( | P (\xi_, \dots, \xi_n) | \le \epsilon \Var (P (\xi_, \dots, \xi_n))^{1/2} ) \le B \epsilon^{1/d} . $$ \end{theorem} Their result has been extended by Mossel, O'donnell and Oleszkiewicz \cite{mossel2010noise} to general variables, at a cost of an extra term on the right hand side, which involves the regularity of $P$ (see Section 3). \vskip2mm The goal of this paper is to further improve these anti-concentration bounds, with several applications in complexity theory. Our new results will be nearly optimal in a wide range of parameters. Let $[n] = \{1, 2, \dots, n\}$. Following \cite{razborov2013real}, we first introduce a definition \begin{definition} For a degree $d$ multi-linear polynomial of the form \eqref{form}, the \emph{rank} of $P$, denoted by $\text{rank}(P)$, is the largest integer $r$ such that there exist disjoint sets $S_1,\ldots,S_r \subseteq [n]$ of size $d$ with $|a_{S_j}| \ge 1$, for $j \in [r]$. \end{definition} Our first main result concerns the Rademacher case. Let $\xi_i, i=1, \dots, n$ be iid Rademacher random variables. \begin{theorem} \label{theorem:LOpoly1} There is an absolute constant $B$ such that the following holds for all $d,n$. Let $P$ be a polynomial of the form \eqref{form} whose rank $r\ge 2$. Then for any interval $I$ of length 1, $$\P (P(\xi_1, \dots, \xi_n) \in I) \le \min \left (\frac{B d^{4/3}\sqrt {\log r}}{ r^{ \frac{1} {4d+1} } }, \frac{\exp(B d^{2}(\log \log r)^{2})}{\sqrt r}\right ) . $$ \end{theorem} For the case when $d$ is fixed, it has been conjectured \cite{nguyen2013small} that $\P (P(\xi_1, \dots, \xi_n) \in I) = O(r^{-1/2} )$. This conjectural bound is a natural generalization of Erdos-Littlewood-Offord result and is optimal, as shown by taking $P= (\xi_1 + \dots + \xi_n)^d$, with $n$ even. For this $P$, the rank $r= \Theta (n)$ and $\P ( |P| \le 1/2 ) = \P ( P= 0) = \Theta (n^{-1/2} )$. Our result confirms this conjecture up to the sub polynomial term $\exp( B d^2 (\log \log r)^2) $. In applications it is important that we can allow the degree $d$ tends to infinity with $n$. Our bounds in Theorem \ref{theorem:LOpoly1} are non-trivial for degrees up to $c \log r/\log \log r$, for some positive constant $c$. Up to the $\log \log$ term, this is as good as it gets, as one cannot hope to get any non-trivial bound for polynomials of degree $\log_2 r$. For example, the degree $d$ polynomial on $2^d \cdot d$ variables defined by $P(\xi) = \sum_{i=1}^{2^d} \prod_{j=1}^d (\xi_{ij}+1)$, where $\xi_{ij}$ are iid Rademacher random variables, has $r = 2^d$ and $\P(P(\xi) = 0) = \Omega(1)$. Next, we generalize our result for non-Rademacher distributions. As a first step, we consider the $p$-biased distribution on the hypercube. For $p \in (0,1)$, let $\mu_p$ denote the Bernoulli variable with $p$-biased distribution: $\P_{x\sim \mu_p}(x=0) = 1-p$, $\P_{x\sim \mu_p}(x=1) = p$ and let $\mu_p^{n}$ be the product distribution on $\{0, 1\}^{n}$. \begin{theorem}\label{th:littlewoodoffordbiasedmain} There is an absolute constant $B$ such that the following holds. Let $P$ be a polynomial of the form \eqref{form} whose rank $r\ge 2$. Let $p$ be such that $\tilde r:= 2^{d}\alpha^{d}r\ge 3$ where $\alpha := \min\{p, 1-p\}$. Then for any interval $I$ of length 1, $$\P_{x\sim \mu_p^n}(P(x) \in I) \leq \min\left (\frac{B d^{4/3}(\log \tilde r )^{1/2}}{(\tilde r )^{1/(4d+1)}}, \frac{\exp(B d^2(\log \log (\tilde r )^2)}{\sqrt{ \tilde r }}\right ).$$ \end{theorem} The distribution $\mu_p^n$ plays an essential role in probabilistic combinatorics. For example, it is the ground distribution for the random graphs $G(N,p)$ (with $n:= { N \choose 2}$). We discuss an application in the theory of random graphs in the next section. Finally, we present a result that applies to virtually all sets of independent random variables, with a weak requirement that these variables do not concentrate on a short interval. \begin{theorem}\label{thm:generalist} There is an absolute constant $B$ such that the following holds. Let $\xi_1, \dots, \xi_n$ be independent (but not necessarily iid) random variables. Let $P$ be a polynomial of the form \eqref{form} whose rank $r\ge 2$. Assume that there are positive numbers $p$ and $\epsilon$ such that for each $1 \le i \le n$, there is a number $y_i$ such that $\min \{ \P(\xi_i \le y_i) , \P(\xi_i > y_i ) \}=p$ and $\P (| \xi_i -y_i| \ge 1) \ge \epsilon $. Assume furthermore that $\tilde r:= (p\epsilon)^d r \ge 3$. Then for any interval $I$ of length 1 $$\P(P(\xi_1, \dots, \xi_n) \in I) \leq \min\left (\frac{B d^{4/3}(\log \tilde r )^{1/2}}{(\tilde r )^{1/(4d+1)}}, \frac{\exp(B d^2(\log \log (\tilde r )^2)}{\sqrt{ \tilde r }}\right ).$$ \end{theorem} Notice that even in the gaussian case, Theorem \ref{thm:generalist} is incomparable to Theorem \ref{thm:CW}. If we use Theorem \ref{thm:CW} to bound $\P ( P \in I)$ for an interval $I$ of length 1, then we need to set $\epsilon = \Var (P)^{-1/2 }$, and the resulting bound becomes $\frac{B}{ (\Var P) ^{1/2d} } $. For sparse polynomials, it is typical that $r $ is much larger than $(\Var P) ^{1/d} $ and in this case our bound is superior. To illustrate this point, let us fix a constant $d > c >0$ and consider $$P := \sum_{S \subset \{1, \dots, n \}, |S|=d} a_S \prod_{ i \in S} x_i $$ where $a_S$ are iid random Bernoulli variables with $\P( a_S=1) = n^{-c}$. It is easy to show that the following holds with probability $1-o(1)$ \begin{itemize} \item For any set $X \subset \{1, \dots, n \}$ of size at least $n/2$, there is a subset $S \subset X, |S|=d$, such that $a_S=1$. \item The number nonzero coefficients is at most $n^{d-c} $. \end{itemize} In other words, these two conditions are typical for a sparse polynomial with roughly $n^{d-c}$ nonzero coefficients. On the other hand, if the above two conditions holds, then we have $\Var (P) \le n^{d-c} $ and $r \ge n/2d$ (by a trivial greedy algorithm). Our bound implies that $$\P ( P \in I ) \le C(d) n^{-1/2 +o(1) } $$ while Cabery-Wright bound only gives $$\P ( P \in I) \le C(d) n^{- 1/2 +c/2d }. $$ The rest of the paper is organized as follows. In Section \ref{app} below, we discuss applications in complexity theory and graph theory, with one long proof delayed to Section \ref{app-proof}. Sections \ref{reg-pol} and \ref{reg-lmm} are devoted to some combinatorial lemmas. In Section \ref{main-proof}, we treat polynomials with Rademacher variables. The generalizations are discussed in Section \ref{gen-proof}. All asymptotic notations are used under the assumption that $n$ tends to infinity. All the constants are absolute, unless otherwise noted. \section{ Applications} \label{app} \subsection{Applications in complexity theory} We use our anti-concentration results to prove lower bounds for approximating Boolean functions by polynomials in the {\sl Hamming metric}. The notion of approximation we consider is as follows. \begin{definition} Let $\epsilon > 0$ and $\mu$ be a distribution on $\{0, 1\}^n$. For a Boolean function $f:\{0, 1\}^n \to \{0, 1\}$ and a polynomial $P:\R^n \to \R$, we say $P$ $\epsilon$-approximates $f$ with respect to $\mu$ \footnote{We drop $\mu$ in the description when it is clear from context or if it is the uniform distribution.} if $$\P_{x\sim\mu}(P(x) = f(x)) > 1-\epsilon.$$ We define $d_{\mu,\epsilon}(f)$ to be the least $d$ such that there is a degree $d$ polynomial which $\epsilon$-approximates $f$ with respect to $\mu$. \end{definition} An alternate ({dual}) way to view the above notion is in terms of distributions over low-degree polynomials---``randomized polynomials''---which approximate the function in the worst-case. In particular, by Yao's min-max principle, $d_{\mu,\epsilon}(f) \leq d$ for every distribution $\mu$ if and only if there exists a distribution $\mathcal D$ over degree at most $d$ polynomials which approximates $f$ in the worst-case: for all $x$, $\P_{P \sim \mathcal D}[P(x) = f(x)] > 1- \epsilon$. Approximating Boolean functions by polynomials in the Hamming metric was first considered in the works of Razborov \cite{razborov1987lower} and Smolensky \cite{smolensky1987algebraic} over fields of finite characteristic as a technique for proving lower bounds for small-depth circuits. This was also studied in a similar context over real numbers by the works of \cite{beigel1991perceptron}, \cite{aspnes1994expressive}; the latter work uses them to prove lower bounds for $AC(0)$. More recently, in a remarkable result, Williams \cite{williams2014faster} (also see \cite{williams2014polynomial, abboud2015more}) used polynomial approximations in Hamming metric for obtaining the best known algorithms for all-pairs shortest path and other related algorithmic questions. Here, we study lower bounds for the existence of such approximations. {\bf Approximating Parity.} Let $par_n:\{0, 1\}^n \to \{0, 1\}$ denote the parity function: $par_n(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n$ (where arithmetic is mod $2$). In \cite{razborov2013real}, Razborov and Viola introduced another way to look at this problem. For two functions $f, g : \{0,1\}^n \rightarrow \R$, define their "correlation" to be the quantity $$ Cor_n (f, g) = \P _x (f(x) =g(x) ) -1/2 , $$ where $x$ is uniformly distributed over $\{0,1\}^n$. They highlighted the following challenge {\bf Challenge.} Exhibit an explicit boolean function $f: \{0,1 \}^n \rightarrow \{0,1 \} $ such that for any real polynomial $P$ of degree $\log_2 n$, one has $$\Cor_n (f,P) \le o( 1/\sqrt n ). $$ This challenge is motivated by studies in complexity theory and has connections to many other problems, such as the famous rigidity problem; see \cite{razborov2013real} for more discussion. The Parity function seems to be a natural candidate in problems like this. Razborov and Viola, using Theorem \ref{theorem:LOpolyRV}, proved \begin{theorem} \label{theorem:RV1} \cite{razborov2013real} For all sufficiently large $n$, $\Cor_n (par_n, P) \le 0$ for any real polynomial $P$ of degree at most $\frac{1}{2} \log_2 \log_2 n $. \end{theorem} With Theorem \ref{theorem:LOpoly1}, we obtain the following improvement, which gets us within the Challenge by a $\log \log n$ factor. \begin{theorem} \label{theorem:RV2} For all sufficiently large $n$, $\Cor_n (par_n, P) \le 0$ for any real polynomial $P$ of degree at most $\frac{ \log n}{15 \log \log n }$. \end{theorem} \begin{proof} Let $d$ be the degree of $P$. Following the arguments in the proof of \cite[Theorem 1.1]{razborov2013real}, we can assume that $P$ contains at least $\sqrt n$ pairwise disjoint subsets $S_i$ each of size $d$ and non-zero coefficients. It suffices to show that the probability that $P$ outputs a boolean value is at most $1/2$. By replacing $P$ by $q(x_1, \dots, x_n) := P((x_1+1)/2, \dots, (x_n+1)/2)$, one can convert the problem into polynomial of the same degree defined on $\{\pm 1\}^{n}$, in other words, on Rademacher variables. Then by Theorem \ref{theorem:LOpoly1}, this probability is bounded by $2B\frac{d^{4/3}\log ^{1/2}n}{n^{1/(8d+2)}}$. This is less than $1/2$ for every $d\le \frac{\log n}{15\log \log n}$ when $n$ is sufficiently large. \end{proof} {\bf Approximating AND/OR.} One of the main building blocks in obtaining polynomial approximations in the Hamming metric is the following result for approximating the OR function\footnote{$OR(x_1,\ldots,x_n)$ is $1$ if any of the bits $x_i$ is non-zero.}. \begin{claim} For all $\epsilon \in (0,1)$ and distributions $\mu$ over $\{0, 1\}^n$, there exists a polynomial $P:\R^n \to \R$ of degree at most $O((\log n)(\log 1/\epsilon))$ such that $\P_{x\sim \mu}(P(x) = OR(x)) > 1 - \epsilon$. \end{claim} By iteratively applying the above claim, Aspnes, Beigel, Furst, and Rudich \cite{aspnes1994expressive} showed that $AC(0)$ circuits of depth $d$ have $\epsilon$-approximating polynomials of degree at most $O(((\log s)(\log(1/\epsilon)))^d \cdot (\log(s/\epsilon))^{d-1})$. We prove that the following lower bound for such approximations: \begin{theorem}\label{th:orapprox} There is a constant $c > 0$ and a distribution $\mu$ on $\{0, 1\}^n$ such that for any polynomial $P:\{0, 1\}^n \to \R$ of degree $d < c (\log \log n)/(\log \log \log n)$, $$\P_{x\sim\mu}(P(x) = OR(x)) < 2/3.$$ \end{theorem} To the best of our knowledge no $\omega(1)$ lower bound was known for approximating the OR function. We give an explicit distribution (directly motivated by the upper bound construction in \cite{aspnes1994expressive}) under which OR has no $1/3$-error polynomial approximation. The distribution $\mu$ on $\{0, 1\}^n$ we consider is as follows: \begin{enumerate} \item With probability $1/2$ output $x =0$. \item With probability $1/2$ pick an index $i \in [D]$ uniformly at random and output $x \gets \mu_{2^{-a^i}}^n$ for some suitably chosen parameters $a, D$. \end{enumerate} The analysis then proceeds at a high level as in the lower bound for parity. However, we need some extra care with the inductive argument as unlike for parity, we can't consider arbitrary fixings of subsets of coordinates of the OR function. We get around this hurdle by instead only considering fixing parts of the input to $0$ and decreasing the bias $p$ to make sure that these coordinates are indeed set to $0$ with high probability. The details are defered to Section 7. \subsection{The number of small subgraphs in a random graph} Consider the Erd\H{o}s-R\'enyi random graph $G(N,p)$. Let $H$ be a small fixed graph (a triangle or $C_4$, say). The problem of counting the number of copies of $H$ in $G(N,p)$ is a fundamental topics in the theory of random graphs (see, for instance, the text books \cite{Bollobasbook, Jansonbook}). In fact, one can talk about a more general problem of counting the number of copies of $H$ in a random subgraph of any deterministic graph $G$ on $N$ vertices, formed by choosing each edges of $G$ with probability $p$. We denote the $F(H,G,p)$ this random variable. In this setting we understand that $H$ has constant size, and the size of $G$ tends to infinity. It has been noticed that $F$ can be written as a polynomial in term of the edge-indicator random variables. For example, the number of $C_4$ (circle of length $4$) is $$ \sum_{ i,j,k,l } \xi_{ij} \xi_{jk} \xi_{ kl} \xi_{li} $$ where the summation is over all quadruple $ijkl$ which forms a $C_4$ in $G$ and the Bernoulli random variable $\xi_{ij}$ represents the edge $ij$. Clearly, any polynomial of this type has $n = e(G)$ iid Bernoulli $p$-bias variables $\xi_{ij}$, and its degree equals the number of edges of $H$. The rank $r$ of $F$ is exactly the size of the largest collection of edge disjoint copies of $H$ in $G$. The polynomial representation has been useful in proving {\it concentration} (i.e.{\it large deviation} ) results for $F$ (see \cite{KimVu, Vusurvey}, for instance). Interestingly, it has turned out that one can also use this to derive anti-concentration result, in particular bounds on the probability that the random graph has exactly $m$ copies of $H$. By Theorem \ref{th:littlewoodoffordbiasedmain}, we have \begin{corollary} \label{Hcopy} Assume that $p$ is a constant in $(0,1)$. Then for fixed $H$ and any integer $m$ which may depend on $G$ $$\P ( F(H,G ,p) = m ) \le r^{-1/2 +o(1)} , $$ where $r$ is the size of the largest collection of edge-disjoint copies of $H$ in $G$. In particular, if $G=K_n$, then $$\P ( F(H,K_n ,p) = m ) \le n^{-1/2 +o(1)}. $$ \end{corollary} A similar argument can be used to deal with the number of {\it induced} copies of $H$, which can be also written as a polynomial with degree at most ${v \choose 2}$, with $v$ being the number of vertices of $H$. Details are left out as an exercise. Finally, let us mention that in a recent paper \cite{GK}, Gilmer and Kopparty obtained a precise estimate for $\P( F(H, K_n, p) =m)$ in the case when $H$ is a triangle. \footnote{We would like to thank J. Kahn for pointing out this reference.} Their approach relies on a careful treatment of the characteristic function. It remains to be seen if this method applies to our more general setting. \section {Regular polynomials}\label{reg-pol} Our proofs of anti-concentration bounds use the techniques developed in the context of bounding the \textit{noise sensitivity} of \textit{polynomial threshold functions} in the works \cite{diakonikolas2014regularity, harsha2014bounding, kane2014pseudorandom}. In particular, we use the concept of \textit{regular polynomials}, the invariance principle of Mossel, O'donnell, and Oleszkiewicz \cite{mossel2010noise}, and the \textit{regularity lemma} of \cite{diakonikolas2014regularity, harsha2014bounding}. In this and the following section, we discuss these tools. To start, we define regular polynomials and discuss an anti-concentration result for them. The \textit{influence} of the $i$-th variable on $P$ is defined to be $\text{Inf}_i = \text{Inf}_i(P)=\sum_{i\in S} a_S^{2}$. Since $ \Var(P) = \sum_{S\neq \emptyset} a_S^{2}$, we have \begin{equation} \Var(P)\le \sum_{i=1}^{n} \text{Inf}_i \le d\Var(P).\label{total_inf} \end{equation} Assume the random variables are ordered such that $ \text{Inf}_1\ge \text{Inf}_2\ge\dots \ge\text{Inf}_n$. Let $ \tau>0 $, the $ \tau$-\textit{critical index} of $ P $ is the least $ i $ such that $ \text{Inf}_{i+1}\le \tau\sum_{j=i+1}^{n}\text{Inf}_j$. If it does not hold for any $i$, we say that the $P$ has $\tau$-critical index $\infty$. If $ P $ has $ \tau $-critical index 0, we say that $ P $ is $ \tau $-regular. The following is a corollary of strong results from \cite{carbery2001distributional} and \cite{mossel2010noise}. \begin{proposition} \label{regular} Let $P$ be a non-constant polynomial of the form \ref{form}. Let $\tau> 0$. If $ P $ is $ \tau $-regular, then $ \P(|P(\xi_1, \dots, \xi_n)|\le \alpha)\le \frac{Cd\alpha^{1/d}}{(\Var(P))^{1/2d}}+ Cd\tau^{1/(4d+1)}$ for every $ \alpha>0 $. \end{proposition} \begin{proof} Let $\tilde \xi_1, \dots, \tilde \xi_n$ be independent standard Gaussian variables. Notice that $$\Var(P(\xi_1\, \dots, \xi_n)) = \Var(P(\tilde \xi_1, \dots, \tilde \xi_n)).$$ Our settings satisfy the Hypothesis {\bf H$4$} of \cite[Theorem 3.19]{mossel2010noise} with $r = 4$. Using that theorem, one obtains \begin{eqnarray}\label{moo} \P(|P(\xi_1, \dots, \xi_n)|\le \alpha)&\le& \P(|P(\tilde \xi_1, \dots,\tilde \xi_n)|\le \alpha) + Cd\tau^{1/(4d+1)}. \end{eqnarray} Now, for Gaussian case, it was proved in \cite[Theorem 8]{carbery2001distributional} that for every $ \alpha>0 $, \begin{equation}\label{gau} \P(|P(\tilde{\xi}_1, \dots, \tilde{\xi_n})|\le \alpha)\le C\frac{d\alpha^{1/d}}{(\Var (P))^{1/2d}}. \end{equation} Combining \eqref{moo} and \eqref{gau}, we get the desired bound. \end{proof} \section{ A regularization lemma}\label{reg-lmm} Proposition \ref{regular} would yield our desired bound in Theorem \ref{theorem:LOpoly1} if $\tau$ is small (say at most $r^{-1}$). However, there is no guarantee for this assumption. In order to go from the regular case to the general case, we will use the following regularization lemma, whose proof is a slight modification of \cite[Theorem 1.1]{diakonikolas2014regularity} (the version below gives us better quantitative bounds in our applications). The main idea is to condition on the random variables with large influence. With high probability, the resulting polynomial is either regular or dominated by its constant part. For a set $S\subset [n]$, we consider a random assignment $\rho\in \{\pm 1\}^{|S|}$ which assigns values $\pm 1$ to variables $(\xi_i)_{i\in S}$. We say that ``$\rho$ fixes $S$". For each such $\rho$, the polynomial $P$ becomes a polynomial of $(\xi_{i})_{i\notin S}$ which is denoted by $P_{\rho}$. We write $P_{\rho} = P^{*}(\rho) + q_{\rho}(\xi_i)_{i\notin S}$ where $P^*$ is the constant part of $P_{\rho}$ consisting of monomials of $(\xi_i)_{i\in S}$ only. For $C>0$ and $0<\beta<1$, we say that $P_{\rho}$ is \textit{$(C,\beta)$-tight} if \begin{equation} \sqrt{\Var_{(\xi_{i})_{i\notin S}} (q_{\rho})} \le |P^*(\rho)|\left(C(\log\frac{1}{\beta})\right)^{-d/2},\label{q1} \end{equation} and \begin{equation} \P_{(\xi_{i})_{i\notin S}}\left (|q_\rho|\le \frac{1}{2}|P^*(\rho)|\right )\ge 1-\beta\label{qp*1}. \end{equation} Note that it is always true that $\E_{(\xi_{i})_{i\notin S}} q_{\rho} = 0$. We shall see later that \eqref{q1} actually implies \eqref{qp*1}. \begin{proposition}\label{regularity lemma} There exist absolute constants $C$ and $C'$ such that the following holds true. Let $P(\xi_1, \dots, \xi_n)$ be a a degree-$d$ polynomial, let $0<\tau, \beta<\frac{1}{3}$. Let $\alpha = C(d\log\log 1/\beta+d\log d)$ and $\tau' = (C'd\log d\log\frac{1}{\tau})^{d}\tau$. Let $M\in \mathbb N$ such that $M\frac{\alpha}{\tau}\le n$. Then, there exists a decision tree of depth at most $M\frac{\alpha}{\tau}$ with $P$ at the root, variables $\xi_i$'s at each internal node, and a degree-$d$ polynomial $P_{\rho}$ at each leaf $\rho$, with the following property: with probability at least $1 - (1-\frac{1}{2C^{d}})^{M}$, a random path from the root $P$ reaches a leaf $\rho$ such that $P_{\rho}$ is either $\tau'$-regular or \textit{$(C,\beta)$-tight}. \end{proposition} \begin{proof} First, we consider the case when the $\tau$-critical index of $P$ is large. For a positive integer $K$, denote by $[K]$ the set $\{1, \dots, K\}$. \begin{lemma}\label{large critical index} There exists a constant $C$ such that the following holds true. Let $0<\tau, \beta<\frac{1}{3}$ be deterministic constants that may depend on $n$. Suppose that $P$ has $\tau$-critical index at least $K = \frac{\alpha}{\tau}$, where $\alpha = C(d\log\log 1/\beta+d\log d)$. Then for at least $\frac{1}{2C^{d}}$ fraction of restrictions $\rho$ fixing $[K]$, the polynomial $P_\rho$ is $(C, \beta)$-tight. \end{lemma} Roughly speaking, the $(C,\beta)$-tightness asserts that the resulting polynomial $P_{\rho}$ has large constant term, compared to the random part, and therefore, it concentrates around the constant part. \begin{proof} Since the proof is completely the same as the proof of \cite[Lemma 3.5]{diakonikolas2014regularity}, we only provide a sketch here. Without loss of generality, assume that $\Var(P) = 1$. We first show that \begin{equation} \P_{\rho}(|P^*(\rho)|\ge \frac{1}{2C^{d}})\ge \frac{1}{C^{d}}\label{lmm4DSTW} \end{equation} where by $\P_\rho$ we mean the probability with respect to $\xi_1, \dots, \xi_K$. Observe that $\Var_{\rho}(P^*(\rho))=\sum_{\emptyset\neq S\subset [K]} a_S^{2}\le \Var(P)=1$. Moreover, by definition of critical index, \begin{equation} \sum_{i\notin [K]} \text{Inf}_i(P)\le(1-\tau)^{K}\sum_{i=1}^{n}\text{Inf}_i(P)\le de^{-\alpha}\le \frac{1}{2}.\label{suminf} \end{equation} Hence, $1\ge \Var_{\rho} (P^*(\rho)) = \Var(P) - \sum_{S\subset [n], S\nsubseteq [K]} a_S^{2}\ge 1 - \sum_{i\notin [K]}\text{Inf}_i(P)\ge\frac{1}{2}$. Then, we use the following Theorem \begin{theorem}(\cite{austrin2011randomly}, \cite{dinur2006fourier}, also \cite[Theorem 2.5]{diakonikolas2014regularity}) There is a universal constant $C_0>1$ such that for any non-zero degree-$d$ polynomial $P: \{-1, 1\}^{n}\to \R$ with $\E (P) = 0$, we have $$\P\left (P>\frac{\sqrt{\Var(P)}}{C_0^{d}}\right )>\frac{1}{C_0^{d}}.$$ \end{theorem} Let $C\ge C_0^{2}$. Applying the above Theorem to $P^*(\rho) - \E_{\rho}P^*(\rho)$ if $\E_{\rho}P^*(\rho)\ge 0$ and $-P^*(\rho) +\E_{\rho}P^*(\rho)$ otherwise gives \eqref{lmm4DSTW}. Next, we show that \begin{equation} \P_{\rho}\left (\Var (q_{\rho})>\frac{1}{(2C^{d})^{2}}\left(C(\log\frac{1}{\beta})\right)^{-d}\right )\le \frac{1}{2C^{d}}.\label{lmm5DSTW} \end{equation} Indeed, let $Q(\rho) = \Var (q_{\rho})$. By triangle inequality and Bonami-Beckner inequality (see, for instance, \cite[Theorem 2.1]{diakonikolas2014regularity}, or \cite{bonami1970etude}, \cite{gross1975logarithmic}), one can show that $||Q(\rho)||_{2} = \sqrt{\E_{\rho} Q^{2}(\rho)} \le 3^{d}\sum_{i>K}\E_{\rho}\text{Inf}_{i}(P_\rho)= 3^{d}\sum_{i> K}\text{Inf}_i(P)\le 3^{d}de^{-\alpha}$ where the last inequality is just \eqref{suminf}. From this, we use the following Theorem \begin{theorem}(\cite{austrin2011randomly}, \cite{dinur2006fourier}, also \cite[Theorem 2.2]{diakonikolas2014regularity})\label{thm6} Let $P:\{-1, 1\}^{n}\to \R$ be a degree-$d$ polynomial. For any $t>e^{d}$, we have \begin{equation} \P(|P|> t||P||_2)\le \exp(-\Omega(t^{2/d})).\nonumber \end{equation} \end{theorem} Using this Theorem for the polynomial $Q$ and $t = d^{d}C^{d}\log^{d}C$, we get \eqref{lmm5DSTW}. From \eqref{lmm4DSTW} and \eqref{lmm5DSTW}, with probability at least $\frac{1}{2C^{d}}$ over all possible $\rho$, \eqref{q1} happens. For each such $\rho$, using Theorem \ref{thm6} for $q$, we obtain \begin{eqnarray} \P_{\xi_{K+1}, \dots, \xi_n}(|q_{\rho}|\ge \frac{1}{2}|P^*(\rho)|)\le \P_{\xi_{K+1}, \dots, \xi_n}\left (|q_{\rho}|\ge \frac{1}{2}\left (C\log\frac{1}{\beta}\right )^{d/2}||q_{\rho}||_2\right )\le \beta,\nonumber \end{eqnarray} which gives \eqref{qp*1} and completes the proof of Lemma \ref{large critical index}. \end{proof} Next, we consider the case when $P$ has small critical index. We'll use the following Lemma \cite[Lemma 3.9]{diakonikolas2014regularity} which asserts that by assigning values to the random variables with large influences, with significant probability, one gets a regular polynomial. \begin{lemma}\label{small critical index} Let $C$ be the constant in Lemma \ref{large critical index}. There exists an absolute constant $C'$ such that the following holds. Let $0<\tau<\frac{1}{3}$. Assume that $P$ has $\tau$-critical index $k\in [n]$. Let $\rho$ be a random restriction fixing $[k]$, and $\tau' = (C'd\log d\log\frac{1}{\tau})^{d}\tau$. With probability at least $\frac{1}{2C^{d}}$ over the choice of $\rho$, the restricted polynomial $P_\rho$ is $\tau'$-regular. \end{lemma} Combining Lemmas \ref{large critical index} and \ref{small critical index}, we get \begin{lemma}\label{tree} Let $P(\xi_1, \dots, \xi_n)$ be a a degree-$d$ polynomial, $0<\tau, \beta<\frac{1}{3}$. Let $\alpha = C(d\log\log 1/\beta+d\log d)$ and $\tau' = (C'd\log d\log\frac{1}{\tau})^{d}\tau$. Assume that $\text{Inf}_1\ge \text{Inf}_2\dots\ge\text{Inf}_n$. Then one of the following holds true. \begin{enumerate} \item $P$ is $\tau$-regular. \item The $\tau$-critical index of $P$ is at least $\frac{\alpha}{\tau}$ and the conclusion of Lemma \ref{large critical index} holds. \item The $\tau$-critical index of $P$ is $k<\frac{\alpha}{\tau}$ and the conclusion of Lemma \ref{small critical index} holds. \end{enumerate} \end{lemma} Now, we are ready for the proof of Proposition \ref{regularity lemma}. The strategy is to apply Lemma \ref{tree} repeatedly $M$ times. At first, if $P$ is not $\tau$-regular, we apply Lemma \ref{tree} to $P$ and obtain an initial tree of depth at most $\frac{\alpha}{\tau}$. We know that at least $\frac{1}{2C^{d}}$ fractions of the restricted $P_{\rho}$ are "good", i.e., either $\tau'$-regular or $(C, \beta)$-tight. We keep them as leaves of our final tree and leave them untouched during the next stages. At the second stage, for each of the remaining "bad" polynomials $P_{\rho}$, we order the unrestricted variables in decreasing order of their influences in $P_{\rho}$, and then apply lemma \ref{tree} to it. Note that probability of reaching a bad leaf in this second tree is at most $(1-\frac{1}{2C^{d}})^{2}$. Continuing in this manner $M$ times, we get the desired tree and complete the proof of Theorem \ref{regularity lemma}. \end{proof} \section {Proof of Theorem \ref{theorem:LOpoly1}}\label{main-proof} The high-level argument for the first bound of \ref{theorem:LOpoly1} is as follows. If the polynomial is sufficiently \textit{regular}, we apply the \textit{anti-concentration} property of regular polynomials; the latter property in turn follows from the invariance principle and a similar anti-concentration property for polynomials with respect to the Gaussian distribution. To complete the argument, we use the regularity lemma which shows that any polynomial can be written as a small-depth decision tree where most leaves are labeled by polynomials which are either (1) Regular or (2) Polynomials which are fixed in sign with high probability over a uniformly random input. In the first case, you get a regular polynomial of high rank (as the tree is shallow) and we apply the previous argument. In the second case, we argue directly that the probability of taking the value $0$ is small. To prove the second bound of \ref{theorem:LOpoly1}, we follow the same conceptual approach but adopt a more careful analysis following the work of Kane \cite{kane2014correct}. We defer the details to the actual proof. \subsection{First bound} Without loss of generality, we can assume that $I$ is centered at 0 and $r$ is larger than some constant. We can also assume that $d\le \frac{2\log r}{\log \log r}$ because otherwise $dr^{-1/(4d+1)}\ge 1$ and the desired bound becomes trivial. Let $\tau \in (0, \frac{1}{3})$ and let $\beta = \frac{1}{r}$. We will use Proposition \ref{regularity lemma} to reduce to the regular case. Let $\alpha$, $\tau'$ be as in that Proposition, i.e., $\alpha = C(d\log\log \frac{1}{\beta}+d\log d)$ and $\tau' = (C'd\log d\log\frac{1}{\tau})^{d}\tau$. Let $M = \lfloor\frac{r\tau}{2\alpha}\rfloor$. Call a leaf of the decision tree \emph{good} if $P_\rho$ is either $\tau$-regular or $(C,\beta)$-tight and \emph{bad} otherwise. Now, following our decision tree, we have \begin{eqnarray} \P(P\in I) &\le& \P(\text{reaching a bad leaf} ) + \sum_{\rho\text{ is a good leaf} }\P(\text{reaching } \rho \text{ and } P_{\rho} \in I)\nonumber\\ &\le& (1-\frac{1}{2C^{d}})^{M} + \sum_{\rho\text{ is a good leaf} }\P(\text{reaching } \rho \text{ and } P_{\rho} \in I)\nonumber\\ &\le& 2\exp\left (-\frac{r\tau}{4\alpha C^d}\right )+ \sum_{\rho\text{ is a good leaf} }\P(\text{reaching } \rho \text{ and } P_{\rho} \in I)\label{mon1}. \end{eqnarray} Now, for each good leaf $\rho$, $P_\rho$ is either $(C,\beta)$-tight or $\tau'$-regular. Let $S$ be the set of indices $i$ of the internal nodes $\xi_i$ that lead to $\rho$. In other words, $\rho$ fixes $S$. Since the depth of the decision tree is at most $M\frac{\alpha}{\tau}\le \frac{r}{2}$, one has $|S|\le \frac{r}{2}$ and so $q_\rho$ contains at least $r/2$ monomials of degree $d$ each, with mutually disjoint sets of random variables, and with coefficients at least 1 in magnitude. Therefore, $\Var_{(\xi_i) _{ i\notin S}} (P_\rho)=\Var_{(\xi_i)_{ i\notin S}} (q_\rho)\ge r/2$. Assume $P_\rho$ is $(C,\beta)$-tight, then by \eqref{q1}, one has $|P^*(\rho)|=\Omega(\sqrt r)\ge 2$. This together with \eqref{qp*1} give \begin{eqnarray} \P(\text{reaching } \rho \text{ and } P_{\rho} \in I) &=& \P_{\xi_i, i\in S}(\text{reaching } \rho)\P_{\xi_i, i\notin S}(P_\rho \in I)\nonumber\\ &\le& \P_{\xi_i, i\in S}(\text{reaching } \rho)\P_{\xi_i, i\notin S}(|q_\rho| \ge |P^*(\rho)|-1>\frac{1}{2}|P^*(\rho)|)\nonumber\\ &\le& \beta \P_{\xi_i, i\in S}(\text{reaching } \rho) = \frac{1}{r} \P_{\xi_i, i\in S}(\text{reaching } \rho).\label{mon2} \end{eqnarray} Next, assume that $P_\rho$ is $\tau'$-regular. By Proposition \ref{regular}, \begin{eqnarray} \P(\text{reaching } \rho \text{ and } P_{\rho} \in I) &=& \P_{\xi_i, i\in S}(\text{reaching } \rho)\P_{\xi_i, i\notin S}(P_\rho \in I)\nonumber\\ &\le& \P_{\xi_i, i\in S}(\text{reaching } \rho) \left (\frac{Cd}{r^{1/2d}} +Cd\tau'^{1/(4d+1)} \right ) \nonumber\\ &\le& \P_{\xi_i, i\in S}(\text{reaching } \rho) \left (\frac{Cd}{r^{1/2d}} + C'd^{4/3}\tau^{1/(4d+1)}\left(\log \frac{1}{\tau}\right)^{1/4}\right )\label{mon3} \end{eqnarray} Since the events that the root $P$ reaches different leaves on the tree are disjoint, from \eqref{mon1}, \eqref{mon2}, and \eqref{mon3}, we get that for any $0<\tau<\frac{1}{3}$, \begin{eqnarray} \P(P \in I) &\le& 2\exp\left (-\frac{r\tau}{4C^{d+1}(d\log\log r+d\log d) }\right )+\frac{Cd}{r^{1/2d}}+ C'd^{4/3}\tau^{1/(4d+1)}\left(\log \frac{1}{\tau}\right)^{1/4}+ \frac{1}{r}.\label{tau1} \end{eqnarray} Set $\tau = \frac{8C^{d+1}\log r(d\log\log r+d\log d) }{r}$ then $\tau<\frac{1}{3}$ because we assumed that $d\le \frac{2\log r}{\log\log r}$. The first term on the right of \eqref{tau1} becomes $2r^{-2}$ and the third term is bounded from above by $B\frac{d^{4/3}\log ^{1/2}r}{r^{1/(4d+1)}}$. This completes the proof of the first bound. \subsection{Second bound} We next build on the arguments in the previous section to prove the second bound in Theorem \ref{theorem:LOpoly1}. The main ingredient in proving the second bound is the following technical lemma of \cite{kane2014pseudorandom} which says that a random restriction of a sufficiently regular polynomial will likely have a much larger expectation compared to its standard-deviation. This is useful because polynomials with large expectation relative to standard-deviation have small probability of vanishing by tail bounds such as Theorem \ref{thm6}. In case the tail bound does not give a sufficiently good bound, we recurse on the new restricted polynomial. To state the lemma we need the following definition: For $\gamma \geq 0$, call a polynomial $P:\R^n \to \R$ $\gamma$-\textit{spread} if $\Var(P(\xi_1, \dots, \xi_n))^{1/2} \geq |\E(P(\xi_1, \dots, \xi_n))|/\gamma$. \begin{proposition}\label{lm:goodblocks} Let $b, n$ be such that $b | n$. Let $P:\R^n \to \R$ be a non-constant $\tau$-regular degree $d$ polynomial. Let $S_1,\ldots,S_b$ be a partition of $[n]$ into equal-sized blocks. For $\ell \in [b]$, and an assignment $\xi^l \in \{1, -1\}^{[n] \setminus S_\ell}$ to the variables not in $S_\ell$, let $P_{\xi^\ell}:\R^{S_\ell} \to \R$ denote the polynomial obtained by fixing the variables not in $S_\ell$ to $\xi^l$. Then, $$\sum_{l=1}^b \P_{\xi^l}(P_{\xi^l} \text{ is $\gamma$-spread}) \leq 2^{O(d)} \cdot (\gamma^2 + 1) \cdot \left (\sqrt{b} + b \tau^{1/8d}\right), $$ where for clarity, the assignments $\xi^{l}$ for different $l$ are independent. In particular, there exists an index $l \in [b]$, such that $$\P_{\xi^l}(P_{\xi^l} \text{ is $\gamma$-spread}) \leq 2^{O(d)} \cdot (\gamma^2+ 1) \cdot \left (1/\sqrt{b} + \tau^{1/8d}\right ).$$ \end{proposition} For the proof, we need the following definitions from \cite{kane2014correct}: \begin{itemize} \item For a function $f:\R^n \to \R$ and a vector $v \in \R^n$, $D_v f(x) = v \cdot \nabla f(x)$. \item Let $\zeta = (\zeta_1,\ldots,\zeta_n)$ and $\xi = (\xi_1,\ldots,\xi_n)$ be independent collections of Rademacher random variables. For a polynomial $P:\R^n \to \R$, define $$\alpha(P) = \E_{\zeta, \xi}\left(\min\left(1, \frac{|D_{\zeta} P(\xi)|^2}{|P(\xi)|^2}\right)\right).$$ \end{itemize} The following claims are implicit in \cite{kane2014correct}. \begin{lemma}\label{lm:alphaspread} For any polynomial $P:\R^n \to \R$, $\Var(P) \leq 2^{O(d)} (\E(P)^2 + \Var(P)) \cdot \alpha(P)$. \end{lemma} \begin{proof} The claim is proved in \cite[Lemma 21]{kane2014correct}. \end{proof} \begin{lemma}\label{lm:alphabound} Let $b, n$ be such that $b | n$. Let $P:\R^n \to \R$ be a non-constant $\tau$-regular degree $d$ polynomial. Let $S_1,\ldots,S_b$ be a partition of $[n]$ into equal-sized blocks. For $\ell \in [b]$, and an assignment $\xi^l \in \{1, -1\}^{[n] \setminus S_\ell}$ to the variables not in $S_\ell$, let $P_{\xi^\ell}:\R^{S_\ell} \to \R$ denote the polynomial obtained by fixing the variables not in $S_\ell$ to $\xi^l$. Then, \begin{equation} \sum_{\ell=1}^b \E_{\xi^\ell}(\alpha(P_{\xi^\ell})) = O(d^3 \alpha(P) \sqrt{b} + d^4 b \tau^{1/(8d)}),\label{kan} \end{equation} where for clarity, the assignments $\xi^{l}$ for different $l$ are independent. \end{lemma} \begin{proof} Notice that the right-hand side of \eqref{kan} doesn't change if the assignments $\xi^{l}$ are obtained by choosing $n$ random variables $\xi_1, \dots, \xi_n$ and then looking at the $b$ different restrictions $\xi^{l}$. The lemma is then proved in \cite[Proposition 19]{kane2014correct} (essentially Equation (4)). \end{proof} Combining the above two claims gives us the proposition. \begin{proof}[Proof of Proposition \ref{lm:goodblocks}] For any index $\ell \in [b]$, we have \begin{align*} \P(P_{\xi^\ell} \text{ is $\gamma$-spread}) &= \P(\gamma^2 \Var(P_{\xi^\ell}) \geq \E(P_{\xi^\ell})^2)\\ &= \P\left(\frac{\Var(P_{\xi^\ell})}{\E(P_{\xi^\ell})^2 + \Var(P_{\xi^\ell})} \geq \frac{1}{\gamma^2 + 1}\right)\\ &\leq \P(\alpha(P_{\xi^\ell}) 2^{O(d)} \geq 1/(\gamma^2 + 1)) \text{ (by Lemma \ref{lm:alphaspread} applied to $P_{\xi^\ell}$)}\\ &\leq 2^{O(d)} \cdot (\gamma^2 + 1) \cdot \E(\alpha(P_{\xi^\ell})) \text{ (by Markov's inequality)}. \end{align*} Therefore, by Lemma \ref{lm:alphabound}, \begin{align*} \sum_{\ell=1}^b \P(P_{\xi^\ell} \text{ is $\gamma$-spread}) &\leq 2^{O(d)} \cdot (\gamma^2 + 1) \cdot \sum_{\ell=1}^b\E(\alpha(P_{\xi^\ell})) \\ &= 2^{O(d)} \cdot (\gamma^2 + 1) \cdot O(d^3 \alpha(P) \sqrt{b} + d^4 b \tau^{1/(8d)})\\ &= 2^{O(d)} \cdot (\gamma^2 + 1) \cdot (\alpha(P) \sqrt{b} + b \tau^{1/8d}). \end{align*} The claim now follows as $\alpha(P) \leq 1$ by definition. \end{proof} We are now ready to prove the second bound of Theorem \ref{theorem:LOpoly1}. Similar to the proof of the first bound, without loss of generality, we can assume that $I = [-1, 1]$, $r$ is sufficiently large, and that $d\le \frac{\sqrt{\log r}}{\log \log r}$. Let, \begin{equation} f(r,d) = \max\{\P(P(\xi)\in I): \text{ $P$ degree $d$ polynomial with $\text{rank}(P) \geq r$}\}. \end{equation} Let $P$ be a degree $d$ multi-linear polynomial with $\text{rank}(P) = r$ achieving the minimum $f(r,d)$. For fixed parameters $\tau\in (0, 1/3)$ and $\gamma> 2$ to be chosen later, let $\beta = \frac{1}{r}$ and let $\mathcal{T}$ be a decision tree as guaranteed by Proposition \ref{regularity lemma} with $M = \lceil\frac{r\tau}{2\alpha}\rceil$ where $\alpha$ and $\tau'$ are as in that Proposition. Then the depth of the tree is at most $\frac{r}{2}$, and as in the proof of the first bound, \begin{equation}\label{eq:mainproof1} \P(P(\xi) \in I) \leq 2\exp\left (-\frac{r\tau}{4C^{d}\alpha}\right ) + \frac{1}{r} + \P[P_\rho(\xi) \in I | \text{ $P_\rho$ is $\tau'$-regular}]. \end{equation} Now, consider a leaf $\rho$ so that $Q \equiv P_\rho$ is $\tau'$-regular. Note that $\text{rank}(Q) \geq r/2$ and in particular $Q$ is non-constant. Fix $b < r/4$, a parameter to be chosen later. Fix a partition $S_1,\ldots,S_b$ of the variables of $Q$ such that for $\ell \in [b]$, the restricted polynomials $Q^\ell$ obtained by fixing the variables not in $S_\ell$ each satisfy $\text{rank}(Q^\ell) \geq \lfloor \text{rank}(Q)/b\rfloor$ (this can be done for instance by first partitioning the variables witnessing $\text{rank}(Q)$). Note that if the number of variables in $Q$ is not divisible by $b$, we only need to add a few variables to $Q$ without affecting its output nor its regularity. Now, by Proposition \ref{lm:goodblocks} applied to the polynomial $Q$, there exists $\ell \in [b]$ such that the polynomial $Q^\ell$ obtained by a random assignment to the variables not in $A^\ell$ is $\gamma$-spread with probability at most $$2^{O(d)} \cdot (\gamma^2 + 1) \cdot \left ({1/\sqrt{b} + \tau'^{1/8d}}\right ).$$ Therefore, \begin{eqnarray*} \P(Q(y) \in I) &\leq& 2^{O(d)} \cdot (\gamma^2 + 1) \cdot \left ({1/\sqrt{b} + \tau'^{1/8d}}\right ) \cdot \P( Q^\ell(z) \in I |\text{ $Q^\ell$ is $\gamma$-spread})+ \\ &&\P(Q^\ell(z) \in I | \text{ $Q^\ell$ is not $\gamma$-spread}) \\ &\leq& 2^{O(d)} \cdot (\gamma^2 + 1) \cdot \left ({1/\sqrt{b} + \tau'^{1/8d}}\right ) \cdot f(\lfloor \text{rank}(Q)/b\rfloor,d) + \P(Q^\ell(z) \in I | \text{ $Q^\ell$ is not $\gamma$-spread}). \end{eqnarray*} Finally, to bound the last term, observe that if $Q^\ell$ is not $\gamma$-spread and not identically zero, then \begin{eqnarray*} \P({Q^\ell(z) \in I}) &=& \P(|Q^{\ell}|\le 1) \leq \P({\left|Q^\ell(z) - \E(Q^\ell)\right| \geq |\E(Q^\ell)|-1})\\ &\leq& \P\left ({\left|Q^\ell(z) - \E(Q^\ell)\right| \geq \frac{\gamma \Var(Q^\ell)^{1/2}}{2}}\right )\\ &\leq& 2 \exp\left ({- \Omega(1) \gamma^{2/d}}\right ) \text{ (by Theorem \ref{thm6})}, \end{eqnarray*} where in the next to last inequality, we use the inequalities $|\E(Q^{\ell})|\ge \gamma. \Var(Q^{\ell})^{1/2}\ge \gamma .\text{rank} (Q^{\ell})^{1/2}\ge \gamma.(r/2b)^{1/2}\ge 2$ and so $|\E (Q^{\ell})|-1\ge \frac{|\E (Q^{\ell})|}{2}\ge \frac{\gamma \Var(Q^{\ell})^{1/2}}{2}$. Combining the above arguments, we get that if $b \leq r/4$, $$\P({Q(x) \in I}) \leq 2^{O(d)} \cdot (\gamma^2 + 1) \cdot \left ({1/\sqrt{b} + \tau'^{1/8d}}\right ) \cdot f(\lfloor r/b\rfloor,d) + O(1) \exp\left ({- \Omega(1) \gamma^{1/2d}}\right ) .$$ Hence, by \eqref{eq:mainproof1} we have that \begin{equation} \P({P(x) \in I}) \leq 2\exp\left (-\frac{r\tau}{4C^{d}\alpha}\right ) + \frac{1}{r} + 2^{O(d)} \cdot (\gamma^2 + 1) \cdot \left ({1/\sqrt{b} + \tau'^{1/8d}}\right ) \cdot f(\lfloor r/b\rfloor,d) + O(1) \exp\left ({- \Omega(1) \gamma^{2/d}}\right ).\label{rec} \end{equation} Now, as in the proof of the first bound of Theorem \ref{theorem:LOpoly1}, set $\tau = \frac{8C^{d+1}\log r(d\log\log r+d\log d) }{r}$, $b = r^{1/4d}/(d\log r)^{Cd}$, and $\gamma = (C \log r)^{d/2}$. Then, $$f(r,d) \leq (C \log r))^{Cd} \cdot f(r^{1-1/4d},d) \cdot r^{-1/8d}.$$ (here we used the fact that $f(r, d)\ge \Omega (r^{-1/2})$ by choosing the polynomial $p(\xi_1, \dots, \xi_{rd})=\xi_1\xi_2\dots\xi_d+\xi_{d+1}\dots\xi_{2d}+\dots + \xi_{rd-d+1}\dots\xi_{rd}$, and so all the other terms on the right-high side of \eqref{rec} are dominated by the term $(C \log r))^{Cd} \cdot f(r^{1-1/4d},d) \cdot r^{-1/8d}$.) Let $a = 1 - 1/4d$. Applying this recurrence relation $k$ times with $r^{a^{k}} = C$ (so $k = \Theta(d\log\log r)$), we get \begin{eqnarray} f(r,d) &\leq& (C \log r))^{kCd} \left (\prod_{i=0}^{k-1}a^{i}\right )^{Cd}\cdot f(r^{a^{k}},d) \cdot r^{-(\sum_{i=0}^{k-1}a^{i})/8d}\nonumber\\ &\le& e^{O(d^{2}(\log \log r)^{2})}r^{-(1-a^{k})/2} = Ce^{O(d^{2}(\log \log r)^{2})}r^{-1/2},\nonumber \end{eqnarray} completing the proof of the second bound and hence Theorem \ref{theorem:LOpoly1}. \section{General distributions}\label{gen-proof} \subsection{Proof of Theorem \ref{th:littlewoodoffordbiasedmain}} We reduce the $p$-biased case to the uniform distribution at the expense of a loss in the rank of the polynomial and then apply Theorem \ref{theorem:LOpoly1}. First notice that if $x\sim \mu_p$, then $1-x\sim \mu_{1-p}$. And so, by replacing the polynomial $P$ by $Q(x_1, \dots, x_n) = P(1- x_1, \dots, 1-x_n)$, we can exchange the roles of $p$ and $1-p$. Therefore, without loss of generality, we assume that $\alpha = p\le 1/2$. Our assumption $2^{d}p^{d}r\ge 3$ guarantees that $\log \log (2^{d}p^{d}r) = \Omega(1)$ and hence by choosing the implicit constants on the right-hand side of Theorem \ref{th:littlewoodoffordbiasedmain} to be sufficiently large, we can assume that $2^{d}p^{d}r$ is greater than 100 (say). Let $\eta_1, \dots, \eta_n$ and $\xi_1', \dots, \xi_n'$ be independent Bernoulli random variables with $\P(\eta_i=0) = 1/2$ and $\P(\xi_i' = 0) = 1-2p$. Let $\xi_i = \eta_i\xi_i'$ then $\xi_1, \dots, \xi_n$ are iid Bernoulli variables with $\P(\xi_i = 0) = 1-p$. Therefore, we need to bound $\P(P(\xi_1, \dots, \xi_n)\in I)$. From the definition of $\text{rank}(P)$, there exist disjoint sets $S_1, \dots, S_r$ such that $|a_{S_j}|\ge 1$ for all $j=1, \dots, r$. We have $P(\xi_1, \dots, \xi_n) = \sum_{S\subset [n], |S| \le d} \left (a_{S}\prod_{i\in S}\xi_i'\right )\prod_{i\in S}\eta_i$. Conditioning on the $\xi_i'$'s, $P$ becomes a polynomial of degree $d$ in terms of $\eta_i$ whose coefficients associated with $S_j$ are $b_{S_j}:=a_{S_j}\prod_{i\in S_j}\xi_i'$ accordingly. For each such $j$, one has \begin{eqnarray} \P_{\xi_1', \dots, \xi_n'}(|b_{S_j}|\ge 1)= \P(\xi_i'=1, \forall i\in S_j) = (2p)^{d}.\nonumber \end{eqnarray} Now, since the sets $S_j$ are disjoint, the events $|b_{S_j}|\ge 1$ are independent. Define $X = \sum_{j=1, \dots, r}\textbf{1}_{|b_{S_j}|\ge 1}$. By the classical Chernoff's bound we have, for $0<\gamma<1$, $\P(|X-\E X| \ge \gamma\E X) \le 2e^{-\gamma^{2}\E X/3}$. Thus, we conclude that with probability at least $1 - \exp(-2^{d-1}p^{d}r/6)$, there are at least $2^{d-1}p^{d}r$ indices $j$ with $|b_j|\ge 1$. Conditioning on this event, we obtain a polynomial of degree $d$ in terms of $\eta_1, \dots, \eta_n$ which has rank at least $2^{d-1}p^{d}r$. The theorem now follows from applying Theorem \ref{theorem:LOpoly1} to this polynomial and noting that the additional error of $\exp(-2^{d-1}p^{d}r/6)$ is smaller than both terms from Theorem \ref{theorem:LOpoly1}. \subsection{Proof of Theorem \ref{thm:generalist}} By replacing $P(x_1, \dots, x_n)$ by $Q(x_1, \dots, x_n) = P(x_1+y_1, \dots, x_n+y_n)$ and $\xi_i$ by $\xi_i-y_i$, we can also assume without loss of generality that $y_i=0$ for all $i$. Furthermore, we can assume that $\P(\xi_i\le 0)=p$ for all $i$. Indeed, if for some $i$, $\P(\xi_i>0)=p$, we replace $\xi_i$ by $-\xi_i$ and modify the polynomial $P$ accordingly to reduce to the case $\P(\xi_i<0)=p$. And then the proof runs along the same lines as in the case $\P(\xi_i\le 0)=0$. For each $i=1, \dots, n$, let $\xi_i^{+}$ and $\xi_i^{-}$ be independent random variables satisfying $\P(\xi_i^{+}\in A) = \P(\xi_i\in A|\xi_i>0)$ and $\P(\xi_i^{-}\in A) = \P(\xi_i\in A|\xi_i\le 0)$ for all measurable subset $A\subset \R$. Let $\eta_1, \dots, \eta_n$ be iid random Bernoulli variables (independent of all previous random variables) such that $\P(\eta_i = 0) = p$. Let $\xi_i' = \eta_{i}\xi_i^{+} + (1-\eta_{i})\xi_i^{-}$, then $\xi_i'$ and $\xi$ have the same distribution. Therefore, it suffices to bound the probability that $P(\xi_1', \dots, \xi_n')$ belongs to $I$. One has \begin{equation} P(\xi_1', \dots, \xi_n') = P(\eta_1(\xi_1^{+}-\xi_1^{-}) + \xi_1^{-}, \dots, \eta_n(\xi_n^{+}-\xi_n^{-}) + \xi_n^{-}) = \sum_{S\subset [n], |S| = d} \left (a_{S}\prod_{i\in S}(\xi_i^{+}-\xi_i^{-})\right )\prod_{i\in S}\eta_i + Q,\nonumber \end{equation} where $Q$ is some polynomial which has degree $<d$ in terms of $\eta_i$ when all the $\xi_i^{\pm}$ are fixed. From the definition of $\text{rank}(P)$, let $S_1, \dots, S_r$ be disjoint subsets of $[n]$ with $|a_{S_j}|\ge 1$ for all $1\le j\le r$. Conditioning on the variables $\xi_i^{\pm}$, the polynomial $P$ becomes a polynomial of degree $d$ in terms of $\eta_i$ whose coefficients associated with $S_j$ are $b_{S_j}:=a_{S_j}\prod_{i\in S_j}(\xi_i^{+}-\xi_i^{-})$ accordingly. For each such $j$, one has \begin{eqnarray} \P_{\xi_1^{\pm}, \dots, \xi_n^{\pm}}(|b_{S_j}|\ge 1)\ge \P(\xi_i^{+}-\xi_i^{-}\ge 1, \forall i\in S_j).\nonumber \end{eqnarray} Since $\xi_{i}^{+}\ge 0\ge \xi_i^{-}$ a.e., one has $2\P(\xi_i^{+}-\xi_i^{-}\ge 1)\ge \P(|\xi_i^{+}\ge 1) + \P(|\xi_i^{-}\le -1) = \P(|\xi_i|\ge 1)\ge \epsilon$. Hence, \begin{equation} \P_{\xi_1^{\pm}, \dots, \xi_n^{\pm}}(|b_{S_j}|\ge 1)\ge 2^{-d}\epsilon^{d}.\nonumber \end{equation} Now, since the sets $S_j$ are disjoint, the events $|b_{S_j}|\ge 1$ are independent. Therefore, using a Chernoff-type bound as in the proof of Theorem \ref{th:littlewoodoffordbiasedmain}, one can conclude that with probability at least $1 - \exp(-2^{-d}\epsilon^{d}r/12)$, there are at least $r2^{-d}\epsilon^{d}/2$ indices $j$ with $|b_j|\ge 1$. Conditioning on this event, we obtain a polynomial of degree $d$ in terms of $\eta_1, \dots, \eta_n$ which has rank at least $r2^{-d}\epsilon^{d}/2$. Using Theorem \ref{th:littlewoodoffordbiasedmain}, one obtains the desired bound. \section{Proof of Theorem \ref{th:orapprox}} \label{app-proof} \newcommand{\overline{0}}{\overline{0}} Let $a$ be an integer to be chosen later. Let $D = \lfloor \log_a (\log_2 n-1) \rfloor$ be the largest integer such that $2^{-a^D} \geq 2/n$. Let $\mu$ be the distribution obtained by the following procedure: \begin{enumerate} \item With probability $1/2$ output $x =\overline{0}$ (the all $0$'s vector). \item With probability $1/2$ pick an index $i \in \{1,\ldots,D\}$ uniformly at random and output $x \sim \mu_{2^{-a^i}}^n$. \end{enumerate} We next show that for some constant $c > 0$, there exists no polynomial $P$ of degree $d < c (\log \log n)/(\log \log \log n))$ such that $\P_{x \sim \mu}(P(x) = OR(x)) \geq 2/3$. Let $P$ be such a polynomial. Then, necessarily, $P(\overline{0}) = 0$; as $\P_{x\sim \mu}(P(x)=0) \le 1/2 + 1/2(1-2^{-a^{D}})^n \leq 1/2 + (1/2) (1-2/n)^n <2/3$, there must exist a set of indices $I \subseteq [D]$ with $|I| \geq \Omega(D)$ such that for all $i \in I$, $$\P_{x \sim \mu_{2^{-a^i}}}(P(x) = 1) = \Omega(1).$$ Let $I = \{i_1 < i_2 < \cdots < i_k \}$ and for $\ell \in [k]$, let $p_\ell = 2^{-a^{i_\ell}}$. Now, by Theorem \ref{th:littlewoodoffordbiasedmain} applied to the polynomial $P-1$ and $x \sim \mu_{p_1}^n$, we get that either $\text{rank}(P) \leq (3/2p_1)^d$ or $$\Omega(1) = \P(P(x) = 1) \leq O(d^{4/3}) \frac{\log(\text{rank}(P) (2p_1)^d)^{1/2}}{(\text{rank}(P) (2p_1)^d)^{1/(4d+1)}}.$$ Hence, in any case, $\text{rank}(P) \leq r_1 = (d)^{O(d)}/p_1^d$. This in turn implies that there exists a set of $r_1 \cdot d$ indices $S_1\subseteq [n]$ such that the polynomial $P_1 = P_{S_1}$ obtained by assigning the variables in $S_1$ to $0$ is of degree at most $d-1$. Further, for $x \sim \mu_{p_{2}}^{[n]}$, \begin{eqnarray*} \Omega(1) &=& \P_{x}(P(x) = 1) = \P(x_{S_1} = 0) \cdot \P_{x}(P(x) = 1 | x_{S_1} = 0) + \P(x_{S_1} \neq 0) \cdot \P_x(P(x) = 1 | x_{S_1} \neq 0)\\ &\leq& \P_{x \sim \mu_{p_{2}}^{[n] \setminus [S_1]}}(P_1(x) = 1) + \P(x_{S_1} \neq 0)\\ &\leq& \P_{x \sim \mu_{p_{2}}^{[n] \setminus [S_1]}}(P_1(x) = 1) + |S_1| \cdot p_{2}. \end{eqnarray*} Thus, $$\P_{x \sim \mu_{p_{2}}^{[n] \setminus [S_1]}}(P_1(x) = 1) \geq \Omega(1) - d^{O(d) + 1} \cdot (p_2/p_1^d) = \Omega(1) - d^{O(d)+1} 2^{-a^{i_2} + d a^{i_1}} \geq \Omega(1) - d^{O(d)} 2^{-a^{i_1}},$$ for $a \geq 2d$. Further, note that $P_1(\overline{0}) = 0$. Iterating the argument with $P_1$ and so forth, we get a sequence of polynomials $P_1, P_{2}, \ldots,P_{k-1}$ such that for $1 \leq j \leq \min(d,k-1)$, $P_{j}$ is of degree at most $d-j$, $P_{j}(\overline{0}) = 0$ and for $x \sim \mu_{p_{j+1}}^{[n] \setminus (S_1 \cup \cdots \cup S_{j})}$, $$\P_x(P_{j}(x) = 1) = \Omega(1) - d^{O(d)+j} 2^{-a}.$$ This clearly leads to a contradiction if $k > d$ and $a \geq C d \log d$ for a large enough constant $C$ (so that the right hand side of the above equation is non-zero for $j = d$). Therefore, setting $a = C d \log d$, for a sufficiently big constant $C$, we must have $k = \Omega(D) \leq d$. That is, $\log_2 (n-1) = a^{O(d)} = d^{O(d)}$. Thus, we must have $d = \Omega(1) (\log \log n)/(\log \log \log n)$. \bibliographystyle{siam}

https://arxiv.org/abs/1503.07933

On Limits of Dense Packing of Equal Spheres in a Cube

We examine packing of $n$ congruent spheres in a cube when $n$ is close but less than the number of spheres in a regular cubic close-packed (ccp) arrangement of $\lceil p^{3}/2\rceil$ spheres. For this family of packings, the previous best-known arrangements were usually derived from a ccp by omission of a certain number of spheres without changing the initial structure. In this paper, we show that better arrangements exist for all $n\leq\lceil p^{3}/2\rceil-2$. We introduce an optimization method to reveal improvements of these packings, and present many new improvements for $n\leq4629$.

\section{Introduction} We consider the problem of finding the densest packings of congruent, non-overlapping, spheres in a cube. Equivalently, we can search for an arrangement of points inside a unit cube so that the minimum distance between any two points is as large as possible. The maximum separation distance of $n$ points in $[0,1]^{3}$ we denote by $d_{n}$. To our knowledge, the optimality of $d_{n}$ is proved for $n=2,3,4,5,6,8,9$ \cite{schaer}, $n=10$ \cite{schaer_10} and $n=14$ \cite{joos}. Optimality of $d_{n}$ was conjectured for an infinite family of packings where $\lceil p^{3}/2\rceil$ spheres are arranged in a cubic close-packed (ccp) structure \cite{goldberg}. We denote by $g(p)=\lceil p^{3}/2\rceil$ the number of spheres in these packings, with a maximum separation distance denoted by $d'_{p}=\sqrt{2}/\!\left(p-1\right)$. In this paper, we examine arrangements when $n$ is close, but less than $g(p)$. For this family of packings $d_{n}=d'_{p}$ is often assumed, to mean that the densest known arrangements are derived from ccp by omission of a certain number of spheres without changing the initial structure. Limiting values though were not provided. It was conjectured that $d_{n}$ is constant in the range $12\leq n\leq14$ \cite{goldberg} and $29\leq n\leq32$ \cite{gensane}. A better packing was found for $n=12$ \cite{gensane}. Similarly, previous search results showed the same trend for $60\leq n\leq63$, $103\leq n\leq108$, $\ldots$, $817\leq n\leq864$ \cite{wenqi,locatelli_de,packomania}. In Section 2 we show that most of the listed packings can be improved by proving that $d_{n}>d'_{p}$ for all $n\leq g(p)-2$. We also provide a lower bound for these improvements. In Section 3, we introduce an optimization method and improve the lower bound for $4\leq p\leq21$. We show that the described procedure can be used as a good packing method when $n$ is slightly smaller than $g(p)$. We run search to determine improvements for other packings when $n=g(p)-r$, for $3\leq r\leq6$, $r<p$ and $4\leq p\leq12$. \section{Existence of improved packings} To simplify our notation, we will assume that the radius of all spheres in the packing is 1 and that our task is to determine the smallest size of a cube that contains all spheres. We denote the family of all finite sets of points such that the distance between any two points is at least 2 by \[ \mathcal{F}=\left\{ S\subset\mathbb{R}^{3}:\Vert s_{1}-s_{2}\Vert \geq 2\text{ for all distinct } s_{1},s_{2}\in S \right\}. \] Let $S_{n}=\left\{ s_{1,}\ldots,s_{n}\right\} \in\mathcal{F}$ and let $D_{c}(s_{i},s_{j})$ be the Chebyshev distance between any two points $s_{i},s_{j}\in S_{n}$. The smallest edge length of a cube, with edges parallel to the axes, such that it contains all points $S_{n}$ is equal to \[ D(S_{n})=\max\left\{ D_{c}(s_{i},s_{j}):1\leq i<j\leq n\right\}. \] We notice that the maximum separation distance $d_{n}$ can also be given as \[ d_{n}=\max\frac{2}{D(S_{n})}. \] \begin{theorem*} The maximum separation distance of $n$ points contained in a closed region bounded by a unit cube is larger than $\sqrt{2}/\left(p-1\right)$ for all $n\leq\lceil p^{3}/2\rceil-2$.\end{theorem*} \begin{proof} Let $C_{p}$ be a closed region bounded by a cube with an edge length $D_{p}=2/d'_{p}=(p-1)\sqrt{2}$, where $C_{p}$ is defined by $C_{p}=\left[0,D_{p}\right]^{3}$. Let $G_{p}$ be a set of $g(p)$ sphere centers in a ccp arrangement, such that $G_{p}\in\mathfrak{\mathcal{F}}$, $G_{p}\subset C_{p}$ and $\langle0,0,0\rangle\in G_{p}$ (see Figure 1). \begin{figure}[ht!] \centering \begin{subfigure}{.45\textwidth} \centering \includegraphics[width=.9\linewidth]{./G_2.pdf} \caption{$G_{2}$} \vspace*{10mm} \end{subfigure}% \begin{subfigure}{.45\textwidth} \centering \includegraphics[width=.9\linewidth]{./G_3.pdf} \caption{$G_{3}$} \vspace*{10mm} \end{subfigure} \begin{subfigure}{.6\textwidth} \centering \includegraphics[width=.9\linewidth]{./G_4.pdf} \caption{$G_{4}$} \end{subfigure} \caption{An illustration of arrangements $G_{2}$, $G_{3}$ and $G_{4}$.} \end{figure} For two sets of points $A,B\in\mathfrak{\mathcal{\mathcal{F}}}$ let \[ h(A,B)=\min\left\{ \Vert a-b\Vert-2:a\in A,b\in B\right\} \] and let $L_{p}=G_{p}\setminus G_{p-1}$ (see Figure 2(a) for an example). We denote the improved packing of $g(p)-2$ points by $P_{p}=\left\{ s_{1},\ldots,s_{g(p)-2}\right\} \in\mathcal{\mathcal{F}}$ such that \begin{equation} P_{p}\subset C_{p}\label{eq:p_subset}, \end{equation} \begin{equation} h(P_{p},L_{p+1})>0\label{eq:p_separation}. \end{equation} Equations (\ref{eq:p_subset}) and (\ref{eq:p_separation}) directly imply that if these statements are true, then $D(P_{p})<D_{p}$. \begin{figure}[ht!] \centering \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=.9\linewidth]{./L_5.pdf} \caption{$L_{5}$} \vspace*{10mm} \end{subfigure}% \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=.9\linewidth]{./L_51.pdf} \caption{$L_{5,1}$} \vspace*{10mm} \end{subfigure} \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=.9\linewidth]{./L_52.pdf} \caption{$L_{5,2}$} \vspace*{10mm} \end{subfigure}% \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=.9\linewidth]{./L_53.pdf} \caption{$L_{5,3}$} \vspace*{10mm} \end{subfigure} \caption{An illustration of arrangements $L_{5}$, $L_{5,1}$, $L_{5,2}$ and $L_{5,3}$.} \end{figure} Let $T_{p}=\tau(L_{p})$ be a set of points created by the translation of points from $L_{p}$ by the function $\tau$ such that \begin{equation} T_{p}\in\mathcal{F},T_{p}\subset C_{p},h(P_{p-1},T_{p})\geq0\label{eq:tp_validated}, \end{equation} \begin{equation} h(T_{p},L_{p+1})>0\label{eq:tp_separated} \end{equation} \noindent for all $p>2$. If such a set $T_{p}$ exists, then we can state that \[ P_{p}=P_{p-1}\cup T_{p}\text{, for all }p>2. \] We can easily show that $P_{2}$ exists by constructing it explicitly. If we want to give $P_{2}$ in such a way to maximize the capability to translate points from $L_{3}$ and thus produce a better packing $P_{3}$, we have to find the positive root of the polynomial \[ a^{4}+4a^{3}+8a^{2}-8=0,a>0\Longrightarrow a=0.818425\ldots \] \noindent then $P_{2}$ is given as $P_{2}=\left\{ \langle0,0,a\rangle,\langle b,b,0\rangle\right\}$, where $b=\sqrt{2-a^{2}/2}$. While $P_{p-1}$ and $L_{p}$ are separated (\ref{eq:p_separation}) , we can try to translate all points from $L_{p}$ and keep (\ref{eq:tp_validated}), and (\ref{eq:tp_separated}) true. We split $L_{p}$ into three subsets, and perform the translations on each of them while maintaining the given conditions. Let $L_{p}=L_{p,1}\cup L_{p,2}\cup L_{p,3}$ such that \[ L_{p,1}=\left\{ \langle l_{1},l_{2},l_{3}\rangle\in L_{p}:l_{i},l_{j}\leq(p-3)\sqrt{2}\text{ for some distinct }i,j\in\left\{ 1,2,3\right\}\right\}, \] \[ L_{p,2}=\left\{ \langle l_{1},l_{2},l_{3}\rangle\in L_{p}\setminus L_{p,1}:l_{1},l_{2},l_{3}>0\right\}, \] \[ L_{p,3}=L_{p}\setminus L_{p,1}\setminus L_{p,2}. \] For an illustration of sets $L_{5}$, $L_{5,1}$, $L_{5,2}$ and $L_{5,3}$, see Figure 2. Now we can give $T_{p}$ as \[ T_{p}=T_{p,1}\cup T_{p,2}\cup T_{p,3}, \] \[ T_{p,i}=\left\{ \langle l_{1}-u_{i,1}\tau_{i}(p),l_{2}-u_{i,2}\tau_{i}(p),l_{3}-u_{i,3}\tau_{i}(p)\rangle:\langle l_{1},l_{2},l_{3}\rangle\in L_{p,i}\right\}, \] \[ u_{1,j}=\begin{cases} 1 & \text{if }l_{j}=\left(p-1\right)\sqrt{2}\\ 0 & \text{otherwise} \end{cases},u_{2,j}=1,u_{3,j}=\begin{cases} 1 & \text{if }l_{j}\neq0\\ 0 & \text{otherwise} \end{cases}, \] where $\tau_{1}(p)$, $\tau_{2}(p)$ and $\tau_{3}(p)$ are small numbers such that $\tau_{1}(p)>\tau_{2}(p)>\tau_{3}(p)$ and conditions (\ref{eq:tp_validated}) and (\ref{eq:tp_separated}) hold. If we additionally state that $h(P_{p-1},T_{p,1})=0$, $h(P_{p-1}\cup T_{p,1},T_{p,2})=0$ and $h(P_{p-1}\cup T_{p,1}\cup T_{p,2},T_{p,3})=0$, we can give explicit solutions for $\tau_{1}$, $\tau_{2}$ and $\tau_{3}$ as follows \[ \tau_{1}(3)=2\sqrt{2}-2-a, \] \[ \tau_{2}(p)=\frac{\sqrt{2}}{3}\left(\frac{\tau_{1}(p)}{\sqrt{2}}+2-\sqrt{-\tau_{1}(p)^{2}+2\sqrt{2}\tau_{1}(p)+4}\right)\text{, for all }p\geq3, \] \[ \tau_{3}(p)=\frac{\sqrt{2}}{2}\left(\sqrt{2}\tau_{2}(p)+1-\sqrt{-\tau_{2}(p)^{2}+2\sqrt{2}\tau_{2}(p)+1}\right)\text{, for all }p\geq3, \] \[ \tau_{1}(p)=\sqrt{2}+\tau_{3}(p-1)-\sqrt{-\tau_{3}(p-1)^{2}+2\sqrt{2}\tau_{3}(p-1)+2}\text{, for all }p>3. \] We notice that $D(P_{p})=D_{p}-\tau_{3}(p)$, hence the maximum separation distance can be given as \[ d_{n}\geq\frac{2}{\left(p-1\right)\sqrt{2}-\tau_{3}(p)}\text{, for all }n\leq g(p)-2 \] and while $\tau_{3}(p)>0$, then $d_{n}>\sqrt{2}/\!(p-1)$. \end{proof} We denote the lower bound of improvements by \[ I_{p}=d_{g(p)-2}-d'_{p}. \] By performing the calculations, we get particular values such as $I_{3}>8.235\cdot10^{-11}$, $I_{4}>1.276\cdot10^{-79},\ldots$. This is a rough approximation while condition (\ref{eq:tp_separated}) needs to remain true for all $p$ and does not allow us to further improve the packing for particular $p$. In the next section, we show that $I_{p}$ is usually above this approximation. We did not find a way to improve the packings when $n=g(p)-1$, and we conjecture that in this case $d_{n}=d'_{p}$. \section{Optimization Approach} Most of the existing packing methods focus on searching for a completely new arrangement of spheres, usually performing a search from a randomly given initial position of spheres. Such approach assumes that the packing of higher density can be reached after a certain number of iterations and multiple runs of the search procedure using different initial parameters \cite{gensane,wenqi,locatelli_de}. The large number of iterations often limits the search procedure to the use of double or quadruple floating-point precision, to maintain computation speed. This precision is insufficient to detect improvements in many packings. Many of these approaches are adapted and modified from widely known procedures for packing congruent circles in a square or a circle \cite{csq_graham,cic_locatelli,csq_markot,csq_markot_2,csq_book}. The method we suggest is based on a hypothesis that an improved packing can be reached just by the omission of two or more spheres from the ccp and by performing a translation of spheres using the available space made after we remove the spheres. The initial positions of the sphere centers we denote by $S_{p,r}\subset G_{p}$, as a set of $g(p)-r$ points such that at least one sphere with a center $s_{i}\in S_{p,r}$ can be continuously translated inside a cube container without overlapping with other spheres. More precisely, \[ \bigcup_{i=1}^{g(p)-r}\left\{ q\in C_{p}:q\notin S_{p,r},(S_{p,r}\setminus\left\{ s_{i}\right\} )\cup\left\{ q\right\} \in\mathfrak{\mathcal{\mathcal{F}}}\right\} \neq\varnothing. \] We can see that the construction of $S_{p,r}$ is possible for $r\geq2$, and we experimentally determine solutions based on improvements reached for certain arrangements. To simplify the search procedure, instead of trying to figure out the best performing arrangements $S_{p,r}$ for each pair $(p,r)$, we find removal patterns $R_{r}=G_{p}\setminus S_{p,r}$ and use them to search for improvements for any $p$. Table 1 shows patterns in a simplified notation where $R'_{r}=\left\{ \frac{\sqrt{2}}{2}s:s\in R_{r}\right\} $. We notice that only $R_{2}$ is the most likely an optimal pattern for all $p>1$. \begin{table}[H] \begin{centering} \begin{tabular}{cl} $r$ & $R'_{r}$\tabularnewline \hline 2 & $\left\{ \langle0,1,1\rangle,\langle1,1,0\rangle\right\} $\tabularnewline 3 & $\left\{ \langle0,0,0\rangle,\langle1,0,1\rangle,\langle2,0,0\rangle\right\} $\tabularnewline 4 & $\left\{ \langle0,1,1\rangle,\langle1,1,0\rangle,\langle1,1,2\rangle,\langle2,1,1\rangle\right\} $\tabularnewline 5 & $\left\{ \langle0,1,1\rangle,\langle1,0,1\rangle,\langle1,1,0\rangle,\langle2,0,0\rangle,\langle2,1,1\rangle\right\} $\tabularnewline 6 & $\left\{ \langle0,0,0\rangle,\langle0,1,1\rangle,\langle1,0,1\rangle,\langle1,1,0\rangle,\langle2,0,0\rangle,\langle2,1,1\rangle\right\} $\tabularnewline \hline \end{tabular} \par\end{centering} \caption{Experimentally determined patterns $R'_{r}$} \end{table} Using the initial arrangement $S_{p,r}$ we try to perform the translation of each sphere using the limited set of translation vectors denoted by $T$. This algorithm can be described as follows: \newpage \begin{itemize} \item For a given initial set $S_{n}\gets S_{p,r}$ repeat until $D(S_{n})<D_{p}$: \begin{itemize} \item \vspace{-3mm} For each $s_{i}\in S_{n}$ do: \begin{itemize} \item \vspace{-1mm}Randomly choose $t\in T$, \item Let $v=\left\{ s_{i}+kt:k\in\left[-1,1\right]\right\}$, \item Let $v_{i}=\left\{ q\in v:(S_{n}\setminus\left\{ s_{i}\right\} )\cup\left\{ q\right\} \in\mathcal{F},D(S_{n}\cup\{q\})=D(S_{n})\right\}$, \item Find endpoints $a$ and $b$ of the largest line segment $\overline{ab}\subseteq v_{i}$ such that $s_{i}\in\overline{ab}$, \item New position of $s_{i}$ is given as $\ensuremath{s_{i}\gets(a+b)/2}$. \end{itemize} \end{itemize} \end{itemize} After we try to move all points from $S_{n}$, we say that we completed one iteration. Because of the very limited space to which we translate the spheres, and in order to minimize the number of required iterations, we usually set $T=\left\{ 0,1\right\} ^{3}\setminus\left\{ \langle0,0,0\rangle\right\} $. We also tested performances when $T$ takes different values such as $\left\{ -1,0,1\right\} ^{3}$, more or less reduced sets, but the improvements gained were always slightly worse. In practice, $D(S_{p,r})$ is slightly larger than $D_{p}$ while coordinates are given with finite precision. The described procedure stops when the first improvement is detected, but if we continue the search, we can improve the packing even more. It is also important to set the precision above the expected value of $D_{p}-D(S_{p,r})$, otherwise the improvement cannot be registered. Choosing the higher precision enables us to reach an improvement with less iterations, but only up to a certain level. We usually set the precision 1.5 times higher than the expected improvement. If the precision is too high, the search can be very slow, thus often we have to guess the range of possible improvements using a lower precision at first, and increase it if an improvement cannot be reached. This approach is different from the procedures used for sphere packing in the past, as in \cite{gensane,wenqi,locatelli_de,packomania}, while we focus only on tiny changes/improvements in the high density structure. We cannot consider this approach as a good general packing method for $r\geq p$. Its main weakness is that, because of the small available space where spheres can be moved, random perturbations are hard to implement, or at least we did not find any good method to do it. Still, this method allows us to find improved packings in less than one second for some well examined cases even when high precision is not required, as for example $n=29$, 59 or 60. The improvements attained are shown in the Tables 2 and 3. Table 2 lists values obtained for $I_{p}$ with $4\leq p\leq21$. Because of the slow computation times for $p\geq13$, we ran a search with approximately 5000 iterations when improvement gains started to slow down. Table 3 lists other improved packings for $3\leq r\leq6$, $r<p$ and $4\leq p\leq12$ using $R_{r}$ patterns described in Table 1. The results are listed as the best known values performed after a large number of iterations and multiple runs of the search procedure. \begin{table}[H] \centering{}\textsuperscript{}% \begin{tabular}{lll} $n$ & $p$ & \quad $I_{p}$\tabularnewline \toprule 30 & 4 & $7.34\cdot10^{-68}$\tabularnewline 61 & 5 & $7.18\cdot10^{-80}$\tabularnewline 106 & 6 & $2.26\cdot10^{-314}$\tabularnewline 170 & 7 & $9.09\cdot10^{-622}$\tabularnewline 254 & 8 & $3.74\cdot10^{-629}$\tabularnewline 363 & 9 & $7.51\cdot10^{-629}$\tabularnewline 498 & 10 & $5.00\cdot10^{-2584}$\tabularnewline 664 & 11 & $9.67\cdot10^{-2563}$\tabularnewline 862 & 12 & $1.76\cdot10^{-4988}$\tabularnewline 1097 & 13 & $1.70\cdot10^{-5020}$\tabularnewline 1370 & 14 & $2.01\cdot10^{-5044}$\tabularnewline 1686 & 15 & $1.29\cdot10^{-5076}$\tabularnewline 2046 & 16 & $1.78\cdot10^{-5116}$\tabularnewline 2455 & 17 & $3.30\cdot10^{-5047}$\tabularnewline 2914 & 18 & $1.47\cdot10^{-10118}$\tabularnewline 3428 & 19 & $4.20\cdot10^{-10121}$\tabularnewline 3998 & 20 & $1.16\cdot10^{-20344}$\tabularnewline 4629 & 21 & $2.46\cdot10^{-20582}$\tabularnewline \bottomrule \end{tabular}\caption{Improved values of $I_{p}$} \end{table} \vspace*{-\baselineskip} \begin{table}[H] \begin{centering} \begin{tabular}{llll|llll} $n$ & $p$ & $r$ & $d_{n}-d'_{p}$ & $n$ & $p$ & $r$ & $d_n - d'_p$ \\ \hline 29 & 4 & 3 & $2.23\cdot10^{-12}$ & & & & \\ \hline 59 & 5 & 4 & $1.95\cdot10^{-11}$ & & & & \\ 60 & 5 & 3 & $2.09\cdot10^{-20}$ & & & & \\ \hline 103 & 6 & 5 & $3.38\cdot10^{-14}$ & & & & \\ 104 & 6 & 4 & $9.98\cdot10^{-47}$ & & & & \\ 105 & 6 & 3 & $1.34\cdot10^{-76}$ & & & & \\ \hline 166 & 7 & 6 & $3.08\cdot10^{-21}$ & 494 & 10 & 6 & $4.77\cdot10^{-57}$\\ 167 & 7 & 5 & $7.72\cdot10^{-31}$ & 495 & 10 & 5 & $6.96\cdot10^{-199}$ \\ 168 & 7 & 4 & $1.59\cdot10^{-87}$ & 496 & 10 & 4 & $1.72\cdot10^{-310}$ \\ 169 & 7 & 3 & $8.49\cdot10^{-148}$ & 497 & 10 & 3 & $1.28\cdot10^{-605}$ \\ \hline 250 & 8 & 6 & $2.98\cdot10^{-28}$ & 660 & 11 & 6 & $2.83\cdot10^{-119}$\\ 251 & 8 & 5 & $1.99\cdot10^{-43}$ & 661 & 11 & 5 & $1.99\cdot10^{-173}$\\ 252 & 8 & 4 & $3.11\cdot10^{-102}$ & 662 & 11 & 4 & $1.40\cdot10^{-407}$\\ 253 & 8 & 3 & $6.02\cdot10^{-153}$ & 663 & 11 & 3 & $3.47\cdot10^{-615}$\\ \hline 359 & 9 & 6 & $3.84\cdot10^{-28}$ & 858 & 12 & 6 & $8.40\cdot10^{-248}$\\ 360 & 9 & 5 & $6.45\cdot10^{-44}$ & 859 & 12 & 5 & $2.68\cdot10^{-404}$\\ 361 & 9 & 4 & $2.05\cdot10^{-101}$ & 860 & 12 & 4 & $7.74\cdot10^{-745}$\\ 362 & 8 & 3 & $1.08\cdot10^{-152}$ & 861 & 12 & 3 & $2.91\cdot10^{-1212}$\\ \hline \end{tabular} \par\end{centering} \caption{Improvements reached using patterns $R_{r}$} \end{table} \bibliographystyle{plain}

https://arxiv.org/abs/1406.5295

Rows vs Columns for Linear Systems of Equations - Randomized Kaczmarz or Coordinate Descent?

This paper is about randomized iterative algorithms for solving a linear system of equations $X \beta = y$ in different settings. Recent interest in the topic was reignited when Strohmer and Vershynin (2009) proved the linear convergence rate of a Randomized Kaczmarz (RK) algorithm that works on the rows of $X$ (data points). Following that, Leventhal and Lewis (2010) proved the linear convergence of a Randomized Coordinate Descent (RCD) algorithm that works on the columns of $X$ (features). The aim of this paper is to simplify our understanding of these two algorithms, establish the direct relationships between them (though RK is often compared to Stochastic Gradient Descent), and examine the algorithmic commonalities or tradeoffs involved with working on rows or columns. We also discuss Kernel Ridge Regression and present a Kaczmarz-style algorithm that works on data points and having the advantage of solving the problem without ever storing or forming the Gram matrix, one of the recognized problems encountered when scaling kernelized methods.

\section{Introduction} Solving linear systems of equations is a classical topic and our interest in this problem is fairly limited. While we do compare two algorithms - Randomized Kaczmarz (RK) (see \citet{StrVer09}) and Randomized Coordinate Descent (RCD) (see \citet{LevLew10}) - to each other, we will not be presently concerned with comparing them to the host of other classical algorithms in the literature, like Conjugate Gradient and Gradient Descent methods. Our primary aim will be to understand the algorithmic similarities and differences involved in working with rows and columns (RK and RCD) of a given input. Assume we have a known $n \times p$ matrix $X$ (representing $n$ data points with $p$ features each), an unknown $p$-dimensional vector $\beta$ (regression coefficients), and a known $n$-dimensional vector $y$ (observations). The system of equations that one wants to solve is \begin{equation}\label{eq:syseq} X \beta = y \end{equation} When there exists at least one solution to the above system, we say that it is consistent. When there is a unique consistent solution, solving \eqref{eq:syseq} is a special case of the more general problem of minimizing the residual norm (makes sense when there are no consistent solutions) \begin{equation}\label{eq:LS} \min_{\beta \in \mathbf{R}^p} \tfrac1{2} \|y-X\beta\|^2 =: L(\beta) \end{equation} When there is a unique consistent solution, it is also a special case of the more general problem of finding the minimum norm consistent solution (makes sense when there are infinite solutions) \begin{equation}\label{eq:minnorm} \min_{\beta \in \mathbf{R}^p} \|\beta\|^2 \mbox{ \ s.t. \ } y = X\beta \end{equation} Sparse versions of the above problems, while interesting, and not in the scope of this work. We will later consider the Ridge Regression extension to Eq.\eqref{eq:LS}. We represent the $i$-th row ($i=1,...,n$) of $X$ by $X^i$, and the $j$-th column ($j=1,...,p$) by $X_j$. Similarly, the $i$-th observation is $y^i$, and the $j$-th regression coefficient is $\beta_j$. The reason for the linear regression setup using statistical notation of $n,p$ and $X,\beta,y$, is simply author comfort, and the literature sometimes uses $Ax=b$ instead, with an $m \times n$ matrix A. For the rest of this paper, we refer to a method as being Kaczmarz-like when its updates depend on rows (data points) of $X$, like $$ \beta_{t+1} := \beta_t + \delta_i X^i $$ where stepsize $\delta_i$ could also depend on $X^i$. and we refer to a method as being Coordinate Descent style when its updates are on coordinates of $\beta$ and depend on columns (features) of $X$, like $$ \beta_{t+1} := \beta_t + \delta_j e_j $$ where stepsize $\delta_j$ could depend on $X_j$. Randomized Kaczmarz (and its variations) have been likened to Stochastic Gradient Descent (see \citet{sgdkacz}). Indeed, even before having mentioned the form of $\delta_i$, the update already looks a lot like that of the Perceptron algorithm by \citet{perceptron}, a well known stochastic gradient descent algorithm. However, even though we will later describe some differences from RCD, we argue that RK-style methods are still much more like randomized coordinate descent than stochastic gradient descent algorithms - this is useful not only for interpretation but also for new derivations. We will bring out the intuitive similarity between RK and RCD by establishing striking parallels in proofs of convergence (these proofs are traditionally presented in a different manner), and exploit this in designing a simple RK algorithm for Kernel Ridge Regression. There has been a lot of interest and extensions on both these interesting algorithms, and perhaps connections between the two have been floating around in a subtle manner. One of our aims will be to make these connections explicit, intuitively clear, and enable the reader to build an understanding of ideas and proof techniques by the end. We first need to introduce the aforementioned algorithms before we summarize our contributions. In doing so, we will only refer to papers that are directly relevant to our work. \subsection{Randomized Kaczmarz (RK)} Taking $X, y$ as input and starting from an arbitrary $\beta_0$, it repeats the following in each iteration. First, pick a random row $r \in \{1...n\}$ with probability proportional to its Euclidean norm, i.e. $$ \mathbb{P}(r = i) = \frac{\|X^i\|^2}{\|X\|_F^2} $$ Then, project the current iterate onto that row, i.e. \begin{equation} \beta_{t+1} := \beta_t + \frac{(y^r - X^{rT}\beta_t)}{\|X^r\|^2} X^r \end{equation} Intuitively, this update can be seen as greedily satisfying the $r$th equation in the linear system, because it is easy to see that after the update, \begin{equation} X^{rT}\beta_{t+1} = y^r \end{equation} Alternatively, referring to Eq.\eqref{eq:LS}, since $ L(\beta) = \tfrac1{2} \|y-X\beta\|^2 = \tfrac1{2} \sum_{i=1}^n (y^i - X^{iT} \beta)^2, $ we can interpret this update as stochastic gradient descent (we pick a random data-point on which to update), where the stepsize is the inverse Lipschitz constant of the stochastic gradient $$ \nabla^2 \tfrac1{2} (y^i - X^{iT} \beta)^2 = \|X^i\|^2. $$ \citet{StrVer09} showed that the above algorithm has an expected linear convergence. We will formally discuss the convergence properties of this algorithm in future sections. \subsection{Randomized Coordinate Descent (RCD)} Takeing $X,y$ as input, starting from an arbitrary $\beta_0$, it repeats the following in each iteration. First, pick a random column $c \in \{1...p\}$ with probability proportional to its Euclidean norm, i.e. $$ \mathbb{P}(c = j) = \frac{\|X_j\|^2}{\|X\|_F^2} $$ We then minimize the objective $L(\beta) = \tfrac1{2} \|y-X\beta\|^2$ with respect to this coordinate to get \begin{equation} \beta_{t+1} := \beta_t + \frac{X_c^T(y-X\beta_t)}{\|X_c\|^2} e_c \end{equation} where $e_{c}$ is the $c$th coordinate axis. It can be seen as greedily minimizing the objective with respect to the $c$-th coordinate. Indeed, letting $X_{-c},\beta_{-c}$ represent $X$ without its $c$-th column and $\beta$ without its $c$-th coordinate, \begin{equation} \frac{\partial L}{\partial \beta_c} = -X_c^T(y-X\beta) = -X_c^T(y-X_{-c}\beta_{-c} - X_c\beta_c) \end{equation} Setting this equal to zero for the coordinatewise minimization, we get the aforementioned update for $\beta_c$. Alternately, since $[\nabla L(\beta)]_c = -X_c^T(y-X\beta)$, the above update can intuitively be seen as a univariate descent step where the stepsize is the inverse Lipschitz constant of the gradient along the $c$-th coordinate, since $$[\nabla^2 L(\beta)]_{c,c} = (X^TX)_{c,c} = \|X_c\|^2.$$ \citet{LevLew10} showed that this algorithm has an expected linear convergence. We will discuss the convergence properties of this algorithm in detail in future sections. \section{Main Results} We first examine the differences in behavior of the two algorithms in three distinct but related settings. This will bring out the opposite behaviors of the two similar algorithms. When the system of equations \eqref{eq:syseq} has a unique solution, we represent this by $\beta^*$. This happens when $n \geq p$, and the system is (luckily) consistent. Assuming that $X$ has full column rank, \begin{equation} \beta^* = (X^TX)^{-1}X^T y \end{equation} When \eqref{eq:syseq} does not have any consistent solution, we refer to the least-squares solution of Eq. \eqref{eq:LS} as $\beta_{LS}$. This could happen in the overconstrained case, when $n > p$. Again, assuming that $X$ has full column rank, we have \begin{equation} \beta_{LS} = (X^TX)^{-1}X^T y \end{equation} When \eqref{eq:syseq} has infinite solutions, we call the minimum norm solution to \eqref{eq:minnorm} as $\beta^*_{MN}$. This could happen in the underconstrained case, when $n < p$. Assuming that $X$ has full row rank, we have \begin{equation} \beta_{MN}^* = X^T (XX^T)^{-1} y \end{equation} In the above notation, the $*$ is used to denote the fact that it is consistent, i.e. it solves the system of equations, the $LS$ stands for Least Squares and $MN$ for minimum norm. We shall return to each of these three situations in that order in future sections of this paper. One of our main contributions is to achieve a unified understanding of the behaviour of RK and RCD in these different situations. The literature for RK deals with only the first two settings (see \citet{StrVer09}, \citet{Nee10}, \citet{ZouFre13}), but avoids the third. The literature for RCD typically focuses on more general setups than our specific quadratic least squares loss function $L(\beta)$ (see \citet{Nesterov12} or \citet{RicTak12}). However, for both the purposes of completeness, and for a more thorough understanding the relationship between RK and RCD, it turns out to be crucial to analyse all three settings (for equations \eqref{eq:syseq}-\eqref{eq:minnorm}). \begin{enumerate} \item When $\beta^*$ is a unique consistent solution, we present proofs of the linear convergence of both algorithms - the results are known from papers by \citet{StrVer09} and \citet{LevLew10} but are presented in a novel manner so that their relationship becomes clearer and direct comparison is easily possible. \item When $\beta_{LS}$ is the (inconsistent) least squares solution, we show why RCD iterates converge linearly to $\beta_{LS}$, but RK iterates do not - making RCD preferable. These facts are not hard to see, but we make it more intuitively and mathematically clear why this should be the case. \item When $\beta^*_{MN}$ is the minimum norm consistent solution, we explain why RK converges linearly to it, but RCD iterates do not (both so far undocumented observations) - making RK preferable. \end{enumerate} Together, the above three points complete the picture (with solid accompanying intuition) of the opposing behavior of RK and RCD. We then use the insights thus gained to develop a Kaczmarz style algorithm for Ridge Regression and its kernelized version. It is well known that the solution to \begin{equation} \min_{\beta \in \mathbb{R}^p} \|y-X\beta\|^2 + \lambda \|\beta\|^2 \end{equation} can be given in two equivalent forms (using the covariance and gram matrices) as \begin{equation} \beta_{RR} ~=~ (X^T X + \lambda I)^{-1}X^T y ~=~ X^T(XX^T + \lambda I)^{-1}y \end{equation} The presented algorithms completely avoid inverting, storing or even forming $XX^T$ and $X^TX$. Later, we will show that the following updates take only $O(p)$ computation per iteration like RK (starting with $\delta=0, \alpha=\mathbf{0}_n, \beta = \mathbf{0}_p, r = y$) and have expected linear convergence \begin{eqnarray} \delta_t &=& \frac{y^i - \beta_t^TX^i - \lambda \alpha^i_t}{\|X^i\|^2 + \lambda}\label{eq:RR1}\\ \alpha^i_{t+1} &=& \alpha^i_t + \delta_t\label{eq:RR2} \\ \beta_{t+1} &=& \beta_t + \delta_t X^i\label{eq:RR3} \end{eqnarray} where the $i$-th row is picked with probability proportional to $\|X^i\|^2 + \lambda$. If $\mathcal{H}_k$ is a Reproducing Kernel Hilbert Space (RKHS, see \citet{learningkernels} for an introduction) associated to positive definite kernel $k$ and feature map $\phi_x$, it is well known that the solution to the corresponding Kernel Ridge Regression (see \citet{krr}) problem is \begin{eqnarray} \label{eq:KRRf} f_{KRR} &=& \arg\min_{f \in \mathcal{H}_k} \sum\limits_{i=1}^n (y_i - f(x_i))^2 + \lambda \|f\|^2_{\mathcal{H}_k} \\ &=& \Phi^T (K+\lambda I)^{-1}y \end{eqnarray} where $\Phi = (\phi_{x_1},...,\phi_{x_n})^T$ and $K$ is the gram matrix with $K_{ij}=k(x_i,x_j)$. One of the main problems with kernel methods is as data size grows, the gram matrix becomes too large to store. This has motivated the study of approximation techniques for such kernel matrices, but we have an alternate suggestion. The aim of a Kaczmarz style algorithm would be to solve the problem by never forming $K$ as exmplified in updates for $\beta_{RR}$. For KRR, the update is \begin{eqnarray} \alpha^i_{t+1} &=& \frac{y - \sum_{j\neq i} K(x_i,x_j) \alpha_t^j}{K(x_i,x_i) + \lambda} \end{eqnarray} and costs $O(n)$ per iteration, and results in linear convergence as described later. Note that here RK for Kernel Ridge Regression costs $O(n)$ per iteration and RK for Ridge Regression cost $O(p)$ per iteration due to different parameterization. In the latter, we can keep track of $\beta_t$ as well as $\alpha_t$ easily, see Eq.\eqref{eq:RR2},\eqref{eq:RR3}, but for KRR, calculations can only be performed via evaluations of the kernel only ($\beta_t$ corresponds to a function and cannot be stored), and hence have a different cost. The aforementioned updates and their convergence can be easily derived after we develop a clear understanding of how RK and RCD methods relate to each other and jointly to positive semi-definite systems of equations. We shall see more of this in Sec.\ref{sec:RR}. \newpage \section{Overconstrained System, Consistent} To be clear, here we will assume that $n > p$, $X$ has full column rank, and that the system is consistent, so $y = X\beta^*$. First, let us write the updates used by both algorithms in a revealing fashion. If RK and RCD picked row $i$ and column $j$ at step $t+1$, and $e^i$ is $1$ in the $i$-th position and $0$ elsewhere, then the updates can be rewritten as below: \begin{align}\label{eq:RKupdate} &\mbox{(RK)}& \beta_{t+1} &:= \beta_t + \frac{e^{iT} r_t}{\|X^i\|^2} X^i&\\ &\mbox{(RCD)}& \beta_{t+1} &:= \beta_t + \frac{X_j^T r_t}{\|X_j\|^2}e_j & \end{align} where $r_t = y - X\beta_t = X\beta^*-X\beta_t$ is the residual vector. Then multiplying both equations by $X$ gives \begin{align} &\mbox{(RK)}&X\beta_{t+1} &:= X\beta_t + \frac{X^{iT}(\beta^*-\beta_t)}{\|X^i\|^2} X X^i&\\ &\mbox{(RCD)}& X\beta_{t+1} &:= X\beta_t + \frac{X_j^T X(\beta^*-\beta_t)}{\|X_j\|^2}X_j& \label{eq:RCDupdate} \end{align} We now come to an important difference, which is the key update equation for RK and RCD. Firstly, from the update Eq.(\ref{eq:RKupdate}) for RK, we have $\beta_{t+1}-\beta_t$ is parallel to $X^i$. Also, $\beta_{t+1}-\beta^*$ is orthogonal to $X^i$ (why? because $X^{iT}(\beta_{t+1}-\beta^*)=y^i - y^i = 0$). By Pythagoras, \begin{equation}\label{eq:RKrecursion} \|\beta_{t+1} - \beta^*\|^2 = \|\beta_t - \beta^*\|^2 - \|\beta_{t+1} - \beta_t\|^2 \end{equation} Note that from the update Eq.(\ref{eq:RCDupdate}), we have $X\beta_{t+1} - X\beta_t$ is parallel to $X_j$. Also, $X\beta_{t+1} - X\beta^*$ is orthogonal to $X_j$ (why? because $X_j^T(X\beta_{t+1} - X\beta^*) = X_j^T(X\beta_{t+1} - y) = 0$ by the optimality condition $\partial L/\partial \beta_{j} = 0$). By Pythagoras, \begin{equation}\label{eq:RCDrecursion} \|X\beta_{t+1} - X\beta^*\|^2 = \|X\beta_t - X\beta^*\|^2 - \|X\beta_{t+1} - X\beta_t\|^2 \end{equation} The rest of the proof follows by simply substituting for the last term in the above two equations, and is presented in the following table for easy comparison. Note $\Sigma=X^TX$ is the full-rank covariance matrix and we first take expectations with respect to the randomness at the $t+1$-st step, conditioning on all randomness up to the $t$-th step. We later iterate this expectation. \begin{table}[h] \begin{tabular}{|l|l|} \hline Randomized Kaczmarz $\mathbb{E}\|\beta_{t+1} - \beta^*\|^2$ & Randomized Coordinate Descent $\mathbb{E}\|X\beta_{t+1} - X\beta^*\|^2$ \\ \hline $= \|\beta_t - \beta^*\|^2 - \mathbb{E}\|\beta_{t+1} - \beta_t\|^2$ & $= \|X\beta_t - X\beta^*\|^2 - \mathbb{E}\|X\beta_{t+1} - X\beta_t\|^2$\\ $= \| \beta_{t} - \beta^*\|^2 - \sum_i \frac{\|X^i\|^2}{\|X\|_F^2} \frac{(X^i(\beta_t - \beta^*))^2}{(\|X^i\|^2)^2} \|X^i\|^2$ & $= \|X \beta_{t} - X\beta^*\|^2 - \sum_j \frac{\|X_j\|^2}{\|X\|_F^2} \frac{(X_j^T X(\beta_t - \beta^*))^2}{(\|X_j\|^2)^2}\|X_j\|^2$ \\ $= \|\beta_{t} - \beta^*\|^2 \left(1 - \frac1{\|X\|_F^2} \frac{\| X(\beta_t - \beta^*)\|^2}{\| \beta_{t} - \beta^*\|^2} \right)$ & $= \|X \beta_{t} - X\beta^*\|^2 \left(1 - \frac1{\|X\|_F^2} \frac{\|X^T X(\beta_t - \beta^*)\|^2}{\|X \beta_{t} - X\beta^*\|^2} \right)$ \\ $\leq \|\beta_t-\beta^*\|^2 (1 - \frac{\sigma_{\min}(\Sigma)}{Tr(\Sigma)})$ & $\leq \|X\beta_t-X\beta^*\|^2 (1 - \frac{\sigma_{\min}(\Sigma)}{Tr(\Sigma)})$ \\ \hline \end{tabular} \end{table} Here, $\sigma_{\min}(\Sigma)\| \beta_t - \beta^*\|^2 \leq \|X(\beta_t - \beta^*)\|^2 $ i.e. $\sigma_{\min}(\Sigma)$ is the smallest eigenvalue. It follows tha \begin{align} &\mbox{(RK)}& \mathbb{E}\|\beta_t - \beta^*\|^2 &\leq \left( 1 - \frac{\sigma_{\min}(\Sigma)}{Tr(\Sigma)} \right)^{t}\|\beta_0-\beta^*\|^2&\\ \label{eq:RCDlin} &\mbox{(RCD)}& \mathbb{E}\|\beta_t - \beta^*\|_\Sigma^2 &\leq \left( 1 - \frac{\sigma_{\min}(\Sigma)}{Tr(\Sigma)} \right)^{t}\|\beta_0-\beta^*\|_\Sigma^2 \end{align} Since $\Sigma$ is invertible when $n>p$ and $X$ has full column rank, the last equation also implies linear convergence of $\mathbb{E}\|\beta_t-\beta^*\|^2$. The final results do exist in \citet{StrVer09,LevLew10} but there is utility in seeing the two proofs in a form that differs from their original presentation, side by side. In this setting, both RK and RCD are essentially equivalent (without computational considerations). \newpage \section{Overconstrained System, Inconsistent} Here, we will assume that $n > p$, $X$ is full column rank, and the system is (expectedly) inconsistent, so $y = X\beta_{LS} + z$, where $z$ is such that $X^T z = 0$. It is easy to see this condition, because as mentioned earlier, \[ \beta_{LS} = (X^T X)^{-1}X^T y \] implying that $X^T X \beta_{LS} = X^T y$. Substituting $y=X\beta_{LS}+z$ gives that $X^Tz=0$. In this setting, RK is known to not converge to the least squares solution, as is easily verified experimentally. The tightest convergence upper bounds known are by \citet{Nee10} and \citet{ZouFre13} who show that \[ \mathbb{E}\|\beta_t - \beta_{LS}\|^2 \leq \left( 1 - \frac{\sigma_{\min}^+(\Sigma)}{Tr(\Sigma)} \right)^{t}\|\beta_0-\beta_{LS}\|^2 + \frac{\|w\|^2}{\sigma_{\min}^+(\Sigma)^2} \] If you tried to follow the previous proof, Eq.\eqref{eq:RKrecursion} does not go through - Pythagoras fails because $\beta_{t+1}-\beta_{LS}\not\perp X^i$, since $X^{iT}(\beta_{t+1}-\beta_{LS}) = y^i-X^{iT}\beta_{LS} \neq 0 $. Intuitively, the reason RK does not converge is that every update of RK (say of row $i$) is a projection onto the ``wrong'' hyperplane that has constant $y^i$ (where the ``right'' hyperplane would involve projecting onto a parallel hyperplane with constant $y^i-z^i$ where $z$ was defined above). An alternate intuition is that all RK updates are in the span of the rows, but $\beta_{LS}$ is not in the row span. These intuitive explanations are easily confirmed by experiments seen in \citet{Nee10,ZouFre13}. The same doesn't hold for RCD. Almost magically, in the previous proof, Pythagoras works because $$X_j^T(X\beta_{t+1} - X\beta_{LS}) = X_j^T(X\beta_{t+1} - y) + X_j^T(y-X\beta_{LS}) = 0$$ The first term is 0 by the optimality condition for $\beta_{t+1}$ i.e. $X_j^T(X\beta_{t+1} - y) = \partial L/\partial \beta_j = 0$. The second term is zero by the global optimality of $\beta_{LS}$ i.e. $X^T(y - X\beta_{LS}) = \nabla L = 0$. Also, $\Sigma$ is full rank as before. Hence, since RCD works in the space of fitted values $X\beta$ and not the iterates $\beta$. In summary, RK does not converge to the LS solution, but RCD does at the same linear rate. The Randomized Extended Kaczmarz by \citet{ZouFre13} is a modification of RK designed to converge by randomly projecting out $z$, but its discussion is beyond our scope. We were also alerted to a recent, independent Arxiv paper by \citet{frek}. \section{Underconstrained System, Infinite Solutions} Here, we will assume that $p > n$, $X$ is full row rank and the system is (expectedly) consistent with infinite solutions. As mentioned earlier, it is easy to show that \[ \beta^*_{MN} = X^T(XX^T)^{-1}y \] (which clearly satisfies $X\beta^*_{MN} = y$). Every other consistent solution can be expressed as $$ \beta^* = \beta^*_{MN} + z ~\mbox{~ where ~}~ Xz=0 $$ Clearly any such $\beta^*$ would also satisfy $X\beta^* = X\beta^*_{MN}=0$. Since $Xz=0$, $z \perp \beta^*_{MN}$ implying $\|\beta^*\|^2 = \|\beta^*_{MN}\|^2 + \|z\|^2$, showing that $\beta^*_{MN}$ is indeed the minimum norm solution as claimed. In this case, RK has good behaviour, and starting from $\beta_0=0$, it does converge linearly to $\beta^*_{MN}$. Intuitvely, $\beta^*_{MN} = X^T \alpha$ (for $\alpha = (XX^T)^{-1}y$) and hence is in the row span of $X$. Starting from $\beta_0=0$, RK only adds multiples of rows to its iterates, and hence will never have any component orthogonal to the row span of $X$ (i.e. will never add any $z$ such that $Xz=0$). There is exactly one solution with no component orthogonal to the row span of $X$, and that is $\beta^*_{MN}$, and hence RK converges linearly to the required point. It is important not to start from an arbitrary $\beta_0$ since the RK updates can never wipe out any component of $\beta_0$ that is perpendicular to the row span of $X$. Mathematically, the previous earlier proof works because Pythagoras goes through since it is a consistent system. However, $\Sigma$ is not full rank but note that since both $\beta^*_{MN}$ and $\beta_t$ are in the row span, $\beta_t - \beta^*_{MN}$ has no component orthogonal to $X$ (unless it equals zero and we're done). Hence $\sigma_{\min}^+(\Sigma)\| \beta_t - \beta^*\|^2 \leq \|X(\beta_t - \beta^*)\|^2$ does hold $\sigma_{\min}^+$ being the smallest positive eigenvalue of $\Sigma$. RCD unfortunately suffers the opposite fate. The iterates do not converge to $\beta_{LS}^*$, even though $X\beta_t$ does converge to $X\beta^*$. Mathematically, the convergence proof still carries forward as before till Eq.\eqref{eq:RCDlin}, but in the last step where $X^T X$ cannot be inverted because it is not full rank. Hence we get convergence of the residual to zero, without getting convergence of the iterates to the least squares solution. Unfortunately, when each update is cheaper for RK than RCD (due to matrix size), RCD is preferred for reasons of convergence and when it is cheaper for RCD than RK, RK is preferred. \section{Randomized Kaczmarz for Ridge Regression}\label{sec:RR} Both RK and RCD can be viewed in the following fashion. Suppose we have a positive definite matrix $A$, and we want to solve $Ax=b$. Instead of casting it as $\min_x \|Ax-b\|^2$, we can alternatively pose the different problem $\min_x \tfrac1{2} x^TAx - b^Tx$. Then one could use the update \[ x_{t+1} = x_t + \frac{b_i-A_i^Tx_t}{A_{ii}} e_i \] where the $b_i-A_i^T x_t$ is basically the $i$-th coordinate of the gradient, and $A_{ii}$ is the Lipschitz constant of the $i$-th coordinate of the gradient (related works include \citet{LevLew10}, \citet{Nesterov12},\citet{RicTak12},\citet{SidLee13}). In this light, the RK update can be seen as the randomized coordinate descent rule for the psd system $XX^T \alpha = y$ (substituting $\beta = X^T \alpha$) and treating $A = XX^T$ and $b=y$. Similarly, the RCD update can be seen as the randomized coordinate descent rule for the psd system $X^T X \beta = X^T y$ and treating $A = X^T X$ and $b=X^T y$. Using this connection, we propose the following update rule: \begin{eqnarray} \delta_t &=& \frac{y^i - \beta_t^TX^i - \lambda \alpha^i_t}{\|X^i\|^2 + \lambda} \label{eq:kacz1}\\ \alpha^i_{t+1} &=& \alpha^i_t + \delta_t \label{eq:kacz2}\\ \beta_{t+1} &=& \beta_t + \delta_t X^i \label{eq:kacz3} \end{eqnarray} where the $i$-th row was picked with probability proportional to $\|X^i\|^2 + \lambda$. If all rows are normalized, then this is still a uniform distribution. However, it is more typical to normalize the columns in statistics, and hence one pass over the data must be made to calculate row norms. We argue that this update rule exhibits linear convergence for $\min_\beta \|y - X\beta\|^2 + \lambda \|\beta\|^2$. Similarly, for Kernel Ridge Regression as mentioned in Eq.\eqref{eq:KRRf}, since one hopes to calculate $f_{KRR} = \Phi^T(K + \lambda I_n)^{-1} y$, the RK-style update can be derived from the randomized coordinate descent update rule for the psd system \[ (K + \lambda I) \alpha = y \] by setting $f_{KRR}=\Phi^T \alpha$. The update for $\alpha$ looks like (where $S_a(z) = \frac{z}{1+a}$) \begin{eqnarray}\label{eq:kerkacz1} \alpha^i_{t+1} &=& \frac{K(x_i,x_i)}{K(x_i,x_i)+\lambda}\alpha^i_t + \frac{y_i - \sum_j K(x_i,x_j) \alpha_t^j}{K(x_i,x_i) + \lambda}\\ &=& S_{\frac{\lambda}{K(x_i,x_i)}} \left( \alpha_t^i + \frac{r_i}{K(x_i,x_i)} \right)\label{eq:kerkacz3} \end{eqnarray} where row $i$ is picked proportional to $K(x_i,x_i)+\lambda$ (uniform for translation invariant kernels). Let us contrast this with the randomized coordinate descent update rule for the loss function $\min_x \tfrac1{2} \beta^T (X^TX + \lambda I_p) \beta - y^T X\beta$ i.e. the system $(X^T X + \lambda I_p)\beta = X^T y$. \begin{eqnarray} \beta^i_{t+1} &=&\beta^i_t + \frac{X_i^T y - X_i^T X \beta -\lambda \beta_i}{\|X_i\|^2 + \lambda}\label{eq:rcdridge1}\\ &=& \frac{\|X_i\|^2}{\|X_i\|^2 + \lambda}\beta^i_t + \frac{X_i^T r_t}{\|X_i\|^2 + \lambda} \\ &=& S_{\frac{\lambda}{\|X_i\|^2}} \left(\beta^i_t + \frac{X_i^T r_t}{\|X_i\|^2} \right)\label{eq:rcdridge3} \end{eqnarray} \subsection{Computation and Convergence} The RCD updates in \eqref{eq:rcdridge1}-\eqref{eq:rcdridge3} take $O(n)$ time, since each column (feature) is of size $n$. In contrast, the proposed RK updates in \eqref{eq:kacz1}-\eqref{eq:kacz3} takes $O(p)$ time since that is the length of a data point. Lastly the RK updates in \eqref{eq:kerkacz1}-\eqref{eq:kerkacz3} take $O(n)$ time (to update $r$) not counting time for kernel evaluations. The difference between the two RK updates for Ridge Regression and Kernel Ridge Regression is that for KRR, we cannot maintain $\alpha$ and $\beta$ since the $\beta$ is a function in the RKHS. This different parameterization makes the updates to $\alpha$ cost $O(n)$ instead of $O(p)$. While the RK and RCD algorithms are similar and related, one should not be tempted into thinking their convergence rates are the same. Indeed, with no normalization assumption, using a similar style proof as presented earlier, one can show that the convergence rate of the RK for Kernel Ridge Regression is \begin{eqnarray} \mathbb{E}\|\alpha_t - \alpha^*\|_{K+\lambda I_n}^2 &\leq& \left( 1 - \frac{\sigma_{\min}(K + \lambda I_n)}{Tr(K+\lambda I_n)}\right)^{t} \|\alpha_0 - \alpha^*\|_{K+\lambda I_n}^2\\ &=& \begin{cases} \left( 1 - \frac{\lambda}{\sum_i \sigma_i^2 + n\lambda}\right)^{t} \|\alpha_0 - \alpha^*\|_{K+\lambda I_n}^2 ~\mbox{~if $n>p$ }\\ \left( 1 - \frac{\sigma_1^2 + \lambda}{\sum_i \sigma_i^2 + n\lambda}\right)^{t} \|\alpha_0 - \alpha^*\|_{K+\lambda I_n}^2 ~\mbox{~if $p>n$} \end{cases} \end{eqnarray} and the rate of convergence of the RCD for Ridge Regression is subtly different: \begin{eqnarray} \mathbb{E}\|\beta_t - \beta^*\|_{\Sigma+\lambda I_p}^2 &\leq& \left( 1 - \frac{\sigma_{\min}(\Sigma + \lambda I_p)}{Tr(\Sigma + \lambda I_p)}\right)^{t} \|\beta_0 - \beta^*\|_{\Sigma+\lambda I_p}^2 \\ &=& \begin{cases} \left( 1 - \frac{\sigma_1^2 + \lambda}{\sum_i \sigma_i^2 + p\lambda}\right)^{t} \|\beta_0 - \beta^*\|_{\Sigma+\lambda I_p}^2 ~\mbox{~if $n>p$} \\ \left( 1 - \frac{ \lambda}{\sum_i \sigma_i^2 + p\lambda}\right)^{t} \|\beta_0 - \beta^*\|_{\Sigma+\lambda I_p}^2 ~\mbox{~if $p>n$} \end{cases} \end{eqnarray} Kernel Ridge Regression is used as a subroutine in many kernelized machine learning problems (for example, see section 4.3 of \citet{FukSonGre14} for a recent application to a nonparametric state space models). One of the major issues involved with scaling kernel methods is the formation of the gram matrix, which could be prohibitively large to form, store, invert, etc. Our RK-style algorithm gets around this issue completely by never forming the kernel matrix, just as RCD avoids forming $\Sigma$, making it a great choice for scaling kernel methods. \section{Conclusion} In this paper, we studied the close connections between the RK and RCD algorithms that have both received a lot of recent attention in the literature, and which we show are in some sense instances of each other when appropriately viewed. While RK is often viewed as a stochastic gradient algorithm, we saw that its ties to RCD are much stronger and that it is easier to understand its convergence through the RCD perspective than the SGD viewpoint. We first analysed their opposite behavior with linear systems. If the system was consistent and unique then we showed that both algorithms approached the solution at (approximately) the same linear rate, with extremely similar proofs presented for direct comparison. However, if the system was consistent with infinite solutions then we saw that RK converged to the minimum norm solution but RCD didn't, making RK preferable. In contrast, if the system was inconsistent, RCD converged to the least squares solution but RK did not, making RCD the preferred choice. Unfortunately, in both cases, the preferred choices have costlier updates. We then exploited the connection between RK and RCD to design an RK-style algorithm for Kernel Ridge Regression (KRR) which avoided explicitly forming and inverting the potentially large kernel matrix. We anticipate this and other randomization techniques will help the scalability of kernel methods. \subsection*{Acknowlegements} The authors would like to thank Deanna Needell and Anna Ma for some discussions, Peter Richtarik for pointing out \citet{frek}, and Ryan Tibshirani, for memorable class notes on coordinate descent and for corrections on an earlier version of this manuscript \bibliographystyle{agsm}

https://arxiv.org/abs/1311.4920

Counting elliptic curves with prescribed torsion

Mazur's theorem states that there are exactly 15 possibilities for the torsion subgroup of an elliptic curve over the rational numbers. We determine how often each of these groups actually occurs. Precisely, if $G$ is one of these 15 groups, we show that the number of elliptic curves up to height $X$ whose torsion subgroup is isomorphic to $G$ is on the order of $X^{1/d}$, for some number $d=d(G)$ which we compute.

\section{Introduction} The arithmetic of elliptic curves has long been a central area of study in number theory. One of the first general results, due to Mordell \cite{mordell} (see also \cite[Ch.~VIII]{silverman}), states that the group of rational points of an elliptic curve is finitely generated. In other words, if $E/\mathbf{Q}$ is an elliptic curve then we can write $E(\mathbf{Q})=\mathbf{Z}^r \times E(\mathbf{Q})_{\mathrm{tors}}$, where $r \ge 0$ is an integer (the \emph{rank} of $E$), and $E(\mathbf{Q})_{\mathrm{tors}}$ is a finite group (the \emph{torsion subgroup} of $E$). Given this result, one would like to understand what the possibilities are for the rank and torsion subgroup. Little is known regarding the rank. On the other hand, work of several mathematicians (most notably Mazur \cite{mazur}) established that there are only 15 possibilities for the torsion subgroup, namely: \begin{equation} \label{mazur} \begin{aligned} \mathbf{Z}/N\mathbf{Z} & \quad \textrm{with $1 \le N \le 10$ or $N=12$} \\ \mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/N\mathbf{Z} & \quad \textrm{with $N=2,4,6,8$.} \end{aligned} \end{equation} With this classification in hand, it is natural to ask a more refined question: how often does each of these groups occur? Our purpose here is to provide an answer. \subsection{The main result} To make the question precise, we formulate it as a counting problem. Every elliptic curve $E$ over $\mathbf{Q}$ admits a unique equation of the form \begin{displaymath} y^2=x^3+Ax+B \end{displaymath} where $A$ and $B$ are integers and $\gcd(A^3,B^2)$ is not divisible by any twelfth powers. We call such an equation \emph{minimal}, and define the (\emph{na\"{i}ve}) \emph{height} of $E$ to be $\max(|A|^3, |B|^2)$. Clearly, there are only finitely many elliptic curves (up to isomorphism) of height $<X$, for any real number $X$. We can therefore introduce a meaningful counting function: if $G$ is one of the groups in \eqref{mazur}, we let $N_G(X)$ be the number of (isomorphism classes of) elliptic curves $E/\mathbf{Q}$ of height at most $X$ for which $E(\mathbf{Q})_{\mathrm{tors}}$ is isomorphic to $G$. The following theorem is our main result. \begin{theorem} \label{thm:main} For any group $G$ in \eqref{mazur}, the limit \begin{displaymath} \frac{1}{d(G)}=\lim_{X \to \infty} \frac{\log{N_G(X)}}{\log{X}} \end{displaymath} exists. The value of $d(G)$ is as indicated in Table~\ref{f:main}. \end{theorem} \begin{table}[!h] \caption{The values of $d(G)$.} {\tabulinesep=1.2mm \begin{tabu}{|c|c||c|c||c|c|} \hline $G$ & $d$ & $G$ & $d$ & $G$ & $d$ \\ \hline \hline 0 & 6/5 & $\mathbf{Z}/6\mathbf{Z}$ & 6 & $\mathbf{Z}/12\mathbf{Z}$ & 24 \\ \hline $\mathbf{Z}/2\mathbf{Z}$ & 2 & $\mathbf{Z}/7\mathbf{Z}$ & 12 & $\mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/2\mathbf{Z}$ & 3 \\ \hline $\mathbf{Z}/3\mathbf{Z}$ & 3 & $\mathbf{Z}/8\mathbf{Z}$ & 12 & $\mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/4\mathbf{Z}$ & 6 \\ \hline $\mathbf{Z}/4\mathbf{Z}$ & 4 & $\mathbf{Z}/9\mathbf{Z}$ & 18 & $\mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/6\mathbf{Z}$ & 12 \\ \hline $\mathbf{Z}/5\mathbf{Z}$ & 6 & $\mathbf{Z}/10\mathbf{Z}$ & 18 & $\mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/8\mathbf{Z}$ & 24 \\ \hline \end{tabu}} \label{f:main} \end{table} Since $d(0)<d(G)$ for all non-trivial $G$, we recover the following well-known result due to Duke \cite{duke}. \begin{corollary} Almost all elliptic curves over $\mathbf{Q}$ have trivial torsion. \end{corollary} \subsection{Refined results} \label{refined} We actually prove a stronger result than Theorem~\ref{thm:main}: given $G$ as in \eqref{mazur}, there exist positive constants $K_1$ and $K_2$ such that \begin{displaymath} K_1 X^{1/d(G)} \le N_G(X) \le K_2 X^{1/d(G)} \end{displaymath} holds for all $X \ge 1$. This suggests that the limit \begin{displaymath} c(G)=\lim_{X \to \infty} \frac{N_G(X)}{X^{1/d(G)}} \end{displaymath} might exist. We prove this is the case for $\#G\leq3$. \begin{theorem} \label{thm:asymptotics} Write $c_N$ for $c(\mathbf{Z}/N\mathbf{Z})$. Then the limits defining $c_1$, $c_2$, and $c_3$ exist, and \begin{align*} c_1 &= \frac{4}{\zeta(10)} \approx 3.9960, \\ c_2 &= \frac{1}{\zeta(6)} \left( 2\log(\alpha_-/\alpha_+)+\tfrac{4}{3} (\alpha_++\alpha_-) \right) \approx 3.1969, \\ c_3 &= \frac{1}{\zeta(4)} \left( 2I_+-2I_-+\frac{1}{3}\log\left(\displaystyle\frac{\beta_0\beta_1\beta_5}{\beta_2\beta_3\beta_4}\right)+\frac{9}{4}\left(\beta_0^4+\beta_2^4+\beta_4^4-\beta_1^4-\beta_3^4-\beta_5^4\right) \right) \approx 1.5221, \end{align*} where the $\alpha$'s and $\beta$'s are algebraic numbers and the $I$'s are hyperelliptic integrals (see \S\ref{sec:asymptotics}). \end{theorem} The case $N=1$ is straightforward; see \cite[Lemma~4.3]{brumer}. The case $N=2$ was previously carried out in \cite[\S2]{grant}. The case $N=3$ does not seem to occur in the literature. We, in fact, obtain power-saving error terms (see Theorem~\ref{lem:main_asymp_lemma} for this more precise version). \subsection{Interpretation of $d(G)$} \label{ss:interp} If $f \colon \mathbf{P}^1 \to \mathbf{P}^1$ is a degree $d$ map, then the number of points in $f(\mathbf{P}^1(\mathbf{Q}))$ up to height $X$ is approximately $X^{2/d}$. This suggests that $d(G)$ should be related to the degree of the map $f \colon \mathcal{X}(G) \to \mathcal{X}(1)$, where here $\mathcal{X}(-)$ denotes the appropriate moduli stack. When $\mathcal{X}(G)$ is a scheme, Table~\ref{f:main} shows that \begin{displaymath} d(G)=\tfrac{1}{4} \deg(f)=\tfrac{1}{2} \deg(|f|), \end{displaymath} where $|f|$ denotes the map of coarse spaces. More generally, as long as $G \ne 0$, we have \begin{displaymath} d(G) = \tfrac{1}{2} \deg(|f|)+\tfrac{1}{2} \nu_2 + \nu_3 + \nu_\infty^\prime, \end{displaymath} where $\nu_2$ and $\nu_3$ are the number of elliptic points of orders 2 and 3 on $\mathcal{X}(G)$ and $\nu_\infty^\prime$ is the number of irregular cusps (in the sense of \cite[\S\S 1.2 \& 2.1]{shimura}). We do not have a conceptual explanation of this formula, so it could simply be a numerical coincidence. \begin{remark} If the moduli stacks $\mathcal{X}(G)$ and $\mathcal{X}(1)$ were actually schemes then Theorem~\ref{thm:main} would follow immediately from the aforementioned fact about maps of $\mathbf{P}^1$, and $d(G)$ would be half the degree of $f$. It would be desirable to have a general result for counting points in the image of a map of stacks, from which Theorem~\ref{thm:main} could be deduced; see \S \ref{s:future} for a more precise discussion. \end{remark} \subsection{Overview of the proof} We now go over the main ideas in the proof of Theorem~\ref{thm:main} and make some preliminary reductions. \subsubsection{The function $N'_G(X)$} Let $N'_G(X)$ be the number of (isomorphism classes of) elliptic curves $E$ of height at most $X$ such that $E(\mathbf{Q})$ contains a subgroup isomorphic to $G$. We prove the following theorem. \begin{theorem} \label{thm:main2} For any group $G$ in \eqref{mazur}, there exist positive constants $K_1$ and $K_2$ such that $K_1 X^{1/d(G)} \le N'_G(X) \le K_2 X^{1/d(G)}$ for all $X \ge 1$. \end{theorem} Theorem~\ref{thm:main} follows easily from this. Indeed, we have obvious bounds \begin{displaymath} N'_G(X)-\sum_{G \subsetneqq H} N'_H(X) \le N_G(X) \le N'_G(X). \end{displaymath} As $d(G)<d(H)$, Theorem~\ref{thm:main2} gives $N'_H(X)/N'_G(X) \to 0$ as $X \to \infty$. Thus $N_G(X)/N'_G(X) \to 1$, and so Theorem~\ref{thm:main} (and the stronger form stated in \S \ref{refined}) follows from Theorem~\ref{thm:main2}. \subsubsection{Proof of Theorem~\ref{thm:main2} when $2G \ne 0$} \label{sss:oddg} Suppose that $G$ is one of the 12 groups in \eqref{mazur} for which $2G \ne 0$. Then there is a \emph{universal elliptic curve} $\mathcal{E}$ over an open subset of $\mathbf{A}^1$ equipped with a subgroup isomorphic to $G$. (This is not exactly true for $G=\mathbf{Z}/3\mathbf{Z}$, see \S \ref{ss:z3} for details.) The universal property implies that an arbitrary curve $E/\mathbf{Q}$ admits a copy of $G$ in its rational points if and only if $E$ is isomorphic to $\mathcal{E}_t$ for some $t \in \mathbf{Q}$. We can describe $\mathcal{E}$ by an equation of the form \begin{displaymath} y^2=x^3+f(t)x+g(t) \end{displaymath} where $f$ and $g$ are polynomials. In \S \ref{s:general}, we prove the following general theorem for counting elliptic curves that appear in such families. \begin{theorem} \label{general} Let $f,g \in \mathbf{Q}[t]$ be non-zero coprime polynomials of degrees $r$ and $s$, with at least one of $r$ or $s$ positive, and write \begin{displaymath} \max \left( \frac{r}{4}, \frac{s}{6} \right)=\frac{n}{m}, \end{displaymath} with $n$ and $m$ coprime. Assume $n=1$ or $m=1$. Let $\mathcal{E}$ be the family of elliptic curves defined by \begin{displaymath} y^2=x^3+f(t)x+g(t). \end{displaymath} Let $N(X)$ be the number of (isomorphism classes of) elliptic curves $E/\mathbf{Q}$ of height at most $X$ for which $E \cong \mathcal{E}_t$ for some $t \in \mathbf{Q}$. Then there exist positive constants $K_1$ and $K_2$ such that \begin{displaymath} K_1 X^{(m+1)/12n} \le N(X) \le K_2 X^{(m+1)/12n} \end{displaymath} for all $X \ge 1$. \end{theorem} In each case of interest, the hypotheses of the theorem are satisfied. Table~\ref{f:params} lists the invariants for the various groups (see \S \ref{explan} for proofs). This establishes Theorem~\ref{thm:main2} in the cases $2G \ne 0$. \begin{table} \caption{Data for the universal elliptic curve with $G$-structure.} {\small\tabulinesep=1.2mm \begin{tabu}{|c|c|c|c|c|c|} \hline $G$ & $r$ & $s$ & $n$ & $m$ & $12n/(m+1)$ \\ \hline \hline 3 & 1 & 2 & 1 & 3 & 3 \\ \hline 4 & 2 & 3 & 1 & 2 & 4 \\ \hline 5 & 4 & 6 & 1 & 1 & 6 \\ \hline 6 & 4 & 6 & 1 & 1 & 6 \\ \hline 7 & 8 & 12 & 2 & 1 & 12 \\ \hline 8 & 8 & 12 & 2 & 1 & 12 \\ \hline 9 & 12 & 18 & 3 & 1 & 18 \\ \hline 10 & 12 & 18 & 3 & 1 & 18 \\ \hline 12 & 16 & 24 & 4 & 1 & 24 \\ \hline $(2,4)$ & 4 & 6 & 1 & 1 & 6 \\ \hline $(2,6)$ & 8 & 12 & 2 & 1 & 12 \\ \hline $(2,8)$ & 16 & 24 & 4 & 1 & 24\\ \hline \end{tabu}} \label{f:params} \end{table} \subsubsection{Proof of Theorem~\ref{thm:main2} when $2G=0$} If $2G=0$ then there is not a universal curve (over a scheme), and so the above method does not directly apply. When $G$ is trivial or $\mathbf{Z}/2\mathbf{Z}$, the situation is simple enough to understand directly (see \S \ref{sec:asymptotics}). When $G=\mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/2\mathbf{Z}$, we use a variant of Theorem~\ref{general} to understand $N'_G$ (see \S \ref{sec:Zmod2timesZmod2}). In this way, the remaining cases of Theorem~\ref{thm:main2} are established. \subsubsection{Proof of Theorem~\ref{general}} \label{sss:general} We now comment on the proof of Theorem~\ref{general}, which is the heart of the paper. Recall that if $E_i$ (for $i=1,2$) are elliptic curves over $\mathbf{Q}$ defined by equations \begin{displaymath} y^2=x^3+a_ix+b_i, \end{displaymath} then $E_1$ and $E_2$ are isomorphic if and only if there exists a rational number $u \in \mathbf{Q}$ for which $a_1=u^4 a_2$ and $b_1=u^6 b_2$ \cite[Ch.~III \S 1]{silverman}. We therefore see that an elliptic curve $E/\mathbf{Q}$ given by an equation \begin{equation} \label{eq3} y^2=x^3+Ax+B \end{equation} belongs to a family $\mathcal{E}$ as above if and only if there exist $u, t \in \mathbf{Q}$ such that $A=u^4f(t)$ and $B=u^6g(t)$. It follows that $N(X)$ (as in Theorem~\ref{general}) counts the number of pairs $(A,B) \in \mathbf{Z}^2$ satisfying the following conditions: \begin{itemize} \item $4A^3+27B^2 \ne 0$. \item $\gcd(A^3, B^2)$ is not divisible by any 12th power. \item $|A|<X^{1/3}$ and $|B|<X^{1/2}$. \item There exist $u,t \in \mathbf{Q}$ for which $A=u^4f(t)$ and $B=u^6g(t)$. \end{itemize} The first condition ensures that \eqref{eq3} defines an elliptic curve; the second that the equation is minimal; the third that it has height $<X$; and the fourth that it belongs to $\mathcal{E}$. This reduces the analysis of $N(X)$ to an elementary number theory problem, which we solve directly. \subsection{Future directions} \label{s:future} There are several generalizations of our results that would be of interest. The most immediate, perhaps, is to extend our results to rationally defined subgroups. Precisely, let $G$ be a product of two finite cyclic groups, and let $N^0_G(X)$ be the number of elliptic curves $E/\mathbf{Q}$ up to height $X$ admitting a subgroup defined over $\mathbf{Q}$ which (over $\mathbf{C}$) is isomorphic to $G$. One would then like to compute the limit of $\log{N^0_G(X)}/\log{X}$. Again, there are only finitely many $G$ for which one gets a non-zero answer, namely, those for which the moduli space has genus 0. It would also be interesting to generalize Theorem~\ref{general} to allow $f$ and $g$ to have a common factor and to remove the restriction on $n$ and $m$. How does this affect the count? Perhaps the ultimate generalization in this direction is the following. Let $\mathcal{X}$ and $\mathcal{Y}$ be proper smooth Deligne--Mumford stacks over $\mathbf{Q}$ with coarse space $\mathbf{P}^1$, and let $f \colon \mathcal{Y} \to \mathcal{X}$ be a map. Suppose that there is a good notion of height $h_{\mathcal{X}}$ on the set $|\mathcal{X}(\mathbf{Q})|$, where $|\cdot|$ denotes isomorphism classes. Then one would like a formula for \begin{displaymath} \lim_{T \to \infty} \frac{\# \{ x \in f(|\mathcal{Y}(\mathbf{Q})|) \mid h_{\mathcal{X}}(x) \le T \}}{\log{T}} \end{displaymath} in terms of invariants of $\mathcal{X}$, $\mathcal{Y}$, and $f$ (in the style of \S \ref{ss:interp}). More generally, one may ask these questions over general global fields. What kind of dependence is there on the base field? In addition to these generalizations, it would be interesting to complete the results of \S \ref{refined}, and compute the value of $c(G)$ for other $G$'s. \subsection{Notation} If $f$ and $g$ are two functions of a real variable $X$, we write $f \lesssim g$ to mean ``there exists a positive constant $c$ such that $f(X) \le c g(X)$ holds for all $X \ge 1$.'' For a prime number $p$, we let $\val_p$ be the usual $p$-adic valuation on $\mathbf{Q}$ (so $\val_p(p)=1$). For a place $p \le \infty$ of $\mathbf{Q}$, we let $\vert \cdot \vert_p$ be the usual absolute value on $\mathbf{Q}$ (so $\vert p \vert_p=p^{-1}$ if $p<\infty$). We often write $\vert \cdot \vert$ in place of $\vert \cdot \vert_{\infty}$. We write $\lfloor x \rfloor$ and $\lceil x \rceil$ for the floor and ceiling of $x \in \mathbf{Q}$. \subsection*{Acknowledgements} We would like to thank Jordan Ellenberg, Wei Ho, and Melanie Matchett Wood for useful conversations, and Bjorn Poonen for pointing out an innaccuracy in an earlier version. We would also like to acknowledge the Sagemath Cloud and Sage \cite{sage} in which we carried out several helpful computations. The first author was supported by NSA Young Investigator Grant \#H98230-13-1-0223 and NSF RTG Grant ``Number Theory and Algebraic Geometry at the University of Wisconsin''. The second author was supported by NSF Grant DMS-1303082. \section{Counting elliptic curves in families} \label{s:general} The purpose of this section is to prove Theorem~\ref{general}. As explained in \S \ref{sss:general}, it is equivalent to the following proposition. \begin{proposition} \label{bd} Let $f,g \in \mathbf{Q}[t]$ be coprime polynomials of degrees $r$ and $s$. Assume at least one of $r$ or $s$ is positive. Write \begin{displaymath} \max \left( \frac{r}{4}, \frac{s}{6} \right)=\frac{n}{m} \end{displaymath} with $n$ and $m$ coprime. Assume $n=1$ or $m=1$. Let $S(X)$ be the set of pairs $(A,B) \in \mathbf{Z}^2$ satisfying the following conditions: \begin{itemize} \item $4A^3+27B^2 \ne 0$. \item $\gcd(A^3,B^2)$ is not divisible by any 12th power. \item $|A|<X^{1/3}$ and $|B|<X^{1/2}$. \item There exist $u,t \in \mathbf{Q}$ such that $A=u^4 f(t)$ and $B=u^6 g(t)$. \end{itemize} Then $X^{(m+1)/12n} \lesssim \# S(X) \lesssim X^{(m+1)/12n}$. \end{proposition} \subsection{The upper bound} We begin by establishing the upper bound. Let $S_1(X)$ be the set of pairs $(u,t) \in \mathbf{Q}^2$ such that $(A,B)=(u^4f(t), u^6g(t))$ belongs to $S(X)$. \begin{lemma} \label{bd1} For each place $p \le \infty$ of $\mathbf{Q}$ there exists a constant $c_p>0$ such that for all $t \in \mathbf{Q}$ \begin{displaymath} \max(|f(t)|_p, |g(t)|_p) \ge c_p. \end{displaymath} If $p$ is a sufficiently large finite prime, one can take $c_p=1$. \end{lemma} \begin{proof} Let $\ol{\mathbf{Q}}$ be a fixed algebraic closure of $\mathbf{Q}$, and extend $\vert \cdot \vert_p$ to $\ol{\mathbf{Q}}$ in any way. Let $\{\alpha_i\}$ be the roots of $f$ and $\{\beta_j\}$ the roots of $g$ in $\ol{\mathbf{Q}}$. Note that $\alpha_i \ne \beta_j$ for all $i$ and $j$, since $f$ and $g$ are coprime. Let $\delta=\min(|\alpha_i-\beta_j|_p)$. Let $\epsilon>0$ be such that $|f(t)|_p<\epsilon$ implies $|t-\alpha_i|<\delta/2$ for some $i$, and similarly for $g$. If $|f(t)|_p<\epsilon$ and $|g(t)|_p<\epsilon$ then $|t-\alpha_i|_p<\delta/2$ for some $i$ and $|t-\beta_j|_p<\delta/2$ for some $j$, which implies $|\alpha_i-\beta_j|_p<\delta$, a contradiction. We must therefore have $|f(t)|_p \ge \epsilon$ or $|g(t)|_p \ge \epsilon$ for all $t$, and so can take $c_p=\epsilon$. Now let $p$ be a prime large enough so that: (1) the coefficients of $f$ and $g$ are $p$-integral; (2) the leading coefficients of $f$ and $g$ are $p$-units; and (3) $\alpha_i-\beta_j$ are $p$-units, for all $i$ and $j$. If $|f(t)|_p<1$ then $|t-\alpha_i|<1$ for some $i$. Similarly for $g$. Thus if $|f(t)|_p<1$ and $|g(t)|_p<1$ then $|t-\alpha_i|<1$ and $|t-\beta_j|<1$, which implies $|\alpha_i-\beta_j|<1$, a contradicition. We conclude that we can take $c_p=1$ for such $p$. \end{proof} \begin{lemma} \label{bd2} For each prime number $p$ there exists a constant $C_p$ with the following property. Suppose $(u,t) \in S_1(X)$. Then \begin{displaymath} \val_p(u)=\epsilon+\begin{cases} \lceil -\tfrac{n}{m} \val_p(t) \rceil & \textrm{if $\val_p(t)<0$} \\ 0 & \textrm{if $\val_p(t) \ge 0$,} \end{cases} \end{displaymath} where $|\epsilon| \le C_p$. Furthermore, one can take $C_p=0$ for $p \gg 0$. \end{lemma} \begin{proof} Suppose $(u,t) \in S_1(X)$, and let $p$ be a prime. Since $A$ and $B$ are integral, we have $4\val_p(u)+\val_p(f(t)) \ge 0$ and $6\val_p(u)+\val_p(g(t)) \ge 0$. Furthermore, $\val_p(u)$ must be minimal subject to these inequalities, or else $p^{12}$ would divide $\gcd(A^3, B^2)$. We can thus write \begin{displaymath} \val_p(u)=\max(\lceil -\tfrac{1}{4} \val_p(f(t)) \rceil, \lceil -\tfrac{1}{6} \val_p(g(t)) \rceil). \end{displaymath} Equivalently, \begin{equation} \label{eq1} -\val_p(u)=\min(\lfloor \tfrac{1}{4} \val_p(f(t)) \rfloor, \lfloor \tfrac{1}{6} \val_p(g(t)) \rfloor). \end{equation} Suppose $\val_p(t)<0$. Let $K_1$ be a constant such that $|\val_p(f(t))-r\val_p(t)|<K_1$ and $|\val_p(g(t))-s\val_p(t)|<K_1$ for all such $t$. (This exists since the functions $\val_p(f(t)/t^r)$ and $\val_p(g(t)/t^s)$ are bounded on $\mathbf{Q}_p \setminus \mathbf{Z}_p$.) Then for $\val_p(t)<0$ we have \begin{displaymath} \val_p(u)=\epsilon+\max(\lceil -\tfrac{r}{4} \val_p(t) \rceil, \lceil -\tfrac{s}{6} \val_p(t) \rceil) \end{displaymath} where $|\epsilon|<K_2$, for some $K_2$. (One can take $K_2=1+\tfrac{n}{m} K_1$.) Since $\val_p(t)<0$, we have \begin{displaymath} \max(\lceil -\tfrac{r}{4} \val_p(t) \rceil, \lceil -\tfrac{s}{6} \val_p(t) \rceil) =\lceil -\tfrac{n}{m} \val_p(t) \rceil, \end{displaymath} and so \begin{displaymath} \val_p(u)=\epsilon+\lceil -\tfrac{n}{m} \val_p(t) \rceil, \end{displaymath} with $|\epsilon|<K_2$. Now, consider $\val_p(t) \ge 0$. Let $K_3$ be a constant such that $\min(\val_p(f(t)), \val_p(g(t))) \le K_3$ for all such $t$, which exists by Lemma~\ref{bd1}. Appealing to \eqref{eq1}, we find $-\val_p(u) \le K_4$, for an appropirate $K_4$. Let $K_5$ be so that $\val_p(f(t)) \ge K_5$ and $\val_p(g(t)) \ge K_5$ for all $t$ with $\val_p(t) \ge 0$. Then appealing to \eqref{eq1} again, we find $-\val_p(u) \ge K_6$, for an appropriate $K_6$. We thus see that $|\val_p(u)|<K_7$ whenever $\val_p(t) \ge 0$, where $K_7=\max(K_4,K_6)$. Combining the above two paragraphs, we see that the formula in the statement of the lemma holds with $C_p=\max(K_2, K_7)$. Now suppose $p$ is large enough so that: (1) the coefficients of $f$ and $g$ are $p$-integral; (2) the leading coefficients of $f$ and $g$ are $p$-unit; and (3) the constant $c_p$ from the previous lemma can be taken to be 1. For $\val_p(t)<0$ we have equalities $\val_p(f(t))=r\val_p(t)$ and $\val_p(g(t))=s\val_p(t)$, which shows that \begin{displaymath} \val_p(u)=\max(\lceil -\tfrac{r}{4} \val_p(t) \rceil, \lceil -\tfrac{s}{6} \val_p(t) \rceil) =\lceil -\tfrac{n}{m} \val_p(t) \rceil. \end{displaymath} For $\val_p(t) \ge 0$, we have $\val_p(f(t)) \ge 0$ and $\val_p(g(t)) \ge 0$, with at least one inequality being an equality. Thus $\val_p(u)=0$. This proves that we can take $C_p=0$ in this case. \end{proof} \begin{lemma} \label{bd3} There exists a finite set $Q$ of non-zero rational numbers with the following property. Suppose $(u,t) \in S_1(X)$. Then we can write $t=a/b^m$, where $a$ and $b$ are integers such that $b>0$ and $\gcd(a,b^m)$ is not divisible by any $m$th power, and $u=qb^n$, with $q \in Q$. \end{lemma} \begin{proof} Suppose $(u,t) \in S_1(X)$. We then have a unique expression $t=a/b^m$, where $a$ and $b>0$ are integers with $\gcd(a,b^m)$ not divisible by any $m$th power. If $p\mid b$ then $-\val_p(t)=m\val_p(b)-k$, where $0 \le k<m$. If $m=1$ then $k=0$, while if $n=1$ then $0 \le \tfrac{n}{m} k<1$; in either case, $\lceil-\tfrac{n}{m} k \rceil=0$. Thus $\val_p(u)=\epsilon+n\val_p(b)$ by Lemma~\ref{bd2}, with $|\epsilon| \le C_p$. If $p \nmid b$ then $\val_p(t) \ge 0$, and so $|\val_p(u)| \le C_p$. We thus see that $|\val_p(u/b^n)| \le C_p$ for all $p$. Since we can take $C_p=0$ for $p \gg 0$, this means that there are only finitely many possibilities for the rational number $u/b^n$. \end{proof} \begin{lemma} \label{bd4} Suppose $(u,t) \in S_1(X)$ and write $t=a/b^m$ and $u=qb^n$ per Lemma~\ref{bd3}. Then $|a| \lesssim X^{m/12n}$ and $|b| \lesssim X^{1/12n}$. \end{lemma} \begin{proof} Suppose $(u,t) \in S_1(X)$ and write $t=a/b^m$ and $u=qb^n$ as above. The inequality $\max(|A|^3,|B|^2)<X$ translates to \begin{displaymath} |u| \max(|f(t)|^{1/4}, |g(t)|^{1/6})<X^{1/12}. \end{displaymath} Let $K_1>0$ be a constant such that $\max(|f(t)|^{1/4}, |g(t)|^{1/6}) \ge K_1$ for all $t$, which exists by Lemma~\ref{bd1}. Then $|u| \le K_1^{-1} X^{1/12}$, and so \begin{displaymath} |b| \le K_2 X^{1/12n}, \end{displaymath} with $K_2=K_1^{-1/n} \max_{q \in Q}(q^{-1/n})$. Suppose for the moment that $|t| \ge 1$. Let $K_3>0$ be a constant so that $K_3^4 |t|^r \le |f(t)|$ and $K_3^6 |t|^s \le |g(t)|$ holds for all such $t$. We thus have \begin{displaymath} X^{1/12}>|u| \max(|f(t)|^{1/4}, |g(t)|^{1/6}) \ge K_3 |u| \max(|t|^{r/4}, |t|^{s/6}) = K_3 |u| |t|^{n/m} =K_4(q) |a|^{n/m}, \end{displaymath} where $K_4(q)=K_3 q$. We therefore find $|a|<K_5 X^{m/12n}$ with $K_5=\max_{q \in Q}K_4(q)^{-m/n}$. Now suppose $|t|<1$. Then $|a|<|b^m|\leq K_2^m X^{m/12n}$. Thus, in all cases, we have \begin{displaymath} |a|<K_6 X^{m/12n} \end{displaymath} with $K_6=\max(K_5,K_2^m)$. \end{proof} The lemma shows that $\# S_1(X) \lesssim X^{(m+1)/12n}$, which implies the same for $\# S(X)$. We have thus established the upper bound in Proposition~\ref{bd}. \subsection{The lower bound} We now turn to the lower bound. Observe that by changing $u$ to $Mu$ for appropriate $M$, it suffices to consider the case where $f$ and $g$ have integer coefficients, which we now assume. For $(a,b) \in \mathbf{Z}^2$ put $u=b^n$ and $t=a/b^m$, and $A=u^4f(t)$ and $B=u^6g(t)$. Note that $A$ and $B$ are integers. Fix a small constant $\kappa>0$ and let $S_2(X)$ denote the set of pairs $(a,b) \in \mathbf{Z}^2$ satisfying the following conditions: \begin{itemize} \item $a$ and $b$ are coprime and $b>0$. \item $|a|<\kappa X^{m/12n}$ and $|b|<\kappa X^{1/12n}$. \item $4A^3+27B^2 \ne 0$. \end{itemize} We note that for appropriately chosen $\kappa$, if $(a,b) \in S_2(X)$ then $|A|<X^{1/3}$ and $|B|<X^{1/2}$. \begin{lemma} \label{bd5} There exists a non-zero integer $D$ with the following property: if $(a,b) \in S_2(X)$ then $\gcd(A^3, B^2)$ divides $D$. \end{lemma} \begin{proof} It suffices to find constants $e_p$ such that $\val_p(\gcd(A^3,B^2)) \le e_p$ and $e_p=0$ for $p \gg 0$, for then we can take $D=\prod_p p^{e_p}$. Suppose $(a,b) \in S_2(X)$, and let $p$ be a prime. Let $K_1$ be a constant so that $|3\val_p(f(t))-3r\val_p(t)| \le K_1$ and $|2\val_p(g(t))-2s\val_p(t)| \le K_1$ for all $t \in \mathbf{Q}$ with $\val_p(t)<0$. Note that for $p \gg 0$ we can take $K_1=0$. Suppose $\val_p(b)=k>0$, so that $\val_p(t)=-mk$ and $\val_p(u)=nk$. Then \begin{displaymath} \val_p(A^3)=\val_p(u^{12} f(t)^3)=12nk-3rmk+\epsilon=12m(\tfrac{n}{m}-\tfrac{r}{4})k+\epsilon \end{displaymath} and \begin{displaymath} \val_p(B^2)=\val_p(u^{12} g(t)^2)=12nk-2smk+\delta = 12m(\tfrac{n}{m}-\tfrac{s}{6})k+\delta \end{displaymath} where $|\epsilon| \le K_1$ and $|\delta| \le K_1$. However, $\tfrac{n}{m}=\max(\tfrac{r}{4}, \tfrac{s}{6})$, and so $\val_p(\gcd(A^3, B^2)) \le K_1$. Let $K_2$ be a constant so that $\max(3\val_p(f(t)), 2\val_p(g(t))) \le K_2$ for all $t \in \mathbf{Q}$ with $\val_p(t) \ge 0$. This constant exists by Lemma~\ref{bd1}; furthermore, we can take $K_2=0$ for $p$ sufficiently large. Since $p$ does not divide $b$, this gives $\val_p(\gcd(A^3,B^2)) \le K_2$. We thus see that we can take $e_p=\max(K_1, K_2)$, and this can be taken to be 0 for $p \gg 0$. \end{proof} Let $S_3(X)$ denote the set of pairs $(A,B) \in \mathbf{Z}^2$ coming from $S_2(X)$. We have a map $S_3(X) \to S(X)$ taking $(A,B)$ to $(A/d^4,B/d^6)$, where $d^{12}$ is the largest 12th power dividing $\gcd(A^3,B^2)$. \begin{lemma}\label{lem:S3toS} There exists a constant $N$ such that every fiber of the map $S_3(X) \to S(X)$ has cardinality at most $N$. \end{lemma} \begin{proof} Let $(A_0,B_0) \in S(X)$ be given. Suppose $(A,B) \in S_3(X)$ maps to $(A_0, B_0)$, i.e., $A=d^4A_0$ and $B=d^6B_0$ for some $d \in \mathbf{Z}$. By Lemma~\ref{bd5}, $d^{12} \mid D$. Therefore, if $N$ is the number of 12th powers dividing $D$, the fibers have cardinality at most $N$. \end{proof} \begin{lemma} There exists a constant $M$ such that every fiber of the map $S_2(X) \to S_3(X)$ has cardinality at most $M$. \end{lemma} \begin{proof} Let $(A,B) \in S_3(X)$ be given. An element $(a,b) \in S_2(X)$ of the fiber gives a solution to the equations $Ax^4=f(y)$ and $Bx^6=g(y)$ by $x=u^{-1}=b^{-n}$ and $y=t=a/b^m$. Furthermore, $(a,b)$ is determined from $(x,y)$. These two equations define plane curves of degree $\max(4,r)$ and $\max(6,s)$, which are easily seen to have no common components. By B\'ezout's theorem, they therefore have at most $M=\max(4,r) \max(6,s)$ points in common, and this bounds the cardinality of the fibers. \end{proof} We leave to the reader the standard estimate $X^{(m+1)/12n} \lesssim \# S_2(X)$. This gives the lower bound in Proposition~\ref{bd} since we have a map $S_2(X) \to S(X)$ whose fibers have bounded size. We have thus completed the proof of Proposition~\ref{bd}. \section{Equations for universal curves} \label{explan} Let $G$ be one of the 12 groups in \eqref{mazur} for which $2G \ne 0$. Let $E$ be a family of elliptic curves over a base scheme $S/\mathbf{Q}$. By a \emph{$G$-structure} on $E$ we mean an injection of groups $i \colon G_S \to E$. When $G=\mathbf{Z}/N\mathbf{Z}$, a $G$-structure is just a section of $E$ of order $N$. There is a moduli space $Y/\mathbf{Q}$ of elliptic curves equipped with a $G$-structure, over which there is a universal elliptic curve $\mathcal{E}$ with a $G$-structure. (Although for $G=\mathbf{Z}/3\mathbf{Z}$ we must slightly change the definition of $Y$, see below.) In fact, $Y$ can be identified with an open subvariety of $\mathbf{A}^1$; choose such an identification. Then $\mathcal{E}$ is given by an equation of the form \begin{equation} \label{univeq} y^2=x^3+f(t)x+g(t), \end{equation} where $f$ and $g$ are rational functions in $t$ with rational coefficients. Making a change of variables, we can assume that $f$ and $g$ are polynomials and $\gcd(f^3,g^2)$ is not divisible by any 12th powers. Note that neither $f$ nor $g$ is zero, as then every geometric fiber of $\mathcal{E}$ would have CM. We show that $f$ and $g$ are coprime and compute their degree, which allows us to apply Theorem~\ref{general} to prove Theorem~\ref{thm:main2}. \subsection{The case $\mathbf{Z}/4\mathbf{Z}$} Let $G=\mathbf{Z}/4\mathbf{Z}$. We have the following result. \begin{proposition} For an appropriate choice of embedding $Y \to \mathbf{A}^1$, the polynomials $f$ and $g$ are coprime and satisfy $\deg(f)=2$ and $\deg(g)=3$. \end{proposition} \begin{proof} By \cite[Tab.~3]{kubert}, the curve $\mathcal{E}$ is given by \begin{displaymath} y^2+xy-ty=x^3-tx^2, \end{displaymath} for some embedding $Y \to \mathbf{A}^1$. The result is obtained from the change of variables: \[ x\mapsto x+\frac{t}{3}-\frac{1}{12}\quad\text{and}\quad y\mapsto y-\frac{x}{2}+\frac{t}{3}+\frac{1}{24}. \] \end{proof} This establishes the second row of Table~\ref{f:params}, from which Theorem~\ref{thm:main2} follows for $G$. \subsection{The cases with $\# G>4$} We assume now that $G$ is one of the 10 groups in \eqref{mazur} of order greater than 4. \begin{proposition} \label{prop1} The polynomials $f$ and $g$ are coprime, and $\deg(f)=4\ell$ and $\deg(g)=6\ell$ for some integer $\ell$. In fact, if $k$ denotes the number of injective group homomorphisms $G \to (\mathbf{Q}/\mathbf{Z})^2$ then $\ell=k/24$. \end{proposition} We first require a lemma. \begin{lemma} Let $A$ be a DVR with fraction field $K$ and residue characteristic 0. Let $E/K$ be an elliptic curve admitting a $G$-structure. Then $E$ has semi-stable reduction. \end{lemma} We give two proofs. \begin{proof}[First proof] Suppose $E$ has additive reduction. Let $\ol{E}/A$ be the N\'eron model of $E$. The $G$-structure on $E$ extends to one on $\ol{E}$ by the N\'eron mapping property. Thus the special fiber $\ol{E}_0$ of $\ol{E}$ contains a subgroup isomorphic to $G$. The map $G \to \pi_0(\ol{E}_0)$ necessarily has a kernel, by the classification of special fibers of N\'eron models. Thus $G$ meets the identity component $\ol{E}_0^{\circ}$ non-trivially. But this is a contradiction, since $\ol{E}_0^{\circ} \cong \mathbf{G}_a$ is torsion-free. \end{proof} \begin{proof}[Second proof] Let $X$ be the moduli space of generalized elliptic curves with $G$-structure, in the sense of Deligne--Rapoport. Then $X$ is a proper scheme. It follows that $E/K$ extends to a generalized elliptic curve $\ol{E}/A$ with $G$-structure. The identity component of the smooth locus of the special fiber of $\ol{E}$ is either an elliptic curve or a torus, and so $E$ has semi-stable reduction. \end{proof} \begin{proof}[Proof of Proposition~\ref{prop1}] By the lemma, the universal curve $\mathcal{E}$ has semi-stable reduction at all places of $\mathbf{P}^1_{\mathbf{Q}}$. Since the equation \eqref{univeq} for $\mathcal{E}$ is minimal at all finite places (in the sense of \cite[Ch.~VII \S 1]{silverman}), it follows that $f$ and $g$ are coprime (see \cite[Ch.~VII, Prop.~5.1]{silverman} and \cite[Exc.~7.1]{silverman}). Now, \eqref{univeq} is not necessarily minimal at infinity. The minimal equation is given by \begin{displaymath} y^2=x^3+t^{-4\ell} f(t)+t^{-6\ell} g(t) \end{displaymath} where $\ell$ is minimal so that $t^{-4\ell} f(t)$ and $t^{-6\ell} g(t)$ have non-negative valuation at $\infty$. (Note: the valuation at $\infty$ is $-\deg$.) Now, $Y \subset \mathbf{P}^1$ is exactly the locus where $\mathcal{E}$ has good reduction, by universality. (Alternatively, the Deligne--Rapoport theory shows that $\mathcal{E}$ extends to a genuinely generalized elliptic curve over $\infty$). Thus $\mathcal{E}$ has multiplicative reduction at $\infty$, and so both $t^{-4\ell} f(t)$ and $t^{-6\ell} g(t)$ have valuation zero at $\infty$ \cite[Exc.~7.1]{silverman}, which exactly says that $\deg(f)=4\ell$ and $\deg(g)=6\ell$. Now, the $j$-invariant of the family $\mathcal{E}$ is (up to a scalar) \begin{displaymath} \frac{f^3}{4f^3+27g^2}. \end{displaymath} We regard this as a self-map of $\mathbf{P}^1$. Since $f$ and $g$ are coprime, it has degree $12\ell$. On the other hand, the degree of the map $j \colon Y \to \mathbf{P}^1$ is the number of points in a typical fiber over a complex point. If $E/\mathbf{C}$ is an elliptic curve, the fiber of $j$ over $E$ is the set of isomorphism classes of pairs $(E, i)$ where $i \colon G \to E$ is a $G$-structure. Generically, $\Aut(E)=\{\pm 1\}$, and so the pairs $(E, i)$ and $(E, -i)$ are isomorphic and there are no other identifications. Thus the fiber of $j$ over $E$ has $k/2$ points, and so $12\ell=k/2$. \end{proof} Now, in the notation of Table~\ref{f:params}, we have $r=4\ell$, $s=6\ell$, $n=\ell$, and $m=1$. As $\ell=k/24$, it suffices to compute $k$. When $G=\mathbf{Z}/N\mathbf{Z}$, we have $k=N^2 \prod_{p|N} (1-p^{-2})$; and when $G=\mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/N\mathbf{Z}$ we have $k=2N^2 \prod_{p \mid N} (1-p^{-2})$. This establishes the row in Table~\ref{f:params} corresponding to $G$. Theorem~\ref{thm:main2} for $G$ then follows from Theorem~\ref{general}. \subsection{The case $\mathbf{Z}/3\mathbf{Z}$} \label{ss:z3} We now treat the case $G=\mathbf{Z}/3\mathbf{Z}$. There is a minor complication owing to the fact that $\mathcal{Y}_1(3)$ is not a scheme. Since we will treat this group more thoroughly in \S \ref{sec:asymptotics}, we do not include all the details here. Call a pair $(E, P)$ consisting of an elliptic curve $E$ over a field $k$ of characteristic 0 and a 3-torsion point $P$ \emph{exceptional} if there is a non-trivial automorphism of $E$ over $\ol{k}$ fixing $P$. Let $\rho=e^{2\pi i/3}$, let $E=\mathbf{C}/\mathbf{Z}[\rho]$, and let $P$ be the 3-torsion point on $E$ given by $(1-\rho)/3$. Then $(E, P)$ is exceptional. Furthermore, it is the unique exceptional pair over $\mathbf{C}$, up to isomorphism. There is a moduli space $Y$ of unexceptional pairs, which can be identified with an open subvariety of $\mathbf{A}^1$. (In fact, $Y$ is the complement of the single exceptional point in the stack $\mathcal{Y}_1(3)$.) Let $\mathcal{E}$ be the universal elliptic curve over $Y$. \begin{lemma} \label{lem:z3} Let $E/\mathbf{Q}$ be an elliptic curve and let $P$ be a rational point of order 3. Then $E$ admits an equation of the form \begin{displaymath} y^2+axy+by = x^3 \end{displaymath} with $a,b \in \mathbf{Q}$ such that $P=(0,0)$. The pair $(E,P)$ is exceptional if and only if $a=0$. \end{lemma} \begin{proof} The first statement is well-known; see \cite[Tab.~3]{kubert}, for instance. The second is left to the reader \end{proof} \begin{proposition} For an appropriate embedding $Y \to \mathbf{A}^1$, the universal family $\mathcal{E}$ is given by $y^2=x^3+f(t)x+g(t)$ with $f(t)=2t-\tfrac{1}{3}$ and $g(t)=t^2+\tfrac{2}{3}t+\tfrac{2}{27}$. \end{proposition} \begin{proof} The previous lemma can, in fact, be applied to $\mathcal{E}$, and shows that it admits an equation of the form $y^2+2axy+2by=x^3$ where $a$ and $b$ are functions on $Y$, with $a$ invertible. Changing $y$ to $y-(ax+b)$ yields the equation $y^2=x^3+(ax+b)^2$. Changing $(x,y)$ to $(a^2x,a^3y)$ now gives $y^2=x^3+(x+t)^2$ with $t=b/a^3$. Finally, changing $x$ to $x-\tfrac{1}{3}$ yields the stated equation. \end{proof} Let $N^1(X)$ (resp.\ $N^2(X)$) be the number of (isomorphism classes of) elliptic curves $E/\mathbf{Q}$ of height at most $X$ admitting a 3-torsion point $P$ such that $(E,P)$ is not (resp.\ is) exceptional. Combining the above proposition with Theorem~\ref{general}, we see that $X^{1/3} \lesssim N^1(X) \lesssim X^{1/3}$. By Lemma~\ref{lem:z3}, an exceptional curve admits an equation of the form $y^2=x^3+b^2$ with $b \in \mathbf{Z}$. Clearly then, $N^2(X) \lesssim X^{1/4}$. Finally, we have the obvious bounds $N^1(X) \le N'_G(X) \le N^1(X)+N^2(X)$. This proves $X^{1/3} \lesssim N'_G(X) \lesssim X^{1/3}$, which establishes Theorem~\ref{thm:main2} for $G$. \section{The group $\mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/2\mathbf{Z}$}\label{sec:Zmod2timesZmod2} We begin with a variant of Proposition~\ref{bd}. \begin{proposition} \label{bd-var} Let $f,g \in \mathbf{Q}[t]$ be coprime polynomials of degrees $r$ and $s$. Assume one of $r$ or $s$ is positive. Write \begin{displaymath} \max \left( \frac{r}{2}, \frac{s}{3} \right)=\frac{n}{m} \end{displaymath} with $n$ and $m$ coprime. Assume $n=1$ or $m=1$. Let $S(X)$ be the set of pairs $(A,B) \in \mathbf{Z}^2$ satisfying the following conditions: \begin{itemize} \item $4A^3+27B^2 \ne 0$. \item $|A|<X^{1/3}$ and $|B|<X^{1/2}$. \item $\gcd(A^3,B^2)$ is not divisible by any 12th power. \item There exist $u,t \in \mathbf{Q}$ such that $A=u^2 f(t)$ and $B=u^3 g(t)$. \end{itemize} Define \begin{displaymath} h(X) = \begin{cases} X^{(m+1)/6n} & \textrm{if $m+1>n$} \\ X^{1/6} \log(X) & \textrm{if $m+1=n$} \\ X^{1/6} & \textrm{if $m+1<n$} \end{cases} \end{displaymath} Then $h(X) \lesssim \# S(X) \lesssim h(X)$. \end{proposition} \begin{proof} We only sketch the proof, as the details are similar to the proof of Proposition~\ref{bd}. We begin with the upper bound. A version of Lemma~\ref{bd2} holds, but now one only has $C_p=1$ for $p \gg 0$ --- the reason for this is that putting an extra $p$ into $u$ only adds two $p$'s to $A$ and three to $B$, and so does not contribute a $p^{12}$ in $\gcd(A^3, B^2)$. The analog of Lemma~\ref{bd3} then says that we can write $t=a/b^m$, with $\gcd(a, b^m)$ not divisible by any $m$th powers, and $u=qcb^n$, where $q$ belongs to a finite set and $c$ is squarefree. The analog of Lemma~\ref{bd4} yields the inequalities $|ca^{n/m}|\lesssim X^{1/6}$ and $|cb^n| \lesssim X^{1/6}$. To count the number of possibilities, note that for any given $c$ there are at most $X^{m/6n}/c^{m/n}$ possibilities for $a$ and $X^{1/6n}/c^{1/n}$ possibilities for $b$, yielding $X^{(1+m)/6n}/c^{(1+m)/n}$ total possibilities. Now integrate over $c$ (up to $X^{1/6}$) to obtain the upper bound. We now turn to the lower bound. We only treat the case $m+1>n$, as that is the only one we need and the others are a bit more subtle. It turns out that, in this case, we can find enough points with $c=1$ to establish the bound. More precisely, let $\kappa>0$ be a small constant, and consider the set of pairs of coprime integers $(a,b)$ such that $|a^{n/m}|<\kappa X^{1/6}$ and $|b^n|<\kappa X^{1/6}$ and $\Delta \ne 0$. The number of such pairs is $\gtrsim X^{(m+1)/6n}$. The analog of Lemma~\ref{bd5} shows that $\gcd(A^3, B^2)$ is of the form $\alpha^6 \beta$, where $\beta$ divides a fixed integer $D$, and $\alpha$ is square-free. This is sufficient for the application of Lemma~\ref{lem:S3toS}. \end{proof} \begin{proposition} Let $f(t)=-\frac{1}{3} (t^2-t+1)$ and $g(t)=\tfrac{1}{27} (-2t^3+3t^2+3t-2)$. Suppose $E/\mathbf{Q}$ is an elliptic curve given by an equation $y^2=x^3+Ax+B$. Then all 2-torsion points of $E$ are rational if and only if there exist $u,t \in \mathbf{Q}$ such that $A=u^2f(t)$ and $B=u^3g(t)$. \end{proposition} \begin{proof} It is well-known that all of the 2-torsion of $E$ is rational if and only if it admits an equation of the form $y^2=x(x-a)(x-b)$ with $a,b \in \mathbf{Q}$. Changing $x$ to $x+\tfrac{1}{3}(a+b)$, we obtain the equation \begin{displaymath} y^2=x^3-\tfrac{1}{3} (a^2-ab+b^2)+\tfrac{1}{27} (-2a^3+3a^2b+3ab^2-2b^3). \end{displaymath} We thus see that all of the 2-torsion of $E$ is rational if and only if there exist $a,b,v \in \mathbf{Q}$ such that \begin{displaymath} A=-\frac{v^4}{3} (a^2-ab+b^2), \qquad B=\frac{v^6}{27} (-2a^3+3a^2b+3ab^2-2b^3). \end{displaymath} Changing $(a,b)$ to $(a/v^2, b/v^2)$ shows that we can take $v=1$. The above equations are therefore equivalent to $A=u^2f(t)$ and $B=u^3g(t)$ with $u=b$ and $t=a/b$. (Note: a solution with $b=0$ is impossible, as that would give $4A^3+27B^2=0$.) \end{proof} \begin{proposition} Let $G=\mathbf{Z}/2\mathbf{Z} \times \mathbf{Z}/2\mathbf{Z}$. Then $X^{1/3} \lesssim N'_G(X) \lesssim X^{1/3}$. \end{proposition} \begin{proof} This follows from the above two propositions. Note that, in the notation of Proposition~\ref{bd-var}, we have $m=n=1$, and so $(m+1)/6n=1/3$. \end{proof} \section{Asymptotics for the trivial group, $\mathbf{Z}/2\mathbf{Z}$, and $\mathbf{Z}/3\mathbf{Z}$}\label{sec:asymptotics} In this section, we derive asymptotics for $N_G(X)$ when $G$ has order at most 3. The result is stated in Theorem~\ref{lem:main_asymp_lemma} below and is a refinement of Theorem~\ref{thm:asymptotics}. We take a unified approach by first counting integer points in explicit semi-algebraic sets using the Principle of Lipschitz \cite{davenport} and then sieving. Note that these asymptotics imply Theorem~\ref{thm:main} for these cases. \begin{lemma} Let $E/\mathbf{Q}$ be an elliptic curve given by $y^2=x^3+Ax+B$ with $A$ and $B$ in $\mathbf{Z}$. Then, \begin{enumerate} \item\label{lempart:2tors} $E$ has a rational point of order 2 if and only if there exists $b \in \mathbf{Z}$ such that $B=b^3+Ab$; \item\label{lempart:3tors} $E$ has a rational point of order 3 if and only if there exists $a, b \in \mathbf{Z}$ such that $A=6ab+27 a^4$ and $B=b^2-27a^6$. \end{enumerate} \end{lemma} \begin{proof} As is well-known, $E$ has a rational 2-torsion point if and only if $x^3+Ax+B$ has a rational root. This yields part (\ref{lempart:2tors}). Part (\ref{lempart:3tors}) is proved in \cite[\S2]{GST} \end{proof} This lemma suggests we study integer points in the following three semi-algebraic regions: \begin{align*} R_1(X)&=\{(a,b)\in\mathbf{R}^2:|a|<X^{1/3}\text{ and }|b|<X^{1/2}\},\\ R_2(X)&=\{(a,b)\in\mathbf{R}^2:|a|<X^{1/3}\text{ and }|b^3+ab|<X^{1/2}\}, \textrm{and}\\ R_3(X)&=\{(a,b)\in\mathbf{R}^2:|6ab+27a^4|<X^{1/3}\text{ and }|b^2-27a^6|<X^{1/2}\}. \end{align*} The Principle of Lipschitz states that the number of integer points $r_{i}(X)$ in $R_i(X)$ is given by its area up to an error given by the maximal length of its projections onto the coordinate axes. From this we immediately see that $r_1(X)=4X^{5/6}+O(X^{1/2})$. The next lemma will allow us to compute the main term and the error for $i=2,3$. Before stating it, we must introduce several quantities. Let $\alpha_\pm$ be the (unique) real root of $x^3\pm x-1$. Let $f_\pm=3x^4\pm6x^2+12x-1$, each of which has one negative and one positive root. We denote these four roots by \begin{align*} &\alpha_4\approx0.08011\quad\text{and}\quad\alpha_1\approx-1.22259\quad\text{(roots of }f_+);\\ &\alpha_3\approx0.08711\quad\text{and}\quad\alpha_0\approx-2.01637\quad\text{(roots of }f_-). \end{align*} Also, define $\alpha_2=-\sqrt{3}$ and $\alpha_5=\sqrt{3}$. Let $\displaystyle \beta_i=\sgn(\alpha_i)\sqrt{|\alpha_i|/3}$ except take $\beta_3$ to be negative. Then, $\beta_i<\beta_j$ for $i<j$. Define the hyperelliptic integrals \[ I_+=\int_{\beta_3}^{\beta_4}\sqrt{1+27a^6}da\approx0.33383 \] and \[ I_-=\int_{\beta_0}^{\beta_1}\sqrt{-1+27a^6}da\approx0.32030. \] Finally, let \[ A_\pm(a)=\frac{\pm X^{1/3}-27a^4}{6a}\quad\text{ and }\quad B_\pm(a)=\sqrt{\pm X^{1/2}+27a^6}. \] We leave the verification of the following lemma to the reader. \begin{lemma}\mbox{} \begin{enumerate} \item We have $(a,b)\in R_2(X)$ if and only if \begin{align*} & |b|<\alpha_+X^{1/6}\text{ and }|a|^3<X\text{, or} \\ & \alpha_+X^{1/6}\leq|b|<\alpha_-X^{1/6}\text{ and }-X^{1/3}<a<\frac{X^{1/2}}{|b|}-b^2. \end{align*} Thus, \[ \operatorname{Area}(R_2(1))=2\log(\alpha_-/\alpha_+)+\tfrac{4}{3} (\alpha_++\alpha_-). \] \item Let $R^+_3(X)=R_3(X)\cap\{b\geq0\}$. Then, $(a,b)\in R^+_3(X)$ if and only if \begin{displaymath} \beta_iX^{1/12}<a<\beta_{i+1}X^{1/12} \qquad \textrm{and} \qquad g_i(a)<b<f_i(a), \end{displaymath} for some $0\leq i\leq 4$, where $f_i(a)$ and $g_i(a)$ are as in the following table. \begin{center} {\tabulinesep=1.2mm \begin{tabu}{|c||c|c|c|c|c|} \hline $i$ & 0 & 1 & 2 & 3 & 4 \\ \hline $f_i$ & $A_-(a)$ & $A_-(a)$ & $A_-(a)$ & $B_+(a)$ & $A_+(a)$ \\ \hline $g_i$ & $B_-(a)$ & $A_+(a)$ & 0 & 0 & 0 \\ \hline \end{tabu} } \end{center} Thus, \[ \operatorname{Area}(R_3^+(1))=I_+-I_-+\frac{1}{6}\log\left(\displaystyle\frac{\beta_0\beta_1\beta_5}{\beta_2\beta_3\beta_4}\right)+\frac{9}{8}\left(\beta_0^4+\beta_2^4+\beta_4^4-\beta_1^4-\beta_3^4-\beta_5^4\right). \] Furthermore, $(a,b)\in R_3(X)$ if and only if $(-a,-b)\in R_3(X)$. In particular, $|b|<2\sqrt{7}X^{1/4}$ for all $(a,b)\in R_3(X)$. \end{enumerate} \end{lemma} Let $(d_1,d_2,d_3)=(6/5,2,3)$ and $(e_1,e_2,e_3)=(2,3,4)$. Note that for each $i$, the region $R_i(X)$ is homogeneous in $X$ in the sense that \[ \operatorname{Area}(R_i(X))=X^{1/d_i}\operatorname{Area}(R_i(1)). \] Therefore, combining the above lemma with the Principal of Lipschitz yields \begin{equation}\label{eqn:Ri_count} r_i(X)=\operatorname{Area}(R_i(1))X^{1/d_i}+O(X^{1/e_i}). \end{equation} All that remains is to address the overcounting that occurs in $R_i(X)$. We do this in three lemmas. \begin{lemma} Define $(A,B)=T_i(a,b)$ by \begin{enumerate} \item $T_1(a,b)=(a,b)$, if $i=1$; \item $T_2(a,b)=(a,b^3+ab)$, if $i=2$; \item $T_3(a,b)=(6ab+27a^4,b^2-27a^6)$, if $i=3$. \end{enumerate} Then, the number of $(a,b)\in R_i(X)\cap\mathbf{Z}^2$ such that $4A^3+27B^2=0$ is $O(X^{1/6})$. \end{lemma} \begin{proof} Suppose $4A^3+27B^2=0$. First, consider $i=1$. Then, $4a^3=-27b^2$, which implies $a^\prime=-a/3\in\mathbf{Z}$ and $b^\prime=b/2\in\mathbf{Z}$. Since ${a^\prime}^3={b^\prime}^2<X$, the number of $(a,b)$ with discriminant zero is on the order of the number of sixth powers less than $X$, namely it is $O(X^{1/6})$. For $i=2$, B\'{e}zout's theorem says there are at most 3 pairs $(a,b)$ giving $(A,B)$; for $i=3$, the number is 24. Appealing to the argument for $i=1$ again yields the bound $O(X^{1/6})$. \end{proof} In view of this lemma, we may (and do) redefine $R_i(X)$ by removing from it the discriminant zero locus, while preserving the truth of \eqref{eqn:Ri_count}. \begin{lemma} The number of $(A,B)$ for which there is more than one $(a,b)\in R_i(X)\cap\mathbf{Z}^2$ with $(A,B)=T_i(a,b)$ is $O(X^{1/e_i})$. \end{lemma} \begin{proof} This is non-trivial only for $i=2,3$. Consider first $i=2$. Given an $E/\mathbf{Q}$ with equation $y^2=x^3+Ax+B$ and an integer $b$ such that $B=b^3+Ab$, one obtains a non-trivial 2-torsion point $(-b,0)\in E(\mathbf{Q})$. Therefore, if there is another $b^\prime\in\mathbf{Z}$ such that $B=b^{\prime3}+Ab^\prime$, then the torsion subgroup contains $\mathbf{Z}/2\mathbf{Z}\times\mathbf{Z}/2\mathbf{Z}$. But we proved in \S\ref{sec:Zmod2timesZmod2} that the number of such elliptic curves is $O(X^{1/3})$. The case $i=3$ is similar. Indeed, each pair $(a,b)\in\mathbf{Z}^2$ such that $(A,B)=T_3(a,b)$ yields a pair of non-trivial 3-torsion points $(3a^2, \pm(9a^3+b))\in E(\mathbf{Q})$, which are negatives of each other. Ignoring the cases where $ab=0$ (which is $O(X^{1/4})$ many), a second pair $(a^\prime,b^\prime)$ that gives the same $(A,B)$ yields two \emph{new} non-trivial rational 3-torsion points. But this is impossible: an elliptic curve over $\mathbf{Q}$ cannot contain two copies of $\mathbf{Z}/3\mathbf{Z}$ in its Mordell--Weil group. \end{proof} Let $\mathfrak{E}_i(X)=\{\text{equations }E_{A,B}:y^2=x^3+Ax+B:(A,B)=T_i(a,b),(a,b)\in R_i(X)\cap\mathbf{Z}^2\}$. The above lemma implies that $\#\mathfrak{E}_i(X)=r_i(X)+O(X^{1/e_i})$. We now sieve out non-minimal equations to make our way from $\mathfrak{E}_i(X)$ to $N_{\mathbf{Z}/i\mathbf{Z}}(X)$. \begin{theorem}\label{lem:main_asymp_lemma} For $i=1,2,$ and $3$, \[ N_{\mathbf{Z}/i\mathbf{Z}}=\frac{\operatorname{Area}(R_i(1))}{\zeta(12/d_i)}X^{1/d_i}+O(X^{1/e_i}). \] \end{theorem} \begin{proof} Given an equation $E_{A,B}\in\mathfrak{E}_i(X)$, let $d$ be the largest twelfth power dividing $\gcd(A^3,B^2)$. Then, the curve defined by $E_{A,B}$ is also given (uniquely) by the minimal equation $E_{d^{-4}A,d^{-3}B}$. Sieving yields \begin{align*} N_{\mathbf{Z}/i\mathbf{Z}}^\prime(X) &=\sum_{d=1}^{X^{1/12}}\mu(d)\#\mathfrak{E}_i(d^{-12}X)\\ &=\sum_{d=1}^{X^{1/12}}\mu(d)\left(\operatorname{Area}(R_i(1))\frac{X^{1/d_i}}{d^{12/d_i}}+O(d^{-12/d_i}X^{1/e_i})\right)\\ &=\operatorname{Area}(R_i(1))X^{1/d_i}\sum_{d=1}^{X^{1/12}}\frac{\mu(d)}{d^{12/d_i}}+O(X^{1/e_i})\\ &=\frac{\operatorname{Area}(R_i(1))}{\zeta(12/d_i)}X^{1/d_i}+O(X^{1/e_i}). \end{align*} Since $N_{\mathbf{Z}/i\mathbf{Z}}(X)-N_{\mathbf{Z}/i\mathbf{Z}}^\prime(X)$ is within the error, the theorem is proved. \end{proof} As remarked in the introduction, the case $i=1$ already appears in \cite[Lemma~4.3]{brumer}. In \cite[\S2]{grant}, the case $i=2$ is shown with the error bound $O(X^{\frac{1}{3}+\epsilon})$, for all $\epsilon>0$. The case $i=3$ appears to be new, and we have found no further cases in the literature either.

https://arxiv.org/abs/1609.06473

Explicit formulas for enumeration of lattice paths: basketball and the kernel method

This article deals with the enumeration of directed lattice walks on the integers with any finite set of steps, starting at a given altitude $j$ and ending at a given altitude $k$, with additional constraints such as, for example, to never attain altitude $0$ in-between. We first discuss the case of walks on the integers with steps $-h, \dots, -1, +1, \dots, +h$. The case $h=1$ is equivalent to the classical Dyck paths, for which many ways of getting explicit formulas involving Catalan-like numbers are known. The case $h=2$ corresponds to "basketball" walks, which we treat in full detail. Then we move on to the more general case of walks with any finite set of steps, also allowing some weights/probabilities associated with each step. We show how a method of wide applicability, the so-called "kernel method", leads to explicit formulas for the number of walks of length $n$, for any $h$, in terms of nested sums of binomials. We finally relate some special cases to other combinatorial problems, or to problems arising in queuing theory.

