Sharp Bounds on Eigenvalues via Spectral Embedding Based on Signless Laplacians
Extracted source text · arXiv 2111.08777
% !Mode:: "TeX:UTF-8"
% 使用 PDFLaTeXify 编译 使用 Biber处理参考文献
% 双面文档
\documentclass[12pt,letter,usenames,dvipsnames,twoside]{article}
\usepackage{fullpage}
%================================
%=========语 言 辅 助===============
%================================
\usepackage[british]{babel}
\usepackage[utf8]{inputenc}%变音符编译
\usepackage[T1]{fontenc}%变音字符复制
%\usepackage{ctex}
\usepackage{indentfirst}
% 源代码定位 如果不小把状态栏的 --src 按没了, 也有可能返不回tex中去.
%================================
%===========有用工具==============
%================================
\usepackage{enumerate}
\usepackage{multirow}
%================================
%=========随 机 文 本===============
%================================
\usepackage{lipsum}
%================================
%==========页面设置==============
%================================
%\setlength{\oddsidemargin}{1 cm}
%\setlength{\evensidemargin}{-0.3 cm}
%\setlength{\textheight}{21.6true cm}
%\setlength{\textwidth}{15true cm}
%\setlength{\topmargin}{-0.4 in}
%\usepackage{geometry} %设置页边距的宏包
%\geometry{left=2.3cm,right=2.3cm,top=2.2cm,bottom=2.2cm} %设置 上、左、下、右 页边距
%================================
%==== 页眉页脚设置 开始 =====
%================================
\usepackage{fancyhdr}
\usepackage{lastpage}
\usepackage{layout}
\pagestyle{fancy}% 设置页眉
\fancyhead{} %clear all fields
\renewcommand{\headrulewidth}{0pt}
%\fancyhead[CE]{}
%\fancyhead[CO]{}
%\fancyhead[RO]{Spectral Embedding} %奇数页眉的右边
%\fancyhead[RE]{}
%\fancyhead[LE]{} %偶数页眉的左边
%\fancyhead[RE]{} %偶数页眉的左边
\fancyfoot[C]{Page \thepage\ of \pageref{LastPage}}
%================================
%==== 页眉页脚设置 结束 =====
%================================
\setlength{\baselineskip}{12pt plus 2pt minus 1pt}
%================================
%=========数 学 工 具===============
%================================
\usepackage{amsmath,amsthm,amsfonts,amssymb,mathrsfs,mathtools}
\usepackage{latexsym}
\usepackage{bm}%加粗的数学符号
%================================
%=========图 片 工 具===============
%================================
%敬请参考 https://vvingerfly.github.io/2018/02-16-Tex-imageTypesetting/
\usepackage{graphicx}
\usepackage{caption}
\usepackage{subfigure}
%================================
%=====效 果 辅 助 宏 包===============
%================================
\usepackage{microtype}
%================================
%=========字 体 设 置===============
%================================
%\usepackage{fontspec}
%\setmainfont{Calibri}
%\setmainfont{Times New Roman}
%\setCJKmainfont{STFangsong}
%\usepackage{upgreek}
%================================
%========== 彩 色 链接 =============
%================================
\usepackage{xcolor}% 支持更多颜色
\usepackage{hyperref}
\hypersetup{colorlinks,%
linkcolor=RoyalBlue,%
citecolor=Violet,
urlcolor=BrickRed,
bookmarksopen=true % 书签选项 打开
}
%================================
%======== 书签和目录 ==============
%================================
\usepackage{bookmark}
\usepackage[nottoc]{tocbibind}
%================================
%====Citation and Reference======
%================================
%\numberwithin{equation}{chapter}
\usepackage{url}
\usepackage{doi}
\usepackage[
style = philosophy-modern,
square = true,
url = false,
isbn = false,
doi = true,
hyperref=true,
natbib = true,
parentracker=true,
shorthandintro = true,
backend=biber]{biblatex}
\usepackage{cleveref}
\crefname{equation}{Eq.\@}{Eqs.\@}
%================================
%====Codes for Showkeys 开始 =====
%================================
%\usepackage{showkeys}
%\makeatletter
%\SK@def\Cref#1{\SK@\SK@@ref{#1}\SK@Cref{#1}}%
%\makeatother
%================================
%==== Codes for Showkeys 结束 ====
%================================
%================================
%====调整 参考文献列表大小写 开始 ====
%================================
\DeclareFieldFormat[article,inbook,incollection,inproceedings,patent,thesis,unpublished]{titlecase}{\MakeSentenceCase*{#1}}
\usepackage{titlecaps}
\Addlcwords{on and for a is but with of in as the etc to if its de l'Institut}
\DeclareFieldFormat[phdthesis,book]{titlecase}{\itshape\titlecap{#1}}
\DeclareFieldFormat{jtu}{\itshape\titlecap{#1}}%jtu stands for "journal title uppercase"
\newbibmacro*{journal}{%
\iffieldundef{journaltitle}
{}
{\printtext[journaltitle]{%
\printfield[jtu]{journaltitle}%
\setunit{\subtitlepunct}%
\printfield[jtu]{journalsubtitle}}}}
\DeclareFieldFormat[inproceedings]{booktitle}{
\printfield[jtu]{booktitle}%
\setunit{\subtitlepunct}%
\printfield[jtu]{booksubtitle}
}
%================================
%====调整 参考文献列表大小写 结束 ====
%================================
%================================
%===调整 参考文献卷期数字写法 结束 ====
%================================
\DeclareFieldFormat[article]{volume}{\bibstring{jourvol}\addnbspace #1}
\DeclareFieldFormat[article]{number}{\bibstring{number}\addnbspace #1}
\renewbibmacro*{volume+number+eid}{%
\printfield{volume}%
\setunit{\addcomma\space}%<---- was \setunit*{\adddot}%
\printfield{number}%
\setunit{\addcomma\space}%
\printfield{eid}}
%================================
%===调整 参考文献卷期数字写法 结束 ====
%================================
\addbibresource{Spectral.bib} %%%%%%%%%%%%%%%%%%%指定包含了参考文献的 .bib文件
%================================
%========TeX Environment=========
%================================
\usepackage{cases}% 用于分段定义函数
\newtheorem{theorem}{Theorem}
\newtheorem{example}[theorem]{Example}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{definition}[theorem]{Definition}
\theoremstyle{theorem}
\newtheorem*{thrm}{Theorem}
\theoremstyle{theorem}
\newtheorem*{lem}{Lemma}
\theoremstyle{theorem}
\newtheorem*{prop}{Proposition}
\theoremstyle{theorem}
\newtheorem*{corr}{Corollary}
\theoremstyle{theorem}
\newtheorem*{moti}{Motivation}
%%%%%%%%%%%%%%%%% manual lemma 开始
\newtheorem{innercustomthm}{Lemma}
\newenvironment{hlm}[1]
{\renewcommand\theinnercustomthm{#1}\innercustomthm}
{\endinnercustomthm}
\crefname{hlm}{Lemma}{Lemmas}
%%%%%%%%%%%%%%%%% manual lemma 结束
\theoremstyle{definition}
\newtheorem*{exmp}{Example}
\theoremstyle{remark}
\newtheorem*{rmk}{Remark}
\theoremstyle{remark}
\newtheorem*{hmk}{Homework}
\renewcommand{\proofname}{Proof}
\usepackage{algorithm}
\usepackage{algorithmic}
\numberwithin{theorem}{section}
%================================
%====== 概率及数学期望符号 ========
%================================
\DeclareMathOperator{\pb}{\mathbf{P}\mathopen{}}
\DeclareMathOperator{\E}{\mathbf{E}\mathopen{}}
\newcommand\Pleft[1]{\pb\mkern-1mu\left[#1\right]}
\newcommand\Psubleft[2]{\Psub {#1}\mkern-1mu\left[#2\right]} \newcommand\Pbig[1]{\pb\mkern-.5mu\bigl[#1\bigr]}
\newcommand\Pbigg[1]{\pb\mkern-1.5mu\biggl[#1\biggr]}
\newcommand\PBig[1]{\pb\mkern-.5mu\Bigl[#1\Bigr]}
\newcommand\Eleft[1]{\E\mkern-2mu\left[#1\right]}
\newcommand\Esubleft[2]{\E_{#1}\mkern-2mu\left[#2\right]}
\newcommand\Ebig[1]{\E\mkern-1.5mu\bigl[#1\bigr]}
\newcommand\Ebigg[1]{\E\mkern-1.5mu\biggl[#1\biggr]}
\newcommand\EBig[1]{\E\mkern-1.5mu\Bigl[#1\Bigr]}
\newcommand\EBigg[1]{\E\mkern-1.5mu\Biggl[#1\Biggr]}
\newcommand\Psub [1]{\pb_{\! #1}}
\newcommand\Psubbig [2]{\Psub {#1}\mkern-1.5mu\bigl[#2\bigr]}
\newcommand\PsubBig [2]{\Psub {#1}\mkern-1.5mu\Bigl[#2\Bigr]}
\newcommand\Psubbigg [2]{\Psub {#1}\mkern-1.5mu\biggl[#2\biggr]}
\newcommand\PsubBigg [2]{\Psub {#1}\mkern-1.5mu\Biggl[#2\Biggr]}
\newcommand\Esubbig [2]{\E_{#1}\mkern-1.5mu\bigl[#2\bigr]}
\newcommand\EsubBig [2]{\E_{#1}\mkern-1.5mu\Bigl[#2\Bigr]}
\newcommand\Esubbigg [2]{\E_{#1}\mkern-1.5mu\biggl[#2\biggr]}
\newcommand\EsubBigg [2]{\E_{#1}\mkern-1.5mu\Biggl[#2\Biggr]}
%======= 条件期望竖线 ==========
\newcommand\mbig{\bigm\vert}
\newcommand\mBig{\Bigm\vert}
\newcommand\mbigg{\biggm\vert}
\newcommand\mBigg{\Biggm\vert}
%================================
%======= 数学定界符 ==========
%================================
\DeclarePairedDelimiter\abs{\lvert}{\rvert} % 绝对值符号
\DeclarePairedDelimiter\norm{\lVert}{\rVert} % 范数符号
%================================
%======= 概率运算符 ==========
%================================
\DeclareMathOperator{\cfc}{\mathrm{Cov}}% 协方差
\DeclareMathOperator{\fc}{\mathrm{Var}} % 方差
\newcommand{\deq}{\mathbin{\overset{\mathscr{D}}{=}}} % equal in distribution
\newcommand{\eqas}{\mathbin{\overset{\text{a.s.}}{=}}}% equal almost surely
\newcommand{\cvgp}{\mathrel{\xrightarrow{\pb}}} % convergence in probability
\newcommand{\cvgw}{\mathrel{\xrightarrow{W}}} % convergence in measure W
\newcommand{\indep}{\mathbin{\rotatebox[origin=c]{90}{$\models$}}} % 独立
\newcommand{\ps}{{\mathscr P}} % 概率族 花写 P
\newcommand{\ip}{\quad \text{in probability}} % 依概率 作为附加文本使用在 display 或者 equation 环境中
%================================
%======= 数学符号 =============
%================================
\newcommand{\ue}{\mathrm{e}} % 自然对数的底数 e
%\newcommand\restrict{\mathord\upharpoonright} % 运算符限制
\newcommand\comp{\mathbin{\scalebox{0.7}{$\overset{\circ}{\vphantom{,}}$}}}% 函数复合运算符
%======= 正文格式 =============
\newcommand\deef[1]{\textit{\textbf{#1}}}% 概念定义
%======= 测度论符号=============
\newcommand{\NS}{\mathscr N} % 零测度集 N
\newcommand{\B}{\mathscr B} % Borel 集合 B
\newcommand{\EE}{\mathscr E} % Sigma 代数 E
\newcommand{\F}{\mathscr F} % Sigma 代数 F
\newcommand{\GG}{\mathscr G} % Sigma 代数 G （\G已经定义了）
%======= 各种集合 ===============
\newcommand{\NN}{\mathbb{N}}
\newcommand{\ZN}{\mathbb{Z}}
\newcommand{\QN}{\mathbb{Q}}
\newcommand{\RNS}{\mathbb{R}}
\newcommand{\CN}{\mathbb{C}}
\newcommand{\I}{\mathbb{I}} %指标集合 I
%======= 复数相关 ===============
\newcommand\imi{\mathrm{i}} % 虚数单位 i
\DeclareMathOperator{\re}{\mathrm{Re}} % 复数实部
\DeclareMathOperator{\im}{\mathrm{Im}} % 复数虚部
%======= 常用集类符号============
\newcommand{\KK}{\mathcal{K}} %集合类 K
\newcommand{\UU}{\mathcal{U}} %集合类 U
%======= 数学推导 ===============
\newcommand\all[1]{\forall #1\enspace} % for all 任意的
\newcommand\texists[1]{\exists #1\enspace} % there exists 存在
\def\implies{\quad\Longrightarrow\quad} % 蕴含
\def\IFF{\quad\Longleftrightarrow\quad} % 当且仅当
\newcommand\st{\, ; \;} % 使得
%======= 集合运算 ===============
\newcommand{\ind}{\boldsymbol{1}} % 集合的示性函数
\newcommand\xor{\mathbin{\triangle}} % 对称差运算
\newcommand\buji[1]{{#1}^{\mathrm{c}}} % 补集符号
\DeclareMathOperator{\img}{img} %映射 Image
%======= 微分符号 d =============
\newcommand{\ud}{\mathrm{d}} % 竖直的 d 一般不使用这个而使用下面的几个定制化方案
\newcommand{\df}[1]{\,\mathrm{d}#1}
\newcommand{\adf}[1]{\,\abs{\mathrm{d} #1}}
\newcommand{\adfbig}[1]{\,\abs[\big]{\mathrm{d} #1}}
\newcommand{\adfBig}[1]{\,\abs[\Big]{\mathrm{d} #1}}
\newcommand{\adfbigg}[1]{\,\abs[\bigg]{\mathrm{d} #1}}
%======= 图论符号 ================
\def\rdiam{{\mathcal R}_{\textup{diam}}} %resistance diameter
\DeclareMathOperator{\D}{diam} %open% 直径
\DeclareMathOperator{\dist}{\mathsf{dist}} %距离 dist
\DeclareMathOperator{\vol}{\mathsf{wt}} % 权重weight
%%%%%%%
%几个算符
\def\cp{{\cal P}}
\def\cl{{\cal L}}
\def\cq{{\cal Q}}
\def\cE{{\cal E}}
\def\cH{{\cal H}}% 希尔伯特空间
\newcommand{\R}{\mathscr R}
%几个算符
%%%%%%%
\def\be{{\bf e}} %单位向量
%%%%%%
%顶点度数
\def\dmx{d_{\mathrm{max}}}
\def\dmi{d_{\mathrm{min}}}
\def\dav{d_{\mathrm{avg}}}
%顶点度数
\def\kg{\mathscr{K}(G)}%%% Kudos
%%%%%%
\def\trel{t_{\mathrm{rel}}} %relaxation time
\def\tiq{\tau_{\infty}(1/4)} %mixing time
%======= 证明结束符号 开始 =============
\newcommand{\xqed}{
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill
\quad\hbox{\ensuremath{\qedsymbol}}}
%======= 证明结束符号 结束 =============
%================================
%====== 两 个 英 文 字 体 =========
%================================
\usepackage{yfonts} %for font \gothfamily
\usepackage{calligra} % for font \calligra
\newcommand{\setfont}[2]{{\fontfamily{#1}\selectfont #2}} %一个有用的命令
%================================
%=========标 题 设 置===============
%================================
\usepackage{metalogo} %生成 XeLaTeX 的 logo
%\usepackage{times}
\usepackage{datetime}
\usepackage[super]{nth}
\newcommand\tstamp{\thanks{Current version: \today. First draft on Thursday \nth{11} November, 2021.}}
\usepackage{titling} % 参考 https://www.latexstudio.net/archives/8062.html
\pretitle{\begin{center}\LARGE\bf}
\posttitle{\end{center}}
\title{Sharp Bounds on Eigenvalues via Spectral Embedding Based on Signless Laplacians\tstamp}
\preauthor{\vskip 1em\begin{center}
\large \lineskip 0.5em%
\begin{tabular}[t]{c}}
\postauthor{\end{tabular}\par\end{center}}
\usepackage{authblk} %作者和单位宏包
\newcommand\corref[2]{\thanks{Email: \href{mailto:#1}{#1}. #2}}
\newcommand\authref[1]{\thanks{#1}}
\renewcommand\Authsep{\,,\;}
\renewcommand\Authand{\, and }
\renewcommand\Authands{\,,\;and }
%\author{Russell Lyons\authref{rdlyons@indiana.edu}}
\author{Zhi-Feng Wei\corref{zfwei@iu.edu}{Supported by NSF grant DMS-1954086.}}
\renewcommand*{\Affilfont}{\normalsize\it}
\affil{Department of Mathematics, 831 E.\@ 3rd St.,\\ Indiana University, Bloomington, IN 47405-7106}
%\author[a,b]{Zhi-Feng Wei}
%\author[b]{Heping Zhang}
%\author[a]{Russell Lyons \thanks{Corresponding author: email@mail.com}}
%\affil[a]{Indiana University}
%\affil[b]{Lanzhou University}
\predate{}
\date{}
\postdate{}
\providecommand{\keywords}[1]{\flushleft\textbf{\textit{Keywords:}} #1} % 关键字命令
\providecommand{\amsc}[1]{\flushleft\textbf{\textit{MSC 2020:}} #1} % AMS 分类
%================================
%=========文 档 主 体===============
%================================
%\renewcommand{\baselinestretch}{2}
\begin{document}
\maketitle\thispagestyle{empty}
\begin{abstract}
Using spectral embedding based on the signless Laplacian, we obtain bounds on the spectrum of transition matrices on graphs. As a consequence, we bound return probabilities and the uniform mixing time of simple random walk on graphs. In addition, spectral embedding is used in this article to bound the spectrum of graph adjacency matrices. Our method is adapted from [\citeauthor{Lyons-2018}, \citeyear{Lyons-2018}].\citereset
\keywords{return probability, spectral embedding, signless Laplacian, spectral gap, eigenvalues,
minimum eigenvalue, regular graphs.}
\amsc{05C81, 60J10, 05C50.}
\end{abstract}
\clearpage
\setcounter{tocdepth}{2}%目录深度
\tableofcontents
\thispagestyle{empty}
\clearpage
\setcounter{page}{1}
%$\uppi$
\input{Sections/1intro}
\input{Sections/2prep}
\input{Sections/3return}
\input{Sections/4Oli}
\input{Sections/5vol}
\input{Sections/6mix}
\input{Sections/7transitive}
\input{Sections/8average}
\input{Sections/9bipartite}
\input{Sections/10combinatorial}
\section*{Acknowledgments}%\addcontentsline{toc}{section}{\numberline {}{Acknowledgement}}
\addcontentsline{toc}{section}{{Acknowledgments}}
The author is greatly indebted to Prof.\ Russell Lyons for many useful discussions and for his careful guidance. This work was supported by the National Science Foundation grant DMS-1954086.
\printshorthands\addcontentsline{toc}{section}{{Abbreviations}}
\printbibliography[heading = bibintoc]
\end{document}
%================================
%==========Section 1=============
%================================
\section{Introduction}
Spectral embedding is a popular tool in modern data clustering, as summarized in \autocite{Luxburg-2007}.
Also, spectral embedding was exploited to study graphs. In particular, \autocite{Lyons-2018} introduced spectral embedding as a new tool in analyzing reversible Markov chains (random walks on graphs).
For instance,
\autocite[Theorem~4.9]{Lyons-2018} gave a sharp bound on return probabilities of lazy random walk. Here, the lazy random walk on a graph $G$ stays put at a vertex with probability $1/2$ and moves to a random uniform neighbor otherwise.
\begin{theorem}[\cite{Lyons-2018}, Theorem~4.9]
Let $G$ be a regular, simple, connected graph. For each vertex $x$ of $G$ and $t>0$,
\begin{displaymath}
0\leqslant p'_t(x,x)-\frac{1}{\abs{V(G)}}\leqslant \frac{13}{\sqrt{t}},
\end{displaymath}
where $p'_t(x,x)$ is the probability of returning to $x$ at step $t$ for the lazy random walk on $G$ starting from $x$.\qed
\end{theorem}
Now, a question arises naturally: if simple random walk, instead of lazy random walk, is considered, do we still have good bounds on return probabilities? Here, the simple random walk on a graph moves to a random neighbor uniformly.
To get some feeling, we recall the method used in the proof of the above bound in \autocite{Lyons-2018}. The transition matrix $P'$ of lazy random walk on a graph $G$ is a symmetric operator on the space of functions on $V(G)$ square summable with respect to the degree sequence of $G$; the spectrum of $P'$ is real and contained in the interval $[0,1]$. Also, it is well known that the non-trivial (different from $1$) spectrum of $P'$ ``governs'' the ``convergence rate'' of lazy random walk: if the gap between the largest non-trivial eigenvalue of $P'$ and $1$ is large, then intuitively, the convergence will be faster. Furthermore, the spectrum of $P'$ is closely related to the spectrum of the probabilistic Laplacian $\cl' \coloneqq I-P'$ corresponding to $P'$. To prove the aforementioned bound, using the spectral embedding based on $\cl'$, \autocite{Lyons-2018} first revealed upper bounds on the vertex spectral measure of $\cl'$.
It is known that return probabilities are determined by vertex spectral measures; therefore, bounds on return probabilities can be obtained from bounds on vertex spectral measures (see \autocite[Proposition~3.5]{Lyons-2018}).
When it comes to simple random walk, the situation is different. Note that simple random walks on bipartite graphs have period two and simple random walks on non-bipartite graphs are aperiodic. For simplicity of presentation, we discuss non-bipartite case only in this introduction. To begin, the spectrum of $P$, the transition matrix of the simple random walk on a non-bipartite graph $G$, is real and contained in the interval $[-1,1]$, but it is not necessarily non-negative. We still have the intuition that the ``convergence rate'' of simple random walk is related to the non-trivial spectrum of $P$: for instance, \autocite[Proposition~3]{Diaconis-1991} proved that the geometric convergence rate in total variation norm is determined by the maximum non-trivial eigenvalue \emph{in absolute value} when the random walk has finitely many states. Therefore, in order to deal with ``convergence'' of simple random walk, one also has to consider the negative spectrum of $P$. However, we usually study the spectral gap of the probabilistic Laplacian $\cl \coloneqq I-P$, which does not provide much information about the negative spectrum of $P$.
Our solution is to consider another operator: the probabilistic signless Laplacian operator $\cq\coloneqq I+P$. Obviously, the spectrum of $P$ is closely related to the spectrum of $\cq$ by a shift of 1 unit horizontally. The spectrum of $\cq$ is therefore real and non-negative. This brings us some convenience: we have more tools to deal with the operator $\cq$, its associated quadratic form for a start. %In order to estimate the minimal non-trivial eigenvalue of $P$, one only need to consider the minimal non-zero eigenvalue of $\cq$.
We will first consider the vertex spectral measure of $\cq$ using the spectral embedding based on $\cq$, so the negative spectrum of $P$ is bounded. Then we may proceed to
%will still try to exploit comparison tools, modified from \autocite[Proposition~3.5]{Lyons-2018}), to
get a bound on return probabilities of simple random walk. See \Cref{reg_return} for details.
\subsection{Results}\label{ssec_results}
To give an overview of our results, in this subsection, we constrain ourselves to the case of unweighted graphs. Some notation will be needed, which will be explained in more detail in subsequent sections. Consider a locally finite, simple, connected graph $G=(V,E)$. A vertex $x\in V$ has degree $d(x)$ in $G$. When $G$ is finite, we denote the eigenvalues of the transition matrix $P$ on $G$ as
\begin{displaymath}
-1 \leqslant \lambda^P_{\mathrm{min}} = \lambda^P_1
\leqslant \lambda^P_2 \leqslant \lambda^P_3 \leqslant \dotsb \leqslant \lambda^P_{n-1}
<\lambda^P_n = \lambda^P_{\mathrm{max}} = 1,
\end{displaymath}
where $n=\abs{V}$.
It is well known that $\lambda^P_{\mathrm{min}}=-1$ if and only if $G$ is bipartite.
However, for infinite graphs, it is more appropriate to consider their vertex spectral measures. For instance, denoting the probability of returning to $x$ at step $t$ as $p_t(x,x)$ for the simple random walk on $G$ starting at $x$, then we have
\begin{displaymath}
p_t(x,x)=\int_{[0,2]}(1-\lambda)^t\ud\mu_x(\lambda),
\end{displaymath}
where $\mu_x$ is the vertex spectral measure at $x$ of $\cl$. Denote the vertex spectral measure at $x$ of $\cq$ as $\mu_x^\cq$. It is shown in \Cref{spec_loc} that when $G$ is finite,
\begin{displaymath}
\sum_{x\in V(G)}\mu_x^\cq(\delta)=\abs[\big]{\{j\st 1+\lambda_j^P\leqslant \delta\}}.
\end{displaymath}
This enables us to count eigenvalues of $P$ on the interval $[-1,-1+\delta]$. Therefore, we can get lower bounds on eigenvalues of $P$ from upper bounds of the vertex spectral measure of $\cq$. Write $\pi(x)\coloneqq \frac{d(x)}{2\abs{E}}$, which is $0$ when $G$ is an infinite graph.
We first consider non-bipartite graphs. In fact, when the graph is non-bipartite, the simple random walk is aperiodic, so some troubles are avoided. Our result for simple random walk on regular graphs reads as follows.
\begin{theorem}\label{reg_intro}
For a regular, non-bipartite, simple, connected graph $G$, we have
\begin{displaymath}
\mu_x^\cq(\delta)\leqslant 10\sqrt{\delta},\qquad 0\leqslant \delta\leqslant 2,\,x\in V.
\end{displaymath}
For each $x\in V$, simple random walk on $G$ satisfies
\begin{displaymath}\begin{split}
0\leqslant p_t(x,x)-\pi(x)&\leqslant \frac{18}{\sqrt{t}} \qquad \text{for\quad} t\equiv 0\bmod 2,\\
\abs[\big]{p_t(x,x)-\pi(x)}&\leqslant \frac{9}{\sqrt{t}} \qquad \text{for\quad} t\equiv 1\bmod 2.
\end{split}\end{displaymath}
Furthermore, when $G$ is finite, for $1\leqslant k\leqslant n$, $\lambda^P_k\geqslant -1+\frac{k^2}{100n^2}$.
\end{theorem}
The above result is sharp by \Cref{sharp} and its proof is presented in \Cref{reg}.
\Cref{reg_intro} is interesting since the degree and size of the regular graph $G$ are not involved.