\section{Introduction}\label{Section1} While analysing permutations sortable by a stack, Knuth~\cite[Ex.~1--4 in Sec.~2.2.1]{knuth} showed they were counted by Catalan numbers, and were therefore in bijection with Dyck paths (lattice paths with steps $(1,1)$ and $(1,-1)$ in the plane integer lattice, from the origin to some point on the $x$-axis, and never running below the $x$-axis in-between). He used a method to derive the corresponding generating function (see \cite[p.~536ff]{knuth}) which Flajolet coined {\it``kernel method''}. That name stuck among combinatorialists, although the method already existed in the folklore of statistics and statistical physics --- without a name. The method was later generalized to enumeration and asymptotic analysis of directed lattice paths with any set of steps, and many other combinatorial structures enumerated by bivariate or trivariate functional equations (see, e.g., \cite{Fayolle99, bope00, BaBoDeGaGo02, BaFl02, BoJe06, Eynard16}). We refer to the introduction of~\cite{BaWa16} for a more detailed history of the kernel method. The emphasis in \cite{BaFl02} is on asymptotic analysis, for which the derived (exact) enumeration results serve as a starting point. The latter are in a sense {\it implicit}, since they involve solutions to certain algebraic equations. They are nevertheless perfect for carrying out singularity analysis, which in the end leads to very precise asymptotic results. In general, it is not possible to simplify the exact enumeration results from~\cite{BaFl02}. However, for models involving special choices of step sets, this is possible. These potential simplifications are the main focus of our paper. \begin{figure}[hb] {\includegraphics[width=109mm]{queue}} \caption{A queue corresponding to the basketball walk model.} \label{queue} \end{figure} Such models appear frequently in queuing theory. Indeed, birth and death processes and queues, like the one shown in Figure~\ref{queue}, are naturally encoded by lattice paths (see~\cite{Bohm10,Harris08,KrMo10,KrinikShun11,Margolius17,Mohanty07}). In this article, we solve a problem raised during the 2015 International Conference on ``Lattice Path Combinatorics and Its Applications'': to find closed-form formulas for the number of walks of length $n$ from~$0$ to $k$ for a full family of models similar to Figure~\ref{queue}. As it turns out, the essential tool to achieve this goal is indeed the kernel method. \medskip Our paper is organized as follows. We begin with some preliminaries in Section~\ref{Section2}. In particular, we introduce the directed lattice paths that we are going to discuss here, we provide a first glimpse of the kernel method, and we briefly review the Lagrange--B\"urmann inversion formula for the computation of the coefficients of implicitly defined power series. Section~\ref{Section3} is devoted to (old-time) ``basketball walks", which, by definition, are directed lattice walks with steps from the set $\{(1,-2),(1,-1),(1,1),(1,2)\}$ which always stay above the $x$-axis. (They may be seen as the evolution of --- pre~1984 --- basketball games; see the beginning of that section for a more detailed explanation of the terminology.) We provide exact formulas (often several, not obviously equivalent ones) for generating functions and for the numbers of walks under various constraints. At the end of Section~\ref{Section3}, we also briefly address the asymptotic analysis of the number of these walks. Section~\ref{Section4} then considers the more general problem of enumerating directed walks where the allowed steps are of the form $(1,i)$ with $-h\le i\le h$ (including $i=0$ or not). Again, we provide exact formulas for generating functions --- in terms of roots of the so-called kernel equation --- and for numbers of walks --- in terms of nested sums of binomials. All these results are obtained by appropriate combinations of the kernel method with variants of the Lagrange--B\"urmann inversion formula. In the concluding Section~\ref{Section5}, we relate basketball walks with other combinatorial objects, namely \begin{itemize} \item with certain trees coming from option pricing, \item with increasing unary-binary trees which avoid a certain pattern which arose in work of Riehl~\cite{Riehl16}, \item and with certain Boolean bracketings which appeared in work of Bender and Williamson~\cite{bender91}. \end{itemize} \section{The general setup, and some preliminaries}\label{Section2} In this section, we describe the general setup that we consider in this article. We use (subclasses of) so-called {\it \L ukasiewicz paths} as main example(s) which serve to illustrate this setup. We recall here as well the main tools that we shall use in this article: the {\it kernel method\/} and the {\it Lagrange--B\"urmann inversion formula}. We start with the definition of the lattice paths under consideration. \begin{definition} \label{def:LP} A {\it step set} $\mathcal{S} \subset \mathbb{Z}^2$ is a finite set of vectors $$\{ (x_1,y_1), (x_2,y_2), \ldots, (x_m,y_m)\}.$$ An $n$-step \emph{lattice path} or \emph{walk} is a sequence of vectors $v = (v_1,v_2,\ldots,v_n)$, such that $v_j$ is in $\mathcal{S}$. Geometrically, it may be interpreted as a sequence of points $\omega =(\walk{0},\walk{1},\ldots,\walk{n})$, where $\walk{i} \in \mathbb{Z}^2, \walk{0} = (0,0)$ (or another starting point), and $\walk{i}-\walk{i-1} = v_i$ for $i=1,2,\ldots,n$. The elements of $\mathcal{S}$ are called \emph{steps}. The \emph{length} $|\omega|$ of a lattice path is its number $n$ of steps. \end{definition} The lattice paths can have different additional constraints shown in Table~\ref{fig:4types}. \vspace{-1mm} \begin{table*}[ht] \small \begin{center}\renewcommand{\tabcolsep}{3pt} \begin{tabular}{|c|c|c|} \hline & ending anywhere & ending at 0\\ \hline \begin{tabular}{c} unconstrained \\ (on~$\mathbb{Z}$) \end{tabular} & \begin{tabular}{c} {\includegraphics[width=48mm]{Zche}} \\ walk/path ($\mathcal W$) \end{tabular} & \begin{tabular}{c} {\includegraphics[width=48mm]{Zexc}} \\ bridge ($\mathcal B$) \end{tabular} \\ \hline \begin{tabular}{c}constrained\\ (on $\mathbb{N}$) \end{tabular} & \begin{tabular}{c} \includegraphics[width=48mm]{Nche} \\ meander ($\mathcal M$)\\ \end{tabular} & \begin{tabular}{c} {\includegraphics[width=48mm]{Nexc}} \\ excursion ($\mathcal E$)\\ \end{tabular}\\ \hline \end{tabular} \end{center} \vskip2pt \caption{\label{fig:4types} The four types of walks: unconstrained walks, bridges, meanders, and excursions.} \end{table*} \vspace{-7mm} We restrict our attention to \emph{directed paths}, which are defined by the fact that, for each step $(x,y) \in \mathcal{S}$, one has $x \geq 0$. Moreover, we will focus only on the subclass of \emph{simple paths}, where every element in the step set $\mathcal{S}$ is of the form $(1,b)$. In other words, these paths constantly move one step to the right. Thus, they are essentially one-dimensional objects and can be seen as walks on the integers. We introduce the abbreviation $\mathcal{S} = \{ b_1,b_2, \ldots, b_n \}$ in this case. A \emph{{\L}ukasiewicz path} is a simple path where its associated step set $\mathcal{S}$ is a subset of $\{-1,0,1,\ldots\}$ and $-1 \in \mathcal{S}$. \begin{example}[\sc Dyck paths] \label{ex:dyck} A Dyck path is a path constructed from the step set $\mathcal{S} = \{-1,+1\}$, which starts at the origin, never passes below the $x$-axis, and ends on the $x$-axis. In other words, Dyck paths are excursions with step set $\mathcal{S}=\{-1,+1\}$. \end{example} The next definition allows to merge the probabilistic point of view ({\it random walks}) and the combinatorial point of view ({\it lattice paths}). \begin{definition} For a given step set $\mathcal{S} = \{s_1,s_2,\ldots,s_m\}$, we define the corresponding {\it system of weights}\index{lattice path!weights} as $ \{p_1,p_2,\ldots,p_m\}$, where $p_j >0$ is the weight associated with step $s_j$ for $j=1,2,\ldots,m$. The {\it weight of a path} is defined as the product of the weights of its individual steps. \end{definition} Next we introduce the algebraic structures associated with the previous definitions. The \emph{step polynomial\/} of a given step set $\mathcal{S}$ is defined as the Laurent polynomial\footnote{By a {\it Laurent polynomial\/} in $u$ we mean a polynomial in $u$ and $u^{-1}$.} \begin{align*} P(u) := \sum_{j=1}^m p_j u^{s_j}. \end{align*} Let \begin{equation}\label{c_d} c = - \min_j s_j\quad \text{and}\quad d = \max_j s_j \end{equation} be the two extreme step sizes, and assume throughout that $c,d >0$. Note that for {\L}ukasiewicz paths we have $c=1$. We start with the easy case of unconstrained paths. We define their bivariate generating function as $$W(z,u) = \sum_{n=0}^\infty \sum_{k=-\infty}^\infty W_{n,k} z^n u^k,$$ where $W_{n,k}$ is the number of unconstrained paths ending after $n$ steps at altitude $k$. It is well-known and straightforward to derive that \begin{align} \label{eq:unconstrainedpaths} W(z,u) &= \frac{1}{1- zP(u)}. \end{align} We continue with the generating function of meanders: \begin{align*} F(z,u):=\sum_{n=0}^\infty \sum_{k=0}^\infty F_{n,k}z^nu^k, \end{align*} where $F_{n,k}$ is the number of paths ending after $n$ steps at altitude $k$, and constrained to be always at altitude $\geq 0$ in-between. Note that we are mainly interested in solving the counting problem, i.e., determining the numbers $F_{n,k}$ for specific families of paths (see Table~\ref{fig:4types}). The generating function encodes all information we are interested in. We decompose $F(z,u)$ in two ways, namely \begin{align*} F(z,u) = \sum_{k \geq 0} F_k(z) u^k = \sum_{n \geq 0} f_n(u) z^n. \end{align*} Here, the generating functions $F_k(z)$ enumerate paths ending at altitude $k$, i.e., $F_k(z) = \sum_{n \geq 0} F_{n,k} z^n$. In particular, the generating function for excursions is equal to $F_0(z)$. On the other hand, the polynomials $f_n(u)$ enumerate paths of length $n$. The power of $u$ encodes their final altitude. We will use this decomposition for a step-by-step approach, similar to the one in the case of unconstrained paths. For the sake of illustration, we show below how the kernel method can be used to find a closed form for the generating function of a given class of {\L}ukasiewicz paths. \begin{theorem} \label{theo:luka} Let $\mathcal{S}$ be the step set of a class of {\L}ukasiewicz paths, and let $P(u)$ be the associated step polynomial. Then the bivariate generating function of meanders (where $z$ marks length, and $u$ marks final altitude) and excursions are \begin{align} \label{eqFzuLuka} F(z,u) &= \frac{1 - zF_0(z)/u}{1-zP(u)} & \text{and} && F_0(z) &= \frac{u_1(z)}{z}, \end{align} respectively, where $u_1(z)$ is the unique small solution of the implicit equation \begin{align*} 1 - zP(u) =0, \end{align*} that is, the unique solution satisfying $\lim_{z \to 0} u_1(z) = 0$. \end{theorem} \begin{proof} A meander of length $n$ is either empty, or it is constructed from a meander of length $n-1$ by appending a possible step from $\mathcal{S}$. However, a meander is not allowed to pass below the $x$-axis, thus at altitude $0$ it is not allowed to use the step $-1$. This translates into the relations \begin{align*} f_0(u) &= 1, & f_{n+1}(u) &= \{u^{\geq 0}\} \left( P(u) f_n(u) \right), \end{align*} where $\{u^{\geq 0}\}$ is the linear operator extracting all terms in the power series representation containing non-negative powers of $u$. Multiplying both sides of the above equation by $z^{n+1}$ and subsequently summing over all $n \geq 0$, we obtain the functional equation \begin{align*} F(z,u) &= 1+ zP(u) F(z,u) - \frac{z}{u} F_0(z). \end{align*} Equivalently, \begin{align} \label{eq:genFuncEq} {\left(1-zP(u)\right)} F(z,u) &= 1 - \frac{z}{u} F_0(z)\,. \end{align} We write $K(z,u):=1-zP(u)$ and call this factor $K(z,u)$ the \emph{kernel\/}. The above functional equation looks like an underdetermined equation as there are two unknown functions, namely $F(z,u)$ and $F_0(z)$. However, the special structure on the left-hand side will resolve this problem and leads us to the \emph{kernel method}. \smallskip Using the theory of Newton polygons and Puiseux expansions (cf.~\cite[Appendix of Sec.~3]{DieuIC}), we know that the \emph{kernel equation} \begin{align} \label{eq:kerneleq} 1-zP(u)=0, \end{align} has $d+1$ distinct solutions in $u$ (recall that $c=1$, see Equation~\eqref{c_d}). One of them, say $u_1(z)$, maps $0$ to $0$. We call this solution the ``small branch'' of the kernel equation. It is in modulus smaller than the other $d$ branches. These in turn grow to infinity in modulus while $z$ approaches~0. Consequently, we call the latter the ``large branches'' and denote them by $v_1(z),v_2(z),\dots,v_d(z)$. Inserting the small branch into~\eqref{eq:genFuncEq} (this is legitimate as we stay in the integral domain of Puiseux power series: substitution of the small branch always leads to series having a finite number of terms with negative exponents, even for intermediate computations), we get $F_0(z) = u_1(z)/z$. This proves our second claim. Using this result, we can solve~\eqref{eq:genFuncEq} for $F(z,u)$ to get the first claim. \end{proof} The formula~\eqref{eqFzuLuka} in the previous theorem implies that the number $m_n$ of meanders of length $n$ is directly related to the number $e_n$ of excursions of length $n$ via \begin{equation}m_n = P(1)^n - \sum_{k=0}^{n-1} P(1)^{k} e_{n-k-1}.\label{meanderformula} \end{equation} In the sequel, we therefore focus on giving explicit expressions for $e_n$. \medskip A key tool for finding a formula for the coefficients of power series satisfying implicit equations is the Lagrange inversion formula \cite{Lagrange70}, independently discovered in a slightly extended form by B\"urmann \cite{Buermann} (see also \cite{LaLeAA}). In the statement of the theorem and also later, we use the \emph{coefficient extractor} $[z^n] F(z) :=f_n$ for a power series $F(z)=\sum f_n z^n $. \begin{theorem}[\sc Lagrange--B\"urmann inversion formula] \label{thm:LagrangeBurmann} Let $F(z)$ be a formal power series which satisfies $F(z)=z \phi(F(z))$, where $\phi(z)$ is a power series with $\phi(0)\neq 0$. Then, for any Laurent\footnote{Here, by Laurent series we mean a series of the form $H(z)=\sum_{n\ge a} H_n\,z^n$ for some (possibly negative) integer~$a$.} series $H(z)$ and for all non-zero integers $n$, we have $$[z^n] H(F(z)) = \frac{1}{n} [z^{n-1}] H'(z) \phi^n(z) \,.$$ \end{theorem} \begin{proof} See~\cite[Chapter~A.6]{FlSe09} or~\cite[Theorem~5.4.2]{stan99}. \end{proof} \pagebreak \begin{table}[ht]\scalebox{0.94}{\hspace{-2mm} \begin{tabular}{|Sc|Sc|Sc|} \hline \begin{tabular}{c} name and the associated \\ step polynomial $P(u)$\end{tabular} & number $e_n$ of excursions of length $n$ \\ \hline \begin{tabular}{c} Dyck paths \\ $P(u)=\frac{1}{u}+u$ \end{tabular} & $\displaystyle{e_{2n}= \frac{1}{n+1} \dbinom{2n}{n}}$\\ \hline \begin{tabular}{c}Motzkin paths \\ $P(u)=\frac{1}{u}+1+u$\end{tabular} & $\displaystyle{e_n=\frac{1}{n+1} \sum_{k=0}^{\lceil \frac{n+1}{2} \rceil} \dbinom{n+1}{k} \dbinom{n+1-k}{k-1}}$\\ \hline \begin{tabular}{c}weighted Motzkin paths \\ $P(u)=\frac{p_{-1}}{u}+p_0+p_1 u$ \end{tabular} & $\displaystyle{e_{n}=\frac{1}{n+1} \sum_{k=0}^{\lceil \frac{n+1}{2} \rceil} \dbinom{n+1}{k} \dbinom{n+1-k}{k-1} (p_1 p_{-1})^{k-1} p_0^{n+2-2k}}$\\ \hline \begin{tabular}{c}bicoloured Motzkin paths\\ $ P(u)=\frac{1}{u}+2+u $\end{tabular} & $\displaystyle{e_{n+1}=\frac{1}{n+1}\dbinom{2n}{n}}$\\ \hline \begin{tabular}{c}{\L}ukasiewicz paths\\ $P(u)=\frac{1}{u}+1+u+u^2+\cdots$ \end{tabular} & $\displaystyle{e_{n}=\frac{1}{n+1}\dbinom{2n}{n}}$\\ \hline \begin{tabular}{c}$d$-ary trees \\$P(u)=\frac{1}{u}+u^{d-1}$\end{tabular}& $e_{dn+1}=\displaystyle{\frac{1}{(d-1)n+1}\dbinom{dn}{n}}$\\ \hline \begin{tabular}{c} $\{1,2,\dots,d\}$-ary trees \\$P(u)=\frac{1}{u}+1+\dots+u^{d-1}$\end{tabular}& $e_{n}=\displaystyle{ \frac{1}{n}\sum_{j=0}^{\lfloor{\frac{n-1}{d+1}}\rfloor} (-1)^j \dbinom{n}{ j}\dbinom{2n-2-j(d+1)}{n-1} \ }$\\ \hline \begin{tabular}{c}$\{d,d+1\}$-ary trees\\$P(u)=\frac{1}{u}+u^{d-1}+u^d$\end{tabular}& $\displaystyle{e_n=\frac{1}{n}\sum_{k=0}^{\lfloor \frac{n-1}{d} \rfloor} \dbinom{n}{k} \dbinom{k}{n-1-dk}}$\\ \hline \end{tabular} } \vskip8pt \caption{Closed-form formulas for some famous families of lattice paths. } \label{TableLukaPaths} \end{table} \vspace{-5mm} Table~\ref{TableLukaPaths} presents several applications of this Lagrange inversion formula to lattice path enumeration. It leads to the Catalan numbers for Dyck paths, and to the Motzkin numbers for the Motzkin paths, i.e., excursions associated with the step set $\mathcal{S} = \{-1,0,+1\}$. They are two of the most ubiquitous number sequences in combinatorics, see~\cite[Ex.~6.19, 6.25, and 6.38]{stan99} for more information. Table~\ref{TableLukaPaths} also contains an example of weighted paths (namely weighted Motzkin paths and the special case of bicoloured Motzkin paths), as well as an example with an infinite set of steps (namely the \L ukasiewicz paths with all possible steps allowed). All of the examples in Table~\ref{TableLukaPaths} are intimately related to families of trees (as suggested by some of the namings in the table). In order to explain this, we recall that an \emph{ordered tree} is a rooted tree for which an ordering of the children is specified for each vertex, and for which its arity (i.e., the outdegree, the number of children of each node) is restricted to be in a subset $\mathcal{A}$ of $\mathbb{N}$.\footnote{In this article, by convention $0\in \mathbb{N}$.} If $\mathcal{A}=\{0,2\}$, this leads to the classical binary trees counted by the Catalan numbers; if $\mathcal{A}=\{0,1,2\}$, this leads to the unary-binary trees counted by Motzkin numbers, and if $\mathcal{A}=\mathbb{N}$, this gives the ordered trees (also called planted plane trees), which are also counted by Catalan numbers. Any ordered tree can be traversed starting from the root in \emph{prefix order}: one starts from the root and proceeds depth-first and left-to-right. The listing of the outdegrees of nodes in prefix order is called the \emph{preorder degree sequence}. This characterizes a tree unambiguously, see Figure~\ref{lukabijection}, and it is best summarized by the following folklore proposition. \begin{proposition}[\sc {\L}ukasiewicz correspondence] Ordered trees are in bijection with {\L}ukasiewicz excursions \end{proposition} \begin{proof} Given an ordered tree with $n$ nodes, the preorder sequence can be interpreted as a lattice path. Let $(\sigma_j)_{j=1}^{n}$ be a preorder degree sequence. With each $\sigma_j$ we associate a step $(1,\sigma_j-1) \in \mathbb{N}\times\mathbb{Z}$. Note that, as the minimal degree is $0$, our smallest step is $-1$. Starting at the origin, we concatenate these steps for $j=1,2,\dots,n-1$, ignoring the last step. In this way, we obtain a {\L}ukasiewicz excursion of length $n-1$. \end{proof} \begin{figure}[!ht] \scalebox{0.88}{ \includegraphics[height=25mm]{lukabijtree} \qquad \qquad \includegraphics[height=25mm]{lukabijlp} } \caption{The bijection between trees and {\L}ukasiewicz paths. The preorder degree sequence $(3,1,0,3,0,0,1,0,1,2,0,0)$ uniquely characterizes the tree, and gives the corresponding {\L}ukasiewicz path with step sequence $(2,0,-1,2,-1,-1,-1,0,-1,0,1,-1,-1)$. Dropping the last $-1$ step yields an excursion. \label{lukabijection}} \end{figure} As one can see, the combinatorics of the {\L}ukasiewicz paths is well understood (see e.g.~\cite{FlSe09,Stanley86}), and the true challenge is to analyse lattice paths with other negative steps than just $-1$. The smallest non-{\L}ukasiewicz cases are the Duchon lattice paths (steps $\mathcal{S}=\{-2,+3\}$), and the Knuth lattice paths (steps $\mathcal{S}=\{-2,+5\}$). Their enumerative and asymptotic properties are the subject of another article in this volume~\cite{BaWa16}. For these two families of lattice paths, the asymptotics are tricky, because the generating functions involve several dominant singularities. In the next sections, we concentrate on closed formulas which appear for many other non-{\L}ukasiewicz cases. \section{(Old-time) Basketball walks: steps \texorpdfstring{$\mathcal{S}=\{-2,-1,+1,+2\}$}{S=\{-2,-1,+1,+2\}}}\label{Section3} \begin{figure}[h!]\vspace{-3mm} \includegraphics[height=45mm]{James_Naismith}\vspace{-1mm} \caption{Since its creation in 1892 by James Naismith (November 6, 1861 -- November 28, 1939), the rules of basketball evolved. For example, since 1896, field goals and free throws were counted as two and one points, respectively. The international rules were changed in 1984 so that a ``far'' field goal was now rewarded by $3$ points, while ``ordinary'' field goals remained at $2$ points, a free throw still being worth one point.} \label{fig:BB} \end{figure} We now turn our attention to a class of lattice paths (lattice walks) with rich combinatorial properties: \emph{the basketball walks}. They are constructed from the step set $\mathcal{S}=\{-2, -1, +1,+2\}$. This terminology was introduced by Arvind Ayyer and Doron Zeilberger~\cite{ZeilbergerAyyer07}, and these walks were later also considered by Mireille Bousquet-M\'elou~\cite{BousquetMelou08}. They can be seen as the evolution of the score during a(n old-time) basketball game (see Figure~\ref{fig:BB}). Ayyer, Zeilberger, and Bousquet-M\'elou found interesting results on the shape of the algebraic equations satisfied by the excursion generating function, and similar properties when the height of the excursion is bounded. In this article, we analyse a generalization in which the starting point and the end point of the walks do not necessarily have altitude~$0$. Since, in that case, we lose a natural factorization happening for excursions, we are led to variations of certain parts in the kernel method. In addition, we are interested in closed-form expressions for the number of walks of length $n$. This is complementary to the results in~\cite{BaFl02} and in~\cite{BaWa16}. Moreover, contrary to the previous section, these walks are not {\L}ukasiewicz paths any more. This makes them harder to analyse (the easy bijection with trees is lost, for example). Despite all that, the kernel method will strike again, thus illustrating our main motto: $$ \text{\it ``The kernel method is the method of choice for problems on directed lattice paths!''} $$ \subsection{Generating functions for positive (old-time) basketball walks: the kernel method} We define \emph{positive walks} as walks staying strictly above the $x$-axis, possibly touching it at the first or last step. Returning to the basketball interpretation, these correspond to the evolution of basketball scores where one team (the stronger team, the richer team?) is always ahead of the other team. Let $G_{j,n,k}$ be the number of such walks running from $(0,j)$ to $(n,k)$, and define by $G_j(z,u)$ the generating function of positive walks starting at $(0,j)$. We write \begin{equation} G_j(z,u) := \sum_{n,k \geq 0}G_{j,n,k}z^nu^k = \sum_{n=0}^\infty g_{j,n}(u)z^n = \sum_{k=0}^\infty G_{j,k}(z)u^k. \end{equation} Similar to Section~\ref{Section2}, we shall need the polynomial $g_{j,n}(u)$, the generating function for all walks with $n$ steps, and the series $G_{j,k}(z)$, the generating function for all walks ending at altitude $k$. The bivariate generating function $G_j(z,u)$ is analytic for $|z|<1/P(1)$ and $|u|\le1$. A walk is either the single initial point at altitude $j$, or a walk followed by a step not reaching altitude~0 or below. This leads to the functional equation \begin{equation} (1-zP(u))G_j(z,u)=u^j-z\big(G_{j,1}(z)+G_{j,2}(z)+G_{j,1}(z)/u\big), \quad \quad j>0, \label{fund.eq} \end{equation} where the \textit{step polynomial\/} $P(u)$ is given by \begin{equation*} P(u):=u^{-2}+u^{-1}+u+u^2. \end{equation*} Again, we call the factor $1-zP(u)$ on the left-hand side of \eqref{fund.eq} the \textit{kernel\/} of the equation, and denote it by $K(z,u)$. We refer to~\eqref{fund.eq} as the \textit{fundamental functional equation} for $G_j(z,u)$. The equation has a small problem though: this is {\it one} equation with {\it three} unknowns, namely $G_j(z,u), G_{j,1}(z)$, and $G_{j,2}(z)$! The idea of the so-called {\it`kernel method'} is to equate the kernel $K(z,u)$ to $0$, thus binding $u$ and $z$ in such a way that the left-hand side of~\eqref{fund.eq} vanishes. This produces two extra equations. To equate $K(z,u)$ to zero means to put \begin{equation} \label{eq:kernel.equation} 1-zP(u)=0 \quad \text{ or equivalently } \quad u^2-zu^2P(u)=0. \end{equation} We call this equation the \textit{kernel equation}. As an equation of degree~4 in $u$, it has four roots. We call the two small roots (that is, the roots which tend to~0 when $z$ approaches~0) $u_1(z)$ and $u_2(z)$. Then, on the complex plane slit along the negative real axis, we can identify the small roots $u_1(z)$ and $u_2(z)$ as \begin{align*} u_1(z) &= -\frac{1}{4} \left( \frac{z - \sqrt{4z+9z^2}}{z} + \sqrt{\frac{4 - 6z - 2\sqrt{4z+9z^2}}{z}} \right)\\ &=\sqrt{z}+\frac{1}{2}z+\frac{1}{8}z^{3/2}+\frac{1}{2}z^2+\frac{159}{128}z^{5/2}+O(z^3),\\ u_2(z) &= -\frac{1}{4} \left( \frac{z + \sqrt{4z+9z^2}}{z} - \sqrt{\frac{4 - 6z + 2\sqrt{4z+9z^2}}{z}} \right)\\ &=-\sqrt{z}+\frac{1}{2}z-\frac{1}{8}z^{3/2}+\frac{1}{2}z^2-\frac{159}{128}z^{5/2}+O(z^3). \end{align*} Moreover, their Puiseux expansions are related via the following proposition. \begin{proposition}[\sc Conjugation principle for two small roots]\label{prop:conjugation} The small roots $u_1$ and $u_2$ of\/ $1-zP(u)=0$ satisfy \begin{equation} u_1(z)=\sum_{n\geq 1} a_n z^{n/2} \text{ and } u_2(z) =\sum_{n\geq 1} (-1)^n a_n z^{n/2}\,. \end{equation} \end{proposition} \begin{proof} The kernel equation yields $$ u=X(1+u+u^3+u^4)^{1/2}, $$ with $X=z^{1/2}$ or $X=-z^{1/2}$. Since the above equation possesses a unique formal power series solution $u(X)$, the claim follows. \end{proof} By substituting the small roots $u_1(z)$ and $u_2(z)$ of the kernel equation \eqref{eq:kernel.equation} into the fundamental functional equation \eqref{fund.eq}, we see that the left-hand side vanishes. Subsequently, we solve for $G_{j,1}(z)$ and $G_{j,2}(z)$ and get\footnote{In this article, whenever we thought it could ease the reading, without harming the understanding, we write $u_1$ for $u_1(z)$, or $F$ for $F(z)$, etc.} \begin{align} G_{j,1}(z)&=-\dfrac{u_1u_2(u_1^j-u_2^j)}{z(u_1-u_2)}, & j>0, \label{eq:G1}\\ G_{j,2}(z)&=\dfrac{u_1u_2(u_1^j-u_2^j) + u_1^{j+1} - u_2^{j+1}}{z(u_1-u_2)}, & j>0. \label{eq:G2} \intertext{Substitution in the fundamental functional equation \eqref{fund.eq} then yields} G_j(z,u) &= \frac{u^j-z(G_{j,1}(z)+G_{j,2}(z)+G_{j,1}(z)/u) }{1-zP(u)}, & j>0. \label{eq:G} \end{align} By means of the kernel method, we have thus derived an explicit expression for the bivariate generating function $G_j(z,u)$ for walks starting at altitude $j>0$. In the following proposition, we summarize our findings so far. In addition, we express the generating function for walks from altitude~$j$ to altitude~$k$ (with $j,k>0$) explicitly in terms of the small roots $u_1(z)$ and $u_2(z)$, and we also cover the special case $j=0$, which offers some nice simplifications. \begin{proposition} \label{prop:G0k} As before, let $G_{j,k}(z)$ be the generating function for positive basketball walks with steps $-2,-1,+1,+2$ starting at altitude $j$ and ending at altitude $k$. Furthermore, let $u_1(z)$ and $u_2(z)$ be the small roots of the kernel equation $1-zP(u)=0$, with $P(u)=u^{-2}+u^{-1}+u+u^2$. Then, for $j,k > 0$, we have \begin{align} G_{0,k}(z) &= \frac{u_1^{k+1}(z)-u_2^{k+1}(z)}{u_1(z)-u_2(z)} , \label{eq:G0k} \\ G_{j,k}(z) &= -\frac{u_1(z) u_2(z)}{z} \sum_{i=0}^j \frac{u_1^{j-i+1}(z)-u_2^{j-i+1}(z)}{u_1(z)-u_2(z)} \frac{u_1^{k-i+1}(z)-u_2^{k-i+1}(z)}{u_1(z)-u_2(z)}, \label{eq:Gjkhi} \end{align} \end{proposition} \begin{proof} We start with the proof of \eqref{eq:G0k}. The first step of a walk can only be a step of size $+1$ or $+2$. Thus, removing this first step and shifting the origin, we have \begin{equation} G_{0,k}(z)=z \left( G_{1,k}(z)+G_{2,k}(z) \right), \end{equation} where $G_{1,k}(z)$ and $G_{2,k}(z)$ are the generating functions for positive walks running from altitude $1$ to altitude $k$, respectively from altitude $2$ to altitude $k$. This decomposition is illustrated in Figure~\ref{G01walksinitialsteps}. \begin{figure}[!hb] \includegraphics[width=0.47\textwidth]{G01walk1jump} \quad \includegraphics[width=0.47\textwidth]{G01walk2jump} \caption{Two different instances of walks counted by $G_{0,1}(z)$ showing the two possible first steps $+1$ and $+2$. \label{G01walksinitialsteps}} \end{figure} By ``time reversal'' (due to the symmetry of our step set, i.e., $P(u)=P(u^{-1})$), we also have \begin{equation} G_{1,k}(z)=G_{k,1}(z),\quad\text{ and }\quad G_{2,k}(z)=G_{k,2}(z), \end{equation} where $G_{k,1}(z)$ and $G_{k,2}(z)$ are known from Equations~\eqref{eq:G1} and~\eqref{eq:G2}. Now notice that \begin{align*} G_{k,2}(z) &= \dfrac{u_1u_2(u_1^k-u_2^k) + u_1^{k+1} - u_2^{k+1}}{z(u_1-u_2)} =\dfrac{u_1u_2(u_1^k-u_2^k)}{z(u_1-u_2)}+\dfrac{u_1^{k+1} - u_2^{k+1}}{z(u_1-u_2)} \\ &=\frac{u_1^{k+1}-u_2^{k+1}}{z(u_1-u_2)}-G_{k,1}(z). \end{align*} This leads directly to~\eqref{eq:G0k}. \smallskip For computing $G_{j,k}(z)$ with $j,k > 0$, we use a first passage decomposition with respect to minimal altitude of the walk. Combining \eqref{eq:G0k} with time reversal, we see that $h_m(z) := \frac{u_1^{m+1}-u_2^{m+1}}{u_1-u_2}$ is the generating function for basketball walks starting at altitude $m$, staying always above the $x$-axis, but ending on the $x$-axis. Furthermore, by \eqref{eq:G1} with $j=1$, the series $E(z) = -\frac{u_1 u_2}{z}$ is the generating function for excursions (allowed to touch the $x$-axis). Then the walks from altitude~$j$ to altitude~$k$ can be decomposed into three sets, as illustrated by Figure~\ref{Gjkdecomp}: \begin{enumerate} \item The walk starts at altitude $j$, and continues until it hits for the first time altitude~$i$ (the lowest altitude of the walk, so $1 \leq i \leq j$). This part is counted by $h_{j-i}(z)$. \item The second part is the one from that point to the last time reaching altitude~$i$. In other words, this part is an excursion on level $i$ counted by $E(z)$. \item The last part runs from altitude $i$ to altitude $j$ without ever returning to altitude~$i$. By time reversal one sees that this is counted by $h_{k-i}(z)$. \end{enumerate} Summing over all possible $i$'s, we get~\eqref{eq:Gjkhi}. \end{proof} \begin{figure}[!hb] \includegraphics[width=0.87\textwidth]{Gjkdecomp} \caption{The decomposition for $G_{j,k}$} \label{Gjkdecomp} \end{figure} There is an alternative expression for the generating function $G_{j,k}(z)$, which we present in the next proposition. \begin{proposition}[\sc Formula for walks from altitude $j$ to altitude $k$] \label{prop:Gjk} Let $u_1(z)$ and $u_2(z)$ be the small roots of the kernel equation $1-zP(u)=0$, with $P(u)=u^{-2}+u^{-1}+u+u^2$, and let $G_{j,k}(z)$ be the generating function for positive basketball walks starting at altitude $j$ and ending at altitude $k$. Then \begin{equation} \label{eq:uW} G_{j,k}(z)=W_{j-k} + h_j(u_1,u_2) W_{-k} + u_1 u_2 h_{j-1}(u_1,u_2) W_{-k+1}, \end{equation} where $$W_i(z) = z \left( \frac{u_1'}{u_1^{i+1}} + \frac{u_2'}{u_2^{i+1}} \right)$$ is the generating function of unconstrained walks starting at the origin and ending at altitude $i$, and $$h_i(x_1, x_2) = \frac{x_1^{i+1}-x_2^{i+1}}{x_1-x_2}$$ is the complete homogeneous symmetric polynomial of degree $i$ in $x_1$ and $x_2$. \end{proposition} \begin{proof} Since $G_{j,k}(z) = G_{k,j}(z)$, without loss of generality we may assume that $j\leq k$. We start with~\eqref{eq:G}. Extraction of the coefficient of $u^k$ on the left-hand side gives $G_{j,k}(z)$. As coefficient extraction is linear, we need to find expressions for \begin{align*} [u^i] \frac{1}{1-zP(u)}. \end{align*} By~\eqref{eq:unconstrainedpaths}, these are the generating functions $W_i(z)$ for unconstrained walks starting at the origin and ending at altitude $i$. For basketball walks, we have $P(u)=P(u^{-1})$, hence $W_i(z) = W_{-i}(z)$. Using a straightforward contour integral argument, using Cauchy's integral formula and the residue theorem, we have $$ W_i(z)= [u^i] \frac{1}{1-zP(u)}=\frac {1} {2\pi \sqrt{-1}}\int_{\mathcal C} \frac{du}{u^{i+1}(1-zP(u))} =z \left( \frac{u_1'(z)}{u_1^{i+1}(z)} + \frac{u_2'(z)}{u_2^{i+1}(z)} \right). $$ Thus, we obtain the claimed expression for $W_i(z)$ in terms of the small branches. Finally, the remaining factors in \eqref{eq:uW} are obtained by simplifications in~\eqref{eq:G}. \end{proof} Thus, by~\eqref{eq:G0k}, walks starting at the origin are given by complete homogeneous symmetric polynomials in the small branches. In particular, we have \begin{align} G_{0,1}(z) &=u_1(z)+u_2(z),\label{eq:GGGG01} \\ G_{0,2}(z) &=u_1^2(z)+u_1(z)u_2(z)+u_2^2(z).\label{eq:GG02} \end{align} We now derive an explicit expression for $G_{0,1}(z)$ and $G_{0,2}(z)$. Note that, as~\eqref{eq:GGGG01} is not defined on the negative real axis, we apply analytic continuation in order to derive an expression which is defined for every $|z|<\frac{1}{4}$, which is the radius of convergence of $G_{0,1}(z)$. The function $G_{0,1}(z)$ is an algebraic function since it is the sum of two algebraic functions (namely, $u_1(z)$ and $u_2(z)$). Using a computer algebra package, it is easy to derive an algebraic equation for $G_{0,1}(z)$. For example, the following {\sl Maple} commands (see~\cite{SalvyZimmermann94} for more on these aspects) gives the desired equation: \medskip \begin{maplegroup} \begin{mapleinput} \mapleinline{active}{1d}{AllRoots:=allvalues(solve(1-z*P(u),u)):}{}\vspace{0.3\baselineskip} \mapleinline{active}{1d}{u1:=AllRoots[2]: u2:=AllRoots[3]:}{}\vspace{0.3\baselineskip} \mapleinline{active}{1d}{algeq:=algfuntoalgeq(u1+u2,u(z));}{}\vspace{0.3\baselineskip} \end{mapleinput} \end{maplegroup} \vspace{-1.3\baselineskip} \begin{align} z u^4+2z u^3 +(3z-1)u^2+(2z-1)u+z. \label{eq:G01eqalg} \end{align} \noindent In particular, $G_{0,1}(z)$ is uniquely determined by the previous equation and the fact that its expansion at $z=0$ is a power series with non-negative coefficients. Solving this equation, we arrive at an analytic expression for $G_{0,1}(z)$ for $|z|<1/4$: \begin{align} \label{eq:G01analytic} G_{0,1}(z)&= -\frac{1}{2} + \frac{1}{2} \sqrt{\frac{2 - 3z - 2 \sqrt{1 - 4z}}{z}}\\ &=z + z^2 + 3 z^3 + 7 z^4 + 22 z^5 + 65 z^6 + 213 z^7 + \cdots. \end{align} \medskip Using a computer algebra package again, we find that $G_{0,2}(z)$ satisfies \begin{align} \label{eq:G02eq} z^3 u^4 -3 z^2 u^3 - (z^2-3z)u^2+(z-1)u+z=0. \end{align} Among its four branches, only one is a power series at $z=0$ with non-negative coefficients, namely \begin{align} \label{eq:G02analytic} G_{0,2}(z)&=\frac{3-\sqrt{1-4 z}-\sqrt{2 + 12 z + 2\sqrt{1-4z}}}{4 z}\\ &= z+z^2+4z^3+9z^4+31z^5+ 91z^6+ 309z^7+\cdots. \end{align} \smallskip In order to undertake a small digression on complexity of computation: these explicit forms are not the fastest way to access the coefficients. A better way is to take advantage of the theory of holonomic functions (as, e.g., implemented in the {\sl gfun} {\sl Maple} package, see~\cite{SalvyZimmermann94}). To begin with, the kernel method gave us an algebraic equation. Applying the derivative to both sides of this equation and using the obtained new relations, we are led to a linear differential equation satisfied by the function $G(z)$ (where we write $G(z)$ instead of $G_{0,1}(z)$ for short): \begin{mapleinput}\mapleinline{active}{1d}{diffeq:=algeqtodiffeq(subs(u=G,algeq),G(z),{G(0)=0}):}{}\end{mapleinput} $$\begin{cases} G(0)=0,\\ \vspace{0.9\baselineskip} 6\,z+6+12\, \left( z+1 \right) G \left( z \right) +2\, \left( 162\,{z}^{3}+66\,{z}^{2}+z-3 \right) {\frac {d}{dz}}G \left( z \right) \\ \vspace{0.6\baselineskip} +z \left( 9\,z+4 \right) \left( 4\,z-1 \right) \left( 6\,z+1 \right) { \frac {d^{2}}{d{z}^{2}}}G \left( z \right) =0 \label{eq:G01diffeq} \end{cases}$$ \noindent Then, extraction of $[z^n]$ on both sides of the differential equation yields a linear recurrence satisfied by the coefficients $g(n)$ of $G$, namely \begin{mapleinput}\mapleinline{active}{1d}{rec:=diffeqtorec(diffeq,G(z),g(n)):}{}\end{mapleinput} $$\begin{cases} g(0) = 0, g(1) = 1, g(2) = 1\,,\\ \vspace{0.6\baselineskip} 0=108\,n \left( 2\,n+1 \right) g \left( n \right) +6\, \left( 13\,{n}^{2}+35\,n+24 \right) g \left( n+1 \right) \\ \vspace{0.3\baselineskip} - \left( 17\,{n}^{2}+49\,n+18 \right) g \left( n+2 \right) -2\, \left( 2\,n+7 \right) \left( n+3 \right) g \left( n+3 \right) \,. \label{eq:G01rec} \end{cases}$$ \smallskip \noindent From this recurrence, a binary splitting approach introduced by the Chudnovskys gives a procedure which surprisingly computes $g(n)$ in only $O(\sqrt{n})$ operations (and $O(n \ln n \ln(\ln n ))$ bit complexity): \begin{mapleinput} \mapleinline{active}{1d}{g:=rectoproc(rec,g(n)):}{}\vspace{0.3\baselineskip} \mapleinline{active}{1d}{g(10^5): #a 6014-digits number computed in only 2 seconds!}{} \end{mapleinput} \medskip The same approach applies to all our directed lattice path models. This approach is much faster than the naive approach by means of dynamic programming (which would compute the bivariate generating function, and would then extract the desired $G(z)$ from it: this would cost $O(n^2)$ in time and $O(n^3)$ in memory). \medskip We just saw how to efficiently compute $g(n)$, for any given value of $n$, but is there a closed-form formula holding for all $n$ at once? We now further investigate this question. \subsection{How to get a closed form for coefficients: Lagrange--B\"urmann inversion} In Section~\ref{Section4}, we present a closed form for the numbers of lattice walks with step polynomial $P(u)=u^{-h}+u^{-h+1}+\dots+u^{h-1}+u^h$, for any $h$. In the case $h=2$ that we are dealing with in the current section, a nice miracle occurs: a more ad hoc approach allows one to derive simpler expressions. \subsubsection{Closed form for coefficients of $G_{0,1}(z)$} The generating function $G_{0,1}(z)$ of walks starting at the origin, ending at altitude~$1$, and never touching the $x$-axis, satisfies the algebraic equation \eqref{eq:G01eqalg}. We rewrite it in the form \begin{align*} G_{0,1}(z) + G_{0,1}^2(z) = z(1 + G_{0,1}(z) + G_{0,1}^2(z))^2. \end{align*} Here, substitution of $G_{0,1}(z) + G_{0,1}^2(z)$ by $C(z)-1$ gives the striking equation \begin{align} \label{eq:LagrangeG01C} 1+ G_{0,1}(z)+G_{0,1}^2(z) &= C(z) , \end{align} where $C(z)=1+zC(z)^2$ is the generating function for Catalan numbers. A recursive bijection for this identity was found by Axel Bacher and (independently) by J\'er\'emie Bettinelli and \'Eric Fusy (personal communication, see also~\cite{BettinelliFusy16}). It remains a challenge to find a more direct simple bijection. This identity is the key to get nice closed-form expressions for the coefficients, via the following variant of Lagrange inversion. \begin{lemma}[\sc Lagrange--B\"urmann inversion variant] \label{lemma:LagrangeComp} Let $F(z)$ and $H(z)$ be two formal power series satisfying the equations \begin{align} F(z) &= z \phi(F(z)), & H(z) &= z \psi(H(z)), \end{align} where $\phi(z)$ and $\psi(z)$ are formal power series such that $\phi(0) \neq 0$ and $\psi(0) \neq 0$. Then, \begin{align}\label{LagrangeComp} [z^n]H(F(z)) &= \frac{1}{n} \sum_{k=1}^{n} \left([z^{k-1}] \psi^{k}(z)\right)\left( [z^{n-k}] \phi^n(z)\right). \end{align} \end{lemma} \begin{proof} By the Lagrange--B\"urmann inversion (Theorem~\ref{thm:LagrangeBurmann}), we have $$ [z^n]H(F(z)) = \frac{1}{n} [z^{n-1}] H'(z) \phi^n(z). $$ Now we apply the Cauchy product formula $[z^m] A(z) B(z) = \sum_{k=0}^m a_k b_{m-k}$ with $m=n-1$, $A(z)=H'(z)$, and $B(z)= \phi^n(z)$. This leads to \begin{align*} [z^n]H(F(z)) &= \frac{1}{n} \sum_{k=0}^{n-1} \left([z^k] H'(z)\right)\left( [z^{n-1-k}] \phi^n(z)\right) \\ &= \frac{1}{n} \sum_{k=1}^{n} \left([z^{k-1}] H'(z)\right)\left( [z^{n-k}] \phi^n(z) \right). \end{align*} This gives Formula~\eqref{LagrangeComp}, after observing $ [z^{k-1}]H'(z) = k [z^{k}] H(z) = [z^{k-1}] \psi^{k}(z)\,, $ where we used Lagrange--B\"urmann inversion again. \end{proof} \begin{proposition}\label{prop:G011} The number of basketball walks of length $n$ from the origin to altitude~$1$ with steps in $\mathcal{S} =\{-2,-1,+1,+2\}$ and never returning to the $x$-axis equals \begin{equation}\label{eq:G011} \frac{1}{n} \sum_{k=1}^{n} (-1)^{k+1} \dbinom{2k-2}{k-1} \dbinom{2n}{n-k}\\ = \frac{1}{n}\sum_{i=0}^{n}\binom{n}{i}\binom{n}{2n+1-3i}\,. \end{equation} \end{proposition} \begin{proof} Equation~\eqref{eq:LagrangeG01C} implies that $ G_{0,1}(z) = H(C(z)-1), $ where $H(z)$ is the functional inverse of the polynomial $x^2+x$. Thus $H(z) = z \psi(H(z))$, with $\psi(z) = \frac{1}{1+z}$. Furthermore, it is well-known that $C_0(z):=C(z)-1$ satisfies $C_0(z) = z \phi(C_0(z))$ with $\phi(z) = (1+z)^2$. Hence, Equation~\eqref{LagrangeComp} yields \begin{align} [z^n] G_{0,1}(z) &= \frac{1}{n} \sum_{k=1}^{n} \left([z^{k-1}] \frac{1}{(1+z)^k}\right)\left( [z^{n-k}] (1+z)^{2n}\right)\\ &= \frac{1}{n} \sum_{k=1}^{n} (-1)^{k+1} \dbinom{2k-2}{k-1} \dbinom{2n}{n-k}\,. \end{align} The alternative expression without the $(-1)^{k+1}$ factors comes from Formula~\eqref{eq:GGGG01}, to which we apply the Lagrange--B\"urmann inversion formula for $u_1$, remembering that $u_1$ satisfies $u^2 = zu^2 P(u)$, and that the conjugation property of the small roots from Proposition~\ref{prop:conjugation} holds: \begin{align} [z^n] G_{0,1}(z)&=[z^n] (u_1(z)+u_2(z)) =2[z^{n}] u_1(z) =\frac{1}{n}\sum_{k=0}^{n}\binom{n}{k}\binom{n}{2n+1-3k}\,. \qedhere \end{align} \end{proof} The last closed-form expression can also be explained via the so-called cycle lemma (cf.\ \cite[Ex.~5.3.8]{stan99}). Namely, by~\eqref{eq:unconstrainedpaths} combined with the factorization $u^{-2}+u^{-1}+u+u^2=u^{-2}(1+u^3)(1+u)$, the number of unrestricted walks from $0$ to $1$ in $n$ steps is given by \begin{align} [u^1z^n] W(z,u) = [u^1] P(u)^n = [u^1] \left(\frac{(1 + u^3) (1 + u) }{u^2} \right)^n =\sum_{i=0}^n\binom{n}{i}\binom{n}{2n+1-3i}. \end{align} \begin{figure}[!hb] \centering \includegraphics[scale=1.05]{G01walk} \qquad \includegraphics[scale=1.05]{W01walk} \caption{Transforming a walk counted by $G_{0,1}(z)$ into a walk counted by $W_{0,1}(z)$. } \label{fig:G01cycle} \end{figure} From the formulas, we see that $[z^n]G_{0,1}(z)=\frac{1}{n}[z^n]W_{0,1}(z)$. There exists indeed a $1$-to-$n$ correspondence between walks counted by $G_{0,1}(z)$ and those counted by $W_{0,1}(z)$. For each walk $\omega$ counted by $G_{0,1}(z)$, decompose $\omega$ into $\omega=\omega_{\ell}B\omega_{r}$ where $B$ is any point in the walk. A new walk $\omega'$ counted by $W_{0,1}(z)$ is constructed by putting $B$ at the origin and adjoining $\omega_\ell$ at the end of $\omega_r$, i.e., $\omega'=B\omega_r\omega_\ell$, see Figure~\ref{fig:G01cycle}. If $\omega$ is of length $n$, then there are $n$ choices for $B$. All these walks are different because there are no walks from altitude $0$ to altitude $1$ which are the concatenation of several copies of one and the same walk. (This is not true for walks from altitude~$0$ to altitude~$2$. For example, the walk $(0,2,1,3,2)$ is the concatenation of two copies of the walk $(0,2,1)$.) Conversely, given a walk $\tau$ of length $n$ counted by $W_{0,1}(z)$, we decompose $\tau$ into $\tau=\tau_{\ell}B\tau_{r},$ where $B$ is the right-most minimum of $\tau$. Then, $\tau'=B\tau_{r}\tau_{\ell}$ is a walk of length $n$ counted by $G_{0,1}(z)$. \subsubsection{Closed form for the coefficients of $G_{0,2}(z)$} Recall that, by means of the kernel method, we derived a closed form expression for the generating function $G_{0,2}(z)$ in \eqref{eq:G02analytic}. \begin{proposition}\label{prop:G012} The number of basketball walks of length $n$ from the origin to altitude~$2$ with steps in $\mathcal{S} =\{-2,-1,+1,+2\}$ and never returning to the $x$-axis equals \begin{equation} \frac{1}{2n+1} \sum_{k=0}^{n+1} (-1)^{n+k+1} \dbinom{2n+1}{n+k} \dbinom{n+2k-1}{k}\,. \label{eq:G0222} \end{equation} \end{proposition} \begin{proof} We define the series $F(z)$ by \begin{equation} \label{eq:G02F} -\frac {1} {F(z)}=G_{0,2}(z)-\frac {1} {z}. \end{equation} It is straightforward to see from this equation that $F(z)=z+z^3+\cdots$. The equation~\eqref{eq:G02eq} translates into the equation \begin{equation} \label{eq:Feq} (F^3(z)-zF(z))(1+F(z))+z^2=0 \end{equation} for $F(z)$. We may rewrite this equation in the form $$ \left(F^2-\frac {z} {2}\right)^2 = \frac {z^2} {4}\cdot\frac {1-3F(z)} {1+F(z)}. $$ Next we take the square root on both sides. In order to decide the sign, we have to observe that $F^2(z)=z^2+\cdots$, hence $$ F^2(z)-\frac {z} {2} = -\frac {z} {2}\sqrt{\frac {1-3F(z)} {1+F(z)}}, $$ or, equivalently, $F(z)$ satisfies $F^2(z)=zB(F(z))$, where $$ B(z)=\frac {1} {2}\left(1 -\sqrt{\frac {1-3z} {1+z}}\right). $$ It is straightforward to verify that $B(z)$ satisfies the equation $B(z) = z A(B(z))$ with $A(z) = \frac{1}{1-z} - z$, and it is the only power series solution of this equation. Hence, for $n\ge1$, by \eqref{eq:G02F}, Lagrange--B\"urmann inversion (Theorem~\ref{thm:LagrangeBurmann}) with $H(z)=z^{-1}$, we have \begin{equation*} [z^n] G_{0,2}(z) = -[z^n]\frac {1} {F(z)} =\frac {1} {n}[z^{n-1}] z^{-2}\left(\frac {B(z)} {z}\right)^n =\frac {1} {n}[z^{2n+1}] B^n(z). \end{equation*} Now we apply Lagrange--B\"urmann inversion again, this time with $F(z)$ replaced by $B(z)$, $n$ replaced by $2n+1$, and $H(z)=z^n$. This yields \begin{align} [z^n] G_{0,2}(z) &= \frac{1}{n(2n+1)} [z^{2n}] n z^{n-1} A^{2n+1}(z) \\ &= \frac{1}{2n+1} [z^{n+1}] \left(\frac{1}{1-z} - z\right)^{2n+1}\,. \end{align} By applying the binomial theorem, we then obtain \begin{align} [z^n] G_{0,2}(z) &= \frac{1}{2n+1} [z^{n+1}] \sum_{k=0}^{2n+1} (-1)^{k+1} \dbinom{2n+1}{k} z^{2n+1-k} \left(\frac{1}{1-z}\right)^{k}. \end{align} Since \begin{align} \left(\frac{1}{1-z}\right)^{k} &= \sum_{\ell \geq 0} \dbinom{k+\ell-1}{\ell}z^{\ell}\,, \end{align} we get \begin{align} [z^n] G_{0,2}(z) &= \frac{1}{2n+1} [z^{n+1}] \sum_{\ell \geq 0} \sum_{k=0}^{2n+1} (-1)^{k+1} \dbinom{2n+1}{k} \dbinom{k+\ell-1}{\ell} z^{2n+1-k+\ell}\\ &= \frac{1}{2n+1} \sum_{k=n}^{2n+1} (-1)^{k+1} \dbinom{2n+1}{k} \dbinom{2k-n-1}{k-n}\\ &= \frac{1}{2n+1} \sum_{k=0}^{n+1} (-1)^{n+k+1} \dbinom{2n+1}{n+k} \dbinom{n+2k-1}{k}\,,\label{eq:G02} \end{align} as desired. \end{proof} The idea of the above proof was to ``build up'' a chain of dependencies between the actual series of interest, $G_{0,2}(z)$, and several auxiliary series, namely the series $F(z)$, $B(z)$, and $A(z)$, so that repeated application of Lagrange--B\"urmann inversion could be applied to provide an explicit expression for the coefficients of the series of interest. This raises the question whether this example is just a coincidence, or whether there exists a general method to transform a power series into a Laurent series with the same positive part, and a ``nice'' algebraic expression, allowing multiple Lagrange--B\"urmann inversions to get ``nice'' closed forms for the coefficients. We have no answer to this question and therefore leave this to future research. \subsubsection{Closed form for the coefficients of basketball excursions} Here, we enumerate basketball {\it excursions}, that is, basketball walks which start at the origin, return to altitude~$0$, and in between do not pass below the $x$-axis. A main difference to the previously considered {\it positive} basketball walks is that the excursions are allowed to touch the $x$-axis anywhere. \begin{proposition}[\sc Enumeration of basketball excursions] \label{prop:baskexc} The number of basketball walks with steps in $\mathcal{S} =\{-2,-1,+1,+2\}$ of length $n$ from the origin to altitude $0$ never passing below the $x$-axis is \begin{equation} e_n := \frac{1}{n+1}\sum_{k=0}^{n}(-1)^{n+k}\binom{2n+2}{n-k}\binom{n+2k+1}{k} =\frac{1}{n+1}\sum_{i=0}^{\lfloor n/2 \rfloor}\binom{2n+2}{i}\binom{n-i-1}{n-2i}. \label{eq2} \end{equation} \end{proposition} \begin{remark} The first few values of the sequence defined by \eqref{eq2} are $$ 1, 0, 2, 2, 11, 24, 93, 272, 971, 3194, 11293, 39148, 139687, 497756,\dots$$ \end{remark} \begin{proof}[Proof of Proposition \ref{prop:baskexc}] By the kernel method, we know that the generating function for excursions, $E(z)$ say, is given by $E(z)=-\frac{u_1u_2}{z}$, and that it satisfies the algebraic equation \begin{align} &z^4 E^4-(2z^3+z^2)E^3+(3z^2+2z)E^2-(2z+1)E+1=0\, \label{eq:poly.for.excursion} \end{align} Among the branches of this algebraic equation, only one has a power series expansion. The equation may be rewritten in the form $$ zE(z)= z\left(\frac 1{(1-zE(z))^2} -\frac{ 2zE(z)}{1-zE(z)} + z^2E^2(z)\right) = z\left(\frac 1{1-zE(z)} - zE(z)\right)^2. $$ This shows that we may apply Lagrange--B\"urmann inversion (Theorem~\ref{thm:LagrangeBurmann}) with $\phi(z)= (\frac{1}{1-z}-z)^2$. So we have \begin{align} [z^n]E(z)& =\frac{1}{n+1}[z^{n}]\phi^{n+1}(z) =\frac{1}{n+1}[z^{n}]\left(\frac{1}{1-z}-z\right)^{2n+2}\\ &=\frac{1}{n+1}\sum_{k=0}^{n}(-1)^{n+k}\binom{2n+2}{n-k}\binom{n+2k+1}{k}. \end{align} It is possible to get an expression involving only positive summands by making use of the rewriting $\phi(z)=(1+\frac{z^2}{1-z})^2$. This leads to~\eqref{eq2}. \end{proof} The trick used in this proof can in fact be translated into an algorithm of wider use: \leavevmode \begin{center} \fbox{\begin{minipage}{0.9\textwidth} {\tt The ``Lagrangean scheme'' algorithm }\medskip {\tt input:} an algebraic power series (given in terms of its algebraic equation $P(z,F)=0$, plus the first terms of the expansion of $F$, so that we can uniquely identify the correct branch of the equation) \medskip {\tt output:} a ``Lagrangean equation'' satisfied by $F$\newline (i.e., $H(z^aF) = z \phi(z^aF)$, where $z^a F$ has valuation\footnote{The valuation of a power series $\sum_{n\ge0}f_nz^n$ is the least~$n$ such that $f_n\ne0$.}~1.)\medskip {\tt way to process:} if we assume that $H=H_1/H_2$ and $\phi=\phi_1/\phi_2$ are rational functions, then we identify them via an indeterminate coefficient approach, by substituting the {\em{polynomials}} $H_1, H_2, \phi_1, \phi_2$ in the equation $P(z,F)=0$. \end{minipage}} \end{center} \medskip This algorithm therefore provides a way to get multiple-binomial-sum representations. See~\cite{Egorychev84,Xin04,BostanLairezSalvy15} for other approaches not relying on the algebraic nature of $F$, but designed for the class of functions which can be written as diagonals of rational functions (these two classes coincide in the bivariate case). For example, Formula~\eqref{eq2} for $e_n$ has the following alternative representation: \begin{equation} (n+1) e_n=[t^n] \operatorname{diagonal} \left( \frac{(1+u)^6 u t^2}{1- ( u (u+1)^2 t+u(1+u)^4 t^2)}+(u+1)^2 \right)\,. \end{equation} The rational function on the right-hand side has the striking feature that its bivariate series expansion has only non-negative coefficients. In fact, it is even a bivariate $\mathbb{N}$-rational function (i.e., a function obtained as iteration of addition, multiplication, and quasi-inverse,\footnote{The {\it quasi-inverse} of a power series $f(z)$ of positive valuation is $1/(1-f(z))$.} starting from polynomials in $u$ and $t$ with positive integer coefficients). Given a multivariate rational function, it is a hard task to write it as an $\mathbb{N}$-rational expression (an algorithm is known in the univariate case), so some human computations were needed here to get the above expression. \smallskip In fact (and we believe that it was not observed before), these multivariate rational functions appearing in the computation of diagonals related to nested sums of binomials are always $\mathbb{N}$-rational: this follows from the closure properties of $\mathbb{N}$-rational functions. It is an open question to give a combinatorial interpretation (in terms of the initial structure counted by the diagonal) of the other diagonals of this rational function. It is also not easy to extrapolate from this rational function a general pattern which could appear for more general sets of steps: we shall see in Section~\ref{Section4} which type of formulas generalize the rich combinatorics that we had for $P(u)=u^{-2}+u^{-1}+u+u^2$. \subsection{How to derive the corresponding asymptotics: singularity analysis} We close this section by briefly addressing how to find the asymptotics of numbers of basketball walks. Indeed, standard techniques from singularity analysis suffice to get the asymptotic growth of the coefficients of $z^n$ in the generating functions that we consider here for $n \to \infty$. The interested reader is referred to~\cite{FlSe09} for more details on this subject (see Figure VI.7 therein for an illustration of singularity analysis). \begin{theorem} Let $G_{0,1}(z)$ and $G_{0,2}(z)$ be the generating functions for positive basketball walks with steps $-2,-1,+1,+2$ starting at the origin and ending at altitude $1$, respectively at $2$. Then, as $n \to \infty$, the coefficients are asymptotically equal to \begin{align} [z^n]G_{0,1}(z) &= \frac{1}{\sqrt{5 \pi}} \frac{4^n}{n^{3/2}} \left( 1 - \frac{81}{200} \frac{1}{n} + O \left(\frac{1}{n^2} \right) \right)\,, \\ [z^n] G_{0,2}(z) &= \frac{5+\sqrt{5}}{10 \sqrt{\pi}} \frac{4^n}{n^{3/2}} \left( 1 - \frac{201+24\sqrt{5}}{200} \frac{1}{n} + O \left(\frac{1}{n^2} \right) \right)\,. \end{align} \end{theorem} \begin{proof} The asymptotic growth of the coefficients is governed by the location of the dominant singularity (the singularity closest to the origin). The dominant singularity of~\eqref{eq:G01analytic} and~\eqref{eq:G02analytic} is given by $1/4$, since the square root becomes singular at this point. Next, we compute the singular expansion for $z\to 1/4$, which is a Puiseux series: \begin{align} G_{0,1}(z) &= -\frac{1-\sqrt{5}}{2} - \frac{2}{\sqrt{5}} \sqrt{1-4z} + O\left(1-4z\right),\\ G_{0,2}(z) &= \left(3-\sqrt{5}\right) - \frac{5+\sqrt{5}}{5} \sqrt{1-4z} + O\left(1-4z\right). \end{align} Finally, we apply the standard function scale from~\cite[Theorem~VI.1]{FlSe09} and the transfer for the error term~\cite[Theorem~VI.3]{FlSe09} to get the asymptotics. \end{proof} More generally, asymptotics for the number of walks from altitude $i$ to altitude $j$ in $n$ steps can be obtained via singularity analysis of the small roots, similarly to what was done in~\cite{BaFl02}. Note that it is easy derive as many terms as needed in the asymptotic expansion of the coefficients by including more terms in the Puiseux expansion. We also want to point out that this process was implemented in {\sl SageMath} (see~\cite{Kauers15}) or in {\sl Maple} by Bruno Salvy (as a part of the {\tt algolib} package). There, the {\tt equivalent} command directly gives the above result: \begin{maplegroup} \begin{mapleinput} \mapleinline{active}{1d}{equivalent(G01,z,n,3); }{} \end{mapleinput} \mapleresult \begin{maplelatex} \mapleinline{inert}{2d}{}{\[\displaystyle \frac{1}{5}\,{\frac { \sqrt{5}\, \, {4}^{n}}{ \sqrt{\pi }\, {n}^{3/2}}}-{\frac {81}{1000}}\,{\frac { \sqrt{5}\, \, {4}^{n}}{ \sqrt{\pi }\, {n}^{5/2}}}\\ \mbox{}+O \left( {\frac {{4}^{n}}{{n}^{7/2}}} \right) \,.\]} \end{maplelatex} \end{maplegroup} \section{General case: Lattice walks with arbitrary steps}\label{Section4} We first prove a theorem which holds for any symmetric set of steps, i.e., when the step polynomial satisfies $P(u) = P(1/u)$. \begin{theorem}[\sc Positive walk enumeration] \label{prop:G0kgeneral} Consider walks with a symmetric step polynomial $P(u)$. Let $G_{0,k}(z)$ be the generating function for positive walks, i.e., walks starting at the origin, ending at altitude~$k$, and always staying strictly above the $x$-axis in-between, and let $M_{> 0}(z)$ be the generating function of positive meanders, i.e., positive walks ending at any altitude $>0$. Then \begin{align} M_{> 0}(z) &= \sum_{k>0} G_{0,k}(z)=\prod_{i=1}^h \frac{1}{1-u_i(z)}, \label{eq:posmeanders}\\ G_{0,k}(z) &= h_k\left(u_1(z),u_2(z), \ldots, u_h(z) \right), \label{eq:G0kgeneral} \end{align} where $u_1(z),u_2(z),\dots,u_h(z)$ are the small roots of the kernel equation $1-zP(u)=0$, and \begin{align*} h_k(x_1,x_2,\ldots,x_h) &= \underset{i_1 +\dots+ i_h =k}{\sum_{i_1,\dots,i_h\ge0}} x_1^{i_1} x_2^{i_2} \cdots x_h^{i_h} \end{align*} is the complete homogeneous symmetric polynomial of degree $k$ in the variables $x_1,x_2,\break\ldots,x_h$. \end{theorem} \begin{proof} The formula for positive meanders follows from the expression for meanders (which are allowed to touch the $x$-axis!) in~\cite[Corollary~1]{BaFl02}, \begin{align*} M_{\geq 0 }(z)= -\frac{1}{z}\prod_{i=1}^h \frac{1}{1-v_i(z)}\,, \end{align*} where $v_1(z),v_2(z),\dots,v_h(z)$ are the large roots of $1-zP(u)=0$, i.e., those roots $v(z)$ for which $\lim_{z\to0}\vert v(z)\vert=\infty$. Every meander starts with an initial excursion, and later never returns to the $x$-axis any more. This simple fact implies the generating function equation $M_{\geq 0}(z)=E(z) M_{>0}(z)$. Hence, we need to divide the above expression for $M_{\ge0}(z)$ by the generating function for excursions --- which, by~\cite[Theorem~2]{BaFl02}, is given by $$E(z) = \frac{(-1)^{h-1}}{z} \prod_{i=1}^h u_i(z).$$ Finally, due to $P(u)=P(u^{-1})$, we have $u_i(z) = 1/v_i(z)$, which gives the final expression for $M_{>0}$, while the formula for $G_{0,k}(z)$ is proven in~\cite{BaWa16b}. \end{proof} This proof shows, in particular, that generating functions for strictly positive walks, respectively for weakly positive walks, are intimately related, and are therefore given by similar expressions. (The price of positivity is a division by $E(z)$, which encodes the excursion prefactor.) The proof also extends to non-symmetric steps, but then the formulas involve one more factor. It is possible to deal with them exactly in the way we proceed for symmetric steps, but this leads to slightly less nice formulas. \medskip In the sequel, we focus on positive walks with symmetric steps. We show in which way we can use the obtained expressions for the generating functions in order to get nice closed-form expressions for their coefficients. \subsection{Counting walks with steps in \texorpdfstring{$\mathcal{S}=\{0,\pm1,\dots,\pm h\}$}{S=\{-h,-h+1,\dots,h-1,h\}}} In Section~\ref{Section3} on basketball walks, we had a taste of what the kernel method could do for us when combined with Lagrange--B\"urmann inversion. This was, however, only for the case $\mathcal{S}=\{\pm1,\pm2\}$. In this section, we illustrate again the power of the kernel method, when applied to more general step sets~$\mathcal{S}$. We first start with a generalization of Section~\ref{Section2} to $\mathcal{S}=\{0,\pm1,\dots,\pm h\}$. In order to have a convenient notation, we introduce {\em $m$-nomial coefficients} by defining \begin{equation} \binom{n}{k}_m := [ u^k ] (1+u+\cdots+u^{m-1})^n\,, \end{equation} where $k$ is between $0$ and $(m-1) n$. \begin{proposition} The $m$-nomial coefficient equals \begin{equation}\label{mnomial} \binom{n}{k}_m=\sum_{i=1}^n(-1)^i\binom{n}{i}\binom{n+k-mi-1}{n-1}. \end{equation} \end{proposition} \begin{proof} Coefficient extraction in the defining expression for $\binom nk_m$ yields \begin{align} \binom{n}{k}_m &=[u^k](1+u+\cdots+u^{m-1})^n =[u^k](1-u^{m})^n\dfrac{1}{(1-u)^n}\\ &=[u^k]\left( \sum_{i=0}^n\binom{n}{i}(-1)^iu^{mi} \right) \left( \sum_{j\ge0}^{}\binom{n+j-1}{n-1}u^j\right) \\ &=\sum_{i=0}^{\left\lfloor (n+k-1)/m \right\rfloor} (-1)^i\binom{n}{i}\binom{n+k-mi-1}{n-1}. \end{align} The upper bound in the sum can be taken more naturally to be $i=n$, using the convention that binomials $\binom{n}{k}$ are $0$ for $n<0$ or $k>n$ (the reader should be warned that this is not the convention of {\sl Maple} or {\sl Mathematica}). This gives Formula~\eqref{mnomial} \end{proof} {\sl Historical remark.} These $m$-nomial coefficients appear in more than fifty articles (many of them focusing on trinomial coefficients) dealing with their rich combinatorial aspects (see e.g.~\cite{AndrewsBaxter87,Renzi92,Andrews07,Blaziak08}). We use the notation $\binom{n}{k}_m$ promoted by George Andrews~\cite{Andrews90}. It should be noted that they were previously called {\it polynomial coefficients} by Louis Comtet~\cite[p.~78] {Comtet74}, who is mentioning early work of D\'esir\'e Andr\'e (with a typo in the date) and Paul Montel~\cite{Andre1876,Montel1942}, and who was himself using another notation for these numbers, namely $\binom{n,m}{k}$. \smallskip These coefficients have a direct combinatorial interpretation in terms of lattice walk enumeration. \begin{theorem}[\sc Unconstrained walk enumeration] The number of unconstrained\/\footnote{Unconstrained means that the walks are allowed to have both positive and negative altitudes.} walks running from the origin to altitude $k$ in $n$ steps taken from $\{0,\pm1,\pm2,\break\dots,\pm h\}$ equals $\binom{n}{k+hn}_{2h+1}$. \end{theorem} \begin{proof} By~\eqref{eq:unconstrainedpaths}, the generating function for unconstrained walks is $$W(z,u)=\frac{1}{1-zP(u)} = \sum_{n=0}^\infty P^n(u)z^n\,.$$ Then a simple factorization shows that \begin{align} [u^k]P^n(u) &=[u^k]\left(\sum_{i=-h}^{h}u^i\right)^n =[u^k]u^{-hn}\left(\sum_{i=0}^{2h}u^i\right)^n=\binom{n}{k+hn}_{2h+1}. \label{eq:steppoly}\qedhere \end{align} \end{proof} Now we will see how to link these coefficients with {\em constrained\/} lattice walks. To this end, we first state the general version of the conjugation principle that we encountered in Proposition~\ref{prop:conjugation}. \begin{proposition}[\sc Conjugation principle for small roots] \label{prop:conjugationgeneral} Let $$P(u) = \sum_{i=-c}^d p_i u^i$$ be the step polynomial, and let $\omega = e^{ 2 \pi i / c}$ be a $c$-th root of unity. The small roots $u_i(z)$, $i=1,2,\ldots,c$, of\/ $1-zP(u)=0$ satisfy \begin{equation} u_i(z)=\sum_{n\geq 1} \omega^{n (i-1)} a_n z^{n/c} \end{equation} for certain ``universal" coefficients $a_n$, $n=1,2,\dots$. \end{proposition} \begin{proof} The kernel equation yields $$ u=X \left( p_{-c} + p_{-c+1}u + p_{-c+2} u^2 + \cdots + p_{d-1}u^{c+d-1} + p_{d}u^{c+d} \right)^{1/c}, $$ with $X=\omega^{j} {z}^{1/c}$ for $j=0,1,\ldots,c-1$. Since the above equation possesses a unique formal power series solution $u(X)$, the claim follows. \end{proof} Next, we apply Lagrange--B\"urmann inversion to the small roots given by the kernel method, and combine it with the conjugation principle. \begin{proposition}[\sc Explicit expansion of the roots $u_i$] \label{prop:coeffuigeneral} For lattice walks with step polynomial given by $P(u)=u^{-h}+u^{-h+1}+\dots+u^{h-1} +u^h$, let $U(z)$ be the root of\/ $1-z^h P(U) = 0$ whose Taylor expansion at~$0$ starts $U(z)=z+\cdots$. The series $U(z)$ is a power series, not a genuine Puiseux series. Then all small and large roots can be expressed in terms of $U(z)$, namely we have \begin{equation} u_i(z)=U(\omega^{i-1}z^{1/h}) \text{\quad and \quad} v_i(z)=1/U(\omega^{i-1}z^{1/h}), \quad i=1,2,\dots,h, \end{equation} where $\omega=e^{2\pi i/h}$ is a primitive $h$-th root of unity. The expansion of a power of the series $U(z)$ is explicitly given by \begin{equation} U^m(z)=\sum_{n=m}^\infty \frac{m}{n}\binom{n/h}{n-m}_{2h+1}z^n. \end{equation} \end{proposition} \begin{proof We want to solve $1-zP(u)=0$ for $u$. We may rewrite this equation as $$ z=\frac {u^h} {1+u+\cdots+u^{2h}}. $$ Taking the $h$-th root, we get $$ \omega^{i-1}z^{1/h}=\frac {u} {(1+u+\cdots+u^{2h})^{1/h}}, $$ for some $i$ with $1\le i\le h$. Since an equation of the form $Z=u\phi(u)$, where $\phi(u)$ is a power series in~$u$, has a unique power series solution $u(Z)$, the above equation has a unique solution $u_i(z)$, which turns out to have exactly the form described in the proposition. The equation for $v_i$ follows from $u_i=1/v_i$ as we have $P(u)=P(1/u)$. The equation for $U^m$ comes from Lagrange--B\"urmann inversion: \begin{align} [z^n]U^m(z) &=\frac{1}{n}[z^{-1}](z^m)^\prime P^{n/h}(z)\\ &=\frac{m}{n}[z^{-m}]\sum_{k}z^k\binom{n/h}{k+n}_{2h+1}\\ &=\frac{m}{n}\binom{n/h}{n-m}_{2h+1}. \qedhere \end{align} \end{proof} \begin{theorem}[\sc Closed-form expression for walks with $\mathcal{S}=\{0,\pm1,\dots,\pm h\}$] The numbers of positive walks and meanders from the origin to altitude~$k$ in $n$~steps from $\mathcal{S}\!=\!\{0,\pm1,\dots,\pm h\}$ admit the closed-form expressions \begin{align*} [z^n] G_{0,k}(z) &= \sum_{n_1+\dots+n_h=nh}\,\sum_{i_1+\dots+i_h=k} \frac {i_1} {n_1}{\binom {n_1/h}{n_1-i_1}}_{2h+1}\\ &\kern6cm \cdots \frac {i_h} {n_h}{\binom {n_h/h}{n_h-i_h}}_{2h+1} \omega^{\sum_{j=1}^h(j-1)n_j}, \\ [z^n] M_{>0}(z) &= \sum_{n_1+\dots+n_h=nh}\,\sum_{i_1,\dots,i_h\ge0} \frac {i_1} {n_1}{\binom {n_1/h}{n_1-i_1}}_{2h+1} \cdots \frac {i_h} {n_h}{\binom {n_h/h}{n_h-i_h}}_{2h+1} \omega^{\sum_{j=1}^h(j-1)n_j}. \end{align*} \end{theorem} \begin{proof} We use the expansions from Proposition~\ref{prop:coeffuigeneral} in the generating function formulas from Theorem~\ref{prop:G0kgeneral}. \end{proof} Here are some sequences of numbers of positive walks with steps $\mathcal{S}=\{0,\pm1,\dots,\pm h\}$, starting at the origin, and ending at altitude~$1$, for different values of~$h$: \begin{align*} h &=1 \quad \OEIS{A168049}\footnotemark{}: && 0, 1, 1, 2, 4, 9, 21, 51, 127, 323, 835, \ldots \\ h &=2 \quad \OEIS{A104632}: && 0, 1, 2, 6, 20, 73, 281, 1125, 4635, 19525, 83710, \ldots \\ h &=3 \quad \OEIS{A276902}: && 0, 1, 3, 12, 56, 284, 1526, 8530, 49106, 289149, 1733347, \ldots \\ h &=4 \quad \OEIS{A277920}: && 0, 1, 4, 20, 120, 780, 5382, 38638, 285762, 2162033, 16655167, \ldots \end{align*} \footnotetext{Axxxxxx refers to the corresponding sequence in the On-Line Encyclopedia of Integer Sequences, available electronically at \url{https://oeis.org}.} Furthermore, here are some sequences of numbers of positive walks with steps $\mathcal{S}=\{0,\pm1,\dots,\pm h\}$, starting at the origin, and ending at altitude~$2$, for small values of~$h$: \begin{align*} h &=1 \quad \OEIS{A105695}: && 0,0,1,2,5,12,30,76,196,512,1353, \ldots \\ h &=2 \quad \OEIS{A276903}: && 0,1,2,7,25,96,382,1567,6575,28096,121847,534953, \ldots \\ h &=3 \quad \OEIS{A276904}: && 0,1,3,14,68,358,1966,11172,65104,387029,2337919, \ldots \\ h &=4 \quad \OEIS{A277921}: && 0,1,4,23,142,950,6662,48420,361378,2753687,21334313,\ldots \intertext{\indent Here are the corresponding sequences for positive meanders:} h &=1 \quad \OEIS{A005773}: && 1, 1, 2, 5, 13, 35, 96, 267, 750, 2123, 6046, 17303, \ldots \\ h &=2 \quad \OEIS{A278391}: && 1, 2, 7, 29, 126, 565, 2583, 11971, 56038, 264345, \ldots \\ h &=3 \quad \OEIS{A278392}: && 1,3, 15, 87, 530, 3329, 21316, 138345, 906853, \ldots \\ h &=4 \quad \OEIS{A278393}: && 1,4,26,194,1521,12289,101205,844711,7120398,\ldots \intertext{\indent Here are the corresponding sequences for meanders (allowed to touch $0$): } h &=1 \quad \OEIS{A005773}: && 1, 2, 5, 13, 35, 96, 267, 750, 2123, 6046, 17303, 49721, \ldots \\ h &=2 \quad \OEIS{A180898}: && 1, 3, 12, 51, 226, 1025, 4724, 22022, 103550, 490191, \ldots \\ h &=3 \quad \OEIS{A180899}: && 1, 4, 22, 130, 803, 5085, 32747, 213419, 1403399, \ldots \\ h &=4 \quad \OEIS{A180900}: && 1, 5, 35, 265, 2100, 17075, 141246, 1182719, 9994086,\ldots \intertext{\indent Here are the corresponding sequences for excursions:} h &=1 \quad \OEIS{A001006}: && 1, 1, 2, 4, 9, 21, 51, 127, 323, 835, 2188, 5798, 15511, 41835, \ldots \\ h &=2 \quad \OEIS{A104184}: && 1, 1, 3, 9, 32, 120, 473, 1925, 8034, 34188, 147787, 647141, \ldots \\ h &=3 \quad \OEIS{A204208}: && 1, 1, 4, 16, 78, 404, 2208, 12492, 72589, 430569, 2596471, \ldots \\ h &=4 \quad \OEIS{A204209}: && 1, 1, 5, 25, 155, 1025, 7167, 51945, 387000, 2944860,\ldots \end{align*} \begin{remark} Most of the above sequences for $h\geq 3$ were not contained in the On-Line Encyclopedia of Integer Sequences (OEIS) before we added them. In Section~\ref{Section5}, we discuss the combinatorial structures related to the sequences which were already in the OEIS. \end{remark} \subsection{Counting walks with steps in \texorpdfstring{$\mathcal{S}=\{\pm1,\dots,\pm h\}$}{S=\{-h,\dots,-1,1,\dots,h\}}}\label{4.2} In this Section~\ref{4.2}, we consider the same steps as in the previous one, except that we drop the $0$-step. Certainly, for any type of walks consisting of $k$ steps with $0$-step included, enumerated by $f_k$ say, the number of walks of the same type consisting of $n$ steps, all of which different from the $0$-step, can be obtained by the inclusion-exclusion principle. The result is $\sum_{k=0}^n(-1)^{n-k}\binom nk f_k$. Here, our way to derive the corresponding formulas is more ad hoc and relies on the shape of the considered steps in~$\mathcal{S}$. This offers the advantage of leading to positive sum formulas, as opposed to the alternating sums produced by inclusion-exclusion. For convenience, we introduce the mock-$m$-nomial coefficients by \begin{equation} \binom{n}{k}_{2m}^{\hspace{-1mm}*} := [u^k] (1+\cdots+u^{m-1}+u^{m+1}+\cdots+u^{2m})^n\,. \end{equation} \begin{proposition} The mock-$m$-nomial coefficients can be expressed in terms of the (ordinary) $m$-nomial coefficients in the form\footnote{Here, the $^{\hspace{0mm}*}$ is a mnemonic to remind us that we do not have the 0 step.} \begin{equation}\label{mockmnomial} \binom{n}{k}_{2m}^{\hspace{-1mm}*}=\sum_{i=0}^n\binom{n}{i}\binom{n}{k-(m+1)i}_{m}. \end{equation} \end{proposition} \begin{proof} Factoring the expression and extracting coefficients, we obtain \begin{align} \binom{n}{k}_{2m}^{\hspace{-1mm}*}&= [u^k](1+\cdots+u^{m-1}+u^{m+1}+\cdots+u^{2m})^n\\ &=[u^k](1+u^{m+1})^n(1+u+\cdots+u^{m-1})^n\\ &=[u^k]\left(\sum_{i\ge0}\binom{n}{i}u^{(m+1)i} \right)\left(\sum_{j\ge0}\binom{n}{j}_{m}u^j\right)\\ &=\sum_{i=0}^n\binom{n}{i}\binom{n}{k-(m+1)i}_{m}. \qedhere \end{align} \end{proof} These mock-$m$-nomial coefficients have also a direct combinatorial interpretation in terms of lattice walk enumeration. \begin{theorem}[\sc Unconstrained walk enumeration] The mock-$m$-nomial coefficient $\binom{n}{k+hn}_{2h}^{\hspace{-1mm}*}$ is the number of unconstrained walks running from~$0$ to $k$ in $n$ steps taken from $\{\pm1,\pm2,\dots,\pm h\}$. \end{theorem} \begin{proof} We have \begin{align} [u^k]P^n(u) &=[u^k]\left(\sum_{i=-h}^{-1}u^i+\sum_{i=1}^{h}u^i\right)^n \\ &=[u^k]u^{-hn}\left(\sum_{i=0}^{h-1}u^i+\sum_{i=h+1}^{2h}u^i\right)^n =\binom{n}{k+hn}_{2h}^{\hspace{-1mm}*}. \qedhere \end{align} \end{proof} \begin{proposition}[\sc Explicit expansion of the roots $u_i$] \label{prop:smallbranches} For lattice walks with step polynomial given by $P(u)=u^{-h}+\cdots+u^{-1}+u^1+\cdots+u^{h}$, let $U(z)$ be the root of\/ $1-z^h P(U) = 0$ whose Taylor expansion at~$0$ starts $U(z)=z+\cdots$. Again, $U(z)$ is a power series, not a genuine Puiseux series. Then $U(z)$ satisfies \begin{equation}\label{smallroot2} U^m(z)=\sum_{n=1}^\infty \frac{m}{n}\binom{n/h}{n-m}^{\hspace{-1mm}*}_{2h}z^n, \end{equation} and all small and large roots are expressed in terms of $U(z)$ as \begin{equation} u_i(z)=U(\omega^{i-1}z^{1/h}) \text{\quad and \quad} v_i(z)=1/U(\omega^{i-1}z^{1/h}), \quad \text{for $i=1,2,\dots,h$}, \end{equation} where $\omega=e^{2\pi i/h}$ is a primitive $h$-th root of unity. \end{proposition} \begin{proof} We apply Lagrange--B\"urmann inversion to get \begin{align*} [z^n]U^m(z) =\frac{1}{n}[z^{-1}](z^m)^\prime P^{n/h}(z) =\frac{m}{n}[z^{-m}]\sum_{k}u^k\binom{n/h}{k+n}^{\hspace{-1mm}*}_{2h} =\frac{m}{n}\binom{n/h}{n-m}^{\hspace{-1mm}*}_{2h}. \qquad \qedhere \end{align*} \end{proof} \begin{theorem}[\sc Closed-form expression for walks with $\mathcal{S}\!=\!\{\pm1,\dots,\pm h\}$] The numbers of positive walks and meanders from the origin to altitude~$k$ in $n$~steps from $\mathcal{S}=\{\pm1,\dots,\pm h\}$ admit the closed-form expressions \begin{align*} [z^n] G_{0,k}(z) &= \sum_{n_1+\dots+n_h=nh}\,\sum_{i_1+\dots+i_h=k} \frac {i_1} {n_1}{\binom {n_1/h}{n_1-i_1}}_{2h}^{\hspace{-1mm}*} \cdots \frac {i_h} {n_h}{\binom {n_h/h}{n_h-i_h}}_{2h}^{\hspace{-1mm}*} \omega^{\sum_{j=1}^h(j-1)n_j},\\ [z^n] M_{>0}(z) &= \sum_{n_1+\dots+n_h=nh}\,\sum_{i_1,\dots,i_h\ge0} \frac {i_1} {n_1}{\binom {n_1/h}{n_1-i_1}}_{2h}^{\hspace{-1mm}*} \cdots \frac {i_h} {n_h}{\binom {n_h/h}{n_h-i_h}}_{2h}^{\hspace{-1mm}*} \omega^{\sum_{j=1}^h(j-1)n_j}. \end{align*} \end{theorem} \begin{proof} We use the expansions from Proposition~\ref{prop:smallbranches} in the generating function formulas from Theorem~\ref{prop:G0kgeneral}. \end{proof} Here are some sequences of numbers of walks with steps in $\mathcal{S}=\{\pm1,\pm2,\dots,\pm h\}$, starting at the origin, and ending at altitude~$1$, for different values of~$h$: \begin{align*} h &=1 \quad \OEIS{A000108}: && 0, 1, 0, 1, 0, 2, 0, 5, 0, 14, 0, \ldots \\ h &=2 \quad \OEIS{A166135}: && 0, 1, 1, 3, 7, 22, 65, 213, 693, 2352, 8034, \ldots \\ h &=3 \quad \OEIS{A276852}: && 0, 1, 2, 7, 28, 121, 560, 2677, 13230, 66742, 343092, \ldots \\ h &=4 \quad \OEIS{A277922}: && 0, 1, 3, 13, 71, 405, 2501, 15923, 104825, 704818, 4827957, \ldots \end{align*} Furthermore, here are some sequences of numbers of walks with steps in $\mathcal{S}=\{\pm1,\pm2,\break\dots,\pm h\}$, starting at the origin, and ending at altitude~$2$, for different values of~$h$: \begin{align*} h &=1 \quad \OEIS{A000108}: && 0, 0, 1, 0, 2, 0, 5, 0, 14, 0, 42, \ldots \\ h &=2 \quad \OEIS{A111160}: && 0,1,1,4,9,31,91,309,1009,3481,11956, \ldots \\ h &=3 \quad \OEIS{A276901}: && 0,1,2,9,34,159,730,3579,17762,90538,467796, \ldots \\ h &=4 \quad \OEIS{A277923}: && 0, 1, 3, 16, 84, 505, 3121, 20180, 133604, 904512, 6224305,\ldots \end{align*} Here are the corresponding sequences for positive meanders: \begin{align*} h &=1 \quad \OEIS{A001405}: && 1, 1, 1, 2, 3, 6, 10, 20, 35, 70, 126, 252, 462, 924, \ldots \\ h &=2 \quad \OEIS{A278394}: && 1, 2, 5, 17, 58, 209, 761, 2823, 10557, 39833, 151147 \ldots \\ h &=3 \quad \OEIS{A278395}: && 1, 3, 12, 60, 311, 1674, 9173, 51002, 286384, 1620776, \ldots \\ h &=4 \quad \OEIS{A278396}: && 1, 4, 22, 146, 1013, 7269, 53156, 394154, 2951950,\ldots \end{align*} Here are the corresponding sequences for meanders (allowed to touch~$0$): \begin{align*} h &=1 \quad \OEIS{A001405}: && 1, 1, 2, 3, 6, 10, 20, 35, 70, 126, 252, 462, 924, 1716, 3432, \ldots \\ h &=2 \quad \OEIS{A047002}: && 1,2,7,23,83,299,1107,4122,15523,58769,223848, \ldots \\ h &=3 \quad \OEIS{A278398}: && 1,3,15,75,400,2169,11989,66985,377718,2144290, \ldots \\ h &=4 \quad \OEIS{A278416}: && 1,4,26,174,1231,8899,65492,487646,3664123,\ldots \end{align*} Here are the corresponding sequences for excursions: \begin{align*} h &=1 \quad \OEIS{A126120}: && 1, 0, 1, 0, 2, 0, 5, 0, 14, 0, 42, 0, 132, 0, 429, 0 \ldots \\ h &=2 \quad \OEIS{A187430}: && 1, 0, 2, 2, 11, 24, 93, 272, 971, 3194, 11293, 39148, 139687 \ldots \\ h &=3 \quad \OEIS{A205336}: && 1, 0, 3, 6, 35, 138, 689, 3272, 16522, 83792, 434749, \ldots \\ h &=4 \quad \OEIS{A205337}: && 1, 0, 4, 12, 82, 454, 2912, 18652, 124299, 841400,\ldots \end{align*} \begin{remark} The cases with $h=1$ lead to famous sequences, having many links with the combinatorics of trees, via the {\L}ukasiewicz correspondence (see Section~\ref{Section2}). It is surprising that the cases with $h=2$ also offer many links with trees, as we show in the next section. \end{remark} \pagebreak \section{Some links with other combinatorial problems}\label{Section5} In this section, we establish some links between our lattice walks and other combinatorial problems. Thereby we prove several conjectures issued in the On-Line Encyclopedia of Integer Sequences. \subsection{Trees and basketball walks from 0 to 1} First, we prove that the sequence \oeis{A166135} from the On-Line Encyclopedia of Integer Sequences, coming from the enumeration of certain tree structures used in financial mathematics, is in fact related to basketball walks, and corresponds more precisely to the coefficients of $G_{0,1}(z)$. The $m$-nomial tree is a lattice-based computational model used in financial mathematics to price options. It was developed by Phelim Boyle~\cite{Boyle86} in 1986. For example, for $m=3$, the underlying stock price is modelled as a recombining tree, where, at each node, the price has three possible paths: an up, down, or stable path. The case $m=2$ has a long history going back to one of the founding problems of financial mathematics and probability theory, the ``ruin problem'', analysed in the XVIIIth and XIXth century by de Moivre, Laplace, Huygens, Amp\`ere, Rouch\'e, before to be revisited by combinatorialists like Catalan, Whitworth, Bertrand, Andr\'e, Delannoy (see~\cite{BaSc05} for more on these aspects). Figure~\ref{4nomialTree} illustrates a 4-nomial tree. \begin{figure}[!hb] \includegraphics[width=.6\textwidth]{4nomialtree.png} \caption{Cutting a 4-nomial tree at one unit from its root gives the above picture, which thus naturally corresponds to the lattice supporting our lattice basketball walks. The numbers near each node indicate the number of walks from the root to this node.\label{4nomialTree}} \end{figure} The following proposition gives the exact link between these trees and a generalization of basketball walks. \begin{proposition}[\sc Link between lattice walks and $m$-nomial trees] Consider the step sets $$ \mathcal{S}_{2n} =\{-n,\dots,-1,1,\dots,n\} \text{ and } \mathcal{S}_{2n+1}=\mathcal{S}_{2n}\cup\{0\}.$$ For each step set $S_{m}$, define $T_m(z)$ to be the generating function for walks using steps from $\mathcal{S}_{m}$, starting at the origin and getting absorbed at $-1$. (By this, we mean that the walks may never touch $y=-1$ except, possibly, at the very last step.) Then the coefficients of $T_2(z)$ are the Catalan numbers, the coefficients of $T_3(z)$ are the Motzkin numbers, while the coefficients of $T_4(z)$ count our basketball walks from 0 to 1 (walks with steps $\pm 2, \pm 1$, starting at the origin and ending at altitude~1, and never touching 0 in-between). \end{proposition} \begin{proof}While the correspondence is direct for $m\leq 3$, it follows for $m=4$ from a time reversion, as each walk from $T_4$ can then be obtained from $G_{0,1}$ and vice versa (see Table~\ref{fig:4types2}). Thus, $T_4(z)=G_{0,1}(z)$. \end{proof} \medskip \begin{table*}[!hb] \small \begin{center}\renewcommand{\tabcolsep}{3pt} \begin{tabular}{|c|c|} \hline $T_4$ & $G_{0,1}$\\ \hline \begin{tabular}{c} {\includegraphics[width=71mm]{G01walk1jumpreverse}} \\ last step is a 1-step down \end{tabular} & \begin{tabular}{c} {\includegraphics[width=71mm]{G01walk1jump}} \\ first step is a 1-step up \end{tabular} \\ \hline \begin{tabular}{c} \includegraphics[width=71mm]{G01walk2jumpreverse} \\ last step is a 2-step down\\ \end{tabular} & \begin{tabular}{c} {\includegraphics[width=71mm]{G01walk2jump}} \\ first step is a 2-step up\\ \end{tabular}\\ \hline \end{tabular} \end{center} \caption{\label{fig:4types2} By time reversal, $T_4(z)=G_{0,1}(z)$.} \end{table*} \subsection{Increasing trees and basketball walks} A \textit{unary-binary tree} is an ordered tree such that each node has $0,1$, or~$2$ children. An \textit{increasing unary-binary tree on $n$ vertices} is a unary-binary tree with $n$ vertices labelled $1,2,\dots,n$ such that the labels along each walk from the root are increasing (cf.~\cite[p.~51]{Stanley86}). Given an increasing unary-binary tree~$T$, we associate with $T$ the \textit{permutation} $\sigma_T$ constructed by reading the tree left to right, level by level, starting at the root. A permutation $\sigma$ is said to \textit{contain the pattern $\pi$} if there exists a subsequence of $\sigma$ that has the same relative order as $\pi$. Otherwise, $\sigma$ is said to \textit{avoid the pattern $\pi$}. For example, the permutation $\sigma=14235$ contains the pattern $213$ because $\sigma$ contains the subsequence $425$, in which the numbers have the same relative order as in $213$, while the permutation $12453$ avoids $213$. Manda Riehl initiated studies of increasing trees for which the associated permutation avoids a given pattern (see also~\cite{Riehl16}). By a computer program, she obtained the first terms of the corresponding sequences for patterns of length~3. She observed that ``the number of increasing unary-binary trees with associated permutation avoiding 213'' seems to coincide with sequence \oeis{A166135}, which we proved to count basketballs walks from altitude~0~to altitude~1. Figure~\ref{trees} shows a verification of this claim for $n=5$: there are $39$ increasing unary-binary trees on $5$ vertices, among them, $22$ correspond to permutations avoiding the pattern $213$. (The forbidden subsequences are highlighted in red. The trees in black all avoid $213$. The trees are grouped according to their associated permutations. Tree labels are read left to right.) \begin{figure} \vspace{.1cm} \begin{center} \includegraphics[scale=0.66]{iubtrees1.pdf} \vspace{.2cm} \includegraphics[scale=0.66]{iubtrees2.pdf} \vspace{.2cm} \includegraphics[scale=0.66]{iubtrees3.pdf} \end{center} \vspace{-5mm} \noindent\rule{\textwidth}{1pt} \begin{center} \includegraphics[scale=0.66]{incbad1.pdf} \vspace{.2cm} \includegraphics[scale=0.66]{incbad2.pdf} \end{center} \caption{All increasing unary-binary trees with 5 nodes, where patterns $213$ are marked in red. There are 22 trees (drawn in black) for which the associated permutation avoids this pattern.\label{trees}} \end{figure} Here is the reformulation of Riehl's conjecture which takes into account our findings. \begin{conjecture} \label{conj:Riehl} The number of basketball walks of length $n$ starting at the origin and ending at altitude~$1$ that never touch or pass below the $x$-axis equals the number of increasing unary-binary trees on $n$ vertices with associated permutation avoiding $213$. \end{conjecture} After the first version of this article was circulated via the {\tt arXiv}, Bettinelli, Fusy, Mailler, and Randazzo~\cite{BettinelliFusy16} found a nice bijective proof of this conjecture. \begin{comment} \begin{proof} \medskip \noindent Fact 1 (strong structural property): If one has a descent (i.e., a {\em consecutive} pattern 21), then 2 is a leaf. (Because 1 can't be a child of 2 as our tree is increasing. So the only possibility left for a child of 2 is a 3. But this would make our tree have the forbidden pattern 213.)\\ Fact 2: All labels before $n$ are increasing. (Otherwise, the tree would contain a pattern 213.)\\ Fact 3: All labels after $n$ are a permutation of a consecutive set of integers. (Otherwise, there exists label $b$ before $n$ and labels $a$ and $c$ after $n$ such that $a<b<c$ and where $a$ appears before $c$ (because if $c$ appears before $a$, then the tree would not be increasing since $b,n,c$ are all larger than $a$). Then, $bac$ would form the forbidden 213 pattern.) \\ Fact 4: Consecutive internal nodes (same level or not) have consecutive labels in increasing order. Indeed, assume we have nonconsecutive labels on a pair of consecutive internal nodes. Consider the first such occurrence. We have $a<b<c<d$ where $a$ and $c$ are labels on consecutive internal nodes, label $b$ occurs after $a$ and $c$ (otherwise, $bac$ forms a 213 pattern), and $d$ is the label on a child of $c$. Notice that b must appear before $d$ in the associated permutation, since each internal node after $c$ is larger than $c$, precluding $b$ from being its child. Then $cbd$ is a 213 pattern.\\ Fact 5: There are no restrictions on the labels for the leaves, except that the permutation must avoid 213.\\ Fact 6: $G_{01}$ can be written as a composition in terms of the Catalan generating function: $G_{01}(z)=A(C(z)-1)$, with $A(z)= -z C(-z)$, and $C(z)=1+zC(z)^2$, i.e. $G_{01}(z)=(C(z)-1) C(1-C(z))$.\\ All of this leads to $$z(T+T^2)= (z+zT+zT^2 )^2.$$ Introducing $C:=T+T^2= z+2z^2+5z^3+14z^4+42z^5+\cdots$, this is thus equivalent to $$C= z (1+C)^2.$$ We therefore get the algebraic equation which was characterizing $G_{0,1}$, with the same first coefficients, thus proving our claim. \end{proof} \end{comment} \medskip How strong is the constraint of avoiding the pattern 213? For this, we need to compute the probability that an increasing unary-binary tree avoids the pattern 213. Due to Conjecture~\ref{conj:Riehl}, proved in~\cite{BettinelliFusy16}, we know the number of increasing unary-binary trees which avoid $213$. Hence, the question is to compute the total number $t_n$ of increasing unary-binary trees, which can be done via the so-called boxed product. The boxed product (written $\hspace{1mm}{}^\square\hspace{-1mm}\times$) is the combinatorial construction corresponding to a labelled product, in which the minimal label is forced to be in the first component of this product (see~\cite{FlSe09}). This leads the following recursive decomposition for binary-ternary increasing trees ${\mathcal T}$: $${\mathcal T}= leaf + root \hspace{1mm}{}^\square\hspace{-1mm}\times {\mathcal T} + root \hspace{1mm}{}^\square\hspace{-1mm}\times {\mathcal T} \times {\mathcal T} \,,$$ which translates into the following functional equation for the corresponding exponential generating function: $$T(z)= z + \int_0^z T(t) dt + \int_0^z T^2(t) dt\,.$$ By solving the associated differential equation $T'(z) = 1+ T(z)+T^2(z)$, we obtain $$T(z)= \frac{\sqrt{3}}{2} \tan\left(\frac{\pi}{6}+\frac{\sqrt{3}}{2}z \right) -\frac{1}{2}\,.$$ The corresponding Taylor expansion is $$T(z)=\sum_{n\geq 1} t_n \frac{z^n}{n!} = z + \frac{z^2}{2!} + 3 \frac{z^3}{3!} + 9 \frac{z^4}{4!}+ 39 \frac{z^5}{5!}+189\frac{z^6}{6!}+ 1107\frac{z^7}{7!}+O(z^8) \,.$$ Singularity analysis on the dominant poles of the $\tan$ function implies that $$t_n \sim 3 \sqrt{\frac{3}{2\pi}} \left( \frac{3^{3/2}}{2 e \pi}\right)^n \sqrt{n} \, n^n\,.$$ In conclusion, increasing unary-binary trees grow like $n^{1/2} A^n n^n$, while the same trees avoiding the pattern 213 grow like $n^{-3/2}4^n$. This observation suggests the following natural conjecture. \begin{conjecture}[\sc A Stanley--Wilf-like conjecture for pattern avoidance in increasing trees] Let ${\mathcal T}$ be a class of increasing trees of prescribed arity encoded by a power series $\phi$, i.e., one has ${\mathcal T}'=z \phi({\mathcal T})$. Then the number $a_n$ of such trees avoiding a given pattern satisfies $a_n=O(C^n)$, for some $C$ depending on the pattern and on $\phi$. \end{conjecture} This conjecture shares the spirit of the Stanley--Wilf conjecture (proven by a combination of \cite{KlazAA} and~\cite{MarcusTardos04}), which asserted that any class of pattern-avoiding permutations has an exponential growth rate. \subsection{Boolean trees and basketball walks from 0 to 2} In~\cite{bender91}, Bender and Williamson considered the problem of bracketing some binary operations (objects that are in bijection with the Boolean trees that we present in Figure~\ref{booleantrees}). It turns out that this problem is doubly related to our basketball walks (walks with steps $\pm 1, \pm 2$, always positive). This is what we address in the next two propositions. \begin{figure}[H] \includegraphics[scale=0.88]{binarytreetruefalse.pdf} \includegraphics[scale=0.88]{binarytree.pdf} \caption{Boolean trees (i.e., binary trees where each node is labelled either ``false'' or ``true'') such that a node having children with Boolean value $A$ and $B$ will have the Boolean value ``$B \Rightarrow A$''. }\label{booleantrees} \end{figure} \begin{proposition} Under the conventions $1^1=1^0=0^0=1$ and $0^1=0$, the number of bracketings of $n+1$ zeroes $0$\^{}$\cdots$\^{}$0$ giving result $1$ is equal to the number of basketball walks from altitude~$0$ to altitude~$2$ of length $n$. \end{proposition} \begin{proof} Let $W(z)$ {\em(}respectively $Z(z)${\em)} be the generating function for the number of bracketings of $n$ zeroes 0\^{}0\^{}$\cdots$\^{}0 producing result $1$ (respectively $0$). The objects that are counted by $W(z)$ are of the form (``1'')\^{}(``1''), (``1'')\^{}(``0''), or (``0'')\^{}(``0''), where ``1'' stands for a bracketing producing the result~1, and ``0'' stands for a bracketing producing the result~0. This observation translates into the generating function equation \begin{equation} W(z)=W^2(z)+Z(z)W(z)+Z^2(z). \label{eq:W} \end{equation} Similarly, a bracketing producing $0$ may either be a single $0$ or a bracketing of the form (``0'')\^{}(``1''). This yields the equation \begin{equation} Z(z)=z+Z(z)W(z). \label{eq:Z} \end{equation} Let $C(z):=Z(z)+W(z)$. Equations~\eqref{eq:W} and~\eqref{eq:Z} imply $ C(z)=1+zC^2(z)$, i.e., $ C(z)=\frac{1}{2z}-\frac{1}{2z}\sqrt{1-4z}. $ This is not a surprise because $W+Z$ corresponds to well parenthesized words, known to be counted by Catalan numbers. We may ``replace'' $W(z)$ by $C(z)$ in Equation~\eqref{eq:Z}. This leads to \begin{equation} Z(z)=z+Z(z)(C(z)-Z(z)). \end{equation} Solving for $Z(z)$, we obtain \begin{align}\label{eq:ZZ} \begin{aligned} Z(z) &=\frac{C(z)-1+\sqrt{(C(z)-1)^2+4z}}{2} \\ &=-\frac{1}{4}-\frac{1}{4}\sqrt{1-4z}+\frac{1}{4}\sqrt{2+12z+2\sqrt{1-4z}}. \end{aligned} \end{align} Therefore, we get \begin{equation} W(z)=C(z)-Z(z) =\frac{3}{4}-\frac{1}{4}\sqrt{1-4 z}-\frac{1}{4}\sqrt{2 + 12 z + 2\sqrt{1-4 z}}. \end{equation} Comparison of this expression with Expression~\eqref{eq:G02analytic} for $G_{0,2}(z)$ shows that $W(z)=zG_{0,2}(z)$. \end{proof} We leave it to the reader to find a bijective proof between bracketings of 0\^{}\dots\^{}0 having value~1 and basketball walks from altitude~0 to altitude~2. \begin{proposition} The number of basketball walks of length $n$ starting at the origin, ending at altitude~$1$, never running below the $x$-axis in-between, is equal to the number of bracketings of $n+2$ zeroes $0$\^{}$0$\^{}$\cdots$\^{}$0$ producing result~$0$. \end{proposition} \begin{proof} The generating function $F_1(z)$ for walks ending at $1$ is given by~\eqref{eq:G2} in the form \begin{equation} F_1(z) =G_{1,2}(z) = \frac{u_1(z)u_2(z)+u_1(z)+u_2(z)}{z}. \end{equation} The generating function $Z(z)$ for the number of bracketings of $n$ zeroes 0\^{}$\cdots$\^{}0 having value $0$ is given by \eqref{eq:ZZ}. Substitution of the closed-form expressions for the small roots into $F_1(z)$ yields $z^2F_1(z)=Z(z)$. This establishes the claim. \end{proof} \section{Conclusion} In this article, we show how to derive closed-form expressions for the enumeration of lattice walks satisfying various constraints (starting point, ending point, positivity, allowed steps, \dots). The key is a proper use of the Lagrange--B\"urmann inversion in combination with the expressions given by the kernel method. This technique admits many extensions, which will work in a similar way: it is possible to extend it to walks in which we want to keep track of some parameters (marking a specific step, pattern, altitude, \dots), allowing an infinite set of steps, or unbounded steps (this would encode what is called catastrophes in queuing theory language). It is also possible to consider other constraints, such as to force the walk to live in some cone or to have some forbidden patterns. In all these cases, the kernel method will give a closed-form expression for the generating function, in terms of the roots of the kernel, and thus, our mix of kernel method and Lagrange--B\"urmann inversion will lead in these situations also to some closed-form expression for the coefficients of the generating function (in terms of nested sums of binomials). In several cases, these nested sums of binomials provide the nice challenge of finding bijective proofs. It is satisfying to find {\it some} formula for the enumeration of certain lattice paths which is efficient (in terms of algorithmic complexity), but the fact that many of these sums involve only positive terms is an indication that combinatorics has still its word to say on these formulas. The holonomic approach, as well illustrated by the book of Petkov{\v s}ek, Wilf, and Zeilberger~\cite{pewz96}, or Kauers and Paule~\cite{KauersPaule11}, is a way to prove that different binomial expressions correspond in fact to the same sequence. It remains an open question to know which methods can lead to the most concise formula: the platypus algorithms and the Flajolet--Soria formula~\cite{BaFl02,BanderierDrmota15}, or the cycle lemma, and extraction of diagonals of rational functions seem to indicate that we could in fact need an arbitrarily large amount of nested sums. In some cases, one can reduce the number of nested sums with techniques from symbolic summation theory (e.g., by $\Sigma \Pi$ extension theory~\cite{Schneider07}, or geometric simplifications in diagonal extractions of rational functions~\cite{BostanLairezSalvy15}), but it is still unknown if, for the directed lattice path models we considered, there is a miraculous simple formula (with just one or two nested sums). \bigskip {\em Acknowledgments:} We thank the organizers of the 8th International Conference on Lattice Path Combinatorics \& Applications, which provided the opportunity for this collaboration. Sri Gopal Mohanty played an important role in the birth of this sequence of conferences, and his book~\cite{Mohanty79} was the first one (together with the book of his Ph.D. advisor Tadepalli Venkata Narayana~\cite{Narayana79}) to spur strong interest in lattice path enumeration. We are therefore pleased to dedicate our article to him. \bibliographystyle{plain}

https://arxiv.org/abs/2108.02275

Admissibility and Frame Homotopy for Quaternionic Frames

We consider the following questions: when do there exist quaternionic frames with given frame spectrum and given frame vector norms? When such frames exist, is it always possible to interpolate between any two while fixing their spectra and norms? In other words, the first question is the admissibility question for quaternionic frames and the second is a generalization of the frame homotopy conjecture. We give complete answers to both questions. For the first question, the existence criterion is exactly the same as in the real and complex cases. For the second, the non-empty spaces of quaternionic frames with specified frame spectrum and frame vector norms are always path-connected, just as in the complex case. Our strategy for proving these results is based on interpreting equivalence classes of frames with given frame spectrum as adjoint orbits, which is an approach that is also well-suited to the study of real and complex frames.

\section{Introduction}\label{sec:intro} In a finite-dimensional real or complex Hilbert space $\mathfrak{H}$, an (ordered) \emph{frame} is simply a collection $f_1, \dots , f_N \in \mathfrak{H}$ that spans $\mathfrak{H}$. When $N=d:=\dim \mathfrak{H}$, this is just a basis for $\mathfrak{H}$, but when $N > d$ a frame gives a redundant representation of a signal $v \in \mathfrak{H}$ by \begin{equation}\label{eq:analysis} (\langle v, f_1 \rangle, \dots , \langle v, f_N \rangle) \end{equation} which can be more robust to erasures and other corruption of the data than a basis representation~\cite{Casazza:2003vp,Goyal:2001cd,Holmes:2004iv}. Identifying the frame with the matrix $F = [f_1 | \dots | f_N]$ whose columns are the frame vectors, the redundant representation~\eqref{eq:analysis} corresponds to evaluation of the \emph{analysis operator} $v \mapsto F^\ast v$. Composition of the analysis operator with its adjoint \emph{synthesis operator} $w \mapsto F w$ gives the \emph{frame operator} \[ v \mapsto FF^\ast v. \] For orthonormal bases---or more generally, and by definition, \emph{Parseval frames}---the frame operator is the identity and the synthesis operator provides a simple method for reconstructing the signal $v$ from the data $F^\ast v$. As in the case of Parseval frames, it is often desirable to choose frames with a fixed spectrum of the frame operator, for example to provide optimal reconstruction in a given noise model~\cite{Goyal:2001cd,Casazza:2011ev,Viswanath:2002bv}. Likewise, the squared norms $\|f_1\|^2,\dots,\|f_N\|^2$ of the frame vectors are often fixed for both practical and theoretical reasons~\cite{Kovacevic:2007fja,Viswanath:1999hf,Rupf:1994fl,Casazza:2006cx}. In other words, given vectors $\boldsymbol{\lambda}=(\lambda_1, \dots , \lambda_d)$ and $\boldsymbol{r}=(r_1, \dots , r_N)$ of positive numbers, we are often interested in selecting frames from the space $\F^{\mathfrak{H},N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ of frames $f_1, \dots, f_N \in \mathfrak{H}$ so that $\boldsymbol{\lambda}$ is the spectrum of the frame operator $FF^\ast$ and $\|f_i\|^2 = r_i$ for $i=1, \dots , N$. In particular, two natural questions immediately present themselves: \begin{enumerate} \item Does there exist a frame for $\mathfrak{H}$ with prescribed data $\boldsymbol{\lambda}$ and $\boldsymbol{r}$? In other words, is $\F^{\mathfrak{H},N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ non-empty? \item Can we interpolate between arbitrary elements of $\F^{\mathfrak{H},N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$? In other words, is $\F^{\mathfrak{H},N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ path-connected? \end{enumerate} Question 1 has been completely answered by Casazza and Leon~\cite{Casazza:2010ti}, who give a simple compatibility criterion for $\boldsymbol{\lambda}$ and $\boldsymbol{r}$ which determines whether or not $\F^{\mathfrak{H},N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is empty. Moreover, the same criterion applies in both the real and the complex cases. Question 2 is a generalization of the well-known \emph{frame homotopy conjecture}, which was posed by Larson in a 2002 REU and first appeared in the literature in Dykema and Strawn's 2006 paper~\cite{Dykema:2006ux}. This conjecture says that when $\boldsymbol{\lambda}$ and $\boldsymbol{r}$ are constant vectors, the space $\F^{\mathfrak{H},N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is path-connected (when considered with the natural subspace topology). Up to scale one can assume that the constant vector $\boldsymbol{r} = (1, \dots , 1)$, and hence the frames under consideration are \emph{unit-norm tight frames}, so the frame homotopy conjecture says that the space of unit-norm tight frames in $\mathfrak{H}$ is path-connected. The frame homotopy conjecture was proved for both real and complex frames by Cahill, Mixon, and Strawn in a 2017 paper~\cite{Cahill:2017gv}. In previous work~\cite{NeedhamSGC}, we gave a complete answer to Question 2 for complex frames, showing that the space $\mathcal{F}^{\C^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is always path-connected. For real frames with nonconstant $\boldsymbol{\lambda}$ or $\boldsymbol{r}$, Question 2 remains open. There has been a recent flourishing of interest in \emph{quaternionic frames}~\cite{iverson_note_2021,Waldron:2020tp,Cohn:2016bz,EtTaoui:2020kf,Waldron:2020ti,Khokulan:2017ic,Sharma:2019js,Virender:2020ua}; that is, frames in $\mathbb{H}^d$ for $d \geq 1$ and $\mathbb{H}$ being the 4-dimensional skew field of quaternions. Since the group $\operatorname{Sp}(1)$ of unit quaternions is isomorphic to $\operatorname{Spin}(3)$ (the universal cover of the rotation group $\operatorname{SO}(3)$), the quaternions can be interpreted as a cone over $\operatorname{Spin}(3)$, so they are well-suited to parameterizing rotations in $\mathbb{R}^3$~\cite{Hanson:2006tr}. Just as complex numbers consist of a magnitude and a phase, quaternions consist of a magnitude and a versor, which determines a particular 3D rotation. Consequently, vectors $v \in \mathbb{H}^d$ can be used to record both magnitude and orientation information, for example in framed curves like polymers or inflatable elastic rods~\cite{Cantarella:2013bla,Hanson:2012vb,Howard:2011fj,Needham:2017vd}. Moreover, quaternions have become an increasingly popular tool for representing signals in data-driven applications. For example, quaternion-valued signals have recently been used to encode RGB images \cite{ell2006hypercomplex,fletcher2017development}, measurements in industrial machinery \cite{yi2017quaternion}, and multicomponent seismic measurements \cite{zhao2020quaternion}; see also the special issue of \emph{Signal Processing} on Hypercomplex Signal Processing~\cite{SPspecialissue}. This motivates the extension of signal processing techniques for classical (i.e., real- or complex-valued) signals to quaternionic signals. Despite the difficulties introduced by the non-commutativity of the quaternions, it is still possible to define frames in $\mathbb{H}^d$, the norms of the frame vectors, and the spectrum of the frame operator (see \Cref{sub:quaternionic_frames}). In other words, the space $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ of length-$N$ frames in $\mathbb{H}^d$ with fixed frame spectrum $\boldsymbol{\lambda}$ and fixed frame vector norms $\|f_i\|^2 = r_i$ exists, and the goal of this paper is to give complete answers to Questions 1 and 2 in this case. Specifically, we show: \begin{thm}\label{thm:existence} Let $N$, $d$, $\boldsymbol{\lambda}$, and $\boldsymbol{r}$ be as above, and additionally assume that $\boldsymbol{\lambda}$ and $\boldsymbol{r}$ are sorted in non-increasing order: $\lambda_1 \geq \dots \geq \lambda_d > 0$ and $r_1 \geq \dots \geq r_N > 0$. Then $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is non-empty if and only if \[ \sum_{i=1}^k r_i \leq \sum_{i=1}^k \lambda_i \text{ for all } k=1, \dots , d \qquad \text{and} \qquad \sum_{i=1}^N r_i = \sum_{i=1}^d \lambda_i. \] \end{thm} \begin{thm}\label{thm:connected} For any $\boldsymbol{r}$ and $\boldsymbol{\lambda}$, the space $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is path-connected. \end{thm} Since the empty set is trivially path-connected, the substance of \Cref{thm:connected} is that all of the non-empty $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ spaces are path-connected. Note that the condition in \Cref{thm:existence} is exactly the same as that given by Casazza and Leon in the real and complex cases~\cite{Casazza:2010ti}. In other words, for given $\boldsymbol{\lambda}$ and $\boldsymbol{r}$ either all three of $\mathcal{F}^{\mathbb{R}^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$, $\mathcal{F}^{\C^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$, and $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ are empty, or all three are non-empty. Likewise, the result in \Cref{thm:connected} is the same as in the complex case~\cite{NeedhamSGC}, so that $\mathcal{F}^{\C^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ and $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ are both path-connected. In particular, \Cref{thm:existence,thm:connected} imply that all spaces of quaternionic unit-norm tight frames are non-empty and path-connected, so the standard frame homotopy conjecture is true in the quaternionic setting. Our strategy for proving the complex version of \Cref{thm:connected} in~\cite{NeedhamSGC} was based on symplectic geometry, which is extremely well-suited to the complex setting, but not to either the real or quaternionic settings. In this paper, we instead identify (equivalence classes of) frames with fixed frame spectrum $\boldsymbol{\lambda}$ with an adjoint orbit of the action of the symplectic group\footnote{The (compact) symplectic group is the quaternionic analog of the unitary group; see \Cref{sub:quaternions} for the definition.} $\operatorname{Sp}(N)$ on the space of $N \times N$ Hermitian matrices, which is the natural home of the Gram matrix $F^\ast F$ associated to a frame $F$. This is not special to quaternionic frames: one can make an analogous identification for real and complex frames, so we view this as a unifying perspective on all three classes of frames. The advantage is that such adjoint orbits are extremely nice geometrically. In particular, they are motivating special cases of both Kostant's convexity theorem~\cite{Kostant:1973ff} and of the theory of isoparametric submanifolds~\cite[Chapter~6]{Palais:1988ks}, and the basic strategy for proving \Cref{thm:existence,thm:connected} is to apply those very general tools to these specific problems. Specifically, since $\boldsymbol{\lambda}$ and $\boldsymbol{r}$ are essentially the spectrum and diagonal entries, respectively, of the Gram matrix $F^\ast F$, the real and complex versions of \Cref{thm:existence} can be proved using the Schur--Horn theorem~\cite{Antezana:2007ci,Tropp:2005do}. Likewise, we will prove \Cref{thm:existence} using a quaternionic Schur--Horn theorem. This quaternionic version of Schur--Horn is an easy consequence of Kostant's convexity theorem, but its statement does not seem to be readily accessible in the literature (though see~\cite[Example~8.6]{kobert_spectrahedral_2021}). Since it may be of some independent interest, we give a statement and proof in \Cref{thm:quaternionic Schur-Horn}. On the other hand, adjoint orbits (and, more generally, isoparametric submanifolds) share many of the nice features of Hamiltonian manifolds~\cite{Terng:1986hg,Mare:2005eo}. In particular, in this setting we have access to a version of Atiyah's connectedness theorem~\cite{Atiyah:1982ih} (see \Cref{thm:Mare}), and with that tool in hand the proof of \Cref{thm:connected} goes much as it did in the complex case. \section{Quaternions and Quaternionic Frames} \label{sec:background} \subsection{Quaternions} \label{sub:quaternions} In thinking about $\mathbb{H}^m$ as a vector space over $\mathbb{H}$, some care is required. First, recall that $\mathbb{H}$ is a skew field whose elements can be written as \[ a + b \mathbf{i} + c \mathbf{j} + d \mathbf{k} \] with multiplication given by the identities \[ \mathbf{i}^2 = \mathbf{j}^2 = \mathbf{k}^2 = \mathbf{i} \mathbf{j} \mathbf{k} = -1, \] which in particular implies $\mathbf{i} \mathbf{j} = \mathbf{k}$, $\mathbf{j} \mathbf{i} = -\mathbf{k}$, etc. The products of distinct elements of $\{\mathbf{i},\mathbf{j},\mathbf{k}\}$ are encoded in the diagram \[ \begin{tikzpicture}[decoration = {markings, mark = at position 0.999 with {\arrow[>=stealth]{>}} } ] \path (90 :1cm) node (i) {$\mathbf{i}$} (330:1cm) node (j) {$\mathbf{j}$} (210:1cm) node (k) {$\mathbf{k}$}; \draw[postaction = decorate] (i) to[bend left=45] (j); \draw[postaction = decorate] (j) to[bend left=45] (k); \draw[postaction = decorate] (k) to[bend left=45] (i); \end{tikzpicture} \] which for example says that $\mathbf{k} \mathbf{i} = \mathbf{j}$ since the product follows arrows with their correct orientation and $\mathbf{k} \mathbf{j} = -\mathbf{i}$ since this product reverses directions of arrows. Instead of thinking in terms of quadruples of real numbers, it is also often convenient to interpret quaternions as pairs of complex numbers: \[ q = a+b\mathbf{i}+c\mathbf{j}+d\mathbf{k} = (a+b\mathbf{i}) + (c+d\mathbf{i})\mathbf{j}. \] This makes it easy to identify an element of $\mathbb{H}$ with a $2 \times 2$ complex matrix: \begin{equation}\label{eq:H to C map} z + w\mathbf{j} \mapsto \begin{bmatrix} z & w \\ -\overline{w} & \overline{z} \end{bmatrix}. \end{equation} This map is an algebra isomorphism between $\mathbb{H}$ and the collection of $2 \times 2$ complex matrices of the given form. The \emph{conjugate} of a quaternion $q =a + b\mathbf{i}+c\mathbf{j}+d \mathbf{k} \in \mathbb{H}$ is defined by \[ \overline{q} = a - b\mathbf{i}-c\mathbf{j}-d\mathbf{k}, \] and the \emph{modulus} by \[ |q| = \sqrt{q \overline{q}} = \sqrt{a^2+b^2+c^2+d^2}. \] Since quaternionic multiplication is non-commutative, we have to distinguish between left and right vector spaces over $\mathbb{H}$. In this paper we will follow Waldron's conventions~\cite{Waldron:2020ti} and consider $\mathbb{H}^m$ as a right vector space over $\mathbb{H}$, meaning that scalar multiplication happens on the right. This is done so that left matrix multiplication is linear: for $v_1,\dots , v_N \in \mathbb{H}^d$ thought of as column vectors, $\alpha_1,\dots , \alpha_N \in \mathbb{H}$, and a matrix $A \in \mathbb{H}^{m \times d}$, we have \[ A(v_1 \alpha_1 + \dots + v_N \alpha_N) = (A v_1) \alpha_1 + \dots + (A v_N) \alpha_N. \] Much of linear algebra over $\mathbb{H}$ goes through as over $\mathbb{R}$ or $\C$; see~\cite{Zhang:1997cd,Rodman:2014ue} for more, but in particular recall that $\mathbb{H}^m$ has a standard $\mathbb{H}$-valued Hermitian inner product given by \begin{equation}\label{eq:inner product} \langle v, w \rangle = \sum_i \overline{w}_j v_j \in \mathbb{H} \end{equation} for all $v,w \in \mathbb{H}^m$. The \emph{Frobenius inner product} on matrices $A,B \in \mathbb{H}^{m \times k}$ is given by \[ \langle A, B \rangle_F := \operatorname{tr}(B^\ast A), \] where, as usual, $B^\ast$ is the conjugate transpose of the matrix $B$. This agrees with the Hermitian inner product~\eqref{eq:inner product} on vectorized versions of the matrices. The quaternionic analog of the unitary group is usually called the \emph{symplectic group} (sometimes the \emph{compact symplectic group}) and denoted $\operatorname{Sp}(m)$. That is, a $m \times m$ quaternionic matrix $U \in \mathbb{H}^{m \times m}$ is symplectic if \[ \langle U v, U w \rangle = \langle v , w \rangle \] for all $v,w \in \mathbb{H}^m$; equivalently, $U^\ast U = \mathbb{I}_m$, the $m \times m$ identity matrix. $\operatorname{Sp}(m)$ is a compact, semisimple Lie group with type-$C$ Lie algebra $\mathfrak{sp}(m)$ consisting of the skew-Hermitian $m \times m$ quaternionic matrices; that is, those matrices $A$ so that $A^\ast = -A$. A simple parameter count shows that \[ \dim_\mathbb{R} \operatorname{Sp}(m) = \dim_\mathbb{R} \mathfrak{sp}(m) = 3m+4\frac{m(m-1)}{2} = 2m^2+m. \] Finally, we introduce the notation \[ \mathcal{H}(m) = \{A \in \mathbb{H}^{m \times m} : A^\ast = A\} \] for the $(2m^2-m)$-dimensional real vector space of $m \times m$ Hermitian quaternionic matrices, and let $\mathcal{H}_0(m)$ be the subspace of traceless Hermitian matrices. \subsection{Quaternionic Frames} \label{sub:quaternionic_frames} Frames in $\mathbb{H}^d$ are defined in the usual way~\cite{Waldron:2020ti,Khokulan:2017ic,Sharma:2019js,Virender:2020ua}: a sequence $(f_i)$ of vectors in $\mathbb{H}^d$ is a \emph{frame} if there exist $0<A\leq B < \infty$ so that \begin{equation}\label{eq:frame inequality} A\|v\|^2 \leq \sum_i |\langle v, f_i\rangle|^2 \leq B \|v\|^2. \end{equation} For finite collections of vectors the upper bound is automatically satisfied with $B = \sum_i \|f_i\|^2$ and the lower bound is satisfied when the $f_i$ span $\mathbb{H}^d$. In other words, a finite collection of vectors $f_1, \dots , f_N \in \mathbb{H}^d$ is a frame for $\mathbb{H}^d$ if and only if $\{f_1, \dots , f_N\}$ is a spanning set for $\mathbb{H}^d$. We will use $\F^{\quat^d,N}$ to denote the collection of all frames consisting of $N$ vectors in $\mathbb{H}^d$. A frame is called \emph{tight} if we can choose $A=B$ in \eqref{eq:frame inequality}, and \emph{Parseval} if $A=B=1$. We get a nice alternative characterization of these frames by introducing some operators. To do so, we will usually identify a frame $f_1, \dots , f_N$ with the $d \times N$ matrix $F = [f_1 | \dots | f_N ] \in \mathbb{H}^{d \times N}$ whose columns are the frame vectors, and we will often just write $F \in \F^{\quat^d,N}$. The \emph{analysis operator} associated to the frame is the map $\mathbb{H}^d \to \mathbb{H}^N$ given by \[ v \mapsto (\langle v, f_1 \rangle , \dots , \langle v, f_N\rangle ) = F^\ast v, \] and the \emph{synthesis operator} is the map $\mathbb{H}^N \to \mathbb{H}^d$ given by \[ w \mapsto \sum_{i=1}^N f_i w_i = F w. \] Composing these two operators one way gives the \emph{frame operator} $S:\mathbb{H}^d \to \mathbb{H}^d$ given by \[ S(v) = \sum_{i=1}^N f_i \langle v, f_i \rangle = FF^\ast v. \] In this interpretation, a tight frame has frame operator $S = FF^\ast = A \mathbb{I}_d$, so in particular a Parseval frame satisfies $FF^\ast = \mathbb{I}_d$. Composing in the other way gives the \emph{Gram matrix} $F^\ast F$ whose entries are the pairwise inner products of the frame vectors. For real and complex frames, cyclic invariance of the trace immediately implies that $\operatorname{tr}(FF^\ast) = \operatorname{tr}(F^\ast F)$, which is an important and frequently-used identity. Over the quaternions, the trace of a product of matrices is not generally invariant under cyclic permutations of terms, but the real part is: $\Re(\operatorname{tr}(AB)) = \Re(\operatorname{tr}(BA))$ for quaternionic matrices $A$ and $B$ of dimensions for which these products make sense. Since the frame operator and Gram matrix are both Hermitian, and hence have real eigenvalues and real trace, it follows that \begin{equation}\label{eq:cyclic trace} \operatorname{tr}(FF^\ast) = \Re(\operatorname{tr}(FF^\ast)) = \Re(\operatorname{tr}(F^\ast F)) = \operatorname{tr}(F^\ast F) = \sum_{i=1}^N \|f_i\|^2. \end{equation} In fact, since there is a singular value decomposition for quaternionic matrices~\cite{Zhang:1997cd,Rodman:2014ue}, the usual argument from the real and complex cases shows that the frame operator $FF^\ast \in \mathcal{H}(d)$ and the Gram matrix $F^\ast F \in \mathcal{H}(N)$ have the same nonzero (right) eigenvalues. In particular: \begin{prop}\label{prop:spectrum} If $F \in \F^{\quat^d,N}$ has frame operator $FF^\ast$ with spectrum $\boldsymbol{\lambda} = (\lambda_1, \dots , \lambda_d) \in \mathbb{R}_+^d$,\footnote{Since the columns of $F$ are a spanning set for $\mathbb{H}^d$, the frame operator $FF^\ast$ must be nonsingular, so we know $\lambda_i > 0$ for all $i=1,\dots, d$.} then the Gram matrix $F^\ast F$ has spectrum $\widetilde{\boldsymbol{\lambda}} := (\lambda_1, \dots , \lambda_d, 0 , \dots , 0)$. \end{prop} Since we are interested in frames with specified spectrum of the frame operator, we introduce some terminology and notation: \begin{definition}\label{def:frame spectrum} If $F \in \F^{\quat^d,N}$, call the spectrum of $FF^\ast$ the \emph{frame spectrum} of $F$. For given $\boldsymbol{\lambda} = (\lambda_1, \dots , \lambda_d) \in \mathbb{R}_+^d$, let $\F^{\quat^d,N}_{\boldsymbol{\lambda}}$ be the set of all frames in $\F^{\quat^d,N}$ with frame spectrum equal to $\boldsymbol{\lambda}$. \end{definition} Since Gram matrices are positive semidefinite and the spectral decomposition of Hermitian quaternionic matrices works as expected~\cite{Zhang:1997cd,Rodman:2014ue}, the same proof as in the real or complex case shows that every positive semidefinite Hermitian quaternionic matrix is the Gram matrix of a collection of vectors which is a frame for its span, and that this collection is uniquely determined up to the action of the symplectic group. In other words, if $M \in \mathcal{H}(N)$ is positive semidefinite of rank $d$, then there exists $F = [f_1 | \dots | f_N] \in \F^{\quat^d,N} \subset \mathbb{H}^{d \times N}$ so that $M = F^\ast F$. Notice that, for $U \in \operatorname{Sp}(d)$, the frame $UF$ has the same Gram matrix: $(UF)^\ast (UF) = F^\ast U^\ast U F = F^\ast F = M$. Conversely, if $G \in \mathbb{H}^{d \times N}$ so that $G^\ast G = M$, then $G = UF$ for some $U \in \operatorname{Sp}(d)$. We summarize the above discussion in the following proposition: \begin{prop}\label{prop:equivalence classes} $\operatorname{Sp}(d)$-equivalence classes of frames in $\F^{\quat^d,N}$ are uniquely determined by their Gram matrices, which consist of all rank-$d$ positive semidefinite elements of $\mathcal{H}(N)$. \end{prop} \section{Adjoint Orbits, Convexity, and Isoparametric Submanifolds} \label{sec:isoparametric} Our next goal is to connect the story of Gram matrices to adjoint orbits and isoparametric submanifolds. We will largely follow Mare's discussion~\cite[Example~5.4]{Mare:2005eo}; see \cite[\S III.7]{Helgason:1978vb} for a more general exposition of Cartan decompositions. Thinking of a square quaternionic matrix $A \in \mathbb{H}^{m \times m}$ as \[ A = Z + W \mathbf{j} \] for $Z,W \in \C^{m \times m}$, we can define a map $\Psi_m:\mathbb{H}^{m \times m} \to \C^{2m \times 2m}$ analogously to~\eqref{eq:H to C map}: \[ \Psi_m(Z+W\mathbf{j}) := \begin{bmatrix} Z & W \\ -\overline{W} & \overline{Z} \end{bmatrix}. \] If $A \in \mathbb{H}^{m \times m}$ is invertible, then $\Psi_m(A)$ is as well. Moreover, the image $\Psi_m(\operatorname{Sp}(m))$ of $\operatorname{Sp}(m)$ under this map is a subgroup of $\operatorname{SU}(2m)$, the group of $2m \times 2m$ unitary matrices with determinant 1, and in fact $\Psi_m(\operatorname{Sp}(m))$ is precisely the fixed point set inside $\operatorname{SU}(2m)$ of the involution $\sigma:\C^{2m \times 2m} \to \C^{2m \times 2m}$ given by \[ \sigma(M) = \Omega^\ast \overline{M} \Omega, \] where $\Omega = \begin{bmatrix} 0 & \mathbb{I}_m \\ -\mathbb{I}_m & 0 \end{bmatrix}$. Passing to the Lie algebra gives the Cartan decomposition \[ \mathfrak{su}(2m) = \mathfrak{k} \oplus \mathfrak{p}, \] where $\mathfrak{k}$ is the $(+1)$-eigenspace of the linearization $d_{\mathbb{I}_m}\sigma: \mathfrak{su}(2m) \to \mathfrak{su}(2m)$ and $\mathfrak{p}$ is the $(-1)$-eigenspace.\footnote{Since $d_{\mathbb{I}_m}\sigma(A) = -A^\ast$ is an involution, its only eigenvalues are $\pm 1$.} The subalgebra $\mathfrak{k}$ is just the Lie algebra of $\Psi_m(\operatorname{Sp}(m)) \simeq \operatorname{Sp}(m)$, so $\mathfrak{k} \simeq \mathfrak{sp}(m)$, with an explicit isomorphism given by $\Psi_m|_{\mathfrak{sp}(m)}: \mathfrak{sp}(m) \to \mathfrak{k}$. Notice that $[\mathfrak{k},\mathfrak{p}] \subseteq \mathfrak{p}$, and so the adjoint action of $\Psi_m(\operatorname{Sp}(m))$ on $\mathfrak{su}(2m)$ restricts to an action on $\mathfrak{p}$. At first glance this appears slightly esoteric, but we can use $\Psi_m$ to relate $\mathfrak{p}$ to a more familiar collection of matrices. Specifically, $\Psi_m|_{\mathcal{H}_0(m)}$ gives an $\operatorname{Sp}(m)$-equivariant linear isomorphism between the traceless Hermitian matrices $\mathcal{H}_0(m)$ and $\mathfrak{p}$, and the adjoint action of $\Psi_m(\operatorname{Sp}(m))$ on $\mathfrak{p}$ corresponds to the conjugation action of $\operatorname{Sp}(m)$ on $\mathcal{H}_0(m)$. Under this isomorphism, we can identify the standard maximal abelian subspace of $\mathfrak{p}$ with the set $\mathfrak{a} \subset \mathcal{H}_0(m)$ of real diagonal $m \times m$ matrices with trace $0$. The orbit $\mathcal{O}_\Lambda := \operatorname{Sp}(m)\cdot \Lambda$ of a point $\Lambda = \operatorname{diag}(\lambda_1, \dots , \lambda_m) \in \mathfrak{a}$ is simply \[ \mathcal{O}_\Lambda = \{U \Lambda U^\ast : U \in \operatorname{Sp}(m)\}, \] which is precisely the collection of Hermitian $m \times m$ quaternionic matrices with spectrum $(\lambda_1, \dots , \lambda_m)$. At this point, it is hopefully clear that we intend to use \Cref{prop:spectrum,prop:equivalence classes} to relate $\operatorname{Sp}(d)$-equivalence classes of frames to orbits of this form. Before doing so, let's see what desirable features these orbits have. Needless to say, we could have introduced these orbits without recourse to Cartan decompositions or adjoint actions. The point of approaching things in this slightly roundabout way, though, is that we can now easily see that $\mathcal{O}_\Lambda$ fits into both \emph{Kostant's convexity theorem} and the story of so-called \emph{isoparametric submanifolds}. \subsection{Kostant's Convexity Theorem} \label{sub:Kostant} Consider the Cartan decomposition \[ \mathfrak{G} = \mathfrak{K} \oplus \mathfrak{P} \] of the Lie algebra $\mathfrak{G}$ of a semisimple Lie group $G$.\footnote{In defiance of the usual convention, we are using capital fraktur letters here for the Lie algebra and its subspaces, so as not to confuse the general $\mathfrak{K}$ and $\mathfrak{P}$ discussed here with the specific $\mathfrak{k}$ and $\mathfrak{p}$ defined above.} Then, as above, $\mathfrak{K}$ is the Lie algebra of a compact Lie subgroup $K \subset G$, and the adjoint action of $K$ on $\mathfrak{G}$ restricts to an action on $\mathfrak{P}$. Let $\mathfrak{A} \subset \mathfrak{P}$ be a maximal abelian subspace and let $P:\mathfrak{P} \to \mathfrak{A}$ be the orthogonal (with respect to the Killing form) projection. The \emph{Weyl group} $W$ associated with the pair $(\mathfrak{A},\mathfrak{G})$ is the finite group $N_K(\mathfrak{A})/Z_K(\mathfrak{A})$, where $N_K(\mathfrak{A})$ is the normalizer of $\mathfrak{A}$ in $K$, and $Z_K(\mathfrak{A})$ is its centralizer. Now, let $a \in \mathfrak{A} \subset \mathfrak{P}$ and let $\mathcal{O}_a$ be the orbit of $a$ under the adjoint action of $K$ on $\mathfrak{P}$. Then Kostant's convexity theorem characterizes the image of $\mathcal{O}_a$ under $P$: \begin{thm}[Kostant~\cite{Kostant:1973ff}]\label{thm:Kostant} $P(\mathcal{O}_a) = \operatorname{conv}(W\cdot a)$, the convex hull of the Weyl orbit of $a \in \mathfrak{A}$. \end{thm} In the case where $G=\operatorname{SL}_n(\C)$ and $K = \operatorname{SU}(n)$, this is essentially the Schur--Horn theorem~\cite{schur1923uber,horn1954doubly}, which says that the diagonal entries of Hermitian matrices with fixed spectrum fill out the convex hull of the collection of vectors given by all possible re-orderings of the spectrum. With $G=\operatorname{SU}(2m)$, $K=\operatorname{Sp}(m)$, and the Cartan decomposition $\mathfrak{su}(2m) = \mathfrak{k} \oplus \mathfrak{p}$ as above, \Cref{thm:Kostant} implies the following quaternionic analog of the Schur--Horn theorem (compare to~\cite[Example~8.6]{kobert_spectrahedral_2021}): \begin{thm}\label{thm:quaternionic Schur-Horn} Let $\boldsymbol{\lambda}= (\lambda_1 , \dots , \lambda_m) \in \mathbb{R}^m$. Let $\mathcal{H}_{\boldsymbol{\lambda}}(m)$ be the collection of quaternionic Hermitian $m \times m$ matrices with spectrum $\boldsymbol{\lambda}$. Let $\Delta: \mathcal{H}(m) \to \mathbb{R}^m$ record the diagonal entries of a matrix. Then \[ \Delta(\mathcal{H}_{\boldsymbol{\lambda}}(m)) = \operatorname{conv}(S_m \cdot \boldsymbol{\lambda}), \] the convex hull of the permutation orbit of $\boldsymbol{\lambda}$. \end{thm} \begin{proof} First of all, if $\lambda_1 + \dots + \lambda_m = 0$, then $\mathcal{H}_{\boldsymbol{\lambda}}(m) \subset \mathcal{H}_0(m)$ corresponds to the orbit of $\Lambda = \operatorname{diag}(\lambda_1, \dots , \lambda_m)$ under the adjoint action of $\operatorname{Sp}(m)$ on $\mathfrak{p} \simeq \mathcal{H}_0(m)$. Moreover, the Weyl group associated to $(\mathfrak{a},\mathfrak{su}(2m))$ is the symmetric group $S_m$ and the orthogonal projection $\mathfrak{p} \to \mathfrak{a}$ corresponds to the diagonal entry map $\Delta$, so \Cref{thm:Kostant} says exactly that \[ \Delta(\mathcal{H}_{\boldsymbol{\lambda}}(m)) = \operatorname{conv}(S_m \cdot \boldsymbol{\lambda}). \] If $\lambda_1 + \dots + \lambda_m \neq 0$, then we just need to re-center, which we do using the map $\tau: \mathcal{H}_{\boldsymbol{\lambda}}(m) \to \mathcal{O}_\Lambda$ defined in \Cref{lem:re-center} below. In the notation of that lemma, \Cref{thm:Kostant} implies that \[ (\Delta \circ \tau)(\mathcal{H}_{\boldsymbol{\lambda}}(m)) = \operatorname{conv}(t(S_m \cdot \boldsymbol{\lambda})). \] Since taking the convex hull is $t$-equivariant and since $\Delta \circ \tau = t \circ \Delta$, it follows that $\Delta(\mathcal{H}_{\boldsymbol{\lambda}}(m)) = \operatorname{conv}(S_m \cdot \boldsymbol{\lambda})$, as desired. \end{proof} \begin{lem}\label{lem:re-center} Let $\boldsymbol{\lambda} = (\lambda_1, \dots , \lambda_m) \in \mathbb{R}^m$, let $\sigma = \lambda_1 + \dots + \lambda_m$, and define $\Lambda := \operatorname{diag}(\lambda_1, \dots , \lambda_m)-\frac{\sigma}{m} \mathbb{I}_m$. Then the map $B \mapsto B - \frac{\sigma}{m} \mathbb{I}_m$ defines a diffeomorphism $\tau: \mathcal{H}_{\boldsymbol{\lambda}}(m) \to \mathcal{O}_\Lambda$ which is equivariant with respect to the $\operatorname{Sp}(m)$ actions on domain and range and makes the following diagram commute: \[\begin{tikzcd} {\mathcal{H}_{\boldsymbol{\lambda}}(m)} && {\mathcal{O}_\Lambda} \\ \\ {\mathbb{R}^m} && {\mathbb{R}^m} \arrow["\tau", from=1-1, to=1-3] \arrow["\Delta"', from=1-1, to=3-1] \arrow["t", from=3-1, to=3-3] \arrow["\Delta", from=1-3, to=3-3] \end{tikzcd}\] Here $t: \mathbb{R}^m \to \mathbb{R}^m$ is the translation map $t(\boldsymbol{x}) := \boldsymbol{x}-\left(\frac{\sigma}{m}, \dots , \frac{\sigma}{m}\right)$. \end{lem} \begin{proof} First, $\tau$ is certainly well-defined and smooth with smooth inverse, so it is a diffeomorphism. If $B \in \mathcal{H}_{\boldsymbol{\lambda}}(m)$ and $U \in \operatorname{Sp}(d)$, then \[ U \tau(B) U^\ast = U\left(B - \frac{\sigma}{m}\mathbb{I}_m\right)U^\ast = UBU^\ast - \frac{\sigma}{m}\mathbb{I}_m = \tau(UBU^\ast), \] so $\tau$ is $\operatorname{Sp}(m)$-equivariant. Finally, for $B \in \mathcal{H}_{\boldsymbol{\lambda}}(m)$, \[ (\Delta \circ \tau)(B) = \Delta\left(B - \frac{\sigma}{m}\mathbb{I}_m\right) = \Delta(B) - \left(\frac{\sigma}{m},\dots , \frac{\sigma}{m}\right) = (t \circ \Delta)(B), \] so the diagram commutes. \end{proof} \subsection{Isoparametric Submanifolds} \label{sub:isoparametric submanifolds} Isoparametric hypersurfaces were first studied by Cartan~\cite{Cartan:1938kh,Cartan:1939gv,Cartan:1939wj,Cartan:tw,Nomizu:1975bq}; for a more modern and general introduction see \cite[Chapter~6]{Palais:1988ks} or~\cite{Terng:1985eu}. First, we give the definition, the details of which will not concern us greatly: \begin{definition}\label{def:isoparametric} A submanifold $M$ of a Riemannian manifold $N$ is \emph{isoparametric} if its normal bundle $\nu(M)$ is flat and the principal curvatures along any parallel normal field of $M$ are constant. \end{definition} Principal orbits of isotropy representations are classic examples of isoparametric submanifolds. In the context of our story, we have: \begin{prop}[{see, e.g.,~\cite[Example~6.5.6]{Palais:1988ks}}]\label{prop:orbits are isoparametric} If $\Lambda \in \mathfrak{a}$ is generic (i.e., $\lambda_i \neq \lambda_j$ for all $i \neq j$), then $\mathcal{O}_\Lambda$ is an isoparametric submanifold of $\mathcal{H}_0(N)$, thought of as a Riemannian manifold by taking the Frobenius inner product on each tangent space. Moreover, the normal space to $\mathcal{O}_\Lambda$ at $\Lambda$ is just $\mathfrak{a}$. If $\Lambda$ is not generic, then $\mathcal{O}_\Lambda$ is parallel to an isoparametric submanifold of $\mathcal{H}_0(N)$. \end{prop} To explain the terminology in the previous sentence, if $M \subset \mathbb{R}^n$ is isoparametric and $\xi$ is a parallel section of the normal bundle $\nu(M)$, then $M_\xi := \{p + \xi(p) : p \in M\}$ is a \emph{parallel submanifold} to $M$, and these parallel submanifolds give a singular foliation of the ambient space. In the special case of isotropy orbits, the orbit foliation and the parallel foliation coincide. Isoparametric submanifolds and their parallels have some of the nice features of symplectic manifolds admitting Hamiltonian torus actions without necessarily being symplectic. For example, the following fundamental result in symplectic geometry has an analog in the isoparametric setting. \begin{thm}[{Atiyah~\cite{Atiyah:1982ih} and Guillemin--Sternberg~\cite{Guillemin:1982gx}}]\label{thm:AGS} Let $(M,\omega)$ be a compact symplectic manifold admitting a Hamiltonian action of a torus $T\simeq \operatorname{U}(1)^n$. Let $\mu:M \to \mathbb{R}^n$ be the associated momentum map. \begin{itemize} \item For any $v \in \mathbb{R}^n$, $\mu^{-1}(v)$ is either empty or connected. \item $\mu(M)$ is the convex hull of the images of the fixed points of the $T$-action. \end{itemize} \end{thm} \begin{example} Let $K$ be a compact Lie group of rank $n$ with Lie algebra $\mathfrak{K}$. Let $\mathcal{O} \subset \mathfrak{K}^\ast$ be an orbit of the coadjoint action of $K$ on $\mathfrak{K}^\ast$. This action of $K$ on $\mathcal{O}$ is Hamiltonian with momentum map being the inclusion $\mathcal{O} \hookrightarrow \mathfrak{K}^\ast$~\cite[Example~5.3.11]{mcduff2017introduction}. If $T \subset K$ is a maximal torus, then $T \simeq U(1)^n$ and the coadjoint action of $T$ on $\mathcal{O}$ is also Hamiltonian, with moment map given by the restriction of the projection $\mathfrak{K}^\ast \to \mathfrak{T}^\ast \simeq \mathbb{R}^n$ induced by the inclusion $T \hookrightarrow K$~\cite[Proposition~II.1.10]{Audin:2004bh}. Then \Cref{thm:AGS} implies that the image of this map is convex and its non-empty level sets are connected. In this case, convexity actually follows from \Cref{thm:Kostant}: let $G = K_\C$ be the complexification of $K$, with Lie algebra $\mathfrak{G} = \mathfrak{K} \oplus \mathbf{i} \mathfrak{K}$. This is a Cartan decomposition and the dual Lie algebra $\mathfrak{K}^\ast$ can be identified with the complementary subspace $\mathbf{i} \mathfrak{K}$ so that the coadjoint action of $K$ on $\mathfrak{K}^\ast$ corresponds to the adjoint action of $K$ on $\mathbf{i} \mathfrak{K}$ and $\mathfrak{T}^\ast \subset \mathfrak{K}^\ast$ corresponds to a maximal abelian subalgebra $\mathfrak{A} \subset \mathbf{i} \mathfrak{K}$. Hence, the projection map that \Cref{thm:Kostant} says has convex image is just the momentum map of the torus action. Of course, Kostant's convexity theorem was a major inspiration for Atiyah and Guillemin--Sternberg, whose result can be interpreted as generalizing this case of Kostant's theorem to arbitrary Hamiltonian torus actions. On the other hand, if $\mathcal{O}$ is a principal orbit, then $\mathcal{O}$ is an isoparametric submanifold of $\mathfrak{K}^\ast$ and the normal space at a point can be identified with $\mathfrak{A} \simeq \mathfrak{T}^\ast$. This gives yet another interpretation of the momentum map as the orthogonal projection onto the normal space at a point. Terng showed that, under this interpretation, the convexity part of \Cref{thm:AGS} generalizes to arbitrary isoparametric submanifolds: \end{example} \begin{thm}[{Terng~\cite{Terng:1986hg}}]\label{thm:Terng convexity} Let $M_\xi \subset \mathbb{R}^{n}$ be parallel to an isoparametric submanifold $M$. Let $p \in M$ so that $p + \xi(p) \in M_\xi$, let $\nu_p(M)$ be the normal space to $M$ at $p$, and let $P:M_\xi \to \nu_p(M)$ be orthogonal projection. Then $P(M_\xi)$ is a convex polytope. \end{thm} Given this, the projection map $P$ is like a momentum map for isoparametric submanifolds and their parallels, even though these submanifolds need not be symplectic. It is then reasonable to ask whether, as in the first part of \Cref{thm:AGS}, the level sets of $P$ are connected. They needn't be in general, but Mare gave a sufficient condition for all non-empty level sets of $P$ to be connected: \begin{thm}[{Mare~\cite[Theorem~1.2 and Remark~1.3(a)]{Mare:2005eo}}]\label{thm:Mare} Let $M \subset \mathbb{R}^n$ be an isoparametric submanifold with all multiplicities $\geq 2$ and let $M_\xi$ be parallel to $M$. If $p \in M$ and $b \in \nu_p(M)$ is in the image of the projection $P: M_\xi \to \nu_p(M)$, then $P^{-1}(b)$ is connected. \end{thm} See Mare's paper for a general definition of multiplicities; in the setting of \Cref{thm:Kostant}, where we are considering orbits of the adjoint action of $K \subset G$ on the complementary subspace $\mathfrak{P} \subset \mathfrak{G}$, the multiplicities are the differences in dimension between a principal orbit and subprincipal orbits~\cite{Palais:1988ks,Hsiang:1988ka}. Recall from \Cref{prop:orbits are isoparametric} that the normal space can be identified with the maximal abelian subspace $\mathfrak{A}$. \section{Existence} \label{sec:admissibility} Let $N \geq d \geq 1$ be integers and let $\boldsymbol{\lambda}=(\lambda_1, \dots, \lambda_d)$ and $\boldsymbol{r} = (r_1, \dots , r_N)$ be lists of positive numbers. Our main object of interest is the collection $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ of frames $F =[f_1 | \dots | f_N] \in \F^{\quat^d,N} \subset \mathbb{H}^{d \times N}$ so that $FF^\ast$ has spectrum $\boldsymbol{\lambda}$ and $\|f_i\|^2 = r_i$ for $i=1, \dots , N$. Our first goal is to answer the question: are there any such frames? The analogous question for real and complex frames was answered by Casazza and Leon~\cite{Casazza:2010ti}, and the answer in the quaternionic case is essentially the same: \begin{thm}\label{thm:admissibility} Let $N$, $d$, $\boldsymbol{\lambda}$, and $\boldsymbol{r}$ be as above. The space $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is non-empty if and only if $\boldsymbol{r} \in \operatorname{conv}(S_N \cdot \widetilde{\boldsymbol{\lambda}})$, the convex hull of the collection of vectors in $\mathbb{R}^N$ given by permuting the entries of $\widetilde{\boldsymbol{\lambda}} := (\lambda_1, \dots , \lambda_d,0,\dots,0)$. \end{thm} \begin{proof} \Cref{thm:quaternionic Schur-Horn} will be the key to the proof, so the goal is to relate $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ to a space of the form $\mathcal{H}_{\boldsymbol{\eta}}(m)$ for some choice of $\boldsymbol{\eta}$ and $m$. Recall from \Cref{prop:equivalence classes} that $\operatorname{Sp}(d)$-equivalence classes of frames in $\F^{\quat^d,N}$ are determined by their Gram matrices. Specifically, the space $\F^{\quat^d,N}_{\boldsymbol{\lambda}}/\operatorname{Sp}(d)$ of $\operatorname{Sp}(d)$-equivalence classes of frames with frame spectrum $\boldsymbol{\lambda}$ is diffeomorphic to the space of all possible Gram matrices of frames in $\F^{\quat^d,N}_{\boldsymbol{\lambda}}$. Now, if $F \in \F^{\quat^d,N}_{\boldsymbol{\lambda}}$, then definitionally $FF^\ast$ has spectrum $\boldsymbol{\lambda}$, and hence the Gram matrix $F^\ast F$ has spectrum $\widetilde{\boldsymbol{\lambda}}=(\lambda_1, \dots , \lambda_d, 0, \dots , 0) \in \mathbb{R}^N$. Conversely, any matrix in $\mathcal{H}_{\widetilde{\boldsymbol{\lambda}}}(N)$ can be realized as the Gram matrix of a frame in $\F^{\quat^d,N}_{\boldsymbol{\lambda}}$, so we see that \[ \F^{\quat^d,N}_{\boldsymbol{\lambda}}/\operatorname{Sp}(d) \simeq \mathcal{H}_{\widetilde{\boldsymbol{\lambda}}}(N). \] In turn, if $\Delta:\mathcal{H}_{\widetilde{\boldsymbol{\lambda}}}(N) \to \mathbb{R}^N$ is the map which records diagonal entries, then the level set $\Delta^{-1}(\boldsymbol{r})$ is the collection of $\operatorname{Sp}(d)$-equivalence classes of frames with frame spectrum $\boldsymbol{\lambda}$ and squared frame norms $\boldsymbol{r}$; that is, \[ \Delta^{-1}(\boldsymbol{r}) \simeq \F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})/\operatorname{Sp}(d). \] This means that $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is non-empty if and only if $\boldsymbol{r}$ is in the image of $\Delta$. But now \Cref{thm:quaternionic Schur-Horn} tells us that the image of $\Delta$ is exactly the convex hull of $S_N \cdot \widetilde{\boldsymbol{\lambda}}$, as desired. \end{proof} Since permuting entries doesn't change their sum, $\operatorname{conv}(S_N \cdot \widetilde{\boldsymbol{\lambda}})$ lies in the affine hyperplane $\{(x_1, \dots , x_N) : \sum_{i=1}^N x_i = \sum_{i=1}^d \lambda_i\}$. Hence, $\boldsymbol{r} \in \operatorname{conv}(S_N \cdot \widetilde{\boldsymbol{\lambda}})$ only if \begin{equation}\label{eq:admissibility equality} \sum_{i=1}^N r_i = \sum_{i=1}^d \lambda_i. \end{equation} This is just a restatement of~\eqref{eq:cyclic trace}, which said the frame operator and the Gram matrix have the same trace. In practice, the order of the frame vectors is no more than a bookkeeping convenience, so it is no problem to permute the frame vectors. Likewise, we can freely permute the numbers comprising the frame spectrum. In particular, we can get a more straightforward criterion for non-emptiness of $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ by sorting $\boldsymbol{\lambda}$ and $\boldsymbol{r}$ in non-increasing order. Specifically, if we make the additional assumption that $\lambda_1 \geq \dots \geq \lambda_d > 0$ and $r_1 \geq \dots \geq r_N > 0$, then we see that $\boldsymbol{r} \in \operatorname{conv}(S_N \cdot \widetilde{\boldsymbol{\lambda}})$ if and only if it satisfies~\eqref{eq:admissibility equality} and \begin{equation}\label{eq:admissibility inequality} \sum_{i=1}^k r_i \leq \sum_{i=1}^k \lambda_i \quad \text{for all }k=1, \dots , d. \end{equation} We call the (sorted) $\boldsymbol{r}$ satisfying~\eqref{eq:admissibility equality} and~\eqref{eq:admissibility inequality} \emph{$\boldsymbol{\lambda}$-admissible}. Notice that the admissibility criterion is exactly the same as that given by Casazza and Leon in the real and complex cases~\cite{Casazza:2010ti}. Thus, \Cref{thm:admissibility} is equivalent to \Cref{thm:existence}, which we restate in a more compact form: \begin{existence} Let $N$, $d$, $\boldsymbol{\lambda}$, and $\boldsymbol{r}$ be as above, so that $\boldsymbol{\lambda}$ and $\boldsymbol{r}$ are sorted in non-increasing order. Then $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is non-empty if and only if $\boldsymbol{r}$ is $\boldsymbol{\lambda}$-admissible. \end{existence} \section{Connectedness} \label{sec:connectedness} Next, we turn to the question of when $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is connected. As in the previous section, we will focus on $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})/\operatorname{Sp}(d)$, since the following lemma combined with the fact that $\operatorname{Sp}(d)$ is connected implies the quotient is connected if and only if $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is connected. \begin{lem}\label{lem:connected quotient} Let $X$ be a topological space and let $G$ be a connected topological group acting continuously on $X$. If $X/G$ is connected, then $X$ is connected. \end{lem} This lemma follows from a standard point-set topology argument (see, e.g., \cite[Exercise~5.5]{Manetti:2015ep}) since connectedness of $G$ implies the fibers of the quotient map $X \to X/G$ are connected. The strategy is to prove connectedness using \Cref{thm:Mare}, which applies to adjoint orbits of a compact group acting on the complementary subspace of the Lie algebra of some larger group $G$. Following the setup in \Cref{sec:isoparametric}, let $G=\operatorname{SU}(2N)$, so that $K=\operatorname{Sp}(N)$ and the complementary subspace $\mathfrak{p} \subset \mathfrak{su}(2N)$ can be identified with $\mathcal{H}_0(N)$, the space of traceless Hermitian quaternionic $N \times N$ matrices, and the standard maximal abelian subspace of $\mathfrak{p}$ corresponds to the subset $\mathfrak{a} \subset \mathcal{H}_0(N)$ of real diagonal $N \times N$ matrices with trace 0. We saw in the proof of \Cref{thm:admissibility} that $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})/\operatorname{Sp}(d) \simeq \mathcal{H}_{\widetilde{\boldsymbol{\lambda}}}(N)$, the space of quaternionic Hermitian $N \times N$ matrices with spectrum $\widetilde{\boldsymbol{\lambda}} = (\lambda_1, \dots , \lambda_d, 0, \dots 0)$. Since $\lambda_1 + \dots + \lambda_d \neq 0$, this is not quite an orbit of the adjoint action of $\operatorname{Sp}(N)$ on $\mathcal{H}_0(N)$, but it is a simple translation of such an orbit. Letting $\sigma:=\lambda_1 + \dots + \lambda_d$ and \[ \widetilde{\Lambda} := \operatorname{diag}\left(\lambda_1 ,\dots , \lambda_d, 0, \dots, 0\right) - \frac{\sigma}{N} \mathbb{I}_N \in \mathcal{H}_0(N), \] define $\tau: \mathcal{H}_{\widetilde{\boldsymbol{\lambda}}}(N) \to \mathcal{O}_{\widetilde{\Lambda}}$ by $\tau(B) = B - \frac{\sigma}{N} \mathbb{I}_N$. If $\boldsymbol{r}$ is $\boldsymbol{\lambda}$-admissible, then $\Delta^{-1}(\boldsymbol{r}) \subset \mathcal{H}_{\boldsymbol{\lambda}}(N)$ is non-empty, and hence so is \[ \tau(\Delta^{-1}(\boldsymbol{r})) = \Delta^{-1}(t(\boldsymbol{r})) \subset \mathcal{O}_{\widetilde{\Lambda}}, \] where $t: \mathbb{R}^N \to \mathbb{R}^N$ is defined by $t(\boldsymbol{x}) := \boldsymbol{x} - \left(\frac{\sigma}{N},\dots , \frac{\sigma}{N}\right)$ and the above equality follows from the fact that the diagram in \Cref{lem:re-center} commutes. But now \[ \Delta^{-1}(t(\boldsymbol{r})) = P^{-1}(\operatorname{diag}(t(\boldsymbol{r}))), \] where $P: \mathcal{O}_{\widetilde{\Lambda}} \to \mathfrak{a}$ is the orthogonal projection. Since $\mathcal{O}_{\widetilde{\Lambda}}$ is an isotropy orbit\footnote{In fact, it is an example of a quaternionic flag manifold, and much is known about its cohomology~\cite{Mare:2008iz,Hsiang:1988ka,Mare:2006bf}.} and hence also a parallel submanifold to some principal orbit, and since all multiplicities in this setting are equal to 4~\cite[\S 3.3]{Hsiang:1988ka}, connectedness of $\Delta^{-1}(t(\boldsymbol{r}))$ follows from \Cref{thm:Mare}. Since $\tau$ is a diffeomorphism, we conclude that \[ \Delta^{-1}(\boldsymbol{r}) \simeq \F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})/\operatorname{Sp}(d) \] is connected. Finally, applying \Cref{lem:connected quotient} shows that $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is connected whenever it is non-empty; i.e., whenever $\boldsymbol{r}$ is $\boldsymbol{\lambda}$-admissible. In turn, since $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is a real algebraic set in $\mathbb{H}^{d \times N} \simeq \mathbb{R}^{4dN}$, it is locally path-connected (in fact, triangulable by \L{}ojasiewicz's triangulation theorem \cite{lojasiewicz1964triangulation}) so that connectivity implies path-connectivity. Thus, we have proved \Cref{thm:connected}, which we now restate: \begin{connectedness} For any $\boldsymbol{r}$ and $\boldsymbol{\lambda}$, the space $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is path-connected. \end{connectedness} If we prefer to focus on the space \[ \F^{\quat^d,N}_S(\boldsymbol{r}) := \{F \in \F^{\quat^d,N} : FF^\ast = S \text{ and } \|f_i\|^2 = r_i \text{ for all } i = 1 , \dots , N\} \] of frames with fixed frame operator (rather than fixed frame spectrum) and fixed frame vector norms, connectivity still holds: \begin{cor}\label{cor:fixed frame operator} If $S \in \mathcal{H}(N)$ is positive definite with spectrum $\boldsymbol{\lambda}$ and $\boldsymbol{r} \in \mathbb{R}_+^N$, the space $\F^{\quat^d,N}_S(\boldsymbol{r})$ is (i) non-empty if and only if $\boldsymbol{r}$ is $\boldsymbol{\lambda}$-admissible and (ii) path-connected. \end{cor} \begin{proof} If $\F^{\quat^d,N}_S(\boldsymbol{r}) \subset \F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is non-empty, then certainly $\boldsymbol{r}$ is $\boldsymbol{\lambda}$-admissible by \Cref{thm:admissibility}. Conversely, if $\boldsymbol{r}$ is $\boldsymbol{\lambda}$-admissible, then $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is non-empty, so that there exists $F \in \F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$, and $FF^\ast = V S V^\ast$ for some $V \in \operatorname{Sp}(d)$. But then $V^\ast F \in \F^{\quat^d,N}_S(\boldsymbol{r})$, so we see that $\F^{\quat^d,N}_S(\boldsymbol{r})$ is non-empty as well. Since the empty set is trivially path-connected, the only thing to prove is that $\F^{\quat^d,N}_S(\boldsymbol{r})$ is path-connected whenever it is non-empty. In this case, we know from \Cref{thm:connected} that $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ is path-connected. So if $F_0, F_1 \in \F^{\quat^d,N}_S(\boldsymbol{r}) \subset \F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$, then there exists a path $\widetilde{\gamma}: [0,1] \to \F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ so that $\widetilde{\gamma}(0) = F_0$ and $\widetilde{\gamma}(1) = F_1$. Of course, the path $\widetilde{\gamma}$ need not stay in $\F^{\quat^d,N}_S(\boldsymbol{r})$, but we can easily fix it up to do so. Indeed, the frame operator \[ \widetilde{\gamma}(t)\widetilde{\gamma}(t)^\ast = U_t S U_t^\ast \] determines a continuous path $\{U_t: t \in [0,1]\} \subset \operatorname{Sp}(d)$, and hence \[ \gamma(t) := U_t^\ast \widetilde{\gamma}(t) \] is a path in $\F^{\quat^d,N}_S(\boldsymbol{r})$ connecting $F_0$ and $F_1$. \end{proof} \section{Discussion} \Cref{cor:fixed frame operator} is the exact analog of the main theorem in our previous paper~\cite{NeedhamSGC} on complex frames. While our argument in the complex case was based on symplectic geometry, the strategy employed here for proving \Cref{thm:connected} and \Cref{cor:fixed frame operator} can be adapted to the complex setting, so we view this approach as somewhat more general. Indeed, much of the setup goes through in the real case as well: orthogonal equivalence classes of real frames with fixed frame spectrum correspond to adjoint orbits by way of their Gram matrices, and hence are parallel to isoparametric submanifolds of the space of symmetric matrices. Unfortunately, all multiplicities in this case are equal to 1~\cite[\S 3.2]{Hsiang:1988ka}, so \Cref{thm:Mare} does not apply. Indeed, while some of the real frame spaces $\mathcal{F}^{\mathbb{R}^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$ are connected~\cite{Cahill:2017gv}, others are not~\cite{Kapovich:1995wg} (compare with~\cite[Theorem~2.7]{Goyal:2001cd}), and it seems challenging (but interesting!) to characterize which $\boldsymbol{\lambda}$ and $\boldsymbol{r}$ lead to which outcome. Since the language of isoparametric submanifolds provides a common framework for understanding real, complex, and quaternionic frames, this perspective seems ripe for further exploration. \Cref{thm:existence,thm:connected} have the same statements as the corresponding results in the complex case, which gives some reason to hope that direct translations of other results about complex frames might give true statements about quaternionic frames. For example, it seems likely that there is a quaternionic extension of the eigenstep method~\cite{Cahill:2013jv} which would construct all elements of $\F^{\quat^d,N}_{\boldsymbol{\lambda}}(\boldsymbol{r})$, and more generally give a constructive proof of \Cref{thm:quaternionic Schur-Horn} (cf.~\cite{Fickus:2013ki}). \subsection*{Acknowledgments} We are very grateful to Augustin-Liviu Mare, Emily King, and Colin Roberts for providing inspiration and sharing their knowledge and insight. This work was partially supported by grants from the National Science Foundation (DMS--2107808, Tom Needham; DMS--2107700, Clayton Shonkwiler).

https://arxiv.org/abs/1703.09618

On generalization of D'Aurizio-Sándor trigonometric inequalities with a parameter

In this work, we generalize the D'Aurizio-Sándor inequalities (\cite{D'Aurizio,Sandor}) using an elementary approach. In particular, our approach provides an alternative proof of the D'Aurizio-Sándor inequalities. Moreover, as an immediate consequence of the generalized D'Aurizio-Sándor inequalities, we establish the D'Aurizio-Sándor-type inequalities for hyperbolic functions.

\section{Introduction} Based on infinite product expansions and inequalities on series and the Riemann's zeta function, D'Aurizio (\cite{D'Aurizio}) proved the following inequality: \begin{equation}\label{eqn: D'Aurizio ineq} \frac{1-\displaystyle\frac{\cos x}{\cos \frac{x}{2}}}{x^2}<\frac{4}{\pi^2}, \end{equation} where $x\in(0,\pi/2)$. Using an elementary approach, S\'andor (\cite{Sandor}) offered an alternative proof of \eqref{eqn: D'Aurizio ineq} by employing trigonometric inequalities and an auxiliary function. In the same paper, S\'andor also provided the converse to \eqref{eqn: D'Aurizio ineq}: \begin{equation} \frac{1-\displaystyle\frac{\cos x}{\cos \frac{x}{2}}}{x^2}>\frac{3}{8}, \end{equation} where $x\in(0,\pi/2)$. In addition, S\'andor found the following analogous inequality \eqref{eqn: Sandor ineq} holds true for the case of sine functions: \begin{theorem}[D'Aurizio-S\'andor inequalities (\cite{D'Aurizio,Sandor})]\label{thm: Daurizio-Sandor ineq original} The two double inequalities \begin{equation} \frac{3}{8}<\frac{1-\displaystyle\frac{\cos x}{\cos \frac{x}{2}}}{x^2}<\frac{4}{\pi^2} \end{equation} and \begin{equation}\label{eqn: Sandor ineq} \frac{4}{\pi^2}(2-\sqrt{2})<\frac{\displaystyle 2-\frac{\sin x}{\sin \frac{x}{2}}}{x^2}<\frac{1}{4} \end{equation} hold for any $x\in(0,\pi/2)$. \end{theorem} Throughout this paper, we denote $\frac{1-\frac{\cos x}{\cos \frac{x}{p}}}{x^2}$ and $\frac{p-\frac{\sin x}{\sin \frac{x}{p}}}{x^2}$ by $f_c(x)$ and $f_s(x)$, respectively: \begin{align} f_{p}^{c}(x)= & \frac{1-\displaystyle\frac{\cos x}{\cos \frac{x}{p}}}{x^2}, \\ f_{p}^{s}(x)= & \frac{p-\displaystyle\frac{\sin x}{\sin \frac{x}{p}}}{x^2}. \end{align} Our aim is to generalize the D'Aurizio-S\'andor inequalities for the case of $f_{p}^{c}(x)$ and $f_{p}^{s}(x)$ as follows: \begin{theorem}[Generalized D'Aurizio-S\'andor inequalities]\label{thm: Daurizio-Sandor ineq generalized} Let $0<x<\pi/2$. Then the two double inequalities \begin{equation}\label{eqn: cos ineq p>2} \frac{4}{\pi^2}<\frac{1-\displaystyle\frac{\cos x}{\cos \frac{x}{p}}}{x^2}<\frac{p^2-1}{2\,p^2} \end{equation} and \begin{equation}\label{eqn: sin ineq p>2} \frac{4}{\pi^2}\left(p-\csc \left(\frac{\pi}{2p}\right)\right)<\frac{\displaystyle p-\frac{\sin x}{\sin \frac{x}{p}}}{x^2}<\frac{p^2-1}{6\,p} \end{equation} hold for $p=3, 4 , 5,\cdots$. In particular, the double inequality \eqref{eqn: sin ineq p>2} remains true when $p=2$ while the double inequality \eqref{eqn: cos ineq p>2} is reversed when $p=2$. \end{theorem} The remainder of this paper is organized as follows. Section~\ref{sec: proof of the main results} is devoted to the proof of Theorem~\ref{thm: Daurizio-Sandor ineq generalized} and an alternative proof of Theorem~\ref{thm: Daurizio-Sandor ineq original}. In Section~\ref{sec: D'Aurizio-Sandor inequalities for hyperbolic functions}, we establish analogue of Theorem~\ref{thm: Daurizio-Sandor ineq generalized} for hyperbolic functions. As an application of Theorem~\ref{thm: Daurizio-Sandor ineq generalized}, we apply in Section~\ref{sec: application} inequality \eqref{eqn: sin ineq p>2} to the Chebyshev polynomials of the second kind and establish a trigonometric inequality. \section{Proof of the main results}\label{sec: proof of the main results} At first we will prove the following lemma. The lemma provides expressions of the higher-order derivative $\frac{d^2}{dx^2}(x^3\frac{d}{dx} f_{p}^{\triangle}(x))$ involving $f_{p}^{\triangle}(x)$ $(\triangle=c, s)$, which are helpful in proving Theorem~\ref{thm: Daurizio-Sandor ineq generalized}. We note that the sign of $\frac{d^2}{dx^2}(x^3\frac{d}{dx} f_{p}^{\triangle}(x))$ plays a crucial role in proving Theorem~\ref{thm: Daurizio-Sandor ineq generalized}. \begin{lemma}\label{lem: lots of f'''} Let $0<x<\pi/2$ and $k=1,2,3,\cdots$. Then when $p\in\mathbb{R}$ and $p\neq0$, we hav \begin{description} \item[$(i)$] \begin{align}\label{eqn: f''' cos} \nonumber \frac{d^2}{dx^2}& \left(x^3\frac{d}{dx} f_{p}^{c}(x)\right) =-\frac{x\,\csc ^4\left(\frac{x}{p}\right)}{8\,p^3} \bigg( \left(p+1\right)^3 \sin \Big(x-\frac{3 x}{p}\Big)+\left(p-1\right)^3 \sin \Big(x+\frac{3 x}{p}\Big)\\ +& \left(3 p^3+3 p^2-15 p-23\right) \sin \Big(x-\frac{x}{p}\Big)+\left(3 p^3-3 p^2-15 p+23\right) \sin\Big(x+\frac{x}{p}\Big)\bigg); \end{align} \item[$(ii)$] \begin{align}\label{eqn: f''' sin} \nonumber \frac{d^2}{dx^2}& \left(x^3\frac{d}{dx} f_{p}^{s}(x)\right) =\frac{x\,\csc ^4\left(\frac{x}{p}\right)}{8\,p^3} \bigg( \left(p+1\right)^3 \sin \Big(x-\frac{3 x}{p}\Big)-\left(p-1\right)^3 \sin \Big(x+\frac{3 x}{p}\Big)\\ +& \left(-3 p^3-3 p^2+15 p+23\right) \sin \Big(x-\frac{x}{p}\Big)+\left(3 p^3-3 p^2-15 p+23\right) \sin\Big(x+\frac{x}{p}\Big)\bigg). \end{align} In particular, \item[$(iii)$] when $p=2\,k$, \begin{equation}\label{eqn: 2k sin} \frac{d^2}{dx^2} \left(x^3\frac{d}{dx} f_{p}^{s}(x)\right)=-\frac{x}{4\,k^3}\sum_{j=0}^{k-1} (2\,j+1)^3\,\sin\left(\frac{2\,j+1}{2\,k}\,x\right); \end{equation} \item[$(iv)$] when $p=2\,k+1$, \begin{align} \label{eqn: 2k+1 cos} \frac{d^2}{dx^2} \left(x^3\frac{d}{dx} f_{p}^{c}(x)\right)=&-\frac{16\,x}{(2\,k+1)^3}\sum_{j=1}^{k} j^3\,\sin\left(\frac{2\,j}{2\,k+1}\,x\right)\,(-1)^{j-1}, \\ \label{eqn: 2k+1 sin} \frac{d^2}{dx^2} \left(x^3\frac{d}{dx} f_{p}^{s}(x)\right)=&-\frac{16\,x}{(2\,k+1)^3}\sum_{j=1}^{k} j^3\,\sin\left(\frac{2\,j}{2\,k+1}\,x\right). \end{align} \item[$(v)$] For $\triangle=c, s$ and $p\in\mathbb{R}\setminus\{0\}$, \begin{equation} \displaystyle \lim_{x\to0} \frac{d}{dx}\left(x^3\frac{d}{dx} f_{p}^{\triangle}(x)\right)=\lim_{x\to0} x^3\frac{d}{dx} f_{p}^{\triangle}(x)=0. \end{equation} \end{description} \end{lemma} \begin{proof} $(i)$, $(ii)$ and $(v)$ follows directly from calculations using elementary Calculus. In particular, trigonometric addition formulas are used in proving $(i)$ and $(ii)$. To prove \eqref{eqn: 2k sin}, we claim \begin{equation -\frac{x}{4\,k^3}\sum_{j=0}^{k-1} (2\,j+1)^3\,\sin\left(\frac{2\,j+1}{2\,k}\,x\right)=-\,x\,\frac{d^3}{dx^3} \left( \frac{\sin x}{\sin\left(\frac{x}{2\,k}\right)}\right). \end{equation} Indeed, we rewrite \begin{equation}\label{eqn: trans f''' p=2k} \frac{1}{4\,k^3}\sum_{j=0}^{k-1} (2\,j+1)^3\,\sin\left(\frac{2\,j+1}{2\,k}\,x\right) = 2\frac{d^3}{dx^3} \left(\sum_{j=0}^{k-1}\cos\left(\frac{2\,j+1}{2\,k}\,x\right)\right). \end{equation} On the other hand, making use of Euler's formula $e^{i\,z}=\cos z+i\,\sin z$ leads to an alternative expression of the left-hand side of \eqref{eqn: trans f''' p=2k}: \begin{align} \label{} \sum_{j=0}^{k-1}\cos\left(\frac{2\,j+1}{2\,k}\,x\right) &= \sum_{j=0}^{k-1} \displaystyle\Re\left\{ e^{i(\frac{x}{2\,k}+\frac{x}{k}\,j)}\right\} = \Re\left\{e^{i\frac{x}{2\,k}}\sum_{j=0}^{k-1} \displaystyle \left(e^{i\frac{x}{k}}\right)^j\right\} \\ &= \Re\left\{e^{i\frac{x}{2\,k}}\displaystyle\frac{1-e^{i\,x}}{1-e^{i\frac{x}{k}}} \right\} = \Re\left\{e^{i\frac{x}{2\,k}}\displaystyle\frac{e^{\frac{i\,x}{2}}(e^{-\frac{i\,x}{2}}-e^{\frac{i\,x}{2}})}{e^{\frac{i\,x}{2\,k}}(e^{-\frac{i\,x}{2\,k}}-e^{\frac{i\,x}{2\,k}})} \right\}\\ &= \Re\left\{ e^{\frac{i\,x}{2}}\,\frac{\sin \left(\frac{x}{2}\right)}{\sin \left(\frac{x}{2\,k}\right)}\right\} =\cos \left(\frac{x}{2}\right)\frac{\sin \left(\frac{x}{2}\right)}{\sin \left(\frac{x}{2\,k}\right)}=\frac{\sin x}{2\,\sin \left(\frac{x}{2\,k}\right)}, \end{align} where $\Re\left\{z\right\}$ is the real part of $z$ and $i=\sqrt{-1}$. Now it suffices to show \begin{equation \frac{d^2}{dx^2} \left(x^3\frac{d}{dx} f_{p}^{s}(x)\right)=-x\frac{d^3}{dx^3} \left( \frac{\sin x}{\sin\left(\frac{x}{2\,k}\right)}\right). \end{equation} Using \eqref{eqn: f''' sin} in $(ii)$, this can be achieved by straightforward calculations. Thus $(iii)$ is true. The proof of $(iv)$ is similar, and we omit the details. We complete the proof of Lemma~\ref{lem: lots of f'''}.\qed \end{proof} We provide here an alternative proof of the two double inequalities in Theorem~\ref{thm: Daurizio-Sandor ineq original}. \begin{proof}[Proof of Theorem~\ref{thm: Daurizio-Sandor ineq original}] To this end, we show that for $x\in(0,\pi/2)$, $f_{2}^{c}(x)=\frac{1-\frac{\cos x}{\cos \frac{x}{2}}}{x^2}$ is strictly increasing while $f_{2}^{s}(x)=\frac{2-\frac{\sin x}{\sin \frac{x}{2}}}{x^2}$ is strictly decreasing. These lead to the desired inequalities since it is easy to see that \begin{alignat}{4 \lim_{x\to0}f_{2}^{c}(x) & =\frac{3}{8}, & \quad \lim_{x\to\pi/2}f_{2}^{c}(x) & =\frac{4}{\pi^2}, & \qua \\ \lim_{x\to0}f_{2}^{s}(x) & =\frac{1}{4}, & \quad \lim_{x\to\pi/2}f_{2}^{s}(x) & =\frac{4}{\pi^2}(2-\sqrt{2}). & \qua \end{alignat} To see $f_{2}^{c}(x)$ is strictly increasing, we employ \eqref{eqn: f''' cos} in Lemma~\ref{lem: lots of f'''} to obtain \begin{align}\label{eqn: f''' lemma1}\nonumber \frac{d^2}{dx^2}\left(x^3\frac{d}{dx} f_{2}^{c}(x)\right) &=-\frac{x}{64} \sec^4\left(\frac{x}{2}\right) \left(-44 \sin\left(\frac{x}{2}\right)+5 \sin \left(\frac{3 x}{2}\right)+\sin \left(\frac{5 x}{2}\right)\right) \\ &=-\frac{x}{16} \sec^4\left(\frac{x}{2}\right) \sin \left(\frac{x}{2}\right) (\cos x-2)(\cos x+5)>0. \end{align} As $\lim_{x\to0} \frac{d}{dx}\left(x^3\frac{d}{dx} f_{2}^{c}(x)\right)=0$, it follows that $\frac{d}{dx}\left(x^3\frac{d}{dx} f_{2}^{c}(x)\right)>0$. We are led to $x^3\frac{d}{dx} f_{2}^{c}(x)>0$ or $\frac{d}{dx} f_{2}^{c}(x)>0$ since $\lim_{x\to0} \left(x^3\frac{d}{dx} f_{2}^{c}(x)\right)=0$. This that shows $f_{2}^{c}(x)$ is strictly increasing. By using \eqref{eqn: 2k sin} in Lemma~\ref{lem: lots of f'''}, we have \begin{equation} \frac{d^2}{dx^2}\left(x^3\frac{d}{dx} f_{2}^{s}(x)\right)=-\frac{x}{4}\,\sin \left(\frac{x}{2}\right)<0, \end{equation} from which we infer that $\frac{d}{dx}\left(x^3\frac{d}{dx} f_{2}^{s}(x)\right)<0$ since $\lim_{x\to0} \frac{d}{dx}\left(x^3\frac{d}{dx} f_{2}^{s}(x)\right)=0$ by $(v)$ of Lemma~\ref{lem: lots of f'''}. Then \begin{equation} \displaystyle \frac{d}{dx}\left(x^3\frac{d}{dx} f_{2}^{s}(x)\right)<0 \end{equation} together with the fact $\lim_{x\to0} \left(x^3\frac{d}{dx} f_{2}^{s}(x)\right)=0$ from $(v)$ of Lemma~\ref{lem: lots of f'''} yields $x^3\frac{d}{dx} f_{2}^{s}(x)<0$ or $\frac{d}{dx} f_{2}^{s}(x)<0$. Thus we have shown that $f_{2}^{s}(x)$ is strictly decreasing. This completes the proof of the theorem.\qed \end{proof} We are now in the position to give the proof of Theorem~\ref{thm: Daurizio-Sandor ineq generalized}. \begin{proof}[Proof of Theorem~\ref{thm: Daurizio-Sandor ineq generalized}] The proof of the case when $p=2$ has been given in Theorem~\ref{thm: Daurizio-Sandor ineq generalized}. For $p\ge3$, we prove the desired inequalities by showing that $\frac{d}{dx} f_{p}^{\triangle}(x)<0$ for $\triangle=c, s$. Due to $(i)$ of Lemma~\ref{lem: lots of f'''}, we see that $\frac{d^2}{dx^2}\left(x^3\frac{d}{dx} f_{p}^{c}(x)\right)<0$ for $p\ge3$. Instead of employing $(ii)$ of Lemma~\ref{lem: lots of f'''}, we use \eqref{eqn: 2k sin} and \eqref{eqn: 2k+1 sin} in Lemma~\ref{lem: lots of f'''} to conclude that $\frac{d^2}{dx^2}\left(x^3\frac{d}{dx} f_{p}^{s}(x)\right)<0$. Thus we have for $\triangle=c, s$, \begin{equation} \frac{d^2}{dx^2}\left(x^3\frac{d}{dx} f_{p}^{\triangle}(x)\right)<0. \end{equation} Because of the first vanishing limit in $(v)$ of Lemma~\ref{lem: lots of f'''}, it follows that \begin{equation} \frac{d}{dx}\left(x^3\frac{d}{dx} f_{p}^{\triangle}(x)\right)<0, \end{equation} which, together with the fact that the second limit in $(v)$ of Lemma~\ref{lem: lots of f'''} vanishes, implies that $x^3\frac{d}{dx} f_{p}^{\triangle}(x)<0$ or $\frac{d}{dx} f_{p}^{\triangle}(x)<0$ for $\triangle=c, s$. It remains to find the following limits: \begin{alignat}{4 \lim_{x\to0}f_{p}^{c}(x) & =\frac{p^2-1}{2\,p^2}, & \quad \lim_{x\to\pi/2}f_{p}^{c}(x) & =\frac{4}{\pi^2}, & \qua \\ \lim_{x\to0}f_{p}^{s}(x) & =\frac{p^2-1}{6\,p}, & \quad \lim_{x\to\pi/2}f_{p}^{s}(x) & =\frac{4}{\pi^2}\left(p-\csc \left(\frac{\pi}{2p}\right)\right). & \qua \end{alignat} We immediately have \begin{equation \frac{4}{\pi^2}=\lim_{x\to\pi/2}f_{p}^{c}(x)<\frac{1-\displaystyle\frac{\cos x}{\cos \frac{x}{p}}}{x^2}<\lim_{x\to0}f_{p}^{c}(x)=\frac{p^2-1}{2\,p^2} \end{equation} and \begin{equation \frac{4}{\pi^2}\left(p-\csc \left(\frac{\pi}{2p}\right)\right)=\lim_{x\to\pi/2}f_{p}^{s}(x)<\frac{\displaystyle p-\frac{\sin x}{\sin \frac{x}{p}}}{x^2}<\lim_{x\to0}f_{p}^{s}(x)=\frac{p^2-1}{6\,p}. \end{equation} The proof is completed.\qed \end{proof} \section{Generalized D'Aurizio-S\'andor inequalities for hyperbolic functions}\label{sec: D'Aurizio-Sandor inequalities for hyperbolic functions} In this section, we show an analogue of Theorem~\ref{thm: Daurizio-Sandor ineq generalized} for the case of hyperbolic functions holds true. Let \begin{align} h_{p}^{c}(x)& = \frac{1-\displaystyle\frac{\cosh x}{\cosh \frac{x}{p}}}{x^2}, \\ h_{p}^{s}(x)& = \frac{p-\displaystyle\frac{\sinh x}{\sinh \frac{x}{p}}}{x^2}. \end{align} Following the same arguments for proving Lemma~\ref{lem: lots of f'''}, it can be shown that Lemma~\ref{lem: lots of f'''} with $\cos x$, $\sin x$ and $f_{p}^{\triangle}(x)$ $(\triangle=c, s)$ replaced by $\cosh x$, $\sinh x$ and $h_{p}^{\triangle}(x)$ $(\triangle=c, s)$ respectively, remains true. It follows that we can prove $\frac{d}{dx} h_{p}^{\triangle}(x)<0$ for $\triangle=c, s$ as in the proof of Theorem~\ref{thm: Daurizio-Sandor ineq generalized}. It remains to calculate the following limits: \begin{alignat}{4 \lim_{x\to0}f_{p}^{c}(x) & =\frac{1-p^2}{2\,p^2}, & \quad \lim_{x\to\pi/2}f_{p}^{c}(x) & =\frac{4}{\pi^2}\left(1-\cosh \left(\frac{\pi }{2}\right)\sech\left(\frac{\pi }{2 p}\right)\right), & \qua \\ \lim_{x\to0}f_{p}^{s}(x) & =\frac{1-p^2}{6\,p}, & \quad \lim_{x\to\pi/2}f_{p}^{s}(x) & =\frac{4}{\pi ^2}\left(p-\sinh \left(\frac{\pi }{2}\right)\csch\left(\frac{\pi }{2p}\right)\right). & \qua \end{alignat} Thus, we have the following analogue of Theorem~\ref{thm: Daurizio-Sandor ineq generalized} for $\cosh x$ and $\sinh x$. \begin{theorem}\label{thm: Daurizio-Sandor ineq generalized hyperbolic fcn} Let $0<x<\pi/2$. Then the two double inequalities \begin{equation}\label{eqn: cosh ineq p>2} \frac{4}{\pi^2}\left(1-\cosh \left(\frac{\pi }{2}\right)\sech\left(\frac{\pi }{2 p}\right)\right)<\frac{1-\displaystyle\frac{\cosh x}{\cosh \frac{x}{p}}}{x^2}<\frac{1-p^2}{2\,p^2} \end{equation} and \begin{equation}\label{eqn: sinh ineq p>2} \frac{4}{\pi ^2}\left(p-\sinh \left(\frac{\pi }{2}\right)\csch\left(\frac{\pi }{2p}\right)\right)<\frac{\displaystyle p-\frac{\sinh x}{\sinh \frac{x}{p}}}{x^2}<\frac{1-p^2}{6\,p} \end{equation} hold for $p=3, 4 , 5,\cdots$. In particular, the double inequality \eqref{eqn: cosh ineq p>2} is reversed when $p=2$ while the double inequality \eqref{eqn: sinh ineq p>2} remains true when $p=2$. \end{theorem} \section{Application of the generalized D'Aurizio-S\'andor inequalities to the Chebyshev polynomials of the second kinds}\label{sec: application} The first few Chebyshev polynomials of the second kind $U_n(x)$ $(n=0,1,2,\cdots)$ are (\cite{AS2,Chebyshev}) \begin{align} \label{eqn: Chebyshev poly n=0} U_0(x)&=1, \\ \label{eqn: Chebyshev poly n=1} U_1(x)&=2\,x, \\ \label{eqn: Chebyshev poly n=1} U_2(x)&=4\,x^2-1, \\ \label{eqn: Chebyshev poly n=1} U_3(x)&=8\,x^3-4\,x, \\ \label{eqn: Chebyshev poly n=1} U_4(x)&=16\,x^4-12\,x^2+1, \\ \label{eqn: Chebyshev poly n=1} U_5(x)&=32\,x^5-32\,x^3+6\,x, \\ \label{eqn: Chebyshev poly n=1} U_6(x)&=64\,x^6-80\,x^4+24\,x^2-1. \end{align} In this section, we apply Theorem~\ref{thm: Daurizio-Sandor ineq generalized} to $U_n(x)$ with $x=\cos \theta$. By means of the formula $U_{n}(\cos \theta)=\frac{\sin ((n+1)\,\theta)}{\sin \theta}$, we obtain the following corollary. \begin{corollary}\label{cor: Daurizio-Sandor ineq applied to Chebyshev poly} Let $y\in(0,\frac{\pi}{2\,p})$. The double inequality \begin{equation \frac{p}{6}\left((1-p^2 )y^2+6\right)<U_{p-1}(\cos y)<p-\frac{4}{\pi^2}\left(p-\csc \left(\frac{\pi}{2p}\right)\right)\,p\,y^2 \end{equation} holds for $p=2, 3, 4 , 5,\cdots$. \end{corollary} \begin{proof} The double inequality \eqref{eqn: sin ineq p>2} in Theorem~\ref{thm: Daurizio-Sandor ineq generalized} can be written as \begin{equation p-\frac{p^2-1}{6\,p}\,x^2<\frac{\sin x}{\sin \frac{x}{p}}<p-\frac{4}{\pi^2}\left(p-\csc \left(\frac{\pi}{2p}\right)\right)\,x^2, \quad x\in(0,\pi/2). \end{equation} Letting $x/p=y$, we have \begin{equation \frac{p}{6}\left((1-p^2 )y^2+6\right)<\frac{\sin (p\,y)}{\sin y}<p-\frac{4}{\pi^2}\left(p-\csc \left(\frac{\pi}{2p}\right)\right)\,p\,y^2, \quad y\in(0,\frac{\pi}{2\,p}). \end{equation} Since $\frac{\sin (p\,y)}{\sin y}=U_{p-1}(\cos y)$, the proof is completed.\qed \end{proof} \begin{example} Letting $p=7$ in Corollary~\ref{cor: Daurizio-Sandor ineq applied to Chebyshev poly} results in the following inequality \begin{equation 7-56\,y^2<64\,\cos ^6 y-80\,\cos ^4 y+24\,\cos ^2 y-1<7-\frac{196 \left(7-\csc \left(\frac{\pi}{14}\right)\right)}{\pi ^2}\,y^2, \end{equation} where $y\in(0,\frac{\pi}{14})\approx (0,0.2244)$ and $\frac{196 \left(7-\csc \left(\frac{\pi}{14}\right)\right)}{\pi ^2}\approx 49.7673$. \end{example} \textit{Acknowledgements}. The authors wish to express sincere gratitude to Tom Mollee for his careful reading of the manuscript and valuable suggestions to improve the readability of the paper. Thanks are also due to Chiun-Chuan Chen and Mach Nguyet Minh for the fruitful discussions. The authors are grateful to the anonymous referee for many helpful comments and valuable suggestions on this paper. \Refs \bibitem{AS2} \by M. Abramowitz and I. A. Stegun (Eds) \book Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables \publ National Bureau of Standards, Applied Mathematics Series {\bf 55}, 9th printing \publaddr Washington \year 1970 \endref \bibitem{D'Aurizio} \by J. D'Aurizio \paper Refinements of the {S}hafer-{F}ink inequality of arbitrary uniform precision \jour Math. Inequal. Appl. \vol 17 \issue 4 \yr 2014 \pages 1487--1498 \endref \bibitem{Chebyshev} \by T. J. Rivlin \book Chebyshev Polynomials \publ Wiley \publaddr New York \year 1990 \endref \bibitem{Sandor} \by J. S\'andor \paper On D'Aurizio's trigonometric inequality \jour J. Math. Inequal. \vol 10 \issue 3 \yr 2016 \pages 885--888 \endref \endRefs \end{document}

https://arxiv.org/abs/1610.00084

Spectral asymptotics for Kac-Murdock-Szegő matrices

Szegő's First Limit Theorem provides the limiting statistical distribution (LSD) of the eigenvalues of large Toeplitz matrices. Szegő's Second (or Strong) Limit Theorem for Toeplitz matrices gives a second order correction to the First Limit Theorem, and allows one to calculate asymptotics for the determinants of large Toeplitz matrices. In this paper we survey results extending the first and strong limit theorems to Kac-Murdock-Szegő (KMS) matrices. These are matrices whose entries along the diagonals are not necessarily constants, but modeled by functions. We clarify and extend some existing results, and explain some apparently contradictory results in the literature.

\section{Introduction} \subsection{The problem and its history} For any function $a$ of two variables with Fourier series \begin{equation} \label{FS1} a(x,t) =\sum_{k \in {\mathbb Z}} \hat{a}_k(x) e^{ikt}, \end{equation} Kac, Murdock and Szeg\H{o} \cite{kamusz53, grsz58} introduced in 1953 what they called \textit{generalized Toeplitz} matrices as the matrices of the form \begin{equation} T_n(a) = \left[ \hat{a}_{j-i} \left(\frac{i+j}{2n+2}\right) \right]_{i,j=0}^{n} \label{KMStype} \end{equation} We will call matrices of this form Kac-Murdock-Szeg\H{o} \ (KMS) matrices, although this is not the generally accepted term, as we will explain below. Such matrices have sometimes been called \textit{generalized Toeplitz}, \textit{Toeplitz-like}, \textit{twisted Toeplitz}, \textit{variable coefficient Toeplitz} matrices, and probably some other terms which we have yet to run across. In addition, Tilli \cite{ti98b}, motivated by applications to differential equations, introduced what he called \textit{locally Toeplitz} matrices, which are closely related to a special class of KMS matrices (see more below). As the above mentioned terms are somewhat ambiguous, we prefer the term KMS. Note that when $a(x,t)=a(t)$ is independent of $x$, $T_n(a)$ is Toeplitz. Toeplitz matrices can be characterized by the condition that the differences between elements on the $k$th diagonal are zero: \[T_n(a)_{j+1,j+k+1} - T_n(a)_{j,j+k} = 0 \] for all $j,k$ where the terms are defined. KMS matrices have the property that the differences between elements on the $k$th diagonal approach zero as $n\rightarrow\infty$: \begin{align*} T_n(a)_{j+1,j+k+1} - T_n(a)_{j,j+k} & = \hat{a}_{k}\left(\frac{2j+k}{2n+2} + \frac{1}{n+1}\right) - \hat{a}_{k}\left(\frac{2j+k}{2n+2}\right) \\ & = o(1) \end{align*} as long as the $\hat{a}_k$ are continuous. KMS matrices are thus natural generalizations of Toeplitz matrices, and they are `locally Toeplitz' in this sense. As long as the functions $\hat{a}_k$ are continuous, or even Riemann integrable, locally they are not too far from being Toeplitz. For this reason, much of the machinery used in the study of Toeplitz matrices can be applied to KMS matrices, as long as one is careful to keep track of the error terms. \begin{rem} The definition \eqref{KMStype} has the advantage that $T_n(a)$ will be Hermitian exactly when $a(x,t)$ is real valued. However, in many applications, the indexing is not always so neat. The indexing in \eqref{KMStype} of $(i+j)/(2n+2)$ along the $(j-i)$th diagonal is not necessary for the First Theorem \eqref{First KMS} on eigenvalue distributions to hold. Sampling the functions $\hat{a}_k$ along the diagonals at any partition whose mesh size approaches zero will lead to the same \emph{limiting statistical distribution (LSD)} of the eigenvalues (see \S\ref{partition}). Indeed, Kac, Murdock and Szeg\H{o} also considered matrices of the form \[ \left[ \hat{a}_{j-i} \left(\frac{\min\{i,j\}}{n+1}\right) \right]_{i,j=0}^{n} \quad \text{ and } \quad \left[ \hat{a}_{j-i} \left(\frac{\max\{i,j\}}{n+1}\right) \right]_{i,j=0}^{n} \] These matrices, as the one in \eqref{KMStype}, will also be Hermitian if and only if the symbol $a$ is real-valued. And, they have the same LSD as those defined by \eqref{KMStype}. However, while the asymptotic eigenvalue distribution is independent of how the indexing is done, the asymptotics of the determinants of $T_n(a)$ are extremely sensitive to the way in which the indexing is done (see \S\S\ref{DSOsection} and Remark~\ref{shiftremark}). \end{rem} The main result of the original paper of Kac, Murdock and Szeg\H{o} is concerned with proving a generalized First Szeg\H{o}'s Limit Theorem for matrices of the type \eqref{KMStype}. In that paper they also included a chapter on extreme eigenvalues of Toeplitz matrices of the form $$\left[\rho^{|j-i|}\right]_{i,j=0}^n, $$ where $0<\rho < 1$. It is this kind of Toeplitz matrix that has often been referred to as a KMS matrix in the literature. In particular, in a series of papers, Trench \cite{tr02} generalized this notion to a broader class of Toeplitz matrices and obtained asymptotic results for the spectra in some cases where the First Szeg\H{o}'s Limit Theorem does not apply. However, it is, in our opinion, a misnomer to refer to these as KMS matrices, when they are, after all, Toeplitz matrices. \smallskip KMS and related matrices have received a lot of attention lately with applications to such various fields as statistical mechanics \cite{ka13,demc98}, differential equations \cite{bosh08, ti98b, ti99b, se00, mcboag09, shtata12, comase16}, quantum integrable systems \cite{agbo08, bomc09, bo10} and the Heisenberg group \cite{budihiwi16, budihiwi16b}, among others. There has been a renewed interest in matrices of this type which followed a renewed interest in Toeplitz forms and their connections to orthogonal polynomials. The history of citations of \cite{kamusz53} illustrates the general trend. MathSciNet currently lists 77 citations for \cite{kamusz53}. Seven of these were from papers published from 1958 to 1965. There were no citations between the years 1965 and 1999. After these 34 years of dormancy, from 1999 on there have been 70 citations. Curiously, only a few of those papers citing \cite{kamusz53} dealt with matrices of the form \eqref{KMStype}. Google Scholar, which keeps a more comprehensive collection of citations including from math and non-math journals, lists 209 citations for \cite{kamusz53}. Google Scholar shows more citations in the years 1965 to 1999, but the same general trend of relatively few citations until near the end of the 20th century, followed by renewed interest at the turn of the century. Figure~\ref{cites} shows the number of citations per year from each source. \smallskip \begin{figure}[h] \begin{center} \includegraphics*[height=1.8in]{cites1.pdf} \hskip .1in \includegraphics*[height=1.8in]{cites2.pdf} \end{center} \caption{Citations of \cite{kamusz53} by year, according to MathSciNet (left) and Google Scholar (right).} \label{cites} \end{figure} In the early 1960's, Mejlbo and Schmidt \cite{mesc61, mesc62} proved a Strong Szeg\H{o}'s Theorem for KMS matrices, which we discuss at length below and in \S\ref{sec.strong}. After the mid-1960's, interest in such matrices waned. One notable exception is the paper of Trotter \cite{tr84}, who in 1984 generalized the results of Kac, Murdock and Szeg\H{o} to symbols in which the functions $\hat{a}_k$ are only Riemann integrable. Interest in spectral asymptotics of KMS matrices and their like renewed toward the end of the 20th century. In 1998, Shao \cite{sh98} generalized the First Limit Theorem to symbols of bounded variation. That same year, Deift and McLaughlin \cite{demc98} considered sequences of Jacobi matrices with variable coefficients in the context of the Toda lattice. Also in 1998, Tilli \cite{ti98b} introduced locally Toeplitz matrices. Tilli originally considered matrices that arise in the discretization of certain differential equations, which are, modulo finite rank perturbations, asymptotic to KMS matrices with symbols $a(x,t)=f(x)g(t)$. He derived a First Szeg\H{o}'s Theorem for such matrices, in effect rederiving the result of Kac, Murdock and Szeg\H{o} for a special case. This notion of locally Toeplitz has been generalized by several authors and applied quite fruitfully to the study of differential equations, notably by Serra-Capizzano. (See \cite{se03,se06}, and \cite{gase16} for a review of recent work on locally Toeplitz matrices and applications to differential equations.) \smallskip In 1999, Kuijlaars and Van Assche \cite{kuva99} published a paper on the asymptotic density of zeros of orthogonal polynomials with continuously varying coefficients. Such zeros are eigenvalues of symmetric tridiagonal KMS matrices whose diagonal entries are modeled by continuous functions. This has become a highly cited and influential paper. This was followed by a paper by Kuijlaars and Serra Capizzano \cite{kuse01} that extended the result to diagonals modeled by Riemann integrable functions. Kuijlaars et.al. obtained a formula for the asymptotic density of eigenvalues, whereas Kac, Murdock and Szeg\H{o} calculated the LSD via an integral (eqn \ref{First KMS} below). However, both results give essentially the same information, and it is a trivial exercise to extract the density from the LSD in the Jacobi case. The proof of Kuijlaars and Van Assche used powerful tools from potential theory, and were completely different from the moments method employed by Kac, Murdock and Szeg\H{o}. Still, their results were essentially a special case of results obtained nearly 50 years earlier using more elementary tools. \smallskip It is apparent that the results of Kac, Murdock and Szeg\H{o} were not well known in the mathematical community. This includes, until quite recently, the authors of the present paper, who undertook to write this article partly in order to rectify this situation. Indeed, the first author would have benefitted greatly from the knowledge of \cite{kamusz53} when writing the articles \cite{agbo08, agbo09, bo12} on generalizations of tridiagonal KMS matrices. The main purpose of the present paper is to collect and clarify results on the statistical distribution of eigenvalues of KMS matrices as $n\rightarrow\infty$ (the limiting statistical distribution, or LSD). We will be able to extend some of the known results. We will also explain some of the results in the literature that appear to be contradictory. The paper is mostly self-contained. For completeness, we will include most of the proofs. We have endeavored to present the proofs that are the clearest and most powerful. \subsection{The First and Second Szeg\H{o}'s Limit Theorems} Suppose the symbol $a$ in \eqref{FS1} is real-valued, so that $T_n(a)$ is Hermitian. Suppose, further, that the functions $\hat{a}_k$ are continuous on $[0,1]$ and satisfy the decay condition \begin{equation} \label{decay} \vvvert a \vvvert := \sum_{k \in {\mathbb Z}} \|\hat{a}_k\|_\infty < \infty. \end{equation} By Gershgorin's disks Theorem, the spectrum of $T_n(a)$ lies inside the closed interval $[-\vvvert a \vvvert, \vvvert a \vvvert ] $. Under these condition, Kac, Murdock and Szeg\H{o} \cite{kamusz53} proved the following First Limit Theorem for KMS matrices in 1953: \begin{equation} \label{First KMS} \text{Tr}[\varphi(T_n(a))] = \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} \varphi(a(x,t)) \, dt \, dx + o(n) \qquad (n \to \infty) \end{equation} for any continuous function $\varphi$ on $ [-\vvvert a \vvvert, \vvvert a \vvvert]$. When the symbol $a$ is independent of $x$, i.e. $a(x,t)=a(t)$, the result in \eqref{First KMS} reduces to the First Szeg\H{o}'s Limit Theorem for Toeplitz matrices. \smallskip The formula \eqref{First KMS} can be written in the equivalent form \[ \lim_{n\rightarrow\infty} \frac{\sum_{j=1}^{n+1}\varphi\left(\lambda_j(T_n(a))\right) }{n+1} = \frac{1}{2\pi} \int_0^1 \int_0^{2\pi} \varphi(a(x,t)) \, dt \, dx \] where $\lambda_j(T_n(a)) \ (j=1,\dots, n+1)$ are the eigenvalues of $T_n(a)$. The result \eqref{First KMS} thus gives us the LSD of the eigenvalues of $T_n(a)$. It says, roughly, that as $n \to \infty$, the eigenvalues of $T_n(a)$ distribute like the values of $a(x,t)$ sampled at regularly spaced points in the rectangle $0\leq t\leq 2\pi, \ 0\leq x\leq 1$. Alternatively, \eqref{First KMS} is equivalent to the following. Let $\alpha < \beta$ and let $N(n;\alpha,\beta)$ be the number of eigenvalues $\lambda$ of $T_n(a)$ satisfying $\alpha\leq \lambda \leq \beta$. Then \[ \lim_{n\rightarrow\infty} \frac{N(n;\alpha,\beta)}{n+1} = \frac{\sigma}{2\pi} \] where $\sigma$ is the area of the sub-domain in the rectangle $0\leq t\leq 2\pi, \ 0\leq x\leq 1$ such that $\alpha\leq a(x,t)\leq\beta$. \smallskip If one makes the additional assumptions $a(x,t) >0$ and $\log a(x,t) \in L^1([0,1] \times [0,2\pi])$, then \eqref{First KMS} for $\varphi(x) = \log x$ reads as \begin{equation*} \log \det T_n(a) = \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} \log(a(x,t)) \, dt \, dx + o(n) \end{equation*} When $a(x,t)$ is independent of $x$, the Szeg\H{o}'s Second--or Strong--Limit Theorem gives the more precise statement \cite{grsz58, bogr05}: \begin{equation} \label{Strong Szego1} \log \det (T_n(a)) = \frac{n+1}{2\pi} \int_0^{2\pi} \log(a(t)) \, dt + E(a) + o(1) \end{equation} where $E(a)$ is the constant \[E(a)= \sum_{k =1}^{\infty} k \widehat{(\log a)}_k\widehat{(\log a)}_{-k} \] and \[\widehat{(\log a)}_k = \frac{1}{2\pi}\int_0^{2\pi} \log(a(t)) e^{-ikt} dt \] denotes the $k$th Fourier coefficient of $\log a$. The Strong Szeg\H{o}'s Theorem \eqref{Strong Szego1} is more often expressed in the form \begin{equation} \lim_{n \to \infty} \frac{\det T_n(a)}{ G(a)^{n+1}} = \exp\left[ E(a) \right] \label{SS2} \end{equation} where \[G(a) = \exp\left[ \frac{1}{2\pi} \int_0^{2\pi} \log a(t) \, dt\right] \] is the geometric mean of $a$. In the early 60's, Mejlbo and Schmidt \cite{mesc61, mesc62} gave the following extension of Szeg\H{o}'s Strong Theorem to KMS matrices. Suppose $a$ is a complex-valued symbol such that $\hat{a}_k \in C^2([0,1])$ and satisfy the conditions \begin{equation*} \label{MSC} \sum_{k \in {\mathbb Z}} \| \hat{a}_k\|_\infty<1, \ \sum_{k \in {\mathbb Z}} \|\hat{a}'_k\|_\infty < \infty, \ \sum_{k \in {\mathbb Z}} \|\hat{a}^{''}_k\|_\infty < \infty, \text{ and } \sum_{k \in {\mathbb Z}} |k|^\alpha \|\hat{a}_k\|_\infty < \infty \end{equation*} for some $\alpha>2$. Suppose, further, that $a$ satisfies the diagonal dominance condition $\min |\hat{a}_0| > \sum_{ k \neq 0} \| \hat{a}_k\|_\infty$. Then, we have \begin{eqnarray} \label{Strong KMS} \lim_{n \to \infty} \frac{\det T_n(a)}{G(a)^{n+1}} = \exp \frac{1}{2}\{ e(a;0)-e(a;1) +E(a;0)+E(a;1) \} \end{eqnarray} where $G(a)$, $e(a;x)$ and $E(a;x)$ are now given by \begin{equation} G(a) = \exp\left[\frac{1}{2\pi} \int_0^1 \int_0^{2\pi} \log(a(x,t)) \, dt \, dx \right], \label{GeE1} \end{equation} and \begin{align} \label{GeE2} e(a;x) = \frac{1}{2\pi} \int_0^{2\pi} \log a(x,t) \, dt, \end{align} \begin{align} E(a;x) = \sum_{k =1}^{\infty} k \, \widehat{(\log a)}_k(x) \, \widehat{(\log a)}_{-k}(x). \label{GeE3} \end{align} Of course, when $a$ does not depend on $x$, $E(a;x) = E(a)$, and Mejlbo and Schmidt's result coincides with the Szeg\H{o} Strong Theorem. Quite remarkably, the error term obtained by Mejlbo and Schmidt in \eqref{Strong KMS} depends only on the values of the symbol $a$ at the endpoints $x=0$ and $x=1$. \subsection{Content of the present paper} In the next two sections, we prove a First Limit Theorem and a Strong Limit Theorem for KMS matrices under some minor improvements of the results mentioned above. For the sake of consistency and simplicity, both proofs use the moments method rather than the probabilistic method of Trotter \cite{tr84} in the First Limit Theorem or the operator method of Ehrhardt and Shao \cite{ehsh01} for the Strong Theorem. \S\ref{sec.first} deals with the First Limit Theorem. In addition to proving this result for KMS matrices, we include a section on the asymptotic distribution of singular values. When the matrices $T_n(a)$ are normal, we obtain the LSD of the eigenvalues. When $T_n(a)$ is not normal, we cannot obtain the LSD. However, we can say that most of the eigenvalues will accumulate in the extended range of the symbol $a$. This is the content of \S\ref{sec.clustering}. \smallskip \S\ref{sec.strong} deals with the Strong Limit Theorem. We prove this theorem along the same lines as Mejlbo and Schmidt. We then explain the results of Ehrhardt and Shao, who proved a similar theorem, but obtained a different formula from that of Mejlbo and Schmidt. We explain where the difference arises, and why their methods cannot be applied to KMS matrices. The last subsections of \S\ref{sec.strong} deal with the discrete Schr\"{o}dinger operator, i.e the special KMS matrix ``discrete Laplacian $+$ diagonal'' of the form \[ -\left[ \delta_{j,j+1} + \delta_{j+1,j} \right] + \mbox{diag}\left( f(1/n), f(2/n), \dots, f(n/n)\right) \] We report here on recent results for the asymptotics of determinants of such matrices. These results are extensions of a beautiful result of Kac from 1969 \cite{ka69}. One such extension is that if one shifts the indexing, one can adjust the limit of the determinant without changing the LSD of the eigenvalues. Another extension deals with asymptotics when $f$ has jump discontinuities. In that case, the determinant approaches a value (modulo $o(1)$ terms) that depends on $n$ in a peculiar way. These results demonstrate the extreme sensitivity of the determinant to small changes, and why the limit of $\det(T_n(a))/G(a)^{n+1}$ cannot exist when the symbol has discontinuities. We conclude with some open problems and conjectures. \section{First Limit Theorems} \label{sec.first} We present several First Limit Theorems for sequences of normal and non-normal KMS matrices. Our presentation follows the original approach of Kac, Murdock and Szeg\H{o} for some obvious reasons. First, their argument is in our opinion the most elegant and simplest one. Second, it allows straightforward generalizations to sequences of matrices with alternative indexing (see \S\ref{partition}) and to sequences of block matrices (see \S\ref{block}). Their arguments can also be easily modified to obtain the LSD of the singular values (see \S\ref{singular}). We end in \S\ref{sec.clustering} with a clustering property for the spectrum of $T_n(a)$. \smallskip \subsection{Normal matrices} \label{normal} Let $a(x,t)$ be a complex-valued function on $[0,1] \times [0,2\pi]$ and let $a_t$, $a_x$ be the functions defined by $a_t(x)=a(x,t)=a_x(t)$. Throughout \S\ref{normal}-\S\ref{sec.clustering}, we assume $a_t$ is Riemann integrable on $[0,1]$ and $a_x \in L^\infty ([0,2\pi])$. As a consequence of Gershgorin's disks Theorem and condition \eqref{decay}, the spectrum of every $T_n(a)$ lies inside the disk $D_a=\{z \in {\mathbb C} : |z| \leq \vvvert a \vvvert\}$. Our first result is concerned with the LSD of sequences of normal KMS matrices. This was first obtained by Trotter \cite{tr84} using some non-trivial probability argument. We give here a more elementary proof based on Kac-Murdock-Szeg\H{o} approach. \begin{theo} \label{normal} Let $\{ T_n(a) \}$ be a sequence of normal KMS matrices with symbol $a$ as in \eqref{FS1} and for which \eqref{decay} holds. Then, \begin{equation} \label{KMS1eq} \lim_{n \to \infty} \frac{\text{Tr}[\varphi(T_n(a))] }{n+1} = \frac{1}{2\pi} \int_0^1 \int_0^{2\pi} \varphi(a(x,t)) \, dt \, dx \end{equation} for any $\varphi \in C(D_a)$. \end{theo} \begin{proof} The normality of $T_n(a)$ ensures that $$ \mbox{Tr}[ T_n^p T_n^{*q} ] = \sum_{j=1}^{n+1} \lambda_j^p (T_n(a)) \, \overline{\lambda_j^q(T_n(a))} \qquad (p,q \in {\mathbb N}).$$ By the Stone-Weierstrass Theorem, the polynomials in $z$ and $\bar{z}$ are dense in the space of continuous functions on $D_a$. Consequently, it suffices to prove the result when $\varphi(z)=z^p \bar{z}^q$ with $p, q \in {\mathbb N}$. For simplicity, we write $a_{ij}$ for the entries of $T_n(a)$, and $T_n$ for $T_n(a)$. We have \begin{equation} \label{trace 1} \mbox{Tr}[ T_n^p T_n^{*q} ] = \sum_{i=0}^{n} \sum_{j_1,...,j_{p+q}=0}^n a_{i,j_1} a_{j_1,j_2} \cdots a_{j_{p-1},j_p} \bar{a}_{j_{p+1},j_p} \cdots \bar{a}_{i, j_{p+q}}. \end{equation} Let $h=(h_1,...,h_p)$ and $k=(k_1,...,k_q)$ be the multi-indices defined by $$h_1=j_1-i, \ h_2=j_2-j_1, \ldots, h_\alpha=j_p - j_{p-1}$$ and $$k_1=j_{p+1}-j_p, \ldots, k_{q-1}= j_{p+q}-j_{p+q-1}, \ k_q= j-i_{p+q}.$$ Adding the above relations, we see that $h$ and $k$ must satisfy the condition $|h|=|k|$. Hence, we can rewrite \eqref{trace 1} as \begin{equation} \label{trace 2} \text{Tr}[ T_n^r T_n^{*s} ] = \sum_{i=0}^{n} \sideset{}{'}\sum_{|h|=|k|} \prod_{l=1}^p \hat{a}_{h_l} \left( \frac{2i+\nu_l}{2n+2} \right) \, \prod_{m=1}^q \bar{\hat{a}}_{k_m} \left( \frac{2i+\nu_{p+m}}{2n+2} \right) \end{equation} where $$ \nu_j = \begin{cases} 2h_1+2h_2+\cdots + 2h_{j-1}+h_j & \text{ if } 1 \leq j \leq p \\ \nu_p -2k_1 -2k_2 - \cdots - 2k_{j-1} - k_j & \text{ if } p+1 \leq j \leq p+q. \end{cases}$$ The prime on the middle sum of \eqref{trace 2} is used to indicate that only multi-indices satisfying $ -2i \leq \nu_j \leq n-2i$ must enter into the summation. Let $\varepsilon>0$. By condition \eqref{decay}, there exists $r_0$ large enough so that $$ \sum_{|k| >r} \| \hat{a}_k \|_\infty < \varepsilon \qquad (\forall r>r_0).$$ Pick $r>r_0$ and let $M(h,k)=\max\{|h_l|,|k_m|\}$. It follows that \begin{align} & \sum_{i=0}^{n} \sideset{}{'}\sum_{ \substack{ |h|=|k| \\mathcal{M}(h,k) >r}} \prod_{l=1}^p \hat{a}_{h_l} \left( \frac{2i+\nu_l}{2n+2} \right) \, \prod_{m=1}^q \bar{\hat{a}}_{k_m} \left( \frac{2i+\nu_{p+m}}{2n+2} \right) \nonumber \\ & \qquad \qquad \leq \sum_{i=0}^n \left( \sum_{|k|>r} \|\hat{a}_k \|_\infty \right) \ \left( \sum_{k \in {\mathbb Z}} \|\hat{a}_k\|_\infty \right)^{p+q-1} \nonumber \\ & \qquad \qquad \leq (n+1) \, \vvvert a \vvvert^{p+q-1} \ \varepsilon, \end{align} i.e. the above sum is $o(n)$. Hence, it suffices to consider in \eqref{trace 2} multi-indices $h$ and $k$ for which $|h|=|k|$ and $M(h,k) \leq r$. For such multi-indices, we have \begin{align*} \sum_{|h|=|k|} & = \sideset{}{'}\sum_{|h|=|k|} + \sum_{ \substack{|h| = |k| \\ \min \nu_j <-2i \\ \max \nu_j <n-2i}} + \sum_{\substack{|h|=|k| \\ \max \nu_j > n-2i}} \\ & = \sideset{}{'}\sum_{|h|=|k|} +\, o(n) \end{align*} since $M(h,k) \leq r$. Hence, we can replace \eqref{trace 2} by \begin{equation*} \text{Tr}[ T_n^p T_n^{*q} ] = \sum_{i=0}^{n} \sum_{ \substack{|h|=|k| \\ M(h,k) \leq r}} \prod_{l=1}^p \hat{a}_{h_l} \left( \frac{2i+\nu_l}{2n+2} \right) \, \prod_{m=1}^q \bar{\hat{a}}_{k_m} \left( \frac{2i+\nu_{p+m}}{2n+2} \right) +o(n). \end{equation*} We now use the fact that the functions $\hat{a}_k$ are Riemann integrable on $[0,1]$ and $|\nu_j| ={\cal O}(1)$, so we can shift - up to an error of $o(n)$ - each of the expressions inside the functions $\hat{a}_k$ to obtain \begin{align*} \text{Tr}[ T_n^p T_n^{*q} ] & = \sum_{i=0}^{n} \sum_{ \substack{|h|=|k| \\ M(h,k) \leq r}} \prod_{l=1}^p \hat{a}_{h_l} \left( \frac{i}{n+1} \right) \, \prod_{m=1}^q \bar{\hat{a}}_{k_m} \left( \frac{i}{n+1} \right) +o(n) \\ & = (n+1) \int_0^1 \sum_{ \substack{|h|=|k| \\ M(h,k) \leq r}} \prod_{l=1}^p \hat{a}_{h_l} \left( x \right) \, \prod_{m=1}^q \bar{\hat{a}}_{k_m} (x) \, dx + o(n) \label{riemsum} \\ & = \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} \ a^p_r (x,t) \ \overline{a^q_r (x,t)} \, dt \, dx +o(n) \end{align*} where $a_r$ is the trigonometric polynomial given by $$ a_r(x,t) = \sum_{|k| \leq r} \hat{a}_k(x) \, e^{ikt}.$$ By condition \eqref{decay}, $a_r$ converges uniformly to $a$ on $[0,1] \times [0,2\pi]$ as $r \to \infty$. Therefore, we conclude \begin{align*} \text{Tr}[ T_n^p T_n^{*q} ] = \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} \ a^p (x,t) \ \overline{a^q (x,t)} \, dt \, dx +o(n) \end{align*} as desired. \end{proof} Obviously, the above argument breaks down when $T_n(a)$ is not normal, and hence we are unable to compute the LSD of arbitrary sequences of KMS matrices. However, we can easily prove the following weaker result. \begin{theo} \label{First Analytic} Let $\{T_n(a)\}$ be a sequence of KMS matrices that satisfies condition \eqref{decay}. Then, we have \begin{equation} \label{analyticeq1} \lim_{n \to \infty} \frac{ \text{Tr}[\varphi(T_n(a))] }{n+1}= \frac{1}{2\pi} \int_0^1 \int_0^{2\pi} \varphi(a(x,t)) \, dt \, dx \end{equation} for any analytic function $\varphi$ on the closed disk $D_{a}$. \end{theo} \begin{proof} From the proof of Theorem \ref{normal} with $q=0$, we deduce that $$ \text{Tr}[ T_n^p(a) ] = \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} \ a^p (x,t) \, dt \, dx +o(n)$$ for every $p \in {\mathbb N}$. Note that the normality is not needed in this case. The conclusion then follows from Mergelyan's Theorem that asserts that polynomials are dense in the space of analytic functions on $D_a$. \end{proof} \subsection{Alternative indexing} \label{partition} We start with a result that should be fairly obvious from the proof of Theorem ~\ref{normal}. There we used the indexing by $(i+j)/(2n+2)$ along the $(j-i)$th diagonal to form a Riemann sum. But, any partition whose mesh size approaches zero will accomplish the same, so the strict indexing in the definition \eqref{KMStype} is not necessary for Theorem \ref{normal} to hold. Indeed, any partition will do. The partition points can also be shifted by finite amounts. The following theorems, whose proofs are omitted, should thus be intuitive. We refer the reader to \cite[Cf. Theorem~3.6 and Corollary~3.7]{bomc15} for detailed proofs of some of the results. \smallskip \begin{defn} Let $\{ \mathcal{P}_n \}$ be a sequence of partitions of $[0,1]$ such that $$ \mathcal{P}_n=\left\{x_{0}^{(n)},x_{1}^{(n)},...,x_{n+1}^{(n)} \right\}$$ and whose meshes satisfy $\| \mathcal{P}_n \|= o(1)$ as $n \to \infty$. To every $\mathcal{P}_n$, we associate $(2n+1)$-tagged partitions $\{\mathcal{P}_{n;k}\}$ with $$\mathcal{P}_{n;k} = \left\{ [x_j^{(n)},x_{j+1}^{(n)}] ; \xi_{j;k}^{(n)} \right\} \qquad (-n \leq k \leq n) . $$ We defined generalized KMS matrices as the matrices $$ T_n(a) = \bigg[ \hat{a}_{j-i} \left( \xi_{i \wedge j;j-i}^{(n)} \right) \bigg]_{i,j=0}^n.$$ \end{defn} For instance, it is not hard to see that KMS matrices and their variant (see Remark 1.1) can all be expressed in that form. \smallskip The next results are straightforward extensions of Theorem \ref{normal} to sequences of generalized normal KMS matrices. They also extend previous results of Kuiljaars, Van Assche and Serra Capizzano \cite{kuva99,kuse01} for tridiagonal matrices. \begin{theo} \label{altindex} Let $\{T_n(a)\}$ be a sequence of generalized normal KMS matrices. Then, for any $\varphi \in C(D_{a})$, we have \begin{equation*} \lim_{n \to \infty} \frac{ \text{Tr}\left[\varphi(T_n(a))\right] }{n+1} = \frac{1}{2\pi} \int_0^1 \int_{0}^{2\pi} \varphi( a(x,t)) \, dt \, dx. \end{equation*} In the non-normal case, $\varphi$ must be chosen to be analytic in $D_a$. \end{theo} \smallskip It is possible to give a slightly more general result for small perturbations of $T_n(a)$. Indeed, let $\{A_n\}$ and $\{B_n\}$ be two sequences of normal matrices whose entries satisfy the estimate $$ |a_{ij}^{(n)} - b_{ij}^{(n)} | = o \left( n^{-1/2} \right).$$ The Wielandt-Hoffman inequality \cite{bh97} then implies $$ \frac{1}{n+1} \sum_{k=1}^{n+1} | \lambda_k(A_n) - \lambda_k(B_n) | \leq \frac{1}{\sqrt{n+1}} \| A_n-B_n\|_F = o\left(1 \right).$$ In addition, if we assume that $\{A_n\}$ and $\{B_n\}$ satisfy a decay condition as \eqref{decay}, then $\{A_n\}$ and $\{B_n\}$ have the same LSD. \begin{theo} \label{altindex2} Let $\left\{A_n \right\}$ with $A_n=\left[ a_{ij}^{(n)}\right]_{i,j=0}^n$ be a sequence of normal matrices that satisfies the decay condition \[ \alpha =\sup_n \left[ \sum_{k=0}^n \max_{0\leq j\leq n-k} \left|a_{j+k,j}^{(n)} \right| + \sum_{k=1}^n \max_{0 \leq j \leq n-k} \left|a_{j,j+k}^{(n)} \right| \right] < \infty \] Suppose there exists a sequence $\{T_n(a)\}$ of generalized normal KMS matrices for which \eqref{decay} holds and \[ \left|a_{ij}^{(n)} - \hat{a}_{j-i}\left(\xi_{i \wedge j;j-i}^{(n)} \right)\right| = o\left( n^{-1/2} \right). \] If $M=\max\{ \alpha, \vvvert a \vvvert\}$, then we have \begin{equation*} \lim_{n \to \infty} \frac{ \text{Tr}\left[\varphi(A_n)\right] }{n+1} = \frac{1}{2\pi} \int_0^1 \int_{0}^{2\pi} \varphi( a(x,t)) \, dt \, dx \end{equation*} for every $\varphi \in C(D_a)$. \end{theo} \subsection{Block matrices} \label{block} The proofs of Theorems \ref{normal} and \ref{First Analytic} -- and their extensions given above -- make no use of the product commutativity of the entries of $T_n(a)$. This naturally suggests to extend our class of symbols to matrix-valued ones. That is, $a(x,t)$ is now a matrix $A(x,t)$ of some fixed order $k$. In this context, $T_n(A)$ is then a block matrix of order $k(n+1)$ whose entries $A_{pq}$ are given by the $k \times k$ matrices $$ A_{pq} = \hat{A}_{q-p} \left( \xi^{(n)}_{p \wedge q;q-p} \right) = \frac{1}{2\pi} \int_0^{2\pi} A\left( \xi_{p \wedge q;q-p}^{(n)},t \right) \, e^{-i(q-p)t} \, dt.$$ Obviously, one can also define generalized block KMS matrices by allowing the entries to be evaluated on some tagged partitions of $[0,1]$. Note that our definition includes as special cases the class of locally Toeplitz matrices \cite{ti98b} and their generalizations \cite{se03}. As before, we assume $A_t$ is Riemann integrable on $[0,1]$ and $A_x \in L^\infty ([0,2\pi])$ for any given $x$ and $t$. In the block case, this is equivalent to say that each entry of $A(x,t)$ is Riemann integrable on $[0,1]$ in $x$ and bounded in $t$ on $[0,2\pi]$. Condition \eqref{decay} now reads as \begin{equation} \label{decay matrix} \vvvert A \vvvert = \sum_{k \in {\mathbb Z}} \ \sup_{x \in [0,1]} \| \hat{A}_k(x) \|_F <\infty. \end{equation} Of course, the choice of matrix norm is arbitrary -- for convenience, we picked the Frobenius norm -- any other matrix norm could also be chosen. We can now restate Theorems \ref{normal} and \ref{First Analytic} in the context of generalized block KMS matrices. \begin{theo} Let $\{T_n(A)\}$ be a sequence of normal block generalized KMS matrices that satisfies condition \eqref{decay matrix}. If $A(x,t)$ is normal for all $x$ and $t$, then \begin{equation*} \lim_{n \to \infty} \frac{\text{Tr}[\varphi(T_n(A))] }{n+1}= \frac{1}{2\pi} \int_0^1 \int_0^{2\pi} \text{Tr} [\varphi(A(x,t))] \, dt \, dx \end{equation*} for any $\varphi \in C(D_{kA})$. In the non-normal case, the function $\varphi$ must be analytic in the closed disk $D_{kA} = \{ z : |z| \leq k \vvvert A \vvvert\}$. \end{theo} Similarly, we can extend the perturbation result in Theorem \ref{altindex2} to the block matrix case. Once again, we choose the Frobenius norm for convenience. \begin{theo} \label{block2} Let $\left\{ A_n \right\}$ with $A_n=\left[ A_{ij}^{(n)}\right]_{i,j=0}^n$ be a sequence of normal block matrices. Suppose the entries $A_{ij}^n$ are normal and satisfies the decay condition \begin{equation} \label{decay2} \mathcal{A}:=\sup_n \left[ \sum_{k=0}^n \max_{0\leq j\leq n-k} \left\| A_{j+k,j}^{(n)} \right\|_F + \sum_{k=1}^n \max_{0\leq j\leq n-k} \left\| A_{j,j+k}^{(n)} \right\|_F \right] < \infty \end{equation} Suppose, further, there exists a sequence $\{T_n(A)\}$ of generalized normal block KMS matrices for which \eqref{decay matrix} holds and \[ \left\| A_{ij}^{(n)} - \hat{A}_{j-i}\left(t_{i \wedge j}^{(n)} \right)\right\|_F = o \left( n^{-1/2} \right). \] Let $\mathcal{M}=\max\{ \mathcal{A}, \vvvert A \vvvert \}$. If $A(x,t)$ is normal for all $x$ and $t$, then we have \begin{equation*} \lim_{n \to \infty} \frac{ \text{Tr}\left[\varphi(A_n)\right] }{n+1} = \frac{1}{2\pi} \int_0^1 \int_{0}^{2\pi} \text{Tr} \left[ \varphi( A(x,t)) \right] \, dt \, dx \end{equation*} for every $\varphi \in C(D_{k\mathcal{M}})$. In the non-normal case, $\varphi$ must be analytic in the closed disk $D_{k\mathcal{M}} =\{ z: |z| \leq k\mathcal{M}\}$. \end{theo} \subsection{Finite rank perturbations} The results above also hold for finite rank perturbations. In the Toeplitz case, this was first observed by Tyrtyshnikov \cite{ty96}. Let $\{A_n\}$ and $\{B_n\}$ be two sequences of normal block-matrices. Suppose the spectrum of $A_n-B_n$ satisfies the bound $$\|A_n -B_n \|_\infty = \sup_{1 \leq k \leq k(n+1)} |\lambda_k(A_n-B_n)| =o( \sqrt{n}) ,$$ and $$ \text{rank}(A_n-B_n) \leq r$$ for some $r$ independent of $n$. Then, the sequences $\{A_n\}$ and $\{B_n\}$ have the same LSD. Indeed, the Wielandt-Hoffman inequality implies \begin{align*} & \frac{1}{n+1} \sum_{j=1}^{k( n+1) } | \lambda_j(A_n) - \lambda_j(B_n)| \\ & \qquad \qquad \leq \frac{\sqrt{k} }{\sqrt{n+1}} \ \| A_n-B_n \|_F = o \left( 1 \right) \end{align*} since $\|A_n-B_n\|_F = o( r \sqrt{n}) = o(\sqrt{n})$. \begin{theo} \label{block3} Let $\{A_n\}$ be as in Theorem~\ref{block2} and let $\{C_n\}$ be a sequence of block matrices with $\|C_n\|_\infty= o (\sqrt{n})$ and $ \mbox{rank}\,(C_n) \leq r$ for some positive constant $r$. If $B_n = A_n + C_n$ is normal for all $n$, then \[ \lim_{n \to \infty} \frac{ \text{Tr}\left[\varphi(B_n)\right] }{n+1} = \frac{1}{2\pi} \int_0^1\int_{0}^{2\pi} \text{Tr} [ \varphi( A(x,t))] \, dt \, dx \] for any $\varphi \in C(D_{k\mathcal{M}})$. \end{theo} \subsection{Singular value distribution} \label{singular} The distribution of the singular values of sequences of KMS matrices can be obtained fairly easily once we have the preceding machinery. A similar result for sequences of Toeplitz matrices goes back to Avram \cite{av88} and Parter \cite{pa86} (see also \cite{bogr05}). Shao \cite{sh04} obtained a similar result using operator methods. \begin{theo} \label{SVDth} Let $T_{m,n}(a)$ be the $(m+1) \times (n+1)$ rectangular KMS matrix whose $(j,k)$ entry is given by \[ \hat{a}_{j-i} \left(\frac{i+j}{2\max\{m,n\}+2}\right) \] where the $\hat{a}_k$'s satisfy condition \eqref{decay}. Let $$\sigma_1(T_{m,n}(a)) \geq \sigma_2(T_{m,n}(a)) \geq \cdots \geq \sigma_{m \wedge n}(T_{n,m}(a)) \geq 0 $$ be the $m \wedge n = \min\{m+1,n+1\}$ singular values of $T_{m,n}(a)$. Then, we have $$ \lim_{m,n \to \infty} \frac{1}{ m \wedge n } \sum_{k=1}^{m \wedge n} \varphi(\sigma_k(T_{m,n}(a))) = \frac{1}{2\pi} \int_0^1 \int_{0}^{2\pi} \varphi(|a(x,t)|) \, dt \, dx$$ for every $\varphi \in C([0,\vvvert a \vvvert])$. \end{theo} \begin{proof} One can easily modify the trace formula in Theorem \ref{normal} to obtain $$ \text{Tr}[ (T_{m,n}(a) T_{m,n}^{*}(a))^p ] = \frac{m\wedge n}{2\pi} \int_0^1 \int_0^{2\pi} |a (x,t)|^p \, dt \, dx +o(n)$$ for any $p \in {\mathbb N}$. Since the singular values of $T_{m,n}(a)$ are the eigenvalues of the Hermitian matrix $T_{m,n}(a)T_{m,n}^{*}(a)$, it suffices to apply Weierstrass Approximation Theorem to obtain the desired result. \end{proof} Theorems \ref{block2} and \ref{block3} on sequences of block matrices can also be restated for the singular values. \begin{theo} (i) Let $\left\{ A_{m,n} \right\}$ be a sequence of block matrices whose entries $A_{ij}^{(n)}$ are $k \times k$ matrices that satisfy the condition \eqref{decay2}. Suppose there exists a sequence $\{T_{m,n}(A)\}$ of generalized block KMS matrices for which \eqref{decay matrix} holds and \[ \left\| A_{ij}^{(n)} - \hat{A}_{j-i}\left(\xi_{i\wedge j}^{(n)} \right)\right\|_F = o \left( n^{-1/2} \right). \] Then, we have $$ \lim_{m,n \to \infty} \frac{1}{ m \wedge n } \sum_{j=1}^{k(m \wedge n)} \varphi(\sigma_j(A_{m,n})) = \frac{1}{2\pi} \int_0^1 \int_{0}^{2\pi} \sum_{j=1}^k \varphi( \sigma_j(A(x,t))) \, dt \, dx$$ for every $\varphi \in C([0,k \mathcal{M}])$. \smallskip (ii) Let $A_{m,n}$ be as above and let $\{C_{m,n}\}$ be a sequence of block matrices of bounded rank satisfying \[ \sigma_1(C_{m,n}) = {\cal O}(1) \qquad \text{and} \qquad \mbox{rank}\,(C_{m,n}) \leq r \] for some constant $r$. Let $B_{m,n} = A_{m,n}+C_{m,n}$. Then $$ \lim_{m,n \to \infty} \frac{1}{ m \wedge n } \sum_{j=1}^{k(m \wedge n)} \varphi(\sigma_j(B_{m,n})) = \frac{1}{2\pi} \int_0^1 \int_{0}^{2\pi} \sum_{j=1}^k \varphi( \sigma_j(A(x,t))) \, dt \, dx$$ for every $\varphi \in C([0,k \mathcal{M}])$. \end{theo} \subsection{Eigenvalue clustering for non-normal matrices} \label{sec.clustering} When $T_n(a)$ is not normal, \eqref{analyticeq1} holds, but only for analytic $\varphi$. Since continuous functions cannot be approximated by analytic functions in the complex plane, this does not give us the LSD. Indeed, it is easy to construct matrices where \eqref{analyticeq1} fails for continuous $\varphi$. For example, consider the Toeplitz matrix with symbol $a(t) = e^{it}+e^{-2it}$, i.e. the matrix $[ \delta_{j+1,j} + \delta_{j,j+2}]$. The eigenvalues of $T_n(a)$ all lie on the star $\{ r\omega^j : \omega = e^{2\pi i/3}, r\geq 0, j=0,1,2\}$ in the complex plane \cite{mc09}. The spectrum remains bounded away from the range of the symbol (see figure~\ref{toeplitz_ex}). Thus, we can find a continuous $\varphi$ that is zero on the spectrum of $T_n(a)$ and positive on a portion of the range of $a$ with positive measure, which would contradict \eqref{analyticeq1} if it held for continuous $\varphi$. \begin{figure}[h] \begin{center} \includegraphics*[width=2in]{toeplitz_ex.pdf} \end{center} \caption{Eigenvalues of $T_{50}(a)$ (circles) for $a(t) = e^{it}+e^{-2it}$, and curve $a(t)$ for $t\in [0,2\pi]$.} \label{toeplitz_ex} \end{figure} While we cannot deduce the LSD for non-normal KMS matrices, the symbol does give us some information about how the eigenvalues accumulate in the long run. In this subsection, we will extend a result of Tilli \cite{ti99} about the clustering properties of the eigenvalues of Toeplitz matrices to KMS matrices. That is, we will find a set in the complex plane such any of its neighborhoods contains all of the eigenvalues of $T_{n}(a)$, except at most $o(n)$ of them. We begin by defining the \textit{essential range} of the symbol $a(x,t)$, denoted $\mathcal{R}(a)$, by : \begin{equation*} \mbox{$\mathcal{R}(a):=\left\lbrace z \in \mathbb{C}: \mbox{meas} \left\lbrace a^{-1}(D_r(z)) \right\rbrace>0,\forall r>0\right\rbrace$}, \end{equation*} \noindent That is, $\mathcal{R}(a)$ is the set of those points $z \in \mathbb{C}$ with the property that the Lebesgue measure of the inverse image of any open set containing $z$ is positive. $D_r(z)$ denotes an open disk in the complex plane with radius $r$ centered at $z$. As before, we consider symbols that satisfy the decay condition \eqref{decay} and such that $a_t$ is Riemann integrable on $[0,1]$ and $a_x \in L^\infty( [0,2\pi])$. In such case, $\mathcal{R}(a)$ is a compact set; hence its complement has just one unbounded connected component, and we can write \begin{equation} \mbox{$\mathbb{C} \backslash \mathcal{R}(a)=: U_{0}$ $\cup$ $\bigcup \limits_{j=1}^{\infty} U_{j}, \hspace{.25in} U_{i} \cap U_{j}=\emptyset$ if $ i\neq j $}, \label{UNION} \end{equation} where each $U_{j}$, $j\geq 1$, is a connected bounded open set, and $U_{0}$ is an unbounded connected open set. Using (\ref{UNION}) we define the \textit{extended range} of the symbol $a$ as $$\mathcal{ER}(a):=\mathbb{C} \backslash U_{0} $$ Hence, the \textit{extended range} $\mathcal{ER}(a)$ is the union of the range of $a$ and all the bounded components of its complement. We can now state the main result of this section, which was proven by Tilli \cite{ti99} for Toeplitz matrices. The proof that follows is only a minor modification of Tilli's. \begin{theo} \label{ClusterTheorem} Let $a(x,t)$ be as above. Then, the extended range $\mathcal{ER}(a)$ is a cluster of the eigenvalues of $T_{n}(a)$. That is, for any open set $V$ containing $\mathcal{ER}(a)$ there holds \begin{equation} \lim \limits_{n\rightarrow \infty} \dfrac{\mathcal{N}(V,n)}{n+1}=1 , \label{Cluster} \end{equation} where $\mathcal{N}(V,n)$ is the number of eigenvalues of $T_{n}(a)$ that lie inside $V$. In other words, any $\epsilon$-neighborhood of $\mathcal{ER}(a)$ contains all of the eigenvalues of $T_{n}$ except at most $o(n)$ of them. \eth \begin{proof} Choose some $z \notin \mathcal{ER}(a)$. Since $\mathcal{ER}(a)$ is closed, there exists some small open disk $D$ centered at $z$ such that $\overline{D}$ $\cap$ $\mathcal{ER}(a)=\emptyset$. Let $K= \mathcal{ER}(a) \cup \overline{D}$ and define $F$ on $K$ as $$F(\xi)=\left\{ \begin{array}{ll} 1 & $if $ \xi \in \overline{D}\\ 0 & $if $ \xi \in \mathcal{ER}(a). \end{array} \right. $$ \noindent Since $\mathbb{C}\backslash K$ is connected, by the Mergelyan theorem we can uniformly approximate $F$ with a polynomial $P$, so that for any given $\varepsilon>0$, $$|P(\xi)-F(\xi)| \leq \epsilon \qquad \qquad (\xi \in K).$$ Let $\mathcal{N}(D,n)$ be the number of eigenvalues of $T_{n}(a)$ inside $D$ and let $\chi_{D}$ be the characteristic function of $D$. Since $1-\epsilon \leq |P(\lambda)|$ whenever $\lambda \in D$, we have \begin{align*} (1-\epsilon) \ \mathcal{N}(D,n) &\leq \sum \limits_{j=1}^{n+1} \chi_{D} (\lambda_{j} (T_{n}(a))) \, |P(\lambda_{j}(T_{n}(a)))| \\ &\leq \left(\sum \limits_{j=1}^{n+1} \chi_{D} (\lambda_{j} (T_{n}(a))) \right)^{1/2} \ \left( \sum \limits_{j=1}^{n} |P(\lambda_{j}(T_{n}(a)))|^{2} \right)^{1/2} \\ & = \mathcal{N}(D,n)^{1/2} \left( \sum \limits_{j=1}^{n+1} |P(\lambda_{j}(T_{n}(a)))|^{2} \right)^{1/2}. \end{align*} Since $P(\lambda_{j}(T_{n}(a)))=\lambda_{j}(P(T_{n}(a)))$, we can square both sides to obtain $$(1-\epsilon)^{2} \ \mathcal{N}(D,n) \leq \sum\limits_{j=1}^{n+1} |P(\lambda_{j}(T_{n}(a)))|^{2} \leq \ ||P(T_{n}(a))||_{F}^{2}. $$ As noted in the previous section, one can modify the proof of Theorem \ref{normal} to obtain \begin{align} ||P(T_{n}(a))||_{F}^{2} & = \text{Tr}[ P(T_n(a)) \, (P(T_n(a)))^* ] \nonumber\\ & = \frac{n+1}{2\pi} \int_{0}^{1} \int_{0}^{2\pi} |P(a(x,t))|^{2} dt\, dx +o(n). \label{clust estimate} \end{align} The assumption of normality is not needed here since we only consider polynomials. Similarly, condition \eqref{decay} yields that $T_n(a)$ does not have to be banded for \eqref{clust estimate}. Consequently, we deduce $$\limsup_{n\rightarrow\infty} \frac{ (1-\epsilon)^{2} \, \mathcal{N}(D,n)}{n+1} \leq \frac{1}{2\pi} \int_{0}^{1} \int_{0}^{2\pi} |P(a(x,t))|^{2} \,dt \, dx \leq \epsilon^{2} $$ The last inequality holds since $|P(\xi)|\leq \epsilon$ whenever $\xi \in \mathcal{ER}(a)$. From the arbitrariness of $\epsilon$, the last inequality implies $\mathcal{N}(D,n)=o(n)$. Now consider an arbitrary open set $V\supset \mathcal{ER}(a)$. By Gershgorin's Theorem, the spectrum of $T_n(a)$ lies in the union of some disks, say $G$. Let $C:=G \cap (\mathbb{C}\backslash V)$. If $C$ is empty, then there is nothing to prove. If $C \neq \emptyset$, then every $z \in C$ lies within some open disk $D(z)$ centered at $z$ for which $$ \mathcal{N}(D(z),n)=o(n).$$ Thus, $C$ can be covered by a family of open disks $\left\lbrace D(z) \right\rbrace_{z\in C}$, each of which contains at most $o(n)$ eigenvalues. $C$ being a compact set, it can be cover by a finite sub-covering of those disks; hence $C$ itself must contain at most $o(n)$ eigenvalues of $T_{n}(a)$. Since $C \cup V$ contains all the spectrum of $T_{n}(a)$, equation \eqref{Cluster} follows and the proof is complete. \end{proof} To illustrate the above theorem, and also its limitations, consider the matrices $B_n$ whose entries on the $+1$ and $-2$ diagonals are given by \[ b_{j,j-1} = 1 - \frac{\mu_{j-2}}{\mu_n}, \quad b_{j,j+2} = \frac{j(j+1)}{\mu_n} \] where \[ \mu_n = n(n-1 + 6 \rho) \] The eigenvalue problem for $B_n$ arises when one seeks polynomial solutions to the generalized Lam\'{e} equation. See \cite{mcboag09} for an interpretation of the eigenvalues of $B_n$ in terms of charges in a logarithmic potential. It is easy to see that the $B_n$ are asymptotic to the KMS matrix with symbol \[ a(x,t) = \left(1-x^2\right) e^{it} + x^2 e^{-2it} \] Theorem~\ref{ClusterTheorem} thus implies that the eigenvalues will cluster in the extended range of the symbol, which in this case is just the unit disk. (Note that in this case the range and the extended range are the same.) This is true, but it doesn't give us much detailed information about the spectrum. In fact, the eigenvalues of $B_n$ all lie on the star $\{ r\omega^j : \omega = e^{2\pi i /3}, r\geq 0, j=0,2,2\}$ in the complex plane \cite{mc09}. One thing Theorem~\ref{ClusterTheorem} does give us is a bound on the magnitude of the eigenvalues (at least most of them). This is illustrated in figure~\ref{lame_ex}. \begin{figure}[h] \begin{center} \includegraphics*[width=2in]{lame_ex.pdf} \end{center} \caption{Eigenvalues of $B_n$ (circles) for $n=50$ and $\rho = 1$, and curves $a(s,t)$ for $x=0,.05,\dots, 1$, $t\in [0,2\pi]$. Compare with figure~\ref{toeplitz_ex}.} \label{lame_ex} \end{figure} As another example, where the extended range is more informative, consider the symbol \begin{equation} a(x,t) = \left(\frac{1}{2} + \sqrt{x - \frac{1}{2}}\right) e^{-2it} + e^{-it} + e^{it} \label{clustersymbol} \end{equation} Eigenvalues and curves $a(x,t)$ for fixed values of $x$ are shown in figure~\ref{cluster_ex}. We see that not all eigenvalues lie in the range of $a$, but they all lie in the extended range ${\cal ER}(a)$, which is roughly the interior of the envelope of the blue curves. \begin{figure}[h] \begin{center} \includegraphics*[height=1.5in]{cluster_ex3.pdf} \end{center} \caption{Eigenvalues of $T_{100}(a)$ (circles) for $a$ as in \eqref{clustersymbol}, and curves $a(x,t)$ for $x=0,.025,\dots, 1$, $t\in [0,2\pi]$.} \label{cluster_ex} \end{figure} \section{Strong Limit Theorems} \label{sec.strong} Now we turn to the Strong Limit Theorem, which can be viewed as a first order correction to the First Limit Theorem. In order to prove a Strong Theorem for KMS matrices based on the moments method, we need to obtain a more precise form of the error term in the proof of the First Limit Theorem. One way is to impose faster decay on the Fourier coefficients of the symbol $a$, or equivalently some smoothness on $a$. \subsection{Mejlbo and Schmidt's result and its generalizations} We start by proving the following lemma whose proof follows essentially the same lines as the one found in \cite{mesc62}. We make some minor adjustments allowing us to relax the regularity conditions on the symbol. \begin{lem} \label{main lemma} Let $a$ be a complex-valued symbol on $[0,1] \times [0,2\pi]$ such that $$a_x \in C^{1+\alpha}([0,2\pi]) \qquad \text{ and } \qquad a_t \in C^{1+\lceil 1/\alpha \rceil } ([0,1])$$ for some $0<\alpha \leq 1$. Moreover, suppose the following two conditions hold: \begin{equation} \label{decay 2} \sum_{|k| \in {\mathbb Z}} |k|^{ 1+1/\alpha} \| \hat{a}_k \|_\infty < \infty \qquad \text{ and } \qquad \sum_{|k| \in {\mathbb Z}} \| \partial_x{\hat{a}}_k \|_{\varepsilon} < \infty. \end{equation} Then, for any $p \in {\mathbb N}$, we have \begin{align*} \text{Tr}[T_n^p( a) ] & = \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} a^p(x,t) \, dt \, dx + \frac{1}{4\pi} \int_0^{2\pi} a^p(0,t) \, dt - \frac{1}{4\pi} \int_0^{2\pi} a^p(1,t) \, dt \\ & \quad + \sum_{|h|=0} N(h) \, \left[ \prod_{j=1}^p \hat{a}_{h_j}(0) + \prod_{j=1}^p \hat{a}_{h_j}(1) \right] + p \, \vvvert a \vvvert^{p-1} \, o(1) \end{align*} where $N(h) = \max\{0,h_1,h_1+h_2,....,h_1+\cdots + h_{p-1} \} $. \end{lem} \begin{proof} Pick $0<\delta < \alpha/(1+\alpha)$ and let $b(x,t)$ be the truncated symbol defined by $$ b(x,t) = \sum_{|k| \leq (n+1)^\delta} \hat{a}_k(x) \ e^{ikt}.$$ In order to simplify several of the expressions below, we introduce some useful notations. For every multi-index $h=(h_1,...,h_p)$ such that $|h|=0$, we denote by $$H=\{ h_1, h_1+h_2,...,h_1+ \cdots + h_{p-1} \}$$ and by $m(H)=\min H$ and $M(H)=\max H$. We also denote by $S_{i,\delta}$ the set $$ S_{i,\delta} = \{ h : \, \max_j |h_j| \leq (n+1)^\delta, \ -i \leq m(H) \leq M(H) \leq n-i\}.$$ Finally, we introduce the functions $$ f_{h,i} (t) = \prod_{k=1}^p \hat{a}_{h_k} \left( \frac{i}{n+1} + t \, \frac{2h_1+\cdots + 2h_{k-1} +h_k}{2n+2} \right).$$ With $\vvvert a \vvvert $ be as in \eqref{decay}, the trace formula \eqref{trace 2} together with the bounds \eqref{decay 2} yield \begin{align*} \left| \text{Tr}[T^p_n(a)] - \text{Tr}[T^p_n(b) ]\right| & = \left| \sum_{i=0}^n \sum_{h \in S_{i,1}} f_{h,i}(1) - \sum_{i=0}^n \sum_{h \in S_{i,\delta}} f_{h,i}(1) \right| \\ & \leq \sum_{i=0}^n \sum_{h \in S_{i,1} - S_{i\delta}} |f_{h,i}(1)| \\ & \leq p \vvvert a \vvvert^{p-1} (n+1) \sum_{|k|>(n+1)^{\delta}} \| \hat{a}_k\|_\infty\\ & \leq p \vvvert a \vvvert^{p-1} \sum_{|k|> (n+1)^{\delta}} |k|^{1+1/\alpha} \| \hat{a}_k \|_\infty. \end{align*} Thus, condition \eqref{decay 2} implies for $n$ large enough that \begin{equation*} \left| \text{Tr}[T^p_n(a)] - \text{Tr}[T^p_n(b) \right| = p \vvvert a \vvvert^{p-1} \, o(1). \end{equation*} If we denote by $T^{tr}_n(b)$ the transpose matrix -- not the conjugate-transpose -- of $T_n(b)$, then another application of \eqref{trace 2} gives us \begin{align*} \text{Tr}[(T^{tr}_n (b))^p ] & = \prod_{k=1}^p \hat{a}_{h_k} \left( \frac{i}{n+1} - \, \frac{2h_1+\cdots + 2h_{k-1} +h_k}{2n+2} \right) \\ & = \sum_{i=0}^n \sum_{h \in S_{i,\delta} } f_{h,i}(-1). \end{align*} Consequently, we obtain \begin{align*} \text{Tr}[T_{n}^p(b)] - \sum_{i=1}^n \sum_{h \in S_{i,\delta}} f_{h,i}(0) & = \frac{1}{2} \, \left( \text{Tr}[T_n^p(a_{(n+1)^\delta})] + \text{Tr}[((T_n^{tr})^p)(a_{(n+1)^\delta})] \right)\\ & \quad - \sum_{i=0}^n \sum_{|h|=0} f_{h,i}(0) \\ & = \frac{1}{2} \sum_{i=0}^n \sum_{ h \in S_{i,\delta}} \left( f_{h,i}(-1) + f_{h,i}(1) - 2 f_{h,i}(0) \right) \end{align*} By applying the Mean Value Theorem to $f_{h,i}(-1)-f_{h,i}(0)$ and $f_{h_i}(1)-f_{h,i}(0)$ and using the H\"older continuity of $\partial_x a$, we can bound the previous expression by \begin{align*} \left| \text{Tr}[T_{n}^p(b)] - \sum_{i=0}^n \sum_{h \in S_{i,\delta}} f_{h,i}(0) \right| & \leq \frac{(n+1) \, p \, \vvvert a \vvvert^{p-1}}{2} \, \left( \frac{ (n+1)^\delta}{n+1} \right)^{1+1/\alpha} \\ &= p \, \vvvert a \vvvert^{p-1} o(1) \end{align*} from our choice of $\delta$. By a similar argument as above, we also obtain \begin{equation*} \left| \sum_{i=0}^n \sum_{h \in S_{i,1}} f_{h,i}(0) - \sum_{i=0}^n \sum_{h \in S_{i,\delta}} f_{h,i}(0) \right| = p \vvvert a \vvvert^{p-1} \, o(1). \end{equation*} Therefore, it suffices to prove the result for the sum $ \sum_{i=0}^n \sum_{h \in S_{i,\delta}} f_{h,i}(0)$. We have \begin{equation} \label{MS1} \sum_{i=0}^n \sum_{h \in S_{i,1}} f_{h,i}(0) = \sum_{i=0}^n \sum_{|h|=0} f_{h,i}(0) - \sum_{i=0}^n \sum_{h \in R_{i,1}} f_{h,i}(0) - \sum_{i=0}^n \sum_{h \in R_{i,2}} f_{h,i}(0) \end{equation} where $R_{i,1}=\{ h: |h|=0, \ m(H)<-i, \ M(H) \leq n-i\} $ and $R_{i,2} =\{ h : \ |h|=0, \ M(H)>n-i\}$. Arguing as in the proof of the First Theorem and applying the Euler-Maclaurin formula, we see \begin{align} \label{MS2} \sum_{i=0}^n \sum_{h \in S_{i,1}} f_{h,i}(0) & = \sum_{i=0}^n \int_0^{2\pi} a^p(i/n,t) \, dt \nonumber\\ & = (n+1) \int_0^1 \int_0^{2\pi} a^p(x,t) \, dt \, dx + \frac{1}{4\pi} \int_0^{2\pi} a^p(0,t) \, dt \nonumber \\ & \quad - \frac{1}{4\pi} \int_0^{2\pi} a^{p}(1,t) \, dt + p \vvvert a \vvvert^{p-1} o(1). \end{align} On the other hand, \[ \sum_{i=0}^{n} \sum_{h \in R_{i,1}} |f_{h,i}(0)| \leq \sum_{|h|=0} \left[ \sum_{k=1}^p |h_k| \right] \prod_{k=1}^p \|\hat{a}_k\|_\infty \leq p \left[ \sum_{k \in {\mathbb Z}} |k| \|\hat{a}_k\|_\infty \right] \vvvert a \vvvert^{p-1}, \] so we can apply the Dominated Convergence Theorem to conclude \begin{align} \label{MS3} \sum_{i=0}^n \sum_{h \in R_{i,1}} f_{h,i}(0) & = \sum_{i=0}^n \sum_{h \in R_{i,1}} \prod_{k=1}^p \hat{a}_{h_k}(0) + p \, \vvvert a \vvvert^{p-1} o(1) \nonumber\\ & = \sum_{|h|=0} N(h) \prod_{k=1}^p \hat{a}_{h_k}(0) + p \, \vvvert a \vvvert^{p-1} o(1). \end{align} In the same way, we prove \begin{align} \label{MS4} \sum_{i=0}^n \sum_{h \in R_{i,2}} f_{h,i}(0) & = \sum_{i=0}^n \sum_{ |h|=0 \atop M(H)>i} \prod_{k=1}^p \hat{a}_{h_k}\left( \frac{n-i}{n+1} \right) \nonumber \\ & = \sum_{|h|=0} N(h) \prod_{k=1}^p \hat{a}_{h_k}(1) + p \, \vvvert a \vvvert^{p-1} o(1). \end{align} The result follows by combining \eqref{MS1}, \eqref{MS2}, \eqref{MS3} and \eqref{MS4}. \end{proof} The Strong Theorem for KMS matrices below is a modest improvement of the one's obtained by Mejlbo and Schmidt. We use the same notation as in the introduction. \begin{theo} Let $a$ be as in Lemma \ref{main lemma}. If $\text{Re} \, (a) > 0$, then \begin{equation} \label{Strong KMS eq} \lim_{n \to \infty} \frac{\det T_n(a)}{(G(a))^n} = \exp \,\frac{1}{2} \{ e(a;0)-e(a;1) +E(a;0)+E(a;1) \}. \end{equation} \end{theo} \begin{proof} Let $b=a-1$ and let $\vvvert a \vvvert $ and $\vvvert b \vvvert$ be the bounds \eqref{decay} associated to $a$ and $b$. The left hand side of \eqref{Strong KMS eq} is invariant under scaling of the symbol, therefore we may assume without loss of generality that $\vvvert a \vvvert<1$ and $\|a\|_\infty < 1$ with similar bounds for $b$. In particular, the eigenvalues of $T_n(b)$ satisfy $|\lambda_k(T_n(b))|<1$. Thus, \begin{align*} \log (\det T_n(a) ) & = \sum_{k=0}^n \log \lambda_k(T_n(a))\\ & = \sum_{k=0}^n \sum_{p=1}^\infty \frac{1}{p} \ (\lambda_k(T_n(a))-1)^p\\ & = \sum_{k=0}^n \sum_{p=1}^\infty \frac{1}{p} \ (\lambda_k(T_n(b)))^p. \end{align*} Since $\vvvert b \vvvert<1$, the double sum is absolutely convergent, and hence we can switch the order of summation. By previous lemma applied to the symbol $b$, it then follows \begin{align*} \log (\det T_n(a) ) & = \sum_{p=1}^\infty \frac{1}{p} \left\{ \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} b^p(x,t) \, dt \, dx \right.\\ & \quad + \frac{1}{4\pi} \int_0^{2\pi} b^p(0,t) \, dt - \frac{1}{4\pi} \int_0^{2\pi} b^p(1,t) \, dt \\ & \quad + \left. \sum_{|h|=0} N(h) \, \left[ \prod_{j=1}^p \hat{b}_{h_j}(0) + \prod_{j=1}^p \hat{b}_{h_j}(1) \right] \right\}+o(1). \end{align*} That is, \begin{align*} \log (\det T_n(a) ) & = \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} \log a(x,t) \, dt \, dx \\ & \quad +\frac{1}{4\pi} \int_0^{2\pi} \log a(0,t) \, dt - \frac{1}{4\pi} \int_0^{2\pi} \log a(1,t) \, dt \\ & \quad + \sum_{p=1}^\infty \frac{1}{p} \sum_{|h|=0} N(h) \, \left[ \prod_{j=1}^p \hat{b}_{h_j}(0) + \prod_{j=1}^p \hat{b}_{h_j}(1) \right] +o(1). \end{align*} Note that for $x=0$ or $x=1$, the matrices $T_n(b)$ are Toeplitz. By the Strong Szeg\H{o}'s Theorem for Toeplitz matrices, we deduce \begin{eqnarray*} \lefteqn{ \sum_{p=1}^\infty \frac{1}{p} \sum_{|h|=0} N(h) \, \left[ \prod_{j=1}^p \hat{b}_{h_j}(0) + \prod_{j=1}^p \hat{b}_{h_j}(1) \right] }\\ & & = \frac{1}{2} \sum_{k \in Z} \left[ k\, \widehat{(\log a)}_k(0) \, \widehat{(\log a)}_{-k}(0)+ \sum_{k \in {\mathbb Z}} k \, \widehat{(\log a)}_k(1) \, \widehat{(\log a)}_{-k}(1) \right] +o(1). \end{eqnarray*} The desired conclusion is then an immediate consequence of last two estimates. \end{proof} It is well-know that the Fourier coefficients decay faster when more regularity of the symbol is assumed. To this extent, recall the estimate \begin{equation*} \label{fourier coeff} \| \hat{a}_k\|_\infty \leq \frac{\| a\|_{r+\alpha}}{(1+|k|)^{r+\alpha}} \end{equation*} where $\| \cdot \|_{r+\alpha}$ denotes the H\"older norm with $r \in {\mathbb N}$ and $0 < \alpha \leq 1$. For instance, both conditions in \eqref{decay 2} hold if we assume that $a_x \in C^{2+\alpha}([0,2\pi])$ and $a_t \in C^{2+\lceil 1/\alpha \rceil}([0,1])$. \begin{cor} Let $a$ be a complex-valued symbol on $[0,1] \times [0,2\pi]$ with $\text{Re}\, (a) > 0$. Suppose $a_t \in C^{2+\alpha}([0,1])$ and $a_x \in C^{2+\lceil 1/\alpha \rceil } ([0,2\pi])$ for $0<\alpha \leq 1$. Then, $\{T_n(a)\}$ satisfies the conclusion of the Strong Limit Theorem above. \end{cor} Similarly, if $a$ is merely a trigonometric polynomial, or equivalently the matrices $T_n(a)$ have fixed band size, then $a_x$ is smooth on $[0,2\pi]$. Hence, we obtain the following consequence. \begin{cor} The Strong Theorem holds if $a$ is a trigonometric polynomial on $[0,1] \times [0,2\pi]$ with $\text{Re}(a)> 0$, and $a_t \in C^{1+\alpha}([0,1])$ for any $\alpha >0$. \label{STtrig} \end{cor} \subsection{Widom's result for KMS matrices} We include here an interesting extension of the Strong Limit Theorem due to Widom \cite{wi76} in the Toeplitz case. The Strong Limit Theorem \eqref{Strong KMS eq} is written using $\varphi = \log$. However, this can be generalized. If $\{T_n(a) \}$ is a sequence of Toeplitz matrices whose symbol is absolutely continuous on $[0,2\pi]$, then Widom obtained a Strong Theorem for arbitrary analytic functions other than the logarithm. Namely, \begin{eqnarray*} \lefteqn{ \lim_{n \to \infty} \left[ \text{Tr}[ \varphi(T_n(a)) ] - \frac{n+1}{2\pi} \int_0^{2\pi} \varphi(a(t)) \, dt \right] }\\ & & = \frac{1}{8\pi^2} \sum_{k=1}^\infty \int_0^{2\pi} \int_0^{2\pi} \sin(k(t-\tau)) \, \frac{\varphi(a(t)) - \varphi(a(\tau))}{a(t)-a(\tau)} \, ( a'(t) - a'(\tau)) \, dt \, d\tau. \end{eqnarray*} In the theorem below, we extend Widom's result to sequence of KMS matrices. \begin{theo} Let $a$ be as in Lemma \ref{main lemma}. Then, for any $\varphi \in C^\omega(D_{M})$ \begin{align*} \lim_{n \to \infty} & \left[ \text{Tr}[ \varphi(T_n(a)) ] - \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} \varphi(a(x,t)) \, dt \, dx \right] \\ & = \frac{1}{4\pi} \int_0^{2\pi} \varphi(a(0,t)) - \varphi(a(1,t)) \, dt \\ & \quad + \frac{1}{4\pi^2} \sum_{k=1}^\infty \int_0^{2\pi} \int_0^{2\pi} \, \left( \Phi(0,t,\tau)+ \Phi(1,t,\tau) \right) \, \sin (k(t-\tau)) \, dt \, d\tau \end{align*} where $$\Phi(x,t,\tau) = \frac{\varphi(a(x,t)) - \varphi(a(x,\tau))}{a(x,t)-a(x,\tau)} \, ( \partial_t a(x,t) - \partial_t a(x,\tau)).$$ \end{theo} \begin{proof} Write $\varphi(z) = \sum_{r=0}^\infty c_r z^r$ for $|z| < \vvvert a \vvvert$. From Lemma \ref{main lemma}, we deduce that \begin{align*} \text{Tr} [\varphi (T_n(a) )] & - \frac{n+1}{2\pi} \int_0^1 \int_0^{2\pi} \varphi(a(x,t)) \, dt \, dx \\ & = \frac{1}{4\pi} \int_0^{2\pi} \varphi (a(0,t)) \, dt - \frac{1}{4\pi} \int_0^{2\pi} \varphi( a(1,t)) \, dt \\ & \quad + \sum_{r=0}^\infty c_r \sum_{|h|=0} N(h) \, \left[ \prod_{j=1}^r \hat{a}_{h_jd}(0) + \prod_{j=1}^r \hat{a}_{h_j}(1) \right] +o(1). \end{align*} Using the fact that $T_n(a)$ is Toeplitz for fixed $s$, we obtain from Widom's reult \begin{align*} \ \sum_{r=0}^\infty c_r & \sum_{|h|=0} N(h) \, \left[ \prod_{j=1}^r \hat{a}_{h_jd}(0) + \prod_{j=1}^r \hat{a}_{h_j}(1) \right] \\ & = \frac{1}{8\pi^2} \sum_{k=1}^\infty \int_0^{2\pi} \int_0^{2\pi} \sin (k(t-\tau)) \, (\Phi(0,t,\tau)+\Phi(1,t,\tau)) \, dt \, d\tau + o(1) \end{align*} from which the desired result follows. \end{proof} \subsection{Ehrhardt and Shao's result} In a series of papers Shao \cite{sh98,sh03} and Ehrhardt \cite{ehsh01} generalized the elegant operator method of Widom \cite{wi74, wi76} to extend the results of Kac, Murdock and Szeg\H{o}, and Mejlbo and Schmidt to symbols with less regularity. First, Shao \cite{sh98} proved a first limit theorem for symbols of bounded variation. Later Ehrhardt and Shao \cite{ehsh01} proved a strong theorem for symbols with less regularity than Mejlbo and Schmidt required. Shao \cite{sh03} then generalized some of these results to block matrices. \smallskip Ehrhardt and Shao define their matrices differently than Kac, Murdock and Szeg\H{o}. In particular, given a function $a$ with Fourier series \eqref{FS1}, Ehrhardt and Shao define the matrices \newcommand{\mbox{op}_n}{\mbox{op}_n} \begin{equation} \mbox{op}_n a = \left[ \hat{a}_{j-i}\left(\frac{i}{n}\right)\right]_{i,j=0}^n \label{opdef} \end{equation} These are not KMS matrices due to the indexing by $i/n$, as opposed to $(i+j)/(2n+2)$. A peculiarity of this definition is that it is difficult to find a condition on the symbol $a$ so that $\mbox{op}_n a$ will be Hermitian. Hence, it is a somewhat unnatural definition. By Theorem~\ref{altindex}, the LSD of such matrices is the same as for KMS matrices (a result derived in \cite{sh98}). However, Ehrhardt and Shao found the following for the determinant. Suppose $a$ has winding number zero, $a$ is $C^{1+\alpha}$ in $x$ and $C^{\alpha}$ in $t$. Then \begin{equation} \lim_{n\rightarrow\infty} \frac{\det (\mbox{op}_n a)}{G(a)^{n+1}} = \exp\frac{1}{2}\left\{e(a;0)+e(a;1) + E(a;0) + E(a;1) + {\cal F}(a)\right\} \label{ES1} \end{equation} where $G, e$ and $E$ are as in (\ref{GeE1}-\ref{GeE3}), and \[ {\cal F}(a) = \int_0^1 \left(\sum_{k=-\infty}^{\infty} k \widehat{(\log a)}_k(x) \widehat{(\partial_x \log a)}_{-k}(x)\right) dx \] \begin{rem} The paper of Ehrhardt and Shao \cite{ehsh01} contains a typo in the formula for $\lim_{n\rightarrow\infty} \det (\mbox{op}_n a)/G(a)^{n+1}$. Their formula (eqn (1.5) in \cite{ehsh01}) contains $ G(a)^{-1}$. This would imply that the exponent of $G(a)$ in the denominator of \eqref{ES1} is $n$, instead of $n+1$, which is incorrect. \end{rem} The formula \eqref{ES1} of Ehrhardt and Shao appears to contradict the result \eqref{Strong KMS eq} of Mejlbo and Schmidt. This difference was noted (but not explained) by Ehrhardt and Shao. We will attempt to explain the difference here. First there is the sign difference between $e(a;0)$ and $e(a;1)$. However, this is only due to the fact that the KMS matrices index from $0$ to $n/(n+1)$ along the diagonals, whereas Ehrhardt and Shao index from $0$ to $1$. This sensitivity was pointed out by Kac \cite{ka69}. Changing the indexing in \eqref{opdef} from $i/n$ to $i/(n+1)$ would have the effect of changing the sign in \eqref{ES1} to $e(a;0)-e(a;1)$. This difference arises from the Euler-Maclaurin approximation of the integral. \smallskip The more serious difference, though, is the presence of the ${\cal F}(a)$ term in \eqref{ES1}, which does not appear in \eqref{Strong KMS eq}. Indeed, whereas Mejlbo and Schmidt's formula depends on the symbol $a$ only at $x=0$ and $x=1$, the formula of Ehrhardt and Shao depends on $a$ for \emph{all} $x\in[0,1]$. And, as Ehrhardt and Shao point out, $\exp\{{\cal F}(a)\}$ can take on any nonzero constant $c$ by choosing $a$ so that \[ \log a(x,t) = 2e^{it}\log c + e^{-it} x \] Thus, the two formulas are incompatible. \smallskip The formula \eqref{ES1} does not contradict \eqref{Strong KMS eq} because $\mbox{op}_n a$ and $T_n(a)$ are different matrices. While the LSD of the eigenvalues is the same for $\mbox{op}_n a$ and $T_n(a)$, the determinants behave differently in the limit. Determinants are much more sensitive to small changes than are eigenvalue distributions. The ${\cal F}(a)$ term in \eqref{ES1} is an artifact of the way Ehrhardt and Shao define their matrices. The formula \eqref{ES1} comes from a generalization of the operator results of Widom. Recall that the standard Toeplitz and Hankel operators acting on $l^2$ are defined for symbols of one variable by \[ T(a) = \left[\hat{a}_{j-i}\right]_{i,j=0}^{\infty} \quad \mbox{ and } \quad H(a) = \left[\hat{a}_{i+j+1}\right]_{i,j=0}^{\infty} \] The product of Toeplitz operators is Toeplitz+Hankel, i.e. \[ T(ab) - T(a)T(b) = H(a) H(\tilde{b}) \] where $\tilde{b}(z) = b(z^{-1})$. Widom \cite{wi74, wi76} derived a finite dimensional analogue: \begin{equation} T_n(ab) - T_n(a)T_n(b) = P_n H(a) H(\tilde{b}) P_n + Q_n H(\tilde{a}) H(b) Q_n \label{w1} \end{equation} where $P_n$ and $Q_n$ are the projection and flip operators: \begin{align*} P_n (f_0, f_1, \dots) &= (f_0, \dots, f_n, 0, \dots) \\ Q_n (f_0, f_1, \dots) &= (f_n, f_{n-1}, \dots, f_0, 0, \dots) \end{align*} Widom then employed this formula in a beautiful argument using operator theory to prove the Strong Szeg\H{o} Theorem \eqref{SS2}, showing that the error term can be written as \[ \exp\{E(a)\} = \det T(a)T(a^{-1}) \] Ehrhardt and Shao generalize this result to matrices of the form \eqref{opdef}. First, they generalize \eqref{w1} to \begin{equation} \mbox{op}_n a(x)b(y) - (\mbox{op}_n a)(\mbox{op}_n \overline{b})^* = H_n(a) H_n(\tilde{b})^{tr} + J_n(a)J_n(\tilde{b})^{tr} \label{ES3} \end{equation} where the matrix $\mbox{op}_n a(x)b(y)$ is the matrix whose $(p,q)-$entry is given by \[ \frac{1}{2\pi} \int_0^{2\pi} a\left(\frac{p}{n},t\right) b\left(\frac{q}{n},t \right) e^{-i(p-q)t} dt = \sum_{k=-\infty}^{\infty}\hat{a}_{p-k}\left(\frac{p}{n}\right) \hat{b}_{k-q}\left(\frac{q}{n}\right) \] and $J_n(a)$ and $H_n(a)$ are the half-infinite matrices given by \[ J_n(a) = \left[ \hat{a}_{-1-n+j-i}\left(\frac{i}{n}\right) \right], \ H_n(a) = \left[\hat{a}_{1+i+j}\left(\frac{i}{n}\right)\right], \ \left( \begin{array}{l} 0\leq i \leq n, \\[.05in] j=0,1,2,\dots\end{array}\right) \] Once the formula \eqref{ES3} is established they are able, with some difficulty, to carry through the operator argument to establish the generalized Strong Szeg\H{o} Theorem \eqref{ES1}. \smallskip It appears that this argument does not work for KMS matrices. When one indexes by $(i+j)/(2n+2)$ along the diagonals, the formula \eqref{ES3} breaks down, and there does not seem to be a consistent way to define something analogous to $\mbox{op}_n a(x)b(y)$ to make it work. Ehrhardt and Shao's way of defining their matrices works since one only indexes by the row. This makes the definitions used in \eqref{ES3} consistent, and allows them to carry through the computation. The term ${\cal F}(a)$ arises as a perhaps unfortunate side effect. \medskip \subsection {The discrete Schr\"{o}dinger operator } \label{DSOsection} In this section we report on results for the important special case of the discrete Schr\"{o}dinger operator \begin{equation} T_n(f) = \begin{bmatrix} f(\frac{1}{n}) & -1 & 0 & \cdots & 0 \\ -1 & f(\frac{2}{n}) & -1 & \cdots & 0 \\ 0 & -1 & f(\frac{3}{n}) & \cdots & 0 \\ & & & \ddots \\ 0 & 0 & 0 & \cdots & f(\frac{n}{n}) \end{bmatrix} \label{DSOdef} \end{equation} In this subsection we use a slightly different notation from the rest of the paper, in order to be consistent with that of M. Kac, whose paper \cite{ka69} contains the results on which the results of this subsection are based on. As the proofs, which are found in \cite{bomc16}, are somewhat lengthy and technical, they are omitted. They are obtained by modifying Kac's proof in \cite{ka69} for $T_n(f)$ when $f$ is twice differentiable. The first theorem \eqref{First KMS} obviously holds for $T_n(f)$, as long as $f$ is Riemann integrable, and tells us that the eigenvalues accumulate like the values of $a(x,t) = f(x)-2\cos t$ sampled at regularly spaced points in $[0,1]\times [0,2\pi]$. The second theorem \eqref{Strong KMS} holds as long as $f\in C^{1+\alpha}([0,1])$ for some $\alpha > 0$. Let \[ D_n(f) = \det T_n(f) \] Then, in the notation of this section, \[ \lim_{n\rightarrow\infty} \frac{ D_n(f)}{G(f)^n} = E(f) \] where \[G(f) = \exp \left\{ \int_0^1 \log\left(\frac{f(x) + \sqrt{f^2(x)-4}}{2}\right) dx \right\} \] is the geometric mean of $f(x)-2\cos t$, and $E(f)$ is given in terms of the series of Fourier coefficients of $f(x)-2\cos t$. (The terms in \eqref{Strong KMS} have to be modified slightly for the different convention used by Kac.) \smallskip In 1969 Kac \cite{ka69} derived a beautiful and simple formula for $E(f)$ for this case. The following theorem, dealing with families of matrices with any shift in the indexing, is obtained by modifying Kac's original proof. \begin{theo} \label{scaletheorem} Let $f$ be twice differentiable on some open interval $I$ containing $[0,1]$. Suppose $f$ has a bounded second derivative and satisfies $f>2$. Let $\varepsilon\in \mathbb{R}$ and define the matrices \[ T_{n}(f;\varepsilon ) = \begin{bmatrix} f(\frac{\varepsilon }{n}) & -1 & 0 & \cdots & 0 \\ -1 & f(\frac{1+\varepsilon }{n}) & -1 & \cdots & 0 \\ 0 & -1 & f(\frac{2+\varepsilon }{n}) & \cdots & 0 \\ & & & \ddots \\ 0 & 0 & 0 & \cdots & f(\frac{n-1+\varepsilon }{n}) \end{bmatrix} \] Then \begin{equation} \lim_{n \to \infty} \frac{\det {T}_{n}(f;\varepsilon)}{G(f)^n} =\frac{\displaystyle \left( f(0)+\sqrt{f^2(0 ) - 4}\right)^{1-\varepsilon} \left( f(1)+\sqrt{f^2(1 ) - 4}\right)^{\varepsilon} }{ 2 \displaystyle {\sqrt[4]{(f^2(0)-4)(f^2(1)-4)}}} \label{dsoe} \end{equation} \end{theo} \medskip \begin{rem} As long as $f(0)\neq f(1)$, the above limit can be adjusted to any positive number just by choosing the correct shift $\varepsilon$. By Theorem~\ref{altindex2}, the LSD of the eigenvalues of $T_n(f;\varepsilon)$ does not depend on $\varepsilon$. The formula \eqref{First KMS} holds for $T_n(f;\varepsilon)$ for any $\varepsilon$: \[ \lim_{n\rightarrow\infty} \frac{ \text{Tr}[\varphi(T_n(f;\varepsilon ))]}{n} = \frac{1}{2\pi} \int_0^1 \int_0^{2\pi} \varphi(f(x)-2\cos t) \, dt \, dx \] If one scales by $G(f)$, one thus obtains a family of matrices whose asymptotic eigenvalue distribution is invariant, but the determinant can be made to converge to any positive number. \label{shiftremark} \end{rem} The above results can also be modified for the case when $f$ has a finite number of jump discontinuities. \begin{theo} (i) Let $f$ be twice differentiable on $[0,1]$, with a bounded second derivative, except for $r<\infty$ jump discontinuities at $c_1,\dots, c_r \in (0,1)$, where both sided limits exist and are finite, and $f$ is left-continuous at $c_j$: $f(c_j)=f(c_j-)$. Suppose, also, that $f>2+\epsilon$ for some $\epsilon > 0$. Then \begin{equation} \frac{D_n(f)}{G(f)^n} = \alpha \prod_{j=1}^r \beta_j \gamma_j^{\{nc_j\}} + o(1) \label{jumpeqn} \end{equation} where $\{x\}=x-\lfloor x \rfloor$ is the fractional part of $x$, \begin{align*} \alpha &= \frac{1}{2} \frac{f(1)+\sqrt{f^2(1) - 4}}{\sqrt[4]{(f^2(0)-4)(f^2(1)-4)}} \\[.1in] \beta_j &= \frac{f(c_j-)-f(c_j+) + \sqrt{f^2(c_j+)-4} + \sqrt{f^2(c_j-)-4}}{2\sqrt[4]{(f^2(c_j+)-4)(f^2(c_j-)-4)}} \\[.1in] \mbox{and } \quad \gamma_j &= \frac{f(c_j+)+\sqrt{f^2(c_j+)-4}}{f(c_j-)+\sqrt{f^2(c_j-)-4}} \end{align*} \smallskip \noindent (ii) If $f$ is right-continuous at $c_j$, then the formula \eqref{jumpeqn} holds with $\{c_jn\}$ replaced by $\{c_jn\}'$, where \[ \{x\}' = 1+x-\lceil x\rceil \] is the fractional part of $x$, but equal to $1$ if $x$ is an integer. \label{jumptheorem} \end{theo} \begin{rem} Note that $\beta_j$ and $\gamma_j$ are $1$ if $f$ is continuous at $c_j$, so \eqref{jumpeqn} reduces to \eqref{dsoe} for $\varepsilon=1$ when $f$ is twice differentiable. Since $\{ c_jn\} =\{c_jn\}'$ if $c_j$ is irrational, the difference between cases (i) and (ii) of the above theorem only occurs when $c_j$ is rational. In that case the difference arises when $f$ is evaluated at the point $c_j$. \end{rem} \begin{rem} Obviously, if there is a discontinuity in $f$, the limit \[ \lim_{n\rightarrow\infty} \frac{D_n(f)}{G(f)^n} \] does not exist. However, we can calculate the $\limsup$ and $\liminf$. For example, if there is one jump discontinuity at $c=p/q$, \begin{align*} \limsup_{n\rightarrow\infty} \frac{D_n(f)}{G(f)^n} &= \alpha\cdot \beta\cdot \max\{ \gamma^{1/q}, \gamma\} \\ \liminf_{n\rightarrow\infty} \frac{D_n(f)}{G(f)^n} &= \alpha\cdot \beta \cdot \min\{ \gamma^{1/q}, \gamma\} \end{align*} If $c$ is irrational, the same is true with $\gamma^{1/q}$ replaced by $1$. Analogous statements hold when there are $r$ jump discontinuities. \end{rem} To illustrate the asymptotic behavior of $D_n(f)/G(f)^n$, we consider the case of a single jump discontinuity at $c\in (0,1)$. If $c =p/q$ is rational, $\{D_n(f)/G(f)^n\}$ (modulo an $o(1)$ term) is cyclic of order $q$. When $c$ is irrational, $\{D_n(f)/G(f)^n\}$ is dense on the interval between $\alpha\beta$ and $\alpha\beta\gamma$. This is another indication of how exquisitely sensitive $D_n(f)/G(f)^n$ is. The slightest irrational perturbation of the point of discontinuity from $c=1/2$, causes the values of $D_n(f)/G(f)^n$ (modulo the $o(1)$ term) to go from alternating between two values to taking on infinitely many values. This behavior is illustrated in figure~\ref{jumpfig}. There we calculate $D_n(f)/G(f)^n$ for the piecewise function \begin{equation} f(x) = \begin{cases} 3+x^2 +\sqrt{x} \sin(13 x) & x < c \\ 4.5-\cos(20x)/x & x \geq c \end{cases} \label{ff} \end{equation} We compare the values of $D_n(f)/G(f)^n$ with $\alpha\beta\gamma^{\{cn\}'}$ in the case when $c$ is rational and $c$ is irrational. Agreement is quite good for moderately large $n$. \begin{figure}[h] \begin{center} \includegraphics*[width=2.3in]{newdetA.pdf} \includegraphics*[width=2.3in]{newdetB.pdf} \end{center} \caption{$n$ vs. $D_n(f)/G(f)^n$ for $f$ as in \eqref{ff}. Left: $c=1/2$; right: $c=.9-1.5/\pi\approx .4225$. The values of $D_n(f)/G(f)^n$ are marked with circles; the values of $\alpha\beta\gamma^{\{cn\}'}$ are marked with $+$'s. } \label{jumpfig} \end{figure} \section{Some open problems, conjectures and other remarks} \label{conclusion} It is perhaps not surprising that Szeg\H{o}'s limit theorems can be generalized to KMS matrices. These matrices begin to look very much like Toeplitz matrices, locally, when $n$ is large. There are a whole host of problems revolving around generalizing other results for Toeplitz matrices to KMS matrices. While some results for Toeplitz matrices have natural generalizations to KMS matrices, there are some challenges for KMS matrices that are fundamentally different from their Toeplitz counterparts. \smallskip The most powerful results for KMS matrices are for when the matrices are normal. In this case, we can derive the LSD. In effect, we can completely characterize how the spectrum behaves in the limit of large $n$. When $T_n(a)$ is not normal, all we can do is delimit a region in the complex plane where most of the eigenvalues will lie. And, as figure~\ref{lame_ex} shows, this estimate may not give much detail for the spectrum. Normal matrices also have other desirable properties such as numerical stability that are lacking in non-normal matrices. For these reasons, it is of interest to be able to characterize when a matrix, or sequence of matrices, is normal. For KMS matrices, it is easy to tell if they are Hermitian: $T_n(a)$ is Hermitian if and only of the symbol $a(x,t)$ is real valued. A condition along these lines for normality would be quite valuable. \smallskip By a result of Brown and Halmos \cite{brha64} (see also \cite{nore09,nore11}), a Toeplitz matrix is normal if and only if its symbol $b$ can be obtained by rotation and translation of a real-valued symbol $a$, i.e. $b(t)=c+e^{i\theta} a(t)$. (It follows that the spectrum of a normal Toeplitz matrix lies on a line in the complex plane.) However, no such characterization for KMS matrices has been found. In fact, we were unable to construct sequences of normal KMS matrices outside the obvious Toeplitz and Hermitian ones. This leads to our first problem: \begin{prob} Characterize normal KMS matrices. Give a condition on $a(x,t)$ to guarantee that $T_n(a)$ will be normal. \end{prob} It has long been known that some non-normal Toeplitz matrices are so-called \textit{canonically distributed}, as Widom calls them. This means that the spectra accumulate on the range of the symbols. The question of which Toeplitz matrices are canonically distributed is tied to the types of singularities in the symbol \cite{wi90, wi94}. Thus, another avenue of research is to study asymptotics when the symbol $a(x,t)$ has a singularity in $t$. The famous Fisher-Hartwig conjecture was recently solved for Toeplitz matrices \cite{deitkr11}. The first step is the following: \begin{prob} Formulate a generalized Fisher-Hartwig conjecture for KMS matrices. \end{prob} \noindent The second step, obviously, is to prove the conjecture. \smallskip In the general non-normal, non canonically distributed case, the spectrum of KMS matrices still appears to have some structure. Schmidt and Spitzer \cite{scsp60} showed that if $T_n(a)$ is Toeplitz where $a(t)$ is a trigonometric polynomial, then the spectrum of $T_n(a)$ accumulates on a set $\Lambda(a)$ consisting of the union of a finite number of pairwise disjoint open analytic arcs and a finite number of exceptional points (branch points and points $\lambda$ such that for some open neighborhood $U$ of $\lambda$, $\Lambda(a)\cap U$ is an analytic arc starting and terminating on $\partial U$). Ullman \cite{ul67} proved that $\Lambda(a)$ is connected. Based on numerical experiments such as that seen in figure~\ref{cluster_ex}, we conjecture that the same holds for KMS matrices when the symbol is smooth in $x$. Hirschmann \cite{hi67} obtained an implicit formula for the asymptotic density of eigenvalues for banded Toeplitz matrices. It may be possible to do the same for KMS matrices: \begin{prob} Determine the LSD for non-normal KMS matrices. \end{prob} \smallskip Necessary conditions for the first and second theorems for KMS matrices are still lacking. For the first theorem, it was Trotter who first proved that it was enough for the $\hat{a}_k$'s to be Riemann integrable. In their paper on the Jacobi case, Kuijlaars and Serra Capizzano \cite{kuse01} reproved this result for tridiagonal KMS matrices using potential theory. Their result was for coefficients satisfying a closeness condition for functions only in $L^1([0,1])$. Clearly, it is not sufficient to take the $\hat{a}_k(x)$ to be only in $L^1([0,1])$. However, it may be enough for the $\hat{a}_k$ to be in $L^1([0,1])$, with the condition that the set of discontinuities is nowhere dense. \begin{prob} Determine necessary conditions on the $\hat{a}_k$ for the first theorem to hold. \end{prob} Consider the banded case where the symbol $a$ is a trigonometric polynomial \[ a(x,t) = \sum_{k=p}^q \hat{a}_k(x) e^{ikt} \] Then, by Corollary~\ref{STtrig}, as long as the $\hat{a}_k$ are $C^{1+\epsilon}$, the limit \[ \lim_{n\rightarrow\infty}\frac{\det T_n(a)}{G(a)^{n+1}} \] exists and can be calculated. However, it is not known if this condition is necessary. In figure~\ref{DSOfig} below we show the fraction $D_n(f)/G(f)^{n}$ for the discrete Schr\"{o}dinger operator \eqref{DSOdef} where $f$ has decreasing levels of regularity. Numerical evidence suggests that it is not sufficient for $f(x)$ to be merely $C^1$, although we have no proof of this result. \begin{figure}[h] \begin{center} \includegraphics*[width=1.6in]{FSx2sinB.pdf} \includegraphics*[width=1.6in]{FSx32sinB.pdf} \includegraphics*[width=1.6in]{FSxsinB.pdf} \end{center} \caption{$\displaystyle \frac{D_n(f)}{G(f)^{n}}$ for the discrete Schr\"{o}dinger operator \eqref{DSOdef}. Left: $f(x)=3.5+x^2\sin(1/x)$; middle: $f(x)=3.5+x^{3/2}\sin(1/x)$; right: $f(x)=3.5+x\sin(1/x)$.} \label{DSOfig} \end{figure} \begin{prob} Determine necessary and sufficient conditions on the $\hat{a}_k$ for the second theorem to hold. \end{prob} One of the shortcomings of the Strong Szeg\H{o} Theorem is that the error term is computed in terms of a series of the Fourier coefficients of the logarithm. It can thus be somewhat opaque to deduce properties of the error based on the symbol itself. The only case where we have a simple formula for the error is for the discrete Schr\"{o}dinger operator: the formula \eqref{dsoe} in the continuous case, and \eqref{jumpeqn} when there are jump discontinuities. It may be possible to use similar methods to obtain results for other operators. \begin{prob} Extend the formulas \eqref{dsoe} and \eqref{jumpeqn} to tridiagonal KMS matrices. \end{prob} For Toeplitz band matrices, the $o(1)$ error term in \eqref{Strong Szego1} is in fact ${\cal O}(q^n)$ for some $0<q<1$ (see e.g. \cite{bogr05}). Consequently, Szeg\H{o}'s Strong Theorem gives the complete asymptotic expansion of $ \det(T_n(a))$ when $a$ is a trigonometric polynomial. All of the existing proofs of the strong theorem for KMS matrices make use of the Euler-Maclaurin formula, and hence the $o(1)$ error term is the best known result. \begin{prob} Improve the error term in the strong theorem for KMS band matrices. \end{prob} In applications (e.g. in determining stability) one is often interested in the bounds for the eigenvalues. There is still the open problem: \begin{prob} Determine the extreme eigenvalues of a sequence of KMS matrices. \end{prob} The present paper is concerned with the spectra of KMS matrices. However, in applications it is often more important to know the pseudospectra. Also of great importance are the eigenvectors of KMS matrices, of which we have said nothing. While a great deal is known about the pseudospectra of Toeplitz matrices (see \cite{bogr05} and many references therein), not much is known about the pseudospectra of KMS matrices. The paper of Trefethan and Chapman \cite{trch04} contains some interesting results on the pseudospectra and eigenvectors of KMS (called ``twisted Toeplitz'' in that paper) matrices. However, there is much more to be discovered. We can thus pose the somewhat general problem: \begin{prob} Derive bounds on the pseudospectra of KMS matrices. \end{prob} There is by now an enormous, and growing, literature on Toeplitz matrices. Only a few of the results for Toeplitz matrices, importantly the first and second Szeg\H{o}'s theorems, have been extended to KMS matrices. We look forward with great anticipation to the development of theory for these natural generalizations.

https://arxiv.org/abs/1807.02194

An algorithm for enumerating difference sets

The DifSets package for GAP implements an algorithm for enumerating all difference sets in a group up to equivalence and provides access to a library of results. The algorithm functions by finding difference sums, which are potential images of difference sets in quotient groups of the original group, and searching their preimages. In this way, the search space can be dramatically decreased, and searches of groups of relatively large order (such as order 64 or order 96) can be completed.

\section{Introduction}\label{sec:introduction} Let $G$ be a finite group of order $v$ and $D$ a subset of $G$ with $k$ elements. Then $D$ is a \emph{$(v, k, \lambda)$-difference set} if each nonidentity element of $G$ can be written as $d_id_j^{-1}$ for $d_i, d_j \in D$ in exactly $\lambda$ different ways. Difference sets were first studied in relation to finite geometries \cite{Singer1938} and have connections to symmetric designs, coding theory, and many other fields of mathematics \cite{MoorePollatsek2013, DavisJedwab1996, CRC, BeJuLe1999}. Large libraries of difference sets are useful for developing conjectures and building examples. Gordon \cite{LaJolla} provides an extensive library of difference sets in abelian groups, but has no results for non-abelian groups, which do show distinct behavior \cite{Smith1995}. A wide variety of techniques can be used to construct difference sets for these libraries (see, for example, \cite{Dillon1985} and \cite{DavisJedwab1997}), but fully enumerating all difference sets in a given group requires some amount of exhaustive search, which can quickly become computationally infeasible. Kibler \cite{Kibler1978} performed the first major exhaustive enumeration of difference sets, and considered groups where difference sets could be found with $k < 20$. In recent years, AbuGhneim \cite{AbuGhneim2013, AbuGhneim2016} performed almost complete enumerations for all groups of order 64, and several authors have found difference sets in groups of order 96 \cite{GoMaVu2005, GoMaVu2007, AbuGhneimSmith2007}. The {\tt DifSets} package for the {\tt GAP} \cite{GAP4} computer algebra system efficiently and generally implements the techniques used by these and many other authors to exhaustively enumerate all difference sets up to equivalence in a group. With the package loaded, a search of a given group can be performed with a single command. \begin{verbatim} gap> DifferenceSets(CyclicGroup(7)); [ [ 1, 2, 4 ] ] \end{verbatim} The package has been used to give the first complete enumeration of all difference sets up to equivalence in groups of order 64 and 96, and in total provides a library of results for 1006 of the 1032 groups of order less than 100. Results are organized by their ID in the {\tt SmallGroups} \cite{SmallGrp} library and can be easily loaded by {\tt GAP}. \begin{verbatim} gap> LoadDifferenceSets(16, 5); # results for SmallGroup(16, 5) [ [ 1, 2, 3, 4, 8, 15 ], [ 1, 2, 3, 4, 11, 13 ] ] \end{verbatim} The ease of use of these top-level functions is the primary interface difference between the {\tt DifSets} package and the {\tt RDS} \cite{RDS} package, a similar {\tt GAP} package that provides a variety of tools to search for difference sets. The functions involving coset signatures in {\tt RDS} provide similar functionality to the {\tt DifSets} package, but require substantial user interaction to perform efficient searches, and are not feasible for searching most groups of order 64 and 96. In addition, {\tt RDS} provides no precomputed results, though it does provide significant additional functionality related to relative difference sets, partial difference sets, and projective planes. \section{Difference Sets}\label{sec:difsets} For notational purposes it is useful to consider a subset $D \subseteq G$ as an element of the group ring $\mathbb{Z}[G]$. We will abuse notation to define the group ring elements \[ G = \sum_{g \in G} g, \quad D = \sum_{d \in D} d, \quad D^{(-1)} = \sum_{d \in D} d^{-1}, \quad gD = \sum_{d \in D} gd, \quad D^\phi = \sum_{d \in D} \phi(d) \] where $g \in G$ and $\phi$ is a homomorphism from $G$. Then the statement that $D$ is a $(v, k, \lambda)$-difference set is equivalent to the equation \[ DD^{(-1)} = (k-\lambda)1_{G} + \lambda G \] where $D$ is an element of $\mathbb{Z}[G]$ with coefficients in $\{0, 1\}$. With this definition it is a quick exercise to prove the following (see page 298 of \cite{BeJuLe1999} and Theorem 4.2 and 4.11 of \cite{MoorePollatsek2013}). \begin{prop}\label{prop:equivsets} Let $G$ be a group of order $v$. Then \begin{enumerate} \item Any one element subset of $G$ is a $(v, 1, 0)$-difference set. \item The complement of a $(v, k, \lambda)$-difference set in $G$ is a $(v, v-k, \lambda+v-2k)$-difference set in $G$. \item If $D$ is a $(v, k, \lambda)$-difference set in $G$, $g \in G$, and $\phi \in \mathrm{Aut}(G)$, then $gD^\phi$ is also a $(v, k, \lambda)$-difference set in $G$. \end{enumerate} \end{prop} In addition, an immediate consequence of the definition is that $k(k-1) = \lambda (v-1)$ for any valid set of parameters of a difference set, so that for a given value of $v$ there are typically only a few possible values of $k$ and $\lambda$. More sophisticated results, such as the Bruck-Ryser-Chowla theorem, can reduce the number of possibilities even further. As a result of Proposition~\ref{prop:equivsets}, in enumerating difference sets we ignore the trivial one element difference sets, only take the smaller of each complementary pair of sets, and only consider sets distinct up to an equivalence given by part (3). \begin{defi}\label{defi:equivsets} Let $D_1$ and $D_2$ be difference sets in $G$. Then $D_1$ and $D_2$ are \emph{equivalent difference sets} if $D_1 = gD_2^\phi$ for some $g \in G$ and $\phi \in \mathrm{Aut}(G)$. \end{defi} In the {\tt DifSets} package, difference sets are stored as lists of integers. These integers represent indices in the list returned by the {\tt GAP} function {\tt Elements(G)}, which is a sorted\footnote{Element comparison (and thus the list {\tt Elements(G)}) is instance-independent in {\tt GAP} for permutation and pc groups, which includes, for example, all groups in the {\tt SmallGroups} library.} list of elements of the group {\tt G}. For example, consider the group $C_7 = \langle x | x^7 = 1 \rangle$. In {\tt GAP} we have \begin{verbatim} gap> C7 := CyclicGroup(7);; gap> Elements(C7); [ <identity> of ..., f1, f1^2, f1^3, f1^4, f1^5, f1^6 ] \end{verbatim} where clearly {\tt f1} is the generator corresponding to our $x$. Then the subset $D = \{x, x^2, x^4\}$ corresponds to the set consisting of the second, third, and fifth elements of {\tt Elements(C7)}, which we can represent in indices as {\tt [2, 3, 5]}. We can check that this is a difference set and also note that it is equivalent to the difference set $xD = \{x^2, x^3, x^5\}$, which is represented as {\tt [3, 4, 6]}. \begin{verbatim} gap> IsDifferenceSet(C7, [2, 3, 5]); true gap> IsEquivalentDifferenceSet(C7, [2, 3, 5], [3, 4, 6]); true \end{verbatim} \section{Difference Sums} A basic method for enumerating all difference sets in a group $G$ is to enumerate all subsets of $G$ and check if each is a difference set by definition. But since the number of subsets in a group is exponential in its order, we cannot feasibly enumerate and test all subsets for groups of even a modest size. The key to decreasing the search space is the following well-known lemma, which motivates our definition of a \emph{difference sum}\footnote{Concepts similar to difference sums are elsewhere referred to as difference lists, intersection numbers, or signatures. However, difference sums require both a group $G$ and normal subgroup $N$, not just the group structure of the quotient $G/N$ used in some other definitions. This precision is needed for specifying induced automorphisms in Definition~\ref{defi:equivsums} so that we can prove Lemma~\ref{lem:equivsums}.}. \begin{lem}\label{lem:image} Suppose $D$ is a $(v, k, \lambda)$-difference set in $G$ and $\theta$ is a homomorphism of $G$ with $|\mathrm{ker}(\theta)| = w$. Let $S = D^\theta$ and $H = G^\theta$. Then \[ SS^{(-1)} = (k - \lambda)1_H + \lambda w H. \] \end{lem} \begin{defi}\label{defi:difsum} Given a finite group $G$ and normal subgroup $N$, a $(v, k, \lambda)$-\emph{difference sum} is an element $S$ of $\mathbb{Z}[G/N]$ such that $SS^{(-1)} = (k - \lambda)1_{G/N} + \lambda |N| G/N$ and the coefficients of $S$ have values in $\{0, 1, \dots, |N|\}$. \end{defi} By construction, any difference set in $G$ induces difference sums under the natural projection in quotients of $G$, as seen in Figure~\ref{fig:induce}. Precisely, we have \begin{figure} \renewcommand{\arraystretch}{1.3} \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|cc} \cline{1-15} 1 & 1 & 1 & 0 & 0 & 1 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & \quad & $G/N_3$ \\ \cline{1-15} \multicolumn{3}{|c|}{3} & \multicolumn{3}{|c|}{1} & \multicolumn{3}{|c|}{1} & \multicolumn{3}{|c|}{1} & \multicolumn{3}{|c|}{1} & \quad & $G/N_2$ \\ \cline{1-15} \multicolumn{15}{|c|}{7} & \quad & $G/N_1$ \\ \cline{1-15} \end{tabular} \caption{A difference set of size 7 in the group $G$ of order 15 and the difference sums it induces in $G/N_i$ where $G = N_1 \triangleright N_2 \triangleright N_3 = \{1\}$. Each row in the diagram is a group, with each block a coset. \label{fig:induce}} \end{figure} \begin{lem}\label{lem:setinduce} Suppose $G$ is a finite group with normal subgroup $N$ and natural projection $\pi : G \to G/N$. Then any $(v, k, \lambda)$-difference set $D$ in $G$ induces a $(v, k, \lambda)$-difference sum $D^\pi$ in $G/N$. \end{lem} \begin{lem}\label{lem:suminduce} Suppose $G$ is a finite group with normal subgroups $N_1$ and $N_2$ such that $N_2 \subseteq N_1$ and $\pi : G/N_2 \to G/N_1$ is the natural projection. Then any $(v, k, \lambda)$-difference sum $S$ in $G/N_2$ induces a $(v, k, \lambda)$-difference sum $S^\pi$ in $G/N_1$. \end{lem} Lemma~\ref{lem:setinduce} means that our search for difference sets only requires checking the subsets of $G$ that induce difference sums in some quotient. In finding these difference sums, Lemma~\ref{lem:suminduce} additionally allows us to only test sums that induce difference sums in further quotients. In each case the search space is dramatically decreased. Since our search is for difference sets up to equivalence, we also define a complementary equivalence of difference sums such that equivalent difference sums are induced by equivalent collections of difference sets. \begin{defi}\label{defi:equivsums} Let $S_1$ and $S_2$ be difference sums in $G/N$. Then $S_1$ and $S_2$ are \emph{equivalent difference sums} if $S_1 = gS_2^\phi$ for some $g \in G/N$ and $\phi$ an automorphism of $G/N$ induced by an automorphism of $G$. \end{defi} \begin{lem}\label{lem:equivsums} Suppose $S_1$ and $S_2$ are equivalent difference sums in $G/N$. Then if $D_1$ is any difference set in $G$ that induces $S_1$, there exists a difference set $D_2$ in $G$ that induces $S_2$ such that $D_1$ and $D_2$ are equivalent. \end{lem} In the {\tt DifSets} package, difference sums are stored as lists of integers representing the coefficients of the group ring elements, with position in the list given by the position of the coset in the list returned by the {\tt GAP} function {\tt Elements(G/N)}. For example, {\tt [3, 1, 1, 1, 1]} represents a difference sum in {\tt SmallGroup(15, 1)} mod its normal subgroup of order 3 with coefficient 3 on the identity coset and coefficient 1 on all other cosets. \begin{verbatim} gap> G := SmallGroup(15, 1);; N := NormalSubgroups(G)[2];; gap> IsDifferenceSum(G, N, [3, 1, 1, 1, 1]); true \end{verbatim} \section{Algorithm}\label{sec:algorithm} The basic structure of the algorithm is to start at the bottom of Figure~\ref{fig:induce} and travel upwards. Given a group $G$, first compute $v = |G|$ and then find all values of $k$ that give solutions satisfying the Bruck-Ryser-Chowla theorem to the equation $k(k-1) = \lambda (v-1)$ mentioned in Section~\ref{sec:difsets}. For example, \begin{verbatim} gap> G := SmallGroup(15, 1);; gap> PossibleDifferenceSetSizes(G); [ 7 ] \end{verbatim} Each value of $k$ will be handled separately. The algorithm starts with the normal subgroup $N_1 = G$, where the only difference sum of size $k$ in $G/N_1 = \{1\}$ is {\tt [k]}. \begin{verbatim} gap> N1 := G;; gap> difsums := [ [7] ];; \end{verbatim} Given a normal subgroup $N_2$ of $G$ such that $N_2 \subseteq N_1$, first enumerate all preimages in $G/N_2$ of current difference sums in $G/N_1$ and return those that are themselves difference sums. Then remove all but one representative of each equivalence class from this collection. \begin{verbatim} gap> N2 := NormalSubgroups(G)[2];; gap> difsums := AllRefinedDifferenceSums(G, N1, N2, difsums); [ [ 1, 1, 1, 1, 3 ], [ 1, 1, 1, 3, 1 ], [ 1, 1, 3, 1, 1 ], [ 1, 3, 1, 1, 1 ], [ 3, 1, 1, 1, 1 ] ] gap> difsums := EquivalentFreeListOfDifferenceSums(G, N2, difsums); [ [ 3, 1, 1, 1, 1 ] ] \end{verbatim} In the general case, the above step is repeated along a chief series $G = N_1 \triangleright \dots \triangleright N_r = \{1\}$ of $G$ with $N_{r-1}$ a nontrivial normal subgroup of minimal possible size in $G$. At $N_{r-1}$, enumerate sets and remove equivalents to leave the final result. \begin{verbatim} gap> difsets := AllRefinedDifferenceSets(G, N2, difsums); [ [ 1, 2, 4, 3, 8, 11, 12 ], [ 1, 2, 4, 3, 10, 13, 12 ], [ 1, 2, 4, 5, 6, 9, 14 ], [ 1, 2, 4, 5, 10, 13, 14 ], [ 1, 2, 4, 7, 6, 9, 15 ], [ 1, 2, 4, 7, 8, 11, 15 ] ] gap> difsets := EquivalentFreeListOfDifferenceSets(G, difsets); [ [ 1, 2, 4, 7, 8, 11, 15 ] ] \end{verbatim} These steps are encapsulated in the function {\tt DifferenceSets} mentioned in Section~\ref{sec:introduction}, with two modifications. First, since every difference set is equivalent to some difference set containing the identity, the algorithm does not enumerate some preimages that are guaranteed to be equivalent to others. Second, the final elimination of all but one representative of equivalence classes of difference sets uses the {\tt SmallestImageSet} function \cite{Linton2004} from the {\tt GAP} package {\tt GRAPE} \cite{GRAPE}. Although roughly 20\% slower than the function given above for most cases, {\tt SmallestImageSet} gives a unique minimal result and handles groups with large automorphism groups much more efficiently. \section{Results} The {\tt DifSets} package successfully computed results for 1006 of the 1032 groups of order less than 100, including all groups of order 64 and 96. Full results with timings and comments can be found in the package and its documentation. Here we include a summary for order 64 and 96. All computations were performed with {\tt GAP} 4.9.1 on a 4.00GHz i7-6700K using 8GB of RAM. \begin{center} \begin{tabular}{ccccc} Order & Groups & Difference Sets & Median Time per Group & Total Time \\ \hline 64 & 267 & 330159 & 0.415 hours & 295.811 hours \\ 96 & 231 & 2627 & 3.133 hours & 1568.746 hours \end{tabular} \end{center} Timing comparisons with the {\tt RDS} package mentioned in Section~\ref{sec:introduction} are difficult since {\tt RDS} provides a variety of tools rather than a single algorithm. Ordered coset signatures in {\tt RDS} correspond to difference sums in {\tt DifSets}, but, unlike difference sums, coset signatures cannot be refined through multiple stages, which makes the generation of good coset signatures in {\tt RDS} infeasible for most order 64 and order 96 groups. However, if an ordered signature is available, building difference sets through partial difference sets in {\tt RDS} can in some cases be much faster than searching the corresponding difference sum using {\tt DifSets}. In particular, replacing the final step in Section~\ref{sec:algorithm} with a search using {\tt RDS} can significantly improve times for some groups of order 96. Further work to combine the refining of difference sums used by {\tt DifSets} and the generation of difference sets through partial difference sets used by {\tt RDS} could lead to significantly better times than either package could manage alone. \section*{Acknowledgements} The author thanks Ken Smith, Alexander Hulpke, and an anonymous reviewer for helpful comments that improved the {\tt DifSets} package and this article. \bibliographystyle{plain}

https://arxiv.org/abs/1401.2597

Multivariate Density Estimation via Adaptive Partitioning (I): Sieve MLE

We study a non-parametric approach to multivariate density estimation. The estimators are piecewise constant density functions supported by binary partitions. The partition of the sample space is learned by maximizing the likelihood of the corresponding histogram on that partition. We analyze the convergence rate of the sieve maximum likelihood estimator, and reach a conclusion that for a relatively rich class of density functions the rate does not directly depend on the dimension. This suggests that, under certain conditions, this method is immune to the curse of dimensionality, in the sense that it is possible to get close to the parametric rate even in high dimensions. We also apply this method to several special cases, and calculate the explicit convergence rates respectively.

\section{Estimation of functions of bounded variation} In image analysis, the denoised image is usually assumed to be a function of bounded variation. Obtaining an approximation is a crucial procedure before any downstream analysis. Currently, nonlinear approximation, such as wavelet compression \cite{DeVore} and wavelet shrinkage or thresholding \cite{donoho1995}, has been widely used in this field, contributing to problems including image compression and image segmentation. In this section, we treat the denoised image as a density function, and apply this multivariate density estimation method to obtain the approximation. We evaluate the performance of this method by calculating the convergence rate when the density function is of bounded variation (Section \ref{subsec:bdfunrate}). Actually, its performance is comparable to that of wavelet thresholding. This point becomes clearer in the companion papar \cite{linxiliu2015}, where under Bayesian settings we show that minimax convergence rate can be achieved for the space of bounded variation (BV) up to a logarithmic term. To begin with, we briefly introduce the space BV in Section \ref{subsec:bdfun}. \subsection{The space BV} \label{subsec:bdfun} Let $\Omega= [0, 1)^2$ be a domain in $\mathbb{R}^2$. We define the space $BV(\Omega)$ of functions of bounded variation on $\Omega$ as follows. For a vector $\nu \in \mathbb{R}^2$, the difference operator $\Delta_{\nu}$ along the direction $\nu$ is defined by \begin{equation*} \Delta_{\nu}(f, y) := f( y+\nu) - f(y). \end{equation*} For functions $f$ defined on $\Omega$, $\Delta_{\nu}(f, y)$ is defined whenever $y \in \Omega(\nu)$, where $\Omega(\nu) := \{ y: [y, y+\nu] \subset \Omega \}$ and $[y, y+\nu]$ is the line segment connecting $y$ and $y+\nu$. Denote by $e_l, l=1,2$ the two coordinate vectors in $\mathbb{R}^2$. We say that a function $f \in L_1 (\Omega)$ is in $BV(\Omega)$ if and only if \begin{equation*} V_{\Omega}(f) := \sup_{h>0} h^{-1} \sum_{l=1}^2 \| \Delta_{h e_l} (f, \cdot)\|_{L_1 (\Omega(h e_l) )} = \lim_{h \rightarrow 0} h^{-1} \sum_{l=1}^2 \| \Delta_{h e_l} (f, \cdot)\|_{L_1 (\Omega(h e_l) )} \end{equation*} is finite. The quantity $V_{\Omega}(f)$ is the \emph{variation} of $f$ over $\Omega$. $p$-dimensional bounded variation function can be defined similarly. In \cite{donoho1993}, the author demonstrated that, in one dimension wavelet representations of bounded variation balls are optimal. The result of optimality can be extended to two-dimensional cases as well. Given the fact that the density functions defined on binary partitions are essentially equivalent to the Haar bases, it is natural to ask what the convergence rate will be if we apply this multivariate density estimator to $BV(\Omega)$. This result will be presented in the section below. \subsection{Convergence rate} \label{subsec:bdfunrate} In this section, we still use the Haar basis defined in Section \ref{subsec:Haarbasis1}. Let $\Lambda$ be the set of indices for the wavelet basis. Each element in $\Lambda$ is a pair of scale and location parameters. We will denote by $\Sigma_N$ the spaces consisting of $N$-term approximation in the Haar system, in other words, \begin{equation*} \Sigma_N := \{ \sum_{\lambda \in E} c_{\lambda} \psi_{\lambda}: E \subset \Lambda, |E| \leq N \}, \end{equation*} where $|E|$ denotes the cardinality of the discrete set $E$. First, we cite a theorem from \cite{cohen1999}. It provides a result on the approximation rate to a function of bounded variation by $\Sigma_N$. \begin{lemma} If $f \in BV(\Omega)$ has mean value zero on $\Omega$, we have \begin{equation} \label{eq:bdapproximation} \inf_{g \in \Sigma_N} \| f - g \|_{L_2(\Omega)} \leq C N^{-1/2} V_{\Omega} (f), \end{equation} with $C=2592(3\sqrt{5} +\sqrt{3})$. \end{lemma} Assume $f_0$ is a density function on $\Omega$ of bounded variation. By subtracting the mean, we can always assume that $\sqrt{f_0}$ has mean value zero over $\Omega$. For the square root of $f_0$, applying the lemma above, we can find an $N$-term approximation $g$ in the Haar system, such that $\|\sqrt{f_0} -g \|_{L_2(\Omega)} \lesssim N^{-1/2}$. Translating this inequality into the size of partition, we reach the conclusion that for a density function in $BV(\Omega)$, we can find an approximation in $\Theta_I$, such that $\rho(f_0, f_I) \lesssim I^{-1/2}$. Now we are ready to state the result of the convergence rate. \begin{theorem} Assume that $f_0 \in BV(\Omega)$. If we apply the multivariate density estimator based on adaptive partitioning here to estimate $f_0$, the convergence rate is $n^{-1/4} (\log n)^{3/4}$. \end{theorem} \begin{proof} The proof follows Theorem \ref{th:convergencerate} and the previous approximation result directly. \end{proof} \section{Introduction} Density estimation is a fundamental problem in statistics. Once an explicit estimate of the density function is obtained, various kinds of statistical inference can follow, including non-parametric testing, clustering, and data compression. Previous research focused on both parametric and nonparametric density estimation methods. However, currently, increasing dimension and data size impose great difficulty on these traditional methods. For instance, a fixed parametric family, such as multivariate Gaussian, may fail to capture the spatial features of the true density function under high dimensions. On the other hand, a traditional nonparametric method, like the kernel density estimator, may suffer from the difficulty of choosing appropriate bandwidths \cite{jones1996}. In this paper, we study a nonparametric method for multivariate density estimation. This is a sieve maximum likelihood method which employs simple, but still flexible, binary partitions to adapt to the data distribution. In the paper, we will carry out thorough analyses of the convergence rate to quantify the performance of this method as the dimension increases and the regularity of the true density function varies. These analyses demonstrate the major advantages of this method, especially when the dimension is moderately large (e.g. 5 to 50). \subsection{Challenges in multivariate density estimation} Most of the established methods for density estimation were initially designed for the estimation of univariate or low-dimensional density functions. For example, the popular kernel method (\cite{rosenblatt1956} and \cite{parzen1962}), which approximates the density by the superposition of windowed kernel functions centering on the observed data points, works well for estimating smooth low-dimensional densities. As the dimension increases, the accuracy of the kernel estimates becomes very sensitive to the choice of the window size and the shape of the kernel. To obtain good performance, both of these choices need to depend on the data. However, the question of how to adapt these parameters to the data has not been adequately addressed. This is especially the case for the kernel which itself is a multidimensional function. As a result, the performance of current kernel estimators deteriorates rapidly as the dimension increases. The difficulty caused by high dimensionality is also revealed in a classic result by Charles Stone \cite{stone1980}. In this paper, it was shown that the optimal rate of convergence for density in $d$-dimensional space, when the density is assumed to have $p$ bounded derivatives, is of the order $n^{-\alpha}$, where $\alpha = p/(2p+d)$. When $d$ is small and the density is smooth (i.e. $p$ is large), then methods such as kernel density estimation can achieve a convergence rate almost as good as the parametric rate of $n^{-1/2}$. However, when $d$ is large, then even if the density has many bounded derivatives, the best possible rate will still be unacceptably slow. Thus standard smoothness assumptions on the density will not protect us from the ``curse of dimensionality''. Instead, we must seek alternative conditions on the underlying class of densities that are general enough to cover some useful applications under high dimensions, and yet strong enough to enable the construction of density estimators with fast convergence. More specifically, suppose $r$ is a parameter that controls the complexity (in a sense to be made precise) of the density class, with large value of $r$ indicating low complexity. We would like to construct density estimators with a convergence rate of the order $n^{-\gamma(r)}$, where $\gamma(\cdot)$ is an increasing function not as sensitive to $d$ as that of the traditional methods, and satisfying the property that $\gamma(r) \uparrow \frac{1}{2}$ as $r \uparrow \infty$. Since this rate is not sensitive to $d$, it is possible to obtain fast convergence even in high dimensional cases. For density estimators based on adaptive partitioning, such a result is established in Theorem \ref{th:convergencerate} below. \subsection{Adaptive partitioning} \label{subsec:reviewofmethod} The most basic method for density estimation is the histogram. With appropriately chosen bin width, the histogram density value within each bin is proportional to the relative frequency of the data points in that bin. Further developments of the method allow the bins to depend on data, and substantial improvement can be obtained by such ``data-adaptive'' histograms (\cite{scott1979}). This idea has been naturally extended to multivariate cases. Multivariate histograms with data-adaptive partitions have been studied in \cite{shang1994} and \cite{Ooi2002}. The breakthrough work of Lugosi and Nobel \cite{lugosi1996} presented general sufficient conditions for the almost sure $L_1$-consistency based on data-dependent partitions. Later in \cite{barron1999}, the authors constructed a multivariate histogram which achieves asymptotic minimax rates over anisotropic H\"older classes for the $L_2$ loss. Another closely-related type of methods is multivariate density estimation based on wavelet expansions (\cite{donoho1996} and \cite{tribouley1995}). Along this line, in \cite{neumann2000} and \cite{klemela2009} the authors showed that estimators based on wavelet expansions achieve minimax convergence rates up to a logarithmic factor over a large scale of anisotropic Besov classes. Apart from these two types of methods, in a recent work \cite{wong2010} by Wong and Ma, the authors proposed a Bayesian formulation to learn the data-adaptive partition in multi-dimensional cases. By employing sequential importance sampling (\cite{KLW} and \cite{liu2001monte}), they designed efficient algorithms (\cite{LJW} and \cite{jiang2015}) to sample from the posterior distribution. The methods are also shown to perform well empirically in a range of continuous and discrete problems, and achieves satisfying performance. \subsection{Contribution of the paper} In this paper we study the asymptotic properties of density estimators based on adaptive partitioning. The data-adaptive partition is obtained by maximizing the likelihood of the corresponding histogram on that partition. We start by formulating a complexity index (denoted by $r$) for a density, with large value of $r$ indicating low complexity in the sense that the density can be approximated at a fast rate by piecewise constant density functions as the size of the underlying partition increases. We assume that the complexity of the true density is known, and study how the size of the partition of our density estimator should be chosen in order to achieve fast convergence to the true density. Our analysis shows that roughly (i.e. up to $\log n$ factors) the achievable rate is $n^{-(r/(2r+1))}$. Thus when $r$ is large, our estimate will converge to the true density at a rate close to the parametric rate of $n^{-1/2}$, not directly depending on the dimension $d$ of the sample space. This is in contrast with the achievable convergence rate under smoothness condition (\cite{stone1980}), which deteriorates rapidly as the dimensional $d$ increases. In order to gain a deeper understanding of our complexity index, in this paper we also perform explicit computation of $r$ for density functions with certain spatial sparsity properties as well as functions of bounded variation. Our results also imply that, when the true density depends only on a subset of the variables, our density estimator will automatically exhibit a ``variable selection'' property. Indeed, the function classes studied in this paper correspond to Besov classes and anisotropic H\"older classes studied by the previous literature (\cite{barron1999}, \cite{neumann2000} and \cite{klemela2009}). But the density functions are characterized by the conditions which are easier to verify and closer to statistical models. In addition to this, in most of the previous literature regarding convergence rates, the estimator is constructed for a specific function space. Here, we introduce the estimator under a quite general formulation, and then derive the convergence rates by calculating complexity indices $r$ for different function classes. While the results in this paper can provide insights given the complexity index $r$ of the true density, in practice we do not know $r$. This raises the question of how to make our density estimator adaptive to the unknown complexity, in the sense that it will automatically achieve the optimal rate $n^{-(r/(2r+1))}$ even when $r$ is not known to us. This question will be taken up in a companion paper (\cite{linxiliu2015}). Extending the present analysis, we will show in the companion paper that adaptation to unknown complexity can be achieved by a fully Bayesian approach with an exponentially decreasing prior distribution on the size of the partition. The rest of the paper is organized in the following way. In Section \ref{sec:notation}, we discuss the partition scheme, introduce the estimation method, and summarize our main results on convergence rate. Section \ref{sec:rate} focuses on the proof of the main theorem. If the reader is not interested in the mathematical details, this section can be skipped without affecting the understanding of the later part. From Section \ref{sec:spatialadaptation} to Section \ref{sec:variableselection}, we apply our main results to spatial adaptation, estimation of functions of bounded variation, and variable selection cases respectively. \section*{Acknowledgements} The authors would like to thank Emmanuel Cand\`es, Matan Gavish and Xiaotong Shen for helpful discussions. \bibliographystyle{imsart-nameyear} \section{The maximum likelihood estimators based on adaptive partitioning} \label{sec:notation} Let $Y_1, Y_2, \cdots, Y_n$ be a sequence of independent random variables distributed according to a density $f_0(y)$ with respect to a $\sigma$-finite measure $\mu$ on a measurable space $(\Omega, \mathcal{B})$. We are interested in the case when $\Omega$ is a bounded rectangle in $\mathbb{R}^p$ and $\mu$ is the Lebesgue measure. After translation and scaling, we may assume that the sample space is the unit cube in $\mathbb{R}^p$, that is, $\Omega=\{(y^1, y^2, \cdots, y^p): y^l \in [0,1]\}$. Let $\mathcal{F}=\{f \mbox{ is a nonnegative measurable function on } \Omega: \int_\Omega f d\mu =1 \}$ be the collection of all the density functions on $(\Omega, \mathcal{B}, \mu)$. $\mathcal{F}$ constitutes the parameter space in this problem. \subsection{Densities on binary partitions} \label{subsec:sieve} Similar to \cite{LJW}, we use \emph{binary partitions} to capture the features of the true density function. By increasing the complexity of the partitions, we construct a sequence of density spaces $\Theta_1, \Theta_2, \cdots, \Theta_I, \cdots$. These spaces are finite dimensional approximations to the infinite dimensional parameter space $\mathcal{F}$ with decreasing approximation error. A detailed description of these spaces is as follows. First, we use a recursive procedure to define a binary partition with $I$ subregions of the unit cube in $\mathbb{R}^p$. Let $\Omega=\{(y^1, y^2, \cdots, y^p): y^l \in [0,1]\}$ be that unit cube. In the first step, we choose one of the coordinates $y^l$ and cut $\Omega$ into two subregions along the midpoint of the range of $y^l$. That is, $\Omega=\Omega^l_0 \cup \Omega^l_1$, where $\Omega^l_0 = \{y \in \Omega: y^l \leq 1/2\}$ and $\Omega^l_1 = \Omega \backslash \Omega^l_0$. In this way, we obtain a partition with two subregions. It is easily observed that the total number of all possible partitions after one cut is equal to the dimension $p$. Suppose after $I-1$ steps of the recursion, we already obtained a partition $\{ \Omega_i \}_{i=1}^I$ with $I$ subregions. In the $I$-th step, further partitioning of the region is defined as follows: \begin{enumerate} \item Choose a region from $\Omega_1, \cdots, \Omega_I$. Denote it as $\Omega_{i_0}$. \item Choose one coodinate $y^l$ and divide $\Omega_{i_0}$ into two subregions along the midpoint of the range of $y^l$. \end{enumerate} Such a partition obtained by $I$ recursive steps is called a binary partition of size $I+1$. Figure~\ref{fig:binarypartition} displays all the two dimensional binary partitions when $I$ is 1, 2 and 3. \begin{figure} \label{fig:binarypartition} \centering \includegraphics[scale=0.4]{partition.jpg} \caption{Binary partitions} \end{figure} Now, we define \begin{eqnarray*} \Theta_I = \{f \in \mathcal{F}: f=\sum_{i=1}^I \beta_i \mathbbm{1}_{\Omega_i},\ \mbox{where\ }\sum_{i=1}^I \beta_i \mu(\Omega_i) =1,\\\mbox{and\ } \{\Omega_i\}_{i=1}^I\mbox{ is\ a\ binary\ partition\ of\ $\Omega$\ of\ size\ $I$.}\} \end{eqnarray*} This is to say, $\Theta_I$ is the collection of piecewise constant density functions supported by the binary partitions of size $I$. Then these $\Theta_I$ constitute a sequence of approximating spaces to $\mathcal{F}$ (i.e. a sieve, see \cite{Grenander} and \cite{shen1994} for background on sieve theory). We take the metric on $\mathcal{F}$ and $\Theta_I$ to be Hellinger distance, which is defined by \begin{equation} \label{al:hellinger} \rho(f,g)=\left(\int_{\Omega} \left(\sqrt{f(y)} - \sqrt{g(y)} \right)^2 dy \right)^{1/2},\ f,g \in \mathcal{F}. \end{equation} For $f, g \in \Theta_I$, let $f= \sum_{i=1}^I \beta_i^1 \mathbbm{1}_{\Omega_i^1}$ and $g=\sum_{i=1}^I \beta_i^2 \mathbbm{1}_{\Omega_i^2}$, where $\{ \Omega_i^1\}_{i=1}^I$ and $\{\Omega_i^2\}_{i=1}^I$ are binary partitions of $\Omega$. Then the Hellinger distance between $f$ and $g$ can be written more explicitly as \begin{equation} \rho^2(f, g)=\sum_{i=1}^I \sum_{j=1}^I \left(\sqrt{\beta_i^1} - \sqrt{\beta_j^2}\right)^2 \mu \left(\Omega_i^1 \cap \Omega_j^2 \right). \end{equation} We also introduce the Kullback-Leibler divergence, which is defined to be \begin{equation*} K(f,g) = \mathbb{E}_{f} \left( \log \frac{f(Y)}{g(Y)} \right). \end{equation*} Note the Kullback-Leibler divergence is a stronger distance compared to the Hellinger distance, in the sense that for any $f, g \in \mathcal{F}$, $ \rho^2(f, g) \leq K(f,g)$. We further restrict our interest to a subsect $\mathcal{F}_0 \subset \mathcal{F}$ of densities which satisfy the following two conditions: First, $\int_{\Omega} f^2 < \infty$. Second, for any $f \in \mathcal{F}_0$, there exists a sequence of approximations $f_I\in \Theta_I$ such that $\rho(f, f_I ) \leq A I^{-r}$, where $r$ is a parameter characterizing the decay rate of the approximation error, $A$ is a constant that may depend on $f$. In order to demonstrate that $\mathcal{F}_0$ is still a rich class, from Section \ref{sec:spatialadaptation} to Section \ref{sec:variableselection}, we study several specific density classes belonging to $\mathcal{F}_0$, which are frequently occurred in statistical modelding. \subsection{The sieve MLE} For any $f \in \Theta_I$, the log-likelihood is defined to be \begin{equation} \label{eq:loglikeli} L_n(f) = \sum_{j=1}^n \log f(Y_j) = \sum_{i=1}^I N_i \log \beta_i, \end{equation} where $N_i$ is the count of data points in $\Omega_i$, i.e., $N_i= \mbox{card}\{j: Y_j \in \Omega_i, 1\leq j\leq n\}$. The maximum likelihood estimator on $\Theta_I$ is defined to be \begin{equation} \label{eq:sievemle} \hat{f}_{n,I} = \arg\max_{f \in \Theta_I} L_n(f). \end{equation} We claim that $\hat{f}_{n,I}$ is well defined. This is true because given the binary partition $\{ \Omega_i\}_{i=1}^I$, the underlying distribution becomes a multinomial one, and $(\beta_1, \cdots, \beta_I)$ can be determined by maximizing the log-likelihood. And within each $\Theta_I$, the number of possible binary partitions is finite. Since $\Theta_I$ constitute a sieve to $\mathcal{F}_0$, this sequence of estimators is also called the sieve maximum likelihood estimator (sieve MLE). In order to illustrating how the idea of partition learning is incorporated in this framework, given the binary partition $\mathcal{A} =\{ \Omega_i \}_{i=1}^I$, we can derive the maximum of the likelihood values achieved by histograms on that partition, which has a closed-form expression \begin{equation*} L_n (\mathcal{A}) = \sum_{i=1}^I N_i \log \left( \frac{N_i}{n \mu(\Omega_i)} \right). \end{equation*} We treat this as a \emph{score} of the partition $\mathcal{A}$. By maximizing the score over all the binary partitions of a fixed size, we learn a most promising one which adapts to the true density function. Then $\hat{f}_{n,I}$ is simply the histogram based on that partition. \subsection{Main results on convergence rate} Having defined a sequence of maximum likelihood estimators $\hat{f}_{n,I}$, we are now ready to state the main results on the rate at which $\hat{f}_{n,I}$ converges to $f_0$. \begin{theorem} \label{th:convergencerate} For any $f_0 \in \mathcal{F}_0$, $\hat{f}_{n,I}$ is the maximum likelihood estimator over $\Theta_I$. $A$ and $r$ are the parameters that characterize achievable approximation errors to $f_0$ by the elements in $\Theta_I$. Assume that $n$ and $I$ satisfy \begin{equation} \label{eq:thmC2} I = \left( (2^8 A^2 r /c_1) \frac{n}{\log n} \right)^{\frac{1}{2r+1}}, \end{equation} where the constant $c_1$ can be chosen to be in $(0, 1)$. Then the convergence rate of the sieve MLE is $ n^{-\frac{r}{2r+1}} (\log n) ^{(\frac{1}{2}+\frac{r}{2r+1})}$, in the sense that \begin{equation*} \mathbb{P}_{f_0} \left( \rho (\hat{f}_{n,I} , f_0 ) \geq D n^{-\frac{r}{2r+1}} (\log n) ^{(\frac{1}{2}+\frac{r}{2r+1})} \right) \rightarrow 0, \end{equation*} where $D$ is a constant. \end{theorem} \begin{remark} It is possible to partition the sample space in a more flexible way. In particular, the analysis and resulting rate remain the same if we replace binary partition at the mid-point by binary partition at a point chosen from a fixed sized grid (e.g. regular equi-spaced grid). \end{remark} From this theorem, we see that the convergence rate does not directly depend on the dimension of the problem. Instead, it only depends on how well the true density can be approximated by the sieve. As $r$ increases, up to a $\log n$ term, the rate gets close to the parametric rate of $n^{-1/2}$. The analysis in this paper will demonstrate that the optimal convergence rate can be achieved by balancing the sample size with the complexity of the approximating spaces. On one hand, the complexity of $\Theta_I$ affects the convergence rate in a way that, the richer the approximating spaces the lower bias the estimators have. Conversely, given a sample $Y_1, Y_2, \cdots, Y_n$ of fixed size, there is a point beyond which the limited amount of information conveyed in the data may be overwhelmed by the overly-complex approximating spaces. A major contribution of the main result is that it clarifies how to strike the balance between the sample size and the complexity of the approximating spaces. \section{Proof of Theorem \ref{th:convergencerate}} \label{sec:rate} This section is devoted to the proof of the main theorem. Studies of convergence rate invariably rely on the previous results from studies of empirical process indexed by log-likelihood ratios. However, while the results in the classical work by Wong and Shen (\cite{shen1994} and \cite{wong1995}) are the most applicable ones to our study, they must be modified to adapt to the current settings. The following is an outline of the proof. In Section \ref{subsec:entropy} we briefly discuss {\em metric entropy with bracketing}, which measures the complexity of the approximating spaces by ``counting'' how many pairs of functions in an $\epsilon$-net are needed to provide simultaneous upper and lower bounds of all the elements. An important result of this section is an upper bound for the bracketing metric entropy of $\Theta_I$. In Section \ref{subsec:likelihoodratio}, a large-deviation type inequality for the likelihood ratio surface follows. Previous results on the convergence rate of sieve MLE assume that the true parameter can be approximated by the sieve under Kullback-Leibler divergence. Here, we switch to the weaker Hellinger distance, because under the current settings, we can obtain an explicit bound for the Kullback-Leibler divergence in terms of the Hellinger distance. Thus our results are more general than those in \cite{wong1995} for this type of density estimation problems. Finally in Section \ref{subsec:rate}, we establish the main result of the convergence rate. After splitting the tail probability into two parts corresponding to variance and bias, we apply the results obtained in Section \ref{subsec:likelihoodratio} and \ref{subsec:KLbound} to bound each of them respectively. \subsection{Calculation of the metric entropy with bracketing} \label{subsec:entropy} A general discussion of {\em metric entropy} can be found in \cite{kolmogorov1992selected}. In this section, we introduce a form of metric entropy with bracketing corresponding to current parameter space, and provide an upper bound for the bracketing metric entropy of the approximating spaces defined in Section \ref{subsec:sieve}. \begin{definition} Let $ (\Theta, \rho)$ be a seperable pseudo-metric space. $\Theta(\epsilon)$ is a finite set of pairs of functions $\{ (f_j^L, f_j^U), j=1,\cdots, N\}$ satisfying \begin{equation} \label{eq:entropyDefC1} \rho( f_j^L, f_j^U) \leq \epsilon\ for\ j=1,\cdots,N, \end{equation} and for any $f \in \Theta$, there is a $j$ such that \begin{equation} \label{eq:entropyDefC2} f_j^L \leq f \leq f_j^U. \end{equation} Let \begin{equation} \label{eq:entropyDefN} N(\epsilon, \Theta, \rho) = min \{card\ \Theta(\epsilon): (\ref{eq:entropyDefC1})\ and\ (\ref{eq:entropyDefC2})\ are\ satisfied \}. \end{equation} Then, we define the metric entropy with bracketing of $\Theta$ to be \begin{equation} \label{eq:entropyDef} H(\epsilon, \Theta, \rho) = \log N(\epsilon, \Theta, \rho). \end{equation} \end{definition} Recall that $\Theta_1, \cdots, \Theta_I, \cdots$ are the approximating spaces defined in section \ref{subsec:sieve}. The next two lemmas are devoted to an upper bound for the bracketing metric entropy of $\Theta_I$. \begin{lemma} \label{lemma:boundofH} Take $\rho$ to be the Hellinger distance. Let $\Theta_I^{\mathcal{A}, d} = \{ f \in \Theta_I:\ f\ is\ supported\ by\ the\ binary\ partition\ \mathcal{A} = \{ \Omega_i\}_{i=1}^I,\ and\ \rho(f, f_0) \leq d \}$. Then, \begin{equation} \label{eq:boundofH} H(u, \Theta_I^{\mathcal{A}, d}, \rho) \leq \frac{I}{2} \log I + I \log\frac{d}{u} + b', \end{equation} where $b'$ is a constant not dependent on the binary partition. \end{lemma} \begin{proof} Assume $f=\sum_{i=1}^I \beta_i \mathbbm{1}_{\Omega_i}$. When the binary partitions $\{\Omega_i\}_{i=1}^I$ are fixed, there exits a one-to-one correspondance between any $f \in \Theta_I^{\mathcal{A}, d}$ and an $I$-dimensional vector $(\sqrt{\beta_1 \mu(\Omega_1)}, \cdots, \sqrt{\beta_I \mu(\Omega_I)})$. As a consequence of Cauchy-Schwartz inequality, \begin{eqnarray*} \rho(f, f_0)^2&=&\sum_{i=1}^I \int_{\Omega_i} ( \sqrt{\beta_i} -\sqrt{f_0(x)} )^2 \mu(dx)\\ &\geq&\sum_{i=1}^I \mu(\Omega_i) (\int_{\Omega_i} (\sqrt{\beta_i}-\sqrt{f_0(x)}) \frac{\mu(dx)}{\mu(\Omega_i)})^2\\ &=&\sum_{i=1}^I \mu(\Omega_i) (\sqrt{\beta_i} - \frac{\int_{\Omega_i} \sqrt{f_0(x)} \mu(dx)} {\mu(\Omega_i)})^2. \end{eqnarray*} Then, we have, \begin{eqnarray*} & & \{(\sqrt{\beta_1 \mu(\Omega_1)}, \cdots, \sqrt{\beta_I \mu(\Omega_I)}): \sum_{i=1}^I \int_{\Omega_i} ( \sqrt{\beta_i} -\sqrt{f_0(x)} )^2 \mu(dx) \leq d^2\}\\ &\subset& \{(\sqrt{\beta_1 \mu(\Omega_1)}, \cdots, \sqrt{\beta_I \mu(\Omega_I)}): \sum_{i=1}^I (\sqrt{\beta_i \mu(\Omega_i)} - \frac{\int_{\Omega_i} \sqrt{f_0(x)} \mu(dx)} {\sqrt{\mu(\Omega_i)}} )^2 \leq d^2 \} \\ &=:& B_I ^{\mathcal{A}, d}. \end{eqnarray*} If we treat the element in $\Theta_I^{\mathcal{A},d}$ as the $I$-dimensional vector $( \sqrt{\beta_1 \mu(\Omega_1)}, \cdots, \sqrt{\beta_I \mu(\Omega_I )} )$, then from the above inclusion relation we learn that, $\Theta_I^{\mathcal{A},d} \subset B_I^{\mathcal{A},d}$. We also note that the Hellinger distance on $B_I^{\mathcal{A},d}$ is equivalent to the $L_2$ norm on the $I$-dimensional Euclidean space. Thus, \begin{equation} \label{eq:hellingertol2} N(u, \Theta_I^{\mathcal{A}, d}, \rho) \leq N(u, B_I^{\mathcal{A}, d},\| \cdot \|_2). \end{equation} Because the metric entropy is invariant under translation, calculating the bracketing metric entropy of $B_I^{\mathcal{A}, d}$ is equivalent to calculating that of \begin{equation*} \tilde{B}_I^{\mathcal{A}, d} := \{ \left(\sqrt{\beta_1 \mu(\Omega_1)}, \cdots, \sqrt{\beta_I \mu(\Omega_I)} \right): \sum_{i=1}^I \left( \sqrt{\beta_i \mu(\Omega_i)} \right)^2 \leq d^2 \}. \end{equation*} The unit sphere under $L_2$-norm is \begin{equation*} S=\{ \left(\sqrt{\beta_1 \mu(\Omega_1)}, \cdots, \sqrt{\beta_I \mu(\Omega_I)} \right): \sum_{i=1}^I \left(\beta_i \mu(\Omega_i) \right) \leq 1 \}. \end{equation*} The unit sphere under $L_{\infty}$-norm is \begin{equation*} S_{\infty} = \{ \left(\sqrt{\beta_1 \mu(\Omega_1)}, \cdots, \sqrt{\beta_I \mu(\Omega_I)} \right): \max_{1 \leq i \leq I} \sqrt{\beta_i \mu(\Omega_i)} \leq 1 \}. \end{equation*} Note that $ \max_{1 \leq i \leq I} \sqrt{\beta_i \mu(\Omega_i)} \leq 1/ \sqrt{I}$ implies that $\sum_{i=1}^I \beta_i \mu(\Omega_i) \leq 1$, and $\sum_{i=1}^I \beta_i \mu(\Omega_i) \leq d^2$ implies that $\max_{1 \leq i \leq I}\sqrt{\beta_i \mu(\Omega_i)} \leq d$, we have \begin{equation} \label{eq:relation} S \subset \tilde{B}_I^{\mathcal{A}, d} \subset d S_{\infty} \mbox{\ and\ } S_{\infty} \subset \sqrt{I} S. \end{equation} Therefore, \begin{eqnarray*} N(u, \Theta_I ^{\mathcal{A}, d}, \rho) &\leq& N(u, \tilde{B}_I^{\mathcal{A}, d}, \| \cdot \|_2) \\ &\leq& \left(\frac{d \sqrt{I}}{u} +2 \right)^I \\ &\leq& b' I^{I/2} (d/u)^I, \end{eqnarray*} where $b'$ is a constant not dependent on the partiton. The desired result follows. \end{proof} \begin{lemma} \label{lemma:dballupperbound} Under the same assumptions as in Lemma \ref{lemma:boundofH}, let $\Theta_I^d = \{ f \in \Theta_I: \rho(f, f_0) \leq d\}$. Then, \begin{eqnarray} \label{eq:dballupperbound} \nonumber & &H(u, \Theta_I^d, \rho) \\ &\leq& I \log p + (I+1) \log(I+1) + \frac{I}{2} \log I + I \log \frac{d}{u} +b, \end{eqnarray} where $b$ is a constant not dependent on $I$ or $d$. \end{lemma} \begin{proof} According to the construction of the sieve, given the size $I$, the number of possible binary partitions is upper bounded by $p^I I!$ ($p$ is the dimension of the Euclidean space). Therefore, \begin{eqnarray*} \nonumber N(u, \Theta_I^d, \rho) &\leq& p^I I! N(u, \Theta_I^{\mathcal{A}, d}, \rho) \\ &\leq& b' p^I I! I^{I/2} (d/u)^I, \end{eqnarray*} and, \begin{eqnarray} \label{eq:upperE5} \nonumber & &\ H(u, \Theta_I^d,\rho) \\ &\leq& I \log p + (I+1) \log(I+1) + \frac{I}{2} \log I + I \log \frac{d}{u} +b. \end{eqnarray} \end{proof} \subsection{An inequality for the likelihood ratio surface} \label{subsec:likelihoodratio} In this section, we focus on bounding the tail behavior of the likelihood ratio. First, we cite a theorem in \cite{wong1995}, which gives a uniform exponential bound for likelihood ratios. \begin{theorem}[Wong and Shen (1995)] \label{thm:likelihoodratio} Let $\rho$ be the Hellinger distance and $\mathcal{P}_n$ be a space of densities. There exist positive constants $a>0$, $c$, $c_1$ and $c_2$, such that, for any $\epsilon>0$, if \begin{equation} \label{eq:entropycondition} \int_{\epsilon^2/2^8}^{\sqrt{2}\epsilon}H^{1/2} ( u/a, \mathcal{P}_n, \rho) du \leq c n^{1/2} \epsilon^2, \end{equation} then \begin{equation*} \mathbb{P}_{f_0} \Big( \sup_{\{ \rho(f, f_0) \geq \epsilon, f \in \mathcal{P}_n \}} \prod_{i=1}^n \frac{f(Y_i)}{f_0(Y_i)} \geq \exp( -c_1 n \epsilon^2) \Big) \leq 4 \exp(-c_2 n \epsilon^2), \end{equation*} where $\mathbb{P}_{f_0}$ is understood to be the outer probability mesure under $f_0$. The constants $c_1$ and $c_2$ can be chosen in $(0,1)$ and $c$ can be set as $(2/3)^{5/2} /512$. \end{theorem} Combining the theorem with the entropy bounds in Section \ref{subsec:entropy} gives the desired inequality, which is summarized in the next corollary. \begin{corollary} \label{cor:largedeviation} Let $\delta_{n,I} = (\frac{I \log I}{n / \log n}) ^{1/2}$. When $n$ and $I$ are sufficiently large, we have \begin{equation*} \mathbb{P}_{f_0} \Big( \sup_{\{ \rho(f, f_0) \geq \delta_{n,I}, f \in \Theta_I \}} \prod_{i=1}^n \frac{f(Y_i)}{f_0(Y_i)} \geq \exp( -c_1 n \delta_{n,I}^2) \Big) \leq 4 \exp(-c_2 n \delta_{n,I}^2). \end{equation*} \end{corollary} \begin{proof} The key to the proof is to check condition (\ref{eq:entropycondition}) so that we can apply Theorem \ref{thm:likelihoodratio} to the sequence of approximating spaces $\Theta_I$. By Lemma \ref{lemma:dballupperbound}, \begin{eqnarray} \label{eq:entropybound} \nonumber & &\int_{\delta_{n,I}^2/2^8}^{\sqrt{2} \delta_{n,I}} H^{1/2} ( u/a, \Theta_I, \rho) du \\ \nonumber &\leq& \int_{\delta_{n,I}^2/2^8}^{\sqrt{2} \delta_{n,I}} \left( I \log (4pa) + 2 (I+1) \log(I+1) -I \log u \right)^{1/2} du \\ &\sim& I^{1/2} \int_{\delta_{n,I}^2/2^8}^{\sqrt{2} \delta_{n,I}} \left( \log\frac{I^2}{u} \right)^{1/2} du. \end{eqnarray} We calculate the integral and obtain \begin{eqnarray*} & & \eqref{eq:entropybound}\\ &\leq& I^{1/2} \left( u \sqrt{ \log \frac{I^2}{u} } - \frac{\sqrt{\pi}}{2} I^2 \mbox{erf} \left( \sqrt{ \log \frac{I^2}{u} } \right) \right)|_{\delta_{n,I}^2/2^8}^{\sqrt{2} \delta_{n,I}} \\ &\leq& I^{1/2} ( \sqrt{2} \delta_{n,I} \sqrt{ \log \frac{I^2}{ \sqrt{2} \delta_{n,I}} } - \frac{\sqrt{\pi}}{2} I^2 \mbox{erf} \left( \sqrt{ \log \frac{I^2}{ \sqrt{2}\delta_{n,I} } } \right) -\frac{\delta_{n,I}^2}{2^8} \sqrt{ \log \frac{2^8 I^2}{ \delta_{n,I}^2} }\\ & & + \frac{\sqrt{\pi}}{2} I^2 \mbox{erf} \left( \sqrt{ \log \frac{2^8 I^2}{\delta_{n,I}^2} } \right) ) \\ &\sim& I^{1/2} \left( \delta_{n,I} \sqrt{ \log \frac{I^2}{\delta_{n,I}} } + \frac{\sqrt{\pi}}{2} I^2 \left(\mbox{erf} ( \sqrt{ \log \frac{2^8 I^2}{\delta_{n,I}^2} } ) -\mbox{erf} ( \sqrt{ \log \frac{I^2}{\sqrt{2} \delta_{n,I}} }) \right) \right) \\ &\sim& I^{1/2} \left( \delta_{n,I} \sqrt{ \log \frac{I^2}{\delta_{n,I}} } + \frac{\sqrt{\pi}}{2} I^2 \left( 1 - \frac{ \frac{ \delta_{n,I}^2 }{ 2^8 I^2}} {\sqrt{ \pi \log \frac{2^8 I^2}{\delta_{n,I}^2}}} -1 + \frac{ \frac{\sqrt{2} \delta_{n,I} }{I^2} } { \sqrt{\pi \log \frac{I^2}{\sqrt{2} \delta_{n,I} }} } \right) \right) \\ &\sim& I^{1/2} \delta_{n,I} \sqrt{ \log \frac{I^2}{\delta_{n,I}} } \\ &\leq & c \sqrt{n} \delta_{n,I}^2. \end{eqnarray*} Therefore, condition (\ref{eq:entropycondition}) is satisfied. The desired result follows from Theorem \ref{thm:likelihoodratio}. \end{proof} \subsection{An inequality for the Kullback-Leibler divergence} \label{subsec:KLbound} It is well known that Hellinger distance can be bounded by Kullback-Leibler divergence. In \cite{wong1995}, the authors showed that the other direction also holds under mild conditions. This type of result becomes quite useful in this paper because it would allow us to study the convergence rate under the weaker Hellinger distance. We first cite the result from their paper. It enables us to obtain a more explicite bound for density functions in $\mathcal{F}_0$ in the later part of this section. \begin{lemma} [Wong and Shen (1995) Theorem 5] \label{lemma:hellingertokl} Let $f$, $f_0$ be two densities, $\rho^2(f_0, f) \leq \epsilon^2$. Suppose that $M_{\lambda} ^2 = \int_{\{f_0/f \geq e^{1/\lambda} \} } f_0 (f_0 /f)^{\lambda} < \infty$ for some $\lambda \in (0, 1]$. Then for all $\epsilon^2 \leq \frac{1}{2} ( 1-e^{-1})^2$, we have \begin{eqnarray*} \int f_0 \log (\frac{f_0}{f}) \leq \big[ 6 + \frac{2 \log 2}{(1-e^{-1})^2} + \frac{8}{\lambda} \max \big( 1 , \log (\frac{M_{\lambda}}{\epsilon}) \big) \big] \epsilon^2,\\ \int f_0 \big( \log (\frac{f_0}{f}) \big)^2 \leq 5\epsilon^2 \big[ \frac{1}{\lambda} \max \big( 1, \log (\frac{M_{\lambda}}{\epsilon} ) \big) \big]^2. \end{eqnarray*} \end{lemma} When $f_0$ is the true density function and $f$ is an approximation to $f_0$ in $\Theta_I$, we can obtain a more explicit bound of the Kullback-Leibler divergence. The result is summarized in the lemma below. \begin{lemma} \label{lemma:KLbound} $f_0$ is a density function on $\Omega$. If $f_0 \in \mathcal{F}_0$, then we can find $g \in \Theta_I$, such that \begin{eqnarray*} \int f_0 \log (f_0/g) \leq 128 A^2 r I^{-2r} \log I,\\ \int f_0 (\log ( f_0/g) )^2 \leq 320 A^2 r^2 I^{-2r} (\log I)^2. \end{eqnarray*} \end{lemma} \begin{proof} Assume that $f= \sum_{i=1}^I \beta_i \mathbbm{1}_{\Omega_i}$, where $\{ \Omega_i \}_{i=1}^I$ is a binary partition of $\Omega$. From the property of $L_2$-projection, we have \begin{eqnarray*} \rho^2(f_0, f) &=& \sum_{i=1}^I \int_{\Omega_i} ( \sqrt{f_0(x)} - \sqrt{\beta_i})^2 \mu(dx) \\ &\geq & \sum_{i=1}^I \int_{\Omega_i} \left( \sqrt{f_0} - \frac{\int_{\Omega_i} \sqrt{f_0} } { \mu (\Omega_i)} \right)^2. \end{eqnarray*} Let $h=\sum_{i=1}^I \left( \frac{\int_{\Omega_i} \sqrt{f_0}}{\mu(\Omega_i)} \right)^2 \mathbbm{1}_{\Omega_i}$, then \begin{equation*} \int_{\Omega} h = \sum_{i=1}^I \frac{ (\int_{\Omega_i} \sqrt{f_0} )^2 } {\mu(\Omega_i)} \leq \sum_{i=1}^I \frac{( \int_{\Omega_i} f_0 )\mu(\Omega_i)}{\mu(\Omega_i)} =1. \end{equation*} Let $g= \frac{h} {\int_{\Omega} h}$, then $g$ is a density function, and \begin{eqnarray*} \rho^2 (f_0, g) &=& \| \sqrt{f_0} -\sqrt{h} + \sqrt{h} -\frac{ \sqrt{h}}{ \| \sqrt{h} \|_2 } \|_2^2\\ &\leq& 2 \| \sqrt{f_0} -\sqrt{h} \|_2^2 + 2 (1-\frac{1}{\| \sqrt{h} \|_2} )^2 \| \sqrt{h} \|_2^2 \\ &\leq & 2\rho^2(f_0, f) + 2(1- \| \sqrt{h} \|_2^2) \\ &=& 2\rho^2(f_0,f) + 2 \| \sqrt{f_0} -\sqrt{h} \|_2^2 \\ &\leq& 4 \rho^2(f_0,f). \end{eqnarray*} For density functions $f_0$ and $g$, \begin{eqnarray*} M_{1/4}^2 &=& \sum_{i=1}^I \int_{\Omega_i \cap \{ f_0 / g > e^4 \} } f_0 \left( \frac{f_0}{ \left( \frac{\int_{\Omega_i} \sqrt{f_0} } { \| \sqrt{h} \|_2 \mu(\Omega_i)} \right) ^2} \right)^{1/4} \\ &\leq& \sum_{i=1}^I \frac{ \int_{\Omega_i} f_0^{1+1/4}} { \left( \frac{\int_{\Omega_i} \sqrt{f_0} }{\mu(\Omega_i)} \right)^{1/2}} \\ &\leq& \sum_{i=1}^I \frac{ (\int_{\Omega_i} f_0 ^{1/2})^{1/2} (\int_{\Omega_i} f_0 ^2 )^{1/2} }{ \left( \frac{\int_{\Omega_i} \sqrt{f_0} }{\mu(\Omega_i)} \right)^{1/2}}= \sum_{i=1}^I (\int_{\Omega_i} f_0^2 ) ^{1/2} (\mu(\Omega_i))^{1/2} \\ &\leq& \left(\int_{\Omega} f_0 ^2 \right) ^{1/2}. \end{eqnarray*} Therefore, if we set $\lambda=1/4$, when $I$ is large enough \begin{eqnarray*} \int f_0 \log (f_0/g) &\leq& [6 + \frac{2 \log2}{(1-e^{-1})^2} +32 \max (1, \log \frac{ (\int f_0^2)^{1/4}}{ 2 A I^{-r} }) ] \cdot 4 A^2 I^{-2r} \\ &\leq& 128 A^2 r I^{-2r} \log I. \end{eqnarray*} Similarly, \begin{equation*} \int f_0 (\log ( f_0/g) )^2 \leq 320 A^2 r^2 I^{-2r} (\log I)^2. \end{equation*} \end{proof} \subsection{Convergence rate} \label{subsec:rate} In this section, we apply the above large-deviation type inequality and the bound of the Kullback-Leibler divergence to derive convergence rate of the sieve MLE for $f_0 \in \mathcal{F}_0$. \begin{proof}[proof of Theorem \ref{th:convergencerate}] For any $g \in \Theta_I$ and $D>1$, we have \begin{equation*} \mathbb{P} \left( \rho(f_0, \hat{f}_{n,I} ) \geq D \delta_{n,I} \right) \leq \mathbb{P}_{f_0} \left( \sup_{ \{ \rho(f_0, f) \geq D \delta_{n,I}, f \in \Theta_I \}} \prod_{j=1}^n f(Y_j)/g(Y_j) \geq 1 \right), \end{equation*} where $\mathbb{P}_{f_0}$ is understood to be the outer probability measure under $f_0$. Let $C= \{ f \in \Theta_I: \rho(f_0, f) \geq D \delta_{n,I} \}$. Then for $g \in \Theta_I$, \begin{equation*} \mathbb{P}_{f_0} \left( \sup_{f \in C} \prod_{j=1}^n \frac{f(Y_j)}{g(Y_j)} \geq 1 \right) \leq P_1 +P_2, \end{equation*} where \begin{equation*} P_1 = \mathbb{P}_{f_0} \left( \sup_{f \in C} \prod_{j=1}^n \frac{f(Y_j)}{f_0(Y_j)} \geq \exp\left(-c_1 n (D \delta_{n,I})^2 \right) \right), \end{equation*} \begin{equation*} P_2 = \mathbb{P} \left( \prod_{j=1}^n \frac{f_0(Y_j)}{g(Y_j)} \geq \exp \left( c_1 n ( D \delta_{n,I})^2 \right) \right). \end{equation*} In regards of $P_1$, Corollary \ref{cor:largedeviation} still applies here if we replace $\delta_{n,I}$ by $D \delta_{n,I}$. Thus, $P_1 \leq 4 \exp( - c_2 n D^2 \delta^2_{n,I})$. To bound $P_2$, we write \begin{eqnarray*} P_2 &=& \mathbb{P} \left( \sum_{j=1}^n \log \left( \frac{f_0}{g} \right) (Y_j) \geq c_1 n (D \delta_{n,I})^2 \right) \\ &=& \mathbb{P} \left( \sum_{j=1}^n \left[ \log \left(\frac{f_0}{g} \right) (Y_j) - \mathbb{E} \log \left(\frac{f_0}{g} \right) \right] \geq c_1 n (D \delta_{n,I} )^2 - n \int f_0 \log \left( \frac{f_0}{g} \right) \right). \end{eqnarray*} If $\int f_0 \log (f_0/g) < c_1 D^2 \delta^2_{n,I}$, then \begin{equation*} P_2 \leq \frac { n \int f_0 \log \left(\frac{f_0}{g} \right)^2 }{ n^2 \left( c_1 D^2 \delta^2_{n,I} - \int f_0 \log \left(\frac{f_0}{g} \right) \right)^2 }. \end{equation*} Based on our assumption, there exists $f_I \in \Theta_I$, such that $\rho(f_0, f_I) \leq A I^{-r}$. Then from Lemma \ref{lemma:KLbound}, we know that \begin{eqnarray*} \int f_0 \log (f_0/f_I) \leq 128 A^2 r I^{-2r} \log I,\\ \int f_0 (\log ( f_0/f_I) )^2 \leq 320 A^2 r^2 I^{-2r} (\log I)^2. \end{eqnarray*} Therefore, \begin{equation*} \inf_{g \in \Theta_I} P_2 \leq \frac{320 A^2 r^2 I^{-2r} (\log I)^2} { n \left( c_1 D^2 \delta^2_{n,I} - 128 A^2 r I^{-2r} \log I \right)^2}. \end{equation*} If we take $I =\left( (2^8 A^2 r /c_1) \frac{n}{\log n} \right) ^{\frac{1}{2r+1} }$, then the condition $\int f_0 \log (f_0/f_I) < c_1 D^2 \delta^2_{n,I}$ is satisfied. If $n$ and $I$ are matched in this way, the order of $\delta_{n,I}$ determines the final convergence rate, which is $n^{-\frac{r}{2r+1}} (\log n) ^{(\frac{1}{2}+\frac{r}{2r+1})}$. This finishes the proof. \end{proof} \section{Application to spatial adaptation} \label{sec:spatialadaptation} In this section, we assume that the density concentrates spatially. Mathematically, this implies the density function satisfies a type of \emph{spatial sparsity}. In the past two decades, sparsity has become one of the most discussed types of structure under which we are able to overcome the curse of dimensionality. A remarkable example is that it allows us to solve high-dimensional linear models, especially when the system is underdetermined. It would be interesting to study how we could benefit from the sparse structure when performing density estimation. This section is devoted to this purpose. Under current settings, it is natural to characterize the spatial sparsity by controlling the decay rate of the ordered Haar wavelet coefficients. After introducing the high-dimensional Haar basis in Section \ref{subsec:Haarbasis1}, we provide a rigorous characterization of the sparsity condition in Section \ref{subsec:sparsity}, and illustrate this characterization by several examples. In Section \ref{subsec:spatialadaptationrate}, we demonstrate that the sparse structure allows fast convergence by calculating the explicit convergence rate. \subsection{High-dimensional Haar basis} \label{subsec:Haarbasis1} Haar basis is the simplest but widely used wavelet basis. In one dimension, the Haar wavelet's mother wavelet function is \begin{equation*} \psi(y) = \begin{cases} 1 &\mbox{if } 0\leq y < 1/2, \\ -1 & \mbox{if } 1/2 \leq y < 1, \\ 0 & \mbox{otherwise.} \end{cases} \end{equation*} And its scaling function is \begin{equation*} \phi(y) = \begin{cases} 1 &\mbox{if } 0 \leq y < 1, \\ 0 &\mbox{otherwise.} \end{cases} \end{equation*} Here, we take the two-dimensional case to illustrate how the system is built. This construction can be extended to high dimensional cases as well. The two-dimensional scaling function is defined to be \begin{equation*} \phi \phi(y^1,y^2) :=\phi(y^1) \phi(y^2), \end{equation*} and three wavelet functions are \begin{equation*} \phi \psi(y^1,y^2) := \phi(y^1) \psi(y^2), \end{equation*} \begin{equation*} \psi \phi(y^1, y^2) := \psi(y^1) \phi(y^2), \end{equation*} \begin{equation*} \psi \psi(y^1, y^2) := \psi(y^1) \psi(y^2). \end{equation*} If we use a superscript $l$ to index the scaling level of the wavelet function and subscripts $i$ and $j$ ($i$ and $j$ can be 0 or 1) to denote the horizontal and vertical translations respectively, then the scales and translates of the three wavelet functions $\phi \psi$, $\psi \phi$ and $\psi \psi$ are defined to be \begin{eqnarray*} &\phi \psi_{ij}^l (y^1, y^2) := (\sqrt{2} )^{2 \cdot l} \phi \psi(2^l y^1-i, 2^l y^2 -j),\\ &\psi \phi_{ij}^l (y^1, y^2) := (\sqrt{2} )^{2 \cdot l} \psi \phi(2^l y^1 -i, 2^l y^2 -j),\\ & \psi \psi_{ij}^l (y^1, y^2) := (\sqrt{2} )^{2 \cdot l} \psi \psi (2^l y^1 -i, 2^l y^2 -j). \end{eqnarray*} These functions together with the single scaling function $\phi \phi$ define the two-dimensional Haar wavelet basis \textbf{$\Psi$}. \subsection{Spatial sparsity} \label{subsec:sparsity} Let $f$ be a $p$ dimensional density function and \textbf{$\Psi$} the $p$-dimensional Haar basis constructed as above. We will work with $g=\sqrt{f}$ first. Note that $g \in L^2([0,1]^p)$. Thus we can expand $g$ with respect to \textbf{$\Psi$} as $g = \sum_{\psi \in \Psi} <g,\psi> \psi$ (here $\psi$ is a basis function instead of the Haar wavelet's mother wavelet function defined above). We rearrange this summation by the size of wavelet coefficients. In other words, we order the coefficients as the following \begin{equation*} |<g, \psi_{(1)} >| \geq | <g, \psi_{(2)}>| \geq \cdots \geq |<g, \psi_{(k)}>| \geq \cdots, \end{equation*} then the sparsity condition imposed on the density functions is that the decay of the wavelet coefficients follows a power law, \begin{equation} \label{eq:powlawdecay} |<g,\psi_{(k)}> | \leq C k^{-\beta} \mbox{ for all $k \in \mathbb{N}$ and $\beta > 1/2$, } \end{equation} where $C$ is a constant. This condition has been widely used to characterize the sparsity of signals and images (\cite{abramovich2006} and \cite{Candes}). In particular, in \cite{DeVore}, it was shown that for two-dimensional cases, when $\beta > 1/2$, this condition reasonably captures the sparsity of real world images. Next we use several examples to illustrate how this condition implies the spatial sparsity of the density function. \begin{example} Assume that the we are studying a density function in three dimensions. All the mass concentrates in a dyadic cube. Without loss of generality, we assume that $f_0 = 64 \mathbbm{1}_{ \{0\leq y^1, y^2, y^3 < 1/4 \}}$, so that all the mass is concentrated in a small cubical subregion. In the three-dimensional case, the single scaling function is $\phi \phi \phi$, and the seven wavelet functions are defined as \begin{eqnarray*} &\chi_1 = \psi \phi \phi,\ \chi_2 =\phi \psi \phi,\ \chi_3 = \phi \phi \psi, \\ & \chi_4 = \psi \psi \phi,\ \chi_5 = \psi \phi \psi,\ \chi_6 = \phi \psi \psi, \\ &\chi_7=\psi \psi \psi. \end{eqnarray*} If we still use the superscript to denote the scaling level and the subscripts to denote the spatial translations, then the expansion of $f_0$ with respect to the Haar basis is \begin{equation*} f_0 = \phi \phi \phi + \sum_{k=1}^7 \chi_{k,000}^{(0)} + 2 \sqrt{2} \sum_{k=1}^7 \chi_{k, 000}^{(1)}. \end{equation*} The coefficients display a decaying trend, although the number of them is finite. More generally, any density function whose expansion only has finite terms will satisfy the condition (\ref{eq:powlawdecay}). \end{example} \begin{example} Assume that the two-dimensional true density function is \begin{equation*} \begin{pmatrix} Y_1 \\ Y_2 \\ \end{pmatrix} \sim \frac{2}{5} \mathcal{N} \left( \begin{pmatrix} 0.25 \\ 0.25 \\ \end{pmatrix} , 0.05^2 I_{2 \times 2} \right) + \frac{3}{5} \mathcal{N} \left( \begin{pmatrix} 0.75 \\ 0.75 \\ \end{pmatrix} , 0.05^2 I_{2 \times 2} \right). \end{equation*} We perform the Haar transform. The heatmap of the density function is displayed in Figure~\ref{fig:heatmap1} and the plot of the Haar coefficients is shown in Figure~\ref{fig:Haar1}. The left panel in Figure~\ref{fig:Haar1} is the plot of all the coefficients to level ten from low resolution to high resolution. The middle one is the sorted coefficients according to their absolute value. And the right one is the same as the middle plot but with the abcissa in log-scale. From this we can clearly see that the power-law decay is satisfied, and an empirical estimation of the corresponding $\beta$ can be obtained in this case. \begin{figure} \centering \includegraphics[scale=0.4]{heatmap1.jpg} \caption{Heatmap of the density} \label{fig:heatmap1} \end{figure} \begin{figure} \centering \includegraphics[scale=0.4]{plot1.jpg} \caption{Plots of the 2-dimensional Haar coefficients. The left panel is the plot of all the coefficients from low resolution to high resolution. The middle one is the plot of the sorted coefficients. And the right one is the same as the middle plot but with the abcissa in log scale.} \label{fig:Haar1} \end{figure} \end{example} \begin{example} Let the three-dimensional density function be \begin{equation*} \begin{pmatrix} Y_1 \\ Y_2 \\ Y_3 \\ \end{pmatrix} \sim \frac{2}{5} \mathcal{N} \left( \begin{pmatrix} 0.25 \\ 0.25 \\ 0.25 \\ \end{pmatrix}, \begin{pmatrix} 0.05^2 & 0.03^2 & 0 \\ 0.03^2 & 0.05^2 & 0 \\ 0 & 0 & 0.05^2 \\ \end{pmatrix} \right) + \frac{3}{5} \mathcal{N} \left( \begin{pmatrix} 0.75 \\ 0.75 \\ 0.75 \\ \end{pmatrix}, 0.05^2 I_{3 \times 3} \right). \end{equation*} In this example, we impose some correlation structure in one component. Haar transform is performed and the behavior of the Haar coefficients is summarized in Figure~\ref{fig:Haar2}. The arrangement of the plots is the same as that in the previous example. For this three-dimensional example, the power-law decay is still satisfied. \begin{figure} \centering \includegraphics[scale=0.4]{3dnormal_new.jpg} \caption{Plots of the 3-dimensional Haar coefficients. The order of the plots is the same as that in Figure~\ref{fig:Haar1}.} \label{fig:Haar2} \end{figure} \end{example} \subsection{Convergence rate} \label{subsec:spatialadaptationrate} Assume that $f_0$ is the $p$-dimensional density function satisfy the spatial sparsity condition (\ref{eq:powlawdecay}). Now we calculate the convergence rate of the sieve MLE. \begin{lemma} \label{lemma:sparsedensity} Suppose $f_0$ is a $p$-dimensional density function. $g_0 =\sqrt{f_0}$ satisfies the condition (\ref{eq:powlawdecay}). Then there exists a sequence of $f_I \in \Theta_I$, such that $\rho(f_0, f_I) \lesssim I^{-(\beta -1/2)}$, or equivalently, $\rho(f_0, f_I) \leq c I^{-(\beta- 1/2)}$, where $c$ may depend on $\beta$ and $p$ but not $I$. \end{lemma} \begin{proof} Let $g_K = \sum_{k=1}^{K} <g_0,\psi_{(k)}> \psi_{(k)}$. From condition (\ref{eq:powlawdecay}) we have \begin{eqnarray} \label{eq:approxsparse} \nonumber \rho^2(f_0, g_K^2) &=& \| g_0-g_K\|_2^2 = \| \sum_{k=K+1}^{+ \infty} <g_0, \psi_{(k)}> \psi_{(k)} \|_2^2 \\ \nonumber &=& \sum_{k=K+1}^{+ \infty} <g_0,\psi_{(k)}>^2 \\ &\leq& C^2 \sum_{k=K+1}^{+ \infty} k^{-2\beta} \leq \frac{C^2}{2\beta -1} K^{-(2\beta-1)}. \end{eqnarray} Then we normalize $g_K$ to $\tilde{g}_K$ and obtain \begin{eqnarray} \label{eq:normalizesparse} \nonumber \rho^2(f_0, \tilde{g}_K^2) &=& \| g_0 - \tilde{g}_K \|_2^2 \\ \nonumber &=& \|g_0 -g_K \|_2^2 + (1-\frac{1}{\|g_K \|_2})^2 \|g_K \|_2^2 \\ \nonumber &\leq& \|g_0 -g_K \|_2^2 +1- \|g_K \|_2^2 \\ \nonumber &=&2\|g_0 -g_K \|_2^2 \\ & \leq& \frac{2C^2}{2\beta -1} K^{-(2\beta-1)}. \end{eqnarray} Note that given a supporting rectangle, the positive and negative parts of the Haar basis function defined on it can further divide the original rectangle into smaller subregions, and the total number of such subregions is upper bounded by $2^p$. Therefore, $2^p K$ is the largest possible sized binary partition on which the density function $\tilde{g}_K$ is piecewise constant. Replacing $K$ in (\ref{eq:normalizesparse}) by $I/2^p$, we get the desired result of approximation rate. \end{proof} The next theorem calculates the convergence rate in this case. \begin{theorem} (Application to spatial adaptation) Assume $f_0$ is the same as defined in Lemma (\ref{lemma:sparsedensity}). If we apply the maximum likelihood density estimator based on adaptive partitioning here to estimate the true density function, the convergence rate is $ n^{- \frac{\beta- 1/2}{2\beta}} (\log n)^{\frac{1}{2}+\frac{\beta- 1/2}{2\beta} }$. \end{theorem} \begin{proof} This follows Theorem \ref{th:convergencerate} and Lemma \ref{lemma:sparsedensity} directly. \end{proof} From the theorem we see that the convergence rate only depends on how fast the coefficients decay as opposed to the dimension of the sample space. Thus for large $\beta$, the density estimate is able to take advantage of spatial sparsities to achieve fast convergence rate even in high dimensions. \section{Application to variable selection} \label{sec:variableselection} For high dimensional data analysis, selecting significant variables greatly contributes to simplifying the model, improving model interpretability, and reducing overfitting. In the context of density estimation, the variable selection problem is formulated as follows. Assume $f_0$ is a $p$-dimensional density function and it only depends on $\tilde{p}$ variables, but we do not know in advance which $\tilde{p}$ variables. We apply the multivariate density estimation method here to the $p$-dimensional density. The essential part of our method is to learn a partition of the support of the true density function, and then to estimate the density on each subregion separately. Because the true density lies in a $\tilde{p}$-dimensional subspace, we may surmise that the corresponding convergence rate only depends on the effective dimension. The goal of this section is to carry out exact calculations to reveal that this is indeed the case. Here, we consider a class of density functions satisfying certain type of continuity. We first provide a mathematical description of this density class in Section \ref{subsec:functionclass}. This description still depends on the Haar transform of the true density function. However, an alternative construction of the high-dimensional Haar basis via tensor product is introduced in Section \ref{subsec:Haarbasis2} because of some technical issue. We provide the result on the explicite convergence rate in Section \ref{subsec:variableselectionrate}. \subsection{Tensor Haar basis} \label{subsec:Haarbasis2} In one dimension, the Haar wavelet's mother wavelet function $\psi$ and its scaling function $\phi$ are the same as those defined in Section \ref{subsec:Haarbasis1}. For any $l \in \mathbb{N}$ and $0\leq k < 2^l$, the Haar function is \begin{equation*} \psi^l_k (y) = 2^{l/2} \psi(2^l y -k). \end{equation*} Then Haar basis \textbf{$\Psi$} is the collection of all Haar functions together with the scale function. Namely, \begin{equation*} \textbf{$\Psi$} = \{ \phi \} \cup \{ \psi^l_k, l \in \mathbb{N}, 0\leq k < 2^l \}. \end{equation*} It is an orthonormal basis for Hilbert space $L^2([0,1])$. Turning to high-dimensional settings, we can obtain an orthonormal basis for $L^2 ([0,1]^p)$ by using the fact that the Hilbert space $L^2 ([0,1]^p)$ is isomorphic to the tensor product of $p$ one-dimensional spaces. In detail, if $\mathcal{X}_1, \cdots, \mathcal{X}_p$ are $p$ copies of $L^2 ([0,1])$ and \textbf{$\Psi_1$}, $\cdots$, \textbf{$\Psi_p$} are Haar bases of these spaces respectively, then $L^2 ([0,1]^p)$ is isomorphic to $\bigotimes_{i=1}^p \mathcal{X}_i$. Define tensor Haar basis \textbf{$\Psi$} by \begin{equation*} \textbf{$\Psi$} = \{\psi: \psi= \prod_{i=1}^p \psi_i, \psi_i \in \textbf{$\Psi_i$} \}. \end{equation*} From the property of tensor product of Hilbert spaces, we know that \textbf{$\Psi$} is an orthonormal basis for $L^2 ([0,1]^p)$. \subsection{Mixed-H\"older continuity} \label{subsec:functionclass} In this section, we assume the density function satisfies the \emph{Mixed-H\"older} condition. A similar condition first appeared in \cite{stromberg1998}. The author showed that tensor Haar basis with large support is efficient in representing certain type of functions on $[0,1]^p$. More precisely, if the function satisfies the \emph{bounded mixed variation} condition, then it can be approximated to error $\epsilon>0$ using no more than $ \frac{1}{\epsilon} (\log (\frac{1}{\epsilon}) )^{p-1}$ terms, and the volume of the supporting rectangle of the each wavelet basis function involved in the approximation is greater than $\epsilon$. However, the result is restrictive in application since the mixed derivative is not rotationally invariant. In \cite{gavish2012}, the authors extended the previous approximation scheme to matrix. A significant improvement is that, in their paper, the bounded mixed variation condition is replaced by the Mixed-H\"older condition, which is more natural and accessible. For two-dimensional discrete analyses of matrices, they show that the approximation is still efficient for this class of matrices, and the Mixed-H\"older condition is connected to the decay rate of wavelet coefficients. The idea of controlling the decay rate of wavelet coefficients will be further developed here. It leads to the characterization of the density class under consideration. For any $f \in \mathcal{F}$, $\sqrt{f} \in L^2 ([0,1]^p)$. Therefore, we can expand $\sqrt{f}$ with respect to the tensor Haar basis. Let $g= \sqrt{f}$. Then \begin{equation*} g = \sum_{\psi \in \textbf{$\Psi$}} <g, \psi> \psi,\ \mbox{where} <g,\psi>=\int_{\Omega} g(y) \psi(y) dy. \end{equation*} For each tensor Haar function $\psi$, let $R(\psi)$ denote its supporting rectangle. The density class under consideration is defined to be \begin{equation*} \mathcal{F}_H = \{ f \in \mathcal{F}: |<\sqrt{f}, \psi>| \leq C |R(\psi)|^{\alpha+1/2} \mbox{ for all } \psi \in \textbf{$\Psi$} \}, \end{equation*} where $C$ is a constant which may depend on $f$, $|R(\psi)|$ denotes the volume of the rectangle and $\alpha$ is a positive constant. Next, we provide several examples to illustrate how this condition relates to the Mixed-H\"older continuity. \begin{example} A real-valued function $f$ on $\mathbb{R}$ is \emph{H\"older continuous}, if there exist nonnegative constants $C$ and $\alpha \in (0, 1]$, such that $|f(x) - f(y)| \leq C |x-y|^{\alpha}$, for all $x, y \in \mathbb{R}$. If square root of the true density function $f_0$ is H\"older continuous for some constants $C, \alpha$, then for any Haar basis function $\psi$, $|<\sqrt{f_0}, \psi>| \leq C |R(\psi)|^{\alpha+1/2}$. \begin{proof} From H\"older continuity, we know that for any $x, y \in [0,1]$, $|\sqrt{f_0(x)}- \sqrt{f_0(y)}| \leq C |x-y|^{\alpha}$. For any point $x_0 \in R$, we have \begin{eqnarray*} |<\sqrt{f_0}, \psi>| ^2 &=& \left( \int_{R} \sqrt{f_0(x)} \psi(x) dx \right)^2 \\ &=& \left( \int_{R} \left(\sqrt{f_0(x)} -\sqrt{f_0(x')} \right) \psi(x) dx \right)^2 \\ &\leq& \int_{R} \left(\sqrt{f_0(x)} -\sqrt{f_0(x')} \right)^2 dx \cdot \int_{R} \psi(x)^2 dx \\ &\leq& C^2 \int_{R} |x-x_0|^{2\alpha} dx \\ &=& C^2 |R|^{2\alpha+1}. \end{eqnarray*} The desired results follows. \end{proof} \end{example} \begin{example} A real-valued function $f$ on $\mathbb{R}^2$ is called Mixed-H\"older continuous for some nonnegative constant $C$ and $\alpha \in (0, 1]$, if for any $(x_1, y_1), (x_1, y_2) \in \mathbb{R}^2$, \begin{equation*} |f(x_2, y_2) - f(x_2, y_1) - f(x_1, y_2) + f(x_1, y_1) | \leq C |x_1 -x_2|^{\alpha} |y_1 -y_2 |^{\alpha}. \end{equation*} If square root of the two-dimensional true density function $f_0$ is Mixed-H\"older continuous for some constants $C, \alpha$, then $f_0 \in \mathcal{F}_H$. The proof is similar to the one-dimensional case. \end{example} More generally, this type of continuity condition can be extended to high-dimensional cases \cite{stromberg1998}. The corresponding density functions also belong to the space $\mathcal{F}_H$. Therefore, in this section, we use the more general condition $|<\sqrt{f}, \psi>| \leq C |R(\psi)|^{\alpha+1/2}$ for all $\psi$ to characterize the density class. \subsection{Convergence rate} \label{subsec:variableselectionrate} First, we provide a result on the rate at which the approximation error decreases to zero as the complexity of the approximating spaces increases. \begin{lemma} \label{th:approximationrate} $\mathcal{F}_H$ and $\Theta_I$ are defined as above. For any $f_0 \in \mathcal{F}_H$, there exists a sequence of $f_I \in \Theta_I$, such that $\rho(f_0, f_I) \lesssim I^{-\alpha/p} (\log I)^{p/2}$, where $\alpha$ is as defined in the Section \ref{subsec:functionclass}, and $p$ is the dimension of the Euclidean space. \end{lemma} \begin{proof} \label{th:approxrate} Let $g_0 = \sqrt{f_0}$. We can expand $g_0$ with respect to the tensor Haar basis. The expansion can be written as $g_0 = \sum_{\psi} <g_0, \psi> \psi$. Let $g_{\epsilon} = \sum_{\psi: |R(\psi)| > \epsilon} <g_0, \psi> \psi$. Then $g_{\epsilon}$ is an approximation to $g_0$ obtained by requiring that the volumes of the supporting rectangles of the involved wavelet basis functions are greater than $\epsilon$. We will derive an approximation rate as a function of $\epsilon$ first, and then convert the lower bound on the volume to an upper bound on the size of the partition. This yields an approximation rate as a function of the size of the partition. Note that $g_{\epsilon}$ is not a density function, but it is easier to work with. Let $\tilde{g}_{\epsilon} = g_{\epsilon}/ \|g_{\epsilon} \|_2$ be the normalization of $g_{\epsilon}$. The upper bounds for the approximation errors $\rho(f_0, g_{\epsilon} ^2)$ and $\rho(f_0, \tilde{g}_{\epsilon}^2)$ will be derived successively. Before delving into the proof, we introduce some notations first. For each supporting rectangle $|R(\psi)|$, the lengths of its edges should be powers of $1/2$. We may assume that $\psi=\prod_{i=1}^p \psi_i$, and for each $\psi_i$ the length of its supporting interval is $(1/2)^{l_i}$. Let $\mathcal{R}^{l_1, \cdots, l_p}$ denote the collection of the rectangles for which the lengths of the edges are $(1/2)^{l_1}, \cdots, (1/2)^{l_p}$. Recall that $f_0$ satisfies the condition \begin{equation} \label{eq:densityfuncond} | <\sqrt{f_0}, \psi>| \leq C |R(\psi)|^{\alpha + 1/2} \mbox{ for all } \psi. \end{equation} Then, \begin{eqnarray} \label{eq:approxI} \nonumber \rho^2(f_0, g_{\epsilon}^2 ) &=& \| g_0 - g_{\epsilon} \| ^2 = \| \sum_{\psi: |R(\psi)| < \epsilon } <g_0, \psi> \psi \|_2^2\\ \nonumber &=& \sum_{\psi: |R(\psi)| <\epsilon} <g_0, \psi>^2 \\ \nonumber &\leq& C^2 \sum_{\psi: |R(\psi)|< \epsilon } |R(\psi)|^{2\alpha+1} \\ &\leq& 2^p C^2 \sum_{l_1, \cdots, l_p} \sum_{R \in \mathcal{R}^{l_1, \cdots, l_p}, |R| <\epsilon } |R|^{2\alpha+1}. \end{eqnarray} The last inequality follows from the fact that, given a supporting rectangle, there are at most $2^p$ basis functions defined on it. Let $N= \lceil \log_{\frac{1}{2}} \epsilon \rceil$, \begin{eqnarray} \label{eq:approxII} \nonumber (\ref{eq:approxI}) &=& 2^p c^2 \sum_{l_1 + \cdots + l_p \geq N } \sum_{R \in \mathcal{R}^{l_1, \cdots, l_p}} |R|^ {2\alpha+1} \\ \nonumber &=& 2^p C^2 \sum_{l_1+ \cdots+ l_p \geq N} (\frac{1}{2})^{2\alpha(l_1 +\cdots +l_p)} \sum_{R \in \mathcal{R}^{l_1, \cdots, l_p}} |R| \\ &=& 2^p C^2 \sum_{l_1+ \cdots +l_p \geq N} (\frac{1}{2})^{2\alpha(l_1 +\cdots +l_p)}. \end{eqnarray} The last equality is obtained by plugging in $\sum_{R \in \mathcal{R}^{l_1, \cdots, l_p}} |R| =1$. Note that \begin{eqnarray} \label{eq:approxIII} & & \sum_{l_1+\cdots +l_p \geq N} ( \frac{1}{2} ) ^{2\alpha(l_1+\cdots +l_p)} \\ \nonumber &\leq& \sum_{l_1=0}^{N} \sum_{l_2=0}^{N-l_1} \cdots \sum_{l_p=N-(l_1+\cdots +l_{p-1})}^{+ \infty} (\frac{1}{2})^{2\alpha(l_1 +\cdots +l_p)} \\ \nonumber & &+ \sum_{l_1=0}^N \sum_{l_2=0}^{N-l_1} \cdots \sum_{l_{p-1}=N-(l_1 +\cdots +l_{p-2})} ^{+\infty} \sum_{l_p=0}^{+\infty} (\frac{1}{2})^{2\alpha(l_1 +\cdots +l_p)} \\ \nonumber & &+ \cdots \\ \nonumber & &+ \sum_{l_1=N}^{+\infty} \sum_{l_2=0}^{+\infty} \cdots \sum_{l_p=0}^{+\infty} (\frac{1}{2})^{2\alpha(l_1 +\cdots +l_p)} \\ \nonumber &\leq& (N+1)^{p-1} \frac{ (\frac{1}{2})^{2\alpha N}}{1-2^{-2\alpha}} + (N+1)^{p-2} \frac{ (\frac{1}{2})^{2\alpha N}}{(1-2^{-2\alpha})^2} + \cdots + \frac{ (\frac{1}{2})^{2\alpha N} } {(1-2^{-2\alpha})^p} \\ \nonumber &=& (\frac{1}{2})^{2\alpha N} \frac{(N+1)^p - (1-2^{-2\alpha})^{-p}} {(N+1)(1-2^{-2\alpha}) -1} \\ \nonumber &\leq& C' \epsilon^{2\alpha} (\log_{\frac{1}{2}} \epsilon)^p. \end{eqnarray} From this, we know that \begin{equation} \label{eq:approxIV} \rho^2(f_0, g_{\epsilon}^2 ) = \|g_0 - g_{\epsilon} \|_2^2 \leq 2^p C' C^2 \epsilon^{2\alpha} (\log_{\frac{1}{2}} \epsilon)^p. \end{equation} We normalize $g_{\epsilon}$ to $\tilde{g}_{\epsilon}$, then \begin{eqnarray*} \rho^2(f_0, \tilde{g}_{\epsilon}^2) &=& \| g_0 - \tilde{g}_{\epsilon} \|_2^2 \\ &=& \|g_0 -g_{\epsilon} \|_2^2 + (1-\frac{1}{\|g_{\epsilon} \|_2})^2 \|g_{\epsilon} \|_2^2 \\ &\leq& \|g_0 -g_{\epsilon} \|_2^2 +1- \|g_{\epsilon} \|_2^2 \\ &=&2\|g_0 -g_{\epsilon} \|_2^2 . \end{eqnarray*} The last equality is obtained by using $\|g_0 - g_{\epsilon} \|_2^2 + \|g_{\epsilon} \|_2^2 =\|g_0 \|_2^2 =1$.Therefore, \begin{eqnarray} \label{eq:approxrate} \rho^2(f_0, \tilde{g}_{\epsilon}^2 ) = \|g_0 - \tilde{g}_{\epsilon} \|_2^2 \leq 2^p C'' C^2 \epsilon^{2\alpha} (\log_{\frac{1}{2}} \epsilon)^p, \end{eqnarray} where $C''$ is a constant. Next, we will convert the lower bound on the volume of the supporting rectangles to an upper bound on the size of the partition, and derive the approximation rate in terms of the latter one. If we require the volumes of the supporting rectangles be greater than $\epsilon$, then the lengths of the edges can not be smaller than $2^{-\lfloor \log_{\frac{1}{2}} \epsilon \rfloor}$. The size of the partition supporting $\tilde{g}_{\epsilon}$ can be bounded by $2^p 2^{p \log_{\frac{1}{2}} \epsilon} = 2^p \epsilon^{-p}$. There is a coefficient $2^p$ in front. This is the case because given a supporting rectangle, the positive and negative parts of the tensor Haar basis defined on it will further divide the original rectangle into smaller subregions and the number of such subregions is at most $2^p$. Given the size of the partition $I$, we can determine $\epsilon$ by solving $2^p \epsilon^{-p} =I$ and define $\tilde{g}_{\epsilon} \in \Theta_I$ as above. Then from (\ref{eq:approxrate}) we reach a conclusion that $\tilde{g}_{\epsilon}$ is an approximation satisfying $\rho(f_0, \tilde{g}_{\epsilon}^2) \lesssim I^{-\alpha/p} (\log I)^{p/2}$. This finishes the proof. \end{proof} \begin{theorem} \label{th:variableselectionrate} Assume that $f_0 \in \mathcal{F}_H$ is a $p$-dimensional density function. It only depends on $\tilde{p}$ arguments which are not specified in advance. If we apply the multivariate density estimation method to this problem, the convergence rate is $n^{-\frac{\alpha}{2\alpha +\tilde{p}} } (\log n)^{1+\frac{\tilde{p} (\tilde{p}-1)/2}{2\alpha+\tilde{p}}}$. \end{theorem} \begin{proof} Because $f_0$ lies in a $\tilde{p}$-dimensional subspace, the decay rate of the approximation error only depends on $\tilde{p}$ instead of $p$. Then the result follows Theorem \ref{th:convergencerate} and Theorem \ref{th:approximationrate}. \end{proof} From this theorem, we learn that only the effective dimension affects the rate of our method. This implies that in extreme cases of $\tilde{p} \ll p$, our method can still achieve stable performances. The advantage of our method is demonstrated by the following facts: in the partition learning stage, it can quickly restrict our attention to those relevant variables. Ideally, it can estimate the density as a function of the effective variables alone, although they are not specified in advance.

https://arxiv.org/abs/1102.1984

Deformation Retracts of Neighborhood Complexes of Stable Kneser Graphs

In 2003, A. Bjorner and M. de Longueville proved that the neighborhood complex of the stable Kneser graph SG_{n,k} is homotopy equivalent to a k-sphere. Further, for n=2 they showed that the neighborhood complex deformation retracts to a subcomplex isomorphic to the associahedron. They went on to ask whether or not, for all n and k, the neighborhood complex of SG_{n,k} contains as a deformation retract the boundary complex of a simplicial polytope.Our purpose is to give a positive answer to this question in the case k=2. We also find in this case that, after partially subdividing the neighborhood complex, the resulting complex deformation retracts onto a subcomplex arising as a polyhedral boundary sphere that is invariant under the action induced by the automorphism group of SG_{n,2}.