If we consider general finite graphs and introduce the relaxation time $\trel$, we have the following bound on return probabilities.
\begin{theorem}
For $t\geqslant t_{\mathrm{rel}}$, simple random walk on a non-bipartite, finite, simple, connected graph $G$ satisfies
\begin{displaymath}\begin{split}
0\leqslant p_{2t}(x,x)-\pi(x)&\leqslant \frac{8d(x)\sqrt{\trel+1}}{(t+1)\dmi},\\
\abs[\big]{p_{2t+1}(x,x)-\pi(x)}&\leqslant \frac{8d(x)\sqrt{\trel+1}}{\sqrt{(t+1)(t+2)}\,\dmi},
\end{split}\end{displaymath}
where $\dmi$ is the minimum degree of $G$ and\/ $\trel\coloneqq \bigl(1-\abs{\lambda^P_1}\vee\abs{\lambda^P_{n-1}}\bigr)^{-1}$.
\end{theorem}
For lazy random walk, a similar result was obtained in \autocite{Oliveira-2019}. We use a similar method as theirs in \Cref{sec_Oli} to prove the above theorem.
The following proposition is for graphs satisfying volume growth conditions.
\begin{proposition}
Let $G$ be a non-bipartite, infinite, simple, connected graph. Suppose that $c > 0$ and $D \geqslant 1$ are constants such that\/ $\vol(x, r) \geqslant c (r+1)^D$ for all
$r \geqslant 0$. Then for all $\delta\in(0,2)$,
\begin{displaymath}\begin{split}
\mu^*_x(\delta) &\leqslant C w(x) \delta^{D/(D+1)},\\
\mu^\cq _x(\delta) &\leqslant C w(x) \delta^{D/(D+1)},
\end{split}\end{displaymath}
where
\begin{displaymath}
C \coloneqq \frac{(D+1)^2}{c^{1/(D+1)} D^{2D/(D+1)}}.
\end{displaymath}
Hence for all $t \geqslant 1$, simple random walk satisfies
\begin{displaymath}\begin{split}
p_t(x,x)
&\leqslant
2C' w(x) t^{-D/(D+1)}\qquad \text{for\quad} t\equiv 0\bmod 2,\\
p_t(x,x)
&\leqslant
C' w(x) t^{-D/(D+1)}
\qquad \text{for\quad} t\equiv 1\bmod 2,
\end{split}\end{displaymath}
where
\begin{displaymath}
C' \coloneqq \frac{D+1}{c^{1/(D+1)} D^{(D-1)/(D+1)}} \Gamma\Big(\frac{D}{D+1}\Big).
\end{displaymath}
\end{proposition}
Using the spectral embedding based on $\cq$, we also bound the vertex spectral measure of $\cq$ in another way and get bounds on the uniform mixing time.
\begin{proposition}\label{mix_intro}
For any non-bipartite, finite, simple, connected graph $G$, we have
\begin{displaymath}
\mu_x^\cq(\delta) \leqslant \frac{d(x)\delta}{\kg},\qquad \delta\in[0,2),\,x\in V,
\end{displaymath}
where\/ $\kg$ is defined as
\begin{displaymath}
\min\biggl\{
\frac{\sum_{(v,u)\in E(G)}\abs{f(v)+f(u)}^2}{\max_{y\in V}\abs{f(y)}^2}
\st
\min_{y\in V}f(y)<0< \max_{y\in V}f(y)
\biggr\}
\end{displaymath}
and satisfies $\kg\geqslant \frac{1}{\D(G)+1}$. Consequently, for\/ $1\leqslant k\leqslant n$,
\begin{displaymath}
\lambda^P_k
\geqslant
-1+\frac{k\kg}{\sum_{x\in V}d(x)}
\geqslant
-1+\frac{k}{\bigl(\D(G)+1\bigr)\sum_{x\in V}d(x)}.
\end{displaymath}
Furthermore, $\tiq\leqslant 8n^3$. If $G$ is also regular, then we have $\tiq\leqslant 24n^2$.
\end{proposition}
Note that for a finite, simple, connected graph $G$, \autocite[Proposition~4.2]{Lyons-2018} implies
\begin{displaymath}
\lambda^P_{n-k}\leqslant 1-\frac{k}{\rdiam(G)\sum_{x\in V}d(x)},\qquad 0\leqslant k\leqslant n-1,
\end{displaymath}
where $\rdiam(G)$ is the resistance diameter of $G$.
This bound combined with the lower bound on eigenvalues of $P$ in \Cref{mix_intro} improves \autocite{Landau-1981}, which asserted that each eigenvalue $\lambda$ of $P$ that is neither $1$ nor $-1$ satisfies
\begin{displaymath}
\abs{\lambda}\leqslant 1-\frac{1}{\bigl(\D(G)+1\bigr)n\dmx},
\end{displaymath}
where $\dmx$ is the maximum degree of $G$.
The bound $\tiq= O(n^2)$ in \Cref{mix_intro} is sharp by the example of cycles: there is a constant $c>0$ such that for all odd number $n$, simple random walk on
an $n$-cycle satisfies $\tiq\geqslant cn^2$ (cf.\@ \autocite[Example~3.11]{Tetali-2005}). See \Cref{sec_mix} for more details.
As a special class of regular graphs, vertex-transitive graphs are of interest, since they are intuitively ``homogeneous'' and especially well studied.
\begin{theorem}
Let $G$ be a non-bipartite, simple, connected, vertex-transitive graph with degree $d<\infty$. For each $x\in V$, $c\in(0,1)$, and $\delta\in (0,2]$, we have
\begin{displaymath}
\mu^\cq_x(\delta)\leqslant
\frac{1}{c^2N^\#\Bigl(\frac{\sqrt{1-c}}{\sqrt{d\delta}}\Bigr)},
\end{displaymath}
where $N^\#(r)$ denotes the number of vertices in a ball of radius $r$. In addition, if $G$ is finite,
\begin{displaymath}
\lambda^\cq_{\mathrm{min}}
\geqslant
\frac{2}{d}\Bigl(\sin\frac{\pi}{4\bigl(\D(G)+1\bigr)}\Bigr)^2.
\end{displaymath}
\end{theorem}
As usual, a bound on return probabilities of simple random walk follows immediately from a bound on vertex spectral measures. See \Cref{trans_return} as an example.
Apart from vertex spectral measure at one single vertex, we may consider average spectral measure for finite graphs. Given a finite graph $G$, the average spectral measure of $\cq$ is defined as $\mu^\cq\coloneqq \sum_{x\in V}\mu_x^\cq/n$.
\begin{theorem}\label{avg_intro}
For any non-bipartite, finite, simple, connected graph $G$ and
$\delta\in(0,2)$, we have
\begin{displaymath}
\mu^\cq(\delta)<(4000\delta)^{1/3}.
\end{displaymath}
Consequently, $\lambda^P_k\geqslant -1+\frac{k^3}{4000n^3}$ for $1\leqslant k\leqslant n$.
Furthermore,
\begin{displaymath}\begin{split}
0\leqslant
\frac{\sum_{x\in V(G)}p_t(x,x)-1}{n}
&\leqslant
\frac{30}{t^{1/3}} \qquad \text{for\quad} t\equiv 0\bmod 2,\\
\frac{\abs[\big]{\sum_{x\in V(G)}p_t(x,x)-1}}{n}
&\leqslant
\frac{15}{t^{1/3}} \qquad \text{for\quad} t\equiv 1\bmod 2.
\end{split}\end{displaymath}
\end{theorem}
\Cref{avg_intro} is treated in \Cref{sec_avg}.
Non-bipartite graphs are considered in previous paragraphs. But the bipartite case is a bit different, because simple random walk on a bipartite graph has period two. Fortunately, by \autocite[Theorem~4.8]{Mohar-1989}, the vertex spectral measures of $\cl$ and $\cq$ coincide. Therefore, only the estimates for the vertex spectral measure of $\cl$ from \autocite{Lyons-2018} will be enough for us to get bounds on return probabilities of simple random walk. For example, we have the following result for bipartite graphs in \Cref{sec_bip}.
\begin{theorem}
Consider simple random walk on a regular, bipartite, simple, connected graph $G$. Then for each $x\in V$,
\begin{displaymath}
0\leqslant p_t(x,x)-2\pi(x)\leqslant \frac{18}{\sqrt{t}} \qquad \text{for\quad} t\equiv 0\bmod 2.
\end{displaymath}
\end{theorem}
Our method work as well for analyzing the spectrum of the adjacency matrix $A$ of a finite graph $G$. Suppose the eigenvalues of $A$ are
\begin{displaymath}
-d_{\mathrm{max}} \leqslant \lambda^A_{\mathrm{min}} = \lambda^A_1
\leqslant \lambda^A_2 \leqslant \lambda^A_3 \leqslant \dotsb \leqslant \lambda^A_{n-1} < \lambda^A_n= \lambda^A_{\mathrm{max}} \leqslant d_{\mathrm{max}},
\end{displaymath}
where $\dmx$ is the maximum degree of $G$.
\begin{proposition}\label{Alon_ext_intro}
Let $G$ be a non-bipartite, finite, simple, connected graph. For\/ $1\leqslant k\leqslant n$,
\begin{displaymath}
\dmx+\lambda_k^A
\geqslant
\frac{k\kg}{n}
\geqslant
\frac{k}{\bigl(\D(G)+1\bigr)n}.
\end{displaymath}
\end{proposition}
\Cref{Alon_ext_intro} improves \autocite[Theorem~1.1]{Alon-2000}, which obtained that $\dmx+\lambda_1^A \geqslant \frac{1}{(\D(G)+1)n}$. To get \Cref{Alon_ext_intro}, in \Cref{sec_comb}, we first consider the spectral embedding based on the combinatorial signless Laplacian $\Theta\coloneqq D+A$, where $D$ is the diagonal degree matrix of $G$. Then \Cref{Alon_ext_intro} follows from the bound below on the vertex spectral measure of $\Theta$.
%by a linear algebra argument.
\begin{proposition}
Let $G$ be a non-bipartite, finite, simple, connected graph. Then for each
$\delta\in [0,\lambda^\Theta_{\mathrm{max}})$ and $x\in V$, we have
\begin{displaymath}
\mu_x^\Theta(\delta)
\leqslant
\frac{\delta}{\kg}
\leqslant
\bigl(\D(G)+1\bigr)\delta.
\end{displaymath}
\end{proposition}
\subsection{Related Works}
In the field of spectral graph theory, the combinatorial signless Laplacian has already drawn wide attention: \autocite{Cvetkovic-2009,Cvetkovic-2010a,Cvetkovic-2010b} are surveys on the study of the combinatorial signless Laplacian. In fact, \autocite[Theorem~1.1]{Alon-2000} also used the combinatorial signless Laplacian implicitly by considering its associated quadratic form. We use the method of spectral embedding, so not only the minimum eigenvalue but all eigenvalues of the graph adjacency matrices are bounded from below in \Cref{Alon_ext_intro}.
One highlight of this article is the introduction of the probabilistic signless Laplacian $\cq$, enabling us to deal with the negative spectrum of the transition matrix $P$ on a graph. Indeed, the quadratic form associated to $\cq$ was used in \autocite{Desai-1993,Diaconis-1991} to give lower bounds on the minimum eigenvalue of $P$. We consider the operator $\cq$ explicitly and exploit the spectral embedding based on it. Therefore, the entire spectrum of $P$ is treated, rather than only the minimum eigenvalue. For instance, \autocite{Landau-1981} is improved as we discussed after \Cref{mix_intro}.
In the literature, there are many fewer results on the probabilistic signless Laplacian $\cq$ than on the probabilistic Laplacian $\cl$. For instance, tools from electrical network theory are pretty useful in the study of $\cl$, but they are not readily available to deal with $\cq$. We have to make a ``detour'' and adapt the existing tools.
There has been extensive study of return probability bounds of random walks, both on finite and infinite graphs. The asymptotics of large-time return probabilities correspond to the asymptotics of the spectral measures of $\cl$ and $\cq$ near $0$, as suggested in \Cref{ssec_prel}. Therefore, our bounds on the vertex spectral measure of $\cq$, combined with the bounds on the vertex spectral measure of $\cl$ from \autocite{Lyons-2018}, will result in return probability bounds for simple random walks, which are parallel to the bounds for lazy random walks in \autocite{Lyons-2018}.
\subsection{Structure of this Article}
We review notation for graphs and introduce spectral embedding based on the signless Laplacian in
\Cref{ssec_notation,ssec_se}. Some preliminaries are included in \Cref{ssec_prel}.
Return probabilities of simple random walk on regular graphs are considered in \Cref{sec_reg}.
A bound on return probabilities involving the relaxation time is presented in \Cref{sec_Oli}. Return probability bounds based on volume growth conditions are discussed in \Cref{sec_vol}. The case of transitive graphs is treated in \Cref{sec_trans}. We also consider average return probabilities for finite graphs in \Cref{sec_avg}. In \Cref{sec_mix}, we bound the uniform mixing time. Bipartite graphs are discussed in \Cref{sec_bip}. The tool of spectral embedding is exploited to study eigenvalues of graph adjacency matrices in \Cref{sec_comb}.
%================================
%==========Section 2=============
%================================
\section{Graph Notation and Spectral Embedding}\label{sec_prep}
\subsection{Graph Notation and Laplacian Operators}\label{ssec_notation}
Let $G=\bigl(V(G),E(G)\bigr)$ be a finite or infinite, undirected, simple, connected, weighted graph. For an edge of $G$, say $e=(x,y)\in E(G)$, let $w(e)=w(x,y)>0$ be its weight. We say $G$ is \emph{unweighted} if $w(e)=1$ for each edge $e\in E(G)$.
We assume $G$ has weighted adjacency matrix $A(G)$, (unweighted) diameter $\D(G)$, and (weighted) resistance diameter $\rdiam(G)$. Also, minimum, maximum, and average degrees in $G$ are denoted by $\dmi(G)$, $\dmx(G)$, and $\dav(G)$, respectively.
When $G$ is finite, we denote $\abs[\big]{V(G)}=n(G)$. If $G$ is understood, reference to $G$ may be omitted.
For $x\in V(G)$, we use standard graph notation $N(x)\coloneqq \bigl\{y\in V(G)\st (x,y)\in E(G)\bigr\}$ to denote the collection of all neighbors of $x$. Throughout this article, we require that $G$ is \emph{locally finite}, i.e., $\sum_{y\in N(x)}w(x,y)<\infty$ for each $x\in V(G)$. We say that ${w(x)\coloneqq \sum_{y\in N(x)}w(x,y)}$ is the weight of $x\in V(G)$ in $G$, and the weight of a vertex subset $S\subseteq V$ in $G$ is ${\vol(S;G)\coloneqq \sum_{x\in S}w(x)}$. If $G$ is unweighted, $w(x)$ equals the degree $d(x)$ of $x$. Again, when $G$ is understood, the reference to $G$ may be omitted. In particular, when $G$ is vertex-transitive, all vertices have the same weight, denoted by $w$. Write $\pi(x)= \frac{w(x)}{\vol(V(G))}$. For a vertex $x\in V(G)$ and $r\geqslant0$, let
\begin{displaymath}
B(x,r;G)=\bigl\{y\in V(G)\st \dist(x,y)\leqslant r\},
\end{displaymath}
where $\dist$ is the distance on $G$. Set
\begin{displaymath}
\vol(x,r;G)
\coloneqq
\vol\bigl(B(x,r;G);G\bigr)
%=
%\sum_{y\colon \dist(x,y)\leqslant r}w(y).
\end{displaymath}
We have a series of useful operators: the probability transition matrix $P$ on $G$, the probabilistic Laplacian $\cl = I-P$, and the probabilistic signless Laplacian $\cq=I+P$.
We know that $P$, $\cl$, and $\cq$ are bounded self-adjoint operators on Hilbert space $\ell^2(V, w)$. Recall that we write
$\ell^2(V, w)$ for the (real or complex) Hilbert space of functions $f \colon V \to \RNS \text{ or } \CN$ with inner product
\begin{displaymath}
\langle f,g \rangle_w \coloneqq \sum_{x \in V} w(x) f(x)\overline{g(x)}
\end{displaymath}
and squared norm $\norm{f}_w^2 \coloneqq \langle f,f\rangle_w$.
As in \autocite{Lyons-2018}, we reserve $\langle \cdot , \cdot \rangle$ and $\|\cdot\|$ for the standard inner product and norm on $\RNS^k\,(k \in \NN)$ and $\ell^2(V)$, and also for the norm on a generic Hilbert space.
For a vertex $x\in V$, we use $\ind_x$ to denote the indicator vector of $x$:
\begin{numcases}
{\ind_x(y)\coloneqq }
1
\quad \nonumber &if\/ $y=x$,\\
0
\quad \nonumber &otherwise.
\end{numcases}
We also write $\be_x \coloneqq \frac{\ind_x}{\sqrt{w(x)}}$. Note that $\be_x\in \ell^2(V, w)$ is of unit norm: $\norm{\be_x}_w=1$.
Obviously, whether $G$ is finite or infinite, the spectrum of $P$ and $\cq$ are related by a shift of 1 unit horizontally.
Denote the resolution of identity for $\cl$ as $I_\cl$.
As in \autocite{Lyons-2018}, the vertex spectral measure of $\cl$ at $x\in V$ is defined by
\begin{displaymath}
\mu_x(\delta)
\coloneqq \bigl\langle I_\cl\bigl([0,\delta]\bigr)\be_x,\be_x\bigr\rangle_w
=\bigl\langle I_\cl\bigl([0,\delta]\bigr)\ind_x,\ind_x\bigr\rangle,
\qquad 0\leqslant \delta\leqslant 2.
\end{displaymath}
For convenience, we will also use
\begin{displaymath}
\mu_x^*(\delta)
\coloneqq \bigl\langle I_\cl\bigl((0,\delta]\bigr)\be_x,\be_x\bigr\rangle_w
=\bigl\langle I_\cl\bigl((0,\delta]\bigr)\ind_x,\ind_x\bigr\rangle,
\qquad 0\leqslant \delta\leqslant 2,\,x\in V.
\end{displaymath}
It is easy to see that $\mu_x(\delta)=\mu_x^*(\delta)+\pi(x)$ for $0\leqslant \delta\leqslant 2$
(cf.\ \autocite[Section~3.1]{Lyons-2018}).
When $G$ is finite, the spectra of $P$ and $\cq$ consist of eigenvalues (point spectrum) only:
%In this case, we may list the eigenvalues in an increasing order:
\begin{itemize}
\item Denote the eigenvalues of the transition matrix $P$ on $G$ as
\begin{displaymath}
-1 \leqslant \lambda^P_{\mathrm{min}} = \lambda^P_1
\leqslant \lambda^P_2 \leqslant \lambda^P_3 \leqslant \dotsb \leqslant \lambda^P_{n-1}
<\lambda^P_n = \lambda^P_{\mathrm{max}} = 1.
\end{displaymath}
\item Denote the eigenvalues of the probabilistic signless Laplacian $\cq$ as
\begin{displaymath}
0\leqslant \lambda^\cq_{\mathrm{min}} = \lambda^\cq_1
\leqslant \lambda^\cq_2 \leqslant \lambda^\cq_3 \leqslant \dotsb \leqslant \lambda^\cq_{n-1}
< \lambda^\cq_n = \lambda^\cq_{\mathrm{max}} = 2.
\end{displaymath}
\end{itemize}
\subsection{Spectral Embedding Based on the Signless Laplacian}\label{ssec_se}
The spectral embedding based on $\cl$ is introduced in \autocite[Section~3.4]{Lyons-2018} as a powerful tool in analyzing random walk on graphs. In this subsection, we will introduce the spectral embedding based on $\cq$ in parallel. Denote the resolution of identity for $\cq$ as $I_\cq$, with vertex spectral measure at $x$
\begin{displaymath}
\mu_x^\cq(\delta)\coloneqq \bigl\langle I_\cq(\delta)\be_x,\be_x\bigr\rangle_w,\qquad 0\leqslant \delta\leqslant 2,\,x\in V,
\end{displaymath}
where $I_\cq(\delta)\coloneqq I_\cq\bigl([0,\delta]\bigr)$.
\begin{lemma}\label{q_form}
Let $f\in \ell^2(V, w)$. We have
\begin{displaymath}
\langle\cq f,f\rangle_w=\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2.
\end{displaymath}
For $\delta\in[0,2]$ and $f\in\img\bigl(I_\cq(\delta)\bigr)$,
\begin{displaymath}
\langle\cq f,f\rangle_w\leqslant \delta\norm{f}_w^2=\delta\sum_{v\in V}w(v)\abs[\big]{f(v)}^2.
\end{displaymath}
\end{lemma}
\begin{proof}
We need prove the first assertion only. First, we claim that
\begin{displaymath}
\sum_{v,u\in V}w(v,u)\abs{f(v)}\abs{f(v)+f(u)}<\infty.
\end{displaymath}
In fact, we have
\begin{displaymath}\begin{split}
\sum_{v,u\in V}w(v,u)\abs{f(v)}\abs{f(v)+f(u)}
&\leqslant \sum_{v,u\in V}w(v,u)\bigl(\abs{f(v)}^2+\abs{f(v)}\abs{f(u)}\bigr)\\
&=\sum_{v,u\in V}w(v,u)\abs{f(v)}^2 + \sum_{v,u\in V}w(v,u)\abs{f(v)}\abs{f(u)}.
\end{split}\end{displaymath}
By the Cauchy--Schwarz inequality, we may proceed and get
\begin{displaymath}\begin{split}
\sum_{v,u\in V}w(v,u)\abs{f(v)}\abs{f(v)+f(u)}&\leqslant
\sum_{v,u\in V}w(v,u)\abs{f(v)}^2\\
&{\qquad}+ \Bigl(\sum_{v,u\in V}w(v,u)\abs{f(v)}^2\Bigr)^{1/2}\Bigl(\sum_{v,u\in V}w(v,u)\abs{f(u)}^2\Bigr)^{1/2}\\
&= \sum_{v,u\in V}w(v,u)\abs{f(v)}^2 + \sum_{v,u\in V}w(v,u)\abs{f(v)}^2\\
&= 2\sum_{v,u\in V}w(v,u)\abs{f(v)}^2
= 2\sum_{v\in V}\abs{f(v)}^2\sum_{u\in V}w(v,u)\\
&= 2\sum_{v\in V}\abs{f(v)}^2w(v)=2\norm{f}_w<\infty.
\end{split}\end{displaymath}
The claim is proved. By this claim, using Fubini's theorem, we have
\begin{displaymath}\begin{split}
\sum_{v,u\in V}w(v,u)f(v)\overline{\bigl(f(v)+f(u)\bigr)}
&=\sum_{v\in V}f(v)\overline{\sum_{u\in V}w(v,u)\bigl(f(v)+f(u)\bigr)}\\
&=\sum_{v\in V}w(v)f(v)\overline{\bigl(f(v)+\sum_{u\in V}p(v,u)f(u)\bigr)}\\
&=\sum_{v\in V}w(v)f(v)\overline{\bigr((I+P)f\bigl)(v)}=\langle f,(I+P)f\rangle_w \\
&=\langle f,\cq f\rangle_w.
\end{split}\end{displaymath}
By interchanging $u$ and $v$, we have
\begin{displaymath}
\sum_{v,u\in V}w(v,u)f(u)\overline{\bigl(f(v)+f(u)\bigr)}
=\langle f,\cq f\rangle_w.
\end{displaymath}
Therefore,
\begin{displaymath}\begin{split}
\sum_{(v,u)\in E(G)}&w(v,u)\abs[\big]{f(v)+f(u)}^2
=
\frac{1}{2}\sum_{v,u\in V}w(v,u)\abs[\big]{f(v)+f(u)}^2\\
&=\frac{1}{2}\Bigl(
\sum_{v,u\in V}w(v,u)f(v)\overline{\bigl(f(v)+f(u)\bigr)}
+\sum_{v,u\in V}w(v,u)f(u)\overline{\bigl(f(v)+f(u)\bigr)}
\Bigr)\\
&= \tfrac{1}{2}\bigl(\langle f,\cq f\rangle_w+\langle f,\cq f\rangle_w\bigr)\\
&=\langle f,\cq f\rangle_w.\qedhere
\end{split}\end{displaymath}
\end{proof}
\begin{corollary}\label{0span}
If $G$ is connected and non-bipartite, then\/ $\img\bigl(I_\cq(\{0\})\bigr)$ contains only the zero function. Therefore in this case, $I_\cq(\delta)=I_\cq\bigl([0,\delta]\bigr)=I_\cq\bigl((0,\delta]\bigr)$. It also follows that\/ $0$ is not an eigenvalue of $\cq$ if $G$ is non-bipartite.
\end{corollary}
\begin{proof}
Assume $f\in\img\bigl(I_\cq(\{0\})\bigr)$. We see that
\begin{displaymath}
\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2
=\langle\cq f,f\rangle_w
\leqslant 0\norm{f}_w^2=0.
\end{displaymath}
Therefore, for $(v,u)\in E(G)$, we have $f(v)=-f(u)$.
Since $G$ is non-bipartite, there is an odd cycle $C=v_1v_2\cdots v_sv_1$, where $s\equiv 1\bmod 2$. Thus, $f(v_1)=-f(v_1)$, implying that $f(v_1)=0$. By connectedness, $f(v)=0$ for all $v\in V$. This corollary is proved.
\end{proof}
For a fixed $\delta\in[0,2]$, we define spectral embedding $F^\cq$ based on $\cq$ as
\begin{displaymath}\begin{split}
F^\cq\colon V(G)&\rightarrow \ell^2(V(G), w)\\
x&\mapsto F^\cq_x\coloneqq \frac{I_\cq(\delta)\be_x}{\sqrt{w(x)}}=\frac{I_\cq(\delta)\ind_x}{w(x)}.
\end{split}\end{displaymath}
It is clear that for each $x\in V$, $F^\cq_x\in\ell^2(V, w)$ is a real-valued function on $V$.
\begin{lemma}\label{spec_embed_norm}
For each finite or infinite graph $G$ and $x\in V(G)$,
\begin{displaymath}
\pushQED{\qed}
\norm[\big]{F^\cq_x}_w^2=F^\cq_x(x)=\frac{\mu_x^\cq(\delta)}{w(x)}.\qedhere
\popQED
\end{displaymath}
\end{lemma}
\begin{lemma}\label{def_f}
If $\mu_x^\cq(\delta)>0$, define $f\colon V\rightarrow\CN$ as $f\coloneqq \frac{F^\cq_x}{\norm{F^\cq_x}_w}$. Then
\begin{enumerate}[\upshape(1)]
\item $\norm{f}_w=1$;
\item $f(x)=\sqrt{\mu_x^\cq(\delta)/w(x)}$;
\item $f\in\img\bigl(I_\cq(\delta)\bigr)$.\hfill\qedsymbol
\end{enumerate}
\end{lemma}
We need only mimic the proofs of \autocite[Lemmas~3.11 and~3.12]{Lyons-2018} to prove {\Cref{spec_embed_norm,def_f}}.