\section{Introduction and Main Result} In 1978, L. Lov\'{a}sz proved in \cite{LovaszChromaticNumberHomotopy} M. Kneser's conjecture that if one partitions all the subsets of size $n$ of a $(2n + k)$-element set into $(k+1)$ classes, then one of the classes must contain two disjoint subsets. Lov\'{a}sz proved this conjecture by modeling the problem as a graph coloring problem: see Section~\ref{Background} for definitions of the following objects. For the Kneser graphs $KG_{n,k}$, Kneser's conjecture is equivalent to the statement that the chromatic number of $KG_{n,k}$ is equal to $k+2$. Lov\'{a}sz's proof methods actually provided a general lower bound on the chromatic number of any graph $G$ as a function of the topological connectivity of an associated simplicial complex called the neighborhood complex of $G$. Of particular interest in his proof was the critical role played by the Borsuk-Ulam theorem. Later that year, A. Schrijver identified in \cite{Schrijvergraphs} a vertex-critical family of subgraphs of the Kneser graphs called the stable Kneser graphs $SG_{n,k}$, or Schrijver graphs, and determined that the chromatic number of $SG_{n,k}$ is equal to $k+2$. In 2003, A. Bj\"{o}rner and M. de Longueville gave in \cite{BjornerDeLongueville} a new proof of Schrijver's result by applying Lov\'{a}sz's method to the stable Kneser graphs; in particular, they proved that the neighborhood complex of $SG_{n,k}$ is homotopy equivalent to a $k$-sphere. In the final section of their paper, Bj\"{o}rner and De Longueville showed that the neighborhood complex of $SG_{2,k}$ contains the boundary complex of a $(k+1)$-dimensional associahedron as a deformation retract. Their paper concluded with the following: \begin{question}\label{BDQ}{\rm (Bj\"{o}rner and De Longueville, \cite{BjornerDeLongueville}) } For all $n$ and $k$, does the neighborhood complex of $SG_{n,k}$ contain as a deformation retract the boundary complex of a simplicial polytope? \end{question} Our main contribution in this paper is to provide a positive answer to Question~\ref{BDQ} in the case $k=2$. Specifically, we show the following: \begin{theorem}\label{mainthm} For every $n\geq 1$, the neighborhood complex of $SG_{n,2}$ simplicially collapses onto a subcomplex arising as the boundary of a three-dimensional simplicial polytope. \end{theorem} The subcomplex for $\mathcal{N}(SG_{3,2})$ is shown in Figure~\ref{p}. \begin{figure}[ht] \label{p} \begin{center} \includegraphics[width=5in]{Nsquare5.eps} \end{center} \caption{} \end{figure} In \cite{BraunAutStableKneser}, the first author proved that for $k\geq 1$ and $n\geq 1$ the automorphism group of $SG_{n,k}$ is isomorphic to the dihedral group of order $2(2n+k)$. It is natural to ask if there exist spherical subcomplexes of the neighborhood complex of $SG_{n,k}$ that are invariant under the induced action of this group. While our spheres arising in Theorem~\ref{mainthm} are not invariant, we are able to show the following: \begin{theorem}\label{dihedralthm} For every $n\geq 1$, there exists a partial subdivision of the neighborhood complex of $SG_{n,2}$ that simplicially collapses onto a subcomplex invariant under the action induced by the automorphism group of $SG_{n,2}$ arising as the boundary of a three-dimensional simplicial polytope. \end{theorem} In addition to its aesthetic attraction, there are two primary reasons we are interested in Question~\ref{BDQ}. First, any polytopes found in response to Question~\ref{BDQ} will be common generalizations of simplices, associahedra, and $1$-spheres given as odd cycles, due to the following observations: for $SG_{1,k}=K_{k+2}$, the neighborhood complex is a simplex boundary; for $SG_{n,1}$, the neighborhood complex is an odd cycle, hence a one-dimensional sphere; for $SG_{2,k}$, the neighborhood complex deformation retracts to an associahedron. A family of polytopes generalizing these objects would be interesting to identify. Second, a broad extension of the neighborhood complex construction is the graph homomorphism complex $HOM(H,G)$ studied in \cite{BabsonKozlovComplexes,BabsonKozlovLovaszConjecture,DochtermannEngstromCellular,DochtermannSchultz,SchultzStableKneserNotTest,SchultzSpacesOfCircuits}. The complex $HOM(K_2,G)$ is known to be homotopy equivalent to the neighborhood complex of $G$. The homomorphism complex construction leads to interesting phenomena, yet at present the lower bounds on graph chromatic numbers obtained by these are no better than those provided by the neighborhood complex. We believe it is appropriate to continue to focus attention on the neighborhood complex construction along with the $HOM$ construction. The rest of this paper is as follows. In Section~\ref{Background}, we introduce the necessary background and notation regarding neighborhood complexes and stable Kneser graphs as well as discrete Morse theory, the primary tool in our proofs. In Sections~\ref{Matching} and~\ref{Sphere1}, we provide a proof of Theorem~\ref{mainthm}. In Section~\ref{Sphere2} we provide a proof of Theorem~\ref{dihedralthm}. \section{Definitions and Background}\label{Background} Let $[n]:=\{1,2,\ldots,n\}$. The material in this section is adapted from the texts \cite{JonssonBook} and \cite{KozlovBook}, where more details may be found. \subsection{Neighborhood Complexes and Stable Kneser Graphs} The following definition is due to Lov\'{a}sz. \begin{definition} Given a graph $G=(V,E)$, the \emph{neighborhood complex of $G$} is the simplicial complex $\mathcal{N}(G)$ with vertex set $V$ and faces given by subsets of $V$ sharing a common neighbor in $G$, i.e. $\mathcal{N}(G) := \{ F\subset V: \exists v\in V \textrm{ s.t. } \forall u\in F, \{u,v\}\in E\}$. \end{definition} The graphs we are interested in are the following. \begin{definition}\label{KG} For $n\ge 1$ and $k\ge 0$ the \emph{Kneser graph}, denoted $KG_{n, k}$, is the graph whose vertices are the subsets of $[2n +k]$ of size $n$. We connect two such vertices with an edge when they are disjoint as sets. We call an $n$-set $\alpha$ of $[2n+k]$ \emph{stable} if $\alpha$ does not contain the subset $\{1, 2n+k\}$ or any of the subsets $\{i, i+1\}$ for $i=1,\ldots , 2n+k-1$. The \emph{stable Kneser graph}, denoted $SG_{n, k}$, is the induced subgraph of $KG_{n, k}$ whose vertices are the stable subsets of $[2n+k]$. \end{definition} Our focus in this paper is on the case $k=2$; we will assume through the rest of the paper that this holds. In order to handle different stable $n$-sets, we distinguish between them as follows, with all addition on elements being modulo $2n + 2$. \begin{definition} We call a stable $n$-set $\alpha$ \emph{tight} if $\alpha = \{i, i + 2, i + 4, \ldots, i + 2(n-1) \}$ for some $i \in [2n + 2]$. Otherwise, we call $\alpha$ a \emph{loose} stable $n$-set. For $\alpha=\{\alpha_1, \ldots, \alpha_n\}$ and $\beta$ stable $n$-sets, we call $\alpha$ and $\beta$ \emph{immediate neighbors} if $\alpha \oplus 1 = \beta$ or $\alpha \ominus 1 = \beta$, where $\alpha \oplus j:= \{\alpha_1 + j, \ldots, \alpha_n + j\}$ and $\alpha \ominus j$ is defined similarly. We call $\alpha$ and $\beta$ \emph{outer neighbors} if there is an ordering of the elements of $\alpha$ such that $\beta =(\alpha_1 + 1, \alpha_2 + 1, \ldots, \alpha_{i-1} + 1 , \alpha_i + 2, \alpha_{i+1} + 1, \ldots, \alpha_n + 1)$ and $\alpha$ and $\beta$ are neighbors in $SG_{n, 2}$. \end{definition} The following remarks provide some insight into the structure of these graphs; further discussion, including proofs of these remarks, can be found in \cite{BraunIndComplexKneser,BraunAutStableKneser}. \begin{itemize} \item A cycle is formed in $SG_{n,2}$ with vertices a stable $n$-set $\alpha$ and the stable $n$-sets $\alpha\oplus1$, $\alpha \oplus 2$, etc, with the edges $\{\alpha\oplus i,\alpha\oplus (i+1)\}$. Thus, $\alpha$ and $\beta$ are immediate neighbors if they are neighbors on such a cycle in $SG_{n, 2}$. \item Stable $n$-sets $\alpha$ and $\beta$ are outer neighbors in $SG_{n, 2}$ if they are neighbors and lie on two different cycles created via the immediate neighbor process. \item A loose stable $n$-set has degree $4$ in $SG_{n,2}$. Two of its neighbors are immediate neighbors while the other two are outer neighbors. \item The tight stable $n$-sets correspond to vertices that together induce a complete bipartite subgraph in $SG_{n,2}$. \end{itemize} \subsection{Discrete Morse Theory} We now introduce some tools from discrete Morse theory. Discrete Morse theory was first developed by R. Forman in \cite{FormanMorseTheory} and has since become a powerful tool for topological combinatorialists. The main idea of the theory is to systematically pair off faces within a simplicial complex in such a way that we obtain a collapsing order for the complex, yielding a homotopy equivalent cell complex. \begin{definition} A \textit{partial matching} in a poset $P$ is a partial matching in the underlying graph of the Hasse diagram of $P$, i.e., it is a subset $M\subseteq P \times P$ such that \begin{itemize} \item $(a,b)\in M$ implies $b \succ a;$ i.e. $a<b$ and no $c$ satisifies $a<c<b$. \item each $a\in P$ belongs to at most one element in $M$. \end{itemize} When $(a,b) \in M$, we write $a=d(b)$ and $b=u(a)$. \item A partial matching on $P$ is called \emph{acyclic} if there does not exist a cycle \[ a_1 \prec u(a_1) \succ a_2 \prec u(a_2) \succ \cdots \prec u(a_m) \succ a_1 \] with $m\ge 2$ and all $a_i\in P$ being distinct. \end{definition} Given an acyclic partial matching $M$ on a poset $P$, we call an element $c$ \emph{critical} if it is unmatched. If every element is matched by $M$, we say $M$ is \emph{perfect}. We are now able to state the main theorem of discrete Morse theory. \begin{theorem} Let $\Delta$ be a simplicial complex and let $M$ be an acyclic matching on the face poset of $\Delta$. Let $c_i$ denote the number of critical $i$-dimensional cells of $\Delta$. The space $\Delta$ is homotopy equivalent to a cell complex $\Delta_c$ with $c_i$ cells of dimension $i$ for each $i\ge 0$, plus a single $0$-dimensional cell in the case where the emptyset is paired in the matching. \end{theorem} \begin{remark} If the critical cells of an acyclic matching on $\Delta$ form a subcomplex $\Gamma$ of $\Delta$, then $\Delta$ simplicially collapses to $\Gamma$, implying that $\Gamma$ is a deformation retract of $\Delta$. \end{remark} It is often useful to create acyclic partial matchings on several different sections of the face poset of a simplicial complex and then combine them to form a larger acyclic partial matching on the entire poset. This process is detailed in the following theorem known as the \textit{Cluster lemma} in \cite{JonssonBook} and the \textit{Patchwork theorem} in \cite{KozlovBook}. \begin{theorem}\label{patchwork} Assume that $\varphi : P \rightarrow Q$ is an order-preserving map. For any collection of acyclic matchings on the subposets $\varphi^{-1}(q)$ for $q\in Q$, the union of these matchings is itself an acyclic matching on $P$. \end{theorem} \section{Construction of the acyclic matching}\label{Matching} In this section we use discrete Morse theory to describe a simplicial collapsing of $\mathcal{N}(SG_{n,2})$. Section~\ref{Sphere1} contains an analysis of the complex of critical cells of our discrete Morse matching. Our approach will be to produce poset maps from subposets of the face poset of $\mathcal{N}(SG_{n,2})$ to various target posets, construct acyclic matchings on inverse images of these poset maps, and apply Theorem~\ref{patchwork} to obtain an acyclic matching on the entire face poset. In our construction of these poset maps, we will consider facets of $\mathcal{N}(SG_{n,2})$, which by definition arise in the following way. \begin{definition} For $\gamma$ a vertex of $SG_{n,2}$, let $\Sigma_{\gamma}$ be the facet in $\mathcal{N}(SG_{n,2})$ formed by the neighbors of $\gamma$. \end{definition} A key role in our simplicial collapsing is played by the two simplices formed by the collections of all vertices of $SG_{n,2}$ of the form $\{\alpha_1,\ldots,\alpha_n\}$, where in each simplex the $\alpha_i$ have all even or all odd entries, respectively. These all even and all odd simplices may be viewed as North and South poles for the complex. As these pole simplices are not two-dimensional, we must collapse them to smaller dimension. The facets $\Sigma_{\gamma}$ where $\gamma$ is loose then collapse to pairs of triangles that interpolate between these two poles, forming our sphere. \subsection{Collapsing in facets of loose stable $n$-sets} For any loose stable $n$-set $\alpha$, $\Sigma_{\alpha}$ is a $3$-dimensional simplex in $\mathcal{N}(SG_{n,2})$ formed by the outer and immediate neighbors of $\alpha$. A routine check reveals that the edge consisting of $\alpha$'s outer neighbors is free in $\mathcal{N}(SG_{n,2})$. Thus, for each such facet we may perform the following collapse. Label the vertices of $\Sigma_\alpha$ by $a,b,c,d$ with the outer neighbors of $\alpha$ labeled $b$ and $d$. Let $P_\alpha$ be the face poset of $\Sigma_{\alpha}$ and $Q_\alpha:=A<B_\alpha$ a chain of length $2$. Let $\theta_\alpha : P_\alpha \rightarrow Q_\alpha$ be defined by \[ \theta_\alpha (x) = \left \{ \begin{array}{ll} A& \mbox{if } \{ b, d\} \nsubseteq x \\ B_\alpha & \mbox{if } \{ b, d\} \subseteq x \end{array} \right. \] It is immediate that $\theta_\alpha$ is a poset map and that $\theta^{-1}(B_\alpha)$ yields a perfect acyclic matching when we match an element $x$ in the inverse not containing $a$ with $x \cup \{a\}$. This matching collapses each facet given by a loose stable $n$-set to two triangles that share a common edge. \subsection{Collapsing in facets of tight stable $n$-sets}\label{collapse} Consider a tight stable $n$-set $\alpha$ in $[2n+2]$, and observe that all elements of $\alpha$ are of the same parity. \begin{lemma} $\alpha$ has a unique outer neighbor. \end{lemma} \begin{proof} Observe that $[2n+2] \setminus \alpha$ consists of the $n+1$ elements of the opposite parity of the elements of $\alpha$ and the one remaining element of the same parity as the elements of $\alpha$. An outer neighbor of $\alpha$ must contain the one element of the same parity as the elements of $\alpha$, which we denote $p$. As the outer neighbor is a stable $n$-set, it cannot contain $p\pm 1$. Since there are only $n-1$ viable elements left in $[2n+2] \setminus \alpha$, an outer neighbor of $\alpha$ must contain them all. Hence, $\alpha$ has a unique outer neighbor. \end{proof} To simplify our presentation we introduce additional notation. Lexicographically assign the neighbors of $\alpha$ the labels $ v^1, v^2, \ldots, v^{n+1}$ and $\eta_{\alpha}$ where the $v^i$'s are all tight and of the opposite parity of $\alpha$. The remaining vertex, $\eta_\alpha$, denotes $\alpha$'s unique outer neighbor. Let $\Sigma_{\alpha}$ denote the $(n+1)$-simplex formed by the neighbors of $\alpha$ and let $P_\alpha$ denote the face poset of $\Sigma_\alpha$. Given $\alpha$ and its unique outer neighbor $\eta_\alpha$, let $p$ denote the element in the outer neighbor $\eta_\alpha$ of identical parity to the elements of $\alpha$. For some $j$, we obtain $v^j$ and $v^{j+1}$ from $\eta_\alpha$ by replacing $p$ with $p-1$ or $p+1$, respectively. \begin{lemma} $\Sigma_\alpha$ collapses to the simplicial complex $N_\alpha$ where $N_\alpha$ consists of the following facets and their subsets: \[ \{v^1,v^2,v^3\} , \{v^1,v^3,v^4\}, \{v^1,v^4,v^5 \}, \ldots , \{v^1, v^n, v^{n+1} \}, \{v^j, v^{j+1}, \eta_\alpha \} \] where if $j= n + 1$ then the last set listed above is replaced by $\{ v^1 , v^{n+1} , \eta_\alpha\}$. In other words, $\Sigma_\alpha$ collapses to a triangulated $(n+1)$-gon where all diagonals in the triangulation emanate from the vertex labeled $v^1$ and the triangle $\{v^j, v^{j+1}, \eta_\alpha \}$ is attached to the $(n+1)$-gon. \end{lemma} The idea behind our matching in the following proof is that the intersection of any two facets corresponding to tight sets is the simplex $\{v^1,v^2,v^3,\ldots,v^{n+1}\}$. To collapse $\Sigma_\alpha$ to $N_\alpha$, we will pair unwanted faces contained in $\{v^1,v^2,v^3,\ldots,v^{n+1}\}$ with $v^1$, and pair unwanted faces containing $\eta_\alpha$ with $v^j$. Separating these matchings allows us to patch the relevant poset maps together in a coherent way in the following subsection. \begin{proof} Fix a tight stable $n$-set $\alpha$, with outer neighbor $\eta_\alpha$ and associated $v^j$. Let $Q_\alpha:=A<B<C_\alpha$ be a three element chain. Consider the map $\varphi_\alpha : P_\alpha \rightarrow Q_\alpha$ defined by \[ \varphi_\alpha (x) = \left \{ \begin{array}{ll} A & \mbox{if } |x|=1, \mbox{ } x= \{v^r, v^s\}, x= \{v^1, v^r, v^s\}, \mbox{ or } x\subseteq\{v^j,v^{j+1},\eta_\alpha\}\\ B & \mbox{for all other } x \mbox{ such that } \eta_\alpha \notin x \\ C_\alpha & \mbox{otherwise} \\ \end{array} \right. \] where either $r =1$ and $s\in [n+1]\setminus \{1\}$ or $s=r+1$ for $r\in [n]\setminus \{1\}$. Observe that $\varphi^{-1}_\alpha(A)$ is exactly the complex $N_\alpha$ defined above. We now construct acyclic matchings on the posets $\varphi_\alpha^{-1}(B)$ and $\varphi_\alpha^{-1}(C_\alpha)$. We claim that matching each $x\in \varphi_{\alpha}^{-1}(B)$ not containing $v^1$ with $x \cup \{v^1\}$ yields a perfect acyclic matching. One first needs to check that no element is paired with an element of $\varphi_{\alpha}^{-1}(A)$ or $\varphi_{\alpha}^{-1}(C_\alpha)$, which is clear from the definitions. That every face is matched is similarly clear. To verify acyclicity, suppose a cycle exists, say $x_1 \prec u(x_1) \succ x_2 \prec u(x_2) \succ \cdots \prec u(x_m) \succ x_1$, for $m$ minimal. Then, both $u(x_1)$ and $u(x_m)$ contain $x_1$ (as sets). However, our matching dictates that we match $x_1$ and $u(x_1)$ if and only if they are in $\varphi_\alpha^{-1} (B)$ and $u(x_1) = x_1 \cup \{v^1\}$. If $v^1\in u(x_m)$, then $u(x_m) = u(x_1)$ implying $x_m = x_1$, a contradiction. Otherwise, $v^1\notin u(x_m)$ implies $u(x_m)$ is also matched with $u(x_m) \cup \{v^1\}$, a contradiction. We claim that matching each $x\in \varphi_{\alpha}^{-1}(C_\alpha)$ such that $j\notin x$ with $x\cup \{j\}$ yields a perfect acyclic matching in $\varphi_{\alpha}^{-1}(C_\alpha)$. It is clear from the definitions that no element is paired with something outside $\varphi_{\alpha}^{-1}(C_\alpha)$, keeping in mind the observation that the pairs $(\eta_\alpha,\{v^j,\eta_\alpha\})$ and $(\{v^{j+1},\eta_\alpha\},\{v^j,v^{j+1},\eta_\alpha\})$ are not included in this preimage; they are included in the preimage $\varphi_\alpha^{-1}(A)$. Verifying that this is a perfect acyclic matching is similar to the previous case. \end{proof} \subsection{Combining the loose and tight cases to form a single poset map} Our matchings were all defined by studying poset maps with domains the facets of $\mathcal{N}(SG_{n,2})$. To apply Theorem~\ref{patchwork}, we need to show that these maps may be combined into a single poset map in a coherent manner. Consider the poset $Q(n,2)$ formed by identifying along commonly named elements the posets $Q_\alpha$ from the constructions of our matchings. In other words, $Q(n,2)$ has a unique minimal element $A$, a maximal chain on two vertices labeled $A<B_\alpha$ for each loose $n$-set $\alpha$, and a maximal chain of length three labeled $A<B<C_\alpha$ for each tight $n$-set $\alpha$ that all share the common subchain $A<B$. For each of the poset maps $\theta_{\alpha}$ and $\varphi_{\alpha}$ defined in the previous subsection, we view them as a map from $P_\alpha$ to $Q(n,2)$. Let $P(n, 2)$ denote the face poset of $\mathcal{N}(SG_{n,2})$. We define a map $\Phi$ from $P(n,2)$ to $Q(n,2)$ by mapping a face $x\in \Sigma_{\alpha}$ to \[ \Phi(x) = \left\{ \begin{array}{ll} \theta_{\alpha}(x) & \mbox{if } \alpha \mbox{ is loose} \\ \varphi_{\alpha}(x) & \mbox{if } \alpha \mbox{ is tight } \\ \end{array} \right. \] \begin{lemma} $\Phi$ is a well-defined poset map. \end{lemma} \begin{proof} Assuming that $\Phi$ is well-defined, that it is a poset map is immediate since $\theta$ and $\varphi$ are poset maps. To verify $\Phi$ is well-defined, we need to check that faces contained in more than one facet are mapped coherently by $\Phi$. Let $\alpha^1$ and $\alpha^2$ be two stable sets that yield the facets $\Sigma_{\alpha^1}$ and $\Sigma_{\alpha^2}$ in $\mathcal{N}(SG_{n,2})$. \underline{Case 1:} Suppose $\alpha^1$ and $\alpha^2$ are both loose sets. We consider the size of the intersection of their respective facets. If $|\Sigma_{\alpha^1} \cap \Sigma_{\alpha^2}| = 4$, then $\alpha^1 = \alpha^2$ and we are done. Suppose $|\Sigma_{\alpha^1} \cap \Sigma_{\alpha^2}| = 3$. Say $\{v^1, v^2, v^3\} \subset \Sigma_{\alpha^1} \cap \Sigma_{\alpha^2}$ along with all their subsets for some vertices $v^1, v^2,$ and $v^3$. Consider the support of these vertices, $supp(v^1, v^2, v^3)$, where $supp(v^1,\ldots,v^k):=\cup_i v^i$ as sets. We know each of these vertices avoid the stable $n$-sets $\alpha^1$ and $\alpha^2$, thus there are at most $n+1$ viable elements remaining in $[2n+2]$. However, $|supp(v^1, v^2, v^3)| \ge n+2$, since the intersection of any two of the vertices can have at most $n-1$ elements in common. Thus, this case does not occur. Suppose $|\Sigma_{\alpha^1} \cap \Sigma_{\alpha^2}| \le 2$. In this case, $\Sigma_{\alpha^1} \cap \Sigma_{\alpha^2}$ is either a single vertex or an edge between an inner and an outer neighbor. As any such face is sent to $A$ by both $\theta_{\alpha^1}$ and $\theta_{\alpha^2}$, we see that $\Phi$ is well-defined on the intersection of pairs of loose sets. \underline{Case 2:} If $\alpha^1$ and $\alpha^2$ are both tight sets with $\alpha^1 \ne \alpha^2$, then $|\alpha^1 \cap \alpha^2|= n-1$, implying that $\Sigma_{\alpha^1} \cap \Sigma_{\alpha^2}$ is an $n$-dimensional simplex $\Sigma$. Using our previous notation, $\Sigma=\{v^1,v^2,v^3,\ldots,v^{n+1}\}$. For a given face $x\in \Sigma$, every map $\varphi_\alpha$ maps $x$ to either $A$ or $B$ in a coherent manner, as the definitions of the $\varphi$-maps are the same on $\Sigma$. Thus, $\Phi$ is well-defined on the intersection of pairs of tight sets. \underline{Case 3:} Suppose $\alpha^1$ is a tight set and $\alpha^2$ is a loose set. If $|\alpha^1 \cap \alpha^2| \le n-2$, then $|supp(\alpha^1,\alpha^2)| \ge n+2$. Hence, $|[2n+2] \setminus supp(\alpha^1,\alpha^2)|\le n$. Thus, $\Sigma_{\alpha^1}$ and $\Sigma_{\alpha^2}$ intersect in a vertex $x$, and $\Phi (x) =A$ is well-defined. If $|\alpha^1 \cap \alpha^2| = n-1$ consider $F=[2n+2]\setminus supp(\alpha^1, \alpha^2)$. We know $|F|=n+1$ as $|supp(\alpha^1, \alpha^2)|=n+1$. Moreover, $F$ consists of $n$ elements of the opposite parity of $\alpha^1$ and one element, say $p$, of the same parity. From this we know that $p \pm 1$ is in $F$, but not both. Hence, $F$ contains only two stable $n$-sets, one set $\beta$ which is tight and whose elements are of opposite parity of $\alpha^1$ and another set $\gamma$, which consists of $p$ and $n-1$ elements of opposite parity of $\alpha^1$ not including $p\pm 1$. Thus, $\Sigma_{\alpha^1} \cap \Sigma_{\alpha^2} = \{\beta, \gamma\}$, all faces of which are mapped to $A$ by $\varphi_{\alpha^1}$. We next show that all faces of $F$ are mapped to $A$ by $\theta_{\alpha^2}$ as well. As $\alpha^2$ is loose and $\beta$ is a tight neighbor, we know that $\beta$ is an outer neighbor of $\alpha^2$. In addition, $\gamma$ is an immediate neighbor of $\alpha^2$ by construction. The only edge sent to $B_{\alpha^2}$ by $\theta_{\alpha^2}$ is the one formed by both outer neighbors of $\alpha^2$, which this edge is not, thus it is sent to $A$. Hence, $\Phi$ is well-defined. \end{proof} To use Theorem~\ref{patchwork}, we now need to verify that our previous matchings are valid. Recall that \[\Phi^{-1}(A)=\cup_\alpha\varphi^{-1}_\alpha(A) \bigcup \cup_\alpha\theta_\alpha^{-1}(A),\] and that none of these preimages carried matchings. The other preimages of $\Phi$ correspond to the preimages of $\theta_\alpha$ or $\varphi_\alpha$ depending whether $\alpha$ is loose or tight. Our previous matchings may therefore be applied, after noting that on $\Phi^{-1}(B)$ the matching is independent of choice of $\alpha$. Hence by Theorem~\ref{patchwork} and the remark following it we have that $\mathcal{N}(SG_{n,2})$ simplicially collapses onto the complex whose face poset is $\Phi^{-1}(A)$. \section{Analysis of the complex of critical faces}\label{Sphere1} Throughout this section it will be useful to refer to Figure~\ref{p}, illustrating the case $n=3$. Denote by $\widetilde{\mathcal{N}}(SG_{n,2})$ the complex of critical faces given by $\Phi^{-1}(A)$. By construction, $\widetilde{\mathcal{N}}(SG_{n,2})$ is two-dimensional and pure; in this section we prove that it is the boundary of a three-dimensional simplicial polytope. Our approach is to first construct a planar graph inducing a triangulation of $S^2$ that realizes $\widetilde{\mathcal{N}}(SG_{n,2})$, then to apply the following theorem. Recall that a graph $G$ is \emph{$3$-connected} if for any pair of vertices $v$ and $w$ in $G$, there exist three disjoint paths from $v$ to $w$. \begin{theorem}\label{Steinitz}{\rm (Steinitz' theorem, see \cite{ZieglerLectures})} A simple graph $G$ is the one-skeleton of a three-dimensional polytope if and only if it is planar and $3$-connected. \end{theorem} \subsection{Construction of $\widetilde{\mathcal{N}}(SG_{n,2})$} We want to realize $\widetilde{\mathcal{N}}(SG_{n,2})$ as a triangulation of $S^2$; we will do so by constructing its one-skeleton in the plane. We begin with notation and several lemmas. For a stable $n$-set $\alpha$, let $\alpha_{\mbox{\small{odd}}}$ be the set of all odd elements of $\alpha$ and let $\alpha_{\mbox{\small{even}}}$ be the set of all even elements of $\alpha$. Throughout this subsection, unless otherwise indicated, we assume for a stable $n$-set $\alpha=\{\alpha_1, \ldots , \alpha_n\}$ that $\alpha_{\mbox{\small{odd}}} =\{\alpha_1, \ldots, \alpha_i\}$ and $\alpha_{\mbox{\small{even}}} = \{\alpha_{i+1}, \ldots, \alpha_n\}$. Let $P_i$ denote the set of stable $n$-sets consisting of $i$ even elements and $n-i$ odd elements. \begin{lemma} $P_i=\{\alpha^0,\ldots,\alpha^n\}$ is lexicographically ordered by setting \[\alpha^0:=\{1,3,\ldots,2(n-i)-1,2(n-i)+2,\ldots,2n\}\] and $\alpha^j:=\alpha^0\ominus 2j$. Also, $\alpha^n \ominus 2 = \alpha^0$. \end{lemma} \begin{proof} For $P_0$, it is immediate that $\{1,3,5,\ldots,2n-1\}\leq \{1,3,5,\ldots,2n-3,2n+1\} \leq \{1,3,5,\ldots,2n-5,2n-1,2n+1\}\leq \cdots$ orders $P_0$ lexicographically. Given an $\alpha\in P_0$, it follows by inspection that $\alpha\ominus 2$ is the next term in the sequence. The set $P_n$ is handled similarly. For the case of $P_i$, $0<i<n$, it is immediate that $\{1,3,5,\ldots,2(n-i)-1,2(n-i)+2,\ldots,2n\}\leq \{1,3,5,\ldots,2(n-i)-3,2(n-i),\ldots,2n,2n+1\}\leq \cdots \leq \{3,5,7,\ldots,2(n-i)+1,2(n-i)+4,\ldots,2n+2\}$ orders $P_i$ lexicographically. Given an $\alpha\in P_i$, it follows by inspection that $\alpha\ominus 2$ is the next term in the sequence. \end{proof} \begin{lemma}\label{neigbors_cycle} If $\alpha \in P_i$ and $\beta, \gamma \in P_{i+1}$ such that $|\alpha \cap \beta| = |(\alpha\ominus 2) \cap \beta| = n-1$, $|(\alpha\ominus 2) \cap \beta| = |(\alpha\ominus 2) \cap \gamma| = n-1$, and $|(\alpha\ominus 2) \cap (\beta \ominus 2)| = |(\alpha\ominus 4) \cap \beta\ominus 2| = n-1$, then $\gamma=\beta\ominus 2$. \end{lemma} \begin{proof} We maintain our ordering of the elements for a stable $n$-set $\alpha=\{\alpha_1, \ldots , \alpha_n\}$ as $\alpha_{\mbox{\small{odd}}} =\{\alpha_1, \ldots, \alpha_i\}$ and $\alpha_{\mbox{\small{even}}} = \{\alpha_{i+1}, \ldots, \alpha_n\}$. As $\alpha \in P_i$ we have \begin{eqnarray} \alpha_{\mbox{\small{odd}}} & = & \{\alpha_1, \ldots, \alpha_i\} \nonumber \\ (\alpha \ominus 2)_{\mbox{\small{odd}}} & = & \{\alpha_1-2, \ldots, \alpha_i - 2\}\nonumber\\ & = & \{\alpha_1, \ldots, \alpha_{i-1}, \alpha_n +1\}\nonumber\\ (\alpha \ominus 4)_{\mbox{\small{odd}}} & = & \{\alpha_1-4, \ldots, \alpha_i - 4\}\nonumber\\ & = & \{\alpha_1, \ldots, \alpha_{i-2}, \alpha_n -1 , \alpha_n + 1\}\nonumber \end{eqnarray} By our assumptions about $\beta$ and $\gamma$ we have \begin{eqnarray} \beta_{\mbox{\small{odd}}} &=& \alpha_{\mbox{\small{odd}}} \cap (\alpha \ominus 2)_{\mbox{\small{odd}}} = \{\alpha_1, \ldots, \alpha_{i-1}\}\nonumber\\ \gamma_{\mbox{\small{odd}}} &=& (\alpha \ominus 2)_{\mbox{\small{odd}}} \cap (\alpha \ominus 4)_{\mbox{\small{odd}}} = \{\alpha_1, \ldots, \alpha_{i-2}, \alpha_n +1\}\nonumber \end{eqnarray} Thus, $\beta_{\mbox{\small{odd}}} \ominus 2 = \gamma_{\mbox{\small{odd}}}$. By a similar argument we see that $\beta_{\mbox{\small{even}}} \ominus 2 = \gamma_{\mbox{\small{even}}}$ and hence $\beta \ominus 2 =\gamma$. So $\beta$ and $\gamma$ are neighbors in $P_{i+1}$. \end{proof} To construct our planar graph, order the elements of $P_0$ lexicographically and denote them $v^0, \ldots , v^n$. Draw a regular $(n+1)$-gon, which we will also refer to as $P_0$, and cyclically label its vertices by $v^0, \ldots , v^n$. Triangulate $P_0$ so that each diagonal in the triangulation has the vertex $v^0$ as an endpoint. Next, we draw a second regular $(n+1)$-gon, denoted $P_1$, around $P_0$, satisfying two conditions: \begin{itemize} \item The vertices of $P_1$ lie outside $P_0$ on lines through the center point of $P_0$ and the midpoints of the edges of $P_0$, and \item The edges of $P_1$ do not intersect $P_0$. \end{itemize} Label the vertices of the polygon $P_1$ by the elements of the set $P_1$, where the labels are placed cyclically about the circle in the lexicographic order; the lexicographically first label for $P_1$ is placed on the ray between the center of $P_0$ and the edge between $v^0$ and $v^1$. Connect a vertex $v$ of $P_0$ to a vertex $w$ of $P_1$ if $|v\cap w| = n-1$, i.e. connect $w$ to the endpoints of the edge of $P_0$ that it is nearest to. We inductively continue this process for $i\leq n$ by drawing an $(n+1)$-gon denoted $P_i$ around $P_{i-1}$. Label the vertices of the polygon $P_i$ with the elements of the set $P_i$ in such a way that one may connect a vertex $v$ of $P_{i-1}$ to a vertex $w$ of $P_i$ exactly when $|v\cap w| = n-1$. This results in the vertices of $P_i$ being labeled cyclically with respect to lexicographic order, with the requirement that the lex-first label for $P_i$ is placed on the ray between the center of $P_0$ and the edge between the lexicographically first and second vertices of $P_{i-1}$. To complete the construction, once $P_n$ has been drawn and connected to $P_{n-1}$, draw arcs representing the edges $\{\{2,4, 6, \ldots , 2n \}, e \}$ for all even, tight stable $n$-sets $e$. It is immediate from our lemmas that this construction is legitimate, and also it is clear that it yields a triangulation of the sphere. To finish our proof, we must show that the facets of $\widetilde{\mathcal{N}}(SG_{n,2})$ are the same as the facets of this triangulation, i.e. that this triangulation is actually a realization of $\widetilde{\mathcal{N}}(SG_{n,2})$. Observe that both $P_0$ and $P_n$ bound triangulated $(n+1)$-gons where the vertices of $P_0$ are the odd tight sets while the vertices of $P_n$ are the even tight sets and the diagonals in the triangulations all emanate from the lexicographically smallest tight stable set in each of $P_0$ and $P_n$. These triangulated polygons correspond exactly to the triangulated polygons contained in the $N_\alpha$ complexes defined in regards to facets of tight $n$-sets. What remains is to show that every other facet of our triangulation corresponds to a two-dimensional simplex in $\widetilde{\mathcal{N}}(SG_{n,2})$ and vice versa. Let $\Sigma$ be a simplex in $\widetilde{\mathcal{N}}(SG_{n,2})$. We will show that $\partial \Sigma$ exists in our constructed graph. We consider three cases. \underline{Case 1:} Suppose $\Sigma$ consists of only tight vertices. Then $\Sigma = \{v^1, v^j, v^{j+1}\}$ for some $j= 2, \ldots, n.$ As $v^j$ and $v^{j+1}$ are lexicographically ordered, we have $v^j= v^{j+1} \ominus 2 $ In the construction of our graph we cyclically connected vertices ordered lexicographically, hence the edge $\{v^j, v^{j+1}\}$ exists in our graph. Moreover, in the construction of our graph we connected all vertices of $P_0$ to the vertex $\{1, 3, 5, \ldots, 2n-1\}$ and all vertices of $P_n$ to the vertex $\{2,4, 6, \ldots , 2n \}$. These are precisely the edges $\{v^1, v^j\}$ and $\{v^1, v^{j+1}\}$. \underline{Case 2:} Suppose $\Sigma$ consists of tight and loose vertices. This case follows easily from the following lemma. \begin{lemma}\label{neighbors} For $\alpha \in P_i$, there exists a unique vertex $\pi\in P_{i+1}$ such that $|\alpha\cap \pi|=|\alpha\ominus 2 \cap \pi| = n-1$. \end{lemma} \begin{proof} Let $\alpha$, $\alpha \ominus 2 \in P_i$ be two neighboring vertices. We consider two cases. Suppose $\alpha$ and $\alpha \ominus 2$ are both tight sets. Without loss of generality we may assume $\alpha, (\alpha \ominus 2) \in P_0$ and we have $|\alpha \cap (\alpha \ominus 2)| = n-1$. If $\pi$ is a common neighbor to both $\alpha$ and $\alpha \ominus 2$ in $\widetilde{\mathcal{N}}(SG_{n,2})$ then, by construction, $\pi = \alpha \cap (\alpha \ominus 2) \cup \{p\}$ for some $p \in [2n+2]$. We claim $p$ must be $\alpha_n - 1$. By definition, $p$ cannot be any element in $\alpha\cup (\alpha \ominus 2)$. There are $n+1$ such elements. Additionally, $p$ cannot be any of the $n$ elements adjacent to an element in $\alpha\cap (\alpha \ominus 2)$. Thus, we are left with only one choice for $p$ as claimed. Suppose $\alpha$ and $\alpha \ominus 2$ are both loose sets. Set $\pi = (\alpha_{\mbox{\small{odd}}} \cap (\alpha_{\mbox{\small{odd}}} \ominus 2)) \cup (\alpha_{\mbox{\small{even}}} \cup (\alpha_{\mbox{\small{even}}} \ominus 2))$. From our definition of $\pi$ it is immediate that $\pi$ is a stable $n$-set. Moreover, the definitions of $\alpha, (\alpha \ominus 2),$ and $\pi$ we have $|\alpha \cap \pi| = |(\alpha \ominus 2)\cap \pi| = n-1$. Finally, as $|\alpha_{\mbox{\small{odd}}} \cap (\alpha_{\mbox{\small{odd}}} \ominus 2)| = i-1$ and $|\alpha_{\mbox{\small{even}}} \cup (\alpha_{\mbox{\small{even}}} \ominus 2)| = n -i + 1$ we have that $\pi\in P_{i+1}$. The uniqueness of $\pi$ follows from the definitions of $\alpha$ and $\alpha \ominus 2$. Thus our claim holds. \end{proof} A similar argument shows that if $\alpha, (\alpha \ominus 2) \in P_i$, then there exists a unique vertex $\pi \in P_{i-1}$ that is a neighbor to both $\alpha$ and $(\alpha \ominus 2)$ for $i = 1, \ldots, n+1$, where $\pi= (\alpha_{\mbox{\small{even}}} \cap (\alpha_{\mbox{\small{even}}} \ominus 2)) \cup (\alpha_{\mbox{\small{odd}}} \cup (\alpha_{\mbox{\small{odd}}} \ominus 2))$. \underline{Case 3:} Suppose $\Sigma$ consists of only loose vertices. By construction of our poset map, we know that two of these vertices, $v^r$ and $v^s$, are immediate neighbors to some vertex $\alpha$ and the other vertex, $v^t$ is an outer neighbor of $\alpha$. Set $\alpha =\{\alpha_1, \ldots, \alpha_n\}$ where $\alpha$ is a concatenation of $\alpha_{\mbox{\small{odd}}}$ and $\alpha_{\mbox{\small{even}}}$. Then, without loss of generality, $\alpha \ominus 1 = v^r$ and $\alpha \oplus 1 = v^s$. This implies that $v^s \ominus 2 = v^r$ which is an edge in our graph. By definition of an outer neighbor, $v^t = \{(\alpha_1 +1 , \alpha_2 +1, \ldots, \alpha_{j-1} +1 , \alpha_j \pm 2, \alpha_{j+1}+1 , \ldots, \alpha_n +1 )$ where $\alpha_i$ is odd for $i=1, \ldots, j-1$ and is even for $i= j+1, \ldots, n$. The parity of $\alpha_j$ is unknown. If $\alpha_j$ is odd, then $v^t = \{(\alpha_1 +1 , \alpha_2 +1, \ldots, \alpha_{j-1} +1 , \alpha_j + 2, \alpha_{j+1}+1 , \ldots, \alpha_n +1 )$. From this we immediately see that $v^t \setminus v^s = \{\alpha_j + 2\}$, so $|v^s\cap v^t| = n-1$. Consider $v^t \setminus v^r$. We claim that $v^t\setminus v^r = \{\alpha_n +1\}$ implying $|v^r\cap v^t| = n-1$ so that the edges $\{v^r,v^t\}$ and $\{v^s,v^t\}$ exist in our graph by claim 2 of Lemma~\ref{neighbors}. It is enough to show that $\alpha_j + 2 \in v^r$ as we know $\alpha_n +1 \notin v^r$ and the remaining elements of $v^t$ are in $v^r$ by the definitions of $v^t$ and $v^r$. Since, by assumption, $\alpha_j$ is odd, we know that there is a gap of size two between $\alpha_j$ and $\alpha_{j+1}$. Now, $\alpha_{j+1} -1 \in v^r$ by definition and is odd. As $\alpha_j$ is odd, it is also the case that $\alpha_j +2$ is odd. Moreover, there is only one odd number between $\alpha_j$ and $\alpha_{j+1}$. Thus $\alpha_j +2 = \alpha_{j+1} -1$. The case when $\alpha_j$ is even follows similarly. Now consider a simplex $\sigma$ in our constructed complex. If $\sigma$ consists of only tight vertices, then it is immediate from Case 1 that $\tau \in \widetilde{\mathcal{N}}(SG_{n,2})$, as we constructed it to be so. If $\tau$ consists of any loose vertices then the fact that it is also in $\widetilde{\mathcal{N}}(SG_{n,2})$ is immediate from Lemma~\ref{neighbors} or Case 3 above. \subsection{Proof that $\widetilde{\mathcal{N}}(SG_{n,2})$ is a simplicial polytope}\label{polytopeproof} Let $G_n$ be the $1$-skeleton of $\widetilde{\mathcal{N}}(SG_{n,2})$. By definition, $G_n$ is simple; the planarity of $G_n$ is shown by the our construction. To apply Theorem~\ref{Steinitz} and complete our proof, we must show that $G_n$ is $3$-connected. Let $x$ and $y$ be any two vertices of $G_n$. We will show that there exist (at least) three disjoint paths from $x$ to $y$. The above construction shows us that $G_n$ is built from $n+1$ concentric $(n+1)$-cycles, labeled from inside out $P_0, \ldots P_n$. Recall that each vertex $v$ on a given cycle $P_i$, with the exception of the two cycles formed by tight vertices, is connected to two pairs of adjacent vertices off $P_i$, one pair on each of $P_{i-1}$ and $P_{i+1}$. Each vertex $v$ on either $P_0$ or $P_n$ is connected to only one vertex on an adjacent cycle, either $P_1$ or $P_{n-1}$, respectively. Suppose first that $x$ and $y$ lie on the same cycle $P_i$. Traverse $P_i$ from $x$ to $y$ in opposite directions to obtain two edge-independent paths. The third path can be found by first moving from $x$ to an adjacent cycle, $P_{i+1}$ or $P_{i-1}$, then traveling around this cycle in either direction until a neighbor of $y$ is reached. Next, suppose $x$ and $y$ lie on different cycles, say $x$ on $P_j$ and $y$ on $P_k$ with $j<k$; we begin by finding a pair of disjoint paths from $x$ to $y$. We first construct a pair of disjoint paths from $x$ to $P_k$. Let $v^{1}$ and $w^{1}$ be the neighbors of $x$ that lie on $P_{j+1}$. If $j+1=k$, stop at this point having constructed paths $x,v^1$ and $x,w^1$, otherwise proceed. Let $r^2$ and $v^2$ be the neighbors of $v^{1}$ on $P_{j+2}$ and $v^2$ and $w^2$ be the neighbors of $w^1$ on $P_{j+2}$, noting that $v^2$ is a common neighbor of $v^1$ and $w^1$. If $j+2=k$, stop at this point having constructed paths $x,v^1,v^2$ and $x,w^1,w^2$, otherwise proceed. Now we are in the same situation with $v^2$ and $w^2$ as we were in with $v^1$ and $w^1$, in that we may denote the neighbors of $v^2$ on $P_{j+4}$ by $r^3$ and $v^3$ and the neighbors of $w^3$ by $v^3$ and $w^3$, which allows us to construct paths $x,v^1,v^2,v^3$ and $x,w^1,w^2,w^3$ from $x$ to $P_{j+3}$. If $y$ is not on $P_{j+3}$, then as in the previous cases, we may extend these two paths by setting $v^4$ equal to the unique common neighbor of $v^3$ and $w^3$ on $P_{j+4}$ and setting $w^4$ equal to the other neighbor of $w^3$ on $P_{j+4}$. We continue in this fashion, creating two paths that curve side-by-side through the graph $G_n$, until we reach $P_k$ with paths $x,v^1,\ldots,v^{k-j}$ and $x,w^1,\ldots,w^{k-j}$. Note that $v^{k-j}$ and $w^{k-j}$ are neighbors on $P_k$ by construction. If $v^{k-j}$ and $w^{k-j}$ are both on $P_k$ and neither is $y$, then we may extend these two paths along $P_k$ in opposite directions until we meet $y$. If either $v^{k-j}$ or $w^{k-j}$ is $y$, then we may complete the other path by connecting via one edge. Having completed two disjoint paths from $x$ to $y$, we now need to find a third path disjoint from the first two. If $j+1=k$, let $z$ be a neighbor of $y$ on $P_j$; we may create a third path by considering the path in $P_j$ from $x$ to $z$ followed by the edge from $z$ to $y$. If $k>j+1$, then let $z^0$ be a neighbor of $x$ on $P_j$ not connected by a diagonal. There exists a common neighbor $t$ of $z^0$ and $x$ on $P_1$; let $z^1$ be the other neighbor of $z^0$ on $P_1$. We may choose $z^2$ to be the common neighbor of $z^1$ and $v^1$ on $P_2$. Continue in this fashion, choosing $z^m$ to be the common neighbor of $z^{m-1}$ and $v^{m-1}$ on $P_m$, until one reaches $z^{k-1}$ on $P_{k-1}$. If neither $v^{k-j}$ nor $w^{k-j}$ are equal to $y$, choose a neighbor $s$ of $y$ on $P_{k-1}$ such that $s$ is not $v^{k-1}$ or $w^{k-1}$. Extend the path $x,z^1,z^2,\ldots,z^{k-1}$ to $s$ by traversing $P_{k-1}$, then connect to $y$. If one of $v^{k-j}$ or $w^{k-j}$ is equal to $y$, then extend $z^{k-1}$ to $z^k$ on $P_k$ and connect $z^k$ to $y$ on $P_k$ to complete the path. Our result is a third path that is disjoint from the first two, connecting $x$ to $y$. Thus, $G_n$ is $3$-connected and planar, hence the one-skeleton of a $3$-dimensional polytope. \section{Invariant subcomplexes}\label{Sphere2} In \cite{BraunAutStableKneser}, the first author proved that for $k\geq 1$ and $n\geq 1$ the automorphism group of $SG_{n,k}$ is isomorphic to the dihedral group of order $2(2n+k)$, which we denote $D_{2n+k}$. This action arises naturally, as $D_{2n+k}$ acts on $[2n+k]$ thought of as a regular $(2n+k)$-gon with vertices labeled cyclically; this action preserves stable $n$-sets and disjointness, hence induces an action on $SG_{n,k}$. It is clear from the example in Figure~\ref{p} that this action does not restrict to simplicial automorphisms of $\widetilde{\mathcal{N}}(SG_{n,2})$, because the vertices $\{1,3,5\}$ and $\{3,5,7\}$ are in the same $D_{2n+2}$-orbit but do not have simplicially isomorphic neighborhoods. In general, the vertices $\{1,3,5,\ldots,2n-1\}$ and $\{3,5,7,\ldots,2n+1\}$ share this behavior. It is interesting to search for a polytopal boundary sphere contained in $\mathcal{N}(SG_{n,k})$ that is invariant under this group action. In the case $k=2$, we can find such a sphere after passing to a partial subdivision. \subsection{Subdividing and collapsing $\mathcal{N}(SG_{n,2})$}\label{subdivide} We subdivide $\mathcal{N}(SG_{n,2})$ into a complex we call $\overline{\mathcal{N}}(SG_{n,2})$ by leaving the facets of loose vertices unchanged and subdividing only the facets of tight vertices. We shall consider the case where $\alpha$ is a tight vertex consisting of even elements. For any such $\Sigma_\alpha$, $n+1$ of its vertices are the even, tight vertices and the remaining vertex is a loose vertex consisting of $n-1$ even elements and one odd element. Order the tight even sets in $\mathcal{N}(SG_{n,2})$ lexicographically, denoted by $\alpha^1, \ldots , \alpha^{n+1}$, and label the loose set in $\Sigma_{\alpha^i}$ by $\eta_{\alpha^i}$. For the facet $\Sigma_{\alpha^i}$, using the notation of Subsection~\ref{collapse}, we have distinguished vertices $v^j$ and $v^{j+1}$. Note that $(v^j \cap v^{j+1}) \subset \eta_{\alpha^i}$. Recall that since each of these facets $\Sigma_{\alpha^i}$ contain all the odd, tight vertices, they intersect in a common $n$-dimensional face which we will denote by $F_o$. Barycentrically subdivide $F_o$, and subdivide $\Sigma_{\alpha^i}$ by coning over the subdivision of $F_o$ with $\eta_{\alpha^i}$. To form $\overline{\mathcal{N}}(SG_{n,2})$, apply this subdivision and an identical procedure to the odd tight vertices; denote by $F_e$ the $n$-dimensional face given by the even, tight vertices. The complex we collapse onto will arise as a subcomplex of $\overline{\mathcal{N}}(SG_{n,2})$, which we denote $M(SG_{n,2})$. We will first produce a simplicial collapsing on $\mathcal{N}(SG_{n,2})$ that preserves $F_o$ and $F_e$, then subdivide the $F$'s and some adjacent cells, and finally complete the collapsing on this subdivided complex. Our strategy is very similar to the one used to create $\widetilde{\mathcal{N}}(SG_{n,2})$, and consists of the following steps: \begin{enumerate} \item In facets of loose vertices we collapse on the free edge formed by the outer neighbors of the vertex. \item In each $\Sigma_{\alpha^i}$, we collapse the faces of $\Sigma_{\alpha^i}$ containing $\eta_{\alpha^i}$, except for the triangle $\{v^j,v^{j+1},\eta_{\alpha^i}\}$. \item On the $F$'s, we barycentrically subdivide and then collapse all faces except the triangles $\{\{v^i,v^{i+1}\},v^i,b\}$ and $\{\{v^i,v^{i+1}\},v^{i+1},b\}$, where $b$ is the barycenter of $F$ and the $v^i$'s are the same notation introduced in Subsection~\ref{collapse}. We also subdivide the triangles $\{v^j,v^{j+1},\eta_{\alpha^i}\}$ by subdividing the edge $\{v^j,v^{j+1}\}$. \end{enumerate} Via these collapses, the facets of loose simplices will collapse to our previous pairs of triangles sharing an edge, while the union of the subdivided $\Sigma_{\alpha^i}$'s will deformation retract to complexes given as a barycentrically subdivided polygon with a triangle glued to each boundary edge. For the first two steps of our process, we use the poset map $\Phi$ from Subsection~\ref{collapse}, and apply the matchings used there on the preimages $\Phi^{-1}(B_\alpha)$ and $\Phi^{-1}(C_\alpha)$ ranging over all stable $n$-sets $\alpha$. The resulting matching induces a simplicial collapse onto a subcomplex of $\mathcal{N}(SG_{n,2})$ consisting of a pair of triangles for each loose vertex, the simplices $F_o$ and $F_e$, and a triangle of the form $\{v^j,v^{j+1},\eta_{\alpha^i}\}$ for each tight set $\alpha^i$. For the third step in our process, we will subdivide and collapse $F_o$ and $F_e$, along with the $\{v^j,v^{j+1},\eta_{\alpha^i}\}$ triangles. We illustrate this only for $F_o$; $F_e$ is handled identically. Label the odd, tight stable $n$-sets $v^1, \ldots, v^{n+1}$ as before. Apply this labeling to $F_o$; barycentrically subdivide $F_o$, relabeling the remaining vertices in the standard way except we use the label of $b$ for the barycenter. To ensure that our subdivision remains a simplicial complex, we must also subdivide each $\{v^j,v^{j+1},\eta_{\alpha^i}\}$ into two triangles, $\{\{v^j,v^{j+1}\},v^j,\eta_{\alpha^i}\}$ and $\{\{v^j,v^{j+1}\},v^{j+1},\eta_{\alpha^i}\}$. Let $\Psi$ be the poset map from the face poset of $F_o$ to the $2$-chain $Q:=0<1$ such that \[ \Psi(x) = \left\{ \begin{array}{ll} 0 & \mbox{ if } x\subseteq\{\{v^m,v^{m+1}\}, m, b\} \mbox{ or } x\subseteq\{\{v^m,v^{m+1}\},v^{m+1}, b\} \\ 1_o & \mbox{ if } \mbox{ otherwise }\\ \end{array} \right. \] where $m\in [n+1]$ and $m+1 = 1$ if $m=n+1$. For $x\in \Psi^{-1}(1)$, we match $x$ with $x \cup b$ if $b \notin x$. This matching is clearly acyclic, and it is perfect since if $w\in \Psi^{-1}(0)$ and $b\notin w$, then $b\cup w$ is contained in $\Psi^{-1}(0)$. It is straightforward to paste the $\Psi$-maps for $F_o$ and $F_e$ together into a single poset map into the poset consisting of two $2$-element chains sharing a common minimal element, i.e. $\{0,1_o,1_e\}$ such that $0<1_o$ and $0<1_e$. If a face of our collapsed, then subdivided, complex from the first two steps is not mapped by $\Psi$ for $F_o$ or $F_e$, then map it to $0$. The resulting poset map allows the application of Theorem~\ref{patchwork}. The proof that $M(SG_{n,2})$ is the boundary of a $3$-dimensional polytope is almost identical to the proof in Subsection~\ref{polytopeproof}. One only needs to observe that the complex resulting from the current analysis is obtained from our previous case by removing the edges inside $P_0$ and $P_n$, barycentrically subdividing $P_0$ and $P_n$, and then subdividing the remaining triangles sharing an edge with $P_0$ or $P_n$. The proof that the one-skeleton is $3$-connected is the same aside from handling the situation where vertices arising from the subdivision are involved, which is a straightforward modification of the argument given in the previous case. \subsection{Action of $D_{2n+2}$ on $M(SG_{n,2})$} Our goal in this subsection is to show that $D_{2n+2}$ acts simplicially on $M(SG_{n,2})$. Consider $[2n+2]$ as the set of vertices of a regular $(2n+2)$-gon on which $D_{2n+2}$ acts in the usual way. Let $\alpha = \{\alpha_1 , \ldots, \alpha_n\}$ be a loose stable set with $\alpha_{\mbox{\small{odd}}} = \{\alpha_1, \ldots, \alpha_i\}$ and $\alpha_{\mbox{\small{even}}} = \{\alpha_{i+1}, \ldots, \alpha_n\}$ Let the immediate neighbors of $\alpha$ be denoted $i(1):=\alpha\oplus1$ and $i(2):=\alpha\ominus1$. Let the outer neighbors of $\alpha$ be denoted \begin{eqnarray*} o(1) &:=& \{\alpha_1 + 1 , \ldots , \alpha_{i-1} +1 , \alpha_i + 2, \alpha_{i+1} +1 , \ldots , \alpha_n +1\} \\ o(2) &:=& \{\alpha_1 +1 , \ldots , \alpha_{n-1} +1 , \alpha_n + 2\}. \end{eqnarray*} There are two simplices associated to $\alpha$ in $M(SG_{n,2})$, given by $\{i(1),i(2),o(j)\}$ for $j=1,2$. As $D_{2n+2}$ is the automorphism group of $SG_{n,2}$, the neighbors of $\alpha$ are mapped by an element $g\in D_{2n+2}$ to neighbors of $g(\alpha)$. Since $\alpha$ is a loose set and $D_{2n+2}$ clearly preserves the loose and tight conditions, $g(\alpha)$ is also a loose set. We will show that the outer neighbors of $\alpha$ are carried by $g$ to the outer neighbors of $g(\alpha)$, hence each of these two simplices associated to $\alpha$ are taken to one of the simplices associated to $g(\alpha)$. For an element $g\in D_{2n+2}$, $g$ is either a rotation or a flip of $[2n+2]$. If $g$ is a rotation, then \begin{eqnarray*} g(o(1)) &=& \{g(\alpha_1 + 1) , \ldots , g(\alpha_{i-1} +1) , g(\alpha_i + 2), g(\alpha_{i+1} +1) , \ldots , g(\alpha_n +1)\}, \\ &=& \{g(\alpha_1) + 1 , \ldots , g(\alpha_{i-1}) +1 , g(\alpha_i) + 2, g(\alpha_{i+1}) +1 , \ldots , g(\alpha_n) +1\}. \end{eqnarray*} Otherwise, $g$ is a flip and \begin{eqnarray*} g(o(1)) &=& \{g(\alpha_1 + 1) , \ldots , g(\alpha_{i-1} + 1) , g(\alpha_i + 2), g(\alpha_{i+1} +1) , \ldots , g(\alpha_n + 1)\} \\ &=& \{g(\alpha_1) - 1 , \ldots , g(\alpha_{i-1}) -1 , g(\alpha_i) - 2, g(\alpha_{i+1}) -1 , \ldots , g(\alpha_n) -1\} \\ &=& \{g(\alpha_1) + 1 , \ldots , g(\alpha_{i-1}) +1 , g(\alpha_i) +1 , g(\alpha_{i+1}) +1 , \ldots , g(\alpha_n) +2\} \end{eqnarray*} In either case, $g(o(1))$ is an outer neighbor of $\alpha$ and by a similar argument, $g(o(2))$ is an outer neighbor of $\alpha$. Thus $D_{2n+2}$ sends the associated simplices in $M(SG_{n,2})$ of a loose set $\alpha$ to the associated simplices of $g(\alpha)$. Let $\alpha$ be a tight stable set and consider the four triangles $T_1:=\{v^j, \{v^j,v^{j+1}\}, b\}$, $T_2:=\{v^{j+1}, \{v^j,v^{j+1}\}, b\}$, $T_3:=\{v^j, \{v^j,v^{j+1}\}, \eta_{\alpha}\}$, and $T_4:=\{v^{j+1}, \{v^j,v^{j+1}\}, b\}$ in $M(SG_{n,2})$ where $v^{j+1}$ is a neighboring tight vertex, $\{v^j,v^{j+1}\}$ is the barycenter of the edge $v^j v^{j+1}$, $\eta_{\alpha}$ is the unique vertex that is both a stable set and a neighbor of $v^j$ and $v^{j+1}$ in $M(SG_{n,2})$, and $b$ is the barycenter of the $(n+1)$-gon. Every remaining facet in $M(SG_{n,2})$ can be associated to a tight stable set $\alpha$ in this way, e.g. every remaining facet contains some tight stable set as a vertex. We now apply $g$ to these triangles and show that their image is contained in $M(SG_{n,2})$. For $T_1$, we have $g(T_1) = \{ \{g(v^j), g(\{v^j,v^{j+1}\}), g(b)\}$. As $g$ is either parity preserving or reversing for all elements of $[2n+2]$, we know that $g(b)$ is either $b$ or the corresponding element of opposite parity. In either case, $g(v^j)$ and $g(b)$ are neighbors in our complex as well as $g(\{v^j,v^{j+1}\})$ and $g(b)$. Finally, $g(\{v^j, v^{j+1}\})$ and $g(v^j)$ are neighbors in our complex as $g(\{v^j, v^{j+1}\})= \{g(v^j), g(v^{j+1})\}$. So $T_1$ (as well as $T_2$ by symmetry) maps to a corresponding triangle in our complex for any $g\in D_{2n+2}$. To see that $T_3$ and $T_4$ map to appropriate triangles, we need only check that $g(\eta_{\alpha})$ is a neighbor of $g(v^j)$ and $g(v^{j+1})$ in our complex. By definition and construction the set $\eta_{\alpha} = \{v^j\}\cap \{v^{j+1}\} \cup \{p\}$ where $p$ is the unique element of opposite parity of the elements of $v^j$ and $v^{j+1}$ that allows $\eta_{\alpha}$ to remain stable. Then $g(\eta_{\alpha}) = g(v^j) \cap g(v^{j+1}) \cup g(p)$. As $g(v^j)$ and $g(v^{j+1})$ are connected via $g(\{v^j,v^{j+1}\})$, we know that they have exactly $n-1$ elements in common. Moreover, $g(p)$ is of opposite parity of the elements of $g(v^j)$. Hence, $g(\eta_{\alpha})$ is stable and a neighbor of both $g(v^j)$ and $g(v^{j+1})$ in $M(SG_{n,2})$, thus $D_{2n+2}$ preserves triangles of this form as well. \bibliographystyle{plain}

https://arxiv.org/abs/1804.11003

Gradient Sampling Methods for Nonsmooth Optimization

This paper reviews the gradient sampling methodology for solving nonsmooth, nonconvex optimization problems. An intuitively straightforward gradient sampling algorithm is stated and its convergence properties are summarized. Throughout this discussion, we emphasize the simplicity of gradient sampling as an extension of the steepest descent method for minimizing smooth objectives. We then provide overviews of various enhancements that have been proposed to improve practical performance, as well as of several extensions that have been made in the literature, such as to solve constrained problems. The paper also includes clarification of certain technical aspects of the analysis of gradient sampling algorithms, most notably related to the assumptions one needs to make about the set of points at which the objective is continuously differentiable. Finally, we discuss possible future research directions.

\section{Introduction}\label{sec.introduction} The Gradient Sampling (GS) algorithm is a conceptually simple descent method for solving nonsmooth, nonconvex optimization problems, yet it is one that possesses a solid theoretical foundation and has been employed to substantial success in a wide variety of applications. Since the appearance of the fundamental algorithm and its analysis little over a dozen years ago, GS has matured into a comprehensive methodology. Various enhancements have been proposed that make it a competitive approach in many nonsmooth optimization contexts, and it has been extended in various interesting ways, such as for nonsmooth optimization on manifolds and for constrained problems. The purpose of this work is to provide background and motivation for the development of the GS method, discuss its theoretical guarantees, and provide an overview of the enhancements and extensions that have been the subject of research over recent years. The underlying philosophy of GS is that virtually any nonsmooth objective function of interest is differentiable almost everywhere; in particular, this is true if the objective $f : \R{n} \to \R{}$ is either locally Lipschitz continuous or semialgebraic. In such cases, when $f$ is evaluated at a randomly generated point $x \in \R{n}$, it is differentiable there with probability one. This means that an algorithm can rely on an ability to obtain the objective function value $f(x)$ and gradient~$\nabla f(x)$, as when $f$ is smooth, rather than require an oracle to compute a subgradient. In most interesting settings, $f$ \emph{is not} differentiable at its local minimizers, but, under reasonable assumptions, the carefully crafted mechanisms of the GS algorithm generate a sequence of iterates---at which~$f$ \emph{is} differentiable---converging to stationarity. At the heart of GS is a stabilized steepest descent approach. When $f$ is differentiable at~$x$, the negative gradient $-\nabla f(x)$ is, of course, the traditional steepest descent direction for~$f$ at~$x$ in the 2-norm in that \bequation\label{eq.steepest_descent} -\frac{\nabla f(x)}{\|\nabla f(x)\|_2} = \arg\min_{\|d\|_2 \leq 1}\ \nabla f(x)^Td. \eequation However, when $x$ is near a point where $f$ is not differentiable, it may be necessary to take a very short step along $-\nabla f(x)$ to obtain decrease in $f$. It is for this reason that the traditional steepest descent method may converge to nonstationary points when $f$ is nonsmooth.\footnote{Although this fact has been known for decades, most of the examples that appear in the literature are rather artificial since they were designed with exact line searches in mind. Analyses showing that the steepest descent method with inexact line searches converges to nonstationary points of some simple convex nonsmooth functions have appeared recently in~\cite{AslOver17,GuoLewi17}.} The GS algorithm stabilizes the choice of the search direction to avoid this issue. In each iteration, a descent direction from the current iterate $x^k$ is obtained by supplementing the information provided by $\nabla f(x^k)$ with gradients evaluated at randomly generated points $\{x^{k,1},\dots,x^{k,m}\} \subset \Bmbb(x^k,\epsilon_k) := \{x \in \R{n} : \|x - x^k\|_2 \leq \epsilon_k\}$, which are \emph{near} $x^k$, and then computing the minimum-norm vector $g^k$ in the convex hull of these gradients. This choice can be motivated by the goal that, with $\clarkesub_\epsilon f(x)$ denoting the Clarke $\epsilon$-subdifferential of $f$ at $x$ (see \S\ref{sec.theory}), \bequation -\frac{g^k}{\|g^k\|_2} \approx \arg\min_{\|d\|_2 \leq 1} \max_{g\in{\bar\partial} f_\epsilon(x)} g^Td; \eequation i.e., $-g^k$ can essentially be viewed as a steepest descent direction for $f$ from $x^k$ in a more ``robust'' sense. A line search is then used to find a positive stepsize $t_k$ yielding decrease in $f$, i.e., $f(x^k - t_kg^k) < f(x^k)$. The sampling radius $\epsilon_k$ that determines the meaning of ``\emph{near} $x^k$'' may either be fixed or adjusted dynamically. A specific instance of the GS algorithm is presented in \S\ref{sec.algorithm}. Its convergence guarantees are summarized in~\S\ref{sec.theory}. We then present various enhancements and extensions of the approach in \S\ref{sec.enhancements} and \S\ref{sec.extensions}, respectively, followed by a discussion of some successful applications of the GS methodology in \S\ref{sec.applications}. Throughout this work, our goal is to emphasize the \emph{simplicity} of the fundamental GS strategy. We believe that this, in addition to its strong convergence properties for locally Lipschitz optimization, makes it an attractive choice when attempting to solve difficult types of nonsmooth optimization problems. Although the first convergence analysis of a GS algorithm was given by Burke, Lewis, and Overton in \cite{BurkLewiOver05}, an earlier version of the method was presented by these authors in~\cite{BurkLewiOver02b}. That algorithm, originally called a ``gradient bundle'' method, was applied to a function that was not only nonconvex and nonsmooth, but also non-locally-Lipschitz, namely, the spectral abscissa---i.e., the largest of the real parts of the eigenvalues---of a linear matrix function $A$ mapping a parameter vector $x$ to the space of nonsymmetric square matrices. The spectral abscissa is not locally Lipschitz at a matrix $\bar X$ when an eigenvalue of $\bar X$ with largest real part has multiplicity two or more \cite{BurkOver01}, but it is semialgebraic and, hence, differentiable almost everywhere, so a GS algorithm was applicable. The method was surprisingly effective. As anticipated, in most cases the apparent local minimizers that were approximated had the property that the eigenvalues of $A$ with largest real part had multiplicity two or more. An illustration that appeared in \cite{BurkLewiOver02b} is reproduced in Figure~\ref{grad_bundle}; the extremely ``steep" contours of the objective function indicate its non-Lipschitzness. Obtaining theoretical results for a GS algorithm applied to non-locally-Lipschitz problems seems very challenging; we discuss this issue further in \S\ref{sec.nonlip}, after describing the substantial body of theory that has been developed for the locally Lipschitz case in \S\ref{sec.global} and \S\ref{sec.local}. \begin{figure}[ht] \centering \includegraphics[height=2.5in]{turncorner} \caption{Contours of the spectral abscissa of an affine matrix family given in \cite{BurkLewiOver02b}. Iterates of the ordinary gradient (``steepest descent") method with a line search are shown (small circles) along with those of the gradient sampling (``gradient bundle") algorithm (asterisks). Both start at $(-1,-1)$.} \label{grad_bundle} \end{figure} \section{Algorithm GS}\label{sec.algorithm} We now state a specific variant of the GS algorithm. We start by assuming only that the objective function~$f$ is locally Lipschitz over $\R{n}$, which implies, by Rademacher's theorem \cite{Clar83}, that $f$ is differentiable almost everywhere. As previously mentioned, at the heart of the algorithm is the computation of a descent direction by finding the minimum norm element of the convex hull of gradients obtained about each iterate. The remaining procedures relate to the line search to obtain decrease in $f$ and the selection of a subsequent iterate so that $f$ is differentiable at $\{x^k\}$. \balgorithm[ht] \renewcommand{\thealgorithm}{GS} \caption{(Gradient Sampling)} \label{alg.gs} \balgorithmic[1] \Require initial point $x^0$ at which $f$ is differentiable, initial sampling radius $\epsilon_0 \in (0,\infty)$, initial stationarity target $\nu_0 \in [0,\infty)$, sample size $m \geq n+1$, line search parameters $(\beta,\gamma) \in (0,1) \times (0,1)$, termination tolerances $(\epsilon_\mathrm{opt},\nu_\mathrm{opt}) \in [0,\infty) \times [0,\infty)$, and reduction factors $(\theta_\epsilon,\theta_\nu) \in (0,1] \times (0,1]$ \For{$k \in \N{}$} \State independently sample $\{x^{k,1},\dots,x^{k,m}\}$ uniformly from $\Bmbb(x^k,\epsilon_k)$ \label{step.sample} \State compute $g^k$ as the solution of $\min_{g\in\Gcal^k} \thalf \|g\|_2^2$, where $\Gcal^k := \conv\{\nabla f(x^k),\nabla f(x^{k,1}),\dots,\nabla f(x^{k,m})\}$ \label{step.qp} \State \textbf{if} $\|g^k\|_2 \leq \nu_\mathrm{opt}$ and $\epsilon_k \leq \epsilon_\mathrm{opt}$ \textbf{then} terminate \label{step.terminate} \State \textbf{if} $\|g^k\|_2 \leq \nu_k$ \State \qquad \textbf{then} set $\nu_{k+1} \gets \theta_\nu\nu_k$, $\epsilon_{k+1} \gets \theta_\epsilon\epsilon_k$, and $t_k \gets 0$ \State \qquad \textbf{else} set $\nu_{k+1} \gets \nu_k$, $\epsilon_{k+1} \gets \epsilon_k$, and \label{step.armijo} \bequation\label{eq.armijo} t_k \gets \max\left\{ t \in \{1,\gamma,\gamma^2,\dots\} : f(x^k - t g^k) < f(x^k) - \beta t \|g^k\|_2^2 \right\} \eequation \State \textbf{if} $f$ is differentiable at $x^k - t_k g^k $ \label{step.check} \State \qquad \textbf{then} set $x^{k+1} \gets x^k - t_k g^k$ \State \qquad \textbf{else} set $x^{k+1}$ randomly as any point where $f$ is differentiable such that \label{step.update} \bequation\label{eq.update} f(x^{k+1}) < f(x^k) - \beta t_k \|g^k\|_2^2\ \ \text{and}\ \ \|x^k - t_k g^k - x^{k+1}\|_2 \leq \min\{t_k,\epsilon_k\} \|g^k\|_2 \eequation \EndFor \ealgorithmic \ealgorithm While the essence of the methods from \cite{BurkLewiOver02b} and \cite{BurkLewiOver05} remains intact, Algorithm~\ref{alg.gs} differs in subtle yet important ways from the methods presented in these papers, as we now explain. \begin{enumerate} \item Algorithm~\ref{alg.gs} incorporates a key modification proposed by Kiwiel in \cite[Alg.~2.1]{Kiwi07}, namely, the second inequality in \eqref{eq.update}; the version in \cite{BurkLewiOver05} used $\epsilon_k$ instead of $\min\{t_k,\epsilon_k\}$. As Kiwiel explained, this minor change allowed him to drop the assumption in \cite{BurkLewiOver05} that the level set $\{ x: f(x) \leq f(x^0)\}$ is compact, strengthening the convergence results for the algorithm. \item A second change suggested in \cite[Alg.~2.1]{Kiwi07} is the introduction of the termination tolerances $\nu_\mathrm{opt}$ and $\epsilon_\mathrm{opt}$. These were used in the computational experiments in \cite{BurkLewiOver05}, but not in the algorithm statement or analysis. Note that if $\epsilon_\mathrm{opt}$ is set to zero, then Algorithm~\ref{alg.gs} never terminates since $\epsilon_k$ can never be zero, though it may happen that one obtains $\|g^k\|_2 = 0$. \item A third change, also suggested by Kiwiel, is the usage of the nonnormalized search direction $-g^k$ (originally used in \cite{BurkLewiOver02b}) instead of the normalized search direction $-g^k/\|g^k\|_2$ (used in \cite{BurkLewiOver05}). The resulting inequalities in \eqref{eq.armijo} and \eqref{eq.update} are taken from \cite[Sec.~4.1]{Kiwi07}. This choice does not affect the main conclusions of the convergence theory as in both cases it is established \cite{BurkLewiOver05,Kiwi07} that the stepsize $t_k$ can be determined by a finite process. However, since Theorem \ref{th.eps_to_zero} below shows that a subsequence of $\{g^k\}$ converges to zero under reasonable conditions, one expects that fewer function evaluations should be required by the line search asymptotically when using the nonnormalized search direction, whereas using the normalized direction may result in the number of function evaluations growing arbitrarily large \cite[Sec.~4.1]{Kiwi07}. Furthermore, our practical experience is consistent with this viewpoint. \item Another aspect of Algorithm~\ref{alg.gs} that is different in both \cite{BurkLewiOver05} and \cite{Kiwi07} concerns the randomization procedure in Step~\ref{step.sample}. In the variants given in those papers, it was stated that the algorithm terminates if $f$ is not \emph{continuously} differentiable at the randomly sampled points $\{x^{k,1},\dots,x^{k,m}\}$. In the theorem stated in the next section, we require only that $f$ is differentiable at the sampled points. Since by Rademacher's theorem and countable additivity of probabilities this holds for every sampled point with probability one, we do not include a termination condition here. \item Finally, Steps \ref{step.check}--\ref{step.update} in Algorithm~\ref{alg.gs} \emph{do} require explicit checks that ensure that $f$ is differentiable at~$x^{k+1}$, but unlike in the variants in \cite{BurkLewiOver05} and \cite{Kiwi07}, it is not required that $f$ be \emph{continuously} differentiable at $x^{k+1}$. This differentiability requirement is included since it is not the case that $f$ is differentiable at $x^k - t_k g^k$ with probability one, as is shown via an example in \cite{HeloSantSimo16}, discussed further in \S\ref{subsec.avoidcheck}. For a precise procedure for implementing Step~\ref{step.update}, see \cite{Kiwi07}. \end{enumerate} The computation in Step~\ref{step.qp} requires solving a strongly convex quadratic optimization problem (QP) to compute the minimum-norm element of the convex hull of the current and sampled gradients, or, equivalently, to compute the 2-norm projection of the origin onto this convex hull. It is essentially the same operation required in every iteration of a bundle method. To see this, observe that solving the QP in Step~\ref{step.qp} can be expressed, with \bequationNN G_k := \bbmatrix \nabla f(x^k) & \nabla f(x^{k,1}) & \cdots & \nabla f(x^{k,m}) \ebmatrix, \eequationNN as computing $(z_k,d^k,\lambda^k) \in \R{} \times \R{n} \times \R{m+1}$ as the primal-dual solution of the QPs \bequation\label{prob.qp} \left\{ \baligned \min_{(z,d) \in \R{} \times \R{n}} &\ z + \thalf \|d\|_2^2 \\ \st &\ G_k^Td \leq z \mathds{1} \ealigned \right\} \quad \text{and} \quad \left\{ \baligned \min_{\lambda \in \R{m+1}} &\ \thalf \|G_k\lambda\|_2^2 \\ \st &\ \mathds{1}^T\lambda = 1,\ \lambda \geq 0 \ealigned \right\}. \eequation The latter problem, yielding $G_k\lambda^k = g^k$, can easily be seen to be equivalent to solving the subproblem $\min_{g\in\Gcal^k} \thalf \|g\|_2^2$ stated in Step~\ref{step.qp}, whereas the former problem, yielding $d^k = -g^k$, can be seen to have the same form as the subproblems arising in bundle methods. Normally, the initial stationarity target $\nu_0$ is chosen to be positive and the reduction factors $\theta_\nu$ and $\theta_\epsilon$ are chosen to be less than one so that the stationarity target and sampling radius are reduced every time the condition $\|g^k\|\leq\nu_k$ is satisfied. However, it is also interesting to consider the variant with $\nu_0=0$ and $\theta_\epsilon=1$, forcing the algorithm to run forever with $\epsilon$ fixed unless it terminates with $g^k=0$ for some $k \in \N{}$. We consider both of these variants in the global convergence theory given in the next section. \section{Convergence Theory for Algorithm~\ref{alg.gs}}\label{sec.theory} We begin with some definitions. In the locally Lipschitz case, the Clarke subdifferential of $f$ at $x \in \R{n}$ is defined by the convex hull of the limits of gradients of $f$ on sequences converging to $x$ \cite[Def.\ 1.1]{Clar75}, i.e., \bequation\label{eq.clarkesub} {\bar\partial} f(x) = \conv \left\{ \lim_{j\to\infty} \nabla f(y^j) : \{y^j\} \to x\ \text{where } f \text{ is differentiable at } y^j \text{ for all } j \in \N{} \right\}. \eequation A point $x$ is Clarke stationary for $f$ if $0 \in {\bar\partial} f(x)$. A more ``robust'' sense of stationarity with respect to $f$ can be defined by considering the limits of gradients corresponding to limiting points \emph{near} $x$; in particular, given a radius $\epsilon \in [0,\infty)$, the Clarke $\epsilon$-subdifferential \cite{Gold77} is given by\footnote{The definition in \cite{BurkLewiOver05} includes a closure operation but this is unnecessary.} \bequation\label{eq.clarkeepssub} \clarkesub_\epsilon f(x) = \conv \{{\bar\partial} f(\Bmbb(x,\epsilon))\}. \eequation A point $x$ is Clarke $\epsilon$-stationary for $f$ if $0 \in \clarkesub_\epsilon f(x)$. For all practical purposes, one cannot generally evaluate ${\bar\partial} f$ (or $\clarkesub_\epsilon f$) at (or near) any point where $f$ is not differentiable. That said, Algorithm~\ref{alg.gs} is based on the idea that one can approximate the minimum norm element in $\clarkesub_\epsilon f(x^k)$ through random sampling of gradients in the ball $\Bmbb(x^k,\epsilon)$. To a large extent this idea is motivated by \cite{BurkLewiOver02a} which investigates how well the entire Clarke subdifferential ${\bar\partial} f(x)$ can be approximated through random sampling. However, the results in \cite{BurkLewiOver02a} cannot be directly exploited in the analysis of the GS algorithm because the gradients are sampled only at a finite number of points near any given iterate. \subsection{Global Convergence Guarantees}\label{sec.global} A critical aspect of theoretical convergence guarantees for Algorithm~\ref{alg.gs} concerns the set of points where~$f$ is \emph{continuously} differentiable, which we denote by $D$. Consideration of $D$ played a crucial role in the analysis in both \cite{BurkLewiOver05} and \cite{Kiwi07}, but there were some oversights concerning both the requirements of the algorithm with respect to $D$ and the assumptions on $D$. Regarding the requirements of the algorithm with respect to $D$, there is actually no need, from a theoretical point of view, for either the iterates $\{x^k\}$ or the randomly generated sampled points $\{x^{k,j}\}$ to lie in $D$; all that is needed is that $f$ is differentiable at these points. Most implementations of GS algorithms do not attempt to check any form of differentiability in any case, but if one were to attempt to implement such a check, it is certainly more tractable to check for differentiability than continuous differentiability. Regarding the assumptions on $D$, in the theorems that we state below, we assume that $D$ is an open set with full measure in $\R{n}$. In contrast, the relevant assumption stated in \cite{BurkLewiOver05,Kiwi07} is weaker, namely, that $D$ is an open dense subset of $\R{n}$. However, the proofs of convergence actually require the full measure assumption on $D$ that we include below.\footnote{This oversight went unnoticed for 12 years until J.~Portegies and T.~Mitchell brought it to our attention recently.} There are three types of global convergence guarantees of interest for Algorithm~\ref{alg.gs}: one when the input parameters ensure that $\{\epsilon_k\} \searrow 0$, one when $\epsilon_k $ is repeatedly reduced but a positive stopping tolerance prevents it from converging to zero, and one when $\epsilon_k=\epsilon > 0$ for all $k$. These lead to different properties for the iterate sequence. The first theorem below relates to cases when the stationarity tolerance and sampling radius tolerance are both set to zero so that the algorithm can never terminate. \btheorem\label{th.eps_to_zero} Suppose that $f$ is locally Lipschitz in $\R{n}$ and continuously differentiable on an open set $D$ with full measure in $\R{n}$. Suppose further that Algorithm~\ref{alg.gs} is run with $\nu_0 > 0$, $\nu_\mathrm{opt} = \epsilon_\mathrm{opt} = 0$, and strict reduction factors $\theta_\nu < 1$ and $\theta_\epsilon < 1$. Then, with probability one, Algorithm~\ref{alg.gs} is well defined in the sense that the gradients in Step \ref{step.qp} exist in every iteration, the algorithm does not terminate, and either \benumerate \item[(i)] $f(x^k) \searrow -\infty$, or \item[(ii)] $\nu_k \searrow 0$, $\epsilon_k\searrow 0$, and every cluster point of $\{x^k\}$ is Clarke stationary for $f$. \eenumerate \etheorem Theorem~\ref{th.eps_to_zero} is essentially the same as \cite[Thm 3.3]{Kiwi07} (with the modifications given in~\cite[\S4.1]{Kiwi07} for nonnormalized directions), except for two aspects: \benumerate \item The proof given in \cite{Kiwi07} implicitly assumes that $D$ is an open set with full measure, as does the proof of \cite[Thm 3.4]{BurkLewiOver05} on which Kiwiel's proof is based, although the relevant assumption on $D$ in both papers is the weaker condition that $D$ is an open dense set. Details are given in Appendix \ref{app.fullmeasure}. \item In the algorithms analyzed in \cite{Kiwi07} and \cite{BurkLewiOver05}, the iterates $\{x^k\}$ and the randomly sampled points $\{x^{k,j}\}$ were enforced to lie in the set $D$ where $f$ is continuously differentiable. We show in Appendix \ref{app.fdifOK} that the theorem still holds when this requirement is relaxed to ensure only that~$f$ is differentiable at these points. \eenumerate As Kiwiel argues, Theorem~\ref{th.eps_to_zero} is essentially the best result that could be expected. Furthermore, as pointed out in \cite[Remark 3.7(ii)]{Kiwi07}, it leads immediately to the following corollary. \bcorollary Suppose that $f$ is locally Lipschitz in $\R{n}$ and continuously differentiable on an open set $D$ with full measure in $\R{n}$. Suppose further that Algorithm~\ref{alg.gs} is run with $\nu_0 > \nu_\mathrm{opt} > 0$, $\epsilon_0 > \epsilon_\mathrm{opt} > 0$, and strict reduction factors $\theta_\nu < 1$ and $\theta_\epsilon < 1$. Then, with probability one, Algorithm~\ref{alg.gs} is well defined in the sense that the gradients in Step \ref{step.qp} exist at every iteration, and either \benumerate \item[(i)] $f(x^k) \searrow -\infty$, or \item[(ii)] Algorithm~\ref{alg.gs} terminates by the stopping criteria in Step~\ref{step.terminate}. \eenumerate \ecorollary The final result that we state concerns the case when the sampling radius is fixed. A proof of this result is essentially given by that of \cite[Thm 3.5]{Kiwi07}, again taking into account the comments in Appendices \ref{app.fullmeasure} and \ref{app.fdifOK}. \btheorem\label{th.eps_pos} Suppose that $f$ is locally Lipschitz in $\R{n}$ and continuously differentiable on an open set $D$ with full measure in $\R{n}$. Suppose further that Algorithm~\ref{alg.gs} is run with $\nu_0 = \nu_\mathrm{opt} = 0$, $\epsilon_0 = \epsilon_\mathrm{opt} = \epsilon > 0$, and $\theta_\epsilon = 1$. Then, with probability one, Algorithm~\ref{alg.gs} is well defined in the sense that the gradients in Step~\ref{step.qp} exist at every iteration, and one of the following occurs: \benumerate \item[(a)] $f(x^k) \searrow -\infty$, or \item[(b)] Algorithm~\ref{alg.gs} terminates for some $k \in \N{}$ with $g^k = 0$, or \item[(c)] there exists $\Kcal \subseteq \N{}$ with $\{g^k\}_{k\in \Kcal} \to 0$ and every cluster point of $\{x^k\}_{k\in \Kcal}$ is Clarke $\epsilon$-stationary for $f$. \eenumerate \etheorem Of the five open questions regarding the convergence analysis for gradient sampling raised in \cite{BurkLewiOver05}, three were answered explicitly by Kiwiel in \cite{Kiwi07}. Another open question was: ``Under what conditions can one guarantee that the GS algorithm terminates finitely?" This was posed in the context of a fixed sampling radius and therefore asks how one might know whether outcome (b) or (c) occurs in Theorem \ref{th.eps_pos}, assuming $f$ is bounded below. This remains open, but Kiwiel's introduction of the termination tolerances in the GS algorithm statement led to Corollary 3.1 which guarantees that when the sampling radius is reduced dynamically and the tolerances are nonzero, Algorithm \ref{alg.gs} must terminate if $f$ is bounded below. The only other open question concerns extending the convergence analysis to the non-Lipschitz case. Overall, Algorithm~\ref{alg.gs} has a very satisfactory convergence theory in the locally Lipschitz case. Its main weakness is its per-iteration cost, most notably due to the need to compute $m \geq n+1$ gradients in every iteration and solve a corresponding~QP. However, enhancements to the algorithm have been proposed that can vastly reduce this per-iteration cost while maintaining these guarantees. We discuss these and other enhancements in \S\ref{sec.enhancements}. \subsection{A Local Linear Convergence Result}\label{sec.local} Given the relationship between gradient sampling and a traditional steepest descent approach, one might ask if there are scenarios in which Algorithm~\ref{alg.gs} can attain a linear rate of local convergence. The study in~\cite{HeloSantSimo17} answers this in the affirmative, at least in a certain probabilistic sense. If $(i)$ the set of sampled points is \emph{good} in a certain sense described in \cite{HeloSantSimo17}, $(ii)$ the objective function $f$ belongs to a class of functions defined by the maximum of a finite number of smooth functions (``finite-max" functions), and $(iii)$ the input parameters are set appropriately, then Algorithm~\ref{alg.gs} will produce a step yielding a reduction in $f$ that is significant. This analysis involves $\Vcal\Ucal$-decomposition ideas~\cite{LemaOustSaga00,Lewi02,MiffSaga05}, where in particular it is shown that the reduction in $f$ is comparable to that achieved by a steepest descent method restricted to the smooth $\Ucal$-space of $f$. This means that a linear rate of local convergence can be attained over any infinite subsequence of iterations in which the sets of sampled points are \emph{good}. \subsection{The Non-Lipschitz Case}\label{sec.nonlip} In the non-locally Lipschitz case, the Clarke subdifferential ${\bar\partial} f$ is defined in \cite[p.~573]{BurkLewiOver02a}; unlike in the Lipschitz case, this set may be unbounded, presenting obvious difficulties for approximating it through random sampling of gradients. In fact, more than half of \cite{BurkLewiOver02a} is devoted to investigating this issue, relying heavily on modern variational analysis as expounded in \cite{RockWets98}. Some positive results were obtained, specifically in the case that $f$ is ``directionally Lipschitz" at $\bar x$, which means that the ``horizon cone" \cite[p.~572]{BurkLewiOver02a} of $f$ at $\bar x$ is pointed, that is, it does not contain a line. For example, this excludes the function on $\R{}$ defined by $f(x)=|x|^{1/2}$ at $\bar x=0$, but it applies to the case $f(x)=(\max(0,x))^{1/2}$ even at $\bar x=0$. The discussion of the directionally Lipschitz case culminates with Corollary 6.1, which establishes that the Clarke subdifferential can indeed be approximated by convex hulls of gradients. On the more negative side, Example 7.2 shows that this approximation can fail badly in the general Lipschitz case. Motivated by these results, Burke and Lin have recently extended the GS convergence theory to the directionally Lipschitz case \cite{BurkLin17,Lin09}. However, it would seem difficult to extend these results to the more general non-Lipschitz case. Suppose $f:\R{n\times n}\rightarrow \R{}$ is defined by \[ f(X) = \max\{\mathrm{Re~} \lambda: \det(\lambda I - X) = 0\}, \] the spectral abscissa (maximum of the real parts of the eigenvalues) of $X$. Assume that $\bar X$ has the property that its only eigenvalues whose real parts coincide with $f$ make up a zero eigenvalue with multiplicity $q$ associated with a single Jordan block (the generic case). In this case the results in \cite{BurkOver01} tell us that the horizon cone of $f$ is pointed at $\bar X$ if and only if the multiplicity $q\leq 2$; on the other hand $f$ is locally Lipschitz at $X$ if and only if $q=1$. In the light of the previous paragraph, one might expect much greater practical success in applying GS to minimize the spectral abscissa of a parameterized matrix if the optimal multiplicities are limited to 1 or 2. However, this seems not to be the case. The results reported in \cite{BurkLewiOver02b} for unconstrained spectral abscissa minimization, as well as results for applying the algorithm of \cite{CurtOver12} (see \S\ref{subsec.constrained} below) for constrained nonsmooth, nonconvex optimization to problems with spectral radius objective and constraints, as reported in \cite[Sec.~4.2 \red{and Appendix~A.1}]{CurtMitcOver17}, do not show any marked deterioration as the optimal multiplicities increase from 2 or 3, although certainly the problems are much more challenging for larger multiplicities. We view understanding the rather remarkably good behavior of the GS algorithm on such examples as a potentially rewarding, though certainly challenging, line of investigation. \section{Enhancements}\label{sec.enhancements} As explained above, the statement of Algorithm~\ref{alg.gs} differs in several ways from the algorithms stated in \cite{BurkLewiOver02b}, \cite{BurkLewiOver05}, and \cite{Kiwi07}. Other variants of the strategy have also been proposed in recent years, in some cases to pose new solutions to theoretical obstacles (such as the need, in theory, to check for differentiability of $f$ at each new iterate), and in others to enhance the practical performance of the approach. In this section, we discuss a few of these enhancements. \subsection{Restricting the Line Search to within a Trust Region} Since the gradient information about $f$ is obtained only within the ball $\Bmbb(x^k,\epsilon_k)$ for all $k \in \N{}$, and since one might expect that smaller steps should be made when the sampling radius is small, an argument can be made that the algorithm might benefit by restricting the line search to within the ball $\Bmbb(x^k,\epsilon_k)$ for all $k \in \N{}$. In \cite[\S4.2]{Kiwi07}, such a variant is proposed where in place of $-g^k$ the search direction is set as $-\epsilon_k g^k/\|g^k\|_2$. With minor corresponding changes to conditions \eqref{eq.armijo} and \eqref{eq.update}, all of the theoretical convergence guarantees of the algorithm are maintained. Such a variant with the trust region radius defined as a positive multiple of the sampling radius $\epsilon_k$ for all $k \in \N{}$ would have similar properties. This variant might perform well in practice, especially in situations when otherwise setting the search direction as $-g^k$ would lead to significant effort being spent in the line search. \subsection{Avoiding the Differentiability Check}\label{subsec.avoidcheck} The largest distraction from the fundamentally simple nature of Algorithm~\ref{alg.gs} is the procedure for choosing a perturbed subsequent iterate if $f$ is not differentiable at $x^k - t_kg^k$; see Steps~\ref{step.check}--\ref{step.update}. This procedure is necessary for the algorithm to be well defined since, to ensure that $-g^k$ is a descent direction for all $k \in \N{}$, the algorithm relies on the existence of and ability to compute $-\nabla f(x^k)$ for all $k \in \N{}$. One might hope that this procedure, while necessary for theoretical convergence guarantees, could be ignored in practice. However, due to the deterministic nature of the line search, situations exist in which landing on a point of nondifferentiability of $f$ occurs with positive probability. For example, consider the function $f : \R{2} \to \R{}$ given by \bequationNN f(w,z) = \max\{ 0.5w^2 + 0.1z,w + 0.1z + 1, -w + 0.1z + 1, -0.05z - 50 \}; \eequationNN see Figure~\ref{lucas_example}. As shown by Helou, Santos, and Sim\~oes in \cite{HeloSantSimo16}, if Algorithm~\ref{alg.gs} is initialized at $x^0 = (w_0,z_0)$ chosen anywhere in the unit ball centered at $(10,10)$, then there is a positive probability that the function $f$ will not be differentiable at $x^0 - t_0g^0$. This can be explained as follows. At any point in the unit ball centered at $(10,10)$, the function $f$ is continuously differentiable and $\nabla f(x^0) = [w_0;\ 0.1]^T$. Moreover, there is a positive probability that the sampled points obtained at this first iteration will yield $g^0 = \nabla f(x^0)$. Therefore, given a reasonable value for the parameter $\beta$ that appears in~\eqref{eq.armijo} (e.g., $\beta = 10^{-4}$), the sufficient decrease of the function value is attained with $t_0 = 1$. This guarantees that the function $f$ will not be differentiable at the next iterate, since the first coordinate of $x^1 = (w_1,z_1)$ will be zero. \begin{figure}[ht] \centering \includegraphics[height=2.5in]{lucas_example} \caption{Contours of a function illustrating the necessity of the differentiability check in Algorithm~\ref{alg.gs}. Initialized uniformly within the illustrated ball, there is a positive probability that $x^0 - t_0g^0 \not\in D$.} \label{lucas_example} \end{figure} The authors of \cite{HeloSantSimo16} propose two strategies to avoid the issues highlighted by this example. The first is that, rather than perturb the iterate after the line search, one could perturb the search direction before the line search. It is shown that if the random perturbation of the search direction is sufficiently small such that, for one thing, the resulting direction is still one of descent for $f$, then $f$ will be differentiable at all iterates with probability one. Their second proposed strategy involves the use of a nonmonotone line search. In particular, it is shown that if a strictly positive value $\Delta_k$ is added on the right-hand side of the sufficient decrease condition in \eqref{eq.armijo} such that $\{\Delta_k\}$ is summable, then one can remove $\nabla f(x^k)$ from the set $\Gcal_k$ for all $k \in \N{}$ and maintain convergence guarantees even when $f$ is \emph{not} differentiable at $\{x^k\}$. This can be shown by noticing that, due to the positive term $\Delta_k$, the line search procedure continues to be well defined even if~$-g^k$ is not a descent direction for $f$ at $x^k$ (which may happen since $\nabla f(x_k)$ is no longer involved in the computation of~$g^k$). However, the summability of $\{\Delta_k\}$ implies that the possible increases in~$f$ will be finite and, when the sampled points are good enough to describe the local behavior of $f$ around the current iterate, the function value will necessarily decrease if the method has not reached (approximate) stationarity. Overall, the sufficient reductions in $f$ achieved in certain iterations will ultimately exceed any increases. Another proposal for avoiding the need to have $f$ differentiable at $x^k$ for all $k \in \N{}$ is given in \cite[\S4.3]{Kiwi07}, wherein a technique is proposed for using a limited line search, potentially causing the algorithm to take null steps in some iterations. In fact, this idea of using a limited line search can also be used to avoid the need to sample a new set of $m \geq n+1$ gradients in each iteration, as we discuss next. \subsection{Adaptive Sampling}\label{sec.adaptive} As previously mentioned, the main weakness of Algorithm~\ref{alg.gs} is the cost of computing $m \geq n+1$ gradients in every iteration and solving a corresponding QP. Loosely speaking, this lower bound on the number of gradients is required in the analysis of the algorithm so that one can use Carath\'eodory's theorem to argue that, with at least $n+1$ gradients, there is a sufficiently good chance that the combination vector $-g^k$ represents a sufficiently good direction from $x^k$. However, in many situations in practice, the sampling of such a large number of gradients in each iteration can lead to a significant amount of wasted computational effort. One can instead sample \emph{adaptively}, attempting to search along directions computed using fewer gradients and proceeding as long as a sufficient reduction is attained. In \cite{CurtQue13}, Curtis and Que show how such an adaptive sampling strategy can be employed so that the convergence guarantees of Algorithm~\ref{alg.gs} are maintained while only a constant number (independent of $n$) of gradients need to be sampled in each iteration. A key aspect that allows one to maintain these guarantees is the employment of a limited line search, as first proposed in \cite{Kiwi07}, potentially leading to a null step when fewer than $n+1$ gradients are currently in hand and when the line search is not successful after a prescribed finite number of function evaluations. See also \cite{CurtQue15} for further development of these ideas, where it is shown that one might not need to sample \emph{any} gradients as long as a sufficient reduction is attained. The work in \cite{CurtQue13} also introduces the idea that, when adaptive sampling is employed, the algorithm can exploit a practical feature commonly used in bundle methods. This idea relates to warm-starting the algorithm for solving the QP subproblems. In particular, suppose that one has solved the primal-dual pair of QPs in \eqref{prob.qp} for some $m \in \N{}$ to obtain $(z_k,d^k,\lambda^k)$, yielding $g^k = -d^k$. If one were to subsequently aim to solve the pair of QPs corresponding to the augmented matrix of gradients \bequationNN \overline{G}_k = \bbmatrix G_k & \nabla f(x^{k,m+1}) & \cdots & \nabla f(x^{k,m+p}) \ebmatrix, \eequationNN then one obtains a viable feasible starting point for the latter QP in \eqref{prob.qp} by augmenting the vector $\lambda^k$ with~$p$ zeros. This can be exploited, e.g., in an active-set method for solving this QP; see \cite{Kiwi86}. As a further practical enhancement, the work in \cite{CurtQue13,CurtQue15} also proposes the natural idea that, after moving to a new iterate $x^k$, gradients computed in previous iterations can be ``re-used'' if they correspond to points that lie within $\Bmbb(x^k,\epsilon_k)$. This may further reduce the number of sampled points needed in practice. \subsection{Second-Order-Type Variants} The solution vector $d^k$ of the QP in \eqref{prob.qp} can be viewed as the minimizer of the model of $f$ at $x^k$ given by \bequationNN q_k(d) = f(x^k) + \max_{g\in\Gcal_k}\ g^Td + \thalf d^TH_kd \eequationNN with $H_k=I$, the identity matrix. As in other second-order-type methods for nonlinear optimization, one might also consider algorithm variants where $H_k$ is set to some other symmetric positive definite matrix. Ideas of this type have been explored in the literature. For example, in \cite{CurtQue13}, two techniques are proposed: one in which $H_k$ is set using a quasi-Newton updating strategy and one in which the matrix is set in an attempt to ensure that the model $q_k$ represents an upper bounding model for $f$. The idea of employing a quasi-Newton approach, inspired by the success of quasi-Newton methods in practice for nonsmooth optimization (see \cite{LewiOver13}), has also been explored further in \cite{CurtQue15,CurtRobiZhou17}. Another approach, motivated by the encouraging results obtained when employing spectral gradient methods to solve smooth~\cite{Flet05,BirgMartRayd14} and nonsmooth~\cite{CremLoreRayd07} optimization problems, has been to employ a Barzilai-Borwein (BB) strategy for computing initial stepsizes in a GS approach; see \cite{LoreAponCoreRayd17} and the background in~\cite{BarzBorw88,Rayd93,Rayd97}. Using a BB strategy can be viewed as choosing $H_k = \alpha_k I$ for all $k \in \N{}$ where the scalar $\alpha_k$ is set according to iterate and gradient displacements in the latest iteration. In all of these second-order-type approaches, one is able to maintain convergence guarantees of the algorithm as long as the procedure for setting the matrix $H_k$ is safeguarded during the optimization process. For example, one way to maintain guarantees is to restrict each $H_k$ to the set of symmetric matrices whose eigenvalues are contained within a fixed positive interval. One might also attempt to exploit the self-correcting properties of BFGS updating; see \cite{CurtRobiZhou17}. \section{Extensions}\label{sec.extensions} In this section, we discuss extensions to the GS methodology to solve classes of problems beyond unconstrained optimization on $\R{n}$. \subsection{Riemannian GS for Optimization on Manifolds} Hosseini and Uschmajew in \cite{HossUsch17} have extended the GS methodology for minimizing a locally Lipschitz $f$ over a set $\Mcal$, where $\Mcal$ is a complete Riemannian manifold of dimension $n$. The main idea of this extension is to employ the convex hull of gradients from tangent spaces at randomly sampled points \emph{transported} to the tangent space of the current iterate. In this manner, the algorithm can be characterized as a generalization of the Riemannian steepest descent method just as GS is a generalization of traditional steepest descent. Assuming that the vector transport satisfies certain assumptions, including a \emph{locking condition}, the algorithm attains convergence guarantees on par with those for Algorithm~\ref{alg.gs}. \subsection{SQP-GS for Constrained Optimization} \label{subsec.constrained} Curtis and Overton in \cite{CurtOver12} proposed a combination sequential quadratic programming (SQP) and gradient sampling method for solving constrained optimization problems involving potentially nonsmooth constraint functions, i.e., \bequation\label{eq.huatulco} \min_{x\in\R{n}}\ f(x)\ \ \st\ \ c(x) \leq 0, \eequation where $f : \R{n} \to \R{}$ and $c : \R{n} \to \R{n_c}$ are locally Lipschitz. A key aspect of the proposed SQP-GS approach is that sampled points for the objective and each individual constraint function are generated independently. With this important feature, it is shown that the algorithm, which follows a penalty-SQP strategy (e.g., see~\cite{Flet87}), attains convergence guarantees for minimizing an exact penalty function that are similar to those in \S\ref{sec.global}. Moreover, with the algorithm's penalty parameter updating strategy, it is shown that either the penalty function is driven to $-\infty$, the penalty parameter settles at a finite value and any limit point will be feasible for the constraints and stationary for the penalty function, or the penalty parameter will be driven to zero and any limit point of the algorithm will be stationary for a constraint violation measure. As for other exact penalty function methods for nonlinear optimization, one can translate between guarantees for minimizing the exact penalty function and solving the constrained problem~\eqref{eq.huatulco}; in particular, if the problem is \emph{calm} \cite{Clar83,Rock82}, then at any local minimizer~$x_*$ of~\eqref{eq.huatulco} there exists a threshold for the penalty parameter beyond which $x_*$ will be a local minimizer of the penalty function. Tang, Liu, Jian, and Li have also proposed in \cite{TangLiuJianLi14} a \emph{feasible} variant of the SQP-GS method in which the iterates are forced to remain feasible for the constraints and the objective function is monotonically decreasing throughout the optimization process. This opens the door to employing a two-phase approach common for solving some optimization problems, where phase 1 is responsible for attaining a feasible point and phase 2 seeks optimality while maintaining feasibility. \subsection{Derivative-Free Optimization} Given its simple nature, gradient sampling has proved to be an attractive basis for the design of new algorithms even when gradient information cannot be computed explicitly. Indeed, there have been a few variants of \emph{derivative-free} algorithms that have been inspired by gradient sampling. The first algorithm for derivative-free optimization inspired by gradient sampling was proposed by Kiwiel in~\cite{Kiwi10}. In short, in place of the gradients appearing in Algorithm~\ref{alg.gs}, this approach employs Gupal's estimates of gradients of the Steklov averages of $f$. In this manner, function values only---specifically, $\Ocal(mn)$ per iteration---are required for convergence guarantees. A less expensive \emph{incremental} version is also proposed. Another derivative-free variant of GS, proposed by Hare and Nutini in \cite{HareNuti13}, is specifically designed for minimizing finite-max functions. This approach exploits knowledge about which of these functions are \emph{almost active}---in terms of having value close to the objective function---at a particular point. In so doing, rather than attempt to approximate gradients at nearby points, as in done in \cite{Kiwi10}, this approach only attempts to approximate gradients of almost active functions. The convergence guarantees proved for the algorithm are similar to those for GS methods, though the practical performance is improved by the algorithm's tailored gradient approximation strategy. Finally, we mention the \emph{manifold sampling} algorithm, proposed by Larson, Menickelly, and Wild in \cite{LarsMeniWild16}, for solving nonconvex problems where the objective function is the $\ell_1$-norm of a smooth vector function $F : \R{n} \to \R{r}$. While this approach does not employ a straightforward GS methodology in that it does not randomly sample points, it does employ a GS-type approach in the way that the gradient of a model of the objective function is constructed by solving a QP of the type in \eqref{prob.qp}. Random sampling can be avoided in this construction since the algorithm can exploit knowledge of the signs of the elements of $F(x)$ at any $x \in \R{n}$ along with knowledge of ${\bar\partial} \|\cdot\|_1$. \section{Applications}\label{sec.applications} We mentioned in the introduction that the original gradient sampling paper \cite{BurkLewiOver02b} reported results for spectral abscissa optimization problems that had not been solved previously. The second gradient sampling paper \cite{BurkLewiOver05} reported results for many more applications that again had not been solved previously: these included Chebyshev approximation by exponential sums, eigenvalue product minimization for symmetric matrices, spectral and pseudospectral abscissa minimization, maximization of the ``distance to instability", and fixed-order controller design by static output feedback. Subsequently, the GS algorithm played a key role in the {\sc hanso}\ (Hybrid Algorithm for Non-Smooth Optimization)\footnote{www.cs.nyu.edu/overton/software/hanso/} and {\sc hifoo}\ (H-Infinity Fixed-Order Optimization)\footnote{www.cs.nyu.edu/overton/software/hifoo/} toolboxes. The former is a stand-alone code for unconstrained nonsmooth optimization while the latter is a more specialized code used for the design of low-order controllers for linear dynamical systems with input and output, computing fixed-order controllers by optimizing stability measures that are generally nonsmooth at local minimizers \cite{BurkHenrLewiOver06}. {\sc Hifoo}\ calls {\sc hanso}\ to carry out the optimization. The use of ``hybrid" in the expansion of the {\sc hanso}\ acronym indicated that, from its inception, {\sc hanso}\ combined the use of both a quasi-Newton algorithm (BFGS) and Gradient Sampling. The quasi-Newton method was used in an initial phase which, rather surprisingly, typically worked very effectively even in the presence of nonsmoothness, very often providing a fast way to approximate a local minimizer. This was followed by a GS phase to refine the approximation, typically verifying a loose measure of local optimality. The {\sc hifoo}\ toolbox has been used successfully in a wide variety of applications, including synchronization of heterogeneous multi-agent systems and networks, design of motorized gimbals that stabilize an angular motion of an optical payload around an axis, flight control via static output feedback, robust observer-based fault detection and isolation, influence of tire damping on control of quarter-car suspensions, flexible aircraft lateral flight dynamic control, optimal control of aircraft with a blended wing body, vibration control of a fluid/plate system, controller design of a nose landing gear steering system, bilateral teleoperation for minimally invasive surgery, design of an aircraft controller for improved gust alleviation and passenger comfort, robust controller design for a proton exchange membrane fuel cell system, design of power systems controllers, and design of winding systems for elastic web materials --- for a full list of references, see \cite{CurtMitcOver17}. The successful use of BFGS in {\sc hanso}\ and {\sc hifoo}\ led to papers on the use of quasi-Newton methods in the nonsmooth context, both for unconstrained \cite{LewiOver13} and constrained \cite{CurtMitcOver17} optimization. The latter paper introduced a new BFGS-SQP method for nonsmooth constrained optimization and compared it with the SQP-GS method discussed in \S~\ref{subsec.constrained} on a suite of challenging static output feedback controller design problems, half of them non-Lipschitz (spectral radius minimization) and half of them locally Lipschitz (pseudospectral radius minimization). It was found that although the BFGS-SQP method was much faster than SQP-GS, nonetheless, if the latter method based on Gradient Sampling was allowed sufficient running time, it frequently found better approximate solutions than the former method based on BFGS in a well defined sense, evaluated using ``relative minimization profiles". Interestingly, this was particularly pronounced on the non-Lipschitz problems, despite the fact that the GS convergence theory does not extend to this domain. See \S\ref{sec.nonlip} for further discussion of this issue. Finally, we mention an interesting application of GS to robot path planning \cite{TrafMitc16}. This work is based on the observation that shortest paths generated through gradient descent on a value function have a tendency to chatter and/or require an unreasonable number of steps to synthesize. The authors demonstrate that the GS algorithm can largely alleviate this problem. For systems subject to state uncertainty whose state estimate is tracked using a particle filter, they proposed the Gradient Sampling with Particle Filter (GSPF) algorithm, which uses the particles as the locations in which to sample the gradient. At each step, the GSPF efficiently finds a consensus direction suitable for all particles or identifies the type of stationary point on which it is stuck. If the stationary point is a minimum, the system has reached its goal (to within the limits of the state uncertainty) and the algorithm terminates; otherwise, the authors propose two approaches to find a suitable descent direction. They illustrated the effectiveness of the GSPF on several examples using well known robot simulation environments. This work was recently extended and modified in \cite{EstrMitc18}, where the practical effectiveness of both the GSPF algorithm and the new modification was demonstrated on a Segway Robotic Mobility Platform. \section{Conclusion and Future Directions}\label{sec.conclusion} Gradient sampling is a conceptually straightforward approximate steepest descent method. With a solid convergence theory, the method has blossomed into a powerful methodology for solving nonsmooth minimization problems. The theme of our treatment of GS in this work has been to emphasize the fact that, even though the basic algorithm has been enhanced and extended in various ways, the foundation of the approach is fundamentally simple in nature. We have also corrected an oversight in the original GS theory (i.e., that the convergence results depend on assuming that the set of points over which the Lipschitz function $f$ is continuously differentiable has full measure, although we do not have a counterexample to convergence of GS in the absence of this assumption). At the same time we have loosened the requirements of the algorithm (showing that $f$ need only be differentiable at the iterates and sampled points). An open question that still remains is whether one can extend the GS theory to broader function classes, such as the case where $f$ is assumed to be semi-algebraic but not necessarily locally Lipschitz or directionally Lipschitz. Opportunities for extending GS theory for broader function classes may include connecting the algorithm to other randomized/stochastic optimization methods. For example, one might view the algorithm as a stochastic-gradient-like method applied to a smoothed objective. (A similar philosophy underlies the analysis by Nesterov and Spokoiny in \cite{NestSpok17}).) More precisely, given a locally Lipschitz objective $f$, consider a smoothing~$f_\epsilon$ whose value at any point $x$ is given by the mean value of $f$ over the ball $\Bmbb(x,\epsilon)$. The GS algorithm uses gradients of $f$ at uniformly distributed random points in this ball. Notice that each such gradient can also be viewed as a stochastic gradient for the smoothing $f_\epsilon$ in the sense that its expectation is the gradient of $f_\epsilon$ at $x$. Thus, one might hope to prove convergence results for a GS algorithm (with predetermined stepsizes rather than line searches) that parallel convergence theory for stochastic gradient methods. Recent work by Davis, Drusvyatskiy, Kakade and Lee \cite{DaviDrusKakaLee18} gives convergence results for stochastic \emph{subgradient} methods on a broad class of problems. Another potentially interesting connection is with the work of Davis and Drusvyatskiy \cite{DaviDrus18} on stochastic model-based optimization. Consider a GS variant that successively minimizes stochastic models of the objective function $f$, where we assume for simplicity that $f$ is a globally Lipschitz convex function. In this variant, rather than moving along the direction $-g^k$, consider instead the construction of a cutting plane approximation of $f$ from its affine minorants at the current iterate $x^k$ and the sampled points $\{x^{k,i}\}$, augmented by the proximal term $\beta_k \|x - x^k\|^2$, where $\{\beta_k\}$ is a predetermined sequence. Suppose that the next iterate is chosen as the minimizer of this model; for a given $k$ and with $\beta_k = 1$, by equation \eqref{prob.qp}, this scheme and GS produce similar descent directions as the sampling radius tends to zero. It follows from the results of \cite{DaviDrus18} that the expected norm of the gradient of the Moreau envelope \cite[p.~6]{DaviDrus18} is reduced below $\epsilon$ in $\Ocal(\epsilon^{-4})$ iterations. In fact, the assumptions on $f$ in \cite{DaviDrus18} are substantially weaker than convexity, and do not require any property of the set on which $f$ is continuously differentiable. Connecting the convergence theory for GS to stochastic methods as suggested in the previous two paragraphs could be enlightening. However, while stochastic methods are often designed for settings in which it is intractable to compute function values exactly---a feature reflected in the fact that the analyses for such methods are based on using predetermined stepsize sequences---the GS methodology has so far been motivated by problems for which functions and gradients are tractable to compute. In such settings, the line search in Algorithm~GS is an ingredient that is crucial to its practical success.

https://arxiv.org/abs/2205.04734

Brezzi--Douglas--Marini interpolation on anisotropic simplices and prisms

The Brezzi--Douglas--Marini interpolation error on anisotropic elements has been analyzed in two recent publications, the first focusing on simplices with estimates in $L^2$, the other considering parallelotopes with estimates in terms of $L^p$-norms. This contribution provides generalized estimates for anisotropic simplices for the $L^p$ case, $1\leq p\leq\infty$, and shows new estimates for anisotropic prisms with triangular base.

\section{Introduction} The Brezzi--Douglas--Marini (BDM) finite element \cite{Nedelec1986} was introduced to approximate $\vec{H}_{\mathrm{div}}$ by polynomials. This proves useful for problems of incompressible fluid flow, where recent approaches employ $\vec{H}_{\mathrm{div}}$-conforming discretizations to approximate the velocity solution. The corresponding interpolation operator can also be used as reconstruction operator to gain pressure-robust methods in the spirit of \cite{Linke2014}. Boundary layers or edge singularities in these problems require the use of anisotropic, i.e., highly stretched elements, so interpolation error estimates for such settings are required, see \cite{ApelKempf2021}. In \cite{ApelKempf2020}, these estimates were shown for simplicial elements in terms of the $\vec{L}^2$-norm, and \cite{Franz2021} contains estimates for anisotropic parallelotopes in terms of $\vec{L}^p$-norms. The main focus in \cite{Franz2021} is on estimates in the $\vec{H}_{\mathrm{div}}$-norm and while the technique for these proofs can also be used for simplices, it is not applicable for the case of prisms as defined in \cite{Nedelec1986}, as here the commuting diagram property \cite[(3)]{Franz2021} is not satisfied. Interestingly, it is neither satisfied on cubes as defined in \cite{Nedelec1986}; here \cite{Franz2021} uses the original definition of the BDM elements, see, e.g., \cite{BoffiBrezziFortin2013}. In the following sections we provide a generalization of the results from \cite{ApelKempf2020} to $\vec{L}^p$ spaces, $1\leq p \leq \infty$, and show new interpolation error estimates for anisotropic triangular prisms. The results with detailed proofs are contained in the author's PhD thesis \cite{Kempf2022:PhD} that is to appear. \textsc{Notation:} Vectors, vector valued functions and the spaces of such functions are set in bold and we use $I_n = \{1,\ldots,n\}$ for index sets. The spatial dimension is denoted by $d$ and the expression $a \lesssim b$ means there is a positive constant $C$ so that $a \leq C b$. The norms of the Sobolev spaces $W^{m,p}(G)$ are denoted by $\norm{\cdot}_{m,p,G}$, multi-indices by $\vec{\alpha}$. The directional derivative in direction $\vec{l}$ is denoted by $\pdv{}{\vec{l}}$. \section{Error estimates on anisotropic simplices} We state the anisotropic interpolation error estimates on simplices in terms of $\vec{L}^p$-norms, which is a generalization of the results from \cite{ApelKempf2020}, where only $p=2$ was considered. As the proofs are largely analogous to those in \cite{ApelKempf2020}, we omit them here. The principal idea of the proof is to first show stability estimates on the reference elements and then transfer them in two steps first to an element of a reference family using an affine transformation with a diagonal matrix and then to the general element. Recall that a simplex is said to satisfy the \emph{maximum angle condition} if all angles within and between facets are bounded by a constant $\bar{\phi}<\pi$. In addition, it satisfies the \emph{regular vertex property} if for one vertex there is a constant $\bar{c}>0$ so that $\abs{\det N} \geq \bar{c}$ holds for the matrix $N$ whose columns consist of the unit vectors $\vec{l}_i$, $i\in I_d$, along the outgoing edges from this vertex. The lengths of the edges corresponding to the vectors $\vec{l}_i$ are the element size parameters and are denoted by $h_i$. The degrees of freedom for the BDM interpolation operator $\BDMIk{k}$ of order $k$ on simplices can be found in \cite[p. 59]{Nedelec1986}. The error estimates for the BDM interpolation depending on the geometric regularity are given by the two following theorems, cf. \cite[Theorems 4.3, 4.4]{ApelKempf2020} and \cite[Theorems 6.2, 6.3]{AcostaApelDuranLombardi2011}. \begin{theorem} Let a simplicial element $T$ satisfy the regular vertex property with constant $\bar{c}$. Then for $k\geq 1$, $0\leq m\leq k$ and $\vec{v}\in \vec{W}^{m+1,p}(T)$, $1\leq p\leq\infty$, the estimate \begin{equation*} \norm{\vec{v}-\BDMIk{k}\vec{v}}_{0,p,T} \lesssim \sum_{\abs{\vec{\alpha}}=m+1} h^{\vec{\alpha}} \norm{D^{\vec{\alpha}}_{\vec{l}}\vec{v}}_{0,p,T} + h_T \sum_{\abs{\vec{\alpha}}=m} h^{\vec{\alpha}} \norm{D^{\vec{\alpha}}_{\vec{l}} \div\vec{v}}_{0,p,T} \end{equation*} holds, the constant only depends on $\bar{c}$ and $k$. Here $h_T={\operatorname{diam}\,} T$, $h^{\vec{\alpha}} = \prod_{j\in I_d}h_j^{\alpha_j}$ and $D_{\vec{l}}^{\vec{\alpha}} = \frac{\partial^{\abs{\vec{\alpha}}}}{\partial_{\vec{l}_1}^{\alpha_1} \cdots \partial_{\vec{l}_d}^{\alpha_d}}$. \end{theorem} \begin{theorem} Let a simplicial element $T$ satisfy the maximum angle condition with constant $\bar{\phi}$. Then for $k\geq 1$, $0\leq m \leq k$ and $\vec{v}\in \vec{W}^{m+1,p}(T)$, $1\leq p\leq\infty$, the estimate \begin{equation*} \norm{\vec{v}-\BDMIk{k}\vec{v}}_{0,p,T} \lesssim h_T^{m+1} \norm{D^{m+1} \vec{v}}_{0,p,T} \end{equation*} holds, the constant only depends on $\bar{\phi}$ and $k$. The notation $D^n$ means the sum of the absolute values of all derivatives of order $n$. \end{theorem} \section{Error estimates on anisotropic prisms} Similar results can be achieved for triangular prisms with some modifications. The prismatic reference element $\hat{P}$ with the notation for the vertices is given on the left hand side of \Cref{fig:prism_element}. \begin{figure}[t] \input{Figures/src/Prism_reference_element.tikz} \hfill \begin{tikzpicture} \draw[white] (0,0) -- (1,0); \draw[-stealth] (0,2) -- node[above] {$J_{\tilde{P}}$} (2,2); \end{tikzpicture} \hfill \input{Figures/src/Prism_element.tikz} \caption{Reference prism $\hat{P}$ with vertex numbering, and a transformed prism of the reference family $\mathcal{R}_P$.}\label{fig:prism_element} \end{figure} The element $\hat{P}$ is transformed to an element $\tilde{P}$ of the reference family $\mathcal{R}_P$ by the transformation $\tilde{\vec{x}} = J_{\tilde{P}} \hat{\vec{x}}$, where \begin{equation}\label{eq:trafo} J_{\tilde{P}} = \begin{pmatrix}h_1&0&0\\0&h_2&0\\0&0&h_3\end{pmatrix}, \end{equation} see \Cref{fig:prism_element}. The vertical facet of $\hat{P}$ opposite of the vertices $\vec{p}_i$ and $\vec{p}^i$ is denoted by $e_i$, the horizontal facet at $\hat{x}_3 = 0$ by $e_b$ and the one at $\hat{x}_3 = 1$ by $e_t$. The facet normals of $\hat{P}$ are thus given by \begin{equation*} \vec{n}_{\hat{P},e_1} = \begin{pmatrix}-1\\0\\0\end{pmatrix}, \quad \vec{n}_{\hat{P},e_2} = \begin{pmatrix}0\\-1\\0\end{pmatrix}, \quad \vec{n}_{\hat{P},e_3} = \frac{1}{\sqrt{2}}\begin{pmatrix}1\\1\\0\end{pmatrix}, \quad \vec{n}_{\hat{P},e_b} = \begin{pmatrix}0\\0\\-1\end{pmatrix}, \quad \vec{n}_{\hat{P},e_t} = \begin{pmatrix}0\\0\\1\end{pmatrix}. \end{equation*} The local BDM interpolation operator $\BDMIk{k}$ of order $k$ on a prism $P$ maps into the space $\Polyvec{k,k}{P}$ of vector valued polynomials of total degree $k$ in $x_1$ and $x_2$ and degree $k$ in $x_3$. It is defined by the functionals, see \cite[p. 64]{Nedelec1986}, \begin{subequations}\label{eq:interpolation_relations} \begin{align} \int_{e_i} \BDMIk{k}\vec{v} \cdot \vec{n}_{P,e_i} z \dd{\vec{s}} &= \int_{e_i} \vec{v} \cdot \vec{n}_{P,e_i} z \dd{\vec{s}} &&\forall z\in \Poly{k}{e_i}, \quad i\in \{b,t\}, \\ \int_{e_i} \BDMIk{k}\vec{v} \cdot \vec{n}_{P,e_i} z \dd{\vec{s}} &= \int_{e_i} \vec{v} \cdot \vec{n}_{P,e_i} z \dd{\vec{s}} &&\forall z\in Q_k(e_i), \quad i\in I_3, \\ \int_P (\BDMIk{k}\vec{v})_3 z_3 \dd{\vec{x}} &= \int_P v_3 z_3 \dd{\vec{x}} &&\forall z_3\in \Poly{k,k-2}{P}, \\ \int_P (\BDMIk{k}\vec{v})_1 z_1 + (\BDMIk{k}\vec{v})_2 z_2 \dd{\vec{x}} &= \int_P v_1 z_1 + v_2 z_2 \dd{\vec{x}} &&\forall (z_1,z_2)\in \mathcal{P}_{k-1,k}(P), \end{align} \end{subequations} where $P_k(e)$ is the space of polynomials with maximal degree $k$, $Q_k(e)$ is the space of polynomials with degree $k$ in each of the dimensions of the facet $e$, $\Poly{m,n}{P}$ is the space of polynomials with total degree $m$ in $x_1$ and $x_2$, and degree $n$ in $x_3$. The space $\mathcal{P}_{m,n}(P)$ consists of pairs of polynomials with degree $n$ in $x_3$, which are for fixed $x_3$, i.e., on the triangle $T_{x_3}$, in the space $\vec{N}_{m}(T_{x_3}) = \vec{P}_{m-1}(T_{x_3}) \oplus \vec{S}_{m}(T_{x_3})$, where $\vec{S}_{m}(T_{x_3}) = \{\vec{p}\in \vec{P}_{m}(T_{x_3}):\vec{p}(\vec{x})\cdot \vec{x}=0\ \forall\vec{x}\in T_{x_3}\}$. We first require the analog of \cite[Lemma 3.1]{ApelKempf2020} for the prism reference element. The proof is omitted here for brevity since it follows along the same steps as the proof for the analogous lemma for simplices. \begin{lemma}\label{lem:BDMI_technical_prism} Let $\hat{P}$ be the reference element from \Cref{fig:prism_element}, $\hat{f}_j\in L^p(e_j)$, $j\in I_2$, $\hat{f}_3\in L^p(e_b)$, $1\leq p\leq\infty$, and \begin{align*} &\hat{\vec{u}}(\hat{\vec{x}}) = \begin{pmatrix} \hat{f}_1(\hat{x}_2,\hat{x}_3),0,0\end{pmatrix}^T,&&\hat{\vec{v}}(\hat{\vec{x}}) = \begin{pmatrix}0, \hat{f}_2(\hat{x}_1,\hat{x}_3),0\end{pmatrix}^T,&&\hat{\vec{w}}(\hat{\vec{x}}) = \begin{pmatrix}0,0,\hat{f}_3(\hat{x}_1,\hat{x}_2)\end{pmatrix}^T. \end{align*} Then there are functions $\hat{q}_j\in P_k(e_j)$, $j\in I_2$, $\hat{q}_3\in P_k(e_b)$, so that \begin{align*} &\hBDMIk{k}\hat{\vec{u}} = \begin{pmatrix} \hat{q}_1(\hat{x}_2,\hat{x}_3),0,0\end{pmatrix}^T,&&\hBDMIk{k}\hat{\vec{v}} = \begin{pmatrix}0,\hat{q}_2(\hat{x}_1,\hat{x}_3),0\end{pmatrix}^T,&& \hBDMIk{k}\hat{\vec{w}} = \begin{pmatrix}0,0,\hat{q}_3(\hat{x}_1,\hat{x}_2)\end{pmatrix}^T. \end{align*} \end{lemma} With this lemma, we get the stability estimate on the reference element. \begin{lemma}\label{lem:stability_BDM_reference_element_prism} Let $\hat{\vec{u}}\in\vec{W}^{1,p}(\hat{P})$, $1\leq p\leq\infty$. Then we have the estimates \begin{subequations} \begin{align*} \norm{(\hBDMIk{k} \hat{\vec{u}})_i}_{0,p,\hat{P}} &\lesssim \norm{\hat{u}_i}_{1,p,\hat{P}} + \norm{{\widehat{\operatorname{div}}\,}\hat{\vec{u}}}_{0,p,\hat{P}} + \norm{\pdv{\hat{u}_3}{\hat{x}_3}}_{0,p,\hat{P}}, \qquad i\in I_2,\\ \norm{(\hBDMIk{k} \hat{\vec{u}})_3}_{0,p,\hat{P}} &\lesssim \norm{\hat{u}_3}_{1,p,\hat{P}} + \norm{{\widehat{\operatorname{div}}\,}\hat{\vec{u}}}_{0,p,\hat{P}}. \end{align*} \end{subequations} \end{lemma} \begin{proof} The proofs for $i\in I_2$ are analogous so we only show the details for the first and third components, starting with $i=1$. Let \begin{align*} &\hat{\vec{u}}_* = \begin{pmatrix}0\\\hat{u}_2(\hat{x}_1,0,\hat{x}_3)\\\hat{u}_3(\hat{x}_1,\hat{x}_2,0)\end{pmatrix}, &&\hat{\vec{v}}=\hat{\vec{u}}-\hat{\vec{u}}_* = \begin{pmatrix}\hat{u}_1\\\hat{u}_2-\hat{u}_2(\hat{x}_1,0,\hat{x}_3)\\\hat{u}_3-\hat{u}_3(\hat{x}_1,\hat{x}_2,0)\end{pmatrix}. \end{align*} The previous \Cref{lem:BDMI_technical_prism} thus yields $(\hBDMIk{k} \hat{\vec{v}})_1 = (\hBDMIk{k}\hat{\vec{u}})_1$, and it holds ${\widehat{\operatorname{div}}\,}\hat{\vec{v}} = {\widehat{\operatorname{div}}\,}\hat{\vec{u}} - {\widehat{\operatorname{div}}\,}\hat{\vec{u}}_* = {\widehat{\operatorname{div}}\,}\hat{\vec{u}}$. We define the two functions \begin{align*} &\hat{\vec{v}}_* = \begin{pmatrix}0\\\hat{x}_2 \hat{q}_2\\\hat{x}_3\hat{q}_3\end{pmatrix}, &&\hat{\vec{w}}=\hat{\vec{v}}-\hat{\vec{v}}_* = \begin{pmatrix}\hat{v}_1\\\hat{v}_2-\hat{x}_2\hat{q}_2\\\hat{v}_3-\hat{x}_3\hat{q}_3\end{pmatrix}, \end{align*} where $\hat{q}_2\in \Poly{k-1,k}{\hat{P}}$, $\hat{q}_3\in \Poly{k,k-1}{\hat{P}}$, so that \begin{align*} \int_{\hat{P}} \hat{w}_2 z \dd{\vec{x}}=\int_{\hat{P}} (\hat{v}_2 - \hat{x}_2\hat{q}_2)z \dd{\vec{x}}=0 \qquad \forall z\in \Poly{k-1,k}{\hat{P}}, \\ \int_{\hat{P}} \hat{w}_3 z \dd{\vec{x}}=\int_{\hat{P}} (\hat{v}_3 - \hat{x}_3\hat{q}_3)z \dd{\vec{x}}=0 \qquad \forall z\in \Poly{k,k-1}{\hat{P}}. \end{align*} This means that the functions $\hat{q}_2$ and $\hat{q}_3$ are well defined. Since $\hat{\vec{v}}_*\in\Polyvec{k,k}{\hat{P}}$ it follows that $\hBDMIk{k}\hat{\vec{v}}_* = \hat{\vec{v}}_*$ and thus $(\hBDMIk{k} \hat{\vec{w}})_1 = (\hBDMIk{k} \hat{\vec{v}})_1 = (\hBDMIk{k} \hat{\vec{u}})_1$. The interpolated function $\hBDMIk{k} \hat{\vec{w}} = \hat{\vec{t}} = (\hat{t}_1,\hat{t}_2,\hat{t}_3)^T$ is then defined by the relations, see \eqref{eq:interpolation_relations}, \begingroup \allowdisplaybreaks \begin{align*} \int_{e_b} \hat{t}_3 z \dd{\vec{s}} &= \int_{e_b} \hat{w}_3 z \dd{\vec{s}} = 0 && \forall z\in \Poly{k}{e_b}, \\ \int_{e_t} \hat{t}_3 z \dd{\vec{s}} &= \int_{e_t} \hat{w}_3 z \dd{\vec{s}} = \int_{e_t} \hat{w}_3 z \dd{\vec{s}} - \int_{e_b} \hat{w}_3 z \dd{\vec{s}} = \int_{\hat{P}} \pdv{\hat{w}_3}{\hat{x}_3} z \dd{\vec{x}} && \forall z\in \Poly{k}{e_t}, \\ \int_{e_1} \hat{t}_1 z \dd{\vec{s}} &= \int_{e_1} \hat{w}_1 z \dd{\vec{s}} && \forall z \in Q_k(e_1),\\ \int_{e_2} \hat{t}_2 z \dd{\vec{s}} &= \int_{e_2} \hat{w}_2 z \dd{\vec{s}} = 0 && \forall z \in Q_k(e_2), \\ \int_{e_3} (\hat{t}_1 + \hat{t}_2) z \dd{\vec{s}} &= \int_{e_3} (\hat{w}_1 + \hat{w}_2) z \dd{\vec{s}} && \forall z \in Q_k(e_3), \\ \int_{\hat{P}} \hat{t}_3 z_3 \dd{\vec{x}} &= \int_{\hat{P}} \hat{w}_3 z_3 \dd{\vec{x}} = \int_{\hat{P}} (\hat{v}_3 - \hat{x}_3 \hat{q}_3) z_3 \dd{\vec{x}} = 0 &&\forall z_3 \in \Poly{k,k-2}{\hat{P}}, \\ \int_{\hat{P}} \hat{t}_1 z_1 + \hat{t}_2 z_2 \dd{\vec{x}} &= \int_{\hat{P}} \hat{w}_1 z_1 + \hat{w}_2 z_2 \dd{\vec{x}} = \int_{\hat{P}} \hat{w}_1 z_1 \dd{\vec{x}} &&\forall (z_1,z_2) \in \mathcal{P}_{k-1,k}(\hat{P}), \end{align*} \endgroup where the definitions of $\hat{q}_2$, $\hat{q}_3$, and that $\hat{w}_2|_{e_2} \equiv 0$, $\hat{w}_3|_{e_b} \equiv 0$ were used. Additional computations for the relation on $e_3$ yield \begin{align*} \frac{1}{\sqrt{2}}\int_{e_3} (\hat{t}_1 + \hat{t}_2) z \dd{\vec{s}} &= \frac{1}{\sqrt{2}}\int_{e_3} (\hat{w}_1 + \hat{w}_2) z \dd{\vec{s}} = \int_{e_3} \hat{\vec{w}} \cdot \vec{n}_{e_3} \dd{\vec{s}} \\ &= \int_{\hat{P}} \left(\pdv{\hat{w}_1}{\hat{x}_1} + \pdv{\hat{w}_2}{\hat{x}_2}\right) z \dd{\vec{x}} + \int_{\hat{P}} \hat{w}_1 \pdv{z}{\hat{x}_1} \dd{\vec{x}} - \int_{\partial\hat{P}\setminus e_3} \hat{\vec{w}}\cdot \vec{n}_{\partial\hat{P}} z \dd{\vec{s}}\\ &= \int_{\hat{P}} \left( {\widehat{\operatorname{div}}\,}\hat{\vec{w}} -2\pdv{\hat{w}_3}{\hat{x}_3}\right) \dd{\vec{x}} + \int_{\hat{P}} \hat{w}_1 \pdv{z}{\hat{x}_1} \dd{\vec{x}} - \int_{e_1} \hat{w}_1 z\dd{\vec{s}}. \end{align*} Thus, the terms \begin{subequations} \begin{align*} \int_{e_1} \hat{w}_1 z \dd{\vec{s}} &= \int_{e_1} \hat{u}_1 z \dd{\vec{s}},& \int_{\hat{P}} \hat{w}_1 z \dd{\vec{x}} &= \int_{\hat{P}} \hat{u}_1 z \dd{\vec{x}}, \\ \int_{\hat{P}} \pdv{\hat{w}_3}{\hat{x}_3} z \dd{\vec{x}} &= \int_{\hat{P}} \pdv{\hat{u}_3-\hat{x}_3\hat{q}_3}{\hat{x}_3} z \dd{\vec{x}},& \int_{\hat{P}} ({\widehat{\operatorname{div}}\,} \hat{\vec{w}}) z \dd{\vec{x}} &= \int_{\hat{P}} {\widehat{\operatorname{div}}\,} (\hat{\vec{u}} - \hat{\vec{v}}_*) z\dd{\vec{x}} \end{align*} \end{subequations} define the interpolant. We get the desired estimate \begin{equation*} \norm{(\hBDMIk{k} \hat{\vec{u}})_1}_{0,p,\hat{P}} \lesssim \norm{\hat{u}_1}_{1,p,\hat{P}} + \norm{{\widehat{\operatorname{div}}\,}\hat{\vec{u}}}_{0,p,\hat{P}} + \norm{\pdv{\hat{u}_3}{\hat{x}_3}}_{0,p,\hat{P}} \end{equation*} using a trace theorem and steps similar to those in the proof of \cite[Lemma 3.3]{AcostaApelDuranLombardi2011} to estimate the terms involving $\hat{x}_3\hat{q}_3$ and ${\widehat{\operatorname{div}}\,} \hat{\vec{v}}_*$. For the third component, with the definitions \begin{align*} &\hat{\vec{v}}=\hat{\vec{u}}-\hat{\vec{u}}_* = \begin{pmatrix}\hat{u}_1-\hat{u}_1(0,\hat{x}_2,\hat{x}_3)\\\hat{u}_2-\hat{u}_2(\hat{x}_1,0,\hat{x}_3)\\\hat{u}_3\end{pmatrix}, && \hat{\vec{v}}_* = \begin{pmatrix} \hat{x}_1\hat{q}_1\\ \hat{x}_2\hat{q}_2\\ 0 \end{pmatrix}, &&\hat{\vec{w}}=\hat{\vec{v}} - \hat{\vec{v}}_* = \begin{pmatrix}\hat{v}_1-\hat{x}_1\hat{q}_1\\\hat{v}_2-\hat{x}_2\hat{q}_2\\\hat{v}_3\end{pmatrix}, \end{align*} where the functions $\hat{q}_j \in \Poly{k-1,k}{\hat{P}}$, $j\in I_2$, are now defined by \begin{align*} \int_{\hat{P}} \hat{w}_j z \dd{\vec{x}}=\int_{\hat{P}} (\hat{v}_j - \hat{x}_j\hat{q}_j)z \dd{\vec{x}}=0 \qquad \forall z\in \Poly{k-1,k}{\hat{P}}, \end{align*} the relevant terms of the interpolation relations are \begin{align*} \int_{e_b} \hat{w}_3 z \dd{\vec{s}} &= \int_{e_b} \hat{u}_3 z \dd{\vec{s}}, & \int_{e_t} \hat{w}_3 z \dd{\vec{s}} &= \int_{e_t} \hat{u}_3 z \dd{\vec{s}}, \\ \int_{\hat{P}} \hat{w}_3 z_3 \dd{\vec{x}} &= \int_{\hat{P}} \hat{u}_3 z_3 \dd{\vec{x}}, & \int_{\hat{P}} {\widehat{\operatorname{div}}\,} \hat{\vec{w}} z \dd{\vec{x}} &= \int_{\hat{P}} {\widehat{\operatorname{div}}\,} (\hat{\vec{u}} - \hat{\vec{v}}_*) z\dd{\vec{x}}. \end{align*} From here the same steps as for the first component yield the desired estimate. \end{proof} Using the transformation \eqref{eq:trafo} we bring the stability estimate to an element of the reference family. \begin{lemma} Let $\tilde{P} = J_{\tilde{P}} \hat{P} + \vec{x}_0$, $\vec{x}_0 \in {\mathbb R}^3$, and $\tilde{\vec{v}}\in\vec{W}^{1,p}(\tilde{P})$, $1\leq p \leq \infty$. Then on the prism $\tilde{P}$ the estimate \begin{equation*} \norm{\tBDMIk{k}\tilde{\vec{v}}}_{0,p,\tilde{P}} \lesssim \sum_{\abs{\vec{\alpha}}\leq 1} h^{\vec{\alpha}} \norm{D^{\vec{\alpha}}\tilde{\vec{v}}}_{0,p,\tilde{P}} + h_{\tilde{P}} \norm{{\widetilde{\operatorname{div}}\,} \tilde{\vec{v}}}_{0,p,\tilde{P}} + (h_1+h_2)\norm{\pdv{\tilde{v}_3}{\tilde{x}_3}}_{0,p,\tilde{P}} \end{equation*} holds, where $h_{\tilde{P}} = \max\{h_1, h_2, h_3\}$. \end{lemma} \begin{proof} Using \Cref{lem:stability_BDM_reference_element_prism} and the relations \begin{align*} \norm{\tilde{\vec{w}}}_{0,p,\tilde{P}} &= \bigg(\int_{\tilde{P}} \sum_{i\in I_d} \abs{\tilde{w}_i}^p \dd{\vec{x}}\bigg)^{\nicefrac{1}{p}} \\ &\leq (\det J_{\tilde{P}})^{\nicefrac{1}{p}} \sum_{i\in I_d} {}_i h ^{-1} \left(\int_{\hat{P}} \abs{\hat{w}_i}^p \dd{\vec{x}}\right)^{\nicefrac{1}{p}} = (\det J_{\tilde{P}})^{\nicefrac{1}{p}} \sum_{i\in I_d} {}_i h ^{-1} \norm{\hat{w}_i}_{0,p,\hat{P}}, \end{align*} \begin{equation*} (\det J_{\tilde{P}})^{\nicefrac{1}{p}}\norm{\hat{v}_i}_{1,p,\hat{P}} = {}_i h \sum_{\abs{\vec{\alpha}}\leq 1} h^{\vec{\alpha}} \norm{D^{\vec{\alpha}} \tilde{v}_i}_{0,p,\tilde{P}}, \end{equation*} where ${}_i h = \prod_{j\in I_3\setminus\{i\}} h_j$, we compute \begingroup \allowdisplaybreaks \begin{align*} \norm{\tBDMIk{k}\tilde{\vec{v}}}_{0,p,\tilde{P}} &\leq (\det J_{\tilde{P}})^{\nicefrac{1}{p}} \sum_{i\in I_3} {}_i h^{-1} \norm{(\hBDMIk{k}\hat{\vec{v}})_i}_{0,p,\hat{P}} \\ &\lesssim (\det J_{\tilde{P}})^{\nicefrac{1}{p}} \left[\sum_{i\in I_3} {}_i h^{-1} \left(\norm{\hat{v}_i}_{1,p,\hat{P}} + \norm{{\widehat{\operatorname{div}}\,}\hat{\vec{v}}}_{0,p,\hat{P}}\right) + \frac{h_1+h_2}{h_1 h_2 h_3} \norm{\pdv{\hat{v}_3}{\hat{x}_3}}_{0,p,\hat{P}}\right]\\ &= \sum_{i\in I_3} \left(\sum_{\abs{\vec{\alpha}}\leq 1} h^{\vec{\alpha}} \norm{D^{\vec{\alpha}}\tilde{v}_i}_{0,p,\tilde{P}} + h_i\norm{{\widetilde{\operatorname{div}}\,} \tilde{\vec{v}}}_{0,p,\tilde{P}}\right) + (h_1+h_2)\norm{\pdv{\tilde{v}_3}{\tilde{x}_3}}_{0,p,\tilde{P}}\\ &\lesssim \sum_{\abs{\vec{\alpha}}\leq 1} h^{\vec{\alpha}} \norm{D^{\vec{\alpha}}\tilde{\vec{v}}}_{0,p,\tilde{P}} + h_{\tilde{P}} \norm{{\widetilde{\operatorname{div}}\,} \tilde{\vec{v}}}_{0,p,\tilde{P}} + (h_1+h_2)\norm{\pdv{\tilde{v}_3}{\tilde{x}_3}}_{0,p,\tilde{P}}. \qedhere \end{align*} \endgroup \end{proof} The estimate from the previous lemma can be brought to the general prism $P$ where the transformation is assumed to be reasonable in a certain sense. The stability estimate on $P$ can be used to get the interpolation error estimate by a Bramble--Hilbert type argument. The proofs for the stability and interpolation error estimates on the element $P$ follow the same steps as their analogs on simplices, see \cite[Theorems 3.5, 4.3]{ApelKempf2020} and cf. \cite[Theorems 3.1, 6.2]{AcostaApelDuranLombardi2011}, which is why they are omitted for brevity. \begin{theorem}\label{th:BDM_prism_stability} Let $P$ be a prism element that emerges by the affine transformation $\vec{x} = J_P\tilde{\vec{x}}$, with $\norm{J_P}_{\infty},\norm{J_P^{-1}}_{\infty} \leq C$, of the element $\tilde{P}\in\mathcal{R}_P$. Then for $\vec{v}\in\vec{W}^{1,p}(P)$, $1\leq p\leq\infty$, the estimate \begin{equation*} \norm{\BDMIk{k} \vec{v}}_{0,p,P} \lesssim \norm{\vec{v}}_{0,p,P} + \sum_{j\in I_3} h_j\norm{\pdv{\vec{v}}{\vec{l}_j}}_{0,p,P} + h_P \norm{\div\vec{v}}_{0,p,P} + (h_1+h_2)\norm{\pdv{v_3}{\vec{l}_3}}_{0,p,P} \end{equation*} is satisfied. The vectors $\vec{l}_j$ are the outgoing unit vectors along the edges adjacent to the transformed vertex $\vec{p}_i$. \end{theorem} \begin{theorem} Let a prism $P$ satisfy the same condition as in \Cref{th:BDM_prism_stability}. Then for $k\geq 1$, $0\leq m\leq k$ and $\vec{v}\in \vec{W}^{m+1,p}(P)$, $1\leq p\leq\infty$, the estimate \begin{align*} \norm{\vec{v}-\BDMIk{k}\vec{v}}_{0,p,P} &\lesssim \sum_{\abs{\vec{\alpha}}=m+1} h^{\vec{\alpha}} \norm{D^{\vec{\alpha}}_{\vec{l}}\vec{v}}_{0,p,P} + h_P \sum_{\abs{\vec{\alpha}}=m} h^{\vec{\alpha}} \norm{D^{\vec{\alpha}}_{\vec{l}} \div\vec{v}}_{0,p,P}\\ &\hphantom{\lesssim\ }+(h_1+h_2) \sum_{\abs{\vec{\alpha}}= m}h^{\vec{\alpha}} \norm{\frac{\partial^{m+1}v_3}{\partial\vec{l}_1^{\alpha_1}\partial\vec{l}_2^{\alpha_2}\partial\vec{l}_3^{\alpha_3 +1}}}_{0,p,P} \end{align*} holds and the constant only depends on $k$. \end{theorem} \printbibliography \end{document}

https://arxiv.org/abs/1903.00465

A note on congruence properties of the generalized bi-periodic Horadam sequence

In this paper, we consider a generalization of Horadam sequence {w_n} which is defined by the recurrence w_n = aw_n-1 + cw_n-2; if n is even, w_n = bw_n-1 + cw_n-2; if n is odd with arbitrary initial conditions w_0, w_1 and nonzero real numbers a, b, and c. We investigate some congruence properties of the generalized Horadam sequence {w_n}.

\section{Introduction} The generalized bi-periodic Horadam sequence $\left\{ w_{n}\right\} $ is defined by the recurrence relatio \begin{equation*} w_{n}=\left\{ \begin{array}{ll} aw_{n-1}+cw_{n-2}, & \mbox{ if }n\mbox{ is even} \\ bw_{n-1}+cw_{n-2}, & \mbox{ if }n\mbox{ is odd \end{array \right. ,\text{ }n\geq 2 \end{equation* with arbitrary initial conditions $w_{0},w_{1}$ and nonzero real numbers a,b $ and $c$. It is emerged as a generalization of the best known sequences in the literature, such as the Horadam sequence, the Fibonacci\&Lucas sequence, the $k$-Fibonacci\&$k$-Lucas sequence, the Pell\&Pell-Lucas sequence, the Jacobsthal\& Jacobsthal-Lucas sequence, etc. Similar to the notation of the classical Horadam sequence \cite{Horadam}, we write $\left\{ w_{n}\right\} := $ $\{w_{n}\left( w_{0},w_{1};a,b,c\right) \}.$ In particular, using this notation, we define $\{u_{n}\}=\{w_{n}\left( 0,1;a,b,c\right) \}$ and $\{v_{n}\}=\{w_{n}\left( 2,b;a,b,c\right) \}$ as the \textit{generalized bi-periodic Fibonacci sequence} and the \textit generalized bi-periodic Lucas sequence, }respectively\textit{.} For the basic properties of the generalized bi-periodic Horadam sequence and some special cases of this sequence, see \cite{Edson, Yayenie, Sahin, Panario, Bilgici, Tan1, Tan3, Tan5, Tan-Leung}. On the other hand, it is important to investigate the congruence properties of different integer sequences. Several methods can be applied to produce identities for the Fibonacci and Lucas sequences. For example, Carlitz and Ferns \cite{Carlitz} used polynomial identities in conjunction with the Binet formula to generate new identities for these sequences. The method of Carlitz and Ferns was used by several authors to obtain analogous results for the generalized Fibonacci and Lucas sequences, see \cite{Zhang0, Kilic}. On the other hand, Keskin and Siar \cite{Siar} obtained some number theoretic properties of the generalized Fibonacci and Lucas numbers by using matrix method. Morover, Hsu and Maosen \cite{Hsu} and Zhang \cite{Zhang1} applied an operator method to establish some of these properties. Recently, Yang and Zhang \cite{Zhang} have studied some congruence relations for the bi-periodic Fibonacci and Lucas sequences by using operator method. But some of the results that are obtained by the operator method are incorrect. In this study, by using the method of Carlitz and Ferns \cite{Carlitz}, we give more general identities involving the generalized bi-periodic Horadam sequences and derive some congruence properties of the generalized bi-periodic Horadam numbers. In particular, our results include the corrected version of some of the results in \cite{Zhang}. The outline of this paper as follows: In Section $2$, we give some basic properties of the generalized bi-periodic Horadam sequence. In Section $3$, we give some binomial identities and congruence relations for the generalized bi-periodic Horadam sequence by using the method of Carlitz and Ferns \cite{Carlitz}. \section{Some preliminary results for the sequence $\{w_{n}\}$} In this section, we give some basic properties of the bi-periodic Horadam sequences. The Binet formula of the sequence $\{u_{n}\}$ i \begin{equation} u_{n}=\frac{a^{\xi \left( n+1\right) }}{\left( ab\right) ^{\left\lfloor \frac{n}{2}\right\rfloor }}\left( \frac{\alpha ^{n}-\beta ^{n}}{\alpha -\beta }\right) \label{3} \end{equation which can be obtained by \cite[Theorem 8]{Yayenie}. Here $\alpha $ and \beta $ are the roots of the polynomial $x^{2}-abx-abc,$ that is, $\alpha \frac{ab+\sqrt{a^{2}b^{2}+4abc}}{2}$ and $\beta =\frac{ab-\sqrt a^{2}b^{2}+4abc}}{2}$, and $\xi \left( n\right) =n-2\left\lfloor \frac{n}{2 \right\rfloor $ is the parity function, i.e., $\xi \left( n\right) =0$ when n$ is even and $\xi \left( n\right) =1$ when $n$ is odd. Let assume $\Delta :=a^{2}b^{2}+4abc\neq 0.$ Also we have $\alpha +\beta =ab,$ $\alpha -\beta \sqrt{a^{2}b^{2}+4abc}$ and $\alpha \beta =-abc.$ \begin{lemma} \label{l1}For any integer $n>0,$ we hav \begin{equation*} w_{n}=u_{n}w_{1}+c\left( \frac{b}{a}\right) ^{\xi \left( n\right) }u_{n-1}w_{0}. \end{equation*} \end{lemma} By\ using Lemma \ref{l1} and the Binet formula of $\left\{ u_{n}\right\} $ in (\ref{3}), we can easily obtain the Binet formula of the sequence \{w_{n}\}.$ We note that the extended Binet formula for the general case of this sequence was given in \cite[Theorem 9]{Panario}. But here we express the Binet formula of the sequence $\{w_{n}\}$ in a different manner, that is, our $\alpha $ and $\beta $ are different from the roots which are used in \cite{Panario}. \begin{theorem} \label{t1}(Binet Formula) For $n>0,$ we hav \begin{equation*} w_{n}=\frac{a^{\xi \left( n+1\right) }}{\left( ab\right) ^{\left\lfloor \frac{n}{2}\right\rfloor }}\left( A\alpha ^{n}-B\beta ^{n}\right) , \end{equation* where $A:=\frac{w_{1}-\frac{\beta }{a}w_{0}}{\alpha -\beta }$and $B:=\frac w_{1}-\frac{\alpha }{a}w_{0}}{\alpha -\beta }.$ \end{theorem} \begin{proof} By\ using Lemma \ref{l1} and (\ref{3}), we get the desired result. \end{proof} By taking initial conditions $w_{0}=2,w_{1}=b$ in Theorem \ref{t1}, we obtain the Binet formula for the sequence $\{v_{n}\}$ as follows \begin{equation} v_{n}=\frac{a^{-\xi \left( n\right) }}{\left( ab\right) ^{\left\lfloor \frac n}{2}\right\rfloor }}\left( \alpha ^{n}+\beta ^{n}\right) . \label{4} \end{equation Also by\ using Lemma \ref{l1}, we have $v_{n}=bu_{n}+2c\left( \frac{b}{a \right) ^{\xi \left( n\right) }u_{n-1}.$ Thus we get a relation between the generalized bi-periodic Fibonacci and the generalized bi-periodic Lucas numbers as \begin{equation} v_{n}=\left( \frac{b}{a}\right) ^{\xi \left( n\right) }\left( u_{n+1}+cu_{n-1}\right) . \label{5} \end{equation It should be noted that the generalized Lucas sequence $\left\{ t_{n}\right\} $ in \cite{Zhang} is a special case of the generalized bi-periodic Horadam sequence. That is, $\{t_{n}\}=\{w_{n}\left( 2a,ab;a,b,1\right) \}.$ The generating function of the sequence $\left\{ w_{n}\right\} $ i \begin{equation} G\left( x\right) =\frac{w_{0}+w_{1}x+\left( aw_{1}-\left( ab+c\right) w_{0}\right) x^{2}+c\left( bw_{0}-w_{1}\right) x^{3}}{1-\left( ab+2c\right) x^{2}+c^{2}x^{4}}, \label{6} \end{equation which can be obtained from \cite[Theorem 6]{Panario}. Also we need the following identity which can be found in \cite{Tan-Leung} \begin{equation} u_{mn+r}=\frac{a^{1-\xi \left( mn+r\right) }}{\left( ab\right) ^{\left\lfloor \frac{mn+r}{2}\right\rfloor }}\sum_{i=0}^{n}\dbinom{n}{i c^{n-i}u_{m}^{i}u_{m-1}^{n-i}u_{i+r}\delta \lbrack m,n,r,i] \label{7} \end{equation where $\delta \lbrack m,n,r,i]:=\left( ab\right) ^{\left\lfloor \frac{i+r}{2 \right\rfloor +n\left\lfloor \frac{m}{2}\right\rfloor }a^{-\xi \left( m+1\right) i-1+\xi \left( i+r\right) }b^{\xi \left( m\right) \left( n-i\right) }.$ \section{Main results} To extend the results in \cite[Theorem 4.7, Theorem 4.9, Theorem 4.11, Theorem 4.13]{Zhang}, we give the following theorem. Also we assume that a,b $ and $c$ are positive integers. \begin{theorem} \label{t2}For any nonnegative integers $n,r$ and $m$ with $m>1$, we hav \begin{equation*} w_{mn+r}=\frac{a^{1-\xi \left( mn+r\right) }}{\left( ab\right) ^{\left\lfloor \frac{mn+r}{2}\right\rfloor }}\sum_{i=0}^{n}\dbinom{n}{i c^{n-i}u_{m}^{i}u_{m-1}^{n-i}w_{i+r}\delta \lbrack m,n,r,i] \end{equation* where $\delta \lbrack m,n,r,i]:=\left( ab\right) ^{\left\lfloor \frac{i+r}{2 \right\rfloor +n\left\lfloor \frac{m}{2}\right\rfloor }a^{-\xi \left( m+1\right) i-1+\xi \left( i+r\right) }b^{\xi \left( m\right) \left( n-i\right) }.$ \end{theorem} \begin{proof} Similar to the relation $\gamma ^{n}=\gamma F_{n}+F_{n-1}$ for the classical Fibonacci numbers, where $\gamma $ is one of the root of the equation x^{2}-x-1=0,$ we have \begin{equation*} \alpha ^{m}=a^{-1}a^{\frac{m+\xi (m)}{2}}b^{\frac{m-\xi (m)}{2}}\alpha u_{m}+ca^{\frac{m-\xi (m)}{2}}b^{\frac{m+\xi (m)}{2}}u_{m-1} \end{equation* an \begin{equation*} \beta ^{m}=a^{-1}a^{\frac{m+\xi (m)}{2}}b^{\frac{m-\xi (m)}{2}}\beta u_{m}+ca^{\frac{m-\xi (m)}{2}}b^{\frac{m+\xi (m)}{2}}u_{m-1}. \end{equation* By using the binomial theorem, we hav \begin{equation*} \alpha ^{mn}=\sum_{i=0}^{n}\dbinom{n}{i}a^{-i}a^{i\frac{m+\xi (m)}{2}+\left( n-i\right) \frac{m-\xi (m)}{2}}b^{i\frac{m-\xi (m)}{2}+\left( n-i\right) \frac{m+\xi (m)}{2}}c^{n-i}u_{m}^{i}u_{m-1}^{n-i}\alpha ^{i}, \end{equation* \begin{equation*} \beta ^{mn}=\sum_{i=0}^{n}\dbinom{n}{i}a^{-i}a^{i\frac{m+\xi (m)}{2}+\left( n-i\right) \frac{m-\xi (m)}{2}}b^{i\frac{m-\xi (m)}{2}+\left( n-i\right) \frac{m+\xi (m)}{2}}c^{n-i}u_{m}^{i}u_{m-1}^{n-i}\beta ^{i}. \end{equation*} Multiplying both sides of the above equalities by $A\alpha ^{r}$ and $B\beta ^{r},$ respectively, and using the Binet formula of $\left\{ w_{n}\right\} ,$ we ge \begin{eqnarray*} &&\left( A\alpha ^{mn+r}-B\beta ^{mn+r}\right) \\ &=&a^{-\xi \left( mn+r+1\right) }\left( ab\right) ^{\left\lfloor \frac{mn+r} 2}\right\rfloor }w_{mn+r} \\ &=&\sum_{i=0}^{n}\dbinom{n}{i}\left( ab\right) ^{\left\lfloor \frac{i+r}{2 \right\rfloor +n\left\lfloor \frac{m}{2}\right\rfloor }a^{-\xi \left( m+1\right) i-1+\xi \left( i+r\right) }b^{\xi \left( m\right) \left( n-i\right) }c^{n-i}u_{m}^{i}u_{m-1}^{n-i}w_{i+r}. \end{eqnarray*} which gives the desired result. It can be expressed a \begin{equation*} w_{mn+r}=\sum_{i=0}^{n}\dbinom{n}{i}\left( \frac{b}{a}\right) ^{\frac{\xi \left( mn+r\right) -2i\xi (m)+i-\xi \left( i+r\right) +n\xi (m)}{2 }c^{n-i}u_{m}^{i}u_{m-1}^{n-i}w_{i+r}. \end{equation*} We note that it can also be obtained by using Lemma \ref{l1} and the identity (\ref{7}) as \begin{eqnarray*} &&w_{mn+r} \\ &=&w_{1}\frac{a^{1-\xi \left( mn+r\right) }}{\left( ab\right) ^{\left\lfloor \frac{mn+r}{2}\right\rfloor }}\sum_{i=0}^{n}\dbinom{n}{i c^{n-i}u_{m}^{i}u_{m-1}^{n-i}u_{i+r}\delta \lbrack m,n,r,i] \\ &&+w_{0}c\left( \frac{b}{a}\right) ^{\xi \left( mn+r\right) }\frac{a^{1-\xi \left( mn+r-1\right) }}{\left( ab\right) ^{\left\lfloor \frac{mn+r-1}{2 \right\rfloor }}\sum_{i=0}^{n}\dbinom{n}{i c^{n-i}u_{m}^{i}u_{m-1}^{n-i}u_{i+r-1}\delta \lbrack m,n,r-1,i]\text{ \ \ \ \ } \end{eqnarray* \begin{eqnarray*} &=&\frac{a^{1-\xi \left( mn+r\right) }}{\left( ab\right) ^{\left\lfloor \frac{mn+r}{2}\right\rfloor }}\sum_{i=0}^{n}\dbinom{n}{i c^{n-i}u_{m}^{i}u_{m-1}^{n-i}\delta \lbrack m,n,r,i]\left( w_{1}u_{i+r}+cw_{0}\left( \frac{b}{a}\right) ^{\xi \left( i+r\right) }u_{i+r-1}\right) \\ &=&\frac{a^{1-\xi \left( mn+r\right) }}{\left( ab\right) ^{\left\lfloor \frac{mn+r}{2}\right\rfloor }}\sum_{i=0}^{n}\dbinom{n}{i c^{n-i}u_{m}^{i}u_{m-1}^{n-i}w_{i+r}\delta \lbrack m,n,r,i]. \end{eqnarray*} \end{proof} By considering the identity $\xi (mn+r)=\xi (mn)+\xi (r)-2\xi (mn)\xi (r),$ we have the following corollary. \begin{remark} When $m$ and $r$ are all even and $c=1$ in Theorem \ref{t2}, we obtain the identit \begin{equation*} w_{2mn+2r}=\dsum\limits_{i=0}^{n}\binom{n}{i}\left( \frac{b}{a}\right) ^ \frac{i-\xi \left( i\right) }{2}}u_{2m}^{i}u_{2m-1}^{n-i}w_{i+2r}. \end{equation* Thus, the result in \cite[Theorem 4.7]{Zhang} can be corrected by multiplying the right side of the equation by $\left( \frac{b}{a}\right) ^ \frac{i-\xi \left( i\right) }{2}}.$ The other results in \cite[Theorem 4.9, Theorem 4.11, Theorem 4.13]{Zhang} can be corrected similarly. \end{remark} \begin{corollary} \label{c1}For $m,n,r>0,$ we hav \begin{equation*} w_{mn+r}-\left( \frac{b}{a}\right) ^{\xi (m)\left( \frac{n+\xi (n)}{2 \right) -\xi \left( mn\right) \xi \left( r\right) }c^{n}u_{m-1}^{n}w_{r \text{ }\equiv 0\left( \func{mod}u_{m}\right) . \end{equation*} \end{corollary} Now we give a generalization of the Ruggles identity \cite{Ruggles} which also generalizes the identities in \cite[Theorem 2.2 (3-4)]{Zhang} and \cite Theorem 1]{Yayenie}. Then we give a related binomial identity for the generalized bi-periodic Horadam sequence\textit{.} For $n\geq 0$ and $k\geq 1,$ the Ruggles identity \cite{Ruggles} is given b \begin{equation*} F_{n+2k}=L_{k}F_{n+k}+\left( -1\right) ^{k+1}F_{n}, \end{equation* where $\{F_{n}\}$ and $\{L_{n}\}$ are the Fibonacci and Lucas numbers, respectively. Horadam \cite{Horadam} generalized this result to a general second order recurrence relatio \begin{equation*} W_{n+2k}=V_{k}W_{n+k}+\left( -1\right) ^{k+1}q^{k}W_{n}, \end{equation* where $W_{k}=pW_{k-1}+qW_{k-2}$ with arbitrary initial conditions $W_{0}$ and $W_{1}.$ The sequence $\left\{ V_{k}\right\} $ satisfies the same recurrence relation as the sequence $\left\{ W_{k}\right\} ,$ but it begins with $V_{0}=2,V_{1}=p.$ A generalization of Ruggles identity can be given in the following lemma. \begin{lemma} \label{l2}For integers $n\geq 0$ and $k\geq 1,$ we hav \begin{equation*} w_{n+2k}=\left( \frac{a}{b}\right) ^{\xi (n+1)\xi (k)}v_{k}w_{n+k}-\left( -c\right) ^{k}w_{n} \end{equation* where $\left\{ w_{n}\right\} $ is the \textit{generalized bi-periodic Horadam sequence} and $\{v_{n}\}$ is the \textit{generalized bi-periodic Lucas sequence}. \end{lemma} \begin{proof} It can be obtained simply by the Binet formula of $\left\{ w_{n}\right\} .$ \end{proof} \begin{theorem} \label{t3}For nonnegative integers $n,r$ and $m$ with $m>1$, we hav \begin{equation*} w_{2mn+r}=\dsum\limits_{i=0}^{n}\binom{n}{i}\left( -1\right) ^{\left( m+1\right) \left( n-i\right) }\left( \frac{a}{b}\right) ^{\xi (m)\left( \frac{i+\xi \left( i\right) }{2}\right) -\xi (im)\xi (r)}c^{m\left( n-i\right) }v_{m}^{i}w_{im+r}. \end{equation*} \end{theorem} \begin{proof} From the Binet formula of $\left\{ v_{k}\right\} $ and $\alpha \beta =-abc,$ it is clear to see that \begin{equation*} \alpha ^{2m}=a^{\frac{m+\xi (m)}{2}}b^{\frac{m-\xi (m)}{2}}v_{m}\alpha ^{m}-\left( -abc\right) ^{m}. \end{equation* By using the binomial theorem, we hav \begin{equation*} \alpha ^{2mn}=\sum_{i=0}^{n}\dbinom{n}{i}a^{i\frac{m+\xi (m)}{2}}b^{i\frac m-\xi (m)}{2}}\left( -1\right) ^{\left( m+1\right) \left( n-i\right) }\left( abc\right) ^{m\left( n-i\right) }v_{m}^{i}\alpha ^{im}. \end{equation* Similarly, we hav \begin{equation*} \beta ^{2mn}=\sum_{i=0}^{n}\dbinom{n}{i}a^{i\frac{m+\xi (m)}{2}}b^{i\frac m-\xi (m)}{2}}\left( -1\right) ^{\left( m+1\right) \left( n-i\right) }\left( abc\right) ^{m\left( n-i\right) }v_{m}^{i}\beta ^{im}. \end{equation*} Multiplying both sides of the above equalities by $A\alpha ^{r}$ and $B\beta ^{r},$ respectively, and using the Binet formula of $\left\{ w_{n}\right\} ,$ we ge \begin{eqnarray*} &&\left( A\alpha ^{2mn+r}-B\beta ^{2mn+r}\right) \\ &=&a^{-\xi \left( 2mn+r+1\right) }\left( ab\right) ^{\left\lfloor \frac{2mn+ }{2}\right\rfloor }w_{2mn+r} \\ &=&\sum_{i=0}^{n}\dbinom{n}{i}a^{i\frac{m+\xi (m)}{2}}b^{i\frac{m-\xi (m)}{2 }\left( -1\right) ^{\left( m+1\right) \left( n-i\right) }\left( abc\right) ^{m\left( n-i\right) }v_{m}^{i}\left( A\alpha ^{im+r}-B\beta ^{im+r}\right) . \end{eqnarray* Thus, again by using the identity $\xi (mn+r)=\xi (mn)+\xi (r)-2\xi (mn)\xi (r),$ we hav \begin{eqnarray*} w_{2mn+r} &=&\sum_{i=0}^{n}\dbinom{n}{i}\left( -1\right) ^{\left( m+1\right) \left( n-i\right) }c^{m\left( n-i\right) }v_{m}^{i}w_{im+r} \\ &&\times a^{i\frac{m+\xi (m)}{2}}b^{i\frac{m-\xi (m)}{2}}\left( ab\right) ^{m\left( n-i\right) -\left\lfloor \frac{2mn+r}{2}\right\rfloor +\left\lfloor \frac{im+r}{2}\right\rfloor }a^{\xi \left( 2mn+r+1\right) }a^{-\xi \left( im+r+1\right) } \end{eqnarray* which gives the desired result. \end{proof} \begin{remark} Since $\left( \frac{a}{b}\right) ^{\xi (r+1)\xi (m)i+\left( -1\right) ^{r+1}\xi (m)\left( \frac{i-\xi \left( i\right) }{2}\right) }=\left( \frac{ }{b}\right) ^{\xi (m)\left( \frac{i+\xi \left( i\right) }{2}\right) -\xi (im)\xi (r)},$ the results in \cite[Theorem 4.3, Theorem 4.5]{Zhang} can be corrected by multiplying the right side of the equations by $\left( \frac{a} b}\right) ^{\left( -1\right) ^{r+1}\left( \frac{i-\xi \left( i\right) }{2 \right) }.$ \end{remark} \begin{corollary} \label{c2}For $m,n,r>0,$ we hav \begin{equation*} w_{2mn+r}-\left( -1\right) ^{\left( m+1\right) n}c^{mn}w_{r}\equiv 0\left( \func{mod}v_{m}\right) . \end{equation*} \end{corollary} Note that for $m=2$ and $m=3$, Corollary \ref{c2} gives the results in \cite Corollary 4.2, Corollary 4.4, Corollary 4.6]{Zhang}. Also for the case of generalized Fibonacci and Lucas sequences, it gives the results in \cite 3.3. Corollary]{Siar}. \begin{lemma} \label{l3}For $m,r>0$, we hav \begin{eqnarray*} &&-\left( -abc\right) ^{m+r}+a^{\frac{r+\xi (r)}{2}}b^{\frac{r-\xi (r)}{2 }v_{r}\left( -abc\right) ^{m}z^{r}+z^{2\left( m+r\right) } \\ &=&z^{m+2r}a^{\frac{m+\xi (m)}{2}}b^{\frac{m-\xi (m)}{2}}v_{m} \end{eqnarray* where $z$ is either $\alpha $ or $\beta .$ \end{lemma} \begin{proof} From the Binet formula of $\left\{ v_{n}\right\} $ and $\alpha \beta =-abc,$ it is clear to see that $z^{2r}=a^{\frac{r+\xi (r)}{2}}b^{\frac{r-\xi (r)}{2 }v_{r}z^{r}-\left( -abc\right) ^{r}.$ Thus we hav \begin{eqnarray*} &&-\left( -abc\right) ^{m+r}+a^{\frac{r+\xi (r)}{2}}b^{\frac{r-\xi (r)}{2 }v_{r}\left( -abc\right) ^{m}z^{r}+z^{2\left( m+r\right) } \\ &=&\left( -abc\right) ^{m}\left( a^{\frac{r+\xi (r)}{2}}b^{\frac{r-\xi (r)}{ }}v_{r}z^{r}-\left( -abc\right) ^{r}\right) +z^{2\left( m+r\right) } \\ &=&\left( -abc\right) ^{m}z^{2r}+z^{2\left( m+r\right) } \\ &=&z^{m+2r}\left( \left( -abc\right) ^{m}z^{-m}+z^{m}\right) \\ &=&z^{m+2r}\left( \beta ^{m}+\alpha ^{m}\right) \\ &=&z^{m+2r}a^{\frac{m+\xi (m)}{2}}b^{\frac{m-\xi (m)}{2}}v_{m}. \end{eqnarray*} \end{proof} \begin{theorem} \label{t4}For $n,m,r>0$, we hav \begin{eqnarray*} &&-\left( -c\right) ^{m+r}w_{n}+\left( -c\right) ^{m}\left( \frac{a}{b \right) ^{\xi (r)\xi (n+1)}v_{r}w_{r+n}+w_{2\left( m+r\right) +n} \\ &=&\left( \frac{a}{b}\right) ^{\xi (m)\xi (n+1)}v_{m}w_{m+2r+n}. \end{eqnarray*} \end{theorem} \begin{proof} From Lemma \ref{l3}, we hav \begin{eqnarray} &&-\left( -abc\right) ^{m+r}+a^{\frac{r+\xi (r)}{2}}b^{\frac{r-\xi (r)}{2 }v_{r}\left( -abc\right) ^{m}\alpha ^{r}+\alpha ^{2\left( m+r\right) } \notag \\ &=&\alpha ^{m+2r}a^{\frac{m+\xi (m)}{2}}b^{\frac{m-\xi (m)}{2}}v_{m}. \label{*} \end{eqnarray Similarly, we hav \begin{eqnarray} &&-\left( -abc\right) ^{m+r}+a^{\frac{r+\xi (r)}{2}}b^{\frac{r-\xi (r)}{2 }v_{r}\left( -abc\right) ^{m}\beta ^{r}+\beta ^{2\left( m+r\right) } \notag \\ &=&\beta ^{m+2r}a^{\frac{m+\xi (m)}{2}}b^{\frac{m-\xi (m)}{2}}v_{m}. \label{**} \end{eqnarray By multiplying both sides of the equations (\ref{*}) and (\ref{**}) by A\alpha ^{n}$ and $B\beta ^{n},$ respectively, we ge \begin{eqnarray*} &&-\left( A\alpha ^{n}-B\beta ^{n}\right) \left( -abc\right) ^{m+r} \\ &&+a^{\frac{r+\xi (r)}{2}}b^{\frac{r-\xi (r)}{2}}v_{r}\left( -abc\right) ^{m}\left( A\alpha ^{r+n}-B\beta ^{r+n}\right) \\ &&+\left( A\alpha ^{2\left( m+r\right) +n}-B\beta ^{2\left( m+r\right) +n}\right) \\ &=&a^{\frac{m+\xi (m)}{2}}b^{\frac{m-\xi (m)}{2}}v_{m}\left( A\alpha ^{m+2r+n}-B\beta ^{m+2r+n}\right) . \end{eqnarray* Then by using the Binet formula of $\left\{ w_{n}\right\} $, we have \begin{eqnarray*} &&-\left( ab\right) ^{m+r}\left( -c\right) ^{m+r}a^{\frac{n+\xi (n)}{2}}b^ \frac{n-\xi (n)}{2}}w_{n} \\ &&+\left( -c\right) ^{m}\left( ab\right) ^{m+r}a^{\frac{n+\xi (r+n)+\xi (r)} 2}}b^{\frac{n-\xi (r+n)-\xi (r)}{2}}v_{r}w_{r+n} \\ &&+\left( ab\right) ^{m+r}a^{\frac{n+\xi (n)}{2}}b^{\frac{n-\xi (n)}{2 }w_{2\left( m+r\right) +n} \\ &=&\left( ab\right) ^{m+r}a^{\frac{n+\xi (m+n)+\xi (m)}{2}}b^{\frac{n-\xi (m+n)-\xi (m)}{2}}v_{m}w_{m+2r+n}. \end{eqnarray* By considering the identity $\xi (mn+r)=\xi (mn)+\xi (r)-2\xi (mn)\xi (r),$ we get the desired result. \end{proof} If we take $r=1,m=2,c=1$ in Theorem \ref{t4}, we ge \begin{equation*} \left( ab+2\right) w_{n+4}=w_{n}+a^{\xi (n+1)}b^{\xi (n)}w_{n+1}+w_{n+6} \end{equation* which reduces to the identit \begin{equation*} \left( ab+1\right) w_{n+4}=w_{n}+a^{\xi (n+1)}b^{\xi (n)}w_{n+1}+a^{\xi (n+1)}b^{\xi (n)}w_{n+5} \end{equation* in \cite[Theorem 4.15, Theorem 4.17]{Zhang}. \begin{theorem} \label{t5}The symbol $\dbinom{n}{i,j}$is defined by $\dbinom{n}{i,j}:=\frac n!}{i!j!\left( n-i-j\right) !}.$ For $n,m,r,d>0,$ we hav \begin{eqnarray} w_{\left( m+2r\right) n+d} &=&v_{m}^{-n}\sum_{i+j+s=n}\dbinom{n}{i,j}\left( -1\right) ^{s}\left( -c\right) ^{mj+\left( m+r\right) s}v_{r}^{j}w_{2\left( m+r\right) i+rj+d} \notag \\ &&\times \left( \frac{a}{b}\right) ^{\xi (r)\frac{j+\xi \left( j\right) }{2 -\xi (m)\frac{n+\xi \left( n\right) }{2}-\xi \left( rj\right) \xi \left( d\right) +\xi \left( mn\right) \xi \left( d\right) } \label{s2} \end{eqnarray an \begin{eqnarray} w_{2\left( m+r\right) n+d} &=&\sum_{i+j+s=n}\dbinom{n}{i,j}\left( -1\right) ^{j}\left( -c\right) ^{s\left( m+r\right) +mj}v_{m}^{i}v_{r}^{j}w_{\left( m+2r\right) i+rj+d} \notag \\ &&\times \left( \frac{a}{b}\right) ^{\xi (m)\frac{i+\xi \left( i\right) }{2 +\xi (r)\frac{j+\xi \left( j\right) }{2}-\xi \left( mi\right) \xi \left( rj\right) -\xi \left( mi+rj\right) \xi \left( d\right) }. \label{s3} \end{eqnarray} \end{theorem} \begin{proof} By using Lemma \ref{l3} and the multinomial theorem, we obtain the following identities \begin{eqnarray*} &&a^{n\frac{m+\xi (m)}{2}}b^{n\frac{m-\xi (m)}{2}}v_{m}^{n}z^{\left( m+2r\right) n} \\ &=&\sum_{i+j+s=n}\dbinom{n}{i,j}\left( -1\right) ^{s}\left( -abc\right) ^{s\left( m+r\right) +mj}a^{j\frac{r+\xi (r)}{2}}b^{j\frac{r-\xi (r)}{2 }v_{r}^{j}z^{2\left( m+r\right) i+rj}\text{\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ } \end{eqnarray* an \begin{eqnarray*} &&z^{2\left( m+r\right) n} \\ &=&\sum_{i+j+s=n}\dbinom{n}{i,j}\left( -1\right) ^{j}\left( -abc\right) ^{s\left( m+r\right) +mj}a^{\frac{im+i\xi (m)+jr+j\xi (r)}{2}}b^{\frac im-i\xi (m)+jr-j\xi (r)}{2}}v_{m}^{i}v_{r}^{j}z^{\left( m+2r\right) i+rj}. \end{eqnarray* By multiplying both sides in the preceding equalities by $z^{d}$ and using the Binet formula of $\left\{ w_{n}\right\} ,$ we have (\ref{s2}) and (\re {s3}), respectively. \end{proof} We note that for the computational simplicity, the equation (\ref{s3}) is more practical than the equation (\ref{s2}). From (\ref{s3}), by using the decompositio \begin{equation*} \sum_{i+j+s=n}=\sum_{i+j+s=n,i=0}+\sum_{i+j+s=n,i\neq 0} \end{equation* and Theorem \ref{t3}, we get the following corollary. \begin{corollary} \label{c3}For $n,m,r,d>0$, we hav \begin{equation*} w_{2\left( m+r\right) n+d}-\left( -1\right) ^{n\left( m+1\right) }c^{mn}w_{2rn+d}\equiv 0\left( \func{mod}v_{m}\right) . \end{equation*} \end{corollary} \section{Acknowledgement} The first author is grateful to Dr. Mohamed Salim for the arrangement of her visit to United Arab Emirates University (UAEU) in February 2019. It is supported by UAEU UPAR Grant G00002599 (Fund No. 31S314).

https://arxiv.org/abs/1310.2691

An incomplete variant of Wilson's congruence

This article examines the nontrivial solutions of the congruence \[ (p-1)\cdots(p-r) \equiv -1 \pmod p. \] We discuss heuristics for the proportion of primes $p$ that have exactly $N$ solutions to this congruence. We supply numerical evidence in favour of these conjectures, and discuss the algorithms used in our calculations.

\section{Heuristics and conjectures} Wilson's Theorem \cite[Theorems 80 and 81]{HW} states that \[ (p-1)! \equiv -1 \pmod p \] if and only if $p$ is a prime. Now truncate the factorial after $r$ terms. For which primes $p$ is there an $r$ for which \begin{equation}\label{wit} (p-1)\cdots(p-r) \equiv -1 \pmod p ? \end{equation} Certainly $r=1$ is trivial; $r=p-1$ follows from Wilson's Theorem, whence $r=p-2$ follows trivially. Henceforth we consider only $2 \leq r \leq p - 3$. Initially we proceeded as follows. The congruence \eqref{wit} has no solutions if and only if none of the $p - 4$ integers $(p-1)\cdots(p-r) + 1$, for $2\leq r \leq p-3$, are divisible by $p$. The probability that a prime $p$ divides a `random' integer $N$ is $1/p$. Given $m$ random integers chosen independently, the probability that $p$ does not divide any of them is then $(1 - 1/p)^m$. Thus, under heroic randomness and independence assumptions, we expect the proportion of $p$ for which \eqref{wit} has no solutions to be roughly \[ (1-1/p)^{p-4}\rightarrow e^{-1} \approx 0.36788 \] when $p$ is large. Turning to numerical experiment, we find that 429 of the 1229 primes less than $10^4$ have no solutions to \eqref{wit}. The proportion is 0.349, reasonably close to our initial guess. It turns out that this guess is almost certainly wrong. The remainder of this paper may serve as yet another cautionary tale about the dangers of heuristic probabilistic reasoning in number theory. First, our independence assumption is not justified. Denoting by $T_r$ the partial product $T_r = (p - 1) \cdots (p - r)$, we have: \begin{Lem}\label{Lemma1} For any $2 \leq r \leq p - 3$, \[ T_r T_{p-r-1} \equiv (-1)^{r+1} \pmod p. \] \end{Lem} \begin{proof} Observe that $T_r \equiv (-1)^r r! \pmod p$, and apply Wilson's Theorem. \end{proof} Thus the cases $r$ and $s = p - r - 1$ are not independent. For odd $r$, we see that \eqref{wit} holds for $r$ if and only if it holds for $s$. For even $r$, we see that \eqref{wit} holds for either $r$ or $s$, but not both; and it holds for one of them if and only if $T_r \equiv \pm 1 \pmod p$. Taking these observations into account, we should posit $p/4 + O(1)$ independent events with probability $1 - 1/p$ corresponding to the odd $r < p/2$, and $p/4 + O(1)$ independent events with probability $1 - 2/p$ corresponding to the even $r < p/2$. Our revised estimate for the proportion of primes for which \eqref{wit} has no solutions is thus \[ (1 - 1/p)^{p/4} (1 - 2/p)^{p/4} \rightarrow e^{-3/4} \approx 0.47237. \] This `improved' heuristic is an even worse match for the observed data! In fact we are on the right track. We have simply forgotten the following arithmetic gem. \begin{Lem}\label{Lemma2} Let $p \equiv 3 \pmod 4$. Then \begin{equation*} \left(\frac{p-1}2\right)! \equiv (-1)^{\nu_p} \pmod p, \end{equation*} where $\nu_p$ is the number of quadratic non-residues $1 < x < p/2$. \end{Lem} \begin{proof} See \cite[Theorem 114]{HW}. \end{proof} When $\nu_p$ is even, the congruence \eqref{wit} automatically has the solution $r = (p-1)/2$. To incorporate this into our model, we must address the question as to how often $\nu_p$ is even. Numerical evidence (see Section \ref{sec:computations}) suggests the following conjecture: \begin{con}\label{con1} For $p \equiv 3 \pmod 4$, the proportion of $p$ for which $\left(\frac{p-1}2\right)! \equiv 1 \pmod p$ approaches $\frac{1}{2}$ as $p \to \infty$. \end{con} We are not aware of this conjecture having appeared before in print, but it has been raised on the Mathoverflow discussion forum \cite{MO}. The problem has been recast by Mordell \cite{Mordell1961} in terms of the class number $h(-p)$ of $\QQ(\sqrt{-p})$. Namely, for $p > 3$ we have \[ \nu_p = \begin{cases} 0 \pmod 2 & \text{if $h(-p) \equiv 3 \pmod 4$}, \\ 1 \pmod 2 & \text{if $h(-p) \equiv 1 \pmod 4$}. \end{cases} \] We do not know if this interpretation sheds any light on Conjecture \ref{con1}. We now revise our model a second time, taking into account Lemma \ref{Lemma2} and Conjecture \ref{con1}. Of those primes satisfying $p \equiv 1 \pmod 4$, asymptotically half the primes, our estimate for the proportion of primes for which \eqref{wit} has no solution is still $e^{-3/4}$. For those primes $p \equiv 3 \pmod 4$ with $\nu_p$ odd, the estimate is again $e^{-3/4}$. According to Conjecture \ref{con1} this accounts for another quarter of the primes. However, for the remaining primes, where $\nu_p$ is even, our estimate is zero. This leads to our main conjecture. \begin{con}\label{Conmain} The proportion of primes for which \eqref{wit} has no nontrivial solutions is \[ \frac{3}{4} e^{-3/4} \approx 0.3542749. \] \end{con} Using the same model, we may develop a more refined conjecture that estimates the proportion of primes $p$ for which there are exactly $N$ values of $r$ satisfying \eqref{wit}. \begin{con}\label{ConN} Let $N \geq 0$. The proportion of primes $p$ for which \eqref{wit} has exactly $N$ nontrivial solutions is \[ \frac{e^{-3/4}}{2^{N+1}} \left( \frac{3}{2} \sum_{k=0}^{\lfloor N/2 \rfloor} \frac{1}{k! (N-2k)!} + \sum_{k=0}^{\lfloor (N-1)/2 \rfloor} \frac{1}{k! (N-1-2k)!}\right). \] \end{con} This formula is derived as follows. For $k \geq 0$, denote by $P_k$ the probability that \eqref{wit} has exactly $k$ odd solutions in the range $3 \leq r < (p-1)/2$. By the discussion following Lemma \ref{Lemma1}, and the usual properties of the binomial distribution, for large $p$ our model suggests that \[ P_k = \binom{p/4 + O(1)}{k} \left(\frac1p\right)^k \left(1 - \frac1p\right)^{p/4 - k + O(1)} \rightarrow \frac{e^{-1/4}}{4^k k!}. \] Similarly, for $\ell \geq 0$ denote by $Q_\ell$ the probability that $T_r \equiv \pm 1 \pmod p$ has exactly $\ell$ even solutions in the range $2 \leq r < (p-1)/2$. Then \[ Q_\ell = \binom{p/4 + O(1)}{\ell} \left(\frac2p\right)^\ell \left(1 - \frac2p\right)^{p/4 - \ell + O(1)} \rightarrow \frac{e^{-1/2}}{2^\ell \ell!}. \] Assuming that the behaviour for odd and even $r$ is independent, the probability of observing exactly $N$ solutions for $2 \leq r \leq p - 3$, $r \neq (p-1)/2$, should be \[ \sum_{2k + \ell = N} P_k Q_\ell = \sum_{k=0}^{\lfloor N/2\rfloor} \frac{e^{-1/4}}{4^k k!} \cdot \frac{e^{-1/2}}{2^{N-2k} (N-2k)!} = \frac{e^{-3/4}}{2^N} \sum_{k=0}^{\lfloor N/2\rfloor} \frac1{k!(N-2k)!}. \] Finally, for $p \equiv 1 \pmod 4$, and for $p \equiv 3 \pmod 4$ with $\nu_p$ odd, the probability that \eqref{wit} has exactly $N$ solutions is given by the above formula (the exceptional value $r = (p-1)/2$ makes a negligible contribution asymptotically). For $p \equiv 3 \pmod 4$ with $\nu_p$ even, we must replace $N$ by $N-1$ to account for the automatic solution $r = (p-1)/2$. Our final estimated probability is thus \[ \frac34 \left(\frac{e^{-3/4}}{2^N} \sum_{k=0}^{\lfloor N/2\rfloor} \frac1{k!(N-2k)!} \right) + \frac 14 \left(\frac{e^{-3/4}}{2^{N-1}} \sum_{k=0}^{\lfloor (N-1)/2\rfloor} \frac1{k!(N-1-2k)!}\right). \] \section{Algorithms and computations} \label{sec:computations} We first consider the motivating problem, counting the number of nontrivial solutions to \eqref{wit}. For this the na\"ive algorithm appears to be the best available. For each prime $p$ up to some bound, we compute $T_2, T_3, \ldots$, by successive multiplication modulo $p$, and count how many times we see $-1$. We wrote a simple C implementation of this algorithm, paying some attention to efficient modular arithmetic. We ran it for all primes up to $10^8$. The running time was 22 hours on a 16-core 2.6 GHz Intel Xeon server. Table \ref{tab1} summarises the results. The last column shows the probabilities for each $N$ proposed in Conjecture \ref{ConN}; they are a superb fit for the observed proportions in the previous column. It is difficult to push the search bound higher. The running time for each prime is essentially linear in $p$, so the cost of handling all $p < x$ grows essentially quadratically in $x$. For example, to increase the search bound to $10^9$ would take about three months on the same hardware. We do not know of any asymptotically faster algorithms for this problem. \begin{table}[h] \centering \caption{Statistics of nontrivial solutions to \eqref{wit} for $p < 10^8$} \begin{tabular}{rrrr} \toprule $N$ & \# Primes with $N$ solutions & Proportion & Conjecture \ref{ConN} \\ \midrule 0 & 2041117 & 0.3542711 & 0.3542749 \\ 1 & 1701240 & 0.2952796 & 0.2952291 \\ 2 & 1104376 & 0.1916835 & 0.1918989 \\ 3 & 553921 & 0.0961426 & 0.0959495 \\ 4 & 232308 & 0.0403211 & 0.0402865 \\ 5 & 87019 & 0.0151037 & 0.0151612 \\ 6 & 29037 & 0.0050399 & 0.0050358 \\ 7 & 8887 & 0.0015425 & 0.0015638 \\ 8 & 2631 & 0.0004567 & 0.0004423 \\ 9 & 692 & 0.0001201 & 0.0001190 \\ 10 & 165 & 0.0000286 & 0.0000298 \\ 11 & 42 & 0.0000073 & 0.0000071 \\ 12 & 17 & 0.0000030 & 0.0000016 \\ 13 & 3 & 0.0000005 & 0.0000004 \\ \midrule Total & 5761455 & 1.0000000 & 1.0000000 \\ \bottomrule \end{tabular} \label{tab1} \end{table} Next we consider the problem of computing $((p-1)/2)! \pmod p$ for $p \equiv 3 \pmod 4$, in order to test Conjecture \ref{con1}. For this there is a greater variety of algorithms available. The na\"ive approach leads to an $O(p)$ algorithm as above (with a comparable implied constant). An algorithm with complexity $p^{1/2+o(1)}$ can be deduced from \cite{BGS}. We opted to implement an algorithm with average complexity only $(\log p)^{4 + o(1)}$ per prime, using the ``accumulating remainder tree'' technique introduced in \cite{CGH}. We give a brief sketch of this algorithm. Suppose that we wish to compute $r_p = ((p-1)/2)! \bmod p$ for all primes $p \equiv 3 \pmod 4$ in some interval $2M < p < 2N$, where $M$ and $N$ are positive integers. Consider the binary tree, with nodes indexed by pairs $(a, b)$, where $b > a > 0$ are integers, defined as follows. The root node is $(M, N)$. The children of a given node $(a, b)$ are $(a, c)$ and $(c, b)$, where $c = \lfloor (a + b)/2\rfloor$. For each node let \[ I_{a,b} = \{k \in \ZZ : \text{$k$ odd, } 2a < k < 2b\}. \] Thus at level $d$, the intervals $I_{a,b}$ partition $I_{M,N}$ into $2^d$ subintervals of roughly equal size. We stop at level $\ell = \lceil \log_2(N - M)\rceil$; at this level each $I_{a,b}$ has cardinality either zero or one. The algorithm now proceeds as follows. First, for each node let \[ P_{a,b} = \prod_{\substack{p \in I_{a,b} \\ p \equiv 3 \bmod 4 \\ \text{$p$ prime}}} p, \qquad V_{a,b} = \prod_{k \in I_{a,b}} \frac{k+1}2. \] Compute $V_{a,b}$ and $P_{a,b}$ for each node, working from the bottom of the tree to the top, using the identities $V_{a,b} = V_{a,c} V_{c,b}$ and $P_{a,b} = P_{a,c} P_{c,b}$. Second, for each node let \[ X_{a,b} = a! \bmod {P_{a,b}}. \] Compute $X_{M,N} = M! \bmod P_{M,N}$ using the method of Sch\"onhage (see for example \cite[Prop.~2.3]{CGH}). Then compute $X_{a,b}$ for each node, now working from the top of the tree downwards, using the formulae $X_{a,c} = X_{a,b} \bmod P_{a,c}$ and $X_{c,b} = X_{a,b} V_{a,c} \bmod P_{c,b}$ to descend from each node to its children. Finally, for each $p \equiv 3 \pmod 4$ in the interval $2M < p < 2N$, there is a unique node $(a, b)$ at level $\ell$ such that $p \in I_{a,b}$; for this node we have $I_{a,b} = \{p\}$, $P_{a,b} = p$ and $X_{a,b} = ((p-1)/2)! \pmod p$. For more details, including a complexity analysis, see \cite{CGH}. We mention here only that the complexity bound depends essentially on asymptotically fast algorithms for multiplication and division of large integers. Using a straightforward implementation of the above algorithm in the Sage computer algebra system \cite{sage}, we computed $r_p$ for all $p < 10^{10}$. To keep memory usage under control, we split the work into intervals $(M, N)$ of size $1.5 \times 10^8$. The total CPU time expended was 4.4 days. The results, shown in Table \ref{tab2}, are in excellent agreement with Conjecture \ref{con1}. \begin{table}[h] \centering \caption{Statistics of $r_p$ for $p < 10^{10}$, $p \equiv 3 \pmod 4$} \begin{tabular}{lrrr} \toprule $X$ & $\# \{p < X\}$ & $\# \{ p < X: r_p = 1 \}$ & proportion \\ \midrule $10^1$ & 2 & 1 & 0.5000000000 \\ $10^2$ & 13 & 6 & 0.4615384615 \\ $10^3$ & 87 & 43 & 0.4942528736 \\ $10^4$ & 619 & 310 & 0.5008077544 \\ $10^5$ & 4\,808 & 2\,418 & 0.5029118136 \\ $10^6$ & 39\,322 & 19\,704 & 0.5010935354 \\ $10^7$ & 332\,398 & 166\,270 & 0.5002135994 \\ $10^8$ & 2\,880\,950 & 1\,440\,268 & 0.4999281487 \\ $10^9$ & 25\,424\,042 & 12\,713\,329 & 0.5000514474 \\ $10^{10}$ & 227\,529\,235 & 113\,772\,462 & 0.5000344769 \\ \bottomrule \end{tabular} \label{tab2} \end{table} \section*{Acknowledgements} We are indebted to Professor Roger Heath-Brown for suggesting part of the argument leading to Conjectures \ref{Conmain} and \ref{ConN}. \bibliographystyle{plain}

https://arxiv.org/abs/1310.6013

Matchings and Hamilton Cycles with Constraints on Sets of Edges

The aim of this paper is to extend and generalise some work of Katona on the existence of perfect matchings or Hamilton cycles in graphs subject to certain constraints. The most general form of these constraints is that we are given a family of sets of edges of our graph and are not allowed to use all the edges of any member of this family. We consider two natural ways of expressing constraints of this kind using graphs and using set systems.For the first version we ask for conditions on regular bipartite graphs $G$ and $H$ for there to exist a perfect matching in $G$, no two edges of which form a $4$-cycle with two edges of $H$.In the second, we ask for conditions under which a Hamilton cycle in the complete graph (or equivalently a cyclic permutation) exists, with the property that it has no collection of intervals of prescribed lengths whose union is an element of a given family of sets. For instance we prove that the smallest family of $4$-sets with the property that every cyclic permutation of an $n$-set contains two adjacent pairs of points has size between $(1/9+o(1))n^2$ and $(1/2-o(1))n^2$. We also give bounds on the general version of this problem and on other natural special cases.We finish by raising numerous open problems and directions for further study.

\section{Introduction} Many results in graph theory concern establishing conditions on a graph $G$ which guarantee that $G$ must contain some particular spanning structure. A classical example of this is Dirac's theorem \cite{Dirac}: \begin{theorem}[Dirac]\label{dirac:thm} Every graph on $n$ vertices with minimum degree at least $n/2$ has a Hamilton cycle. \end{theorem} Demetrovics, Katona, and Sali \cite{DKS} proved an extension of this in which there is a second graph $H$ on the same vertex set as $G$, and we are required to find a Hamilton cycle in $G$ which satisfies the condition that no two edges of it form an ``alternating cycle'' with two edges of $H$. More precisely they proved: \begin{theorem}[Demetrovics, Katona, and Sali]\label{dks:thm} Let $G$ and $H$ be graphs with $V(G)=V(H)$ and $E(G)\cap E(H)=\emptyset$. Let $|V(G)|=n$, $r$ be the minimum degree of $G$, and $s$ be the maximum degree of $H$. Provided that \[ 2r-8s^2-s-1>n \] there is a Hamilton cycle in $G$ such that if $(a,b)$ and $(c,d)$ are both edges of the cycle, then $(b,c),(d,a)$ are not both edges of $H$. \end{theorem} We mention briefly that this theorem was proved with an application involving pairing up sets in mind. In this application, the graph $G$ has as vertices the $r$-element subsets of $[m]$ with disjoint subsets being adjacent. By a suitable choice of $H$ it was shown that the $r$-element subsets of $[m]$ can be paired up into disjoint pairs with distinct pairs being significantly different in some natural quantitative sense. See \cite{DKS} and \cite{K} for more details. Some further results on finding Hamilton cycles in graphs under constraints of this kind were proved and some questions were raised in \cite{K}. The aim of this paper is to answer, completely or partially, some of these questions, and to give some further results, questions and conjectures which may lead towards a more general theory of problems of this kind. In Section 2 we consider the following question and related problems. \begin{q}\label{regularmatch:q} Let $A,B$ be fixed $n$-sets and let $G,H$ be bipartite graphs on $A\cup B$ with bipartition $(A,B)$. Further, suppose that $G$ is $r$-regular, $H$ is $s$-regular and $G$ and $H$ have disjoint edge sets. For which $n,r,s$ are we guaranteed that there exists a perfect matching in $G$ in which no two edges of the matching form a 4-cycle with two edges of $H$. \end{q} Answers to this can be thought of as bipartite analogues of Theorem \ref{dks:thm}. The replacement of the Hamilton cycle in a graph of Theorem \ref{dks:thm} with a perfect matching in a bipartite graph is natural particularly given the pairing up sets applications from \cite{DKS} and \cite{K}. We give conditions on $r$ and $s$ for a perfect matching of this form to exist. In the case that $G$ and $H$ are bipartite complements (in the sense that they are edge disjoint and their union is $K_{n,n}$) we are able to bound the smallest $s$ for which a suitable perfect matching is guaranteed to exist between $c_1n^{1/2}$ and $c_2n^{3/4}$. Section 3 concerns generalisations of the following question from \cite{K}. \begin{q}\label{F:q} What is the smallest family $\mathcal{F}\subseteq[n]^{(4)}$ with the property that every Hamilton cycle in $K_n$ contains a pair of edges whose union is an element of $\mathcal{F}$. \end{q} In the general question $\mathcal{F}$ is a family of $r$-sets and we require that our Hamilton cycle does not have a set of intervals of given lengths whose union is an element of $\mathcal{F}$. To describe this generalisation more precisely, let $\mathcal{F}\subseteq[n]^{(r)}$, and ${\bf x}=(x_1,\dots,x_k)$ with $x_i$ positive integers and $\sum x_i=r$. We say that a Hamilton cycle in $G$ is ${\bf x}$-acceptable for $\mathcal{F}$ if we do not have intervals of vertices of lengths $x_1,\dots,x_k$ in the cycle whose union is an element of $\mathcal{F}$. We are interested in how small $|\mathcal{F}|$ can be if there is no ${\bf x}$-acceptable Hamilton cycle for $\mathcal{F}$. Although this question is implicitly present in \cite{K}, no general results or conjectures are made there. We give upper and lower bounds for the smallest such $|\mathcal{F}|$, each of the form $c n^{r-k}$, and conjecture that our general upper bound is asymptotically tight. We prove results on a number of special cases including improved bounds when ${\bf x}=(2,2)$, an asymptotic determination of the extremal function when ${\bf x}=(2,\underbrace{1,\dots,1}_{r-2})$, and an answer to a question from \cite{K} on the ${\bf x}=(3,1)$ case. In the final section we suggest a number of further questions and directions for study. We note that a more general context for expressing this kind of problem is the notion of finding structures under constraints which forbid certain sets of edges all being used. We are given a graph $G$ and a family of sets of edges. Under what conditions on $G$ and this family can we guarantee the existence of some structure (typically a Hamilton cycle or a perfect matching) the edge set of which does not contain any member of the forbidden family. For example, the conditions of Theorem \ref{dks:thm} can be phrased as: for every pair of edges $(a,b),(c,d)\in E(G)$ for which $(b,c),(d,a)\in E(H)$ we cannot use both $(a,b)$ and $(c,d)$ in our Hamilton cycle. Both the constraints coming from a graph looked at in Section 2 and the constraints coming from a set system considered in Section 3 are also of this form. It may be interesting to consider different ways of expressing constraints of this kind in addition to the two considered here. Finally we remark that although the case that $G$ is a complete graph considered as \ref{F:q} looks rather special, we feel this may be a potentially good first step towards understanding the general behaviour. Another justification for this special case is that it naturally generalises some classical graph theoretic results. Dirac's theorem for instance can be stated (somewhat perversely) as: ``If $F\subseteq V^{(2)}$ is a set of forbidden edges then provided that no vertex is incident with more than $n/2$ elements of $F$ the graph $K_n$ contains a Hamilton cycle which does not contain any element of $F$.'' In our problem we replace the condition that certain edges are not allowed to be used with the condition that none of a certain family of \emph{sets} of edges is allowed to be all used . \section{Matchings with Constraints given by a graph} Throughout this section, by bipartite graph we mean a graph on vertex set $V=A\cup B$ with bipartition $(A,B)$ and $|A|=|B|=n$. In other words when we speak of several bipartite graphs they always have a single fixed bipartition. If $G$ and $H$ are bipartite graphs we say that a perfect matching $M$ in $G$ is \emph{acceptable} for $H$ if no two edges of $M$ form a 4-cycle with two edges of $H$. The general question is: \begin{q}\label{match:q} Under what conditions on the degrees of $G$ and $H$ can we guarantee the existence of a perfect matching in $G$ which is acceptable for $H$? \end{q} It is natural to consider the case that $G$ and $H$ are regular graphs. It will be convenient to assume further that the edge sets of $G$ and $H$ are disjoint. This brings us to Question \ref{regularmatch:q} from the introduction. We have the following positive result which applies when $r$ is reasonably large. \begin{theorem}\label{match-deg:thm} Let $G$ be an $r$-regular bipartite graph and $H$ be an $s$-regular bipartite graph with $E(G)\cap E(H)=\emptyset$. Provided that $r>\frac{n}{2}+s^2$ there exists a perfect matching in $G$ which is acceptable for $H$. \end{theorem} \begin{proof} Since $G$ is regular it certainly contains at least one perfect matching. Suppose, for a contradiction, that $G$ does not contain a perfect matching which is acceptable for $H$. Remove edges from $H$ one by one until one of the perfect matchings in $G$ becomes acceptable. Let $H'$ be the subgraph we are left with, and $e$ be the last edge removed. Let $M$ be any perfect matching in $G$ which is acceptable for $H'$. Label the vertices so that $A=\{a_1,\dots,a_n\}$, $B=\{b_1,\dots,b_n\}$ with $M=\{(a_i,b_i): 1\leq i\leq n\}$ and $e=(a_n,b_{n-1})$. Adding $e$ to $H'$ makes $M$ unacceptable but this can only be because of the $4$-cycle $a_n,b_n,a_{n-1},b_{n-1}$. Hence, the matching $M\setminus (a_n,b_n)$ of $n-1$ edges is acceptable for $H'\cup e$. Let $Y=\Gamma_H(a_n)\subseteq B$ and $X=\Gamma_H(b_n)\subseteq A$. Let $X'=\{b_i: a_i\in X\}$, $Y'=\{a_i:b_i\in Y\}$. Finally, let $X''=\Gamma_H(X')\subseteq A$ and $Y''=\Gamma_H(Y')\subseteq B$. Since $H$ is $s$-regular $|X''|, |Y''|\leq s^2$. Since $deg_G(a_n)-|Y''|\geq r-s^2>n/2$ and similarly $deg_G(b_n)-|X''|\geq r-s^2>n/2$ we can find $t$ so that $(a_n,b_t),(a_t,b_n)\in E(G)$, $a_t\not\in X''$, $b_t\not\in Y''$. We claim that $M'=M\setminus \{(a_n,b_n),(a_t,b_t)\}\cup\{(a_t,b_n),(a_n,b_t)\}$ is acceptable for $H'\cup e$. This will contradict the definition of $H'$ and $e$. Suppose that we have a bad $4$-cycle containing two edges of $M'$ and two edges of $H'\cup e$. Since $M\setminus (a_n,b_n)$ is acceptable for $H'\cup e$ our bad cycle must contain at least one edge of $(a_t,b_n),(a_n,b_t)$. Further, since $a_n,b_n\not\in E(H)$ (as $E(G)\cap E(H)=\emptyset$) it must contain exactly one of the edges. Suppose that our bad cycle is $a_n,b_t,a_s,b_s$ with $(a_n,b_s),(a_s,b_t)\in E(H)$. This means that $b_s\in Y$, $a_s\in Y'$ and $b_t\in Y''$, contradicting our choice of $t$. Similarly, by the definition of $X,X'$ and $X''$, we cannot have that our bad cycle is $b_n,a_t,b_s,a_s$. It follows that $M'$ is acceptable for $H'\cup e$ and this contradiction completes the proof. \end{proof} A special case of this result gives a condition for an acceptable matching when $E(H)=E(K_{n,n})\setminus E(G)$. To express this we introduce some notation. Suppose that $G$ is a bipartite graph. We write $\overline{G}$ for the bipartite graph with edge set $\{ab: a\in A, b\in B, ab\not\in E(G)\}$. Let $b(n)$ be the minimum number $s$ such that there exists an $s$-regular bipartite graph $H$ for which there is no perfect matching in $\overline{H}$ which is acceptable for $H$. \begin{cor}\label{match-comp-lb:cor} For all $n$ we have that \[ b(n)>\frac{1}{2}\left(\sqrt{2n+1}-1\right)=(\sqrt{2}+o(1))n^{1/2}. \] \end{cor} \begin{proof} If $H$ is an $s$-regular graph then $G=\overline{H}$ is an $n-s$ regular graph. Theorem \ref{match-deg:thm} shows that we have a perfect matching in $G$ which is acceptable for $H$ if $n-s>\frac{n}{2}+s^2$. That is if $s<\frac{1}{2}\left(\sqrt{2n+1}-1\right)$. \end{proof} A construction gives an upper bound for $b(n)$. \begin{theorem}\label{match-comp-ub:thm} For infinitely many $n$ we have that \[ b(n)<(2+o(1))n^{3/4}. \] \end{theorem} The construction we give works for all $n$ of the form $m\lfloor m^{1/3}\rfloor$ where $m=\frac{q^4-1}{q-1}$ with $q$ a prime power. It appears to be surprisingly difficult to deduce anything valid for all values of $n$ from this. The fact that we do not even know whether the function $b(n)$ is non-decreasing in $n$ contributes to this difficulty. An ingredient of this construction is a bipartite graph with good expansion properties as described in the following result of Alon \cite{Alon}. \begin{theorem}[Alon \cite{Alon}]\label{alon:thm} For any integer $d\geq 1$ and $n=\frac{q^{d+1}-1}{q-1}$ with $q$ a prime power, there is an $r$-regular bipartite graph with $r=\frac{q^d-1}{q-1}=(1+o(1))n^{1-1/d}$ such that for all $0<x<n$ and any $X\subseteq A$ with $|X|=x$ we have $|\Gamma(X)|\geq n-\frac{n^{1+1/d}}{x}$. \end{theorem} \begin{proof}[Proof of Theorem \ref{match-comp-ub:thm}] Let $m=\frac{q^4-1}{q-1}$ and $n=m\lfloor m^{1/3}\rfloor$. We need to construct a regular bipartite graph $H$ of degree $(2+o(1))n^{3/4}$ with the property that no perfect matching in $G=\overline{H}$ is acceptable for $H$. We first partition the vertices in $A$ into $k=\lfloor m^{1/3}\rfloor$ parts $A_1,A_2,\dots,A_k$ each of size $m$. Similarly we partition the vertices in $B$ into $k=\lfloor m^{1/3}\rfloor$ parts $B_1,B_2,\dots,B_k$ each of size $m$. Let $B(m)$ be a bipartite graph with $m$ vertices in each part which satisfies the conditions of the $d=3$ case of Theorem \ref{alon:thm}. This graph is regular of degree $(1+o(1))m^{2/3}$. We form $H$ by putting a copy of $K_{m,m}$ between each pair $A_i,B_i$ for $1\leq i\leq k$ and a copy of $B(m)$ between each pair $A_i,B_j$ with $i\not=j$. The graph $H$ is clearly regular of degree $m+(k-1)(1+o(1))m^{2/3}=(2+o(1))n^{3/4}$. Suppose that $M$ is any perfect matching in $G$. Let $Y=\{b\in B: (a,b)\in M \text{ for some } a\in A_1\}$ and $X=\{a\in B: (a,b)\in M \text{ for some } b\in B_1\}$. Since $|X|=|Y|=m$ and $X\cap A_1=Y\cap B_1=\emptyset$ there is some $s$ with $|X\cap A_s|\geq m/(k-1)>m^{2/3}$ and some $t$ with $|Y\cap B_t|\geq m/(k-1)>m^{2/3}$. Now, if $s=t$ we have a $4$-cycle consisting of an edge of $M$ between $A_1$ and $B_t$, an edge of $M$ between $A_s$ and $B_1$, and two edges from copies of $K_{m,m}$. If $s\not=t$ then the construction of $B(m)$ implies that there is some edge of $H$ between $X\cap A_s$ and $Y\cap B_t$. If not then $X\cap A_s$ has less than $m-|Y\cap B_t|\leq m-m^{2/3}$ neighbours in $B_s$ contradicting the defintion of $B(m)$. It follows that we have a $4$-cycle consisting of this edge, an edge of $M$ between $A_1$ and $B_t$, an edge of $M$ between $A_s$ and $B_1$, and an edge from the copy of $K_{m,m}$ between $A_1$ and $B_1$. We conclude that $H$ does have the property that no perfect matching in $G$ is acceptable for $H$. \end{proof} In fact, if we only require an $H$ with small maximum degree with the property that no perfect matching in $\overline{H}$ is acceptable for $H$ a slightly simpler construction works. Assuming $n$ is of the form $\frac{q^4-1}{q-1}$ start with a copy of $B(n)$ as defined in the proof above. Take subsets $X\subseteq A$, $Y\subseteq B$ with $|X|=|Y|=n^{2/3}+1$ and add all edges between $X$ and $Y$ to form a graph $H$. Since there is an edge of $B(n)$ between any two subsets of size $n^{2/3}+1$, there is an edge of $B(n)$ between the vertices matched to $X$ and the vertices matched to $Y$ for any perfect matching in $\overline{H}$. It follows that no perfect matching in $\overline{H}$ is acceptable for $H$. The graph $H$ has maximum degree at most $2n^{2/3}+1$ although it is not regular. Turning now to the more general question in which $G$ and $H$ are not required to be complementary Theorem \ref{match-deg:thm} is rather weak since the condition only holds when the degree of $G$ is quite large. For smaller degrees of $G$ we have a negative result obtained by taking copies of the graph constructed in the proof of Theorem \ref{match-comp-ub:thm}. \begin{cor} For any $r$ there exists, for infinitely many $n$, an $r$-regular graph $G$ and a $(2+o(1))r^{3/4}$-regular graph $H$ with disjoint edge sets and with the property that no perfect matching in $G$ is acceptable for $H$. \end{cor} \begin{proof} Take vertex disjoint copies of the graphs constructed in the proof of Theorem \ref{match-comp-ub:thm}. \end{proof} Positive results in this small degree of $G$ case seem to be harder to prove. Indeed, we can not answer even the following apparently simple question: \begin{q}\label{match-small:q} Is there an integer $k$ such that for all sufficiently large $n$, if $G$ is a $k$-regular bipartite graph and $H$ is a 1-regular bipartite graph, we are guaranteed to have a perfect matching in $G$ which is acceptable for $H$? \end{q} We remark briefly that the conclusion does not hold if $k=2$ since such a 2-regular $G$ may have only 2 perfect matchings and it is easy to choose a 1-regular $H$ so that neither of them is acceptable (provided that $n\geq7$). However, it may be that even $k=3$ is sufficient for a positive answer. \section{Constraints given by a set system} \subsection{Hamilton cycles} We turn now to constraints described in a different form. If $\mathcal{F}\subseteq[n]^{(4)}$ and $G$ is a graph with vertex set $[n]$ we say that a Hamilton cycle in $G$ is a $(2,2)$-acceptable Hamilton cycle with respect to $\mathcal{F}$ if we do not have two disjoint edges of the cycle whose union is an element of $\mathcal{F}$. We will mainly be interested in the case when $G=K_n$. This leads us to Question \ref{F:q} from the introduction which can now be rephrased as: \begin{q}\label{F22:q} What is the smallest family $\mathcal{F}\subseteq[n]^{(4)}$ with the property that $K_n$ does not contain a $(2,2)$-acceptable Hamilton cycle with respect to $\mathcal{F}$? \end{q} A Hamilton cycle in $K_n$ can be thought of as a cyclic ordering of the vertices and we shall represent our Hamilton cycles by such a string of vertices. For a more general question, suppose that $G$ is a graph with vertex set $[n]$, $\mathcal{F}\subseteq[n]^{(r)}$, and ${\bf x}=(x_1,\dots,x_k)$ with $x_i$ positive integers and $\sum x_i=r$. We say that a Hamilton cycle $c_1,c_2,c_3,\dots,c_n$ in $G$ is ${\bf x}$-acceptable for $\mathcal{F}$ if we do not have $t_1,\dots,t_k$ with \[ \bigcup_{i=1}^k\{c_{t_i+1},c_{t_i+2},\dots,c_{t_i+x_i}\}\in\mathcal{F}. \] (Here and elsewhere we interpret suffices modulo $n$.) We generalise Question \ref{F22:q} in the obvious way. \begin{q}\label{Fx:q} Given ${\bf x}$ as above what is the smallest family $\mathcal{F}\subseteq[n]^{(r)}$ with the property that $K_n$ does not contain an ${\bf x}$-acceptable Hamilton cycle with respect to $\mathcal{F}$? \end{q} We will denote the size of this smallest family $\mathcal{F}$ by $m({\bf x},n)$. This question of determining $m({\bf x},n)$ was essentially raised (with slightly different notation) in \cite{K} where a number of constructions relating to particular cases of it are given. Our aim is to improve the bounds in some of these and other special cases, and also to consider what general results may hold. We will generally be interested in asymptotic results in which $n$ tends to infinity with ${\bf x}$ (and hence $r$) being fixed. A simple averaging argument gives a lower bound on $m({\bf x},n)$ for any ${\bf x}$. \begin{theorem}\label{mxn-lb:thm} Given ${\bf x}$ as above we have that \[ m({\bf x},n)\geq\left(\frac{1}{c({\bf x})x_1!x_2!\dots x_k!}+o(1)\right)n^{r-k}, \] where $c({\bf x})$ is the number of unordered partitions of an $r$-set into $k$ sets of sizes $x_1,x_2,\dots,x_k$ (and in particular does not depend on $n$). \end{theorem} Let $S$ be the set of all Hamilton cycles in $K_n$ (cyclic orderings of the vertices). Clearly $|S|=(n-1)!$. Given a set $F\in[n]^{(r)}$ we denote by $H(F)$ the set of all elements of $S$ which are not acceptable with respect to the single set $F$ (that is all cyclic orderings in which $F$ is a union of disjoint intervals of the appropriate size). Of course $H(F)$ also depends on ${\bf x}$ but it will always be clear from the context what ${\bf x}$ is and so this notation should cause no confusion. \begin{proof} If $\mathcal{F}$ is such that $K_n$ has no ${\bf x}$-acceptable Hamilton cycle with respect to $\mathcal{F}$ then $\bigcup_{F\in\mathcal{F}}H(F)=S$ and so $\sum_{F\in\mathcal{F}}|H(F)|\geq(n-1)!$. Now $|H(F)|$ does not depend on $F$ and $|H(F)|\leq c({\bf x})x_1!x_2!\dots x_k!(n-r+k-1)!$ (for certain ${\bf x}$ this will be an equality but in some cases it will be possible to have a Hamilton cycle for which $F$ can be represented as the union of suitable intervals in more than one way). It follows that \[ c({\bf x})x_1!x_2!\dots x_k!(n-r+k-1)!|\mathcal{F}|\geq(n-1)! \] and so \[ |\mathcal{F}|\geq\frac{(n-1)!}{c({\bf x})x_1!\dots x_k!(n-r+k-1)!}=\left(\frac{1}{c({\bf x})x_1!x_2!\dots x_k!}+o(1)\right)n^{r-k} \] as required. \end{proof} In fact this lower bound gives the correct order of magnitude for $m({\bf x},n)$ as the following theorem shows. \begin{theorem}\label{mxn-ub:thm} Given ${\bf x}$ as above and $n\geq r$ we have that \[ m({\bf x},n)\leq\binom{n-k}{r-k}=\left(\frac{1}{(r-k)!}+o(1)\right)n^{r-k}. \] \end{theorem} \begin{proof} Let $\mathcal{F}=\{X\in[n]^{(r)}: \{1,2,\dots,k\}\subseteq X\}$. We will show that in any permutation of $[n]$ it is possible to find $k$ disjoint intervals of lengths $x_1,x_2,\dots,x_k$ such that the union of these intervals contains $[k]$. This clearly shows that there is no ${\bf x}$-acceptable Hamilton cycle with respect to $\mathcal{F}$. We will prove this claim by induction on $n$. If $n=1$ then the claim obviously holds. It is also clearly true if $n=r$. Suppose that $n>1$, $r<n$ and $c_1,\dots,c_n$ is our permutation. If $c_1\not\in[k]$ then applying the induction hypothesis to the permutation $c_2,\dots,c_n$ gives the result. If $c_1\in[k]$ then we will take our first interval to be $c_1,\dots,c_{x_1}$ and consider the permutation $c_{x_1+1},\dots,c_n$. Applying the induction hypothesis to this permutation with vector of interval lengths $(x_2,\dots,x_k)$ gives the result. (It may be that we have fewer than $k-1$ elements of $[k]$ contained in this permutation but this can only weaken the condition we need to satisfy.) \end{proof} If some of the $x_i$ are equal to 1 then the bound of Theorem \ref{mxn-ub:thm} is not sharp. In this situation we have the following stronger result. \begin{theorem}\label{mxn-ub2:thm} If ${\bf x}=(x_1,\dots,x_k)$ with $x_1\geq\dots\geq x_t>x_{t+1}=\dots=x_k=1$ then \[ m({\bf x},n)\leq(1+o(1))\frac{\binom{n-t}{r-k}}{\binom{r-t}{r-k}}=\left(\frac{(k-t)!}{(r-t)!}+o(1)\right)n^{r-k}. \] \end{theorem} \begin{proof} Let $\mathcal{M}$ be the smallest family of $(r-t)$-subsets of $[n]\setminus[t]$ with the property that every $(r-k)$-subset of $[n]\setminus[t]$ is contained in at least one set in $\mathcal{M}$. By R\"odl's proof of the Erd\H os-Hanani conjecture \cite{Rodl} we have that $|\mathcal{M}|=(1+o(1))\frac{\binom{n-t}{r-k}}{\binom{r-t}{r-k}}=\left(\frac{(k-t)!}{(r-t)!}+o(1)\right)n^{r-k}$. Now let $\mathcal{F}=\{X\in[n]^{(r)}: [t]\subseteq X, X\setminus[t]\in \mathcal{M}\}$. We will show that there is no ${\bf x}$-acceptable Hamilton cycle with respect to $\mathcal{F}$. From which the required upper bound on $m({\bf x},n)$ follows. As in the proof of Theorem \ref{mxn-ub:thm}, in any permutation of $[n]$ it is possible to find $t$ disjoint intervals $X_1,\dots,X_t$ of lengths $x_1,\dots,x_t$ such that the union of these intervals contains $[t]$. Now $\left(X_1\cup\dots\cup X_t\right)\setminus[t]$ is an $(r-k)$-subset of $[n]\setminus[t]$ and so there is some $(r-t)$-set $M\in\mathcal{M}$ which contains it. It follows that the intervals $X_1,\dots,X_t$ together with $k-t$ singleton intervals form a set in $\mathcal{F}$. It follows that no Hamilton cycle is ${\bf x}$-acceptable with respect to $\mathcal{F}$. \end{proof} We tentatively conjecture that the upper bounds of Theorems \ref{mxn-ub:thm} and \ref{mxn-ub2:thm} are asymptotically sharp under the appropriate conditions on ${\bf x}$. \begin{conj}\label{mxn:conj} If $x_i\not=1$ for all $i$ then \[ m({\bf x},n)=\left(\frac{1}{(r-k)!}+o(1)\right)n^{r-k}. \] If ${\bf x}=(x_1,\dots,x_k)$ with $x_1\geq\dots\geq x_t>x_{t+1}=\dots=x_k=1$ then \[ m({\bf x},n)=\left(\frac{(k-t)!}{(r-t)!}+o(1)\right)n^{r-k}. \] \end{conj} As we shall see later the exact upper bound of Theorem \ref{mxn-ub:thm} is not always correct even when $x_i\not=1$ for all $i$. This slight improvement suggest that even if the conjecture is correct, the extremal families may have quite a complicated structure. The next result improves the bound given by the averaging argument of Theorem \ref{mxn-lb:thm} (except in the trivial case when $x_i=1$ for all $i$). \begin{theorem}\label{mxn-lb2:thm} Given ${\bf x}$ as above with the $x_i$ not all equal to 1 we have that \[ m({\bf x},n)\geq\frac{4kr+1}{4kr}\left(\frac{1}{c({\bf x})x_1!x_2!\dots x_k!}+o(1)\right)n^{r-k}. \] \end{theorem} The proof is by showing that the assumptions made in the proof of Theorem \ref{mxn-lb:thm} cannot hold with equality. The main aim of this result is to demonstrate that Theorem \ref{mxn-lb:thm} is not sharp and we have not made a particular effort to obtain the strongest bound this method will give. We will however go through the details more carefully in one special case later. If $C$ is a Hamilton cycle (cyclic ordering) then let $d(C)$ be the number of $F\in\mathcal{F}$ for which $F$ consists of $k$ disjoint intervals of lengths $x_1,\dots,x_k$ in $S$ (in other words the number of $F\in\mathcal{F}$ for which $C\in H(F)$). Strictly $d(C)$ depends on $\mathcal{F}$ and ${\bf x}$ as well as $C$ but it will always be clear from the context what these are. \begin{proof} We will assume that $x_1=t$ is the maximum of the $x_i$. Let $\mathcal{F}$ be such that there is no ${\bf x}$-acceptable Hamilton cycle with respect to $\mathcal{F}$. We may assume also that $|\mathcal{F}|\leq\binom{n-k}{r-k}$ since we know that this is an upper bound for $m({\bf x},n)$. We have that $\sum_{F\in\mathcal{F}}|H(F)|=\sum_{C\in S}d(C)$. The property that there is no ${\bf x}$-acceptable Hamilton cycle with respect to $\mathcal{F}$ is equivalent to having $d(C)\geq1$ for all $C\in S$. Previously we used this bound on $d(C)$ to deduce a bound on $\mathcal{F}$. Here we will show that it is not possible for all the $d(C)$ to be this small and so obtain a stronger bound on their sum. Let \[ U'=\{C\in S : d(C)=1\}. \] For a cycle in $U'$ there is a unique $F\in\mathcal{F}$ for which $C\in H(F)$. Let $U$ be the set of all cycles in $U'$ for which the end points of any two of the $k$ intervals in $C$ whose union is this $F$ are at distance at least $r$ around $C$. We have that each $F\in \mathcal{F}$ gives rise to at most $c(n-r+k-2)!$ cycles in $U'\setminus U$ where $c$ is some constant depending only on ${\bf x}$. It follows that $|U'\setminus U|\leq |\mathcal{F}|c(n-r+k-2)!=o((n-1)!)$. Hence, if we can show that $|U|\leq\alpha(n-1)!$ then we will have $|U'|\leq(\alpha+o(1))(n-1)!$. Our aim is to prove such a bound by showing that each cycle in $U$ gives rise to a cycle not in $U'$. Suppose that $C\in U$. Without loss of generality we may assume that $C=1,2,\dots,n$ and that $C\in H(F)$ where $F=\bigcup_{i=1}^k\{a_i+1,\dots,a_i+x_i\}$ with $a_1=1$. Consider the cycle $C'=1,2,\dots,t-1,t+1,t,t+2,\dots,n$ (that is $C$ with $t$ and $t+1$ exchanged). This is not in $H(F)$ since $F$ meets $C'$ in $k+1$ intervals (we are using here that $C\not\in U'$). Because there is no ${\bf x}$-acceptable Hamilton cycle we must have $C'\in H(F')$ for some $F'\in\mathcal{F}$. Now since $C\not\in H(F')$ we must have that either $t-1,t+1\in F'$, $t\not\in F'$ (Case 1) or $t,t+2\in F'$, $t+1\not\in F'$ (Case 2). If we are in Case 1 then one of the $k$ intervals in $C'$ which is contained in $F'$ must be the length $t-i+1$ interval $i,i+1,\dots,t-1,t+1$. By our choice of $x_1$ to be the largest $x_i$ we have that $i\geq1$ and so the interval must be of this form. Let $C''$ be the cycle \[ 1,2,\dots,i-1,t,i,i+1,\dots,t-1,t+1,t+2,\dots, n. \] Now it is clear that $C''\in H(F)$ and also $C''\in H(F')$ and so $d(C'')\geq2$. If we are in Case 2 then suppose that one of the $k$ intervals in $C'$ which is contained in $F'$ is the length $i-t$ interval $t,t+2,t+3,\dots,i$. Let $C''$ be the cycle \[ 1,2,\dots,t-1,t,t+2,t+3,\dots,i,t+1,i+1,\dots,n. \] Now it is clear that $C''\in H(F')$. We also know, since the intervals of $C$ whose union is $F$ are at distance at least $r$ round the cycle, that none of $t+1,\dots,i$ are elements of $F$ (this is where we use the fact that we are working in $U$ and not just $U'$). It follows that $C''\in H(F)$ and so $d(C'')\geq2$. Notice finally that in each case $C''$ was constructed from $C$ by moving one vertex by $x_i$ places in the cycle for some $i$. It follows that each cycle in $C''\in S\setminus U'$ can arise in this way from at most $4kr$ different cycles in $U$ (if $F,G\in\mathcal{F}$ with $C''\in H(F)\cap H(G)$ then the moved vertex must be in $F\cup G$ and there are at most $2k$ choices for where to move it). Hence \[ |U|\leq 4kr|S\setminus U| \] and so \[ |U|\leq\frac{4kr}{4kr+1}|S| \] Now since $U'\setminus U=o((n-1)!)$ we have that $U'\leq\left(\frac{4kr}{4kr+1}+o(1)\right)(n-1)!$. Using the same approach as in Theorem \ref{mxn-lb:thm} \[ c({\bf x})x_1!x_2!\dots x_k!(n-r+k-1)!|\mathcal{F}|\geq(n-1)!+\left(\frac{1}{4kr+1}+o(1)\right)(n-1)! \] That is \[ |F|\geq\left(\frac{4kr+1}{4kr}\frac{1}{c({\bf x})x_1!x_2!\dots x_k!}+o(1)\right)n^{r-k} \] \end{proof} We now address some natural special cases of our problem of determining $m({\bf x},n)$. One, mentioned in \cite{K}, is the case ${\bf x}=(2,2)$ which we referred to earlier. Theorems \ref{mxn-ub:thm} and \ref{mxn-lb:thm} show that \[ \left(\frac{1}{12}+o(1)\right)n^2\leq m((2,2),n)\leq\left(\frac{1}{2}+o(1)\right)n^2. \] Going through a similar argument to Theorem \ref{mxn-lb2:thm} with more care and refining the construction of Theorem \ref{mxn-ub:thm} we are able to improve the constant in the lower bound and the $o(n^2)$ term in the upper bound as follows \begin{theorem}\label{22:thm} For some constant $c>0$ we have \[ \left(\frac{1}{9}+o(1)\right)n^2\leq m((2,2),n)\leq\frac{1}{2}n^2-\frac{1}{2}n^{3/2} \] \end{theorem} \begin{proof} For the upper bound let $H$ be a graph with $V(H)=\{3,4,\dots,n\}$ which is $C_4$-free, triangle-free, and contains no Hamilton path. We will show that the there are no $(2,2)$-acceptable Hamilton cycles with respect to the family \[ \mathcal{F}=\{X\in[n]^{(4)}:1,2\in X, X\setminus\{1,2\}\not\in E(H)\}. \] The simplest case of the K\H ov\'ari-S\'os-Tur\'an Theorem \cite{KST} shows that there is a $C_4$-free bipartite graph with at most $1/2n^{3/2}+n$ edges. A suitable $H$ may be constructed from such a graph by deleting at most $n$ edges. It follows that we may take $|E(H)|=1/2n^{3/2}$ which will establish the upper bound. To prove that $\mathcal{F}$ is as required we show that any Hamilton cycle has two disjoint edges whose union is an element of $\mathcal{F}$. Suppose that our Hamilton cycle is $C=c_1,\dots,c_n$ with $c_i=1$, $c_j=2$ and (without loss of generality) $1<i<j<n$. We consider three cases: \begin{enumerate} \item[Case 1:] If $j-i>2$ then consider the pairs $(c_{i-1},c_{j-1})(c_{i-1},c_{j+1}),(c_{i+1},c_{j-1}),(c_{i+1},c_{j+1})$. Since $H$ is $C_4$-free at least one of these pairs is not an edge of $H$. It follows that at least one of these pairs together with 1,2 forms a 4-set in $\mathcal{F}$ and so $C$ is not $(2,2)$-acceptable with respect to $\mathcal{F}$. \item[Case 2:] If $j-i=2$ the consider the pairs $(c_{i-1},c_{i+1}),(c_{i-1},c_{i+3}),(c_{i+1},c_{i+3})$. Since $H$ is triangle-free at least one of these pairs is not an edge of $H$. It follows that at least one of these pairs together with 1,2 forms a 4-set in $\mathcal{F}$ and so $C$ is not $(2,2)$-acceptable with respect to $\mathcal{F}$. \item[Case 3:] If $j-i=1$ then consider the pairs $(c_{j+1},c_{j+2}),(c_{j+2},c_{j+3}),\dots,(c_{i-2},c_{i-1})$. Since $H$ does not have a Hamilton path at least one of these pairs is not an edge of $H$. It follows that at least one of these pairs together with 1,2 forms a 4-set in $\mathcal{F}$ and so $C$ is not $(2,2)$-acceptable with respect to $\mathcal{F}$. \end{enumerate} For the lower bound let $\mathcal{F}$ be such that there is no $(2,2)$-acceptable Hamilton cycle with respect to $\mathcal{F}$ and assume that $|\mathcal{F}|\leq\binom{n-2}{2}$. As before we have that $\sum_{F\in\mathcal{F}}|H(F)|=\sum_{C\in S}d(C)$. Let \[ U'=\{C\in S : d(C)=1\}. \] For a cycle $C\in U'$ there is a unique $F\in\mathcal{F}$ for which $C\in H(F)$. Let $U$ be the set of all cycles in $U'$ for which the end points of the 2 edges whose union is this $F$ are at distance at least 3 around $C$. As in the proof of Theorem \ref{mxn-lb2:thm} $|U'\setminus U|=o((n-1)!)$. We also define \begin{align*} D&=\{C\in S : d(C)=2\}\\ T&=\{C\in S : d(C)\geq3\}. \end{align*} If $i$ and $j$ are consecutive elements of a cyclic ordering $C$ we will denote by $\pi_{i,j}(C)$ the cyclic ordering formed by swapping $i$ and $j$.We will refer to cycles $B$ and $C$ with $C=\pi_{i,j}(B)$ as being neighbouring. Suppose that the cycle $C$ has $d(C)=1$ and let $F=\{a,b,x,y\}\in\mathcal{F}$ with $a$ and $b$ consecutive in $C$ and $x$ and $y$ consecutive in $C$. Suppose that $a,b,c,d$ are consecutive in $C$ and consider the cycle $C'=\pi_{b,c}(C)$. There is a set $F'\in\mathcal{F}$ with $C'\in H(F')$ and we must have either $a,c\in F'$ (Case 1) or $b,d\in F'$ (Case 2). Since $C\not\in H(F')$ we cannot have that $F'=\{a,b,c,d\}$ and so exactly one of these cases occurs. If we are in Case 1 then the cycle $C''=\pi_{a,b}(C)$ is in both $H(F)$ and $H(F')$ and so $d(C'')\geq2$. If we are in Case 2 then the cycle $C''=\pi_{d,c}(C)$ is in both $H(F)$ and $H(F')$ and so $d(C'')\geq2$ (note that here we need the condition which distinguishes $U$ from $U'$ to ensure that $C''\in H(F)$). Repeating the same argument starting from $C'=\pi_{z,a}(C)$ where $z$ is the predecessor of $a$ in $C$ we obtain that $C$ has either at least 2 neighbouring cycles in $D\cup T$ or at least 1 neighbouring cycle in $T$. Repeating the same argument starting with the pair $x,y$ we obtain that $C$ has one of the following: \begin{itemize} \item at least 2 neighbouring cycles in $T$, \item at least 3 neighbouring cycle in $D\cup T$ of which at least 1 is in $T$, \item at least 4 neighbouring cycles in $D\cup T$. \end{itemize} Writing $e(U,T)$ for the number of pairs of neighbouring cycles with one in $U$ and one in $T$ and similarly for $e(U,D)$ we obtain that \[ 4|U|\leq 2e(U,T)+e(U,D). \] It is easy to see that a cycle in $D\cup T$ can neighbour at most 8 cycles in $U$ and so $e(U,T)\leq 8|T|$, $e(U,D)\leq8|D|$. From this we get that \[ |U|\leq 4|T|+2|D|. \] Now \[ 12(n-3)!|\mathcal{F}|=\sum_{F\in\mathcal{F}}|H(F)|=\sum_{C\in S}d(C)\geq|S|+|D|+2|T|\geq|S|+\frac{1}{2}|U|. \] Also \[ 12(n-3)!|\mathcal{F}|\geq|S|+(|S|-|U'|)=2|S|-|U|+o((n-1)!). \] Taking whichever of these bounds is stronger depending on $|U|$ we conclude that \[ 12(n-3)!|\mathcal{F}|\geq\frac{4}{3}|S|+o((n-1)!) \] and so \[ |\mathcal{F}|\geq\left(\frac{1}{9}+o(1)\right)n^2. \] \end{proof} Another natural case is ${\bf x}=(2,1,..,1)$ . Here we have a lower bound which agrees asymptotically with the upper bound from Theorem \ref{mxn-ub2:thm}. \begin{theorem}\label{211:thm} Let ${\bf x}=(2,\underbrace{1,\dots,1}_{r-2})$ with $r$ fixed. We have \[ m({\bf x},n)=\left(\frac{1}{r-1}+o(1)\right)n. \] \end{theorem} It is worth noting that in this case $m({\bf x},n)$ can be expressed more directly in graph theoretic terms; it is the smallest number of copies of $K_r$ whose removal makes $K_n$ non-Hamiltonian. We will use this formulation in the proof below. We will also the use the well known Bondy-Chv\'atal Theorem \cite{BC} which characterises Hamiltonian graphs. \begin{theorem}[Bondy-Chv\'atal] Let $G$ be a graph on $n$ vertices and $x,y$ be two non-adjacent vertices in $G$ with $\deg(x)+\deg(y)\geq n$. Then $G$ is Hamiltonian if and only if the graph formed by adding the edge $xy$ to $G$ is Hamiltonian. \end{theorem} \begin{proof}[Proof of Theorem \ref{211:thm}] The upper bound follows from Theorem \ref{mxn-ub2:thm}. In fact it is also rather easy to describe the construction directly. Let $P$ be a family of $(r-1)$-subsets of $\{2,\dots,n\}$ with $|P|=\lceil\frac{n-1}{r-1}\rceil$ whose union is $\{2,\dots,n\}$ (if $r-1$ divides $n-1$ we simply take a partition if not then we keep the overlap as small as possible). Now let $\mathcal{F}=\{\{1\}\cup X: X\in P\}$. If $C=c_1,\dots,c_n$ is a Hamilton cycle with $c_1=1$ then $c_2\in X$ for some $X\in P$ and then $\{1\}\cup X$ contains two consecutive elements of $C$. This means that $\mathcal{F}$ has no ${\bf x}$-acceptable Hamilton cycle and so $m({\bf x},n)\leq\lceil\frac{n-1}{r-1}\rceil$. For the corresponding lower bound we use the fact that $m({\bf x},n)$ is equal to the smallest number $t$ for which we can delete $t$ copies of $K_r$ from $K_n$ and be left with a non-Hamiltonian graph. Suppose that the graph we are left with after deleting these copies of $K_r$ is $G$ and let $\deg_G(i)=d_i$. We will say that a pair $x,y\in[n]$ with $xy\not\in E(G)$ is \emph{bad} if $d_x+d_y<n$. By the Bondy-Chv\'atal Theorem $G$ must remain non-Hamiltonian even after we add to it all pairs which are not bad. So we must, by Dirac's theorem, have a vertex $x$ which is incident with at least $n/2$ bad edges. Let $f=n-1-d_x$ be the number of edges in $K_n$ incident with $x$ which are deleted (that is those that are contained in some deleted $K_r$). Each $K_r$ removed from $K_n$ contributes at most $r-1$ to $f$ and so \[ t\geq\frac{f}{r-1}. \] For the pair $x,y$ to be bad we must have $d_y\leq f$ and so at least $n-1-f$ edges incident with $y$ must be deleted. In total we have $n/2$ vertices $y$ for which this holds and each edge deleted is incident with at most 2 of them. It follows that at least $\frac{n(n-1-f)}{4}$ edges must be deleted. Since each $K_r$ contributes $\binom{r}{2}$ to the total number of deleted edges, we have that \[ t\geq\frac{n(n-1-f)}{2r(r-1)}. \] The first bound is increasing with $f$, the second bound is decreasing with $f$. They are equal when $f=\frac{n(n-1)}{n+2r}$ at which point their common value is $\frac{n(n-1)}{(n+2r)(r-1)}$. It follows that \[ t\geq\frac{n(n-1)}{(n+2r)(r-1)}=(1-o(1))\frac{n}{r-1}. \] \end{proof} A further instance of the general problem which we mention briefly is the case ${\bf x}=(r-1,1)$. The upper bound we get here from Theorem \ref{mxn-ub2:thm} is \[ m((r-1,1),n)\leq\left(\frac{1}{(r-1)!}+o(1)\right)n^{r-2}. \] In particular $m((3,1),n)\leq(\frac{1}{6}+o(1))n^2$. This answers in the negative a question of Katona (Problem 5 from \cite{K}) which essentially asked ``Is it true that if $|\mathcal{F}|\leq(\frac{1}{4}+o(1))n^2$ then $K_n$ contains a Hamilton cycle which is both $(3,1)$-acceptable and $(2,2)$-acceptable for $\mathcal{F}$?'' \subsection{Other graphs} It is possible to raise similar questions in which the structure we are interested in is something other than a Hamilton cycle. Probably the most natural sort of structure is a spanning subgraph of some simple form. One simple variant is to consider Hamilton paths. \begin{q}\label{F22-path:q} What is the smallest family $\mathcal{F}\subseteq[n]^{(4)}$ with the property that $K_n$ does not contain a $(2,2)$-acceptable Hamilton path with respect to $\mathcal{F}$? \end{q} Where, naturally, a Hamilton path is $(2,2)$-acceptable if we do not have two disjoint edges of the path whose union is an element of $\mathcal{F}$. We denote the size of this smallest family $\mathcal{F}$ by $p((2,2),n)$. The main reason for raising this question is that there are indications (see below) that the extremal families for this variant may have a simpler structure. This suggests that it may be good test case for developing approaches to this type of problem. \begin{theorem}\label{22hpath:thm} \[ \frac{1}{6}(n-3)(n-2)\leq p((2,2),n)\leq\binom{n-2}{2} \] \end{theorem} \begin{proof} For the upper bound note that if $\mathcal{F}=\{X\in[n]^{(4)}:1,2\in X\}$ there is no $(2,2)$-acceptable Hamilton path with respect to $\mathcal{F}$. For the lower bound consider, as in the proof of Theorem \ref{mxn-lb:thm}, the set $S$ of all Hamilton cycles in $K_n$ and denote by $H(F)$ the set of all elements of $S$ which are not acceptable with respect to the single set $F$. If $\mathcal{F}$ is such that $K_n$ has no ${\bf x}$-acceptable Hamilton path with respect to $\mathcal{F}$ then every element of $S$ is contained in at least $2$ of the $H(F)$ with $F\in\mathcal{F}$ (if a Hamilton cycle was contained in a unique $H(F)$ then deleting one of the edges whose union is $F$ would give an acceptable Hamilton path). It follows that $\sum_{F\in\mathcal{F}}|H(F)|\geq2(n-1)!$ and since $|H(F)|=12(n-3)!$ the bound follows. \end{proof} The upper bound construction is the same as that given for the Hamilton cycle case in Theorem \ref{mxn-ub:thm}. However, in contrast to the Hamilton cycle problem this construction is minimal in that we cannnot remove any $4$-set from it without making some Hamilton path $(2,2)$-acceptable. It is possible that this upper bound is exactly sharp and if this is the case then the simple structure of the extremal family may make the problem easier that the Hamilton cycle case. \section{Further Questions} Several questions and conjectures have been mentioned in earlier sections. In this section we summarise these and collect a few other possible questions and directions for further study, concentrating mainly on the situation where our constraints are given by a set system. For matchings under constraints given by a graph the most obvious question is to bound $b(n)$ more tightly. We have no feeling for where the true order of magnitude should lie between $n^{1/2}$ and $n^{3/4}$. In addition it would be nice to know more about the general behaviour of $b$; for instance is it a non-decreasing function. The $k=3$ case of Question \ref{match-small:q} is also an appealing problem which we repeat here. \begin{q}\label{match-3:q} Is it true that for all sufficiently large $n$, if $G$ is a $3$-regular bipartite graph and $H$ is a 1-regular bipartite graph, we are guaranteed to have a perfect matching in $G$ which is acceptable for $H$? \end{q} For Hamilton cycles under constraints given by a set system the main open problem is Conjecture \ref{mxn:conj} on the asymptotic behaviour of $m({\bf x},n)$. As well as this asymptotic behaviour we could ask for exact values of $m({\bf x},n)$. This is probably a much harder problem even for our main example ${\bf x}=(2,2)$ (at least if Conjecture \ref{mxn:conj} is correct). However, as we indicated earlier, there may be variants of the problem for which this is a more approachable question. For instance: \begin{q} Is it true that if $\mathcal{F}\subseteq[n]^{(4)}$ with $|\mathcal{F}|<\binom{n-2}{2}$ then $K_n$ does not contain a $(2,2)$-acceptable Hamilton path with respect to $\mathcal{F}$? \end{q} It may be worth seeking out other variants along these lines for which the conjectured extremal family has a simple structure. Another direction also raised in \cite{K} is to consider degree versions. Given $0\leq t<r$ and $F\subseteq[n]^{(r)}$ we let $d_t(\mathcal{F})$ be the maximum $t$-degree of $\mathcal{F}$; that is the maximum over all $t$-subsets $D\in[n]^{(t)}$ of the number of elements of $\mathcal{F}$ which contain $D$. We now define $m_t({\bf x},n)$ to be the smallest $t$-degree $d_t(\mathcal{F})$ where $\mathcal{F}\subseteq[n]^{(r)}$ is a family with the property that $K_n$ does not contain an ${\bf x}$-acceptable Hamilton cycle with respect to $\mathcal{F}$. This is a generalisation of our earlier question since $d_0(\mathcal{F})=|\mathcal{F}|$ and so $m({\bf x},n)=m_0({\bf x},n)$. For the family $\mathcal{F}=\{\{1,x,x+1,y\}: 2\leq x\leq n-1, y\not=1,x,x+1\}\cup\{\{1,2,y,n\}: y\not=1,2,n\}$ there are no $(2,2)$-acceptable Hamilton cycles in $K_n$. Also no pair is contained in more that $3n-13$ of the $4$-sets in $\mathcal{F}$. It follows that $m_2((2,2),n)\leq 3n-13$. It seems plausible that $m_2((2,2),n)$ is linear in $n$ but we could not prove this. For $m_1({\bf x},n)$ it is trivial that $\frac{m_0({\bf x},n)}{n}\leq m_1({\bf x},n)\leq m_0({\bf x},n)$ so the order of magnitude of $m_1({\bf x},n)$ is between $n^{r-k-1}$ and $n^{r-k}$. In all our constructions $\mathcal{F}$ is very asymmetric. This leads us to the following question: \begin{q} What is the smallest vertex-transitive $\mathcal{F}$ with the property that $K_n$ does not contain a $(2,2)$-acceptable Hamilton cycle with respect to $\mathcal{F}$? In particular is there such a family with $|\mathcal{F}|=cn^2$ for some $c$? \end{q} We suspect that the answer to the second of these questions is no. The connection between Hamilton cycles and symmetry is tantalising (consider for instance Lov\'asz's famous question of whether all but finitely many connected vertex-transitive graphs are Hamiltonian \cite{Lov}) and questions like the one above may give some insight into it. There is also a connection between this problem and the degree variants mentioned above since a vertex-transitive $\mathcal{F}$ would have $d_1(\mathcal{F})=\frac{|\mathcal{F}|}{n}$ and so if the above question has a positive answer then $m_1((2,2),n)$ is linear in $n$. Recalling Katona's question on the minimum size of an $\mathcal{F}$ with no Hamilton cycle which is both $(2,2)$-acceptable and $(3,1)$-acceptable it would be possible to pose the more general problem involving more than one ${\bf x}$. Specifically, if ${\bf x}=(x_1,\dots,x_k)$, ${\bf y}=(y_1,\dots,y_l)$ with $x_i$ positive integers and $\sum x_i=\sum y_i=r$ then what is the smallest $\mathcal{F}$ for which $K_n$ has no Hamilton cycle which is both ${\bf x}$-acceptable and ${\bf y}$-acceptable. The size of such an $\mathcal{F}$ is necessarily smaller than both $m({\bf x},n)$ and $m({\bf y},n)$. It would be interesting to know whether behaviour analogous to principality of the graph Tur\'an density exists. That is, whether the size of the minimal $\mathcal{F}$ is essentially equal to whichever of of $m({\bf x},n)$ and $m({\bf y},n)$ is smallest or whether it can be significantly smaller than both of them. Finally, could these results on the existence of ${\bf x}$-acceptable Hamilton cycle in $K_n$ be extended to ${\bf x}$-acceptable Hamilton cycles in an arbitrary $r$-regular graph $G$? \section{Acknowledgement} Part of this work was carried out during the author's visit to Budapest in March 2011. I am grateful to Gyula Katona for several useful and interesting conversations. I also thank the London Mathematical Society and the Alfr\'ed R\'enyi Institute for providing funding for this visit.

https://arxiv.org/abs/1111.5749

Accuracy analysis of the box-counting algorithm

Accuracy of the box-counting algorithm for numerical computation of the fractal exponents is investigated. To this end several sample mathematical fractal sets are analyzed. It is shown that the standard deviation obtained for the fit of the fractal scaling in the log-log plot strongly underestimates the actual error. The real computational error was found to have power scaling with respect to the number of data points in the sample ($n_{tot}$). For fractals embedded in two-dimensional space the error is larger than for those embedded in one-dimensional space. For fractal functions the error is even larger. Obtained formula can give more realistic estimates for the computed generalized fractal exponents' accuracy.