\subsection{Some Preliminaries}\label{ssec_prel}
Lemma~4.5 in \autocite{Lyons-2018} gave an estimate for diameter of regular graphs. Using a similar argument, we get the following \Cref{reg_diam}.
\begin{lemma}\label{reg_diam}
Let $G$ be a finite, simple, connected graph. We have\/ $\D(G)\leqslant \frac{3n}{\dmi+1}-1$.
% 笔记 55条
%\begin{enumerate}[\upshape (1)]
%\item $\D(G)\leqslant \frac{3n}{d+1}-1$;
%\item $\abs[\big]{B(x,r)}\geqslant \frac{r}{3}(d+1)+2$.
%\end{enumerate}
\end{lemma}
\begin{proof}
Suppose that $u_0$ and $u_{\D(G)}$ are two vertices of $G$ with
\begin{displaymath}
\dist\bigl(u_0,u_{\D(G)}\bigr)=\D(G),
\end{displaymath}
and $\mathcal{P}=u_0u_1\ldots u_{\D(G)}$ is a shortest path joining these two vertices. Let
\begin{displaymath}
S=\{u_0,u_3,\ldots,u_{\lfloor\D(G)\rfloor_{(3)}}\},
\end{displaymath}
where $\lfloor\D(G)\rfloor_{(3)}$ is the largest multiple of $3$ that is smaller than or equal to $\D(G)$. It is easy to see that
\begin{displaymath}
\abs{S}=\frac{\lceil\D(G)+1\rceil^{(3)}}{3},
\end{displaymath}
where $\lceil\D(G)+1\rceil^{(3)}$ is the smallest multiple of $3$ that is greater than or equal to $\D(G)+1$. Because $\mathcal{P}$ is a shortest path, no vertex in $S$ is adjacent to any other vertex in $S$, and no pair of vertices in $S$ have any common neighbors. Hence,
\begin{displaymath}\begin{split}
n&\geqslant \abs{S}+\dmi\abs{S}=\abs{S}(\dmi+1)\\
&=\frac{\lceil\D(G)+1\rceil^{(3)}}{3}(\dmi+1)\geqslant \frac{\D(G)+1}{3}(\dmi+1).
\end{split}\end{displaymath}
Thus,
\begin{displaymath}
\D(G)\leqslant \frac{3n}{\dmi+1}-1.\qedhere
\end{displaymath}
\end{proof}
The following two results will be useful when we are dealing with return probabilities in subsequent sections.
\begin{lemma}\label{ibp}
Let $t$ be a positive integer and $\eta$ be an increasing and right-continuous function on\/ $[0,2]$ with $\eta(0)=0$. Then we have
\begin{displaymath}
t\int_{0}^{2}\eta(\lambda)(1-\lambda)^{t-1}\df{\lambda}
=
(-1)^{t+1}\eta(2)+\int_{(0,2]}(1-\lambda)^t\,\ud\eta(\lambda).
\end{displaymath}
\end{lemma}
\begin{proof}
Using integration by parts, we have
\begin{displaymath}\begin{split}
t\int_{0}^{2}\eta(\lambda)(1-\lambda)^{t-1}\df{\lambda}
&=-\int_{0}^{2}\eta(\lambda)\df{(1-\lambda)^t}\\
&=-\left.\eta(\lambda)(1-\lambda)^t\right|^2_0+\int_{(0,2]}(1-\lambda)^t\ud\eta(\lambda)\\
&=(-1)^{t+1}\eta(2)+\int_{(0,2]}(1-\lambda)^t\ud\eta(\lambda).\qedhere
\end{split}\end{displaymath}
\end{proof}
\begin{lemma}\label{spec_bound}
Consider simple random walk on a non-bipartite, simple, connected, weighted graph $G$.
%笔记 15条
\begin{enumerate}[\upshape(1)]
\item
Let $\varphi$ be an increasing and right-continuous function on\/ $[0,2]$. Assume further that $\varphi$ satisfies the following conditions:
\begin{displaymath}\begin{split}
&\varphi(0)=0=\mu_x^*(0),\qquad \varphi(2)=1-\pi(x)=\mu_x^*(2),\\
&\mu_x^*(\lambda)\leqslant \varphi(\lambda) \qquad \text{for\quad} \lambda\in[0,1),\\
&\mu_x^*(\lambda)\geqslant \varphi(\lambda) \qquad \text{for\quad} \lambda\in[1,2].
\end{split}\end{displaymath}
Then for $t\equiv 0\bmod 2$,
\begin{displaymath}
\pi(x)\leqslant p_t(x,x)\leqslant \pi(x)+\int_{(0,2]}(1-\lambda)^t\,\ud\varphi(\lambda).
\end{displaymath}
\item Let $\psi_1$ and $\psi_2$ be increasing and right-continuous functions on\/ $[0,2]$. Assume further that $\psi_1$ and $\psi_2$ satisfy the following conditions:
\begin{displaymath}\begin{split}
&\psi_1(0)=\psi_2(0)=0=\mu_x^*(0),\\
&\psi_1(2)=\psi_1(2)=1-\pi(x)=\mu_x^*(2),\\
&\mu_x^*(\lambda)\leqslant \psi_2(\lambda) \qquad \text{for\quad} \lambda\in[0,2],\\
&\mu_x^*(\lambda)\geqslant \psi_1(\lambda) \qquad \text{for\quad} \lambda\in[0,2].
\end{split}\end{displaymath}
Then for $t\equiv 1\bmod 2$, we have
\begin{displaymath}
\pi(x)+\int_{(0,2]}(1-\lambda)^t\,\ud\psi_1(\lambda)
\leqslant
p_t(x,x)
\leqslant \pi(x)+\int_{(0,2]}(1-\lambda)^t\,\ud\psi_2(\lambda).
\end{displaymath}
\end{enumerate}
\end{lemma}
\begin{proof}
a) This part is essentially a mimic of the proof of \autocite[Lemma~3.5]{Lyons-2018}. We consider a non-negative integer $t$ in this part.
Since $P=I-\cl$, we have
\begin{displaymath}
p_t(x,x)=\bigl\langle(I-\cl)^t\ind_x,\ind_x\bigr\rangle.
\end{displaymath}
Symbolic calculus gives
\begin{displaymath}
(I-\cl)^t=\int_{[0,2]}(1-\lambda)^tI_\cl(\ud\lambda).
\end{displaymath}
Therefore, by the definition of the vertex spectral measure of $\cl$, we have
\begin{displaymath}
p_t(x,x)=\int_{[0,2]}(1-\lambda)^t\df{\bigl\langle I_\cl(\lambda)\ind_x,\ind_x\bigr\rangle}
=\int_{[0,2]}(1-\lambda)^t\mu_x(\ud\lambda).
\end{displaymath}
It follows that
\begin{equation}\label{rep}
p_t(x,x)=\pi(x)+\int_{[0,2]}(1-\lambda)^t\mu^*_x(\ud\lambda).
\end{equation}
Furthermore, using integration by parts, we have
\begin{displaymath}\begin{split}
p_t(x,x)&=\pi(x)+\int_{[0,2]}(1-\lambda)^t\mu^*_x(\ud\lambda)\\
&=\pi(x)+\left.(1-\lambda)^t\mu^*_x(\lambda)\right|^2_0-\int_{0}^{2}\mu^*_x(\lambda)\df{(1-\lambda)^t}\\
&=\pi(x)+(-1)^t\mu^*_x(2)+t\int_{0}^{2}\mu^*_x(\lambda)(1-\lambda)^{t-1}\df{\lambda}.
\end{split}\end{displaymath}
\par
b) When $t\equiv 0\bmod 2$, because $(1-\lambda)^t$ is always non-negative, by \Cref{rep}, $p_t(x,x)\geqslant \pi(x)$. On the other hand, by the result in part a),
\begin{displaymath}\begin{split}
p_t(x,x)&=\pi(x)+\bigl(1-\pi(x)\bigr)+t\int_{0}^{2}\mu^*_x(\lambda)(1-\lambda)^{t-1}\df{\lambda}\\
&\leqslant 1+t\int_{0}^{2}\varphi(\lambda)(1-\lambda)^{t-1}\df{\lambda}\\
&= 1-\bigl(1-\pi(x)\bigr)+\int_{(0,2]}(1-\lambda)^t\,\ud\varphi(\lambda)\\
&=\pi(x)+\int_{(0,2]}(1-\lambda)^t\,\ud\varphi(\lambda),
\end{split}\end{displaymath}
where the second equality is by \Cref{ibp}. The first assertion is proved.
\par
c) When $t\equiv 1\bmod 2$, by the result in part a), we have
\begin{displaymath}
p_t(x,x)=2\pi(x)-1+t\int_{0}^{2}\mu^*_x(\lambda)(1-\lambda)^{t-1}\df{\lambda}.
\end{displaymath}
Therefore, by \Cref{ibp},
\begin{displaymath}\begin{split}
p_t(x,x)&\geqslant 2\pi(x)-1+\bigl(1-\pi(x)\bigr)+t\int_{0}^{2}\psi_1(\lambda)(1-\lambda)^{t-1}\df{\lambda}\\
&=\pi(x)+\int_{(0,2]}(1-\lambda)^t\,\ud\psi_1(\lambda).
\end{split}\end{displaymath}
Similarly,
\begin{displaymath}
p_t(x,x)
\leqslant
\pi(x)+\int_{(0,2]}(1-\lambda)^t\,\ud\psi_2(\lambda).\qedhere
\end{displaymath}
\end{proof}
%\begin{rmk}
%参看菲赫金哥尔茨第三本75页
%\end{rmk}
%================================
%==========Section 3=============
%================================
\section{Return Probability on Regular Graphs}\label{sec_reg}
In this section, we consider regular graphs.
\subsection{Estimate of Spectral Measure and Convergence Rate}\label{reg}
%谱测度估计与收敛速度
\begin{theorem}\label{reg_return_spec}
Let $G$ be a non-bipartite, connected, regular, unweighted, simple graph. For each $x\in V(G)$, we have
\begin{displaymath}
\mu_x^\cq(\delta)\leqslant 10\sqrt{\delta},\qquad 0\leqslant \delta\leqslant 2.
\end{displaymath}
\end{theorem}
\begin{proof}
Assume $G$ is $d$-regular. Now fixing a vertex $x\in V(G)$, we define $f$ as in \Cref{def_f}.
\par
a) If $f(y)\geqslant 0$ for all $y\in V(G)$, then
\begin{displaymath}\begin{split}
10\sqrt{\delta}&\geqslant \delta\geqslant \langle\cq f,f\rangle_w=\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2\\
&\geqslant
\sum_{y\in N(x)}w(x,y)\abs[\big]{f(x)+f(y)}^2
\geqslant
w(x)\abs[\big]{f(x)}^2\\
&=\mu_x^\cq(\delta).
\end{split}\end{displaymath}
\par
b) We may now assume without loss of generality that $\bigl\{v\in V(G)\st f(v)<0\bigr\}$ is non-empty. Let $S\coloneqq \bigl\{s\in V\st f(s)>0\bigr\}$ and $T\coloneqq \bigl\{t\in V\st f(t)<0\bigr\}$. By our assumptions, both $S$ and $T$ are non-empty. Recall that for each edge $e\in E(G)$, we have $w(e)=1$ because $G$ is assumed to be unweighted. However, we do the following construction of an auxiliary graph $G'$ for general weighted graphs because this will also be useful subsequently in the proof of \Cref{vgrow_spec}:
\begin{enumerate}[\hspace{3em}(1)]
\item The vertex set $V'$ of $G'$ includes $V$. Also, if $(u,v)=e\in E(G)$ and $u,v\in S$, we introduce two vertices $u^{(e)}$ and $v^{(e)}$ in $V'$. Similarly, if $(u,v)=e\in E(G)$ and $u,v\in T$, we introduce two vertices $u^{(e)}$ and $v^{(e)}$ in $V'$.
\item Construct the edge set $E'$ of $G'$ and their weights: If $e=(u,v)\in E(G)$ with $u$ and $v$ both in $S$, we introduce three edges $(u,u^{(e)})$, $(u^{(e)},v^{(e)})$, and $(v^{(e)},v)$ in $E'$; if $e=(u,v)\in E(G)$ with $u$ and $v$ both in $T$, we introduce three edges $(u,u^{(e)})$, $(u^{(e)},v^{(e)})$, and $(v^{(e)},v)$ in $E'$. Suppose $e=(u,v)$ has weight $w(u,v)$ in $G$. For edges introduced above, we assign $w(u,v)$ as their weights $w'$ in $G'$, i.e.,
\begin{displaymath}
w'(u,u^{(e)}) = w'(u^{(e)},v^{(e)}) = w'(v^{(e)},v) = w(u,v).
\end{displaymath}
If $e=(u,v)\in E(G)$ is not of the aforementioned forms, introduce one edge $(u,v)$ in $E'$ and set its weight $w'(u,w)$ in $G'$ as $w(u,v)$.
\end{enumerate}
It is obvious that $G'$ is connected. By the above construction, for $v\in V\subseteq V'$, $v$ has the same weight in $G$ and in $G'$. For $v\in V'$, denote its weight in $G'$ as $w'(v)$. Since $G$ is unweighted in the current setup, $G'$ is also unweighted. So for a vertex $v'$ of $G'$, $w'(v')$ equals the degree of $v'$ in $G'$.
\par
c) Define a function $g\colon V'\rightarrow \RNS$:
\begin{numcases}
{g(v)\coloneqq }
\abs[\big]{f(v)},
\quad \nonumber &$v\in S\cup T$,\\
0, \quad \nonumber &\text{otherwise}.
\end{numcases}
It follows that
\begin{displaymath}\begin{split}
g(x)=f(x)&=\sqrt{\mu_x^\cq(\delta)/w(x)},\\
\sum_{v\in V'}\abs[\big]{g(v)}^2w'(v)&=\sum_{v\in V}\abs[\big]{f(v)}^2w(v)=1,\\
\sum_{(v_1,v_2)\in E'}w'(v_1,v_2)\abs[\big]{g(v_1)-g(v_2)}^2&\leqslant \sum_{(v_1,v_2)\in E}w(v_1,v_2)\abs[\big]{f(v_1)+f(v_2)}^2.
\end{split}\end{displaymath}
We claim that $g(v')=0$ for some $v'\in V'$. Otherwise, if $V'\setminus{(S\cup T)}=\varnothing$, all edges in $G$ would be between $S$ and $T$, contradicting the assumption that $G$ is non-bipartite.
Set $B\coloneqq \bigl\{y\in V'\st \abs[\big]{g(y)-g(x)}\leqslant \frac{g(x)}{2}\bigr\}$. It follows that $B\subseteq S\cup T\subsetneq V'$. Since $G'$ is connected, there exists a shortest path $\cp$ in $G'$ from $x$ to $V'\setminus B$.
\par
d) If $\abs{\cp}=1$ and at least half of the neighbors of $x$ are outside of $B$, then we have
\begin{displaymath}\begin{split}
\delta
&\geqslant
\langle\cq f,f\rangle=
\sum_{(v_1,v_2)\in E}\abs[\big]{f(v_1)+f(v_2)}^2\\
&\geqslant \sum_{(v_1,v_2)\in E'}\abs[\big]{g(v_1)-g(v_2)}^2\\
&\geqslant\frac{d}{2}\cdot\frac{1}{4}\abs[\big]{g(x)}^2
=\frac{d}{2}\cdot\frac{\mu_x^\cq(\delta)}{4w(x)}\\
&=\frac{\mu_x^\cq(\delta)}{8}.
\end{split}\end{displaymath}
\par
Therefore, when $\delta\leqslant 1$, $\mu_x^\cq(\delta)\leqslant 8\delta\leqslant 8\sqrt{\delta}\leqslant 10\sqrt{\delta}$; when $\delta>1$, $\mu_x^\cq(\delta)\leqslant1<\delta<10\sqrt{\delta}$. The result holds in this case.
\par
e) If $\abs{\cp}=1$ and at least half of the neighbors of $x$ are inside of $B$, then the neighbors of $x$ inside of $B$ are in $S\cup T$ so of degree $d$ in $G'$. Thus,
\begin{displaymath}
\vol(B;G')\geqslant \frac{d}{2}\cdot d=\frac{d^2}{2}\abs{\cp}
\geqslant \frac{d^2}{6}\abs{\cp}.
\end{displaymath}
If $\abs{\cp}\geqslant 2$, we claim that $\vol(B;G')\geqslant \frac{d^2}{6}\abs{\cp}$ still holds. To justify this claim, we assume that $\cp=u_0u_1\cdots u_ru_{r+1}$, where $r=\abs{P}-1$. Consider $\widetilde{B}\coloneqq \bigl\{y\in V'\st \dist(x,y;G')\leqslant r \bigr\}\subseteq B\subseteq S\cup T$. Since $\cp$ is a shortest path, $u_0$, $u_1$, $\ldots$, $u_{r-2}$, and $u_{r-1}$ are all in $S\cup T$, and are therefore of degree $d$ (in both $G$ and $G'$). Setting $K\coloneqq \{u_0,u_3,\ldots, u_{\lfloor r-1\rfloor_{(3)}}\}$, we have
\begin{displaymath}
\abs{K}=\frac{\lceil r\rceil^{(3)}}{3},
\end{displaymath}
where ${\lceil r\rceil^{(3)}}$ is the smallest multiple of 3 that is greater than or equal to $r$.
By a similar argument as in the proof of \Cref{reg_diam}, it follows that
\begin{displaymath}
\abs{\widetilde{B}}\geqslant (r+1)+\abs{K}(d-2),
\end{displaymath}
where $r+1$ is counting the vertices $\{u_0,u_1,\ldots, u_r\}$, and $\abs{K}(d-2)$ is counting the neighbors of $K$ that are not on $\cp$. Hence, we get that
\begin{displaymath}\begin{split}
\vol(B;G')
&\geqslant
\vol(\widetilde{B};G')
=
d\abs{\widetilde{B}}\\
&\geqslant
d\cdot\bigl((r+1)+\abs{K}(d-2)\bigr)\\
&\geqslant
d\cdot \bigl((r+1)+\frac{r}{3}(d-2)\bigr)\\
&\geqslant \frac{d^2r}{3}
=\frac{d^2}{3} \bigl(\abs{\cp}-1\bigr)\\
&\geqslant \frac{d^2}{6}\abs{\cp}.
\end{split}\end{displaymath}
So our claim holds.
\par
Therefore, we may assume that $\vol(B;G')\geqslant \frac{d^2}{6}\abs{\cp}$.
\par
f) Note that $\sum_{v\in V'}\abs[\big]{g(v)}^2w'(v)=1$. It is easy to get
\begin{displaymath}
\vol(B;G')\leqslant \frac{1}{\abs[\big]{\frac{1}{2}g(x)}^2}
=\frac{4w(x)}{\mu_x^\cq(\delta)}
=\frac{4d}{\mu_x^\cq(\delta)}.
\end{displaymath}
Therefore,
\begin{displaymath}
\frac{4d}{\mu_x^\cq(\delta)}\geqslant \vol(B,G')\geqslant \frac{d^2}{6}\abs{\cp}.
\end{displaymath}
Thus,
\begin{displaymath}
\abs{\cp}\leqslant \frac{24}{d\mu_x^\cq(\delta)}.
\end{displaymath}
Hence,
\begin{displaymath}\begin{split}
\delta&\geqslant \langle\cq f,f\rangle_w
=\sum_{(v_1,v_2)\in E}\abs[\big]{f(v_1)+f(v_2)}^2\\
&\geqslant \sum_{(v_1,v_2)\in E'}\abs[\big]{g(v_1)-g(v_2)}^2
\geqslant \sum_{i=0}^{\abs{\cp}-1}\abs[\big]{g(u_i)-g(u_{i+1})}^2\\
&\geqslant \frac{1}{\abs{\cp}}\biggl(\sum_{i=0}^{\abs{\cp}-1}\abs[\big]{g(u_i)-g(u_{i+1})}\biggr)^2
\geqslant \frac{1}{\abs{\cp}}\abs[\big]{g(u_0)-g(u_{\abs{\cp}})}^2\\
&\geqslant \frac{1}{\abs{\cp}}\frac{\abs[\big]{g(x)}^2}{4}
=\frac{\mu_x^\cq(\delta)}{4\abs{\cp}w(x)}\\
&=\frac{\mu_x^\cq(\delta)}{4d\abs{\cp}}.
\end{split}\end{displaymath}
Proceeding further, we have
\begin{displaymath}\begin{split}
\delta\geqslant \frac{\mu_x^\cq(\delta)}{4d\abs{\cp}}\geqslant
\frac{\mu_x^\cq(\delta)}{4d\frac{24}{d\mu_x^\cq(\delta)}}
=\frac{\bigl(\mu_x^\cq(\delta)\bigr)^2}{96}.
\end{split}\end{displaymath}
Therefore,
\begin{displaymath}
\mu_x^\cq(\delta)\leqslant \sqrt{96\delta},\qquad 0\leqslant \delta\leqslant 2.\qedhere
\end{displaymath}
\end{proof}
For \Cref{reg_return_spec}, we remark that the same upper bound holds with an extra factor of the maximum degree over the minimum degree for general unweighted graphs; it is similar for the return probability bound in \Cref{reg_return}. The upper bound in \Cref{reg_return_spec} could be easily used to get lower bounds on the eigenvalues of $P$. To this end, we need \Cref{spec_loc}:
\begin{lemma}\label{spec_loc}
Let $G$ be a finite, connected, weighted graph. We have
\begin{displaymath}
\sum_{x\in V}\mu_x^\cq(\delta)=\abs[\big]{\{j\st \lambda_j^\cq\leqslant \delta\}}.
\end{displaymath}
\end{lemma}
\begin{proof}
Recall that the eigenvalues of $\cq$ are
\begin{displaymath}
0\leqslant \lambda^\cq_{\mathrm{min}} = \lambda^\cq_1
\leqslant \lambda^\cq_2 \leqslant \lambda^\cq_3 \leqslant \dotsb \leqslant \lambda^\cq_{n-1}
< \lambda^\cq_n = \lambda^\cq_{\mathrm{max}} = 2.
\end{displaymath}
Let $h_1$, $h_2$, $\ldots$, $h_n$ be an orthonormal basis of $\ell^2(V,w)$ such that
\begin{displaymath}
\cq h_j=\lambda^\cq_j h_j,\qquad 1\leqslant j\leqslant n.
\end{displaymath}
It follows that
\begin{displaymath}
\sum_{x\in V}\abs{\langle h_j,\be_x\rangle_w}^2
=\norm{h_j}_w^2=1.
\end{displaymath}
Therefore,
\begin{displaymath}\begin{split}
\sum_{x\in V}\mu_x^\cq(\delta)
&=
\sum_{x\in V}\bigl\langle I_\cq(\delta)\be_x,\be_x\bigr\rangle_w
=
\sum_{x\in V}\norm{I_\cq(\delta)\be_x}^2_w \\
&=
\sum_{x\in V}\norm[\Big]{\sum_{j\st \lambda_j^\cq\leqslant \delta}\langle\be_x,h_j\rangle_w h_j}^2
=
\sum_{x\in V}\sum_{j\st \lambda_j^\cq\leqslant \delta}\abs[\big]{\langle\be_x,h_j\rangle_w}^2\\
&=
\sum_{j\st \lambda_j^\cq\leqslant \delta}\sum_{x\in V}\abs[\big]{\langle\be_x,h_j\rangle_w}^2
=
\sum_{j\st \lambda_j^\cq\leqslant \delta}\norm{h_j}^2_w
=
\sum_{j\st \lambda_j^\cq\leqslant \delta}1\\
&=
\abs[\big]{\{j\st \lambda_j^\cq\leqslant \delta\}}.\qedhere
\end{split}\end{displaymath}
\end{proof}
\begin{corollary}\label{lbb_reg}
Let $G$ be a regular, non-bipartite, finite, simple, connected, unweighted graph. For\/ $1\leqslant k\leqslant n$, we have $\lambda^\cq_k\geqslant \frac{k^2}{100n^2}$. Therefore, $\lambda^P_k\geqslant -1+\frac{k^2}{100n^2}$.
\end{corollary}
\begin{proof}
By \Cref{spec_loc} and \Cref{reg_return_spec},
\begin{displaymath}
\abs[\big]{\{j\st \lambda_j^\cq\leqslant \delta\}}=\sum_{x\in V}\mu_x^\cq(\delta)=10n\sqrt{\delta}.
\end{displaymath}
Therefore, if $10n\sqrt{\delta}<k$, $\abs[\big]{\{j\st \lambda_j^\cq\leqslant \delta\}}<k$. In other words, $\lambda^\cq_k\geqslant \frac{k^2}{100n^2}$.
\end{proof}
Our main interest is to get bounds on return probabilities of simple random walk on regular graphs.
\begin{theorem}\label{reg_return}
Let $G$ be a regular, non-bipartite, simple, connected, unweighted graph. For each $x\in V(G)$, simple random walk on $G$ satisfies
\begin{displaymath}\begin{split}
&0\leqslant p_t(x,x)-\pi(x)\leqslant \frac{18}{\sqrt{t}} \qquad \text{for\quad} t\equiv 0\bmod 2,\\
&\abs[\big]{p_t(x,x)-\pi(x)}\leqslant \frac{9}{\sqrt{t}} \qquad \text{for\quad} t\equiv 1\bmod 2.
\end{split}\end{displaymath}
\end{theorem}
In the proof of \Cref{reg_return}, we will need the following elementary calculation.
\begin{lemma}\label{calc_aux}
For $t>0$, we have
\begin{displaymath}
\int_{0}^{1}(1-\lambda)^t\frac{1}{\sqrt{\lambda}}\df{\lambda}\leqslant \frac{1.8}{\sqrt{t}}.
\end{displaymath}
\end{lemma}
\begin{proof}
We have
\begin{displaymath}
\int_{0}^{1}(1-\lambda)^t\frac{1}{\sqrt{\lambda}}\df{\lambda}
\leqslant \int_{0}^{1}\exp\{-\lambda t\}\lambda^{-1/2}\df{\lambda}
\leqslant \int_{0}^{\infty}\exp\{-\lambda t\}\lambda^{-1/2}\df{\lambda}.
\end{displaymath}
Taking a change of variable $\lambda t=s$, we get that
\begin{displaymath}\begin{split}
\int_{0}^{\infty}\ue^{-s}(s/t)^{-1/2}\df{(s/t)}
&\leqslant \frac{1}{\sqrt{t}}\int_{0}^{\infty}\ue^{-s}s^{-1/2}\df{s}\\
&=\frac{\Gamma(1/2)}{\sqrt{t}}\leqslant \frac{1.8}{\sqrt{t}}.
\end{split}\end{displaymath}
Hence, the inequality follows immediately.
\end{proof}
\begin{proof}[Proof of \Cref{reg_return}]
a) Set
\begin{numcases}
{\varphi(\lambda)\coloneqq }
10\sqrt{\lambda}
\quad \nonumber &if\/ $\lambda\geqslant0\text{ and }10\sqrt{\lambda}\leqslant \mu_x^*(1)$,\\
\mu_x^*(1) \quad \nonumber &\text{for intermediate values of $\lambda$},\\
\mu_x^*(2)-10\sqrt{2-\lambda}
\quad \nonumber &if\/ $\lambda\leqslant2\text{ and }\mu_x^*(2)-10\sqrt{2-\lambda}\geqslant \mu_x^*(1)$.
\end{numcases}
We claim that the function $\varphi$ defined above satisfies the conditions in \Cref{spec_bound}(1). In fact, it is known that $\mu^*_x(\lambda)<10\sqrt{\lambda}$ from \autocite[Theorem~4.9]{Lyons-2018}. On the other hand, by \Cref{reg_return_spec},
\begin{displaymath}\begin{split}
\mu^*_x(\lambda)&=\bigl\langle I_\cl\bigl((0,\lambda]\bigr)\be_x,\be_x\bigr\rangle_w \\
&=\bigl\langle I_\cl\bigl((0,2]\bigr)\be_x,\be_x\bigr\rangle_w
-\bigl\langle I_\cl\bigl((\lambda,2]\bigr)\be_x,\be_x\bigr\rangle_w \\
&= 1-\bigl\langle I_\cl({0})\be_x,\be_x\bigr\rangle_w
-\bigl\langle I_\cl\bigl((\lambda,2]\bigr)\be_x,\be_x\bigr\rangle_w \\
&\geqslant 1-\bigl\langle I_\cl({0})\be_x,\be_x\bigr\rangle_w
-\bigl\langle I_\cl\bigl([\lambda,2]\bigr)\be_x,\be_x\bigr\rangle_w \\
&= 1-\pi(x)
-\bigl\langle I_\cq(2-\lambda)\be_x,\be_x\bigr\rangle_w \\
&\geqslant 1-\pi(x)-10\sqrt{2-\lambda}\\
&=\mu_x^*(2)-10\sqrt{2-\lambda}.
\end{split}\end{displaymath}
%where we are using \Cref{0span} to calculate out $\bigl\langle I_\cl(\{0\})\be_x,\be_x)\bigr\rangle_w$.
The claim is now proved.
Therefore, for $t\equiv 0\bmod 2$, by \Cref{spec_bound},
\begin{displaymath}\begin{split}
\pi(x)\leqslant p_t(x,x)&\leqslant \pi(x)+\int_{0}^{2}(1-\lambda)^t\varphi'(\lambda)\df{\lambda}\\
&\leqslant \pi(x)+2\int_{0}^{1}(1-\lambda)^t\frac{5}{\sqrt{\lambda}}\df{\lambda}\\
&= \pi(x)+10\int_{0}^{1}(1-\lambda)^t\frac{1}{\sqrt{\lambda}}\df{\lambda}\\
&\leqslant \pi(x)+\frac{18}{\sqrt{t}}.
\end{split}\end{displaymath}
The first assertion is proved.
\par
b) Set
\begin{displaymath}
\psi_1(\lambda)\coloneqq \bigl(\mu_x^*(2)-10\sqrt{2-\lambda}\,\bigr)\vee 0,\qquad \lambda\in [0,2],
\end{displaymath}
and
\begin{displaymath}
\psi_2(\lambda)\coloneqq (10\sqrt{\lambda}\,)\wedge \mu_x^*(2),\qquad \lambda\in [0,2].
\end{displaymath}
Then they satisfy the conditions of \Cref{spec_bound}(2). Consequently, we have
\begin{displaymath}
\pi(x)+\int_{0}^{2}(1-\lambda)^t\psi_1'(\lambda)\df{\lambda}\leqslant p_t(x,x)\leqslant \pi(x)+\int_{0}^{2}(1-\lambda)^t\psi_2'(\lambda)\df{\lambda}.
\end{displaymath}
By some elementary calculation, we get that
\begin{displaymath}
\pi(x)-\int_{0}^{1}(1-\lambda)^t\frac{5}{\sqrt{\lambda}}\df{\lambda}\leqslant p_t(x,x)\leqslant \pi(x)+\int_{0}^{1}(1-\lambda)^t\frac{5}{\sqrt{\lambda}}\df{\lambda}.
\end{displaymath}
The second assertion follows immediately from \Cref{calc_aux}.
\end{proof}
\subsection{Sharpness and Spectral Radius}
The order of $\frac{1}{\sqrt{t}}$ in \Cref{reg_return} is sharp. We show this by the following \Cref{sharp}.
\begin{example}\label{sharp}
We will exploit a local limit theorem. Consider an unweighted graph $G$ with $V(G)=\ZN$: $(i,j)\in E(G)$ if and only if\/ $0<\abs{i-j}\leqslant 2$. Obviously, $G$ is non-bipartite, connected, and\/ $4$-regular. We also consider an i.i.d.\ sequence $\{\xi_k\}_{k=1}^\infty$ of random variables that are uniform on\/ $\{-2,-1,1,2\}$. Then for the simple random walk on $G$,
\begin{displaymath}
p_t(0,0)=\PBig{\sum_{k=1}^{t}\xi_k=0}.
\end{displaymath}
Using the notation in \autocite[Theorem 1.1]{Davis-1995}, we have
\begin{displaymath}\begin{split}
a_t=0,\qquad t\geqslant 1,\\
b_t=\sqrt{\frac{5t}{2}},\qquad t\geqslant 1.
\end{split}\end{displaymath}
Therefore, by \autocite[Theorem~1.1]{Davis-1995}, we have
\begin{displaymath}
\abs[\bigg]{\sqrt{\frac{5t}{2}}\PBig{\sum_{k=1}^{t}\xi_k=0}-\frac{1}{\sqrt{2\pi}}}\to 0,\qquad\text{as }t\to\infty.
\end{displaymath}
In other words,
\begin{displaymath}
p_t(0,0)\sim \frac{1}{\sqrt{5\pi t}},\qquad\text{as }t\to\infty.
\end{displaymath}
Hence, the sharpness is demonstrated.\qed
\end{example}
In \Cref{reg}, we focused on the negative spectrum of $P$ by considering $\cq=I+P$. As a comparison, in \autocite[Theorem~4.9]{Lyons-2018}, $\cl=I-P$ is exploited and so essentially the positive spectrum of $P$ is focused on. Recalling that the spectrum $\sigma(P)$ of $P$, as an operator on $\ell^2(V, w)$, is contained in $[-1,1]$, we set
\begin{displaymath}\begin{split}
\gamma_-&\coloneqq \inf\sigma(P)-(-1),\\
\gamma_+&\coloneqq 1-\sup\bigl(\sigma(P)\setminus\{1\}\bigr).
\end{split}\end{displaymath}
It is obvious that $\gamma_-$ and $\gamma_+$ are both non-negative.
But which is larger between $\gamma_-$ and $\gamma_+$? When $G$ is finite, it depends; but when $G$ is infinite, we have the following simple fact:
\begin{proposition}\label{gap}
Let $G=(V,E)$ be a connected, weighted, infinite, locally finite graph, with\/ $\vol(V)=\infty$. Then we have $\gamma_+\leqslant \gamma_-$. In other words, the spectral radius of $P$ is achieved by the positive spectrum.
\end{proposition}
\begin{proof}
To begin, since $\vol(V)=\infty$, we see that the constant function is not in $\ell^2(V, w)$ and $1$ is not in the point spectrum of $P$.
\par
For an arbitrarily small number $\varepsilon>0$, there exists $f\in \ell^2(V, w)$, with $\norm{f}_w=1$, satisfying
\begin{displaymath}
-1+\gamma_-+\varepsilon
\geqslant
\langle Pf,f\rangle_w
=\sum_{x\in V}w(x)(Pf)(x)\overline{f(x)}
=\sum_{x\in V}w(x)\Bigl(\sum_{y\in N(x)}p(x,y)f(y)\Bigr)\overline{f(x)}.
\end{displaymath}
Hence, we have
\begin{displaymath}\begin{split}
1-\gamma_--\varepsilon &\leqslant \abs{\langle Pf,f\rangle_w}
=
\abs[\bigg]{\sum_{x\in V}w(x)\Bigl(\sum_{y\in N(x)}p(x,y)f(y)\Bigr)\overline{f(x)}}\\
&\leqslant
\sum_{x\in V}w(x)\Bigl(\sum_{y\in N(x)}p(x,y)\abs{f(y)}\Bigr)\abs{f(x)}\\
&= \langle P\abs{f},\abs{f}\rangle_w.
\end{split}\end{displaymath}
Since $1$ is not in the point spectrum of $P$, we further have
\begin{displaymath}
\langle P\abs{f},\abs{f}\rangle_w\leqslant 1-\gamma_+.
\end{displaymath}
Hence, $1-(\gamma_-+\varepsilon)\leqslant 1-\gamma_+$. It follows that
\begin{displaymath}
\gamma_+\leqslant \gamma_-+\varepsilon.
\end{displaymath}
Because $\varepsilon>0$ is arbitrary, we get $\gamma_+\leqslant \gamma_-$.
\end{proof}
\begin{rmk}
\Cref{gap} is stated in the language of graphs, but essentially it also applies to irreducible, non-ergodic, symmetrizable Markov chains with invariant measure, $\pi$. Note that the condition of ``non-ergodic'' means the total mass of $\pi$ is infinity.
\end{rmk}
%================================
%==========Section 4=============
%================================
\section{A Slight Improvement of Return Probability Bound}\label{sec_Oli}
In this section, we consider finite graphs only. We use essentially the same method as in \autocite{Oliveira-2019}. Let $G$ be a non-bipartite, finite, simple, connected, unweighted graph.
Then $\lambda^P_{\mathrm{min}} = \lambda^P_1>-1$. Set $\Lambda\coloneqq \abs{\lambda^P_1}\vee\abs{\lambda^P_{n-1}}$, $\trel\coloneqq (1-\Lambda)^{-1}$, $t'\coloneqq \lceil\trel\rceil-1$.
\begin{theorem}\label{reg_return_imp}
Let $G$ be a non-bipartite, finite, simple, connected, unweighted graph. For $t\geqslant \trel$, simple random walk on $G$ satisfies
\begin{displaymath}\begin{split}
&0\leqslant p_{2t}(x,x)-\pi(x)\leqslant \frac{8d(x)\sqrt{\trel+1}}{(t+1)\dmi},\\
&\abs[\big]{p_{2t+1}(x,x)-\pi(x)}\leqslant \frac{8d(x)\sqrt{\trel+1}}{\sqrt{(t+1)(t+2)}\,\dmi}.
\end{split}\end{displaymath}
\end{theorem}
%{\color{red} When $t\geqslant t_{\text{rel}}$, the bound in \Cref{reg_return_imp} is comparable to that in \Cref{reg_return}.}
To prove \Cref{reg_return_imp}, we need some preparation. Recall that the hitting time of $A\subseteq V$ is $\tau_A\coloneqq \inf\{t\geqslant 0\st X_t\in A\}$ and Green's function is
\begin{displaymath}
g_t(x,y)\coloneqq \sum_{s=0}^{t}p_t(x,y),\qquad t\geqslant 0,\ x,y\in V.
\end{displaymath}
\begin{lemma}\label{lm31}
For $t\geqslant t'$ and $x\in V$, we have
\begin{displaymath}
0\leqslant p_{2t}(x,x)-\pi(x)\leqslant \frac{1}{1-\ue^{-2}}\cdot\frac{g_{2t'}(x,x)}{t+1}.
\end{displaymath}
\end{lemma}
\begin{proof}
By \Cref{rep}, $p_{2t}(x,x)-\pi(x)$ is nonnegative and decreasing in $t$. Therefore, we have
\begin{equation}\label{green1}
0\leqslant p_{2t}(x,x)-\pi(x)\leqslant \frac{1}{t+1}\Bigl(\sum_{s=0}^{t}p_{2s}(x,x)-(t+1)\pi(x)\Bigr).
\end{equation}
In addition,
\begin{equation}\label{green2}\begin{split}
\sum_{s=0}^{t}p_{2s}(x,x)-(t+1)\pi(x)&=\sum_{s=0}^{t}\int_{(-1,1)}\lambda^{2s}\norm{I_P(\ud\lambda)\be_x}_w^2
=\int_{(-1,1)}\Bigl(\sum_{s=0}^{t}\lambda^{2s}\Bigr)\norm{I_P(\ud\lambda)\be_x}_w^2\\
&=\int_{(-1,1)}\frac{1-\lambda^{2t+2}}{1-\lambda^2}\norm{I_P(\ud\lambda)\be_x}_w^2\\
&\leqslant \int_{(-1,1)}\frac{1}{1-\lambda^2}\norm{I_P(\ud\lambda)\be_x}_w^2,
\end{split}\end{equation}
where $I_P$ is the resolution of identity for $P$.
On the other hand, since for $1\leqslant i \leqslant n$,
\begin{displaymath}
(\lambda_i^P)^{2t'+2}\leqslant \Lambda^{2t'+2}\leqslant \Lambda^{2/(1-\Lambda)}\leqslant \ue^{-2},
\end{displaymath}
we have
\begin{equation}\begin{split}\label{green3}
\int_{(-1,1)}\frac{1-\ue^{-2}}{1-\lambda^2}\norm{I_P(\ud\lambda)\be_x}_w^2
&\leqslant \int_{(-1,1)}\frac{1-\lambda^{2t'+2}}{1-\lambda^2}\norm{I_P(\ud\lambda)\be_x}_w^2\\
&\leqslant
\sum_{s=0}^{t'}p_{2s}(x,x)-(t'+1)\pi(x)
\leqslant \sum_{s=0}^{t'}p_{2s}(x,x)\\
&\leqslant g_{2t'}(x,x).
\end{split}\end{equation}
By \Cref{green1,green2,green3},
\begin{displaymath}\begin{split}
0&\leqslant p_{2t}(x,x)-\pi(x)
\leqslant
\frac{1}{t+1}\Bigl(\sum_{s=0}^{t}p_{2s}(x,x)-(t+1)\pi(x)\Bigr)\\
&\leqslant
\frac{1}{t+1}\int_{(-1,1)}\frac{1}{1-\lambda^2}\norm{I_P(\ud\lambda)\be_x}_w^2\\
&\leqslant
\frac{1}{(t+1)(1-\ue^{-2})}\int_{(-1,1)}\frac{1-\ue^{-2}}{1-\lambda^2}\norm{I_P(\ud\lambda)\be_x}_w^2\\
&\leqslant
\frac{1}{(t+1)(1-\ue^{-2})}g_{2t'}(x,x).
\end{split}\end{displaymath}
\Cref{lm31} is proved.
\end{proof}
\begin{proposition}\label{trel_bound}
We have
\begin{displaymath}
\trel
\leqslant \max_{x,y\in V}\bigl(\Esubleft{x}{\tau_y}+\Esubleft{y}{\tau_x}\bigr)
\leqslant 3n^2\frac{\dav}{\dmi}-2.
\end{displaymath}
\end{proposition}
\begin{proof}
The first inequality follows from \autocite[Lemma~12.17]{Levin-2017}. For the second inequality, by \autocite[Corollary 2.21]{Lyons-2017} and a simple network reduction argument, we have
\begin{displaymath}
\max_{x,y\in V}\bigl(\Esubleft{x}{\tau_y}+\Esubleft{y}{\tau_x}\bigr)
\leqslant
\D(G)\vol(V).
\end{displaymath}
Since \Cref{reg_diam} gives $\D(G)\leqslant \frac{3n}{\dmi}-1$, we may proceed and get
\begin{displaymath}\begin{split}
\max_{x,y\in V}\bigl(\Esubleft{x}{\tau_y}+\Esubleft{y}{\tau_x}\bigr)
&\leqslant \Bigl(\frac{3n}{\dmi}-1\Bigr)\vol(V)\\
&= \frac{3n\vol(V)}{\dmi}-\vol(V)\\
&\leqslant 3n^2\frac{\dav}{\dmi}-2.\qedhere
\end{split}\end{displaymath}
\end{proof}
For non-empty$A\subsetneq V$ and $x\in V\setminus A$, define $ G(x,x;A)\coloneqq \Esubbig{x}{\sum_{s=0}^{\tau_A-1}\ind_{\{X_s\notin A\}}}$.
\begin{proposition}\label{prp32}
We have
\begin{displaymath}
\frac{G(x,x;A)}{\pi(x)}\leqslant \frac{9}{2}\Bigl(\frac{\dav n}{\dmi}\Bigr)^2\bigl(1-\pi(A)\bigr).
\end{displaymath}
\end{proposition}
\begin{proof}
By \autocite[Eq.~(2.5)]{Lyons-2017} and network reduction, we have
\begin{displaymath}
\frac{G(x,x;A)}{\pi(x)}
\leqslant \vol(V){\mathcal R}_{\text{eff}}(x\leftrightarrow A)
=\dav n {\mathcal R}_{\text{eff}}(x\leftrightarrow A),
\end{displaymath}
where ${\mathcal R}_{\text{eff}}(x\leftrightarrow A)$ is the effective resistance between $x$ and $A$. Then we need only follow the proof of \autocite[Proposition~3.2]{Oliveira-2019} to get
\begin{displaymath}
{\mathcal R}_{\text{eff}}(x\leftrightarrow A)
\leqslant
\frac{9}{2}\frac{\dav n}{\dmi^2}\bigl(1-\pi(A)\bigr).\qedhere
\end{displaymath}
\end{proof}
Now fix $x\in V$. For $\alpha>1$, let
\begin{displaymath}
A_\alpha\coloneqq \bigl\{y\in V\st g_{2t'}(y,x)\leqslant \alpha\pi(x)(2t'+1)\bigr\}.
\end{displaymath}
We claim that $A_\alpha\neq \varnothing$ for $\alpha>1$. In fact,
\begin{equation}\label{alpha}\begin{split}
1-\pi(A_\alpha)&=\sum_{y\notin A_\alpha}\pi(y)<\sum_{y\notin A_\alpha}\pi(y)\frac{g_{2t'}(y,x)}{\alpha\pi(x)(2t'+1)}\\
&=\sum_{y\notin A_\alpha}\frac{\pi(x)g_{2t'}(x,y)}{\alpha\pi(x)(2t'+1)}
\leqslant \sum_{y\in V}\frac{g_{2t'}(x,y)}{\alpha(2t'+1)}\\
&=\frac{1}{\alpha}<1.
\end{split}\end{equation}
So for $\alpha>1$, $A_\alpha$ is non-empty.
\begin{lemma}\label{lm32}
For $x\in V$,
\begin{displaymath}
\frac{g_{2t'}(x,x)}{\pi(x)}
\leqslant \frac{9\dav n}{2\dmi}\sqrt{2t'+1}.
\end{displaymath}
\end{lemma}
\begin{proof}
a) Set $\alpha_0\coloneqq \frac{3\dav n}{\dmi}\cdot\frac{1}{\sqrt{2t'+1}}$. We claim $\alpha_0>1$. In fact, \Cref{trel_bound} ensures
\begin{displaymath}
\lceil\trel\rceil\leqslant 3\frac{\dav}{\dmi}n^2.
\end{displaymath}
Therefore, we have
\begin{displaymath}\begin{split}
\alpha_0
&=
\frac{3\dav n}{\dmi}\cdot\frac{1}{\sqrt{2\lceil\trel\rceil-1}}\\
&\geqslant
\frac{3\dav n}{\dmi}\cdot\frac{1}{\sqrt{\frac{6\dav n^2}{\dmi}}}
=
\frac{3}{\sqrt{6}}\cdot\sqrt{\frac{\dav}{\dmi}}\\
&>1.
\end{split}\end{displaymath}
As a consequence, $A_{\alpha_0}$ is non-empty.
\par
b) If $x\in A_{\alpha_0}$, by the definition of $A_{\alpha_0}$,
\begin{displaymath}
\frac{g_{2t'}(x,x)}{\pi(x)}\leqslant \alpha_0(2t'+1)=\frac{3\dav n}{\dmi}\sqrt{2t'+1}.
\end{displaymath}
\par
c) If $x\notin A_{\alpha_0}$, by the strong Markov property, \Cref{prp32}, and the definition of $A_{\alpha_0}$,
\begin{displaymath}\begin{split}
\frac{g_{2t'}(x,x)}{\pi(x)}
&\leqslant \frac{G(x,x;A_{\alpha_0})}{\pi(x)} + \Esubbigg{x}{\frac{g_{2t'}(X_{\tau_{A_{\alpha_0}}},x)}{\pi(x)}}\\
&\leqslant \frac{9}{2}\cdot\Bigl(\frac{\dav n}{\dmi}\Bigr)^2\bigl(1-\pi(A_{\alpha_0})\bigr)+\alpha_0(2t'+1)\\
&\leqslant \frac{9}{2}\cdot\Bigl(\frac{\dav n}{\dmi}\Bigr)^2\frac{1}{\alpha_0}+\alpha_0(2t'+1)\\
&= (\tfrac{3}{2}+3)\sqrt{2t'+1}\frac{\dav n}{\dmi}\\
&=\frac{9\dav n}{2\dmi}\sqrt{2t'+1},
\end{split}\end{displaymath}
where the third inequality is by \Cref{alpha}.
\end{proof}
We are now in position to prove \Cref{reg_return_imp}.
\begin{proof}[Proof of \Cref{reg_return_imp}]
a) Recalling that $t'=\lceil\trel\rceil-1$, we have $\trel\geqslant t'$. Therefore, $t\geqslant \trel$ implies $t\geqslant t'$.
\par
b) By \Cref{lm31,lm32}, for $t\geqslant \trel$ and $x\in V$,
\begin{displaymath}\begin{split}
0&\leqslant p_{2t}(x,x)-\pi(x)\leqslant \frac{1}{1-\ue^{-2}}\cdot\frac{g_{2t'}(x,x)}{t+1}\\
&\leqslant
\frac{9\dav n\pi(x)}{2(1-\ue^{-2})(t+1)\dmi}\sqrt{2t'+1}
= \frac{9d(x)}{2(1-\ue^{-2})(t+1)\dmi}\sqrt{2\lceil\trel\rceil-1}\\
&\leqslant \frac{9d(x)}{2(1-\ue^{-2})(t+1)\dmi}\sqrt{2(\trel+1)}
\leqslant \frac{8d(x)}{\dmi}\frac{\sqrt{\trel+1}}{t+1}.
\end{split}\end{displaymath}
\par
c) Our calculation in part b) implies that for $t\geqslant \trel$,
\begin{displaymath}
p_{2t}(x,x)-\pi(x)=\int_{(-1,1)}\lambda^{2t}\norm{I_P(\ud\lambda)\be_x}_w^2
\leqslant \frac{8d(x)\sqrt{\trel+1}}{(t+1)\dmi}.
\end{displaymath}
Therefore, when $t\geqslant \trel$, we have
\begin{displaymath}\begin{split}
\abs[\big]{p_{2t+1}(x,x)-\pi(x)}
&=\abs[\Big]{\int_{(-1,1)}\lambda^{2t+1}\norm{I_P(\ud\lambda)\be_x}_w^2}
\leqslant \int_{(-1,1)}\abs{\lambda}^{2t+1}\norm{I_P(\ud\lambda)\be_x}_w^2\\
&= \int_{(-1,1)}\abs{\lambda}^{t}\abs{\lambda}^{t+1}\norm{I_P(\ud\lambda)\be_x}_w^2.
\end{split}\end{displaymath}
Hence, the Cauchy--Schwarz inequality gives
\begin{displaymath}\begin{split}
\abs[\big]{p_{2t+1}(x,x)-\pi(x)}
&\leqslant \sqrt{\int_{(-1,1)}\abs{\lambda}^{2t}\norm{I_P(\ud\lambda)\be_x}_w^2}
\sqrt{\int_{(-1,1)}\abs{\lambda}^{2t+2}\norm{I_P(\ud\lambda)\be_x}_w^2}\\
&\leqslant \frac{8d(x)\sqrt{\trel+1}}{\sqrt{(t+1)(t+2)}\,\dmi}.\qedhere
\end{split}\end{displaymath}
\end{proof}
%================================
%==========Section 5=============
%================================
\section{Volume Growth Conditions}\label{sec_vol}
For lazy random walk, \autocite[Corollaries 4.10 and 4.11]{Lyons-2018} presented return probability bounds depending on volume growth conditions. We have parallel results for simple random walk. Let us begin with the following proposition, which is comparable to the first assertionof \autocite[Proposition 4.7]{Lyons-2018}.
\begin{proposition}\label{vgrow_spec}
Let $G$ be a non-bipartite, finite or infinite, simple, connected, weighted graph with weight at least\/ $1$ for each edge. For each vertex $x\in V(G)$, $\delta\in(0,2)$, $\alpha\in(0,1)$, and $r\geqslant 0$,
$\vol(x,r)>\frac{w(x)}{(1-\alpha)^2\mu_x^\cq(\delta)}$ implies that $\mu_x^\cq(\delta)\leqslant \frac{\delta w(x)}{\alpha^2}r$.
\end{proposition}
\begin{proof}
We may assume $\mu_x^\cq(\delta)>0$. Fixing a vertex $x\in V(G)$, we define $f$ as in \Cref{def_f}.
\par
a) If $f(y)\geqslant 0$ for all $y\in V(G)$, then
\begin{displaymath}
\delta
\geqslant
\langle\cq f,f\rangle_w
=
\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2
\geqslant
\abs[\big]{f(x)}^2
=\frac{\mu_x^\cq(\delta)}{w(x)}.
\end{displaymath}
Note that $\vol(x,r)>\frac{w(x)}{(1-\alpha)^2\mu_x^\cq(\delta)}$ implies $r\geqslant 1$. Therefore,
\begin{displaymath}
\mu_x^\cq(\delta)
\leqslant
\delta w(x)
\leqslant
\frac{\delta w(x)}{\alpha^2}r.
\end{displaymath}
\par
b) From now on, we may assume that $T\coloneqq \bigl\{t\in V(G)\st f(t)<0\bigr\}$ is non-empty. We construct $G'$ and $g$ as in the proof of \Cref{reg_return_spec} and use notations there. Set
\begin{displaymath}
B^{(\alpha)}
\coloneqq
\bigl\{
y\in V' \st \abs{g(y)-g(x)}\leqslant \alpha g(x)
\bigr\}.