\section{Introduction} In last decades computations of fractal dimensions (exponents) have become very popular in various areas of physics, as well as in interdisciplinary research. Fractal structures have been found in wide spectrum of problems, ranging from high energy physics \cite{Bialas1988} to cosmology \cite{Chmaj1991} and from medicine \cite{AZG-Skrzat} to econophysics \cite{AZG2002}. In spite of its popularity accuracy of obtained results is usually either not discussed or overestimated. Moreover, it has been found that in quite a few papers wrong numerical results and conclusions have been published \cite{AZG-Skrzat,AZGpreprint,AZG2001,McCauley2002}. The aim of this paper is to calculate fractal exponents for several well known mathematical fractals with the box-counting algorithm to estimate real accuracy of these computations. The dependence of accuracy with respect to the size of the available data set ($n_{tot}$) is also discussed and its simple scaling properties are found. It should be stressed that accuracy of fractal exponent computations in principle depends on many factors. For example, the accuracy can be degraded by presence of noise in the data. Also, one can get different results using different box-counting algorithms (see {\it e.g.} \cite{Theiler1990}) or using different digital representations of the investigated physical object (picture). Furthermore, one should be very careful translating Hurst exponents into fractal exponents as, in general, there is no simple reation of both \cite{Jaffard1997}. In Sec.~2 we calculate the box-counting fractal exponents for six different fractal sets for various numbers of data points ($n_{tot}$). The generalized fractal exponent is defined in the standard way \cite{AZG2001} \begin{equation} \label{dqdef} d(q) = \frac{1}{1-q} \, \lim_{N\to\infty} \frac{\sum_i \log p_i^q(N)}{\log N} \ , \end{equation} where $N$ denotes the total number of boxes and $p_i(N)$ is the measure of the subset in the $i$-th box for the given division $N$. Where the box size $\varepsilon = 1/N$. To find accuracy estimates the obtained results are compared with precise mathematical values of the exponents determined analytically. Furthermore, we calculate standard errors for the linear fits in the log-log plots used to calculate the exponents. Finally, the inverse power fits were found to give fair approximation of accuracy dependence on the size of the data set ($n_{tot}$). The final Section contains summary and conclusions. \section{Accuracy estimates} To start with we calculate fractal exponents for fractal sets embedded in one-dimensional space, namely the classical Cantor set (CS) \cite{Mandelbrot} and the (multifractal) asymmetric Cantor set (ACS) \cite{Tel1987}. The calculations have been repeated for different sizes of the sets, ranging from less than $10^2$ up to $10^5$ data points. The final results are given in Fig.~1. Crosses indicate the real accuracy of the box-countong algorithm computations, \textit{i.e.} the absolute value of the difference between the calculated exponent and exact analytical result. Circles give standard errors obtained for the linear fits in the corresponding log-log plots. In addition, the inverse power fit for accuracy as the function of the number of available data points ($n_{tot}$) is given by the dashed line. The parameter $\alpha$ denotes the exponent of the inverse power fit \begin{equation} \label{alphadef} \text{real error} \ \sim \ \frac{1}{n_{tot}^\alpha} \ . \end{equation} At first glimp it is clear that the standard error, that is often treated as the accuracy of the algorithm, considerably (up to the order of magnitude) underestimates the real error. In the case of Cantor set (A) we have $\alpha_{CS} \approx 0.50$. Similar result was obtained for the case of asymmetric Cantor set, $\alpha_{ACS} \approx 0.48$. Hence, in these cases one can expect the error of the size $\sim 1/\sqrt{n_{tot}}$. \begin{figure} \begin{center} \includegraphics[width=8.0cm,angle=0]{Fig1.eps} \caption{Real accuracy of the fractal exponent $d(0)$ (crosses) and the standard errors obtained for the linear fit in the log-log plots used to determine fractal exponents (circles). The dashed line is the inverse power fit. The upper plot (A) is for the Cantor set and the lower plot (B) is for the ACS.} \end{center} \label{fig:fig1} \end{figure} \begin{figure} \begin{center} \includegraphics[width=8.0cm,angle=0]{Fig2.eps} \caption{Real accuracy of the fractal exponent $d(0)$ (crosses) and the standard errors obtained for the linear fit in the log-log plots used to determine fractal exponents (circles). The dashed line is the inverse power fit. The upper plot (A) is for the Sierpi\'nski triangle and the lower plot (B) is for the Koch curve.} \end{center} \label{fig:fig2} \end{figure} \begin{figure} \begin{center} \includegraphics[width=8.0cm,angle=0]{Fig3.eps} \caption{Real accuracy of the fractal exponent $d(0)$ (crosses) and the standard errors obtained for the linear fit in the log-log plots used to determine fractal exponents (circles). The dashed line is the inverse power fit. The upper plot (A) is for the Weierstrass-Mandelbrot curve with parameter $D=1.5$ and the lower plot (B) is for the curve with $D=1.8$.} \end{center} \label{fig:fig3} \end{figure} As the second step we analyze fractal sets embedded in two-dimensional space: the Sierpi\'nski triangle and the Koch curve. The results are given in Fig.~2, with the same notation as for the Fig.~1. In this case we have $\alpha_{ST}\approx 0.31$ and $\alpha_{KC}\approx 0.18$. Hence, the error scales approximately as $\sim 1/n_{tot}^{1/4}$. This results is intuitively clear, as to have the same accuracy as for the one-dimensional case the squared number of data points has to be used. In these examples the actual error is also much bigger than the estimated standard error. For fractals embedded in two-dimensional space special attention should be paid to fractals that are of the function type, \textit{i.e.} there exists a reference frame in which for a given value of one coordinate there is at most one point of the fractal set. Hence, in one direction number of points in a given box is limited and this may cause slower convergence of the box-counting scheme (scaling can be observed in one direction only) resulting in bigger errors, slower convergence. Because this type of fractals has wide applicability, \textit{e.g.} in the time series analysis, we will consider this case separately. A good example of such fractal sets with precisely known fractal dimensions are the Weierstrass-Mandelbrot (WM) functions \cite{BerryLewis} \begin{equation} \label{WMdef} W(t) = \sum_{n=-\infty}^{n=+\infty} \ \frac{1}{\gamma^{(2-D)n}} \ \left[ 1 - \cos (\gamma^n t) \right] \ . \end{equation} We investigate two WM fractal functions with dimensions $d=1.5$ (A) and $d=1.8$ (B). The results are shown in Fig.~3. Again one can find a fair inverse power fit for the error with the exponents $\alpha_{WM}$ equal to $0.14$ and $0.12$ ($\sim 1/8$), respectively. Hence, to have similar accuracy as for ordinary fractals embedded in two-dimensional space one has to use squared number of data points ($n_{tot}$). This is intuitively clear, as in this case the fractal scaling can be observed in only one (instead of two) dimensions. In effect, to have reasonable accuracy for the fractal exponent, a very large number of data points has to be taken into account ($>10^5$) even though we deal with perfect mathematical fractals without any external noise. \section{Summary and conclusions} It has been shown that for the box-counting algorithm there is a fair inverse power scaling of the actual error of the computed fractal exponents of the type (\ref{alphadef}). The standard error calculated for the fit in the log-log plot strongly underestimates the actual error leading to the overestimated accuracy. Furthermore, to obtain a given level of accuracy number of data points ($n_{tot}$) used for fractals embedded in two-dimensional space should be the squared number of data points sufficient for fractals embedded in one-dimensional space. The corresponding exponents $\alpha$ are roughly equal to $1/4$ and $1/2$, respectively. Similar phenomenon occurs for fractal functions embedded in two-dimensional space, where the exponent $\alpha$ was found around $1/8$. Hence, up to an overall factor, to have the same accuracy as for the Cantor set and $10^2$ data points one has to use of order $10^8$ data points for the fractal WM function. The formula (\ref{alphadef}) and plots in Figs.~1--3 can be used to estimate accuracy of such computationion in much higher accuracy than the estimated standard errors for the linear fits. The estimated accuracy of the box-counting algorithm for various sizes of the data sets ($n_{tot}$) is also presented in Table~1. It should be stressed that for fractals with additional external noise one can expect even worse results --- the errors will be greater. \begin{table} \begin{center} \caption{\label{tab:table1} Absolute errors. } \begin{tabular}{|lccc|}% \hline % \hline % $n_{tot}$ & $1000$ & $10~000$ & $100~000$\\ \hline% \hline% 1-D fractals & $\pm$0.020 & $\pm$0.006 & $\pm$0.002 \\ 2-D fractals & $\pm$0.060 & $\pm$0.030 & $\pm$0.020 \\ 2-D W-M curve & $\pm$0.200 & $\pm$0.150 & $\pm$0.100 \\ \hline % \hline % \end{tabular} \end{center} \end{table}

https://arxiv.org/abs/1912.09763

Optimizing Sparsity over Lattices and Semigroups

Motivated by problems in optimization we study the sparsity of the solutions to systems of linear Diophantine equations and linear integer programs, i.e., the number of non-zero entries of a solution, which is often referred to as the $\ell_0$-norm. Our main results are improved bounds on the $\ell_0$-norm of sparse solutions to systems $A x = b$, where $A \in \mathbb{Z}^{m \times n}$, $b \in \mathbb{Z}^m$ and $x$ is either a general integer vector (lattice case) or a non-negative integer vector (semigroup case). In the lattice case and certain scenarios of the semigroup case, we give polynomial time algorithms for computing solutions with $\ell_0$-norm satisfying the obtained bounds.

\section{Introduction} This paper discusses the problem of finding sparse solutions to systems of linear Diophantine equations and integer linear programs. We investigate the $\ell_0$-norm $\|\boldsymbol x\|_0 := |\setcond{ i }{x_i \ne 0}|$, a function widely used in the theory of {\em compressed sensing} \cite{CS_survey,candestao}, which measures the sparsity of a given vector $\boldsymbol x = (x_1,\ldots,x_n)^\top \in \mathbb R^n$ (it is clear that the $\ell_0$-norm is actually not a norm). Sparsity is a topic of interest in several areas of optimization. The $\ell_0$-norm minimization problem over reals is central in the theory of the classical compressed sensing, where a linear programming relaxation provides a guaranteed approximation \cite{candesetal2006stable,candestao}. Support minimization for solutions to Diophantine equations is relevant for the theory of compressed sensing for discrete-valued signals \cite{PROMP,Lenya,Konyagin}. There is still little understanding of discrete signals in the compressed sensing paradigm, despite the fact that there are many applications in which the signal is known to have discrete-valued entries, for instance, in wireless communication \cite{MIMO} and the theory of error-correcting codes \cite{ECC}. Sparsity was also investigated in integer optimization \cite{Support,EisenbrandShmonin2006,MR3950898}, where many combinatorial optimization problems have useful interpretations as sparse semigroup problems. For example, the edge-coloring problem can be seen as a problem in the semigroup generated by matchings of the graph \cite{LOVASZ1987187}. Our results provide natural out-of-the-box sparsity bounds for problems with linear constraints and integer variables in a general form. \subsection{Lattices: sparse solutions of linear Diophantine systems} Each integer matrix $A \in \mathbb Z^{m \times n}$ determines the lattice $\mathcal{L}(A):= \setcond{A \boldsymbol x }{\boldsymbol x \in \mathbb Z^n}$ generated by the columns of $A$. By an easy reduction via row transformations, we may assume without loss of generality that the rank of $A$ is $m$. Let $[n]:= \{1,\ldots,n\}$ and let $\binom{[n]}{m}$ be the set of all $m$-element subsets of $[n]$. For $\gamma \subseteq [n]$, consider the $m \times |\gamma|$ submatrix $A_\gamma$ of $A$ with columns indexed by $\gamma$. One can easily prove that the determinant of $\mathcal{L}(A)$ is equal to \[ \gcd(A):= \gcd \setcond{\det(A_\gamma)}{\gamma \in \binom{[n]}{m}}. \] Since $\mathcal{L}(A_\gamma)$ is the lattice spanned by the columns of $A$ indexed by $\gamma$, it is a sublattice of $\mathcal{L}(A)$. We first deal with a natural question: {\em Can the description of a given lattice $\mathcal{L}(A)$ in terms of $A$ be made sparser by passing from $A$ to $A_\gamma$ with $\gamma$ having a smaller cardinality than $n$ and satisfying $\mathcal{L}(A) = \mathcal{L}(A_\gamma)$? } That is, we want to discard some of the columns of $A$ and generate $\mathcal{L}(A)$ by $|\gamma|$ columns with $|\gamma|$ being possibly small. For stating our results, we need several number-theoretic functions. Given $z \in \mathbb Z_{>0}$, consider the prime factorization $z = p_1^{s_1} \cdots p_k^{s_k}$ with pairwise distinct prime factors $p_1,\ldots,p_k$ and their multiplicities $s_1,\ldots,s_k \in \mathbb Z_{>0}$. Then the number of prime factors $\sum_{i=1}^k s_i$ counting the multiplicities is denoted by $\Omega(z)$. Furthermore, we introduce $\Omega_m(z) := \sum_{i=1}^k \min \{s_i,m\}$. That is, by introducing $m$ we set a threshold to account for multiplicities. In the case $m=1$ we thus have $\omega(z):=\Omega_1(z) = k$, which is the number of prime factors in $z$, not taking the multiplicities into account. The functions $\Omega$ and $\omega$ are called \emph{prime $\Omega$-function} and \emph{prime $\omega$-function}, respectively, in number theory \cite{hardy2008introduction}. We call $\Omega_m$ the \emph{truncated prime $\Omega$-function}. \begin{thm} \label{thm:sparsifying:lattice:A} Let $A \in \mathbb Z^{m \times n}$, with $m \le n$, and let $\tau \in \binom{[n]}{m}$ be such that the matrix $A_\tau$ is non-singular. Then the equality $\mathcal{L}(A) = \mathcal{L}(A_\gamma)$ holds for some $\gamma$ satisfying $\tau \subseteq \gamma \subseteq [n]$ and \begin{equation} \label{card:gamma:bound} |\gamma| \le m + \Omega_m \left( \frac{|\det(A_\tau)|}{\gcd(A)} \right). \end{equation} Given $A$ and $\tau$, the set $\gamma$ can be computed in polynomial time. \end{thm} One can easily see that $\omega(z) \le \Omega_m(z) \le \Omega(z) \le \log_2 (z)$ for every $z \in \mathbb Z_{>0}$. The estimate using $\log_2(z)$ gives a first impression on the quality of the bound \eqref{card:gamma:bound}. It turns out, however, that $\Omega_m(z)$ is much smaller on the average. Results in number theory \cite[\S22.10]{hardy2008introduction} show that the average values $\frac{1}{z}(\omega(1) + \cdots + \omega(z))$ and $\frac{1}{z}(\Omega(1) + \cdots + \Omega(z))$ are of order $\log \log z$, as $z \to \infty$. As an immediate consequence of Theorem~\ref{thm:sparsifying:lattice:A} we obtain \begin{cor}\label{cor:sparsity:dioph:sys} Consider the linear Diophantine system \begin{equation}\label{lin:dioph:sys} A \boldsymbol x = \boldsymbol b, \ \boldsymbol x \in \mathbb Z^n \end{equation} with $A \in \mathbb Z^{m \times n}$, $\boldsymbol b \in \mathbb Z^{m }$ and $m \le n$. Let $\tau \in \binom{[n]}{m}$ be such that the $m \times m$ matrix $A_\tau$ is non-singular. If \eqref{lin:dioph:sys} is feasible, then \eqref{lin:dioph:sys} has a solution $\boldsymbol x$ satisfying the sparsity bound \begin{equation* \|\boldsymbol x \|_0 \le m + \Omega_m \left( \frac{| \det(A_\tau)| }{\gcd(A)} \right). \end{equation*} Under the above assumptions, for given $A,\boldsymbol b$ and $\tau$, such a sparse solution can be computed in polynomial time. \end{cor} From the optimization perspective, Corollary~\ref{cor:sparsity:dioph:sys} deals with the problem \begin{equation* \min \setcond{ \|\boldsymbol x\|_0 }{ A \boldsymbol x = \boldsymbol b, \ \boldsymbol x \in \mathbb Z^n} \end{equation*} of minimization of the $\ell_0$-norm over the affine lattice $\setcond{\boldsymbol x \in \mathbb Z^n}{A \boldsymbol x = \boldsymbol b}$. \subsection{Semigroups: sparse solutions in integer programming} Consider next the standard form of the feasibility constraints of integer linear programming \begin{equation}\label{ILP:feasibility} A \boldsymbol x =\boldsymbol b, \ \boldsymbol x \in \mathbb Z_{\ge 0}^n. \end{equation} For a given matrix $A$, the set of all $\boldsymbol b$ such that \eqref{ILP:feasibility} is feasible, is the \emph{ semigroup} $ \icone(A) = \{ A \boldsymbol x : \boldsymbol x \in \mathbb{Z}_{\ge0}^n \}$ generated by the columns of $A$. If \eqref{ILP:feasibility} has a solution, i.e., $\boldsymbol b\in\icone(A)$, \emph{ how sparse can such a solution be?} In other words, we are interested in the $\ell_0$-norm minimization problem \begin{equation}\label{l0:IP} \min \setcond{ \|\boldsymbol x\|_0 }{A \boldsymbol x = \boldsymbol b, \ \boldsymbol x \in \mathbb Z_{\ge 0}^n}. \end{equation} It is clear that Problem \eqref{l0:IP} is NP-hard, because deciding the feasibility of \eqref{ILP:feasibility} \cite[\S~18.2]{Schrijver} or even solving the relaxation of \eqref{l0:IP} with the condition ${\boldsymbol x} \in \mathbb Z_{\ge 0}^n$ replaced by ${\boldsymbol x} \in \mathbb R^n$ \cite{natarajan} is NP-hard. Taking the NP-hardness of Problem (\ref{l0:IP}) into account, our aim is to \emph{estimate} the optimal value of \eqref{l0:IP} under the assumption that this problem is feasible. In \cite[Theorem 1.1 (i)]{ADOO} (see also \cite[Theorem 1]{Support}), it was shown that for any $\boldsymbol b \in \icone(A)$, there exists a $\boldsymbol x \in \mathbb Z^n$, such that $A\boldsymbol x = \boldsymbol b$ and \begin{equation}\label{ADNO:bound} \|\boldsymbol x\|_0 \le m + \floor{ \log_2 \left( \frac{\sqrt{\det(A A^\top)}}{\gcd(A)} \right) }. \end{equation} We show that in some special cases we can significantly improve this bound. As a consequence of Theorem~\ref{thm:sparsifying:lattice:A} we obtain the following. \begin{cor} \label{cor:pos:spanning:columns} Let $A \in \mathbb Z^{m \times n}$ be a matrix whose columns positively span $\mathbb R^m$ and let $\boldsymbol b \in \mathbb Z^m$. Then $ \mathcal{L}(A)=\icone(A)$. Furthermore, if $\boldsymbol b \in \mathcal{L}(A)$, and $\tau \in \binom{[n]}{m}$ is a set, for which the matrix $A_\tau$ is non-singular, then there is a solution $\boldsymbol x$ of the integer-programming feasibility problem $A \boldsymbol x = \boldsymbol b, \boldsymbol x \in \mathbb Z_{\ge 0}^m$ that satisfies the sparsity bound \begin{equation} \label{pos:spanning:bound} \|\boldsymbol x\|_0 \le 2m + \Omega_m \left( \frac{|\det(A_\tau)|}{\gcd(A)} \right). \end{equation} Under the above assumptions, for given $A, \boldsymbol b$ and $\tau$, such a sparse solution $\boldsymbol x$ can be computed in polynomial time. \end{cor} Note that for a fixed $m$, \eqref{pos:spanning:bound} is usually much tighter than \eqref{ADNO:bound}, because the function $\Omega_m(z)$ is bounded from above by the logarithmic function $\log_2(z)$ and is much smaller than $\log_2(z)$ on the average. Furthermore, $|\det(A_\tau)| \le \sqrt{\det(A A^\top)}$ in view of the Cauchy-Binet formula. We take a closer look at the case $m=1$ of a single equation and tighten the given bounds in this case. That is, we consider the \emph{knapsack feasibility problem} \begin{equation}\label{knapsack_problem} \boldsymbol a^\top \boldsymbol x = b, \ \boldsymbol x \in \mathbb Z_{\ge 0}^n, \end{equation} where $\boldsymbol a \in \mathbb Z^n$ and $b \in \mathbb Z$. Without loss of generality we can assume that all components of the vector $\boldsymbol a$ are not equal to zero. It follows from (\ref{ADNO:bound}) that a feasible problem (\ref{knapsack_problem}) has a solution ${\boldsymbol x}$ with \begin{eqnarray}\label{previous_knapsack_bound} \|\boldsymbol x\|_0 \le 1 + \left\lfloor\log\left (\frac{\|{\boldsymbol a}\|_2}{\gcd({ \boldsymbol a})}\right )\right\rfloor\,. \end{eqnarray} If all components of $\vec a$ have the same sign, without loss of generality we can assume $\vec a \in \mathbb Z_{>0}^n$. In this setting, Theorem 1.2 in \cite{ADOO} strengthens the bound (\ref{previous_knapsack_bound}) by replacing the $\ell_2$-norm of the vector $\boldsymbol a$ with the $\ell_{\infty}$-norm. It was conjectured in \cite[page~247]{ADOO} that a bound $\|\boldsymbol x\|_0 \le c + \left\lfloor\log_2\left (\|{ \boldsymbol a}\|_{\infty}/\gcd({ \boldsymbol a})\right )\right\rfloor\,$ with an absolute constant $c$ holds for an \emph{arbitrary} $\boldsymbol a \in \mathbb Z^n$. We obtain the following result, which covers the case that has not been settled so far and yields a confirmation of this conjecture. \begin{cor} \label{cor:mixed:sign:knapsack} Let $\boldsymbol a = (a_1,\ldots,a_n)^\top \in (\mathbb Z \setminus \{0\})^n$ be a vector that contains both positive and negative components. If the knapsack feasibility problem $\boldsymbol a^\top \boldsymbol x = b, \ \boldsymbol x \in \mathbb Z_{\ge 0}^n$ has a solution, then there is a solution $\boldsymbol x$ satisfying the sparsity bound \[ \|\boldsymbol x\|_0 \le 2 + \min \setcond{ \omega \left( \frac{|a_i|}{\gcd(\boldsymbol a)}\right) }{i \in [n]}. \] Under the above assumptions, for given $\boldsymbol a$ and $b$, such a sparse solution $\boldsymbol x$ can be computed in polynomial time. \end{cor} Our next contribution is that, given additional structure on $A$, we can improve on \cite[Theorem 1.1 (i)]{ADOO}, which in turn also gives an improvement on \cite[Theorem 1.2]{ADOO}. For $\boldsymbol a_1,\ldots,\boldsymbol a_n \in \mathbb R^m$, we denote by $\operatorname{cone}(\boldsymbol a_1,\ldots,\boldsymbol a_n)$ the convex conic hull of the set $\{\boldsymbol a_1,\ldots,\boldsymbol a_n\}$. Now assume the matrix $A = (\boldsymbol a_1,\ldots,\boldsymbol a_n) \in \mathbb Z^{m \times n}$ with columns ${\boldsymbol a}_i$ satisfies the following conditions: \begin{align} & \boldsymbol a_1,\ldots,\boldsymbol a_n \in \mathbb Z^m \setminus \{\boldsymbol 0\}, \label{non_zero_col} \\ & \text{$\operatorname{cone}(\boldsymbol a_1,\ldots,\boldsymbol a_n)$ is an $m$-dimensional pointed cone}, \label{m_dim_pointed_cone} \\ & \text{$\operatorname{cone}(\boldsymbol a_1)$ is an extreme ray of $\operatorname{cone}(\boldsymbol a_1,\ldots,\boldsymbol a_n)$}. \label{extreme_ray} \end{align} Note that the previously best sparsity bound for the general case of the integer-programming feasibility problem is \eqref{ADNO:bound}. Using the Cauchy-Binet formula, \eqref{ADNO:bound} can be written as \[ \|\boldsymbol x\|_0 \le m + \log_2 \frac{\sqrt{\sum_{I \in \binom{[n]}{m}}\det(A_I)^2}}{\gcd(A)}. \] The following theorem improves this bound in the \emph{``pointed cone case''} by removing a fraction of $m/n$ of terms in the sum under the square root. \begin{thm}\label{thm:pointed:cone} Let $A = (\boldsymbol a_1,\ldots,\boldsymbol a_n) \in \mathbb Z^{m \times n}$ satisfy \eqref{non_zero_col}--\eqref{extreme_ray} and, for $\boldsymbol b \in \mathbb Z^m$, consider the integer-programming feasibility problem \begin{equation} \label{IP_feas} A \boldsymbol x =\boldsymbol b, \ \boldsymbol x \in \mathbb Z_{\ge 0}^n. \end{equation} If \eqref{IP_feas} is feasible, then there is a feasible solution $\boldsymbol x$ satisfying the sparsity bound \[ \|\boldsymbol x\|_0 \le m + \floor{\log_2 \frac{q(A)}{\gcd(A)}}, \] where \begin{equation*} q(A):= \sqrt{\sum_{I \in \binom{[n]}{m} \,:\, 1 \in I} \det(A_I)^2}. \end{equation*} \end{thm} We omit the proof of this result due to the page limit for the IPCO proceedings. Instead we focus on the particularly interesting case $m=1$. In this case, assumption~\eqref{m_dim_pointed_cone} is equivalent to $\boldsymbol a \in \mathbb Z_{>0}^n \cup \mathbb Z_{<0}^n$. Without loss of generality, one can assume $\boldsymbol a \in \mathbb Z_{>0}^n$. \begin{thm} \label{thm:positive:knapsack} Let $\boldsymbol a = (a_1,\ldots,a_n)^\top \in \mathbb Z_{>0}^n$ and $b \in \mathbb Z_{\ge 0}$. If the knapsack feasibility problem $\boldsymbol a^\top \boldsymbol x = b, \ \boldsymbol x \in \mathbb Z_{\ge 0}^n$ has a solution, there is a solution $\boldsymbol x$ satisfying the sparsity bound \begin{equation*} \|\boldsymbol x\|_0 \le 1 + \floor{\log_2 \left( \frac{\min \{a_1,\ldots,a_n\}}{\gcd(\boldsymbol a)} \right)}. \end{equation*} \end{thm} \noindent When dealing with bounds for sparsity it would be interesting to understand \emph{the worst case scenario among all members of the semigroup}, which is described by the function \begin{equation}\label{def:ICR} \mathsf{ICR}(A) = \max_{ \boldsymbol b \in \icone(A)} \min\{ \| {\boldsymbol x} \|_0 \,:\, A \boldsymbol x = \boldsymbol b, \ {\boldsymbol x} \in \ZZ_{\ge 0}^n\}. \end{equation} We call $\mathsf{ICR}(A)$ the \emph{integer Carath\'eodory rank} in resemblance to the classical problem of finding the integer Carath\'eodory number for Hilbert bases \cite{SeboCaratheodory}. Above results for the problem $A \boldsymbol x =b, \ \boldsymbol x \in \mathbb Z_{\ge 0}^n$ can be phrased as upper bounds on $\mathsf{ICR}(A)$. We are interested in the complexity of computing $\mathsf{ICR}(A)$. The first question is: \emph{can the integer Carath\'eodory rank of a matrix $A$ be computed at all?} After all, remember that the semigroup has infinitely many elements and, despite the fact that $\mathsf{ICR}(A)$ is a finite number, a direct usage of \eqref{def:ICR} would result into the determination of the sparsest representation $A \boldsymbol x = \boldsymbol b$ for all of the infinitely many elements $\boldsymbol b$ of $\icone(A)$. It turns out that $\mathsf{ICR}(A)$ is computable, as the inequality $\mathsf{ICR}(A) \le k$ can be expressed as the formula $\forall \boldsymbol x \in \mathbb Z_{\ge 0}^n \, \exists \boldsymbol y \in \mathbb Z_{\ge 0}^n \,: \, (A \boldsymbol x = A \boldsymbol y) \wedge (\|\boldsymbol y\|_0 \le k )$ in \emph{Presburger arithmetic} \cite{haase2018survival}. Beyond this fact, the complexity status of computing $\mathsf{ICR}(A)$ is largely open, even when $A$ is just one row: \begin{prob} \label{prob:ICR} Given the input $\boldsymbol a = (a_1,\ldots,a_n)^\top \in \mathbb Z^n$, is it NP-hard to compute $\mathsf{ICR}(\boldsymbol a^\top)$? \end{prob} The \emph{Frobenius number} $\max\, \mathbb Z_{\ge 0} \setminus \icone(\boldsymbol a^\top)$, defined under the assumptions $\boldsymbol a \in \mathbb Z_{>0}^n$ and $\gcd({\boldsymbol a})=1$, is yet another value associated to $\icone(\boldsymbol a^\top)$. The Frobenius number can be computed in polynomial time when $n$ is fixed \cite{barvinok-woods-2003,kannan1992lattice} but is NP-hard to compute when $n$ is not fixed \cite{ramirez1996complexity}. It seems that there might be a connection between computing the Frobenius number and $\mathsf{ICR}(\boldsymbol a^\top)$. \section{Proofs of Theorem~\ref{thm:sparsifying:lattice:A} and its consequences} The proof Theorem~\ref{thm:sparsifying:lattice:A} relies on the theory of finite Abelian groups. We write Abelian groups additively. An Abelian group $G$ is said to be a \emph{direct sum} of its finitely many subgroups $G_1,\ldots,G_m$, which is written as $G = \bigoplus_{i=1}^m G_i$, if every element $x \in G$ has a unique representation as $x = x_1 + \cdots + x_m$ with $x_i \in G_i$ for each $i \in [m]$. A \emph{primary cyclic group} is a non-zero finite cyclic group whose order is a power of a prime number. We use $G / H$ to denote the quotient of $G$ modulo its subgroup $H$. The fundamental theorem of finite Abelian groups states that every finite Abelian group $G$ has a \emph{primary decomposition}, which is essentially unique. This means, $G$ is decomposable into a direct sum of its primary cyclic groups and that this decomposition is unique up to automorphisms of $G$. We denote by $\kappa(G)$ the number of direct summands in the primary decomposition of $G$. For a subset $S$ of a finite Abelian group $G$, we denote by $\gen{S}$ the subgroup of $G$ generated by $S$. We call a subset $S$ of $G$ \emph{non-redundant} if the subgroups $\gen{T}$ generated by proper subsets $T$ of $S$ are properly contained in $\gen{S}$. In other words, $S$ is non-redundant if $\gen{S \setminus \{x\}}$ is a proper subgroup of $\gen{S}$ for every $x \in S$. The following result can be found in \cite[Lemma~A.6]{Geroldinger+HalterKoch}. \begin{thm} \label{thm:max:card:nonredundant} Let $G$ be a finite Abelian group. Then the maximum cardinality of a non-redundant subset $S$ of $G$ is equal to $\kappa(G)$. \end{thm} We will also need the following lemmas. \begin{lemma} \label{lem:kappa:for:sum:cyclic} Let $G$ be a finite Abelian group representable as a direct sum $G = \bigoplus_{j=1}^m G_j$ of $m \in \mathbb Z_{>0}$ cyclic groups. Then $\kappa(G) \le \Omega_m(|G|)$. \end{lemma} \begin{proof} Consider the prime factorization $|G| = p_1^{n_1} \cdots p_s^{n_s}$. Then $|G_j| = p_1^{n_{i,j}} \cdots p_s^{n_{i,j}}$ with $0 \le n_{i,j} \le n_i$ and, by the Chinese Remainder Theorem, the cyclic group $G_j$ can be represented as $G_j = \bigoplus_{i=1}^s G_{i,j}$, where $G_{i,j}$ is a cyclic group of order $p_i^{n_{i,j}}$. Consequently, $G = \bigoplus_{i=1}^s \bigoplus_{j=1}^m G_{i,j}$. This is a decomposition of $G$ into a direct sum of primary cyclic groups and, possibly, some trivial summands $G_{i,j}$ equal to $\{0\}$. We can count the non-trivial direct summands whose order is a power of $p_i$, for a given $i \in [s]$. There is at most one summand like this for each of the groups $G_j$. So, there are at most $m$ non-trivial summands in the decomposition whose order is a power of $p_i$. On the other hand, the direct sum of all non-trivial summands whose order is a power of $p_i$ is a group of order $p_i^{n_{i,1} + \cdots + n_{i,s}} = p_i^{n_i}$ so that the total number of such summands is not larger than $n_i$, as every summand contributes the factor at least $p_i$ to the power $p_i^{n_i}$. This shows that the total number of non-zero summands in the decomposition of $G$ is at most $\sum_{i=1}^s \min \{m, n_i\} = \Omega_m(|G|)$. \end{proof} \begin{lemma} \label{lem:group:from:lattice} Let $\Lambda$ be a sublattice of $\mathbb Z^m$ of rank $m \in \mathbb Z_{>0}^m$. Then $G = \mathbb Z^m / \Lambda$ is a finite Abelian group of order $\det(\Lambda)$ that can be represented as a direct sum of at most $m$ cyclic groups. \end{lemma} \begin{proof} The proof relies on the relationship of finite Abelian groups and lattices, see \cite[\S4.4]{Schrijver}. Fix a matrix $M \in \mathbb Z^{m \times m}$ whose columns form a basis of $\Lambda$. Then $|\det(M)| = \det(\Lambda)$. There exist unimodular matrices $U \in \mathbb Z^{m \times m}$ and $V \in \mathbb Z^{m \times m}$ such that $D:= U M V$ is diagonal matrix with positive integer diagonal entries. For example, one can choose $D$ to be the Smith Normal Form of $M$ \cite[\S4.4]{Schrijver}. Let $d_1,\ldots,d_m \in \mathbb Z_{>0}$ be the diagonal entries of $D$. Since $U$ and $V$ are unimodular, $d_1 \cdots d_m = \det(D) = \det(\Lambda)$. We introduce the quotient group $G' := \mathbb Z^m / \Lambda' = ( \mathbb Z / d_1 \mathbb Z) \times \cdots \times (\mathbb Z / d_m \mathbb Z)$ with respect to the lattice $\Lambda' := \mathcal{L}(D) = (d_1 \mathbb Z) \times \cdots \times (d_m \mathbb Z)$. The order of $G'$ is $d_1 \cdots d_m = \det(D) = \det(\Lambda)$ and $G'$ is a direct sum of at most $m$ cyclic groups, as every $d_i > 1$ determines a non-trivial direct summand. To conclude the proof, it suffices to show that $G'$ is isomorphic to $G$. To see this, note that $\Lambda' = \mathcal{L}(D) = \mathcal{L}(UMV) = \mathcal{L}(UM) = \setcond{U z}{z \in \Lambda}$. Thus, the map $z \mapsto U z$ is an automorphism of $\mathbb Z^m$ and an isomorphism from $\Lambda$ to $\Lambda'$. Thus, $z \mapsto U z$ induces an isomorphism from the group $G = \mathbb Z^m / \Lambda$ to the group $G'=\mathbb Z^m / \Lambda'$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:sparsifying:lattice:A}] Let $\vec a_1,\ldots,\vec a_n$ be the columns of $A$. Without loss of generality, let $\tau = [m]$. We use the notation $B:=A_\tau$. \emph{Reduction to the case $\gcd(A)=1$.} For a non-singular square matrix $M$, the columns of $M^{-1} A$ are representations of the columns of $A$ in the basis of columns of $M$. In particular, for a matrix $M$ whose columns form a basis of $\mathcal{L}(A)$, the matrix $M^{-1} A$ is integral and the $m \times m$ minors of $M^{-1} A$ are the respective $m \times m$ minors of $A$ divided by $\det(M) = \gcd(A)$. Thus, replacing $A$ by $M^{-1} A$, we pass from $\mathcal{L}(A)$ to $\mathcal{L}(M^{-1} A) = \setcond{M^{-1} z}{z \in \mathcal{L}(A)}$, which corresponds to a change of a coordinate system in $\mathbb R^m$ and ensures that $\gcd(A)=1$. \emph{Sparsity bound \eqref{card:gamma:bound}.} The matrix $B$ gives rise to the lattice $\Lambda := \mathcal{L}(B)$ of rank $m$, while $\Lambda$ determines the finite Abelian group $\mathbb Z^m / \Lambda$. Consider the canonical homomorphism $\phi : \mathbb Z^m \to \mathbb Z^m / \Lambda$, sending an element of $\mathbb Z^m$ to its coset modulo $\Lambda$. Since $\gcd(A)=1$, we have $\mathcal{L}(A)= \mathbb Z^m$, which implies \( \gen{ T } = \mathbb Z^m / \Lambda \) for $T:= \{\phi(\vec a_{m+1}),\ldots, \phi(\vec a_n) \}$. For every non-redundant subset $S$ of $T$, we have \begin{align*} |S| & \le \kappa( \mathbb Z^m / \Lambda) & & \text{(by Theorem~\ref{thm:max:card:nonredundant})} \\ & \le \Omega_m(|\det(A_\tau)|) & & \text{(by Lemmas~\ref{lem:kappa:for:sum:cyclic} and \ref{lem:group:from:lattice}).} \end{align*} Fixing a set $I \subseteq \{m+1,\ldots,n\}$ that satisfies $|I|=|S|$ and $S = \setcond{\phi(\vec a_i)}{i \in I}$, we reformulate $\gen{S} = \mathbb Z^m / \Lambda$ as \( \mathbb Z^m = \mathcal{L}(A_I) + \Lambda = \mathcal{L}(A_I) + \mathcal{L}(A_\tau) = \mathcal{L}(A_{I \cup \tau}). \) Thus, \eqref{card:gamma:bound} holds for $\gamma = I \cup \tau$. \emph{Construction of $\gamma$ in polynomial time.} The matrix $M$ used in the reduction to the case $\gcd(A) = 1$ can be constructed in polynomial time: one can obtain $M$ from the Hermite Normal Form of $A$ (with respect to the column transformations) by discarding zero columns. For the determination of $\gamma$, the set $I$ that defines the non-redundant subset $S = \setcond{\phi(\vec a_i)}{i \in I}$ of $\mathbb Z^m / \Lambda$ needs to be determined. Start with $I = \{m+1,\ldots,n\}$ and iteratively check if some of the elements $\phi(\vec a_i) \in \mathbb Z^m / \Lambda$, where $i \in I$, is in the group generated by the remaining elements. Suppose $j \in I$ and we want to check if $\phi(\vec a_j)$ is in the group generated by all $\phi(\vec a_i)$ with $i \in I\setminus \{j\}$. Since $\Lambda = \mathcal{L}(A_\tau)$, this is equivalent to checking $\vec a_j \in \mathcal{L}(A_{I \setminus \{j\} \cup \tau})$ and is thus reduced to solving a system of linear Diophantine equations with the left-hand side matrix $A_{I \setminus \{j\} \cup \tau}$ and the right-hand side vector $\vec a_j$. Thus, carrying the above procedure for every $j \in I$ and removing $j$ from $I$ whenever $\vec a_j \in \mathcal{L}(A_{I \setminus \{j\} \cup \tau})$, we eventually arrive at a set $I$ that determines a non-redundant subset $S$ of $\mathbb Z^m / \Lambda$. This is done by solving at most $n-m$ linear Diophantine systems in total, where the matrix of each system is a sub-matrix of $A$ and the right-hand vector of the system is a column of $A$. \end{proof} \begin{remark}[Optimality of the bounds] For a given $\Delta \in \mathbb Z_{\ge 2}$ let us consider matrices $A \in \mathbb Z^{m \times n}$ with $\Delta = |\det(A_\tau)| / \gcd(A)$. We construct a matrix $A$ that shows the optimality of the bound \eqref{card:gamma:bound}. As in the proof of Theorem~\ref{thm:sparsifying:lattice:A}, we assume $\tau =[m]$ and use the notation $B = A_\tau$. Consider the prime factorization $\Delta = p_1^{n_1} \cdots p_s^{n_s}$. We will fix the matrix $B$ to be a diagonal matrix with diagonal entries $d_1,\ldots,d_m \in \mathbb Z_{>0}$ so that $\det(B) = d_1 \cdots d_m = \Delta$. The diagonal entries are defined by distributing the prime factors of $\Delta$ among the diagonal entries of $B$. If the multiplicity $n_i$ of the prime $p_i$ is less than $m$, we introduce $p_i$ as a factor of multiplicity $1$ in $n_i$ of the $m$ diagonal entries of $B$. If the multiplicity $n_i$ is at least $m$, we are able distribute the factors $p_i$ among \emph{all} of the diagonal entries of $B$ so that each diagonal entry contains the factor $p_i$ with multiplicity at least $1$. The group $\mathbb Z^m / \Lambda = \mathbb Z^m / \mathcal{L}(B)$ is a direct sum of $m$ cyclic groups $G_1,\ldots, G_m$ of orders $d_1,\ldots,d_m$, respectively. By the Chinese Remainder Theorem, these cyclic groups can be further decomposed into the direct sum of primary cyclic groups. By our construction, the prime factor $p_i$ of the multiplicity $n_i < m$ generates a cyclic direct summand of order $p_i$ in $n_i$ of the subgroups $G_1,\ldots,G_m$. If $n_i \ge m$, then each of the groups $G_1,\ldots,G_m$ has a direct summand, which is a non-trivial cyclic group whose order is a power of $p_i$. Summarizing, we see that the decomposition of $\mathbb Z^m / \Lambda$ into primary cyclic groups contains $n_i$ summands of order $p_i$, when $n_i < m$, and $m$ summands, whose order is a power of $p_i$, when $n_i \ge m$. The total number of summands is thus $\sum_{i=1}^s \min \{m, n_i\} = \Omega_m(\Delta)$. Now, fix $n= m + \Omega_m(\Delta)$ and choose columns $\vec a_{m+1},\ldots,\vec a_n$ so that $\phi(\vec a_{m+1}),$ $\ldots, \phi(\vec a_n)$ generate all direct summands in the decomposition of $\mathbb Z^m / \Lambda$ into primary cyclic groups. With this choice, $\phi(\vec a_{m+1}),\ldots, \phi(\vec a_n)$ generate $\mathbb Z^m / \Lambda$, which means that $\mathcal{L}(A) = \mathbb Z^m$ and implies $\gcd(A)=1$. On the other hand, any proper subset $\{\phi(\vec a_{m+1}),\ldots, \phi(\vec a_n)\}$ generates a proper subgroup of $\mathbb Z^m / \Lambda$, as some of the direct summands in the decomposition of $\mathbb Z^m / \Lambda$ into primary cyclic groups will be missing. This means $\mathcal{L}(A_{[m] \cup I}) \varsubsetneq \mathbb Z^m$ for every $I \varsubsetneq \{m+1,\ldots,n\}$. \end{remark} \begin{proof}[Proof of Corollary~\ref{cor:sparsity:dioph:sys}] Feasiblity of \eqref{lin:dioph:sys} can be expressed as $\vec b \in \mathcal{L}(A)$. Choose $\gamma$ from the assertion of Theorem~\ref{thm:sparsifying:lattice:A}. One has $\vec b \in \mathcal{L}(A) = \mathcal{L}(A_\gamma)$ and so there exists a solution $\vec x$ of \eqref{lin:dioph:sys} whose support is a subset of $\gamma$. This sparse solution $\boldsymbol x$ can be computed by solving the Diophantine system with the left-hand side matrix $A_\gamma$ and the right-hand side vector $\boldsymbol b$. \end{proof} \begin{proof}[Proof of Corollary~\ref{cor:pos:spanning:columns}] Assume that the Diophantine system $A \vec x = \vec b, \ \vec x \in \mathbb Z^n$ has a solution. It suffices to show that, in this case, the integer-programming feasibility problem $A \vec x = \vec b, \ \vec x \in \mathbb Z_{\ge 0}^n$ has a solution, too, and that one can find a solution of the desired sparsity to the integer-programming feasibility problem in polynomial time. One can determine $\gamma$ as in Theorem~\ref{card:gamma:bound} in polynomial time. Using $\gamma$, we can determine a solution $\vec x^\ast = (x^\ast_1,\ldots,x^\ast_n)^\top \in \mathbb Z^n$ of the Diophantine system $A \vec x = b, \ \vec x \in \mathbb Z^n$ satisfying $x^\ast_i =0$ for $i \in [n] \setminus \gamma$ in polynomial time, as described in the proof of Corollary~\ref{cor:sparsity:dioph:sys}. Let $\vec a_1,\ldots,\vec a_n$ be the columns of $A$. Since the matrix $A_\tau$ is non-singular, the $m$ vectors vectors $\vec a_i$, where $i \in \tau$, together with the vector $\vec v= -\sum_{i \in \tau} \vec a_i$ positively span $\mathbb R^n$. Since all columns of $A$ positive span $\mathbb R^n$, the conic version of the Carath\'eodory theorem implies the existence of a set $\beta \subseteq [m]$ with $|\beta| \le m$, such that $\vec v$ is in the conic hull of $\setcond{\vec a_i}{i \in \beta}$. Consequently, the set $\setcond{\vec a_i}{i \in \beta \cup \tau}$ and by this also the larger set $\setcond{\vec a_i}{i \in \beta \cup \gamma}$ positively span $\mathbb R^m$. Let $I = \beta \cup \gamma$. By construction, $|I| \le |\beta| + |\gamma| \le m + |\gamma|$. Since the vectors $\vec a_i$ with $i \in I$ positively span $\mathbb R^m$, there exist a choice of rational coefficients $\lambda_i>0$ ($i \in I$) with $\sum_{i \in I} \lambda_i \vec a_i = 0$. After rescaling we can assume $\lambda_i \in \mathbb Z_{>0}$. Define $\vec x' = (x_1',\ldots,x_n')^\top \in \mathbb Z_{\ge 0}^n$ by setting $x'_i = \lambda_i$ for $i \in I$ and $x'_i=0$ otherwise. The vector $\vec x'$ is a solution of $A \vec x = \vec 0$. Choosing $N \in \mathbb Z_{>0}$ large enough, we can ensure that the vector $\boldsymbol x^\ast + N \boldsymbol x'$ has non-negative components. Hence, $\boldsymbol x = \boldsymbol x^\ast + N \boldsymbol x'$ is a solution of the system $A \vec x = \vec b, \ \vec x \in \mathbb Z_{\ge 0}^n$ satisfying the desired sparsity estimate. The coefficients $\lambda_i$ and the number $N$ can be computed in polynomial time. \end{proof} \begin{proof}[Proof of Corollary~\ref{cor:mixed:sign:knapsack}] The assertion follows by applying Corollary~\ref{cor:pos:spanning:columns} for $m=1$ and all $\tau = \{i\}$ with $i \in [n]$. \end{proof} \section{Proof of Theorem~\ref{thm:positive:knapsack}} \begin{lemma} \label{basic:lemma} Let $a_1,\ldots,a_t \in \mathbb Z_{>0}$, where $t \in \mathbb Z_{> 0}$. If $t > 1 + \log_2 (a_1)$, then the system \begin{align*} & y_1 a_1 + \cdots + y_t a_t = 0, \\ & y_1 \in \mathbb Z_{\ge 0}, \ y_2,\ldots,y_t \in \{-1,0,1\}. \end{align*} in the unknowns $y_1,\ldots,y_t$ has a solution that is not identically equal to zero. \end{lemma} \begin{proof} The proof is inspired by the approach in \cite[\S~3.1]{averkov2012:SIDMA} (used in a different context) that suggests to reformulate the underlying equation over integers as two strict inequalities and then use Minkowski's first theorem \cite[Ch.~VII, Sect.~3]{MR1940576} from the geometry of numbers. Consider the convex set $Y \subseteq \mathbb R^t$ defined by $2t$ strict linear inequalities \begin{align*} -1 < & y_1 a_1 + \cdots + y_t a_t < 1, \\ -2 < & y_i < 2 \text{ for all } i \in \{2,\ldots,t\}. \end{align*} Clearly, the set $Y$ is the interior of a hyper-parallelepiped and can also be described as $Y = \setcond{\boldsymbol y \in \mathbb R^t}{\|M \vec y\|_\infty < 1}$, where $M$ is the upper triangular matrix \[ M = \begin{pmatrix} a_1 & a_2 & \cdots & a_t \\ & 1/2 & & \\ & & \ddots & \\ & & & 1/2 \end{pmatrix}. \] It is easy to see that the $t$-dimensional volume $\operatorname{vol}(Y)$ of $Y$ is \[ \operatorname{vol}(Y)=\operatorname{vol}(M^{-1}[-1,1]^t)=\frac{1}{\det(M)}2^t=\frac{4^t}{2 a_1}.\] The assumption $t > 1 + \log_2 (a_1)$ implies that the volume of $Y$ is strictly larger than $2^t$. Thus, by Minkowski's first theorem, the set $Y$ contains a non-zero integer vector $\boldsymbol y = (y_1,\ldots,y_t)^\top \in \mathbb Z^t$. Without loss of generality we can assume that $y_1 \ge 0$ (if the latter is not true, one can replace $\boldsymbol y$ by $-\boldsymbol y$). The vector $\boldsymbol y$ is a desired solution from the assertion of the lemma. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:positive:knapsack}] Without loss of generality we can assume that $\gcd(\boldsymbol a)=1$. In fact, if $b$ is divisible by $\gcd(\boldsymbol a)$ we can convert $\boldsymbol a^\top \boldsymbol x = b$ to $\overline{\boldsymbol a}^\top \boldsymbol x = \overline{b}$ with $\overline{\boldsymbol a} = \frac{\boldsymbol a}{\gcd(\boldsymbol a)}$ and $\overline{b} = \frac{b}{\gcd(\boldsymbol a)}$, and, if $b$ is not divisible by $\gcd(\boldsymbol a)$, the knapsack feasibility problem $\boldsymbol a^\top \boldsymbol x = b, \ \boldsymbol x \in \mathbb Z_{\ge 0}^n$ has no solution. Without loss of generality, let $ a_1 = \min \{a_1,\ldots,a_n\}$. We need to show the existence of solution of the knapsack feasibility problem satisfying $\|\boldsymbol x\|_0 \le 1 + \log_2( a_1)$. Choose a solution $\boldsymbol x=(x_1,\ldots,x_n)^\top$ of the knapsack feasibility problem with the property that the number of indices $i \in \{2,\ldots,n\}$ for which $x_i \ne 0$ is minimized. Without loss of generality we can assume that, for some $t \in \{2,\ldots,n\}$ one has $x_2 > 0,\ldots, x_t > 0, x_{t+1} = \cdots = x_n = 0$. Lemma~\ref{basic:lemma} implies $t \le 1 + \log_2 (a_1) $. In fact, if the latter was not true, then a solution $\boldsymbol y \in \mathbb R^t$ of the system in Lemma~\ref{basic:lemma} could be extended to a solution $\boldsymbol y \in \mathbb R^n$ by appending zero components. It is clear that some of the components $y_2,\ldots,y_t$ are negative, because $a_2 > 0,\ldots, a_t > 0$. It then turns out that, for an appropriate choice of $k \in \mathbb Z_{\ge 0}$, the vector $\boldsymbol x'= (x_1',\ldots,x'_n)^\top = \boldsymbol x + k \boldsymbol y $ is a solution of the same knapsack feasibility problem satisfying $x_1' \ge 0,\ldots, x'_t \ge 0, \ x'_{t+1} = \cdots = x'_n = 0$ and $x'_i=0$ for at least one $i \in \{2,\ldots,t\}$. Indeed, one can choose $k$ to be the minimum among all $a_i$ with $i \in \{2,\ldots,t\}$ and $y_i=-1$. The existence of $\boldsymbol x'$ with at most $t-1$ non-zero components $x_i'$ with $i \in \{2,\ldots,n\}$ contradicts the choice of $\boldsymbol x$ and yields the assertion. \end{proof} \vskip .3cm \noindent {\bf Acknowledgements} The second author is supported by the DFG (German Research Foundation) within the project number 413995221. The third author acknowledges partial support from the NSF grants 1818969 and CCF-1934568. \bibliographystyle{plain}

https://arxiv.org/abs/1901.03403

Mean Estimation from One-Bit Measurements

We consider the problem of estimating the mean of a symmetric log-concave distribution under the constraint that only a single bit per sample from this distribution is available to the estimator. We study the mean squared error as a function of the sample size (and hence the number of bits). We consider three settings: first, a centralized setting, where an encoder may release $n$ bits given a sample of size $n$, and for which there is no asymptotic penalty for quantization; second, an adaptive setting in which each bit is a function of the current observation and previously recorded bits, where we show that the optimal relative efficiency compared to the sample mean is precisely the efficiency of the median; lastly, we show that in a distributed setting where each bit is only a function of a local sample, no estimator can achieve optimal efficiency uniformly over the parameter space. We additionally complement our results in the adaptive setting by showing that \emph{one} round of adaptivity is sufficient to achieve optimal mean-square error.

\section{Adaptive Estimation \label{sec:sequential}} The first main result of this paper (Theorem~\ref{thm:adpative_lower_bound}) gives that the asymptotic variance of any adaptive estimator must be at least $\eta(0)\sigma^2$, which is precisely the efficiency of the median of the sample $X_1,\ldots,X_n$. Conveniently, the stochastic (sub)gradient estimator for the median---which minimizes $\E[|X - \theta|]$---is a sequence of signs (single bits), so that we can exhibit an asymptotically optimal adaptive estimation scheme. We begin with our first theorem, whose proof we provide in Appendix~\ref{proof:thm:adpative_lower_bound}. \begin{thm}[Fundamental limits]\label{thm:adpative_lower_bound} Let Assumption~\ref{assump:failure_rate} hold. Let ${\theta}_n$ be any estimator of $\theta$ in the adaptive setting of Figure~\ref{fig:setup}(ii). Assume that the prior density $\pi(\cdot)$ on $\theta$ converges to zero at the endpoints of the interval $\Theta$ and define the prior Fisher information $I_0 \defeq \E_\pi[(\pi'(\theta) / \pi(\theta))^2]$. Then \begin{equation*} \E\left[ (\theta-{\theta}_n)^2 \right] \geq \frac{n}{ 4f^2(0) n + I_0}. \end{equation*} \end{thm} It is possible to extend Theorem~\ref{thm:adpative_lower_bound} to other loss functions using a more subtle version of the Van Trees inequality~\cite{efroimovich1980information}; see also \cite{DBLP:journals/corr/abs-1902-08582}. We now turn to asymptotically optimal estimators, first showing how a simple stochastic gradient scheme is asymptotically optimal (in the fully adaptive setting), after which we show that a one-round adaptive scheme can also achieve this optimal efficiency. \subsection{Asymptotically optimal estimator} The starting point for our first estimator is to note that the median of a distribution minimizes $\E[|X - \theta|]$ over $\theta \in \R$, and moreover, we have the familiar result (cf.~\cite{VanDerVaart98}, Ch.~21) that given a sample $X_1,\ldots,X_n \simiid P$, if $\theta = \mbox{med}(P)$ and $P$ has continuous density $f(\cdot - \theta)$ near $\theta$, then \begin{equation*} \sqrt{n}(\mbox{med}(X_1^n) - \theta) \cd \normal\left(0, \frac{1}{4 f(0)^2}\right), \end{equation*} which is precisely the variance lower bound in Theorem~\ref{thm:adpative_lower_bound}. Thus, it is natural to consider a stochastic gradient procedure for minimizing $\E[|X - \theta|]$. To that end, let $\left\{ \gamma_n \right\}_{n\in \mathbb N}$ be a strictly positive sequence of stepsizes, and define the sequence \begin{equation} \label{eq:sgd_alg} \theta_n = \theta_{n-1} + \gamma_n B_n, \quad n = 1,2,\ldots, \end{equation} where \begin{equation*} B_n = \sgn (X_n - \theta_{n-1}). \end{equation*} We make one of two assumptions on the stepsizes $\gamma_n$, which are relatively standard: we always have $\gamma_n$ non-increasing, and \begin{subequations} For some $0 < \lambda \le 1$, \begin{align} \label{eqn:lazy-gamma} \frac{\gamma_n - \gamma_{n+1}}{\gamma_n^2} \to 0, & ~~~ \sum_n \frac{\gamma_n^\frac{1 + \lambda}{2}}{\sqrt{n}} < \infty ~~ \mbox{or} \\ \gamma_n = o(n^{-2/3}), & ~~~ \sum_n \gamma_n = \infty. \label{eqn:stringent-gamma} \end{align} \end{subequations} Then we can adapt the results of Polyak and Juditsky~\cite{PolyakJu92} on the asymptotic normality of averaged stochastic gradient estimators to establish the following theorem. \begin{thm} \label{thm:sgd} Define the average $\bar{\theta}_n \defeq \frac{1}{n} \sum_{i = 1}^n \theta_i$. Assume that in a neighborhood of $\theta = \mbox{med}(P)$, the distribution $P$ has a Lipschitz continuous density $f$. Then \begin{enumerate}[(i)] \item \label{item:normal-sgd} Assume that $\left\{ \gamma_n \right\}_{n\in \mathbb N}$ satisfies condition~\eqref{eqn:lazy-gamma}. Then \begin{equation*} \sqrt{n}\left( \bar{\theta}_n - \theta\right) \cd \normal\left(0,\frac{1}{4 f(0)^2}\right). \end{equation*} % % \item \label{item:sgd-regular} Let $\{P_\theta\}_{\theta \in \R}$ be the family of distributions with density $f(\cdot - \theta)$, where $f$ has median 0. Let $h_n \to h \in \R$, and define the distributions $P_n = P_{\theta + h_n/\sqrt{n}}^n$. Then \begin{equation*} \sqrt{n}\left(\bar{\theta}_n - \theta - h_n / \sqrt{n}\right) \mathop{\cd}_{P_n} \normal\left(0, \frac{1}{4 f(0)^2}\right), \end{equation*} and for any bounded, symmetric, and quasi-convex function $L$, \begin{align} & \sup_{c < \infty} \limsup_{n \to \infty} \sup_{\tau\,:\,|\theta-\tau| \leq \frac{c}{\sqrt{n} }} \E_\tau \left[ L\left( \sqrt{n}(\bar{\theta}_{n} - \tau) \right) \right] \nonumber \\ & \qquad \qquad \qquad \qquad = \mathbb E \left[L(W/ 2 f(0)) \right], \label{eq:attaining_LAM} \end{align} where $W \sim \normal(0,1)$. % % \item \label{item:sgd-ms-convergence} Assume the stepsizes $\gamma_n$ satisfy both conditions~\eqref{eqn:lazy-gamma} and~\eqref{eqn:stringent-gamma}. Let $\pi$ be a distribution on $\R$. Then \begin{equation*} \int \E\Big[( \bar{\theta}_n - \theta )^2\Big] \pi(d\theta) = \frac{1}{4 n f(0)^2} + o(n^{-1}). \end{equation*} \end{enumerate} \end{thm} \noindent We provide the proofs of the items in Appendices~\ref{sec:proof-normal-sgd}, \ref{sec:proof-sgd-regular}, \ref{sec:proof-sgd-ms-convergence}, respectively. As an immediate corollary to Theorem~\ref{thm:sgd}, we obtain the following asymptotic optimality results of the averaged stochastic gradient sequence. Specifically, the average of the stochastic gradient iterates~\eqref{eq:sgd_alg} is locally asymptotically minimax, and they achieve the lower bound of Theorem~\ref{thm:adpative_lower_bound}. \begin{corollary} Let the conditions of Theorem~\ref{thm:adpative_lower_bound} hold and $\theta_n$ be defined by the iteration~\eqref{eq:sgd_alg}. Let $\{P_\theta\}_{\theta \in \R}$ be the family of distributions with densities $f(\cdot - \theta)$. \begin{enumerate}[(i)] \item Define the shorthand $P_n = P_{\theta + h_n/\sqrt{n}}^n$. If the stepsizes satisfy condition~\eqref{eqn:lazy-gamma}, then \begin{equation*} \sqrt{n}(\bar{\theta}_n - \theta - h_n / \sqrt{n}) \mathop{\cd}_{P_n} \normal\left(0, \frac{1}{\eta(0)}\right). \end{equation*} \item If in addition the stepsizes satisfy condition~\eqref{eqn:stringent-gamma}, then they achieve the lower bound of Theorem~\ref{thm:adpative_lower_bound} for any prior $\pi$ on $\R$. \end{enumerate} \end{corollary} \begin{figure} \begin{center} \begin{tikzpicture}[node distance=2cm,auto] \node at (0,0) (source) {$X_1$} ; \node[int1, right of = source, node distance = 1.2cm] (enc1) {$\enc$}; \draw[->,line width = 2pt] (source) -- (enc1); \node[below of = source, node distance = 1.7cm] (source3) {$X_{n_1}$}; \node[int1, right of = source3, node distance = 1.2cm] (enc3) {$\enc$}; \draw[->,line width = 2pt] (source3) -- (enc3); \node[below of = source, node distance = 0.5cm] {$\vdots$}; \node[int1, right of = enc3, node distance = 2.1cm ] (est) {$\est_1$}; \draw[->,line width = 2pt] (enc1) -| node[above, xshift = -1cm] (mes1) {$B_1$} (est); \draw[->,line width = 2pt] (enc3) -- node[above, xshift = 0cm] {$B_{n_1}$} (est); \node[below right = 0.75 and 1.5 of source3] (sourceB) {$X_{n_1 +1}$} ; \node[int1, right of = sourceB, node distance = 1.7cm] (enc1B) {$\enc$}; \draw[->,line width = 2pt] (sourceB) -- (enc1B); \node[below of = sourceB, node distance = 1.7cm] (source3B) {$X_n$}; \node[int1, right of = source3B, node distance = 1.7cm] (enc3B) {$\enc$}; \draw[->,line width = 2pt] (source3B) -- (enc3B); \node[below of = sourceB, node distance = 0.4cm] {$\vdots$}; \node[int1, right of = enc3B, node distance = 2.1cm ] (estB) {$\est_2$}; \draw[->,line width = 2pt] (enc1B) -| node[above, xshift = -0.5cm] {$B_{n_1+1}$} (estB); \draw[->,line width = 2pt] (enc3B) -- node[above] {$B_n$} (estB); \draw[->,line width = 1pt] (est.east) node[above, xshift =0.5cm] {${\theta}_{n_1}$} -| (enc1B.north); \draw[->,line width = 0.5pt] (est.east) -| +(1.3,-0.5) -- +(1.3,-2.5) -| (enc3B.north); \draw[->,line width = 0.5pt] (estB) -- +(0.8,0) node[right] {${\theta}_n$}; \node[below of = enc1B, node distance = 0.5cm] {$\vdots$}; \end{tikzpicture} \end{center} \caption{Distributed encoding with a single interaction: The estimation obtained from the first $n_1$ bits in a distributed manner is to obtain another $n-n_1$ bits in a distributed manner. \label{fig:one_round} } \end{figure} \subsection{Maximal Efficiency using One Round of Threshold Adaptation} In the encoding and estimating procedure \eqref{eq:sgd_alg}, each one-bit message $B_n$ depends on its private sample as well as the current gradient descent estimate $\theta_{n-1}$. In this sense, each encoder in this algorithm interacts with previous one by using the current estimate. This amount of adaptivity is unnecessary: as we now consider, a similar encoding yields an asymptotically normal estimator attaining the lower variance bound $1/\eta(0)$, provided we allow \emph{one} adaptive update to the threshold value $\theta_0$ based on previously observed bits. In this procedure we separate the sample into the disjoint sets $X_1,\ldots,X_{n_1}$ and $X_{n_1+1},\ldots,X_n$ for some $n_1 < n$. We first use the estimator \eqref{eq:estimator_naive} to obtain an estimate ${\theta}_{n_1}$ based on $B_1,\ldots,B_{n_1}$, and then use ${\theta}_{n_1}$ as the new threshold value to obtain messages $B_{n_1+1}, \ldots, B_n$. Figure~\ref{fig:one_round} illustrates a diagram of this procedure. More formally, we consider the following estimation scheme. Given $n_1 \in \{1, \ldots, n\}$, set the individual bits \begin{equation*} B_i = \begin{cases} \indic{X_i \le \theta_0} & i = 1,\ldots,n_1, \\ \indic{X_i \le T_n}& i={n_1+1,\ldots,n}, \end{cases} \end{equation*} where \begin{equation*} T_n \defeq \theta_0 - F^{-1}\left( \frac{1}{n_1} \sum_{i=1}^{n_1} B_i \right) ~~ \mbox{and} ~~ \theta_n \defeq T_n - F^{-1} \left(\frac{1}{n - n_1} \sum_{i = n_1}^n B_i \right). \end{equation*} The intuition here is that the estimator $\theta_n$ is a one-step correction (cf.~\cite[Thm.~6.4.3]{LehmannCa98}) of the initial estimator $T_n$, which approximately estimates $\theta_0 - F^{-1}(\theta_0 - \theta) = \theta$. We then have the following convergence result. \begin{thm} Assume that $X_i = Z_i + \theta$, where $Z_i$ are i.i.d.\ with density $f$ and CDF $F$ and $\mbox{med}(Z_i) = 0$. Assume that $f$ is continuous at 0, and that as $n \to \infty$, $n_1(n) \rightarrow \infty$ and $n_1 / n \to 0$. Then \begin{align*} \sqrt{n} \left( {\theta}_n - \theta \right) \cd \normal\left( 0, \frac{1}{4 f(0)^2}\right). \end{align*} \end{thm} \noindent That is, under Assumption~\ref{assump:failure_rate}, the method is asymptotically optimal. \begin{proof} We abuse notation and instead of assuming we receive $n$ observations, assume we receive the $n + n_1$ observations $X_{-n_1}, \ldots, X_{-1}$ and $X_1, \ldots, X_n$, defining $T_n = \theta_0 - F^{-1}(\frac{1}{n_1} \sum_{i = -n_1}^{-1} B_i)$ and $B_i = \indic{X_i \le T_n}$ for $i \ge 1$. Letting $X_i = Z_i + \theta$ for $Z_i$ i.i.d.\ with fixed density $f = F'$, we have $B_i \cas F(\theta_0 - \theta)$, so that $\frac{1}{n_1} \sum_{i = -n_1}^{-1} B_i \cas F(\theta - \theta_0)$ and by the continuous mapping theorem we have $T_n \cas \theta$ as $n_1 \to \infty$. Now let $E_n = \E[B_i \mid T_n] = P(X_i \le T_n)$, so that $\var(B_i \mid T_n) = E_n(1 - E_n)$. Define also the random variable \begin{equation*} Y_n \defeq \sqrt{n} \frac{1}{\sqrt{E_n (1 - E_n)}} \bigg[\frac{1}{n}\sum_{i = 1}^n B_i - E_n\bigg], \end{equation*} and let $F_n(\cdot \mid T_n)$ be its cumulative distribution function. Then because \begin{equation*} \E\left[|B_i - E_n|^3 \mid T_n\right] \le E_n(1 - E_n) ~~ \mbox{we have} ~~ \E\left[\frac{|B_i - E_n|^3}{(E_n(1 - E_n))^{3/2}} \mid T_n\right] \le \frac{1}{\sqrt{E_n(1 - E_n)}}. \end{equation*} The Berry-Esseen theorem implies that there exists a constant $C \le 1$ such that \begin{equation*} \sup_t \left|F_n(t \mid T_n) - \Phi(t) \right| \le \frac{C}{\sqrt{E_n (1 - E_n)} \sqrt{n}} \wedge 2, \end{equation*} where $\Phi$ is the standard Gaussian CDF. As $E_n (1 - E_n) \cas \frac{1}{4}$ by definition of the median, we have that (with probability 1) \begin{equation*} \sup_t \left|F_n(t \mid T_n) - \Phi(t)\right| \le \frac{C}{\sqrt{n}} ~~ \mbox{eventually}. \end{equation*} By dominated convergence and Jensen's inequality we thus obtain \begin{equation*} \sup_t \left|\mathbb{P}(Y_n \le t) - \Phi(t)\right| \le \E\left[ \sup_t \left|F_n(t \mid T_n) - \Phi(t)\right| \right] \to 0, \end{equation*} which gives that $Y_n \cd \normal(0, 1)$. Now, Slutsky's lemmas imply \begin{equation} \sqrt{n} \cdot \frac{2}{n} \sum_{i = 1}^n (B_i - E_n) = (1 + o_P(1)) \frac{1}{\sqrt{n E_n(1 - E_n)}} \sum_{i = 1}^n \left(B_i - E_n\right) \cd \normal(0, 1). \label{eqn:apply-slutsky} \end{equation} For shorthand we write $P_n B = \frac{1}{n} \sum_{i = 1}^n B_i$, and using that $E_n = \E[B_i \mid T_n] = F(T_n - \theta)$, we may thus use the delta method to write \begin{align*} \sqrt{n}(\theta_n - \theta) & = \sqrt{n}\left(T_n - F^{-1}\left(P_n B\right) - \theta \right) \\ & = \sqrt{n} \left(T_n - F^{-1}\left(F(T_n - \theta) + P_n B - F(T_n - \theta)\right) - \theta \right) \\ & = \sqrt{n} \left(T_n - (T_n - \theta ) + (F^{-1})'(T_n - \theta + o_P(1))\cdot \left( P_n B - E_n\right) - \theta \right) \\ & = \sqrt{n} (F^{-1})'(0) (P_n B - E_n) + o_P(1) \cd \normal\left(0, \frac{1}{4 f(0)^2}\right), \end{align*} where we have used the limiting distribution~\eqref{eqn:apply-slutsky}. \end{proof} \begin{figure} \begin{center} \begin{tikzpicture}[scale = 0.6] \begin{axis}[ width=10cm, height=6cm, xmin = 0, xmax=800, restrict y to domain = 0:3, ymin = 0, ymax = 3.4, samples=10, xlabel= $n$, ylabel = {$n\mathbb E \left[\left(\theta - {\theta}_n \right)^2 \right]$}, ytick={0,1,1.57}, yticklabels={0,1,$\pi/2$}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, solid, smooth] plot table [x = itr, y = SGD, col sep=comma] {./SimRes/sim_res_nMonte5000.csv}; \addlegendentry{asymptotically optimal}; \addplot[color = red, solid, smooth] plot table [x = itr, y = split, col sep=comma] {./SimRes/sim_res_nMonte5000.csv}; \addlegendentry{single interaction}; \end{axis} \end{tikzpicture} \caption{Normalized empirical risk versus number of samples $n$ for $10,000$ Monte Carlo trials with $f(x)$ the standard normal density. In each trial, $\theta$ is chosen uniformly over the interval $(-1.64,1.64)$. The single interaction strategy uses $n_1 = \lfloor \sqrt{n} \rfloor$ samples for its first stage. \label{fig:adaptive_error} } \end{center} \end{figure} \section{Conclusions \label{sec:conclusions}} We considered the risk and efficiency in estimating the mean of a symmetric and log-concave distribution from a sequence of bits, where each bit is obtained by encoding a single sample from this distribution. In an adaptive encoding setting, we showed that no estimator can attain asymptotic relative efficiency larger than that of the median of the samples. We also showed that this bound is tight by presenting two adaptive encoding and estimation procedures that are as efficient as the median. \par In the distributed setting we provided conditions for local asymptotic normality of the encoded samples, which implies asymptotic minimax bound on both the risk and efficiency relative to the mean. These condition implies that, in the distributed setting, the optimal estimation performance derived for the adaptive case can only be attained over a finite number of points, i.e., no scheme is uniformly optimal in this setting. We further considered the special case where the sequence of bits is obtained by comparing against a prescribed sequence of thresholds. We characterized the performance of the optimal estimator from such bit-sequence using the density of the thresholds and considered the density that minimizes the minimax risk. \section{Distributed Estimation \label{sec:distributed}} We now consider the distributed encoding setting in Figure~\ref{fig:setup}-(iii) where each one-bit message $B_i$ is a function only of its private sample $X_i$. In this case, the $i$th encoder is of the form $B_i = \indic{X_i \in A_i}$, where the detection region $A_i$ is a Borel set independent of $X_1, X_2, \ldots$. We begin by making a few restrictions on the collections of the sets $A_i$, which we believe not unreasonable, but which allow us to develop fundamental limits for estimation. We require a bit of notation to define the assumptions. As we work with a location family based on a density $f$ with associated probability distribution $P$ on variables $Z$, we define \begin{equation*} P_\theta(A) \defeq P(Z - \theta \in A) \end{equation*} for $Z$ with density $f$. Whenever $A$ is a collection of disjoint intervals $A = \cup_i \{[t_i^-, t_i^+]\}$, we may define \begin{equation*} \dPtheta(A) \defeq \frac{\partial}{\partial \theta} P_\theta(A) = \sum_i \left(f(t_i^- - \theta) - f(t_i^+ - \theta)\right), \end{equation*} and similarly we define the score function $\score_\theta(A) \defeq \dPtheta(A) / P_\theta(A)$. For $B = \indic{X \in A}$, we abuse notation and also write $\score_\theta(B) = \score_\theta(A)$ and similarly for $\dPtheta$. With this, we may define the variance of the scores $\score_\theta(B_i)$ under $P_\theta$ via \begin{equation} \label{eq:precision_general} L_n(A_1,\ldots,A_n;\theta) \defeq \frac{1}{n} \sum_{i=1}^n \frac{ \dPtheta(A_i)^2}{ P_\theta(A_i) (1 - P_\theta(A_i))}. \end{equation} We then make the following assumption. \begin{assumption} \label{assumption:detection-regions} The density and detection regions satisfy \begin{enumerate}[(i)] \item \label{item:lipschitz-density} The density function $f$ of $X_n - \theta$ is Lipschitz continuous. \item \label{item:finite-intervals} Each set $A_n$ is the finite union of $k_n$ disjoint intervals (which may include $\pm \infty$), where \begin{equation*} \frac{1}{n} \cdot \max_{i \le n} \frac{k_i^3}{P_\theta(A_i)^4 (1 - P_\theta(A_i))^4} \to 0. \end{equation*} \item \label{item:limit-variance} The limit \begin{equation} \label{eqn:LAN-limit} \kappa(\theta) \defeq \lim_{n\to \infty} L_n(A_1,\ldots,A_n; \theta) \end{equation} exists and is finite. \end{enumerate} \end{assumption} Under this assumption, we have the following theorem, which provides a local asymptotic minimax lower bound on the efficiency of \emph{any} non-adaptive estimator. \begin{theorem} \label{theorem:non-adaptive-minimax} Let Assumption~\ref{assumption:detection-regions} hold, and let ${\theta}_n$ be an estimator of $\theta \in \Theta$ from observations $B_i = \indic{X_i \in A_i}$. Then for $Z \sim \normal(0, 1)$ and any symmetric and quasi-convex function $L$, \begin{align*} & \liminf_{c \to \infty} \liminf_{n \to \infty} \sup_{\tau\,:\,|\theta-\tau| \leq \frac{c}{\sqrt{n} }} \E \left[ L\left( \sqrt{n}({\theta}_{n} - \tau) \right) \right] \\ & \qquad \qquad \qquad \qquad \geq \E\left[ L (Z/\sqrt{\kappa(\theta)}) \right]. \end{align*} \end{theorem} \noindent See Appendix~\ref{sec:proof-non-adaptive-minimax} for a proof. Theorem~\ref{theorem:non-adaptive-minimax} shows that the limiting variance term $\kappa(\theta)$ provides strong lower bounds on the efficiency of any non-adaptive estimator, and moreover, that these bounds necessarily depend on $\theta$. This suggests that attaining any type of good (uniform) efficiency with non-adaptive estimators will be challenging. As a particular consequence, for the squared error $L(x) = x^2$, for any $\delta > 0$ and $\theta$, there exists a $c < \infty$ such that $\sup_{|\tau - \theta| \le c / \sqrt{n}} \E_\tau[(\theta_n - \tau)^2] \ge \frac{(1 - \delta)}{n \kappa(\theta)} + o(1/n)$. Yet, Theorem~\ref{theorem:non-adaptive-minimax} limits non-adaptive strategies in stronger ways. Under the density models we have considered, with the additional Assumption~\ref{assump:failure_rate}, we can show stronger optimality results that adaptivity is essential for achieving optimal convergence guarantees. Recall the transformation~\eqref{eq:eta_def} of the hazard rate function, $\eta(x) = \frac{f^2(x)}{F(x)(1 - F(x))}$, which has unique maximum at $x = 0$ under Assumption~\ref{assump:failure_rate}. When each detection region $A_n$ consists of a bounded number of intervals, the next theorem shows that the minimal risk $1/\eta(0)$ can only be attained at finitely many points within $\Theta$. In particular, distinct from the adaptive setting, no distributed estimation scheme can achieve asymptotic variance $\eta(0)$ uniformly in $\theta \in \Theta$. \begin{thm} \label{thm:non_existence} Let Assumptions~\ref{assump:failure_rate} and~\ref{assumption:detection-regions} hold. Additionally, assume that $A_i$ is the union of at most $K$ intervals. The number of points $\theta \in \Theta$ satisfying $\kappa(\theta) = \eta(0)$ is at most $2K$. \end{thm} \noindent See Appendix~\ref{proof:thm:non_existence} for a proof. \subsection{Threshold Detection} \label{subsec:threshold} We now consider consider a restricted case where each detection region is a half-open interval, i.e., the $i$th message is obtained by comparing $X_i$ against a single threshold. Under adaptive signal acquisition, this is sufficient for asymptotic optimality; in non-adaptive settings, it is not sufficient, though we may characterize a few additional optimality results. Assume now that each $B_i$ is of the form \begin{equation} \label{eq:threshold_message} B_i = \sgn(t_i - X_i) = \begin{cases} 1 & X_i< t_i, \\ -1 & X_i > t_i, \end{cases} \end{equation} where $t_i\in\mathbb R$ is the \emph{threshold} of the $i$th encoder. In other words, the detection region of $B_i$ is $A_i = (t_i,\infty)$ and $\mathbb P(X_i \in A_i) = F \left( B_i(t_i-\theta) \right)$. It follows that \begin{equation} L_n(A_1,\ldots,A_n;\theta) = \frac{1}{n} \sum_{i=1}^n \frac{ \left(f(t_i-\theta) \right)^2 }{F\left(t_i-\theta \right) F\left(\theta - t_i \right) } = \frac{1}{n} \sum_{i=1}^n \eta(t_i - \theta). \label{eq:Ln_threshold} \end{equation} A natural condition for the existence of the limit \eqref{eq:Ln_threshold} as $n\to \infty$ is that the empirical distribution of the threshold values converges to a probability measure. Specifically, for an interval $I \subset \mathbb R$, define \begin{equation*} \lambda_n(I) = \frac{ \card \left( I \cap \{t_1,t_2,\ldots \} \right)}{n}. \end{equation*} Then an investigation of the proof of Theorem~\ref{theorem:non-adaptive-minimax} in Section~\ref{sec:proof-non-adaptive-minimax}, specifically Sec.~\ref{sec:proof-lan-bits} and the bounds~\eqref{eqn:h-fourth}, show that as $\eta(t) \le \eta(0)$ for all $t \in \R$ under Assumption~\ref{assump:failure_rate}, the following corollary follows. (The corollary relies on local asymptotic normality~\cite[Ch.~7]{VanDerVaart98}; see Appendix~\ref{sec:proof-sgd-regular} for some brief discussion of such conditions.) \begin{cor} \label{cor:LAN_thresh} Let $\{t_n\}_{n=1}^\infty$ be a sequence of threshold values such that $\lambda_n$ converges (weakly) to a probability measure $\lambda$ on $\mathbb R$. Then the conclusions of Theorem~\ref{theorem:non-adaptive-minimax} apply with \begin{equation*} \kappa(\theta) = \int_{\mathbb R} \eta(t-\theta) \lambda(dt). \end{equation*} Moreover, the family of laws of $\{B_i = \sgn(X_i - t_i)\}_{i = 1}^n$ under $\{P_\theta\}_{\theta \in \Theta}$ is locally asymptotically normal with information $\kappa(\theta)$. \end{cor} The condition that $\lambda_n$ converges to a probability measure is satisfied, for example, whenever the $t_1,\ldots,t_n$s are drawn independently from a probability distribution $\lambda(dt)$ on $\mathbb R$. When the conclusions of Corollary~\ref{cor:LAN_thresh} hold, local asymptotic normality of $\{B_n\}_{n=1}^\infty$ implies that the maximum likelihood estimator (ML) of $\theta$ from $B_1,\ldots,B_n$, denoted here by ${\theta}^{ML}_n$, is local asymptotic minimax in the sense that \begin{equation*} \sqrt{n} \left( {\theta}^{ML}_n - \theta \right) \cd \normal\left(0, 1/\kappa(\theta) \right). \end{equation*} We note that ${\theta}^{ML}_n$ solves \begin{equation} \label{eq:ML} 0 = \sum_{i=1}^n B_i \frac{f \left( t_i-\theta\right) }{F \left(B_i (t_i-\theta)\right) }. \end{equation} If the collection $\{t_1,t_2\ldots\}$ is bounded (for example $\{t_1,t_2\ldots\} \subset \Theta$), then \begin{equation*} \lim_{n\to \infty} n \ex{\left({\theta}^{ML}_n - \theta \right)^2} = 1/\kappa(\theta), \end{equation*} so that the ML estimator attains the local asymptotic MSE of Theorem~\ref{thm:adpative_lower_bound}. By Assumption~\ref{assump:failure_rate}, $\eta(x)$ attains its maximum at the origin, so we conclude that \begin{equation*} \kappa(\theta) \leq \sup_{t\in \mathbb R} \eta \left( t-\theta\right) = \eta(0). \end{equation*} Moreover, this upper bound on $\kappa(\theta)$ is attained only when $\lambda$ is the point mass at $\theta$. Since $\theta$ is \emph{a priori} unknown, we conclude that estimation in the distributed setting using threshold detection is strictly suboptimal compared to the adaptive setting; the ability to choose the thresholds $t_i$ adaptively conditional on previous messages is necessary for optimal efficiency. \subsection{Minimax Threshold Density} We conclude this section by considering the density of the threshold values that maximizes the worst-case information $\inf_\theta \kappa(\theta) = \kappa_\lambda(\theta)$ where $\kappa_\lambda(\theta) = \int \eta(t - \theta) d\lambda(t)$. This distribution $\lambda$ solves the optimization problem \begin{align} \label{eq:var_cvx_minimax} \begin{split} \mathrm{maximize} \quad & \inf_{\theta \in \Theta} \int \eta(t-\theta) \lambda(dt) \\ \mathrm{subject~to} \quad & \lambda(dt)\geq 0,\quad \int \lambda(dt) \leq 1. \end{split} \end{align} The objective function~\eqref{eq:var_cvx_minimax} is concave in $\lambda(dt)$ and continuous in the weak topology over measures on $\Theta$, so that by discretizing, we can approximately solve this problem using convex optimization. We let $\kappa^\star$ denote the maximal value of problem~\eqref{eq:var_cvx_minimax} and $\lambda^\star(dt)$ be the density achieving the maximum. By drawing thresholds $t_i \simiid \lambda^\star$, Corollary~\ref{cor:LAN_thresh} guarantees that for any $\theta \in \Theta$, the maximum likelihood estimator using $\{B_i = \sgn(X_i - t_i)\}_{i \in \N}$ is at least $\kappa^\star$. Figure~\ref{fig:minimax_support} illustrates an approximation to $\lambda^\star(dt)$ obtained by solving a discretized version of \eqref{eq:var_cvx_minimax} for the case when $f(x)$ is the normal density with variance $\sigma^2$ and $\Theta = [-T,T]$. The minimax asymptotic precision parameter $\kappa^\star$ obtained this way is illustrated in Fig.~\ref{fig:minimax_ARE} as a function of half the support size $T$. Also illustrated in these figures is $\kappa_{\unif}$, the precision parameter corresponding to threshold values uniformly distribution over $\Theta = [-T,T]$, \begin{align} & \kappa_{\unif} \triangleq \min_{\theta \in [-T,T]} \frac{1}{2T}\int_{-T}^T \eta\left(t-\theta\right) dt \nonumber \\ & = \frac{1}{2T}\int_{-T}^{T} \eta\left(t\pm T\right) dt = \frac{1}{2T}\int_{0}^{2T} \eta(t) dt \label{eq:uniform_risk}. \end{align} \begin{figure} \begin{center} \begin{tikzpicture}[scale = 0.55] \begin{axis}[ width=8cm, height=6cm, xmin = -0.5, xmax=0.5, restrict y to domain = 0:100, ymin = 0, ymax = 1, samples=10, xlabel= {$dt$}, xtick={-0.5,0,0.5}, xticklabels={-$T$, 0, $T$}, title = {$\sigma/T = 1/2$}, ytick={0,0.1,1}, yticklabels={0,0.1,1}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, smooth, mark = o ] plot table [col sep=comma] {./Figs/minmax_lmd_b0.5_sig0.5.csv}; \addplot[color = black!30!green, smooth, dashed] plot table [x = x, y = z,col sep=comma] {./Figs/minimax_th_b0.5_sig0.5.csv}; \addplot[color = red, smooth] plot table [x = x, y = y,col sep=comma] {./Figs/minimax_th_b0.5_sig0.5.csv}; \end{axis} \end{tikzpicture} \begin{tikzpicture}[scale = 0.55] \begin{axis}[ width=8cm, height=6cm, xmin = -0.5, xmax=0.5, restrict y to domain = 0:100, ymin = 0, ymax = 0.6, samples=10, xlabel= {$dt$}, title = {$\sigma/T = 1/5$}, xtick={-0.5,0,0.5}, xticklabels={-$T$, 0, $T$}, ytick={0,0.5}, ylabel = {$\lambda(dt)$}, yticklabels={0,0.5}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, smooth, mark = o ] plot table [col sep=comma] {./Figs/minmax_lmd_b0.5_sig0.2.csv}; \addplot[color = black!30!green, smooth, dashed] plot table [x = x, y = z,col sep=comma] {./Figs/minimax_th_b0.5_sig0.2.csv}; \addplot[color = red, smooth] plot table [x = x, y = y,col sep=comma] {./Figs/minimax_th_b0.5_sig0.2.csv}; \end{axis} \end{tikzpicture} \begin{tikzpicture}[scale = 0.55] \begin{axis}[ width=8cm, height=6cm, xmin = -0.5, xmax=0.5, restrict y to domain = 0:100, ymin = 0, ymax = 0.3, samples=10, xlabel= {$dt$}, xtick={-0.5,0,0.5}, title = {$\sigma / T = 1/10$}, xticklabels={-$T$, 0, $T$}, ytick={0,0.2}, yticklabels={0,0.2}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, smooth, mark = o ] plot table [col sep=comma] {./Figs/minmax_lmd_b0.5_sig0.1.csv}; \addplot[color = black!30!green, smooth, dashed] plot table [x = x, y = z,col sep=comma] {./Figs/minimax_th_b0.5_sig0.1.csv}; \addplot[color = red, smooth] plot table [x = x, y = y,col sep=comma] {./Figs/minimax_th_b0.5_sig0.1.csv}; \end{axis} \end{tikzpicture} \begin{tikzpicture}[scale = 0.55] \begin{axis}[ width=8cm, height=6cm, xmin = -0.5, xmax=0.5, restrict y to domain = 0:100, ymin = 0, ymax = 0.15, samples=10, xlabel= {$dt$}, title = {$\sigma / T = 1/20$}, xtick={-0.5,0,0.5}, xticklabels={-$T$, 0, $T$}, ytick={0,0.1}, ylabel = {$\lambda(dt)$}, yticklabels={0,0.1}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, smooth, mark = o ] plot table [x = x, y = y,col sep=comma] {./Figs/minmax_lmd_b0.5_sig0.05.csv}; \addplot[color = black!30!green, smooth, dashed] plot table [x = x, y = z,col sep=comma] {./Figs/minimax_th_b0.5_sig0.05.csv}; \addplot[color = red, smooth] plot table [x = x, y = y,col sep=comma] {./Figs/minimax_th_b0.5_sig0.05.csv}; \end{axis} \end{tikzpicture} \caption{\label{fig:minimax_support} Optimal threshold density $\lambda^\star(dt)$ (blue) that maximizes the ARE for $f(x) = \Ncal(\theta,\sigma^2)$ and $\theta \in [-T,T]$. The continuous curve (red) represents the reciprocal of the asymptotic risk for each $\theta \in [-T,T]$ under the optimal density, i.e., the minimax risk is the inverse of the minimal value of this curve. The dashed curve (green) is the reciprocal of the asymptotic risk for each $\theta \in [-T,T]$ when the threshold values are uniformly distributed over $[-T,T]$, hence its minimal value is the inverse of \eqref{eq:uniform_risk}. } \end{center} \end{figure} \begin{figure} \begin{center} \begin{tikzpicture}[scale = 1] \begin{axis}[ width=8cm, height=6cm, xmax=2.5, xmode = log, restrict y to domain = 0:100, ymin = 0, ymax = 1, samples=1, xlabel= {$\sigma/T$}, ytick={0,0.637,1}, yticklabels={0,$2/\pi$,1}, ylabel = {$\ARE$}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = red, smooth, line width = 1pt] plot table [x = x, y = y, col sep=comma] {./Figs/minmax_ARE_b0.5.csv}; \addlegendentry{optimal threshold density}; \addplot[color = black!35!green, smooth, dashed, line width = 1pt] plot table [x = x, y = z, col sep=comma] {./Figs/minmax_ARE_b0.5.csv}; \addlegendentry{uniform threshold desnity}; \end{axis} \end{tikzpicture} \caption{\label{fig:minimax_ARE} Minimax ARE versus $\sigma/T$ for $f(x) = \Ncal(\theta,\sigma^2)$ and $\theta \in \Theta = [-T,T]$. The dashed curve (green) is the ARE under a uniform threshold density over $\Theta$ given by $K_{\unif}\sigma^2$, where $\kappa_{\unif}$ is given by \eqref{eq:uniform_risk}. }. \end{center} \end{figure} \begin{figure} \begin{center} \begin{tikzpicture}[scale = 0.55] \begin{axis}[ width=8cm, height=6cm, xmin = -0.5, xmax=0.5, restrict y to domain = 0:100, ymin = 0, ymax = 1, samples=10, xlabel= {$dt$}, title = {$\sigma/\sigma_{\theta} = 4$}, xtick={-0.5,0,0.5}, xticklabels={-$T$, 0, $T$}, ytick={0,1}, ylabel = {$\lambda(dt)$}, yticklabels={0,1}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, smooth, mark = o ] plot table [col sep=comma] {./Figs/unif_Bayes_lmd_b0.5_sig4.csv}; \end{axis} \end{tikzpicture} \begin{tikzpicture}[scale = 0.55] \begin{axis}[ width=8cm, height=6cm, xmin = -0.5, xmax=0.5, restrict y to domain = 0:100, ymin = 0, ymax = 1, samples=10, xlabel= {$dt$}, xtick={-0.5,0,0.5}, title = {$\sigma/\sigma_{\theta} = 3$}, xticklabels={-$T$, 0, $T$}, ytick={0,1}, yticklabels={0,1}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, smooth, mark = o ] plot table [col sep=comma] {./Figs/unif_Bayes_lmd_b0.5_sig3.csv}; \end{axis} \end{tikzpicture} \begin{tikzpicture}[scale = 0.55] \begin{axis}[ width=8cm, height=6cm, xmin = -0.5, xmax=0.5, restrict y to domain = 0:100, ymin = 0, ymax = 1, samples=10, xlabel= {$dt$}, title = {$\sigma/\sigma_{\theta} = 2$}, xtick={-0.5,0,0.5}, xticklabels={-$T$, 0, $T$}, ytick={0,1}, ylabel = {$\lambda(dt)$}, yticklabels={0,1}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, smooth, mark = o ] plot table [col sep=comma] {./Figs/unif_Bayes_lmd_b0.5_sig2.csv}; \end{axis} \end{tikzpicture} \begin{tikzpicture}[scale = 0.55] \begin{axis}[ width=8cm, height=6cm, xmin = -0.5, xmax=0.5, restrict y to domain = 0:100, ymin = 0, ymax = 1, samples=10, xlabel= {$dt$}, xtick={-0.5,0,0.5}, xticklabels={-$T$, 0, $T$}, title = {$\sigma/\sigma_{\theta} = 1$}, ytick={0,1}, yticklabels={0,1}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, smooth, mark = o ] plot table [col sep=comma] {./Figs/unif_Bayes_lmd_b0.5_sig1.csv}; \end{axis} \end{tikzpicture} \caption{\label{fig:opt_density} Optimal threshold density $\lambda^\star(dt)$ that minimizes the asymptotic Bayes risk \eqref{eq:cvx_average} for a uniform prior with $\sigma/\sigma_\theta=1,2,3,4$, where $\sigma_\theta^2=T^2/3$ is the variance of the prior. } \end{center} \end{figure} When a prior $\pi$ on $\Theta$ is provided, one may be interested in the threshold density $\lambda$ minimizing the Bayes risk $R_n(\pi)$. The resulting minimization problem is \begin{align} \label{eq:cvx_average} \begin{split} \mathrm{minimize} \quad & R_{\pi} = \int \frac{\pi(d\theta)}{ \int \eta \left( t-\theta\right) \lambda(dt)}. \\ \mathrm{subject~to} \quad & \lambda(dt)\geq 0,\quad \int \lambda(dt) =1, \end{split} \end{align} which is convex in $\lambda$ since $x \rightarrow 1/x$, $x>0$, is convex. Figure~\ref{fig:opt_density} illustrates the solution to \eqref{eq:cvx_average} for the case of where $f(x)$ is the normal distribution and $\pi$ is the uniform distribution over $\Theta = [-T,T]$. \section{Introduction} \label{sec:Intro} We consider estimation of parameters from data collected by multiple units under communication constraints between the units. Such scenarios arise in sensor arrays, where sensor motes collect information, which they transmit to a central estimation unit~\cite{LesserOrTa03,LiWoHuSa02}. More generally, communication is substantially more expensive than computation in modern computing infrastructure~\cite{FullerMi11}. It is thus of interest to understand the extent to which communication constraints induce fundamental accuracy and efficiency limits in parametric estimation problems. \begin{figure*} \include{Figs/encoding-settings} \caption{\label{fig:setup} Three encoding settings: (i) Centralized -- an encoder sends $n$ bits after observing $n$ samples. (ii) Adaptive (sequential) -- the $i$th encoder sends the bit $B_i$ depending on its private sample $X_i$ and previous bits $B_1,\ldots,B_{i-1}$. (iii) Distributed -- each encoder send the bit $B_i$ based on its private sample $X_i$ only.} \end{figure*} We answer this question in a sylized version of this problem: the estimation of the mean $\theta$ of a symmetric log-concave distribution under the constraint that only a single bit can be communicated about each observation from this distribution. Different information sharing schemes strongly affect the performance of estimators for $\theta$; we illustrate the three main settings we consider in Figure~\ref{fig:setup}. \begin{enumerate}[(i)] \item \emph{Centralized} encoding: all $n$ encoders confer and produce a single message consists of $n$ bits. \item \emph{Adaptive} or \emph{sequential} encoding: The $n$th encoder observes the $n$th sample and the $n-1$ previous bits. \item \emph{Distributed} encoding: The $n$th message is only a function of the $n$th sample. \end{enumerate} The distributed setting~(iii) is the most restrictive; as it turns out, (ii) is slightly more restrictive than the fully centralized setting~(i), and in our setting, a variant of the adaptive setting~(ii) in which there is only \emph{one} round of adaptivity---as we make formal later---is enough to achieve the same efficiency as the fully sequential setting~(ii). Each setting has natural applications: \begin{itemize} \item {\bf Signal acquisition (i):} A quantity is measured $n$ times at different instances. The results are averaged in order to reduce measurement noise and the averaged result is then stored or communicated using $n$ bits. \item {\bf Analog-to-digital conversion (ii):} A sigma-delta modulator (SDM) converts an analog signal into a sequence of bits by sampling the signal at a very high rate and then using one-bit threshold detector combined with a feedback loop to update an accumulated error state \cite{1092194}. Therefore, the expected error in tracking an analog signal using an SDM falls under our setting (ii) when we assume that the signal at the input to the modulator is a constant (direct current) corrupted by, say, thermal noise \cite{53738}. Since the sampling rates in SDM are usually many times more than the bandwidth of its input, analyzing SDM under a constant input provides meaningful lower bound even for non-constant signals. \item {\bf Privacy (ii)--(iii):} A business entity is interested in estimating the average income of its clients. In order to keep this information as confidential as possible, each client independently provides an answer to a yes/no question related to its income~\cite{DuchiJoWa18}. \end{itemize} Let us provide an informal description of our results and setting. For an estimator $\theta_n$ with finite quadratic risk (mean squared error (MSE)) $R_n = \E_\theta[(\theta_n - \theta)^2]$, we study the limit \begin{equation} \label{eq:ARE_def} \limsup_{n\to \infty} n R_n. \end{equation} By comparing this quantity to achievable rates of convergence without communication constraints, we can evaluate the efficiency losses---asymptotic relative efficiency---of the estimator to appropriately optimal (unconstrained) estimators. (We shall be more formal in the sequel.) By lower bounding the quantity~\eqref{eq:ARE_def}, we also provide limits on estimation of single-bit-per-measurement constrained signals in more general settings~\cite{baraniuk2017exponential, jacques2013robust, plan2013one, li2017channel, choi2016near}. In setting (i), the estimator can evaluate any optimal estimator of location (e.g., the sample mean if the data is Gaussian), then quantize it using $n$ bits. As the accuracy in describing the empirical mean decreases exponentially in $n$, the quantization error is negligible compared to the statistical error in mean estimation~\cite{720540}. That is, centralized encoding induces no asymptotic efficiency loss. The story is different in settings (ii) and (iii). Precisely, we show that in the adaptive setting~(ii), the optimal efficiency of a one-bit scheme is (asypmtotically) precisely that of the sample median, and that this efficiency is achievable. As a concrete example, when $X_i$ are i.i.d.\ Gaussian, we necessarily lose a factor of $\pi/2 \approx 1.57$ in the asymptotic risk; the one-bit constraint decreases the effective sample size by a factor of $\pi/2$ compared to estimating it without the bit constraint. It turns out that, in the settings we consider, only a \emph{single round} of adaptivity (see Fig.~\ref{fig:one_round} for an illustration) is sufficient to achieve optimal convergence rates. In distinction from setting (ii), in setting (iii) when the messages must be independent, there is no distributed estimation scheme that achieves the efficiency of the sample median uniformly over $\theta$. We establish this result via Le Cam's local asyptotic normality theory, allowing us to provide exact characterizations of the asymptotic efficiency of suitably regular encoding schemes. Our asymptotic setting is important in that it allows us to elide difficulties present in finite sample settings. For example, in setting~(i), developing an optimal quantizer at finite $n$ requires choosing a $2^n$ level scalar quantizer, which is non-trivial~\cite{gray1998quantization}. In interactive and sequential settings (e.g.~(ii)), the situation is more challenging, as it is unclear whether any type of compositionality applies, in that an $n-1$-step optimal estimator may be only vaguely related to the $n$-step optimal estimator. Thus, to provide our lower bounds, we rely on stronger information-based inequalities, including the Van Trees inequality~\cite{Tsybakov09} and Le Cam's local asymptotic normality theory~\cite{LeCam86,LeCamYa00,VanDerVaart98}. \subsection*{Related Work} The many challenges of estimation under communication constraints have given rise to a large literature investigating different aspects of constrained estimation. While our setting---in which we observe a single bit per signal $X_i$---is restrictive, it inspires substantial work. Perhaps the most related is that of Wong and Gray~\cite{53738}, who study one-bit analog-to-digital conversion of a constant input corrupted by Gaussian noise using a Sigma-Delta Modulator (SDM). They show almost sure convergence, but provide no rate (and no rates follow from their analysis); in contrast, we provide an optimal procedure and matching lower bound achieving risk $\frac{\pi}{2} \sigma^2$ in the limit~\eqref{eq:ARE_def} when $X_i \simiid \normal(\theta, \sigma^2)$. A growing literature on one-bit measurements in high-dimensional problems \cite{baraniuk2017exponential, DavenportPlVaWo15, PlanVe13} shows how to reconstruct sparse signals, where Baraniuk et al.~\cite{baraniuk2017exponential} show that in noiseless settings, exponential decay in MSE is possible; our results make precise the penalty for noise under one-bit sensing, showing that the error can decay (under Gaussian noise) at best as $\frac{\pi}{2} \frac{\sigma^2}{n}$. In fully distributed settings (iii), the challenges are different, and there is also a substantial literature with one-bit (quantized) measurements~\cite{904560,4244748, 6882252, chen2010performance, 5184907}. We complement these results by providing precise lower bounds and optimality results; previous performance bounds are suboptimal. Work on the remote multiterminal source coding problem, or CEO problem~\cite{berger1996ceo, viswanathan1997quadratic, oohama1998rate, prabhakaran2004rate}, provides lower bounds on the MSE in setting~(iii); because of the somewhat distinct setting, these bounds are looser than ours (which have optimal constants). In settings more similar to our statistical estimation scenario---such as estimation of parameters in a multi-dimensional linear model---a line of work provides lower bounds on statistical estimation~\cite{zhang2013information, duchi2014optimality, GargMaNg14, BravermanGaMaNgWo16, DBLP:journals/corr/abs-1802-08417, zhang1988estimation, han2018distributed, xu2017information}. These results are finite sample and apply more broadly than ours, but as a consequence, they have unusable constants, while our stylized model allows precise identification of exact constants. Work subsequent to the initial draft of this paper~\cite{Barnes2018} uses an approach similar to ours---bounding quantized Fisher information---to derive lower bounds on the error in parametric estimation problems from quantized measurements in non-adaptive settings. Testing (and discrete estimation) problems also enjoy a robust literature, though as a consequence of our results to come, the results for testing, i.e., when the parameter space $\Theta$ is finite, are quite different from those for estimation, as it is possible to construct optimal decision (testing) rules in a completely distributed fashion. In this context, Longo et al.~\cite{52470} propose procedures for distributed testing based on optimizing a Bhattacharyya distance. Tsitsiklis~\cite{tsitsiklis1988decentralized} shows that when the cardinality of $\Theta$ is at most $M$ and the probability of error criterion is used, then no more than $M(M-1)/2$ different detection rules are necessary in order to attain probability of error with optimal exponent. Moreover, in a distributed setting, feedback is unnecessary for optimal testing/detection~\cite{5751320}, in strong distinction to the estimation case we consider. The remainder of this paper is organized as follows. In Section~\ref{sec:problem} we describe the problem, notation, and our basic assumptions. In Section~\ref{sec:preliminary} we provide two simple bounds on the efficiency and MSE. Our main results for the adaptive and distributed cases are given in Sections~\ref{sec:sequential} and \ref{sec:distributed}, respectively. In Section~\ref{sec:conclusions} we provide concluding remarks. \section*{Appendices} \input{uniform-are-weirdos} \input{proof-adaptive-optimality} \input{proof-sgd} \input{proof-lan-distributed} \input{proof-no-distributed} \bibliographystyle{IEEEtran} \section{Consistent Estimation and Off-the-shelf Bounds \label{sec:preliminary}} We begin our technical treatment by deriving a few bounds on the efficiency of estimators in setting~\eqref{item:distributed}. These bounds establish the following facts: \begin{enumerate}[1.] \item A consistent estimator with an asymptotically normal distribution always exists in setting~\eqref{item:distributed}, and hence in the adaptive settings~\eqref{item:adaptive} and (\ref{item:one-step-adaptive}'). \item For the normal distribution, the asymptotic relative efficiency~\eqref{eqn:are-def} in the distributed setting~\eqref{item:distributed} is at most $3/4$. No estimator can be as efficient as the sample mean. \end{enumerate} \subsection{Consistent Estimation} The simplest estimator is simply to invert a quantile. Indeed, fix $\theta_0 \in \mathbb R$ and define the $i$th message by \[ B_i = \mathbf 1_{X_i<\theta_0}, \] where $\mathbf 1_A$ is the indicator of the event $A$. We have \[ P_n \triangleq\frac{1}{n} \sum_{i=1}^n B_i \overset{a.s.}{\rightarrow} F(\theta_0 - \theta), \] so that \begin{equation} \label{eq:estimator_naive} {\theta}_n = \theta_0 - F^{-1}\left( P_n \right) \end{equation} is a consistent estimator for $\theta$ in the distributed setting of Figure~\ref{fig:setup}-(iii), where we note that $F$ is invertible over the support of $f$. As the variance of $P_n$ is $F(\theta_0-\theta)\left(1-F(\theta_0-\theta)\right)$, a delta method calculation~\cite[Ch.~23]{VanDerVaart98} implies that ${\theta}_n$ is asymptotically normal with variance \begin{equation*} \frac{F(\theta_0-\theta)\left(1-F(\theta_0-\theta)\right)}{f^2(\theta_0-\theta)} = \frac{1}{ \eta(\theta_0-\theta)}. \end{equation*} In the Gaussian case where the $X_i \simiid \normal(\theta, \sigma^2)$, the ARE of ${\theta}_n$ is $\eta(\theta_0 - \theta)\sigma^2$. Assumption~\ref{assump:failure_rate} implies that the optimal asymptotic variance for an estimator of the form~\eqref{eq:estimator_naive} is $1 / \eta(0)$, the asymptotic of the sample median. Unfortunately, as $\theta$ is (by definition) \emph{a priori} unknown and $\eta(x)$ monotonically decreases in $|x|$, this naive estimator $\theta_n$ may be very inefficient when $\theta$ is far from the initial guess $\theta_0$. As an example, when $f$ is a the normal density, the ARE of ${\theta}_n$ is less than $0.15$ when $|\theta_0 - \theta| \ge 2\sigma$, and more broadly, $\ARE(\theta_n)$ asymptotes to $|\theta_0| \exp(-\theta_0^2 / 2) / \sqrt{2\pi}$ as $|\theta_0 - \theta|$ gets large. Yet that $\theta_0 = \theta$ minimizes this asymptotic variance, and $\eta$ is continuous, is suggestive: if we can use a suitably good initial estimate $\theta_n\init$ for $\theta$, it is possible that a one-step adaptive estimator (recall~(\ref{item:one-step-adaptive}')) may be asymptotically strong, as we see in Section~\ref{sec:sequential}. \subsection{Multiterminal Source Coding} \label{sec:ceo} A related problem is the CEO problem, which considers the estimation of a sequence $\theta_1,\theta_2\ldots$, where a noisy version of each $\theta_j$ is available at $n$ terminals. At each terminal $i$, an encoder observes the $k$ noisy samples \[ X_{i,j} = \theta_j + Z_{i,j},\qquad j=1,\ldots,k, \qquad i = 1,\ldots,n, \] and transmits $r_i k$ bits to a central estimator~\cite{berger1996ceo}. The central estimator produces estimates ${\hat{\theta}}_1,\ldots,{\hat{\theta}}_k$ with the goal of minimizing the quadratic risk: \[ R_{\CEO} = \frac{1}{k} \sum_{j=1}^k \mathbb E \left[\left(\hat{\theta}_j - {\theta_j} \right)^2 \right]. \] Note that any distributed encoding scheme using one-bit per sample can be replicated $k$ times and thus leads to a legitimate encoding and estimation scheme for the CEO problem with $r_1=\ldots=r_n = 1$. It follows that, assuming that $\theta$ is drawn once from the prior $\pi$, our mean estimation problem from one-bit samples under distributed encoding corresponds to the CEO setting with $k=1$ realization of $\theta$ observed under noise at $n$ different locations, and communicated at each location using an encoder sending a single bit. Consequently, a lower bound on the MSE in estimating $\theta$ in the distributed encoding setting is given by the minimal MSE in the CEO setting as $k \to \infty$. Note that the difference between the CEO setting and ours lays in the privilege of each of the encoders to describe $k$ realizations of $\theta$ using $k$ bits with MSE averaged over these realizations, rather than a single realization using a single bit in ours. When the prior on $\theta$ and the noise corrupting it at each location are Gaussian, Prabhakaran et al.~\cite{prabhakaran2004rate} characterize the optimal encoding and its asymptotic risk as $k \to \infty$. Chen et al.~\cite{chen2004upper} also provide an expression for the quadratic risk in the CEO setting under Gaussian priors. Adapting to our setting, this expression provides the following proposition: \begin{prop} \label{prop:CEO} Assume that $\Theta = \mathbb R$ and $\pi(\theta) = \Ncal(0,\sigma_\theta^2)$ where $\sigma_\theta^2 \in \mathbb R$ is arbitrary. Then any estimator ${\theta}_n$ of $\theta$ in the distributed setting satisfies \begin{equation} \label{eq:ceo_bound} n \ex{\left( \theta - \theta_n \right)^2} \geq \frac{4}{3} \sigma^2 + O(n^{-1}), \end{equation} where the expectation is with respect to $\theta$ and $X_1,\ldots,X_n$. \end{prop} \noindent See Appendix~\ref{app:proof:CEO} for a proof. As we shall see, this bound is loose: the difference between the MSE lower bound~\eqref{eq:ceo_bound} and the actual MSE in the distributed setting (case~\eqref{item:distributed}) occurs because in the CEO setting, each encoder may encode an arbitrary number of $k$ independent realizations of $\theta$ using $k$ bits; in our situation, $k = 1$. That blocking allows more efficient encoding and exploiting the high-dimensional geometry of the product probability space in the CEO problem is perhaps unsurprising, and our goal in the sequel will be to characterize the performance degradation one bit encoding engenders. \section{Problem Formulation and Notation} \label{sec:problem} Let $f : \R \to \R_+$ be a symmetric and log-concave probability density, which necessarily has finite second moment $\sigma^2$, and let $\Theta \subset \R$ be closed and convex. For $\theta \in \R$, let $P_\theta$ be the probability distribution with density $f(x-\theta)$, so that $\theta$ indexes the location family $\{P_\theta\}_{\theta \in \Theta}$. The log-concavity and symmetry $f(x)$ imply that $P_\theta$ has a unique mean and median at $\theta$~\cite{ibragimov1956composition}. We observe a sample $X_1, \ldots, X_n \simiid P_\theta$, where $\theta$ is unknown, and wish to estimate $\theta$ given only binary messages $B_1, \ldots, B_n \in \{0, 1\}$ about each $X_i$. We study this under three distinct computational scenarios, which we illustrate in Figure~\ref{fig:setup}: \begin{enumerate}[(i)] \item \label{item:centralized} Centralized, where $B_i = B_i(X_1,\ldots,X_n)$, $i=1,\ldots,n$. \item \label{item:adaptive} Adaptive, where $B_i = B_i(X_i,B_1,\ldots,B_{i-1})$, $i=2,\ldots,n$. \item \label{item:distributed} Distributed, where $B_i = B_i(X_i)$, $i=1,\ldots,n$. \end{enumerate} \noindent We also consider a hybrid of the fully distributed setting (where the bits $B_i$ are independent) and the adaptive setting (where each bit $B_i$ may depend on the previous bits) to a \emph{one-step} adaptive setting, where the quantization scheme may be modified to depend on one fixed function of the previous information. \begin{enumerate}[(i')] \setcounter{enumi}{1} \item \label{item:one-step-adaptive} One-step adaptive, where for some function $g$ and a (fixed) $t$, if $i \le t$ then $B_i = B_i(X_i)$ while if $i > t$, then $B_i = B_i(X_i, g(B_1, \ldots, B_t))$. \end{enumerate} We measure the performance of an estimator $\theta_n \defeq \theta_n(B_1^n)$ by one of a few notions. In the simplest case, we assume a prior $\pi$ on $\theta$ (which may be a point mass) and consider the quadratic risk \begin{equation} \label{eq:error_def} R_n = R_n(\pi) \defeq \int \E_\theta\left({\theta}_n - \theta \right)^2 d\pi(\theta), \end{equation} where the expectation is taken with respect to the distribution of $X_1,\ldots,X_n \simiid P_\theta$. The main problems we consider in this paper are the minimal value of the risk~\eqref{eq:error_def} as a function of the sample size $n$ and the density $f$, under different choices of the encoding functions in cases~\eqref{item:centralized}--\eqref{item:distributed}. The quadratic risk~\eqref{eq:error_def} may be infinite in some cases; we defer discussion of this case to later sections, as it is technically demanding and detracts from the presentation here. Now, let $\sigma_f^2 \defeq \E[\frac{f'(X)^2}{f(X)^2}]$ be the Fisher information for the location in the family $\{P_\theta\}$, which is finite when $f$ is log-concave and symmetric. We give particular attention to the asymptotic relative efficiency (ARE) of estimators with respect to asymptotically normal efficient estimators achieving the information bound~\cite{VanDerVaart98}. In this case, if $\{m(n), n \in \N\}$ is a sequence such that \begin{equation*} \sqrt{m(n)} (\theta_n - \theta) \cd \normal(0, \sigma_f^2), \end{equation*} then the ARE of the estimator is~\cite[Def.~6.6.6]{LehmannCa98} \begin{equation} \label{eqn:are-def} \ARE({\theta}_n) \defeq \liminf_{n\rightarrow \infty} \frac{m(n)}{n}. \end{equation} In the special case where there exists $V \in \mathbb R$ such that \begin{equation*} m(n) R_n = m(n) \mathbb E_\theta\left({\theta}_n - \theta \right)^2 = V + o(1), \end{equation*} the ARE of ${\theta}_n$ is $\sigma_f^2/V$, so that $\theta_n$ requires a sample $V / \sigma_f^2$-times larger than that of an efficient estimator for comparable accuracy to the (information) efficient estimator. \subsection*{Notation and basic assumptions} To describe our results and make them formal, we require some additional notation and one main assumption, which restricts the class of distributions we consider. We use the typical notation that $F(x) = \int_{-\infty}^x f(t) dt$ is the cumulative distribution function of the $X_i$, and we let \begin{equation*} h(x) \triangleq \frac{f(x)}{1-F(x)} = \frac{f(x)}{F(-x)} \end{equation*} be the \emph{hazard} function (or the \emph{failure rate} or \emph{force of mortality}), which is monotone increasing as $f$ is log-concave~\cite{bagnoli2005log}. Given the centrality of the median to our efficiency bounds, it is unsurprising that the quantity \begin{equation} \label{eq:eta_def} \eta(x) \triangleq \frac{f^2(x)}{F(x)(1-F(x))} \stackrel{(\star)}{=} \frac{f(x)f(-x)}{F(x)F(-x)} \end{equation} appears throughout our development. (Here equality~$(\star)$ is immediate by the symmetry of $f$.) For $p \in (0, 1)$ and $x = F^{-1}(p)$, \begin{equation} \label{eqn:variance-quantiles} \frac{1}{\eta(x)} = \frac{1}{\eta(F^{-1}(p))} = \frac{p (1 - p)}{f(F^{-1}(p))^2} \end{equation} is of course the familiar asymptotic variance of the $p$th quantile of the sample $X_1,\ldots,X^n$ (cf.~\cite{VanDerVaart98}, Ch.~21). For $f$ the normal density, classical results~\cite{Samford1953, hammersley1950estimating} show that $\eta(x)$ is a strictly decreasing function of $|x|$, as we illustrate in Fig.~\ref{fig:eta}. We consider log-concave symmetric distributions sharing this property. Specifically, we require the following. \begin{assumption} \label{assump:failure_rate} The density $f$ is log-concave and symmetric. Additionally, the origin $x = 0$ uniquely maximizes $\eta(x)$, and $\eta(x)$ is non-increasing in $|x|$. \end{assumption} Under this assumption, \begin{equation*} 4 f^2(x) \leq \eta(x) \leq \eta(0), \end{equation*} where $\eta(0) = 4 f^2(0)$ is the asymptotic variance of the sample median (Eq.~\eqref{eqn:variance-quantiles} at $p = 1/2$). Combined with log-concavity of $f(x)$, Assumption~\ref{assump:failure_rate} implies that $\eta(x)$ vanishes as $|x|\rightarrow \infty$. Several distributions satisfy Assumption~\ref{assump:failure_rate}, including the generalized normal distributions with a shape parameter between $1$ and $2$ (including the normal and Laplace distributions). Symmetric log-concave distributions failing Assumption~\ref{assump:failure_rate} include the uniform distribution and the generalized normal distribution with shape parameter greater than $2$. Some restriction on the class of distributions is necessary to develop our results; indeed, in Appendix~\ref{sec:uniform-weirdos} we provide a brief discussion on the uniform distribution, where a one-step adaptive estimator with single bit observations can achieve convergence rates faster than the familiar $\sqrt{n}$ paramateric rate. \begin{figure} \begin{center} \begin{tikzpicture}[scale = 0.6] \begin{axis}[ width=8cm, height=6cm, xmin = -3, xmax=3, restrict y to domain = 0:100, ymin = 0, ymax = 0.9, samples=10, xlabel= $x$, xtick={-2,-1,0,1,2}, xticklabels={-2,-1,0,1,2}, ytick={0,0.3989423,0.6366198}, yticklabels={0,$\frac{1}{\sqrt{2 \pi}}$,$2/\pi$}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, solid, smooth] plot table [x = x, y = y, col sep=comma] {./Figs/eta.csv}; \addlegendentry{$\eta(x)$}; \addplot[domain = -5:5, samples = 50, color = red, solid, smooth] {exp(-x^2/2) / sqrt(2*3.14159)}; \addlegendentry{$\phi(x)$}; \addplot[domain = -5:5, samples = 50, color = black, solid, dashed] {4*exp(-x^2) / (2*3.14159)}; \addlegendentry{$4\phi^2(x)$}; \end{axis} \end{tikzpicture} \begin{tikzpicture}[scale = 0.6] \begin{axis}[ width=8cm, height=6cm, xmin = -4, xmax=4, restrict y to domain = -10:0, ymin = -10, samples=10, xlabel= $x$, xtick={-3,-2,-1,0,1,2,3}, xticklabels={-3,-2,-1,0,1,2,3}, ytick={0,-0.45158,-0.919}, yticklabels={0,,}, line width=1.0pt, mark size=1.5pt, ymajorgrids, xmajorgrids, legend style= {at={(1,1)},anchor=north east,draw=black,fill=white,align=left} ] \addplot[color = blue, solid, smooth] plot table [x = x, y = logy, col sep=comma] {./Figs/eta.csv}; \addlegendentry{$\log \eta(x)$}; \addplot[domain = -5:5, samples = 30, color = red, solid, smooth] {-(x)^2/2 -0.9189}; \addlegendentry{$\log \phi(x)$}; \end{axis} \end{tikzpicture} \caption{ The function $\eta(x) = f^2(x) / F(x)F(-x)$ for $f(x) = \phi(x)$ the standard normal density. \label{fig:eta} } \end{center} \end{figure} \section{Proof of Theorem~\ref{thm:adpative_lower_bound}} \label{proof:thm:adpative_lower_bound} We begin with two technical lemmas. \begin{lem} \label{lem:bound_intervals} Let $f$ be a log-concave and symmetric density function for which Assumption~\ref{assump:failure_rate} holds. For any $x_1 \geq \ldots \geq x_n \in \R$, \begin{equation} \frac{ \left| \sum_{k=1}^n (-1)^{k+1} f(x_k) \right|^2 }{ \left( \sum_{k=1}^n (-1)^{k+1} F(x_k) \right) \left(1- \sum_{k=1}^n (-1)^{k+1} F(x_k) \right) } \leq 4f(0)^2. \label{eq:lem_bound_intervals} \end{equation} \end{lem} \begin{lem} \label{lem:fisher_bound} Let $X$ be a random variable with a symmetric, log-concave, and continuously differentiable density function $f(x)$ such that Assumption~1 holds. For a Borel measurable set $A$, define \begin{equation*} B(x) \defeq \begin{cases} 1 & \mbox{if} ~x \in A, \\ -1 & \mbox{if}~ x \notin A. \end{cases} \end{equation*} Then the Fisher information of $B$ with respect to $\theta$ is bounded from above by $\eta(0)$. \end{lem} Lemma~\ref{lem:bound_intervals} is the special case $\delta = 0$ of Lemma~\ref{lem:bound_intervals_delta} to come in Section~\ref{sec:bound_intervals_delta}. We now prove Lemma~\ref{lem:fisher_bound}. \begin{proof-of-lemma}[\ref{lem:fisher_bound}] When $f$ is the normal density function, this lemma follows from \cite[Thm.~3]{Barnes2018}. The proof below is based on a different techique than in \cite{Barnes2018}, and is valid for any log-concave symmetric density satisfying Assumption~\ref{assump:failure_rate}. \\ The Fisher information of $B$ with respect to $\theta$ is given by \begin{align} I_\theta & = \ex{ \left( \frac{d}{d\theta} \log P\left( B | \theta \right) \right)^2 |\theta } \nonumber \\ & = \frac{ \left(\frac{d}{d\theta} P(B=1|\theta) \right)^2}{P(B=1| \theta)} + \frac{ \left(\frac{d}{d\theta} P(B=-1|\theta) \right)^2} {P(B=-1| \theta)} \nonumber \\ & = \frac{ \left( \frac{d}{d\theta} \int_A f \left( x-\theta\right)dx \right)^2} { P(B=1| \theta) } + \frac{ \left( \frac{d}{d\theta}\int_A f \left( x-\theta \right)dx \right)^2} { P(B=-1| \theta) } \nonumber \\ & \overset{(a)}{=} \frac{ \left( - \int_A f' \left( x-\theta \right)dx \right)^2} { P(B=1| \theta) } + \frac{ \left(- \int_A f' \left( x-\theta \right)dx \right)^2} { P(B=-1| \theta) } \nonumber \\ & = \frac{\left( \int_A f'\left( x-\theta \right) dx \right)^2 }{ P(B=1 | \theta) \left(1-P(B=1|\theta) \right) }, \nonumber \\ & = \frac{\left( \int_A f'\left( x-\theta \right) dx \right) \left( \int_A f'\left( x-\theta \right) dx \right)}{ \left( \int_A f \left( x-\theta \right) dx \right) \left(1- \int_A f \left( x-\theta \right) dx \right) }, \label{eq:lem_fisher_bound_proof1} \end{align} where differentiation under the integral sign in $(a)$ is possible since $f$ is log-concave and so a.e.\ differentiable (cf.~\cite{Bertsekas73}), with a.e.\ derivative $f'(x)$. Regularity of the Lebesgue measure implies that for any $\epsilon>0$, there exists a finite number $k$ of disjoint open intervals $I_1,\ldots I_k$ such that \begin{equation*} \int_{A\setminus \cup_{j=1}^k I_j } dx < \epsilon, \end{equation*} which implies that for any $\epsilon' > 0$, the set $A$ in \eqref{eq:lem_fisher_bound_proof1} can be replaced by a finite union of disjoint intervals without increasing $I_\theta$ by more than $\epsilon'$. It is therefore enough to proceed in the proof assuming that $A$ is of the form \begin{equation*} A = \cup_{j=1}^k (a_j,b_j), \end{equation*} with $\infty \leq a_1 \leq \ldots a_k$, $b_1 \leq b_k \leq \infty$ and $a_j \leq b_j$ for $j=1,\ldots,k$. Under this assumption we have \begin{align*} \mathbb P(B_n=1| \theta) & = \sum_{j=1}^k \mathbb P\left(X_n \in (a_j,b_j) \right) \\ & = \sum_{j=1}^k \left( F \left(b_j-\theta\right) - F \left(a_j-\theta\right) \right), \end{align*} so we may rewrite Eq.~\eqref{eq:lem_fisher_bound_proof1} as \begin{align*} I_\theta & = \frac { \left( \sum_{j=1}^{k} f \left(a_j-\theta \right) - f \left( b_j-\theta \right) \right)^2 } { \left( \sum_{j=1}^k F \left( b_j-\theta \right) - F \left( a_j-\theta \right) \right) } \nonumber \\ & \times \frac {1} {1- \left( \sum_{j=1}^k F \left( b_j-\theta \right) - F \left( a_j-\theta \right) \right) } \end{align*} It follows from Lemma~\ref{lem:bound_intervals} that for any $\theta \in \R$ and any choice of the intervals' endpoints, \begin{equation*} I_\theta \le \max_{t \in \{a_j,b_j, j=1,\ldots,k\} } 4f(t)^2 \leq 4 f(0)^2, \end{equation*} where the last transition is by Assumption~\ref{assump:failure_rate}. \end{proof-of-lemma} We now finish the proof of Theorem~\ref{thm:adpative_lower_bound}. To bound the Fisher information of any set of $n$ one-bit messages with respect to $\theta$, we consider the conditional distribution $P({B_1,\ldots,B_n|\theta})$ of $(B_1,\ldots,B_n)$ given $\theta$. We have \begin{align*} P\left( B_1,\ldots,B_n | \theta \right) & = \prod_{i=1}^n P\left(B_i | \theta, B_1,\ldots,B_{i-1} \right), \end{align*} where $P\left(B_i =1 | \theta, B_1,\ldots,B_{i-1} \right) = \mathbb P\left( X_i \in A_i\right)$, when $B_i = \indic{X_i \in A_i}$ by independence, so that the Fisher information of $B_1,\ldots,B_n$ with respect to $\theta$ is \begin{align} I_\theta(B_1,\ldots,B_N) = \sum_{i=1}^n I_\theta (B_i|B_1,\ldots,B_{i-1}), \label{eq:fisher_information} \end{align} where $I_\theta (B_i|B_{i-1},\ldots,B_1)$ is the Fisher information of the distribution of $B_i$ given $B_1,\ldots,B_{i-1}$. Lemma~\ref{lem:fisher_bound} implies that $I_\theta (B_i|B_{i-1},\ldots,B_1) \leq 4f(0)^2$. The Van Trees inequality \cite{van2004detection, gill1995applications} now implies \begin{align*} \ex{ \left( \theta_n - \theta \right)^2} & \geq \frac{1}{ \ex{ I_\theta(B_1,\ldots,B_n)} + I_0} \\ & = \frac{1}{ \sum_{i=1}^n I_\theta (B_i | B^{i-1} ) + I_0} \\ & \geq \frac{1}{ 4f^2(0) n + I_0} \end{align*} as desired. \subsection{An Isoperimetric Lemma} \label{sec:bound_intervals_delta} The following lemma is essential to the proofs of Theorems~\ref{thm:adpative_lower_bound} and \ref{thm:non_existence}. \begin{lem} \label{lem:bound_intervals_delta} Let $f$ be a log-concave and symmetric density function. Let $\delta\geq 0$. Assume that the function \begin{equation*} \eta_\delta(x) \triangleq \eta^{1+\delta}(x)/f^\delta(x) = \frac{ \left( f(x) \right)^{2+\delta}}{\left(F(x)(1-F(x))\right)^{1+\delta}} \end{equation*} is non-increasing in $|x|$. Then for any $x_1 \ge \ldots \ge x_n \in \R^n$, \begin{equation} \frac{ \left| \sum_{i=1}^n (-1)^{i+1} f(x_i) \right|^{2+\delta} } {\left| \sum_{i=1}^n (-1)^{i+1} F(x_i) \right|^{1+\delta} \left|1- \sum_{k=1}^n (-1)^{i+1} F(x_i) \right|^{1+\delta} } \leq \max_{i} \eta_\delta(x_i). \label{eq:lem_bound_intervals_delta} \end{equation} \end{lem} In particular, \begin{equation*} \frac{ \left| \sum_{i=1}^n (-1)^{i+1} f(x_i) \right|^{2+\delta} } {\left| \sum_{i=1}^n (-1)^{i+1} F(x_i) \right|^{1+\delta} \left|1- \sum_{i=1}^n (-1)^{i+1} F(x_i) \right|^{1+\delta} } \leq \eta_\delta(0) = 4^{1+\delta} f^{2+\delta}(0). \end{equation*} \begin{proof-of-lemma}[\ref{lem:bound_intervals_delta}] Denote \begin{equation*} \delta_n(x_1,\ldots,x_n) \triangleq \sum_{i=1}^{n} s_i f(x_i), \end{equation*} \begin{equation*} \Delta_n(x_1,\ldots,x_n) \triangleq \sum_{i=1}^n s_i F(x_i), \end{equation*} where $s_i \triangleq (-1)^{i+1}$. We use induction on $n \in \N$ to show that \begin{align} \label{eq:lemm:interval_bounds:to_show} \frac{ \left| \delta_n(x_1,\ldots,x_n) \right|^{2+\delta}} {\left|\Delta_n(x_1,\ldots,x_n)\left(1- \Delta_n(x_1,\ldots,x_n) \right) \right|^{1+\delta} } \leq \max_{i}\eta_{\delta}(x_i). \end{align} Since \begin{equation*} \eta_\delta(x) = \frac{ \left|\delta_1(x) \right|^{2+\delta}}{\left|\Delta_1(x) (1-\Delta_1(x)) \right|^{1+\delta}}, \end{equation*} The case $n=1$ is trivial. Assume that \eqref{eq:lemm:interval_bounds:to_show} holds for all integers up to $n = N$ and for any $x_1 \ge \ldots \ge x_N$. Consider the case $n = N+1$. Let $i^*$ be the index such that $x_{i^*}$ has minimal absolute value among $x_1,\ldots,x_N$. The assumption on $\eta_\delta(x)$ implies that \begin{equation*} \eta_\delta(x_{i^*}) = \max_i \eta_\delta(x_i). \end{equation*} Since the LHS of \eqref{eq:lem_bound_intervals_delta} is invariant to a sign flip of all $x_1,\ldots,x_{N+1}$, we may assume that $x_{i^*}$ is positive without loss of generality. Set $x^* = x_{i^*}$ and let $k=i^*-1$. In what follows, variables with subscript of non-positive index are ignored in summations and in lists of arguments to functions. Consider the function \begin{align} & g(y_1,\ldots,y_N) \triangleq g(y_1,\ldots,y_N|x^*,k) \\ & \triangleq \frac{\left| \delta_{N+1}(y_1,\ldots,y_k,x^*,y_{k+1}\ldots,y_N) )\right|^{2+\delta}}{ \left|\Delta_{N+1}(y_1,\ldots,y_k,x_{i^*},y_{k+1}\ldots,y_N) (1 -\Delta_{N+1}(y_1,\ldots,y_k,x^*,y_{k+1}\ldots,y_N) \right|^{1+\delta} } \label{eq:g_def}. \end{align} The LHS of \eqref{eq:lemm:interval_bounds:to_show} is obtained by taking $y_i=x_{k_i}$ where $k_i$ is the $i$th element in $\{1,\ldots,N+1\}\setminus \{i^*\}$. It is therefore enough to prove that \begin{equation*} \max_{(y_1,\ldots,y_N) \in A_N} g(y_1,\ldots,y_N) \leq \eta_{\delta}(x^*), \end{equation*} where \begin{equation*} A_N(x^*,k) \triangleq \left\{ (y_1,\ldots,y_N) \in \R^N\, : \, y_1 \ge y_k \ge x^* \ge -x^* \ge y_{k+1} \ldots \ge y_N \right\}. \end{equation*} % Since $f(x)$ is log-concave and symmetric, we may write $f(x) = e^{c(x)}$ where $c(x)$ is concave, symmetric, and superdifferentiable on the interior of its domain with supergradient set $\partial c(x) = \{v \in \R \mid c(y) \le c(x) + v (y - x) ~ \mbox{for~all}~ y\}$; $c$ is also differentiable a.e.\ with derivative \begin{equation*} c'(x) \triangleq \frac{f'(x)}{f(x)} \end{equation*} (when it exists), and we otherwise simply treat $f'(x) / f(x) = c'(x) \in \partial c(x)$ as an arbitrary element of the superdifferential. The supergradient sets $\partial c(x)$ are increasing, in that $v_0 \in \partial c(x_0)$ and $v_1 \in \partial c(x_1)$ implies that $(v_1 - v_0) (x_1 - x_0) \le 0$. We first prove the lemma under the assumption that $c$ is strictly concave, or, equivalently, that $v_i \in \partial c(x_i)$ implies that $(v_1 - v_0)(x_1 - x_0) < 0$ whenever $x_1 \neq x_0$; that is, $c'$ is strictly decreasing. The maximal value of $g(y_1,\ldots,y_N)$ is attained for the same $(y_1,\ldots,y_N) \in A_N(x^*,k)$ that maximizes \begin{align*} \log(g)(y_1,\ldots, y_N) = (2+\delta) \log \left( \delta_N \right) - (1+\delta) \log \left( \Delta_N \right) - (1+\delta) \log \left(1 - \Delta_N \right), \end{align*} where in the last display and henceforth we suppress the arguments $y_1,\ldots,y_k,x^*,y_{k+1},\ldots, y_N$ of the functions $\delta_N$ and $\Delta_N$. % Within the interior of $A_N(x^*,k)$, all three expressions in \eqref{eq:g_def} within an absolute value are positive. It follows that partial derivative of $\log(g)(y_1,\ldots,y_N)$ with respect to $y_i$ within the interior of $A_N(x^*,k)$ is \begin{equation*} \frac{\partial \log(g)}{\partial y_i} = \frac{(2+\delta) s_i f'(x_i)}{\delta_N } -\frac{(1+\delta) s_i f(x_i)}{\Delta_N } + \frac{(1+\delta)s_i f(y_i)}{1-\Delta_N }. \end{equation*} We conclude that the gradient of $\log(g)$ vanishes if and only if \begin{equation} \label{eq:gradient_zero} c'(y_i) = \frac{f'(y_i)}{f(y_i)} = \frac{1+\delta}{2+\delta} \frac{\delta_N}{2} \left( \frac{1}{\Delta_N} - \frac{1}{1-\Delta_N } \right),\quad i=1,\ldots,N. \end{equation} % Since we assumed that $\partial c(x)$ is injective, equality~\eqref{eq:gradient_zero} holds if and only if $y_1 = \ldots = y_N$. In this case, $g(y_1,\ldots,y_N) = \eta_\delta(x_{i^*})$ if $N$ is even. If $N$ is odd and $y_1 = \ldots = y_N > x^*$, then \begin{align*} & g(y_1,\ldots,y_N) = \frac{\left| f(y_1)-f(x_{i^*}) \right|^{2+\delta}} { \left| F(y_1) - F(x_{i^*}) \right|^{1+ \delta} \left| 1 - (F(y_1) -F(x_{i^*})) \right|^{1+ \delta} } \end{align*} which is bounded from above by $\eta_\delta(x_{i^*})$ by the induction hypothesis. The case where $N$ is odd and $-x^* \leq y_1 = \ldots = y_N$ is similar. % We now consider the possibility that the maximum of $g(y_1,\ldots,y_N)$ is attained at the boundaries of $A_N(x^*,k)$. At boundary points for which $y_i = y_{i+1}$ for some $i$, the contribution of $y_i$ and $y_{i+1}$ to $g(y_1,\ldots,y_N)$ is zero and the induction assumption for $n=N-1$ implies that \begin{equation*} g(y_1,\ldots,y_N) \leq \eta_{\delta}(x^*) \end{equation*} The remaining boundary points of $A_N(x^*,k)$ are covered by the following cases: \begin{itemize} \item[(1)] $y_N \to -\infty$. \item[(2)] $y_1 \to \infty$. \item[(3)] $y_k = x_{i^*}$. \item[(4)] $y_{k+1} = -x_{i^*}$. \end{itemize} For case (1), \begin{align*} g(y_1,\ldots,y_N) \to \frac{ \left| \sum_{i=1}^{k} s_i f(y_i) + s_{i^*} f(x_{i^*}) - \sum_{i=k+1}^{N-1} s_i f(y_i) \right|^{2+\delta}} {\left| \sum_{i=1}^{k} s_i F(y_i) + s_{i^*} F(x_{i^*}) - \sum_{i=k+1}^{N-1} s_i F(y_i) \right|^{1+\delta}\left|1- \left( \sum_{i=1}^{k} s_i F(y_i) + s_{i^*} F(x_{i^*}) - \sum_{i=k+1}^{N-1} s_i F(y_i) \right) \right|^{1+\delta} }, \end{align*} which is smaller than $\eta_\delta(x_{i^*})$ by the induction hypothesis. Similarly, under case (2), \begin{align*} & g(y_1,\ldots,y_N) \to \frac{ \left| \sum_{i=2}^{k} s_i f(y_i) + s_{i^*} f(x_{i^*}) - \sum_{i=k+1}^{N} s_i f(y_i) \right|^{2+\delta}} {\left| 1+ \sum_{i=2}^{k} s_i F(y_i) + s_{i^*} F(x_{i^*}) - \sum_{i=k+1}^{N} s_i F(y_i) \right|^{1+\delta}\left|-\left( \sum_{i=2}^{k} s_i F(y_i) + s_{i^*} F(x_{i^*}) - \sum_{i=k+1}^{N} s_i F(y_i) \right) \right|^{1+\delta} }, \\ & = \frac{ \left| -\sum_{i=2}^{k} s_i f(y_i) - s_{i^*} f(x_{i^*}) + \sum_{i=k+1}^{N} s_i f(y_i) \right|^{2+\delta}} { \left|1 - \left(-\sum_{i=2}^{k} s_i F(y_i) - s_{i^*} F(x_{i^*}) + \sum_{i=k+1}^{N} s_i F(y_i) \right) \right|^{1+\delta} \left|-\sum_{i=2}^{k} s_i F(y_i) - s_{i^*} F(x_{i^*}) + \sum_{i=k+1}^{N} s_i F(y_i) \right|^{1+\delta} } \end{align*} which is smaller than $\eta_{\delta}(x_{i^*})$ by the induction hypothesis. Under case (3), the terms in $\delta_N$ and $\Delta_N$ corresponding to $y_k$ and $x_{i^*}$ cancel each other. As a result, $g(y_1,\ldots,y_N)$ reduces to an expression with $n=N-1$ variables hence this case is handled by the induction hypothesis. % Finally, under case (4), set \begin{equation*} d \triangleq s_k F(-x^*) + s_{i^*} F(x^*) = s_{i^*}\left(1-2F(-x^*) \right), \end{equation*} \begin{equation*} \sigma \triangleq \sum_{i=1}^{k-1} s_i f(y_i) - \sum_{i=k+1}^{N} s_i f(y_i). \end{equation*} and \begin{equation*} \Sigma \triangleq \sum_{i=1}^{k-1} s_i F(y_i) - \sum_{i=k+1}^{N} s_i F(y_i). \end{equation*} We have \begin{align*} & g(y_1,\ldots,y_N) =\nonumber \\ & = \frac{ \left| \sum_{i=1}^{k-1} s_i f(y_i) - \sum_{i=k+1}^{N} s_i f(y_i) \right|^{2+\delta}} {\left| \sum_{i=1}^{k-1} s_i F(y_i) + d(x^*) - \sum_{i=k+1}^{N} s_i F(y_i) \right|^{1+\delta} \left|1- \sum_{i=1}^{k-1} s_i F(y_i) - d(x^*) + \sum_{i=k+1}^{N} s_i F(y_i) \right|^{1+\delta} }, \\ & = \frac{ \left| \sigma \right|^{2+\delta}} {\left| \Sigma+d \right|^{1+\delta} \left|1- \Sigma - d \right|^{1+\delta} } = \frac{ \left| \sigma \right|^{2+\delta}} {\left| \Sigma \right|^{1+\delta} \left|1- \Sigma \right|^{1+\delta} } \left| \frac{\Sigma(1-\Sigma) } { \Sigma(1-\Sigma) + d(1-2\Sigma)-d^2} \right|^{1+\delta}. \end{align*} By the induction hypothesis, \begin{equation*} \frac{ \left| \sigma \right|^{2+\delta}} {\left| \Sigma \right|^{1+\delta} \left|1- \Sigma \right|^{1+\delta} } \leq \eta_\delta(x^*), \end{equation*} hence it is left to show that \begin{equation*} \frac{\Sigma(1-\Sigma) } { \Sigma(1-\Sigma) + d(1-2\Sigma)-d^2} \leq 1. \end{equation*} Whenever $d>0$, \begin{align*} & \frac{ \Sigma(1-\Sigma) + d(1-2\Sigma)-d^2} {\Sigma(1-\Sigma)} \geq 1 \Leftrightarrow 1-2\Sigma \geq d, \end{align*} while for $d<0$, \begin{align*} & \frac{ \Sigma(1-\Sigma) + d(1-2\Sigma)-d^2} {\Sigma(1-\Sigma)} \geq 1 \Leftrightarrow 1-2\Sigma \leq d. \end{align*} Therefore, it is enough to show that $\Sigma \leq F(-x^*)$ if $s_{i^*}=1$ and $\Sigma \geq F(-x^*)$ if $s_{i^*}=-1$. % Indeed, if $s_{i^*}=1$, then $s_{k+1}=-1$ and monotonicity of $F(x)$ implies that \begin{equation*} \Sigma + d \leq F(y_1) - F(y_k) + F(x^*) - F(-x^*) + F(y_{k+2}) - F(y_N), \end{equation*} and hence \begin{equation*} \Sigma \leq 1-F(x^*) = F(-x^*). \end{equation*} Similarly, if $s_{i^*}=-1$ then \begin{equation*} 1 - \Sigma \leq 1 - F(-x^*). \end{equation*} This conclude the proof in the case where $c'(x)$ is an injection. In the case where $c(x)$ is not strictly concave, so that $c'$ does not strictly decrease, we approximate $c$ using another concave symmetric function with decreasing derivative. We assume w.l.o.g.\ that $c(0) = 0$ maximizes $c$. For $\alpha>0$ consider the function $f_\alpha(x) = \kappa(\alpha) e^{-|c(x)|^{1+\alpha}}$, where $\kappa(\alpha)$ normalizes $f_\alpha$. Then $c_\alpha(x)$ is concave, symmetric, and a.e.\ differentiable with \begin{equation*} c_\alpha'(x) \triangleq \frac{f'_\alpha(x)}{f_\alpha(x)} = (1+\alpha)|c(x)|^{\alpha} c'(x). \end{equation*} Now $c_\alpha'(x)$ is non-increasing since it is the derivative of a concave function. Furthermore, since $c(x)$ is non-constant on any interval and $c'(x)$ is non-increasing, $c_\alpha'(x)$ is non-constant on any interval hence an injection. It follows from the first part of the proof that, for any $\alpha>0$, \begin{align} \label{eq:proof:lem:bound_intervals} \frac{(\delta_{n,\alpha})^{2+\delta}}{\left(\Delta_{n,\alpha}(1-\Delta_{n,\alpha})\right)^{1+\delta}} \leq \max_i \eta_{\delta,\alpha}(x_i), \end{align} where \begin{equation*} \delta_{n,\alpha} \triangleq \sum_{k=1}^{n} (-1)^{k+1} f_{\alpha}(x_k), \end{equation*} \begin{equation*} \Delta_{n,\alpha} \triangleq \sum_{k=1}^n (-1)^{k+1} F_{\alpha}(x_k), \end{equation*} and \begin{equation*} \eta_{\delta,\alpha}(x) \triangleq \frac{(f_\alpha(x))^{2+\delta}}{\left(F_{\alpha}(x)(1-F(x)) \right)^{1+\delta}}. \end{equation*} The proof is completed by noting that \begin{align*} \lim_{\alpha \to 0} \frac{(\delta_{n,\alpha})^{2+\delta} }{ \left(\Delta_{n,\alpha}(1-\Delta_{n,\alpha})\right)^{1+\delta}} = \frac{(\delta_{n})^{2+\delta }}{\left(\Delta_{n}(1-\Delta_{n}) \right)^{1+\delta}}, \end{align*} and, since the maximum is over a finite set, \begin{align*} \lim_{\alpha \to 0} \max_i \eta_{\delta,\alpha}(x_i) = \max_i\eta_\delta(x_i). \end{align*} \end{proof-of-lemma} \section{Proof of Theorem~\ref{theorem:non-adaptive-minimax}} \label{sec:proof-non-adaptive-minimax} We follow a similar outline to the optimality results we establish in the proof of Theorem~\ref{thm:sgd}\eqref{item:sgd-regular} in Sec.~\ref{sec:proof-sgd-regular}. Roughly, we establish that the family $P_\theta$ of distributions on the bits $B_i$ is locally asymptotically normal (Definition~\ref{definition:lan}) via a quadratic mean differentiability argument. After this, the result follows by standard local asymptotic minimax theory. We begin with an argument on the smoothness properties of the densities, which is important for our Taylor expansions to come. \begin{lemma} \label{lemma:derivative-bounds} Let Assumption~\ref{assumption:detection-regions}\eqref{item:lipschitz-density} and \eqref{item:finite-intervals} hold. Then for any $A = \cup_{i = 1}^k \{[a_i, b_i]\}$ and $h \in \R$, \begin{equation*} P_{\theta + h}(A) = P_\theta(A) + \dPtheta(A) h \pm k \cdot \lip(f) h^2, \end{equation*} where \begin{equation} \dPtheta(A) = \sum_{i = 1}^k f(a_i - \theta) - f(b_i - \theta). \label{eqn:expansion-dPtheta} \end{equation} Additionally, we have the bounds \begin{equation} \label{eqn:bound-density-diffs} |f(b) - f(a)| \le 2 \sqrt{\lip(f) P([a, b])} ~~ \mbox{and} ~~ |\dPtheta(A)| \le 2 \sqrt{k \lip(f)}. \end{equation} \end{lemma} \noindent See Section~\ref{sec:proof-derivative-bounds} for a proof. The second lemma provides the local asymptotic normality we require. \begin{lemma} \label{lemma:lan-bits} Let Assumption~\ref{assumption:detection-regions}\eqref{item:lipschitz-density} and~\eqref{item:finite-intervals} hold, and let $B_i = \indic{X_i \in A_i}$. Let $h_n \to h \in \R$. Then for any $\theta \in \mbox{int}\Theta$, \begin{equation*} \sum_{i = 1}^n \log \frac{P_{\theta + h_n/\sqrt{n}}(B_i)}{ P_\theta(B_i)} = \frac{h}{\sqrt{n}} \sum_{i = 1}^n \score_\theta(B_i) - \frac{h^2}{4n} \sum_{i = 1}^n \var(\score_\theta(B_i)) - \frac{h^2}{4n} \sum_{i = 1}^n \score_\theta(B_i)^2 + o_P(1). \end{equation*} If additionally Assumption~\ref{assumption:detection-regions}\eqref{item:limit-variance} holds, then \begin{equation*} \sum_{i = 1}^n \log \frac{P_{\theta + h_n/\sqrt{n}}(B_i)}{ P_\theta(B_i)} = \frac{h}{\sqrt{n}} \sum_{i = 1}^n \score_\theta(B_i) - \frac{h^2}{2} \kappa(\theta) + o_P(1). \end{equation*} \end{lemma} \noindent The proof of Lemma~\ref{lemma:lan-bits} is quite technical, so we defer it to Section~\ref{sec:proof-lan-bits}. With this lemma, it is not too challenging to demonstrate the local asymptotic normality (Definition~\ref{definition:lan}) of the family $\{P_\theta\}$. Indeed, Lemma~\ref{lemma:derivative-bounds} guarantees that $|\dPtheta(A_n)| \le 2\sqrt{k_n \lip(f)}$ for all $n$, so that $\E_\theta[|\score_\theta(B_i)|^3] \le C \frac{k_i^{3/2} \lip(f)^{3/2}}{P_\theta(A_i)^2 (1 - P_\theta(A_i))^2}$, while Assumption~\ref{assumption:detection-regions}\eqref{item:finite-intervals} guarantees that $\frac{1}{n^3} \sum_{i = 1}^n \E_\theta[|\score_\theta(B_i)|^3] \to 0$. Because $\E[\score_\theta(B_i)] = 0$, the Lyapunov central limit theorem applies to give \begin{equation*} \frac{1}{\sqrt{n}} \sum_{i = 1}^n \score_\theta(B_i) \cd \normal\left(0, \kappa(\theta)\right) \end{equation*} under Assumption~\ref{assumption:detection-regions}\eqref{item:limit-variance}, so that the family $\{P_\theta\}$ is locally asymptotically normal (Def.~\ref{definition:lan}). We now recall the familiar H\'{a}jek-Le-Cam local asymptotic minimax result~\cite[Thm.~8.11]{VanDerVaart98}: if the family $\{P_\theta\}$ is LAN with precision $\kappa(\theta)$, then \begin{equation*} \liminf_{c \to \infty} \liminf_n \sup_{\norm{\tau - \theta} \le c / \sqrt{n}} \E_\tau\left[L(\sqrt{n}(\theta_n - \tau))\right] \ge \E[L(Z / \sqrt{\kappa(\theta)})] \end{equation*} for any symmetric quasi-convex loss $L$, where $Z \sim \normal(0, 1)$. This immediately gives Theorem~\ref{theorem:non-adaptive-minimax}. \subsection{Proof of Lemma~\ref{lemma:derivative-bounds}} \label{sec:proof-derivative-bounds} To see the first claim of the lemma, we consider the simpler special case that $A = [a, b]$. Then as $f$ is Lipschitz (and hence absolutely continuous and a.e.\ differentiable with $\linf{f'} \le \lip(f)$), we have \begin{align*} P_{\theta + h}(A) - P_\theta(A) & = \int_a^b (f(z - \theta - h) - f(z - \theta)) dz \\ & = -\int_a^b \int_0^h f'(z - \theta - u) du dz \\ & = -\int_0^h \int_a^b f'(z - \theta - u) dz du \\ & = \int_0^h f(a - \theta - u) - f(b - \theta - u) du \\ & = \int_0^h (f(a - \theta) - f(b - \theta)) du \pm 2 \int_0^h \lip(f) u du \\ & = \left[f(a - \theta) - f(b - \theta)\right] h \pm \lip(f) h^2. \end{align*} This gives the first two claims of the lemma. For the second, we require a bit more work. Let $L = \lip(f)$ for shorthand. Let $a < b$. Then we always have \begin{equation} \label{eqn:interval-length-thing} P([a, b]) \ge \int_a^b f(z) dz \ge \int_a^b \max\{f(b) - L (b - z), f(a) - L(z - a), 0\} dz. \end{equation} If $f(a) + f(b) \ge L (b - a)$, then the point $\hat{z} = \frac{a + b}{2} - \frac{f(b) - f(a)}{2L}$ satisfies both $f(b) - L(b - \hat{z}) \ge 0$ and $f(a) - L(\hat{z} - a) \ge 0$. The integral~\eqref{eqn:interval-length-thing} then becomes \begin{align*} \lefteqn{\int_a^{\hat{z}} (f(a) - L (z - a) dz) + \int_{\hat{z}}^b (f(b) - L (b - z)) dz} \\ & = \frac{f(a) + f(b)}{2} \left(\frac{b - a}{2}\right) - L \left(\frac{b - a}{2} \right)^2 + \frac{(f(b) - f(a))^2}{4L}, \end{align*} and using the assumption that $\frac{f(a) + f(b)}{2} \ge L(b - a)$, we obtain \begin{align*} \frac{(f(b) - f(a))^2}{4L} & \le \frac{f(b) + f(b)}{2} \frac{b - a}{2} - L \left(\frac{b - a}{2}\right)^2 + \frac{(f(b) - f(a))^2}{4L} \le P([a, b]). \end{align*} That is, $|f(b) - f(a)| \le 2 \sqrt{\lip(f) P([a, b])}$. In the converse case that $f(a) + f(b) \le L(b - a)$, then the integral~\eqref{eqn:interval-length-thing} becomes \begin{align*} P([a, b]) & \ge \int_a^{a + \frac{f(a)}{L}} (f(a) - L (z - a)) dz + \int_{b - \frac{f(b)}{L}}^b (f(b) - L (b - z)) dz \\ & = \frac{f(a)^2}{L} - \frac{f(a)^2}{2L} + \frac{f(b)^2}{L} - \frac{f(b)^2}{2L}, \end{align*} so that $\sqrt{f(a)^2 + f(b)^2} \le \sqrt{2 \lip(f) P([a, b])}$. In sum, we have demonstrated that always the first bound~\eqref{eqn:bound-density-diffs} holds. To show the second inequality in expression~\eqref{eqn:bound-density-diffs}, note that $\sum_i P([a_i, b_i]) \le 1$, and apply Cauchy-Schwarz. \subsection{Proof of Lemma~\ref{lemma:lan-bits}} \label{sec:proof-lan-bits} Our proof follows that of \cite[Thm.~7.2]{VanDerVaart98} closely. We first demonstrate a type of uniform quadratic mean differentiability (Definition~\ref{definition:qmd}) for sets $A$ that are finite unions of intervals. By a Taylor approximation and concavity of $\sqrt{\cdot}$, we have \begin{equation*} \sqrt{a} + \frac{b}{2 \sqrt{a}} - \frac{b^2}{4\sqrt{a}} \le \sqrt{a + b} \le \sqrt{a} + \frac{b}{2 \sqrt{a}} \end{equation*} for any $a > 0$ and $|b| \le 3a/4$. Consequently, recalling that $\score_\theta(A) = \dPtheta(A) / P_\theta(A)$, for any $h \in \R$ and $A = \cup_{i = 1}^k \{[t_i^-, t_i^+]\}$ the union of $k$ intervals, the expansion~\eqref{eqn:expansion-dPtheta} yields \begin{equation*} \left(\sqrt{P_{\theta + h}(A)} - \sqrt{P_\theta(A)} - \half h \score_\theta(A) \sqrt{P_\theta(A)} \right)^2 \le \left( \frac{k \lip(f)}{2 \sqrt{P_\theta(A)}} h^2 + \frac{(|\dPtheta(A) h| + h^2 \lip(f))^2}{P_\theta(A)^{3/2}} \right)^2, \end{equation*} valid for $h$ such that $|\dPtheta(A) h| \le P_\theta(A) / 4$ and $k h^2 \lip^2(f) \le P_\theta(A) / 4$. Thus, under Assumption~\ref{assumption:detection-regions}\eqref{item:finite-intervals}, there exists a numerical constant $C < \infty$ such that \begin{subequations} \label{eqn:h-fourth} \begin{align} \nonumber \left(\sqrt{P_{\theta + h}(A)} - \sqrt{P_\theta(A)} - \half h \score_\theta(A) \sqrt{P_\theta(A)}\right)^2 & \le \left(\frac{h^2 k \cdot \lip(f)}{2 \sqrt{P_\theta(A)}} + \frac{(|\dPtheta(A) h| + k h^2 \lip(f))^2}{ P_\theta(A)^{3/2}}\right)^2 \\ & \le \frac{C}{P_\theta(A)} \left[k^2 \lip^2(f) + \score_\theta(A)^2 + \frac{k^4 h^4 \lip^4(f)}{P_\theta(A)^2} \right] \cdot h^4, \end{align} valid whenever $|\dPtheta(A) h| \le P_\theta(A) / 4$ and $k h^2 \lip^2(f) \le P_\theta(A) / 4$, and similarly, we have \begin{equation} \left(\sqrt{P_{\theta + h}(A^c)} - \sqrt{P_\theta(A^c)} - \half h \score_\theta(A^c) \sqrt{P_\theta(A^c)}\right)^2 \le \frac{C}{P_\theta(A^c)} \left[k^2 \lip^2(f) + \score_\theta(A^c)^2 + \frac{k^4 h^4 \lip^4(f)}{P_\theta(A^c)^2} \right] \cdot h^4. \end{equation} \end{subequations} That is, the family $\{P_\theta\}$ with bit observations $B_n$ satisfies a uniform type of quadratic-mean differentiability (Def.~\ref{definition:qmd}). For shorthand, define $P_n = P_{\theta + h_n / \sqrt{n}}$ and $P = P_\theta$, and let $p_n, p$ be shorthand for the p.m.f.s of the two distributions. For the sets $A_i$ we recall that $B_i = \indic{X_i \in A_i}$. The random variables \begin{equation*} W_{n,i} \defeq 2 \left(\sqrt{\frac{p_n}{p}}(B_i) - 1\right) \end{equation*} are with $P$-probability 1 well-defined, and by the inequalities~\eqref{eqn:h-fourth}, we have that \begin{align} \lefteqn{\var\left(W_{n,i} - \frac{h_n}{\sqrt{n}} \score_\theta(B_i)\right) \le C \frac{k_i^2 \lip^2(f) + \score_\theta(A_i)^2 + \score_\theta(A_i^c)^2}{P_\theta(A_i) P_\theta(A_i^c)} \cdot \frac{h_n^4}{n^2} + C \frac{k^4 \lip^4(f)}{ P_\theta(A_i)^3 P_\theta(A_i^c)^3} \frac{h_n^8}{n^4}} \nonumber \\ & \qquad\qquad\qquad\qquad\qquad \le C \frac{k_i^2 \lip^2(f) + \score_\theta(A_i)^2 + \score_\theta(A_i^c)^2}{P_\theta(A_i) P_\theta(A_i^c)} \cdot \frac{h^4}{n^2} + C \frac{k^4 \lip^4(f)}{ P_\theta(A_i)^3 P_\theta(A_i^c)^3} \frac{h^8}{n^4} \label{eqn:var-wni-expansion} \end{align} whenever \begin{equation*} \frac{h}{\sqrt{n}} \max\{\score_\theta(A_i), \score_\theta(A_i^c)\} \le \frac{1}{4} ~~ \mbox{and} ~~ \frac{k_i h^2}{n} \lip^2(f) \le \frac{\min\{P_\theta(A_i), P_\theta(A_i^c)\}}{4} \end{equation*} Now, we use Assumption~\ref{assumption:detection-regions}\eqref{item:finite-intervals}, coupled with Lemma~\ref{lemma:derivative-bounds} to show that the summed variances converge to zero. Indeed, Lemma~\ref{lemma:derivative-bounds} and inequality~\eqref{eqn:var-wni-expansion} give that \begin{equation*} \var\left(W_{n,i} - \frac{h_n}{\sqrt{n}} \score_\theta(B_i)\right) \le C \cdot \left[\frac{k_i^2}{P_\theta(A_i) P_\theta(A_i^c)} \frac{1}{n} + \frac{k_i}{P_\theta(A_i) P_\theta(A_i^c)} \frac{1}{n} + \frac{k_i^4}{P_\theta(A_i)^3 P_\theta(A_i^c)^3} \frac{1}{n^3}\right] \frac{1}{n}, \end{equation*} where $C < \infty$ depends only on $\lip(f)$ and $h$ (both of which are uniformly bounded) whenever \begin{equation*} \frac{k_i}{P_\theta(A_i) P_\theta(A_i^c)} \frac{1}{n} \le \frac{1}{C}. \end{equation*} Assumption~\ref{assumption:detection-regions}\eqref{item:finite-intervals} thus implies that $\E[\score_\theta(B_i)] = 0$ and \begin{equation} \label{eqn:summed-variances-to-zero} \var\left(\sum_{i = 1}^n W_{n,i} - \frac{h_n}{\sqrt{n}} \score_\theta(B_i) \right) = \sum_{i = 1}^n \var\left(W_{n,i} - \frac{h_n}{\sqrt{n}} \score_\theta(B_i) \right) \to 0. \end{equation} We now control the expectation of the $W_{n,i}$. Defining $\mu_i$ to be the induced counting measure on $B_i = \indic{X_i \in A_i}$, \begin{align*} \sum_{i = 1}^n \E[W_{n,i}] & = 2\sum_{i = 1}^n \left(\int \sqrt{p_n(b)} \sqrt{p(b)} d\mu_i(b) - 1 \right) = -\sum_{i = 1}^n \int \left(\sqrt{p_n(b)} - \sqrt{p(b)}\right)^2 d\mu_i(b) \\ & = -\frac{h_n^2}{4 n} \sum_{i = 1}^n \E[\score_\theta(B_i)^2] - \sum_{i = 1}^n \int \left(\sqrt{p_n(b)} - \sqrt{p(b)} - \frac{h_n}{\sqrt{n}} \score_\theta(b) \sqrt{p(b)}\right)^2 d\mu_i(b) \\ & \qquad \qquad ~ - \sum_{i = 1}^n \int \left(\sqrt{p_n(b)} - \sqrt{p(b)} - \frac{h_n}{\sqrt{n}} \score_\theta(b) \sqrt{p(b)}\right) \frac{h_n}{\sqrt{n}} \score_\theta(b) \sqrt{p(b)}d\mu_i(b) \\ & = -\bigg(\frac{h^2}{4n} \sum_{i = 1}^n \E[\score_\theta(B_i)^2]\bigg) - o(1) \end{align*} uniformly in $h$, with a derivation completely paralleling that above. Therefore, we obtain \begin{equation*} \sum_{i = 1}^n W_{n,i} = \sum_{i = 1}^n \left(W_{n,i} - \frac{h_n}{\sqrt{n}} \score_\theta(B_i)\right) + \frac{h_n}{\sqrt{n}} \sum_{i = 1}^n \score_\theta(B_i) = -\frac{h^2}{4n} \sum_{i = 1}^n \E[\score_\theta(B_i)^2] + \frac{h}{\sqrt{n}} \sum_{i = 1}^n \score_\theta(B_i) + o_P(1), \end{equation*} where we have used that $h_n \to h$. Now, we write the log-likelihood ratio. We have \begin{align*} \sum_{i = 1}^n \log \frac{p_n(B_i)}{p(B_i)} & = 2 \sum_{i = 1}^n \log\left(1 + \half W_{n,i}\right) \\ & = \sum_{i = 1}^n W_{n,i} - \frac{1}{4} \sum_{i = 1}^n W_{n,i}^2 + \half \sum_{i = 1}^n W_{n,i}^2 R(W_{n,i}) \end{align*} where the remainder $|R(W_{n,i})| \le |W_{n,i}|$ for $|W_{n,i}| \le 1$. Using the Taylor expansions of $\sqrt{\cdot}$ and Lemma~\ref{lemma:derivative-bounds}, we have \begin{align} \nonumber \half W_{n,i} & = \half \score_\theta(B_i) \frac{h_n}{\sqrt{n}} \pm \left|\frac{h_n^2}{n} \frac{k_i \lip(f)}{p(B_i)} + \frac{h_n^2}{n} \score_\theta(B_i)^2 + \frac{ h_n^4}{n^2} \frac{k_i^2 \lip(f)^2}{p(B_i)^2}\right| \\ & \in \pm C \left|\frac{\sqrt{k_i}}{\sqrt{n} p(B_i)} + \frac{k_i}{n p(B_i)} + \frac{\sqrt{k_i}}{p(B_i)^2 n} + \frac{k_i^2}{p(B_i)^2 n^2} \right| \to 0, \label{eqn:max-wni-zero} \end{align} where $C < \infty$ depends only on $\lip(f)$ and $h_n$ and so is uniformly bounded; the limit follows from Assumption~\ref{assumption:detection-regions}\eqref{item:finite-intervals}. Consequently $\max_i W_{n,i} \to 0$, so that \begin{equation} \label{eqn:almost-at-the-end} \sum_{i = 1}^n \log \frac{p_n(B_i)}{p(B_i)} = \frac{h}{\sqrt{n}} \sum_{i = 1}^n \score_\theta(B_i) - \frac{1}{4} \sum_{i = 1}^n \E[\score_\theta(B_i)^2] - \frac{1}{4} \sum_{i = 1}^n W_{n,i}^2 + o_P(1). \end{equation} It remains to compute $\E[W_{n,i}^2]$. Using the bounds that $|\score_\theta(B_i)| \le C \sqrt{k_i} / p(B_i)$ from Lemma~\ref{lemma:derivative-bounds}, the expansion~\eqref{eqn:max-wni-zero} yields \begin{align*} \lefteqn{\left|\E\left[W_{n,i}^2 - \frac{h_n^2}{2n} \score_\theta(B_i)^2\right] \right|} \\ & \le \frac{C}{n} \left[ \frac{k_i^{3/2}}{p(A_i)(1 - p(A_i)) \sqrt{n}} + \frac{k_i^{3/2}}{p(A_i)^2 (1 - p(A_i))^2 \sqrt{n}} + \frac{k_i^2}{p(A_i)(1 - p(A_i))} \frac{1}{n^{3/2}} \right] \\ & \qquad ~ + \frac{C}{n} \left[\frac{k_i^2}{p(A_i)(1 - p(A_i))} \frac{1}{n} + \frac{k_i}{p(A_i)^3(1 - p(A_i))^3} \frac{1}{n} + \frac{k_i^4}{p(A_i)^3 (1 - p(A_i))^3} \frac{1}{n^3} \right], \end{align*} where $C$ depends only on $h$ and $\lip(f)$. Thus \begin{equation*} \sum_{i = 1}^n W_{n,i}^2 = \frac{h_n^2}{n} \sum_{i = 1}^n \score_\theta(B_i)^2 + o(1), \end{equation*} giving Lemma~\ref{lemma:lan-bits}. \section{Proof of Theorem~\ref{thm:non_existence}} \label{proof:thm:non_existence} Let $\Xi$ be the set of points $\theta \in \Theta$ for which $\kappa(\theta) = \eta(0)$. Since $B_1,B_2,\ldots$ satisfy the conditions in Theorem~\ref{theorem:non-adaptive-minimax}, $\theta$ is in $\Xi$ if and only if $\lim_{n\to \infty} L_n(A_1,\ldots,A_n;\theta) = \eta(0)$. By assumption, we have $B_i = \indic{X_i \in A_i}$ where each $A_i = \cup_{i=1}^K \{[a_{i,k},b_{i,k}]\}$, where $a_{i,1} \leq b_{i,1} \leq \ldots \leq a_{i,K}, b_{i,K}$, and $a_{i,1}$ and $b_{i,K}$ may take the values $-\infty$ and $\infty$, respectively. Denote the endpoints \begin{equation*} E_i = \cup_{k=1}^{K}\{a_{i,k},b_{i,k}\}, \end{equation*} and for $\theta$ and $\epsilon>0$, define \begin{equation*} S_n(\theta, \epsilon) \triangleq \left\{ i\leq n ~ \mbox{s.t.}~ (\theta-\epsilon,\theta+\epsilon) \cap E_i \neq \emptyset \right\} \end{equation*} In words, $S_n$ contains all integers smaller than $n$ in which an $\epsilon$-ball around $\theta$ contains an endpoint of one of the intervals defining $A_i$. We now claim that if $\theta \in \Xi$ then $\card(S_n(\theta, \epsilon))/n \to 1$. Indeed, for such $\theta$ we have \begin{align} & L_n(A_1,\ldots,A_n; \theta) \nonumber \\ & = \frac{1}{n} \sum_{i \in S_n(\epsilon,\theta)} \frac{ \left(\sum_{k=1}^{K} f(\theta - b_{i,k})- f(\theta - a_{i,k}) \right)^2}{ \sum_{k=1}^{K} \left( F(\theta - b_{i,k})- F(\theta - a_{i,k}) \right) \left(1-\sum_{k=1}^{K} \left( F(\theta - b_{i,k})- F(\theta - a_{i,k}) \right)\right)} \nonumber \\ & + \frac{1}{n}\sum_{i \notin S_n(\epsilon,\theta) } \frac{ \left(\sum_{k=1}^{K} f(b_{i,k}-\theta) - f(a_{i,k}-\theta) \right)^2} { \sum_{k=1}^{K} \left( F(\theta - b_{i,k})- F(\theta - a_{i,k}) \right) \left(1-\sum_{k=1}^{K} \left( F(\theta - b_{i,k})- F(\theta - a_{i,k}) \right)\right)} \nonumber \\ & \overset{(a)}{\leq} \frac{\card\left(S_n(\theta,\epsilon)\right)}{n} \eta(0) + \frac{n-\card\left(S_n(\theta,\epsilon) \right) }{n} \eta(\epsilon) \label{eq:non_existence_proof1} \end{align} where $(a)$ follows from Lemma~\ref{lem:bound_intervals_delta} with $\delta =0$, and the fact that for $i \in S_n(\theta, \epsilon)$, \begin{equation*} \max\left\{ \max_k \eta(b_{i,k}-\theta) , \max_k \eta(a_{i,k}-\theta) \right\} \leq \eta(\epsilon) < \eta(0). \end{equation*} Unless $\card \left(S_n(\theta, \epsilon) \right)/n \to 1$, we get that \eqref{eq:non_existence_proof1}, hence $L_n(A_1,\ldots,A_n ; \theta)$, are bounded from above by a constant that is smaller then $\eta(0)$ in contradiction to the fact that $\theta \in \Xi$. Assume for the sake of contradiction that there exists $N \geq 2K + 1$ distinct elements $\theta_1,\ldots,\theta_N \in \Xi$. Since each $A_i$ consists of at most $K$ intervals, we have that \begin{equation} \label{eq:few_optimality_points_proof} \card (\cup_{i=1}^n \mathcal B_i) \leq 2 n K. \end{equation} Fix $\epsilon>0$ such that \begin{equation*} \epsilon < \frac{1}{2}\min_{i\neq j} |\theta_i - \theta_j|. \end{equation*} Since for each $\theta \in \Theta$ we have $S_n(\theta, \epsilon) \to 1$, there exists $n$ large enough such that \begin{equation*} \card \left(S_n(\theta_i, \epsilon) \right) \geq n \left(1-\frac{1}{2N} \right) \end{equation*} for all $i=1,\ldots,N$. However, $S_n(\theta_1,\epsilon), \ldots S_n(\theta_N,\epsilon)$ are disjoint, so the cardinality of their union is at least $n\left(1-\frac{1}{2N} \right)N > 2nK + n/2$, a contradiction to inequality~\eqref{eq:few_optimality_points_proof}. \section{Proof of Theorem~\ref{thm:sgd}} \label{proof:sgd} The estimation algorithm~\eqref{eq:sgd_alg} is a special case of the stochastic gradient procedures in the papers \cite{PolyakJu92, polyak1990new}. We rely on several of their results. Throughout this proof, we assume without loss of generality that the median $\theta = \mbox{med}(P) = 0$. \subsection{Proof of Theorem~\ref{thm:sgd}\eqref{item:normal-sgd}} \label{sec:proof-normal-sgd} Consider the following simplified version of \cite[Thm. 4]{polyak1992acceleration}: \begin{corollary}{\cite[Thms. 3 \& 4]{PolyakJu92}} \label{corollary:polyak-juditsky} Let $\varphi : \R \to \R$ and $Z_i$ be i.i.d.\ random variables, and \begin{equation*} X_i = \theta + Z_i. \end{equation*} Define \begin{align} \begin{split} \theta_i & = \theta_{i-1} + \gamma_i \varphi(X_i - \theta_{i-1}), \\ \bar{\theta}_n & = \frac{1}{n} \sum_{i=0}^{n-1} \theta_i, \end{split} \label{eq:Polyak_Juditsky_alg} \end{align} where in addition, \begin{enumerate}[(i)] \item There exists $K_1$ such that $\left| \varphi(x) \right| \leq K_1(1+|x|)$ for all $x\in \R$. \item The sequence $\left\{ \gamma_i \right\}_{i=1}^\infty$ satisfies condition~\eqref{eqn:lazy-gamma}. \item \label{item:zero-gradient} The function $\psi(x) \triangleq \ex{ \varphi(x+Z_1)}$ satisfies $\psi(0) = 0$ and $x\psi(x) > 0$ for $x\neq 0$. Moreover, $\psi$ is differentiable at 0 with $\psi'(0) > 0$ and there exists $K_2$ and $0 < \lambda \leq 1$ such that \begin{equation} \label{eqn:local-hessian-psi} \left| \psi(x) - \psi'(0)x \right|\leq K_2 |x|^{1+\lambda}. \end{equation} \item The function $\chi(x) \triangleq \ex{\varphi^2(x+Z_1)}$ is continuous at zero. \end{enumerate} Then $\bar{\theta}_n \cas \theta$ and $ \sqrt{n}({\theta}_n - \theta) \cd \normal(0,V)$ for $V = \frac{ \chi(0)} {\psi'(0)^2}$. \end{corollary} Using the notation in Corollary~\ref{corollary:polyak-juditsky}, we set $\varphi(x) = \sgn(x)$ and $Z_i = X_i - \theta$, where $\theta = \mbox{med}(P)$. Without loss of generality and for notational convenience, we assume for the remainder of this derivation that $\theta = 0$. As a consequence, we have $\mbox{med}(Z) = 0$, and $\chi(x) = \ex{ \sgn^2(x+Z_1) }= 1$, so $\chi(0) = 1$. In addition, \begin{align*} \psi(x) & = \ex{ \sgn(x+ Z_1) } = P(Z \ge -x) - P(Z < -x) = 1 - 2 P(Z \le -x). \end{align*} Using that $P$ has a density $f$ near its median, it follows that $\psi'(x) = 2f(-x)$ and thus $\psi'(0) = 2f(0) > 0$. We may now verify that the conditions in Corollary~\ref{corollary:polyak-juditsky} hold for $\lambda = 1$. Condition~(i) is obvious, and the convexity of $|\cdot|$ gives most of condition~(iii) excepting inequality~\eqref{eqn:local-hessian-psi}. For that, note that as $f$ is Lipschitz near 0 with constant $\lip_0(f)$, we have for small $x$ that \begin{equation*} \psi(x) = 2 \int_0^x f(-t) dt = 2 \int_0^x \left[f(0) \pm \lip_0(f) t \right] dt = 2 f(0) x \pm \frac{\lip_0(f)}{2} x^2 = \psi'(0) x \pm \frac{\lip_0(f)}{2} x^2, \end{equation*} so that condition~(iii) holds. As evidently $\chi(0) / \psi'(0)^2 = \frac{1}{4 f(0)^2}$, Corollary~\ref{corollary:polyak-juditsky} gives Theorem~\ref{thm:sgd}\eqref{item:normal-sgd}. \subsection{Proof of Theorem~\ref{thm:sgd}\eqref{item:sgd-regular}} \label{sec:proof-sgd-regular} This proof requires somewhat more technicality than the first part of the theorem, including a brief detour into local asymptotic normality theory, regular estimators, and quadratic-mean differentiability~\cite[cf.]{VanDerVaart98}. We assume without loss of generality that the median of the density $f$ is 0, so that if $P_\theta$ has density $f(\cdot - \theta)$, the median of $P_\theta$ is $\theta$. We begin by recalling the statistical concepts we require. \begin{definition} \label{definition:regular-estimator} A sequence of estimators $T_n$ for a parameter $\theta$ in the parametric family $\{P_\theta\}_{\theta \in \Theta}$ is \emph{regular at $\theta$} if there exists a distribution $Q$ such that for any bounded sequence $h_n$, \begin{equation*} \sqrt{n}(T_n - (\theta + h_n / \sqrt{n})) \mathop{\cd}_{P_{\theta + h_n/\sqrt{n}}} Q. \end{equation*} \end{definition} \begin{definition} \label{definition:qmd} Let $\{P_\theta\}_{\theta \in \Theta}$ have densities $p_\theta$ with respect to a base measure $\mu$. The family is \emph{quadratic mean differentiable} (QMD) at $\theta$ with score $\score_\theta$ if \begin{equation} \label{eqn:qmd} \int \left(\sqrt{p_{\theta + h}} - \sqrt{p_\theta} - \half h^\top \score_\theta\sqrt{p_\theta} \right)^2 d\mu = o(\norm{h}^2) \end{equation} as $h \to 0$. \end{definition} \begin{definition} \label{definition:lan} A family of distributions $\{P_\theta\}_{\theta \in \Theta}$ is \emph{locally asymptotically normal with information matrix $I_\theta$} (LAN) at $\theta$ if there exists a sequence of random vectors $Z_n$ such that for all $h_n \to h$, \begin{equation*} \sum_{i = 1}^n \log \frac{dP_{\theta + h_n/\sqrt{n}}}{dP_\theta} (X_1, \ldots, X_n) = h^\top Z_n - \half h^\top I_\theta h + o_P(1) \end{equation*} where $Z_n \cd \normal(0, I_\theta)$ under $P_\theta$, where $X_i \simiid P_\theta$. \end{definition} These three definitions are linked in our case by a few important results. First~\cite[Theorem 7.2]{VanDerVaart98}, if $\{P_\theta\}$ is QMD (Def.~\ref{definition:qmd}) at the point $\theta$, then it is locally asymptotically normal with $Z_n = \frac{1}{\sqrt{n}} \sum_{i = 1}^n \score_\theta(X_i)$ and information matrix $I_\theta = \E_\theta[\score_\theta \score_\theta^\top]$. Moreover, in any family $\{P_\theta\}$ that is LAN (Def.~\ref{definition:lan}) at $\theta$, if $T_n$ is a regular estimator (Def.~\ref{definition:regular-estimator}) at $\theta$ with limiting distribution $Q$, then for any bounded, symmetric, quasi-convex loss $L$ and $c < \infty$, \begin{equation} \label{eqn:limit-law-regular} \limsup_n \sup_{\norm{h} \le c} \E_{P_{\theta + h/\sqrt{n}}} \left[L(\sqrt{n}(T_n - \theta - h / \sqrt{n}))\right] = \E[L(W)] ~~ \mbox{for~}W \sim Q \end{equation} (see Beran~\cite{beran1995role}, Eq.~(4.2)). Thus, we show two results: first, that that the family $\{P_\theta\}$ of distributions defined by the shifted densities $\{f(\cdot - \theta)\}_{\theta \in \R}$ is quadratic-mean-differentiable at any $\theta$, and second, that $\bar{\theta}_n$ is regular and asymptotically normal. The combination evidently gives the theorem. For quadratic mean differentiability, we have the following lemma, somewhat more general than we need; we defer proof to Sec.~\ref{sec:proof-qmd}. \begin{lemma}[Extension of \cite{VanDerVaart98}, Lemma 7.6] \label{lemma:qmd} Let $p_\theta$ be a density with respect to $\mu$, and assume that $\theta \mapsto s_\theta(x) \defeq \sqrt{p_\theta(x)}$ is absolutely continuous for all $x$. Let $\dot{p}_\theta(x) = \nabla_\theta p_\theta(x)$ (when it exists), and assume that \begin{equation*} \mu(\{x : \dot{p}_\theta(x) ~ \mbox{fails~to~exist}\}) = 0. \end{equation*} Assume that $I_\theta \defeq \E_{P_\theta}[\dot{p}_\theta \dot{p}_\theta^\top / p_\theta^2]$ is continuous at $\theta_0$. Then $P_\theta$ is QMD (Definition~\ref{definition:qmd}) at $\theta = \theta_0$ with $\score_\theta = \dot{p}_\theta / p_\theta$. \end{lemma} By the assumption in Theorem~\ref{thm:sgd} that the density $f$ is Lipschitz continuous, we see that the location family $\{P_\theta\}_{\theta \in \R}$ defined by $dP_\theta(x) = f(x - \theta)$ satisfies the conditions of Lemma~\ref{lemma:qmd}. It remains to show that the average $\bar{\theta}_n$ is regular: \begin{lemma} \label{lemma:sgd-median-regular} Let $h_n \to h \in \R$, and define $P_n = P_{\theta + h_n / \sqrt{n}}^n$. Then $\bar{\theta}_n$ is regular with \begin{align} \label{eqn:sgd-median-regular} \sqrt{n}\left( \bar{\theta}_n - \theta - h_n / \sqrt{n}\right) \mathop{\cd}_{P_n} \normal\left( h,\frac{1}{4 f(0)^2}\right). \end{align} \end{lemma} \begin{proof} To show the convergence~\eqref{eqn:sgd-median-regular} we use the following refinement of Corollary~\ref{corollary:polyak-juditsky}, which provides a generalized convergence result for iteratively defined $\theta_n$, and whose proof we defer to Section~\ref{proof:normal-expansion}. % \begin{corollary} \label{corollary:normal-expansion} Let the conditions of Corollary~\ref{corollary:polyak-juditsky} hold, meaning that $\theta_i = \theta_{i-1} + \gamma_i \varphi(X_i - \theta_{i-1})$ for $X_i = \theta + Z_i$, where $Z_i$ are i.i.d.\ with $\E[\varphi(Z)] = 0$. Additionally assume the local smoothness condition that there exist $0 < \lambda \le 1$ and $K < \infty$ such that \begin{equation} \label{eqn:additional-local-smooth} \E[|\varphi(x + Z) - \varphi(Z)|^2] \le K (|x|^\lambda + x^2). \end{equation} Set $\Delta_i = \theta_i - \theta$ and $\bar{\Delta}_i = \frac{1}{n} \sum_{i=1}^n \Delta_i$. Then \begin{enumerate}[(i)] \item \label{item:regularity} The sequence $\Delta_i$ is regular, that is, \begin{equation} \sqrt{n} \bar{\Delta}_n = -\frac{1}{\sqrt{n}} \frac{1}{\psi'(0)} \sum_{i=1}^{n-1} \varphi(Z_i) + o_{P,n}(1). \label{eq:normal_expansion_lem} \end{equation} \item \label{item:apply-le-cam} Let $Z_i$ as in Corollary~\ref{corollary:polyak-juditsky} have absolutely continuous density $p$ with median $0$, define $\score_h(z) = \frac{p'(z - h)}{p(z - h)}$, and assume that $I_h \defeq \E_\theta[\score_h(Z)^2]$ is continuous in $h$ near 0. Then for any converging sequence $h_n \to h$, \begin{equation*} \sqrt{n} \bar{\Delta}_n \mathop{\cd}_{P_{\theta + h_n/\sqrt{n}}^n} \normal\left( \frac{-h}{\psi'(0)} \E_p[\varphi(Z) \score_0(Z)], \frac{\chi(0)}{\psi'^2(0)} \right). \end{equation*} \end{enumerate} \end{corollary} We now verify that the setting of Theorem~\ref{thm:sgd} (and Lemma~\ref{lemma:sgd-median-regular}) satisfies the conditions of Corollary~\ref{corollary:polyak-juditsky}. First, we have the obvious fact that \begin{equation*} |\sgn(z) - \sgn(x + z)| \le 2 \cdot \indic{|x| \ge |z|}. \end{equation*} Recalling that the density $f$ is Lipschitz with median 0, for $\varphi(z) = \sgn(z)$, and $Z = X - \theta$ distributed with density $f$, we have \begin{align*} \ex{ \left| \varphi(Z) - \varphi(x + Z) \right| } \leq 2 \Prob\left( |Z_1| \le |x| \right) = 2 \int_{-|x|}^{|x|} f(t) dt = 4 f(0)|x| \pm 2 \int_{-|x|}^{|x|} \lip(f) t dt = 4 f(0) |x| + 2 \lip(f) x^2 \end{align*} where $\lip(f)$ is the Lipschitz constant of $f$. For large $x$ we always have $\E[|\varphi(Z) - \varphi(x + Z)|] \le 2$, so that condition~\eqref{eqn:additional-local-smooth} holds. In addition, we have \begin{equation*} \int_{\R} \varphi(x) f'(x ) dx = \int_{\R} \sgn(x) f'(x ) dx = \int_0^\infty f'(x ) dx - \int_{-\infty}^0 f'(x ) = -2 f(0) = -\psi'(0). \end{equation*} Corollary~\ref{corollary:normal-expansion} now implies the convergence \eqref{eqn:sgd-median-regular}. \end{proof} Combining Lemmas~\ref{lemma:qmd} and~\ref{lemma:sgd-median-regular} with the limit~\eqref{eqn:limit-law-regular} gives Theorem~\ref{thm:sgd}\eqref{item:sgd-regular}. \subsection{Proof of Theorem~\ref{thm:sgd}\eqref{item:sgd-ms-convergence}} \label{sec:proof-sgd-ms-convergence} We begin with the following result from \cite{polyak1990new}: \begin{corollary}[\cite{polyak1990new}, Theorem 2] \label{corollary:polyak-mse} Define the iteration \begin{align} \label{eq:polyak_new_measurements} \begin{cases} U_n = U_{n-1} - \gamma_n \varphi(Y_n), & Y_n = g'(U_{n-1})+Z_n \\ \bar{U}_n= \frac{1}{n} \sum_{i=1}^n U_n, & n=1,2,\ldots. \end{cases} \end{align} Assume that the function $g$ is $\mc{C}^2$, strictly convex, has Lipschitz derivative, and is minimized by $x^\star$. Moreover, assume that the noises $Z_n$ are i.i.d.\ with density $p$ and that the Fisher information $\E[(p'(Z))^2 / p(Z)^2]$ exists and is finite. Let $\psi(x)$ and $\chi(x)$ be defined as in Corollary~\ref{corollary:polyak-juditsky} and satisfy the conditions in the corollary. Assume in addition that $\chi(0)>0$, condition \eqref{eqn:local-hessian-psi} with $\lambda = 1$, and there exits $K_3$ such that \begin{equation*} \ex{ | \varphi(x+Z_1) |^4 } \leq K_3(1+|x|^4). \end{equation*} Finally, assume that the sequence $\{\gamma_n \}$ satisfies conditions \eqref{eqn:lazy-gamma} and \eqref{eqn:stringent-gamma}. Then \begin{equation*} V_n \defeq \E\Big[(\bar{U}_n-x^\star )^2\Big] = n^{-1}\frac{\chi(0)} { (\psi'(0))^2 (g''(x^\star))^2 } + o(n^{-1}). \end{equation*} \end{corollary} We apply Corollary~\ref{corollary:polyak-mse} with $g(x) = 0.5(x-\theta)^2$, $\varphi(x) = \sgn(x)$, $Z_n = \theta-X_n$. The update~\eqref{eq:polyak_new_measurements} gives \begin{align*} U_n & = U_{n-1} + \gamma_n \sgn(X_n-U_{n-1} ), \end{align*} so the estimator $\bar{U}_n$ is identical to the stochastic gradient estimator~\eqref{eq:sgd_alg} with $\bar{\theta}_n = \frac{1}{n} \sum_{i = 1}^n \theta_i$. We have $\E[\varphi(x + Z)^4] = 1$ and by assumption the Fisher information $\E[(f'(Z))^2 / f(Z)^2]$ exists, and the functions $\psi$ and $\chi$ have the desired conditions of Corollary~\ref{corollary:polyak-juditsky} (as we verify in Section~\ref{sec:proof-normal-sgd}). Finally, the function $\theta \mapsto \E_{P_\theta}[(\bar{\theta}_n - \theta)^2]$ is continuous in $\theta$, so that for $x^\star = \theta$ and $g'' \equiv 1$, we may apply Corollary~\ref{corollary:polyak-mse} and a finite approximation to the prior measure $\pi$ to obtain Theorem~\ref{thm:sgd}\eqref{item:sgd-ms-convergence}. \subsection{Proof of Corollary~\ref{corollary:normal-expansion}} \label{proof:normal-expansion} \subsubsection*{Proof of Corollary~\ref{corollary:normal-expansion}\eqref{item:regularity}} The proof of part~\eqref{item:regularity} requires two additional lemmas of Polyak and Juditsky~\cite{PolyakJu92}. \begin{lem}[\cite{PolyakJu92}, Lemma 2] \label{lemma:polyak-expansion} Define the process $\Delta_i^1 = \Delta_{i-1}^i - \gamma_i (A \Delta_{i-1} + \xi_i)$ for $i = 1, 2, \ldots$. Assume that $A>0$ and the stepsizes $\gamma_i$ satisfy condition~\eqref{eqn:lazy-gamma}. Then for $\bar{\Delta}_n^1 = \frac{1}{n} \sum_{i = 1}^n \Delta_i^1$, we have \begin{equation} \label{eqn:polyak-expansion} \sqrt{n} \bar{\Delta}_n^1 = \frac{\alpha_n \Delta_0^1}{\sqrt{n} \gamma_0} + \frac{1}{\sqrt{n} A} \sum_{i=1}^{n-1} \xi_i + \frac{1}{\sqrt{n}}\sum_{i=1}^{n-1} w_i^n \xi_i, \end{equation} where $\alpha_n$ and $w_i^n$ are real numbers such that $|\alpha_n| \leq K$ and $|w_i^n|\leq K$ for some $K< \infty$, and $\lim_{n\to \infty} \frac{1}{n} \sum_{i=1}^{n-1} |w_i^n| = 0$. \end{lem} \begin{lem}[\cite{PolyakJu92}] \label{lemma:converging-power-sum} Under the conditions of Corollary~\ref{corollary:normal-expansion}, with probability 1, \begin{equation*} \sum_{i=1}^\infty \frac{|\Delta_{i}|^{1+\lambda}}{\sqrt{i}} < \infty. \end{equation*} \end{lem} \noindent Lemma~\ref{lemma:converging-power-sum} follows from the proof of Theorem 2 in \cite[page 851]{polyak1992acceleration}. We separate the proof of part~\eqref{item:regularity} into two lemmas, which mirror the proofs of Polyak and Juditsky~\cite{PolyakJu92}; together they immediately give the result. \begin{lemma} The expansion~\eqref{eq:normal_expansion_lem} holds for the process $\bar{\Delta}^1_n$ defined by the iteration \begin{align} \label{eqn:polyak-expansion_lem1_alg} & \Delta_i^1 = \Delta_{i-1}^1 - \gamma_i \psi'(0) \Delta_{i-1}^1 - \gamma_i \varphi(Z_i), \qquad \Delta_0^1 = \Delta_0\\ & \bar{\Delta}^1_n = \frac{1}{n}\sum_{i=0}^{n-1} \Delta^1_i. \nonumber \end{align} \end{lemma} \begin{proof} To prove this claim, use Lemma~\ref{lemma:polyak-expansion} with $A = \psi'(0)$ and $\xi_i = -\varphi(Z_i)$, which by condition~\eqref{item:zero-gradient} in Corollary~\ref{corollary:polyak-juditsky} gives that $\E[\xi_i] = 0$ and that the $\xi_i$ are independent. The first term $\alpha_n \Delta_0^1 / \gamma_0 \sqrt{n} \to 0$ in Eq.~\eqref{eqn:polyak-expansion}. In addition, by independence and that the $\xi_i$ are mean-zero, we have \begin{align*} \ex{ \left( \frac{1}{\sqrt{n}} \sum_{i=1}^{n-1} w_i^n \xi_i \right)^2 } & = \frac{1}{n} \sum_{i=1}^n (w_i^n)^2 \ex{ \xi_i^2} + \frac{1}{n} \sum_{i\neq j}^n w_i^n w_j^n \ex{ \xi_i \xi_j} \\ & = \frac{1}{n} \sum_{i=1}^n (w_i^n)^2 \ex{ \varphi(Z_i)^2} = \chi(0) \frac{1}{n} \sum_{i=1}^n (w_i^n)^2 \to 0 \end{align*} by Lemma~\ref{lemma:polyak-expansion}. Thus, the expansion~\eqref{eqn:polyak-expansion} in Lemma~\ref{lemma:polyak-expansion} gives \begin{equation*} \sqrt{n} \bar{\Delta}^1_n = -\frac{1}{\sqrt{n}} \frac{1}{\psi'(0)} \sum_{i=1}^{n-1} \varphi(Z_i)+ o_{P,n}(1) \end{equation*} as desired. \end{proof} We then have the following asymptotic equivalence. \begin{lemma} The sequences $\bar{\Delta}_n$ and $\bar{\Delta}^1_n$ are asymptotically equivalent, meaning that $\sqrt{n} (\bar{\Delta}_n - \bar{\Delta}_n^1) \cp 0$. \end{lemma} \begin{proof} From the recursions~\eqref{eq:Polyak_Juditsky_alg} and \eqref{eqn:polyak-expansion_lem1_alg}, the difference $\delta_i = \Delta_i - \Delta_i^1$ satisfies \begin{equation*} \delta_i = \delta_{i-1} - \gamma_i \psi'(0) \delta_{i-1} + \gamma_i \left( \psi'(0) \Delta_{i-1} + \varphi(Z_i) - \varphi(\Delta_{i-1} + Z_i) \right), \end{equation*} where $\delta_0 = 0$. Applying Lemma~\ref{lemma:polyak-expansion} with the choices $\xi_i = \psi'(0) \Delta_{i-1} + \varphi(Z_i) - \varphi(\Delta_{i-1} + Z_i)$ yields \begin{align} \sqrt{n}\bar{\delta}_n & = \frac{1}{\sqrt{n}} \sum_{i=1}^{n-1} \left( \frac{1}{\psi'(0)} + w_i^n \right) \xi_i \nonumber \\ & = \frac{1}{\sqrt{n}} \sum_{i=1}^{n-1} \left( \frac{1}{\psi'(0)} + w_i^n \right) \left( \psi'(0) \Delta_{i-1} - \psi(\Delta_{i-1}) \right) \label{eq:PJ_proof1} \\ & \qquad ~ + \frac{1}{\sqrt{n}} \sum_{i=1}^{n-1} \left( \frac{1}{\psi'(0)} + w_i^n \right) \left( \psi(\Delta_{i-1}) + \varphi(Z_i) - \varphi(\Delta_{i-1}+Z_i) \right) \label{eq:PJ_proof2} \end{align} For the term \eqref{eq:PJ_proof1}, the assumption~\eqref{eqn:local-hessian-psi} that $|\psi(x) - \psi'(0) x| = O(x^{1 + \lambda})$ and that $\sup_{i,n} |w_i^n| < \infty$ by Lemma~\ref{lemma:polyak-expansion} give that there exists $K < \infty$ such that $|\psi'(0)^{-1} + w_i^n| |\psi'(0) \Delta_{i-1} - \psi(\Delta_{i-1})| \le K |\Delta_i|^{1 + \lambda}$. Lemma~\ref{lemma:converging-power-sum} gives that $\sum_{i = 1}^n \frac{1}{\sqrt{i}} |\Delta_i|^{1 + \lambda} < \infty$, and so the Kronecker lemma gives that \begin{equation*} \frac{1}{\sqrt{n}} \sum_{i=1}^{n-1} \left( \frac{1}{\psi'(0)} + w_i^n \right) \left( \psi'(0) \Delta_{i-1} - \psi(\Delta_{i-1}) \right) \cas 0. \end{equation*} The term \eqref{eq:PJ_proof2} is somewhat more challenging to control. We define \begin{equation*} \epsilon_i \triangleq \psi(\Delta_{i-1}) + \varphi(Z_i) - \varphi(\Delta_{i-1}+Z_i), \end{equation*} and let $\mc{F}_i = \sigma(Z_1, \ldots, Z_i)$ be the $\sigma$-field of the randomness through time $i$. We use a square integrable martingale convergence theorem~\cite[Exercise~5.3.35]{Dembo16}. Noting that $\Delta_i \in \mc{F}_i$, we have \begin{align} \E[\epsilon_i^2 \mid \mc{F}_{i-1}] & = \E[(\psi(\Delta_{i-1}) + \varphi(Z_i) - \varphi(\Delta_{i-1} + Z_i))^2 \mid \mc{F}_{i-1}] \nonumber \\ & \le 2 \psi(\Delta_{i-1})^2 + 2 \E[(\varphi(\Delta_{i-1} + Z_i) - \varphi(Z_i))^2 \mid \mc{F}_{i-1}] \nonumber \\ & \le K \left[|\Delta_{i-1}|^{1 + \lambda} + |\Delta_{i-1}|^\lambda + \Delta_{i-1}^2 \right], \label{eqn:bound-psi-error-expectations} \end{align} where inequality~\eqref{eqn:bound-psi-error-expectations} follows by the conditions~\eqref{eqn:local-hessian-psi} and~\eqref{eqn:additional-local-smooth}, and $\E[\varepsilon_i \mid \mc{F}_{i-1}] = 0$ for all $i$ by definition of $\psi(x) = \E[\varphi(x + Z)]$ and that $\psi(0) = 0$. We now control the expectations of these quantities. For $R < \infty$, define the stopping time $\tau_R \defeq \inf \{i : |\Delta_i| > R\}$, which satisfies $\{\tau_R \le i\} \in \mc{F}_i$ for each $i$. Then using~\cite[Eq.~(A14)]{PolyakJu92}, we have \begin{equation*} \E[\Delta_i^2 \indic{\tau_R > i}] \le K \gamma_i, \end{equation*} and so inequality~\eqref{eqn:bound-psi-error-expectations} gives that \begin{equation*} \E\bigg[\sum_{i = 1}^\infty \frac{1}{i} |\varepsilon_i|^2 \indic{\tau_R > i} \bigg] \le K \sum_{i = 1}^\infty \frac{\gamma_i^\lambda}{i} < \infty ~~ \mbox{so} ~~ \sum_{i = 1}^\infty \frac{1}{i} \varepsilon_i^2 \indic{\tau_R > i} < \infty ~ \mbox{a.s.} \end{equation*} by Condition~\eqref{eqn:lazy-gamma}. As with probability 1 there exists some (random) $R < \infty$ such that $\tau_R = \infty$, we obtain that \begin{equation*} \sum_{i = 1}^\infty \frac{1}{i} \varepsilon_i^2 < \infty ~ \mbox{a.s.} \end{equation*} Applying the square integrable martingale convergence theorem of \cite[Ex.~5.3.35]{Dembo16}, we have \begin{equation*} \frac{1}{\sqrt{n}} \sum_{i = 1}^n \left(\frac{1}{\psi'(0)} + w_i^n\right) \epsilon_i \cas 0, \end{equation*} so that both equations~\eqref{eq:PJ_proof1} and~\eqref{eq:PJ_proof2} converge almost surely to 0. \end{proof} \subsubsection*{Proof of Corollary~\ref{corollary:normal-expansion}\eqref{item:apply-le-cam}} This is essentially an immediate consequence of Le Cam's third lemma~\cite[Example 6.7]{VanDerVaart98}. Recall~\cite[Thm.~7.2]{VanDerVaart98} that if a family $\{P_\theta\}_{\theta \in \Theta}$ is quadratic mean differentiable at $\theta$ with score $\score_\theta$, then it is LAN at $\theta$ (Definition~\ref{definition:lan}) with information matrix $I_\theta = \E[\score_\theta \score_\theta^\top]$. The regularity result~\eqref{eq:normal_expansion_lem} gives \begin{equation*} \sqrt{n} \bar{\Delta}_n = -\frac{1}{\sqrt{n}} \sum_{i = 1}^n \frac{\varphi(Z_i)}{\psi'(0)} + o_{P,n}(1). \end{equation*} The conditions in Corollary~\ref{corollary:normal-expansion}\eqref{item:apply-le-cam} imply that the Fisher information $I_h = \E_h[\score_h(Z)^2]$ exists and is continuous for $\score_h(z) = \frac{p'(z - h)}{p(z - h)}$, and the asymptotic expansion Definitions~\ref{definition:qmd} and~\ref{definition:lan} combined with the preceding display, give the joint convergence \begin{equation*} \left(\sqrt{n} \bar{\Delta}_n, \log \frac{P^n_{h_n/\sqrt{n}}}{P^n_0}(Z_1^n) \right) \cd \normal \left( \left(0,-\frac{h^2}{2} I_0 \right), \begin{pmatrix} \frac{\chi(0)}{\psi'(0)^2} & \frac{-h}{ \psi'(0)} \E_p[\varphi(Z) \score_0(Z)] \\ \frac{-h}{ \psi'(0)} \E_p[\varphi(Z) \score_0(Z)] & h^2 I_0 \end{pmatrix} \right). \end{equation*} Le Cam's third lemma \cite[Exm. 6.7]{VanDerVaart98} then implies the convergence \begin{equation*} \sqrt{n}\bar{\Delta}_n \mathop{\cd}_{P_{h_n/\sqrt{n}}^n} \Ncal\left(\frac{-h}{ \psi'(0)} \E_p[\varphi(Z) \score(Z)], \frac{\chi(0)}{\psi'(0)^2} \right) \end{equation*} under the alternatives $P_{h_n/\sqrt{n}}$, which gives Corollary~\ref{corollary:normal-expansion}\eqref{item:apply-le-cam}. \subsection{Proof of Lemma~\ref{lemma:qmd}} \label{sec:proof-qmd} The proof is essentially completely parallel to that of \cite[Lemma 7.6]{VanDerVaart98}. Define $\dot{s}_\theta = \half \frac{\dot{p}_\theta}{p_\theta} \sqrt{p_\theta}$, which exists $\mu$-almost surely, so that $\int \dot{s}_\theta \dot{s}_\theta^\top d\mu$ is well-defined (though it may be infinite). By Lebesgue's integration theorem, we have \begin{equation*} s_{\theta + h}(x) - s_\theta(x) = \int_0^1 h^\top \dot{s}_{\theta + t h}(x) dt, \end{equation*} and so Jensen's inequality (or Cauchy-Schwarz) we have \begin{equation*} (s_{\theta + h}(x) - s_\theta(x))^2 \le \int_0^1 h^\top \dot{s}_{\theta + t h}(x) \dot{s}_{\theta + t h}(x) ^\top h dt. \end{equation*} Thus for any $h_t$ we have \begin{align*} \int \left(\frac{s_{\theta + t h_t}(x) - s_\theta(x)}{t}\right)^2 d\mu(x) & \le \int \int_0^1 (h_t^\top \dot{s}_{\theta + u t h_t})^2 du d\mu \\ & = \int_0^1 h_t^\top \int \dot{s}_{\theta + u t h_t}\dot{s}_{\theta + u t h_t}^\top d\mu(x) h_t du = \frac{1}{4} h_t^\top \left(\int_0^1 I_{\theta + u t h_t} du\right) h_t. \end{align*} By continuity, as $h_t \to h$ and $t \to 0$ the assumed continuity of $\theta \mapsto I_\theta$ gives that the final display converges to $h^\top I_\theta h$. Now, we note that \begin{equation*} \lim_{t \downarrow 0} \left(\frac{s_{\theta + t h_t}(x) - s_\theta(x)}{t} - h^\top \dot{s}_\theta(x)\right)^2 = 0 \end{equation*} for all $x$ excepting a $\mu$-null set, and the variant dominated convergence theorem in \cite[Prop.~2.29]{VanDerVaart98} implies that \begin{equation*} \lim_{t \to 0} \frac{1}{t^2} \int \left(s_{\theta + t h_t}(x) - s_\theta(x) - t h^\top \dot{s}_\theta(x)\right)^2 d\mu(x) = \lim_{t \to 0} \int \left(\frac{s_{\theta + t h_t}(x) - s_\theta(x)}{t} - h^\top \dot{s}_\theta(x)\right)^2 d\mu(x) = 0, \end{equation*} completing the proof. \section{Fast convergence of uniform estimators under bit constraints} \label{sec:uniform-weirdos} Here we consider the uniform distribution as our location family, demonstrating that in the adaptive setting~\eqref{item:adaptive} or even the one-step adaptive setting~(\ref{item:one-step-adaptive}'), constrained estimators can attain rates faster than the $1 / \sqrt{n}$ rates regular estimands allow. Indeed, define $c(x) = -\log 2$ for $x \in [-1, 1]$ and $c(x) = -\infty$ for $x \not \in [-1, 1]$. Then $f(x) = e^{-c(x)}$ is log-concave and symmetric, and we may consider the location family with densities $f(x - \theta)$. For notational simplicity, we assume we have a sample of size $2n$. We provide a proof sketch that there is a one-step adaptive estimator $\theta_n$ such that \begin{equation} \label{eqn:uniforms-are-easy} \sup_{|\theta| \le \log n} P_\theta\left(|\theta_n - \theta| \ge \frac{8 \log n}{n^{3/4}}\right) \le \frac{2}{n^2}. \end{equation} for all large $n$, and so (by the Borel-Cantelli lemmas), for any $\theta \in \R$ we have $P_\theta(|\theta_n - \theta| \le \sqrt{2 \log n} / n^{3/4} ~ \mbox{eventually}) = 1$. This is of course faster than the $1/\sqrt{n}$ rates we prove throughout. To prove inequality~\eqref{eqn:uniforms-are-easy}, we proceed in two steps, both quite similar. First, we define an initial estimator $\theta\init_n$. Let $\epsilon > 0$, which we will determine presently, though we will take $n \epsilon \to \infty$ as $n \to \infty$, so that we may assume w.l.o.g.\ that $\theta \in [-n\epsilon/2, n \epsilon/2]$. Take the interval $[-n\epsilon, n\epsilon]$, and construct $m$ thresholds at intervals of size $2 n \epsilon / m$; let the $j$th such threshold be \begin{equation*} t_j \defeq -n \epsilon + \frac{2 n (j-1) \epsilon}{m} \end{equation*} Then we ``assign'' observations to each pair of thresholds, so that threshold $j$ corresponds to observations $I_j \defeq \{\frac{n (j - 1)}{m} + 1, \ldots, \frac{n j}{m}\}$, of which there are $n/m$. For each index $i \in I_j$, we set \begin{equation*} B_i = \begin{cases} 1 & \mbox{if~} X_i \ge t_j \\ 0 & \mbox{otherwise}. \end{cases} \end{equation*} Then we simply set $\theta_n\init$ to be the maximal threshold for which $B_i = 0$ for all observations $X_i$ corresponding to that threshold. Let us now consider the probability that $\theta_n\init$ is substantially wrong. For notational simplicity, let $U_i = (1 + X_i - \theta) / 2$, so that the $U_i$ are uniform on $[0, 1]$. First, note that we always have $\theta_n\init \ge \theta - \frac{2 n \epsilon}{m}$, because no observations will be below the appropriate threshold. Let $j\opt$ be the smallest index $j$ for which $\theta \le t_{j\opt}$, and consider the index sets $I_{j\opt}$, $I_{j\opt + 1}$, and so on. The event $\theta\init_n \ge t_{j\opt} + \frac{2 n \epsilon}{m}$ may occur only if for each of the $n/m$ observations in the set $I_{j\opt + 1}$, we have $U_i \ge \frac{n \epsilon}{m}$. Thus, \begin{equation*} P_\theta\left(\theta_n\init \ge t_{j\opt} + \frac{2 n \epsilon}{m} \right) \le \left(1 - \frac{n \epsilon}{m}\right)^\frac{n}{m} \le \exp\left(-\frac{n^2\epsilon}{m^2}\right). \end{equation*} Setting the number of bins $m = \sqrt{n}$, the resolution $\epsilon = 2 \log n / n$, we obtain $P_\theta(\theta_n\init \ge t_{j\opt} + 4 \log n / \sqrt{n}) \le n^{-2}$. Thus we have \begin{equation} \label{eqn:quality-of-initial-estimate} \sup_{|\theta| \le \log n} P_\theta\left(|\theta_n\init - \theta| \ge \frac{8 \log n}{\sqrt{n}}\right) \le \frac{1}{n^2}. \end{equation} The second stage estimator follows roughly the same strategy, except that the resolution of the bins is tighter. In particular, let us assume that $|\theta_n\init - \theta| \le \frac{8 \log n}{\sqrt{n}}$, which happens eventually by inequality~\eqref{eqn:quality-of-initial-estimate}. (We will assume this tacitly for the remainder of the argument.) Consider the interval $\Theta_n \defeq \theta_n\init + [-\frac{16 \log n}{\sqrt{n}}, \frac{16 \log n}{\sqrt{n}}]$ centered at $\theta_n\init$; we know that the interval includes $[\theta - \frac{8 \log n}{\sqrt{n}}, \theta + \frac{8 \log n}{\sqrt{n}}]$. Without loss of generality we assume $\theta_n\init = 0$. Following precisely the same discretization strategy as that for $\theta_n\init$, we divide $\Theta_n$ into $m$ equal intervals, with thresholds $t_j = -\frac{16 \log n}{\sqrt{n}} + \frac{32 (j - 1) \log n}{m \sqrt{n}}$; let $\epsilon_n = \frac{32 \log n}{m \sqrt{n}}$ be the width of these intervals. Then following exactly the same reasoning as above, we assign indices $I_j = \{\frac{n(j - 1)}{m} + 1, \ldots, \frac{n j}{m}\}$ and for $i \in I_j$, set $B_i = 1$ if $X_i \ge t_j$. We define $\theta_n$ to be the maximal threshold $t_j$ for which $B_i = 0$ for all observations $X_i \in I_j$. Then following precisely the reasoning above, we have (on the event that $|\theta_n\init - \theta| \le \frac{8 \log n}{\sqrt{n}}$) \begin{equation*} P_\theta(|\theta_n - \theta| \ge 2 \epsilon_n) \le (1 - \epsilon_n)^\frac{n}{m} \le \exp\left(-\frac{n \epsilon_n}{m}\right) = \exp\left(-\frac{32 \sqrt{n} \log n}{m^2}\right). \end{equation*} Set $m = 4 n^{1/4}$ to obtain the claimed result~\eqref{eqn:uniforms-are-easy}. \subsection{Proof of Proposition~\ref{prop:CEO} \label{app:proof:CEO}} Denote by $D^\star$ the optimal MSE in the Gaussian CEO with $L$ observers and under a total sum-rate $r = r_1 + \ldots +r_L$. An expression for $D^\star$ as a function of $r$ is give as \cite[Eq. 10]{chen2004upper}: \begin{equation} \label{eq:ceo_optimal_sumrate} r = \frac{1}{2} \log^+ \left[ \frac{\sigma_\theta^2}{D^\star} \left( \frac{D^\star L}{ D^\star L - \sigma^2 + D^\star \sigma^2 / \sigma_\theta^2 }\right)^L \right]. \end{equation} For the special case where $r = n$ and $L=n$, we have \begin{equation} \label{eq:ceo_optimal_sumrate2} n = \frac{1}{2} \log_2 \left[ \frac{\sigma_\theta^2}{D^\star} \left(\frac{ D^\star n }{D^\star n - \sigma^2 + D^\star \sigma^2/\sigma_\theta^2 } \right)^n \right]. \end{equation} Consider the distributed encoding setting (iii) in the case where $f(x) = \Ncal(0,\sigma^2)$ and the prior on $\Theta$ is $\pi = \Ncal(0,\sigma_\theta^2)$. The Gaussian CEO problem of \cite{viswanathan1997quadratic} with a unit bitrate $r_1=\ldots = r_n =1$ at each terminal and blocklength $k=1$ reduces to our distributed setting (iii). Since $D^\star$ satisfying \eqref{eq:ceo_optimal_sumrate2} describes the MSE in the CEO setting under an optimal allocation of the sum-rate $r = n$ among $n$ encoders, it provides a lower bound to the minimal MSE in estimating $\theta$ in the distributed setting. By considering the limit $n\rightarrow \infty$ in \eqref{eq:ceo_optimal_sumrate2}, we see that \[ D^\star = \frac{ 4\sigma^2 }{3n + 4 \sigma^2 / \sigma_\theta^2 } + o(n^{-1}) = \frac{4\sigma^2}{3n} + o(n^{-1}). \] This implies Proposition~\ref{prop:CEO}.

https://arxiv.org/abs/1508.06170

Recent advances on Dirac-type problems for hypergraphs

A fundamental question in graph theory is to establish conditions that ensure a graph contains certain spanning subgraphs. Two well-known examples are Tutte's theorem on perfect matchings and Dirac's theorem on Hamilton cycles. Generalizations of Dirac's theorem, and related matching and packing problems for hypergraphs, have received much attention in recent years. New tools such as the absorbing method and regularity method have helped produce many new results, and yet some fundamental problems in the area remain unsolved. We survey recent developments on Dirac-type problems along with the methods involved, and highlight some open problems.