\end{displaymath}
It follows that $B^{(\alpha)}\subseteq S\cup T\subsetneq V'$. Since $G'$ is connected, there exists a shortest path $\cp=u_0u_1\cdots u_{\abs{\cp}}$ in $G'$ from $x$ to $V'\setminus B^{(\alpha)}$ with $u_0=x$. Hence,
\begin{displaymath}\begin{split}
\delta&\geqslant
\langle\cq f,f\rangle_w
=\sum_{(v_1,v_2)\in E(G)}w(v_1,v_2)\abs[\big]{f(v_1)+f(v_2)}^2\\
&\geqslant \sum_{(v_1,v_2)\in E'}w'(v_1,v_2)\abs[\big]{g(v_1)-g(v_2)}^2
\geqslant \sum_{i=0}^{\abs{\cp}-1}\abs[\big]{g(u_i)-g(u_{i+1})}^2\\
&\geqslant \frac{1}{\abs{\cp}}\biggl(\sum_{i=0}^{\abs{\cp}-1}\abs[\big]{g(u_i)-g(u_{i+1})}\biggr)^2
\geqslant \frac{1}{\abs{\cp}}\abs[\big]{g(u_0)-g(u_{\abs{\cp}})}^2\\
&> \frac{\alpha^2\abs[\big]{g(x)}^2}{\abs{\cp}}
=\frac{\alpha^2\mu_x^\cq(\delta)}{w(x)\abs{\cp}}.
\end{split}\end{displaymath}
\par
c) We claim that
\begin{equation}\label{aux_equal}
\bigl\{
y\in V' \st \dist(y,x;G')\leqslant \abs{\cp}-1
\bigr\}
=
\bigl\{
y\in V \st \dist(y,x;G)\leqslant \abs{\cp}-1
\bigr\}.
\end{equation}
In fact,
\begin{displaymath}
\bigl\{
y\in V' \st \dist(y,x;G')\leqslant \abs{\cp}-1
\bigr\}
\subseteq
B^{(\alpha)}\subseteq S\cup T\subseteq V.
\end{displaymath}
By the construction of $G'$, $\dist(y_1,y_2;G')\geqslant \dist(y_1,y_2;G)$ for $y_1,y_2\in V$. Therefore,
\begin{displaymath}
\bigl\{
y\in V' \st \dist(y,x;G')\leqslant \abs{\cp}-1
\bigr\}
\subseteq
\bigl\{
y\in V \st \dist(y,x;G)\leqslant \abs{\cp}-1
\bigr\}.
\end{displaymath}
Suppose the above inclusion is strict. Then there will be a vertex $\widetilde{v}\in V(G)$ that is not in the left-hand side of \Cref{aux_equal}, and a path $\widetilde{\cp}=\widetilde{u}_0\widetilde{u}_1\cdots\widetilde{u}_{\abs{\widetilde{\cp}}}$ in $G$ with $\abs{\widetilde{\cp}}\leqslant \abs{\cp}-1$, $\widetilde{u}_0=x$, and $\widetilde{u}_{\abs{\widetilde{\cp}}}=\widetilde{v}$. Set
\begin{displaymath}
L\coloneqq
\bigl\{
0\leqslant i\leqslant \abs{\widetilde{\cp}}-1
\st
\{\widetilde{u}_i,\widetilde{u}_{i+1}\}\subseteq S \text{ or } \{\widetilde{u}_i,\widetilde{u}_{i+1}\}\subseteq T
\bigr\}.
\end{displaymath}
Then $L$ must be non-empty, otherwise $\widetilde{v}\in V(G)$ would be in the left-hand side of \Cref{aux_equal}. Pick the smallest number $i^*$ in $L$ and write $e=(\widetilde{u}_{i^*},\widetilde{u}_{i^*+1})$. By the construction of $G'$ and $g$, we have $g(\widetilde{u}_{i^*}^{(e)})=0$.
Thus, $\widetilde{u}_0\widetilde{u}_1\cdots\widetilde{u}_{i^*}\widetilde{u}_{i^*}^{(e)}$ is a path in $G'$ linking $x$ and $V'\setminus B^{(\alpha)}$, whose length is $i^*+1\leqslant \abs{\widetilde{\cp}}<\abs{\cp}$. This is a contradiction. Therefore, \Cref{aux_equal} holds.
\par
d) If $\vol(x,r;G)>\frac{w(x)}{(1-\alpha)^2\mu_x^\cq(\delta)}$, we must have $r\geqslant \abs{\cp}$. In fact, we notice that
\begin{displaymath}
\vol(B^{(\alpha)};G')(1-\alpha)^2g^2(x)
\leqslant
\sum_{y\in B^{(\alpha)}}\abs[\big]{g(y)}^2w'(y)
\leqslant
\sum_{y\in V'}\abs[\big]{g(y)}^2w'(y)
=1.
\end{displaymath}
Therefore,
\begin{displaymath}
\vol(x,\abs{\cp}-1;G')
\leqslant
\vol(B^{(\alpha)};G')
\leqslant
\frac{1}{(1-\alpha)^2g^2(x)}
=
\frac{w(x)}{(1-\alpha)^2\mu_x^\cq(\delta)}.
\end{displaymath}
Hence, \Cref{aux_equal} gives
\begin{displaymath}
\vol(x,\abs{\cp}-1;G)
=
\vol(x,\abs{\cp}-1;G')
\leqslant
\frac{w(x)}{(1-\alpha)^2\mu_x^\cq(\delta)}.
\end{displaymath}
Consequently, $\vol(x,r)>\frac{w(x)}{(1-\alpha)^2\mu_x^\cq(\delta)}$ implies $r\geqslant \abs{\cp}$.
\par
e) By the results in b) and d), we have
\begin{displaymath}
\delta
\geqslant
\frac{\alpha^2\mu_x^\cq(\delta)}{w(x)\abs{\cp}}
\geqslant
\frac{\alpha^2\mu_x^\cq(\delta)}{w(x)r}.
\end{displaymath}
Hence, we arrive at the conclusion that $\mu_x^\cq(\delta)\leqslant \frac{\delta w(x)}{\alpha^2}r$.
\end{proof}
\begin{corollary}\label{prp472}
Let $G$ be a non-bipartite, finite or infinite, simple, connected, weighted graph with weight at least\/ $1$ for each edge. For each vertex $x\in V(G)$ and $\delta<\frac{1}{r\vol(x,r)}$, we have $\mu_x^\cq(\delta)\leqslant \frac{4w(x)}{\vol(x,r)}$.\qed
\end{corollary}
\Cref{prp472} is comparable to the second assertion of \autocite[Proposition 4.7]{Lyons-2018}. To get a proof, one need only exploit \Cref{vgrow_spec} and mimic the proof for the second assertion of \autocite[Proposition 4.7]{Lyons-2018}.
Now we present return probability bounds based on volume growth conditions.
\begin{corollary}\label{vgrow_pro}
Let $G$ be a non-bipartite, infinite, simple, connected, weighted graph with weight at least\/ $1$ for each edge. Suppose that $c > 0$ and $D \geqslant 1$ are constants such that\/ $\vol(x, r) \geqslant c (r+1)^D$ for all
$r \geqslant 0$. Then for all $\delta\in(0,2)$,
\begin{displaymath}\begin{split}
\mu^*_x(\delta) &\leqslant C w(x) \delta^{D/(D+1)},\\
\mu^\cq _x(\delta) &\leqslant C w(x) \delta^{D/(D+1)},
\end{split}\end{displaymath}
where
\begin{displaymath}
C \coloneqq \frac{(D+1)^2}{c^{1/(D+1)} D^{2D/(D+1)}}.
\end{displaymath}
Hence for all $t \geqslant 1$, simple random walk satisfies
\begin{displaymath}\begin{split}
p_t(x,x)
&\leqslant
2C' w(x) t^{-D/(D+1)}\qquad \text{for\quad} t\equiv 0\bmod 2,\\
p_t(x,x)
&\leqslant
C' w(x) t^{-D/(D+1)}
\qquad \text{for\quad} t\equiv 1\bmod 2,
\end{split}\end{displaymath}
where
\begin{displaymath}
C' \coloneqq \frac{D+1}{c^{1/(D+1)} D^{(D-1)/(D+1)}} \Gamma\Big(\frac{D}{D+1}\Big).
\end{displaymath}
\end{corollary}
\begin{proof}
The bound on $\mu^*_x$ is proved in \autocite[Corollary 4.10]{Lyons-2018}. A similar argument can be used to prove the bound on $\mu^\cq_x$: in lieu of \autocite[Eq.\@ (4.4)]{Lyons-2018}, we use \Cref{prp472}. To prove the bound on return probabilities, we set
\begin{numcases}
{\varphi(\lambda)\coloneqq }
C w(x) \lambda^{D/(D+1)}
\; \nonumber &if\/ $\lambda \in [0,1)\text{ and }C w(x) \lambda^{D/(D+1)}\leqslant \mu_x^*(1)$,\\
\mu_x^*(1) \; \nonumber &\text{for intermediate values of $\lambda$},\\
\mu_x^*(2) - C w(x) (2-\lambda)^{D/(D+1)}
\; \nonumber &if\/ $\lambda\in [1,2)\text{ and }C w(x) (2-\lambda)^{D/(D+1)}\geqslant \mu_x^*(1)$,
\end{numcases}
and
\begin{numcases}
{\psi_2(\lambda)\coloneqq }
C w(x) \lambda^{D/(D+1)}\wedge \mu_x^*(2)
\; \nonumber &if\/ $\lambda \in [0,1)$,\\
\mu_x^*(2)
\; \nonumber &if\/ $\lambda\in [1,2]$.
\end{numcases}
It is easy to see that $\varphi$ and $\psi_2$ satisfy the conditions in \Cref{spec_bound}. Therefore, by \Cref{spec_bound} and some elementary calculation, we have
\begin{displaymath}\begin{split}
p_t(x,x)&\leqslant
\frac{2 C w(x)D}{D+1}\int_{0}^{1}\lambda^{-1/(D+1)}(1-\lambda)^t\df{\lambda}
\qquad \text{for\quad} t\equiv 0\bmod 2,\\
p_t(x,x)&\leqslant
\frac{C w(x)D}{D+1}\int_{0}^{1}\lambda^{-1/(D+1)}(1-\lambda)^t\df{\lambda}
\qquad \text{for\quad} t\equiv 1\bmod 2.
\end{split}\end{displaymath}
But we have
\begin{displaymath}\begin{split}
\int_{0}^{1}\lambda^{-1/(D+1)}(1-\lambda)^t\df{\lambda}
&\leqslant
\int_{0}^{1}\lambda^{-1/(D+1)}\ue^{-\lambda t}\df{\lambda}
\leqslant
\int_{0}^{\infty}\lambda^{-1/(D+1)}\ue^{-\lambda t}\df{\lambda}\\
&=
\int_{0}^{\infty}(s/t)^{-1/(D+1)}\ue^{-s}\df{(s/t)}
=
t^{-D/(D+1)}\int_{0}^{\infty}s^{-1/(D+1)}\ue^{-s}\df{s}\\
&=
t^{-D/(D+1)}\Gamma\bigl(\frac{D}{D+1}\bigr),
\end{split}\end{displaymath}
where we introduce a change of variable $\lambda t=s$ to get the first equality. Hence, the bound on return probabilities is proved.
\end{proof}
\Cref{vgrow_pro} is comparable to \autocite[Corollary~4.10]{Lyons-2018}. Using a similar method, one may get an analogue of \autocite[Corollarie~4.11]{Lyons-2018}. Details are omitted.
%================================
%==========Section 6=============
%================================
\section{Mixing Time Bound}\label{sec_mix}
We are concerned in this section with mixing time bounds, which are based on the bounds on vertex spectral measures in \Cref{mix_spec_prep}. As a preparation, we introduce the following graph parameter.
\begin{definition}\label{def_kudos}
Let $G$ be a finite and weighted graph.
We call
\begin{displaymath}
\min\biggl\{
\frac{\sum_{(v,u)\in E(G)}w(v,u)\abs{f(v)+f(u)}^2}{\max_{y\in V}\abs{f(y)}^2}
\st
\min_{y\in V}f(y)<0< \max_{y\in V}f(y)
\biggr\}
\end{displaymath}
the \deef{energy efficiency} of $G$ and denote it as $\kg$.
\end{definition}
For bipartite graphs, it is easy to see that their energy efficiency must be $0$. For non-bipartite graphs, we bound their energy efficiency in the following lemma.
\begin{lemma}\label{est}
Let $G$ be a connected and weighted graph, with weight at least\/ $1$ for each edge.
\begin{enumerate}[\upshape(1)]
\item Assume that $f$ is a function on $V(G)$ and $\cp=z_0z_1\cdots z_{\abs{\cp}}$ is an edge-simple path. Then
\begin{displaymath}
\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2\geqslant \frac{1}{\abs{P}}\bigl(f(z_0)-(-1)^{\abs{\cp}}f(z_{\abs{\cp}})\bigr)^2.
\end{displaymath}
\item If $G$ is also non-bipartite, finite, and simple, then $\kg\geqslant \frac{1}{\D(G)+1}$.
\end{enumerate}
\end{lemma}
\begin{proof}
a) Note that
\begin{displaymath}
\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2
\geqslant \sum_{i=0}^{\abs{P}-1}\abs[\big]{f(z_i)+f(z_{i+1})}^2
\end{displaymath}
By the Cauchy--Schwarz inequality, we have
\begin{displaymath}\begin{split}
\sum_{i=0}^{\abs{P}-1}\abs[\big]{f(z_i)+f(z_{i+1})}^2
&=
\sum_{i=0}^{\abs{P}-1}\abs[\big]{(-1)^if(z_i)-(-1)^{i+1}f(z_{i+1})}^2\\
&\geqslant \frac{1}{\abs{P}}\biggl(\sum_{i=0}^{\abs{P}-1}\bigl((-1)^if(z_i)-(-1)^{i+1}f(z_{i+1})\bigr)\biggr)^2\\
&= \frac{1}{\abs{P}}\bigl(f(z_0)-(-1)^{\abs{\cp}}f(z_{\abs{\cp}})\bigr)^2.
\end{split}\end{displaymath}
The first assertion is proved.
\par
b) Now we deal with the second assertion. Take $f$ satisfying the constraints in \Cref{def_kudos} such that
\begin{displaymath}
\kg
=
\frac{\sum_{(v,u)\in E(G)}w(v,u)\abs{f(v)+f(u)}^2}{\max_{y\in V}\abs{f(y)}^2}.
\end{displaymath}
Assume that $\abs{f}$ attains its maximum at $x$, then
\begin{displaymath}
\kg
=
\frac{\sum_{(v,u)\in E(G)}w(v,u)\abs{f(v)+f(u)}^2}{\abs{f(x)}^2}.
\end{displaymath}
We may assume $f(x)> 0$ without loss of generality.
Set
\begin{displaymath}
S\coloneqq \bigl\{s\in V\st f(s)\geqslant 0\bigr\},\qquad T\coloneqq \bigl\{t\in V\st f(t)<0\bigr\}.
\end{displaymath}
By the assumptions, $S$ and $T$ are both non-empty. Because $G$ is non-bipartite, there exists an edge $(s_1,s_2)\in E$ with $s_1,s_2\in S$, or an edge $(t_1,t_2)\in E$ with $t_1,t_2\in T$. \vadjust{\kern2pt}%
\par
c) If there is an edge $(s_1,s_2)\in E$ with $s_1,s_2\in S$, let $\widehat{\cp}_0$ be a shortest path from \vadjust{\kern2pt}%
$x$ to $\{s_1,s_2\}$.
Without loss of generality, we may assume $\widehat{\cp}_0$ is from $x$ to $s_1$. If $\abs{\widehat{\cp}_0}$ is odd, \vadjust{\kern2pt}%
we set $\widehat{\cp}\coloneqq \widehat{\cp}_0$; if $\abs{\widehat{\cp}_0}$ is even, we set $\widehat{\cp}\coloneqq \widehat{\cp}_0.(s_1,s_2)$, \vadjust{\kern2pt}%
the concatenation of $\widehat{\cp}_0$ and the edge $(s_1,s_2)$.
Hence, $\widehat{\cp}$ is a path of odd length in any case. Assume
\begin{displaymath}
\widehat{\cp}=u_0u_1\cdots u_k,
\end{displaymath}
with $u_0=x$. Then $k$ is odd and $k\leqslant \D(G)+1$. Hence, by the first assertion,
\begin{displaymath}\begin{split}
\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2
\geqslant \frac{1}{k}\bigl(f(u_0)+f(u_k)\bigr)^2
\geqslant \frac{1}{k}\abs[\big]{f(x)}^2
\geqslant \frac{\abs[\big]{f(x)}^2}{\D(G)+1}.
\end{split}\end{displaymath}
Therefore, in this case,
\begin{displaymath}
\kg\geqslant \frac{1}{\D(G)+1}.
\end{displaymath}
\par
d) If there is an edge $(t_1,t_2)\in E$ with $t_1,t_2\in T$, \vadjust{\kern2pt}%
let $\overline{\cp}_0$ be a shortest path from $x$ to $\{t_1,t_2\}$.
Without loss of generality, we may assume $\overline{\cp}_0$ is from $x$ to $t_1$. If $\abs{\overline{\cp}_0}$ is even, \vadjust{\kern2pt}%
we set $\overline{\cp}\coloneqq \overline{\cp}_0$; if $\abs{\overline{\cp}_0}$ is odd, we set $\overline{\cp}\coloneqq \overline{\cp}_0.(t_1,t_2)$. Hence, $\overline{\cp}$ is a path of even length in any case. Assume
\begin{displaymath}
\overline{\cp}=v_0v_1\cdots v_\ell,
\end{displaymath}
with $v_0=x$. Then $\ell$ is even and $\ell\leqslant \D(G)+1$. Hence, by the first assertion,
\begin{displaymath}\begin{split}
\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2
\geqslant \frac{1}{\ell}\bigl(f(v_0)-f(v_{\ell})\bigr)^2
> \frac{1}{\ell}\abs[\big]{f(x)}^2
\geqslant \frac{\abs[\big]{f(x)}^2}{\D(G)+1}.
\end{split}\end{displaymath}
Therefore, in this case,
\begin{displaymath}
\kg\geqslant \frac{1}{\D(G)+1}.\qedhere
\end{displaymath}
\end{proof}
Now, we present our bounds on vertex spectral measures.
\begin{lemma}\label{mix_spec_prep}
Let $G$ be a finite, simple, connected, weighted graph, with weight at least\/ $1$ for each edge.
\begin{enumerate}[\upshape (1)]
\item For $\delta\in[0,2)$ and $x\in V$, $\mu_x^*(\delta) + \pi(x)\leqslant \rdiam(G)w(x)\delta$.
\item If $G$ is also non-bipartite, for $\delta\in[0,2)$ and
$x\in V$, $\mu_x^\cq(\delta) \leqslant \frac{w(x)\delta}{\kg}$. Moreover, by \Cref{est}, we have
$\mu_x^\cq(\delta)\leqslant \bigl(\D(G)+1\bigr)w(x)\delta$ for $\delta\in[0,2)$.
\end{enumerate}
\end{lemma}
\begin{proof}
The first assertion follows immediately from \autocite[Proposition~4.2]{Lyons-2018}. We now deal with the second assertion. Fix a vertex $x\in V$ and define $f$ as in \Cref{def_f}.
\par
%a) When $\delta=2$, $(\D(G)+1)w(x)\delta\geqslant 4$ and $\mu_x^\cq(\delta)\leqslant 1$. The desired inequality holds. UNDER CONSTRUCTION.
%obviously $f(y)\geqslant 0$ for all $y\in V$. So,
%\begin{displaymath}\begin{split}
%\bigl(\D(G)+1\bigr)\delta
%&\geqslant \delta
%\geqslant
%\langle\cq f,f\rangle_w \\
%&=\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2\\
%&\geqslant \abs[\big]{f(x)}^2
%=\frac{\mu_x^\cq(\delta)}{w(x)}.
%\end{split}\end{displaymath}
\par
When $\delta\in[0,\lambda_{\mathrm{min}}^\cq)$, $\mu_x^\cq(\delta)=0$ by definition. So the inequality holds automatically.
\par
When $\lambda^\cq_{\mathrm{min}}\leqslant \delta<\lambda^\cq_{\mathrm{max}}=2$, we know that $f$ is orthogonal to the eigenspace of $\cq$ corresponding to $\lambda^\cq_{\mathrm{max}}=2$, which is spanned by $(1,1,\ldots, 1)$, the constant vector. Therefore, by \Cref{def_kudos}, we have
\begin{displaymath}
\delta \geqslant \langle f,\cq f\rangle_w
=
\sum_{(v,u)\in E(G)}w(v,u)\abs{f(v)+f(u)}^2
\geqslant f(x)^2\kg
=\frac{\mu_x^\cq(\delta)\kg}{w(x)}.\qedhere
\end{displaymath}
\end{proof}
By \Cref{est}(2), \Cref{mix_spec_prep}(2), and \Cref{spec_loc}, we have the following corollary.
\begin{corollary}\label{Landau_imp}
For any non-bipartite, finite, simple, connected, weighted graph $G$ with weight at least\/ $1$ for each edge, we have
\begin{displaymath}
\lambda^\cq_k\geqslant \frac{k\kg}{\vol(V)}\geqslant \frac{k}{\bigl(\D(G)+1\bigr)\vol(V)}
\end{displaymath}
for\/ $1\leqslant k\leqslant n$. Therefore,
\begin{displaymath}
\pushQED{\qed}
\lambda^P_k
\geqslant
-1+\frac{k\kg}{\vol(V)}
\geqslant
-1+\frac{k}{\bigl(\D(G)+1\bigr)\vol(V)}.\qedhere
\popQED
\end{displaymath}
\end{corollary}
As mentioned in \Cref{ssec_results}, \Cref{Landau_imp}, combined with \autocite[Proposition~4.2]{Lyons-2018}, improves the result in \autocite{Landau-1981}.
To get a mixing time bound, we first give a bound on return probabilities using \Cref{mix_spec_prep}.
\begin{corollary}\label{mix_even_prep}
Let $G$ be a non-bipartite, finite, simple, connected, unweighted graph, and $t\equiv 0\bmod 2$. Then we have
\begin{displaymath}
0\leqslant
\frac{p_{t}(x,x)-\pi(x)}{\pi(x)}
\leqslant
\frac{2\bigl(\D(G)+1\bigr)\vol(V)}{t}.
\end{displaymath}
\end{corollary}
\begin{proof}
Set
\begin{numcases}
{\widetilde{\varphi}(\lambda)\coloneqq }
\rdiam(G)w(x)\lambda
\; \nonumber &if\/ $\lambda\in [0,1)\text{ and }\rdiam(G)w(x)\lambda\leqslant \mu_x^*(1)$,\\
\mu_x^*(1)\; \nonumber &\text{for intermediate values of $\lambda$},\\
\mu_x^*(2)-\bigl(\D(G)+1\bigr)w(x)(2-\lambda)
\; \nonumber
&\hspace{-5pt}$\begin{array}{l}
\text{if }\lambda\in[1,2]\text{ and }\\
\mu_x^*(2)-\bigl(\D(G)+1\bigr)w(x)(2-\lambda)\geqslant \mu_x^*(1).
\end{array}$
\end{numcases}
By \Cref{mix_spec_prep}, using the argument as in part a) of the proof of \Cref{reg_return}, the function $\widetilde{\varphi}$ defined above satisfies the conditions in \Cref{spec_bound}(1). Therefore, for $t\equiv 0\bmod 2$,
\begin{displaymath}\begin{split}
0\leqslant p_t(x,x)-\pi(x)&\leqslant \int_{0}^{2}(1-\lambda)^t\widetilde{\varphi}'(\lambda)\df{\lambda}\\
&\leqslant \rdiam(G)w(x)\int_{0}^{1}(1-\lambda)^t\df{\lambda}+\bigl(\D(G)+1\bigr)w(x)\int_{0}^{1}(1-\lambda)^t\df{\lambda}\\
&\leqslant 2\bigl(\D(G)+1\bigr)w(x)\int_{0}^{1}(1-\lambda)^t\df{\lambda}.
\end{split}\end{displaymath}
Hence,
\begin{displaymath}
\frac{p_{t}(x,x)-\pi(x)}{\pi(x)}
\leqslant 2\bigl(\D(G)+1\bigr)\frac{w(x)}{\pi(x)}\int_{0}^{1}(1-\lambda)^t\df{\lambda}.
\end{displaymath}
But we know $\frac{w(x)}{\pi(x)}=\vol(V)$ and
\begin{displaymath}
\int_{0}^{1}(1-\lambda)^t\df{\lambda}
\leqslant \int_{0}^{1}\exp\{-\lambda t\}\df{\lambda}
\leqslant\int_{0}^{\infty}\exp\{-\lambda t\}\df{\lambda}=\frac{1}{t}.
\end{displaymath}
Therefore,
\begin{displaymath}
\frac{p_{t}(x,x)-\pi(x)}{\pi(x)}\leqslant \frac{2\bigl(\D(G)+1\bigr)\vol(V)}{t}.\qedhere
\end{displaymath}
\end{proof}
We are now almost in position to give the following mixing time bound.
\begin{corollary}\label{mixx}
For a non-bipartite, finite, simple, connected, unweighted graph $G$, the uniform mixing time of the simple random walk on $G$ satisfies
\begin{displaymath}
\tiq\leqslant 8n^3.
\end{displaymath}
If we further assume that $G$ is regular, then $\tiq\leqslant 24n^2$.
\end{corollary}
In the proof of \Cref{mixx}, the following lemma will be useful.
\begin{lemma}\label{mix_mono_prep}
For a reversible Markov chain with stationary distribution $\pi$, we have
\begin{displaymath}
\abs[\bigg]{\frac{p_{2t}(x,y)-\pi(y)}{\pi(y)}}
\leqslant
\sqrt{\frac{p_{2t}(x,x)-\pi(x)}{\pi(x)}}\sqrt{\frac{p_{2t}(y,y)-\pi(y)}{\pi(y)}}.
\end{displaymath}
In addition, $\max_{x,y}{\abs[\big]{\frac{p_{t}(x,y)-\pi(y)}{\pi(y)}}}$ is decreasing in $t$.
\end{lemma}
\begin{proof}
One may refer to the proof of \autocite[Proposition~A.1]{Lyons-2018}. The monotonicity is mentioned in \autocite{Lyons-2021}.
\end{proof}
\begin{proof}[Proof of \Cref{mixx}]
a) Combining \Cref{mix_mono_prep} and \Cref{mix_even_prep}, for $t\equiv 0\bmod 2$, we have
\begin{displaymath}
\max_{x,y}{\abs[\Big]{\frac{p_{t}(x,y)-\pi(y)}{\pi(y)}}}\leqslant \max_x\frac{p_{t}(x,x)-\pi(x)}{\pi(x)}\leqslant
\frac{2\bigl(\D(G)+1\bigr)\vol(V)}{t}.
\end{displaymath}
In addition, we notice that $\vol(V)\leqslant n(n-1)$.
Hence, the second assertion of \Cref{mix_mono_prep} ensures
\begin{displaymath}
\tiq\leqslant \Bigl\lceil8\bigl(\D(G)+1\bigr)\vol(V)\Bigr\rceil
\leqslant \Bigl\lceil8n\cdot n(n-1)\Bigr\rceil
\leqslant 8n^3.