\section{Introduction} Given two (hyper)graphs $F$ and $H$, which conditions guarantee $H$ contains $F$ as a subgraph? When $|V(F)|= |V(H)|$, the decision problem of whether $H$ contains $F$ is often NP-complete, {\em e.g.}, deciding if a graph $H$ contains a Hamilton cycle is a well-known NP-complete problem. Therefore it is natural to look for sufficient conditions for such problems. A classical result of Dirac \cite{Dirac} states that every graph on $n\ge 3$ vertices with minimum degree $n/2$ contains a Hamilton cycle. Problems that relate the minimum degree (in general, minimum $d$-degree in hypergraphs) to the structure of the (hyper)graphs are often referred to as \emph{Dirac-type problems}. The Dirac-type problems for hypergraphs have received much attention in recent years. In this survey we concentrate on three such problems: matching problems (Section~1), packing problems (Section~2), and Hamilton cycles (Section~3). Many problems in this survey were already considered in the survey of R\"odl and Ruci\'nski \cite{RR-survey}. However, since this is a fast-growing area, there are new developments in the last few years and we will emphasize these new advances. Since we only consider Dirac-type problems, we do not discuss matching, packing, or Hamilton cycles in random or quasi-random hypergraphs. We also omit corresponding results in graphs and digraphs. Many results that we omit can be found in other surveys, {\em e.g.}, K\"uhn and Osthus \cite{KuOs-survey, KOsurvey12, KuOs14ICM}, and Gould \cite{Gou03, Gou14}. \section{Matching problems} Given $k\ge 2$, a \emph{$k$-uniform hypergraph} (\emph{$k$-graph}) consists of a vertex set $V$ and an edge set $E$, where each edge is a $k$-element subset ($k$-subset) of $V$. Thus a $2$-graph is simply a graph. In this survey a \emph{hypergraph} refers to a $k$-graph with $k\ge 3$. Given a $k$-graph $H$ with $k\ge 2$, a \emph{matching of size $s$} is a collection of $s$ disjoint edges; a \emph{perfect matching} is a matching that covers the vertex set of $H$ (thus it is necessary that $k$ divides $|V(H)|$). Many open problems in combinatorics can be formulated as a problem of finding perfect matchings in hypergraphs, {\em e.g.}, Ryser's conjecture that every Latin square of odd order has a transversal, and the existence of combinatorial designs (recently solved by Keevash \cite{Ke-design}). A well-known result of Tutte \cite{Tu47} characterized all the graphs with perfect matchings and there are efficient algorithms (e.g., Edmond's algorithm \cite{Edmonds}) that determine if a graph has a perfect matching. However, deciding if a 3-partite 3-graph contains a perfect matching is among the first 21 NP-complete problems given by Karp~\cite{Karp}. Therefore it is natural to look for sufficient conditions that guarantee a perfect matching. \subsection{Degree Conditions for Perfect Matchings} \label{ss:1} There are multiple ways to define degrees in hypergraphs. Given a $k$-graph $H$ with a set $S$ of $d$ vertices, where $0 \leq d \leq k-1$, the \emph{degree} of $S$, denoted by ${\rm deg}_H(S)$ or simply ${\rm deg}(S)$, is the number of edges containing $S$. The \emph{minimum $d$-degree $\delta _{d} (H)$} of $H$ is the minimum of ${\rm deg}(S)$ over all $d$-subsets $S$ of $V(H)$. Hence $\delta_0(H)= e(H)$ is the number of edges in $H$. We refer to $\delta _1 (H)$ as the \emph{minimum vertex degree} of $H$ and $\delta _{k-1} (H)$ as the \emph{minimum codegree} of $H$. The simple monotonicity \[ \frac{\delta_0(H)}{\binom{n}{k}} \ge \frac{\delta_1(H)}{\binom{n-1}{k-1}} \ge \cdots \ge \frac{\delta_{k-1}(H)}{n-k+1} \] suggests that a codegree condition is stronger than other degree conditions. Bollob\'as, Daykin and Erd\H{o}s \cite{BDE} first related the minimum (vertex) degree to the existence of a large (but far from perfect) matching in $k$-graphs. Daykin and H\"aggkvist \cite{DaHa} extended this result by showing that every $k$-graph with $\delta_1(H)\ge (1 - 1/k)\binom{n-1}{k-1}$ contains a perfect matching. Given integers $d<k\le n$ such that $k$ divides $n$, define the \emph{minimum $d$-degree threshold} $\mathbf{m_{d}(k, n)}$ as the smallest integer $m$ such that every $k$-graph $H$ on $n$ vertices with $\delta_d(H)\ge m$ contains a prefect matching. A simple greedy argument shows that $m_1(2, n)=n/2$ for all $n\in 2\mathbb{N}$. Given $k\ge 3$, a result of R\"odl, Ruci\'nski and Szemer\'edi \cite{RRS06} on Hamilton cycles implies that $m_{k-1}(k, n)\le n/2 + o(n)$. K\"uhn and Osthus \cite{KO06mat} sharpened this bound to $m_{k-1}(k, n)\le n/2 + 3k^2 \sqrt{n\log n}$ by reducing the problem to the one for $k$-partite $k$-graphs. R\"odl, Ruci\'nski and Szemer\'edi \cite{RRS06mat} improved it further to $m_{k-1}(k, n)\le n/2 + O(\log n)$ by using the absorbing method. R\"odl, Ruci\'nski, Schacht, and Szemer\'edi \cite{RRSS08} found a simple proof of $m_{k-1}(k, n)\le n/2 + k/4$. Finally R\"odl, Ruci\'nski and Szemer\'edi \cite{RRS09} determined $m_{k-1}(k, n)$ exactly for all $k\ge 3$ and sufficiently large $n$ (again by the absorbing method). In order to state this and later results, let us describe a class of extremal configurations that are usually referred to as \emph{divisibility barriers}. \begin{construction}\cite{TrZh12} \label{cons:md} Define $\mathcal{H}_{\rm ext} (n,k)$ to be the family of all $k$-graphs $H=(V, E)$, in which there is a partition of $V$ into two parts $A, B$ and $i\in \{0, 1\}$ such that $|A|\ne i |V| /k \mod 2$ and $|e\cap A| = i \mod 2$ for all edges $e\in E$. \end{construction} It is easy to see that no hypergraph $H\in \mathcal{H}_{\text{ext}} (n,k)$ contains a perfect matching. Indeed, suppose $H$ contains a perfect matching $M$, then $|A|= \sum_{e\in M} |e\cap A| = i |V|/ k \mod 2$, contradicting the definition of $H$. Define $\delta (n,k, d)$ to be the maximum of the minimum $d$-degrees among all the hypergraphs in $\mathcal H_{\text{ext}} (n,k)$ and note that $m_{d}(k, n)> \delta(n, k, d)$. It is easy to see that \begin{equation} \label{eq:nk1} \delta(n, k, k-1) = \left\{\begin{array}{ll} {n}/{2} - k + 2 & \text{if $k/2$ is even and $n/k$ is odd}\\ {n}/{2} - k + {3}/{2} & \text{if $k$ is odd and $(n-1)/{2}$ is odd}\\ {n}/{2} - k + {1}/{2} & \text{if $k$ is odd and $(n-1)/{2}$ is even} \\ {n}/{2} - k + 1 & \text{otherwise.} \end{array} \right. \end{equation} \begin{theorem}\cite{RRS09} \label{thm:RRS} For $k\ge 3$, $m_{k-1}(k, n)=\delta(n, k, k-1) + 1$ for sufficiently large $n$. \end{theorem} With more case analysis, one may determine $\delta (n,k, k-2)$, {\em e.g.}, it is shown in \cite{TrZh12} that \[ \delta(n, 4, 2) \le \frac{n^2}{4}-\frac{5n}{4} - \frac{\sqrt{n-3}}{2}+\frac{3}{2} \] and equality holds for infinitely many $n$. In general, $\delta (n,k, d)= (1/2+o(1))\binom{n- d}{k- d}$ for any fixed $k>d$ but the general formula of $\delta (n,k, d)$ is unknown -- this is related to the open problem of finding the minima of binary Krawtchouk polynomials. Nevertheless, Treglown and the author \cite{TrZh12, TrZh13} determined $m_{d}(k, n)$ in terms of $\delta(n, k, d)$ for all $d\ge k/2$. \begin{theorem} \cite{TrZh12, TrZh13} \label{thm:TZ12} Let $k\ge 3$ and $d \geq k/2$. Then $m_{d}(k, n)=\delta(n, k, d) + 1$ for sufficiently large $n$. \end{theorem} Previously Pikhurko~\cite{Pik} showed that $m_d(k, n)= (1/2+ o(1))\binom{n-d}{k-d}$ for all $d \geq k/2$. Independently Czygrinow and Kamat \cite{CzKa} determined $m_2(4, n)$ for sufficiently large $n$. Another class of extremal constructions are known as \emph{space barriers}. \begin{construction}\label{cons:ms} Given $s, k, n\in \mathbb{N}$ such that $s\le \lceil n/k \rceil$ ($k$ may not divide $n$), let $H^0_s(n,k)$ be the $k$-graph on $n$ vertices whose vertex set is partitioned into two parts $A$ and $B$ such that $|A|= s-1$, and whose edge set consists of all those edges with at least one vertex in $A$. When $k$ divides $n$, let $H^0(n, k) := H^0_{n/k}(n, k)$. \end{construction} Since each edge contains at least one vertex from $A$ and $|A|< s$, $H^0_s(n,k)$ contains no matching of size $s$; in particular, $H^0(n,k)$ contains no prefect matching. Note that \[ \delta _{d} (H^0(n,k)) = \binom{n-d}{k-d}-\binom{(1-1/k)n -d +1}{k-d}\approx \left(1-\left(\frac{k-1}{k}\right) ^{k-d} \right ) \binom{n-d}{k-d}. \] H\`an, Person and Schacht~\cite{HPS} proved that $m_1(3, n) = (5/9 + o(1))n\approx \delta _{1} (H^0(n,3))$. Khan~\cite{khan1} and independently K\"uhn, Osthus and Treglown~\cite{KOT} obtained that $m_1(3, n) = \delta _{1} (H^0(n,3)) +1$ for sufficiently large $n$. Khan~\cite{khan2} also proved that $m_1(4, n) = \delta _{1} (H^0(n,4)) +1$ for sufficiently large $n$. Alon, Frankl, Huang, R\"odl, Ruci\'nski and Sudakov~\cite{AFHRRS} determined $m_{d}(k, n)$ asymptotically for all $d\ge k-4$, including the new cases when $(k, d)= (5, 1)$, $(5, 2)$, $(6, 2)$, and $(7, 3)$. Very recently Treglown and the author \cite{TrZh15} determined $m_2(5, n)$ and $m_3(7, n)$ exactly for sufficiently large $n$. All these results point to the following conjecture, whose asymptotic version \eqref{eq:md} has appeared earlier, {\em e.g.}, \cite{HPS,KuOs-survey}. \begin{conjecture}\cite{TrZh15} \label{generalconj} Let $k, d \in \mathbb N$ such that $d \leq k-1$. Then for sufficiently large $n\in k\mathbb{N}$, \[ m_{d}(k, n)= \max \left \{ \delta (n,k,d), \ \binom{n-d}{k-d}-\binom{(1-1/k)n -d +1}{k-d}\right \} + 1. \] In particular, \begin{equation}\label{eq:md} m_{d}(k, n)= \left(\max \left \{ \frac12, \ 1-\left(\frac{k-1}{k}\right)^{k-d} \right\} + o(1) \right) \binom{n-d}{k-d}. \end{equation} \end{conjecture} Note that for all $1\leq d \leq k-1$, \[ \left( \frac{k-1}{k} \right) ^{k-d} < \left ( \frac{1}{e} \right)^{1-\frac{d}{k}} \ \ \text{ and } \ \ 1- \left (\frac{k-1}{k} \right) ^{k \ln 2} \rightarrow \frac{1}{2} \ \ \text{ as } \ \ k \rightarrow \infty, \] where $ln$ denotes the natural logarithm. Thus, for $1\ll k\ll n$, if $d$ is significantly bigger than $(1-\ln 2)k \approx 0.307k$, then $\delta(n,k, d) > \binom{n-d}{k-d}-\binom{(1-1/k)n -d +1}{k-d}$. On the other hand, if $d$ is smaller than $(1-\ln 2)k$ then $\delta(n,k, d) < \binom{n-d}{k-d}-\binom{(1-1/k)n -d +1}{k-d}$ for sufficiently large $n$. Other than the aforementioned results, no other asymptotic or exact value of $m_{d}(k, n)$ is known. When $k\ge 3$ and $1\le d< k/2$, H\`an, Person and Schacht~\cite{HPS} gave a general bound: $m_{d}(k, n)\le ((k-d)/k + o(1)) \binom{n-d}{k-d}$. This was improved by Markstr\"om and Ruci\'nski \cite{MaRu} to $m_{d}(k, n)\le ((k-d)/k - 1/k^{k-d}+ o(1)) \binom{n-d}{k-d}$ and by K\"uhn, Osthus and Townsend~\cite{KOTo} to \[ m_{d}(k, n)\le \left( \frac{k-d}k - \frac{k-d-1}{k^{k-d}}+ o(1) \right) \binom{n-d}{k-d}. \] \medskip Let us discuss proof techniques. Most aforementioned results were obtained by the \emph{absorbing method}, initiated by R\"odl, Ruci\'nski, and Szemer\'edi \cite{RRS06}. Roughly speaking, the absorbing method reduces the task of finding a spanning sub(hyper)graph to that of finding a near spanning sub(hyper)graph by using some \emph{absorbing structure}. Given a $k$-graph that contains a matching $M$ and a vertex set $S$ such that $V(M)\cap S= \emptyset$, we say that $M$ \emph{absorbs} $S$ if there is another matching $M'$ with $V(M') = V(M)\cup S$. Suppose we want to prove \eqref{eq:md} for some $d<k$. Let $a= \max \left\{ \frac12, 1 - (\frac{k-1}{k})^{k-d} \right\}$. Let $H$ be a $k$-graph with $\delta_d(H)\ge (a + 2\gamma) \binom{n-d}{k-d}$ for some $\gamma> 0$. We first apply the following absorbing lemma of H\`an, Person and Schacht~\cite[Lemma 2.4]{HPS}. \begin{lemma}\cite{HPS} \label{lem:HPS} For all $\gamma>0$ and positive integers $k> d$ there exists $n_0$ such that the following holds for all $n\ge n_0$. Suppose $H$ is a $k$-graph on $n$ vertices with $\delta_d(H)\ge (\frac12 + 2\gamma) \binom{n-d}{k-d}$, then $H$ contains a matching $M$ of size $\gamma^k n/k$ that can absorb any vertex set $W\subseteq V(H)\setminus V(M)$ with $|W|\in k\mathbb{N}$ and $|W|\le \gamma^{2k} n$. \end{lemma} We next remove $V(M)$ from $H$ and let $H'=H[V(H)\setminus V(M)]$. Then $\delta_d(H)\ge (a + \gamma) \binom{n-d}{k-d}$. If we can show that $H'$ contains a matching that covers all but at most $\gamma^{2k} n$ vertices, then $M$ can absorb these vertices and we obtain the desired perfect matching of $H$. Thus it suffices to prove that \emph{every $k$-graph on $n$ vertices with $\delta_d(H)\ge \left(a + o(1) \right) \binom{n-d}{k-d}$ contains a matching that covers all but $o(n)$ vertices}. We call this assertion the \emph{almost perfect matching lemma}, and note that it is weaker than an asymptotic version of Conjecture~\ref{conj:ms} from Section~\ref{ss:2}. When $d\ge k/2$, this lemma essentially follows from a greedy argument (see {\em e.g.}, \cite[Theorem 1.3]{HPS}), but it appears hard to prove the lemma in general when $d < k/2$. As shown in \cite{AFHRRS, RR-survey}, it suffices to find an almost perfect \emph{fractional matching} instead (see Section~\ref{ss:3} for details). This helps the authors of \cite{AFHRRS} to obtain $m_d(k, n)$ asymptotically for $d\ge k-4$. However, finding an almost perfect fractional matching for smaller $d$ is still an open problem (see Conjecture~\ref{conj:fpm}). Now suppose we want to find $m_d(k, n)$ \emph{exactly}. Naturally we separate the \emph{extremal case} (when $H$ is close to the extremal configuration) from the non-extremal case. The proof for the non-extremal case follows the procedure described above, except that we may assume that $H$ is \emph{not} close to the extremal configuration when proving the absorbing lemma and the almost perfect matching lemma. For example, suppose $1 - (\frac{k-1}{k})^{k-d}< 1/2$, we need the following refinement of Lemma~\ref{lem:HPS}. Let $\varepsilon>0$ and $H$ and $H'$ be two $k$-graphs on the same $n$ vertices. We say that $H$ is \emph{$\varepsilon$-close to $H'$} if $H$ becomes a copy of $H'$ after adding and deleting at most $\varepsilon n^k$ edges. Let $\mathcal B_{n,k}$ ($\overline{\mathcal B}_{n,k}$) denote the $k$-graph whose vertex set can be partitioned into $A, B$ with $|A|=\lfloor n/2 \rfloor$ and $|B|=\lceil n/2 \rceil$ such that all its edges intersect $A$ in an odd (even) number of vertices. \begin{theorem}\cite[Theorem 5]{TrZh15} \label{thm:TZ3} Given any $\varepsilon >0$ and integer $k \geq 2$, there exist $0< \alpha, \xi < \varepsilon$ and $n_0 \in \mathbb N$ such that the following holds. Suppose that $H$ is a $k$-graph on $n \geq n_0$ vertices with $\delta _{1} (H) \geq \left( \frac{1}{2}-\alpha \right) \binom{n - 1}{k- 1}$. Then $H$ is $\varepsilon$-close to $\mathcal B_{n,k}$ or $\overline{\mathcal B}_{n,k}$, or $H$ contains a matching $M$ of size $|M| \le \xi n/k$ that absorbs any set $W\subseteq V(H) \setminus V(M)$ such that $|W| \in k\mathbb{N}$ with $|W| \le \xi^2 n$. \end{theorem} When $1 - (\frac{k-1}{k})^{k-d}< 1/2$, a general extremal case was solved in \cite[Theorem 4.1]{TrZh12}: \emph{for $1\le d< k$, there exists $n_0$ such that every $k$-graph $H$ on $n\ge n_0$ vertices contains a perfect matching if $H$ is $\varepsilon$-close to $\mathcal B_{n,k}$ or $\overline{\mathcal B}_{n,k}$ and $\delta_d(H)\ge \delta(n, k, d)+1$}. \subsection{Smaller matchings} \label{ss:2} In this section we discuss the connection between $\nu(H)$, the size of the largest matching in $H$, and the minimum $d$-degree $\delta_d(H)$. Given $0 \le d< k\le n$ and $s\le n/k$, let $\mathbf{m_d^s(k, n)}$ be the smallest $m$ such that every $k$-graph $H$ on $n$ vertices with $\delta_d(H)\ge m$ contains a matching of size $s$. When $d=k-1$, R\"odl, Ruci\'nski and Szemer\'edi \cite{RRS06, RRS09} noticed a striking contrast between the codegree threshold for perfect matchings (about $n/2$ as seen in Theorem~\ref{thm:RRS}) and the one for almost perfect matchings (about $n/k$ as shown below). We present the simple proof from \cite{RRS09} because it is a good example of a greedy argument. \begin{proposition}\cite[Fact 2.1]{RRS09} \label{prop:RRS} Given integers $n\ge k\ge 2$, every $k$-graph $H$ with $\delta_{k-1}(H)\ge s$ for some $s \le \lfloor n/k \rfloor - k + 2$ contains a matching of size $s$. \end{proposition} \begin{proof} Let $M$ be the largest matching of $H$ and $U= V(H)\setminus V(M)$. Suppose to the contrary, that $|M|< s$ and thus $|V(M)|\le k(\lfloor n/k \rfloor - k + 1)\le n - (k-1)k$. Hence $|U|\ge (k-1)k$. Let $S_1, \dots, S_k$ denote (arbitrary) $k$ disjoint $(k-1)$-subsets of $U$. By the minimum codegree condition, we have $\sum_{i=1}^k {\rm deg}(S_i)\ge ks$. All the neighbors of $S_i$ are in $V(M)$ -- otherwise we can enlarge $M$. Since $|M|< s$, there exists an edge $e\in M$ that contains at least $k+1$ neighbors of $S_1, \dots, S_k$. Consequently there are two vertices $v_1, v_2\in e$ and $i\ne j$ such that $ v_1\in N(S_i)$ and $v_2\in N(S_j)$. By replacing $e$ with $\{v_1\}\cup S_i$ and $\{v_2\}\cup S_j$, we obtain a matching of size $|M|+1$, contradiction. \end{proof} When $k$ does not divide $n$, the largest matching in a $k$-graph on $n$ vertices is of size $\lfloor n/k \rfloor$. Such matching is called a \emph{near perfect} matching. R\"odl, Ruci\'nski and Szemer\'edi \cite{RRS09} proved that every $k$-graph $H$ on $n$ vertices with $\delta_{k-1}(H)\ge n/k + O(\log n)$ contains a near perfect matching and conjectured that $\delta_{k-1}(H)\ge \lfloor n/k \rfloor$ suffices. Using the absorbing method, Han \cite{Han14} recently proved this conjecture. \begin{theorem}\cite{Han14} \label{thm:Han1} Let $n\ge k\ge 2$ be integers such that $k \nmid n$ and $n$ is sufficiently large. Then every $k$-graph $H$ on $n$ vertices with $\delta_{k-1}(H)\ge \lfloor n/k \rfloor$ contains a matching of size $\lfloor n/k \rfloor$. \end{theorem} A corollary of Theorem~\ref{thm:Han1} is $m^s_{k-1}(k, n)\le s$ for \emph{all} $s< n/k$ and sufficiently large $n$, which generalizes Proposition~\ref{prop:RRS}. To see this, let $H=(V, E)$ be a $k$-graph on $n$ vertices with $\delta_{k-1}(H)\ge s$ for some $s< n/k$. By adding a set $T$ of about $\frac{n - ks}{k-1}$ vertices to $V$ and all the $k$-subsets of $V\cup T$ that intersect $T$ to $E$, we obtain a $k$-graph $H'$ on $n'$ vertices with $\delta_1(H)\ge \lfloor n'/k \rfloor$ and $n'\not\in k\mathbb{N}$. Applying Theorem~\ref{thm:Han1}, we obtain a near perfect matching of $H'$. Removing the vertices of $T$ from this matching, we obtain a matching in $H$ of size at least $s$. For any integer $s\le n/k$, the $k$-graph $H^0_s(n, k)$ defined in Construction~\ref{cons:ms} satisfies $\nu(H^0_s(n, k))=s-1$ and $\delta_d(H)= \binom{n-d}{k-d}-\binom{ n-s-d +1}{k-d}$ for all $0\le d\le k-1$. This implies that \begin{equation}\label{eq:H0} m_d^s(k, n) \ge \binom{n-d}{k-d}-\binom{ n-s-d +1}{k-d}+1 \quad \text{for all } s\le n/k. \end{equation} Combining this with the aforementioned corollary of Theorem~\ref{thm:Han1}, we obtain that for all $k\ge 2$ and sufficiently large $n$, \[ m^s_{k-1}(k, n)= s \qquad \text{for all $s< n/k$}. \] This prompts us to conjecture that $H^0_s(n, k)$ provides the correct value of $m_d^s(k, n)$ for all $d<k$ and most $s< n/k$. As noted in Section~\ref{ss:1}, an asymptotic version of this conjecture already implies \eqref{eq:md}, which determines $m_d(k, n)$ asymptotically for all $d<k$. \begin{conjecture}\label{conj:ms} Given $1\le d\le k-2$, there exist $n_0$ and $C$ such that \begin{equation}\label{eq:msd} m_d^s(k, n) = \binom{n-d}{k-d}-\binom{ n-s-d +1}{k-d}+1 \end{equation} for all $n\ge n_0$ and all $s\le n/k -C$. \end{conjecture} When $d=1$, a result of Bollob\'as, Daykin and Erd\H{o}s \cite{BDE} showed that \eqref{eq:msd} holds for all $n> 2k^3 (s+1)$. When $d=1$ and $k=3$, K\"uhn, Osthus and Treglown \cite{KOT} proved \eqref{eq:msd} for all $s\le n/3$ and sufficiently large $n$ (not necessarily divisible by $3$). K\"uhn, Osthus and Townsend~\cite[Conjecture 1.3]{KOTo} proposed an asymptotic version of Conjecture~\ref{conj:ms}, \[ m_d^s(k, n) = \left(1-\left(1 - \frac{s}{n} \right) ^{k-d} + o(1) \right ) \binom{n-d}{k-d}, \] and proved it for all $s\le \min\{ \frac{n}{2(k-d)}, \frac{n - o(n)}{k} \}$. \subsection{Erd\H{o}s Conjecture and fractional matching} \label{ss:3} How many edges of a $k$-graph guarantee a matching of size $s$? This question dates back to an old conjecture of Erd\H{o}s \cite{Erd65mat} that has received much attention lately. \begin{conjecture}\cite{Erd65mat} \label{conj:Erd} Let $s, k, n$ be integers such that $2\le k\le n$ and $1\le s\le n/k$. Then \[ m_0^s(k, n) = \max \left\{ \binom{ks-1}{k}, \binom{n}{k}-\binom{ n-s +1}{k} \right\}+1. \] \end{conjecture} The lower bound comes from \eqref{eq:H0} and the $k$-graph consisting of a complete $k$-graph $K^k_{ks-1}$ on $ks-1$ vertices and $n-ks+1$ isolated vertices. The $s=2$ case of Conjecture~\ref{conj:Erd} is the well-known Erd\H{o}s-Ko-Rado theorem \cite{EKR}. A classic theorem of Erd\H{o}s and Gallai \cite{ErGa59} confirms the conjecture for $k=2$. Erd\H{o}s \cite{Erd65mat} proved the conjecture for $n\ge n_0(k, s)$. Bollob\'as, Daykin and Erd\H{o}s \cite{BDE} proved the conjecture for $n> 2k^3 (s-1)$ and Huang, Loh, and Sudakov \cite{HLS12} recently improved it to $n\ge 3k^2 s$. When $k=3$, Frankl, R\"odl, and Ruci\'nski \cite{FRR12} proved the conjecture for $n\ge 4s$ while {\L}uczak and Mieczkowska \cite{LuMi} proved it for sufficiently large $s$. Recently Frankl \cite{Frankl12} proved the conjecture for $k=3$. Frankl \cite{Frankl13} also proved the conjecture for $s\le n/(2k)$. Alon et al. \cite{AFHRRS} considered a fractional version of the Erd\H{o}s conjecture. Let $H=(V, E)$ be a $k$-graph on $n$ vertices. A \emph{fractional matching} in $H$ is a function $w: E\rightarrow [0,1]$ such that for each $v \in V$ we have $\sum _{e \ni v} w(e)\leq1$. The \emph{size} of $w$, denoted by $\nu^*(H)$, is $\sum _{e \in E} w(e)= \frac1{k}\sum_{v} \sum _{e \ni v} w(e) \le n/k$. If $\nu^*(H) = n/k$ then we call $w$ a \emph{perfect fractional matching}. Determining $\nu^*(H)$ is a linear programming problem. Its dual problem is finding a minimum \emph{fractional vertex cover} $\tau^*(H)= \sum_{v\in V} w(v)$ over all functions $w: V\rightarrow [0,1]$ such that $\sum_{v\in e} w(v) \ge 1$ for all $e\in E$. Correspondingly $\tau(H)$ is the minimum number of vertices in a vertex cover of $H$. Given integers $n\ge k> d\ge 0$ (with $k\ge 2$) and a real number $0< s\le n/k$, define $\mathbf{f^s_d(k, n)}$ to be the smallest integer $m$ such that every $k$-graph $H$ on $n$ vertices with $\delta _d (H) \geq m$ contains a fractional matching of size $s$. We let $f_d(k, n):= f^{n/k}_d(k, n)$ (note that $n/k$ may not be an integer). The hypergraph $H^0_{\lceil s \rceil}(n,k)$ defined in Construction~\ref{cons:ms} contains no fractional matching of size $s$ because \[ \nu^*(H^0_{\lceil s \rceil}(n, k))= \tau^*(H^0_{\lceil s \rceil}(n, k)) \le \tau(H^0_{\lceil s \rceil}(n, k))= {\lceil s \rceil}-1<s. \] When $s$ is an integer, the complete $k$-graph $K^k_{ks-1}$ contains no fractional matching of size $s$. When $s= n/k$, it was shown in \cite{RRS06mat} that $f_{k-1}(k, n) = \delta_{k-1}(H^0_{\lceil n/k \rceil}(n, k)) + 1 =\lceil n/k \rceil$. The following two conjectures were given in \cite{AFHRRS}. \begin{conjecture}\cite{AFHRRS} \label{conj:fpm} \begin{enumerate}[{\rm (i)}] \item For all $1\le d\le k-1$, $\displaystyle f_d(k, n)= \left(1-\left(\frac{k-1}{k}\right)^{k-d} + o(1) \right) \binom{n-d}{k-d}. $ \item For all positive integers $n, k, s$ such that $k\ge 2$ and $s\le n/k$, \[ f^s_0(k, n)= \max \left\{ \binom{ks-1}{k}, \binom{n}{k}-\binom{ n-s +1}{k} \right\}+1. \] \end{enumerate} \end{conjecture} Alon et al. \cite{AFHRRS} first proved Conjecture~\ref{conj:fpm} (ii) asymptotically for $k\in \{3, 4\}$ by reducing it to an old probabilistic conjecture of Samuels \cite{Sam} and then they applied this result to prove (i) with $k-d\in \{3, 4\}$. One can convert an almost perfect fractional matching to an almost perfect integer matching by applying either a result of Frankl and R\"odl \cite{FrRo85} or the weak Regularity Lemma (see \cite[Section 4]{AFHRRS} and \cite[Section 5]{KOTo} for details of these two approaches). This implies the following connection between $m_d(k, n)$ and $f_d(k, n)$. \begin{theorem}\cite{AFHRRS} \label{thm:afh} For $1\le d\le k-1$ if there exists $c>0$ such that $f_d(k, n)=( c+ o(1)) \binom{n-d}{k-d}$ then \begin{equation*} m_{d}(k, n)= \left( \max \left \{ \frac12, \ c \right \} + o(1) \right) \binom{n-d}{k-d}. \end{equation*} \end{theorem} This theorem implies that to determine $m_d(k, n)$, it suffices to prove Conjecture~\ref{conj:fpm} (i). Theorem~\ref{thm:afh} was slightly improved in \cite{TrZh15} to \begin{equation*} m_{d}(k, n)= \max \left \{ \delta (n,k,d)+1, \ (c + o(1)) \binom{n- d}{k- d} \right \}, \end{equation*} where $\delta(n, k, d)$ was defined in Section~\ref{ss:1}. \subsection{Other Matching problems} First we consider matching problems in multipartite hypergraphs. Let $\mathcal{K}_k(n)$ denote the family of all $k$-partite $k$-graphs $H$ with parts $V_1, \cdots, V_k$ such that $|V_1| = \cdots = |V_k|=n$. Given $I\subseteq [k]$, a vertex set $S$ is \emph{$I$-crossing} if $|S\cap V_i|=1$ for $i\in I$ and $S\cap V_i= \emptyset$ otherwise. Let $\delta_I(H)$ be the minimum of ${\rm deg}(S)$ over all $I$-crossing sets $S$ and let $\delta'_d(H) := \min \delta_I(H)$ over all $I\subseteq [k]$ of size $d$. K\"uhn and Osthus \cite{KO06mat} proved that every $H\in \mathcal{K}_k(n)$ contains a perfect matching if $\delta'_{k-1}(H)\ge n/2 + \sqrt{2n \log n}$. Aharoni, Georgakopoulos, and Spr\"ussel \cite{AGS} found an elegant proof of the following theorem. \begin{theorem} \label{thm:AGS}\cite{AGS} Let $H\in \mathcal{K}_k(n)$. If ${\rm deg}(S)> n/2$ for all $[k-1]$-crossing sets $S$ and ${\rm deg}(S)\ge n/2$ for all $\{2, 3, \dots, k\}$-crossing sets $S$, then $H$ contains a perfect matching. \end{theorem} This theorem implies that $H$ contains a perfect matching if $\delta'_{k-1}(H)> n/2$ for \emph{all} values of $n$. This is tight when $k$ is even and $n=2 \pmod 4$ but may not be tight in other cases (off by at most one). Pikhurko \cite{Pik} proved that every $H\in \mathcal{K}_k(n)$ contains a perfect matching if there exists nonempty $L\subsetneq [k]$ such that \[ \frac{\delta_L(H)}{n^{k-|L|}} + \frac{\delta_{[k]\setminus L}(H)}{n^{|L|}}= 1 + \Omega\left(\sqrt{\tfrac{\log n}{n}}\right). \] This implies that $H$ contains a perfect matching if $\delta'_d(H)\ge (1/2 + o(1))n^{k-d}$ for some $d\ge k/2$. Recently Lo and Markstr\"om \cite{LoMa-3p} determined the minimum $\delta'_1(H)$ that guarantees a perfect matching when $H\in \mathcal{K}_3(n)$ and $n$ is sufficiently large. \medskip Keevash and Mycroft \cite{KeMy1} investigated the structure of $k$-graphs satisfying $\delta_{k-1}(H)\ge n/k - o(n)$ but containing no perfect matching. They showed that such $k$-graphs are close to either space or divisibility barriers (see Theorem~\ref{thm:KM} in the next section). Applying the results in \cite{KeMy1}, Keevash, Knox, and Mycroft recently \cite{KKM15} found a necessary and sufficient condition for the existence of a perfect matching in all $k$-graphs $H$ with $\delta_{k-1}(H)\ge n/k + o(n)$. Given $k\ge 3$ and $0\le c\le 1$, let \textbf{PM}$(k, c)$ denote the decision problem of determining whether a $k$-graph $H$ contains a perfect matching when $\delta_{k-1}(H) \ge cn$. The result of Karp \cite{Karp} says that \textbf{PM}$(k,0)$ is NP-complete. On the other hand, Theorem~\ref{thm:RRS} implies that \textbf{PM}$(k, 1/2)$ can be decided in constant time. Szyma\'nska \cite{Szy13} proved that for $c < 1/k$ the problem \textbf{PM}$(k,0)$ admits a polynomial-time reduction to \textbf{PM}$(k, c)$ and hence \textbf{PM}$(k, c)$ is also NP-complete. Karpi\'nski, Ruci\'nski and Szyma\'nska \cite{KRS10} showed that there exists $\varepsilon > 0$ such that \textbf{PM}$(k, 1/2-\varepsilon)$ is in P. Applying the aforementioned structure result, Keevash, Knox, and Mycroft \cite{KKM15} proved that \textbf{PM}$(k, 1/k + \varepsilon)$ is in P for all $\varepsilon>0$ (furthermore, their algorithm provides a perfect matching if it exists). Very recently Han \cite{Han14_Poly} solved the remaining case \textbf{PM}$(k, 1/k)$. In fact, he proved that \textbf{PM}$(k, c)$ is in P for all $c\ge 1/k$ by using some theory developed in \cite{KKM15} and a lattice-based absorbing method. \section{Packing Problems} \subsection{Results} \emph{(Hyper)graph packing}, alternatively called \emph{(hyper)graph tiling}, is a natural extension of the matching problem. Given two $k$-graphs $H$ and $F$, an \emph{$F$-packing} is a sub(hyper)graph of $H$ that consists of vertex-disjoint copies of $F$. An $F$-packing is \emph{perfect} (called an $F$-\emph{factor}) if it is a spanning sub(hyper)graph of $H$. In this case it is necessary that $|V(F)|$ divides $|V(H)|$. Packing problems have been studied extensively for graphs. The celebrated Hajnal-Szemer\'edi theorem \cite{HaSz} states that for all $n\in t\mathbb{N}$ every $n$-vertex graph with minimum degree at least $(1-1/t)n$ contains a $K_t$-factor. Given a $k$-graph $F$ of order $f$ and an integer $n$ divisible by $f$, we define \emph{the $F$-packing threshold} $\mathbf{\delta_d(n,F)}$ as the smallest integer $t$ such that every $n$-vertex $k$-graph $H$ with $\delta_{d}(H)\ge t$ contains an $F$-factor. We simply write $\delta(n, F)$ for $\delta_{k-1}(n, F)$. When $k=2$, K\"{u}hn and Osthus \cite{KuOs09} determined $\delta(n,F)$ up to an additive constant for all graphs $F$. This improves the earlier results of Alon and Yuster \cite{AY96}, Koml\'os \cite{Kom}, and Koml\'os, S\'ark\"ozy, and Szemer\'edi \cite{KSS-AY}. For more results on graph packing, see the survey \cite{KuOs-survey}. It is not surprising that packing problems become harder in hypergraphs. Other than the matching problems mentioned in Section~1, only a few packing thresholds are known. Recall that $K_t^k$ is the complete $k$-graph on $t$ vertices. The first step towards a hypergraph Hajnal-Szemer\'edi theorem is determining $\delta(n, K_4^3)$ ($K_4^3$-packing is also interesting because the corresponding Tur\'an problem is a famous conjecture of Tur\'an \cite{Turan}). Czygrinow and Nagle \cite{CzNa} showed that $\delta(n, K^3_4)\ge 3n/5 + o(n)$. Keevash and Sudakov observed that $\delta(n, K^3_4)\ge 5n/8 + o(n)$. Pikhurko \cite{Pik} proved $3n/4 -2 \le \delta(n, K^3_4)\le 0.861 n$; independently Keevash and the author (unpublished, see \cite{KeMy1}) proved that $2n/3 -1\le \delta(n, K^3_4)\le 4n/5 + o(n)$. Lo and Markstr\"om~\cite{LoMa-fa} showed that $\delta(n, K^3_4)= 3n/4 + o(n) $ by the absorbing method. Independently and simultaneously Keevash and Mycroft \cite{KeMy1} determined $\delta(n, K^3_4)$ \emph{exactly}. \begin{theorem}\cite{KeMy1} For sufficiently large $n\in 4\mathbb{N}$, \[ \delta(n, K_4^3) = \begin{cases} 3n/4 -2 & \text{if $n\in 8\mathbb{N}$} \\ 3n/4 - 1 & \text{otherwise.} \end{cases} \] \end{theorem} We will elaborate on the approaches used in \cite{LoMa-fa, KeMy1} in the next subsection. It is desirable to find the packing threshold $\delta(n, K_t^k)$ for all $3\le k< t$. However, at present such a hypergraph Hajnal-Szemer\'edi theorem seems out of reach. When $t=k+1$, Lo and Markstr\"om~\cite{LoMa-fa} showed that $\delta(n, K_{k+1}^k)\le (1 - 1/2k)n$ for $k\ge 3$. It is plausible that one can prove $\delta(n, K^k_{k+1}) \le \frac{k}{k+1}n + o(n)$ by applying the approach of \cite{KeMy1}. Unfortunately we do not know a matching lower bound (it was shown in \cite{LoMa-fa} that $\delta(n, K_{k+1}^k)\ge 2n/3$ for even $k$). For arbitrary $t$, it was shown in \cite{LoMa-fa} that \[ \left( 1 - \frac{193\log (t-1)}{(t-1)^2}\right)n \le \delta(n, K_t^3) \le \left( 1 - \frac{2}{t^2 - 3t + 4} + o(1) \right)n \] and $\delta(n, K_t^k)\le (1 - \binom{t-1}{k-1}^{-1} + o(1))n$ for $k\ge 6$ and $t\ge (3+\sqrt{5})k/2$. \begin{problem} \begin{enumerate} \item Prove or disprove that $\delta(n, K^k_{k+1}) = \frac{k}{k+1}n + o(n)$ for $k\ge 4$. \item Improve the existing bounds for $\delta(n, F)$ when $F=K_t^3$, $t\ge 5$ and $F=K_t^k$, $4\le k<t$. \end{enumerate} \end{problem} Let $K_4^-$ be the (unique) 3-graph with four vertices and three edges. Lo and Markstr\"om~\cite{LoMa-k4} showed that $\delta(n, K^-_4)= n/2 + o(n) $ by using the absorbing method. All the remaining success was on packing with $k$-partite $k$-graphs. Given positive integers $m_1\le \cdots \le m_k$, let $K^k_{m_1, \dots, m_k}$ denote the complete $k$-partite $k$-graph with parts of sizes $m_1, \dots, m_k$. In particular, let $K^k_k(m)= K^k_{m, \dots, m}$. It is clear that $\delta_d(n, K^k_k(m))\ge m_d(k, n)$ but it is possible to have $\delta_d(n, K^k_{m_1, \dots, m_k})< m_d(k, n)$ for certain $m_1, \dots, m_k$. Other than the matching problems, perhaps the earliest result on hypergraph packing was on $K^3_{1,1,2}-packing$ (note that $K^3_{1,1,2}$ is the unique 3-graph with four vertices and 2 triples). As a corollary of their main result on loose Hamilton cycles, K\"{u}hn and Osthus \cite{KuOs-hc} proved that $\delta(n, K^3_{1,1,2})= n/4+ o(n)$. Recently Czygrinow, DeBiasio, and Nagle \cite{CDN} determined this threshold exactly for sufficiently large $n$. Mycroft \cite{Myc14} recently determined the codegree packing thresholds asymptotically for all complete $k$-partite $k$-graphs, as well as a large class of non-complete $k$-partite $k$-graphs. Given a complete $k$-partite $k$-graph $K= K^k_{m_1, \dots, m_k}$, define $\gcd(K)$ to be $\gcd\{ m_j - m_i: 1\le i<j\le k\}$ and $\sigma(K)= m_1/(m_1 + \dots + m_k)$. We say that $K$ is type 0 if $\gcd(m_1, \dots, m_k)> 1$ or all $m_i =1$; $K$ is type $d\ge 1$ if $\gcd(m_1, \dots, m_k)= 1$ and $\gcd(K)=d$. It was shown in \cite{Myc14} that \begin{equation*} \delta(n, K) = \begin{cases} n/2 + o(n) & \text{if $K$ is type 0;} \\ \sigma(K)n + o(n) & \text{if $K$ is type 1;} \\ \max \{ \sigma(K)n, n/p \} + o(n) & \text{if $K$ is type $d\ge 2$,} \end{cases} \end{equation*} where $p$ is the smallest prime factor of $d$. Mycroft \cite{Myc14} also answered a question of R\"odl and Ruci\'nski \cite{RR-survey} by determining $\delta(n, C^k_s)$ asymptotically for all $k\ge 3$ and $s\ge 2$, where $C^k_s$ is a $k$-uniform loose cycle with $s$ edges (see Section~\ref{sec:hc} for its definition). The proof of \cite{Myc14} makes use of hypergraph regularity, including the Regularity Approximation Lemma of R\"odl and Schacht \cite{RS} and the Blow-up Lemma of Keevash \cite{Ke-bu}. Let us consider hypergraph packing under vertex degree conditions.\footnote{We are not aware of any packing results on $d$-degree conditions for $1<d<k-1$.} Very little is known beyond the matching problem. Lu and Sz\'ekely \cite{LuSz} used the Local Lemma to derive a general upper bound for $\delta_1(n, F)$ for arbitrary $k$-graph $F$. \begin{theorem}\cite{LuSz} \label{thm:LS} Let $F$ be a $t$-vertex $m$-edge $k$-graph in which each edge intersects at most $d$ other edges. Then \begin{equation*} \delta_1(n, F) \le \left( 1 - \frac{1}{e(d+1+ \frac{m}{t}k^2)} \right)\binom{n-1}{k-1}, \quad \text{where } e= 2.718.... \end{equation*} \end{theorem} For example, when $F=K^k_t$, we have $m=\binom{t}{k}$, $d< k\binom{t-1}{k-1}$ and consequently $ \delta_1(n, K_t^k) \le \left(1- \frac{1}{2ek \binom{t-1}{k-1}}\right)\binom{n-1}{k-1}$. On the other hand, we have $\delta_1(n, K_t^k)> \binom{n-1}{k-1} - \binom{(k-1)n/t}{k-1}$ by considering $H^0_{(t- k+1)n/t}(n, k)$ defined in Construction~\ref{cons:ms}. Therefore, when $k$ is fixed and $t\to \infty$, we have \begin{equation}\label{eq:d1K} \delta_1(n, K_t^k) = \left(1- \Theta \left(\frac{1}{t^{k-1}}\right)\right)\binom{n-1}{k-1}. \end{equation} By considering complement $r$-graphs, we can translate \eqref{eq:d1K} to the following corollary on equitable colorings of hypergraphs. An \emph{equitable $\ell$-coloring} of a (hyper)graph is a proper vertex coloring with $\ell$ colors such that the sizes of any two color classes differ by at most one. \begin{corollary}\label{cor} For every $k\ge 3$ there exists $c_k>0$ such that every $k$-graph with maximum vertex degree $d$ has an equitable $\ell$-coloring for any $\ell\ge c_k d^{1/(k-1)}$. \end{corollary} \begin{problem} Improve the constant $c_k$ in Corollary~\ref{cor}, {\em i.e.}, improve the constants hidden in \eqref{eq:d1K}. \end{problem} Lo and Markstr\"om \cite{LoMa-fa} applied the results in \cite{khan1,khan2} to determine $\delta_1(n, K^3_3(m))$ and $\delta_1(n, K^4_4(m))$ asymptotically. Recently Han and the author \cite{HZ3} and independently Czygrinow \cite{Czy14} determined $\delta_1(n, K^3_{1, 1, 2})$ exactly for sufficiently large $n$. More recently Han, Zang, and the author \cite{HZZ_tiling} determined $\delta_1(n, K^3_{a, b, c})$ asymptotically for arbitrary $a\le b\le c$. Given $a\le b\le c$, let $d=\gcd(b-a,c-b)$ and define \begin{equation}\label{absbound} f(a,b,c) := \begin{cases} 1/4, & \text{if $a=1$, $\gcd(a,b,c)=1$ and $d=1$;}\\ 6-4\sqrt{2}\approx 0.343, & \text{if $a\geq 2$, $\gcd(a,b,c)=1$ and $d=1$;}\\ 4/9 , & \text{if $\gcd(a,b,c)=1$ and $d\geq 3$ is odd;}\\ 1/2, & \text{otherwise.} \end{cases} \end{equation} \begin{theorem}\cite{HZZ_tiling} \label{thm:HZZ} \begin{equation*} \delta_{1}(n, K_{a,b,c})= \left(\max\left\{f(a,b,c), 1-\left(\frac {b+c}{a+b+c} \right)^2, \left(\frac {a+b}{a+b+c} \right)^2 \right\} + o(1) \right)\binom n2. \end{equation*} \end{theorem} It is interesting to note that Theorem~\ref{thm:HZZ} contains a case where the coefficient of the packing threshold is \emph{irrational}. In fact, as far as we know, all the previously known tiling thresholds have rational coefficients. The lower bound in Theorem \ref{thm:HZZ} follows from six constructions: three of them are {divisibility barriers}, two are {space barriers}, and the last construction was called a \emph{tiling barrier} because there exists a vertex that is not contained in any copy of $K_{a, b, c}$. In general, given a $k$-graph $F$, let $\tau_d(n, F)$ denote the minimum integer $t$ such that every $k$-graph $H$ of order $n$ with $\delta_d(H) \ge t$ has the property that every vertex of $H$ is covered in some copy of $F$. When $F$ is a graph, it is not hard to see that $\tau_1(n, F) = (1 - 1/(\chi(F) - 1) + o(1))n$ (see the concluding remarks of \cite{HZZ_tiling}). Given a $k$-graph $F$, trivially \begin{equation} \label{eq:tau} \mbox{\rm ex}_d(n, F) < \tau_d(n, F) \le \delta_{d}(n, F), \end{equation} where $\mbox{\rm ex}_d(n, F)$ is the \emph{$d$-degree Tur\'an number} of $F$, defined as the smallest integer $t$ such that every $r$-graph $H$ of order $n$ with $\delta_d(H) \ge t+1$ contains a copy of $F$. It was shown \cite{HZZ_tiling} that $\tau_1(n, K^3_{a,b,c})\le (6-4\sqrt{2}+o(1)) \binom{n}{2}$, and equality holds when $a\ge 2$. It will be interesting to know $\tau_1(n, F)$ for other 3-graphs. \subsection{Methods} Most packing thresholds on graphs, {\em e.g.}, \cite{KSS-AY, KuOs09} were obtained by the regularity method using the Regularity Lemma of Szemer\'edi \cite{Sze} and the Blow-up Lemma of Koml\'os, S\'ark\"ozy, and Szemer\'edi \cite{Blowup}. As the hypergraph versions of these two lemmas are now available, it is possible to attack hypergraph packing problems by the same approach though the proofs become long and technical -- \cite{KeMy1, Myc14} are two examples. Keevash and Mycroft \cite{KeMy1} derived a theorem on perfect matchings for simplicial complexes that is very useful for packing problems. To state this result precisely, we need several definitions. A \emph{$k$-system} $J$ is a set system in which the largest set of $J$ has size $k$ and $\emptyset \in J$. A \emph{$k$-complex} is a {downward closed} $k$-system, namely, every subset of a set in $J$ is also in $J$. Given a $k$-system $J$, let $J_r$ denote the family of $r$-sets in $J$ for $0\le r\le k$. The \emph{minimum $r$-degree} of $J$, denoted by $d_r(J)$, is the minimum ${\rm deg}_{J_{r+1}}(e)$ among all $e\in J_r$. (Note that this is different from $\delta_r(J_{r+1})$, which is the minimum ${\rm deg}_{J_{r+1}}(S)$ among all $r$-sets $S\subseteq V(J)$.) The \emph{degree sequence} of $J$ is $\mathbf{d}(J)= (d_0(J), d_1(J), \dots, d_{k-1}(J))$. Given a vector $\mathbf{a}= (a_0, a_1, \dots, a_{k-1})$, we write $\mathbf{d}(J)\ge \mathbf{a}$ if $d_i(J)\ge a_i$ for $0\le i\le k-1$. Let $H=(V, E)$ be a $k$-graph and let $\mathcal{P}$ be an ordered partition of $V$ into $V_1, \dots, V_t$. The \emph{index vector} $\mathbf{i}_{\mathcal{P}}(S)$ of a set $S\subseteq V$ is defined as $ ( | S\cap V_1|, \dots, |S\cap V_t| )$. A vector of $\mathbb{Z}^t$ is referred to as a \emph{$k$-vector} if all its coordinates are non-negative and sum to $k$ (thus all $\mathbf{i}_{\mathcal{P}}(e)$, $e\in E$, are $k$-vectors). A \emph{lattice} in $ \mathbb{Z}^t$ is an additive subgroup of $ \mathbb{Z}^t$. We let $L_{\mathcal{P}}(H)$ denote the lattice generated by the index vectors $\mathbf{i}_{\mathcal{P}}(e)$ for all edges $e$ of $H$. Given $\mu>0$, let $L_{\mathcal{P}}^{\mu}(H)$ denote the lattice generated by all vectors $\mathbf{x}\in \mathbb{Z}^t$ such that there are at least $\mu n^k$ edges $e$ of $H$ with $\mathbf{i}_{\mathcal{P}}(e) = \mathbf{x}$. A lattice is \emph{complete} if it contains all $k$-vectors; otherwise it is \emph{incomplete}. A lattice is \emph{transferral-free} if it contains no $\mathbf{u}_i - \mathbf{u}_j$ for any $i\ne j$, where $\mathbf{u}_i$ is the 1-vector whose $i$th coordinate is one. \begin{theorem}\cite[Theorem 2.9]{KeMy1} \label{thm:KM} Suppose that $1/n \ll \alpha \ll \mu, \beta \ll 1/k$ and $k$ divides $n$. Let $J$ be a $k$-complex on $n$ vertices with degree sequence \begin{equation} \label{eq:dJ} \mathbf{d}(J) \ge \left(n, (\tfrac{k-1}{k} - \alpha)n, (\tfrac{k-2}{k} - \alpha)n, \dots, (\tfrac{1}{k} - \alpha)n \right). \end{equation} Then at least one of the following properties holds: \begin{description} \item[1] $J_k$ contains a perfect matching. \item[2 (Space barrier)] There exists a set $S\in V(J)$ with $|S|= jn/ k$ for some $1\le j\le k-1$ such that all but at most $\beta n^k$ edges of $J_k$ intersect $S$ with at most $j$ vertices. \item[3 (Divisibility barrier)] There exists a partition $\mathcal{P}$ of $V(J)$ into $t\le k$ parts of size at least $d_{k-1}(J) - \beta n$ such that $L_{\mathcal{P}}^{\mu}(J_k)$ is incomplete and transferral-free. \end{description} \end{theorem} After obtaining this theorem, Keevash and Mycroft \cite{KeMy1} attacked the $K_4^3$-packing problem as follows. Let $H=(V, E)$ be a $3$-graph with $\delta(H)\ge 3n/4$. Consider the so-called \emph{clique $4$-complex} $J$ of $H$, in which $J_i= \binom{V}{i}$ for $i\le 2$, $J_3= H$, and $J_4$ is the family of all $4$-sets that span copies of $K_4^3$. It is easy to see that \[ d_0(J)= n, \quad d_1(J)= n-1, \quad d_2(J)=\delta_2(H)\ge \frac34 n, \quad and \quad d_3(J)\ge \frac{n}4 + 3. \] Indeed, to find $d_3(J)$, we fix a 3-set $abc\in E(H)$; each of the pairs $ab, ac, bc$ has at least $3n/4 -1$ neighbors in $V\setminus \{a, b, c\}$ and thus at least $n/4+3$ vertices $d\in V$ are the neighbors of all $ab, ac, bc$.\footnote{Here we see why we need to consider the 4-complex $J$ instead of the 4-graph $J_4$ alone: $\delta_3(J_4)=0$ because a 3-set $abc\not\in E(H)$ has degree zero in $J_4$.} Next we apply Theorem~\ref{thm:KM} to conclude that either $J_4$ contains a perfect matching or there is a space or divisibility barrier. What remains is to show that $H$ contains a $K_4^3$-factor if it is a space or divisibility barrier satisfying the minimum codegree condition. Most of the aforementioned packing thresholds were obtained by the absorbing method. As described in Section~\ref{ss:1} for the matching problems, our goal is first obtaining a small absorbing $F$-packing that can absorb any smaller set of vertices, and then finding an $F$-packing that covers most of the remaining vertices. Given a $k$-graph $F$ of order $f$, suppose we want to find an $F$-factor in an $n$-vertex $k$-graph $H=(V, E)$ with certain degree conditions. Given $\varepsilon>0$, $i\in \mathbb{N}$, and two vertices $x, y\in V(H)$, we say that $x$ and $y$ are \emph{$(\varepsilon, i)$-reachable} if there are at least $\varepsilon n^{if-1}$ $(if-1)$-subsets of $W\subset V$ such that both $H[\{x\} \cup W]$ and $H[\{y\} \cup W]$ contain $F$-factors. A vertex set $U\subseteq V$ is \emph{$(\varepsilon, i)$-closed} if any two vertices of $U$ are $(\varepsilon, i)$-reachable. In order to find the desired absorbing $F$-packing, it suffices to show that $V$ is $(\varepsilon, i)$-closed for some $\varepsilon>0$ and $i\in \mathbb{N}$ -- see \cite[Lemma 1.1]{LoMa-fa}. Sometimes this is straightforward but sometimes this is done in two steps, referred to as a \emph{lattice-based absorbing method} in \cite{Han14_Poly, HZZ_tiling}. In Step 1 we find a partition $\mathcal{P}= (V_1, \dots, V_t)$ of $V$ such that each $V_i$ is not small and $(\varepsilon, i)$-closed for some $\varepsilon>0$ and $i\in \mathbb{N}$. In Step 2 we show that $L_{\mathcal{P}, F}^{\mu}$ is complete for some $\mu>0$, where $L_{\mathcal{P}, F}^{\mu}$ is the lattice generated by all vectors $\mathbf{v}\in \mathbb{Z}^t$ such that there are at least $\mu n^f$ $f$-sets $S$ that span copies of $F$ and satisfy $\mathbf{i}_{\mathcal{P}}(S)= \mathbf{v}$. Once these steps are done, we can easily show that $V$ is $(\varepsilon', i')$-closed for some $\varepsilon'>0$ and $i'\in \mathbb{N}$. \section{Hamilton Cycles} \label{sec:hc} Cycles in hypergraphs have been studied since the 1970s. There are several notions of cycles. A $k$-graph $(V, E)$ is called a \emph{Berge cycle} if $E$ consists of $k$ distinct edges $e_1, \dots, e_t$ and $V$ contains distinct vertices $v_1, \dots, v_t$ (and possibly other vertices) such that each $e_i$ contains $v_i$ and $v_{i+1}$, where $v_{t +1} = v_1$. Bermond, Germa, Heydemann, and Sotteau \cite{BGHS} proved a Dirac-type theorem for Berge cycles. In recent years a more structured notion of cycles has become more popular. Given $1\le l< k$, a $k$-graph $C$ is a called an \emph{$l$-cycle} if its vertices can be ordered cyclically such that each of its edges consists of $k$ consecutive vertices and every two consecutive edges (in the natural order of the edges) share exactly $l$ vertices. In a $k$-graph, a $(k-1)$-cycle is often called a \emph{tight} cycle while a $1$-cycle is often called a \emph{loose} cycle (sometimes called \emph{linear} cycle). We say that a $k$-graph contains a Hamilton $l$-cycle if it contains an $l$-cycle as a spanning subhypergraph. Note that a $k$-uniform $l$-cycle of order $n$ contains exactly $n/(k - l)$ edges, implying that $k- l$ divides $n$. We define the threshold $\mathbf{h_d^{l}(k, n)}$ as the smallest integer $m$ such that every $k$-graph $H$ on $n$ vertices with $\delta_d(H)\ge m$ contains a Hamilton $l$-cycle, provided that $k- l$ divides $n$. As before, we may omit the subscript when $d=k-1$. Unless stated otherwise, we assume that $n$ is sufficiently large in this section. Katona and Kierstead \cite{KaKi} first studied $h^{k-1}(k, n)$ and proved that \[ \left\lfloor \frac{n-k+3}{2} \right\rfloor \le h^{k-1}(k, n) \le \left( 1 - \frac{1}{2k} \right) n + O(1) \] The following construction provides the lower bound. \begin{construction}\label{cons:KK} Let $H=(V, E)$, where $V=X\cup Y\cup \{v\}$ such that $|X|= \lfloor \frac{n-1}2 \rfloor$ and $|Y|= \lceil \frac{n-1}2 \rceil$, and $E$ consists of all $k$-subsets of $V$ containing $v$ and all $k$-sets $S\subset X\cup Y$ such that $|S\cap X|\ne \lfloor k/2 \rfloor$. \end{construction} It is easy to check that $\delta_{k-1}(H)\ge \lfloor (n-k+1)/2 \rfloor$. Suppose $H$ contains a tight Hamilton cycle. Then there exists an ordering $v_1, \dots, v_{n-1}$ of the vertices of $X\cup Y$ such that all $e_i:= \{v_i, \dots, v_{i+k-1}\}$, $1\le i\le n-k$, are edges. Let $a_i = | X\cap e_i |$. Then $\sum_{i=1}^{n-k} a_i$ is about $|X| k$ when $n$ is sufficiently large because all the vertices of $X$ are counted $k$ times except for those $v_i$ with $i<k$ and $i> n-k$. Thus the average $a_i$ is about $k/2$. On the other hand, we have $a_i\ne \lfloor k/2 \rfloor$ and $|a_i - a_{i+1}|\le 1$ for all $i$. Thus either all $a_i\le \lfloor k/2 \rfloor -1$ or all $a_i\ge \lfloor k/2 \rfloor +1$, a contradiction. Using the absorbing method, R\"odl, Ruci\'nski and Szemer\'edi \cite{RRS06, RRS08} showed that \begin{equation} \label{eq:hk-1} h^{k-1}(k, n) = \frac{n}2 + o(n) \end{equation} for all $k\ge 3$. With long and involved arguments, they \cite{RRS11} were able to obtain an exact result when $k=3$. \begin{theorem}\cite{RRS11} For sufficiently large $n$, $h^{2}(3, n) = \lfloor n/2 \rfloor$. \end{theorem} Assume that $1\le l< k$ and $k-l$ divides $n$. Since every $(k-1)$-cycle of order $n$ contains an $l$-cycle on the same vertices, \eqref{eq:hk-1} implies that $h^{l}(k, n) \le \frac{n}2 + o(n)$. On the other hand, it is not hard to see that $h^{l}(k, n) \ge \frac{n}{2} - k$ when $k-l$ divides $k$. In fact, when $k$ divides $n$, a Hamilton $l$-cycle contains a perfect matching thus Construction~\ref{cons:md} provides this bound; when $k$ does not divide $n$, this was proven by Markstr\"om and Ruci\'nski \cite[Proposition 2]{MaRu}. Consequently, \begin{equation} \label{eq:RRSc} h^{l}(k, n) = \frac{n}2 + o(n) \quad \text{if } k-l \mid k. \end{equation} Very recently Han and the author \cite{HZk2} determined the $d$-degree threshold $h^{k/2}_d(k, n)$ exactly for all even $k\ge 6$ and all $d\ge k/2$. The value of $h_d^{k/2}(k, n)$ turns out to be close (but not always equal) to $\delta(n, k, d)$ defined in Section~\ref{ss:1}. When $k-l$ does not divide $k$, the threshold $h^l(k, n)$ is much smaller. K\"uhn and Osthus \cite{KuOs-hc} proved that $h^{1}(3, n) = n/4 + o(n)$. This was generalized to arbitrary $k$ and $l=1$ by Keevash, K\"uhn, Mycroft, and Osthus \cite{KKMO} and to arbitrary $k$ and arbitrary $l< k/2$ by H\`{a}n and Schacht \cite{HaSc}. Later K\"uhn, Mycroft, and Osthus \cite{KMO} showed that \begin{equation} \label{eq:KOc} h^{l}(k, n) = \frac{n}{\lceil \frac{k}{k-l} \rceil (k-l)} +o(n) \quad \text{if } k-l \nmid k. \end{equation} The following simple construction supports the lower bound in \eqref{eq:KOc}. \begin{construction}\label{cons:KMO} Suppose $1\le l< k$, $k-l \nmid k$, and $k-l \mid n$. Let $t= n/(k-l)$ and $s= \lceil {k}/(k-l) \rceil$. Let $H_0=(V, E)$ be an $n$-vertex $k$-graph in which $V$ is partitioned into sets $A$ and $B$ such that $|A| = \left\lceil t/s \right\rceil - 1$. The edge set $E$ consists of all the $k$-sets that intersect $A$. \end{construction} We have $\delta_{k-1}(H_0)= |A|= \left\lceil {n}/ \left(\lceil \frac{k}{k-l} \rceil (k-l) \right) \right\rceil - 1$. If $H_0$ contains a Hamilton $l$-cycle $C$, then each vertex is contained in at most $s$ edges of $C$. Since $A$ is a vertex cover of $C$, $|C|\le |A| s< t$, a contradiction. Recently Czygrinow and Molla \cite{CzMo} showed that $h^{1}(3, n)= \lceil n/4 \rceil$. Independently Han and the author \cite{HZ2} proved that $h^{l}(k, n)= \lceil n/(2k - 2l) \rceil$ for all $l< k/2$. It was conjectured \cite{HZ2} that Construction~\ref{cons:KMO} is an extremal configuration for Hamilton $l$-cycles whenever $k-l$ does not divide $k$. On the other hand, since the extremal cases in \cite{HZk2, RRS11} require involved work, it seems harder to determine the exact value of $h^{l}(k, n)$ when $k-l$ divides $k$. \begin{problem} Determine $h^{l}(k, n)$ exactly for all $l\ge k/2$, in particular, prove that $h^{l}(k, n) = \left\lceil {n}/ \left(\lceil \frac{k}{k-l} \rceil (k-l) \right) \right\rceil$ when $k-l \nmid k$. \end{problem} Much less is known on the value of $h_d^{l}(k, n)$ when $d\le k-2$. Bu\ss, H\`{a}n, and Schacht \cite{BHS} showed that $h_1^{1}(3, n) = (\frac{7}{16} + o(1)) \binom{n}{2}$. Recently Han and the author \cite{HZ1} improved this to an exact result. \begin{theorem}\cite{HZ1} There exists $n_0$ such that the following holds. If $H$ is a 3-graph $H$ on $n\ge n_0$ vertices with $n\in 2\mathbb{N}$ and $\delta_1(H)\ge \binom{n-1}2 - \binom{\lfloor\frac34 n\rfloor}2 + c$, where $c=2$ if $n\in 4\mathbb N$ and $c=1$ otherwise. Then $H$ contains a loose Hamilton cycle. \end{theorem} It is conjectured in \cite{BHS} that $h_1^{1}(k, n) = \delta_1(H_0) + o(n^{k-1})$ for $H_0$ defined in Construction~\ref{cons:KMO}. \begin{problem} Determine, asymptotically or exactly, the other values of $h_d^{l}(k, n)$, in particular, prove or disprove that $h_1^1(k, n) = \delta_1(H_0) + o(n^{k-1})$ for all $k\ge 4$. \end{problem} It was conjectured in \cite{RR-survey} that $ h_d^{k-1}(k, n) = m_d(k, n) + o(n^{k-d})$ for all $d<k$ and the aforementioned result \cite{RRS09} confirmed this for $d=k-1$. However, Han and the author \cite{HaZh15} recently disproved this conjecture. Nevertheless, it might be true that $h_1^{2}(3, n) \approx m_1(3, n) = (\frac59 + o(1))\binom{n}{2}$. This problem seems (much) harder than the corresponding matching problem. The best known bound $h_1^{2}(3, n) \le \frac13 (5 - \sqrt{5} + o(1)) \binom{n}{2}$ was given by R\"odl and Ruci\'nski \cite{RoRu14} recently. For arbitrary $d<k$, Glebov, Person, and Weps \cite{GPW} proved that $h_d^{k-1}(k, n) \le (1- c_{k, d}) \binom{n-d}{k-d}$ for some small $c_{k, d}>0$. \medskip As with the matching and packing problems, the absorbing method is the main tool of finding a Hamilton $l$-cycle in $k$-graphs. Since \cite[Section 2.2]{RR-survey} gave a detailed sketch on this approach (in the case of $l=d= k-1$), we only highlight the main ideas. In a $k$-graph, an \emph{$l$-path} consists of vertices $v_0, v_1, \dots, v_{t-1}$ for some $t= (k-l)q + l$ and edges $E_0, E_1, \dots, E_{q-1}$, where $E_i= v_{(k-l)i} \dots v_{(k-l)i+k-1}$. We call $v_0 v_1 \dots v_{l-1}$ and $v_{t-l} \dots v_{t-1}$ two \emph{ends} of the path. Given an $l$-path $P$ and a vertex set $S$ such that $V(P)\cap S= \emptyset$, we say that $P$ \emph{absorbs} $S$ if there is an $l$-path on $V(P)\cup S$ with the same ends of $P$. To find a Hamilton $l$-cycle in a $k$-graph $H$, we proceed in the following steps. \begin{description} \item[Step 1] Find an \emph{absorbing $l$-path} $P$ with $|P|\ll n$; namely, $P$ can absorb any vertex set $W\subseteq V(H)\setminus V(P)$ such that $|W|\ll |P|$ and $|W|\in (k-l)\mathbb{N}$. Denote the two ends of $P$ by $L_1$ and $L_2$ and let $V' = (V\setminus V(P))\cup L_1\cup L_2$. \item[Step 2] Find a \emph{reservoir set} $R\subset V'$ of size $|R|\ll |P|$ such that any two $l$-subsets of $V'$ can be connected to an $l$-path via many $s$-subsets of $R$ for some constant $s$ ({\em i.e.}, independent of $n$). Let $V''= V\setminus (V(P)\cup R)$. \item[Step 3] Cover all but $o(n)$ vertices of $V''$ with constant many vertex-disjoint $l$-paths. \end{description} Once these three steps are done, we connect all the $l$-paths that we have found by using the vertices of $R$ and finally absorb the leftover vertices in Step 3 and the remaining vertices in $R$ with $P$. If $\delta_d(H)\ge m_d(k, n) + o(n^{k-d})$, then Step 3 can be easily done by first applying the weak Regularity Lemma and then finding an almost perfect matching in the reduced $k$-graph. (Thus the difficulty of proving $h^2_1(3, n)\le \frac59\binom n2 + o(n^2)$ resides on the first two steps.) On the other hand, when $l< k-1$, a smaller $\delta_d(H)$ may suffice for Step 3. For example, when $2l< k$ and $\delta_{k-1}(H)\ge n/(2k - 2l)$, the authors of \cite{HZ2} accomplished Step 3 by finding an almost $Y_{k, 2l}$-factor in the reduced $k$-graph, where $Y_{k, 2l}$ consists of two $k$-sets that share exactly $2l$ vertices. This is possible because one can convert a copy of $Y_{k, 2l}$ in the reduced $k$-graph to an almost spanning $l$-path in the original $k$-graph, due to the fact that, when $2l< k$, every (long) $l$-path is a $k$-partite $k$-graph whose first $2l$ parts are of size about $m$ and the remaining parts are of size about $2m$ for some integer $m$. \section*{Acknowledgements} The author would like to thank Jie Han for valuable discussion when preparing this manuscript. He also thanks Albert Bush, Jie Han, Allan Lo, Richard Mycroft, and Andrew Treglown for their comments that improved the presentation of the manuscript. \bibliographystyle{abbrv}

https://arxiv.org/abs/2008.03385

Solving two-parameter eigenvalue problems using an alternating method

We present a new approach to compute selected eigenvalues and eigenvectors of the two-parameter eigenvalue problem. Our method requires computing generalized eigenvalue problems of the same size as the matrices of the initial two-parameter eigenvalue problem. The method is applicable for right definite problems, possibly after performing an affine transformation. This includes a class of Helmholtz equations when separation of variables is applied. We provide a convergence proof for extremal eigenvalues and empirical evidence along with a local convergence proof for other eigenvalues.

\section{Introduction}\label{setting} In this work we consider the \emph{two-parameter eigenvalue problem} \begin{equation}\label{2paraev} \begin{aligned} (A_1+\lambda B_1+\mu C_1)u&=0,\\ (A_2+\lambda B_2+\mu C_2)v&=0 \end{aligned} \end{equation} with matrices $A_1,B_1,C_1\in\mathbb{R}^{n\times n}$ and $A_2,B_2,C_2\in\mathbb{R}^{m\times m}$. A solution to this problem is given by possibly complex $(\lambda ,\mu,u,v)$ if they fulfill \eqref{2paraev} and $u,v\neq 0$. We call the pair $(\lambda ,\mu)$ an \emph{eigenvalue of the two-parameter eigenvalue problem} if $u,v\neq 0$ exist such that $(\lambda ,\mu,u,v)$ satisfies \eqref{2paraev} and we call the tensor product $u\otimes v$ an \emph{eigenvector of the two parameter eigenvalue problem} (in the literature the pair $(u,v)$ is often referred to as the eigenvector; the terminology here is simply for convenience). Two-parameter eigenvalue problems have been extensively studied~\cite{Atkinson72,MR973644} and naturally arise in mathematical physics when separation of variables is applied. Consider for example the Helmholtz equation \begin{align*} \Delta u+\lambda u&=0 \quad\text{in } \Omega,\\ u&=0 \quad\text{on } \partial\Omega \end{align*} where $\Omega=\{(x,y)\in \mathbb{R}^2:\left(\frac{x}{a}\right)^2+\left(\frac{y}{b}\right)^2<1, y>0\}$ is half of an open ellipse. Using elliptical coordinates, the problem can be reformulated into \begin{align} \begin{aligned}\label{ex: elliptic} v''(r)+(\lambda c^2\sinh^2(r)+\mu) v(r)&=0, \quad v(0)=0=v(R)\\ w''(\varphi)+ (\lambda c^2\sin^2(\varphi)-\mu )w(\varphi)&=0, \quad w(0)=0=w(\pi) \end{aligned} \end{align} with $u(c\cosh(r)\cos(\varphi),c\sinh(r)\sin(\varphi))=v(r)w(\varphi)$ and $\mu \in \mathbb{R}$ (this will be explained in Section \ref{examples}). This is of the form~\eqref{2paraev} after discretization. There are also many other applications that lead to~\eqref{2paraev}, including delay differential equations \cite{jarlebring2009polynomial} and optimization \cite{Shinsaku16}. There are several numerical methods for solving this problem. The traditional approach~\cite{Atkinson72} solves an $n^2\times n^2$ generalized eigenvalue problem \eqref{matrixev} to find all solutions. Another possibility is Jacobi-Davidson type methods discussed in~\cite{JacobiDavidson02,JacobiDavidson04}. These methods work well in finding eigenvalues close to a given target value. In~\cite{SilArnTwopara} a Sylvester-Arnoldi type method is described that can be used to find a small subset of eigenvalues and uses that Sylvester equations can be solved efficiently. Another possibility is based on homotopy continutation, for example discussed in~\cite{Continuation2001,Homotopy16}. These aim to find all eigenvalues. We present a new algorithm that can be seen as an \emph{alternating method} as they are for example considered in~\cite{Holtz2012} for tensors. A similar approach is used in \cite{Bailey_alternating} and \cite{Xingzhi_alternating}. In these methods they alternatingly solve the first eigenvalue problems for $\lambda $ while fixing $\mu$ and the second eigenvalue problem for $\mu$ while fixing $\lambda $, and repeating the process. Another similar approach was taken in~\cite{Nonlineartwopara}. In there it is described how to transform a two-parameter eigenvalue problem with $n\gg m$ into a nonlinear eigenvalue problem and using techniques for nonlinear eigenvalue problems. By contrast, our method will choose either $\lambda $ or $\mu$ and solve two coupled linear eigenvalue problems alternatingly. This aims to reduce the complexity of finding one solution to~\eqref{2paraev}. The complexity is the one of solving a generalized eigenvalue problem with matrices of size $n\times n$. Hence, the number of operations is $o(n^3)$ for one solution. While we only establish global convergence for extremal eigenpairs, empirically our algorithm also finds all solutions, with $o(n^5)$ cost. It will turn out that this method can find eigenvalues based on their \emph{index}. The index of an eigenvalue of multiparameter eigenvalue problems is a generalisation of ordering real eigenvalues of a standard symmetric eigenvalue problems~\cite[Ch.~1]{MR973644}. We will need the following assumptions to hold: \begin{enumerate}[label=\textbf{A.\arabic*}] \item \label{Ass1} All matrices are symmetric; \item The matrices $C_2$ and $C_1\otimes B_2-B_1\otimes C_2$ are positve definite and $C_1$ is negative definite. \label{Ass2} \end{enumerate} Notice that in example \eqref{ex: elliptic} the second derivative is indeed a symmetric operator, and $\sinh^2(r)>0$ and $\sin^2(\varphi)>0$ almost everywhere. A discrete version of these equations hence leads to a problem of the type \eqref{2paraev} that satisfies the assumptions \ref{Ass1} and \ref{Ass2} up to signs. \begin{remark} We only require the matrices $C_1$, $C_2$, and $C_1\otimes B_2-B_1\otimes C_2$ to be definite, however we chose these specific definiteness assumptions for convenience as to not have a list of all cases in the following theorems. \end{remark} \begin{remark} The assumptions are not restrictive for right definite two parameter eigenvalue problems, that is for two parameter problems with positive or negative definite $C_1\otimes B_2-B_1\otimes C_2$. Indeed we can always find an affine transformation of \eqref{2paraev} satisfying the definiteness assumptions. First we can always find an affine transformation to make one matrix definite. Assume $C_2$ is indefinite. We can then perform the affine transformation \begin{align*} B_1&\leftarrow B_1-\frac{v^\top B_2 v}{v^\top C_2 v} C_1\\ B_2&\leftarrow B_2-\frac{v^\top B_2 v}{v^\top C_2 v} C_2 \end{align*} with $v$ such that $v^\top C_2 v >0$. The new $B_1$ is negative definite, as the positive definiteness of $C_1\otimes B_2-B_1\otimes C_2$ implies that \[ (u^\top C_1u )(v^\top B_2v)-(u^\top B_1u)( v^\top C_2v)>0 \] for any $u$ and $v$. If $C_2$ is definite we can do the following in a similar fashion, and if $C_2$ is semidefinite but not definite, then $C_1$ is definite. We may therefore assume $B_1$ to be negative definite. Now let $u=\argmin \frac{u^\top C_1 u}{u^\top B_1 u}$ be a minimizer. We can perform the affine transformation \begin{align*} C_1&\leftarrow C_1-(\frac{u^\top C_1 u}{u^\top B_1 u}-\epsilon) B_1\\ C_2&\leftarrow C_2-(\frac{u^\top C_1 u}{u^\top B_1 u} -\epsilon)B_2 \end{align*} for a sufficiently small $\epsilon>0$. Now the positive definiteness of $C_1\otimes B_2-B_1\otimes C_2$ and negative definiteness of $B_1$ implies that the new $C_2$ is positive definite and negative definiteness of $B_1$ and minimality of $u$ implies negative definiteness of the new $C_1$. Note that when replacing $B_1,B_2$ and $C_1,C_2$ by the linear combination of matrices \[\tilde B_1=b_1 B_1+b_2 C_1,\quad\tilde B_2=b_1 B_2+b_2 C_2 \] and \[ \tilde C_1=c_1 B_1+c_2 C_1,\quad\tilde C_2=c_1 B_2+c_2 C_2 \] and find the eigenvalue $(\tilde\lambda,\tilde\mu)$, we can recover the original eigenvalue $(\lambda,\mu)$ via \[ \lambda=b_1\tilde\lambda+c_1\tilde\mu ,\quad\mu=b_2\tilde\lambda+c_2\tilde\mu. \] \end{remark} This article is organized as follows: In Section~\ref{Sec:Main}, we first motivate our method and state useful results from multiparameter eigenvalue theory. Afterwards we discuss fixed point properties of the algorithm and prove convergence for extremal eigenvalues. Finally, we examine the time complexity of our method briefly. In Section~\ref{examples}, we show that a class of boundary value problems satisfies our assumptions if properly discretized. Finally, we exhibit results of numerical experiments in Section~\ref{Sec: Numerics}. \section{An alternating algorithm for the two-parameter problem}\label{Sec:Main} In this section, we derive our algorithm and prove local convergence for all eigenvalues and global convergence for extremal eigenvalues. First, let us fix our notation. By $\otimes$ we denote the \emph{tensor product}. For vectors this can be seen as the outer product, i.e., $u\otimes v$ corresponds to the rank-one matrix $uv^\top$. The tensor product of matrices then corresponds to a linear operator acting on matrices, i.e., the product $A\otimes B$ acts on the matrix $X$ by $AXB^\top$. This means that $(A\otimes B) (u\otimes v)=(Au\otimes Bv)$. It can also be seen as the Kronecker product for matrices~\cite[Ch.~12]{golubbook4th}. Then the product of two vectors $u\otimes v$ can be reshaped into a rank-one matrix. \subsection{Derivation of the algorithm}\label{derivation} The two-parameter eigenvalue problem \eqref{2paraev} can be reduced to two generalized eigenvalue problems~\cite{Atkinson72}. For this define the operators acting on $n\times m$ matrices \begin{align} \begin{aligned} M_0=B_1\otimes C_2-C_1 \otimes B_2,\\ M_1=A_1\otimes C_2-C_1 \otimes A_2,\\ M_2=B_1\otimes A_2-A_1 \otimes B_2. \end{aligned} \end{align} A straightforward computation shows that a solution of \eqref{2paraev} satisfies \begin{align*} \begin{aligned} M_1 (u\otimes v)+\lambda M_0(u\otimes v)=0,\\ M_2 (u\otimes v)+\mu M_0(u\otimes v)=0. \end{aligned} \end{align*} Hence, solutions of \eqref{2paraev} are also rank-one solutions $X$ of the system of generalized matrix eigenvalue problems \begin{align} \begin{aligned}\label{matrixev} M_1(X)+\lambda M_0(X)=0,\\ M_2(X)+\mu M_0(X)=0, \end{aligned} \end{align} which are of size $nm\times nm$. Note that the eigenvectors are shared between the two generalized eigenvalue problems. Also under our assumptions solutions of~\eqref{matrixev} lead to solutions of~\eqref{2paraev}. This is a well known consequence of classical theory~\cite{Atkinson72,MR973644}. We provide a self contained proof for completeness. \begin{lemma} Under the assumptions~\ref{Ass1} and~\ref{Ass2}, the eigenvectors of the generalized eigenvalue problem $M_1(X)+\lambda M_0(X)=0$ are either rank one or are a linear combination of rank-one eigenvectors. A rank-one eigenvector is then also an eigenvector of the corresponding two-parameter eigenvalue problem. \end{lemma} \begin{proof} The assumption \ref{Ass2} ensures that the eigenspaces of each subproblem in \eqref{matrixev} is spanned by rank-one matrices. To see this, first note that the operator $M_0$ is negative definite, and $M_1$ and $M_2$ are symmetric. Therefore all eigenvalues of $\eqref{matrixev}$ are real. Second, we can without loss of generality assume $-C_1$ and $C_2$ to be identity matrices, else we just transform the matrices to $\Tilde{A_1}=(-C_1)^{-\frac{1}{2}}A_1(-C_1)^{-\frac{1}{2}}$, $\Tilde{B_1}=(-C_1)^{-\frac{1}{2}}B_1(-C_1)^{-\frac{1}{2}}$, etc., and $\Tilde{u}=(-C_1)^\frac{1}{2}u$, $\Tilde{v}=C_2^\frac{1}{2}v$. Then $M_0$ and $M_1$ are operators in the form of a \emph{Sylvester equation}, i.e., they have the form \begin{align*} M_0=B_1\otimes I_m+I_n \otimes B_2,\\ M_1=A_1\otimes I_m+I_n \otimes A_2. \end{align*} A solution to \eqref{matrixev} then satisfies \begin{align*} \left( (A_1+\lambda B_1)\otimes I_m+I_n\otimes (A_2+\lambda B_2)\right)(X)=0, \end{align*} i.e., $X$ is a zero solution of a Sylvester equation. Such operators possess an orthonormal basis of rank-one eigenvectors, namely $u\otimes v$, where $u$ is an eigenvector of $(A_1+\lambda B_1)$ and $v$ of $(A_2+\lambda B_2)$ \cite[Theorem 4.4.5]{Horn_Topics}, both of which are real symmetric and therefore possess an orthonormal basis of eigenvectors. It follows that $X$ is a sum of rank-one eigenvectors corresponding to the eigenvalue $0$. The dimension equals that of the null space of $M_1(X)+\lambda M_0(X)$, i.e., the geometric multiplicity of $\lambda $. We next show that a rank-one solution to just one of the eigenvalue problems in \eqref{matrixev} suffices to get a solution of the initial problem~\eqref{2paraev}. Indeed, assume that $X=u\otimes v$ solves the second eigenvalue problem in \eqref{matrixev}. We then get \begin{align*} 0=&M_1 (u\otimes v)+\lambda M_0(u\otimes v)\\ = &A_1u\otimes C_2v-C_1u \otimes A_2v+\lambda B_1u\otimes C_2v-\mu C_1u \otimes B_2v\\ =&(A_1+\lambda B_1)u\otimes C_2v- C_1u \otimes(A_2+\lambda B_2)v, \end{align*} which holds true if and only if there is a $\mu$ such that \begin{align*} - \mu C_1u&=(A_1+\lambda B_1)u,\\ - \mu C_2v&=(A_2+\lambda B_2)v. \end{align*} This implies \eqref{2paraev}. \end{proof} We can conclude that \eqref{2paraev} has essentially $n m$ solutions which can be obtained by computing the rank-one solutions of one of the eigenvalue problems in \eqref{matrixev}. In the following, we select the first of these eigenvalue problems. The assumptions imply that the operators $M_0$ and $M_1$ are symmetric and $M_0$ is negative definite. Thus, the solution of \begin{align} M_1(X)+\lambda M_0(X)=0 \label{2meigenvalue} \end{align} with \emph{maximal} eigenvalue $\lambda $ can be obtained by maximizing the \emph{Rayleigh quotient} \begin{align}\label{rquotient} \mathfrak{R}(X)=\frac{\langle X,M_1(X) \rangle}{\langle X,-M_0(X) \rangle}, \end{align} and since the solution is a rank-one matrix, we can just maximize $ \mathfrak{R}(u\otimes v)$ over $u$ and $v$. For convenience, define the functions \begin{align*} a_1(u)&=u^\top A_1 u,&\quad b_1(u)&=u^\top B_1 u, &\quad c_1(u)&=u^\top C_1 u,\\ a_2(v)&=v^\top A_2 v,&\quad b_2(v)&=v^\top B_2 v, &\quad c_2(v)&=v^\top C_2 v. \end{align*} The assumption~\ref{Ass2} assures that $b_2(v)C_1-c_2(v) B_1$ and $ c_1(u) B_2-b_1 (u)C_2$ are positive definite. We can then write \begin{align*} \mathfrak{R}(u\otimes v) =-\frac{u^\top (c_2(v) A_1-a_2(v) C_1)u}{u^\top (c_2(v) B_1-b_2(v) C_1)u}. \end{align*} For fixed $v$ the matrix in the denominator is negative definite. Hence, the maximal value is given by the maximal eigenvalue of the generalized eigenvalue problem \begin{align} \left(c_2(v) A_1-a_2(v) C_1\right)u=\lambda (b_2(v) C_1-c_2(v) B_1)u,\label{ev1} \end{align} and respectively fixing $u$, the maximal value of \begin{align*} \mathfrak{R}(u\otimes v) =-\frac{v^\top (a_1(u)C_2-c_1(u)A_2 )v}{v^\top (b_1(u)C_2 -c_1(u)B_2 )v} \end{align*} is given by the maximal eigenvalue of \begin{align} \left(a_1(u)C_2 -c_1(u)A_2\right)v=\lambda (c_1(u)B_2-b_1(u)C_2 )v.\label{ev2} \end{align} These are generalized eigenvalue problems with matrices of size $n\times n$ and $m\times m$, while~\eqref{2meigenvalue} is a generalized eigenvalue problem of size $nm\times nm$. A similar alternating procedure is of course obtained when minimizing the Rayleigh quotient in \eqref{rquotient}, i.e., when aiming at the smallest eigenvalue of \eqref{2meigenvalue}. More generally, nothing even prevents us from updating $u$ and $v$ with non-extremal eigenpairs of the subproblems \eqref{ev1} and \eqref{ev2} in the hope of finding non-extremal eigenpairs of \eqref{2meigenvalue}. For instance, we can solve \eqref{ev1} for the $i$-th eigenvalue and \eqref{ev2} for the $j$-th eigenvalue, for some fixed $i$ and $j$. This idea motivates Algorithm \ref{algo}. We will call a pair of eigenvectors $u$ and $v$ a \emph{fixed point of Algorithm~\ref{algo}} if it simultaneously solves the eigenvalue problems~\eqref{ev1} and \eqref{ev2}. Then choosing the corresponding index $(i,j)$ the algorithm will not change $u$ and $v$ anymore, provided eigenvalues are simple. \begin{algorithm}[!tb] \caption{Alternating Algorithm for solving two-parameter eigenvalue problems.}\label{algo} \SetKwInOut{Input}{Input} \SetKwInOut{Output}{Output} \Input{Matrices $A_i,B_i,C_i$ for $i=1,2$ satisfying \ref{Ass1} and \ref{Ass2} and index $(i,j)$.} \Output{Eigenvalue $(\lambda ,\mu )$ of index $(i,j)$ with corresponding eigenvector $u\otimes v$.} select random nonzero $u_0\in\mathbb{R}^{n}$\; \For{k=1,2,3,\dots}{ $a_1:=a_1(u_{k-1}),\quad b_1:=b_1(u_{k-1}),\quad c_1:=c_1(u_{k-1})$\; compute the eigenvector $v_k$ corresponding to the $j$-th smallest eigenvalue of the symmetric right definite generalized eigenvalue problem\begin{align*} \left(a_1 C_2 -c_1 A_2\right)v=\lambda ( c_1 B_2-b_1 C_2)v; \end{align*}\\ $a_2:=a_2(v_k);\quad b_2:=b_2(v_k);\quad c_2:=c_2(v_k)$\; compute the eigenpair $(u_k,\lambda_k )$ corresponding to the $i$-th smallest eigenvalue of the symmetric right definite generalized eigenvalue problem\begin{align*} \left(c_2 A_1-a_2 C_1\right)u=\lambda (b_2 C_1-c_2 B_1)u; \end{align*}\\ } $\lambda:=\lambda_k$, \quad $u:=u_k$,\quad $v:=v_k$,\quad $\mu :=-\frac{a_2+\lambda b_2}{c_2}$\; \KwRet{$(\lambda ,\mu) $ and $u\otimes v$.} \end{algorithm} \subsection{Fixed point properties} Our aim in this section is to show that in principle we can find all rank-one eigenpairs of the problem~\eqref{2meigenvalue}, and hence of~\eqref{2paraev}, by finding fixed points of the subproblems~\eqref{ev1} and~\eqref{ev2} solved in Algortihm~\ref{algo} for all possible input indices $(i,j)\in\{1,\dots,n\}\times\{1,\dots,m \}$. The first step in this direction is the following lemma, which shows that a fixed point of the algorithm indeed provides a solution to~\eqref{2meigenvalue} and~\eqref{2paraev}. \begin{lemma}\label{lem: fixed point} Let $u$ and $v$ simultaneously solve \eqref{ev1} and \eqref{ev2}. Under the assumptions {\upshape \ref{Ass1}} and {\upshape \ref{Ass2}} both eigenvalues coincide. Moreover, setting $\mu :=-\frac{a_2(v)+\lambda b_2(v)}{c_2(v)}$, $(u\otimes v,\lambda )$ is an eigenpair of \eqref{2meigenvalue} and $(u,v,\lambda ,\mu )$ is a solution for the two-parameter problem \eqref{2paraev}. \end{lemma} \begin{proof} Denote the eigenvalue in \eqref{ev1} and \eqref{ev2} by $\hat \lambda$ and $\tilde\lambda$, respectively. Then multiplying \eqref{ev1} by $u^\top$ and \eqref{ev2} by $v^\top$, we get \begin{align*} c_2(v)a_1(u)-a_2(v)c_1(u)=-\hat\lambda (c_2(v)b_1(u)-b_2(v)c_1(u)) \end{align*} and \begin{align*} c_2(v)a_1(u)-a_2(v)c_1(u)=-\tilde\lambda (c_2(v)b_1(u)-b_2(v)c_1(u)). \end{align*} Since \ref{Ass2} implies $c_2(v)b_1(u)-b_2(v)c_1(u)<0$, we have $\hat \lambda=\tilde\lambda=:\lambda $. Collecting terms in \eqref{ev1} and \eqref{ev2} gives \begin{align*} c_2(v)A_1u +c_2\lambda B_1u -(a_2+\lambda b_2(v)) C_1 u&=0,\\ -c_1(u) A_2 v -c_1\lambda B_2 v+ (a_1 +\lambda b_1(u)) C_2v&=0. \end{align*} Dividing the first equation by $c_2(v)>0$ and the second by $-c_1(u)>0$, we get \eqref{2paraev} with $\mu =-\frac{a_1(u)+\lambda b_1(u)}{c_1(u)}$, if $$\frac{a_2(v)+\lambda b_2(v) }{c_2(v)}=\frac{a_1(u)+\lambda b_1(u) }{c_1(u)}.$$ This equation is however just a consequence of \begin{align*} c_2(v)a_1(u)-a_2(v)c_1(u)=-\lambda (c_2(v)b_1(u)-b_2(v)c_1(u)). \end{align*} By the considerations in Section \ref{derivation} $(u\otimes v,\lambda )$ is also a solution of \eqref{2meigenvalue}. \end{proof} The previous lemma does not yet let us conclude that all solutions of \eqref{2paraev} occur as fixed points of Algorithm \ref{algo} when varying the input index $(i,j)$. To show this, we need to introduce the notion of the \emph{index of an eigenvalue} $(\lambda ,\mu )$ of the two-parameter eigenvalue problem. \begin{definition}\label{defindex} An eigenvalue $(\lambda ,\mu )$ of the two-parameter eigenvalue problem has the index $(i,j)$ if $0$ is the $i$-th smallest eigenvalue of $A_1+\lambda B_1+ \mu C_1$ and the $j$-th smallest eigenvalue of $A_2+\lambda B_2+ \mu C_2$. \end{definition} If $0$ is a multiple eigenvalue of $A_1+\lambda B_1+ \mu C_1$ or $A_2+\lambda B_2+ \mu C_2$, then the corresponding eigenvalue of the two-parameter eigenvalue problem has multiple indices as well. Under the assumption \ref{Ass1} every real valued eigenvalue of the two-parameter eigenvalue problem has an index since the matrices $A_1+\lambda B_1+ \mu C_1$ and $A_2+\lambda B_2+ \mu C_2$ are symmetric. More important for us is the following result, which immediately follows from~\cite[Theorem 1.4.1]{MR973644}. \begin{theorem}\label{thindex} Under the assumptions {\upshape \ref{Ass1}} and {\upshape \ref{Ass2}} there is a unique eigenvalue $(\lambda , \mu )$ to every index $(i,j)\in\{1,\dots,n\}\times \{1,\dots,m\}$. \end{theorem} The idea is now to show that the index in the sense of Definition \ref{defindex} of the eigenvalue solution provided by a fixed point of Algorithm \ref{algo} coincides with the given input index $(i,j)$. Together with Theorem \ref{thindex} this then implies that all solutions can be obtained this way. For this we need the following version of Sylvester's law of inertia. \begin{lemma}\label{evlemma} Let $A,B$ be symmetric matrices and $I+B$ be positive definite. Then $\lambda _i$ is the $i$-th largest eigenvalue of $A$ if and only if it is the $i$-th largest eigenvalue of the generalized eigenvalue problem \begin{align*} (A+\lambda _i B)u=\lambda (I+B) u. \end{align*} \end{lemma} \begin{proof} Consider the matrices $M_1=A-\lambda _i I$ and $M_2=(I+B)^{-\frac{1}{2}}(A-\lambda _i I)(I+B)^{-\frac{1}{2}}$. The matrix $M_2$ is well defined since $I+B$ is positive definite and since $(I+B)^{-\frac{1}{2}}=((I+B)^{-\frac{1}{2}})^T)$ is invertible, $M_1$ and $M_2$ are congruent and therefore have the same number of positive and respectively negative eigenvalues by Sylvester's law of inertia \cite[Theorem 4.5.8]{MR2978290}. We can rewrite $M_2$ in the following way: \begin{align*} M_2&=(I+B)^{-\frac{1}{2}}(A-\lambda _i I)(I+B)^{-\frac{1}{2}}\\ &=(I+B)^{-\frac{1}{2}}(A+\lambda _i B-\lambda _iB-\lambda _i I)(I+B)^{-\frac{1}{2}}\\ &=(I+B)^{-\frac{1}{2}}(A+\lambda _i B)(I+B)^{-\frac{1}{2}}-\lambda _i I. \end{align*} This implies that $A$ and $(I+B)^{-\frac{1}{2}}(A+\lambda _i B)(I+B)^{-\frac{1}{2}}$ have the same number of eigenvalues that are smaller or respectively larger than $\lambda _i$. Since the eigenvalues of $(I+B)^{-\frac{1}{2}}(A+\lambda _i B)(I+B)^{-\frac{1}{2}}$ are the same as the eigenvalues of the generalized eigenvalue problem \begin{align*} (A+\lambda _i B)u=\lambda (I+B) u \end{align*} the claim is proven. \end{proof} We are now in a position to prove the main result. \begin{theorem}\label{mainth} Let $(i,j)$ be the input index of Algorithm \ref{algo} and let $(u,v,\lambda ,\mu)$ be a fixed point of Algorithm \ref{algo}. Then the output $(\lambda ,\mu )$ is the eigenvalue of problem \eqref{2paraev} with index $(i,j)$ in the sense of Defintion \ref{defindex} and $u\otimes v$ is the corresponding eigenvector. \end{theorem} \begin{proof} Let $(u,v)$ be a fixed point of Algorithm \ref{algo} and let $(\lambda ,\mu )$ be the corresponding eigenvalue, i.e., $\lambda $ is the $i$-th smallest eigenvalue of the generalized eigenvalue problem \begin{align*} \left(c_2(v) A_1-a_2(v) C_1\right)u=\lambda (b_2(v) C_1-c_2(v) B_1)u. \end{align*} The Assumption~\ref{Ass2} guarantees that $b_2(v) C_1-c_2(v) B_1$ is positive definite, we may therefore apply Lemma~\ref{evlemma}. It follows that $\lambda$ is the $i$-th smallest eigenvalue of the matrix \[ c_2(v) A_1-a_2(v) C_1 -\lambda (b_2(v) C_1-c_2(v) B_1)+\lambda I_n=c_2(v)( A_1 +\lambda B_1 + \mu C_1)+\lambda I_n, \] where we substituted $a_2(v)=-\lambda b_2(v)-\mu c_2(v)$. This implies that $0$ is the $i$-th smallest eigenvalue of $A_1 +\lambda B_1 + \mu C_1$ since $c_2(v)>0$ is positive by Assumpotion~\ref{Ass2}. Similarly, $\lambda $ is the $j$-th smallest eigenvalue of the generalized eigenvalue problem \begin{align*} \left(a_1(u) C_2-c_1(u) A_2\right)v=\lambda (c_1(u) B_2-b_1(u) C_2)v. \end{align*} We can again use Lemma~\ref{evlemma} to conclude that $\lambda$ is the $j$-th smallest eigenvalue of the matrix \[ a_1(u) C_2-c_1(u) A_2-\lambda (c_1(u) B_2-b_1(u) C_2)+\lambda I_m=-c_1(u)( A_2+\lambda B_2 + \mu C_2)+\lambda I_m, \] where we substituted $a_1(u)=-\lambda b_1(u)-\mu c_1(u)$. Note that the values for $\mu$ coincide in a fixed point, as was shown in Lemma~\ref{lem: fixed point}. The Assumption~\ref{Ass2} implies that $c_1(u)<0$ and therefore $0$ is the $j$-th smallest eigenvalue of $A_2+\lambda B_2 + \mu C_2$. \end{proof} \begin{remark} This result made use of the definiteness of $C_1$ and $C_2$. If instead $B_1$ and $B_2$ are definite one obtains a similar correspondence of input indices and indices of the eigenvalue when considering an alternating method resulting from the eigenvalue problem $(M_2+\mu M_0)(X)=0$ instead of $(M_1+\lambda M_0)(X)=0$. \end{remark} Theorem~\ref{mainth} implies that if Algorithm \ref{algo} converges, then it computes a solution to the two-parameter eigenvalue problem~\eqref{2paraev}. \begin{corollary}\label{maincorollry} Let $(u_k,v_k,\lambda_k)$ be a sequence generated by Algorithm \ref{algo}. If $u_k$ and $v_k$ converge in the projective sense to $u$ and $v$, i.e., convergence is up to sign flip if we choose normalized $u_k$ and $v_k$ in each step, then $\lambda_k$ converge to $\lambda $ and $u,v,\lambda $ are fixed by Algorithm~\ref{algo}. Therefore the output $(\lambda ,\mu )$ is the eigenvalue of problem \eqref{2paraev} with index $(i,j)$ in the sense of Defintion \ref{defindex} and $u\otimes v$ is the corresponding eigenvector. \end{corollary} \begin{proof} Let $X_{2k}=u_k\otimes v_k$ and $X_{2k+1}=u_{k+1}\otimes v_k$. Since $u_k$ and $v_k$ converge to $u$ and $v$ respectively $x_k$ converges to $u\otimes v$. Define $\hat\lambda_k=\mathfrak{R}(X_{2k+1})$. Notice that $\lambda _k=\mathfrak{R}(X_{2k})$ is the eigenvalue corresponding to~\eqref{ev1} and $\hat\lambda_k$ is the eigenvalue corresponding to~\eqref{ev2} . By continuity of the Rayleigh quotient $\mathfrak{R}$, we get $\lim_{k\to\infty}\lambda _k=\lim_{k\to\infty}\hat\lambda _k=\lambda $. By continuity of the functions $a_1,b_1,c_1,a_2,b_2,c_2$ and continuity of eigenvalues of a nonsingular generalized eigenvalue problem, we get that $u$ and $v$ are eigenvectors of the eigenvalue problems \eqref{ev1} and \eqref{ev2} with eigenvalue $\lambda $ and continuity also ensures that the eigenvalue is still the $i$-th or respectively $j$-th largest. Hence, $(u,v,\lambda ,\mu )$ satisfies the conditions of Theorem~\ref{mainth}. \end{proof} \subsection{Geometric interpretation and rate of convergence} Algorithm~\ref{algo} can be interpreted geometrically. For a given input index $(i,j)$ we look for the intersection of the curves \begin{align*} \gamma_i=\{(\lambda,\mu): \text{ $0$ is the $i$-th largest eigenvalue of }A_1+\lambda B_1 +\mu C_1 \},\\ \zeta_j=\{(\lambda,\mu): \text{ $0$ is the $j$-th largest eigenvalue of }A_2+\lambda B_2 +\mu C_2 \}. \end{align*} As a consequence of~\cite[Theorem~II.6.1]{Kato_Pert} and definiteness of $C_1$ and $C_2$, these curves are continuous and piecewise analytic, and the corresponding eigenvectors are also piecewise analytic. In a fashion similar to~\cite[Lemma~2.2]{Binding_89} we can derive the tangent line of $\gamma_i$ and $\zeta_j$ at an analytic point $(\lambda,\mu)$ with the corresponding eigenvectors. We can choose $\lambda$ as a local variable and obtain \begin{align*} (A_1 +\lambda B_1 +\mu(\lambda) C_1 )u(\lambda)=0 \end{align*} which implies \begin{align*} (A_1 +\lambda B_1 +\mu(\lambda) C_1 )u'(\lambda)=-(B_1 +\mu'(\lambda) C_1 )u(\lambda). \end{align*} Multiplying both equations to with $u(\lambda)^\top$ on the left results in \begin{align} a_1(u(\lambda)) +\lambda b_1(u(\lambda)) +\mu(\lambda) c_1(u(\lambda)) &=0,\nonumber\\ b_1(u(\lambda)) +\mu'(\lambda) c_1 (u(\lambda)&=0.\label{eq: eigencurve derivative} \end{align} The tangent line $T_\gamma$ of $\gamma_i$ at a point $(\lambda_0,\mu(\lambda_0))$ is therefore given by \[ T_\gamma=\{(\lambda,\mu):\mu {c_1(u(\lambda_0))}=-\big({a_1(u(\lambda_0))+\lambda b_1(u(\lambda_0))\big) }\} \] and similarly the tangent line $T_\zeta$ of $\zeta_j$ at a point $(\lambda_0,\mu(\lambda_0))$ is given by \[T_\zeta=\{(\lambda,\mu):\mu{c_2(v(\lambda_0)) =-\big({a_2(v(\lambda_0))+\lambda b_2(v(\lambda_0))\big) }}\}, \] where $v(\lambda_0)$ is the corresponding eigenvector of $A_2+\lambda_0 B_2 +\mu(\lambda_0) C_2 $. Examining Algorithm~\ref{algo}, it can therefore be interpreted as follows: \begin{enumerate} \item Start at a point $(\lambda,\mu)$ of $\gamma_i$ and compute it's tangent line $T_\gamma$; \item compute the intersection point $(\lambda,\mu)$ of $T_\gamma$ and $\zeta_j$ and compute the tangent line $T_\zeta$; \item compute the intersection point $(\lambda,\mu)$ of $T_\zeta$ and $\gamma_i$ and compute the tangent line $T_\gamma$; \item go to step 2. \end{enumerate} We can use this interpretation to estimate the local rate of convergence. Notably, this interpretation is related to Newton's method for computing an intersection point of two functions $f$ and $g$. While Newton's method would compute a new iterate via \[ f(x_k)-g(x_k)+(f'(x_k)-g'(x_k))(x_{k+1}-x_k)=0, \] our method performs two steps via \begin{align} f(x_k)+f'(x_k)(x_{k+1}-x_k)=g(x_{k+1})\quad \text{and}\quad g(x_{k+1})+g'(x_{k+1})(x_{k+2}-x_{k+1})=f(x_{k+2}). \label{quasinewton} \end{align} This allows us to prove local quadratic convergence in a similar fashion to a standard Newton's method. \begin{proposition}\label{prop: local convergence} Let $f,g:U\subset \mathbb{R}\to\mathbb{R}$ be differentiable with Lipschitz continuous derivatives $f'$ and $g'$ and corresponding constants $\alpha$ and $\beta$. Furthermore, let $|f'(x)-g'(x)|\geq\gamma>0$ for all $x$. Then a sequence $x_k$ generated by~\eqref{quasinewton} satisfies \begin{align*} |\Delta x_k|(1-\frac{\beta}{2}|\Delta x_k|) &\leq \frac{\alpha}{2\gamma}|\Delta x_{k-1}|^2,\\ |\Delta x_{k+1}|(1-\frac{\alpha}{2}|\Delta x_{k+1}|) &\leq \frac{\beta}{2\gamma}|\Delta x_{k}|^2, \end{align*} where $\Delta x_k=x_{k+1}-x_k$. \end{proposition} \begin{proof} We start with~\eqref{quasinewton} and with a Taylor expansion for $g$ we get \begin{align*} f(x_k)+f'(x_k)\Delta x_k=g(x_{k+1})=g(x_k)+g'(x_k)\Delta x_k +\int_0^1 \big(g'(x_k+s\Delta x_k)-g'(x_k)\big) \Delta x_k ds. \end{align*} A Taylor expansion of $f$ and $g(x_k)=f(x_{k-1})+f'(x_{k-1})\Delta x_{k-1}$ leads to \[ f(x_k)=g(x_k)+\int_0^1 \big(f'(x_{k-1}+s\Delta x_{k-1})-f'(x_{k-1})\big) \Delta x_{k-1} ds. \] Combining both equations, we get \[ \big(f'(x_k)-g'(x_k)\big)\Delta x_k =\int_0^1 \big(g'(x_k+s\Delta x_k)-g'(x_k)\big) \Delta x_k - \big(f'(x_{k-1}+s\Delta x_{k-1})-f'(x_{k-1})\big) \Delta x_{k-1} ds. \] Using Lipschitz continuity and $|f'(x)-g'(x)|>\gamma$, we arrive at \[ \gamma |x_k|\leq \frac{\beta}{2}|x_k|^2+\frac{\alpha}{2}|x_{k-1}|^2. \] The first inequality follows directly, and the second one can be derived analogously. \end{proof} We can apply this result, when the intersection point of $\gamma_i$ and $\zeta_j$ is an analytic point of both curves. Then we can describe both locally via smooth functions $\mu=f(\lambda)$ for $\gamma_i$ and $\mu=g(\lambda)$ for $\zeta_j$. The difference $|f'-g'|$ can be bounded with~\eqref{eq: eigencurve derivative}. We get $|f'-g'|\geq |\frac{b_1(u)}{c_1(u)}-\frac{b_2(v)}{c_2(v)}|>0$ by the definiteness assumption~\ref{Ass2}. \subsection{Global convergence for extremal eigenpairs} So far, we have shown that if the algorithm converges, it returns the eigenvalue of the two-parameter eigenvalue problem with the desired index. For extremal indices, we can actually prove global convergence. \begin{theorem}\label{convtheorem} The sequences $u_k$ and $v_k$ generated by Algorithm \ref{algo} converge (up to normalization and sign flip) to a solution of~\eqref{2paraev} if the input index is either $(1,1)$ or $(n,m)$ and under the assumption that the respective eigenvalue is simple. \end{theorem} We use the following lemma for proving convergence when each step is at least as good as a line search. \begin{lemma}\label{gradlemma} Let $f:D\to\mathbb{R}$ be continuously differentiable, where $D\subset\mathbb{R}^n$ is open. Let $\{x_k\}_{k\in \mathbb{N}}\subset K\subset D$ and $g_k=\nabla f(x_k)$ satisfy \begin{align*} f(x_{k+1})\leq f(x_k-\sigma g_k) \end{align*} for every $\sigma\in\mathbb{R}$ such that $x_k-\sigma g_k\in D$, where $K$ is compact. Then $g_k$ converges to zero as $k\to\infty$. \end{lemma} \begin{proof}[Proof of Lemma~\ref{gradlemma}] First note that $ f(x_{k+1})\leq f(x_k-\sigma g_k)$ for all $\sigma$ implies that $f(x_k)$ is monotonically nonincreasing and $\{x_k\}_{k\in \mathbb{N}}\subset K$ implies that $f(x_k)$ is bounded from below. Hence, $f(x_k)$ converges to some value $\tilde{f}$ and since $x_k$ lies in a compact set, the gradients $g_k$ are bounded as well. Now towards a contradiction assume that $\|g_k\|\geq 2\epsilon> 0$ for some $\epsilon > 0$ and every $k$. Else either $g_k$ converges to zero or we can choose a subsequence $g_{k'}$ such that $\|g_{k'}\|\geq 2\epsilon> 0$. Since $g_k$ are bounded, there exists $\delta_1>0$ such that $x_k-\sigma g_k\in \tilde{K}\subset D$ for every $\sigma\in [0,\delta_1]$, where $\tilde{K}$ is also compact. Since $f$ is continuously differentiable, $\nabla f$ is uniformly continuous on $\tilde{K}$. Therefore, there is $\delta_2 \in (0, \delta_1]$ such that \begin{align*} \|\nabla f(x_k-\xi g_k)-g_k\|<\epsilon. \end{align*} for all $\xi \in [0,\delta_2]$. By the mean value theorem, Cauchy-Schwarz inequality, and the assumption $\|g_k\|\geq2\epsilon$, we have \begin{align*} f(x_k) - f(x_k-\delta_2 g_k)&=\delta_2\nabla f(x_k-\xi g_k)^\top g_k\\ &\geq\delta_2 \|g_k\|^2-\delta_2\|\nabla f(x_k-\xi g_k) - g_k\|\|g_k\|\\ &\geq\delta_2\|g_k\|^2-\delta_2\|g_k\|\epsilon\\ &\geq \frac{\delta_2}{2} \| g_k \|^2 \geq2\delta_2\epsilon^2. \end{align*} This yields the contradiction \begin{align*} \infty>f(x_0)-\tilde{f}=\sum_{k=0}^\infty f(x_k)-f(x_{k+1})\geq\sum_{k=0}^\infty f(x_k)-f(x_k-\delta_2 g_k)\geq \sum_{k=0}^\infty 2\delta_2\epsilon^2=\infty, \end{align*} therefore $g_k$ converges to zero. \end{proof} \begin{proof}[Proof of Theorem~\ref{convtheorem}] Without loss of generality, assume that $C_1=-I_{n}$ and $C_2=I_{m}$. Finding the vectors $u$ and $v$ for the indices $(1,1)$ and $(n,m)$ corresponds to either minimizing or maximizing the Rayleigh quotient \begin{align*} \mathfrak{R}(X)=-\frac{\langle X,M_1(X) \rangle}{\langle X,M_0(X) \rangle}. \end{align*} Its gradient is given by \begin{align*} \nabla \mathfrak{R}(X)=-\frac{2}{\langle X,M_0(X)\rangle}\left( M_1(X)+\mathfrak{R}(X)M_0(X) \right) \end{align*} and if $X=u\otimes v$, then since $C_1=-I_{n}$ and $C_2=I_{m}$, we have \begin{align*} \nabla \mathfrak{R}(u\otimes v)=-\frac{2}{\langle u\otimes v,M_0( u\otimes v)\rangle}\left( u\otimes (A_2+\mathfrak{R}(u\otimes v)B_2)v+(A_1+\mathfrak{R}(u\otimes v)B_1)u\otimes v \right). \end{align*} This shows that the gradient is an element in the tangent space of the rank-one matrix manifold at $u\otimes v$, which is given by \begin{align*} T_{u\otimes v}\mathcal{M}_1=\{x\otimes v+u\otimes y:x\in \mathbb{R}^{n},y\in \mathbb{R}^{m}\}. \end{align*} Now assume every iterate of $u$ and $v$ are unit vectors. Then $-c_1(u)=1=c_2(v)$ and the Rayleigh quotient reads \begin{align*} \mathfrak{R}(u\otimes v)=\frac{a_1(u)+a_2(v)}{b_1(u)+b_2(v)}. \end{align*} Now let $(\lambda ,v)$ be an eigenpair of \begin{align*} \left(a_1(u)I_{m} +A_2\right)v=-\lambda (b_1(u)I_{m}+ B_2)v, \end{align*} which is one step of Algorithm \ref{algo}. It follows that $\lambda =\mathfrak{R}(u\otimes v)$ and \begin{align*} - \left(A_2 +\lambda B_2\right)v=(a_1(u)+\lambda b_1(u))v. \end{align*} The gradient is then \begin{align*} \nabla \mathfrak{R}(u\otimes v)=\frac{2}{\langle u\otimes v,M_0( u\otimes v)\rangle}\left(-u\otimes (a_1(u)+\lambda b_1(u))v+(A_1+\lambda B_1)u\otimes v \right), \end{align*} and is orthogonal to the subspace $\{u\otimes y:y\in\mathbb{R}^{m}\}\subset T_{u\otimes v}\mathcal{M}_1$ as \begin{align*} \langle \nabla \mathfrak{R}(u\otimes v),u\otimes y\rangle=\frac{2}{\langle u\otimes v,M_0( u\otimes v)\rangle}\left(- (a_1(u)+\lambda b_1(u))v^\top y+(a_1(u)+\lambda b_1(u)) v^\top y \right)=0. \end{align*} This was to be expected, as in each step we optimize the Rayleigh quotient with respect to the subspace $\{u\otimes y:y\in\mathbb{R}^{m}\}$ or respectively $\{x\otimes v:x\in\mathbb{R}^{n}\}$, which makes the gradient orthogonal to that respective space. This, together with $\nabla \mathfrak{R}(u\otimes v)\in T_{u\otimes v}\mathcal{M}_1 $, implies that the gradient lies in $\{u\otimes y:y\in\mathbb{R}^{m}\}$ or respectively $\{x\otimes v:x\in\mathbb{R}^{n}\}$ after each half step. We now only consider the case of minimizing the Rayleigh quotient, as maximizing can be done analogously. As we can choose our iterates $u_k$ and $v_k$ to be normalized, the iterates $X_k=u_k\otimes v_k$ lie in a compact set. As $$v_{k+1}=\argmin_{\|v\|=1} \mathfrak{R}(u_k\otimes v),$$ we have $\mathfrak{R}(X_{k+1})\leq \mathfrak{R}(X_k-\sigma \nabla \mathfrak{R}(X_k))$ since $X_k-\sigma \nabla \mathfrak{R}(X_k)$ is of the form $u_k\otimes v$, and the $\mathfrak{R}$ is scaling invariant. We can thus use Lemma~\ref{gradlemma} and therefore $\nabla \mathfrak{R}(X^k)$ converges to zero, implying that any convergent subsequence of $X_k$ converges to an eigenvector of $M_1(X)+\lambda M_0(X)=0$. Corollary~\ref{maincorollry} implies that this eigenvector corresponds to the eigenvalue of the correct index. Hence, since this eigenvalue is simple, the sequence $X_k$ only has one accumulation point up to sign flip. \end{proof} \subsection{Complexity} We discuss time complexity of Algorithm~\ref{algo}. We assume $m=n$. Computing the eigendecomposition of a generalized eigenvalue problem needs $O(n^{\omega+\gamma})\subset o(n^3)$ operations, where $\omega$ is the exponent of complexity of matrix multiplication and $\gamma>0$~\cite{Demmel_Fast_2007}. This implies that the complexity for computing one eigenvalue of the two-parameter eigenvalue problem is $o(n^3 k)$, where $k$ is the number of iterations in Algorithm~\ref{algo} until a sufficient accuracy is achieved (empirically we observe $k=O(1)$). When computing all $n^2$ eigenvalues of the two-parameter eigenvalue problem, we therefore get a complexity of $o(n^5 k)$ (and with $k=O(1)$ we get a complexity of $o(n^5)$). However, a full eigenvalue decomposition is not always necessary. For extremal eigenvalues, we only need to compute the eigenvector corresponding to the largest or smallest eigenvalue, which is typically possible in $O(n^2)$ operations, for example using Lanczos or LOBPCG~\cite{lobpcg}. Similarly, the $i$-th largest eigenvalue can often be computed in $O(n^2i)$ operations. In summary, computing an eigenvalue with index $(i,j)$ needs $O(n^2k\max(\min(i,n-i+1),\min(j,n-j+1))$ operations, where $k$ is the number of iterations necessary for the desired accuracy. If the matrices allow for fast matrix vector multiplication, for example if they are sparse with $O(n)$ nonzero entries, the complexity can be reduced further up to $O(nk\max(\min(i,n-i+1),\min(j,n-j+1))$ operations. \section{A class of PDE eigenvalue problems}\label{examples} We now present a class of PDE eigenvalue problems, which can be separated into an appropriate two-parameter eigenvalue problem. Let $\phi:(a,b)\times(c,d) \to U\subset \mathbb{R}^2$ be a diffeomorphism, with \begin{align*} g(x,y):=(D\phi(x,y))^\top D\phi(x,y)=\begin{pmatrix} g_1(x)+g_2(y)&0\\ 0&g_1(x)+g_2(y) \end{pmatrix}, \end{align*}and choose $g_1,g_2 > 0$. This is always possible for such a function $g$ as $g(x,y)$ is positive definite and therefore $g_1(x)+g_2(y)> 0$. Now let $\Omega=\phi((a,b)\times(c,d))$. We consider the Helmholtz equation \begin{align*} \Delta u+\lambda u=0 \quad&\text{on } \Omega, \\ u=0 \quad&\text{at } \partial\Omega. \end{align*} We can now write $\Delta u$ in coordinates given by $\phi$. Then \begin{align*} (\Delta u)(\phi(x,y))=&\frac{1}{\sqrt{\det(g(x,y))}}\nabla\cdot\left(\sqrt{\det(g(x,y))}g^{-1}(x,y) \nabla(u(\phi(x,y))\right)\\ =&\frac{1}{g_1(x)+g_2(y)}\Delta(u(\phi(x,y))). \end{align*} Making an ansatz $u(x,y)=v(x)w(y)$, we get \begin{align*} 0&= v''(x)w(y)+v(x)w''(y)+\lambda g_1(x)v(x)w(y)+\lambda g_2(y)v(x)w(y)\\ &=v(x)\big(w''(y) +\lambda g_1(x)w(y)-\mu w(y) \big)+w(y)\big(v''(x)+\lambda g_2(y)v(x)+\mu v(x)\big). \end{align*} Now let $v$ and $w$ satisfy \begin{align}\label{pde2para} v''(x)+\lambda g_1(x)v(x)+\mu v(x)&=0\\ w''(y)+\lambda g_2(y)w(y)-\mu w(y)&=0, \end{align} and $0=v(a)=v(b)=w(c)=w(d)$. Then $u(x,y)=v(x)w(y)$ solves the original eigenvalue problem, and on a given discretrization \eqref{pde2para} satisfies \ref{Ass1} and \ref{Ass2}. \begin{example} The complex function $\phi:\mathbb{C}\to \mathbb{C},z\mapsto z^2$. Then $D\phi(z)=2z$. Therefore with $z=x+iy$, we have \begin{align*} g(x,y)=\begin{pmatrix}2x&2y\\-2y&2x\end{pmatrix}\begin{pmatrix}2x&-2y\\2y&2x\end{pmatrix}=\begin{pmatrix}4x^2+4y^2&0\\0&4x^2+4y^2\end{pmatrix}. \end{align*} \end{example} \begin{example} The complex function $\phi:\mathbb{C}\to \mathbb{C},z\mapsto e^z$. Then $D\phi(z)=e^z$. Therefore with $z=x+iy$, we have \begin{align*} g(x,y)=\begin{pmatrix}e^x \cos(y)&e^x \sin(y)\\-e^x \sin(y)&e^x \cos(y)\end{pmatrix}\begin{pmatrix}e^x \cos(y)&-e^x \sin(y)\\e^x \sin(y)&e^x \cos(y)\end{pmatrix}=\begin{pmatrix}e^{2x}&0\\0&e^{2x}\end{pmatrix}. \end{align*} These coordinates describe an annulus. \end{example} \begin{example} The complex function $\phi:\mathbb{C}\to \mathbb{C},z\mapsto \cosh(z)$. Then $D\phi(z)=\sinh(z)$. Therefore with $z=x+iy$, we have \begin{align*} g(x,y)=&\begin{pmatrix}\sinh(x) \cos(y)&\cosh(x) \sin(y)\\-\cosh(x) \sin(y)&\sinh(x) \cos(y)\end{pmatrix}\begin{pmatrix}\sinh(x) \cos(y)&-\cosh(x) \sin(y)\\\cosh(x) \sin(y)&\sinh(x) \cos(y)\end{pmatrix}\\ =&\begin{pmatrix}\sinh^2(x)+\sin^2(y)&0\\0&\sinh^2(x)+\sin^2(y)\end{pmatrix}. \end{align*} These are elliptical coordinates. \end{example} \section{Numerical Experiments} \label{Sec: Numerics} \begin{figure}[t] \begin{subfigure}{0.48\textwidth} \resizebox{\textwidth}{!}{ \begin{tikzpicture} \begin{semilogyaxis}[ymin=2*10^-11,ymax=5*10^8,grid=major,xlabel=Iterations,ylabel=error ] \addplot[mark=*,,line width=0.8pt ] file[] {data1000gradrand.txt}; \end{semilogyaxis} \end{tikzpicture} } \caption{}\label{grad1000rand} \end{subfigure} \begin{subfigure}{0.48\textwidth} \resizebox{\textwidth}{!}{ \begin{tikzpicture} \begin{semilogyaxis}[ymin=2*10^-11,ymax=5*10^8,grid=major,xlabel=Iterations,ylabel=erro ] \addplot[mark=*,line width=0.8pt ] file[] {data1000grad.txt}; \end{semilogyaxis} \end{tikzpicture} } \caption{}\label{grad1000} \end{subfigure} \caption{Testing Algorithm~\ref{algo} with input index $(1,1)$. Figure~\subref{grad1000rand} depicts results for randomly generated $1000\times 1000$ matrices fulfilling Assumptions~\ref{Ass1} and \ref{Ass2}\ and Figure~\subref{grad1000} depicts results for a discretization of \eqref{ex: elliptic} on {a} $1000\times 1000$ grid. }\label{gradtest} \end{figure} \begin{figure}[t] \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width= \textwidth]{Plots/newrandfig2.pdf} \caption{Two iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width=\textwidth]{Plots/newrandfig3.pdf} \caption{Three iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width= \textwidth]{Plots/newrandfig4.pdf} \caption{Four iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width=\textwidth]{Plots/newrandfig5.pdf} \caption{Five iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width= \textwidth]{Plots/newrandfig6.pdf} \caption{Six iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width=\textwidth]{Plots/newrandfig7.pdf} \caption{Seven iterations} \end{subfigure} \caption{Testing Algorithm \ref{algo} with random $100\times 100$ matrices fulfilling the assumptions. Every picture is taken after solving one eigenvalue problem in line 4 or 6 of the algorithm. The axis label the indices of the computed eigenvalue and the colorscale shows the sum of the absolute values of the $i$-th and $j$-th eigenvalue in~\eqref{2paraev} on a logarithmic scale. }\label{exp: random} \end{figure} In this section, we present results from numerical experiments to test the performance of Algorithm~\ref{algo}. For a measure of the error we use Definition~\ref{defindex}. For a given approximate eigenvalue $(\lambda,\mu)$ we compute the $i$-th smallest eigenvalue of $A_1+\lambda B_1+\mu C_1$ and the $j$-th smallest eigenvalue of $A_2+\lambda B_2+\mu C_2$ and take the sum of the respective absolute values. This quantity is zero if and only if $(\lambda,\mu)$ is an eigenvalue with index $(i,j)$. This is also the sum of the corresponding residual norms with the corresponding normalized eigenvectors. In a first experiment we generated matrices satisfying Assumptions~\ref{Ass1} and~\ref{Ass2} randomly. For the matrices $A_1$ and $A_2$ we generated $n\times n$ and $m\times m$ matrices with independent standard Gaussian distributed entries and took the symmetric part. For the matrices $B_1,B_2,C_1,C_2$ we generated matrices $S_1$ and $S_2$ with Gaussian distributed entries and an $n$ dimensional array $b_1$ with values uniformly distributed between $-0.5$ and $0.5$ and an $m$ dimensional array $b_2$ with values uniformly distributed between $-1.5$ and $-0.5$. We then chose the matrices \[ B_1=S_1^{}\Diag(b_1)S_1^\top,\quad B_2=S_2^{}\Diag(b_2)S_2^\top,\quad C_1=-S_1^{}S_1^\top,\quad C_2=S_2^{}S_2^\top, \] where $\Diag(b_i)$ is the matrix with the entries of $b_i$ on its diagonal. \begin{figure}[t] \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width= \textwidth]{Plots/newellfig2.pdf} \caption{Two iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width=\textwidth]{Plots/newellfig3.pdf} \caption{Three iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width= \textwidth]{Plots/newellfig4.pdf} \caption{Four iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width=\textwidth]{Plots/newellfig5.pdf} \caption{Five iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width= \textwidth]{Plots/newellfig6.pdf} \caption{Six iterations} \end{subfigure} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[width=\textwidth]{Plots/newellfig7.pdf} \caption{Seven iterations} \end{subfigure} \caption{Testing Algorithm \ref{algo} with a discretization of~\eqref{ex: elliptic} on a $100 \times 100$ grid. Every picture is taken after solving one eigenvalue problem in line 4 or 6 of the algorithm. The axis label the indices of the computed eigenvalue and the colorscale shows the sum of the absolute values of the $i$-th and $j$-th eigenvalue in~\eqref{2paraev} on a logarithmic scale. }\label{exp: ellipse} \end{figure} \begin{figure}[t] \begin{subfigure}{0.48\textwidth} \resizebox{\textwidth}{!}{ \begin{tikzpicture} \begin{loglogaxis}[grid=major,xlabel=n,ylabel=time in seconds, legend pos=south east] \addplot[mark=*,color=blue,line width=0.8pt ] file[skip first] {timedata3.txt}; \addlegendentry{alternating method} \addplot[mark=*,color=red,line width=0.8pt ] file[skip first] {timedata2.txt}; \addlegendentry{twopareig} \end{loglogaxis} \end{tikzpicture} } \caption{}\label{subfig: times} \end{subfigure} \begin{subfigure}{0.48\textwidth} \resizebox{\textwidth}{!}{ \begin{tikzpicture} \begin{semilogyaxis}[grid=major,xlabel=n,ylabel=sum of errors, legend pos=south east] \addplot[mark=*,color=blue,line width=0.8pt ] file[skip first] {errsdata.txt}; \addlegendentry{alternating method} \addplot[mark=*,color=red,line width=0.8pt ] file[skip first] {errsdata2.txt}; \addlegendentry{twopareig} \end{semilogyaxis} \end{tikzpicture} } \caption{}\label{subfig: errors} \end{subfigure} \caption{Comparing the computational time~\subref{subfig: times} and precision~\subref{subfig: errors} for randomly generated examples with matrices of size $n\times n$ satisfying the assumption~\ref{Ass1} and~\ref{Ass2} for our method and \texttt{twopareig} from \cite{twopareig}. }\label{timecomp} \end{figure} \begin{figure}[t] \begin{subfigure}{0.48\textwidth} \resizebox{\textwidth}{!}{ \begin{tikzpicture} \begin{loglogaxis}[grid=major,xlabel=n,ylabel=time in seconds, legend pos=south east] \addplot[mark=*,color=blue,line width=0.8pt ] file[skip first] {timedatad.txt}; \addlegendentry{alternating method} \addplot[mark=*,color=red,line width=0.8pt ] file[skip first] {timedatad2.txt}; \addlegendentry{twopareig} \end{loglogaxis} \end{tikzpicture} } \caption{}\label{subfig: timesdiag} \end{subfigure} \begin{subfigure}{0.48\textwidth} \resizebox{\textwidth}{!}{ \begin{tikzpicture} \begin{semilogyaxis}[grid=major,xlabel=n,ylabel=sum of errors, legend pos=south east] \addplot[mark=*,color=blue,line width=0.8pt ] file[skip first] {errsdatad.txt}; \addlegendentry{alternating method} \addplot[mark=*,color=red,line width=0.8pt ] file[skip first] {errsdatad2.txt}; \addlegendentry{twopareig} \end{semilogyaxis} \end{tikzpicture} } \caption{}\label{subfig: errorsdiag} \end{subfigure} \caption{Comparing the computational time~\subref{subfig: timesdiag} and precision~\subref{subfig: errorsdiag} for randomly generated examples with matrices of size $n\times n$ satisfying the assumption~\ref{Ass1} and~\ref{Ass2} for our method and \texttt{twopareig} from \cite{twopareig}. For this example the matrices $B_1,B_2,C_1,C_2$ are diagonal matrices.}\label{timecompdiag} \end{figure} As a second example, we discretized the Helmholtz equation on half of an ellipse as in the example in Section \ref{setting}, i.e., \begin{align}\label{ellipseforexperiment} \begin{aligned} v''(r)+(\lambda c^2\sinh^2(r)+\mu) v(r)&=0, \quad v(0)=0=v(1)\\ w''(\varphi)+ (\lambda c^2\sin^2(\varphi)-\mu )w(\varphi)&=0, \quad w(0)=0=w(\pi). \end{aligned} \end{align} For Figure~\ref{gradtest}, we generated $1000\times 1000$ matrices randomly in~\subref{grad1000rand} and used a discretization on a $1000\times 1000$ grid in~\subref{grad1000} as described above and used Algorithm~\ref{algo} to find an eigenvalue with minimal $\lambda$, i.e., with input index $(1,1)$. After solving 6 and 7 eigenvalue problems respectively as described in lines 4 and 6 of Algorithm~\ref{algo}, we found an eigenvalue with an error of approximately $10^{-9}$ and $10^{-10}$ respectively, confirming the result of Theorem~\ref{convtheorem}. For Figure~\ref{exp: random} we generated $100\times 100$ matrices satisfying Assumptions~\ref{Ass1} and~\ref{Ass2} as described above and use Algorithm~\ref{algo} to find every eigenvalue, i.e., for every input index $(i,j)\in\{1,\dots,100\}\times\{1,\dots,100\}$. The axis describe the input index of Algorithm~\ref{algo} and the greyscale describes the base ten logarithm of the error described above. An iteration is computing the solution of one eigenvalue problem in lines 4 and 6 of Algorithm~\ref{algo}. After 7 iterations the highest error was $4\cdot 10^{-8}$ and the smallest errors were in the range of machine precision. In Figure~\ref{exp: ellipse} we used a discretization of the Helmholtz equation~\eqref{ellipseforexperiment} on a $100\times 100$ grid and repeated the experiment of Figure~\ref{exp: random} for the resulting matrices. Again, after 7 iterations the highest error was $3\cdot 10^{-8}$ and the smallest errors were in range of machine precision. These experiments suggest that Algorithm~\ref{algo} can indeed be used to find every eigenvalue of right definite two-parameter eigenvalue problems. In Figure~\ref{timecomp}\subref{subfig: times} we compared the time of both Algorithm~\ref{algo} and the algorithm {\verb|twopareig|}\cite{twopareig} for computing every eigenvalue of a two-parameter eigenvalue problem of varying sizes $n=m$. We generated matrices satisfying the assumptions as above. Notice that Algorithm~\ref{algo} can be run in parallel for different input indices to further improve efficiency. The experiment was run on a Intel Xeon Gold 6144 at 3.5 GHz with 384 GB RAM. We used 8 cores with 16 threads. We chose to solve 10 small eigenvalue problems in Algorithm~\ref{algo}, corresponding to $k=1,\dots,5$ in line 2. For larger $n$ our method was indeed faster, and the asymptotic slope of the time -$n$ graph on a loglog scale is smaller, indicating that the asymptotic computational cost is lower. In Figure~\ref{timecomp}\subref{subfig: errors} we measure the sum of the $n^2$ errors as described above. We observe that our method computed eigenvalues with higher accuracy. Finally, we repeated the last experiment, but we chose $S_1=I_n=S_2$. This made the matrices $B_1,B_2,C_1$ and $C_2$ diagonal, which effectively transformed the generalized eigenvalue problems in Algorithm~\ref{algo} into ordinary eigenvalue problems to further improve efficiency. When solving for every eigenvalue of the two-parameter problem, we can perform the left and right actions of $(-C_1)^{-\frac{1}{2}}$ and $C_2^{-\frac{1}{2}}$ respectively and afterwards diagonalize $B_1$ and $B_2$. This justifies this experiment. The results are depicted in Figure~\ref{timecompdiag}. In Figure~\ref{timecompdiag}\subref{subfig: timesdiag} we see that the alternating method is faster than the method {\verb|twopareig|}\cite{twopareig} for even smaller $n$. Again our method is more accurate. \section{Conclusion and outlook} We presented a new method for computing eigenvalues of the two-parameter eigenvalue problem. Our approach only requires solving generalized eigenvalue problems of the size of the matrices of the two-parameter problem and can therefore reduce the complexity compared to conventional methods. Our method also uses a search for the eigenvalues by index, which makes it possible to find successive eigenvalues of the two-parameter eigenvalue problem without deflation. So far the technique of our proof only established global convergence for extremal eigenvalues and under definiteness assumptions. The numerical experiments however indicated convergence for every eigenvalue. Proposition~\ref{prop: local convergence} gives insight into the local convergence, but a global convergence proof remains an open problem. Although there are many classes of two-parameter eigenvalue problems that satisfy the assumptions (eventually after performing an affine transformation), many other interesting applications do not. Therefore it would be important to investigate if a generalization to non-singular problems, i.e., the operator $M_0$ is invertible, as in~\cite{JacobiDavidson04} or even to the general case as in~\cite{Hochstenbach2019} is possible. This paper relies heavily on the assumptions \ref{Ass1} and \ref{Ass2}. A natural question is to relax these conditions, in particular the definiteness condition \ref{Ass2}. Indeed under weaker assumptions (such as when the matrices are \emph{almost} definite, with a few eigenvalues of the opposite sign), using inertia laws~\cite{nakatsukasa2019inertia} one can show that the generalized eigenvalue problem~\eqref{matrixev}, and hence also~\eqref{2paraev}, has many real eigenvalues. It would be of interest to investigate the applicability of the results here in such situations. Finally, another interesting generalization that could be considered is to multiparameter eigenvalue problems with more than 2 parameters. The eigenvectors then form rank-one tensors similar to \eqref{matrixev} \cite{Atkinson72}. Again an alternating approach as in \cite{Holtz2012} can be used, however our proof technique will not work and in practice the generalization will not easily assure convergence even for extremal indices. A similar approach using the Tensor-Train format is used in~\cite{TTmulti}. \bibliographystyle{plain}

https://arxiv.org/abs/1311.3644

Inversion and subspaces of a finite field

Let $A$ and $B$ two $F_q$-subspaces of a finite field, of the same size, and let $A^{-1}$ denote the set of inverses of the nonzero elements of $A$. Mattarei proved that $A^{-1}$ can only be contained in $A$ if either $A$ is a subfield, or $A$ is the set of trace zero elements in a quadratic extension of a field. Csajbók refined this to the following quantitative statement: if $A^{-1}\not\subseteq B$, then the bound $|A^{-1}\cap B|\le 2|B|/q-2$ holds. He also gave examples showing that his bound is sharp for $|B|\le q^3$. Our main result is a proof of the stronger bound $|A^{-1}\cap B|\le |B|/q\cdot\bigl(1+O_d(q^{-1/2})\bigr)$, for $|B|=q^d$ with $d>3$. We also classify all examples with $|B|\le q^3$ which attain equality in Csajbók's bound.