\end{displaymath}
\par
b) To prove the second assertion, we assume $G$ is $d$-regular. By \Cref{reg_diam}, we have
\begin{displaymath}
\D(G)+1\leqslant \frac{3n}{d+1}.
\end{displaymath}
In addition, $\vol(V)=nd$.
By the second assertion of \Cref{mix_mono_prep},
\begin{displaymath}
\tiq\leqslant \Bigl\lceil8\bigl(\D(G)+1\bigr)\vol(V)\Bigr\rceil
\leqslant \Bigl\lceil8\cdot\frac{3n}{d+1}\cdot nd\Bigr\rceil
\leqslant 24n^2.\qedhere
\end{displaymath}
\end{proof}
%================================
%==========Section 7=============
%================================
\clearpage
\section{Transitive Case}\label{sec_trans}
We consider vertex-transitive graphs in this section.
\subsection{Estimate of Spectral Measure}
Given a vertex-transitive graph $G$, set
\begin{displaymath}
N^\#(r)\coloneqq \bigl\{v\in V(G)\st \dist(v,x)\leqslant r\bigr\}.
\end{displaymath}
This value does not depend on the choice of $x$ since $G$ is vertex-transitive. Furthermore, as in
\autocite[Lemma~6.4]{Lyons-2018}, for every two vertices $x, y \in V$, it is easy to see that $\norm{F^\cq_x}_w=\norm{F^\cq_y}_w$.
\begin{lemma}
When $G$ is vertex-transitive, for each $x\in V$,
\begin{displaymath}
\cq F_x^\cq=\sum_{v\in N(x)} \frac{w(x,v)}{w}(F_x^\cq+F_v^\cq).
\end{displaymath}
\end{lemma}
\begin{proof}
By the definition of $F_x^\cq$,
\begin{displaymath}
\cq F_x^\cq=\frac{\cq I_\cq(\delta)\ind_x}{w}=\frac{I_\cq(\delta)\cq \ind_x}{w}.
\end{displaymath}
However, by direct calculation,
\begin{displaymath}
\cq \ind_x=\sum_{v\in N(x)}\frac{w(x,v)}{w}(\ind_x+\ind_v).
\end{displaymath}
Thus, we have
\begin{displaymath}\begin{split}
\cq F_x^\cq&=\frac{1}{w}I_\cq(\delta)\sum_{v\in N(x)}\frac{w(x,v)}{w}(\ind_x+\ind_v)\\
&=
\sum_{v\in N(x)}\frac{w(x,v)}{w}\Bigl(\frac{I_\cq(\delta)\ind_x}{w}+\frac{I_\cq(\delta)\ind_v}{w}\Bigr)\\
&=\sum_{v\in N(x)}\frac{w(x,v)}{w}(F_x^\cq+F_v^\cq).\qedhere
\end{split}\end{displaymath}
\end{proof}
\begin{theorem}\label{trans_spec}
Let $G$ be a vertex-transitive, non-bipartite, simple, connected, weighted graph with weight at least\/ $1$ for each edge. For each $x\in V(G)$, $c\in(0,1)$, and $\delta\in\bigl(0,\frac{2}{w}\bigr]$, we have
\begin{displaymath}
\mu^\cq_x(\delta)\leqslant
\frac{1}{c^2N^\#\Bigl(\frac{\arcsin\sqrt{(1-c)/2}}{\arcsin\sqrt{w\delta/2}}\,\Bigr)}.
\end{displaymath}
Furthermore, for $\delta\in (0,2]$,
\begin{displaymath}
\mu^\cq_x(\delta)
\leqslant
\frac{1}{c^2N^\#\Bigl(\frac{\sqrt{1-c}}{\sqrt{w\delta}}\Bigr)}.
\end{displaymath}
\end{theorem}
\begin{proof}
Fix a vertex $x\in V$.
\par
a) Consider the spectral embedding $\{F^\cq_v\st v\in V\}$ based on $\cq$. Let
\begin{displaymath}
\rho\coloneqq \norm{F^\cq_x}_w,\qquad \beta_0\coloneqq \max_{v\in N(x)}\norm{F^\cq_x+F^\cq_v}_w.
\end{displaymath}
We have
\begin{displaymath}\begin{split}
\delta&\geqslant \frac{1}{\norm{F^\cq_x}_w^2}\bigl\langle \cq F^\cq_x,F^\cq_x\bigr\rangle_w
= \frac{1}{\norm{F^\cq_x}_w^2}\biggl\langle \sum_{v\in N(x)}\frac{w(x,v)}{w}(F^\cq_x+F^\cq_v),F^\cq_x \biggr\rangle_w \\
&= \frac{1}{\norm{F^\cq_x}_w^2}\sum_{v\in N(x)}\frac{w(x,v)}{w}\bigl(\norm{F^\cq_x}_w^2+
\bigl\langle F^\cq_x,F^\cq_v\bigr\rangle_w\bigr)\\
&= \frac{1}{\norm{F^\cq_x}_w^2}\sum_{v\in N(x)}\frac{w(x,v)}{w}\frac{\norm{F^\cq_x+F^\cq_v}_w^2}{2}\\
&\geqslant \frac{1}{2w \rho^2}\max_{v\in N(x)}\norm{F^\cq_x+F^\cq_v}^2_w=\frac{1}{2w \rho^2}\beta_0^2.
\end{split}\end{displaymath}
In other words, $\frac{\beta_0}{2\rho}\leqslant \sqrt{w \delta/2}$. Since we assumed that $\delta\in (0,\frac{2}{w}]$, by the monotonicity of $\arcsin$ function on the interval $[0,1]$, we have
\begin{displaymath}
2\arcsin{\frac{\beta_0}{2\rho}}\leqslant 2\arcsin{\sqrt{w \delta/2}}.
\end{displaymath}
\par
b) Set $B_0\coloneqq B\Bigl(x,\frac{\arcsin{\sqrt{(1-c)/2}}}{\arcsin{\sqrt{(w \delta)/2}}}\Bigr)$. \vadjust{\kern2pt}%
For $v\in B_0$, let $\cp=u_0u_1\cdots u_{r-1}u_r$ be a shortest path joining $x$ and $v$, with $u_0=x$ and $u_r=v$. Define
\begin{displaymath}
\widetilde{F}_i\coloneqq (-1)^iF^\cq_{u_i},\qquad i=0,1,2,\ldots,r.
\end{displaymath}
Then $r\leqslant \frac{\arcsin{\sqrt{(1-c)/2}}}{\arcsin{\sqrt{(w \delta)/2}}}$, and
\begin{displaymath}
\norm[\big]{\widetilde{F}_i-\widetilde{F}_{i+1}}_w
=\norm[\big]{F^\cq_{u_i}+F^\cq_{u_{i+1}}}_w, \qquad i=0,1,2,\ldots,r-1.
\end{displaymath}
Write $\theta(h_1,h_2)$ for the angle between $h_1,h_2\in \ell^2(V,w)$. Then using the sphere metric, we get
\begin{equation}\label{trans_cal}\begin{split}
\theta\bigl(\widetilde{F}_0,\widetilde{F}_r\bigr)
&\leqslant \sum_{i=0}^{r-1}\theta\bigl(\widetilde{F}_i,\widetilde{F}_{i+1}\bigr)\\
&=\sum_{i=0}^{r-1}2\arcsin{\frac{\norm[\big]{\widetilde{F}_i-\widetilde{F}_{i+1}}_w}{2\rho}}
=\sum_{i=0}^{r-1}2\arcsin{\frac{\norm[\big]{F^\cq_{u_i}+F^\cq_{u_{i+1}}}_w}{2\rho}}\\
&\leqslant \sum_{i=0}^{r-1}2\arcsin{\frac{\beta_0}{2\rho}}
\leqslant 2r\arcsin{\sqrt{w \delta/2}}\\
&\leqslant 2\frac{\arcsin{\sqrt{(1-c)/2}}}{\arcsin{\sqrt{w \delta/2}}}\arcsin{\sqrt{w \delta/2}}
=2\arcsin{\sqrt{(1-c)/2}}\\
&=\arccos c.
\end{split}\end{equation}
Thus,
\begin{displaymath}
\abs[\big]{\cos{\theta\bigl(F^\cq_x,F^\cq_v\bigr)}}=\cos{\theta\bigl(\widetilde{F}_0,\widetilde{F}_r\bigr)}\geqslant c.
\end{displaymath}
In summary, if $v\in B_0=B\Bigl(x,\frac{\arcsin{\sqrt{(1-c)/2}}}{\arcsin{\sqrt{(w \delta)/2}}}\Bigr)$,
\begin{displaymath}
\abs[\big]{\cos{\theta\bigl(F^\cq_x,F^\cq_v\bigr)}}\geqslant c.
\end{displaymath}
\par
c) By the above discussion,
\begin{displaymath}\begin{split}
\rho^2&
=\norm{F^\cq_x}_w^2
=\sum_{v\in V}w \abs[\big]{F^\cq_x(v)}^2\\
&=\sum_{v\in V}w \abs[\big]{\bigl\langle F^\cq_x,F^\cq_v\bigr\rangle_w}^2
\geqslant w \sum_{v\in B_0}\abs[\big]{\bigl\langle F^\cq_x,F^\cq_v\bigr\rangle_w}^2\\
&= w \sum_{v\in B_0}\abs[\big]{\cos{\theta\bigl(F^\cq_x,F^\cq_v\bigr)}}^2\norm{F^\cq_x}_w^2\norm{F^\cq_v}_w^2 \\
&\geqslant w \rho^4c^2 N^\#\Bigl(\frac{\arcsin\sqrt{(1-c)/2}}{\arcsin\sqrt{w\delta/2}}\,\Bigr).
\end{split}\end{displaymath}
Thus, we have
\begin{displaymath}
\rho^2
\leqslant \frac{1}{wc^2N^\#\Bigl(\frac{\arcsin\sqrt{(1-c)/2}}{\arcsin\sqrt{w\delta/2}}\Bigr)}.
\end{displaymath}
Now, by \Cref{spec_embed_norm},
\begin{displaymath}
\mu_x^\cq(\delta)=w\norm{F^\cq_x}_w^2
=w\rho^2
\leqslant \frac{1}{c^2N^\#\Bigl(\frac{\arcsin\sqrt{(1-c)/2}}{\arcsin\sqrt{w\delta/2}}\Bigr)}.
\end{displaymath}
The first assertion is proved.
\par
d) Now we turn to the second assertion.
If $(1-c)<w \delta$, then
\begin{displaymath}
N^\#\Bigl(\frac{\sqrt{1-c}}{\sqrt{w\delta}}\,\Bigr)=1.
\end{displaymath}
Therefore,
\begin{displaymath}
\frac{1}{c^2N^\#\Bigl(\frac{\sqrt{1-c}}{\sqrt{w\delta}}\Bigr)}>1.
\end{displaymath}
But $\mu_x^\cq(\delta)\leqslant 1$. So the second assertion holds in this case.
If $(1-c)\geqslant w \delta$, then $\sqrt{w \delta/2}<1$. Thus, $\arcsin{\sqrt{w \delta/2}}$ is defined. Note that $\frac{x}{\arcsin{x}}$ is decreasing for $x\in (0,1)$. Therefore,
\begin{displaymath}
\frac{\sqrt{(1-c)/2}}{\arcsin{\sqrt{(1-c)/2}}}
\leqslant
\frac{\sqrt{w \delta/2}}{\arcsin{\sqrt{w \delta/2}}}.
\end{displaymath}
Consequently,
\begin{displaymath}
\frac{\sqrt{1-c}}{\sqrt{w \delta}}
\leqslant
\frac{\arcsin{\sqrt{(1-c)/2}}}{\arcsin{\sqrt{w \delta/2}}}.
\end{displaymath}
Hence, the second assertion follows from the first one immediately.
\end{proof}
By the bounds on vertex spectral measures in \Cref{trans_spec} and \autocite[Theorem~6.1]{Lyons-2018}, using the method as in \Cref{reg_return}, we may get a similar result as \autocite[Corollary~6.6]{Lyons-2018}.
\begin{corollary}\label{trans_return}
Let $G$ be a finite or infinite, vertex-transitive, $d$-regular, non-bipartite, simple, connected, unweighted graph with at least polynomial growth rate $N^\#(r)\geqslant Cr^D$, where $C > 0$ and $D \geqslant 1$ are constants and\/ $0 \leqslant r \leqslant \D(G)$.
Then for each $x\in V$ and $t > 0$,
\begin{displaymath}\begin{split}
&0\leqslant p_t(x,x)-\pi(x)
\leqslant 2\widetilde{C}t^{-D/2} \qquad \text{for\quad} t\equiv 0\bmod 2,\\
&\abs[\big]{p_t(x,x)-\pi(x)}
\leqslant \widetilde{C}t^{-D/2} \qquad \text{for\quad} t\equiv 1\bmod 2,
\end{split}\end{displaymath}
where $\widetilde{C}=\frac{(D+4)^{D/2+2}d^{D/2}}{32CD^{D/2-1}}\Gamma\bigl(\frac{D}{2}\bigr)$.\qed
\end{corollary}
The proof of this corollary is omitted. Note that the rate $t^{-D/2}$ here is the correct decay rate for the simple random walk on $\ZN^D$, $D\in \NN$.
An analogue of \autocite[Corollary~6.7]{Lyons-2018} could also be written down easily.
\subsection{Minimum Eigenvalue}
For a connected, finite graph $G$ and its signless probabilistic Laplacian $\cq=I+P$, we know that $0$ is an eigenvalue of $\cq$ if and only if $G$ is bipartite from \Cref{0span}.
\begin{theorem}
Let $G$ be a vertex-transitive, non-bipartite, finite, simple, connected, weighted graph with weight at least\/ $1$ for each edge. Then
\begin{displaymath}
\lambda^\cq_{\mathrm{min}}
\geqslant
\frac{2}{w}\Bigl(\sin\frac{\pi}{4\bigl(\D(G)+1\bigr)}\Bigr)^2.
\end{displaymath}
\end{theorem}
\begin{proof}
We may assume $\frac{\lambda^\cq_{\mathrm{min}}}{2}<1$: otherwise, the inequality is trivial.
Consider the spectral embedding based on $\cq$ with $\delta=\lambda^\cq_{\mathrm{min}}$. Fix $x\in V$ and let
\begin{displaymath}
S\coloneqq \bigl\{s\in V\st f(s)\geqslant 0\bigr\},\qquad T\coloneqq \bigl\{t\in V\st f(t)<0\bigr\}.
\end{displaymath}
It is easy to see that both $S$ and $T$ are non-empty. Because $G$ is assumed to be non-bipartite, there exists an edge $(s_1,s_2)\in E$ with $s_1,s_2\in S$, or an edge $(t_1,t_2)\in E$ with $t_1,t_2\in T$. \vadjust{\kern2pt}%
\par
a) If there is an edge $(s_1,s_2)\in E$ with $s_1,s_2\in S$, \vadjust{\kern2pt}%
let $\widehat{\cp}_0$ be a shortest path from $x$ to $\{s_1,s_2\}$. Without loss of generality, we may assume $\widehat{\cp}_0$ is from $x$ to $s_1$. \vadjust{\kern2pt}%
If $\abs{\widehat{\cp}_0}$ is odd, we set $\widehat{\cp}\coloneqq \widehat{\cp}_0$; if $\abs{\widehat{\cp}_0}$ is even, we set $\widehat{\cp}\coloneqq \widehat{\cp}_0.(s_1,s_2)$, the concatenation of $\widehat{\cp}_0$ and the edge $(s_1,s_2)$. Hence, $\widehat{\cp}$ is a path of odd length in any case. Assume
\begin{displaymath}
\widehat{\cp}=u_0u_1\cdots u_k,
\end{displaymath}
with $u_0=x$. We define
\begin{displaymath}
\widehat{F}_i\coloneqq (-1)^iF^\cq_{u_i},\qquad i=0,1,2,\ldots,k.
\end{displaymath}
Then $\widehat{F}_0= F^\cq_x$ and
\begin{displaymath}\begin{split}
\bigl\langle\widehat{F}_0,\widehat{F}_k\bigr\rangle_w
&= (-1)^k \bigl\langle F^\cq_x,F^\cq_{u_k}\bigr\rangle_w
=-\bigl\langle F^\cq_x,F^\cq_{u_k}\bigr\rangle_w
=-\langle I_\cq(\delta)\ind_x/w,I_\cq(\delta)\ind_{u_k}/w\bigr\rangle_w \\
&=-\langle I_\cq(\delta)\ind_x/w,\ind_{u_k}/w\bigr\rangle_w
=-\langle I_\cq(\delta)\ind_x/w,\ind_{u_k}\bigr\rangle
=-\langle F^\cq_x,\ind_{u_k}\bigr\rangle\\
&=-F^\cq_x(u_k)\leqslant 0.
\end{split}\end{displaymath}
Hence,
\begin{equation}\label{dunjiao_1}
\frac{\pi}{2}\leqslant \theta\bigl(\widehat{F}_0,\widehat{F}_k\bigr).
\end{equation}
In addition, using the sphere metric as in \Cref{trans_cal}, we have
\begin{equation}\label{angle_est_1}
\theta\bigl(\widehat{F}_0,\widehat{F}_k\bigr)
\leqslant 2k\arcsin{\sqrt{w {\lambda}^\cq_{\mathrm{min}}/2}}
\leqslant 2\bigl(\D(G)+1\bigr)\arcsin{\sqrt{w {\lambda}^\cq_{\mathrm{min}}/2}}.
\end{equation}
Combining \Cref{dunjiao_1,angle_est_1},
\begin{displaymath}
\frac{\pi}{2}\leqslant 2\bigl(\D(G)+1\bigr)\arcsin{\sqrt{w {\lambda}^\cq_{\mathrm{min}}/2}}.
\end{displaymath}
So in this case,
\begin{displaymath}
\lambda^\cq_{\mathrm{min}}
\geqslant
\frac{2}{w}\Bigl(\sin\frac{\pi}{4\bigl(\D(G)+1\bigr)}\Bigr)^2.
\end{displaymath}
\par
b) If there is an edge $(t_1,t_2)\in E$ with $t_1,t_2\in T$, \vadjust{\kern2pt}%
let $\overline{\cp}_0$ be a shortest path from $x$ to $\{t_1,t_2\}$. Without loss of generality, we may assume $\overline{\cp}_0$ is from $x$ to $t_1$. \vadjust{\kern2pt}%
If $\abs{\overline{\cp}_0}$ is even, we set $\overline{\cp}\coloneqq \overline{\cp}_0$; if $\abs{\overline{\cp}_0}$ is odd, we set $\overline{\cp}\coloneqq \overline{\cp}_0.(t_1,t_2)$. Thus, $\overline{\cp}$ is a path of even length in any case. Assume
\begin{displaymath}
\overline{\cp}=v_0v_1\cdots v_\ell,
\end{displaymath}
with $v_0=x$, and define
\begin{displaymath}
\overline{F}_i\coloneqq (-1)^iF^\cq_{v_i},\qquad i=0,1,2,\ldots,\ell.
\end{displaymath}
Then $\overline{F}_0= F^\cq_x$ and
\begin{displaymath}
\bigl\langle\overline{F}_0,\overline{F}_\ell\bigr\rangle_w
= (-1)^\ell \bigl\langle F^\cq_x,F^\cq_{v_\ell}\bigr\rangle_w
=F^\cq_x(v_\ell)< 0.
\end{displaymath}
Hence, we have
\begin{equation}\label{dunjiao_2}
\frac{\pi}{2}<\theta\bigl(\overline{F}_0,\overline{F}_\ell\bigr).
\end{equation}
In addition, using the sphere metric as in \Cref{trans_cal} again, we have
\begin{equation}\label{angle_est_2}
\theta\bigl(\overline{F}_0,\overline{F}_r\bigr)
\leqslant 2\ell\arcsin{\sqrt{w {\lambda}^\cq_{\mathrm{min}}/2}}
\leqslant 2\bigl(\D(G)+1\bigr)\arcsin{\sqrt{w {\lambda}^\cq_{\mathrm{min}}/2}}.
\end{equation}
Combining \Cref{dunjiao_2,angle_est_2}, we get
\begin{displaymath}
\frac{\pi}{2}< 2\bigl(\D(G)+1\bigr)\arcsin{\sqrt{w {\lambda}^\cq_{\mathrm{min}}/2}}.
\end{displaymath}
So in this case,
\begin{displaymath}
\lambda^\cq_{\mathrm{min}}
>
\frac{2}{w}\Bigl(\sin\frac{\pi}{4\bigl(\D(G)+1\bigr)}\Bigr)^2.\qedhere
\end{displaymath}
\end{proof}
%================================
%==========Section 8=============
%================================
\section{Average Return Probability}\label{sec_avg}
In this section, we deal with the average spectral measure of $\cq$ and average return probabilities for simple random walk on generic non-bipartite, simple, finite, connected graphs. Our method is inspired by \autocite[Section~5]{Lyons-2018}.
\subsection{Estimate of Average Spectral Measure}
Define average spectral measure of $\cq$ as
\begin{displaymath}
\mu^\cq(\delta)\coloneqq \frac{1}{n}\sum_{x\in V}\mu^\cq_x(\delta).
\end{displaymath}
We also write $\mu^\cq_S(\delta)\coloneqq \sum_{x\in S}\mu^\cq_x(\delta)$ for $S \subseteq V$.
\begin{theorem}\label{spec_avg}
For any non-bipartite, finite, simple, connected, unweighted graph $G$ and ${\delta\in(0,2)}$, we have
\begin{displaymath}
\mu^\cq(\delta)<(4000\delta)^{1/3}.
\end{displaymath}
\end{theorem}
\begin{corollary}
For any non-bipartite, finite, simple, connected, unweighted graph $G$, we have\/ $\lambda^\cq_k\geqslant \frac{k^3}{4000n^3}$ for\/ $1\leqslant k\leqslant n$. Therefore, $\lambda^P_k\geqslant -1+\frac{k^3}{4000n^3}$.
\end{corollary}
\begin{proof}
By the same argument as in the proof of \Cref{lbb_reg}, this corollary follows easily from \Cref{spec_loc} and \Cref{spec_avg}.
\end{proof}
To prove \Cref{spec_avg}, we need some preparation.
\par
Let $m\coloneqq \bigl\lfloor\frac{n}{2}\mu^\cq(\delta)\bigr\rfloor+1$. For each $x\in V$, set
\begin{displaymath}
\R(x)\coloneqq \Bigl\{
y\in V\st \norm{F_x^\cq-F_y^\cq}_w \leqslant \frac{\norm{F_x^\cq}_w}{4} \text{ or }\norm{F_x^\cq+F_y^\cq}_w \leqslant \frac{\norm{F_x^\cq}_w}{4}
\Bigr\}.
\end{displaymath}
\par
We will use \Cref{alg:setselection} to get some useful sets.
\begin{algorithm}[htbp]
\begin{algorithmic}
\STATE Let $S_0\leftarrow V$.
\FOR {$i=1 \to m$}
\STATE Choose a vertex $x_i$ in $S_{i-1}$ that maximizes $\mu^\cq_{x_i}(\delta)$.
\STATE Let $S_{i}\leftarrow S_{i-1}\setminus \R(x_i)$.
\ENDFOR
\RETURN $\R(x_1),\R(x_2),\ldots,\R(x_m)$.
\end{algorithmic}
\caption{Set-Selection($\alpha$)}
\label{alg:setselection}
\end{algorithm}
\par
For each $x\in V(G)$, set
\begin{displaymath}
\widetilde{N}(x)\coloneqq
\Bigl\{ y\in N(x)\st \norm{F_y^\cq+F_x^\cq}_w \leqslant \frac{\norm{F_x^\cq}_w}{9}\Bigr\}.
\end{displaymath}
Let $\widetilde{T}(x)$ be the star formed by $x$ and $\widetilde{N}(x)$. We pick $T(x)$ as a maximal (with respect to inclusion) connected bipartite graph including $\widetilde{T}(x)$ such that
\begin{displaymath}
\all{y\in V\bigl(T(x)\bigr)}\quad
\norm[\big]{F^\cq_x-(-1)^{\dist(x,y;T(x))}F_y^\cq}_w\leqslant \frac{\norm{F_x^\cq}_w}{9}.
\end{displaymath}
\begin{lemma}\label{lm51}
Let $G$ be a non-bipartite, finite, simple, connected, unweighted graph and $\delta\geqslant \lambda^\cq_\mathrm{min}$.
\begin{enumerate}
\item[\upshape(0)] For $i=1,2,\ldots, m$, we have $\mu^\cq_{\R(x_i)}(\delta)\leqslant \frac{4}{3}$.
\item[\upshape(1)] For $i=1,2,\ldots, m$, we have $\mu^\cq_{x_i}(\delta)\geqslant \frac{\mu^\cq(\delta)}{3}$. In addition, \Cref{alg:setselection} is well designed: each $S_{i-1}$ is non-empty and $x_i$ could be chosen for $i=1,2,\ldots, m$. Therefore, \Cref{alg:setselection} is not stopping before $i=m$.
\item[\upshape(2)] For\/ $1\leqslant i<j\leqslant m$, $V(T(x_i))\cap V(T(x_j))=\varnothing$.
\end{enumerate}
\end{lemma}
\begin{proof}
(0) Since $\delta\geqslant \lambda^\cq_\mathrm{min}$, we must have $\norm{F_{x_i}^\cq}_w>0$. By the proof of \autocite[Lemma~3.13]{Lyons-2018}, for any two non-zero vectors in a Hilbert space $\cH$,
\begin{equation}\label{aux}
\Re\Bigl\langle\frac{f}{\norm{f}_\cH},\frac{g}{\norm{g}_\cH}\Bigr\rangle_\cH
\geqslant
1-\frac{2\norm{f-g}^2_\cH}{\norm{f}_\cH^2},
\end{equation}
where $\Re z$ is the real part of a complex number $z$. Also, for $y\in \R(x_i)$, by the construction of $\R(x_i)$, we have an integer $\sigma_i(y)$ such that
\begin{displaymath}
\norm[\big]{F^\cq_{x_i}-(-1)^{\sigma_i(y)}F_y^\cq}_w\leqslant \frac{\norm{F_{x_i}^\cq}_w}{4}.
\end{displaymath}
This implies $\norm{F^\cq_y}_w>0$ for $y\in\R(x_i)$.
\par
Therefore,
\begin{displaymath}\begin{split}
1&= \sum_{y\in \R(x_i)}w(y)\norm{F^\cq_{y}}_w^2\abs[\Big]{\Bigl\langle \frac{F^\cq_y}{\norm{F^\cq_{y}}_w},\frac{F^\cq_{x_i}}{\norm{F^\cq_{x_i}}_w}\Bigr\rangle_w}^2\\
&= \sum_{y\in \R(x_i)}\mu_y^\cq(\delta)\abs[\Big]{\Bigl\langle \frac{(-1)^{\sigma_i(y)}F^\cq_y}{\norm{(-1)^{\sigma_i(y)}F^\cq_{y}}_w},\frac{F^\cq_{x_i}}{\norm{F^\cq_{x_i}}_w}\Bigr\rangle_w}^2\\
&\geqslant \sum_{y\in\R(x_i)}\mu^\cq_y(\delta)
\Bigl(1-\frac{2\norm{F^\cq_{x_i}-(-1)^{\sigma_i(y)}F^\cq_y}^2_w}{\norm{F^\cq_{x_i}}^2_w}\Bigr)^2\\
&\geqslant \mu^\cq_{\R(x_i)}(\delta)\Bigl(1-\frac{2}{16}\Bigr)^2=\frac{49}{64}\mu^\cq_{\R(x_i)}(\delta),
\end{split}\end{displaymath}
where we use \Cref{aux} to get the first inequality.
It follows that
\begin{displaymath}
\mu^\cq_{\R(x_i)}(\delta)\leqslant \frac{64}{49}< \frac{4}{3}.
\end{displaymath}
\par
(1) We know $\mu_{S_0}^\cq(\delta)=\mu_V^\cq(\delta)=n\mu^\cq(\delta)$. In addition, by assertion (0), the total spectral measure of removed vertices in each iteration of the for loop in \Cref{alg:setselection} is at most $\frac{4}{3}$. We have, for each $i\leqslant m$,
\begin{displaymath}
\mu_{S_{i-1}}^\cq(\delta)\geqslant n\mu^\cq(\delta)-\frac{4}{3}(m-1)\geqslant\frac{n\mu^\cq(\delta)}{3},
\end{displaymath}
where the last inequality holds by the definition of $m$. This implies $S_{i-1}\neq\varnothing$ and $x_i$ could be chosen in \Cref{alg:setselection} for $i=1,2,\ldots, m$. In other words, the algorithm is well designed and is not stopping before $i=m$.
\par
Furthermore, since $x_i$ has the largest vertex spectral measure in $S_{i-1}$ for $i=1,2,\ldots, m$, we have
\begin{displaymath}
\mu_{x_i}^\cq(\delta)\geqslant \frac{\mu_{S_{i-1}}^\cq(\delta)}{n}\geqslant\frac{\mu_{S_{m-1}}^\cq(\delta)}{n}
\geqslant \frac{\mu^\cq(\delta)}{3}.
\end{displaymath}
\par
(2) Suppose that some vertex $y$ lies in $V\bigl(T(x_i)\bigr)\cap V\bigl(T(x_j)\bigr)$. We deal with assertion (2) in the following two cases. For ease of notation, we write $\tau_k\coloneqq \dist(x_k,y;T(x_k))$ for $k=i,j$. By the construction of $T(x_i)$ and $T(x_j)$,
\begin{displaymath}
\norm[\big]{F^\cq_{x_k}-(-1)^{\tau_k}F_y^\cq}_w\leqslant \frac{\norm{F_{x_k}^\cq}_w}{9},\qquad k=i,j.
\end{displaymath}
\par\underline{Case (a)}: $\norm{F^\cq_{x_j}}_w>\frac{5}{4}\norm{F^\cq_{x_i}}_w$. In this case, we have
\begin{displaymath}\begin{split}
\norm[\big]{F^\cq_y}_w
&=
\norm[\big]{(-1)^{\tau_j}F_y^\cq}_w \\
&\geqslant
\frac{8}{9}\norm[\big]{F^\cq_{x_j}}_w
>
\frac{8}{9}\cdot\frac{5}{4}\norm[\big]{F^\cq_{x_i}}_w \\
&=
\frac{10}{9}\norm[\big]{F^\cq_{x_i}}_w.
\end{split}\end{displaymath}
Thus, we have
\begin{displaymath}
\norm[\big]{(-1)^{\tau_i}F_y^\cq}_w
=\norm[\big]{F^\cq_y}_w
>\frac{10}{9}\norm[\big]{F^\cq_{x_i}}_w.
\end{displaymath}
This contradicts our assumption that $y\in V\bigl(T(x_i)\bigr)$.
\par\underline{Case (b)}: $\norm{F^\cq_{x_j}}_w \leqslant \frac{5}{4}\norm{F^\cq_{x_i}}_w$.
In this case, we have
\begin{displaymath}\begin{split}
\norm[\big]{F^\cq_{x_i}-(-1)^{\tau_i+\tau_j}F_{x_j}^\cq}_w
&= \norm[\big]{F^\cq_{x_i}-(-1)^{\tau_i}F_y^\cq+(-1)^{\tau_i}F_y^\cq-(-1)^{\tau_i+\tau_j}F_{x_j}^\cq}_w \\
&\leqslant \norm[\big]{F^\cq_{x_i}-(-1)^{\tau_i}F_y^\cq}_w+\norm[\big]{(-1)^{\tau_i}F_y^\cq-(-1)^{\tau_i+\tau_j}F_{x_j}^\cq}_w \\
&= \norm[\big]{F^\cq_{x_i}-(-1)^{\tau_i}F_y^\cq}_w+\norm[\big]{F_{x_i}^\cq-(-1)^{\tau_j}F_{y}^\cq}_w \\
&\leqslant \frac{\norm{F^\cq_{x_i}}_w}{9}+\frac{\norm{F^\cq_{x_j}}_w}{9}
\leqslant \Bigl(\frac{1}{9}+\frac{1}{9}\cdot\frac{5}{4}\Bigr)\norm{F^\cq_{x_i}}_w \\
&=\frac{\norm{F^\cq_{x_i}}_w}{4}.
\end{split}\end{displaymath}
Thus, $x_j\in\R(x_i)$. But this is impossible.
Therefore, we may conclude that $V(T(x_i))\cap V(T(x_j))=\varnothing$.
\end{proof}
For a subset $E'\subseteq E(G)$ and a mapping $F\colon V\to \cH$ to a Hilbert space $\cH$, we define energy
\begin{displaymath}
\widetilde\cE(E')\coloneqq \sum_{(x,y)\in E'}\norm{F_x+F_y}_\cH^2.
\end{displaymath}
For $S\subseteq V(G)$, let $E(S)$ be the collection of edges that are incident with the vertices in $S$. We define the energy of $S$ as $\widetilde\cE(S)\coloneqq \widetilde\cE\bigl(E(S)\bigr)$.
\begin{lemma}
Assume that the graph $G$ is a non-bipartite, finite, simple, connected, unweighted graph. When $\delta\geqslant \lambda^\cq_\mathrm{min}$, for\/ $1\leqslant i\leqslant n$,
\begin{displaymath}
\widetilde{\cE} \bigl(V(T(x_i))\bigr)
>\frac{\mu^\cq(\delta)}{250\abs[\big]{V(T(x_i))}^2}.
\end{displaymath}
\end{lemma}
\begin{proof}
a) We first consider the case $w(x_i)>\abs[\big]{V(T(x_i))}$. Recall that $w(x_i)$ equals the degree of $x_i$, since we are considering unweighted graphs. In this case,
\begin{displaymath}\begin{split}
\widetilde{\cE} \bigl(V(T(x_i))\bigr)
&\geqslant \sum_{y\in N(x_i)\setminus V(T(x_i))}\norm{F^\cq_{x_i}+F^\cq_y}_w^2\\
&> \abs[\big]{N(x_i)\setminus V(T(x_i))}\cdot\frac{\norm{F^\cq_{x_i}}_w^2}{81}\\
&=\abs[\big]{N(x_i)\setminus V(T(x_i))}\cdot\frac{\mu_{x_i}^\cq(\delta)}{81 w(x_i)}\\
&>\bigl(w(x_i)-\abs[\big]{V(T(x_i))}+1\bigr)\frac{\mu^\cq(\delta)}{250w(x_i)}\\
&\geqslant \frac{\mu^\cq(\delta)}{250\abs[\big]{V(T(x_i))}^2},
\end{split}\end{displaymath}
where the second inequality holds because $V(T(x_i))\supseteq\widetilde{N}(x_i)$, and the third inequality holds thanks to \Cref{lm51}(1).
\par
b) We now turn to the case $w(x_i)\leqslant \abs[\big]{V(T(x_i))}$. In this case, we claim that there must be a path $z_0z_1\cdots z_\ell$, such that $\ell\leqslant \abs[\big]{V(T(x_i))}$, $z_0=x_i$, and
\begin{displaymath}
\norm[\big]{F^\cq_{x_i}-(-1)^{\ell}F_{z_\ell}^\cq}_w> \frac{\norm{F_{x_i}^\cq}_w}{9}.
\end{displaymath}
In fact, if $V\bigl(T(x_i)\bigr)$ is a proper subset of $V(G)$, there is a vertex $y\in V(G)\setminus V(T(x_i))$ with $\dist\bigl(y,T(x_i);G\bigr)=1$. Let $z_0z_1\cdots z_\ell$ be a path joining $z_0\coloneqq x_i$ and $z_\ell\coloneqq y$, with $z_0z_1\cdots z_{\ell-1}$ being a path in $T(x_i)$. Because $T(x_i)$ is assumed maximal and $y\notin V\bigl(T(x_i)\bigr)$, we must have
\begin{displaymath}
\norm[\big]{F^\cq_{x_i}-(-1)^{\ell}F_{z_\ell}^\cq}_w> \frac{\norm{F_{x_i}^\cq}_w}{9}.
\end{displaymath}
If $V\bigl(T(x_i)\bigr) = V(G)$, since $T(x_i)$ is assumed to be a maximal connected bipartite graph and $G$ is a connected non-bipartite graph, there must be an edge $e=(u,v)$, whose addition to $T(x_i)$ results in an odd cycle. To be specific, we have
\begin{displaymath}
\dist\bigl(x_i,u;T(x_i)\bigr)\equiv \dist\bigl(x_i,v;T(x_i)\bigr) \mod 2.
\end{displaymath}
Write $\tau\coloneqq \dist\bigl(x_i,u;T(x_i)\bigr)$. Then
\begin{displaymath}
\norm[\big]{F^\cq_{x_i}-(-1)^\tau F_u^\cq}_w \leqslant \frac{\norm{F_{x_i}^\cq}_w}{9}.
\end{displaymath}
This implies
\begin{equation}\label{1step}
\norm[\big]{F^\cq_{x_i}-(-1)^{\tau+1} F_u^\cq}_w > \frac{\norm{F_{x_i}^\cq}_w}{9}.
\end{equation}
In fact, if otherwise $\norm[\big]{F^\cq_{x_i}-(-1)^{\tau+1} F_u^\cq}_w \leqslant \frac{\norm{F_{x_i}^\cq}_w}{9}$, then
\begin{displaymath}
\norm[\big]{2F^\cq_{x_i}}_w
=\norm[\big]{F^\cq_{x_i}-(-1)^\tau F_u^\cq+F^\cq_{x_i}-(-1)^{\tau+1} F_u^\cq}_w
<\Bigl(\frac{1}{9}+\frac{1}{9}\Bigr)\norm[\big]{F^\cq_{x_i}}_w.
\end{displaymath}
Therefore, $\norm[\big]{F^\cq_{x_i}}_w=0$. But $\norm[\big]{F^\cq_{x_i}}_w \neq0$ for $\delta\geqslant \lambda^\cq_\mathrm{min}$.
Now pick a shortest path $z_0z_1\cdots z_{\ell-1}$ joining $x_i$ and $v$ in $T(x_i)$, where $z_0=x_i$ and $z_{\ell-1}=v$. Let $z_\ell=u$. Then $z_0z_1\cdots z_{\ell-1}z_\ell$ is a path joining $x_i$ and $u$. Hence, by our assumption,
\begin{displaymath}
\ell-1\equiv \tau\mod 2.
\end{displaymath}
Therefore, \Cref{1step} can be written as
\begin{displaymath}
\norm[\big]{F^\cq_{x_i}-(-1)^{\ell} F_{z_\ell}^\cq}_w > \frac{\norm{F_{x_i}^\cq}_w}{9}.
\end{displaymath}
Our claim is proved. By the claim, we have
\begin{displaymath}\begin{split}
\widetilde{\cE} \bigl(V(T(x_i))\bigr)
&\geqslant
\sum_{k=0}^{\ell-1} \norm{F^\cq_{z_k}+F^\cq_{z_{k+1}}}_w^2
=
\sum_{k=0}^{\ell-1} \norm{(-1)^k F^\cq_{z_k}-(-1)^{k+1}F^\cq_{z_{k+1}}}_w^2\\
&\geqslant
\frac{1}{\ell}\Bigl(\sum_{k=0}^{\ell-1} \norm{(-1)^k F^\cq_{z_k}-(-1)^{k+1}F^\cq_{z_{k+1}}}_w\Bigr)^2
\geqslant
\frac{1}{\ell}\norm{ F^\cq_{z_0}-(-1)^\ell F^\cq_{z_{\ell}}}_w^2\\
&>
\frac{1}{\ell}\cdot\frac{1}{81}\norm{F^\cq_{x_i}}^2_w
\geqslant
\frac{1}{\abs[\big]{V(T(x_i))}}\cdot\frac{\norm{F^\cq_{x_i}}^2_w}{81}\\
&=
\frac{1}{\abs[\big]{V(T(x_i))}}\cdot\frac{\mu^\cq_{x_i}(\delta)}{81 w(x_i)}
>
\frac{1}{\abs[\big]{V(T(x_i))}}\cdot\frac{\mu^\cq(\delta)}{250 w(x_i)}\\
&\geqslant
\frac{\mu^\cq(\delta)}{250\abs[\big]{V(T(x_i))}^2},
\end{split}\end{displaymath}
where the sixth inequality follows from \Cref{lm51}(1), and the last inequality holds thanks to our assumption that $w(x_i)\leqslant \abs[\big]{V(T(x_i))}$. The proof is complete.
\end{proof}
\begin{proof}[Proof of \Cref{spec_avg}]
To begin, we claim that
\begin{displaymath}
\delta
\geqslant \frac{\sum_{(y,z)\in E(G)}\norm{F_y^\cq+F_z^\cq}_w^2}{\sum_{x\in V}\norm{F_x^\cq}_w^2w(x)}.
\end{displaymath}
This can be proved exactly in the same way as \autocite[Lemma~3.14]{Lyons-2018}. In fact, for all $x, y \in V$,
\begin{displaymath}
F^\cq_x(y)
=
\big\langle I_\cq(\delta) \ind_x/w(x), \ind_y \big\rangle
=
\big\langle I_\cq(\delta) \ind_x/w(x), \ind_y/w(y)
\big\rangle_w
=
F^\cq_y(x).
\end{displaymath}
Therefore, using \Cref{q_form}, we have
\begin{align*}
\sum_{(y,z)\in E(G)} w(y,z) \norm{F^\cq_y+F^\cq_z}_w^2
&=\sum_{(y,z)\in E(G)} w(y,z)\sum_{x\in V} w(x) \abs[\big]{\bigl(F^\cq_y+F^\cq_z\big)(x)}^2\\
&=\sum_{(y,z)\in E(G)} w(y,z)\sum_{x\in V} w(x) \abs[\big]{F^\cq_x(y)+F^\cq_x(z)}^2.
\end{align*}
Proceeding further, we get that
\begin{align*}
\sum_{(y,z)\in E(G)} w(y,z) \norm{F^\cq_y+F^\cq_z}_w^2
&=\sum_{x\in V} w(x) \sum_{(y,z)\in E(G)} w(y,z)\abs[\big]{F^\cq_x(y)+F^\cq_x(z)}^2\\
&= {\sum_{x\in V} w(x) \langle F^\cq_x, \cq F^\cq_x\rangle_w}\\
&\leqslant {\sum_{x\in V} \delta\,w(x) \norm{F^\cq_x}_w^2}.
\end{align*}
The claim is thus proved.
\par
Now we assume $\delta\geqslant \lambda^\cq_\mathrm{min}$ without loss of generality. Since for different $i$ and $j$,
\begin{displaymath}
V(T(x_i))\cap V(T(x_j))=\varnothing,
\end{displaymath}
we have $\sum_{i=1}^{m}\abs[\big]{V(T(x_i))}\leqslant n$. Therefore,
\begin{displaymath}\begin{split}
\delta
&\geqslant \frac{\sum_{(x,y)\in E(G)}\norm{F_x^\cq+F_y^\cq}_w^2}{\sum_{y\in V}\norm{F_y^\cq}_w^2w(y)}
\geqslant \frac{1}{n\mu^\cq(\delta)}\cdot\frac{1}{2}\sum^{m}_{i=1}\widetilde{\cE} \bigl(V\bigl(T(x_i)\bigr)\bigr)\\
&>\frac{1}{2n\mu^\cq(\delta)}\sum^{m}_{i=1}\frac{\mu^\cq(\delta)}{250\abs[\big]{V(T(x_i))}^2}
\geqslant \frac{m^3}{500n^3}
\geqslant \frac{\mu^\cq(\delta)^3}{4000},
\end{split}\end{displaymath}
where the second inequality follows by \Cref{spec_embed_norm} and the fact
that each edge is counted in at most two sets $T(x_i)$ for energy, the fourth inequality
follows by convexity of the function ${s \mapsto 1/s^2}$,
and the last inequality holds because $m\geqslant n\mu^\cq(\delta)/2$. The proof of
\Cref{spec_avg} is complete.
\end{proof}
\subsection{Average Return Probability}
Using \Cref{spec_avg}, we can bound average return probabilities. \Cref{avg_clc} will be useful in the proofs of \Cref{avg_even,avg_odd}.
\begin{lemma}\label{avg_clc}
For $t>0$, we have
\begin{displaymath}
\frac{\sqrt[3]{4000}}{3}\int_{0}^{1}(1-\lambda)^t\lambda^{-2/3}\df{\lambda}
\leqslant \frac{15}{t^{1/3}}.
\end{displaymath}
\end{lemma}
\begin{proof}
We have
\begin{displaymath}
\int_{0}^{1}(1-\lambda)^t\lambda^{-2/3}\df{\lambda}
\leqslant \int_{0}^{\infty}\exp\{-\lambda t\}\lambda^{-2/3}\df{\lambda}.
\end{displaymath}
Taking $\lambda t=s$, we get
\begin{displaymath}\begin{split}
\int_{0}^{\infty}\exp\{-\lambda t\}\lambda^{-2/3}\df{\lambda}
&=\int_{0}^{\infty}\ue^{-s}(s/t)^{-2/3}\df{(s/t)}\\
&=\frac{1}{t^{1/3}}\int_{0}^{\infty}\ue^{-s}s^{-2/3}\df{s}=\frac{\Gamma(1/3)}{t^{1/3}}.
\end{split}\end{displaymath}
Therefore,
\begin{displaymath}
\frac{\sqrt[3]{4000}}{3}\int_{0}^{1}(1-\lambda)^t\lambda^{-2/3}\df{\lambda}
\leqslant \frac{\sqrt[3]{4000}\,\Gamma(1/3)}{3 t^{1/3}}
\leqslant \frac{15}{t^{1/3}}.\qedhere
\end{displaymath}
\end{proof}
\begin{theorem}\label{avg_even}
Let $G$ be a non-bipartite, finite, simple, connected, unweighted graph. Then for\/ $t\equiv 0\bmod 2$, we have
\begin{displaymath}
0\leqslant
\frac{\sum_{x\in V(G)}p_t(x,x)-1}{n}
\leqslant
\frac{30}{t^{1/3}}.
\end{displaymath}
\end{theorem}
\begin{proof}
Set
\begin{numcases}
{\Phi(\lambda)\coloneqq }
n\sqrt[3]{4000\lambda}
\quad \nonumber &if\/ $\lambda\geqslant0\text{ and }n\sqrt[3]{4000\lambda}\leqslant n\mu^*(1)$,\\
n\mu^*(1) \quad \nonumber &\text{for intermediate values of $\lambda$},\\
n-1-n\sqrt[3]{4000(2-\lambda)}
\quad \nonumber &if\/ $\lambda\leqslant2\text{ and }n-1-n\sqrt[3]{4000(2-\lambda)}\geqslant n\mu^*(1)$.
\end{numcases}
Then as in the proof of \Cref{reg_return}, by \Cref{spec_avg} and \autocite[Theorem~5.1]{Lyons-2018},
\begin{displaymath}\begin{split}
&\Phi(0)=0,\qquad \Phi(2)=n-1,\\
&n\mu^*(\lambda)\leqslant \Phi(\lambda) \qquad \text{for\quad} \lambda\in[0,1],\\
&n\mu^*(\lambda)\geqslant \Phi(\lambda) \qquad \text{for\quad} \lambda\in[1,2].
\end{split}\end{displaymath}
Therefore, by our calculation in part a) of the proof of \Cref{spec_bound}, for $t\equiv 0\bmod 2$,
\begin{displaymath}\begin{split}
\sum_{x\in V}p_t(x,x)
&=\sum_{x\in V}\Bigl(\pi(x)+\bigl(1-\pi(x)\bigr)\Bigr)
+t\int_{0}^{2}\Bigl(\sum_{x\in V}\mu^*_x(\lambda)\Bigr)(1-\lambda)^{t-1}\df{\lambda}\\
&=n+t\int_{0}^{2}n\mu^*(\lambda)(1-\lambda)^{t-1}\df{\lambda}
\leqslant
n+t\int_{0}^{2}\Phi(\lambda)(1-\lambda)^{t-1}\df{\lambda}\\
&=n-(n-1)+\int_{0}^{2}(1-\lambda)^t\Phi'(\lambda)\df{\lambda}
=1+\int_{0}^{2}(1-\lambda)^t\Phi'(\lambda)\df{\lambda}\\
&\leqslant 1+2\int_{0}^{1}(1-\lambda)^t\frac{n\sqrt[3]{4000}}{3}\lambda^{-2/3}\df{\lambda}\\
&\leqslant 1+\frac{30n}{t^{1/3}},
\end{split}\end{displaymath}
where we are using \Cref{ibp} to get the third equality. Therefore,
\begin{displaymath}
\frac{\sum_{x\in V(G)}p_t(x,x)-1}{n}
\leqslant
\frac{30}{t^{1/3}}.
\end{displaymath}
On the other hand, for all $x\in V$ and even $t$, $p_t(x,x)\geqslant \pi(x)$.
Hence, we have
\begin{displaymath}
\frac{\sum_{x\in V(G)}p_t(x,x)-1}{n}
\geqslant 0.\qedhere
\end{displaymath}
\end{proof}
\begin{theorem}\label{avg_odd}
Let $G$ be a non-bipartite, finite, simple, connected, unweighted graph. Then for $t\equiv 1\bmod 2$,
\begin{displaymath}
\frac{\abs[\big]{\sum_{x\in V(G)}p_t(x,x)-1}}{n}
\leqslant
\frac{15}{t^{1/3}}.
\end{displaymath}
\end{theorem}
\begin{proof}
a) Let $t\equiv1\bmod 2$. By our calculation in part a) of the proof of \Cref{spec_bound},
\begin{displaymath}\begin{split}
\sum_{x\in V}p_t(x,x)
&=
\sum_{x\in V}\Bigl(\pi(x)-\bigl(1-\pi(x)\bigr)\Bigr)
+t\int_{0}^{2}\Bigl(\sum_{x\in V}\mu^*_x(\lambda)\Bigr)(1-\lambda)^{t-1}\df{\lambda}\\
&=
2-n+t\int_{0}^{2}n\mu^*(\lambda)(1-\lambda)^{t-1}\df{\lambda}.
\end{split}\end{displaymath}
\par
b) Set
\begin{displaymath}
\Psi_1(\lambda)\coloneqq \bigl(n-1-n\sqrt[3]{4000(2-\lambda)}\,\bigr)\vee0.
\end{displaymath}
Then by \Cref{spec_avg},
\begin{displaymath}\begin{split}
&\Psi_1(0)=0,\qquad \Psi_1(2)=n-1,\\
&\sum_{x\in V}\mu_x^*(\lambda)\geqslant \Psi_1(\lambda) \qquad \text{for } \lambda\in[0,2].
\end{split}\end{displaymath}
By our calculation in part a),
\begin{displaymath}\begin{split}
\sum_{x\in V}p_t(x,x)
&\geqslant
2-n+t\int_{0}^{2}\Psi_1(\lambda)(1-\lambda)^{t-1}\df{\lambda}\\
&=
2-n+(n-1)+\int_{0}^{2}(1-\lambda)^t\Psi'_1(\lambda)\df{\lambda}
=
1+\int_{0}^{2}(1-\lambda)^t\Psi'_1(\lambda)\df{\lambda}\\
&\geqslant
1-\int_{0}^{1}(1-\lambda)^t\frac{n\sqrt[3]{4000}}{3}\lambda^{-2/3}\df{\lambda}
\geqslant
1-\frac{15n}{t^{1/3}},
\end{split}\end{displaymath}
where the first equality holds thanks to \Cref{ibp}.
\par
c) Set
\begin{displaymath}
\Psi_2(\lambda)\coloneqq \bigl(n\sqrt[3]{4000\lambda}\,\bigr)\wedge(n-1).
\end{displaymath}
Then by \autocite[Theorem~5.1]{Lyons-2018},
\begin{displaymath}\begin{split}
&\Psi_2(0)=0,\qquad \Psi_2(2)=n-1,\\
&\sum_{x\in V}\mu_x^*(\lambda)\leqslant \Psi_2(\lambda) \qquad \text{for } \lambda\in[0,2].
\end{split}\end{displaymath}
Using a similar argument as in b), we get
\begin{displaymath}
\sum_{x\in V}p_t(x,x)
\leqslant
1+\frac{15n}{t^{1/3}}.\qedhere
\end{displaymath}
\end{proof}
\subsection{Sum of Eigenvalue Powers in Absolute Value}
Since similar matrices have the same trace, we have
\begin{displaymath}
\frac{\sum_{x\in V(G)}p_t(x,x)-1}{n}
=
\frac{\sum_{i=1}^n(\lambda^P_i)^t-1}{n}=\frac{1}{n}\sum_{i=1}^{n-1}(\lambda^P_i)^t.
\end{displaymath}
Therefore, when $t\geqslant 2$ is even, \Cref{avg_even} gives
\begin{equation}\label{abs_bdd_even}
\frac{1}{n}\sum_{i=1}^{n-1}\abs{\lambda^P_i}^t
=\frac{\sum_{x\in V(G)}p_t(x,x)-1}{n}
\leqslant \frac{30}{t^{1/3}}.
\end{equation}
For odd $t$, we have
\begin{displaymath}
\abs[\Big]{\sum_{x\in V(G)}p_t(x,x)-1}
= \abs[\Big]{\sum^{n-1}_{i=1}(\lambda_i^P)^t}
\leqslant
\sum^{n-1}_{i=1}\abs{\lambda_i^P}^t.