\section{Introduction} In response to a question of Andrea Caranti, for use in~\cite{CDVS:AES}, the author determined in~\cite{Mat:inverse-closed} the additive subgroups of a field which are closed with respect to inverting nonzero elements. The more general question with a division ring instead of a field was independently answered in~\cite{GGSZ}. The proofs depend on Hua's identity~\cite{Hua:sfield_properties}, and on Jordan algebra techniques to cover the noncommutative case. However, a more direct argument based on polynomials was given in~\cite{Mat:inverse-closed} in the special case of finite fields, which appears to have attracted some attention for cryptographic applications. In that special case the result reads as follows: a non-trivial inverse-closed additive subgroup $A$ of a finite field $E$ is either a subfield of $E$ or the set of elements of trace zero in some quadratic field extension contained in $E$. In~\cite{Csajbok:inverse-closed}, Bence Csajb\'{o}k investigated a question which may be thought of as a refinement of this result: can one obtain the same conclusion from the weaker assumption that $A$ is an additive subgroup of a finite field which is {\em almost} inverse-closed, in the sense that {\em most} of the inverses of its nonzero elements belong to $A$? Of course the two words in italics need to be given a precise meaning. A very special case of this occurred in~\cite[Lemma~5.3]{KLS}, where the conclusion was proved under the assumption that $A$ is inverse closed up to at most two nonzero elements. It turns out that this question is better studied in the more general form where two additive subgroups $A$ and $B$ of the same size of a finite field are considered, and one asks for an upper bound on $|A^{-1}\cap B|$ in terms of $|B|$ in case $A^{-1}\not\subseteq B$. Here $S^{-1}$, for $S$ a subset of a field, denotes the set of inverses of the nonzero elements of $S$. Note that the intersection $A^{-1}\cap B$ attains maximal size $q^d-1$ exactly when $A^{-1}\subseteq B$, and in that case $A$ and $B$ are both (one-dimensional) $\F_{q^d}$-subspaces. Because the ambient finite field plays only a minor role, it appears convenient to work in the algebraic closure of a finite field, and so we rather state Csajb\'{o}k's results in the following equivalent form. \begin{theorem}[Theorems~1.2 and~3.1 in~\cite{Csajbok:inverse-closed}]\label{thm:bound} Let $A$ and $B$ be finite nonzero $\F_q$-subspaces of $\barFq$ of the same size, with $A^{-1}\not\subseteq B$. Then $|A^{-1}\cap B|\le 2|B|/q-2$. When $q=2$ and $|B|>2$ the conclusion can be strengthened to $|A^{-1}\cap B|\le 3|B|/4-1$. \end{theorem} In Section~\ref{sec:meet} we present a proof of Theorem~\ref{thm:bound} which is shorter than Csajb\'{o}k's original proof, but also more explicit. This is because, say in case of the former bound of Theorem~\ref{thm:bound}, our proof produces a polynomial $C(x)$, of degree $2|B|/q-2$, explicitly computable from the polynomials defining $A$ and $B$, whose set of roots contains $A^{-1}\cap B$. This can then be effectively used for further study of $A^{-1}\cap B$, as we illustrate next. A natural question which arises at this point is whether the bounds given in Theorem~\ref{thm:bound} are best possible, especially the general bound which holds for arbitrary $q$. In the early draft of~\cite{Csajbok:inverse-closed} which was available to the author during most of the writing of this paper, the few examples provided fell short of showing sharpness of the bound beyond the rather trivial cases where $|B|\le q^2$, which we briefly discuss at the end of Section~\ref{sec:meet}. The present work begun as an attempt to provide examples with $|B|=q^3$ where equality is attained in Csajb\'{o}k's bound. We present such examples in Section~3. They appear in Theorem~\ref{thm:q^3_full}, within a more general situation where $(A^{-1}\cap B)\cup\{0\}$ contains a one-dimensional $\F_{q^2}$-subspace of $\barFq$. In fact, under this assumption, which we will later show not to be restrictive, our proof of Theorem~\ref{thm:bound} is especially effective: it allows us to give a polynomial description of all pairs $(A,B)$ of three-dimensional spaces which attain equality in Csajb\'{o}k's bound, that is, which satisfy $|A^{-1}\cap B|=2q^2-2$. This is possible only for $q$ odd and, in geometric language, it occurs exactly when the image of $A^{-1}\cap B$ in the projective plane $\Proj B\cong\Proj^{2}(\F_q)$ associated with the linear $\F_q$-space $B$ is the union of a (nondegenerate) conic and an external line. The configuration of the union of a conic and a secant line also arises, and the corresponding subspaces then satisfy $|A^{-1}\cap B|=2q^2-2q$. In even characteristic those two configurations collapse to that of a conic and a tangent line, whence $|A^{-1}\cap B|=2q^2-q-1$. The final, published version of~\cite{Csajbok:inverse-closed}, includes a study of the case where $A$ and $B$ have dimension three, providing a presentation of the examples which we have briefly described. However, Csajb\'{o}k's geometric approach limits his results to subspaces of $\F_{q^4}$ for $q$ odd, and there is little overlap with our results, see Remark~\ref{rem:comparison}. The claim we implicitly made above, that we have actually found all pairs $(A,B)$ of three-dimensional $\F_q$-subspaces of $\barFq$ which attain equality in Csajb\'{o}k's bound, relies on removing our additional assumption that $(A^{-1}\cap B)\cup\{0\}$ contains a one-dimensional $\F_{q^2}$-subspace of $\barFq$. We do that in Section~\ref{sec:classification}, as an exceptional case of a more general goal which we now introduce. Csajb\'{o}k speculated in~\cite{Csajbok:inverse-closed} that for $\F_q$-subspaces $A$ and $B$ of the same fixed dimension $d>3$ the stronger bound $|A^{-1}\cap B|\le |B|/q\cdot\bigl(1+O(q^{-1/2})\bigr)$ might hold. Our main result shows that this is indeed the case. \begin{theorem}\label{thm:d_large} Let $A$ and $B$ be finite nonzero $\F_q$-subspaces of $\barFq$, with $|A|=|B|=q^d>q^3$ and $A^{-1}\not\subseteq B$. Then \[ |A^{-1}\cap B|\le q^{d-1}+(d-1)(d-2)q^{d-(3/2)}+C_d\cdot q^{d-2}, \] where $C_d$ only depends on $d$. \end{theorem} It follows, in particular, that Csajb\'{o}k's general bound of Theorem~\ref{thm:bound} can only be sharp for subspaces of dimension up to three (at least for large $q$). \begin{cor}\label{cor:never} Equality in the bound $|A^{-1}\cap B|\le 2|B|/q-2$ of Theorem~\ref{thm:bound} is never attained for $d>3$, where $|B|=q^d$, provided $q$ is sufficiently large with respect to $d$. \end{cor} The bound of Theorem~\ref{thm:d_large}, which we prove in Section~\ref{sec:higher-dim}, results from an application of the Lang-Weil bound to a multivariate polynomial closely related to the polynomial $C(x)$ used in our proof of Theorem~\ref{thm:bound}, when that is irreducible. However, considerable work is necessary, which we postpone to Section~\ref{sec:form}, to show that such polynomial can only be reducible in very special situations. Those exceptional geometric situations generalise the configurations of pairs of two- or three-dimensional subspaces attaining equality in Csajb\'{o}k's bound which we have briefly discussed earlier. \section{Intersecting a subspace with the inverse of another}\label{sec:meet} A finite subset of $\barFq$ is conveniently characterised by the unique monic polynomial in $\barFq[x]$ whose roots are the elements of the subset, each with multiplicity one. Thus, to the $\F_q$-subspaces $A$ and $B$ of $\barFq$, with size $q^d$, throughout the paper we associate the monic polynomials which have the elements of $A$ and $B$, respectively, as their roots, each with multiplicity one. (Using the same letters for the subspaces $A$ and $B$ and their polynomials should create no confusion.) It is well known that $A(x)$ and $B(x)$ are $q$-polynomials, see~\cite[Theorem~3.52]{LN}, which means that they have the form $A(x)=\sum_{i=0}^da_ix^{q^i}$ and $B(x)=\sum_{i=0}^db_ix^{q^i}$, with $a_d=b_d=1$. Also, the simplicity of their roots amounts to $a_0b_0\neq 0$. Hence the roots of $x^{q^d}\,A(1/x)=\sum_{i=0}^da_ix^{q^d-q^i}$ and $B(x)/x$ are the elements of $A^{-1}$ and $B\setminus\{0\}$, respectively. With this notation at hand we now present a very short proof of Csajb\'{o}k's bounds. \begin{proof}[Proof of Theorem~\ref{thm:bound} The idea of the proof is to give an upper bound on the degree of the greatest common divisor of $x^{q^d}\,A(1/x)$ and $B(x)/x$. The fact that the non-leading terms of $B(x)$ have relatively small degree suggests applying a variant of polynomial long division, where one keeps subtracting a scalar multiple of $B(x)/x$ from the current reminder multiplied by the appropriate power of $x$. The final result of this process is condensed in the following argument. The common roots of $x^{q^d}\,A(1/x)$ and $B(x)/x$ are also roots of the polynomial \begin{equation*} \begin{aligned} C(x) &= x^{q^d}A(1/x)\cdot x^{q^{d-1}-1}-B(x)/x \cdot x^{q^{d-1}}\bigl(A(1/x)-1/x^{q^d}\bigr) \\&= x^{q^{d-1}-1} -\bigl(B(x)-x^{q^d}\bigr)/x \cdot x^{q^{d-1}}\bigl(A(1/x)-1/x^{q^d}\bigr), \end{aligned} \end{equation*} which has degree at most $2q^{d-1}-2$. This shows that $|A^{-1}\cap B|\le 2|B|/q-2$, except when the polynomial $C(x)$ vanishes. The latter condition occurs exactly when $A(x)=x^{q^d}+a_0x$ and $B(x)=x^{q^d}+a_0^{-1}x$, which means that $A$ and $B$ are $\F_{q^d}$-subspaces, and $B=A^{-1}\cup\{0\}$. Assuming $d>1$ we now prove a different bound, which holds for arbitrary $q$ but improves on the previous bound only when $q=2$. The polynomial \begin{align*} D(x) &= C(x)\cdot x^{q^d-2q^{d-1}+1}+b_{d-1}x^{q^d}\,A(1/x) \\&= x^{q^d-q^{d-1}}+b_{d-1} -\bigl(B(x)-x^{q^d}-b_{d-1}x^{q^{d-1}}\bigr) \cdot x^{q^d-q^{d-1}}\bigl(A(1/x)-1/x^{q^d}\bigr), \end{align*} has degree at most $q^d-q^{d-1}+q^{d-2}-1$, and is nonzero if $A$ and $B$ are not both $\F_{q^d}$-subspaces. Because the common roots of $x^{q^d}\,A(1/x)$ and $B(x)/x$ are also roots of $D(x)$ we obtain $|A^{-1}\cap B|\le q^d-q^{d-1}+q^{d-2}-1$. This is better than the previous bound only when $q=2$, and then reads $|A^{-1}\cap B|\le 3|B|/4-1$. \end{proof} The above proof offers more advantages over Csajb\'{o}k's original proof than just brevity. It provides us with a polynomial \begin{equation}\label{eq:C(x)} C(x)=x^{q^{d-1}-1}- \bigl(\sum_{i=0}^{d-1}a_ix^{q^{d-1}-q^i}\bigr)\cdot \bigl(\sum_{j=0}^{d-1}b_jx^{q^j-1}\bigr), \end{equation} of degree at most $2q^{d-1}-2$, such that all elements of $A^{-1}\cap B$ are roots of $C(x)$. We will put that to good use in the next sections. We mention in passing that our proof of Theorem~\ref{thm:bound} can be easily modified to deal with affine $d$-dimensional $\F_q$-subspaces of $\barFq$. In fact, such affine subspaces have the form $A+\alpha$ and $B+\beta$, with $A,B$ as in Theorem~\ref{thm:bound} and $\alpha,\beta\in\barFq$, which are the sets of roots of the polynomials $A(x-\alpha)=A(x)-A(\alpha)$ and $B(x)-B(\beta)$. Therefore, the common elements of $(A+\alpha)^{-1}$ and $B+\beta$ are roots of the polynomial \[ xC(x) +A(\alpha)x^{q^{d-1}}\bigl(B(x)-x^{q^d}\bigr) +B(\beta)x^{q^{d-1}}\bigl(A(1/x)-1/x^{q^d}\bigr), \] and hence $|(A+\alpha)^{-1}\cap (B+\beta)|\le 2|B|/q$. However, we will not consider affine subspaces of $\barFq$ any further in this paper. We conclude this section by mentioning an alternate, more direct proof of Csajb\'{o}k's bound in the two-dimensional case. Let $A$ and $B$ be arbitrary two-dimensional subspaces of $\barFq$, and consider a maximal set of $\F_q$-linearly independent elements in $A^{-1}\cap B$. If that set is empty or a singleton, then $|A^{-1}\cap B|$ equals $0$ or $q-1$. Otherwise, that set consists of $1/\xi$ and $1/\eta$, for some $\xi,\eta\in A$. If $|A^{-1}\cap B|>2q-2$, then the inverse of some nontrivial linear combination of $\xi$ and $\eta$ also belongs to $B$. After possibly scaling $\xi$ or $\eta$ we may assume that $1/(\xi+\eta)$ belongs to $B$, and hence $1/(\xi+\eta)=a/\xi+b/\eta$ for some nonzero $a, b\in\F_q$. Therefore, $\xi/\eta$ satisfies a quadratic equation with coefficients in $\F_q$, and hence $\xi/\eta\in\F_{q^2}$. (A generalisation of this argument is presented in~\cite{Mat:caps}.) Consequently, $A$ is an $\F_{q^2}$-subspace, and $B=A^{-1}$ follows. We have shown that $|A^{-1}\cap B|/(q-1)\in\{0,1,2,q+1\}$. In particular, this argument proves an easy fact which is mentioned right after~\cite[Proposition~4.5]{Csajbok:inverse-closed}: any pair $(A,B)$ of two-dimensional $\F_q$-subspaces of $\barFq$ which attain equality in Csajb\'{o}k's bound $|A^{-1}\cap B|\le 2q-2$, is obtained by taking as $A$ an arbitrary two-dimensional $\F_q$-subspace which is not an $\F_{q^2}$-subspace, and as $B$ the $\F_q$-span of the inverses of any two $\F_q$-linearly independent elements of $A$. \section{A special configuration of subspaces}\label{sec:three_special} In this section we consider two $\F_q$-subspaces $A$ and $B$ of $\barFq$, of dimension $d$, in a rather special configuration. That assumption insures that the polynomial $C(x)$ of Equation~\eqref{eq:C(x)} has a factor of the form $x^{q^{d-1}-1}+c$, and this allows precise control over the set of roots of $C(x)$. This may seem like a rather artificial situation but, as we will see later in Theorem~\ref{thm:q^3_count}, when $d=3$ it includes all cases where equality is attained in Csajb\'{o}k's bound. Our proof is purely algebraic, but after the proof we will explain what goes on in geometric terms. Because $(\gamma^{-1}A)^{-1}\cap\gamma B=\gamma(A^{-1}\cap B)$ for any $\gamma\in\barFq^\ast$, we declare the ordered pair of $\F_q$-subspaces $(\gamma^{-1}A,\gamma B)$ to be {\em equivalent} to the pair $(A,B)$. Note that in principle one may consider a weaker equivalence relation which includes the application of field automorphisms, and possibly interchanging $A$ and $B$, but we chose not to do so. \begin{theorem}\label{thm:q^3_full} Let $A$ and $B$ be $\F_q$-subspaces of $\barFq$ of size $q^3$, and suppose that $(A^{-1}\cap B)\cup\{0\}$ contains a one-dimensional $\F_{q^2}$-subspace of $\barFq$. Then the pair $(A,B)$ is equivalent to a pair of subspaces consisting of the roots of the polynomials $A(x)=x^{q^3}+ax^{q^2}-x^q-ax$ and $B(x)=x^{q^3}+bx^{q^2}-x^q-bx$, for some $a,b\in\barFq^\ast$, and we have \begin{enumerate} \item $|A^{-1}\cap B|=2q^2-2$ if $q$ is odd, $a^{q+1}=b^{q+1}=-1$, $a^{(q+1)/2}\neq b^{(q+1)/2}$, and $ab\neq 1$; \item $|A^{-1}\cap B|=2q^2-2q$ if $q$ is odd, $a^{q+1}=-1$, and $a^{(q+1)/2}=b^{(q+1)/2}$; \item $|A^{-1}\cap B|=2q^2-q-1$ if $q$ is even, $a^{q+1}=b^{q+1}=1$, and $ab\neq 1$; \item $|A^{-1}\cap B|\le q^2+2q-3$ otherwise. \end{enumerate} \end{theorem} \begin{proof} With notation as in the proof of Theorem~\ref{thm:bound}, let $A(x)=x^{q^3}+a_2x^{q^2}+a_1x^q+a_0x$ and $B(x)=x^{q^3}+b_2x^{q^2}+b_1x^q+b_0x$ be the monic polynomials with distinct roots, hence with $a_0b_0\neq 0$, which have $A$ and $B$ as their sets of roots. That proof shows that all elements of $A^{-1}\cap B$ are roots of the polynomial \begin{align*} C(x) &= -a_0b_2x^{2q^2-2}-a_1b_2x^{2q^2-q-1}-a_0b_1x^{q^2+q-2} \\&\qquad +(1-a_2b_2-a_1b_1-a_0b_0)x^{q^2-1} \\&\qquad -a_1b_0x^{q^2-q}-a_2b_1x^{q-1}-a_2b_0. \end{align*} By hypothesis $(A^{-1}\cap B)\cup\{0\}$ contains a one-dimensional $\F_{q^2}$-subspace, hence the set of roots of a polynomial of the form $x^{q^2}+cx$. After replacing $A$, $B$ with an equivalent pair we may assume that $c=-1$, which means assuming that the $\F_{q^2}$-subspace under consideration is the subfield $\F_{q^2}$. This means that $x^{q^2}-x$ divides both $A(x)$ and $B(x)$, whence easily $A(x)=x^{q^3}+ax^{q^2}-x^q-ax$ and $B(x)=x^{q^3}+bx^{q^2}-x^q-bx$, and so \begin{align*} C(x) &= abx^{2q^2-2}+bx^{2q^2-q-1}-ax^{q^2+q-2} \\&\qquad -2abx^{q^2-1} \\&\qquad -bx^{q^2-q}+ax^{q-1}+ab. \\&= (x^{q^2-1}-1) (abx^{q^2-1}+bx^{q^2-q}-ax^{q-1}-ab). \end{align*} Because all polynomials involved can be expressed as polynomials in $x^{q-1}$ we conveniently set $y=x^{q-1}$, and so \[ C(x)=(y^{q+1}-1)(aby^{q+1}+by^q-ay-ab). \] The derivative criterion shows that the second factor of $C(x)$ shown above has distinct roots unless $ab=1$, in which case it equals $(y^q-a)(y+a^{-1})$, that is to say, $(y-a^{1/q})^q(y+a^{-1})$. In that case $C(x)$ has at most $q^2+2q-3$ distinct roots in $\barFq$ (as a polynomial in $x$), and hence $|A^{-1}\cap B|\le q^2+2q-3$, as claimed in assertion (4) of the theorem. (This can be improved to $|A^{-1}\cap B|\le q^2+q-2$ when $a^{q+1}=\pm 1$, because then the binomial $y+a^{-1}$ divides one of the other factors $y^{q+1}-1$ and $y^q-a$ of $C(x)$.) Assume $ab\neq 1$ from now on. Because \[ (y^{q+1}+a^{-1}y^q-b^{-1}y-1)\cdot y -(y^{q+1}-1)\cdot (y+a^{-1}) =-b^{-1}y^2+a^{-1} \] we see that the two exhibited factors of $C(x)$ are coprime unless $(b/a)^{(q+1)/2}=1$ for $q$ odd, and unless $(b/a)^{q+1}=1$ for $q$ even, and their greatest common divisor equals $x^{2q-2}-b/a$ in those cases. Note that this has distinct roots when $q$ is odd, but it has $q-1$ double roots when $q$ is even. This will account for the distinction between assertions~(1), (2), and~(3) of the theorem. Our next task is to find the degree of the greatest common divisor of $x^{q^3}A(1/x)$ and $C(x)$. To this goal we compute the remainder of the polynomial $x^{q^3}A(1/x)$ modulo $C(x)/(x^{q^2-1}-1)$. All congruences in the remainder of the proof will tacitly be modulo the latter polynomial, that is, modulo its scalar multiple $y^{q+1}+a^{-1}y^q-b^{-1}y-1$. We have \begin{align*} -bx^{q^3}A(1/x) &= aby^{q^2+q+1}+by^{q^2+q}-aby^{q^2}-b \\&= (aby^{q+1}+by^q-ab)y^{q^2}-b \\&\equiv ay\cdot y^{q^2}-b. \end{align*} Now note that $y^q\equiv (b^{-1}y+1)/(y+a^{-1})$, where our assumption $ab\neq 0$ ensures that the denominator is coprime with the modulus. Consequently, we have \[ y^{q^2} \equiv \frac{b^{-q}y^q+1}{y^q+a^{-q}} \equiv \frac{ b^{-q}\dfrac{b^{-1}y+1}{y+a^{-1}} +1} {\dfrac{b^{-1}y+1}{y+a^{-1}} +a^{-q}} = \frac{(1+b^{-q-1})y+(a^{-1}+b^{-q})} {(a^{-q}+b^{-1})y+(a^{-q-1}+1)}. \] Substituting this into our previous congruence we find \begin{equation}\label{eq:frac} \begin{split} -bx^{q^3}A(1/x) &\equiv ay\cdot \frac{(1+b^{-q-1})y+(a^{-1}+b^{-q})} {(a^{-q}+b^{-1})y+(a^{-q-1}+1)} -b \\&= \frac{a(1+b^{-q-1})y^2 +(ab^{-q}-a^{-q}b)y -(a^{-q-1}+1)b} {(a^{-q}+b^{-1})y+(a^{-q-1}+1)}. \end{split} \end{equation} Hence the greatest common divisor of $x^{q^3-1}A(1/x)$ and $y^{q+1}+a^{-1}y^q-b^{-1}y-1$ divides the numerator of this expression. If that numerator vanishes, that is, if $a^{q+1}=b^{q+1}=-1$, then the factor $y^{q+1}+a^{-1}y^q-b^{-1}y-1$ of $C(x)$ divides $x^{q^3}A(1/x)$, and by assumption so does the other factor $y^{q+1}-1$ of $C(x)$. We conclude that the greatest common divisor of $x^{q^3}A(1/x)$ and $B(x)$, which divides $C(x)$ but has distinct roots, equals the least common multiple of the two factors $x^{q^2-1}-1$ and $x^{q^2-1}+a^{-1}x^{q^2-q}-b^{-1}x^{q-1}-1$ of $C(x)$. According to an earlier calculation, when $q$ is odd this has degree $2q^2-2$ if $a^{(q+1)/2}\neq b^{(q+1)/2}$, and $2q^2-2q$ otherwise, proving assertions~(1) and~(2) of the theorem. When $q$ is even it has degree $2q^2-q-1$, as stated in assertion~(3). If the numerator of the expression found in Equation~\eqref{eq:frac} does not vanish, then the greatest common divisor of $x^{q^3}A(1/x)$ and $B(x)$ divides the product of that numerator and $x^{q^2-1}-1$, whence $|A^{-1}\cap B|\le q^2+2q-3$, as claimed in assertion~(4). \end{proof} We briefly pause to comment on the geometric interpretation of the intersection $A^{-1}\cap B$ in the case considered above, as a subset of the three-dimensional $\F_q$-space $B$. With $A(x)$ and $B(x)$ as in Theorem~\ref{thm:q^3_full}, we have \[ x^{q+2}\cdot C(x) = (x^{q^2}-x) (abx^{q^2+q}+bx^{q^2+1}-ax^{2q}-abx^{q+1}). \] The monomials $x$, $x^q$ and $x^{q^2}$ determine $\F_q$-linear maps $B\to\barFq$, and so they uniquely extend to independent $\barFq$-linear coordinates $x_0$, $x_1$ and $x_2$ on the linear space $B\otimes_{\F_q}\barFq$. With this interpretation, the two factors in the above factorisation of $x^{q+2}\cdot C(x)$ may be viewed as representing a linear form and a quadratic form on $B\otimes_{\F_q}\barFq$, namely, $x_2-x_0$ and $abx_1x_2+bx_0x_2-ax_1^2-abx_0x_1$. The quadratic form is nonsingular provided $ab\neq 1$, as we assume in the rest of this discussion. In the projective plane $\Proj^2(\barFq)$ associated with the linear space $B\otimes_{\F_q}\barFq$, this means that the roots of $C(x)$ represent the $\F_q$-rational points of the union of a line, which is defined over $\F_q$, and a nonsingular conic, which may or may not be defined over $\F_q$. The first three assertions of Theorem~\ref{thm:q^3_full} correspond to the case where the conic is defined over $\F_q$, and the further distinction depends on whether the line is external, secant, or tangent to the conic, with the first two cases occurring only for $q$ odd, and the last case only for $q$ even. An alternate presentation of this configuration for $A^{-1}\cap B$, using ideas from finite geometries and limited to $q$ odd, is given in Section~4 of Csajb\'{o}k's paper~\cite{Csajbok:inverse-closed}. To complete our geometric interpretation of Theorem~\ref{thm:q^3_full}, when the conic under consideration is not defined over $\F_q$, its $\F_q$-rational points belong also to the conic obtained from it by applying the Frobenius map $\alpha\mapsto\alpha^q$ to its coefficients, and so they are at most four, as they lie on the intersection of two distinct nonsingular conics. However, a simple calculation, of which we will sketch a more complex version for cubics in the proof of Theorem~\ref{thm:q^3_Weil}, shows that at most two of the intersection points of the conics over $\barFq$ are $\F_q$-rational. Adding to those the number of points on the line provides a geometric interpretation for the bound $|A^{-1}\cap B|/(q-1)\le q+3$ obtained for that case in~Theorem~\ref{thm:q^3_full}. Part of the argument in the proof of Theorem~\ref{thm:q^3_full} applies to pairs of higher-dimensional subspaces in a special configuration described in the following result, which will be needed later, in the proof of Theorem~\ref{thm:d_large}. \begin{theorem}\label{thm:q^d} Let $A$ and $B$ be $\F_q$-subspaces of $\barFq$ of size $q^d\ge q^3$, and suppose that $(A^{-1}\cap B)\cup\{0\}$ contains a one-dimensional $\F_{q^{d-1}}$-subspace of $\barFq$. Then $|A^{-1}\cap B|\le q^{d-1}+2q^2-3$. \end{theorem} \begin{proof} We argue in a similar way as in the proof of Theorem~\ref{thm:q^3_full}. After replacing $A$, $B$ with an equivalent pair we may assume that $x^{q^d}-x$ divides both $A(x)$ and $B(x)$, whence $A(x)=x^{q^d}+ax^{q^{d-1}}-x^q-ax$ and $B(x)=x^{q^d}+bx^{q^{d-1}}-x^q-bx$, with $ab\neq 0$, and so \begin{align*} C(x) &= (x^{q^{d-1}-1}-1) (abx^{q^{d-1}-1}+bx^{q^{d-1}-q}-ax^{q-1}-ab). \end{align*} We shall compute the remainder of the polynomial $x^{q^4}A(1/x)$ modulo the second factor of $C(x)$. Working modulo that polynomial we have \begin{align*} -bx^{q^d}A(1/x) &= abx^{q^d-1}+bx^{q^d-q}-abx^{q^d-q^{d-1}}-b \\&= (abx^{q^{d-1}-1}+bx^{q^{d-1}-q}-ab)x^{q^d-q^{d-1}}-b \\&\equiv ax^{q-1}\cdot x^{q^d-q^{d-1}}-b. \end{align*} Now we conveniently set $y=x^{q-1}$. Because $y^{q^{d-2}+\cdots+q}\equiv (b^{-1}y+1)/(y+a^{-1})$, we have \begin{align*} y^{q^{d-1}+\cdots+q^2} &\equiv \frac{b^{-q}y^q+1}{y^q+a^{-q}}, \end{align*} and hence \begin{align*} -bx^{q^d}A(1/x) &\equiv ay\cdot y^{q^{d-1}}-b \\&= \frac{1}{y^{q^{d-2}+\cdots+q}} ( ay\cdot y^q\cdot y^{q^{d-1}+\cdots+q^2} )-b \\&\equiv ay^{q+1}\cdot \frac{y+a^{-1}}{b^{-1}y+1}\cdot \frac{b^{-q}y^q+1}{y^q+a^{-q}} -b \\&\equiv \frac{ab^{-q}y^{2q+2}+b^{-q}y^{2q+1}+ay^{q+2}-by^q-a^{-q}y-ba^{-q}} {(b^{-1}y+1)(y^q+a^{-q})}. \end{align*} Because the numerator of this expression is nonzero, its degree $(2q+2)(q-1)=2q^2-2$, in the original indeterminate $x$, is an upper bound for the degree of the greatest common divisor of $x^{q^{d-1}-1}+a^{-1}x^{q^{d-1}-q}-b^{-1}x^{q-1}-1$ and $x^{q^d}A(1/x)$. The desired conclusion follows by taking the other factor $x^{q^{d-1}-1}-1$ of $C(x)$ into account. \end{proof} \section{A better bound for subspaces of dimension at least four}\label{sec:higher-dim} The proof of Theorem~\ref{thm:bound} which we gave in Section~\ref{sec:meet} shows that all elements of $A^{-1}\cap B$ are roots of the polynomial $C(x)$ of Equation~\eqref{eq:C(x)}, and hence of the modified polynomial \begin{align*} C_0(x) &=x^{1+(1+q+q^2+\cdots+q^{d-2})}\cdot C(x) \\&= x^{1+q+q^2+\cdots+q^{d-1}} \biggl(1- \bigl(\sum_{i=0}^{d-1}a_i/x^{q^i}\bigr)\cdot \bigl(\sum_{j=0}^{d-1}b_jx^{q^j}\bigr) \biggr). \end{align*} The latter has the advantage of being a linear combination of monomials, each of whose degrees is a sum of $d$ terms taken from the set $\{1,q,\ldots,q^{d-1}\}$ with at most one repetition, and hence one omission. We now show how being a root of $C_0(x)$ can be interpreted as being a zero of one or more homogeneous forms of degree $d$ on the $\F_q$-space $B$. The $\F_q$-linear maps $x_i:B\to\barFq$ given by $x\mapsto x^{q^i}$, for $0\le i<d$, are $\F_q$-linearly independent, and so they form a complete set of linear coordinates (that is, a basis of the dual space) on the $\barFq$-space $B\otimes_{\F_q}\barFq$. Hence to $C_0(x)$ there corresponds a homogeneous polynomial function \[ E(x_0,\ldots,x_{d-1}) = x_0\cdots x_{d-1}\cdot \biggl(1- \bigl(\sum_{i=0}^{d-1}a_i/x_i\bigr)\cdot \bigl(\sum_{j=0}^{d-1}b_jx_j\bigr) \biggr), \] of degree $d$, defined on the $\barFq$-space $B\otimes_{\F_q}\barFq$. We would rather need a polynomial function on the original $\F_q$-space $B$, but we can obtain that through a linear change of coordinates. In fact, an arbitrary $\F_q$-linear map on $B$ with values in $\F_q$ is given by \begin{multline*} x \mapsto b_0\gamma x +(b_0^q\gamma^q+b_1\gamma)x^q +(b_0^{q^2}\gamma^{q^2}+b_1^q\gamma^q+b_2\gamma)x^{q^2} +\cdots \\ \cdots +(b_0^{q^{d-1}}\gamma^{q^{d-1}}+b_1^{q^{d-2}}\gamma^{q^{d-2}}+\cdots+b_{d-1}\gamma)x^{q^{d-1}}, \end{multline*} where $\gamma$ ranges over the roots of the $q$-polynomial $b_0^{q^d}x^{q^d}+b_1^{q^{d-1}}x^{q^{d-1}}+\cdots+b_dx$. Therefore, a complete set of $\F_q$-linear coordinates $z_0,\ldots,z_{d-1}$ on the $\F_q$-space $B$ is obtained by letting $\gamma$ range over an $\F_q$-basis of the roots of that $q$-polynomial. After expressing each $x_i$ in terms of $z_0,\ldots,z_{d-1}$ (as linear combinations over $\barFq$), the polynomial function $E(x_0,\ldots,x_{d-1})$ on $B\otimes_{\F_q}\barFq$ yields a polynomial function $\tilde E(z_0,\ldots,z_{d-1})$ on $B$, but still with values in $\barFq$, which is homogeneous of degree $d$ in $z_0,\ldots,z_{d-1}$. To recapitulate in slightly different wording, all elements of $A^{-1}\cap B$, once $B$ is identified with $\F_q^d$ via the coordinates $z_i$, are roots of the polynomial $\tilde E(z_0,\ldots,z_{d-1})\in\barF_q[z_0,\ldots,z_{d-1}]$. According to Theorem~\ref{thm:form} below, this polynomial turns out to be usually irreducible for our purposes, with notable exceptions where our geometric problem has already been dealt with in Section~\ref{sec:three_special}. When the polynomial is indeed irreducible, it defines a hypersurface in the projective space $\Proj(B\otimes_{\F_q}\barFq)\cong\Proj^{d-1}(\barFq)$, whose number of $\F_q$-rational points can be bounded using the Lang-Weil bound~\cite{Lang-Weil}. These are the ideas at play in the proof of our Theorem~\ref{thm:d_large} stated in the Introduction, which we give below, but not before stating the result on the possible factorisations of $E$ which we have just mentioned. \begin{theorem}\label{thm:form} The homogeneous polynomial \[ E=E(x_1,\ldots,x_n) = x_1\cdots x_n\cdot \biggl(1+ \bigl(\sum_{i=1}^na_i/x_i\bigr)\cdot \bigl(\sum_{j=1}^nb_jx_j\bigr) \biggr) \] where $a_i,b_j\in\barFq$, has at most two non-monomial (absolutely) irreducible factors. If it has two then at least one of them is a linear combination of exactly two of the indeterminates. Furthermore, if $E$ has $x_1+x_2$ as a factor, then either \begin{multline*} E/(x_3\cdots x_n)= x_1x_2\cdot\bigl( 1+(a/x_1+(a+c)/x_2)\cdot(bx_1+(b-c^{-1})x_2) \bigr) \\ = (x_1+x_2)\bigl((a+c)bx_1+a(b-c^{-1})x_2\bigr) \end{multline*} for some $a,b,c\in\barFq$ with $c\neq 0$, or \begin{multline*} E/(x_4\cdots x_n)= x_1x_2x_3\cdot\bigl( 1+(a/x_1+a/x_2-1/x_3)\cdot(bx_1+bx_2-x_3) \bigr) \\ = (x_1+x_2)(abx_2x_3+abx_1x_3+bx_1x_2-ax_3^2), \end{multline*} up to permuting the indeterminates $x_3,\ldots,x_n$. \end{theorem} The harmless sign change in the definition of $E$ from the previous notation will avoid the occurrence of several minus signs in the proof of Theorem~\ref{thm:form}. We have also conveniently shifted the indices of the indeterminates, which now start from one. The proof of Theorem~\ref{thm:form} is a little technical, and we postpone it to Section~\ref{sec:form} to avoid disrupting the flow of the present argument. Note that the second nontrivial factorisation allowed by Theorem~\ref{thm:form} has already occurred in disguise in the factorisations of $C(x)$ obtained in the proofs of Theorems~\ref{thm:q^3_full} and~\ref{thm:q^d}. \begin{proof}[Proof of Theorem~\ref{thm:d_large}] Continue with the setting introduced above. As we noted in the proof of Theorem~\ref{thm:bound}, the condition $A^{-1}\not\subseteq B$ implies that $C(x)$ is not the zero polynomial, whence $E(x_0,\ldots,x_{d-1})$ is not the zero polynomial. Monomial factors of $E(x_0,\ldots,x_{d-1})$ and, correspondingly, of $C(x)$, clearly give no contribution to estimating $|A^{-1}\cap B|$. Suppose first that $E(x_0,\ldots,x_{d-1})$ has a unique non-monomial irreducible factor $F(x_0,\ldots,x_{d-1})$, of degree $d'\le d$. Let $\tilde F(z_0,\ldots,z_{d-1})$ be the corresponding polynomial written in terms of the coordinates $z_0,\ldots,z_{d-1}$ on the $\F_q$-space $B$. It defines an irreducible algebraic subvariety of the projective space $\Proj(B\otimes_{\F_q}\barFq)\cong\Proj^{d-1}(\barFq)$, of dimension $d-2$ (that is, a hypersurface). If that variety is defined over $\F_q$, which occurs if $\tilde F(z_0,\ldots,z_{d-1})$, after multiplication by a suitable scalar, can be made to have all coefficients in $\F_q$, then according to the Lang-Weil estimate~\cite{Lang-Weil} its number of $\F_q$-rational points is bounded above by \begin{equation}\label{eq:Lang-Weil} q^{d-2}+(d'-1)(d'-2)q^{d-(5/2)}+C_{d,d'}\cdot q^{d-3}, \end{equation} where $C_{d,d'}$ is a constant which depends only on $d$ and $d'$. The desired conclusion is then obtained upon multiplication by $q-1$ and using the fact that $d'\le d$. Now suppose that the variety under consideration is not defined over $\F_q$. Then the Galois-conjugate polynomial $\tilde F^\sigma(z_0,\ldots,z_{d-1})$, obtained from $\tilde F(z_0,\ldots,z_{d-1})$ by applying the Frobenius automorphism $\sigma:a\mapsto a^q$ to each coefficient, is not proportional to $\tilde F(z_0,\ldots,z_{d-1})$, and hence together with the latter it defines a (possibly reducible) algebraic set in $\Proj^{d-1}(\barFq)$, of dimension strictly smaller than $d-2$, and degree at most $(d')^2$. According to a standard fact known as the {\em Schwartz-Zippel lemma,} see~\cite[Lemma A.3]{Tao+:Kakeya} for a proof, the number of $\F_q$-rational points of this algebraic set is at most $(d')^2(q+1)^{d-3}$. After multiplying by $q-1$ we see that the desired conclusion holds in this case as well. Now we may assume that $E(x_0,\ldots,x_{d-1})$ has at least two non-monomial irreducible factors. Then it has exactly two according to Theorem~\ref{thm:form}, and one of them is a linear combination of two of the indeterminates, say $x_i$ and $x_j$ with $i<j$. The corresponding factor of our original polynomial $C(x)$ is then a linear combination of $x^{q^i-1}$ and $x^{q^j-1}$, and hence it accounts for at most $q^{j-i}-1$ distinct nonzero roots of that polynomial. The possible factorisations of $E$ given in Theorem~\ref{thm:form} show that the remaining factor of $C(x)$ has degree at most $q^{d-1}-1$, whence $|A^{-1}\cap B|\le q^{d-1}+q^{j-i}-2$. If $j-i<d-1$ we have reached our goal, hence assume $j-i=d-1$. This means that the former factor of $C(x)$ considered above is a linear combination of $1$ and $x^{q^{d-1}-1}$. Possibly after replacing $(A,B)$ with an equivalent pair we may assume that linear combination to be $x^{q^{d-1}-1}-1$. Now consider, in turn, the two possible factorisations of $E$ stated in Theorem~\ref{thm:form}, and what they entail for the polynomials $A(x)$ and $B(x)$ in our setting. The former factorisation implies $A(x)=x^{q^d}+ax^{q^{d-1}}-(a+c)x$ and $B(x)=x^{q^d}-bx^{q^{d-1}}+(b-c^{-1})x$, whence \begin{align*} C(x) &= (x^{q^{d-1}-1}-1) \bigl(-(a+c)bx^{q^{d-1}-1}+a(b-c^{-1})\bigr). \end{align*} Because $B(x)\equiv x^q-c^{-1}x\pmod{x^{q^{d-1}-1}-1}$, the polynomial $B(x)$ has at most $q-1$ nonzero roots in common with the former factor of $C(x)$, and similarly with the latter factor. Hence in this case we have $|A^{-1}\cap B|\le 2q-2$. The other possible factorisation of $E$ yields $A(x)=x^{q^d}+ax^{q^{d-1}}-cx^{q^e}-ax$ and $B(x)=x^{q^d}+bx^{q^{d-1}}-c^{-1}x^{q^e}-bx$, with $0<e<d-1$ and $abc\neq 0$, and so \begin{align*} C(x) &= (x^{q^{d-1}-1}-1) (abx^{q^{d-1}-1}+bcx^{q^{d-1}-q^e}-ac^{-1}x^{q^e-1}-ab). \end{align*} Because $B(x)\equiv x^q-c^{-1}x^{q^e}\pmod{x^{q^{d-1}-1}-1}$, the polynomial $B(x)$ has at most $q-1$ nonzero roots in common with the former factor of $C(x)$, whence $|A^{-1}\cap B|\le q^{d-1}+q-2$, except when $c=1$ and $e=1$. However, in the latter case Theorem~\ref{thm:q^d} applies and yields $|A^{-1}\cap B|\le q^{d-1}+2q^2-3$. \end{proof} \section{A classification of pairs of three-dimensional subspaces\\ with large intersection $A^{-1}\cap B$}\label{sec:classification} In the case of subspaces of dimension $d=3$, which was excluded from Theorem~\ref{thm:d_large}, parts of its proof still apply. In particular, the Lang-Weil estimate of Equation~\eqref{eq:Lang-Weil} takes the more precise form of the Hasse-Weil bound, and allows us to prove the following result. \begin{theorem}\label{thm:q^3_Weil} Let $A$ and $B$ be $\F_q$-subspaces of $\barFq$, of size $q^3$, with $A^{-1}\not\subseteq B$. If $|A^{-1}\cap B|/(q-1)>q+1+\lfloor 2\sqrt{q}\rfloor$, then $|A^{-1}\cap B|/(q-1)$ equals either $2q+2$ or $2q$ for $q$ odd, and it equals $2q+1$ for $q$ even. \end{theorem} Note that equality in Csajb\'{o}k's bound $|A^{-1}\cap B|\le 2q^2-2$ for three-dimensional spaces cannot be attained in characteristic two. \begin{proof} Arguing as in the proof of Theorem~\ref{thm:d_large}, which we gave in Section~\ref{sec:higher-dim}, and aiming at a contradiction, suppose that $E(x_0,x_1,x_2)$ is irreducible. Writing this in terms of the coordinates $z_0,\ldots,z_{d-1}$ on the $\F_q$-space $B$ we get a polynomial $\tilde E(z_0,z_1,z_2)$ which defines an irreducible cubic in the projective plane $\Proj(B\otimes_{\F_q}\barFq)\cong\Proj^{2}(\barFq)$. If the cubic is defined over $\F_q$, then according to the Hasse-Weil bound its number of $\F_q$-rational points does not exceed $q+1+2\sqrt{q}$, whence $|A^{-1}\cap B|/(q-1)\le q+1+2\sqrt{q}$, which contradicts our hypothesis. If the cubic is not defined over $\F_q$, then its intersection with the irreducible cubic defined by $\tilde E^\sigma(z_0,z_1,z_2)$, where $\sigma$ is the Frobenius automorphism, has at most $3^2=9$ points in $\Proj(B\otimes_{\F_q}\barFq)$ according to B\'{e}zout's theorem (or to the Schwartz-Zippel lemma if we prefer). Hence $|A^{-1}\cap B|/(q-1)\le 9$, and we obtain a contradiction because this number does not exceed Weil's bound $q+1+\lfloor 2\sqrt{q}\rfloor$, except when $q\le 3$. However, when $q=2$ Theorem~\ref{thm:bound} provides the improved bound $|A^{-1}\cap B|\le 3\cdot 2^3/4-1=5$, which yields the desired contradiction. In order to cover the case $q=3$ as well, we sketch how an explicit calculation allows us to strengthen the upper bound of $9$ given by B\'{e}zout's theorem to the bound $|A^{-1}\cap B|/(q-1)\le 6$, for arbitrary $q$. We do that by showing that the intersection of the zero sets of $\tilde E(z_0,z_1,z_2)$ and its Galois-conjugate $\tilde E^\sigma(z_0,z_1,z_2)$ contains at least three non-rational points. One way of computing the Galois-conjugate in terms of the original coordinates $x_0,x_1,x_2$ is raising the polynomial \begin{align*} C_0(x) =x^{q+2}\cdot C(x) &= -a_0b_2x^{2q^2+q}-a_1b_2x^{2q^2+1}-a_0b_1x^{q^2+2q} \\&\qquad +(1-a_2b_2-a_1b_1-a_0b_0)x^{q^2+q+1} \\&\qquad -a_1b_0x^{q^2+2}-a_2b_1x^{2q+1}-a_2b_0x^{q+2} \end{align*} to the $q$-th power and reducing the result modulo the polynomial $B(x)$. Writing both the remainder of this division and the original polynomial $C_0(x)$ in terms of $x_0=x$, $x_1=x^q$, and $x_2=x^{q^2}$, one discovers that both vanish for $(x_0,x_1,x_2)=(1,0,0),(b_1,-b_0,0),(b_2,0,-b_0)$. However, a triple $(x_0,x_1,x_2)$ gives a rational point of our curve, that is, an element of $B$, only when $x_1=x_0^q$, $x_2=x_1^q$, and $x_2^q=-b_2x_2-b_1x_1-b_0x_0$, which is not the case for any of the three triples found. Thus, we have shown that $E(x_0,x_1,x_2)$ cannot be irreducible. Invoking Theorem~\ref{thm:form} and arguing as in the proof of Theorem~\ref{thm:d_large} we see that $|A^{-1}\cap B|$ can possibly be so large only when $C(x)$ has a non-trivial linear combination of $1$ and $x^{q^2}$ as a factor, which may be taken to be $x^{q^2-1}-1$ after passing to an equivalent pair $(A,B)$. The proof of Theorem~\ref{thm:d_large} also shows that we must have the second exceptional factorisation of $E(x_0,x_1,x_2)$ given in Theorem~\ref{thm:form}, and that $A(x)=x^{q^3}+ax^{q^2}-x^q-ax$ and $B(x)=x^{q^3}+bx^{q^2}-x^q-bx$. Consequently, both $A$ and $B$ contain $\F_{q^2}$, hence Theorem~\ref{thm:q^3_full} applies and completes the proof. \end{proof} \begin{rem}\label{rem:comparison} A special version of Theorem~\ref{thm:q^3_Weil} occurs as assertion~(1) of~\cite[Theorem~4.8]{Csajbok:inverse-closed}. That result restricts $A$ and $B$ to be contained in $\F_{q^4}$, which is a rather strong assumption. Note that~\cite[Theorem~4.8]{Csajbok:inverse-closed} is actually proved under the unstated hypothesis that $q$ is odd, but fails to exclude the possibility that $|A^{-1}\cap B|/(q-1)=2q+1$ (except in the special case where $A=B$). At the author's request Csajb\'{o}k has produced a proof which excludes that possibility, based on similar methods as~\cite{Csajbok:inverse-closed}. \end{rem} Theorem~\ref{thm:q^3_Weil}, combined with the special configuration which we investigated in Theorem~\ref{thm:q^3_full}, allows us to classify all pairs $(A,B)$ attaining equality in Csajb\'{o}k's bound, or almost, in the three-dimensional case. \begin{theorem}\label{thm:q^3_count} In the following assertions $A$ and $B$ denote $\F_q$-subspaces of $\barFq$, of size $q^3$, with $A^{-1}\not\subseteq B$. \begin{enumerate} \item There are exactly $(q-1)/2$ equivalence classes of pairs $(A,B)$ such that $|A^{-1}\cap B|=2q^2-2$. \item For odd $q>5$ there are exactly $(q+1)/2$ equivalence classes of pairs $(A,B)$ such that $|A^{-1}\cap B|=2q^2-2q$. \item For even $q>4$ there are exactly $q$ equivalence classes of pairs $(A,B)$ such that $|A^{-1}\cap B|=2q^2-q-1$. \item Each equivalence class described in assertions~(1) and~(2) contains exactly two pairs satisfying $A=B$, and each equivalence class described in assertion~(3) contains exactly one such pair. Each such subspace $A=B$ is contained in $\F_{q^4}$, and equals the kernel of $x\mapsto\Tr_{\F_{q^4}/\F_q}(\alpha x)$, for some $\alpha\in\F_{q^4}$ with $\alpha^2\in\F_{q^2}\setminus\F_q$. \end{enumerate} \end{theorem} \begin{proof}The stated conditions on $q$ insure that $|A^{-1}\cap B|/(q-1)$ exceeds the Hasse-Weil bound in each case. Therefore, as in the proof of Theorem~\ref{thm:q^3_Weil} we conclude that after replacing $(A,B)$ with an equivalent pair we have $A(x)=x^{q^3}+ax^{q^2}-x^q-ax$ and $B(x)=x^{q^3}+bx^{q^2}-x^q-bx$, and Theorem~\ref{thm:q^3_full} gives exact conditions that $a$ and $b$ satisfy. Now the subspaces $\gamma^{-1}A$ and $\gamma B$, which form an equivalent pair to $(A,B)$ for $\gamma\in\barF_q^\ast$, are the sets of roots of the monic polynomials $A_{\gamma^{-1}}(x)=\gamma^{-q^3}A(\gamma x)$ and $B_{\gamma}(x)=\gamma^{q^3}B(x/\gamma)$. Comparing coefficients we see that each of the equalities $\gamma^{-1}A=A$ and $\gamma B=B$ occurs only when $\gamma^{q-1}=1$, that is, when $\gamma\in\F_q^\ast$. However, $A_{\gamma^{-1}}(x)$ has the admissible form $A_{\gamma^{-1}}(x)=x^{q^3}+a'x^{q^2}-x^q-a'x$ considered above (which means that it is a multiple of $x^{q^2}-1$) if and only if $\gamma\in\F_{q^2}^{\ast}$, and so does $B_{\gamma}(x)$. Consequently, each equivalence class of pairs $(A,B)$ under consideration contains exactly $q+1$ pairs for which the corresponding polynomials $A(x)$ and $B(x)$ have the required form. Thus, the number of equivalence classes is obtained after dividing by $q+1$ the number of pairs $(a,b)$ with $a^{q+1}=b^{q+1}=-1$ and $ab\neq 1$, and possibly the further conditions given in Theorem~\ref{thm:q^3_full}. This establishes assertions~(1), (2), and~(3). A similar coefficient comparison shows that for the pairs $(A,B)$ under consideration an equivalent pair $(\gamma^{-1}A,\gamma B)$ satisfies $\gamma^{-1}A=\gamma B$ exactly when $\gamma^{2(q-1)}=a/b$. Consequently, there are two such pairs equivalent to $(A,B)$ when $q$ is odd, and only one when $q$ is even, as claimed in assertion~(4). Furthermore, taking $\gamma^{2(q-1)}=a/b$, whence $\gamma^{q^2-1}=(a/b)^{(q+1)/2}=\pm 1$, we obtain \begin{align*} A_{\gamma^{-1}}(x) =B_{\gamma}(x) &=x^{q^3}+\gamma^{q^3-q^2}bx^{q^2}-\gamma^{q^3-q}x^q+\gamma^{q^3-1}bx \\&=x^{q^3}+\gamma^{q-1}bx^{q^2}-\gamma^{q^2-1}(x^q+\gamma^{q-1}bx) \\&=x^{q^3}+cx^{q^2}+c^{q+1}(x^q+cx) \\&=x^{q^3}+cx^{q^2}+c^{q+1}x^q+c^{q^2+q+1}x, \end{align*} having set $c=\gamma^{q-1}b$ and noted that $c^{q+1}=-\gamma^{q^2-1}=\pm 1$. For $(a,b)$ ranging over all pairs such that $a^{q+1}=b^{q+1}=-1$ and $ab\neq 1$, and with $\gamma$ chosen as above, $c$ takes all the values such that $c^{2(q+1)}=1$ and $c^2\neq 1$. The description of $A$ given in assertion~(4) follows at once by writing $c=\alpha^{-(q-1)}$. \end{proof} \begin{rem} The restrictions on $q$ in assertions~(2) and~(3) of Theorem~\ref{thm:q^3_count} cannot be relaxed. For example, a computer calculation shows that $\F_{5^4}$ contains $31\cdot 8$ pairs $(A,B)$ of three-dimensional $\F_5$-subspaces such that $|A^{-1}\cap B|=2q^2-2q=40$, rather than $(q^2+q+1)\cdot(q+1)/2=31\cdot 3$ as Theorem~\ref{thm:q^3_count} would predict. For those in excess, $A^{-1}\cap B$ yields an irreducible cubic in the projective plane $\Proj B$. \end{rem} \begin{rem}\label{rem:caps} In case of assertion~(1) of Theorem~\ref{thm:q^3_count}, where equality in Csajb\'{o}k's bound is attained, an alternate approach is available and described in~\cite{Mat:caps}. It relies on facts from finite geometries to bypass the arguments of this and the previous section, and hence Theorem~\ref{thm:form}, on which they ultimately depend. Briefly, as a special case of a more general result it is shown in~\cite{Mat:caps} that the image of $A^{-1}\cap B$ in $\Proj B$ is an arc, for three-dimensional $\F_q$-subspaces $A,B$ of $\barFq$, unless $(A^{-1}\cap B)\cup\{0\}$ contains a one-dimensional $\F_{q^2}$-subspace of $\barFq$. Because it is known that an arc in $\Proj^2(\F_q)$ has at most $2q+1$ points if $q>3$, when equality $|A^{-1}\cap B|/(q-1)=2q+2$ holds in Csajb\'{o}k's bound we conclude that $(A^{-1}\cap B)\cup\{0\}$ contains a one-dimensional $\F_{q^2}$-space, and then our Theorem~\ref{thm:q^3_full} applies. If $q$ is odd (and larger than three), at this point one can also deduce that the image of $A^{-1}\cap B$ in $\Proj B$ is the union of a line and a conic without using Theorem~\ref{thm:q^3_full}, appealing instead to the classical result of B.~Segre that an an arc with $q+1$ points in a projective plane is a conic (for $q$ odd), see~\cite[Theorem~5]{Mat:caps}. \end{rem} \begin{rem}\label{rem:q^2} In contrast with the three-dimensional case considered in Theorem~\ref{thm:q^3_count}, for a fixed prime power $q$ there are infinitely many equivalence classes of pairs $(A,B)$ of two-dimensional $\F_q$-subspaces which attain equality in Csajb\'{o}k's bound $|A^{-1}\cap B|\le 2q-2$. This follows at once from their description which we gave at the end of Section~\ref{sec:meet}. \end{rem} \begin{cor}\label{cor:q^3} Assume $q>5$, and set $P(x)=x^{q^3}+x^{q^2}+x^q+x$. If $A$ and $B$ are $\F_q$-subspaces of $\barFq$, of size $q^3$, such that \[ q^2-1+\lfloor 2q^{1/2}\rfloor\cdot(q-1) <|A^{-1}\cap B|<q^3-1, \] then $A$ and $B$ are the sets of roots of $P(\alpha\gamma^{-1}x)$ and $P(\alpha\gamma x)$, respectively, for some $\alpha,\gamma\in\barFq^{\ast}$ with $\alpha^2\in\F_{q^2}\setminus\F_q$. \end{cor} \begin{proof} According to Theorem~\ref{thm:q^3_Weil} the pair $(A,B)$ is one of those described in Theorem~\ref{thm:q^3_count}. According to assertion~(4) of the latter, some equivalent pair $(\gamma^{-1}A,\gamma B)$ satisfies $\gamma^{-1}A=\gamma B$, and that subspace equals the set of roots of $P(\alpha x)$, for some $\alpha\in\barFq^{\ast}$ with $\alpha^2\in\F_{q^2}\setminus\F_q$. \end{proof} The explicit description of the spaces $A$ and $B$ given in Corollary~\ref{cor:q^3} allows one to decide whether and how many of them can be found inside a given finite field. For example, in $\F_{q^4}$ there are exactly $2q(q^2+q+1)$ such pairs $(A,B)$ for $q$ odd, and $q(q^2+q+1)$ for $q$ even, because $\gamma\in\F_{q^4}^{\ast}$ in this case. Those among them with $A=B$ are in number of $2q$ and $q$, respectively, and for $q$ odd they match those described in~\cite[Propositions~4.4 and~4.5]{Csajbok:inverse-closed}. Another example is the following improvement of Csajb\'{o}k's bound for three-dimensional subspaces of finite fields which do not contain $\F_{q^4}$. \begin{cor} Consider the finite field $\F_{q^e}$, where $q>5$ and $e$ is not a multiple of four. If $A$ and $B$ are $\F_q$-subspaces of $\F_{q^e}$ with size $q^3$, and $A^{-1}\not\subseteq B$, then $ |A^{-1}\cap B|\le q^2-1+\lfloor 2q^{1/2}\rfloor\cdot(q-1). $ \end{cor} \begin{proof} Consider subspaces $A$ and $B$ of $\barFq$ as in Corollary~\ref{cor:q^3}. When $\gamma=1$ they are both contained in $\F_{q^4}$, and $(A^{-1}\cap B)\cup\{0\}$ properly contains the one-dimensional $\F_{q^2}$-subspace consisting of the roots of $(\alpha x)^{q^2}+\alpha x$. Therefore, the subfield generated by all the quotients of pairs of elements of $A^{-1}\cap B$ properly contains $\F_{q^2}$, and hence equals $\F_{q^4}$. This last statement carries over to the case of arbitrary $\gamma$. Consequently, the subfield of $\barFq$ generated by $A^{-1}\cap B$ contains $\F_{q^4}$, and hence $A$ and $B$ cannot be both contained in $\F_{q^e}$. \end{proof} \section{Proof of Theorem~\ref{thm:form}}\label{sec:form} Recall that a polynomial is called {\em multilinear} if it has degree at most one in each indeterminate. The following property of the polynomial $E$, which follows from the definition and is inherited by factors, will be crucial in most of our arguments. \begin{property}\label{property} Any factor of $E$ has degree at most two in each indeterminate $x_i$, and joint degree at most three in each pair of indeterminates $x_i$ and $x_j$. \end{property} Thus, for example, $E$ cannot have any term divisible by $x_1^2x_2^2$. As another example, if $E=FG$ and $F$ has degree two in some $x_i$, then $G$ cannot involve $x_i$. \subsection{Linear factors of $E$} We first prove the conclusions of Theorem~\ref{thm:form} under the additional assumption that $E$ has a non-monomial linear factor. \begin{lemma}\label{lemma:linear} Under the hypotheses of Theorem~\ref{thm:form}, any non-monomial linear factor of $E$ is a linear combination of exactly two indeterminates. \end{lemma} \begin{proof} Let $E=FG$, with $G$ a non-monomial linear factor. Because $F$ is homogeneous of degree $n-1$ in $n$ indeterminates, according to Property~\ref{property} each term of $F$ misses at most two indeterminates. Now suppose for a contradiction that $G$ involves at least three indeterminates. Then an arbitrary monomial of $F$ must share at least one indeterminate with $G$, say $x_i$. But then $F$ has degree exactly one in that $x_i$, and joint degree at most two in $x_i$ and $x_j$, for each other indeterminate $x_j$. Consequently, that arbitrary term of $F$ has degree at most one in each indeterminate, and hence $F$ is multilinear. Writing $F=x_1\cdots x_n\cdot\sum_{i=1}^n f_i/x_i$ and $G=\sum_{j=1}^n g_jx_j$ and comparing coefficients of the non-multilinear terms of $E$ on both sides of the equation $FG=E$ we find $f_ig_j=a_ib_j$ for $i\neq j$. Now our assumption that at least three of the coefficients $g_j$ are nonzero implies that the $n$-tuples $(f_1,\ldots,f_n)$ and $(a_1,\ldots,a_n)$ are proportional. In fact, our equations imply $(f_ia_j-f_ja_i)g_kb_k=0$ for any distinct $i,j,k$. If $g_k\neq 0$ for some $k$, then $b_k\neq 0$, otherwise $f_i=a_ib_k/g_k=0$ and $f_j=a_jb_k/g_k=0$, which is impossible. Hence if $g_k\neq 0$ for some $k$, then $f_ia_j=f_ja_i$ for any $i,j\neq k$. Consequently, if $g_k\neq 0$ for at least three values of $k$, then then $f_ia_j=f_ja_i$ for any $i,j$, and hence the $n$-tuples $(f_1,\ldots,f_n)$ and $(a_1,\ldots,a_n)$ are proportional. We conclude that $F$ is a scalar multiple of $x_1\cdots x_n\cdot \sum_{i=1}^{n}a_i/x_i$, in plain contradiction with the definition of $E$ and the fact that $FG=E$. \end{proof} \begin{lemma}\label{lemma:exceptions} Under the hypotheses of Theorem~\ref{thm:form}, if $E$ has $x_1+x_2$ as a factor, then either \begin{multline*} E/(x_3\cdots x_n)= x_1x_2\cdot\bigl( 1+(a/x_1+(a+c)/x_2)\cdot(bx_1+(b-c^{-1})x_2) \bigr) \\ = (x_1+x_2)\bigl((a+c)bx_1+a(b-c^{-1})x_2\bigr) \end{multline*} for some $a,b,c\in\barFq$ with $c\neq 0$, or \begin{multline*} E/(x_4\cdots x_n)= x_1x_2x_3\cdot\bigl( 1+(a/x_1+a/x_2-1/x_3)\cdot(bx_1+bx_2-x_3) \bigr) \\ = (x_1+x_2)(abx_2x_3+abx_1x_3+bx_1x_2-ax_3^2), \end{multline*} up to permuting the indeterminates $x_3,\ldots,x_n$. \end{lemma} \begin{proof} Any non-multilinear factor of $F$ cannot involve either $x_1$ or $x_2$, and hence \[ F=x_1\cdots x_n\cdot\sum_{i=1}^n f_i/x_i+x_3\cdots x_n\cdot\sum_{j=3}^n f'_jx_j. \] Then the product $E=FG$ has no term of the form $x_1\cdots x_n\cdot x_j/x_i$, and so we have $a_ib_j=0$ whenever $i,j>2$ and $i\neq j$. This implies $a_i=b_i=0$ for all indices $i>2$ except possibly one, and possibly after permuting the indeterminates $x_3,\ldots,x_n$ we may assume that $a_i=b_i=0$ for $i>3$. Therefore, each of the indeterminates $x_4,\ldots,x_n$ appears in each monomial of $E$ with exponent exactly one, and hence $f_i=f'_i=0$ for $i>3$. Thus we have $F=(f_1x_2x_3+f_2x_1x_3+f_3x_1x_2+f'_3x_3^2)\cdot x_4\cdots x_n$, and comparing coefficients of each term on both sides of the equation $FG=E$ we find \begin{gather*} f_2=a_2b_1, \quad f_3=a_3b_1, \quad f_1=a_1b_2, \quad f_3=a_3b_2, \\ f_1+f_2=1+a_1b_1+a_2b_2+a_3b_3, \quad f'_3=a_2b_3, \quad f'_3=a_1b_3. \end{gather*} Substituting the first and third equation of the set into the fifth one turns that into $(a_1-a_2)(b_1-b_2)+a_3b_3+1=0$. Hence if $a_3b_3=0$ then $a_1\neq a_2$ and $b_1\neq b_2$, whence $f_3=f'_3=0$, and so \[ F=\bigl( (a+c)bx_1+a(b-c^{-1})x_2 \bigr)\cdot x_3\cdots x_n, \] where we have set $a:=a_1$, $b:=b_1$, and $c:=a_2-a_1=(b_1-b_2)^{-1}$. However, if $a_3b_3\neq 0$, then the displayed equations yield $a_1=a_2$, $b_1=b_2$, and $b_3=-1/a_3$, whence $f_1=f_2=a_1b_1$, $f_3=a_3b_1$, $f'_3=-a_1/a_3$, that is, \[ F=(abx_2x_3+abx_1x_3+bx_1x_2-ax_3^2)\cdot x_4\cdots x_n \] after setting $a:=a_1/a_3$ and $b:=a_3b_1$. \end{proof} \subsection{General plan of the proof} Because of Lemmas~\ref{lemma:linear} and~\ref{lemma:exceptions}, in order to prove Theorem~\ref{thm:form} it remains to show that if $E=FG$ is any factorisation into non-monomial factors, then either $F$ or $G$ is the product of a monomial and a linear factor. This will be our goal from now on. Hence we may set the following assumptions, which will make the subsequent arguments run smoother. \begin{assumptions}\label{ass} Let the polynomial $E$ of Theorem~\ref{thm:form} be the product of two polynomials $F$ and $G$, of degrees $r>1$ and $n-r>1$, neither of which is a monomial. Assume also that $G$ has no non-trivial monomial factor. \end{assumptions} Our assumptions on the degrees are allowed because otherwise either $F$ or $G$ would be the desired non-monomial linear factor. That $G$ has no non-trivial monomial factor can always be achieved by moving any monomial factor from $G$ to $F$. \subsection{The case where either $F$ or $G$ is not multilinear} If $F$ is multilinear but $G$ is not, then we may interchange the roles of $F$ and $G$, after moving any monomial factor so that Assumptions~\ref{ass} remain satisfied. Hence assume that $F$ is not multilinear. Possibly after renumbering the indeterminates, we may assume that $F$ has a term $x_1^2x_2\cdots x_{r-1}$. Then $G$ is a multilinear polynomial of degree $n-r$ in the remaining $n-r+1$ indeterminates $x_r,\ldots,x_n$, otherwise Property~\ref{property} would be contradicted. Because $G$ has no non-trivial monomial factor according to Assumptions~\ref{ass}, each term $x_r\cdots x_n/x_i$ with $i\ge r$ appears in $G$ with a nonzero coefficient. In turn, Property~\ref{property} implies that any non-multilinear term of $F$ can only involve the indeterminates $x_1,\ldots,x_{r-1}$. If some (multilinear) term of $F$ involved at least two of the indeterminates $x_r,\ldots,x_n$, then because $n-r>1$ that term would share at least two indeterminates with some term of $G$, and this would contradict Property~\ref{property}. Therefore, no term of $F$ involves more than one indeterminate from $x_r,\ldots,x_n$. We have seen earlier that any term of $F$ which involves such indeterminate must be multilinear, and so altogether $F$ can be written in the form \[ F=x_1\cdots x_{r-1}\cdot\sum_{i=1}^{n}f_i\,x_i, \] which provides us with the desired linear factor. \subsection{The case where $F$ and $G$ are both multilinear} Now we may suppose that both $F$ and $G$ are multilinear polynomials. Because of Property~\ref{property} each term of $G$ can share at most one indeterminate with each term of $F$. Any two distinct terms of $F$ must involve together exactly $r+1$ or $r+2$ indeterminates. In fact, if they did involve more, then any term of $G$ would involve at least three of them, and hence it would share at least two indeterminates with at least one of the two terms of $F$ under consideration, contradicting Property~\ref{property}. We deal with those two cases separately. \subsection{The subcase $r+1$} Suppose first that $F$ has at least two terms which together involve $r+1$ indeterminates. After renumbering the indeterminates we may assume that two such terms are $f_r\,x_1\cdots x_r+f_{r+1}\,x_1\cdots x_{r-1}\cdot x_{r+1}$, with $f_r,f_{r+1}\neq 0$. Because of Property~\ref{property} each term of $G$ can share at most one indeterminate with each of them, and hence \begin{equation}\label{eq:G} G= \sum_{i=1}^{r+1} g_i\,x_i\cdot x_{r+2}\cdots x_n +\sum_{j=r+2}^n g'_j\,x_r\cdots x_n/x_j \end{equation} for some scalars $g_i,g'_i$. An alternative choice of notation would be restricting the former summation range to $i<r$ and extending the latter to $i\ge r$, provided we set $g'_r=g_{r+1}$ and $g'_{r+1}=g_r$. We conveniently allow this double notation in what follows. Comparing the coefficients of corresponding monomials on both sides of the equality $FG=E$ we find, in particular, \begin{equation}\label{eq:g} f_rg_i=a_{r+1}b_i \quad\text{and}\quad f_{r+1}g_i=a_rb_i \quad\text{for $i<r$}, \end{equation} \begin{equation}\label{eq:g'} f_rg'_j=a_jb_r \quad\text{and}\quad f_{r+1}g'_j=a_jb_{r+1} \quad\text{for $j>r+1$.} \end{equation} In fact, because $F$ is multilinear any term in the product $FG$ where $x_i$ appears with exponent two, for some $i<r$, can only arise in one way, as the product of the term of $G$ with coefficient $g_i$ and a uniquely determined term of $F$, necessarily one of the two considered above. A similar argument applies to any term in the product $FG$ which misses the indeterminate $x_j$, where $j>r+1$. Because $F$ is multilinear, Equation~\eqref{eq:G} implies that $E=FG$ has no term of the form $x_1\cdots x_n\cdot x_i/x_j$ with $i<r$ and $j>r+1$, which means $a_jb_i=0$. Hence either $b_i=0$ for all $i<r$, or $a_j=0$ for all $j>r+1$. However, the latter together with Equation~\eqref{eq:g} yields that $g'_j=0$ for $j>r+1$, whence $G=x_{r+2}\cdots x_n\cdot\sum_{i=1}^{r+1} g_ix_i$ has a non-trivial monomial factor, against Assumptions~\ref{ass}. We conclude that $b_i=0$ for $i<r$, and Equation~\eqref{eq:g} yields $g_i=0$ for $i<r$, and so $G=\sum_{j=r}^n g'_j\,x_r\cdots x_n/x_j$. Because $G$ has no non-trivial monomial factor we have $g'_j\neq 0$ for $j\ge r$. This implies that no term of $F$ can involve more than one indeterminate from the set $\{x_r,\ldots,x_n\}$, and hence \[ F=x_1\cdots x_{r-1}\cdot\sum_{j=r}^nf_j\,x_j, \] providing us with the desired linear factor. \subsection{The subcase $r+2$} Now we may assume that each two distinct terms of $F$ involve together exactly $r+2$ indeterminates. We will deduce a contradiction. Possibly after renumbering the indeterminates we may assume that two of the terms of $F$ are (nonzero) scalar multiples of $x_1x_2\cdot x_5\cdots x_{r+2}$ and $x_3x_4\cdot x_5\cdots x_{r+2}$ (to be appropriately interpreted in case $r=2$). Because each term of $G$ involves exactly $n-r$ indeterminates, and can share at most one indeterminate with each of those two terms of $F$ according to Property~\ref{property}, it must involve all of $x_{r+3},\ldots,x_n$, and exactly one indeterminate from each of the two sets $\{x_1,x_2\}$ and $\{x_3,x_4\}$. Because $G$ has no non-trivial monomial factor we have $r=n-2$, and hence \[ G=g_{13}x_1x_3+g_{14}x_1x_4+g_{23}x_2x_3+g_{24}x_2x_4 \] for certain scalars $g_{ij}$. Again because $G$ has no non-trivial monomial factor we have either $g_{13}g_{24}\neq 0$ or $g_{14}g_{23}\neq 0$. After possibly exchanging $x_3$ and $x_4$ we may assume the former. Now according to Property~\ref{property} each term of $F$ can share at most one indeterminate with each term of $G$, and this leaves only $x_1x_4\cdot x_5\cdots x_n$ and $x_2x_3\cdot x_5\cdots x_n$ as possible terms of $F$ besides those two assumed from the start. However, each of these must be excluded because together with one of the initial terms it involves only $n-1=r+1$ indeterminates, rather than $r+2$. Hence we have \[ F=( f_{12}x_1x_2+ f_{34}x_3x_4 )\cdot x_5\cdots x_n, \] with $f_{12}f_{34}\neq 0$. Comparing coefficients of corresponding terms in the equality $FG=E$ we find $f_{12}g_{13}=a_4b_1$ and $f_{34}g_{13}=a_2b_3$, whence $a_2b_1\neq 0$. This means that $FG$ has a nonzero term $x_1^2x_3x_4\cdot x_5\cdots x_n$, which is clearly impossible. This contradiction completes our proof of Theorem~\ref{thm:form}.

https://arxiv.org/abs/1007.1760

Factorization of banded permutations

We consider the factorization of permutations into bandwidth 1 permutations, which are products of mutually nonadjacent simple transpositions. We exhibit an upper bound on the minimal number of such factors and thus prove a conjecture of Gilbert Strang: a banded permutation of bandwidth $w$ can be represented as the product of at most $2w-1$ permutations of bandwidth 1. An analogous result holds also for infinite and cyclically banded permutations.

\section{Introduction}\label{section:intro} Computational efficiency very often requires us to represent matrices as products of certain special, easily computable, matrices using as few factors as possible. Matrices of bounded bandwidth are often seen in practical applications. In \cite{Strang1} and \cite{Strang2} Gilbert Strang shows that when a matrix and its inverse are of bandwidth $w$, it can always be represented as a product of $O(w^2)$ such matrices of bandwidth $w=1$. In particular this bound is independent of the size of the given matrix. He also conjectures that for permutation matrices this bound is actually $2w-1$. In this paper we will prove this conjecture. This conjecture has been proven independently later also by Albert,Li,Strang and Yu in \cite{AlbertLiStrangYu} and Ezerman and Samson in \cite{SamsonEzerman}. A \textbf{matrix of bandwidth $w$} is a matrix $A$, whose nonzero entries lie within distance $w$ from the main diagonal: $A_{i,j}=0$ whenever $\,|\,i-j\,|\,>w$. In particular, a banded permutation matrix $P$ is a $0-1$ matrix with exactly one $1$ in each row and column and such that $P_{i,j}=0$ if $\,|\,i-j\,|\,>w$. The matrix $P$ corresponds to the permutation $\pi$ defined as $\pi_i=j$ for $P_{i,j}=1$ and vice versa. So $\pi$ is of width $w$ if $\,|\,\pi_i-i\,|\, \leq w$ for every $i$. Our main result is the following. \begin{theorem}\label{mainthrm} For any permutation $\pi$, let $M=\{\pi_j-i| i<j, \pi_i>\pi_j\}$ and let $m=\#M$ be the number of different elements in the set $M$. Then there exist at most $m$ bandwidth 1 permutations $\rho^j$, i.e. $\,|\, \rho^j_i-i\,|\,\leq 1$, such that $\pi=\rho^1\rho^2\cdots$. \end{theorem} If $\pi$ is a permutation of bandwidth $w$ then for the elements of $M$ we have that $\pi_j-i<\pi_i-i\leq w$ and $\pi_j - i> \pi_j - j\geq -w$, so $M\subset \{-w+1,\ldots,w-1\}$. Thus $M$ has no more than $2w-1$ elements and the conjecture follows immediately. \begin{cor}[\textbf{Strang's conjecture}]\label{Strangsconj} If $\pi$ is a permutation of bandwidth $w$ then there exist at most $2w-1$ bandwidth 1 permutations whose product is $\pi$. Moreover,the bound $2w-1$ is exact. \end{cor} A possible extension of banded matrices, also considered by Strang in \cite{Strang2}, are infinite and cyclically banded matrices. Cyclically banded matrices are $n\times n$ matrices $A$, such that $A_{i,j} =0$ if $w<\,|\,i-j\,|\,<n-w$. Here any matrix would have width $n/2$ and so we will require that $w\leq n/2$. In Section \ref{section:cyclic} we consider the analogous question referring to their factorization into cyclically banded matrices of bandwidth 1. We will use the notion of reduced decomposition of a permutation and its visualization called a wiring diagram. We would like to thank Alex Postnikov for suggesting their use. The proofs will rely on the construction of special wiring diagram we call a hook wiring diagram. \section{Hook wiring diagrams} We will consider the simple generators of the symmetric group $S_n$ as a Coxeter group. We will represent a permutation as a certain product of such simple transpositions which will be grouped into the desired bandwidth 1 factors. A simple transposition $s_i=(i,i+1)$ exchanges the $i$th and $(i+1)$st element. As an element of the symmetric group $S_n$, $s_i$ is equal to the permutation $1,2,\ldots, i-1, i+1, i, i+2,\ldots, n$. Note that $i$ determines the transposition, we will say that \textbf{$i$ is the index of the transposition $s_i$}. A reduced decomposition of $\pi$ is a product $s_{i_1}s_{i_2}\cdots s_{i_l}=\pi$ of such transpositions of minimal possible length $l$; see \cite{Coxeter} for the general facts. It follows by inspection of the possible cases that every bandwidth 1 permutation is a product of mutually nonadjacent simple transpositions $s_i$, where two transpositions are adjacent if their indices are consecutive numbers. A \textbf{wiring diagram} (originally appearing in \cite{Goodman}) of a reduced decomposition $s_{i_1}s_{i_2}\cdots s_{i_l}=\pi$ is a planar configuration of $n$ (pseudo-)lines $L_1,\ldots,L_n$ between two columns of the numbers $1,2,\ldots,n$ with the following properties: \begin{itemize} \item Line $L_i$ starts at $i$ and ends at $\pi_i$. \item No two lines intersect more than once and no three lines intersect at a point. \end{itemize} Each wiring diagram depicts a reduced decomposition $s_{i_1}s_{i_2}\cdots s_{i_l}=\pi$ via the following correspondence. Through every intersection between the lines draw a perpendicular ``dashed'' line (as in Figure \ref{example}). Counting from left to right assign to the $k$th dashed line the simple transposition $s_{i_k}=(i_k,i_{k+1})$, whose index $i_k$ is equal to 1 plus the total number of (pseudo-)lines $L_r$ that cross that dashed line above the intersection point. \begin{figure}[ht!] \centering \includegraphics[width=3in]{"wiring_example"} \caption{ Wiring diagram of $\pi=s_2s_4s_3s_1s_4 = 25143$. } \label{example} \end{figure} Figure \ref{example} shows the wiring diagram for $\pi = (2,3)(4,5)(3,4)(1,2)(4,5)$. Notice that line $L_i$ ``carries" the number $i$. A thin vertical slice of a wiring diagram represents an intermediate permutation with the position of $i$ being the relative vertical position of $L_i$ with respect to the other lines at this slice. Two lines crossing simply means that we exchange two adjacent entries. The number of lines vertically above that crossing plus 1 is exactly the index $i$ of the corresponding simple transposition $s_i=(i,i+1)$. \begin{figure}[ht!] \centering \includegraphics[width=2.7in]{"banded_wiring"} \caption{Hook diagram for the permutation $\pi=5$,$4$,$7$,$1$,$9$,$2$,$3$, $10$,$8$,$6$ of bandwidth $w=4$. The dashed lines depict the diagonals $c-r=-3,-2,-1,0,1,2,3$. The numbers at the intersections are the indices of the corresponding simple transposition, e.g. $2$ corresponds to $s_2=(2,3)$.}\label{bandedexample} \end{figure} For any permutation $\pi$ we can also draw (see Figure \ref{bandedexample}) what we'll call a \textbf{hook diagram}. Consider a square grid bounded by $(0,0)$ in the top left corner and vertical and horizontal rays marked with $1,2,\ldots$ going down and to the right following the indexing convention for matrices; so that a point of coordinates $(r,c)$ is at the $r$-th row (counting from the top) and $c$-th column (counting left to right). Place a dot at the points $(i,\pi_i)$ on the grid and connect $(i,\pi_i)$ with $(i,0)$ and $(0,\pi_i)$ by two segments. This way the dots would be at the places of the ones in the permutation matrix of $\pi$ and each $i$ will be connected to the corresponding $\pi_i$ by a hook with corner at the dot $(i,\pi_i)$. \begin{figure}[ht!] \centering \includegraphics[width=1.5in]{"banded_wiring_stretched"} \caption{The wiring diagram obtained from the hook diagram for the permutation $\pi=5,4,7,1,9,2,3,10,8,6$. The numbers at the intersections indicate the indices of the corresponding simple transposition.}\label{bandedwiring} \end{figure} Notice that a hook diagram turns readily into a wiring diagram by extending the horizontal segments through $(0,i)$ and the vertical segments through $(j,0)$, then rotating by $-45^{\circ}$ as shown in Figure \ref{bandedwiring}. The line $L_i$ would be the rotated extended hook through the points $(0,i),(\pi_i,i),(\pi_i,0)$. To determine the index of the simple transposition corresponding to a crossing of $L_i$ and $L_j$ we need to count the number of lines in a thin strip vertically above that crossing in the rotated extended diagram. Assume $i<j$, so since $L_i$ and $L_j$ cross we must have $\pi_i>\pi_j$. The number of lines above the crossing and hence the index will be $r=i+\pi_j - \#\{p:p<i, \pi_p<\pi_j\}+1$, so the transposition is $s_r=(r,r+1)$. \begin{proof}[Proof of Theorem \ref{mainthrm}] Draw the hook diagram of $\pi$ and interpret it as a wiring diagram as described above. We can now read off a reduced decomposition from the hook-wiring diagram as follows. To every intersection of two hooks assign the number $i$ of the corresponding simple transposition as explained above. Let $\sigma^{k}$ be the product of the transpositions on the $k$-th diagonal, i.e. $c-r=k$ (see Figure \ref{bandedexample}), where $k=-n+1,\ldots,n-1$. We have $\sigma^{k} = s_{i_1}\cdots s_{i_l}$ where $i_1,\ldots,i_l$ are the numbers(indices) at the crossings on the $k$th diagonal. These numbers are at least 2 apart, so we have $\sigma^{k}_i = i$ if $i\not \in \{i_1,i_1+1,\ldots,i_l,i_l+1\}$, $\sigma^{k}_{i_j}=i_j+1$ and $\sigma^{k}_{i_j+1}=i_j$ for $j=1,\ldots,l$. Then $\sigma^{k}$ is of bandwidth $1$. The diagram gives a reduced decomposition of $\pi$, where the transpositions corresponding to intersections on the same diagonal $c-r=k$ appear before the ones on the next diagonal, while the relative order of these transpositions on the same diagonal does not matter since they commute. Since their product is $\sigma^k$ we have $\pi = \sigma^{-n+1}\sigma^{-n+2}\cdots \sigma^{n-1}$. Notice that $\sigma^k$ is not trivial if and only if there is an intersection on this diagonal. There is an intersection between $L_i$ and $L_j$, $(i<j)$, if and only if $\pi_i>\pi_j$, and the intersection point is $(i,\pi_j)$. Thus diagonal $k$ has an intersection if and only if $k\in M$ and the number of nontrivial bandwidth 1 factors is equal to the cardinality of $M$, $m=\#M$. If $M=\{k_1,\ldots,k_m\}$, then the nontrivial factors are $\sigma^{k_i}$ for $i=1,\ldots,m$ and we set $\rho^i = \sigma^{k_i}$ to obtain $\pi=\rho^1\rho^2\ldots$ as desired. \end{proof} \begin{proof}[Proof of Corollary \ref{Strangsconj}] As remarked earlier, if $\pi$ is banded then $M\subset\{-w+1,\ldots,w-1\}$, so $\# M\leq 2w-1$ and Strang's conjecture follows from Theorem \ref{mainthrm}. To show that $2w-1$ is the exact bound, consider the permutation $\sigma=(w+1)(w+2)\ldots(2w)123\ldots w \ldots$ of width $w$, where the last $\ldots$ mean the identity $\sigma_i=i$ for $i>2w$. Before we show that this particular $\sigma$ cannot be factored into less than $2w-1$ permutations of bandwidth 1, we need to make a few general observations. \begin{figure}[ht!] \centering \includegraphics[height=1.5in]{"extreme"} \caption{ Any wiring diagram of $\sigma = (w+1)\ldots(2w)12\ldots w \ldots$ is homotopy equivalent to this one. Here $w=3$ and $\sigma = 456123789$.}\label{extreme} \end{figure} For any permutation $\pi$, let $k$ be the minimal number for which $\pi=\pi^{(1)}\cdots \pi^{(k)}$, where $\pi^{(i)}$ are permutations of bandwidth 1. Then there exists a reduced decomposition of $\pi=s^{(1)}_{i_1}s^{(1)}_{i_2}\cdots s^{(k)}_{i_l}$, such that $\pi^{(i)}$ is the product of the $i$th block of transpositions. We will show that there is always such reduced decomposition by decreasing the number of simple transpositions in it. Writing $\pi^{(i)}=s_{i_1}\cdots s_{i_m}$ as a product of transpositions we still have a decomposition of $\pi$ into simple transpositions. We can depict this decomposition graphically like a wiring diagram, without requiring that two lines intersect at most once. The assertion that the decomposition of $\pi^{(i)}=s_{i_1}\cdots s_{i_m}$ is not reduced is equivalent to two lines $L'$ and $L''$ intersecting at least twice at places $r$ and $p$ corresponding to $s_{i_r}$ and $s_{i_p}$. Let $L'=A'B'C'$ and $L''=A''B''C''$ where $A,B,C$ are the portions of the lines obtained after cutting at the two intersections. Substituting $L'$ and $L''$ with $A'B''C'$ and $A''B'C''$ respectively gives us another wiring diagram of $\pi$ for the decomposition $\pi=s_{i_1}\cdots \hat{ s_{i_r} } \cdots \hat{s_{i_p}} \cdots s_{i_m}$. Removing $s_{i_r}$ and $s_{i_p}$ from the $\pi^{(i)}$s to which they belonged gives another factorization of $\pi$ into at most $k$ permutations of width 1 with a smaller number of simple transpositions. Continuing this way we will reach the length $l$ of $\pi$ forcing the underlying decomposition into simple transpositions to be reduced. We can thus assume that our particular $\sigma=\sigma^{(1)}\cdots \sigma^{(k)}$ gives a reduced decomposition. Consider its wiring diagram as depicted in Figure \ref{extreme}: since the transpositions in each $\sigma^{(i)}$ are nonadjacent we can draw the corresponding intersections on the same vertical line. Thus every path from some $i$ to some $\sigma_j$ will pass through at most $k$ intersections. Notice that any wiring diagram of $\sigma$ can be deformed (is ambiently isotopic) to a $w\times w$ grid rotated $45^{\circ}$ like the diagram on Figure \ref{extreme}. Then every path joining $w+1$ with $\sigma_1=w+1$ has $2w-1$ intersection points and so $k\geq 2w-1$. \end{proof} Since our proof is constructive, it leads to an algorithm for the decomposition: find the intersection points in the hook diagram and group them according to the diagonal to which they belong. Let $I_k$ be the set of intersection points on the $k$th diagonal. Assume the inverse permutation $\pi^{-1}$ is known. Then the procedure is as follows: \indent For $i$ from $1$ to $n$: \indent \indent $p:=\pi_i$ \indent \indent For $j$ from $1$ to $p-1$: \indent \indent \indent If $\pi^{-1}_j>i$, then $I_{j-i} \leftarrow (i,j)$ In order to determine which transposition these intersections correspond to, notice that the number of lines $L_r$ intersecting the segment between $(i,j)$ and the origin, and thus the index of the transposition, is $i-1+j-1-\#\{t\mid t <i \,,\,\pi_t <j\}$, which we can count within this algorithm also. Let $s[i,j]=\#\{t\mid t <i \,,\,\pi_t <j\}$, then $$s[i,j+1]=s[i,j] + ((\pi^{-1}_j<i)),$$ where $((statement))$ denotes the logical value $0/1$ of the statement. \section{Infinite and cyclically banded permutations.}\label{section:cyclic} We now consider an extension of banded matrices and their corresponding permutations $\pi=\pi_1\ldots\pi_n$, as defined in Section \ref{section:intro}. In this case bandwidth 1 encompasses more permutations and thus allowed factors. We have a simple transposition $s_0=s_n$ exchanging $\pi_1$ and $\pi_n$, and corresponding to the bandwidth 1 cyclic matrix $A$ with $A_{1n}=A_{n1}=1$. The other additional factor is the shift $S$, acting by cyclic shift on $\pi$ as $S\pi=\pi_2\pi_3\ldots\pi_n\pi_1$, with corresponding matrix $S$ given by $S_{i,i+1}=1$ and 0 otherwise. A bandwidth 1 permutation is either a shift $S$ or $S^{-1}$ or a product of pairwise nonadjacent modulo $n$ simple transpositions $s_0,\cdots,s_{n-1}$. \begin{figure}[ht!] \centering $A = \left[\begin{matrix} 0 &0 &0 &0 & 0& 1\\0&0&0&0&1&0\\0&1&0&0&0&0\\0&0&0&1&0&0\\0&0&1&0&0&0\\1&0&0&0&0&0\end{matrix}\right]$, $B=\begin{tabular}{c@{}c@{\hspace{12pt}}cccccc@{\hspace{12pt}}c@{}c} \dots &0 &1&0&0&0&0&0 &0 &\dots\\[10pt] 0 &1 &0&0&0&0&0&0 &0 &0\\ 1 &0 &0&0&0&0&0&0 &0 &0\\ 0 &0 &0&1&0&0&0&0 &0 &0\\ 0 &0 &0&0&0&1&0&0 &0 &0\\ 0 &0 &0&0&1&0&0&0 &0 &0\\ 0 &0 &0&0&0&0&0&0 &1 &0\\[10pt] 0 &0 &0&0&0&0&0&1 &0 &0\\ \dots&0 &0&0&0&0&1&0 &0 &\dots \end{tabular}$ \caption{ The matrix $A$ of the cyclic banded permutation $\pi = 652431$ of bandwidth $w=3$ and the corresponding infinite $B=\phi(A)$.}\label{cyclicmatrix} \end{figure} A cyclic banded matrix can also be interpreted as a doubly infinite periodic matrix of period $(n,n)$ in the following way. For any $n\times n$ matrix $A$ define $\phi(A) = B$, where $B$ is a doubly infinite matrix given by $B_{i,j}=A_{i \pmod{n},j \pmod{n}}$ for $\,|\,i-j\,|\,\leq n/2$ and 0 otherwise, see Figure \ref{cyclicmatrix} for an example. The map $\phi$ from cyclic banded matrices of bandwidth $w$ to banded doubly infinite matrices of period $(n,n)$ is an isomorphism. We will thus consider the problem of factorization of banded doubly infinite periodic matrices and their corresponding infinite permutations. \begin{figure}[ht!] \centering \includegraphics[height=2in]{"cyclic_hook"} \qquad \includegraphics[height=2in]{"cyclic_wiring"} \qquad \includegraphics[height=2in]{"cyclic_untangled"} \caption{ The hook diagram for $\pi=652431$, produced from its infinite matrix $B$, the corresponding wiring diagram which gives factors $\sigma^1=s_1, \sigma^2=s_0s_4,\sigma^3=s_1$ and the resulting trivial wiring diagram after ``untangling'' giving a shift $S^1$.}\label{cyclicwiring} \end{figure} We can now form the hook wiring diagram of the infinite periodic banded permutation matrix as in the previous Section and as it is shown on Figure \ref{cyclicwiring}. Again the intersections on each diagonal correspond to a bandwidth 1 factor and we can proceed to ``untangle'' them, i.e. multiply by simple transpositions. Let $\sigma^i$ be the product of the simple transpositions corresponding to intersections on the $i$th diagonal, where each transposition is represented by its infinite periodic matrix and thus the number of these transpositions will be finite. After multiplication by $\sigma^1\cdots\sigma^{2w-1}$ we will have a trivial wiring diagram where no lines intersect. Unlike in the finite case where a monotone bijective map $[1,\ldots,n]\rightarrow [1,\ldots,n]$ must be the identity, in the infinite case of $\mathbb{Z}$ these could be shifts, so the permutation/matrix corresponding to a trivial wiring diagram will be $S^k$ for some $k$. We will now show that $k\leq w$. Define the \textbf{shifting index} of an infinite periodic permutation matrix $B$ as follows. Let $p:= \# \{ B_{i,j}=1\,|\,i \leq n \text{ and } n+1\leq j \}$ and $q:=\# \{ B_{i,j}=1\,|\, n+1\leq i \text{ and } j \leq n\}$. Alternatively these are the the number of ones in the upper right $w\times w$ triangle of the original $A$ and the number of ones in the lower left such triangle. Let the \textbf{shifting index} of $B$ be $\si(B) := p-q$. We have that $\si(\id)=0$, $\si(S)=1$, $\si(S^{-1})=-1$. We also have that $\si(SB)=1+ \si(B)$ and $\si(S^{-1}B)=-1+\si(B)$, since $S$ acts on $B$ by shifting one row upwards and the entry 1 on row $n+1$ moving to row $n$ either decreases $q$ by 1 or increases $p$ by 1. Thus every permutation matrix factorizes uniquely as $B = S^{\si(B)}\bar{B}$, where $\si(\bar{B})=0$. Moreover, for any simple transposition $s_i$ and its corresponding infinite matrix $E_i$, we have $\si(E_i)=0$ and $\si(E_iB)=\si(B)$. This can be checked by inspection, $i=0\pmod n$ is the only nontrivial case, where it is still obvious that $p-q$ is preserved after switching row $n$ and $n+1$ in $B$. Also $SE_iS^{-1} = E_{i-1}$, where the indexing is modulo $n$ again. Using the usual inversion index we can show by induction on it that every such $B$ is the product of shifts and simple transpositions (note that any permutation matrix without inversions is a diagonal of 1s, i.e. $S^k$ for some $k$). Thus we can always write $B = S^{\si(B)}E_{i_1}\cdots E_{i_l}$. This also shows that $\si(AB) =\si(A)+\si(B)$, under repeated application of $E_iS=SE_{i+1}$. For a matrix $B$ of bandwidth $w$ we have that $p\leq w$ and $q \leq w$, thus $-w\leq \si(B) \leq w$. Moreover, $B=\sigma^1\cdots\sigma^{2w-1} S^k$, so $\si(B) = \si(\sigma^1)+\cdots+\si(\sigma^{2w-1})+k=k$, and thus $\,|\,k\,|\,\leq w$. We thus have the following analogue of Strang's conjecture, Corollary \ref{Strangsconj}. \begin{theorem}\label{cyclic} Let $\pi \in S_n$ be a cyclic banded permutation of bandwidth $w$, i.e. $\,|\,\pi_i-i\,|\,\leq w$ or $\,|\,\pi_i-i\,|\,\geq n-w$, or alternatively an infinite periodic permutation whose matrix is doubly infinite periodic with only nonzero entries in $\,|\,i-j\,|\,\leq w$. Then $\pi =\sigma^1\ldots\sigma^{2w-1}S^k$, where $\sigma^i$ is a cyclic permutation of bandwidth 1, product of nonadjacent simple transpositions, and $S$ is the cyclic shift by 1 with $\,|\,k\,|\,\leq w$. \end{theorem} \vfil \begin{bibdiv} \begin{biblist} \bib{AlbertLiStrangYu}{article}{ author={Albert, C.}, author={Li, C.-K.}, author={Strang,G.}, author={Yu, G.}, title= {Permutations as products of parallel transpositions}, status={ SIAM J. Discrete Math}, year={2011}, issue ={25}, pages={1412-1417}, } \bib{Coxeter}{book}{ author={Bj{\"o}rner, Anders}, author={Brenti, Francesco}, title={Combinatorics of Coxeter groups}, series={Graduate Texts in Mathematics}, volume={231}, publisher={Springer}, place={New York}, date={2005}, } \bib{SamsonEzerman}{article}{ author={Ezerman,Martianus Frederic}, author={Samson, Michael Daniel}, title={Factoring Permutation Matrices Into a Product of Tridiagonal Matrices}, eprint={arXiv:1007.3467}, year={2010}, } \bib{Goodman}{article}{ author={Goodman, Jacob E.}, title={Proof of a conjecture of Burr, Gr\"unbaum, and Sloane}, journal={Discrete Math.}, volume={32}, date={1980}, number={1}, pages={27--35}, } \bib{Strang1}{article}{ author = {Strang, Gilbert}, title = {Fast transforms: Banded matrices with banded inverses}, journal = { Proc. Natl. Acad. Sciences}, date={2010}, } \bib{Strang2}{article}{ author = {Strang, Gilbert}, title = {Groups of banded matrices with banded inverses}, journal = {Proc. Amer.Math.Soc.}, issue={139}, date = {2011}, pages={4255-4264}, } \end{biblist} \end{bibdiv} \end{document}

https://arxiv.org/abs/2001.05992

Provable Benefit of Orthogonal Initialization in Optimizing Deep Linear Networks

The selection of initial parameter values for gradient-based optimization of deep neural networks is one of the most impactful hyperparameter choices in deep learning systems, affecting both convergence times and model performance. Yet despite significant empirical and theoretical analysis, relatively little has been proved about the concrete effects of different initialization schemes. In this work, we analyze the effect of initialization in deep linear networks, and provide for the first time a rigorous proof that drawing the initial weights from the orthogonal group speeds up convergence relative to the standard Gaussian initialization with iid weights. We show that for deep networks, the width needed for efficient convergence to a global minimum with orthogonal initializations is independent of the depth, whereas the width needed for efficient convergence with Gaussian initializations scales linearly in the depth. Our results demonstrate how the benefits of a good initialization can persist throughout learning, suggesting an explanation for the recent empirical successes found by initializing very deep non-linear networks according to the principle of dynamical isometry.

\section{Conclusion} \label{sec:conclu} In this work, we studied the effect of the initialization parameter values of deep linear neural networks on the convergence time of gradient descent. We found that when the initial weights are iid Gaussian, the convergence time grows exponentially in the depth unless the width is at least as large as the depth. In contrast, when the initial weight matrices are drawn from the orthogonal group, the width needed to guarantee efficient convergence is in fact independent of the depth. These results establish for the first time a concrete proof that orthogonal initialization is superior to Gaussian initialization in terms of convergence time. \section{Experiments} \label{sec:experiment} In this section, we provide empirical evidence to support the results in Sections \ref{sec:ortho} and \ref{sec:gaussian}. To study how depth and width affect convergence speed of gradient descent under orthogonal and Gaussian initialization schemes, we train a family of linear networks with their widths ranging from 10 to 1000 and depths from 1 to 700, on a fixed synthetic dataset $(X, Y)$.\footnote{We choose $X \in \mathbb{R}^{1024\times16}$ and $W^* \in \mathbb{R}^{10\times 1024}$, and set $Y = W^*X$. Entries in $X$ and $W^*$ are drawn i.i.d. from ${\mathcal{N}}(0, 1)$.} Each network is trained using gradient descent staring from both Gaussian and orthogonal initializations. In Figure~\ref{fig:heat-maps}, We lay out the logarithm of the relative training loss $\frac{\ell(t)}{\ell(0)}$, using heap-maps, at steps $t=1258$ and $t=10000$. In each heat-map, each point represents the relative training loss of one experiment; the darker the color, the smaller the loss. Figure~\ref{fig:heat-maps} clearly demonstrates a sharp transition from untrainable to trainable (i.e., from red to black) when we increase the width of the network: \begin{itemize} \item for Gaussian initialization, this transition occurs across a contour characterized by a linear relation between width and depth; \item for orthogonal initialization, the transition occurs at a width that is approximately independent of the depth. \end{itemize} These observations excellently verify our theory developed in Sections~\ref{sec:ortho} and \ref{sec:gaussian}. \begin{figure}[t] \begin{center} \begin{subfigure}[b]{0.4\textwidth} \centering \includegraphics[height=4.5cm]{graphs/gaussian-dataset-16-learning_rate-01-steps-1258.pdf} \caption{Gaussian, steps=1258} \label{fig:tanhphasediag} \end{subfigure}% \begin{subfigure}[b]{0.4\textwidth} \centering \includegraphics[height=4.5cm]{graphs/gaussian-dataset-16-learning_rate-01-steps-10000.pdf} \caption{Gaussian, steps=10000} \label{fig:tanhphasediag} \end{subfigure}% \\ \begin{subfigure}[b]{0.4\textwidth} \centering \includegraphics[height=4.5cm]{graphs/orthogonal-dataset-16-learning_rate-01-steps-1258.pdf} \caption{Orthogonal, steps=1258} \label{fig:tanhphasediag} \end{subfigure}% \begin{subfigure}[b]{0.4\textwidth} \centering \includegraphics[height=4.5cm]{graphs/orthogonal-dataset-16-learning_rate-01-steps-10000.pdf} \caption{Orthogonal, steps=10000} \label{fig:tanhphasediag} \end{subfigure}% \end{center} \caption{$\log \frac{\ell(t)}{\ell(0)}$ at $t=1258$ and $t=10000$, for different depth-width configurations and different initialization schemes. Darker color means smaller loss.} \label{fig:heat-maps} \end{figure} To have a closer look into the training dynamics, we also plot ``relative loss v.s. training time'' for a variety of depth-width configurations. See Figure~\ref{fig:loss}. There again we can clearly see that orthogonal initialization enables fast training at small width (independent of depth), and that the required width for Gaussian initialization depends on depth. \begin{figure}[t] \begin{center} \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[height=3.5cm]{graphs/dataset-16-learning_rate-01-depth-50.pdf} \caption{Depth=50} \label{fig:tanhphasediag} \end{subfigure}% \begin{subfigure}[b]{0.33\textwidth} \centering \includegraphics[height=3.5cm]{graphs/dataset-16-learning_rate-01-depth-200.pdf} \caption{Depth=200} \label{fig:tanhphasediag} \end{subfigure}% \begin{subfigure}[b]{0.33 \textwidth} \centering \includegraphics[height=3.5cm]{graphs/dataset-16-learning_rate-01-depth-400.pdf} \caption{Depth=400} \label{fig:tanhphasediag} \end{subfigure}% \end{center} \caption{Relative loss v.s. training time. For each plot, we vary width from 50 (yellow) to 1200 (purple). Solid and dashed lines represent Gaussian (GS) and orthogonal (OT) initializations.} \label{fig:loss} \end{figure} \section{Exponential Curse of Gaussian Initialization} \label{sec:gaussian} In this section, we show that gradient descent with Gaussian random initialization necessarily suffers from a running time that scales exponentially with the depth of the network, unless the width becomes nearly linear in the depth. Since we mostly focus on the dependence of width and running time on depth, we will assume the depth $L$ to be sufficiently large. Recall that we want to minimize the objective $\ell(W_1, \ldots, W_L) = \frac{1}{2} \norm{\alpha W_{L:1} X - Y}_F^2$ by gradient descent. We assume $Y=W^* X$ for some $W^* \in \mathbb{R}^{d_y\times d_x}$, so that the optimal objective value is $0$. For convenience, we assume $\norm{X}_F = \Theta(1)$ and $\norm{Y}_F=\Theta(1)$. Suppose that at layer $i\in[L]$, every entry of $W_i(0)$ is sampled from ${\mathcal{N}}(0, \sigma_i^2)$, and all weights in the network are independent. We set the scaling factor $\alpha$ such that the initial output of the network does not blow up exponentially (in expectation): \begin{equation} \label{eqn:scaling-not-blow-up} \expect{\norm{f(x; W_1(0), \ldots, W_L(0))}^2} \le L^{O(1)} \cdot \norm{x}^2, \quad \forall x\in \mathbb{R}^{d_x}. \end{equation} Note that $\expect{\norm{f(x; W_1(0), \ldots, W_L(0))}^2} = \alpha^2 \prod_{i=1}^L (d_i\sigma_i^2) \norm{x}^2$. Thus \eqref{eqn:scaling-not-blow-up} means $$ \alpha^2 \prod_{i=1}^L (d_i\sigma_i^2) \le L^{O(1)}. $$ We also assume that the magnitude of initialization at each layer cannot vanish with depth: \begin{equation} \label{eqn:init-not-vanish} d_i\sigma_i^2 \ge \frac{1}{L^{O(1)}}, \quad \forall i\in[L]. \end{equation} Note that the assumptions \eqref{eqn:scaling-not-blow-up} and \eqref{eqn:init-not-vanish} are just sanity checks to rule out the obvious pathological cases -- they are easily satisfied by all the commonly used initialization schemes in practice. Now we formally state our main theorem in this section. \begin{thm} \label{thm:gaussian} Suppose $\max\{d_0, d_1, \ldots, d_L\} \le O(L^{1-\gamma})$ for some universal constant $0< \gamma \le 1$. Then there exists a universal constant $c>0$ such that, if gradient descent is run with learning rate $\eta \le e^{cL^\gamma}$, then with probability at least $0.9$ over the random initialization, for the first $e^{\Omega(L^\gamma)}$ iterations, the objective value is stuck between $0.4\norm{Y}_F^2$ and $0.6\norm{Y}_F^2$. \end{thm} Theorem~\ref{thm:gaussian} establishes that efficient convergence from Gaussian initialization is impossible for large depth unless the width becomes nearly linear in depth. This nearly linear dependence is the best we can hope for, since \cite{du2019width} proved a positive result when the width is larger than linear in depth. Therefore, a phase transition from untrainable to trainable happens at the point when the width and depth has a nearly linear relation. Furthermore, Theorem~\ref{thm:gaussian} generalizes the result of \cite{shamir2018exponential}, which only treats the special case of $d_0=\cdots=d_L=1$. \subsection{Proof of \Theoremref{thm:gaussian}} For convenience, we define a scaled version of $W_i$: let $A_i = W_i / (\sqrt{d_i} \sigma_i)$ and $\beta = \alpha \prod_{i=1}^L (\sqrt{d_i} \sigma_i)$. Then we know $\beta \le L^{O(1)}$ and $\alpha W_{L:1} = \beta A_{L:1}$, where $A_{j:i} = A_j\cdots A_i$. We first give a simple upper bound on $\norm{A_{j:i}(0)}$ for all $1\le i \le j \le L$. \begin{lem} \label{lem:init-product-bound} With probability at least $1-\delta$, we have $\norm{A_{j:i}(0)} \le O\left( \frac{L^3}{\delta} \right)$ for all $1\le i \le j \le L$. \end{lem} The proof of Lemma~\ref{lem:init-product-bound} is given in Appendix~\ref{app:proof-gaussian-init-product-bound}. It simply uses Markov inequality and union bound. Furthermore, a key property at initialization is that if $j-i$ is large enough, $\norm{A_{j:i}(0)}$ will become exponentially small. \begin{lem} \label{lem:init-long-product-bound} With probability at least $1-e^{-\Omega(L^\gamma)}$, for all $1\le i \le j \le L$ such that $j-i \ge \frac{L}{10}$, we have $\norm{A_{j:i}(0)} \le e^{-\Omega(L^\gamma)}$. \end{lem} \begin{proof} We first consider a fixed pair $(i, j)$ such that $j-i \ge \frac{L}{10}$. In order to bound $\norm{A_{j:i}(0)}$, we first take an arbitrary unit vector $v \in \mathbb{R}^{d_{i-1}}$ and bound $\norm{A_{j:i}(0) v}$. We can write $\norm{A_{j:i}(0) v}^2 = \prod_{k=i}^j Z_k$, where $Z_k = \frac{\norm{A_{k:i}(0) v}^2}{\norm{A_{k-1:i}(0)v}^2}$. Note that for any nonzero $v' \in \mathbb{R}^{d_{k-1}}$ independent of $A_k(0)$, the distribution of $ d_k \cdot \frac{\norm{A_k(0) v'}^2}{\norm{v'}^2} $ is $\chi^2_{d_k}$. Therefore, $Z_i, \ldots, Z_j$ are independent, and $d_k Z_k \sim \chi^2_{d_k}$ ($ k =i, i+1, \ldots, j$). Recall the expression for the moments of chi-squared random variables: $\expect{Z_k^\lambda} = \frac{2^\lambda \Gamma(d_k/2 + \lambda)}{d_k^\lambda \Gamma(d_k/2)}$ ($\forall \lambda>0$). Taking $\lambda=\frac12$ and using the bound $\frac{\Gamma(a+\frac12)}{\Gamma(a)} \le \sqrt{a-0.1}\, (\forall a \ge \frac12) $ \citep{qi2012bounds}, we get $\expect{\sqrt{Z_k}} \le \sqrt{\frac{2(d_k/2-0.1)}{d_k}} = \sqrt{1 - \frac{0.2}{d_k}} \le 1 - \frac{0.1}{d_k}$. Therefore we have \begin{align*} \expect{\sqrt{\prod\nolimits_{k=i}^j Z_k}} \le \prod\nolimits_{k=i}^j \left( 1 - \frac{0.1}{d_k} \right) \le \left( 1 - \frac{0.1}{O(L^{1-\gamma})} \right)^{j-i+1} \le \left( 1 - \Omega(L^{\gamma-1}) \right)^{\frac{L}{10}} = e^{-\Omega(L^\gamma)}. \end{align*} Choose a sufficiently small constant $c'>0$. By Markov inequality we have $\Pr\left[ \sqrt{\prod\nolimits_{k=i}^j Z_k} > e^{-c'L^\gamma} \right] \le e^{c'L^\gamma} \expect{\sqrt{\prod\nolimits_{k=i}^j Z_k}} \le e^{c'L^\gamma} e^{-\Omega(L^\gamma)} = e^{-\Omega(L^\gamma)}$. Therefore we have shown that for any fixed unit vector $v \in \mathbb{R}^{d_{i-1}}$, with probability at least $1- e^{-\Omega(L^\gamma)}$ we have $\norm{A_{j:i}(0)v} \le e^{-\Omega(L^\gamma)}$. Next, we use this to bound $\norm{A_{j:i}(0)}$ via an $\epsilon$-net argument. We partition the index set $[d_{i-1}]$ into $[d_{i-1}] = S_1 \cup S_2 \cup \cdots \cup S_q$ such that $|S_l| \le L^{\gamma/2} \, (\forall l\in[q])$ and $q = O(\frac{d_{i-1}}{L^{\gamma/2}})$. For each $l\in[q]$, let ${\mathcal{N}}_l$ be a $\frac{1}{2}$-net for all the unit vectors in $\mathbb{R}^{d_{i-1}}$ with support in $S_l$. Note that we can choose ${\mathcal{N}}_l$ such that $|{\mathcal{N}}_l| = e^{O(|S_l|)} = e^{O(L^{\gamma/2})}$. Taking a union bound over $ \cup_{l=1}^q {\mathcal{N}}_l $, we know that $\norm{A_{j:i}(0)v} \le e^{-\Omega(L^\gamma)} \norm{v}$ simultaneously for all $v \in \cup_{l=1}^q {\mathcal{N}}_l $ with probability at least $1 - \left( \sum_{l=1}^q |{\mathcal{N}}_l| \right) e^{-\Omega(L^\gamma)} \ge 1 - q \cdot e^{O(L^{\gamma/2})} e^{-\Omega(L^\gamma)} = 1 - e^{-\Omega(L^\gamma)}$. Now, for any $u\in \mathbb{R}^{d_{i-1}}$, we write it as $u = \sum_{l=1}^q a_l u_l$ where $a_l $ is a scalar and $u_l$ is a unit vector supported on $S_l$. By the definition of $\frac{1}{2}$-net, for each $l\in[q]$ there exists $v_l \in {\mathcal{N}}_l$ such that $\norm{v_l-u_l}\le \frac{1}{2}$. We know that $\norm{A_{j:i}(0)v_l} \le e^{-\Omega(L^\gamma)} \norm{v_l}$ for all $l\in[q]$. Let $v = \sum_{l=1}^q a_l v_l$. We have \begin{align*} \norm{A_{j:i}(0)v} &\le \sum\nolimits_{l=1}^q |a_l| \cdot \norm{A_{j:i}(0)v_l} \le \sum\nolimits_{l=1}^q |a_l| \cdot e^{-\Omega(L^\gamma)} \norm{v_l} \le e^{-\Omega(L^\gamma)} \sqrt{q \cdot \sum\nolimits_{l=1}^q a_l^2 \norm{v_l}^2} \\ &= \sqrt{q} e^{-\Omega(L^\gamma)} \norm{v} = e^{-\Omega(L^\gamma)} \norm{v}. \end{align*} Note that $\norm{u-v} = \norm{\sum_{l=1}^q a_l (u_l-v_l)} = \sqrt{\sum_{l=1}^q a_l^2 \norm{u_l-v_l}^2} \le \sqrt{\frac14 \sum_{l=1}^q a_l^2 } = \frac12 \norm{u}$, which implies $\norm{v} \le \frac32\norm{u}$. Therefore we have \begin{align*} \norm{A_{j:i}(0)u} &\le \norm{A_{j:i}(0)v} + \norm{A_{j:i}(0)(u-v)} \le e^{-\Omega(L^\gamma)} \norm{v} + \norm{A_{j:i}(0)} \cdot \norm{u-v} \\ &\le e^{-\Omega(L^\gamma)}\cdot \frac32 \norm{u} + \norm{A_{j:i}(0)} \cdot \frac12 \norm{u} = e^{-\Omega(L^\gamma)} \norm{u} + \norm{A_{j:i}(0)} \cdot \frac12 \norm{u}. \end{align*} The above inequality is valid for any $u\in \mathbb{R}^{d_{i-1}}$. Thus we can take the unit vector $u$ that maximizes $\norm{A_{j:i}(0)u} $. This gives us $\norm{A_{j:i}(0)} \le e^{-\Omega(L^\gamma)} + \frac12 \norm{A_{j:i}(0)}$, which implies $\norm{A_{j:i}(0)} \le e^{-\Omega(L^\gamma)} $. Finally, we take a union bound over all possible $(i, j)$. The failure probaility is at most $L^2 e^{-\Omega(L^\gamma)} = e^{-\Omega(L^\gamma)}$. \end{proof} The following lemma shows that the properties in Lemmas~\ref{lem:init-product-bound} and~\ref{lem:init-long-product-bound} are still to some extent preserved after applying small perturbations on all the weight matrices. \begin{lem} \label{lem:perturbation-bound} Suppose that the initial weights satisfy $\norm{A_{j:i}(0)} \le O(L^3)$ for all $1\le i \le j \le L$, and $\norm{A_{j:i}(0)} \le e^{-c_1 L^\gamma}$ if $j-i\ge \frac{L}{10}$, where $c_1>0$ is a universal constant. Then for another set of matrices $A_1, \ldots, A_L$ satisfying $\norm{A_i-A_i(0)} \le e^{-0.6c_1 L^\gamma}$ for all $i\in[L]$, we must have \begin{equation} \label{eqn:product-bound} \begin{aligned} &\norm{A_{j:i}} \le O(L^3), \quad \forall 1\le i \le j \le L, \\ &\norm{A_{j:i}} \le O\left( e^{- c_1 L^\gamma}\right), \quad \forall 1\le i \le j \le L, j-i\ge \frac{L}{4}. \end{aligned} \end{equation} \end{lem} \begin{proof} It suffices to show that the difference $A_{j:i} - A_{j:i}(0)$ is tiny. Let $\Delta_i = A_i - A_i(0)$. We have $A_{j:i} = (A_j(0) + \Delta_j) \cdots (A_{i+1}(0) + \Delta_{i+1}) (A_i(0) + \Delta_i)$. Expanding this product, except for the one term corresponding to $A_{j:i}(0)$, every other term has the form $A_{j:(k_s+1)}(0) \cdot \Delta_{k_s} \cdot A_{(k_s-1):(k_{s-1}+1)}(0) \cdot \Delta_{k_{s-1}} \cdots \Delta_{k_1} \cdot A_{(k_1-1):i}(0) $, where $i \le k_1 < \cdots < k_s \le j$. By assumption, each $\Delta_k$ has spectral norm $e^{-0.6c_1 L^\gamma}$, and each $A_{j':i'}(0)$ has spectral norm $O(L^3)$, so we have $\norm{A_{j:(k_s+1)}(0) \cdot \Delta_{k_s} \cdot A_{(k_s-1):(k_{s-1}+1)}(0) \cdot \Delta_{k_{s-1}} \cdots \Delta_{k_1} \cdot A_{(k_1-1):i}(0) } \le \left(e^{-0.6c_1 L^\gamma}\right)^s \left( O(L^3) \right)^{s+1}$. Therefore we have \begin{align*} &\norm{A_{j:i} - A_{j:i}(0)} \le \sum_{s=1}^{j-i+1} \binom{j-i+1}{s} \left(e^{-0.6c_1 L^\gamma}\right)^s \left( O(L^3) \right)^{s+1} \\ \le\,& \sum_{s=1}^{j-i+1} L^s \left(e^{-0.6c_1 L^\gamma}\right)^s \left( O(L^3) \right)^{s+1} \le O(L^3) \sum_{s=1}^\infty \left( O(L^4) e^{-0.6c_1 L^\gamma} \right)^s \le O(L^3) \sum_{s=1}^\infty (1/2)^s = O(L^3), \end{align*} which implies $\norm{A_{j:i}} \le O(L^3)$ for all $1\le i \le j \le L$. The proof of the second part of the lemma is postponed to Appendix~\ref{app:proof-gaussian-perturbation-bound}. \end{proof} As a consequence of \Lemref{lem:perturbation-bound}, we can control the objective value and the gradient at any point sufficiently close to the random initialization. \begin{lem} \label{lem:loss-gradient-bound} For a set of weight matrices $W_1, \ldots, W_L$ with $A_i = W_i/(\sqrt{d_i}\sigma_i)$ that satisfy~\eqref{eqn:product-bound}, the objective and the gradient satisfy \begin{align*} & 0.4\norm{Y}_F^2 < \ell(W_1, \ldots, W_L) < 0.6\norm{Y}_F^2,\\ & \norm{\nabla_{W_i}\ell(W_1, \ldots, W_L)} \le (\sqrt{d_i}\sigma_i )^{-1} e^{-0.9c_1L^\gamma}, \quad \forall i\in[L]. \end{align*} \end{lem} The proof of Lemma~\ref{lem:loss-gradient-bound} is given in Appendix~\ref{app:proof-gaussian-loss-gradient-bound}. Finally, we can finish the proof of \Theoremref{thm:gaussian} using the above lemmas. \begin{proof}[Proof of \Theoremref{thm:gaussian}] From Lemmas~\ref{lem:init-product-bound} and \ref{lem:init-long-product-bound}, we know that with probability at least $0.9$, we have (i) $\norm{A_{j:i}(0)} \le O(L^3)$ for all $1\le i \le j \le L$, and (ii) $\norm{A_{j:i}(0)} \le e^{-c_1 L^\gamma}$ if $(i, j)$ further satisfies $j-i\ge \frac{L}{10}$. Here $c_1>0$ is a universal constant. From now on we are conditioned on these properties being satisfied. We suppose that the learning rate $\eta$ is at most $e^{0.2c_1L^\gamma}$. We say that a set of weight matrices $W_1, \ldots, W_L$ are in the ``initial neighborhood'' if $\norm{A_i - A_i(0)} \le e^{-0.6c_1L^\gamma}$ for all $i\in[L]$. From Lemmas~\ref{lem:perturbation-bound} and \ref{lem:loss-gradient-bound} we know that in the ``initial neighborhood'' the objective value is always between $0.4\norm{Y}_F^2$ and $0.6\norm{Y}_F^2$. Therefore we have to escape the ``initial neighborhood'' in order to get the objective value out of this interval. Now we calculate how many iterations are necessary to escape the ``initial neighborhood.'' According to \Lemref{lem:loss-gradient-bound}, inside the ``initial neighborhood'' each $W_i$ can move at most $\eta (\sqrt{d_i}\sigma_i )^{-1} e^{-0.9c_1L^\gamma}$ in one iteration by definition of the gradient descent algorithm. In order to leave the ``initial neighborhood,'' some $W_i$ must satisfy $\norm{W_i - W_i(0)} = \sqrt{d_i}\sigma_i \norm{A_i - A_i(0)} > \sqrt{d_i}\sigma_i e^{-0.6c_1L^\gamma}$. In order to move this amount, the number of iterations has to be at least \begin{equation*} \frac{\sqrt{d_i}\sigma_i e^{-0.6c_1L^\gamma}}{ \eta (\sqrt{d_i}\sigma_i )^{-1} e^{-0.9c_1L^\gamma} } = \frac{d_i\sigma_i^2 e^{0.3c_1L^\gamma}}{ \eta } \ge \frac{1}{L^{O(1)}} \cdot \frac{e^{0.3c_1L^\gamma}}{e^{0.2c_1L^\gamma}} \ge e^{\Omega(L^\gamma)}. \end{equation*} This finishes the proof. \end{proof} \section{Introduction} \label{sec:intro} Through their myriad successful applications across a wide range of disciplines, it is now well established that deep neural networks possess an unprecedented ability to model complex real-world datasets, and in many cases they can do so with minimal overfitting. Indeed, the list of practical achievements of deep learning has grown at an astonishing rate, and includes models capable of human-level performance in tasks such as image recognition~\citep{krizhevsky2012}, speech recognition~\citep{hinton2012}, and machine translation~\citep{wu2016}. Yet to each of these deep learning triumphs corresponds a large engineering effort to produce such a high-performing model. Part of the practical difficulty in designing good models stems from a proliferation of hyperparameters and a poor understanding of the general guidelines for their selection. Given a candidate network architecture, some of the most impactful hyperparameters are those governing the choice of the model's initial weights. Although considerable study has been devoted to the selection of initial weights, relatively little has been proved about how these choices affect important quantities such as rate of convergence of gradient descent. In this work, we examine the effect of initialization on the rate of convergence of gradient descent in deep linear networks. We provide for the first time a rigorous proof that drawing the initial weights from the orthogonal group speeds up convergence relative to the standard Gaussian initialization with iid weights. In particular, we show that for deep networks, the width needed for efficient convergence for orthogonal initializations is independent of the depth, whereas the width needed for efficient convergence of Gaussian networks scales linearly in the depth. Orthogonal weight initializations have been the subject of a significant amount of prior theoretical and empirical investigation. For example, in a line of work focusing on \emph{dynamical isometry}, it was found that orthogonal weights can speed up convergence for deep linear networks~\citep{saxe2014exact,advani2017high} and for deep non-linear networks~\citep{pennington2018emergence,xiao2018dynamical,gilboa2019dynamical,chen2018dynamical,pennington2017resurrecting,tarnowski2019dynamical,ling2019spectrum} when they operate in the linear regime. In the context of recurrent neural networks, orthogonality can help improve the system's stability. A main limitation of prior work is that it has focused almost exclusively on model's properties at initialization. In contrast, our analysis focuses on the benefit of orthogonal initialization on the entire training process, thereby establishing a provable benefit for optimization. The paper is organized as follows. After reviewing related work in Section~\ref{sec:related} and establishing some preliminaries in Section~\ref{sec:prelim}, we present our main positive result on efficient convergence from orthogonal initialization in Section~\ref{sec:ortho}. In Section~\ref{sec:gaussian}, we show that Gaussian initialization leads to exponentially long convergence time if the width is too small compared with the depth. In Section~\ref{sec:experiment}, we perform experiments to support our theoretical results. \section{Efficient Convergence using Orthogonal Initialization} \label{sec:ortho} In this section we present our main positive result for orthogonal initialization. We show that orthogonal initialization enables efficient convergence of gradient descent to a global minimum provided that the hidden width is not too small. In order to properly define orthogonal weights, we let the widths of all hidden layers be equal: $d_1=d_2=\cdots=d_{L-1}=m$, and let $m \ge \max\{d_x, d_y\}$. Note that all intermediate matrices $W_2, \ldots, W_{L-1}$ are $m\times m$ square matrices, and $W_1 \in \mathbb{R}^{m\times d_x}, W_L\in\mathbb{R}^{d_y\times m}$. We sample each initial weight matrix $W_i(0)$ independently from a uniform distribution over scaled orthogonal matrices satisfying \begin{equation} \label{eqn:ortho-init} \begin{aligned} &W_1^\top(0) W_1(0) = mI_{d_x},\\ &W_i^\top(0) W_i(0) = W_i(0)W_i^\top(0) = mI_m, \qquad 2\le i \le L-1, \\ &W_L(0)W_L^\top(0)=mI_{d_y}. \end{aligned} \end{equation} In accordance with such initialization, the scaling factor $\alpha$ in~\eqref{eqn:linear-net} is set as $\alpha = \frac{1}{\sqrt{m^{L-1}d_y}}$, which ensures $\expect{\norm{f(x; W_L(0), \ldots, W_1(0))}^2} = \norm{x}^2$ for any $x\in \mathbb{R}^{d_x}$.\footnote{We have $\expect{\norm{f(x; W_L(0), \ldots, W_1(0))}^2} = \alpha^2 \expect{x^\top W_1^\top(0) \cdots W_L^\top(0) W_L(0) \cdots W_1(0) x }$. Note that by our choice \eqref{eqn:ortho-init} we have $\expect{W_L^\top(0) W_L(0)}=d_yI_m$ and $W_i^\top(0) W_i(0)=mI \,(1\le i \le L-1)$, so we have $\expect{\norm{f(x; W_L(0), \ldots, W_1(0))}^2} = \alpha^2 m^{L-1} d_y \norm{x}^2 = \norm{x}^2 $.} The same scaling factor was adopted in \cite{du2019width}, which preserves the expectation of the squared $\ell_2$ norm of any input. Let $W^* \in \argmin_{W \in \mathbb{R}^{d_y \times d_x}} \norm{WX-Y}_F$ and $\ell^* = \frac12 \norm{W^* X - Y}_F^2$. Then $\ell^*$ is the minimum value for the objective~\eqref{eqn:loss-func}. Denote $r = \mathrm{rank}(X)$, $\kappa = \frac{\lambda_{\max}(X^\top X)}{\lambda_r(X^\top X)}$, and $\tilde{r} = \frac{\norm{X}_F^2}{\norm{X}^2}$.\footnote{$\tilde{r}$ is known as the \emph{stable rank} of $X$, which is always no more than the rank.} Our main theorem in this section is the following: \begin{thm}\label{thm:ortho} Suppose \begin{equation} \label{eqn:m-bound-for-ortho} m\ge C \cdot \tilde{r} \kappa^2 \left( d_y(1+\norm{W^*}^2) + \log(r/\delta) \right) \text{ and } m \ge d_x, \end{equation} for some $\delta\in(0, 1)$ and a sufficiently large universal constant $C>0$. Set the learning rate $\eta \le \frac{d_y}{2L \norm{X}^2}$. Then with probability at least $1-\delta$ over the random initialization, we have \begin{align*} &\ell(0) - \ell^* \le O\left( 1 + \frac{\log(r/\delta)}{d_y} + \norm{W^*}^2 \right) \norm{X}_F^2, \\ &\ell(t) - \ell^* \le \left( 1 - \frac{1}{2} \eta L \lambda_r(X^\top X) / d_y \right)^t (\ell(0)-\ell^*), \quad t = 0, 1, 2, \ldots, \end{align*} where $\ell(t)$ is the objective value at iteration $t$. \end{thm} Notably, in \Theoremref{thm:ortho}, the width $m$ need not depend on the depth $L$. This is in sharp contrast with the result of \cite{du2019width} for Gaussian initialization, which requires $m\ge \tilde{\Omega}(L r \kappa^3 d_y)$. It turns out that a near-linear dependence between $m$ and $L$ is necessary for Gaussian initialization to have efficient convergence, as we will show in \Secref{sec:gaussian}. Therefore the requirement in \cite{du2019width} is nearly tight in terms of the dependence on $L$. These results together rigorously establish the benefit of orthogonal initialization in optimizing very deep linear networks. If we set the learning rate optimally according to Theorem~\ref{thm:ortho} to $\eta = \Theta(\frac{d_y}{L\norm{X}^2})$, we obtain that $\ell(t) - \ell^*$ decreases by a ratio of $1 - \Theta(\kappa^{-1})$ after every iteration. This matches the convergence rate of gradient descent on the ($1$-layer) linear regression problem $\min\limits_{W\in\mathbb{R}^{d_y\times d_x}} \frac12 \norm{WX-Y}_F^2$. \subsection{Proof of \Theoremref{thm:ortho}} The proof uses the high-level framework from \cite{du2019width}, which tracks the evolution of the network's output during optimization. This evolution is closely related to a time-varying positive semidefinite (PSD) matrix (defined in \eqref{eqn:P-def}), and the proof relies on carefully upper and lower bounding the eigenvalues of this matrix throughout training, which in turn implies the desired convergence result. First, we can make the following simplifying assumption \emph{without loss of generality}. See Appendix B in \cite{du2019width} for justification. \begin{asmp} \label{asmp:full-rank-data} (Without loss of generality) $X \in \mathbb{R}^{d_x \times r}$, $\mathrm{rank}(X) = r$, $Y = W^* X$, and $\ell^* =0$. \end{asmp} Now we briefly review \cite{du2019width}'s framework. The key idea is to look at the network's output, defined as \begin{align*} U = \alpha W_{L:1} X \in \mathbb{R}^{d_y \times n}. \end{align*} We also write $U(t) = \alpha W_{L:1}(t) X$ as the output at time $t$. Note that $\ell(t) = \frac{1}{2}\norm{U(t)-Y}_F^2$. According to the gradient descent update rule, we write \begin{align*} &W_{L:1}(t+1) = \prod_i \left( W_i(t) - \eta \frac{\partial \ell}{\partial W_i}(t) \right) = W_{L:1}(t) - \sum_{i=1}^L \eta W_{L:i+1}(t) \frac{\partial \ell}{\partial W_i}(t) W_{i-1:1}(t) + E(t), \end{align*} where $E(t)$ contains all the high-order terms (i.e., those with $\eta^2$ or higher). With this definition, the evolution of $U(t)$ can be written as the following equation: \begin{equation} \label{eqn:U-dynamics} \begin{aligned} &\vectorize{U(t+1) - U(t)} = -\eta P(t) \cdot \vectorize{U(t)-Y} + \alpha \cdot \vectorize{E(t)X}, \end{aligned} \end{equation} where \begin{equation} \label{eqn:P-def} \begin{aligned} P(t) = \alpha^2 \sum_{i=1}^L \Big[ \left( \left( W_{i-1:1}(t) X \right)^\top \left( W_{i-1:1}(t)X \right) \right) \otimes \left( W_{L:i+1}(t) W_{L:i+1}^\top(t) \right) \Big] . \end{aligned} \end{equation} Notice that $P(t)$ is always PSD since it is the sum of $L$ PSD matrices. Therefore, in order to establish convergence, we only need to (i) show that the higher-order term $E(t)$ is small and (ii) prove upper and lower bounds on $P(t)$'s eigenvalues. For the second task, it suffices to control the singular values of $W_{i-1:1}(t)$ and $W_{L:i+1}(t)$ ($i\in[L]$).\footnote{Note that for symmetric matrices $A$ and $B$, the set of eigenvalues of $A \otimes B$ is the set of products of an eigenvalue of $A$ and an eigenvalue of $B$.} Under orthogonal initialization, these matrices are perfectly isometric at initialization, and we will show that they stay close to isometry during training, thus enabling efficient convergence. The following lemma summarizes some properties at initialization. \begin{lem} \label{lem:ortho-init} At initialization, we have \begin{equation} \label{eqn:init-spec} \begin{aligned} &\sigma_{\max}(W_{j:i}(0)) = \sigma_{\min}(W_{j:i}(0)) = m^{\frac{j-i+1}{2}} , & \forall 1\le i\le j \le L, (i, j)\not=(1,L). \end{aligned} \end{equation} Furthermore, with probability at least $1-\delta$, the loss at initialization satisfies \begin{equation} \label{eqn:init-loss} \ell(0) \le O\left( 1 + \frac{\log(r/\delta)}{d_y} + \norm{W^*}^2 \right) \norm{X}_F^2. \end{equation} \end{lem} \begin{proof}[Proof sketch] The spectral property~\eqref{eqn:init-spec} follows directly from~\eqref{eqn:ortho-init}. To prove~\eqref{eqn:init-loss}, we essentially need to upper bound the magnitude of the network's initial output. This turns out to be equivalent to studying the magnitude of the projection of a vector onto a random low-dimensional subspace, which we can bound using standard concentration inequalities. The details are given in Appendix~\ref{app:proof-ortho-init}. \end{proof} Now we proceed to prove \Theoremref{thm:ortho}. We define $B = O \left( 1 + \frac{\log(r/\delta)}{d_y} + \norm{W^*}^2 \right) \norm{X}_F^2$ which is the upper bound on $\ell(0)$ from~\eqref{eqn:init-loss}. Conditioned on \eqref{eqn:init-loss} being satisfied, we will use induction on $t$ to prove the following three properties ${\mathcal{A}}(t)$, ${\mathcal{B}}(t)$ and ${\mathcal{C}}(t)$ for all $t=0, 1, \ldots$: \begin{itemize} \item ${\mathcal{A}}(t)$: $\ell(t) \le \left( 1 - \frac12 \eta L \sigma_{\min}^2(X) / d_y \right)^t \ell(0) \le \left( 1 - \frac12 \eta L \sigma_{\min}^2(X) /d_y \right)^t B$. \item ${\mathcal{B}}(t)$: $ \sigma_{\max}(W_{j:i}(t)) \le 1.1 m^{\frac{j-i+1}{2}} , \sigma_{\min}(W_{j:i}(t)) \ge 0.9 m^{\frac{j-i+1}{2}} , \quad \forall 1\le i\le j \le L, (i, j)\not=(1, L). $ \begin{comment} \begin{align*} \begin{cases} &\sigma_{\max}(W_{j:i}(t)) \le \frac54 m^{\frac{L-i+1}{2}}, \, \forall 1<i\le L, \\ &\sigma_{\min}\left( W_{L:i}(t) \right) \ge \frac34 m^{\frac{L-i+1}{2}} , \, \forall 1<i\le L, \\ &\sigma_{\max}(W_{i:1}(t) \cdot X) \le \frac54 m^{\frac{i}{2}} \sigma_{\max}(X), \, \forall 1\le i < L, \\ &\sigma_{\min}\left( W_{i:1}(t) \cdot X \right) \ge \frac34 m^{\frac{i}{2}} \sigma_{\min}(X), \, \forall 1\le i < L, \\ & \norm{W_{j:i}(t)} \le O\left( \sqrt L m^{\frac{j-i+1}{2}}\right), \, \forall 1<i\le j <L. \end{cases} \end{align*} \end{comment} \item ${\mathcal{C}}(t)$: $\norm{W_i(t) - W_i(0)}_F \le \frac{8\sqrt{Bd_y}\norm{X}}{ L \sigma_{\min}^2(X)}, \quad \forall 1\le i\le L$. \end{itemize} ${\mathcal{A}}(0)$ and ${\mathcal{B}}(0)$ are true according to \Lemref{lem:ortho-init}, and ${\mathcal{C}}(0)$ is trivially true. In order to prove ${\mathcal{A}}(t)$, ${\mathcal{B}}(t)$ and ${\mathcal{C}}(t)$ for all $t$, we will prove the following claims for all $t\ge0$: \begin{claim} \label{claim:induction-1} ${\mathcal{A}}(0), \ldots, {\mathcal{A}}(t), {\mathcal{B}}(0), \ldots, {\mathcal{B}}(t) \Longrightarrow {\mathcal{C}}(t+1)$. \end{claim} \begin{claim}\label{claim:induction-2} ${\mathcal{C}}(t) \Longrightarrow {\mathcal{B}}(t)$. \end{claim} \begin{claim}\label{claim:induction-3} ${\mathcal{A}}(t), {\mathcal{B}}(t) \Longrightarrow {\mathcal{A}}(t+1)$. \end{claim} The proofs of these claims are given in Appendix~\ref{app:proof-ortho}. Notice that we finish the proof of \Theoremref{thm:ortho} once we prove ${\mathcal{A}}(t)$ for all $t\ge 0$. \section{Preliminaries} \label{sec:prelim} \subsection{Notation} Let $[n]=\{1, 2, \ldots, n\}$. Denote by $I_d$ the $d\times d$ identity matrix, and by $I$ an identity matrix when its dimension is clear from context. Denote by ${\mathcal{N}}(\mu, \sigma^2)$ the Gaussian distribution with mean $\mu$ and variance $\sigma^2$, and by $\chi^2_k$ the chi-squared distribution with $k$ degrees of freedom. Denote by $\norm{\cdot}$ the $\ell_2$ norm of a vector or the spectral norm of a matrix. Denote by $\norm{\cdot}_F$ the Frobenius norm of a matrix. For a symmetric matrix $A$, let $\lambda_{\max}(A)$ and $\lambda_{\min}(A)$ be its maximum and minimum eigenvalues, and let $\lambda_i(A)$ be its $i$-th largest eigenvalue. For a matrix $B \in \mathbb{R}^{m\times n}$, let $\sigma_i(B)$ be its $i$-th largest singular value ($i=1, 2, \ldots, \min\{m, n\}$), and let $\sigma_{\max}(B) = \sigma_1(B)$, $\sigma_{\min}(B) = \sigma_{\min\{m,n\}}(B)$. Denote by $\vectorize{A}$ be the vectorization of a matrix $A$ in column-first order. The Kronecker product between two matrices $A \in \mathbb{R}^{m_1 \times n_1}$ and $B \in \mathbb{R}^{m_2 \times n_2}$ is defined as $$ A \otimes B = \left ( \begin{matrix} a_{1,1} B& \cdots &a_{1,n_1} B \\ \vdots & \ddots & \vdots \\ a_{m_1,1} B & \cdots & a_{m_1,n_1} B \end{matrix} \right) \in \mathbb{R}^{m_1 m_2 \times n_1 n_2}, $$ where $a_{i,j}$ is the element in the $(i, j)$-th entry of $A$. We use the standard $O(\cdot)$, $\Omega(\cdot)$ and $\Theta(\cdot)$ notation to hide universal constant factors. We also use $C$ to represent a sufficiently large universal constant whose specific value can differ from line to line. \subsection{Problem Setup} Suppose that there are $n$ training examples $\{(x_k,y_k)\}_{k=1}^{n}\subset\mathbb{R}^{d_x}\times\mathbb{R}^{d_y}$. Denote by $X = \left( x_1, \ldots, x_n \right) \in \mathbb{R}^{d_x \times n}$ the input data matrix and by $Y = \left( y_1, \ldots, y_n \right) \in \mathbb{R}^{d_y \times n}$ the target matrix. Consider an $L$-layer linear neural network with weight matrices $W_1, \ldots, W_L$, which given an input $x\in \mathbb{R}^{d_x}$ computes \begin{equation} \label{eqn:linear-net} f(x; W_1, \ldots, W_L) = \alpha W_L W_{L-1} \cdots W_1 x , \end{equation} where $W_i \in \mathbb{R}^{d_i\times d_{i-1}} (i=1, \ldots, L)$, $d_0=d_x$, $d_L=d_y$, and $\alpha$ is a normalization constant which will be specified later according to the initialization scheme. We study the problem of training the deep linear network by minimizing the $\ell_2$ loss over training data: \begin{equation} \label{eqn:loss-func} \begin{aligned} \ell(W_1, \ldots, W_L) = \frac{1}{2} \sum_{k=1}^n \norm{f(x_k; W_1, \ldots, W_L) - y_k}^2 = \frac{1}{2} \norm{\alpha W_L \cdots W_1 X - Y}_F^2. \end{aligned} \end{equation} The algorithm we consider to minimize the objective~\eqref{eqn:loss-func} is gradient descent with random initialization, which first randomly samples the initial weight matrices $\{W_i(0)\}_{i=1}^L$ from a certain distribution, and then updates the weights using gradient descent: for time $t=0, 1, 2, \ldots$, \begin{equation} \label{eqn:gd-update} W_i(t+1) = W_i(t) - \eta \frac{\partial \ell}{\partial W_i}(W_1(t), \ldots, W_L(t)) , \qquad i\in[L], \end{equation} where $\eta>0$ is the learning rate. For convenience, we denote $W_{j:i} = W_jW_{j-1}\cdots W_i\, (1\le i \le j \le L)$ and $W_{i-1:i} = I \, (i\in[L])$. The time index $t$ is used on any variable that depends on $W_1, \ldots, W_L$ to represent its value at time $t$, e.g., $W_{j:i}(t) = W_j(t) \cdots W_i(t)$, $\ell(t) = \ell(W_1(t), \ldots, W_L(t))$, etc. \section{Proofs for \Secref{sec:gaussian}} \label{app:proof-gaussian} \subsection{Proof of Lemma~\ref{lem:init-product-bound}} \label{app:proof-gaussian-init-product-bound} \begin{proof}[Proof of Lemma~\ref{lem:init-product-bound}] Notice that for any $1\le i \le j \le L$ we have $\expect{\norm{A_{j:i}(0)}_F^2} = d_{i-1}$. Then by Markov inequality we have $\Pr\left[ \norm{A_{j:i}(0)}_F^2 \ge \frac{d_{i-1}}{\delta/L^2} \right] \le \delta/L^2$. Taking a union bound, we know that with probability at least $1-\delta$, for all $1\le i \le j \le L$ simultaneously we have $\norm{A_{j:i}(0)} \le \norm{A_{j:i}(0)}_F \le \frac{d_{i-1}}{\delta/L^2} \le O(L^3/\delta)$ (note that $d_{i-1} \le O(L^{1-\gamma}) =O(L)$). \end{proof} \subsection{Proof of Lemma~\ref{lem:perturbation-bound}} \label{app:proof-gaussian-perturbation-bound} \begin{proof}[Proof of Lemma~\ref{lem:perturbation-bound} (continued)] For the second part of the lemma ($j-i\ge \frac L4$), we need to bound the terms of the form $A_{j:k+1}(0) \cdot \Delta_k \cdot A_{k-1:i}(0)$ more carefully. In fact, if $j-i\ge \frac L4$, then $\max\{ j-k-1, k-1-i \} \ge \frac{L}{10}$, which by assumption means either $A_{j:k+1}(0)$ or $A_{k-1:i}(0)$ has spectral norm bounded by $e^{-c_1 L^\gamma}$. This implies $\norm{A_{j:k+1}(0) \cdot \Delta_k \cdot A_{k-1:i}(0)} \le e^{-c_1 L^\gamma} e^{-0.6c_1 L^\gamma} \cdot O(L^3) = e^{-1.6c_1 L^\gamma} \cdot O(L^3)$. Therefore we have \begin{align*} &\norm{A_{j:i} - A_{j:i}(0)} \le (j-i+1) e^{-1.6c_1L^\gamma} \cdot O(L^3) + \sum_{s=2}^{j-i+1} \binom{j-i+1}{s} \left(e^{-0.6c_1 L^\gamma}\right)^s \left( O(L^3) \right)^{s+1} \\ \le\,& e^{- c_1 L^\gamma} + \sum_{s=2}^{\infty} L^s \left(e^{-0.6c_1 L^\gamma}\right)^s \left( O(L^3) \right)^{s+1} \le e^{- c_1 L^\gamma} + \sum_{s=2}^\infty \left( e^{-0.5c_1L^\gamma} \right)^s = O\left( e^{- c_1 L^\gamma} \right). \end{align*} This implies $\norm{A_{j:i}} \le O\left( e^{- c_1 L^\gamma} \right) $. \end{proof} \subsection{Proof of Lemma~\ref{lem:loss-gradient-bound}} \label{app:proof-gaussian-loss-gradient-bound} \begin{proof}[Proof of Lemma~\ref{lem:loss-gradient-bound}] We can bound the network's output as $$\norm{\alpha W_{L:1}(0)X}_F = \norm{\beta A_{L:1}(0)X}_F \le L^{O(1)} \cdot e^{-\Omega(L^\gamma)} \norm{X}_F = e^{-\Omega(L^\gamma)} .$$ Thus the objective value $\ell(W_1, \ldots, W_L) = \frac12 \norm{\alpha W_{L:1}(0)X -Y}_F^2 $ must be extremely close to $\frac{1}{2}\norm{Y}_F^2$ for large $L$, so $0.4 \norm{Y}_F^2 < \ell(W_1, \ldots, W_L) < 0.6 \norm{Y}_F^2$. As for the gradient, for any $i\in[L]$ we have \begin{align*} &\norm{\nabla_{W_i} \ell(W_1, \ldots, W_L)} = \norm{ \alpha W_{L:i+1}^\top \left(\alpha W_{L:1}X -Y\right) X^\top W_{i-1:1}^\top}\\ =\,& \norm{ \beta / (\sqrt{d_i}\sigma_i) \cdot A_{L:i+1}^\top \left(\alpha W_{L:1}X -Y\right) X^\top A_{i-1:1}^\top } \le \frac{L^{O(1)}}{\sqrt{d_i}\sigma_i} \norm{A_{L:i+1}} \cdot O(1) \cdot \norm{A_{i-1:1}} . \end{align*} Using~\eqref{eqn:product-bound}, and noting that either $L-i-1$ or $i-1$ is greater than $\frac L4$, we have \begin{equation*} \norm{\nabla_{W_i} \ell(W_1, \ldots, W_L)} \le \sigma_i^{-1} L^{O(1)} \cdot O\left( e^{-c_1L^\gamma} \right) \cdot O(L^3) \le (\sqrt{d_i}\sigma_i )^{-1} e^{-0.9c_1L^\gamma} . \qedhere \end{equation*} \end{proof} \section{Proofs for \Secref{sec:ortho}} \label{app:proof-ortho} \subsection{Proof of Lemma~\ref{lem:ortho-init}} \label{app:proof-ortho-init} \begin{proof}[Proof of Lemma~\ref{lem:ortho-init}] We only need to prove~\eqref{eqn:init-loss}. We first upper bound the magnitude of the network's initial output on any given input $x\in\mathbb{R}^{d_x}$. Let $z = \frac{1}{\sqrt{m^{L-1}}} W_{L-1:1}(0) \cdot x \in \mathbb{R}^m$. Then we have $\norm{z}=\norm{x}$, and $f(x; W_1(0), \ldots, W_L(0)) = \frac{1}{\sqrt{d_y}} W_L(0) \cdot z = \sqrt{\frac{m}{d_y}} \cdot \frac{1}{\sqrt{m}} W_L(0) \cdot z $. Note that $\frac{1}{\sqrt{m}} W_L(0) \cdot z $ is the (signed) projection of $z$ onto a random subspace in $\mathbb{R}^m$ of dimension $d_y$. Therefore $\norm{\frac{1}{\sqrt{m}} W_L(0) \cdot z}^2 / \norm{z}^2$ has the same distribution as $\frac{g_1^2 + \cdots + g_{d_y}^2}{g_1^2 + \cdots + g_m^2}$, where $g_1, \ldots, g_m $ are i.i.d. samples from ${\mathcal{N}}(0, 1)$. By the standard tail bounds for chi-squared distributions we have \begin{align*} &\Pr\left[ g_1^2 + \cdots + g_{d_y}^2 \le d_y + 2\sqrt{d_y \log(1/\delta')} + 2\log(1/\delta') \right] \ge 1-\delta', \\ &\Pr\left[ g_1^2 + \cdots + g_{m}^2 \ge m - 2\sqrt{m \log(1/\delta') } \right] \ge 1-\delta'. \end{align*} Let $\delta' = \frac{\delta}{2r}$. Note that $m > C\cdot \log(r/\delta)$. We know that with probability at least $1-\frac{\delta}{r}$ we have \begin{align*} \norm{\frac{1}{\sqrt{m}} W_L(0) \cdot z}^2 / \norm{z}^2 \le \frac{d_y + 2\sqrt{d_y \log(2r/\delta)} + 2\log(2r/\delta)}{m - 2\sqrt{m \log(2r/\delta) }} = \frac{O(d_y + \log(r/\delta))}{\Omega(m)}, \end{align*} which implies \begin{equation} \label{eqn:ortho-inproof-1} \begin{aligned} \norm{ f(x; W_1(0), \ldots, W_L(0)) }^2 &= \frac{m}{d_y} \norm{\frac{1}{\sqrt{m}} W_L(0) \cdot z}^2 = \frac{m}{d_y} \cdot O\left( \frac{d_y + \log(r/\delta)}{m} \right) \norm{z}^2 \\ &= O\left( 1 + \frac{\log(r/\delta)}{d_y} \right) \norm{x}^2. \end{aligned} \end{equation} Finally, taking a union bound, we know that with probability at least $1-\delta$, the inequality \eqref{eqn:ortho-inproof-1} holds for every $x\in\{x_1, \ldots, x_r\}$, which implies \begin{align*} \ell(0) &= \frac12 \sum_{k=1}^r \norm{ f(x_k; W_1(0), \ldots, W_L(0)) - y_k }^2 \le \sum_{k=1}^r \left( \norm{ f(x_k; W_1(0), \ldots, W_L(0))}^2 + \norm{ y_k }^2 \right) \\ &\le O\left( 1 + \frac{\log(r/\delta)}{d_y} \right) \sum_{k=1}^r\norm{x_k}^2 + \sum_{k=1}^r \norm{ y_k }^2 = O\left( 1 + \frac{\log(r/\delta)}{d_y} \right) \norm{X}_F^2 + \norm{ Y }_F^2 \\ & \le O\left( 1 + \frac{\log(r/\delta)}{d_y} + \norm{W^*}^2 \right) \norm{X}_F^2. \qedhere \end{align*} \end{proof} \subsection{Proof of Claim~\ref{claim:induction-1}} \label{app:proof-induction-1} \begin{proof}[Proof of Claim~\ref{claim:induction-1}] Let $\gamma = \frac12 L \sigma_{\min}^2(X)/d_y$. From ${\mathcal{A}}(0), \ldots, {\mathcal{A}}(t)$ we have $\ell(s) \le (1-\eta\gamma)^sB$ for all $0\le s \le t$. The gradient of the objective function~\eqref{eqn:loss-func} is $\frac{\partial\ell}{\partial W_i} = \alpha W_{L:i+1}^\top (U - Y) \left( W_{i-1:1} X \right)^\top$. Thus we can bound the gradient norm as follows for all $0\le s\le t$ and all $i\in[L]$: \begin{equation} \label{eqn:gradient-norm-bound} \begin{aligned} &\norm{\frac{\partial\ell}{\partial W_i}(s)}_F \le \alpha \norm{W_{L:i+1}(s)} \norm{U(s) - Y}_F \norm{W_{i-1:1}(s)} \norm{X}\\ \le\,& \frac{1}{\sqrt{m^{L-1}d_y}} \cdot 1.1 m^{\frac{L-i}{2}} \cdot \sqrt{2\ell(s)} \cdot 1.1 m^{\frac{i-1}{2}} \norm{X} \le \frac{2\sqrt{(1-\eta\gamma)^sB}}{\sqrt{d_y}} \norm{X} , \end{aligned} \end{equation} where we have used ${\mathcal{B}}(s)$. Then for all $i\in[L]$ we have: \begin{align*} &\norm{W_i(t+1) - W_i(0)}_F \le \sum_{s=0}^{t} \norm{W_i(s+1) - W_i(s)}_F =\, \sum_{s=0}^{t} \norm{\eta \frac{\partial\ell}{\partial W_i}(s)}_F\\ \le\,& \eta \sum_{s=0}^{t} \frac{2\sqrt{(1-\eta\gamma)^sB}}{\sqrt{d_y}} \norm{X} \le \frac{2\eta \sqrt{B}}{\sqrt{d_y}} \norm{X} \sum_{s=0}^{t-1}(1-\eta\gamma/2)^s \le \frac{2 \eta \sqrt{B}}{ \sqrt{d_y}} \norm{X} \cdot \frac{2}{\eta\gamma} \\ =\,& \frac{8 \sqrt{B d_y} \norm{X}}{ L \sigma_{\min}^2(X)} . \end{align*} This proves ${\mathcal{C}}(t+1)$. \end{proof} \subsection{Proof of Claim~\ref{claim:induction-2}} \label{app:proof-induction-2} \begin{proof}[Proof of Claim~\ref{claim:induction-2}] Let $R = \frac{8\sqrt{Bd_y}\norm{X}}{ L \sigma_{\min}^2(X)}$ and $\Delta_i = W_i(t) - W_i(0)$ $(i\in[L])$. Then ${\mathcal{C}}(t)$ means $\norm{\Delta_i}_F \le R\, (\forall i\in[L])$. For $1\le i \le j \le L$, we have \begin{align*} W_{j:i}(t) = \left( W_j(0) + \Delta_j \right) \cdots \left( W_i(0) + \Delta_i \right) . \end{align*} Expanding this product, each term except $W_{j:i}(0)$ has the form: \begin{equation} \label{eqn:perturbation-terms} \begin{aligned} W_{j:(k_s+1)}(0) \cdot \Delta_{k_s} \cdot W_{(k_s-1):(k_{s-1}+1)}(0) \cdot \Delta_{k_{s-1}} \cdots \Delta_{k_1} \cdot W_{(k_1-1):i}(0) , \end{aligned} \end{equation} where $i \le k_1 < \cdots < k_s \le j$ are locations where terms like $\Delta_{k_l}$ are taken out. Note that every factor in~\eqref{eqn:perturbation-terms} of the form $W_{j':i'}(0)$ satisfies $\norm{W_{j':i'}(0)} = m^{\frac{j'-i'+1}{2}}$ according to~\eqref{eqn:init-spec}. Thus, we can bound the sum of all terms of the form~\eqref{eqn:perturbation-terms} as \begin{align*} &\norm{W_{j:i}(t) - W_{j:i}(0)} \le \sum_{s=1}^{j-i+1} \binom{j-i+1}{s} R^s m^{\frac{j-i+1-s}{2}} = (\sqrt m + R)^{j-i+1} - (\sqrt m)^{j-i+1} \\ =\,& (\sqrt m)^{j-i+1} \left( \left( 1 + {R}/{\sqrt m} \right)^{j-i+1} - 1 \right) \le (\sqrt m)^{j-i+1} \left( \left( 1 + {R}/{\sqrt m} \right)^{L} - 1 \right) \le 0.1 (\sqrt m)^{j-i+1}. \end{align*} Here the last step uses $m > C (LR)^2$ which is implied by~\eqref{eqn:m-bound-for-ortho}. Combined with~\eqref{eqn:init-spec}, this proves ${\mathcal{B}}(t)$. \end{proof} \begin{comment} The proof of~\eqref{eqn:perturbation-left} is very similar. We still look at products of the form~\eqref{eqn:perturbation-terms} with $j=L$. Still, there are at most $s+1$ factors of the form $W_{b:a}(0)$ each satisfying $\norm{W_{b:a}(0)} \le O\left(\sqrt Lm^{\frac{b-a+1}{2}}\right)$. However, there are at most $s$ (instead of $s+1$) such factors such that $1<a\le b<L$. Therefore, the product of spectral norms of all such factors is at most $\left(O(\sqrt{L})\right)^{s} m^{\frac{L-i+1-s}{2}} $. In other words, we can save a factor of $O(\sqrt L)$ compared with~\eqref{eqn:perturbation-bound-middle}. This gives us \begin{align*} \norm{W_{L:i}(t) - W_{L:i}(0)} \le 0.05 m^{\frac{L-i+1}{2}}, \end{align*} proving~\eqref{eqn:perturbation-left}. Next we prove~\eqref{eqn:perturbation-right}. For $1\le i<L$, we need to bound the sum of terms of the following form: \begin{equation*} \begin{aligned} W_{i:(k_s+1)}(0) \cdot \Delta_{k_s} \cdot W_{(k_s-1):(k_{s-1}+1)}(0) \cdot \Delta_{k_{s-1}} \cdots \\ \cdot \Delta_{k_1} \cdot W_{(k_1-1):1}(0) X . \end{aligned} \end{equation*} Again, similar as before and noting $\norm{W_{(k_1-1):1}(0) X} \le \frac54 m^{\frac{k_1-1}{2}}\norm{X}$, we have \small \begin{equation*} \begin{aligned} &\norm{W_{i:1}(t) - W_{i:1}(0)} \\ \le\,& \sum_{s=1}^{i} \binom{i}{s} R^s \left( O\left(\sqrt L\right) \right)^{s} \cdot \frac54 m^{\frac{i-s}{2}} \norm{X} \\ \le\,& \frac54 \sum_{s=1}^{i} L^s R^s \left( O\left(\sqrt L\right) \right)^{s} m^{\frac{i-s}{2}} \norm{X} \\ =\,& \frac54 m^{\frac i2} \sum_{s=1}^{i} \left( \frac{ O\left(L^{3/2}R\right)}{\sqrt m} \right)^s \norm{X} \\ \le\,& 0.05 m^{\frac{i}{2}} \sigma_{\min}(X) , \end{aligned} \end{equation*} \normalsize as long as $ m > C L^{3} R^2 \cdot \frac{\norm{X}^2}{\sigma_{\min}^2(X)} = CL^3R^2 \kappa$ which is implied by~\eqref{eqn:m-bound-implication}. This finishes the proof of~\eqref{eqn:perturbation-right}. \end{comment} \subsection{Proof of Claim~\ref{claim:induction-3}} \label{app:proof-induction-3} \begin{proof}[Proof of Claim~\ref{claim:induction-3}] Recall that we have the dynamics~\eqref{eqn:U-dynamics} for $U(t)$. In order to establish convergence from~\eqref{eqn:U-dynamics} we need to prove upper and lower bounds on the eigenvalues of $P(t)$, as well as show that the high-order term $E(t)$ is small. We will prove these using ${\mathcal{B}}(t)$. Using the definition~\eqref{eqn:P-def} and property ${\mathcal{B}}(t)$, we have \begin{align*} \lambda_{\max}(P(t)) &\le \alpha^2 \sum_{i=1}^L \lambda_{\max}\left( \left( W_{i-1:1}(t) X \right)^\top \left( W_{i-1:1}(t)X \right) \right) \cdot \lambda_{\max}\left( W_{L:i+1}(t) W_{L:i+1}^\top(t) \right) \\ &\le \frac{1}{m^{L-1}d_y} \sum_{i=1}^L \left( 1.1 m^{\frac{i-1}{2}} \sigma_{\max}(X) \right)^2 \left( 1.1 m^{\frac{L-i}{2}} \right)^2 \le 2 L \sigma_{\max}^2(X) / d_y, \\ \lambda_{\min}(P(t)) &\ge \alpha^2 \sum_{i=1}^L \lambda_{\min}\left( \left( W_{i-1:1}(t) X \right)^\top \left( W_{i-1:1}(t)X \right) \right) \cdot \lambda_{\min}\left( W_{L:i+1}(t) W_{L:i+1}^\top(t) \right) \\ &\ge \frac{1}{m^{L-1}d_y} \sum_{i=1}^L \left( 0.9 m^{\frac{i-1}{2}} \sigma_{\min}(X) \right)^2 \left( 0.9 m^{\frac{L-i}{2}} \right)^2 \ge \frac{3}{5} L \sigma_{\min}^2(X) / d_y. \end{align*} In the lower bound above, we make use of the following relation on dimensions: $m\ge d_x \ge r$, which enables the inequality $\lambda_{\min}\left( \left( W_{i-1:1}(t) X \right)^\top \left( W_{i-1:1}(t)X \right) \right) = \sigma_{\min}^2\left( W_{i-1:1}(t) X \right) \ge \sigma_{\min}^2\left( W_{i-1:1}(t) \right) \cdot \sigma_{\min}^2(X)$. Next, we will prove the following bound on the high-order term $E(t)$: \[ \frac{1}{\sqrt{m^{L-1}d_y}} \norm{E(t) X}_F \le \frac16 \eta \lambda_{\min}(P_t) \norm{U(t)-Y}_F . \] Recall that $E(t)$ is the sum of all high-order terms in the product \begin{align*} W_{L:1}(t+1) = \prod_i \left( W_i(t) - \eta \frac{\partial \ell}{\partial W_i}(t) \right). \end{align*} Same as~\eqref{eqn:gradient-norm-bound}, we have $\norm{\frac{\partial \ell}{\partial W_i} (t)}_F \le \frac{2 \sqrt{\ell(t)}\norm{X}}{\sqrt{d_y}}$ ($\forall i\in[L]$). Then we have \begin{align*} & \frac{1}{\sqrt{m^{L-1}d_y}} \norm{E(t) X}_F \\ \le\,& \frac{1}{\sqrt{m^{L-1}d_y}} \sum_{s=2}^L \binom{L}{s} \left(\eta \cdot \frac{2 \sqrt{\ell(t)}\norm{X}}{\sqrt{d_y}} \right)^s m^{\frac{L-s}{2}} \norm{X} \\ \le\,& \sqrt{\frac{m}{d_y}} \norm{X} \sum_{s=2}^L L^s \left(\eta \cdot \frac{2 \sqrt{\ell(t)}\norm{X}}{\sqrt{d_y}} \right)^s m^{-\frac{s}{2}} \\ =\,& \sqrt{\frac{m}{d_y}} \norm{X} \sum_{s=2}^L \left( \frac{2\eta L \sqrt{\ell(t)}\norm{X}}{\sqrt{m d_y}} \right)^s \end{align*} From $\eta \le \frac{d_y}{ 2L \norm{X}^2 }$, we have $\frac{2\eta L \sqrt{\ell(t)}\norm{X}}{\sqrt{m d_y}} \le \frac{ \sqrt{d_y \cdot\ell(t)}}{ \sqrt{m}\norm{X} } $. Note that $m > C \cdot\frac{ d_y B}{\norm{X}^2} \ge C \cdot\frac{ d_y \ell(t)}{\norm{X}^2}$. Thus we have \begin{align*} & \frac{1}{\sqrt{m^{L-1}d_y}} \norm{E(t) X}_F \le \sqrt{\frac{m}{d_y}} \norm{X} \left( \frac{2\eta L \sqrt{\ell(t)}\norm{X}}{\sqrt{m d_y}} \right)^2 \sum_{s=2}^{L-2} 0.5^{s-2} \\ \le\,& 2 \sqrt{\frac{m}{d_y}} \norm{X} \left( \frac{2\eta L \sqrt{\ell(t)}\norm{X}}{\sqrt{m d_y}} \right)^2 \le 2 \sqrt{\frac{m}{d_y}} \norm{X} \cdot \frac{2\eta L \sqrt{\ell(t)}\norm{X}}{\sqrt{m d_y}} \cdot \frac{ \sqrt{d_y \cdot\ell(t)}}{ \sqrt{m}\norm{X} } \\ =\,& \frac{4\eta L \norm{X} \cdot \ell(t)}{\sqrt{md_y}}. \end{align*} It suffices to show that the above bound is at most $\frac16 \eta \lambda_{\min}(P_t) \norm{U(t)-Y}_F = \frac16 \eta \lambda_{\min}(P_t) \sqrt{2\ell(t)} $. Since $\lambda_{\min}(P_t) \ge \frac{3}{5} L \sigma_{\min}^2(X) / d_y$, it suffices to have \begin{align*} \frac{4\eta L \norm{X} \cdot \ell(t)}{\sqrt{md_y}} \le \frac16 \eta \cdot \frac{3 L \sigma_{\min}^2(X) \sqrt{2\ell(t)} }{ 5 d_y}, \end{align*} which is true since $m> C \cdot \frac{d_y B \norm{X}^2}{\sigma_{\min}^4(X)} \ge C \cdot \frac{d_y \ell(t)\norm{X}^2}{\sigma_{\min}^4(X)}$. Finally, from~\eqref{eqn:U-dynamics} and $\eta \le \frac{d_y}{2L \norm{X}^2} \le \frac{1}{\lambda_{\max}(P_t)}$ we have \begin{align*} &\norm{U(t+1)-Y}_F= \norm{\vectorize{U(t+1) - Y} }\\ =\,& \norm{ (I-\eta P(t)) \cdot \vectorize{U(t)-Y} + \frac{1}{\sqrt{m^{L-1}d_y}} \vectorize{E(t)X}} \\ \le\,& (1 - \eta \lambda_{\min}(P(t)) ) \norm{\vectorize{U(t)-Y}} + \frac{1}{\sqrt{m^{L-1}d_y}} \norm{E(t) X}_F \\ \le\,& (1 - \eta \lambda_{\min}(P(t)) ) \norm{U(t)-Y}_F + \frac16 \eta \lambda_{\min}(P_t) \norm{U(t)-Y}_F \\ =\,& \left(1 - \frac56 \eta \lambda_{\min}(P(t)) \right) \norm{U(t)-Y}_F \\ \le\,& \left(1 - \frac12 \eta L \sigma_{\min}^2(X)/d_y \right) \norm{U(t)-Y}_F . \end{align*} Therefore $\ell(t+1) \le \left(1 - \frac12 \eta L \sigma_{\min}^2(X)/d_y \right)^2 \ell(t) \le \left(1 - \frac12 \eta L \sigma_{\min}^2(X)/d_y \right)\ell(t)$. Combined with ${\mathcal{A}}(t)$, this proves ${\mathcal{A}}(t+1)$. \end{proof} \section{Related Work} \label{sec:related} \paragraph{Deep linear networks.} Despite the simplicity of their input-output maps, deep linear networks define high-dimensional non-convex optimization landscapes whose properties closely reflect those of their non-linear counterparts. For this reason, deep linear networks have been the subject of extensive theoretical analysis. A line of work~\citep{kawaguchi2016deep,hardt2016identity,lu2017depth,yun2017global,zhou2018critical,laurent2018deep} studied the landscape properties of deep linear networks. Although it was established that all local minima are global under certain assumptions, these properties alone are still not sufficient to guarantee global convergence or to provide a concrete rate of convergence for gradient-based optimization algorithms. Another line of work directly analyzed the trajectory taken by gradient descent and established conditions that guarantee convergence to global minimum~\citep{bartlett2018gradient, arora2018convergence, du2019width}. Most relevant to our work is the result of \cite{du2019width}, which shows that if the width of hidden layers is larger than the depth, gradient descent with Gaussian initialization can efficiently converge to a global minimum. Our result establishes that for Gaussian initialization, this linear dependence between width and depth is necessary, while for orthogonal initialization, the width can be independent of depth. Our negative result for Gaussian initialization also significantly generalizes the result of \cite{shamir2018exponential}, who proved a similar negative result for $1$-dimensional linear networks. \paragraph{Orthogonal weight initializations.} Orthogonal weight initializations have also found significant success in non-linear networks. In the context of feedforward models, the spectral properties of a network's input-output Jacobian have been empirically linked to convergence speed~\citep{saxe2014exact,pennington2017resurrecting,pennington2018emergence,xiao2018dynamical}. It was found that when this spectrum concentrates around $1$ at initialization, a property dubbed \emph{dynamical isometry}, convergence times improved by orders of magnitude. The conditions for attaining dynamical isometry in the infinite-width limit were established by~\citet{pennington2017resurrecting,pennington2018emergence} and basically require that input-output map to be approximately linear and for the weight matrices to be orthogonal. Therefore the training time benefits of dynamical isometry are likely rooted in the benefits of orthogonality for deep linear networks, which we establish in this work. Orthogonal matrices are also frequently used in the context of recurrent neural networks, for which the stability of the state-to-state transition operator is determined by the spectrum of its Jacobian~\citep{haber2017stable,laurent2016recurrent}. Orthogonal matrices can improve the conditioning, leading to an ability to learn over long time horizons~\citep{le2015simple,henaff2016recurrent,chen2018dynamical,gilboa2019dynamical}. While the benefits of orthogonality can be quite large at initialization, little is known about whether or in what contexts these benefits persist during training, a scenario that has lead to the development of efficient methods of constraining the optimization to the orthogonal group~\citep{wisdom2016full,vorontsov2017orthogonality,mhammedi2017efficient}. Although we do not study the recurrent setting in this work, an extension of our analysis might help determine when orthogonality is beneficial in that setting.

https://arxiv.org/abs/1405.6894

On the number of monotone sequences

One of the most classical results in Ramsey theory is the theorem of Erdős and Szekeres from 1935, which says that every sequence of more than $k^2$ numbers contains a monotone subsequence of length $k+1$. We address the following natural question motivated by this result: Given integers $k$ and $n$ with $n \geq k^2+1$, how many monotone subsequences of length $k+1$ must every sequence of $n$ numbers contain? We answer this question precisely for all sufficiently large $k$ and $n \leq k^2 + c k^{3/2} / \log k$, where $c$ is some absolute positive constant.

\section{Introduction} \label{sec:introduction} A typical problem in extremal combinatorics has the following form: What is the largest size of a structure which does not contain any forbidden configurations? Once this extremal value is known, it is very natural to ask how many forbidden configurations one is guaranteed to find in every structure of a certain size which is larger than the extremal value. There are many results of this kind. Most notably, there is a very large body of work on the problem of determining the smallest number of $k$-vertex cliques in a graph with $n$ vertices and $m$ edges, attributed to Erd{\H{o}}s and Rademacher; see~\cite{Er62, Er69, ErSi83, LoSi83, Ni11, Ra08, Re12}. In extremal set theory there, is an extension of the celebrated Sperner's theorem, where one asks for the minimum number of chains in a family of subsets of $\{1, \ldots, n\}$ with more than $\binom{n}{\lfloor n/2 \rfloor}$ members; see~\cite{DaGaSu13, DoGrKaSe13, ErKl74, Kl68}. Another example is a recent work in \cite{DaGaSu14}, motivated by the classical theorem of Erd\H{o}s, Ko, and Rado. It studies how many disjoint pairs must appear in a $k$-uniform set system of certain size. One can ask analogous questions in Ramsey theory. Once we know the maximum size of a~structure which does not contain some unavoidable pattern, we may ask how many such patterns are bound to appear in every structure whose size exceeds this maximum. This direction of research has also been explored in the past. For example, a well-known problem of Erd{\H{o}}s is to determine the minimum number of monochromatic $k$-vertex cliques in a $2$-coloring of the edges of $K_n$; see, e.g.,~\cite{Co12, FrRo93, Th89}. This may be viewed as a natural extension of Ramsey's theorem. In this paper, we consider a similar generalization of another classical result in Ramsey theory, the famous theorem of Erd{\H{o}}s and Szekeres~\cite{ErSz35}, which states that for every positive integer $k$, any sequence of more than $k^2$ numbers contains a monotone (that is, monotonically increasing or monotonically decreasing) subsequence of length $k+1$. To be more precise, we shall be interested in the following very natural problem. \begin{prob} \label{prob:main} For every $k$ and $n$, determine the minimum number of monotone subsequences of length $k+1$ in a sequence of $n$ numbers. \end{prob} It is not clear when Problem~\ref{prob:main} was originally posed. It appears first in print in a paper of Myers~\cite{My02}, who attributes it to Albert, Atkinson, and Holton. It follows from the aforementioned theorem of Erd{\H{o}}s and Szeker{\'e}s that every sequence of $n$ numbers contains at least $n-k^2$ monotone subsequences of length $k+1$. When $n \le k^2+k$, this is easily seen to be sharp by considering a sequence built from $k$ increasing sequences of lengths $k$ or $k+1$ by concatenating them in decreasing order, such as the sequence $\tau_{k,n}$ defined below. Without loss of generality, we may restrict our attention to sequences that are permutations of the set $\{1, \ldots, n\}$, which we shall from now on abbreviate by~$[n]$. Let us denote by $S_n$ the set of all permutations of $[n]$. Following~\cite{My02}, given a permutation $\sigma \in S_n$, we let $m_k(\sigma)$ denote the number of monotone subsequences of length $k+1$ in $\sigma$ and let \[ m_k(n) = \min\{m_k(\sigma) \colon \sigma \in S_n \}. \] In order to give an upper bound on $m_k(n)$, consider the permutation $\tau_{k,n}$ described by \[ \begin{array}{llll} \lfloor (k-1)n/k \rfloor + 1, & \lfloor (k-1)n/k \rfloor + 2, & \ldots, & n \\ \lfloor (k-2)n/k \rfloor + 1, & \lfloor (k-2)n/k \rfloor + 2, & \ldots, & \lfloor (k-1)n/k \rfloor \\ \vdots & \vdots & & \vdots \\ 1, & 2, & \ldots, & \lfloor n/k \rfloor. \end{array} \] Let $r_{k,n}$ be the unique number $r \in \{0, \ldots, k-1\}$ satisfying $r \equiv n \pmod k$. Since $\tau_{k,n}$ contains no decreasing subsequences of length $k+1$, it is easy to see that \begin{equation} \label{eq:mktau} m_k(\tau_{k,n}) = r_{k,n} \binom{\lceil n/k \rceil}{k+1} + (k-r_{k,n}) \binom{\lfloor n/k \rfloor}{k+1}. \end{equation} It seems quite natural to guess that $m_k(n) = m_k(\tau_{k,n})$ for all $k$ and $n$, that is, that $\tau_{k,n}$ contains the minimum number of monotone subsequences of length $k+1$ among all permutations of $[n]$. This was conjectured by Myers~\cite{My02}, who noticed that Goodman's formula~\cite{Go59}, indeed proves that $m_2(n) = m_2(\tau_{2,n})$ for all $n$ and yields a characterisation of all permutations achieving equality. \begin{conj}[Myers~\cite{My02}] \label{conj:main} Let $n$ and $k$ be positive integers. In any permutation of $[n]$, there are at least $m_k(\tau_{k,n})$ monotone subsequences of length $k+1$. \end{conj} Very recently, Balogh et al.~\cite{BaHuLiPiUdVo} proved this conjecture for $k=3$ and sufficiently large $n$ and described all extremal permutations. Their proof uses computer assistance and is based on the framework of flag algebras. \subsection{Our results} In this paper, we provide first evidence supporting Conjecture~\ref{conj:main} for large $k$. Our main result is that the conjecture holds for all sufficiently large $k$, as long as $n$ is not much larger than $k^2$. This provides a good understanding of how the minimum number of monotone subsequences of length $k+1$ grows in a short interval above the extremal threshold. Our results are similar in spirit to the ones of~\cite{Er62, LoSi83}, which determine the minimum number of cliques in graphs whose number of edges is slightly supercritical (larger than the Tur{\'a}n number for the clique). \begin{thm} \label{thm:main} There exist an integer $k_0$ and a positive real $c$ such that $m_k(n) = m_k(\tau_{k,n})$ for all $k$ and $n$ satisfying $k \ge k_0$ and $n \le k^2 + c k^{3/2}/\log k$. Moreover, if $n \neq k^2+k+1$ and $m_k(\sigma) = m_k(n)$ for some $\sigma \in S_n$, then $\sigma$ contains monotone subsequences of length $k+1$ of only one type (increasing or decreasing). \end{thm} \begin{figure}[h] \centering \begin{tabular}{ccccc} \begin{tikzpicture} \def0.8{0.35} \def\sigma{ {9, 10, 11, 12, 13, 5, 6, 7, 8, 1, 2, 3, 4} } \draw[step=0.8,black,very thin] (0,0) grid (12*0.8,12*0.8); \foreach \x in {0, ..., 12} { \filldraw (\x*0.8, \sigma[\x]*0.8-0.8) circle[radius=2pt]; } \end{tikzpicture} & \quad & \begin{tikzpicture} \def0.8{0.35} \def\sigma{ {10, 6, 11, 12, 13, 2, 7, 8, 9, 1, 3, 4, 5} } \draw[step=0.8,black,very thin] (0,0) grid (12*0.8,12*0.8); \foreach \x in {0, ..., 12} { \filldraw (\x*0.8, \sigma[\x]*0.8-0.8) circle[radius=2pt]; } \end{tikzpicture} & \quad & \begin{tikzpicture} \def0.8{0.35} \def\sigma{ {10, 6, 11, 12, 13, 3, 7, 8, 9, 1, 2, 4, 5} } \draw[step=0.8,black,very thin] (0,0) grid (12*0.8,12*0.8); \foreach \x in {0, ..., 12} { \filldraw (\x*0.8, \sigma[\x]*0.8-0.8) circle[radius=2pt]; } \end{tikzpicture} \end{tabular} \caption{Permutations $\tau_{3,13}$, $\sigma_3^1$, and $\sigma_3^2$.} \label{fig:sigma-ik} \end{figure} Somewhat surprisingly, if $n = k^2+k+1$, then there are $\sigma \in S_n$ with $m_k(\sigma) = m_k(n) = 2k+1$ which contain both increasing and decreasing subsequences of length $k+1$. Two such permutations are $\sigma_k^1$ and $\sigma_k^2$, where $\sigma_k^i$ is described by \[ \begin{array}{llllll} k^2+1, & k^2-k, & k^2+2, & k^2+3, & \ldots, & k^2+k+1, \\ & k^2-2k-1, & k^2-k+1, & k^2-k+2, & \ldots, & k^2,\\ & \vdots & \vdots & \vdots & & \vdots \\ & k+3, & 2k+5, & 2k+6, & \ldots, & 3k+4, \\ & 1+i, & k+4, & k+5, & \ldots, & 2k+3,\\ & 1, & 4-i, & 4, & \ldots, & k+2. \end{array} \] One can check that $\sigma_k^i$ contains $2k+1-i$ increasing subsequences of length $k+1$ and $i$ decreasing subsequences of length $k+1$, see Figure~\ref{fig:sigma-ik}. However, no permutation $\sigma$ with $m_k(\sigma) = m_k(n) = 2k+1$ can have more monotone subsequences of length $k+1$ of the `odd' type than $\sigma_k^2$. It will follow from our proof of Theorem~\ref{thm:main} that for each extremal permutation, at least $2k-1$ out of its $2k+1$ monotone subsequences of length $k+1$ are of the same type. For details, we refer the reader to Theorem~\ref{thm:posets} and Example~\ref{example:extremal-posets}. \subsection{Chains and antichains in posets} Every permutation admits a natural representation as a poset (partially ordered set) in which its increasing and decreasing subsequences are mapped to chains and antichains, respectively. Indeed, given a permutation $\sigma$ of $[n]$, one may define a binary relation $\le_\sigma$ on $[n]$ by letting $i \le_\sigma j$ if and only if $i \le j$ and $\sigma(i) \le \sigma(j)$. It is not hard to see that $P_\sigma = ([n], \le_\sigma)$ is a poset whose chains and antichains are in a one-to-one length-preserving\footnote{By `length' of a chain or an antichain we mean the number of its elements.} correspondence with increasing and decreasing subsequences in $\sigma$, respectively. Via this correspondence, one may easily deduce the theorem of Erd{\H{o}}s and Szekeres from the famous theorem of Dilworth~\cite{Di50}, which says that every poset containing no antichain with $k+1$ elements admits a partition into $k$ chains, or its much easier to prove dual version (due to Mirsky~\cite{Mi71}), which says that every poset containing no chain with $k+1$ elements can be partitioned into $k$ antichains, see Section~\ref{sec:decomposition}. Let us call a set $A$ of elements of a poset \emph{homogenous} if $A$ is a chain or an antichain. A natural generalization of Problem~\ref{prob:main} to posets would be the following. \begin{prob} \label{prob:posets} For every $k$ and $n$, determine the minimum number of homogenous $(k+1)$-element sets in a poset with $n$ elements. \end{prob} Given a poset $P$, we let $h_k(P)$ denote the number of homogenous sets of cardinality $k+1$ in $P$ and \[ h_k(n) = \min\{h_k(P) \colon \text{$P$ is a poset with $n$ elements}\}. \] It follows from the above discussion that $h_k(n) \le m_k(n)$ and we think that it is natural to ask the following. \begin{quest} Is it true that $h_k(n) = m_k(n)$ for all $n$ and $k$? \end{quest} Clearly, not every poset is isomorphic to $P_\sigma$ for some permutation $\sigma$. In fact, this is the case precisely for posets of \emph{order dimension} at most two, that is, posets that are the intersection of two linear orders. Nevertheless, more for the sake of convenience rather than generality, we shall present our arguments using the language of posets. We remark here that several times in the proof, we will use the fact that we can `flip' our poset $P$, exchanging the roles of chains and antichains. That is, we will assume that there is a dual poset $P^*$ defined on the same set as $P$ such that every pair of elements is comparable in either $P$ or $P^*$ but not both of them. This is possible only for posets of order dimension at most two, that is, ones that represent permutations. Indeed, for every permutation $\sigma \in S_n$, we have $P_\sigma^* = P_{\sigma^*}$, where $\sigma^*(i) = n+1-\sigma(i)$. The converse statement was proved by Dushnik and Miller~\cite{DuMi41}. Let us now rephrase Theorem~\ref{thm:main} in the language of posets. \begin{thm} \label{thm:posets} There exist an integer $k_0$ and a positive real $c$ such that the following is true. Let $k$ and $n$ be integers satisfying $k \ge k_0$ and $n \le k^2 + c k^{3/2} / \log k$. If $P$ is an $n$-element poset of order dimension at most two, then \begin{equation} \label{eq:hkP-lower} h_k(P) \ge m_k(\tau_{k,n}). \end{equation} Moreover, if equality holds in~\eqref{eq:hkP-lower}, then $P$ can be decomposed into $k$ chains or $k$ antichains of length $\lfloor n/k \rfloor$ or $\lceil n/k \rceil$ each, unless $n = k^2 + k + 1$ and $P$ (or $P^*$) is one of the posets described in Example~\ref{example:extremal-posets} below. \end{thm} \begin{example} \label{example:extremal-posets} Suppose that $n = k^2+k+1$ and observe that $m_k(\tau_{k,n}) = 2k+1$. We describe two families of $n$-element posets with exactly $2k+1$ homogenous $(k+1)$-sets that contain both chains and antichains with $k+1$ elements. Each of these posets has precisely $k+1$ minimal elements; denote the set of these minimal elements by $A_1$. Moreover, $P \setminus A_1$ can be decomposed into $k$ chains as well as into $k$ antichains. In particular, each chain and each antichain in every such decomposition has precisely $k$ elements. Furthermore, $P \setminus A_1$ contains only $k$ chains of length $k$. Let $A_2$ be the set of minimal elements of $P \setminus A_1$ and note that $|A_2| = k$. The comparability graph of the subposet of $P$ induced by $A_1 \cup A_2$ is either (i) a path with $2k+1$ vertices or (ii) the disjoint union of a path with $2k-1$ vertices and an edge. Moreover, if (ii) holds, then one of the elements of $A_1$ belonging to the path of length $2k-1$ is smaller than the second smallest element of the unique $k$-element chain in $P \setminus A_1$ whose smallest element is the unique element of $A_2$ that does not belong to the path. One can check that if (i) holds, then $P$ has precisely $2k$ chains and one antichain with $k+1$ elements and that if (ii) holds, then $P$ has precisely $2k-1$ chains and $2$ antichains with $k+1$ elements. Finally, in both cases, there exist posets of order dimension at most two fitting the description. Two examples of such posets are $P_{\sigma_k^1}$ and $P_{\sigma_k^2}$, where $\sigma_k^1$ and $\sigma_k^2$ are the permutations defined below Theorem~\ref{thm:main}; see Figure~\ref{fig:posets-sigma-ik}. \end{example} \begin{figure}[h] \centering \begin{tabular}{ccc} \begin{tikzpicture} \def0.8{0.8} \def1.4{1.4} \def\p{ {10,6,2,1,11,7,3,12,8,4,13,9,5} } \foreach \x [evaluate=\x as \lab using ({\p[\x]})] in {0, ..., 3} { \node [label=right:$\lab$] at (2*\x*0.8, 0) {}; \filldraw (2*\x*0.8, 0) circle[radius=2pt]; } \foreach \x in {0, ..., 2} { \draw [very thick] (2*\x*0.8, 0) -- (2*\x*0.8 + 0.8, 1.4) -- (2*\x*0.8 + 2*0.8, 0); \draw (2*\x*0.8 + 0.8, 1.4) -- (2*\x*0.8 + 0.8, 3*1.4); } \foreach \y in {1, ..., 3} { \foreach \x [evaluate=\x as \lab using ({\p[3*\y+\x+1]})] in {0, ..., 2} { \node [label=right:$\lab$] at (2*\x*0.8+0.8, \y*1.4) {}; \filldraw (2*\x*0.8+0.8, \y*1.4) circle[radius=2pt]; } } \end{tikzpicture} & \qquad\qquad\qquad & \begin{tikzpicture} \def0.8{0.8} \def1.4{1.4} \def\p{ {10,6,3,1,11,7,2,12,8,4,13,9,5} } \foreach \x [evaluate=\x as \lab using ({\p[\x]})] in {0, ..., 3} { \node [label=right:$\lab$] at (2*\x*0.8, 0) {}; \filldraw (2*\x*0.8, 0) circle[radius=2pt]; } \foreach \x in {0, ..., 1} { \draw [very thick] (2*\x*0.8, 0) -- (2*\x*0.8 + 0.8, 1.4) -- (2*\x*0.8 + 2*0.8, 0); } \draw (4*0.8, 0) -- (5*0.8, 2*1.4); \draw [very thick] (5*0.8, 1.4) -- (6*0.8, 0); \foreach \x in {0, ..., 2} { \draw (2*\x*0.8 + 0.8, 1.4) -- (2*\x*0.8 + 0.8, 3*1.4); } \foreach \y in {1, ..., 3} { \foreach \x [evaluate=\x as \lab using ({\p[3*\y+\x+1]})] in {0, ..., 2} { \node [label=right:$\lab$] at (2*\x*0.8+0.8, \y*1.4) {}; \filldraw (2*\x*0.8+0.8, \y*1.4) circle[radius=2pt]; } } \end{tikzpicture} \end{tabular} \caption{Hasse diagrams of posets $P_{\sigma_3^1}$ and $P_{\sigma_3^2}$. The edges of the comparability graph induced by $A_1 \cup A_2$ are thickened.} \label{fig:posets-sigma-ik} \end{figure} \subsection{Outline of the paper} The remainder of the paper is organized as follows. In Section~\ref{sec:decomposition}, we describe a canonical decomposition of an arbitrary poset into antichains and introduce several pieces of notation used in the proof of Theorem~\ref{thm:posets} and in Section~\ref{sec:auxiliary-lemmas}, we collect several auxiliary lemmas. In Section~\ref{sec:outline-proof}, we present a brief outline of the proof of Theorem~\ref{thm:posets}. Section~\ref{sec:large-surplus} is devoted to the proof of one of our main lemmas, which provides a lower bound on the number of homogenous sets in posets with large `surplus' (this notion will be defined in Section~\ref{sec:outline-proof}). Finally, Section~\ref{sec:proof} contains the proof of Theorem~\ref{thm:posets}. We close the paper with several concluding remarks. \section{Decomposition into antichains} \label{sec:decomposition} In our arguments, we shall rely on the following canonical decomposition of an arbitrary poset into antichains, cf.\ Mirsky's theorem~\cite{Mi71}. Fix an arbitrary poset $P$. Recall that the height and the width of $P$, which we shall denote by $h(P)$ and $w(P)$, are the cardinalities of the largest chain and the largest antichain in $P$, respectively. For each positive integer $i$, let \[ A_i = \{x \in P \colon \text{the longest chain $L$ with $\max L = x$ has $i$ elements}\}. \] In other words, $A_1$ is the set of minimal elements of $P$ and for every $i \ge 1$, $A_{i+1}$ is the set of minimal elements in $P \setminus (A_1 \cup \ldots \cup A_i)$. For each $i$, the set $A_i$ is an antichain and thus $|A_i| \le w(P)$, and \[ P = \bigcup_{i=1}^{h(P)} A_i. \] For each $i$ with $1 \le i < h(P)$, let $G_i$ be the bipartite graph on the vertex set $A_i \cup A_{i+1}$ whose edges are all pairs $xy$ with $x \in A_i$ and $y \in A_{i+1}$ such that $x \le y$. In other words, $G_i$ is the Hasse diagram of the subposet of $P$ induced by $A_i \cup A_{i+1}$. Observe that each vertex in $A_{i+1}$ has at least one $G_i$-neighbor in $A_i$ (as otherwise it would belong to $A_1 \cup \ldots \cup A_i$). On the other hand, it is possible that some vertices in $A_i$ have degree zero in $G_i$ (as clearly not all elements of $P$ have to belong to some chain of maximum length). Let $h = h(P)$. For every $i \in [h]$ and $x \in A_i$, we let $u_i(x)$ be the number of chains $L \subseteq P$ of length $h - i + 1$ with $\min L = x$. Observe that $u_h(x) = 1$ for every $x \in A_h$ and that for all $i \in [h-1]$ and $x \in A_i$, \[ u_i(x) = |\{(x_{i+1}, \ldots, x_h) \in A_{i+1} \times \ldots \times A_h \colon x \le x_{i+1} \le \ldots \le x_h\}|. \] Upon making this definition, one easily verifies that the number of chains of length $h$ in $P$ is $\sum_{x \in A_1} u_1(x)$ and that for each $i \in [h-1]$ and $x \in A_i$, \begin{equation} \label{eq:ui-uip} u_i(x) = \sum_{xy \in G_i} u_{i+1}(y). \end{equation} Since no element $x \in A_i$ with $u_i(x) = 0$ will be contributing anything to the total count of chains of length $h$, we shall often be focusing our attention on the set $A_i' \subseteq A_i$ defined by \[ A_i' = \{x \in A_i \colon u_i(x) \ge 1\}. \] Let us note here for future reference that~\eqref{eq:ui-uip} implies that there are no edges of $G_i$ between $A_{i+1}'$ and $A_i \setminus A_i'$. As we shall often estimate the sum of $u_i(x)$ over all $x \in A_i$, let us abbreviate it by $\Sigma_i$. That is, for each $i \in [h]$, let \[ \Sigma_i = \sum_{x \in A_i} u_i(x). \] Since each $y \in A_{i+1}$ has at least one $G_i$-neighbor in $A_i$, it follows from~\eqref{eq:ui-uip} that \begin{equation} \label{eq:sum-ui-monotone} \Sigma_{i} \ge \Sigma_{i+1}. \end{equation} Clearly, \eqref{eq:sum-ui-monotone} holds with equality if and only if each $y \in A_{i+1}'$ has exactly one $G_i$-neighbor in $A_i$. This naturally leads to the final definition of this section. Namely, for each $i \in [h-1]$, we let \[ B_{i+1} = \{y \in A_{i+1}' \colon \deg_{G_i}(y) = 1\}. \] \section{Auxiliary lemmas} \label{sec:auxiliary-lemmas} In this section, we collect a few auxiliary lemmas that will be repeatedly used in the proof of Theorem~\ref{thm:posets}. Our first lemma is a straightforward corollary of the Kruskal--Katona theorem~\cite{Ka68, Kr63}. \begin{lemma} \label{lemma:KK} Suppose that $a \ge b > 0$, let $\mathcal{F}$ be an arbitrary family of $a$-element sets, and define \[ \partial_b \mathcal{F} = \{B \colon \text{$|B| = b$ and $B \subseteq A$ for some $A \in \mathcal{F}$}\}. \] Then \[ |\partial_b \mathcal{F}| \ge \min \{ |\mathcal{F}|/2, 2^b \}. \] \end{lemma} \begin{proof} If $\mathcal{F} = \emptyset$, $a=b$, or $a \ge 2b$, then the assertion of the lemma is trivial as \begin{equation} \label{eq:a-choose-b} \binom{2b}{b} \ge 2^b. \end{equation} Otherwise, let $m$ be the smallest integer such that $\binom{m}{a} > |\mathcal{F}|$. By the Kruskal--Katona theorem~\cite{Ka68,Kr63}, $|\partial_b \mathcal{F}| \ge \binom{m-1}{b}$. If $m-1 \ge 2b$, then the conclusion follows from~\eqref{eq:a-choose-b}. Otherwise, since $b < a \le m \le 2b$, then $\binom{m}{a} \le \binom{m}{b+1}$ and consequently, \[ |\partial_b \mathcal{F}| \ge \binom{m-1}{b} > \binom{m-1}{b} \frac{|\mathcal{F}|}{\binom{m}{a}} \ge \frac{\binom{m-1}{b}}{\binom{m}{b+1}} |\mathcal{F}| = \frac{b+1}{m} |\mathcal{F}| \ge \frac{|\mathcal{F}|}{2}.\qedhere \] \end{proof} Our second lemma will be essential in proving a lower bound on the number of chains of length $k+1$ in a poset of height larger than $k+1$ in terms of the number of chains of maximum length. The lemma is somewhat abstract, but it is immediately followed by a much more concrete corollary. \begin{lemma} \label{lemma:signatures} Suppose that $M$ is a positive integer, $X$ and $Y$ are arbitrary sets, and $f_1, \ldots, f_M \colon X \to Y$ are pairwise different functions. There exist sets $X_1, \ldots, X_M \subseteq X$ with $|X_i| \le \log_2 M$ for all $i \in [M]$ such that \begin{equation} \label{eq:fifj} f_i|_{X_i \cup X_j} \neq f_j|_{X_i \cup X_j} \quad \text{for all $i \neq j$}. \end{equation} \end{lemma} \begin{proof} We prove the lemma by induction on $M$. The statement is trivial if $M=1$ (one takes $I_1 = \emptyset$), so we may assume that $M \ge 2$. Since $f_1, \ldots, f_M$ are pairwise different, there is an $x \in X$ such that not all $f_i$ take the same value at $x$. For each $y \in Y$, let \[ I(x, y) = \{i \in [M] \colon f_i(x) = y\} \] and let $y \in Y$ be a value that maximizes $|I(x, y)|$. Note that $|I(x, y)| < M$ by our choice of $x$ and that $|I(x, z)| \le M/2$ for each $z \in Y \setminus \{y\}$. We apply the inductive assumption separately to $\{f_i \colon i \in I(x, z)\}$ for each $z \in Y$ with $I(x,z) \neq \emptyset$ to obtain sets $X_1', \ldots, X_M' \subseteq X$ such that $|X_i'| < \log_2M$ for every~$i$, $|X_i'| \le \log_2 (M/2)$ for every $i \notin I(x,y)$, and~\eqref{eq:fifj} holds for each pair $\{i,j\}$ which is fully contained in one of the sets $I(x,z)$. It is straightforward to check that the sets $X_1, \ldots, X_M$ defined by \[ X_i = \begin{cases} X_i' & \text{if $i \in I(x,y)$,} \\ X_i' \cup \{x\} & \text{if $i \notin I(x, y)$,} \end{cases} \] satisfy the assertion of the lemma. \end{proof} \begin{cor} \label{cor:signatures} Let $k$, $\ell$, and $M$ be positive integers, let $P$ be a poset of height $k+\ell$, and suppose that $m := \log_2 M + 1 \le k/4$. \begin{enumerate}[(i)] \item \label{item:signatures-global} If $P$ contains at least $M$ chains of length $k+\ell$, then it contains at least \[ \exp \left( - \frac{2(\ell-1)m}{k} \right) \cdot M \binom{k+\ell}{k+1} \] chains of length $k+1$. \item \label{item:signatures-local} Given any $y \in P$, (\ref{item:signatures-global}) still holds if we replace `chains' with `chains containing $y$'. \end{enumerate} \end{cor} \begin{proof} We prove (\ref{item:signatures-global}) and (\ref{item:signatures-local}) simultaneously. Suppose that $L_1, \ldots, L_M$ are pairwise distinct chains of length $k+\ell$. For (\ref{item:signatures-local}), assume moreover that each $L_i$ contains $y$. Viewing each chain $L_i$ as a function from $[k+\ell]$ to $P$, we invoke Lemma~\ref{lemma:signatures} to obtain sets $X_1, \ldots, X_M \subseteq [k+\ell]$ such that $|X_i| \le \log_2M$ for each $i$ and $L_i(X_i \cup X_j) \neq L_j(X_i \cup X_j)$ whenever $i \neq j$. For (\ref{item:signatures-local}), add to each $X_i$ the unique index $x$ such that $L_i(x) = y$. By the definition of $m$, we have that $|X_i| \le m$ for each $i$. Let $N$ denote the number of chains of length $k+1$ that are obtained by fixing an $i \in [M]$ and an arbitrary $(k+1)$-set $C_i \subseteq [k+\ell]$ such that $X_i \subseteq C_i$ and considering the set $L_i(C_i)$. Note crucially that for any $i \neq j$ and any choice of $C_i$ and $C_j$ as above, the sets $L_i(C_i)$ and $L_j(C_j)$ are \emph{different} chains (containing $y$). Indeed, since $L_i$ and $L_j$ are chains of maximum length, then for every $z \in L_i \cap L_j$, there is a unique $x \in [k+\ell]$ such that $L_i(x) = z = L_j(x)$. Hence, \[ \begin{split} N & = \sum_{i=1}^M \binom{|L_i| - |X_i|}{k+1 - |X_i|} = \sum_{i=1}^M \binom{k+\ell - |X_i|}{\ell - 1} \ge M \binom{k+\ell-m}{\ell-1} \\ & \ge \left(1 - \frac{1}{k-m}\right)^{(\ell-1) m} \cdot M\binom{k+\ell}{\ell-1} \ge \exp\left(-\frac{2(\ell - 1)m}{k}\right) \cdot M \binom{k+\ell}{k+1}. \end{split} \] Above, we used the fact that $\binom{a}{b} \ge \left(1 - \frac{1}{a-b}\right)^b \binom{a+1}{b}$ and that $1-x \ge e^{-3x/2}$ if $x \le 1/3$. \end{proof} We close this section with a simple lower bound on the number of connected sets in trees. We shall use this bound in the analysis of one of the almost extremal cases in the proof of Theorem~\ref{thm:posets}. \begin{lemma} \label{lemma:connected-sets} If $1 \le c \le t$, then every tree with $t$ vertices contains at least $t-c+1$ connected subsets with $c$ elements. \end{lemma} \begin{proof} We prove the statement by induction on $t-c$. It is certainly true if $t = c$. Assume that $t \ge c+1$ and let $T$ be a tree with $t$ vertices. Let $v$ be an arbitrary leaf of $T$ and set $T' = T - v$. By the inductive assumption, $T'$ contains at least $t-c$ connected subsets with $c$ elements. On the other hand, it is easy to check that $v$ is contained in at least one connected subset of $T$ of any given size between $1$ and $t$. \end{proof} \section{Outline of the proof} \label{sec:outline-proof} Roughly speaking, our proof of Theorem~\ref{thm:posets} is a combination of a stability-type argument and an induction on $n$. More precisely, given an $n$-element poset $P$, we either find an element $x \in P$ which belongs to at least $m_k(\tau_{k,n}) - m_k(\tau_{k,n-1})$ homogenous $(k+1)$-sets, in which case we may simply appeal to the inductive assumption on $P \setminus \{x\}$, or we show more `directly' that $P$ contains at least $m_k(\tau_{k,n})$ homogenous $(k+1)$-sets. Some extra work is needed to deduce the claimed structural description of $P$ when $h_k(P) = m_k(\tau_{k,n})$. At all times, we rely heavily on the assumption that the order dimension of $P$ is at most two and hence $P$, as well as each of its induced subposets, has a dual poset $P^*$. This assumption allows us to focus on counting chains, since we may always replace $P$ with $P^*$, exchanging the roles of chains and antichains. We shall tacitly assume that $h(P) \ge w(P)$, as $h(P^*) = w(P)$, and that $h(P) < \lceil n/k \rceil$, as otherwise each element of a longest chain in $P$ belongs to at least $m_k(\tau_{k,n}) - m_k(\tau_{k,n-1})$ chains of length $k+1$. We first show that if $P$ is `far' from being a union of $k$ chains (or $k$ antichains), then the number of homogenous $(k+1)$-sets is much greater than $m_k(\tau_{k,n})$. To this end, we define a simple parameter termed surplus which measures the distance between a poset and a union of $k$ chains. Let $P$ be an arbitrary poset of height $h$ and let $k$ be an integer. The \emph{$k$-surplus} of $P$, denoted by $s_k(P)$, is defined by $s_k(P) = n - hk$. Observe that \begin{equation} \label{eq:surplus} s_k(P) = \sum_{i = 1}^h (|A_i| - k), \end{equation} where $(A_i)_{i=1}^h$ is the canonical decomposition of $P$ into antichains. In Section~\ref{sec:large-surplus}, we show that if $s_k(P) = \Omega(k)$, then $h_k(P) = 2^{\Omega(\sqrt{k})} \gg m_k(\tau_{k,n})$. On the other hand, $s_k(P) = o(k)$ together with $h(P) < \lceil n/k \rceil$ imply that $h(P) = \lceil n/k \rceil - 1$. In the remainder of the proof, we prove a sequence of lower bounds on $\Sigma_1$, the number of chains of maximum length. The proof of each of these bounds relies on the analysis of the graphs $G_i$ for various indices $i$ such that $A_i \cup A_{i+1}$ contains an antichain with $k+1$ elements. Roughly speaking, we show that for each such $i$, either $\Sigma_i - \Sigma_{i+1}$ is large or $A_i \cup A_{i+1}$ contains many $(k+1)$-element antichains. Each of these bounds gives some sufficient conditions on $P$ which imply that $h_k(P) > m_k(\tau_{k,n})$; here, we use Corollary~\ref{cor:signatures} to translate a lower bound on the number of chains of length $h(P)$ into a lower bound on the number of chains of length $k+1$. If $P$ does not satisfy any of these conditions, then the canonical decomposition of $P$ into antichains becomes greatly restricted. In particular, there are very few indices $i$ with $|A_i| > k$ and $|A_i'| \approx |A_i|$ for all $i$. Finally, some careful case analysis, which involves counting both chains and antichains with $k+1$ elements, shows that $h_k(P) \ge m_k(\tau_{k,n})$ and this inequality is strict unless $n = k^2+k+1$ and $P$ (or $P^*$) is one of the posets described in Example~\ref{example:extremal-posets}. \section{Posets with large surplus} \label{sec:large-surplus} In this section, we prove one of our key lemmas. It says that posets with large $k$-surplus and no `bottlenecks' (small sets whose deletion reduces the height) contain many chains of maximum length or many $(k+1)$-element antichains. \begin{lemma} \label{lemma:large-surplus} Let $d$, $k$, and $s$ be integers satisfying $1 \le d \le k$ and suppose that $P$ is a poset such that $s_k(P) \ge s$ and deletion of no $s/2$ elements reduces the height of $P$. Then $P$ contains either at least $2^d$ antichains with $k+1$ elements or at least $2^{\lfloor s/(2d) \rfloor}$ chains of length $h(P)$. \end{lemma} \begin{proof} We fix $d$ and $k$ with $1 \le d \le k$ and prove the statement by induction on $s$. If $0 \le s < 2d$, then the assertion of the lemma holds vacuously. Suppose now that $s \ge 2d$ and that $P$ satisfies the assumptions of the lemma. Let $(A_j)_{j=1}^{h(P)}$ be the canonical decomposition of $P$ into antichains and let $i$ be the smallest index such that $|A_i| > k$; such $i$ exists since $s_k(P) > 0$, see~\eqref{eq:surplus}. Since $A_i$ is an antichain, we may assume that $|A_i| \le k+d$ since otherwise the number $N$ of $(k+1)$-element antichains in $A_i$ alone satisfies \[ N \ge \binom{k+d+1}{k+1} = \binom{k+d+1}{d} = \prod_{j=1}^d \frac{k+j+1}{j} \ge 2^d, \] where the last inequality follows from our assumption that $d \le k$. Recall the definition of $G_i$ and let $B_{i+1}$ be the set of all $y \in A_{i+1}$ with at most (exactly) one $G_i$-neighbor in $A_i$. For every $Y \subseteq B_{i+1}$, the set $(A_i \setminus N_{G_i}(Y)) \cup Y$ is an antichain with at least $|A_i|$ elements and therefore we may further assume that $|B_{i+1}| < d$ as otherwise $A_i \cup B_{i+1}$ contains at least $2^d$ antichains of size $k+1$. To see this, for every $Y \subseteq B_{i+1}$ with $|Y| \le d$, consider an arbitrary $(k+1)$-element set $L$ with $Y \subseteq L \subseteq (A_i \setminus N_{G_i}(Y)) \cup Y)$. By~\eqref{eq:ui-uip} and~\eqref{eq:sum-ui-monotone}, \begin{equation} \label{eq:chains} \sum_{x \in A_1} u_1(x) \ge \sum_{x \in A_i} u_i(x) = \sum_{y \in A_{i+1}} \deg_{G_i}(y) u_{i+1}(y) \ge 2 \sum_{y \in A_{i+1} \setminus B_{i+1}} u_{i+1}(y). \end{equation} Let $h = h(P)$ and observe that the sum in the right-hand side of~\eqref{eq:chains} is precisely the number of chains of length $h-i$ in the poset $P'$ obtained from $P$ by deleting $A_1 \cup \ldots \cup A_i \cup B_{i+1}$. In order to estimate this sum, we shall apply the inductive assumption to $P'$. First, note that $h(P') \le h-i$, as $A_{i+1} \setminus B_{i+1}, A_{i+1}, \ldots, A_h$ partition $P'$ into $h-i$ antichains. Moreover, if $h(P' \setminus X) < h-i$ for some $X \subseteq P'$, then $h(P \setminus (B_{i+1} \cup X)) < h$ as every chain in $P$ contains at most one element from each of $A_1, \ldots, A_i$. Therefore, not only $h(P') = h-i$, as $|B_{i+1}| < d \leq s/2$, but also the deletion of no $s/2-d$ elements reduces the height of $P'$. Furthermore, \[ \begin{split} s_k(P') & = |P'| - k(h-i) = |P| - \sum_{j=1}^i (|A_j| - k) - |B_{i+1}| - hk \\ & \ge s_k(P) - (|A_i| - k) - |B_{i+1}| > s-2d. \end{split} \] The first inequality above follows from the minimality of $i$ and the second inequality from the assumptions that $|A_i| \le k+d$ and $|B_{i+1}| < d$. Hence, $P'$ satisfies the assumptions of the lemma with $s$ replaced by $s-2d$. This means that either $P'$ contains at least $2^d$ antichains of size $k+1$ or \[ \sum_{y \in A_{i+1} \setminus B_{i+1}} u_{i+1}(y) \ge 2^{\lfloor s/(2d) \rfloor - 1}. \] By~\eqref{eq:chains}, this completes the proof of the lemma. \end{proof} \begin{cor} \label{cor:large-surplus} Let $k$ and $t$ be integers satisfying $0 < t \le k/2$ and suppose that $P$ is a poset of order dimension at most two such that $h(P) \ge w(P)$ and $s_k(P) \ge 3t$. Then $P$ contains at least $2^{\sqrt{t}-1}$ homogenous $(k+1)$-sets. \end{cor} \begin{proof} We first `prune' $P$ by repeatedly performing the following two-step procedure: \begin{enumerate}[(1)] \item \label{item:prune-1} If $P$ contains a set $S$ of at most $t$ elements whose deletion reduces the height of $P$, then remove the smallest such $S$ from $P$. \item \label{item:prune-2} If $h(P) < w(P)$, then replace $P$ with $P^*$, exchanging the roles of chains and antichains. \end{enumerate} Let us list several properties of the `pruning' procedure. First, performing (\ref{item:prune-1}) decreases the height of $P$ by exactly one at the cost of deleting at most $t$ elements. Thus, each time (\ref{item:prune-1}) is executed, the $k$-surplus of $P$ increases by at least $k-t$. Second, since in the beginning and after (\ref{item:prune-2}) is performed, $h(P) \ge w(P)$, step (\ref{item:prune-2}) can be executed only in conjunction with (\ref{item:prune-1}). Third, each time (\ref{item:prune-2}) is performed, it increases the height of $P$ by exactly one, thus reducing the $k$-surplus of $P$ by $k$. Moreover, this cannot happen in two consecutive rounds, since immediately after (\ref{item:prune-2}) is triggered, $h(P) > w(P)$. Therefore, letting $P'$ denote the final outcome of the `pruning' procedure and $r$ the number of rounds, we have (recall that $k \ge 2t$) \[ \begin{split} s_k(P') & \ge s_k(P) + r(k-t) - \lceil r/2 \rceil k \ge 3t + \lfloor r / 2 \rfloor k - rt \\ & \ge 3t + (r-1)t - rt = 2t > 0. \end{split} \] In particular, \[ h(P') \ge w(P') \ge k + \frac{s_k(P')}{h(P')} > k. \] Since $P'$ clearly does not contain a set of $t$ elements whose deletion reduces the height of $P'$, Lemma~\ref{lemma:large-surplus} with $d = \lfloor \sqrt{t} \rfloor$ and $s = 2t$ implies that $P'$ contains either at least $2^{\sqrt{t}-1}$ antichains of size $k+1$ or at least $2^{\sqrt{t}}$ chains of length $h(P')$ and consequently, by Lemma~\ref{lemma:KK}, also at least $2^{\sqrt{t}-1}$ chains of length $k+1$. Finally, since $P$ contains either $P'$ or $(P')^*$, every homogenous set in $P'$ is also homogenous in $P$. \end{proof} \section{Proof of Theorem~\ref{thm:posets}} \label{sec:proof} Let $k$ be a sufficiently large integer. We prove the theorem by induction on $n$. The assertion is trivial when $n \le k^2$ as then $m_k(\tau_{k,n}) = 0$ and every poset with no chain of length $k+1$ can be covered by $k$ antichains, see Section~\ref{sec:decomposition}. Therefore, suppose that $k^2 < n \le k^2 c k^{3/2}/\log_2 k$, where $c = 1/300$, and let $P$ be an arbitrary $n$-element poset of order dimension at most two. Without loss of generality, we may assume that $h(P) \ge w(P)$ as otherwise we may replace $P$ by $P^*$, exchanging the roles of chains and antichains. Let $\ell$ and $q$ be the unique nonnegative integers satisfying $0 < q \le k$ and \begin{equation} \label{eq:n-q-k-ell} n = q \cdot (k+\ell+1) + (k-q) \cdot (k+\ell). \end{equation} In other words, we let $\ell = \lceil n / k \rceil - k - 1$ and $q = n - k(k+\ell)$. Our upper bound on $n$ implies that $\ell \le c\sqrt{k}/\log_2 k$, which we note here for future reference. Observe that $m_k(\tau_{k,n})$ is precisely the number of chains of length $k+1$ in the poset which is the disjoint union of $k$ pairwise incomparable chains: $q$ chains of length $k+\ell+1$ and $k-q$ chains of length $k + \ell$, cf.~\eqref{eq:mktau}. In particular, \begin{equation} \label{eq:mkn-mknm} m_k(\tau_{k,n}) - m_k(\tau_{k,n-1}) = \binom{k+\ell}{k}. \end{equation} With the view of~\eqref{eq:mkn-mknm}, we may and shall assume that $P$ contains no element $x$ that belongs to more than $\binom{k+\ell}{k}$ homogenous $(k+1)$-sets. Indeed, otherwise we could apply the inductive assumption to the poset $P \setminus \{x\}$ and conclude that $h_k(P) > m_k(\tau_{k,n})$. \subsection{Posets of height larger than $k+\ell$} Our inductive assumption allows us to easily deal with the case $h(P) \ge k + \ell + 1$. Indeed, every element of each longest chain in $P$ lies in at least $\binom{h(P)-1}{k}$ chains of length $k+1$ and hence by~\eqref{eq:mkn-mknm}, for any such element $x$, \begin{equation} \label{eq:induction} h_k(P) \ge h_k(P \setminus \{x\}) + \binom{h(P)-1}{k} \ge m_k(\tau_{k,n-1}) + \binom{k+\ell}{k} = m_k(\tau_{k,n}). \end{equation} Since we have promised to characterize all posets satisfying $h_k(P) = m_k(\tau_{k,n})$, we still need to analyze the case when all inequalities in~\eqref{eq:induction} are actually equalities. This means, in particular, that $h(P) = k + \ell + 1$ and that no longest chain intersects a $(k+1)$-element antichain. We claim that $P$ may be partitioned into $k$ chains. Let $x$ be the top element of some longest chain $L$ in~$P$. By the inductive assumption, $P \setminus \{x\}$ can be partitioned into $k$ chains or $k$ antichains. Let us argue that $P \setminus \{x\}$ can actually be partitioned into $k$ chains. When $n-1 = k^2$, this follows from Dilworth's theorem (or its dual version applied to $(P \setminus \{x\})^*$) as $h_k(P \setminus \{x\}) = m_k(\tau_{k,n-1}) = 0$. Otherwise, when $n-1 > k^2$, if $P \setminus \{x\}$ could be partitioned into $k$ antichains, then each of them would have to intersect the chain $L \setminus \{x\}$, which has at least $k$ elements, and one of them would have at least $k+1$ elements, contradicting our assumption above. Suppose that the $k$ chains decomposing $P \setminus \{x\}$ are $L_1, \ldots, L_k$. It suffices to show that $x \ge \max L_i$ for some $L_i$, since then $L_1, \ldots, L_{i-1}, L_i \cup \{x\}, L_{i+1}, \ldots, L_k$ form a partition of $P$ into $k$ chains. Let $y$ be the second largest element of $L$. Clearly, $y \in L_i$ for some $i$. If $y = \max L_i$, then there is nothing left to prove, so suppose that $z = \max L_i > y$ and consider the set $L \cup \{z\} \subseteq P$. It is easy to see that $y$ belongs to at least $\binom{k+\ell}{k} + \binom{k+\ell-1}{k-1}$ chains of length $k+1$, contradicting our assumption. \subsection{Posets of height smaller than $k+\ell$} The case $h(P) \le k+\ell-1$ can be easily resolved with the use of Corollary~\ref{cor:large-surplus}. Indeed, if $h(P) \le k+\ell-1$, then \[ s_k(P) \ge n - k(k+\ell-1) = k+q > k. \] Since $h(P) \ge w(P)$, Corollary~\ref{cor:large-surplus} implies that $h_k(P) \ge 2^{\sqrt{k/3}-1}$. On the other hand, our assumption that $n \le k^2 + ck^{3/2}/\log_2 k$ and~\eqref{eq:mktau} yield \begin{equation} \label{eq:mk-exp-sqrt-k} \begin{split} m_k(\tau_{k,n}) & \le k \binom{\lceil n/k \rceil}{k+1} \le k \binom{k+c\sqrt{k}/\log_2 k+1}{k+1} \\ & \le k \binom{2k}{c\sqrt{k}/\log_2 k} \le \left(\frac{2e\sqrt{k}\log_2k}{c}\right)^{\frac{c \sqrt{k}}{\log_2k}} < 2^{\sqrt{k}/4}. \end{split} \end{equation} The fourth inequality above is $\binom{a}{b} \le (ea/b)^b$ and the last inequality follows since $c < 1/2$ and $k$ is sufficiently large. \subsection{Posets of height $k+\ell$} In view of the above considerations, for the remainder of the proof, we may assume that \begin{equation} \label{eq:height-kpell} h(P) = k + \ell \ge w(P). \end{equation} Since $n \le h(P)w(P)$, this means, in particular, that $n > k^2+k$, as otherwise $\ell = 0$ and we have assumed above that $n > k^2$. As $n > k(k+\ell)$, see~\eqref{eq:n-q-k-ell}, assumption~\eqref{eq:height-kpell} implies that $P$ cannot be decomposed into $k$ chains or $k$ antichains. Thus, we shall show that $h_k(P) > m_k(\tau_{k,n})$, unless $n = k^2+k+1$ and $P$ is one of the posets described in Example~\ref{example:extremal-posets}. Observe that \[ m_k(\tau_{k,n}) = k \binom{k+\ell}{k+1} + q \binom{k+\ell}{k} = \left( q + \frac{k \ell}{k+1} \right) \binom{k+\ell}{k}, \] cf.~\eqref{eq:mktau} and~\eqref{eq:n-q-k-ell}. In particular, since $\ell, q \le k$ and $k$ is sufficiently large, \begin{equation} \label{eq:mktau-k-to-ell} m_k(\tau_{k,n}) \le (q + \ell) \binom{k+\ell}{\ell} \le (k + \ell) \binom{k+\ell}{\ell} < k^{2\ell}. \end{equation} \subsubsection{The key lemma} Let $(A_i)_{i=1}^{k+\ell}$ be the canonical decomposition of $P$ into antichains and recall the definition of $G_i$ from Section~\ref{sec:decomposition}. We shall provide various lower bounds on the number of homogenous sets by analyzing the graphs $G_i$ for various indices $i$ such that $A_i \cup A_{i+1}$ contains an antichain with $k+1$ elements. First and foremost, we shall be looking at $i \in F$, where \begin{equation} \label{eq:F} F = \{i \in [k+\ell] \colon |A_i| \ge k+1\}. \end{equation} We start by establishing a lower bound on the size of $F$. \begin{obs} \label{obs:F-size} For every $I \subseteq [k+\ell]$, \[ |F| \ge \frac{q + \sum_{i \in I} (k-|A_i|)}{\ell}. \] \end{obs} \begin{proof} Recall that the sets $(A_i)_{i = 1}^{k+\ell}$ form a partition of $P$ into antichains and that $|A_i| \le w(P) \le k+\ell$ for all $i$. Hence, \[ \begin{split} n - \sum_{i \in I} |A_i| & = \sum_{i \not\in I \cup F} |A_i| + \sum_{i \in F \setminus I} |A_i| \le (k + \ell - |I \cup F|) \cdot k + |F \setminus I| \cdot (k+\ell) \\ & = (k + \ell - |I|) \cdot k + |F \setminus I| \cdot \ell \le n - q - |I| \cdot k + |F| \cdot \ell.\qedhere \end{split} \] \end{proof} Recall the definitions of $u_i$, $\Sigma_i$, $A_i'$, and $B_i$ from Section~\ref{sec:decomposition}. The following lemma is key \begin{lemma} \label{lemma:F} If $i \in F \cap [k+\ell-1]$, then $A_i \cup B_{i+1}$ contains at least $2^{\min\{k, |B_{i+1}|\}}$ antichains with $k+1$ elements and \[ \Sigma_i \ge \Sigma_{i+1} + \sum_{y \in A_{i+1}' \setminus B_{i+1}} u_{i+1}(y) \ge \Sigma_{i+1} + |A_{i+1}'| - |B_{i+1}|. \] \end{lemma} \begin{proof} Note that for any $Y \subseteq B_{i+1}$, the set $Y \cup (A_i \setminus N_{G_i}(Y))$ is an antichain with at least $|A_i|$ elements. Since $|A_i| \ge k+1$, each of these antichains that additionally satisfies $|Y| \le k+1$ contains a $(k+1)$-element subset $L$ such that $L \cap B_{i+1} = Y$. This proves the first assertion of the lemma. The second assertion holds since \[ \Sigma_i = \sum_{xy \in G_i} u_{i+1}(y) = \sum_{y \in A_{i+1}'} u_{i+1}(y) \deg_{G_i}(y) \ge \sum_{y \in A_{i+1}'} u_{i+1}(y) + \sum_{i \in A_{i+1}' \setminus B_{i+1}} u_{i+1}(y).\qedhere \] \end{proof} \subsubsection{The first lower bound on $\min_i |A_i'|$} In view of Lemma~\ref{lemma:F}, we shall aim at proving a lower bound on the minimum size of $A_{i+1}'$. We first derive a somewhat weak bound on $\min_i |A_i'|$ from Corollary~\ref{cor:large-surplus}. \begin{claim} Either $h_k(P) > m_k(\tau_{k,n})$ or $|A_i'| \ge k/3$ for all $i \in [k+\ell]$, possibly after substituting $P^*$ for $P$. \end{claim} \begin{proof} Suppose that $|A_i'| < k/3$ for some $i$. If $w(P) < h(P)$, then we let $P' = P \setminus A_i'$ and note that $h(P') = h(P) - 1 = k+\ell-1$ as every chain of maximum length in $P$ contains one element of $A_i'$. Thus \[ s_k(P') = |P'| - k \cdot h(P') \ge k(k+\ell) - |A_i'| - k(k+\ell-1) \ge 2k/3. \] Since $w(P') \le w(P) \le h(P) - 1 = h(P')$, we may apply Corollary~\ref{cor:large-surplus} with $t = 2k/9$ to $P'$ and conclude that, recalling~\eqref{eq:mk-exp-sqrt-k}, \[ h_k(P) \ge h_k(P') \ge 2^{\sqrt{2k}/3-1} > m_k(\tau_{k,n}). \] If $w(P) = h(P)$, then we let $(A_i^*)_{i=1}^{k+\ell}$ be the canonical decomposition of $P^*$ into antichains. Now, if $|(A_j^*)'| \ge k/3$ for all $j$, then we work with $P^*$ instead of $P$. Otherwise, if $|(A_j^*)'| < k/3$ for some $j$, then we let $P' = P \setminus (A_i' \cup (A_j^*)')$ and note that $h(P') < h(P)$ and $w(P') < w(P)$. Consequently, letting $P'' = P'$ or $P'' = (P')^*$ so that $w(P'') \le h(P'')$, we have \[ s_k(P'') = |P''| - k \cdot h(P'') \ge k(k+\ell) - |A_i'| - |(A_j^*)'| - k(k+\ell-1) \ge k/3. \] Now, we may again apply Corollary~\ref{cor:large-surplus} with $t = k/9$ to $P''$ to conclude that, again recalling~\eqref{eq:mk-exp-sqrt-k}, \[ h_k(P) \ge h_k(P'') \ge 2^{\sqrt{k}/3 - 1} > m_k(\tau_{k,n}).\qedhere \] \end{proof} \subsubsection{Posets with large $F$} For the remainder of the proof, we may and shall assume that $|A_i'| \ge k/3$ for all $i \in [k+\ell]$. Together with Lemma~\ref{lemma:F}, this bound already allows us to deal with the case $|F| \ge 40 \log_2 k$. The~crucial observation here is the following. \begin{obs} \label{obs:chains-per-element} If some element of $P$ belongs to more than $2k/\ell$ chains of length $k+\ell$, then $h_k(P) > m_k(\tau_{k,n})$. \end{obs} \begin{proof} Let $M = 2k/\ell$ and suppose that some $y \in P$ is contained in $M$ different chains of length $k+\ell$. We shall show that this implies that $y$ belongs to more than $\binom{k+\ell}{k}$ chains of length $k+1$, which, by the inductive assumption, implies that $h_k(P) > m_k(\tau_{k,n})$, see~\eqref{eq:mkn-mknm}. This follows easily from Corollary~\ref{cor:signatures}~(\ref{item:signatures-local}) as, letting $m = \log_2 M + 1$, since $\ell, m \ll \sqrt{k}$, we have that \[ \exp\left(-\frac{2 (\ell-1) m}{k}\right) M \binom{k+\ell}{k+1} \ge \frac{3k}{2\ell} \binom{k+\ell}{k+1} = \frac{3k}{2(k+1)} \binom{k+\ell}{k}.\qedhere \] \end{proof} \begin{claim} \label{claim:F-size} If $|F| \ge 40\log_2 k$, then $h_k(P) > m_k(\tau_{k,n})$. \end{claim} \begin{proof} We may assume that $|B_{i+1}| < 2\ell\log_2k$ for every $i \in F$ as otherwise Lemma~\ref{lemma:F} implies that $h_k(P) \ge k^{2\ell} > m_k(\tau_{k,n})$. We have also assumed that $|A_i'| \ge k/3$ for every $k$ and hence, again by Lemma~\ref{lemma:F}, \[ \Sigma_{i+1} - \Sigma_i \ge k/4 \quad \text{ for every $i \in F \cap [k+\ell-1]$}. \] Partition $F \cap [k+\ell-1]$ into $I_1$ and $I_2$ with $|I_1| \ge 32 \log_2 k$ and $|I_2| \ge 4 \log_2 k$ such that $\min I_1 > \max I_2$. By~\eqref{eq:sum-ui-monotone}, \[ \Sigma_{\min I_1} = \Sigma_{k+\ell} + \sum_{i = \min I_1}^{k+\ell-1} (\Sigma_i - \Sigma_{i+1}) \ge |I_1| \cdot k/4 \ge 8 k \log_2 k. \] Now, consider an arbitrary $i \in I_2$. In accordance with Observation~\ref{obs:chains-per-element}, we may assume that $u_{i+1}(y) \le 2k/\ell$ for every $y \in A_{i+1}$. As $i+1 \le \min I_1$ and $|B_{i+1}| \le 2\ell\log_2k$, \[ \sum_{y \in B_{i+1}} u_{i+1}(y) \le |B_{i+1}| \cdot 2k/\ell \le 4k \log_2 k \le \Sigma_{\min I_1} / 2 \le \Sigma_{i+1}/2, \] where the final inequality follows from~\eqref{eq:sum-ui-monotone}. Hence, by Lemma~\ref{lemma:F}, \[ \Sigma_i \ge \Sigma_{i+1} + \sum_{y \in A_{i+1}' \setminus B_{i+1}} u_{i+1}(y) \ge 3\Sigma_{i+1}/2. \] It follows that \[ \Sigma_1 \ge (3/2)^{|I_2|} \cdot \Sigma_{\min I_1} \ge k^3, \] implying that there is an $x \in A_1$ which belongs to more than $2 k / \ell$ (actually more than $k^2/2$) chains of length $k+\ell$. Consequently Observation~\ref{obs:chains-per-element} implies that $h_k(P) > m_k(\tau_{k,n})$. \end{proof} \subsubsection{A sufficient condition on $\Sigma_1$} For the remainder of the proof, we shall therefore assume that $|F| \le 40\log_2k$ and, as a consequence of Observation~\ref{obs:F-size}, that $q \le 40\ell\log_2k$. In view of Corollary~\ref{cor:signatures}~(\ref{item:signatures-global}), in order to conclude that $h_k(P) > m_k(\tau_{k,n})$, it is enough to provide a sufficiently strong lower bound on $\Sigma_1$, the number of chains of length $k+\ell$ in $P$. To this end, define \begin{equation} \label{eq:S} S = \left(1 + \frac{q}{\ell}\right)k + 50\sqrt{k}\log_2k \end{equation} and note that our assumption on $q$ implies that $S \le 42k\log_2k$. \begin{obs} \label{obs:Sigma-1} If $\Sigma_1 \ge S$, then $h_k(P) > m_k(\tau_{k,n})$. \end{obs} \begin{proof} Since we have assumed that $q \le 40\ell\log_2k$, then \[ S \ge \left(1+\frac{1}{\sqrt{k}}\right) \left(1+\frac{q}{\ell}\right)(k+1). \] As $\binom{k+\ell}{k+1} = \frac{\ell}{k+1}\binom{k+\ell}{k}$, it now follows from Corollary~\ref{cor:signatures}~(\ref{item:signatures-global}) that \[ \begin{split} h_k(P) & \ge \exp\left(-\frac{3\ell\log_2k}{k}\right)\left(1+\frac{1}{\sqrt{k}}\right)\left(1+\frac{q}{\ell}\right)(k+1) \binom{k+\ell}{k+1} \\ & \ge \left(1 - \frac{1}{2\sqrt{k}}\right)\left(1 + \frac{1}{\sqrt{k}}\right) (\ell + q) \binom{k+\ell}{k} > (\ell + q) \binom{k+\ell}{k} \ge m_k(\tau_{k,n}), \end{split} \] where the second inequality holds since $\ell \le \sqrt{k} / (300\log_2 k)$ and the last inequality is~\eqref{eq:mktau-k-to-ell}. \end{proof} \subsubsection{The second lower bound on $\min_i |A_i'|$} We shall now focus on proving the following strong lower bound on $\min_i |A_i'|$. \begin{claim} \label{claim:Ap-2nd} Either $h_k(P) > m_k(\tau_{k,n})$ or $|A_i'| \ge k - 240 \ell \log_2 k$ for all $i \in [k+\ell]$. \end{claim} Recall from Section~\ref{sec:decomposition} that for each $i \in [k+\ell-1]$, the set $(A_i \setminus A_i') \cup A_{i+1}'$ is an antichain and consequently, \begin{equation} \label{eq:Ai-Aip-Aipp} |A_i| - |A_i'| + |A_{i+1}'| \le w(P) \le k+\ell. \end{equation} With foresight, define \begin{equation} \label{eq:Fp} F' = \{i \in [k+\ell-1] \colon |A_i| - |A_i'| + |A_{i+1}'| \ge k+1\}. \end{equation} Let $J$ be an arbitrary subset of $[k+\ell]$. By~\eqref{eq:Ai-Aip-Aipp} and~\eqref{eq:Fp}, \begin{equation} \label{eq:Ap-telescope} \sum_{i \in J} (|A_{i+1}'| - |A_i'|) \le \sum_{i \in J \cap F'} (k+\ell-|A_i|) + \sum_{j \in J \setminus F'} (k - |A_i|) \le \sum_{i \in J} (k-|A_i|) + |F'| \cdot \ell. \end{equation} Now, fix a $j \in [k+\ell]$ and let $J = \{j, \ldots, k+\ell-1\}$. It follows from \eqref{eq:Ap-telescope} and Observation~\ref{obs:F-size} with $I = J \cup \{k+\ell\}$ that (recalling that $A_{k+\ell}' = A_{k+\ell}$) \[ k - |A_j'| = k - |A_{k+\ell}'| + \sum_{i \in J} (|A_{i+1}'| - |A_i'|) \le \sum_{i \in J \cup \{k+\ell\}} (k - |A_i|) + |F'| \cdot \ell \le (|F| + |F'|) \cdot \ell. \] Therefore, in order to establish Claim~\ref{claim:Ap-2nd}, it suffices to prove that $|F'| > 200\log_2 k$ implies that $h_k(P) > m_k(\tau_{k,n})$. This fact is a fairly straightforward consequence of the following lemma, which one may consider as the `dual' version of Lemma~\ref{lemma:F}. \begin{lemma} \label{lemma:Fp} Either $h_k(P) > m_k(\tau_{k,n})$ or $\Sigma_i \ge \Sigma_{i+1} + k/4$ for all $i \in F'$. \end{lemma} \begin{proof} Fix some $i \in F'$ and let \[ C_i = \{x \in A_i' \colon \deg_{G_i}(x, A_{i+1}') \le 1\}. \] Observe that for every $X \subseteq C_i$, the set $(A_i \setminus A_i') \cup X \cup (A_{i+1}' \setminus N_{G_i}(X))$ is an antichain with at least $|A_i| - |A_i'| + |A_{i+1}'|$ elements. Each of these antichains that additionally satisfies $|X| \le k+1$ contains a $(k+1)$-element subset $L$ such that $L \cap A_i' = X$. Therefore, if $|C_i| \ge 2\ell\log_2k$, then $h_k(P) \ge k^{2\ell} > m_k(\tau_{k,n})$. On the other hand, if $|C_i| < 2\ell\log_2k$, then \[ \begin{split} \Sigma_i & = \sum_{xy \in G_i} u_{i+1}(y) = \sum_{y \in A_{i+1}'} u_{i+1}(y)\deg_{G_i}(y) \ge \Sigma_{i+1} + e_{G_i}(A_i', A_{i+1}') - |A_{i+1}'| \\ & \ge \Sigma_{i+1} + 2|A_i'| - |C_i| - |A_{i+1}'| \ge \Sigma_{i+1} + |A_i'| - (|A_{i+1}'| - |A_i'|) - 2\ell\log_2k. \end{split} \] By Observation~\ref{obs:F-size} with $I = \{i\}$, we have $k - |A_i| \le |F| \ell$. Therefore, by~\eqref{eq:Ai-Aip-Aipp} and our assumption on $|F|$, \[ |A_{i+1}'| - |A_i'| \le k+\ell-|A_i| \le (|F| + 1) \cdot \ell \ll k. \] Consequently, as we have assumed that $|A_i'| \ge k/3$, we have $\Sigma_i \ge \Sigma_{i+1} + k/4$. \end{proof} \begin{proof}[Proof of Claim~\ref{claim:Ap-2nd}] Observe first that Lemma~\ref{lemma:Fp} yields $\Sigma_1 \ge |F'| \cdot k/4$. Indeed, similarly as in the proof of Claim~\ref{claim:F-size}, \[ \Sigma_1 = \Sigma_{k+\ell} + \sum_{i=1}^{k+\ell-1} (\Sigma_i - \Sigma_{i+1}) \ge |F'| \cdot k/4. \] As $q \le 40\ell\log_2k$ by our assumption, if $|F'| > 200\log_2k$, then $\Sigma_1 \ge 50k\log_2k \ge S$, see~\eqref{eq:S}, and consequently $h_k(P) > m_k(\tau_{k,n})$ by Observation~\ref{obs:Sigma-1}. \end{proof} \subsubsection{Strengthening Lemmas~\ref{lemma:F} and~\ref{lemma:Fp}} For the remainder of the proof, we shall assume that $|A_i'| \ge k-240\ell\log_2k$ for every $i$. This assumption will allow us to prove the following strengthening of Lemmas~\ref{lemma:F} and~\ref{lemma:Fp}. Recall the definitions of $F$ from~\eqref{eq:F} and $F'$ from~\eqref{eq:Fp}. \begin{lemma} \label{lemma:Fp-strong} Either $h_k(P) > m_k(\tau_{k,n})$ or \begin{enumerate}[(i)] \item \label{item:Fp-strong-one} $\Sigma_i \ge \Sigma_{i+1} + k - \sqrt{k}$ for all $i \in F \cup F'$ and \item \label{item:Fp-strong-two} $\Sigma_i \ge \Sigma_{i+1} + 2k - 3\sqrt{k}$ for all $i \in F$ such that $i < \max F'$. \end{enumerate} \end{lemma} \begin{proof} Suppose first that $i \in F$. We may assume that $|B_{i+1}| \le 2\ell\log_2k$, as otherwise $h_k(P) \ge k^{2\ell} > m_k(\tau_{k,n})$ by Lemma~\ref{lemma:F}. Consequently, \[ \Sigma_i - \Sigma_{i+1} \ge |A_{i+1}'| - |B_{i+1}| \ge k - 242\ell\log_2k \ge k - \sqrt{k}, \] as $\ell \le \sqrt{k}/(300 \log_2 k)$ and we have assumed that $|A_{i+1}'| \ge k - 240\ell\log_2k$. Suppose now that $i \in F'$. As in the proof of Lemma~\ref{lemma:Fp}, let \[ C_i = \{x \in A_i' \colon \deg_{G_i}(x, A_{i+1}') \le 1\} \] and recall that either $|C_i| < 2\ell\log_2k$ or $h_k(P) \ge k^{2\ell} > m_k(\tau_{k,n})$ and that \begin{equation} \label{eq:Fp-strong} \Sigma_i \ge \Sigma_{i+1} + |A_i'| - (|F|+1) \cdot \ell - |C_i| \ge \Sigma_{i+1} +k - \sqrt{k}, \end{equation} as $\ell \le \sqrt{k}/(300 \log_2 k)$ and we have assumed that $|F| \le 40\log_2k$ and that $|A_i'| \ge k - 240\ell\log_2k$. We now turn to proving~(\ref{item:Fp-strong-two}). First, for each $i \in [k+\ell]$, define \[ A_i'' = \{x \in A_i \colon u_i(x) \ge 2\} \] and observe that $A_i'' \supseteq A_i' \setminus C_i$. Indeed, if $x \in A_i' \setminus C_i$, then by~\eqref{eq:ui-uip}, \[ u_i(x) = \sum_{xy \in G_i} u_{i+1}(y) \ge \deg_{G_i}(x,A_{i+1}') \ge 2. \] Consequently, if $i \in F'$ and $h_k(P) \le m_k(\tau_{k,n})$, then $|A_i''| \ge |A_i'| - |C_i| \ge k - \sqrt{k}$, see~\eqref{eq:Fp-strong}. Assume now that $F' \neq \emptyset$ and let $f' = \max F'$. We claim that either $h_k(P) > m_k(\tau_{k,n})$ or $|A_{i+1}''| \ge k - 2\sqrt{k}$ for each $i < f'$. To this end, observe first that for each $i \in [k+\ell-1]$, the set $(A_i \setminus A_i'') \cup A_{i+1}''$ is an antichain and therefore $|A_i| - |A_i''| + |A_{i+1}''| \le k + \ell$. Indeed, if $x \in A_i$ and $y \in A_{i+1}$ satisfy $u_i(x) < u_{i+1}(y)$, then~\eqref{eq:ui-uip} implies that $xy \not\in G_i$. With foresight, let \[ F'' = \{i \in [f' - 1] \colon |A_{i+1}''| - |A_i''| + |A_i| \ge k+1\}. \] Assume that $\min_{i \le f'} |A_i''| < k - 2\sqrt{k}$ and let $i \in [f']$ be the largest index such that $|A_i''| < k - 2\sqrt{k}$. Since, as we have shown above, $|A_{f'}''| \ge k - \sqrt{k}$, then \begin{equation} \label{eq:Fpp} \begin{split} \sqrt{k} & < |A_{f'}''| - |A_i''| = \sum_{j = i}^{f'-1} (|A_{i+1}''| - |A_i''|) \le \sum_{j = i}^{f'-1} (k-|A_i|) + |F'' \cap \{i, \ldots, f'-1\}| \cdot \ell \\ & \le |F| \cdot \ell + |F'' \cap \{i+1, \ldots, f'-1\}| \cdot \ell + \ell, \end{split} \end{equation} where the last inequality follows from Observation~\ref{obs:F-size}. Since $\ell \le \sqrt{k}/(300 \log_2 k)$ and we have assumed that $|F| \le 40\log_2k$, it follows from~\eqref{eq:Fpp} that $|F'' \cap \{i+1, \ldots, f'-1\}| \ge 50 \log_2k$. We claim that this implies that $h_k(P) > m_k(\tau_{k,n})$. To this end, consider some $j \in F''$, let \[ C_j' = \{x \in A_j'' \colon \deg(x, A_{j+1}'') \le 1\}, \] and note that for every $X \subseteq C_j'$, the set $(A_j \setminus A_j'') \cup X \cup (A_{j+1}'' \setminus N_{G_j}(X))$ is an antichain with at least $k+1$ elements. Therefore, if $|C_j''| \ge 2\ell\log_2k$, then $h_k(P) \ge k^{2\ell} > m_k(\tau_{k,n})$. On the other hand, if $|C_j| < 2\ell\log_2k$, then \[ \Sigma_j - \Sigma_{j+1} \ge e_{G_j}(A_j'' + A_{j+1}'') - |A_{j+1}''| \ge 2|A_j''| - (k+\ell) - 2\ell\log_2k. \] Therefore, if $j > i$ and $j \in F''$, then $\Sigma_j - \Sigma_{j+1} \ge k - o(k)$ and consequently, recalling~\eqref{eq:S}, \[ \Sigma_1 \ge |F'' \cap \{i+1, \ldots, f'-1\}| \cdot (k-o(k)) \ge (1-o(1)) \cdot 50k\log_2k \ge S, \] which, by Observation~\ref{obs:Sigma-1}, yields $h_k(P) > m_k(\tau_{k,n})$. Finally, suppose that $i \in F$ and $i < \max F'$. We may now assume that $|A_{i+1}''| \ge k - 2\sqrt{k}$ and therefore, by Lemma~\ref{lemma:F}, \[ \Sigma_i - \Sigma_{i+1} \ge \sum_{y \in A_{i+1}' \setminus B_{i+1}} u_{i+1}(y) \ge |A_{i+1}''| + |A_{i+1}'| - |B_{i+1}| \ge 2k - 3\sqrt{k}.\qedhere \] \end{proof} \subsubsection{Narrowing down to almost extremal posets} The following observation will further narrow down our search for $P$ with $h_k(P) \le m_k(\tau_{k,n})$. \begin{obs} \label{obs:FFp} If $|(F \cup F') \cap [k+\ell-1]| \ge \frac{q+1}{\ell}$, then $h_k(P) > m_k(\tau_{k,n})$. \end{obs} \begin{proof} Recall that $\Sigma_{k+\ell} = |A_{k+\ell}| = |A_{k+\ell}'| \ge k - 240\ell\log_2k \ge k - \sqrt{k}$. By Lemma~\ref{lemma:Fp-strong}~(\ref{item:Fp-strong-one}) and~\eqref{eq:sum-ui-monotone}, \[ \Sigma_1 \ge \Sigma_{k+\ell} + \sum_{i \in (F \cup F') \cap [k+\ell-1]} (\Sigma_i - \Sigma_{i+1}) \ge \left(1 + \frac{q+1}{\ell}\right)\left(k-\sqrt{k}\right). \] Now, since $q \le 40\ell\log_2k$ and $\ell \le \sqrt{k}/(300 \log_2 k)$, then $\Sigma_1 \ge S$ and the conclusion follows from Observation~\ref{obs:Sigma-1}. \end{proof} Recall the definition of $F$ from~\eqref{eq:F}. We shall now split into cases depending on whether or not $k + \ell \in F$, that is, whether or not $|A_{k+\ell}| \ge k+1$. \begin{case} \label{case:k-ell-not-F} $k + \ell \notin F$. \end{case} The assumption that $k+\ell \not\in F$ and Observation~\ref{obs:F-size} imply that \[ |F \cap [k+\ell-1]| = |F| \ge \left\lceil \frac{q + |\{i \in [k+\ell] \colon |A_i| < k\}|}{\ell} \right\rceil. \] By Observation~\ref{obs:FFp}, we may assume that $|A_i| \ge k$ for all $i$, $\ell$ divides $q$, and $|F| = q/\ell$, as otherwise $h_k(P) > m_k(\tau_{k,n})$. This implies that $|A_i| = k+\ell$ for each $i \in F$ and $|A_i| = k$ otherwise. \begin{subcase} $F' \neq \emptyset$ and $\max F' > \min F$. \end{subcase} Let $j = \max F'$ and let $i$ be the largest element of $F$ that is smaller than $j$. By Lemma~\ref{lemma:Fp-strong}~(\ref{item:Fp-strong-two}), \[ \Sigma_i - \Sigma_{i+1} \ge 2k - 3\sqrt{k}. \] Consequently, using Lemma~\ref{lemma:Fp-strong} as in the proof of Observation~\ref{obs:FFp}, \[ \Sigma_1 \ge \Sigma_{k+\ell} + \sum_{j \in F \setminus \{i\}} (\Sigma_j - \Sigma_{j+1}) + \Sigma_i - \Sigma_{i+1} \ge \left(1 + \frac{q}{\ell}\right)\left(k - \sqrt{k}\right) + k - 2\sqrt{k} \ge S, \] which yields $h_k(P) > m_k(\tau_{k,n})$. \begin{subcase} $F' = \emptyset$ or $\max F' \le \min F$. \end{subcase} If $F' \neq \emptyset$, then we may also assume that $\min F' \ge \min F$. Indeed, otherwise \[ |(F \cup F') \cap [k+\ell-1]| \ge |F| + 1 \ge \frac{q}{\ell}+1 \] and hence $h_k(P) > m_k(\tau_{k,n})$ by Observation~\ref{obs:FFp}. Hence $\min F' = \max F' = \min F$. We first claim that $|A_i'| \ge k$ for every $i$. Suppose not and let $i$ be the largest index for which $|A_i'| < k$ and note that $i < k+\ell$ as $A_{k+\ell}' = A_{k+\ell}$ and we have assumed that $|A_j| \ge k$ for each $j \in [k+\ell]$. Consequently, \[ |A_i| + |A_{i+1}'| - |A_i'| \ge |A_i| + k - |A_i'| > |A_i| \ge k, \] implying that $i \in F'$ and thus $i = \min F$. But this is impossible as $|A_i| = k+\ell$ and hence $(A_i \setminus A_i') \cup A_{i+1}'$ would be an antichain with more than $k+\ell$ elements. We now claim that $\max F \le \min F + 1$, that is, that $|F|=1$ or $F = \{f, f+1\}$ for some $f \in [k+\ell-1]$. To see this, note that if $\max F \in F'$, then our assumption implies that $\max F = \min F$, that is, $|F| = 1$. Otherwise, if $\max F \not\in F'$, then \[ |A_{\max F}| - |A_{\max F}'| + |A_{\max F+1}'| \le k \] and hence, as $|A_{\max F+1}'| \ge k$, we have $|A_{\max F}'| = |A_{\max F}| = k+\ell$. Consequently, either $\max F = 1$ or $\max F - 1 \in F'$ and therefore $\max F - 1 = \min F$ by our assumption. Finally, let $f = \min F$. If $f > 1$ and $|A_f'| > k$, then $f - 1 \in F'$, contradicting our assumption. We may thus assume that $f = 1$ or $|A_f'| = k$. Consequently, if $|F| > 1$, then $F = \{1, 2\}$. Indeed, if $|F| > 1$, then $F = \{f, f+1\}$ and $|A_{f+1}'| = k+\ell$. Moreover, since $A_{f+1}' \cup (A_f \setminus A_f')$ is an antichain and $w(P) \le k+\ell$, then $|A_f'| = |A_f| = k+\ell$. Finally, we have shown above that if $f > 1$, then $|A_f'| = k$. \begin{case} $k + \ell \in F$. \end{case} Consider the poset $\hat{P}$ obtained from $P$ by reversing the $\le$ relation, that is, by letting $x \le y$ in $\hat{P}$ if and only if $y \le x$ in $P$. Clearly, the same sets form chains and antichains in both $P$ and $\hat{P}$. Let $(\hat{A}_i)_{i=1}^{k+\ell}$ be the canonical decomposition of $\hat{P}$ into antichains and observe that $\hat{A}_{k+\ell} = A_1'$. Indeed, \[ \begin{split} \hat{A}_{k+\ell} & = \{x \in \hat{P} \colon \text{there is a chain $L \subseteq \hat{P}$ of length $k+\ell$ with $x = \max L$}\} \\ & = \{x \in P \colon \text{there is a chain $L \subseteq P$ of length $k+\ell$ with $x = \min L$}\} = A_1'. \end{split} \] Thus, we may assume that $|A_1'| > k$ as otherwise $\hat{P}$ falls into Case~\ref{case:k-ell-not-F}. Thus $1 \in F$. We show that this implies that $\Sigma_1 \ge S$ and consequently, by Observation~\ref{obs:Sigma-1}, that $h_k(P) > m_k(\tau_{k,n})$. Since $|A_{k+\ell}'| = |A_{k+\ell}| > k$, then $k+\ell-1 \in F'$ and hence, by Lemma~\ref{lemma:Fp-strong}, we may assume that $\Sigma_{k+\ell-1} \ge \Sigma_{k+\ell} + k - \sqrt{k} \ge 2k-\sqrt{k}$ and $\Sigma_i \ge \Sigma_{i+1} + 2k-3\sqrt{k}$ for all $i \in F \cap [k+\ell-2]$. This yields \begin{equation} \label{eq:Sigma-1-case-2} \Sigma_1 \ge \big( 2 + 2|F \cap [k+\ell-2]| + |(F' \setminus F) \cap [k+\ell-2]|\big) (k - 2\sqrt{k}). \end{equation} By our assumption that $1 \in F$ and Observation~\ref{obs:F-size}, \begin{equation} \label{eq:F-case-2} |F \cap [k+\ell-2]| \ge \max\{1, \lceil q / \ell \rceil - 2\}. \end{equation} It follows that either $q = 3\ell$ and we have equality in~\eqref{eq:F-case-2} or \[ \Sigma_1 \ge \left( 1 + \frac{q+1}{\ell} \right)(k - 2\sqrt{k}) = \left(1 + \frac{q}{\ell}\right)k + \frac{k}{\ell} - 2\sqrt{k}\left(1 + \frac{q+1}{\ell}\right) \ge S; \] to see the last inequality, recall that $\ell \le \sqrt{k}/(300 \log_2k)$. The former (i.e., $q = 3\ell$ and equality in~\eqref{eq:F-case-2}) implies that $F = \{1, k+\ell-1, k+\ell\}$, $A_i = k+\ell$ for all $i \in F$, and $|A_i| = k$ for all $i \not\in F$. Since $A_{k+\ell}' \cup (A_{k+\ell-1} \setminus A_{k+\ell-1}')$ is an antichain and $w(P) \le k+\ell = |A_{k+\ell}'|$, we must have $|A_{k+\ell-1}'| = |A_{k+\ell-1}| = k+\ell$ and consequently, $k+\ell-2 \in F'$. Now, \eqref{eq:Sigma-1-case-2} again yields $\Sigma_1 \ge 5(k-2\sqrt{k}) \ge S$. \subsection{Almost extremal posets} Summarizing the above discussion, if $h_k(P) \le m_k(\tau_{k,n})$, then $h(P) = k+\ell$ and either $P$ or $\hat{P}$ (the poset obtained from $P$ by reversing the $\le$ relation) satisfy one of the following two lists of conditions: \begin{enumerate}[(1)] \item \label{item:extremal-1} $F = \{1, 2\}$, $|A_1'| = |A_2'| = k+\ell$, and $|A_i'| = k$ for every $i \ge 3$, \item \label{item:extremal-2} $F = \{f\}$ for some $f \in [k+\ell-1]$ and $|A_i'| \ge k$ for all $i$; if $f > 1$, then $|A_f'| = k$. \end{enumerate} From now on, we shall have to count homogenous sets somewhat more carefully, as there are posets of either of these two types that contain fewer than $m_k(\tau_{k,n})$ chains of length $k+1$ and fewer than $m_k(\tau_{k,n})$ antichains with $k+1$ elements. (So far, we have always managed to show that our poset contains more than $m_k(\tau_{k,n})$ homogenous sets of one of the two types.) \subsubsection{Bounding the number of antichains} We first derive a lower bound for the number of $(k+1)$-element antichains which we shall use in both~(\ref{item:extremal-1}) and~(\ref{item:extremal-2}). To this end, for each $i \in [k+\ell-1]$, let \[ D_{i+1} = \{x \in A_{i+1}' \colon \deg_{G_i}(x) = 2\} \subseteq A_{i+1}' \setminus B_{i+1}. \] Note for future reference that it follows from (the proof of) Lemma~\ref{lemma:F} that for each $i \in F$, \begin{equation} \label{eq:lemma-F} \Sigma_i - \Sigma_{i+1} = \sum_{y \in A_{i+1}'} u_{i+1}(y)(\deg_{G_i}(y)-1) \ge 2|A_{i+1}'| - |D_{i+1}| - 2|B_{i+1}|, \end{equation} which improves the lower bound of $|A_{i+1}'| - |B_{i+1}|$ for $\Sigma_i - \Sigma_{i+1}$ stated in Lemma~\ref{lemma:F} whenever $D_{i+1}$ is smaller than $A_{i+1}' \setminus B_{i+1}$. On the other hand, if $D_{i+1}$ is large, then there are many $(k+1)$-element antichains in $A_i \cup D_{i+1}$, as the following lemma shows. \begin{lemma} \label{lemma:C-antichains} Let $i \in [k+\ell-1]$ and suppose that $|A_i| = k+\ell$ and $|A_{i+1}'| = k$. If $|D_{i+1}| \ge k-k^{2/3}$, then $A_i \cup D_{i+1}$ contains at least $20(\ell-1)\binom{k+\ell}{k+1}$ antichains with $k+1$ elements that intersect $D_{i+1}$. \end{lemma} \begin{proof} For each $z \in A_i$, let $H_z$ be an arbitrary tree with vertex set $N_{G_i}(z) \cap D_{i+1}$ and let $H$ be the multigraph with vertex set $D_{i+1}$ which is the union of all $H_z$ as $z$ ranges over $A_i$. Clearly, \[ e(H) \ge \sum_{z \in A_i} (\deg_{G_i}(z, D_{i+1}) - 1) \ge 2|D_{i+1}| - |A_i| \ge k-3k^{2/3}, \] as $|A_i| = k+\ell$ and $\ell \ll k^{2/3}$. Note crucially that: \begin{enumerate}[(i)] \item \label{item:HX} For every $X \subseteq D_{i+1}$, the set $X \cup (A_i \setminus N_{G_i}(X))$ is an antichain. \item \label{item:eHX} For every $X \subseteq D_{i+1}$, we have $|N_{G_i}(X)| \le 2|X| - e_H(X)$. \end{enumerate} To see~(\ref{item:eHX}), recall that each $H_z$ is a tree and therefore \[ |N_{G_i}(X)| = 2|X| - \sum_{z \in A_i} \max\{\deg_{G_i}(z,X)-1, 0\} \le 2|X| - \sum_{z \in A_i} e_{H_z}(X) = 2|X| - e_H(X). \] We first show that we may assume that $H$ contains fewer than $\sqrt{k}$ cycles (we consider two parallel edges to be a cycle). Since $w(P) \le k+\ell \le |A_i|$, (\ref{item:HX}) implies that $|N_{G_i}(X)| \ge |X|$ for each $X \subseteq D_{i+1}$ and hence each component $T$ of $H$ has at most one cycle as otherwise $|N_{G_i}(T)| < |T|$ by~(\ref{item:eHX}). Suppose now that $X_1, \ldots, X_m$ are cycles in $H$. They belong to different components of $H$, so in particular they are vertex-disjoint. By~(\ref{item:eHX}), for any $J \subseteq [m]$, the set $X = \bigcup_{j \in J} X_j$ satisfies $|N_{G_i}(X)| = |X|$. By~(\ref{item:HX}), $A_i \cup D_{i+1}$ contains at least $2^m$ antichains with $k+\ell$ elements and thus, by Lemma~\ref{lemma:KK}, at least $2^{m-1}$ antichains with $k+1$ elements. Finally, as $\ell < \sqrt{k}/(300 \log_2 k)$, then $2^{\sqrt{k}} \gg \ell \binom{k+\ell}{k+1}$, cf.~\eqref{eq:mk-exp-sqrt-k}. Assume that $H$ has fewer than $\sqrt{k}$ cycles and delete from $H$ one edge in each of its cycles to obtain a forest $H'$ with $e(H') \ge e(H) - \sqrt{k} \ge k - 4k^{2/3}$. Let $N'$ be the number of sets $X \subseteq D_{i+1}$ with $|X| \le 41$ which are connected in $H'$. As $H' \subseteq H$, it follows from~(\ref{item:eHX}) that $|N_{G_i}(X)| \le |X| + 1$ for each such $X$. Thus by~(\ref{item:HX}), the number $N$ of antichains in $A_i \cup D_{i+1}$ satisfies \[ \begin{split} N & \ge \sum_{X \subseteq D_{i+1}} \binom{k+\ell-|N_{G_i}(X)|}{k+1-|X|} \ge N' \binom{k+\ell-42}{k-40} \\ & \ge N' \cdot \left(\frac{k-40}{k+\ell}\right)^{41} \cdot \binom{k+\ell-1}{k+1} \ge \frac{N'}{2} \cdot\frac{\ell-1}{k} \binom{k+\ell}{k+1}, \end{split} \] where we have used the assumption that $\ell = o(k)$. Finally, let $t_1, \ldots, t_m$ be the orders of the trees constituting $H'$. Since $m = k - e(H') \le 4k^{2/3}$, it follows from Lemma~\ref{lemma:connected-sets} that \[ N' \ge \sum_{c=1}^{41} \sum_{j=1}^m (t_j-c) \ge \sum_{c=1}^{41} (k - cm) \ge 41k - O(k^{2/3}) \ge 40k.\qedhere \] \end{proof} \subsubsection{Almost extremal posets of type~(\ref{item:extremal-1})} Recall that $P$ is of type~(\ref{item:extremal-1}) if and only if $|A_i| = |A_i'| = k+\ell$ for $i \in \{1, 2\}$ and $|A_i| = |A_i'| = k$ otherwise. In particular, $q = 2\ell$ and hence $S = 3k + 50\sqrt{k}\log_2k$. We first show that $h_k(P) > m_k(\tau_{k,n})$ for each such $P$. Let $b_2 = |B_2|$ and $b_3 = |B_3|$. By Lemma~\ref{lemma:F} and \eqref{eq:lemma-F}, \[ \Sigma_1 \ge \Sigma_3 + 2|A_3'| - |D_3| - 2b_3 + |A_2'| - b_2 \ge 4k + \ell - |D_3| - b_2 - 2b_3. \] We may assume that $b_2, b_3 \le 2\ell\log_2k$ as otherwise Lemma~\ref{lemma:F} implies that $h_k(P) > m_k(\tau_{k,n})$. If $|D_3| < k - k^{2/3}$, then $\Sigma_1 \ge S$ and, by Observation~\ref{obs:Sigma-1}, $h_k(P) > m_k(\tau_{k,n})$. Thus, we may assume that $|D_3| \ge k - k^{2/3}$. Let us carefully count homogenous $(k+1)$-sets in $P$. First, consider the collection of all chains of length $k+1$ obtained by taking a triple $(x_1,x_2,x_3) \in A_1' \times A_2' \times A_3'$ with $x_1 \le x_2 \le x_3$ and an arbitrary set of $k-2$ elements from some chain of length $k+\ell$ that contains $\{x_1,x_2,x_3\}$. The number of such chains containing a fixed triple is \[ \binom{k+\ell-3}{k-2} = \frac{(k+1)k(k-1)}{(k+\ell)(k+\ell-1)(k+\ell-2)} \binom{k+\ell}{k+1} \ge \left(1 - \frac{\ell-1}{k-1}\right)^3 \binom{k+\ell}{k+1} \] and all chains constructed in this way are distinct. Let $N_t$ be the number of such triples. Since neither $G_1$ nor $G_2$ contain any isolated vertices, $A_2 = A_2'$, $A_3 = A_3'$, $e(G_1) \ge 2|A_2'| - b_2$, and $e(G_2) \ge 2|A_3'| - b_3$, then \[ N_t = \sum_{x \in A_2'} \deg_{G_1}(x) \deg_{G_2}(x) \ge \sum_{x \in A_2'} (\deg_{G_1}(x) + \deg_{G_2}(x) - 1) \ge 2|A_3'| + |A_2'| - b_2 -b_3. \] Therefore, the number $N_c$ of chains of length $k+1$ satisfies \[ N_c \ge (2|A_3'| + |A_2'| - b_2 - b_3) \binom{k+\ell-3}{k+1} \ge \left(1 - \frac{\ell-1}{k-1}\right)^3 (3k+\ell-b_2-b_3) \binom{k+\ell}{k+1}. \] To estimate the number of antichains, note that for any $i \in \{1, 2\}$ and any $Y \subseteq B_{i+1}$, the set $Y \cup (A_i \setminus N_{G_i}(Y))$ is an antichain with $k+\ell$ elements. Hence, the number $N_{a,1}$ of $(k+1)$-element antichains that are contained in either $A_1 \cup B_2$ or $A_2 \cup B_3$ satisfies \[ \begin{split} N_a & \ge \sum_{i=2}^3 \sum_{Y \subseteq B_i} \binom{k+\ell-|Y|}{k+1-|Y|} \ge \sum_{i=2}^3 2^{b_i} \binom{k+\ell-b_i}{k+1-b_i} \ge \sum_{i=2}^3 2^{b_i} \left(1 - \frac{\ell-1}{k+\ell-b_i} \right)^{b_i} \binom{k+\ell}{k+1} \\ & \ge \left[\left(\frac{7}{4}\right)^{b_2} + \left(\frac{7}{4}\right)^{b_3}\right] \binom{k+\ell}{k+1}, \end{split} \] where the last inequality follows since $\ell, b_i \ll k$. Moreover, by Lemma~\ref{lemma:C-antichains}, there are additionally at least $20(\ell-1)\binom{k+\ell}{k+1}$ antichains with $k+1$ elements that contain an element of $D_3$. Thus, using the inequality $\big(1 - \frac{\ell-1}{k-1}\big)^3 \ge 1 - 3\frac{\ell-1}{k-1}$, \begin{equation} \label{eq:example-1} \begin{split} h_k(P) \binom{k+\ell}{k+1}^{-1} & \ge \left(1 - \frac{\ell-1}{k-1}\right)^3 (3k + \ell - b_2 - b_3) + \left(\frac{7}{4}\right)^{b_2} + \left(\frac{7}{4}\right)^{b_3} + 20(\ell-1) \\ & \ge 3k+\ell-b_2-b_3 + 10(\ell-1) + \left(\frac{7}{4}\right)^{b_2} + \left(\frac{7}{4}\right)^{b_3} \\ & \ge 3k + 1 + \left(\frac{7}{4}\right)^{b_2} - b_2 + \left(\frac{7}{4}\right)^{b_3} - b_3. \end{split} \end{equation} Finally, using the fact that $(7/4)^b - b \ge 3/4$ for every integer $b$, we see that the right hand side of~\eqref{eq:example-1} is strictly greater than $3k+2$. This completes the analysis as $m_k(\tau_{k,n}) = (3k+2)\binom{k+\ell}{k+1}$. \subsubsection{Almost extremal posets of type~(\ref{item:extremal-2})} Recall that $P$ is of type~(\ref{item:extremal-2}) if and only if there is some $f \in [k+\ell-1]$ such that $|A_i| = |A_i'| = k$ for all $i \neq f$, $|A_f| = k+\ell$, and $|A_f'| \ge k$. Moreover, $|A_f'| = k$ if $f \neq 1$. In particular, $q = \ell$ and hence $S = 2k + 50\sqrt{k}\log_2k$. We now show that $h_k(P) > m_k(\tau_{k,n})$ for each such $P$, unless $\ell = 1$. Let $b = |B_{f+1}|$. By~\eqref{eq:lemma-F}, \[ \Sigma_f \ge \Sigma_{f+1} + 2|A_{f+1}'| - |D_{f+1}| - 2|B_{f+1}| \ge 3k-|D_{f+1}|-2b. \] As before, by Lemma~\ref{lemma:F}, we may assume that $b \le 2\ell\log_2k$ since otherwise $h_k(P) > m_k(\tau_{k,n})$. If $|D_{f+1}| < k - k^{2/3}$, then $\Sigma_1 \ge \Sigma_f \ge S$, and, by Observation~\ref{obs:Sigma-1}, $h_k(P) > m_k(\tau_{k,n})$. Thus, we may assume that $|D_{f+1}| \ge k-k^{2/3}$. Let us now carefully count homogenous $(k+1)$-sets in $P$. First, consider the collection of all chains of length $k+1$ obtained by taking an edge $xy$ of $G_f$ and an arbitrary set of $k-1$ elements from some chain of length $k+\ell$ that contains both $x$ and $y$ (such a chain exists as $A_f' = A_f$. The number of such chains containing a fixed edge is \[ \binom{k+\ell-2}{k-1} = \frac{(k+1)k}{(k+\ell)(k+\ell-1)} \binom{k+\ell}{k+1} \ge \left(1 - \frac{\ell-1}{k}\right)^2 \binom{k+\ell}{k+1} \] and all chains constructed this way are distinct. Therefore, the number $N_c$ of chains of length $k+1$ satisfies \[ N_c \ge e(G_f) \cdot \binom{k+\ell-2}{k-1} \ge \left(1-\frac{\ell-1}{k}\right)^2(2k-b)\binom{k+\ell}{k+1}. \] To estimate the number of antichains, note that for any $Y \subseteq B_{f+1}$, the set $Y \cup (A_f \setminus N_{G_f}(Y))$ is an antichain with $k+\ell$ elements. Hence, the number $N_{a,1}$ of $(k+1)$-element antichains that are contained in $A_f \cup B_{f+1}$ satisfies \[ N_{a,1} \ge 2^b \binom{k+\ell-b}{k+1-b} \ge 2^b \left(1 - \frac{\ell-1}{k+\ell-b} \right)^b \binom{k+\ell}{k+1} \ge \left(\frac{7}{4}\right)^b \binom{k+\ell}{k+1}, \] where the last inequality holds since $\ell, b \ll k$. Moreover, by Lemma~\ref{lemma:C-antichains}, there are additionally at least $20(\ell-1)\binom{k+\ell}{k+1}$ antichains with $k+1$ elements that contain an element of $D_{f+1}$. It follows that, using the inequality $\big(1 - \frac{\ell-1}{k}\big)^2 \ge 1 - 2\frac{\ell-1}{k}$, \begin{equation} \label{eq:example-2} h_k(P) \binom{k+\ell}{k+1}^{-1} \ge 2k-b - 4(\ell-1) + \left(\frac{7}{4}\right)^b + 20(\ell-1). \end{equation} As $m_k(\tau_{k,n}) = (2k+1)\binom{k+\ell}{k+1}$, we conclude that $h_k(P) > m_k(\tau_{k,n})$ unless $\ell=1$ and $b \le 1$. \subsubsection{Almost extremal posets of type~(\ref{item:extremal-2}) when $\ell = 1$} We finally show that $h_k(P) \ge m_k(\tau_{k,n})$ for each poset $P$ of type~(\ref{item:extremal-2}) and provide a rough structural characterization of such posets which contain precisely $m_k(\tau_{k,n})$ homogeneous $(k+1)$-sets. Since we may assume that $\ell = 1$, then $m_k(\tau_{k,n}) = 2k+1$ and $\Sigma_1$ is the number of chains of length $k+1$ in $P$. Let $b = |B_{f+1}|$ and recall that $b \le 1$. By~\eqref{eq:lemma-F}, \[ \Sigma_1 \ge \Sigma_{f+1} + 2k - |D_{f+1}| - 2b \ge 2k - b. \] Moreover, $A_f \cup B_{f+1}$ contains at least $2^b$ antichains with $k+1$ elements. It follows that $h_k(P) \ge 2k - b + 2^b \ge m_k(\tau_{k,n})$. Moreover, either $h_k(P) > m_k(\tau_{k,n})$ or $\Sigma_{f+1} = k$, $D_{f+1} = A_{f+1} \setminus B_{f+1}$ (i.e., each $x \in A_{f+1} \setminus B_{f+1}$ has degree two in $G_f$), and there are only $2^b$ antichains with $k+1$ elements (plus, they are both contained in $A_f \cup B_{f+1}$). This means that for every $X \subseteq A_{f+1}$ such that $X \neq \emptyset$ and $X \neq B_{f+1}$, we have $|N_{G_f}(X)| \ge |X|+1$, as otherwise $X \cup (A_f \setminus N_{G_f}(X))$ would be an additional antichain with $k+1$ elements (other than $A_f$ and $B_{f+1} \cup (A_f \setminus N_{G_f}(B_{f+1}))$, which were already counted above in the $2^b$ term). In particular, $f=1$ and $A_1' = A_1$, as otherwise $|A_f'|=|A_{f+1}'|=k$ and consequently $|N_{G_f}(A_{f+1}')| = |A_{f+1}'|$. We now show that these conditions uniquely determine the graph $G_1$. \begin{claim} Suppose that $h_k(P) = m_k(\tau_{k,n})$. \begin{enumerate}[(i)] \item \label{item:b0} If $b = 0$, then $G_1$ is a path with $2k+1$ vertices. \item \label{item:b1} If $b = 1$, then $G_1$ is the disjoint union of a path with $2k-1$ vertices and en edge. \end{enumerate} \end{claim} \begin{proof} Let $H$ be the auxiliary multigraph on $A_1$ where the multiplicity of each pair $xy$ is the number of common $G_1$-neighbors of $x$ and $y$ in $A_2$. As we have assumed that $D_2 = A_2 \setminus B_2$, it follows that $e(H) = |A_2 \setminus B_2| = k - b$. Moreover, the condition $|N_{G_1}(X)| \ge |X|+1$ implies that $H$ is a forest. (Here again we consider two parallel edges to be a cycle.) We claim that each tree in $H$ is a path. If this is not the case, then $H$ would have a vertex of degree at least $3$ and thus $G_1$ would contain the $1$-subdivision of $K_{1,3}$ as an induced subgraph. But this is not possible as both $G_1$ and its complement are comparability graphs (since $P$ is a poset of order dimension at most two) and one can check that the $1$-subdivision of $K_{1,3}$ does not have this property. Now, (\ref{item:b0}) follows since $b=0$ implies that $H$ is a path with $k+1$ vertices and hence $G_1$ is a path with $2k+1$ vertices. To see~(\ref{item:b1}), note first that $b=1$ implies that $H$ has two connected components. The unique vertex $z \in B_2$ has a $G_1$-neighbor in one of these components. Denote it by $C$ and let $X = N_{G_1}(C)$. One easily checks that $|X| = |N_{G_1}(X)| = |C|$ and hence $X = \{z\}$ and $|C|=1$, as we have assumed that $X = \emptyset$ and $X = B_2$ are the only subsets of $A_2$ with $|X| \le |N_{G_1}(X)|$. It follows that $H$ is the union of a path with $k$ vertices and an isolated vertex and hence $G_1$ is the union of a path with $2k-1$ vertices and an edge. \end{proof} Finally, the following lemma (where we take $i = 2$ and $j = k+1$) shows that $P \setminus A_1$ may be partitioned into $k$ chains, as otherwise $\Sigma_2 > k$. \begin{lemma} \label{lemma:disjoint-chains} Suppose that $i, j \in [k+\ell]$ with $i \le j$ are such that $|A_i'| = \ldots = |A_j'| = k$. There is some $d \ge 0$ such that: \begin{enumerate}[(i)] \item \label{item:disjoint-chains-one} There are pairwise disjoint chains $L_1, \ldots, L_{k-d}$ of length $j - i + 1$ with $\min L_p \in A_i'$ and $\max L_p \in A_j'$ for each $p \in [k-d]$ and \item $\Sigma_i \ge \Sigma_j + d$. \end{enumerate} \end{lemma} \begin{proof} We prove the claim by reverse induction on $i$. The statement is vacuously true for $i = j$ as $|A_j'| = k$. Suppose now that $i < j$ and, appealing to the inductive assumption, let $L_1', \ldots, L_{k-d'}'$ be a collection of pairwise disjoint chains of length $j-i$ with $\min L_p \in A_{i+1}'$ and $\max L_p \in A_j'$ for each $p \in [k-d']$. Let \[ X = \{\min L_p \colon p \in [k-d']\} \subseteq A_{i+1}' \] and let $k-d$ be the size of the largest $G_i$-matching between $X$ and $A_i'$. Since this matching naturally extends some $k-d$ chains in $\{L_1', \ldots, L_{k-d'}'\}$ to pairwise disjoint chains $L_1, \ldots, L_{k-d}$ satisfying assertion~(\ref{item:disjoint-chains-one}) of the lemma, it is enough to show that $\Sigma_i \ge \Sigma_{i+1} + d-d'$. By Hall's theorem, there is a $Y \subseteq X$ such that $|N_{G_i}(Y)| \le |Y| - (d-d')$. Since $G_i[A_i',A_{i+1}']$ has no isolated vertices, it follows that \[ e_{G_i}(A_i', A_{i+1}') \ge e_{G_i}(A_i',Y) + e_{G_i}(A_i' \setminus N_{G_i}(Y), A_{i+1}') \ge |Y| + |A_i'| - |N_{G_i}(Y)| \ge k + d-d'. \] Consequently, \[ \Sigma_i = \sum_{x \in A_{i+1}'} u_{i+1}(x) \ge \Sigma_{i+1} + e_{G_i}(A_i', A_{i+1}') - |A_{i+1}'| \ge \Sigma_{i+1} + d-d'.\qedhere \] \end{proof} This completes the proof of Theorem~\ref{thm:posets}. \section{Concluding remarks} \label{sec:concluding-remarks} In this paper, we have determined the minimum number of monotone subsequences of length $k+1$ in a sequence of $n \le k^2 + k^{3/2} / (300\log_2k)$ numbers for all sufficiently large $k$. This minimum, which we denoted by $m_k(n)$, is achieved by taking $k$ increasing (decreasing) sequences of lengths $\lfloor n/k \rfloor$ or $\lceil n/k \rceil$ in such a way that there is no decreasing (increasing) subsequence of length $k+1$. One such sequence is $\tau_{k,n}$, defined in Section~\ref{sec:introduction}. Moreover, we have shown that if $n \neq k^2+k+1$, then no extremal sequence contains both increasing and decreasing subsequences of length $k+1$. Our results provide strong evidence supporting Conjecture~\ref{conj:main}, which asserts that the above statements remain true for all pairs of $k$ and $n$. It is also worth mentioning that, although we have not stated it explicitly, our proof establishes a stability statement of the following form: If a sequence of $n$ numbers is not `close' to a union of $k$ increasing (decreasing) sequences of almost equal lengths, then it contains `many' more than $m_k(n)$ monotone subsequences of length $k+1$. Since we are still far from proving Conjecture~\ref{conj:main}, one may ask to determine at least the asymptotic behavior of the function $m_k(n)$. Let $\mu_k(n) = m_k(n) \binom{n}{k+1}^{-1}$. Then Conjecture~\ref{conj:main} suggests that $\mu_k(n) = (1+o(1))k^{-k}$. Standard averaging arguments can be used to show that $\mu_k(n)$ is non-decreasing in $n$. Therefore, the theorem of Erd{\H{o}}s and Szekeres implies that \begin{equation} \label{eq:mukn-lower} \mu_k(n) \ge \mu_k(k^2+1) = \binom{k^2+1}{k+1}^{-1} \sim \sqrt{\frac{2\pi e}{k}} \cdot (ek)^{-k}. \end{equation} Our main theorem yields an improvement of~\eqref{eq:mukn-lower} by a factor of (only) $2^{\Theta(\sqrt{k})}$ and it would be interesting to improve this lower bound further. We find very promising the prospect of studying Erd{\H{o}}s--Rademacher-type problems in other settings. In principle, one can investigate such extensions for any extremal or Ramsey-type result. Some motivation for studying these problems comes from the recently renewed interest in `supersaturation' results, which have been used in conjunction with the `transference' theorems of Conlon and Gowers~\cite{CoGo} and of Schacht~\cite{Sc} as well as the `hypergraph containers' theorems of Balogh, Morris, and the first author~\cite{BaMoSa} and of Saxton and Thomason~\cite{SaTh} to prove numerous `sparse random analogues' of classical extremal and Ramsey-type results. \bibliographystyle{amsplain}

https://arxiv.org/abs/0907.1019

The algebraic crossing number and the braid index of knots and links

It has been conjectured that the algebraic crossing number of a link is uniquely determined in minimal braid representation. This conjecture is true for many classes of knots and links.The Morton-Franks-Williams inequality gives a lower bound for braid index. And sharpness of the inequality on a knot type implies the truth of the conjecture for the knot type.We prove that there are infinitely many examples of knots and links for which the inequality is not sharp but the conjecture is still true. We also show that if the conjecture is true for K and L, then it is also true for the (p,q)-cable of K and for the connect sum of K and L.

\section{Introduction} The {\em braid index} is one of the classical invariants of knots and links. Any knot and link type is presented as a braid closure. The braid index of a link type is the least number of braid strands needed for that. The {\em algebraic crossing number} (or writhe) is an integer associated to an oriented link diagram counting the crossings with weight $+1$ (resp.\ $-1$) for a positive (resp.\ negative) crossing as shown in the left (resp.\ middle) sketch of \fullref{+-0}. Since it is changed under Reidemeister move I, it is {\em not} an invariant of link types. However, it has been asked (see Jones \cite[page 357]{Jones-1} for example): \medskip \noindent\textbf{Question}\qua {\sl Is the algebraic crossing number in a minimal braid representation a link invariant?} \medskip Here ``minimal'' means that the number of braid strands of a link diagram is equal to the braid index of the link type. It is known that the following links have unique algebraic crossing numbers at minimal braid index: torus links, closed positive braids with a full twist, including the Lorenz links (Franks and Williams \cite{FW}), $2$--bridge links and alternating fibered links (Murasugi \cite{Murasugi}) and links with braid index $\leq 3$ (Birman and Menasco \cite{BM3}). In \fullref{chap3} of this paper we approach the above question in three ways. The first way (\fullref{deficit-cable} and its corollaries) is by studying the deficit of the Morton--Franks--Williams (MFW) inequality (Morton \cite{Morton}, Franks and Williams \cite{FW}). It is easy to see that sharpness of the MFW--inequality implies the uniqueness of the algebraic crossing number at minimal index. Then how do we answer the question for links on which the inequality is not sharp? In fact we provide infinitely many examples of non-sharp links having unique algebraic crossing numbers at minimal braid index. The second way is by studying the behavior of the braid index and the algebraic crossing number under the cabling operation. In \fullref{cable-thm} and \fullref{cable-thm-link}, we will prove that the uniqueness property is preserved under cabling. Then we have \fullref{iterated-torus} saying ``yes'' to the question for iterated torus knots. The third way is by studying the connect sum operation. In \fullref{sum}, we will show that the uniqueness property is preserved under taking the connect sum. In \fullref{chap2}, we focus on non-sharpness of the MFW--inequality. To state the MFW--inequality, let ${\cal K}$ be an oriented knot type and let $K$ be a diagram of ${\cal K}$ on a plane. Focus on one crossing of $K$ with sign $\varepsilon$. Denote $K_\varepsilon := K$ and let $K_{-\varepsilon}$ (resp.\ $K_0$) be the closed braid obtained from $K_\varepsilon$ by changing the the crossing to the opposite sign $-\varepsilon$ (resp.\ resolving the crossing), see \fullref{+-0}. % \begin{figure}[ht!]\small \begin{center} \labellist \pinlabel {$K_+$} at 51 15 \pinlabel {$K_-$} at 197 15 \pinlabel {$K_0$} at 340 15 \endlabellist \includegraphics[height=20mm]{\figdir/_-0marta} \caption{Local views of $K_+, K_-, K_0$}\label{+-0} \end{center} \end{figure} The \textit{HOMFLYPT polynomial} $P_{\cal K}(v,z)=P_{K}(v,z)$ satisfies the following relations (for any choice of a crossing): \begin{eqnarray} \frac{1}{v} P_{K_{+}} - v P_{K_{-}} &=& z P_{K_{0}}. \label{skein1} \\ P_{\mathrm{unknot}} &=&1. \notag \end{eqnarray} Now we are ready to state the MFW--inequality. \begin{theorem}[The Morton--Franks--Williams inequality \cite{Morton,FW}] \label{MFW} Let $d_{+}$ and $d_{-}$ be the maximal and minimal degrees of the variable $v$ of $P_{\cal K}(v,z)$. If a knot type ${\cal K}$ has a closed braid representative $K$ with braid index $b_K$ and algebraic crossing number $c_K$, then we have \begin{equation} c_K - b_K +1 \leq d_{-}\leq d_{+}\leq c_K + b_K -1. \label{lower bound-1}% \end{equation}% As a corollary, \begin{equation} \frac{1}{2}(d_{+}-d_{-})+1 \leq b_K, \label{lower bound} \end{equation}% giving a lower bound for the {\em braid index} $b_{\cal K}$ of ${\cal K}$. \end{theorem} In general, it is hard to determine the braid index. This inequality was the first known result of a general nature relating to the computation of braid index, and it appeared to be quite effective. Jones notes, in \cite{Jones-1}, that on all but five knots, $9_{42}, 9_{49}, 10_{132}, 10_{150}, 10_{156}$ in the standard knot table, up to crossing number $10$, the MFW inequality is sharp. Furthermore it has been known that the inequality is sharp on all torus links, closed positive $n$--braids with a full twist \cite{FW}, $2$--bridge links and fibered alternating links \cite{Murasugi}. However, the MFW--inequality is not as strong as it appears to be as above. In \fullref{deficit-thm} we give an infinite class of prime links in which the deficit $D_{\cal K}:= b_{\cal K}-\frac{1}{2}(d_+-d_-)-1$ of the MFW--inequality \eqref{lower bound} can be arbitrarily large. And in \fullref{BM-thm} we see another infinite class of knots, including $9_{42}, 9_{49}, 10_{132}, 10_{150}, 10_{156},$ on which the inequality is not sharp. Then we may ask ``why does non-sharpness occur?'' \fullref{thmA} gives a sufficient condition for non-sharpness of the MFW inequality. In fact all the examples in Theorems \ref{deficit-thm} and \ref{BM-thm} satisfy this sufficient condition. The idea of \fullref{thmA} is to find knots $K_\alpha$ of known braid index $=b$ which have a distinguished crossing such that, after changing that crossing to each of the other two possibilities in \fullref{+-0}, giving knots or links $K_\beta$ and $K_\gamma$, it is revealed that $K_\beta$ and $K_\gamma$ each has braid index $< b.$ Thanks to \fullref{thmA} one can visually observe the ``accumulation'' of deficits (for example under the connect sum operation and other linking operation) by looking only at the distinguished crossings which contribute to deficits. See the proof of \fullref{deficit-thm} for details. \medskip \textbf{Acknowledgment}\qua This paper is part of the author's PhD thesis. She is grateful to her advisor, Joan Birman, for her thoughtful advice and encouragement. She also wishes to thank William Menasco, who told her about the Birman--Menasco diagram and the associated conjecture, when she visited SUNY Buffalo. She appreciates many helpful comments by Walter Neumann, Dylan Thurston, Ilya Kofman and the referee and thanks Alexander Stoimenow for sending a preprint She acknowledges partial support from NSF grants DMS-0405586 and DMS-0306062. Finally, she especially thanks Mikami Hirasawa, who shared many creative ideas and results about fibered knots including the definition and properties of the enhanced Milnor number. \section{Non-sharpness of the Morton--Franks--Williams inequality}\label{chap2} \subsection{Sufficient conditions for non-sharpness} We define the deficit of MFW--inequality (\fullref{deficit-def}) then give sufficient conditions (\fullref{thmA}) for a closed braid on which the inequality is not sharp. Let $b_{\cal K}$ be the braid index of knot type ${\cal K}$, that is the smallest integer $b_{\cal K}$ such that ${\cal K}$ can be represented by a closed $b_{\cal K}$--braid. Let $b_K, c_K$ denote the braid index and the algebraic crossing number of a braid representative $K$ of ${\cal K}$. \begin{definition}\label{deficit-def} Let \begin{equation*} D_{\cal K} : = b_{\cal K} - \frac{1}{2}(d_{+}-d_{-})-1 \end{equation*}% be the difference of the numbers in $\eqref{lower bound},$ ie, of the actual braid index and the lower bound for braid index. Call $D_{\cal K}$ the {\em deficit} of the MFW--inequality for ${\cal K}$. \end{definition} If $D_{\cal K} =0,$ the MFW--inequality is sharp on ${\cal K}$. If $K$ is a braid representative of ${\cal K}$ let $D_K^+ \ := \ (c_K + b_{K} -1)-d_{+}$ and $D_K^- \ := \ d_{-}-(c_K - b_{K} +1).$ When $b_K = b_{\cal K},$ we have \begin{equation}\label{D+D} D_{\cal K} =\frac{1}{2}(D_K^{+} + D_K^{-}). \end{equation}% Note that $D_K^{\pm}$ depends on the choice of braid representative $K$, but the deficit $D_{\cal K}$ is independent from the choice. \begin{theorem}\label{thmA} Assume that $K$ is a closed braid representative of ${\cal K}$ with $b_K = b_{\cal K}$. Focus on one site of $K$ and construct $K_+, K_-, K_0$ (one of the three must be $K$). Let $\alpha, \beta, \gamma \in \{ +, -, 0 \}$ and assume that $\alpha, \beta, \gamma$ are mutually distinct. If $K_\alpha = K$ and positive destabilization is applicable $p$--times to each of $ K_\beta$ and $K_\gamma$, then \begin{equation} D_K^+ \geq 2p; \label{less} \end{equation} and if $K_\alpha = K$ and negative destabilization is applicable $n$--times to each of $ K_\beta$ and $K_\gamma$, then \begin{equation} D_K^- \geq 2n. \label{more} \end{equation} Therefore, by {\em \eqref{D+D}}, the MFW--inequality is not sharp on ${\cal K}$ if $p+n>0$. \end{theorem} Here is a lemma to prove \fullref{thmA}. \begin{lemma}\label{lemma-for-thmA} Let $K$ be a closed braid. Choose one crossing, and construct $K_+, K_-, K_0$ (one of the three must be $K$). We have \begin{eqnarray} d_{+}(P_{K_{+}}) &\leq & \max \{d_{+}(P_{K_{-}})+2, \quad d_{+}(P_{K_{0}})+1\} \label{dplus+} \\ d_{+}(P_{K_{-}}) &\leq & \max \{d_{+}(P_{K_{+}})-2, \quad d_{+}(P_{K_{0}})-1\} \label{dplus-} \\ d_{+}(P_{K_{0}}) &\leq & \max \{d_{+}(P_{K_{+}})-1, \quad d_{+}(P_{K_{-}})+1\} \label{dplus0} \end{eqnarray}% and% \begin{eqnarray*} d_{-}(P_{K_{+}}) &\geq & \min \{d_{-}(P_{K_{-}})+2, \quad d_{-}(P_{K_{0}})+1\} \\ d_{-}(P_{K_{-}}) &\geq & \min \{d_{-}(P_{K_{+}})-2, \quad d_{-}(P_{K_{0}})-1\} \\ d_{-}(P_{K_{0}}) &\geq & \min \{d_{-}(P_{K_{+}})-1, \quad d_{-}(P_{K_{-}})+1\}. \end{eqnarray*} \end{lemma} \begin{proof}[Proof of \fullref{lemma-for-thmA}] By \eqref{skein1}, we have $P_{K_{+}}=v^{2}P_{K_{-}} + vz P_{K_{0}}.$ Thus, $d_+(P_{K_+})=d_+(v^2 P_{K_-} + vz P_{K_0}) \leq \max \{ d_+ ( v^2 P_{K_-}), \ d_+ ( vz P_{K_0})\}$ and we obtain \eqref{dplus+}. The other results follow similarly. \end{proof} Table \eqref{table1} shows the changes of $c_K$, $b_K$, $c_K - b_K + 1$ and $c_K + b_K - 1$ under stabilization and destabilization of a closed braid. \begin{equation} \begin{tabular}{|l|l|l|l|l|} \hline & $c_K$ & $b_K$ & $c_K-b_K+1$ & $c_K+b_K-1$ \\ \hline $+$ stabilization & $+1$ & $+1$ & \multicolumn{1}{|c|}{$0$} & \multicolumn{1}{|c|}{$+2$} \\ \hline $+$ destabilization & $-1$ & $-1$ & \multicolumn{1}{|c|}{$0$} & \multicolumn{1}{|c|}{$-2$} \\ \hline $-$ stabilization & $-1$ & $+1$ & \multicolumn{1}{|c|}{$-2$} & \multicolumn{1}{|c|}{$0$} \\ \hline $-$ destabilization & $+1$ & $-1$ & \multicolumn{1}{|c|}{$+2$} & \multicolumn{1}{|c|}{$0$} \\ \hline \end{tabular} \label{table1} \end{equation} Note that $c_K$ and $b_K$ are invariant under braid isotopy and exchange moves. \begin{proof}[ Proof of \fullref{thmA}] Suppose that $K = K_\alpha = K_+.$ Suppose we can apply positive destabilization $k$--times ($k\geq p$) to $K_{-}$. Let $\tilde{K}_{-}$ denote the closed braid obtained after the destabilization. Then we have: \begin{eqnarray} d_{+}(P_{K_-}) + 2 &=& d_{+}(P_{\tilde{K}_-}) + 2 \notag \\ &\leq &( c_{\tilde{K}_-} + b_{\tilde{K}_-} -1 ) + 2 \notag \\ &=& \{(c_{K_-} + b_{K_-} -1) -2k \} + 2 \label{5eq} \\ &=& (c_{K_+} -2) + b_{K_+} -1 -2k + 2 \notag \\ &=& (c_{K_+} + b_{K_+} -1) -2k = (c_K + b_K -1) -2k. \notag \end{eqnarray} The first equality holds since $K_{-}$ and $\tilde{K}_-$ have the same knot type. The first inequality is the MFW--inequality. The second equality follows from Table \eqref{table1}. Similarly, if we can apply positive destabilization $l$--times $(l \geq p)$ to $K_{0},$ and obtain $\tilde{K_0}$, we have \begin{eqnarray} d_{+}(P_{K_{0}})+1 &=& d_+ (P_{\tilde{K_0}}) + 1 \notag \\ &\leq& (c_{\tilde{K_0}} + b_{\tilde{K_0}} -1 ) + 1 \notag \\ &=& (c_{K_0} + b_{K_0} -1 - 2l ) + 1 \label{5eq'} \\ &=& (c_{K_+} -1 ) + b_{K_+} -1 - 2l + 1 \notag \\ &=& (c_{K_+} + b_{K_+} -1) -2l = (c_K + b_K -1) -2l. \notag \end{eqnarray} By \eqref{dplus+}, \eqref{5eq} and \eqref{5eq'} we get \begin{eqnarray*} d_+ (P_K) &=& d_+ (P_{K_+}) \leq \max \{d_{+}(P_{K_{-}})+2, \quad d_{+}(P_{K_{0}})+1\} \\ &\leq& (c_K + b_{\cal K} -1) - \min \{ 2k, 2l \}, \end{eqnarray*} ie, $D_K^+ \geq \min \{ 2k, 2l \} \geq 2p.$ When $K_\alpha = K_-$ or $K_\alpha = K_0$, the same arguments work (use \eqref{dplus-} or \eqref{dplus0} for these cases in the place of \eqref{dplus+}) and we get \eqref{less}. The other inequality \eqref{more} also holds by the identical argument. \end{proof} \subsection{Deficit growth} Our goal is to exhibit examples (\fullref{deficit-thm}) of prime links on which the deficit of the inequality can be arbitrary large. \begin{theorem}\label{th942} Knot type ${\cal K}=9_{42}$ has a braid representative $K=K_+$ (see \fullref{skein942}) satisfying the sufficient condition in \fullref{thmA}. \end{theorem} \begin{figure}[ht!]\small \labellist \pinlabel {$K_+ =$} at 29 206 \pinlabel {$K_- =$} at 29 120 \pinlabel {$K_0 =$} at 29 35 \pinlabel {negative} <-5pt,15pt> [lB] at 287 95 \pinlabel {destabilization} <-5pt,5pt> [lB] at 287 95 \pinlabel {positive} <-5pt,15pt> [lB] at 548 95 \pinlabel {destabilization} <-5pt,5pt> [lB] at 548 95 \pinlabel {positive} <-5pt,15pt> [lB] at 287 4 \pinlabel {destabilization} <-5pt,5pt> [lB] at 287 4 \endlabellist \begin{center}\includegraphics [height=40mm]{\figdir/skein942} \end{center} \caption{Knot $9_{42}$ satisfies the conditions of \fullref{thmA}}\label{skein942} \end{figure} \begin{proof}[Proof of \fullref{th942}] It is known that $9_{42}$ has braid index $=4$ and deficit $D_{9_{42}}=1.$ Let $K = K_+$ be its braid representative of the minimal braid index as in \fullref{skein942}. Construct $K_-, K_0$ by changing the shaded crossing. Sketches show that both $K_-, K_0$ can be positively destabilized. Thus by \fullref{thmA}, $D_{K}^+ \geq 2$ and $D_{9_{42}}\geq 1.$ \end{proof} \begin{theorem}\label{deficit-thm} For any positive integer $n,$ there exists a prime link $L$ whose deficit $D_L\geq n.$% \end{theorem} \begin{proof}[ Proof of \fullref{deficit-thm}] We prove the theorem by exhibiting examples. For $n \in \mathbb{N}$ let ${\cal A}^n(9_{42})$ be the closure of $n$--copies of $9_{42}$ linked each other by two full twists as in the left sketch of \fullref{942}. Since the braid index $b_{9_{42}} = 4$ and ${\cal A}^n(9_{42})$ is an $n$--component link, we know the braid index of ${\cal A}^n(9_{42})$ is $4n.$ This construction gives a braid representative with $4n$--strands and $n$ distinguished crossings shaded in the left sketch. \begin{figure}[ht!] \begin{center} \includegraphics [height=55mm]{\figdir/942} \end{center} \caption{Prime link ${\cal A}^5(9_{42})$ and $2$--component link ${\cal A}$} \label{942}% \end{figure} In the following we will see that each of the shaded crossing contributes to the deficit. \newpage Let ${\cal K}:={\cal A}^2(9_{42})$ and let $K$ be the braid representative of ${\cal K}$ as in \fullref{942}. Let $K_{- -}, K_{- 0}, K_{0 -}, K_{0 0}$ be the links obtained from $K$ by changing the two shaded crossings. We repeat the discussion of the proof of \fullref{thmA}: We have: \begin{eqnarray*} d_+(P_{K_{- -}}) + (2+2) &=& d_{+}(P_{\tilde{K}_{- -}}) + 4 \\ % &\leq &( c_{\tilde{K}_{- -}} + b_{\tilde{K}_{- -}} -1 ) + 4 \\ &=& \{(c_{K_{- -}} + b_{K_{- -}} -1) -2\cdot 2 \} + 4 \\ &=& (c_{K} -4) + b_{K} -1 -2\cdot 2 + 4 \\ &=& (c_{K} + b_{K} -1) -2\cdot 2 \end{eqnarray*} Similarly: \begin{eqnarray*} d_+(P_{K_{- 0}})+(2+1) &\leq & (c_{K} + b_{K} -1) -2\cdot 2 \\ d_+(P_{K_{0 -}})+(1+2) &\leq & (c_{K} + b_{K} -1) -2\cdot 2 \\ d_+(P_{K_{0 0}})+(1+1) &\leq & (c_{K} + b_{K} -1) -2\cdot 2 \end{eqnarray*} Thus \begin{eqnarray*} d_+ (P_K) &=& \max \{d_+(P_{K_{- -}})+4, \ d_+(P_{K_{- 0}})+3, \ % d_+(P_{K_{0 -}})+3, \ d_+(P_{K_{0 0}})+2 \} \\ % &\leq & (c_K + b_{\cal K} -1) - 2\cdot 2 \end{eqnarray*} and\qquad\quad $D_{\cal K} \geq \frac{1}{2} D_K^+ \geq \frac{1}{2}(2\cdot 2) = 2.$ Similar arguments work when ${\cal K}={\cal A}^n(9_{42})$ for $n\geq 3$ and we have $D_{{\cal A}^n(9_{42})} \geq \frac{1}{2} D_{{\cal A}^n(9_{42})}^+ \geq \frac{1}{2} (2\cdot n) \geq n.$ The $2$--component link ${\cal A}$ of the right sketch is hyperbolic \cite{N}. Pair $(S^3, {\cal A}^n(9_{42})\cup z\mbox{--axis})$ is an $n$--fold cover of $(S^3, {\cal A})$ branched at $z$--axis . Therefore, by Neumann and Zagier \cite{NZ} we can conclude that ${\cal A}^n(9_{42})$'s are all prime except for finitely many $n$'s. \end{proof} \begin{remark} {\rm By taking the connected sum of knots on which the MFW inequality is non-sharp, one can also construct examples of (non-prime) knots with arbitrarily large deficits. This fact follows not only by \fullref{thmA} but also by the definition of HOMFLYPT polynomial \eqref{skein1} and the additivity of braid indices under connected sums (Birman and Menasco \cite{BM4}).} \end{remark} \subsection{Birman--Menasco diagram}\label{section-BM} As an application of \fullref{thmA}, we study another infinite class of knots including all the Jones' five knots ($9_{42}, 9_{49}, 10_{132}, 10_{150}, 10_{156}$) on which the MFW--inequality is not sharp. We call the block-strand diagram (see \cite{MTWS-I} for definition) of \fullref{menasco} the Birman--Menasco (BM) diagram. \begin{figure}[ht!]\small \begin{center} \labellist \pinlabel {$w$} at 214 495 \pinlabel {$z$} at 271 474 \pinlabel {$y$} at 360 474 \pinlabel {$x$} at 407 495 \endlabellist \includegraphics [height=35mm]{\figdir/menasco} \end{center} \caption{The Birman--Menasco diagram $BM_{x,y,z,w}$} \label{menasco} \end{figure} \begin{definition} Let $BM_{x,y,z,w}$, where $x,y,z,w \in \mathbb{Z}$, be the knot (or the link) type which is obtained by assigning $x$ (resp.\ $y, z, w)$ horizontal positive half-twists on two strands to the block $X$ (resp.\ $Y, Z, W)$ of the BM diagram. \end{definition} Recall that on all but only five knots ($9_{42}, 9_{49}, 10_{132}, 10_{150}, 10_{156}$) up to crossing number $10$ the MFW--inequality is sharp. An interesting property of the BM diagram is that it carries all the five knots. Namely, we have $9_{42}=BM_{-1,1,-2,-1}=BM_{-1,-2,-2,2}, \ $ $9_{49}=BM_{-1,1,1,2}\ $, $10_{132} = BM_{-1,-2,-2,-2}, \ $ $10_{150} = BM_{3,-2,-2,2} = BM_{-1,2,-2,2}$ $= BM_{-1,-2,2,2} = BM_{-1,1,2,-1} =BM_{3,1,-2,-1},$ and $10_{156}=BM_{-1,1,1,-2}$. We have the following theorem, which was conjectured informally by Birman and Menasco: % \begin{theorem}\label{BM-thm} There are infinitely many $(x,y,z,w)$'s such that the MFW--inequality is not sharp on $BM_{x,y,z,w}$. \end{theorem} \begin{lemma}\label{D+} We have $D_{BM_{x,y,z,w}}^+ \geq 2$. \end{lemma} \begin{figure}[ht!] \begin{center} \labellist \small \pinlabel (1) [tl] at 130 703 \pinlabel (2-1) [tl] at 0 525 \pinlabel (2-2) [tl] at 308 525 \pinlabel (3-1) [tl] at 0 330 \pinlabel (3-2) [bl] at 129 27 \pinlabel (3-3) [tl] at 308 330 \pinlabel {$K_+$} at 143 431 \pinlabel {$K_-$} at 298 621 \pinlabel {$K_0$} at 143 239 \pinlabel {$x$} at 347 676 \pinlabel {$y$} at 314 662 \pinlabel {$z$} at 258 662 \pinlabel {$w$} at 222 676 \pinlabel {$w$} at 68 485 \pinlabel {$z$} at 105 472 \pinlabel {$y$} at 161 472 \pinlabel {$x$} at 194 485 \pinlabel {$w$} at 376 485 \pinlabel {$z$} at 412 472 \pinlabel {$y$} at 468 472 \pinlabel {$x$} at 500 485 \pinlabel {$w$} at 68 295 \pinlabel {$z$} at 105 282 \pinlabel {$y$} at 161 282 \pinlabel {$x$} at 194 295 \pinlabel {$w$} at 349 210 \pinlabel {$x$} at 382 251 \pinlabel {$z$} at 411 282 \pinlabel {$y$} at 467 281 \pinlabel {$w$} at 195 60 \pinlabel {$x$} at 227 97 \pinlabel {$z$} at 258 129 \pinlabel {$y$} at 314 129 \pinlabel {isotopy} [tr] at 158 142 \endlabellist \includegraphics [height=120mm]{\figdir/menasco2} \end{center} \caption{The BM--diagram satisfies the sufficient condition} \label{menasco2} \end{figure} \begin{proof}[Proof of \fullref{D+}] Change the BM diagram into the diagram in sketch (1) of \fullref{menasco2}% by braid isotopy and denote it by $K$. Focus on the crossing shaded in sketch (1). Regard $K = K_-.$ We can apply positive destabilization once to $K_+$ and obtain the diagram in sketch (2-2). We also can apply positive destabilization once to $K_0$ as we can see in the passage sketch (3-1) $\Rightarrow$ (3-2) $\Rightarrow$ (3-3). Therefore by \fullref{thmA} we have $D_{BM_{x,y,z,w}}^+ \geq 2$ for any $(x,y,z,w)$. \end{proof} It remains to prove that there exist infinitely many $(x,y,z,w)$'s such that the braid index of $BM_{x,y,z,w}$ is $4$. We introduce ${\cal K}_n := BM_{-1, -2, n, 2}$ and will show that for all $m \geq 1$ the braid index of ${\cal K}_{2m}$ is $4$. (Note that ${\cal K}_{-2}= 9_{42}, {\cal K}_2 =10_{150}$ and ${\cal K}_{2m}$ is a knot.) It will then follow, thanks to \fullref{D+}, that the MFW--inequality cannot be sharp on any ${\cal K}_{2m},\ m\geq 1.$ In order to do this, we use the {\em enhanced Milnor number} $\lambda$ defined by Neumann and Rudolph \cite{NR}. Recall that the fiber surface of a fiber knot is obtained by plumbing and deplumbing Hopf bands (see Giroux \cite{Giroux}). This $\lambda$ is an invariant of fibered knots and links counting algebraically the number of negative Hopf bands to get the fiber surface. \begin{figure}[p] \begin{center} \labellist\small \pinlabel (1) [r] at 35 760 \pinlabel (2) [r] at 35 618 \pinlabel (3) [r] at 35 476 \pinlabel (4) [r] at 35 334 \pinlabel (5) [r] at 35 194 \pinlabel (10) [r] at 404 760 \pinlabel (9) [r] at 404 618 \pinlabel (8) [r] at 404 476 \pinlabel (7) [r] at 404 334 \pinlabel (6) [r] at 404 194 \hair 1.5pt \pinlabel {$x=-1$} [t] at 152 673 \pinlabel {$y=-2$} [tl] at 219 657 \pinlabel {$w=2$} [t] at 64 666 \pinlabel {$n$} [t] at 235 695 \pinlabel {$n-1$} [t] at 239 553 \endlabellist \includegraphics [height=160mm]{\figdir/deplum} \end{center} \caption{Deformation of ${\cal K}_n$} \label{deplum} \end{figure} \newpage \begin{lemma}\label{lambda=1} All ${\cal K}_n$ $(n \geq 2)$ are fibered and have enhanced Milnor number $\lambda = 1.$ \end{lemma} \begin{proof}[Proof of \fullref{lambda=1}] Sketch (1) of \fullref{deplum} is the standard Bennequin surface of ${\cal K}_n$. We compress it twice as in the passage sketch $(1) \Rightarrow (2)$ along the disks bounded by dotted circles in sketch (1). Next, deplumb positive Hopf bands as much as possible as in the passage sketch $(2) \Rightarrow (3) \Rightarrow (4)=(5)$. Then isotope the surface until we get sketch (8). These operations do not change the enhanced Milnor number. We apply Melvin and Morton's trick \cite{MM} p.167, as in the passage sketch $(8) \Rightarrow (9).$ We remark that the enhanced Milnor number is invariant under this trick. The surface of sketch $(9)=(10),$ whose boundary is Pretzel link $P(-2,0,2)$, is plumbing of a positive Hopf band and a negative Hopf band. Thus it has $\lambda = 1$ so does ${\cal K}_n$. \end{proof} Here we summarize Xu's classification of $3$--braids \cite{Xu}. Let $\sigma_1, \sigma_2$ be the standard generators of $B_3$ the braid group of $3$--strings satisfying $\sigma_1 \sigma_2 \sigma_1 = \sigma_2 \sigma_1 \sigma_2$. Let $a_1 := \sigma_1, a_2 := \sigma_2$ and $a_3 := \sigma_2 \sigma_1 \sigma_2^{-1}$. We can identify them with the twisted bands in \fullref{bands}. Let $\alpha := a_1 a_3 = a_2 a_1 = a_3 a_2$. If $w \in B_3$ let $\overline{w}$ denote $w^{-1}.$ \begin{figure}[ht!] \labellist\small \pinlabel $a_1$ [t] at 121 460 \pinlabel $a_2$ [t] at 324 460 \pinlabel $a_3$ [t] at 492 460 \endlabellist \begin{center} \includegraphics [height=24mm]{\figdir/bands} \end{center} \caption{Xu's band generators} \label{bands}% \end{figure} \begin{theorem}[Xu \cite{Xu}]\label{Xu-thm} Every conjugacy class in $B_3$ can be represented by a shortest word in $a_1, a_2, a_3$ uniquely up to symmetry. And the word has one of the three forms: $$(1) \alpha^k P, \quad (2) N \overline{\alpha}^{k} , \quad (3) NP. $$ where $k\geq 0$ and $\overline{N}, P$ are positive words and the arrays of subscripts of the words are non-decreasing. \end{theorem} The next is another lemma for \fullref{BM-thm}: \begin{lemma}\label{ABCD} If a closed $3$--braid has $\lambda=1$ and is a knot, then up to symmetry it has one of the following forms: \begin{eqnarray*} A_x &:=& \overline{a_3} \ \overline{a_2} \ (a_1)^x, \quad x\geq 2, \mbox{\ even,} \\% B_{x, y} &:=& \overline{a_3} \ \overline{a_3} \ (a_1)^x (a_2)^y, \quad x, y \geq 3, \mbox{\ odd,} \\% C_{x, y, z} &:=& \overline{a_2} \ (a_1)^x (a_2)^y (a_3)^z, \quad x+z= \mbox{odd, } \ y= \mbox{even, } \ x,y,z\geq 1,\\ % D_{x, y, z, w} &:=& \overline{a_2} \ (a_1)^x (a_2)^y (a_3)^z (a_1)^w, % \quad x,y \geq 2, \ z,w\geq 1. % \end{eqnarray*} \end{lemma} \begin{proof}[Proof of \fullref{ABCD}] Assume we have a word $w \in B_3.$ By \fullref{Xu-thm}, $w$ has one of the following forms: \begin{center} \begin{tabular}{|l|ll|} \hline \vrule width0pt height 12pt depth 5pt Case (1)-1 & $w=\alpha^k$ & $k \geq 1$ \\ \hline \vrule width0pt height 12pt depth 5pt Case (1)-2 & $w=\alpha^k P$ & $k \geq 1$ \\ \hline \vrule width0pt height 12pt depth 5pt Case (1)-3 & $w=P$ & no $\alpha$ part \\ \hline \vrule width0pt height 12pt depth 5pt Case (2)-1 & $w=\overline{\alpha}^k$ & $k \geq 1$ \\ \hline \vrule width0pt height 12pt depth 5pt Case (2)-2 & $w=N\overline{\alpha}^k$ & $k \geq 1$ \\ \hline \vrule width0pt height 12pt depth 5pt Case (2)-2 & $w= N$ & no $\overline\alpha$ part \\ \hline \vrule width0pt height 12pt depth 5pt Case (3) & $w=NP$ & $18$ cases to study \\ \hline \end{tabular} \end{center} In this proof, we use the simplified notations: \begin{center} \begin{tabular}{|c|l|c|} \hline \vrule width0pt height 12pt depth 5pt symbol & meaning & change in $\lambda$ \\ \hline \vrule width0pt height 12pt depth 5pt $i$ & $a_i$ for $i=1,2,3.$ & --- \\ \hline \vrule width0pt height 12pt depth 5pt $=$ & same conjugacy class & $0$ \\ \hline \vrule width0pt height 12pt depth 5pt $\longrightarrow$ & deplumb positive-Hopf bands & $0$ \\ \hline \vrule width0pt height 12pt depth 5pt $\Longrightarrow$ & deplumb negative-Hopf bands & $\geq 1$ \\ \hline \vrule width0pt height 12pt depth 5pt $\approx$ & Melvin--Morton trick \cite{MM} & $0$ \\ \hline \vrule width0pt height 12pt depth 5pt $\leadsto$ &composition of deplumbings of $\pm$ Hopf bands & $\geq 0$ \\ \hline \end{tabular} \end{center} These are formulae we use: \begin{eqnarray} && \alpha^2 \longrightarrow \alpha, \mbox{ \ \fullref{alpha}} \label{1} \\ && \alpha 1 2 3 \longrightarrow \alpha, \mbox{ \ \fullref{alpha2}} \label{2} \\ && \overline{i} (i-1) i \ \approx \ \overline{i}\ \overline{i-1} i, \mbox{ \ Melvin--Morton trick} \label{3} \\ && i (i+1) \overline{i} \ \approx \ i \ \overline{i+1}\ \overline{i}, \mbox{ \ Melvin--Morton trick} \label{4} \end{eqnarray} \begin{figure}[ht!] \begin{center} \labellist\small\hair 1.5pt \pinlabel $\alpha$ [t] at 61 695 \pinlabel $\alpha$ [t] at 107 695 \pinlabel $\alpha$ [t] at 541 695 \hair 0pt \pinlabel deplumb [l] <0pt,-10pt> at 146 685 \pinlabel slide [l] <0pt,-10pt> at 317 685 \pinlabel deplumb [l] <0pt,-10pt> at 459 685 \endlabellist \includegraphics [width=.98\hsize]{\figdir/alpha} \end{center} \caption{$\alpha^2 \longrightarrow \alpha$} \label{alpha}% \end{figure} \begin{figure}[ht!] \begin{center} \labellist\small\hair 1.5pt \pinlabel $\alpha$ [t] at 67 366 \pinlabel $\alpha$ [t] at 515 366 \hair 0pt \pinlabel deplumb [l] <0pt,-10pt> at 145 356 \pinlabel deplumb [l] <0pt,-10pt> at 285 356 \pinlabel deplumb [l] <0pt,-10pt> at 432 356 \pinlabel slide [l] <0pt,-20pt> at 145 356 \pinlabel slide [l] <0pt,-20pt> at 285 356 \endlabellist \includegraphics [width=.98\hsize]{\figdir/alpha2} \end{center} \caption{$\alpha 1 2 3 \longrightarrow \alpha$} \label{alpha2}\end{figure} Now we study each case. \noindent\textbf{Case (1)-1}\qua By \eqref{1}, we have $w=\alpha^k \longrightarrow \alpha (=$ unknot). Thus $w$ has $\lambda=0.$ \noindent\textbf{Case (1)-2}\qua By \eqref{1} up to permutation of $\{1,2,3\}$ we have $\alpha^k P \longrightarrow \alpha P \longrightarrow$\break $\alpha (1 2 3 1 2 3 \cdots \cdots).$ Thanks to \eqref{2} we have $$\alpha \ \overbrace{1 2 3 1 2 3 \cdots \cdots}^{\mbox{length$=l$}} \ \longrightarrow \ \alpha \ \overbrace{1 2 3 1 2 3 \cdots \cdots}^{\mbox{length $=l-3$}} \quad \mbox{for } l\geq 3.$$ If $l=1, 2$, we have $\alpha 1 = 2 1 1 \longrightarrow \alpha$ and $\alpha 1 2 = 2 1 1 2 \longrightarrow \alpha$. Thus $w$ has $\lambda=0.$ \noindent\textbf{Case (1)-3}\qua Assume $w=P$. There are three possible cases: $$P \longrightarrow (1 2 3)^n, \quad P \longrightarrow (1 2 3)^n 1 \quad \mbox{ and } \quad P \longrightarrow (1 2 3)^n 1 2 \quad \mbox{where}\ n\geq 0.$$ If $w$ satisfies the first case, it is proved that $(123)^n$ is not fibered in Theorem 3.2 of \cite{S-preprint}, where Stoimenow determines fibreness of strongly quasi-positive $3$--braid links. Therefore, $w$ is not fibered. The second case can be reduced to the first case, since $(1 2 3)^n 1 = 1 (1 2 3)^n \longrightarrow (1 2 3)^n.$ For the third case, since $(1 2 3)^n 1 2 = 2 (1 2 3)^n 1 = \alpha (2 3 1)^n \longrightarrow \alpha,$ $w$ has $\lambda=0.$ \noindent\textbf{Case (2)-1}\qua By \eqref{1}, $w=\overline{\alpha}^k\Longrightarrow\overline{\alpha}$ and $w$ has $\lambda=2(k-1)\neq 1$. \noindent\textbf{Case (2)-2}\qua Suppose $w=N \overline{\alpha}^k$ where $k\geq 1$. If $w=\overline{i}\ \overline{\alpha},$ we have $\overline{i}\ \overline{\alpha} \Longrightarrow \overline{\alpha}$ and $w$ has $\lambda=1.$ However, the closure of $w$ has more than one component and it does not satisfy the condition of the lemma. If $w\neq \overline{i}\ \overline{\alpha},$ we have $N \overline{\alpha}^k \Longrightarrow \overline{\alpha}$ by \eqref{2}, and $w$ has $\lambda \geq 2.$ \noindent\textbf{Case (2)-3}\qua Suppose $w=N$. There are three possible cases: $$N \Longrightarrow (\overline{3} \ \overline{2}\ \overline{1})^n, \quad N \Longrightarrow (\overline{3} \ \overline{2}\ \overline{1})^n \overline{3} \quad \mbox{and} \quad N \Longrightarrow (\overline{3} \ \overline{2}\ \overline{1})^n \overline{3} \ \overline{2}\quad \mbox{where}\ n\geq 0.$$ For the first case, $w$ is not fibered \cite{S-preprint}. For the second case, if $n=0$ then $w$ has $\lambda=1$ if and only if $w=\overline{3}\ \overline{3}$. However this has two components. If $n\geq 1,$ since $(\overline{3} \ \overline{2}\ \overline{1})^n \overline{3} \Longrightarrow (\overline{3} \ \overline{2}\ \overline{1})^n$ it can be reduced to the first case. For the third case, if $n=0$ then $w$ has $\lambda=1$ if and only if $w=\overline{3}\ \overline{3} \ \overline{2}.$ However it has two components. If $n\geq 1,$ we have $(\overline{3} \ \overline{2} \ \overline{1})^n \overline{3} \ \overline{2} = \overline{2} \ \overline{3} (\overline{2} \ \overline{1} \ \overline{3})^n =\overline{\alpha}\ (\overline{2} \ \overline{1} \ \overline{3})^n \Longrightarrow \overline{\alpha}$ and $w$ has $\lambda \geq 3n.$ \noindent\textbf{Case (3)}\qua Assume $w=NP.$ Let $w'$ be a word obtained from $w$ by deplumbing $\pm$ Hopf bands sufficiently enough times, ie, $w \leadsto w'.$ This $w'$ has one of the following 18 forms up to permutation of $\{1,2,3\}$. $$ \begin{tabular}{|l|lll|} \hline case & $w'$ & & \\ \hline \vrule width0pt height 12pt depth 5pt i & $(\overline{2}\ \overline{1}\ \overline{3}% )^{k}(123)^{l}$ & $k, l \geq 1$ & \\ \hline \vrule width0pt height 12pt depth 5pt ii & $(\overline{2}\ \overline{1}\ \overline{3}% )^{k}(123)^{l}1$ & $k\geq 1,l\geq 0$ & \\ \hline \vrule width0pt height 12pt depth 5pt iii & $(\overline{2}\ \overline{1}\ \overline{3}% )^{k}(123)^{l}12$ & $k\geq 1,l\geq 0$ & not shortest word \\ \hline \vrule width0pt height 12pt depth 5pt iv & $\overline{3}(\overline{2}\ \overline{1}\ \overline{3}% )^{k}(123)^{l}$ & $k\geq 0,l\geq 1$ & not shortest word \\ \hline \vrule width0pt height 12pt depth 5pt v & $\overline{3}(\overline{2}\ \overline{1}\ \overline{3}% )^{k}(123)^{l}1$ & $k, l \geq 0$ & \\ \hline \vrule width0pt height 12pt depth 5pt vi & $\overline{3}(\overline{2}\ \overline{1}\ \overline{3}% )^{k}(123)^{l}12$ & $k, l \geq 0$ & \\ \hline \vrule width0pt height 12pt depth 5pt vii & $\overline{1}\ \overline{3}(\overline{2}\ \overline{1}% \ \overline{3})^{k}(123)^{l}$ & $k\geq 0,l\geq 1$ & \\ \hline \vrule width0pt height 12pt depth 5pt viii & $\overline{1}\ \overline{3}(\overline{2}\ \overline{1}% \ \overline{3})^{k}(123)^{l}1$ & $k, l \geq 0$ & not shortest word \\ \hline \vrule width0pt height 12pt depth 5pt ix & $\overline{1}\ \overline{3}(\overline{2}\ \overline{1}\ \overline{3})^{k}(123)^{l}12$ & $k, l\geq 0$ & \\ \hline \end{tabular}$$ $$ \begin{tabular}{|l|lll|} \hline \vrule width0pt height 12pt depth 5pt i$'$ & $(\overline{1}\ \overline{3}\ \overline{2}% )^{k}(123)^{l}$ & $k, l\geq 1$ & \\ \hline \vrule width0pt height 12pt depth 5pt ii$'$ & $(\overline{1}\ \overline{3}\ \overline{2}% )^{k}(123)^{l}1$ & $k\geq 1,l\geq 0$ & not shortest word \\ \hline \vrule width0pt height 12pt depth 5pt iii$'$ & $(\overline{1}\ \overline{3}\ \overline{2}% )^{k}(123)^{l}12$ & $k\geq 1,l\geq 0$ & \\ \hline \vrule width0pt height 12pt depth 5pt iv$'$ & $\overline{2}(\overline{1}\ \overline{3}\ \overline{2}% )^{k}(123)^{l}$ & $k\geq 0,l\geq 1$ & \\ \hline \vrule width0pt height 12pt depth 5pt v$'$ & $\overline{2}(\overline{1}\ \overline{3}\ \overline{2}% )^{k}(123)^{l}1$ & $k, l\geq 0$ & \\ \hline \vrule width0pt height 12pt depth 5pt vi$'$ & $\overline{2}(\overline{1}\ \overline{3}\ \overline{2}% )^{k}(123)^{l}12$ & $k, l\geq 0$ & not shortest word \\ \hline \vrule width0pt height 12pt depth 5pt vii$'$ & $\overline{3}\ \overline{2}(\overline{1}\ \overline{3}% \ \overline{2})^{k}(123)^{l}$ & $k\geq 0,l\geq 1$ & not shortest word \\ \hline \vrule width0pt height 12pt depth 5pt viii$'$ & $\overline{3}\ \overline{2}(\overline{1}\ \overline{3% }\ \overline{2})^{k}(123)^{l}1$ & $k, l\geq 0$ & \\ \hline \vrule width0pt height 12pt depth 5pt ix$'$ & $\overline{3}\ \overline{2}(\overline{1}\ \overline{3}% \ \overline{2})^{k}(123)^{l}12$ & $k, l\geq 0$ & \\ \hline \end{tabular}$$ Since words of case iii, iv, viii, ii$'$, vi$'$ and vii$'$ are not shortest (reducible) we eliminate them from the list. These are reduction formulae we use: \begin{eqnarray} (\overline{2}\ \overline{1}\ \overline{3})(123) &\stackrel{\eqref{F}}{\longrightarrow} & \overline{2}(3\overline{2}) = 1\overline{2}\ \overline{2} \Longrightarrow 1\overline{2} \label{reduction1} \\ (1 \overline{2})(123) &\stackrel{\eqref{S}}{\longrightarrow}& 11 \overline{2}\longrightarrow 1 \overline{2} \label{A}\\ (\overline{2}\ \overline{1}\ \overline{3})(1 \overline{2}) &\stackrel{\eqref{C}}{\longrightarrow} & 1\overline{3}\ \overline{2} = \overline{2}\ \overline{2}3 \Longrightarrow\ =1 \overline{2} \label{B} \\ (\overline{2}\ \overline{1}\ \overline{3})1 &\approx & \overline{2}\ \overline{1}31=1\overline{2}\ \overline{2}1\Longrightarrow\ = 11\overline{3}\longrightarrow 1 \overline{3} \label{C} \\ (\overline{2}\ \overline{1}\ \overline{3})1 \overline{3} &\stackrel{\eqref{C}}{\longrightarrow}& 1 \overline{3}\ \overline{3} \Longrightarrow 1 \overline{3} \label{D} \\ \overline{1}\ \overline{3}(123) &=& \overline{1}22\overline{1}3\longrightarrow \overline{1}2\overline{1}3 =3\overline{1}\ \overline{1}3 \Longrightarrow\ = 33\overline{2} \longrightarrow 3 \overline{2} \label{F} \\ 3\overline{2}(123) & \stackrel{\eqref{S}}{\longrightarrow}& 3(1\overline{2}) = \overline{1}33\longrightarrow \overline{1}3 = 3\overline{2} \label{G} \\ (123)1\overline{3} &\approx & 123\overline{1}\ \overline{3} = 1\overline{3}22\overline{3} \nonumber \\ &\longrightarrow & 1\overline{3}2\overline{3} = 1\overline{3}\ \overline{3}1\Longrightarrow 1\overline{3}1\longrightarrow 1 \overline{3} \label{H} \\ 12 \overline{1}\ \overline{3} &=& 1\overline{3}\ \overline{3}1\Longrightarrow 1\overline{3}1 \longrightarrow \overline{2}1=1 \overline{3} \label{N} \\ (\overline{1}\ \overline{3}\ \overline{2})(123) &\stackrel{\eqref{Q}}{\longrightarrow}& 3\overline{1}3=33\overline{2} \Longrightarrow 3\overline{2}=\overline{1}3 \label{reduction2} \\ (\overline{1}\ \overline{3}\ \overline{2}) \overline{1}3 &=& \overline{1}\ \overline{3}\ 1 \overline{2}\ \overline{2} \Longrightarrow\ \approx \overline{1} 31 \overline{2} = \overline{1}\ \overline{1} 33 \longrightarrow\ \Longrightarrow \overline{1} 3 \label{O} \\ \overline{1}3 (123) &\approx & \overline{1}\ \overline{3} 123 = \overline{1}22\overline{1}3 \longrightarrow\ \Longrightarrow 3\overline{1}3 \longrightarrow 3 \overline{2} = \overline{1}3 \label{P} \\ (\overline{1}\ \overline{3}\ \overline{2})12 &=& \overline{1}2\overline{3}\ \overline{3}2\Longrightarrow\ = \overline{1}22\overline{1} \longrightarrow \overline{1}2\overline{1}=3\overline{1}\ \overline{1} \Longrightarrow 3\overline{1} \label{Q} \\ (\overline{1}\ \overline{3}\ \overline{2})3\overline{1} &\stackrel{\eqref{B}}{\longrightarrow} & 3\overline{1} \quad \mbox{permutation of \eqref{B}} \label{R} \\ \overline{2}(123) &\approx & \overline{2}\ \overline{1}23 = \overline{2} 33\overline{2} \longrightarrow \overline{2} 3\overline{2} = 1 \overline{2} \ \overline{2} \Longrightarrow 1 \overline{2} \label{S} \\ 1 \overline{2}\ (\overline{1}\ \overline{3}\ \overline{2}) &\approx & 12\overline{1}\ \overline{3}\ \overline{2}= 1\overline{3}\ \overline{3}1\overline{2} \Longrightarrow\ = \overline{2}11\overline{2} \longrightarrow \overline{2}1\overline{2} \Longrightarrow 1 \overline{2} \label{W} \end{eqnarray} \begin{sublemma}\label{red3} For $k, l \geq 1,$ we have: \begin{eqnarray} (\overline{2}\ \overline{1}\ \overline{3})^{k} (123)^l & \leadsto & 1 \overline{2} \label{AA} \\ (\overline{1}\ \overline{3}\ \overline{2})^k (123)^l & \leadsto & \overline{1}3 \label{BB} \end{eqnarray} Either case, the increase of $\lambda$ is $\geq 2.$ \end{sublemma} \begin{proof} From \eqref{reduction1}, \eqref{B} and \eqref{C}, we obtain \eqref{AA}. Similarly, \eqref{BB} follows from \eqref{reduction2}, \eqref{O} and \eqref{P}. \end{proof} \noindent\textbf{Case 3-i}\qua By \eqref{AA}, our $w$ is fibered and $\lambda \geq 2.$ \noindent\textbf{Case 3-ii}\qua If $k\geq 1. l=0,$ $$w \leadsto w' = (\overline{2}\ \overline{1}\ \overline{3})^{k} 1 \ \stackrel{\eqref{C}}{\longrightarrow} \ (\overline{2}\ \overline{1}\ \overline{3})^{k-1} 1 \overline{3} \ \stackrel{\eqref{D}}{\longrightarrow} \cdots \stackrel{\eqref{D}}{\longrightarrow} \ 1 \overline{3} $$ and $w$ has $\lambda = 1$ if and only if $w=\overline{2}\ \overline{1}\ \overline{3}\ 1^x$ for some $x \geq 1.$ However it has $2$ or $3$ components and it does not satisfy the condition of \fullref{ABCD}. If $k, l \geq 1,$ $$w \leadsto w' = (\overline{2}\ \overline{1}\ \overline{3})^{k} (123)^l 1 \ \stackrel{\eqref{AA}}{\longrightarrow} \ (1 \overline{2}) 1 \longrightarrow 1 \overline{2}$$ and $w$ has $\lambda \geq 2.$ \noindent\textbf{Case 3-v}\qua When $k=l=0,$ $w$ has $\lambda=1$ if and only if $w= \overline{3}\ \overline{3}\ 1^x$ for some $x\geq 1,$ which has more than $1$ component. If $k \geq 1$ and $l=0$, $$w \leadsto w' = \overline{3}\ (\overline{2}\ \overline{1}\ \overline{3})^k 1 \ \stackrel{\eqref{D}}{\longrightarrow}\ \cdots\ \stackrel{\eqref{D}}{\longrightarrow}\ 1 \overline{3}$$ and $\lambda \geq 2.$ If $k=0, l\geq 1,$ $$w\leadsto w'=\overline{3}\ (123)^l1=(123)^l1\overline{3}\ \stackrel{\eqref{H}}{\longrightarrow}\ \cdots\ \stackrel{\eqref{H}}{\longrightarrow} \ 1 \overline{3}.$$ Thus $w$ has $\lambda=1$ if and only if $w= \overline{3}\ 1^x\ 2^y\ 3^z\ 1^w$ for $x,y,z,w \geq 1.$ If $k, l \geq 1,$ $$w \leadsto w' = \overline{3}\ (\overline{2}\ \overline{1}\ \overline{3})^k (123)^l 1 \ \stackrel{\eqref{AA}}{\longrightarrow} \ \overline{3} (1 \overline{2}) 1 = \overline{3} 11 \overline{3}\longrightarrow\ \Longrightarrow 1 \overline{3}$$ and $\lambda \geq 2.$ \noindent\textbf{Case 3-vi}\qua If $k=l=0$, $$w \leadsto w' = \overline{3}\ 1 2 = 22\overline{1} \longrightarrow 2 \overline{1}.$$ Therefore, $w$ has $\lambda=1$ if and only if $w= \overline{3}\ \overline{3}\ 1^x\ 2^y$ for $x,y \geq 1.$ If $k=0, l\geq 1$ $$w \leadsto w' = \overline{3} (123)^l 1 2 =(123)^l12\overline{3}\longrightarrow (123)^l 1 \overline{3}\ \stackrel{\eqref{H}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{H}}{\longrightarrow} \ 1 \overline{3}$$ and $w$ has $\lambda=1$ if and only if $w= \overline{3} 1^x\ 2^y\ 3^z\ 1^w\ 2^v$ for some $x,y,z,w,v \geq 1.$ When $k\geq 1, l=0$, $$w \leadsto w' = \overline{3} (\overline{2}\ \overline{1}\ \overline{3})^k 12 \longrightarrow (\overline{2}\ \overline{1}\ \overline{3})^k 1 \overline{3} \stackrel{\eqref{D}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{D}}{\longrightarrow} \ 1 \overline{3}$$ and $\lambda \geq 2$. If $k, l \geq 1,$ $$w \leadsto w'\ \stackrel{\eqref{AA}}{\longrightarrow}\ \overline{3}(1\overline{2})12 =\overline{3}11\overline{3} 2 \longrightarrow\overline{3}1\overline{3} 2 = 2 \overline{3}\ \overline{3}\ 2 \longrightarrow\ \Longrightarrow 2 \overline{3}$$ and $\lambda \geq 2$. \noindent\textbf{Case 3-vii}\qua If $k=0, l \geq 1,$ $$w \leadsto w' = \overline{1}\ \overline{3}\ (123)^l\ \stackrel{\eqref{F}}{\longrightarrow} \ 3 \overline{2} (123)^{l-1} \ \stackrel{\eqref{G}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{F}}{\longrightarrow} \ 3 \overline{2}$$ and $\lambda =1$ if and only if $w=\overline{1}\ \overline{3}\ 1^x\ 2^y\ 3^z$ for some $x,y,z \geq 1.$ If $k, l \geq 1,$ $$w \leadsto w'\ \stackrel{\eqref{AA}}{\longrightarrow} \ \overline{1}\ \overline{3}(1 \overline{2})=\overline{2}\ \overline{1}\ \overline{3}1\ \stackrel{\eqref{C}}{\longrightarrow}\ 1 \overline{3}$$ and $\lambda \geq 2$. \noindent\textbf{Case 3-ix}\qua When $k=l=0,$ $$w \leadsto w' = \overline{1}\ \overline{3}\ 12 = \overline{1}22\overline{1} \longrightarrow \ \Longrightarrow \overline{1}2.$$ Thus $w$ has $\lambda=1$ if and only if $w= \overline{1}\ \overline{3}\ 1^x\ 2^y$ for some $x,y \geq 1.$ When $k=0, l\geq 1$ $$w \leadsto w' = \overline{1}\ \overline{3} (123)^l 12 \ \stackrel{\eqref{F}}{\longrightarrow} \ \stackrel{\eqref{G}}{\longrightarrow} \ \overline{1} 312 = 3 \overline{2} 12 \approx 3 \overline{2}\ \overline{1} 2 = \overline{1} 33 \overline{1} \longrightarrow\ \Longrightarrow \overline{1}3$$ and $\lambda\geq 2.$ When $k\geq 1, l=0$ $$w \leadsto w' = \overline{1}\ \overline{3} (\overline{2}\ \overline{1}\ \overline{3})^k 12 = 12 \overline{1}\ \overline{3} (\overline{2}\ \overline{1}\ \overline{3})^k \ \stackrel{\eqref{N}}{\longrightarrow} \ 1 \overline{3} (\overline{2}\ \overline{1}\ \overline{3})^k \ \stackrel{\eqref{D}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{D}}{\longrightarrow} \ 1 \overline{3}$$ and $\lambda\geq 2.$ When $k,l\geq 1$ $$w \leadsto w' \ \stackrel{\eqref{AA}}{\longrightarrow} \ \overline{1}\ \overline{3} (1\overline{2}) 12 = 12\overline{1}\ \overline{3}1\overline{2} \ \stackrel{\eqref{N}}{\longrightarrow} \ 1 \overline{3}1\overline{2} = \overline{2} 11 \overline{2} \longrightarrow\ \Longrightarrow 1 \overline{2}$$ and $\lambda\geq 2.$ \noindent\textbf{Case 3-i$'$}\qua By \eqref{BB} our $w$ is fibered and $\lambda \geq 2.$ \noindent\textbf{Case 3-iii$'$}\qua When $k>l=0,$ $$w \leadsto w' = (\overline{1}\ \overline{3}\ \overline{2})^k 12 \ \stackrel{\eqref{Q}}{\longrightarrow} \ (\overline{1}\ \overline{3}\ \overline{2})^{k-1} 2 \overline{1} \ \stackrel{\eqref{R}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{R}}{\longrightarrow} \ 3 \overline{1}$$ and $\lambda \geq 2.$ When $k,l\geq 1$, $$w \leadsto w' \ \stackrel{\eqref{BB}}{\longrightarrow} \ (\overline{1}3)12 \approx \overline{1}\ \overline{3}\ 12 = \overline{1}22\overline{1} \longrightarrow\ \Longrightarrow 3 \overline{1}$$ \newpage and $\lambda \geq 2$. \noindent\textbf{Case 3-iv$'$}\qua When $k=0, l\geq 1,$ $$w \leadsto w'= \overline{2}(123)^l \ \stackrel{\eqref{S}}{\longrightarrow} \ 1 \overline{2}(123)^{l-1} \ \stackrel{\eqref{A}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{A}}{\longrightarrow} \ 1 \overline{2}.$$ Thus $w$ has $\lambda = 1$ if and only if $w=\overline{2} 1^x 2^y 3^z$ for $x,y,z\geq 1.$ To make the braid closure one component, we further require $x+z= \mbox{odd.}$ If $k, l \geq 1,$ $$w \leadsto w' \ \stackrel{\eqref{BB}}{\longrightarrow} \ \overline{2}\ (\overline{1}3) = 1 \overline{2}\ \overline{2} \Longrightarrow 1 \overline{2}$$ and $\lambda \geq 2.$ \noindent\textbf{Case 3-v$'$}\qua When $k=l=0,$ $$w \leadsto w'= \overline{2}1.$$ Thus $w$ has $\lambda=1$ if and only if $w=\overline{2}\ \overline{2}1^x,$ which has $2$ or $3$ components. When $k=0$ and $l\geq 1$, $$w \leadsto w'= \overline{2} (123)^l 1 \ \stackrel{\eqref{S}}{\longrightarrow} \ 1 \overline{2}(123)^{l-1}1 \ \stackrel{\eqref{A}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{A}}{\longrightarrow} \ 1 \overline{2}1 = 11\overline{2} \longrightarrow 1 \overline{2}$$ thus $w$ has $\lambda=1$ if and only if $w=\overline{2}\ 1^x\ 2^y\ 3^z\ 1^w$ for some $x,y,z,w\geq 1.$ When $k\geq 1, l=0$, $$w \leadsto w' = \overline{2}\ (\overline{1}\ \overline{3}\ \overline{2})^k 1 = 1 \overline{2}\ (\overline{1}\ \overline{3}\ \overline{2})^k \ \stackrel{\eqref{W}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{W}}{\longrightarrow} \ 1 \overline{2}$$ and $\lambda \geq 2.$ When $k, l\geq 1$ $$w \leadsto w' \ \stackrel{\eqref{BB}}{\longrightarrow} \ \overline{2}(\overline{1}3)1 \Longrightarrow 1\overline{2}1 \longrightarrow 1\overline{2}$$ and $\lambda \geq 2.$ \noindent\textbf{Case 3-viii$'$}\qua When $k=l=0,$ $$w \leadsto w'=\overline{3}\ \overline{2}1 = 1\overline{3}\ \overline{2}=\overline{2}1\overline{2}\Longrightarrow 1\overline{2}$$ Thus $w$ has $\lambda=1$ if and only if $w=\overline{3}\ \overline{2}1^x$ for some $x\geq 1.$ When $k=0$ and $l\geq 1$, $$w \leadsto w'=\overline{3}\ \overline{2}(123)^l 1 = 1\overline{3}\ \overline{2}(123)^l \Longrightarrow 1 \overline{2}(123)^l \ \stackrel{\eqref{A}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{A}}{\longrightarrow} \ 1 \overline{2}$$ and $\lambda \geq 2.$ When $k\geq 1$ and $l=0$, $$w \leadsto w'=\overline{3}\ \overline{2}(\overline{1}\ \overline{3}\ \overline{2})^k 1 \Longrightarrow 1 \overline{2}(\overline{1}\ \overline{3}\ \overline{2})^k \ \stackrel{\eqref{W}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{W}}{\longrightarrow} \ 1 \overline{2}$$ and $\lambda \geq 2.$ When $k, l\geq 1$ $$w \leadsto w' \ \stackrel{\eqref{BB}}{\longrightarrow}\ \overline{3}\ \overline{2}(\overline{1}3)1 \Longrightarrow 1 \overline{2}\ \overline{1}3 = \overline{2} 33\overline{2} \Longrightarrow\ \longrightarrow \overline{2}3 = 1 \overline{2}.$$ \newpage and $\lambda \geq 2.$ \noindent\textbf{Case 3-ix$'$}\qua When $k=l=0,$ $$w \leadsto w'=\overline{3}\ \overline{2}12 = 2 \overline{3}\ \overline{3}2 \Longrightarrow\ \longrightarrow 2 \overline{3}.$$ Thus $w$ has $\lambda=1$ if and only if $w= \overline{3}\ \overline{2}1^x 2^y$ for some $x,y \geq 1.$ When $k=0, l\geq 1$, \begin{eqnarray*} w &\leadsto & w'=\overline{3}\ \overline{2}(123)^l 12=12\overline{3}\ \overline{2}(123)^l=\overline{2}11\overline{2} (123)^l\longrightarrow\ \Longrightarrow \overline{2}3(123)^l \\ &=& 1\overline{2}(123)^l\ \stackrel{\eqref{A}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{A}}{\longrightarrow} \ 1 \overline{2} \end{eqnarray*} and $\lambda \geq 2.$ When $k\geq 1, l=0$, $$ w \leadsto w'=\overline{3}\ \overline{2}(\overline{1}\ \overline{3}\ \overline{2})^k12\longrightarrow\ \Longrightarrow 1\overline{2} (\overline{1}\ \overline{3}\ \overline{2})^k\ \stackrel{\eqref{W}}{\longrightarrow} \ \cdots \ \stackrel{\eqref{W}}{\longrightarrow} \ 1 \overline{2}$$ and $\lambda \geq 2.$ When $k, l\geq 1$ $$w \leadsto w'\ \stackrel{\eqref{BB}}{\longrightarrow}\ \overline{3}\ \overline{2}(\overline{1}3)12\longrightarrow\ \Longrightarrow 1\overline{2}(\overline{1}3)=11\overline{2}\ \overline{2}\longrightarrow\ \Longrightarrow 1\overline{2}$$ and $\lambda \geq 2.$ \fullref{lambda-table} summarizes all the words with $\lambda=1.$ \begin{table}[ht!] \caption{} \label{lambda-table} \begin{center} \begin{tabular}{|l|l|} \hline case & word with $\lambda =1.$ \\ \hline % \vrule width0pt height 12pt depth 5pt i & none. \\ \hline % \vrule width0pt height 12pt depth 5pt ii & $\overline{2}\ \overline{1}\ \overline{3}\ 1^{x}\ $(2 or 3 components.) \\ \hline % v & \begin{tabular}{l} $\overline{3}\ 1^{x}\ 2^{y}\ 3^{z}\ 1^{w}\ =\left\{% \begin{array}{ll} \vrule width0pt height 12pt depth 5pt C_{x+1,y,z} &\ \mbox{when } w=1, \\ D_{x+1,y,z,w-1} & \ \mbox{when } w\geq 2. \end{array} \right. $ \\ \vrule width0pt height 12pt depth 5pt $\overline{3}\ \overline{3}\ 1^{x}\ $(2 or 3 components.)% \end{tabular} \\ \hline% vi & \begin{tabular}{l} \vrule width0pt height 12pt depth 5pt $\overline{3}\ \overline{3}\ 1^{x}\ 2^{y}=:B_{x,y}.$ \\ % $\overline{3}\ 1^{x}\ 2^{y}\ 3^{z}\ 1^{w}\ 2^{v}\ =\left\{% \begin{array}{ll} C_{x+v+1,y,z} & \ \mbox{when } w=1, \\ \vrule width0pt height 12pt depth 5pt D_{x+v+1,y,z,w-1} &\ \mbox{when } w\geq 2.% \end{array} \right. $% \end{tabular} \\ \hline % \vrule width0pt height 12pt depth 5pt vii & $\overline{1}\ \overline{3}\ 1^{x}\ 2^{y}\ 3^{z}\ =% \overline{1}\ \overline{3}\ 1^{x+z}\ 2^{y}$ \\ \hline % ix & $\overline{1}\ \overline{3}\ 1^{x}\ 2^{y}\ = \left\{ \begin{array}{lll} \vrule width0pt height 12pt depth 5pt \overline{3}\ \overline{3}1^{x+1} & \ \mbox{when } y=1 & (\mbox{$2$ or $3$ components)} \\ \vrule width0pt height 12pt depth 5pt B_{x+1,y-1} & \ \mbox{when } y\geq 2 & \end{array}\right. $ \\ \hline % \vrule width0pt height 12pt depth 5pt i$'$ & none. \\ \hline % \vrule width0pt height 12pt depth 5pt iii$'$ & none. \\ \hline \vrule width0pt height 12pt depth 5pt iv$'$ & $\overline{2}\ 1^{x}\ 2^{y}\ 3^{z}\ =:C_{x,y,z}.$ \\ \hline v$'$ & \begin{tabular}{l} $\vrule width0pt height 12pt depth 5pt \overline{2}\ 1^{x}\ 2^{y}\ 3^{z}\ 1^{w}\ =:D_{x,y,z,w}. $ \\ $\vrule width0pt height 12pt depth 5pt \overline{2}\ \overline{2}\ 1^{x}\ $(2 or 3 components.)% \end{tabular}\\ \hline % \vrule width0pt height 12pt depth 5pt viii$'$ & $\overline{3}\ \overline{2}\ 1^{x}\ =:A_{x}.$ \\ \hline ix$'$ & $\overline{3}\ \overline{2}\ 1^{x}\ 2^{y}$ $=\left\{ \begin{array}{lll} \vrule width0pt height 12pt depth 5pt \overline{3}\ \overline{3}\ 2^{y+1} & \ \mbox{when } x=1, & (\mbox{$2$ or $3$ components}) \\ \vrule width0pt height 12pt depth 5pt B_{x-1,y+1} &\ \mbox{when } x\geq 2. & \end{array} \right. $ \\ \hline \end{tabular} \end{center} \end{table} Words $A_x, \cdots, D_{x,y,z,w}$ are defined in \fullref{lambda-table}. We can see that any word with $\lambda=1$ and having one component has one of the forms; $A_x, \cdots, D_{x,y,z,w}.$ \end{proof} \begin{lemma}\label{alexander} The leading terms of the Alexander polynomials of ${\cal K}_{n}$, $A_x$, $B_{x, y}$, $C_{x, y, z}$ and $D_{x, y, z, w}$ are the following: \begin{eqnarray*} {\cal K}_n; && \pm(1 - 4t -6t^2 + 8t^3 - \cdots) \quad \mbox{if $n \geq 2$,} \\ A_x; && \pm(1 - 3t + \cdots) \quad \mbox{if } x\geq 2, \\% B_{x, y}; && \pm(1 - 3t + \cdots) \quad \mbox{if } x, y \geq 3, \\% C_{x, y, z}; && \pm(1 - 5t + \cdots) \quad \mbox{if $x,y,z \geq 2$,}\\ % C_{1, 2, z}, C_{1, y, 2}, C_{2, y, 1}, C_{x, 2, 1}; % && \pm(1 - 4t + 6t^2 -7t^3 + \cdots) % \quad \mbox{if $x, y, z \geq 4$,}\\ % C_{1, y, z}, C_{x, y, 1}; % && \pm(1 - 4t + 7t^2 + \cdots) \quad \mbox{if $x, y, z \geq 3$}\\% D_{x, y, z, w}, D_{x, y, z, 1}; && \pm(1 - 6t + \cdots) \quad \mbox{if } x,y,z,w \geq 2, \\ D_{x, y, 1, w}; && \pm(1 - 5t + \cdots) \quad \mbox{if } x,y,w \geq 2. % \end{eqnarray*} In particular, ${\cal K}_n\neq A_x, B_{x, y},C_{x, y, z}, D_{x, y, z, w}.$ \end{lemma} \begin{figure}[hb!] \begin{center} \labellist\small\hair1.5pt \pinlabel {(1)} at -15 425 \pinlabel {(2)} at 358 348 \pinlabel {(3)} at -5 360 \pinlabel {(4)} at 217 425 \pinlabel {(5)} at 421 390 \pinlabel {$x$} [t] at 44 304 \pinlabel {$y$} [t] at 256 366 \pinlabel {$z$} [t] at 464 300 \endlabellist \includegraphics [height=40mm]{\figdir/2-123} \end{center} \caption{The Bennequin surface $F$ of $C_{x, y, z}=\overline{2} 1^x 2^y 3^z$ and a basis for $H_1(F)$} \label{2-123}% \end{figure} \newpage \begin{proof}[Proof of \fullref{alexander}] We prove that the Alexander polynomial of $C_{x, y, z}$ for $x, y, z \geq 2$ is $\pm(1 - 5t + \cdots)$. Recall that Xu's Bennequin surface is a minimal genus Seifert surface. Let $F$ be the Bennequin surface of $C_{x, y, z}$ and choose a basis $$\{ u^{(1)}, u^{(2)}, u^{(3)}_1, \cdots, u^{(3)}_{x-1}, u^{(4)}_1, \cdots, u^{(4)}_{y-1}, u^{(5)}_1, \cdots, u^{(5)}_{z-1} \}$$ for $H_1(F)$ as in \fullref{2-123}, where $u^{(k)}$ ($k=1,\cdots,5$) corresponds to loop $(k)$. With respect to the basis, let $V_{x,y,z}$ denote the Seifert matrix for $C_{x, y, z}$. $$ V_{x,y,z} = \left[ \begin{array}{c|c|cccc|cccc|cccc} & \scriptstyle{1} & & & & & \scriptstyle{1} & & & & & & & \\ \hline % \scriptstyle{1} & & \scriptstyle{-1} & & & & & & & & \scriptstyle{1} & & & \\ \hline% & & \scriptstyle{-1} & \scriptstyle{1} & & & & & & & & & & \\ & & & \scriptstyle{-1} & \scriptstyle{\ddots} & & & & & & & & & \\ & & & & \scriptstyle{\ddots} & \scriptstyle{1} & & & & & & & & \\ & & & & & \scriptstyle{-1} & & & & & & & & \\ \hline% & & & & & & \scriptstyle{-1} & \scriptstyle{1} & & & & & & \\ & & & & & & & \scriptstyle{-1} & \scriptstyle{\ddots} & & & & & \\ & & & & & & & & \scriptstyle{\ddots} & \scriptstyle{1} & & & & \\ & & & & & & & & & \scriptstyle{-1} & & & & \\ \hline% & & & & & & & & & & \scriptstyle{-1} & \scriptstyle{1} & & \\ & & & & & & & & & & & \scriptstyle{-1} & \scriptstyle{\ddots} & \\ & & & & & & & & & & & & \scriptstyle{\ddots} & \scriptstyle{1} \\ & & & & & & & & & & & & & \scriptstyle{-1}% \end{array} \right] $$ The empty spaces contain only $0$'s. The $3$rd (resp.\ $4$th, $5$th) diagonal block has size $(x-1)\times (x-1)$ (resp.\ $(y-1)\times (y-1)$, $(z-1)\times (z-1)$). The Alexander polynomial satisfies: $\Delta_{x,y,z}(t) = \det(V_{x,y,z}^T - tV_{x,y,z})$ $$ = \det \left[ \begin{array}{c|c|cccc|cccc|cccc} & \scriptstyle{1-t} & & & & & -t & & & & & & & \\ \hline% \scriptstyle{1-t} & & t & & & & & & & & -t & & & \\ \hline% & -1 & \scriptstyle{-1+t} & -t & & & & & & & & & & \\ & & 1 & \ddots & \ddots & & & & & & & & & \\ & & & \ddots & \ddots & -t & & & & & & & & \\ & & & & 1 & \scriptstyle{-1+t} & & & & & & & & \\ \hline% 1 & & & & & & \scriptstyle{-1+t} & -t & & & & & & \\ & & & & & & 1 & \ddots & \ddots & & & & & \\ & & & & & & & \ddots & \ddots & -t & & & & \\ & & & & & & & & 1 & \scriptstyle{-1+t} & & & & \\ \hline% & 1 & & & & & & & & & \scriptstyle{-1+t} & -t & & \\ & & & & & & & & & & 1 & \ddots & \ddots & \\ & & & & & & & & & & & \ddots & \ddots & -t \\ & & & & & & & & & & & & 1 & \scriptstyle{-1+t}% \end{array} \right]. $$ Expanding it by the $(x+1)$th column, we have; $ \Delta_{x,y,z}(t) = (-1+t)\Delta_{x-1,y,z}(t)$ $$ -(-t)\det \left[ \begin{array}{c|c|cccc|ccc|ccc} & \scriptstyle{1-t} & & & & & -t & & & & & \\ \hline% \scriptstyle{1-t} & & t & & & & & & & -t & & \\ \hline% & -1 & \scriptstyle{-1+t} & -t & & & & & & & & \\ & & 1 & \ddots & -t & & & & & & & \\ & & & 1 & \scriptstyle{-1+t} & -t & & & & & & \\ & & & & & 1 & & & & & & \\ \hline% 1 & & & & & & \scriptstyle{-1+t} & -t & & & & \\ & & & & & & 1 & \ddots & -t & & & \\ & & & & & & & 1 & \scriptstyle{-1+t} & & & \\ \hline% & 1 & & & & & & & & \scriptstyle{-1+t} & -t & \\ & & & & & & & & & 1 & \ddots & -t \\ & & & & & & & & & & 1 & \scriptstyle{-1+t}% \end{array} \right] $$ $= (-1+t)\Delta_{x-1,y,z}(t) + t \Delta_{x-2,y,z}(t).$ \eject If $\Delta_{i,y,z}(t)=(-1)^i (\alpha_0 + \alpha_1 t + \alpha_2 t^2 + \cdots )$ for $i=x-1$ and $x-2$, then \begin{eqnarray*} \Delta_{x,y,z}(t) &=& (-1+t) (-1)^{x-1} (\alpha_0 + \alpha_1 t + \alpha_2 t^2 + \cdots )\\ &&\hspace{2in}+ t (-1)^{x-2} (\alpha_0 + \alpha_1 t + \alpha_2 t^2 + \cdots) \\ % &=& (-1)^x (\alpha_0 + \alpha_1 t + \alpha_2 t^2 + \cdots ). \end{eqnarray*} In fact, $\Delta_{x,y,z}(t) = (-1)^{x+y+z} (1-5t+ \cdots )$ for all $x,y,z \in \{2,3\}$. By induction, $\Delta_{x,y,z}(t) = (-1)^{x+y+z} (1-5t+ \cdots )$ for all $x, y, z \geq 2.$ Other cases follow by similar arguments. \end{proof} Finally we are ready to prove the theorem. \begin{proof}[Proof of \fullref{BM-thm}] By Lemmas \ref{lambda=1}, \ref{ABCD}, \ref{alexander}, our knot ${\cal K}_{2m}$ where ($m\geq 1$) cannot be a $3$--braid. Thus by \fullref{D+}, \fullref{BM-thm} follows. \end{proof} \section{Uniqueness of the algebraic crossing number at minimal braid index}\label{chap3} \subsection{Sharpness of the MFW--inequality and conjectures} It has been conjectured (see \cite[page 357]{Jones-1} for example) that the exponent sum in a minimal braid representation is a knot invariant. \begin{conjecture}[Main Conjecture]\label{Jones-conj} Let ${\cal K}$ be a knot type of braid index $b_{\cal K}$. If $K^1$ and $K^2$ are braid representatives of ${\cal K}$ with $b_{K^1}=b_{K^2}=b_{\cal K}$ then their algebraic crossing numbers have $c_{K^1}=c_{K^2}$. \end{conjecture} We deform it into: \begin{conjecture}[Stronger Conjecture]\label{greedy} Let ${\cal B}_{\cal K}$ be the set of braid representatives of ${\cal K}.$ Let $\Phi\co {\cal B}_{\cal K} \to \mathbb{N} \times \mathbb{Z}$ be a map such that $\Phi(K):=\left( b_K, c_K \right)$ for $K\in {\cal B}_{\cal K}.$ Then there exists a unique $c_{\cal K} \in \mathbb{Z}$ with \begin{equation}\label{quadrant equation} \Phi({\cal B}_{\cal K})=\left\{ (b_{\cal K}+x+y, c_{\cal K}+x-y)\ |\ x,y\in \mathbb{N} \right\}, \end{equation} a subset of the infinite quadrant region shaded in \fullref{b-region}. \end{conjecture} \begin{figure}[ht!] \begin{center} \labellist\small\hair2pt \pinlabel {$\gamma_{\min}$} [r] at 49 147 \pinlabel {$\beta_{\max}$} [r] at 49 68 \pinlabel {$c_{\cal K}$} [r] at 49 108 \pinlabel {$b_{\cal K}$} [t] at 92 25 \pinlabel {alg cross number} [b] at 52 192 \pinlabel {braid index} [l] at 275 31 \pinlabel {$x$} [t] at 268 25 \pinlabel {$y$} [l] at 57 184 \endlabellist \includegraphics[height=40mm]{\figdir/b-region} \caption{The region of braid representatives of ${\cal K}$}\label{b-region} \end{center} \end{figure} The inclusion ``$\supset$'' is trivial by the following argument: Let $K_\star \in B_{\cal K}$ be a minimal braid representative with $\Phi(K_\star)=(b_{\cal K}, c_{\cal K}).$ Suppose $K\in B_{\cal K}$ is obtained from $K_\star$ after applying $(+)$--stabilization $x$--times and then $(-)$--stabilization $y$--times. Then $(b_{\cal K}+x+y, c_{\cal K}+x-y)=\Phi(K) \in \Phi({\cal B}_{\cal K}).$ The MFW--inequality \eqref{lower bound-1} says that $c_K \geq -b_K+(d_++1),\ c_K\leq b_K+(d_--1)$ for any $K\in {\cal B}_{\cal K}.$ Thus \begin{eqnarray}\label{quad-eq} \Phi({\cal B}_{\cal K})\subset \left\{(x,y)\ |\ b_{\cal K}\leq x,\ -x+(d_++1)\leq y \leq x+(d_--1)\right\}. \end{eqnarray} Before we provide examples of the conjectures we present: \begin{theorem}\label{imply} Sharpness of the MFW--inequality implies the truth of Conjectures \ref{Jones-conj} and \ref{greedy}. In particular; $$b_{\cal K}=\frac{d_+-d_-}{2}-1, \quad c_{\cal K}=\frac{d_++d_-}{2}.$$ \end{theorem} We remark that the statement in the theorem with regard to \fullref{Jones-conj} has been well known to many people. \begin{proof}[Proof of \fullref{imply}] Let $K_\star \in {\cal B}_{\cal K}$ be a minimal braid representative. Since the MFW inequality \eqref{lower bound-1} is sharp on ${\cal K},$ we have $c_{K_\star}-b_{\cal K}+1=d_-,$ and $d_+=b_{\cal K}+c_{K_\star}-1,$ ie, $c_{K_\star}=(d_++d_-)/2$ which is independent of the choice of $K_\star.$ Thus we denote $c_{K_\star}=:c_{\cal K}.$ In this case, the right side of \eqref{quad-eq} coincides with the right side of \eqref{quadrant equation} and we have the other inclusion ``$\subset$'' of \eqref{quadrant equation}. \end{proof} \begin{example}\label{ex-of-conj}{\rm Both of the conjectures are true for unlinks, torus links, closed positive braids with a full twist (for example, the Lorenz links) \cite{FW}, $2$--bridge links and alternating fibered links \cite{Murasugi}, where the MFW--inequality is sharp and one can apply \fullref{imply}. Also \fullref{Jones-conj} applies to links with braid index $\leq 3$ \cite{BM3}. However, this case has been settled by a completely different way, the classification of $3$--braids. Namely, any link of braid index $3$ admits a unique conjugacy class of $3$--braid representatives or has at most two conjugacy classes of $3$--braid representatives related to each other by a flype move, which does not change the algebraic crossing number of the link. }\end{example} Every transversal knot $TK$ in $S^3$ with the standard contact structure is transversally isotopic to a transversal closed braid $K$ \cite{Ben}. The {\it Bennequin number} $\beta$ is an invariant of transversal knots. By the identification of $TK$ and $K$, we have $\beta (K) = c_K - b_K.$ If \fullref{greedy} is true for ${\cal K},$ then the {\it maximal} Bennequin number $\beta_{\rm max}({\cal K})$ for the knot type ${\cal K}$ is realized on ${\cal B}_{\cal K} \ni K$'s with plotted vertices $\Phi(K)$ on the upper half boundary of the quadrant region of \fullref{b-region}. Let $\gamma (K) := c_K + b_K.$ For any $K$ and its mirror image $\overline{K},$ we have $\gamma(K)=-\beta(\overline{K}).$ Thus $\gamma_{\rm min}({\cal K}) = - \beta_{\rm max} (\overline{\cal K}).$ See \fullref{b-region}. Thus, investigation of $\beta_{\rm max}({\cal K})$ is related to \fullref{greedy}. \subsection{Cabling and the conjectures} In this subsection, we study behavior of the deficit of the MFW inequality under cabling and prove \fullref{deficit-cable}. As a consequence, we observe that the \fullref{Jones-conj} is true for many of the knots and links that appeared in \fullref{chap2}, where the MFW--inequality is not sharp, ie, we cannot apply \fullref{imply}. \vspace{-2pt} We also prove, in Theorems \ref{cable-thm} and \ref{cable-thm-link}, that the truth of \fullref{Jones-conj} is ``inherited'' through cabling operations. \vspace{-2pt} \begin{figure}[ht!] \begin{center} \labellist\small \pinlabel {$k$--times} [t] at 295 396 \endlabellist \includegraphics [height=40mm]{\figdir/_4_q_-trefoil} \end{center} \caption{$(4,q)$--cable ($q=4 \cdot 3 + k$) of the right hand trefoil} \label{(4,q)-trefoil}% \end{figure} Let us fix some notation. Let ${\cal K}$ be a knot type. Denote the $(p,q)$--cable of ${\cal K}$ by ${\cal K}_{p,q}.$ Let $K$ be a braid representative with $b_K = b_{\cal K}$ and with algebraic crossing number $c_K$. Put $$k:= q - p\cdot c_K$$ and let $K_{p,q}$ denote the $p$--parallel copies of $K$ with a $k/p$--twist (see \fullref{(4,q)-trefoil}). We can assume that $K_{p,q}$ is on the boundary of a tubular neighborhood $N$ of $K$ (thus $K$ is the core of solid torus $N$). Then $$q={\rm lk}(K_{p,q}, K)$$ where `lk' is the linking number. \fullref{(4,q)-trefoil} shows that the algebraic crossing number of $K_{p,q}$ is \begin{equation}\label{c_pq} c_{K_{p,q}}=p^2 c_K + k(p-1)=q(p-1)+p\cdot c_K. \end{equation} Thanks to \cite{W}, we know that the braid index of ${\cal K}_{p,q}$ satisfies \begin{equation}\label{cable-index} b_{{\cal K}_{p,q}}=p\cdot b_{\cal K}. \end{equation} Although one can see a similar result in Theorem 7 of \cite{S}, we state the following for completeness of our discussion: \begin{theorem}\label{deficit-cable} Suppose $K^1$ and $K^2$ are braid representatives of ${\cal K}$ with $b_{K^1}=b_{K^2}=b_{\cal K}$ and with distinct algebraic crossing numbers $c_{K^1} < c_{K^2},$ (ie, \fullref{Jones-conj} does not apply to ${\cal K}).$ Then the deficit $D_{{\cal K}_{p,q}}$ of the MFW--inequality for $(p,q)$--cable ${\cal K}_{p,q}$ is; \begin{equation}\label{eq-deficit-cable} D_{{\cal K}_{p,q}} \geq \frac{p}{2} ( c_{K^2}-c_{K^1} ) \geq p. \end{equation} \end{theorem} \begin{proof}[Proof of \fullref{deficit-cable}] Thanks to \eqref{cable-index}, and by the construction of $({K^1})_{p,q}$ and $({K^2})_{p,q}$, they are both minimal braid representatives of ${\cal K}_{p,q}$ ie, $b_{{K^1}_{p,q}}=b_{{K^2}_{p,q}}=b_{{\cal K}_{p,q}}=p\cdot b_{\cal K}.$ Let $k_1, k_2$ be integers satisfying $q=p c_{K^1} + k_1 = p c_{K^2} + k_2.$ By \eqref{c_pq} we have $ c_{{K^1}_{p,q}} = p^2 c_{K^1} + k_1 (p-1)$ and $c_{{K^2}_{p,q}} = p^2 c_{K^2} + k_2 (p-1).$ Therefore, $c_{{K^2}_{p,q}}-c_{{K^1}_{p,q}} = p(c_{K^2}-c_{K^1}).$ By \eqref{lower bound-1} we have $c_{{K^2}_{p,q}} - b_{{\cal K}_{p,q}} +1 \leq d_{-}\leq d_{+}\leq c_{{K^1}_{p,q}} + b_{{\cal K}_{p,q}} -1,$ and by \fullref{deficit-def}, \begin{equation}\label{D(cable)} D_{{\cal K}_{p,q}} \geq \frac{1}{2} (c_{{K^2}_{p,q}}-c_{{K^1}_{p,q}}) = \frac{p}{2} (c_{K^2}-c_{K^1}). \end{equation} This is the first inequality of \eqref{eq-deficit-cable}. Notice that $K^1$ and $K^2$ are related each other by a sequence of Markov moves \cite{B}. Let $K^1 = B_1 \rightarrow B_2 \rightarrow \cdots \rightarrow B_n=K^2$ be a Markov tower. Each arrow corresponds to either braid isotopy, stabilization or destabilization moves. Let $(x_i, y_i)$ be the braid index and the algebraic crossing number of $B_i.$ Then $(x_{i+1}, y_{i+1})-(x_i, y_i) = (0,0), (\pm 1, \pm 1)$ or $(\mp 1, \pm 1)$ depending on the move corresponding to the arrow between $B_{i+1}$ and $B_i$. Since $x_1 = x_n = b_{\cal K}$ the difference $c_{K^1}-c_{K^2}=y_1-y_n\neq 0$ must be an even integer. Therefore, we get the second inequality of \eqref{eq-deficit-cable}. \end{proof} \begin{corollary}\label{conj-5knots} \fullref{Jones-conj} is true for all $9_{42}, 9_{49}, 10_{132}, 10_{150}, 10_{156}.$ \end{corollary} In \cite{K}, it is proved that \fullref{greedy} also holds for the five knots. \begin{proof} Knotscape computes that the deficit of $2$--cable ${\cal K}_{2,2c_K+1}$ is $1$ for each knot. \begin{center} \begin{tabular}{|l|c|c|c|c|c|} \hline \vrule width0pt height 12pt depth 8pt ${\cal K}$ & $b_{\cal K}$ & $D_{\cal K}$ & $K$ & $c_K$ & $D_{{\cal K}_{2,2c_K+1}}$ \\ \hline \vrule width0pt height 12pt depth 5pt $9_{42}$ & $4$ & $1$ & $aaacBAAcB$ & $1$ & $1$ \\ \hline \vrule width0pt height 12pt depth 5pt $9_{49}$ & $4$ & $1$ & $aabbcbAbbcB$ & $7$ & $1$ \\ \hline \vrule width0pt height 12pt depth 5pt $10_{132}$ & $4$ & $2$ & $AbcaaaBBBcb$ & $3$ & $1$ \\ \hline \vrule width0pt height 12pt depth 5pt $10_{150}$ & $4$ & $1$ & $aabbcbABccB$ & $5$ & $1$ \\ \hline \vrule width0pt height 12pt depth 5pt $10_{156}$ & $4$ & $1$ & $aaacBAAcbAb$ & $3$ & $1$ \\ \hline \end{tabular}\end{center} Comparing with \eqref{eq-deficit-cable}, each ${\cal K}$ must have unique algebraic crossing number. \end{proof} Thanks to Knotscape, the $(2, 2c_{{\cal K}_n}+1)$--cable of ${\cal K}_n = BM_{-1, -2, n, 2}$ has deficit $=1$ if $|n|$ is small. ie, \fullref{Jones-conj} is true for ${\cal K}_n$ if $|n|$ is small. \fullref{conj-5knots} implies: \begin{corollary} \fullref{Jones-conj} is true for the prime links ${\cal A}^n(9_{42})$ (see \fullref{942}$).$ \end{corollary} \begin{proof} We know that $9_{42}$ has unique algebraic crossing number $=1$ by \fullref{conj-5knots}. Since each link component of ${\cal A}^n(9_{42})$ is $9_{42},$ we get this corollary. \end{proof} With regard to the deficit of cabled links, we conjecture that: \begin{conjecture}\label{limit} For any $q$, the limit $\displaystyle\lim_{p \rightarrow \infty}D_{{\cal K}_{p,q}}$ of deficits exists. \end{conjecture} \begin{remark}{\rm If \fullref{limit} is true, then \eqref{eq-deficit-cable} of \fullref{deficit-cable} implies the truth of \fullref{Jones-conj}.} \end{remark} We present another property of cabling: \begin{theorem}\label{cable-thm} Let ${\cal K}$ be a non-trivial knot type. If \fullref{Jones-conj} is true for ${\cal K}$ then it is also true for ${\cal K}_{p,q}$ when $p\geq 2.$ In particular, if $c_{\cal K}$ and $c_{{\cal K}_{p,q}}$ denote the unique algebraic crossing numbers of ${\cal K}$ and ${\cal K}_{p,q}$ respectively in their minimal braid representatives then we have $$c_{{\cal K}_{p,q}} =(p-1)q+p\cdot c_{\cal K}.$$ \end{theorem} \begin{remark} {\rm Suppose ${\cal K}$ is the right hand trefoil. The MFW--inequality is sharp on ${\cal K}.$ Since its cable ${\cal K}_{2,7}$ has deficit $D_{{\cal K}_{2,7}}=1$ (see \cite{MS}), we cannot apply \fullref{imply}. However \fullref{cable-thm} guarantees the truth of the conjecture for ${\cal K}_{2,7}.$} \end{remark} The following proof is inspired by the work of Williams \cite{W}, whose main result can be seen in formula \eqref{cable-index}. Note that his result holds not only for cable knots but also for generalized cable links. For the sake of completeness we repeat part of his discussion. \begin{proof}[Proof of \fullref{cable-thm}] Assume \fullref{Jones-conj} is true for ${\cal K}$ and denote the unique algebraic crossing number at minimal braid index by $c_{\cal K}.$ Let $K$ be a braid representative of ${\cal K}.$ Suppose $K'$ is a braid representative of ${\cal K}_{p,q}$ on the boundary of a small tubular (solid torus) neighborhood $N$ of $K.$ We may regard the $z$--axis as the braid axis. Let $\phi\co {\mathbb R}^3 \rightarrow {\mathbb R}^3$ be a diffeomorphism of compact support so that \begin{itemize} \item[$(\star)$] $\phi(K') \subset \partial\phi(N)$ has exactly $p\cdot b_{\cal K}$ maxima and $p\cdot b_{\cal K}$ minima (both non-degenerate critical points) and no other critical points and \item[$(\star\star)$] the ``height'' function $h\co \partial\phi(N)\simeq T^2 \rightarrow \mathbb{R}$ is a Morse function. \end{itemize} In particular, $\phi(K')$ has a braid position with braid index $p\cdot b_{\cal K}.$ By $(\star\star)$, a generic intersection of the horizontal plane with $\partial\phi(N)\simeq T^2$ consists of disjoint simple closed curves. Furthermore, these simple closed curves are either meridians of $T^2$ or trivial in $T^2$ since ${\cal K}$ is knotted (Remark 1 of \cite{W}). Remark 2 in \cite{W} says that there is a plane $\pi$ (parallel to the $(xz)$--plane) intersecting transversely with $T^2$ in a meridian. Let $J$ be an innermost one among such meridians. Then $J$ bounds a disk $d\subset \pi \cap \phi(N)$ which separates $\phi(K')$ into arcs $\{C_i\}$. Close each $C_i$ with aid of some arc $D_i \subset d$ and set $\hat{K_i}:=C_i \cup D_i.$ See \fullref{cable1}. \begin{figure}[ht!] \begin{center} \labellist\small\hair1.5pt \pinlabel {$C_i$} [t] at 105 551 \pinlabel {$C_i$} [t] at 445 551 \pinlabel {$D_i$} at 186 462 \pinlabel {$d$} at 156 454 \pinlabel {$d$} at 493 454 \pinlabel {$D_i'$} at 516 401 \pinlabel {$A$} [b] at 144 429 \pinlabel {$A$} [b] at 483 429 \pinlabel {$\pi$} [r] at 256 465 \pinlabel {$\pi$} [r] at 596 465 \pinlabel {$J$} [t] at 140 480 \pinlabel {$J$} [t] at 479 480 \pinlabel {$J'$} [b] at 126 372 \pinlabel {$J'$} [b] at 464 372 \pinlabel {$\pi'$} [r] at 248 385 \pinlabel {$\pi'$} [r] at 588 385 \pinlabel {$c_i \cap A$} [t] at 224 356 \endlabellist \includegraphics [height=50mm]{\figdir/cable1} \end{center} \caption{Construction of $K_i$ from $\hat{K_i}$}\label{cable1} \end{figure} Thanks to Remark 3 in \cite{W}, $p$ of $\hat{K_i}$'s are non-trivial (ie, do not bound any disk in $\phi(N))$ since the linking number of $J$ and $\phi(K')$ pushed a little bit into the interior of $\phi(N)$ is $p.$ Discard trivial $\hat{K_i}$'s. Our $\hat{K_i}$'s are not in a braid position. As in \cite{W}, we make them have a braid position: Choose another plane $\pi'$ just below $\pi$ and call the annulus between the two planes $A$ (see \fullref{cable1}). We may assume that the other boundary curve $J'\subset \partial A$ is parallel to $J.$ As in the passage of \fullref{cable1} replace the arc $D_i \cup (C_i \cap A)$ (the left sketch) with $D_i' \subset A$ (the right sketch) and construct $p$--parallels; \begin{equation}\label{K_i} K_i:= (C_i - (C_i \cap A)) \cup D_i' \subset \partial\phi(N)\ \ \mbox{ for } i=1,\cdots,p, \end{equation} which is in a braid position. Also the $K_i$'s are disjoint from each other and each is isotopic to the core of the solid torus $\phi(K)\simeq {\cal K},$ thus $b_{\cal K} \leq b_{K_i}.$ Then we have \begin{eqnarray*} p\cdot b_{\cal K} &\leq & \sum_{i=1}^p \{b_{K_i}= \mbox{number of max of $K_i$}\}\\ &\leq & \{ \mbox{number of max of $\phi(K')$} \} = p\cdot b_{\cal K} \end{eqnarray*} where the last equality holds by $(\star\star)$ above. This implies that \begin{itemize} \item[$(\dag)$] there are no trivial $\hat{K_i}$'s (we didn't have to discard anything), \item[$(\dag\dag)$] each knot has $b_{\cal K} = b_{K_i}.$ \end{itemize} Let $n, 0\leq m <p$ be integers such that $$q= p(c_{\cal K}+n)+m.$$ By $(\dag)$, the $p$--component link $L:= K_1 \cup \cdots \cup K_p$ is obtained from ${\bf K'}:= \phi(K')$ by using the meridian disk $d$ to create a cutout and adding an $m/p$--twist along the annulus $A,$ then gluing the end-points. See \fullref{twists}. In other words, $L$ is the $(p, p(c_{\cal K}+n))$--cable of $K$. From $(\dag\dag)$ we have $c_{K_i}=c_K=c_{\cal K}.$ Therefore, $L$ has the algebraic crossing number \begin{figure}[htpb!] \begin{center}\includegraphics [height=30mm]{\figdir/twists} \end{center} \caption{From $A \cap {\bf K'}$ to $A \cap L$, where $p=7, m=2$}\label{twists} \end{figure} \newpage \begin{eqnarray*} c_L &=& \sum_{i=1}^p c_{K_i} + \sum_{i\neq j} {\rm lk}(K_i, K_j) = p\cdot c_{\cal K} + p(p-1)(c_{\cal K}+n) \\ &=& p^2 c_{\cal K} + p(p-1)n \end{eqnarray*} and $c_{\bf K'}-c_L = m (p-1).$ Thus, \begin{eqnarray*} c_{\bf K'} &=& p^2 c_{\cal K} + p(p-1)n + m (p-1) \\ &=& p^2 c_{\cal K} + (p-1)(pn+m) \\ &=& p^2 c_{\cal K} + (q-p\cdot c_{\cal K}) \\ &=& (p-1)q+p\cdot c_{\cal K}, \end{eqnarray*} which is independent of the choice of ${\bf K'}\in {\cal B}_{{\cal K}_{p,q}}.$ Compare with \eqref{c_pq}. This concludes the uniqueness of the algebraic crossing number of ${\cal K}_{p,q}$ at minimal braid index. \end{proof} A similar result to \fullref{cable-thm} holds for links. \begin{theorem}\label{cable-thm-link} Let ${\cal L}={\cal K}^{(1)}\cup\cdots\cup{\cal K}^{(l)}$ be an $l$--component link of braid index $=b_{\cal L}.$ Assume that each ${\cal K}^{(j)}$ is a non-trivial knot. Let ${\cal L}':={\cal K}^{(1)}_{p,q_1}\cup \cdots\cup{\cal K}^{(l)}_{p,q_l}$ be the $p$--cable of ${\cal L}$ such that $q_j={\rm lk}({\cal K}^{(j)}, {\cal K}^{(j)}_{p,q_j})$ for $j=1, \cdots,l.$ If ${\cal L}$ and every component ${\cal K}^{(j)}$ have unique algebraic crossing numbers $c_{\cal L}, c_{{\cal K}^{(j)}}$ in minimal braid representations, then so does ${\cal L}'.$ Furthermore, let $k_j$ satisfy $q_j=p\cdot c_{{\cal K}^{(j)}}+k_j$ then \begin{equation}\label{cable-thm-link-eq} c_{{\cal L}'}=p^2 c_{\cal L} + (p-1)(k_1+\cdots+k_l). \end{equation} \end{theorem} \begin{remark}\label{flype-remark} {\rm The assumption for ${\cal K}^{(j)}$ in the second paragraph of \fullref{cable-thm-link} is essential by the following reason: If $b_{\cal L}=\sum_{j=1}^lb_{{\cal K}^{(j)}},$ then the existence of unique algebraic crossing number of ${\cal L}$ in minimal braid representation implies that each ${\cal K}^{(j)}$ also has unique algebraic crossing number. However, this is not true in general. For instance, assume that ${\cal L}={\cal K}^{(1)}\cup {\cal K}^{(2)}$ is a $2$--component link and has two braid representatives related to each other by a flype move as in Figure\ref{flype}. \begin{figure}[ht!] \begin{center} \labellist\small \pinlabel {$P$} at 60 661 \pinlabel {$P$} at 296 661 \pinlabel {$Q$} at 60 604 \pinlabel {$Q$} at 296 604 \pinlabel {$R$} <0pt,1pt> at 78 634 \pinlabel \rotatebox{180}{$R$} at 358 629 \pinlabel {${\cal K}^{(1)}$} <0pt,4pt> at 199 689 \pinlabel {${\cal K}^{(2)}$} <0pt,4pt> at 199 668 \pinlabel {flype} [b] at 216 585 \endlabellist \includegraphics [height=40mm]{\figdir/flype} \end{center} \caption{A flype move}\label{flype} \end{figure} The thick gray arcs are parallel braid strands of ${\cal L}.$ Braidings occur inside the boxes $P, Q, R.$ In particular, box $R$ contains even number of half twists of ${\cal K}^{(1)}$ (dashed arc) and ${\cal K}^{(2)}$ (black arc). The flype move preserves the number of braid strands and the algebraic crossing number of the link, but it changes the algebraic crossing numbers of link components ${\cal K}^{(1)}, {\cal K}^{(2)}.$ Namely, in the passage from the left sketch to the right sketch, the algebraic crossing number of ${\cal K}^{(1)}$ decreases by $1$ and the one for ${\cal K}^{(2)}$ increases by $1.$ (This means that a flype move in general cannot be a composition of exchange moves.)} \end{remark} \begin{proof}[Proof of \fullref{cable-thm-link}] Suppose $L=K^{(1)}\cup\cdots\cup K^{(l)}$ is a minimal braid representative of ${\cal L}={\cal K}^{(1)}\cup\cdots\cup{\cal K}^{(l)},$ ie, $b_L=b_{\cal L}.$ Let $c_{x,y}:=2\cdot{\rm lk}({\cal K}^{(x)},{\cal K}^{(y)}) \mbox{ for } x\neq y.$ Then \begin{equation}\label{c_L} c_{\cal L}=c_L=\sum_{1\leq x <y\leq l}c_{x,y} + \sum_{j=1}^lc_{{\cal K}^{(j)}}. \end{equation} Let $k_j, n_j, 0\leq m_j<p$ be integers with \begin{equation}\label{q_j} q_j=p\cdot c_{{\cal K}^{(j)}}+k_j=p(c_{{\cal K}^{(j)}}+n_j)+m_j \quad \mbox{ for } j=1,\cdots,l. \end{equation} Williams proved that the braid index of ${\cal L}'={\cal K}^{(1)}_{p,q_1}\cup \cdots\cup{\cal K}^{(l)}_{p,q_l}$ is $p\cdot b_{\cal L}$ \cite{W}. Let $L'$ be a minimal braid representative of ${\cal L}'.$ Let $N_j$ be a tubular neighborhood of $K^{(j)}.$ Let $\phi$ be a compact support diffeomorphism of $\mathbb{R}^3$ such that $\phi(L')=:{K'}^{(1)}\cup\cdots\cup {K'}^{(l)} \subset \partial\phi(N_1\cup\cdots N_l)$ has a minimal braid position. \newpage As we did in \eqref{K_i}, for each $j=1,\cdots,l,$ construct $p$ parallels $K^{(j)}_1,\cdots,K^{(j)}_p \subset \partial\phi(N_j)$ from each ${K'}^{(j)}$ by cutting out an inner most meridian disk $d_j\subset \phi(N_j)$ and adding an $m_j/p$--twist along annulus $A_j \subset \partial\phi(N_j)$ then gluing. Thus, \begin{equation}\label{lk} {\rm lk}(K_i^{(j)},K^{(h)})= \left\{ \begin{array}{ll} c_{{\cal K}^{(j)}}+n_j & \ \mbox{when } j=h, \\ {\rm lk}(K^{(j)},K^{(h)})= \frac{1}{2}c_{j,h} & \ \mbox{otherwise. } \end{array} \right. \end{equation} Let $$L_i:=K^{(1)}_i \cup\cdots\cup =K^{(l)}_i \ \mbox{ for } i=1,\cdots,p.$$ Thanks to \cite{W} we know that $L_i \simeq L$ and $b_{L_i}=b_L=b_{\cal L}.$ By assumption of \fullref{cable-thm-link}, it follows that $c_{L_i}=c_{L}=c_{\cal L}.$ The $(p\cdot l)$--component link $L_1 \cup\cdots\cup L_p$ has the algebraic crossing number; \begin{eqnarray*} c_{L_1 \cup\cdots\cup L_p} &=& \sum_{i=1}^p c_{L_i} + \sum_{x\neq y}{\rm lk}(L_x, L_y) \\ &=& p\cdot c_{\cal L} +\sum_{i=1}^p(p-1){\rm lk}(L_i,L)\\ &=& p\cdot c_{\cal L} +\sum_{i=1}^p(p-1) \{ \sum_{j=1}^l{\rm lk}(K_i^{(j)},K^{(j)}) +\sum_{x\neq y}{\rm lk}(K_i^{(x)},K^{(y)})\} \\ &\stackrel{\eqref{lk}}{=}& p\cdot c_{\cal L} + p(p-1)\{\sum_{j=1}^l(c_{{\cal K}^{(j)}}+n_j)+\sum_{x<y}c_{x,y}\} \\ & \stackrel{\eqref{c_L}}{=}& p\cdot c_{\cal L} + p(p-1)(c_{\cal L}+\sum_{j=1}^ln_j) \\ &=& p^2 c_{\cal L}+ p(p-1)(\sum_{j=1}^ln_j). \end{eqnarray*} Since only the difference between $L_1 \cup\cdots\cup L_p$ and $L'$ occurs on the annuli $A_1,\cdots,A_l,$ we have \begin{eqnarray*} c_{L'}&=&c_{L_1 \cup\cdots\cup L_p}+(p-1)\sum_{j=1}^lm_j \\ &=& p^2 c_{\cal L}+(p-1)\sum_{j=1}^l(pn_j+m_j) \\ &\stackrel{\eqref{q_j}}{=}& p^2 c_{\cal L}+(p-1)\sum_{j=1}^lk_j, \end{eqnarray*} which is independent of the choice of braid representative $L'$. \end{proof} With regard to \fullref{greedy} we have: \begin{theorem}\label{cable-greedy-conj} Let ${\cal L}={\cal K}^{(1)}\cup\cdots\cup{\cal K}^{(l)}$ be an $l$--component link satisfying all the assumptions in \fullref{cable-thm-link}. If \fullref{greedy} is true for ${\cal L}$ then it is also true for its $p$--cable ${\cal L}':={\cal K}^{(1)}_{p,q_1}\cup\cdots\cup{\cal K}^{(l)}_{p,q_l}.$ \end{theorem} \begin{proof}[Proof of \fullref{cable-greedy-conj}] Let $L$ (resp.\ $L'$) be a braid representative of ${\cal L}$ (resp.\ ${\cal L}'$). Take tubular neighborhoods $N=N_1\cup\cdots\cup N_l$ of $L$ (each $N_j$ is a solid torus) and let $\phi\co \mathbb{R}^3 \to\mathbb{R}^3$ be a compact support diffeo morphism such that $\phi(L)=:K^{(1)}\cup\cdots\cup K^{(l)}$, $\phi(L')=:K'^{(1)}\cup\cdots\cup K'^{(l)}$ have braid positions. They are not necessarily minimal braid representatives and in general $b_{\phi(L')}\neq p\cdot b_{\phi(L).}$ We may assume that $K'^{(j)}\subset\partial\phi(N_j)\simeq T^2.$ Let plane $\pi_j=\{(x,y,z)|y=y_0\},$ innermost meridian loop $J_j\subset \pi_j\cap\partial\phi(N_j),$ and meridian disk $d_j\subset \pi_j\cap\phi(N_j)$ be as in the proof of \fullref{cable-thm}. We may assume that the braid axis is {\em not} contained in $\pi_j$ ie, $y_0\neq 0.$ We deform $K'^{(j)}$ in the following way: Suppose sub-arcs $u\subset K'^{(j)}$ and $v\subset J_j$ bound a disk ${\cal D}\subset\partial\phi(N_j)\simeq T^2$. If ${\cal D}$ is innermost, then replace $u$ with $v$. Repeat this until $K'^{(j)}$ and $J_j$ do not bound any disk in $T^2.$ Add up all the linking numbers of $\partial{\cal D}$'s with the $z$--axis and denote it by $x^{(j)}\geq 0.$ Next, from the deformed $K'^{(j)}$ above, construct $p$--parallels $K'^{(j)}_1,\cdots,K'^{(j)}_p$ as in \eqref{K_i}. Since the plane $\pi_j$ does not contain the $z$--axis, the $m_j/p$--twist along a thin annulus does not change the number of braid strands. Suppose that $b_{K^{(j)}}=b_{{\cal K}^{(j)}}+y^{(j)}$ and $b_{K'^{(j)}_i}=b_{{\cal K}^{(j)}}+y^{(j)}+z^{(j)}_i$ with $y^{(j)},\ y^{(j)}+z^{(j)}_i\geq 0.$ Let $L_i:=K^{(1)}_i \cup\cdots\cup =K^{(l)}_i$ then \begin{equation}\label{b_L_i} b_{L_i} =\sum_{j=1}^l b_{K^{(j)}_i} = b_{\cal L} + \sum_{j=1}^l (y^{(j)}+z^{(j)}_i), \end{equation} \begin{equation}\label{b_Phi(L')} b_{\phi(L')}=\sum_{i=1}^p b_{L_i}+ \sum_{j=1}^lx^{(j)} \stackrel{\eqref{b_L_i}}{=} p\cdot b_{\cal L} + \sum_{j=1}^l(x^{(j)}+p\cdot y^{(j)}+\sum_{i=1}^pz^{(j)}_i). \end{equation} Since $L_i\simeq {\cal L},$ our assumption of this theorem and \eqref{b_L_i} give us \begin{equation}\label{c_L_i} c_{\cal L}-\sum_{j=1}^l (y^{(j)}+z^{(j)}_i) \leq c_{L_i} \leq c_{\cal L}+\sum_{j=1}^l (y^{(j)}+z^{(j)}_i). \end{equation} As in the proof of \fullref{cable-thm-link}, let $k_j, n_j, 0\leq m_j<p$ satisfy $ q_j=p\cdot c_{{\cal K}^{(j)}}+k_j=p(c_{{\cal K}^{(j)}}+n_j)+m_j$. Then we have \begin{eqnarray*} c_{L_1 \cup\cdots\cup L_p} &=& \sum_{i=1}^p c_{L_i} + \sum_{x\neq y}{\rm lk}(L_x, L_y) \\ &\stackrel{\eqref{c_L_i}}{\leq}& p(c_{\cal L}+\sum_{j=1}^l y^{(j)})+ \sum_{i=1}^p \sum_{j=1}^lz^{(j)}_i + p(p-1)(c_{\cal L}+\sum_{j=1}^ln_j) \\ &=& p^2 c_{\cal L}+ p(p-1)(\sum_{j=1}^ln_j) + \sum_{j=1}^l (p\cdot y^{(j)}+ \sum_{i=1}^p z^{(j)}_i), \end{eqnarray*} and \begin{eqnarray} c_{\phi(L')} &\leq& c_{L_1 \cup\cdots\cup L_p} +\sum_{j=1}^l m_j + \sum_{j=1}^lx^{(j)} \nonumber \\ &\leq& p^2 c_{\cal L}+ (p-1)\sum_{j=1}^l k_j + \sum_{j=1}^l(x^{(j)}+p\cdot y^{(j)}+ \sum_{i=1}^pz^{(j)}_i). \label{upper} \end{eqnarray} Similarly, \begin{equation}\label{lower} p^2 c_{\cal L}+ (p-1)\sum_{j=1}^l k_j - \sum_{j=1}^l(x^{(j)}+p\cdot y^{(j)}+\sum_{i=1}^pz^{(j)}_i) \leq c_{\phi(L')}. \end{equation} We conclude the theorem by \eqref{cable-thm-link-eq}, \eqref{b_Phi(L')}, \eqref{upper} and \eqref{lower}. \end{proof} \begin{corollary}\label{iterated-torus} Conjectures \ref{Jones-conj} and \ref{greedy} apply to iterated torus knots. \end{corollary} \begin{proof}[Proof of \fullref{iterated-torus}] We know that the both conjectures apply to torus knots (\fullref{ex-of-conj}). Thanks to Theorems \ref{cable-thm}, \ref{cable-thm-link} and \ref{cable-greedy-conj}, we have this corollary. \end{proof} \subsection{Connect sum and the conjecture} We will prove the following: \begin{theorem}\label{sum} If \fullref{Jones-conj} is true for knot types ${\cal K}^1$ and ${\cal K}^2$ then it is also true for the connect sum ${\cal K}^1\sharp{\cal K}^2$. In particular, denoting the unique algebraic crossing numbers of ${\cal K}^i$ in minimal braid representatives by $c_{{\cal K}^i}$ we have $$c_{{\cal K}^1\sharp {\cal K}^2} = c_{{\cal K}^1} + c_{{\cal K}^2}.$$ \end{theorem} Before we prove \fullref{sum} let us recall two important known results: \begin{lemma}\label{unique-prime}{\rm \cite[Theorem 2.12]{Lickorish}}\qua Up to ordering of summands, there is a unique expression for a knot type ${\cal K}$ as a finite connect sum of prime knots. \end{lemma} \begin{lemma}\label{comp-th} {\rm (\textbf{The composite braid theorem,} \cite{BM4}.)} Let ${\cal K}$ be a composite link, and let $K$ be an arbitrary closed $n$--braid representative of ${\cal K}$. Then there is an obvious composite $n$--braid representative $K^\bullet$ of ${\cal K}$ (see \fullref{obvious-sum}) and a finite sequence of closed $n$--braids: $$K=K_0 \rightarrow K_1 \rightarrow \cdots \rightarrow K_m=K^\bullet$$ such that $K_{i+1}$ is obtained from $K_i$ by either braid isotopy or an exchange move. \end{lemma} \begin{figure}[ht!] \begin{center} \labellist\small \pinlabel {$P$} at 240 347 \pinlabel {$Q$} at 241 241 \endlabellist \includegraphics [height=30mm]{\figdir/obvious-sum} \end{center} \caption{An obvious composite braid} \label{obvious-sum} \end{figure} \begin{proof}[Proof of \fullref{sum}] Since an exchange move does not change the algebraic crossing number, Lemmas \ref{unique-prime} and \ref{comp-th} imply the truth of \fullref{sum}. \end{proof} As a corollary of \fullref{sum} we have: \begin{theorem}\label{sum'} If \fullref{greedy} is true for ${\cal K}^1, {\cal K}^2$ then it is also true for ${\cal K}^1 \sharp {\cal K}^2.$ \end{theorem} \begin{proof}[Proof of \fullref{sum'}] Let $K$ be a braid representative of ${\cal K}^1 \sharp {\cal K}^2.$ By \fullref{comp-th}, after applying exchange moves and braid isotopy to $K$ one can get a composite braid representative $K^\bullet=K^1 \sharp K^2.$ Suppose $b_{K^i}=b_{{\cal K}^i}+x_i$ with $x_i\geq 0.$ Then \begin{eqnarray} b_K &=& b_{K^\bullet}=b_{K^1}+b_{K^2}-1 =(b_{{\cal K}^1}+b_{{\cal K}^2}-1)+(x_1+x_2) \nonumber \\ &=& b_{{\cal K}^1 \sharp {\cal K}^2} +(x_1+x_2). \label{b_K} \end{eqnarray} Our assumption gives $c_{{\cal K}^i}-x_i\leq c_{K^i}\leq c_{{\cal K}^i}+x_i.$ Since $c_K=c_{K^\bullet}=c_{K^1}+c_{K^2},$ we have $(c_{{\cal K}^1}+c_{{\cal K}^2})-(x_1+x_2)\leq c_K \leq (c_{{\cal K}^1}+c_{{\cal K}^2})+(x_1+x_2).$ Thanks to \fullref{sum}, \begin{equation}\label{c_K} c_{{\cal K}^1 \sharp {\cal K}^2}-(x_1+x_2)\leq c_K \leq c_{{\cal K}^1 \sharp {\cal K}^2}+(x_1+x_2). \end{equation} The truth of \fullref{greedy} follows by \eqref{b_K}, \eqref{c_K}. \end{proof} \bibliographystyle{gtart}

https://arxiv.org/abs/2204.07885

On the principal minors of the powers of a matrix

We show that if $A$ is an $n\times n$-matrix, then the diagonal entries of each power $A^{m}$ are uniquely determined by the principal minors of $A$, and can be written as universal (integral) polynomials in the latter. Furthermore, if the latter all equal $1$, then so do the former. These results are inspired by Problem B5 on the Putnam contest 2021, and shed a new light on the behavior of minors under matrix multiplication.

\section{Introduction} Let $R$ be a commutative ring. Let $A$ be an $n\times n$-matrix over $R$, where $n$ is a nonnegative integer. A \emph{principal submatrix} of $A$ means a matrix obtained from $A$ by removing some rows and the corresponding columns (i.e., removing the $i_{1 $-th, $i_{2}$-th, $\ldots$, $i_{k}$-th rows and the $i_{1}$-th, $i_{2}$-th, $\ldots$, $i_{k}$-th columns for some choice of $k$ integers $i_{1 ,i_{2},\ldots,i_{k}$ satisfying $1\leq i_{1}<i_{2}<\cdots<i_{k}\leq n$). In particular, $A$ itself is a principal submatrix of $A$ (obtained for $k=0$). A \emph{principal minor} of $A$ means the determinant of a principal submatrix of $A$. In particular, each diagonal entry of $A$ is a principal minor of $A$ (being the determinant of a principal submatrix of size $1\times1$). In total, $A$ has $2^{n}$ principal minors, including its own determinant $\det A$ as well as the trivial principal minor $1$ (obtained as the determinant of a $0\times0$ matrix, which is what remains when all rows and columns are removed). Problem B5 on the \href{https://kskedlaya.org/putnam-archive/2021.pdf}{Putnam contest 2021} asked for a proof of the following: \begin{theorem} \label{thm.putnam}Assume that $R=\mathbb{Z}$. Assume that each principal minor of $A$ is odd. Then, each principal minor of $A^{m}$ is odd whenever $m$ is a nonnegative integer. \end{theorem} Without giving the solution away, it shall be noticed that essentially only one proof is known (see \cite{Kedlaya} or \cite{Putnam-official-2021} for it), and it is not as algebraic as the statement of Theorem \ref{thm.putnam} might suggest. In particular, it is unclear if the theorem remains valid if \textquotedblleft odd\textquotedblright\ is replaced by \textquotedblleft congruent to $1$ modulo $4$\textquotedblright, or if $R$ is replaced by another ring; the official solution (most of which originates in a result by Dobrinskaya \cite[Lemma 3.3]{MasPan}) certainly does not apply to such extensions. An approach that is definitely doomed is to try expressing the principal minors of a power $A^{m}$ in terms of those of $A$. The following example shows that the latter do not uniquely determine the former: \begin{example} \label{exa.CD}Se \[ C: \begin{pmatrix} a & b & 1 & 1\\ c & d & 1 & 1\\ 1 & 1 & p & q\\ 1 & 1 & r & s \end{pmatrix} \ \ \ \ \ \ \ \ \ \ \text{and}\ \ \ \ \ \ \ \ \ \ D: \begin{pmatrix} a & b & 1 & 1\\ c & d & 1 & 1\\ 1 & 1 & p & r\\ 1 & 1 & q & s \end{pmatrix} \] for some $a,b,c,d,p,q,r,s\in R$. Then, the matrices $C$ and $D$ have the same principal minors, but their squares $C^{2}$ and $D^{2}$ differ in their $\left\{ 2,3\right\} $-principal minor (i.e., their principal minor obtained by removing the $1$-st and $4$-th rows and columns) unless $\left( q-r\right) \left( b-c\right) =0$. Thus, the principal minors of the square of a matrix are not uniquely determined by the principal minors of the matrix itself. \end{example} This example is inspired by \cite[Example 3]{HarLoe}, where further related discussion of matrices with equal principal minors can be found. \section{Nevertheless...} However, not all is lost. Among the principal minors of $A^{m}$, the simplest ones (besides $1$) are those of size $1\times1$, that is, the diagonal entries of $A^{m}$. It turns out that these diagonal entries are indeed uniquely determined by the principal minors of $A$, and even better, they can be written as universal polynomials\footnote{A \textquotedblleft universal polynomial\textquotedblright\ means a polynomial with integer coefficients that depends neither on $A$ nor on $R$ (but can depend on $m$ as well as on the location of the diagonal entry).} in the latter. That is, we have the following:\footnote{An \emph{integer polynomial} means a polynomial with integer coefficients.} \begin{theorem} \label{thm.main1}Let $n$ and $m$ be nonnegative integers, and let $i\in\left\{ 1,2,\ldots,n\right\} $. Then, there exists an integer polynomial $P_{n,i,m}$ in $2^{n}$ indeterminates that is independent of $R$ and $A$, and that has the following property: If $A$ is any $n\times n$-matrix over any commutative ring $R$, then the $i$-th diagonal entry of $A^{m}$ can be obtained by substituting the principal minors of $A$ into $P_{n,i,m}$. In particular, the principal minors of $A$ uniquely determine this entry. \end{theorem} Let us verify this for $m=2$: If we denote the $\left( i,j\right) $-th entry of a matrix $B$ by $B_{i,j}$, then each diagonal entry of $A^{2}$ has the for \begin{align*} \left( A^{2}\right) _{i,i} & =\sum_{j=1}^{n}A_{i,j}A_{j,i}=A_{i,i ^{2}+\sum_{j\neq i}\underbrace{A_{i,j}A_{j,i}}_{=A_{i,i}A_{j,j}-\det\left( \begin{array} [c]{cc A_{i,i} & A_{i,j}\\ A_{j,i} & A_{j,j \end{array} \right) }\\ & =A_{i,i}^{2}+\sum_{j\neq i}\left( A_{i,i}A_{j,j}-\det\left( \begin{array} [c]{cc A_{i,i} & A_{i,j}\\ A_{j,i} & A_{j,j \end{array} \right) \right) , \end{align*} which is visibly an integer polynomial in the principal minors of $A$ (since all the $A_{i,i}$ and $A_{j,j}$ and $\det\left( \begin{array} [c]{cc A_{i,i} & A_{i,j}\\ A_{j,i} & A_{j,j \end{array} \right) $ are principal minors of $A$). This verifies Theorem \ref{thm.main1} for $m=2$. Such explicit computations remain technically possible for higher values of $m$, but become longer and more cumbersome as $m$ increases. The goal of this note is to prove Theorem \ref{thm.main1}. We will first show the following theorem, which looks weaker but is essentially equivalent: \begin{theorem} \label{thm.main2}Let $n$ and $m$ be nonnegative integers. Let $R$ be a commutative ring. Let $A$ be an $n\times n$-matrix over $R$. Let $\mathcal{P}$ be the subring of $R$ generated by all principal minors of $A$. Then, all diagonal entries of $A^{m}$ belong to $\mathcal{P}$. \end{theorem} Before we prove this, let us explain how Theorem \ref{thm.main1} can be easily derived from Theorem \ref{thm.main2}: \begin{proof} [Proof of Theorem \ref{thm.main1} using Theorem \ref{thm.main2}.]The notation $B_{i,j}$ shall denote the $\left( i,j\right) $-th entry of any matrix $B$. Let $\mathbf{R}$ be the polynomial ring $\mathbb{Z}\left[ x_{i,j \ \mid\ 1\leq i\leq n\text{ and }1\leq j\leq n\right] $ in $n^{2}$ independent indeterminates $x_{i,j}$ over $\mathbb{Z}$. (For instance, if $n=2$, then $\mathbf{R}=\mathbb{Z}\left[ x_{1,1},\ x_{1,2},\ x_{2,1 ,\ x_{2,2}\right] $.) Let $\mathbf{A}$ be the $n\times n$-matrix over $\mathbf{R}$ whose $\left( i,j\right) $-th entry is $x_{i,j}$ for each $\left( i,j\right) \in\left\{ 1,2,\ldots,n\right\} ^{2}$. This matrix $\mathbf{A}$ is known as \textquotedblleft the general $n\times n -matrix\textquotedblright, since any matrix $A$ over any commutative ring can be obtained from it by substituting appropriate elements (viz., the entries of $A$) for the variables $x_{i,j}$. This very property will be crucial to the argument that follows. Let $\mathbf{p}_{1},\mathbf{p}_{2},\ldots,\mathbf{p}_{2^{n}}$ be the $2^{n}$ principal minors of $\mathbf{A}$ (numbered in some order). Let $\mathcal{P}$ denote the subring of $\mathbf{R}$ generated by all these principal minors of $\mathbf{A}$. Theorem \ref{thm.main2} (applied to $\mathbf{R}$ and $\mathbf{A}$ instead of $R$ and $A$) shows that all diagonal entries of $\mathbf{A}^{m}$ belong to $\mathcal{P}$. In other words, for each $i\in\left\{ 1,2,\ldots,n\right\} $, we have $\left( \mathbf{A}^{m}\right) _{i,i}\in\mathcal{P}$. Fix $i\in\left\{ 1,2,\ldots,n\right\} $. As we just showed, we have $\left( \mathbf{A}^{m}\right) _{i,i}\in\mathcal{P}$. In other words, there exists an integer polynomial $P_{n,i,m}$ in $2^{n}$ indeterminates such that $\left( \mathbf{A}^{m}\right) _{i,i}=P_{n,i,m}\left( \mathbf{p}_{1},\mathbf{p _{2},\ldots,\mathbf{p}_{2^{n}}\right) $ (since $\mathcal{P}$ is the subring of $\mathbf{R}$ generated by $\mathbf{p}_{1},\mathbf{p}_{2},\ldots ,\mathbf{p}_{2^{n}}$). Consider this polynomial $P_{n,i,m}$; note that it is independent of $R$ and $A$ (by its very construction). Now, consider a commutative ring $R$ and an $n\times n$-matrix $A$ over $R$. Let $p_{1},p_{2},\ldots,p_{2^{n}}$ be the $2^{n}$ principal minors of $A$ (numbered in the same order as $\mathbf{p}_{1},\mathbf{p}_{2},\ldots ,\mathbf{p}_{2^{n}}$). Let $f:\mathbf{R}\rightarrow R$ be the $\mathbb{Z $-algebra homomorphism that sends each indeterminate $x_{i,j}$ to the $\left( i,j\right) $-th entry $A_{i,j}$ of $A$. This homomorphism $f$ therefore sends each entry of the matrix $\mathbf{A}$ to the corresponding entry of $A$, and thus also sends each principal minor of $\mathbf{A}$ to the corresponding principal minor of $A$ (since a principal minor is a certain signed sum of products of entries of the matrix). In other words \begin{equation} f\left( \mathbf{p}_{i}\right) =p_{i}\ \ \ \ \ \ \ \ \ \ \text{for each i\in\left\{ 1,2,\ldots,2^{n}\right\} . \label{pf.thm.main1.1 \end{equation} However, $P_{n,i,m}$ is an integer polynomial, and thus \textquotedblleft commutes\textquotedblright\ with any $\mathbb{Z}$-algebra homomorphism -- i.e., if $a_{1},a_{2},\ldots,a_{2^{n}}$ are any $2^{n}$ elements of a commutative ring, and if $g$ is any $\mathbb{Z}$-algebra homomorphism out of that ring, the \[ g\left( P_{n,i,m}\left( a_{1},a_{2},\ldots,a_{2^{n}}\right) \right) =P_{n,i,m}\left( g\left( a_{1}\right) ,\ g\left( a_{2}\right) ,\ \ldots,\ g\left( a_{2^{n}}\right) \right) . \] Applying this to $a_{i}=\mathbf{p}_{i}$ and $g=f$, we obtai \begin{align*} f\left( P_{n,i,m}\left( \mathbf{p}_{1},\mathbf{p}_{2},\ldots,\mathbf{p _{2^{n}}\right) \right) & =P_{n,i,m}\left( f\left( \mathbf{p _{1}\right) ,\ f\left( \mathbf{p}_{2}\right) ,\ \ldots,\ f\left( \mathbf{p}_{2^{n}}\right) \right) \\ & =P_{n,i,m}\left( p_{1},p_{2},\ldots,p_{2^{n}}\right) \ \ \ \ \ \ \ \ \ \ \left( \text{by (\ref{pf.thm.main1.1})}\right) . \end{align*} In view of $\left( \mathbf{A}^{m}\right) _{i,i}=P_{n,i,m}\left( \mathbf{p}_{1},\mathbf{p}_{2},\ldots,\mathbf{p}_{2^{n}}\right) $, we can rewrite this a \begin{equation} f\left( \left( \mathbf{A}^{m}\right) _{i,i}\right) =P_{n,i,m}\left( p_{1},p_{2},\ldots,p_{2^{n}}\right) . \label{pf.thm.main1.2 \end{equation} However, the $\mathbb{Z}$-algebra homomorphism $f$ sends each entry of the matrix $\mathbf{A}$ to the corresponding entry of $A$, and therefore also sends each entry of the matrix $\mathbf{A}^{m}$ to the corresponding entry of $A^{m}$ (since the entries of $\mathbf{A}^{m}$ are certain sums of products of entries of $\mathbf{A}$, whereas the entries of $A^{m}$ are the same sums of products of entries of $A$). In other words, $f\left( \left( \mathbf{A ^{m}\right) _{u,v}\right) =\left( A^{m}\right) _{u,v}$ for any $u,v\in\left\{ 1,2,\ldots,n\right\} $. Thus, in particular, $f\left( \left( \mathbf{A}^{m}\right) _{i,i}\right) =\left( A^{m}\right) _{i,i}$. Comparing this with (\ref{pf.thm.main1.2}), we obtain $\left( A^{m}\right) _{i,i}=P_{n,i,m}\left( p_{1},p_{2},\ldots,p_{2^{n}}\right) $. In other words, the $i$-th diagonal entry of $A^{m}$ can be obtained by substituting the principal minors of $A$ into $P_{n,i,m}$ (since $p_{1},p_{2 ,\ldots,p_{2^{n}}$ are these principal minors of $A$). This proves Theorem \ref{thm.main1}. \end{proof} \section{Notations} In order to prove Theorem \ref{thm.main2}, we will need some more notations regarding matrices and their minors: \begin{itemize} \item If $m\in\mathbb{Z}$, then $\left[ m\right] $ shall denote the set $\left\{ 1,2,\ldots,m\right\} $. \item If $B$ is a $u\times v$-matrix and if $i\in\left[ u\right] $ and $j\in\left[ v\right] $, then $B_{i,j}$ shall denote the $\left( i,j\right) $-th entry of $B$. \item If $u$ and $v$ are two nonnegative integers, and if $a_{i,j}$ is an element of a ring for each $i\in\left[ u\right] $ and $j\in\left[ v\right] $, then the notation $\left( a_{i,j}\right) _{1\leq i\leq u,\ 1\leq j\leq v}$ means the $u\times v$-matrix whose $\left( i,j\right) $-th entry is $a_{i,j}$ for all $i\in\left[ u\right] $ and $j\in\left[ v\right] $. \item If $B$ is a $u\times v$-matrix, and if $\left( i_{1},i_{2},\ldots ,i_{p}\right) \in\left[ u\right] ^{p}$ and $\left( j_{1},j_{2 ,\ldots,j_{q}\right) \in\left[ v\right] ^{q}$ are two sequences of integers, then $\operatorname*{sub}\nolimits_{i_{1},i_{2},\ldots,i_{p} ^{j_{1},j_{2},\ldots,j_{q}}B$ shall denote the $p\times q$-matrix $\left( B_{i_{x},j_{y}}\right) _{1\leq x\leq p,\ 1\leq y\leq q}$. If $i_{1 <i_{2}<\cdots<i_{p}$ and $j_{1}<j_{2}<\cdots<j_{q}$, then this matrix is a submatrix of $B$. \item If $B$ is a $u\times v$-matrix, and if $I$ is a subset of $\left[ u\right] $, and if $J$ is a subset of $\left[ v\right] $, then $\operatorname*{sub}\nolimits_{I}^{J}B$ shall denote the submatrix $\operatorname*{sub}\nolimits_{i_{1},i_{2},\ldots,i_{p}}^{j_{1},j_{2 ,\ldots,j_{q}}B$ of $B$, where $i_{1},i_{2},\ldots,i_{p}$ are the elements of $I$ in increasing order, and where $j_{1},j_{2},\ldots,j_{q}$ are the elements of $J$ in increasing order. Thus, in particular, if $B$ is an $n\times n$-matrix, and if $I$ is a subset of $\left[ n\right] $, then $\operatorname*{sub}\nolimits_{I}^{I}B$ is a principal submatrix of $B$, so that $\det\left( \operatorname*{sub \nolimits_{I}^{I}B\right) $ is a principal minor of $B$. \item If $B$ is an $n\times n$-matrix, and if $i,j\in\left[ n\right] $, then $B_{\sim i,\sim j}$ shall denote the submatrix of $B$ obtained by removing the $i$-th row and the $j$-th column from $B$. In other words, $B_{\sim i,\sim j}$ denotes the matrix $\operatorname*{sub}\nolimits_{\left[ n\right] \setminus\left\{ i\right\} }^{\left[ n\right] \setminus\left\{ j\right\} }B$. \item If $B$ is an $n\times n$-matrix, then $\operatorname*{adj}B$ shall mean the \emph{adjugate matrix} of $B$. This is defined as the $n\times n$-matrix $\left( \left( -1\right) ^{i+j}\det\left( B_{\sim j,\sim i}\right) \right) _{1\leq i\leq n,\ 1\leq j\leq n}$. \item If $m$ is a nonnegative integer, then $I_{m}$ denotes the $m\times m$ identity matrix. \end{itemize} We will need the following properties of determinants: \begin{itemize} \item For any $n\times n$-matrix $B$, we hav \begin{equation} B\cdot\left( \operatorname*{adj}B\right) =\left( \operatorname*{adj B\right) \cdot B=\left( \det B\right) \cdot I_{n}. \label{eq.adj.BadjB \end{equation} (This is the main property of adjugates; see, e.g., \cite[Theorem 6.100]{detnotes} for a proof.) \item For any commutative ring $S$, any $m\times m$-matrix $B$ and any element $x\in S$, we hav \begin{equation} \det\left( B+xI_{m}\right) =\sum_{P\subseteq\left[ m\right] }\det\left( \operatorname*{sub}\nolimits_{P}^{P}B\right) \cdot x^{m-\left\vert P\right\vert }. \label{eq.det(B+xI)0 \end{equation} (This is a folklore result -- essentially the explicit formula for the characteristic polynomial of a matrix in terms of its principal minors. The proof is straightforward: Expand the left hand side into a sum of products, and combine products according to \textquotedblleft which factors come from $B$ and which factors come from $xI_{m}$\textquotedblright. See \cite[Proposition 6.4.29]{21s} or \cite[Corollary 6.164]{detnotes} for detailed proofs.) \end{itemize} For any ring $S$, we consider the univariate polynomial ring $S\left[ t\right] $ as well as the ring $S\left[ \left[ t\right] \right] $ of formal power series. Of course, $S\left[ t\right] $ is a subring of $S\left[ \left[ t\right] \right] $. Note that the ring $S$ needs not be commutative for $S\left[ t\right] $ and $S\left[ \left[ t\right] \right] $ to be defined. \section{Proof of Theorem \ref{thm.main2}} We now finally step to the proof of Theorem \ref{thm.main2}. \begin{proof} [Proof of Theorem \ref{thm.main2}.]We must prove that all diagonal entries of $A^{m}$ belong to $\mathcal{P}$. In other words, we must prove that $\left( A^{m}\right) _{i,i}\in\mathcal{P}$ for each $i\in\left[ n\right] $. It is well-known that a polynomial over a matrix ring is \textquotedblleft essentially the same as\textquotedblright\ a matrix with polynomial entries. In other words, we can identify the ring $R^{n\times n}\left[ t\right] $ with the ring $\left( R\left[ t\right] \right) ^{n\times n}$ using a straightforward ring isomorphism (which sends each $\sum_{i\geq0}C_{i}t^{i}\in R^{n\times n}\left[ t\right] $ to $\sum_{i\geq0}C_{i}t^{i}\in\left( R\left[ t\right] \right) ^{n\times n}$). In the same way, we identify the ring $R^{n\times n}\left[ \left[ t\right] \right] $ with the ring $\left( R\left[ \left[ t\right] \right] \right) ^{n\times n}$. Let $B$ be the matrix $I_{n}-tA$ in the power series ring $R^{n\times n}\left[ \left[ t\right] \right] $. This matrix $B=I_{n}-tA$ is invertible, and its inverse i \begin{equation} B^{-1}=I_{n}+tA+t^{2}A^{2}+t^{3}A^{3}+\cdots. \label{eq.In-tA.inv \end{equation} (This can be proved by directly verifying that $I_{n}+tA+t^{2}A^{2}+t^{3 A^{3}+\cdots$ is inverse to $I_{n}-tA$. Indeed, both products $\left( I_{n}+tA+t^{2}A^{2}+t^{3}A^{3}+\cdots\right) \cdot\left( I_{n}-tA\right) $ and $\left( I_{n}-tA\right) \cdot\left( I_{n}+tA+t^{2}A^{2}+t^{3 A^{3}+\cdots\right) $ turn, upon expanding, into sums that telescope to $I_{n}$.) Since the matrix $B$ is invertible, its determinant $\det B$ is invertible as well (since $\left( \det B\right) \cdot\left( \det\left( B^{-1}\right) \right) =\det\left( \underbrace{BB^{-1}}_{=I_{n}}\right) =\det\left( I_{n}\right) =1$). Now, recall that we must prove that $\left( A^{m}\right) _{i,i \in\mathcal{P}$ for each $i\in\left[ n\right] $. So let us fix $i\in\left[ n\right] $. Then, (\ref{eq.In-tA.inv}) yield \begin{align*} \left( B^{-1}\right) _{i,i} & =\left( I_{n}+tA+t^{2}A^{2}+t^{3 A^{3}+\cdots\right) _{i,i}\\ & =\left( I_{n}\right) _{i,i}+tA_{i,i}+t^{2}\left( A^{2}\right) _{i,i}+t^{3}\left( A^{3}\right) _{i,i}+\cdots. \end{align*} Hence, the $t^{m}$-coefficient of the power series $\left( B^{-1}\right) _{i,i}\in R\left[ \left[ t\right] \right] $ is $\left( A^{m}\right) _{i,i}$. Thus, in order to prove that $\left( A^{m}\right) _{i,i \in\mathcal{P}$ (which is our goal), it suffices to show that all coefficients of the power series $\left( B^{-1}\right) _{i,i}$ belong to $\mathcal{P}$. In other words, it suffices to show that $\left( B^{-1}\right) _{i,i \in\mathcal{P}\left[ \left[ t\right] \right] $. This is what we shall now show. From (\ref{eq.adj.BadjB}), we obtain $B\cdot\left( \operatorname*{adj B\right) =\left( \operatorname*{adj}B\right) \cdot B=\left( \det B\right) \cdot I_{n}$, so that \[ B^{-1}=\dfrac{1}{\det B}\cdot\operatorname*{adj}B. \] Hence \begin{equation} \left( B^{-1}\right) _{i,i}=\dfrac{1}{\det B}\cdot\left( \operatorname*{adj}B\right) _{i,i}. \label{sol.detB-1ii= \end{equation} Our next goal is to show that both factors $\dfrac{1}{\det B}$ and $\left( \operatorname*{adj}B\right) _{i,i}$ on the right hand side of this equality belong to $\mathcal{P}\left[ \left[ t\right] \right] $. This will then entail that $\left( B^{-1}\right) _{i,i}\in\mathcal{P}\left[ \left[ t\right] \right] $ as well, and we will be done. From $B=I_{n}-tA=-tA+1I_{n}$, we obtain \begin{align} \det B & =\det\left( -tA+1I_{n}\right) =\sum_{P\subseteq\left[ n\right] }\det\left( \underbrace{\operatorname*{sub}\nolimits_{P}^{P}\left( -tA\right) }_{=-t\operatorname*{sub}\nolimits_{P}^{P}A}\right) \cdot\underbrace{1^{n-\left\vert P\right\vert }}_{=1}\nonumber\\ & \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left( \begin{array} [c]{c \text{by (\ref{eq.det(B+xI)0}), applied to }n\text{, }R\left[ \left[ t\right] \right] \text{, }-tA\text{ and }1\\ \text{instead of }m\text{, }S\text{, }B\text{ and }x \end{array} \right) \nonumber\\ & =\sum_{P\subseteq\left[ n\right] }\underbrace{\det\left( -t\operatorname*{sub}\nolimits_{P}^{P}A\right) }_{=\left( -t\right) ^{\left\vert P\right\vert }\det\left( \operatorname*{sub}\nolimits_{P ^{P}A\right) }\nonumber\\ & =\sum_{P\subseteq\left[ n\right] }\left( -t\right) ^{\left\vert P\right\vert }\underbrace{\det\left( \operatorname*{sub}\nolimits_{P ^{P}A\right) }_{\substack{\in\mathcal{P}\\\text{(since }\det\left( \operatorname*{sub}\nolimits_{P}^{P}A\right) \text{ is}\\\text{a principal minor of }A\text{)}}}\label{sol.detB=.1}\\ & \in\mathcal{P}\left[ t\right] \subseteq\mathcal{P}\left[ \left[ t\right] \right] .\nonumber \end{align} Thus, $\det B$ is a formal power series over $\mathcal{P}$. Moreover, (\ref{sol.detB=.1}) shows that this power series has constant term $1$ (since the only addend in the sum in (\ref{sol.detB=.1}) that contributes to the constant term is the addend for $P=\varnothing$, but this addend is $\underbrace{\left( -t\right) ^{\left\vert \varnothing\right\vert } _{=1}\underbrace{\det\left( \operatorname*{sub}\nolimits_{\varnothing }^{\varnothing}A\right) }_{\substack{=1\\\text{(since the }0\times 0\text{-matrix}\\\text{has determinant }1\text{)}}}=1$). Thus, this power series is invertible in $\mathcal{P}\left[ \left[ t\right] \right] $. Therefore \begin{equation} \dfrac{1}{\det B}\in\mathcal{P}\left[ \left[ t\right] \right] . \label{sol.1/detB-in \end{equation} Now, recall the definition of an adjugate matrix. This definition yield \begin{align} \left( \operatorname*{adj}B\right) _{i,i} & =\underbrace{\left( -1\right) ^{i+i}}_{=1}\det\left( B_{\sim i,\sim i}\right) =\det\left( B_{\sim i,\sim i}\right) \nonumber\\ & =\det\left( \left( -tA+1I_{n}\right) _{\sim i,\sim i}\right) \ \ \ \ \ \ \ \ \ \ \left( \text{since }B=-tA+1I_{n}\right) \nonumber\\ & =\det\left( -tA_{\sim i,\sim i}+1I_{n-1}\right) \ \ \ \ \ \ \ \ \ \ \left( \text{since }\left( -tA+1I_{n}\right) _{\sim i,\sim i}=-tA_{\sim i,\sim i}+1I_{n-1}\right) \nonumber\\ & =\sum_{P\subseteq\left[ n-1\right] }\det\left( \underbrace{\operatorname*{sub}\nolimits_{P}^{P}\left( -tA_{\sim i,\sim i}\right) }_{=-t\operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) }\right) \cdot\underbrace{1^{n-1-\left\vert P\right\vert } _{=1}\nonumber\\ & \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left( \begin{array} [c]{c \text{by (\ref{eq.det(B+xI)0}), applied to }n-1\text{, }R\left[ \left[ t\right] \right] \text{, }-tA_{\sim i,\sim i}\text{ and }1\\ \text{instead of }m\text{, }S\text{, }B\text{ and }x \end{array} \right) \nonumber\\ & =\sum_{P\subseteq\left[ n-1\right] }\underbrace{\det\left( -t\operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) \right) }_{=\left( -t\right) ^{\left\vert P\right\vert }\det\left( \operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) \right) }\nonumber\\ & =\sum_{P\subseteq\left[ n-1\right] }\left( -t\right) ^{\left\vert P\right\vert }\det\left( \operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) \right) . \label{sol.adjBii=.1 \end{align} Now, let $P$ be an arbitrary subset of $\left[ n-1\right] $. Write this subset $P$ in the form $P=\left\{ p_{1},p_{2},\ldots,p_{r}\right\} $, where $p_{1}<p_{2}<\cdots<p_{r}$. Furthermore, let $g\in\left\{ 0,1,\ldots ,r\right\} $ be the element that satisfie \[ p_{1}<p_{2}<\cdots<p_{g}<i\leq p_{g+1}<p_{g+2}<\cdots<p_{r}. \] (Here, $g$ will be $0$ if all elements of $P$ are $\geq i$, and $g$ will be $r$ if all elements of $P$ are $<i$.) Then, due to the combinatorial nature of removing rows and columns, we hav \[ \operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) =\operatorname*{sub}\nolimits_{P^{\prime}}^{P^{\prime}}A, \] where $P^{\prime}$ is the subset $\left\{ p_{1},p_{2},\ldots,p_{g}\right\} \cup\left\{ p_{g+1}+1,p_{g+2}+1,\ldots,p_{r}+1\right\} $ of $\left[ n\right] $. Hence, $\operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) $ is a principal submatrix of $A$. Therefore, its determinant $\det\left( \operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) \right) $ is a principal minor of $A$, thus belongs to $\mathcal{P}$. Forget that we fixed $P$. We thus have shown that $\det\left( \operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) \right) \in\mathcal{P}$ for each $P\subseteq\left[ n-1\right] $. Therefore, (\ref{sol.adjBii=.1}) become \begin{equation} \left( \operatorname*{adj}B\right) _{i,i}=\sum_{P\subseteq\left[ n-1\right] }\left( -t\right) ^{\left\vert P\right\vert }\underbrace{\det \left( \operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) \right) }_{\in\mathcal{P}}\in\mathcal{P}\left[ t\right] \subseteq \mathcal{P}\left[ \left[ t\right] \right] . \label{sol.adjBii-in \end{equation} Now, (\ref{sol.detB-1ii=}) become \[ \left( B^{-1}\right) _{i,i}=\underbrace{\dfrac{1}{\det B}}_{\substack{\in \mathcal{P}\left[ \left[ t\right] \right] \\\text{(by (\ref{sol.1/detB-in ))}}}\cdot\underbrace{\left( \operatorname*{adj}B\right) _{i,i }_{\substack{\in\mathcal{P}\left[ \left[ t\right] \right] \\\text{(by (\ref{sol.adjBii-in}))}}}\in\mathcal{P}\left[ \left[ t\right] \right] \cdot\mathcal{P}\left[ \left[ t\right] \right] \subseteq\mathcal{P}\left[ \left[ t\right] \right] . \] As explained above, this completes our proof of Theorem \ref{thm.main2}. \end{proof} Somewhat regrettably, the above proof is the slickest I am aware of. A more-or-less equivalent proof can be given avoiding the use of power series (using \cite[Proposition 3.9 and Lemma 3.11]{trach} instead). A more pedestrian (but harder to formalize) proof uses the Cayley--Hamilton theorem and a variant of the inclusion/exclusion principle. \section{Variants} A counterpart of Theorem \ref{thm.main2} for the off-diagonal entries of $A^{m}$ exists as well: \begin{theorem} \label{thm.main2off}Let $n$, $m$, $R$, $A$ and $\mathcal{P}$ be as in Theorem \ref{thm.main2}. Let $i$ and $j$ be two distinct elements of $\left[ n\right] $. An $\left( i,j\right) $\emph{-quasiprincipal minor} of $A$ shall mean a determinant of the form $\det\left( \operatorname*{sub}\nolimits_{I}^{J}A\right) $, where $I$ and $J$ are two subsets of $\left[ n\right] $ satisfyin \[ i\in I\text{ and }j\in J\text{ and }\left\vert I\right\vert =\left\vert J\right\vert \text{ and }J=\left( I\setminus\left\{ i\right\} \right) \cup\left\{ j\right\} . \] (For instance, if $n\geq7$, then $\det\left( \operatorname*{sub \nolimits_{\left\{ 1,2,7\right\} }^{\left\{ 2,5,7\right\} }A\right) $ is a $\left( 1,5\right) $-quasiprincipal minor of $A$.) Let $\mathcal{K}_{i,j}$ be the $\mathbb{Z}$-submodule of $R$ spanned by all $\left( i,j\right) $-quasiprincipal minors of $A$. Then, \[ \left( A^{m}\right) _{i,j}\in\mathcal{P}\cdot\mathcal{K}_{i,j}. \] \end{theorem} \begin{proof} [Proof outline.]This is similar to our above proof of Theorem \ref{thm.main2}, but some changes are needed. Most importantly, instead of proving that $\left( \operatorname*{adj}B\right) _{i,i}\in\mathcal{P}\left[ \left[ t\right] \right] $, we now need to show that $\left( \operatorname*{adj B\right) _{i,j}\in\mathcal{K}_{i,j}\left[ \left[ t\right] \right] $ (that is, that all coefficients of the power series $\left( \operatorname*{adj B\right) _{j,i}$ belong to $\mathcal{K}_{i,j}$). To do so, we apply the definition of the adjugate matrix to see that \begin{equation} \left( \operatorname*{adj}B\right) _{i,j}=\left( -1\right) ^{j+i \det\left( B_{\sim j,\sim i}\right) . \label{pf.thm.main2off.1 \end{equation} We can simplify $B_{\sim j,\sim i}$ further to $-tA_{\sim j,\sim i}+\left( I_{n}\right) _{\sim j,\sim i}$ (since $B=I_{n}-tA=-tA+I_{n}$), but unfortunately this is not the same as $-tA_{\sim j,\sim i}+1I_{n-1}$, and thus we can no longer apply (\ref{eq.det(B+xI)0}). Instead, we use a trick: \begin{itemize} \item We define $A^{\prime}$ to be the matrix obtained from $-tA$ by replacing the $j$-th row by $\left( 0,0,\ldots,0,1,0,0,\ldots,0\right) $, where the only entry equal to $1$ is in the $i$-th position. \item We define $I_{n}^{\prime}$ to be the matrix obtained from $I_{n}$ by replacing the $1$ in the $j$-th row by a $0$. \item We define $B^{\prime}$ to be the matrix obtained from $B$ by replacing the $j$-th row by $\left( 0,0,\ldots,0,1,0,0,\ldots,0\right) $, where the only entry equal to $1$ is in the $i$-th position. \end{itemize} Laplace expansion along the $j$-th row shows that \[ \det\left( B^{\prime}\right) =\left( -1\right) ^{j+i}\det\left( \left( B^{\prime}\right) _{\sim j,\sim i}\right) =\left( -1\right) ^{j+i \det\left( B_{\sim j,\sim i}\right) \] (since the matrix $B^{\prime}$ differs from $B$ only in the $j$-th row, and thus we have $\left( B^{\prime}\right) _{\sim j,\sim i}=B_{\sim j,\sim i}$). Comparing this with (\ref{pf.thm.main2off.1}), we fin \begin{equation} \left( \operatorname*{adj}B\right) _{i,j}=\det\left( B^{\prime}\right) . \label{pf.thm.main2off.2 \end{equation} Furthermore, recall that $B=I_{n}-tA=-tA+I_{n}$. Thus, $B^{\prime}=A^{\prime }+I_{n}^{\prime}$ (based on how $A^{\prime}$, $I_{n}^{\prime}$ and $B^{\prime }$ were constructed). On the other hand, the definition of $I_{n}^{\prime}$ shows that $I_{n}^{\prime}$ is a diagonal $n\times n$-matrix with diagonal entries $1,1,\ldots,1,0,1,1,\ldots,1$, where the only diagonal entry equal to $0$ is in the $j$-th position. However, another classical fact about determinants (\cite[Theorem 6.4.26]{21s}, \cite[Corollary 6.162]{detnotes}) shows that if $C$ is any $n\times n$-matrix, and if $D$ is a diagonal $n\times n$-matrix with diagonal entries $d_{1},d_{2},\ldots,d_{n}$, the \[ \det\left( C+D\right) =\sum_{P\subseteq\left[ n\right] }\det\left( \operatorname*{sub}\nolimits_{P}^{P}C\right) \cdot\prod_{k\in\left[ n\right] \setminus P}d_{k}. \] We can apply this to $C=A^{\prime}$ and $D=I_{n}^{\prime}$ and $\left( d_{1},d_{2},\ldots,d_{n}\right) =\underbrace{\left( 1,1,\ldots ,1,0,1,1,\ldots,1\right) }_{\text{the }0\text{ is in the }j\text{-th position}}$, and thus obtai \begin{align*} & \det\left( A^{\prime}+I_{n}^{\prime}\right) \\ & =\sum_{P\subseteq\left[ n\right] }\det\left( \operatorname*{sub \nolimits_{P}^{P}\left( A^{\prime}\right) \right) \cdot\prod_{k\in\left[ n\right] \setminus P \begin{cases} 1, & \text{if }k\neq j;\\ 0, & \text{if }k=j \end{cases} \\ & \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left( \text{since the }k\text{-th diagonal entry of }I_{n}^{\prime}\text{ is \begin{cases} 1, & \text{if }k\neq j;\\ 0, & \text{if }k=j \end{cases} \right) \\ & =\sum_{P\subseteq\left[ n\right] }\det\left( \operatorname*{sub \nolimits_{P}^{P}\left( A^{\prime}\right) \right) \cdo \begin{cases} 1, & \text{if }j\notin\left[ n\right] \setminus P;\\ 0, & \text{if }j\in\left[ n\right] \setminus P \end{cases} \\ & =\sum_{\substack{P\subseteq\left[ n\right] ;\\j\notin\left[ n\right] \setminus P}}\det\left( \operatorname*{sub}\nolimits_{P}^{P}\left( A^{\prime}\right) \right) =\sum_{\substack{P\subseteq\left[ n\right] ;\\j\in P}}\det\left( \operatorname*{sub}\nolimits_{P}^{P}\left( A^{\prime }\right) \right) \\ & =\sum_{\substack{P\subseteq\left[ n\right] ;\\j\in P\text{ and }i\in P}}\det\left( \operatorname*{sub}\nolimits_{P}^{P}\left( A^{\prime}\right) \right) +\sum_{\substack{P\subseteq\left[ n\right] ;\\j\in P\text{ and }i\notin P}}\underbrace{\det\left( \operatorname*{sub}\nolimits_{P ^{P}\left( A^{\prime}\right) \right) }_{\substack{=0\\\text{(since all entries in the }j\text{-th row of }A^{\prime}\\\text{are }0\text{ except for the }i\text{-th entry, and thus}\\\text{the matrix }\operatorname*{sub \nolimits_{P}^{P}\left( A^{\prime}\right) \text{ has a zero row)}}}\\ & =\sum_{\substack{P\subseteq\left[ n\right] ;\\j\in P\text{ and }i\in P}}\underbrace{\det\left( \operatorname*{sub}\nolimits_{P}^{P}\left( A^{\prime}\right) \right) }_{\substack{=\pm\det\left( \operatorname*{sub \nolimits_{P\setminus\left\{ j\right\} }^{P\setminus\left\{ i\right\} }\left( -tA\right) \right) \\\text{(by Laplace expansion along the row}\\\text{that was the }j\text{-th row of }A^{\prime}\text{)} }=\sum_{\substack{P\subseteq\left[ n\right] ;\\j\in P\text{ and }i\in P }\pm\underbrace{\det\left( \operatorname*{sub}\nolimits_{P\setminus\left\{ j\right\} }^{P\setminus\left\{ i\right\} }\left( -tA\right) \right) }_{=\pm t^{\left\vert P\right\vert -1}\det\left( \operatorname*{sub \nolimits_{P\setminus\left\{ j\right\} }^{P\setminus\left\{ i\right\} }A\right) }\\ & =\sum_{\substack{P\subseteq\left[ n\right] ;\\j\in P\text{ and }i\in P}}\pm t^{\left\vert P\right\vert -1}\underbrace{\det\left( \operatorname*{sub}\nolimits_{P\setminus\left\{ j\right\} }^{P\setminus \left\{ i\right\} }A\right) }_{\substack{\in\mathcal{K}_{i,j}\\\text{(since }\det\left( \operatorname*{sub}\nolimits_{P\setminus\left\{ j\right\} }^{P\setminus\left\{ i\right\} }A\right) \text{ is}\\\text{an }\left( i,j\right) \text{-quasiprincipal minor of }A\text{)}}}\in\sum _{\substack{P\subseteq\left[ n\right] ;\\j\in P\text{ and }i\in P}}\pm t^{\left\vert P\right\vert -1}\mathcal{K}_{i,j}\subseteq\mathcal{K _{i,j}\left[ \left[ t\right] \right] . \end{align*} In view of $B^{\prime}=A^{\prime}+I_{n}^{\prime}$, this rewrites as $\det\left( B^{\prime}\right) \in\mathcal{K}_{i,j}\left[ \left[ t\right] \right] $. Hence, (\ref{pf.thm.main2off.2}) becomes $\left( \operatorname*{adj}B\right) _{i,j}=\det\left( B^{\prime}\right) \in\mathcal{K}_{i,j}\left[ \left[ t\right] \right] $. Having showed this, we can finish the proof as we did for Theorem \ref{thm.main2}. \end{proof} Another variant of Theorem \ref{thm.main2} is the following: \begin{theorem} \label{thm.main3}Let $n$ and $m$ be nonnegative integers. Let $R$ be a commutative ring. Let $A$ be an $n\times n$-matrix over $R$. Assume that all principal minors of $A$ equal $1$. Then, all diagonal entries of $A^{m}$ equal $1$. \end{theorem} \begin{proof} Follow the above proof of Theorem \ref{thm.main2}. From (\ref{sol.detB=.1}), we obtai \[ \det B=\sum_{P\subseteq\left[ n\right] }\left( -t\right) ^{\left\vert P\right\vert }\underbrace{\det\left( \operatorname*{sub}\nolimits_{P ^{P}A\right) }_{\substack{=1\\\text{(by assumption, since \operatorname*{sub}\nolimits_{P}^{P}A\\\text{is a principal minor of }A\text{)}}}=\sum_{P\subseteq\left[ n\right] }\left( -t\right) ^{\left\vert P\right\vert }=\left( 1-t\right) ^{n \] (since the binomial formula yields $\left( 1-t\right) ^{n}=\sum_{k=0 ^{n}\dbinom{n}{k}\left( -t\right) ^{k}=\sum_{P\subseteq\left[ n\right] }\left( -t\right) ^{\left\vert P\right\vert }$). Let $i\in\left\{ 1,2,\ldots,n\right\} $. From (\ref{sol.adjBii=.1}), we obtai \[ \left( \operatorname*{adj}B\right) _{i,i}=\sum_{P\subseteq\left[ n-1\right] }\left( -t\right) ^{\left\vert P\right\vert }\underbrace{\det \left( \operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) \right) }_{\substack{=1\\\text{(by assumption,}\\\text{since \operatorname*{sub}\nolimits_{P}^{P}\left( A_{\sim i,\sim i}\right) \\\text{is a principal minor of }A\text{)}}}=\sum_{P\subseteq\left[ n-1\right] }\left( -t\right) ^{\left\vert P\right\vert }=\left( 1-t\right) ^{n-1 \] (again by the binomial formula). Now, (\ref{sol.detB-1ii=}) become \begin{align*} \left( B^{-1}\right) _{i,i} & =\dfrac{1}{\det B}\cdot\left( \operatorname*{adj}B\right) _{i,i}=\underbrace{\left( \operatorname*{adj B\right) _{i,i}}_{=\left( 1-t\right) ^{n-1}}\diagup\underbrace{\left( \det B\right) }_{=\left( 1-t\right) ^{n}}=\left( 1-t\right) ^{n-1 \diagup\left( 1-t\right) ^{n}\\ & =\dfrac{1}{1-t}=1+t+t^{2}+t^{3}+\cdots. \end{align*} Thus, the $t^{m}$-coefficient of the power series $\left( B^{-1}\right) _{i,i}\in R\left[ \left[ t\right] \right] $ is $1$. However, we have already seen that this coefficient is $\left( A^{m}\right) _{i,i}$. Thus, we conclude that $\left( A^{m}\right) _{i,i}=1$. This shows that all diagonal entries of $A^{m}$ equal $1$, so that Theorem \ref{thm.main3} is proved. \end{proof} \section{Back to Putnam 2021} As already mentioned, we do not know whether Theorem \ref{thm.putnam} can be generalized by replacing \textquotedblleft odd\textquotedblright\ by \textquotedblleft congruent to $1$ modulo $4$\textquotedblright. More generally, we are tempted to ask the following: \begin{question} \label{quest.putnam-R}Fix a commutative ring $R$. Let $A$ be an $n\times n$-matrix over $R$. Let $m$ be a nonnegative integer. Assume that each principal minor of $A$ is $1$. Is it true that each principal minor of $A^{m}$ is $1$ as well? \end{question} For $R=\mathbb{Z}/2$, this would yield Theorem \ref{thm.putnam}; the \textquotedblleft congruent to $1$ modulo $4$\textquotedblright\ variant would follow for $R=\mathbb{Z}/4$. Theorem \ref{thm.main3} corresponds to the case when the principal minor of $A^{m}$ is a diagonal entry. The argument from \cite[Lemma 3.3]{MasPan} shows that Question \ref{quest.putnam-R} has a positive answer whenever $R$ is an integral domain; thus, the answer is also positive when $R$ is a product of integral domains. On the other hand, if $R$ can be arbitrary, then the answer to Question \ref{quest.putnam-R} is negative, but the only counterexample we know is when $R$ is a certain quotient ring of a polynomial ring\footnote{Here are the details: Let $R$ be the quotient rin \[ \mathbb{Q}\left[ x,y\right] \diagup\left( x^{3}+y^{3},xy,x^{4},x^{3 y,x^{2}y^{2},xy^{3},y^{4}\right) , \] and let $A:=\left( \begin{array} [c]{cccc 1 & 1 & 0 & 0\\ 0 & 1 & y & x\\ x & 0 & 1 & y\\ y & 0 & x & 1 \end{array} \right) \in R^{4\times4}$. Then, all principal minors of $A$ are $1$, but the principal minor $\det\left( \operatorname*{sub}\nolimits_{\left\{ 2,3\right\} }^{\left\{ 2,3\right\} }\left( A^{2}\right) \right) =1-x^{3}-xy$ is not $1$ since $x^{3}\neq0$ in $R$. Actually, we can replace $\mathbb{Q}$ by any field here (even by $\mathbb{Z}/2$); then, $R$ becomes a finite ring. (But we cannot turn $R$ into $\mathbb{Z}/n$ without changing the construction of $A$.)} (and $n=4$ and $m=2$). The smallest ring $R$ for which the question remains open is $\mathbb{Z}/4$.