\end{displaymath}
\Cref{avg_odd} will not give a bound on $\sum^{n-1}_{i=1}\abs{\lambda_i^P}^t$ directly. But we can still make a detour and bound $\sum^{n-1}_{i=1}\abs{\lambda_i^P}^t$ by \Cref{avg_even} as follows.
\begin{proposition}\label{abs_bdd}
Let $G$ be a non-bipartite, finite, simple, connected, unweighted graph. We have
\begin{displaymath}\begin{split}
\frac{1}{n}\sum^{n-1}_{i=1}\abs{\lambda_i^P}^t&\leqslant \frac{30}{t^{1/3}} \qquad \text{for\quad} t\equiv 0\bmod 2 \text{ and } t\geqslant 2,\\
\frac{1}{n}\sum^{n-1}_{i=1}\abs{\lambda_i^P}^t&\leqslant \frac{30}{(t^2-1)^{1/6}} \qquad \text{for\quad} t\equiv 1\bmod 2\text{ and } t\geqslant 3.
\end{split}\end{displaymath}
\end{proposition}
The bound in \Cref{avg_odd} is $\frac{15}{t^{1/3}}$; we get the bound $\frac{30}{(t^2-1)^{1/6}}$ for odd $t$ in \Cref{abs_bdd}. These two bounds are not far apart: they are of the same order in $t$.
\begin{proof}[Proof of \Cref{abs_bdd}]
The assertion for even $t$ is nothing but \Cref{abs_bdd_even}. We now consider odd $t$. When $t\geqslant 3$ is odd, by \Cref{abs_bdd_even},
\begin{displaymath}\begin{split}
\sum_{i=1}^{n-1}\abs{\lambda^P_i}^t
&=
\sum_{i=1}^{n-1}\bigl(\abs{\lambda^P_i}^{\frac{t-1}{2}}\abs{\lambda^P_i}^{\frac{t+1}{2}}\bigr)^{1/2}
\leqslant
\Bigl(\sum_{i=1}^{n-1}\abs{\lambda^P_i}^{t-1}\Bigr)^{1/2}
\Bigl(\sum_{i=1}^{n-1}\abs{\lambda^P_i}^{t+1}\Bigr)^{1/2}\\
&\leqslant
\Bigl(\frac{30n}{\sqrt[3]{t-1}}\Bigr)^{1/2}\Bigl(\frac{30n}{\sqrt[3]{t+1}}\Bigr)^{1/2}\\
&=
\frac{30n}{(t^2-1)^{1/6}},
\end{split}\end{displaymath}
where the first inequality holds thanks to the Cauchy--Schwarz inequality.
\end{proof}
%================================
%==========Section 9=============
%================================
\section{Bipartite Case}\label{sec_bip}
Now we deal with bipartite graphs. On bipartite graphs, simple random walk has period two. We have the following result on return probabilities, whose proof uses only the bound on the vertex spectral measure of $\cl$ from \autocite[Theorem~4.9]{Lyons-2018}.
\begin{theorem}\label{reg_return_bi}
Consider a regular, bipartite, simple, connected, unweighted graph $G$. For each $x\in V$, simple random walk on $G$ satisfies
\begin{displaymath}
0\leqslant p_t(x,x)-2\pi(x)\leqslant \frac{18}{\sqrt{t}} \qquad \text{for\quad} t\equiv 0\bmod 2.
\end{displaymath}
\end{theorem}
To prove \Cref{reg_return_bi}, we don't need to get a bound on the vertex spectral measure of $\cq$ as we did in \Cref{sec_reg}, because the following lemma gives us a relation between the vertex spectral measures of $\cl$ and $\cq$.
\begin{lemma}[\cite{Mohar-1989}, Theorem~4.8]\label{mohar}
If $G$ is bipartite, the spectrum of $P$ is symmetric with respect to zero. For each $x\in V$,
the vertex spectral measure of $P$, $\bigl\langle I_P(\ud\delta)\be_x,\be_x\bigr\rangle_w$, is symmetric with respect to\/ $0$. As a consequence, if $G$ is bipartite, for each $x\in V$ and $\delta\in[0,2]$, $\mu_x(\delta)=\mu^\cq_x(\delta)$.\qed
\end{lemma}
\begin{proof}[Proof of \Cref{reg_return_bi}]
a) Set
\begin{numcases}
{\mu_x^\#(\lambda)\coloneqq}
\mu_x^*(\lambda)
\quad \nonumber &if\/ $0\leqslant \lambda<2$,\\
1-2\pi(x)
\quad \nonumber &if\/ $\lambda=2$,
\end{numcases}
and
\begin{numcases}
{\varphi^\#(\lambda)\coloneqq }
10\sqrt{\lambda}
\quad \nonumber &if\/ $\lambda\geqslant0\text{ and }10\sqrt{\lambda}\leqslant \mu_x^\#(1)$,\\
\mu_x^\#(1) \quad \nonumber &\text{for intermediate values of $\lambda$},\\
\mu_x^\#(2)-10\sqrt{2-\lambda}
\quad \nonumber &if\/ $\lambda\leqslant2\text{ and }\mu_x^\#(2)-10\sqrt{2-\lambda}\geqslant \mu_x^\#(1)$.
\end{numcases}
Then $\varphi^\#(0)=\mu_x^\#(0)=0$, $\varphi^\#(2)=\mu_x^\#(2)=1-2\pi(x)$. In addition, by \Cref{mohar} and \autocite[Theorem~4.9]{Lyons-2018}, we have
\begin{displaymath}\begin{split}
&\mu_x^\#(\lambda)\leqslant \varphi^\#(\lambda) \qquad \text{for\quad} \lambda\in[0,1],\\
&\mu_x^\#(\lambda)\geqslant \varphi^\#(\lambda) \qquad \text{for\quad} \lambda\in[1,2].
\end{split}\end{displaymath}
\par
b) Now let $t$ be a positive even number. We have
\begin{equation}\begin{split}
p_t(x,x)
&=\int_{[0,2]}(1-\lambda)^t\mu_x(\ud\lambda)\\
&=2\pi(x)+\int_{[0,2]}(1-\lambda)^t\mu^\#_x(\ud\lambda).
\end{split}\end{equation}
Hence, we see that
\begin{displaymath}
p_t(x,x)-2\pi(x)=\int_{[0,2]}(1-\lambda)^t\mu^\#_x(\ud\lambda)\geqslant 0.
\end{displaymath}
\par
c) On the other hand,
\begin{displaymath}\begin{split}
p_t(x,x)
&=
2\pi(x)+\int_{[0,2]}(1-\lambda)^t\mu^\#_x(\ud\lambda)\\
&=
2\pi(x)+\left.(1-\lambda)^t\mu^\#_x(\lambda)\right|^2_0-\int_{0}^{2}\mu^\#_x(\lambda)\df{(1-\lambda)^t}\\
&=
2\pi(x)+\bigl(1-2\pi(x)\bigr)+t\int_{0}^{2}\mu^\#_x(\lambda)(1-\lambda)^{t-1}\df{\lambda}\\
&\leqslant
1+t\int_{0}^{2}\varphi^\#(\lambda)(1-\lambda)^{t-1}\df{\lambda}\\
&=
1-\varphi^\#(2)+\int_{0}^{2}(1-\lambda)^t(\varphi^\#)'(\lambda)\df{\lambda}\\
&=
2\pi(x)+\int_{0}^{2}(1-\lambda)^t(\varphi^\#)'(\lambda)\df{\lambda}\\
&\leqslant
2\pi(x)+2\int_{0}^{1}\frac{5(1-\lambda)^t}{\sqrt{\lambda}}\df{\lambda}
=
2\pi(x)+10\int_{0}^{1}\frac{(1-\lambda)^t}{\sqrt{\lambda}}\df{\lambda}\\
&\leqslant
2\pi(x)+\frac{18}{\sqrt{t}},
\end{split}\end{displaymath}
where the fourth equality follows from \Cref{ibp}, and we use \Cref{calc_aux} to get the last inequality.
\par
d) Summing up the above discussion, we get, for $x\in V$ and $t\equiv 0\bmod 2$,
\begin{displaymath}
0\leqslant p_t(x,x)-2\pi(x)\leqslant \frac{18}{\sqrt{t}}.\qedhere
\end{displaymath}
\end{proof}
\Cref{reg_return_bi} is an example of treating simple random walk on bipartite graphs. Using the same method, one may get parallel results to what we had
in previous sections. For instance, the following \Cref{trans_return_bi} is parallel to \Cref{trans_return}.
\begin{corollary}\label{trans_return_bi}
Let $G$ be a finite or infinite, vertex-transitive, $d$-regular, bipartite, simple, connected, unweighted graph with at least polynomial growth rate $N^\#(r)\geqslant Cr^D$, where $C > 0$ and $D \geqslant 1$ are constants and\/ $0 \leqslant r \leqslant \D(G)$.
Then for each $x\in V$ and $t\equiv 0\bmod 2$,
\begin{displaymath}
0\leqslant p_t(x,x)-\pi(x)
\leqslant 2\widetilde{C}t^{-D/2},
\end{displaymath}
where $\widetilde{C}=\frac{(D+4)^{D/2+2}d^{D/2}}{32CD^{D/2-1}}\Gamma\bigl(\frac{D}{2}\bigr)$.\qed
\end{corollary}
To prove \Cref{trans_return_bi}, one need only follow our argument in the proof of \Cref{reg_return_bi} and use the bound on the vertex spectral measure of $\cl$ obtained in \autocite[Theorem~6.1]{Lyons-2018}. Details are omitted.
%================================
%==========Section 10=============
%================================
\section{Combinatorial Signless Laplacian}\label{sec_comb}
We used spectral embedding to deal with random walk on graphs in previous sections. In fact, this tool is also powerful in analyzing the spectrum of graph adjacency matrices.
Assume that $G$ is a weighted finite graph. Let $\ell^2(V)$ be the Hilbert space of functions $f \colon V \to \RNS \text{ or } \CN$ with inner product
\begin{displaymath}
\langle f,g \rangle \coloneqq \sum_{x \in V}f(x)\overline{g(x)}
\end{displaymath}
and squared norm $\norm{f}^2 \coloneqq \langle f,f\rangle$. Let $W$ be the diagonal weight matrix of the graph $G$: $W\coloneqq \mathsf{diag}(w(x)\st x\in V)$.
Then it is easy to see that the combinatorial signless Laplacian $\Theta\coloneqq W+A$ is a bounded self-adjoint operator on $\ell^2(V)$. We denote the resolution of identity for $\Theta$ as $I_\Theta$ and define the vertex spectral measure of $\Theta$ at $x\in V$ as
\begin{displaymath}
\mu_x^\Theta(\delta)\coloneqq \langle I_\Theta\bigl([0,\delta]\bigr)\ind_x,\ind_x\rangle,\qquad \delta\geqslant 0.
\end{displaymath}
For ease of notation, we also write $I_\Theta(\delta)\coloneqq I_\Theta\bigl([0,\delta]\bigr)$ for $\delta\geqslant 0$.
\begin{lemma}\label{Theta_form}
For $f\in \ell^2(V)$, we have
\begin{displaymath}
\langle\Theta f,f\rangle=\sum_{(v,u)\in E(G)}w(v,u)\abs[\big]{f(v)+f(u)}^2.
\end{displaymath}
Therefore, the spectrum of $\Theta$ is non-negative. Moreover, if $f\in\img\bigl(I_\Theta(\delta)\bigr)$ for some $\delta\geqslant 0$, then $\langle\Theta f,f\rangle\leqslant \delta \norm{f}^2$.
\end{lemma}
\begin{proof}
We need to prove the first assertion only. Notice that
\begin{displaymath}\begin{split}
\sum_{v,u\in V}w(v,u)f(v)\overline{\bigl(f(v)+f(u)\bigr)}
&=
\sum_{v\in V}f(v)\overline{\sum_{u\in V}w(v,u)f(v)}
+\sum_{v\in V}f(v)\overline{\sum_{u\in V}w(v,u)f(u)}\\
&=
\sum_{v\in V}f(v)\overline{w(v)f(v)}
+\sum_{v\in V}f(v)\overline{(Af)(v)}\\
&=
\langle f,Wf\rangle +\langle f,Af\rangle\\
&=
\langle f,\Theta f\rangle.
\end{split}\end{displaymath}
By interchanging $u$ and $v$, we have
\begin{displaymath}
\sum_{v,u\in V}w(v,u)f(u)\overline{\bigl(f(v)+f(u)\bigr)}
=
\langle f,\Theta f\rangle.
\end{displaymath}
Therefore,
\begin{displaymath}\begin{split}
\sum_{(v,u)\in E(G)}&w(v,u)\abs[\big]{f(v)+f(u)}^2\\
&=\frac{1}{2}\Bigl(
\sum_{v,u\in V}w(v,u)f(v)\overline{\bigl(f(v)+f(u)\bigr)}
+\sum_{v,u\in V}w(v,u)f(u)\overline{\bigl(f(v)+f(u)\bigr)}
\Bigr)\\
&=\tfrac{1}{2}\bigl(\langle f,\Theta f\rangle+\langle f,\Theta f\rangle\bigr)\\
&=\langle f,\Theta f\rangle.\qedhere
\end{split}\end{displaymath}
\end{proof}
For $\delta\geqslant 0$, we define the spectral embedding based on $\Theta$ as
\begin{displaymath}\begin{split}
F^\Theta\colon V(G)&\rightarrow \ell^2(V(G))\\
x&\mapsto F^\Theta_x\coloneqq I_\Theta(\delta)\ind_x.
\end{split}\end{displaymath}
It is clear that $F^\Theta_x$ is a real-valued function on $V$ for each $x\in V$.
\begin{lemma}\label{embed_norm_com}
For each finite graph $G$ and $x\in V(G)$,
\begin{displaymath}
\norm[\big]{F^\Theta_x}^2=F^\Theta_x(x)=\mu_x^\Theta(\delta).
\end{displaymath}
\end{lemma}
\begin{proof}
Since $I_\Theta(\delta)$ is a self-adjoint projection operator on $\ell^2(V)$, we see that
\begin{displaymath}
\norm[\big]{F^\Theta_x}^2
=
\langle F^\Theta_x,F^\Theta_x\rangle
=
\langle I_\Theta(\delta)\ind_x,I_\Theta(\delta)\ind_x\rangle
=
\langle I_\Theta(\delta)\ind_x,\ind_x\rangle.
\end{displaymath}
By the definition of the vertex spectral measure of $\Theta$, we see that
$
\langle I_\Theta(\delta)\ind_x,\ind_x\rangle
=
\mu_x^\Theta(\delta)$.
Moreover, since $I_\Theta(\delta)\ind_x=F^\Theta_x$, we have $\langle I_\Theta(\delta)\ind_x,\ind_x\rangle=F^\Theta_x(x)$.
\end{proof}
\begin{lemma}\label{def_f_com}
If $\mu_x^\Theta(\delta)>0$, define $f\colon V\rightarrow\CN$ as $f\coloneqq \frac{F^\Theta_x}{\norm{F^\Theta_x}}$. We have
\begin{enumerate}[\upshape(1)]
\item $\norm{f}=1$;
\item $f(x)=\sqrt{\mu_x^\Theta(\delta)}$;
\item $f\in\img\bigl(I_\Theta(\delta)\bigr)$.
\end{enumerate}
\end{lemma}
\begin{proof}
The first and third assertions are obvious. As for the second assertion, by \Cref{embed_norm_com},
\begin{displaymath}
f(x)
=
\frac{F^\Theta_x(x)}{\norm{F^\Theta_x}}
=
\frac{\mu_x^\Theta(\delta)}{\sqrt{\mu_x^\Theta(\delta)}}
=
\sqrt{\mu_x^\Theta(\delta)}.\qedhere
\end{displaymath}
\end{proof}
We are now in position to present bounds on vertex spectral measures.
Denote the eigenvalues of the combinatorial signless Laplacian $\Theta$ on $G$ as
\begin{displaymath}
0 \leqslant \lambda^\Theta_{\mathrm{min}} = \lambda^\Theta_1
\leqslant \lambda^\Theta_2 \leqslant \lambda^\Theta_3 \leqslant \dotsb \leqslant \lambda^\Theta_{n-1}
< \lambda^\Theta_n = \lambda^\Theta_{\mathrm{max}}.
\end{displaymath}
Recall that $\kg$ is defined in \Cref{def_kudos}; \Cref{est} shows that $\kg\geqslant \frac{1}{\D(G)+1}$ when $G$ is non-bipartite.
\begin{proposition}\label{comb_spec}
Let $G$ be a non-bipartite, finite, simple, connected, weighted graph with weight at least\/ $1$ for each edge. Then for each $\delta\in [0,\lambda^\Theta_{\mathrm{max}})$ and $x\in V$, we have
\begin{displaymath}
\mu_x^\Theta(\delta)
\leqslant
\frac{\delta}{\kg}
\leqslant
\bigl(\D(G)+1\bigr)\delta.
\end{displaymath}
\end{proposition}
\begin{proof}
This proof is a mimic of the proof of \Cref{mix_spec_prep}. Fixing a vertex $x\in V$, we define $f$ as in \Cref{def_f_com}.
\par
When $0\leqslant \delta<\lambda_{\mathrm{min}}^\Theta$, $\mu_x^\Theta(\delta)=0$ by definition. So the inequality holds automatically.
\par
When $\lambda^\Theta_{\mathrm{min}}\leqslant \delta<\lambda^\Theta_{\mathrm{max}}$, we know that $f$ is orthogonal to the eigenspace of $\Theta$ corresponding to $\lambda^\Theta_{\mathrm{max}}$, which is spanned by a positive vector according to the Perron--Frobenius theorem. By \Cref{def_kudos}, we have
\begin{displaymath}
\delta \geqslant \langle f,\Theta f\rangle
=
\sum_{(v,u)\in E(G)}w(v,u)\abs{f(v)+f(u)}^2
\geqslant f(x)^2\kg
=\mu_x^\Theta(\delta)\kg.\qedhere
\end{displaymath}
\end{proof}
In order to get a lower bound on eigenvalues of $\Theta$ from \Cref{comb_spec}, we need the following \Cref{spec_loc_com}.
\begin{lemma}\label{spec_loc_com}
Let $G$ be a finite, connected, weighted graph. We have
\begin{displaymath}
\sum_{x\in V}\mu_x^\Theta(\delta)=\abs[\big]{\{j\st \lambda_j^\Theta\leqslant \delta\}}.
\end{displaymath}
\end{lemma}
\begin{proof}
Let $h_1$, $h_2$, $\ldots$, $h_n$ be an orthonormal basis of $\ell^2(V)$ such that
\begin{displaymath}
\Theta h_j=\lambda^\Theta_j h_j,\qquad 1\leqslant j\leqslant n.
\end{displaymath}
It follows that
\begin{displaymath}
\sum_{x\in V}\abs{\langle h_j,\ind_x\rangle}^2
=\norm{h_j}^2=1.
\end{displaymath}
Therefore,
\begin{displaymath}
\sum_{x\in V}\mu_x^\Theta(\delta)
=
\sum_{x\in V}\bigl\langle I_\Theta(\delta)\ind_x,\ind_x\bigr\rangle.
\end{displaymath}
This is the trace of $I_\Theta(\delta)$, which equals the dimension of its image.
\end{proof}
\begin{corollary}\label{com_s_lap}
Let $G$ be a non-bipartite, finite, simple, connected, weighted graph with weight at least\/ $1$ for each edge. For\/ $1\leqslant k\leqslant n$, we have
\begin{displaymath}
\lambda^\Theta_k
\geqslant \frac{k\kg}{n}
\geqslant\frac{k}{\bigl(\D(G)+1\bigr)n}.
\end{displaymath}
\end{corollary}
\begin{proof}
By \Cref{comb_spec},
\begin{displaymath}
\sum_{x\in V}\mu_x^\Theta(\delta)
\leqslant
\sum_{x\in V}\frac{\delta}{\kg}=\frac{n\delta}{\kg}.
\end{displaymath}
Hence, \Cref{spec_loc_com} gives
\begin{displaymath}
\abs[\big]{\{j\st \lambda_j^\Theta\leqslant \delta\}}
\leqslant
\frac{n\delta}{\kg}.
\end{displaymath}
Therefore, we must have
$\lambda^\Theta_k\geqslant\frac{k\kg}{n}$.
\end{proof}
For the adjacency matrix $A$ of $G$, we denote its eigenvalues as
\begin{displaymath}
-w_{\mathrm{max}} \leqslant \lambda^A_{\mathrm{min}} = \lambda^A_1
\leqslant \lambda^A_2 \leqslant \lambda^A_3 \leqslant \dotsb \leqslant \lambda^A_{n-1} < \lambda^A_n= \lambda^A_{\mathrm{max}} \leqslant w_{\mathrm{max}},
\end{displaymath}
where $w_{\mathrm{max}}\coloneqq \max_{x\in V}w(x)$. The following \Cref{Alon_ext} improves \autocite[Theorem~1.1]{Alon-2000}, which obtained that $\dmx+\lambda_1^A\geqslant\frac{1}{(\D(G)+1)n}$ for unweighted graphs.
\begin{corollary}\label{Alon_ext}
Let $G$ be a non-bipartite, finite, simple, connected, weighted graph with weight at least\/ $1$ for each edge. For\/ $1\leqslant k\leqslant n$, we have
\begin{displaymath}
w_{\mathrm{max}}+\lambda_k^A
\geqslant
\frac{k\kg}{n}
\geqslant
\frac{k}{\bigl(\D(G)+1\bigr)n}.
\end{displaymath}
\end{corollary}
\begin{proof}
Let $X_1$ be the linear subspace of $\ell^2(V)$ spanned by the eigenvectors of $A$ corresponding to $\lambda^A_1$, $\lambda^A_2$, $\ldots$$, \lambda^A_k$, and $X_2$ be the linear subspace of $\ell^2(V)$ spanned by the eigenvectors of $\Theta$ corresponding to $\lambda^\Theta_k$, $\lambda^\Theta_{k+1}$, $\ldots$, $\lambda^\Theta_n$. Then
\begin{displaymath}
\dim X_1\geqslant k,\quad \dim X_2\geqslant n-k+1.
\end{displaymath}
It follows that
\begin{displaymath}
\dim{X_1}+\dim{X_2}>n=\dim{\ell^2(V)}.
\end{displaymath}
Therefore, the intersection of $X_1$ and $X_2$ contains a non-zero vector $h$ of unit norm. Hence,
\begin{displaymath}
\frac{k\kg}{n}\leqslant \lambda^\Theta_k
\leqslant (\Theta h, h)=(Wh,h)+(Ah,h)\leqslant w_{\mathrm{max}}+\lambda_k^A.\qedhere
\end{displaymath}
\end{proof}
\begin{rmk}
The combinatorial signless Laplacian is also related to line graphs. Let $\rho_1\geqslant \rho_2\geqslant \cdots\geqslant \rho_r$ be the positive combinatorial signless Laplacian eigenvalues of $G$. Then by \autocite[Proposition~1.4.1]{Brouwer-2011}, the eigenvalues of the line graph of $G$ are
\begin{displaymath}\begin{split}
\theta_i&=\rho_i-2,\qquad i=1,2,\ldots,r,\\
\theta_i&=-2,\qquad i=r+1,r+2,\ldots,\abs{E(G)}.
\end{split}\end{displaymath}
Using this relation and \Cref{com_s_lap}, one may do some quick analysis on the spectrum of the line graph of $G$.
\end{rmk}
% $ biblatex auxiliary file $
% $ biblatex bbl format version 3.1 $
% Do not modify the above lines!
%
% This is an auxiliary file used by the 'biblatex' package.
% This file may safely be deleted. It will be recreated by
% biber as required.
%
\begingroup
\makeatletter
\@ifundefined{ver@biblatex.sty}
{\@latex@error
{Missing 'biblatex' package}
{The bibliography requires the 'biblatex' package.}
\aftergroup\endinput}
{}
\endgroup
\refsection{0}
\datalist[list]{shorthand:shorthand/global//global/global}
\entry{Lyons-2018}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=768832bf8a096d6166cf6869e8f25cfa}{%
family={Lyons},
familyi={L\bibinitperiod},
given={Russell},
giveni={R\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=b2cdc229226b321211a521789842b3f8}{%
family={Oveis\bibnamedelima Gharan},
familyi={O\bibinitperiod\bibinitdelim G\bibinitperiod},
given={Shayan},
giveni={S\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{fullhash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{bibnamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authorbibnamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authornamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authorfullhash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\field{extraname}{1}
\field{sortinit}{L}
\field{sortinithash}{7c47d417cecb1f4bd38d1825c427a61a}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{abstract}{Spectral embedding of graphs uses the top $k$ non-trivial eigenvectors of the random walk matrix to embed the graph into $\mathbb{R}^k$. The primary use of this embedding has been for practical spectral clustering algorithms [39, 43]. Recently, spectral embedding was studied from a theoretical perspective to prove higher order variants of Cheeger’s inequality [27, 30].We use spectral embedding to provide a unifying framework for bounding all the eigenvalues of graphs. For example, we show that for any finite connected graph with $n$ vertices and all $k \ge 2$, the $k$th largest eigenvalue is at most $1-\Omega(k^3/n^3)$, which extends the only other such result known, which is for $k=2$ only and is due to [26]. This upper bound improves to $1-\Omega(k^2/n^2)$ if the graph is regular. We generalize these results, and we provide sharp bounds on the spectral measure of various classes of graphs, including vertex-transitive graphs and infinite graphs, in terms of specific graph parameters like the volume growth.As a consequence, using the entire spectrum, we provide (improved) upper bounds on the return probabilities and mixing time of random walks with considerably shorter and more direct proofs. Our work introduces spectral embedding as a new tool in analysing reversible Markov chains. Furthermore, building on [33], we design a local algorithm to approximate the number of spanning trees of massive graphs.}
\field{issn}{1073-7928 1687-0247}
\field{journaltitle}{International Mathematics Research Notices}
\field{number}{24}
\field{shorthand}{LOG18}
\field{title}{Sharp Bounds On Random Walk Eigenvalues Via Spectral Embedding}
\field{type}{Journal Article}
\field{volume}{2018}
\field{year}{2018}
\field{pages}{7555\bibrangedash 7605}
\range{pages}{51}
\verb{doi}
\verb 10.1093/imrn/rnx082
\endverb
\verb{urlraw}
\verb https://doi.org/10.1093/imrn/rnx082 https://watermark.silverchair.com/rnx082.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAqIwggKeBgkqhkiG9w0BBwagggKPMIICiwIBADCCAoQGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMhQRrmRLL7Ch-IWqgAgEQgIICVVBkabO29YTtlbfXnBznTpzktIa8EOlnOCDVjjfOAAwxPjm9os7k_UiZAa9RiqY7gSMTkNQfyIPwhwFzUMRcV9icWDnBom-jnxOYOqRBa4_dT8DRNHsPX8nh59yslE6RNiFX51in32G6DNSVD6-igyrQqfGhiLYonRuA-lm_dn3mDD4sjP6ZAyC0sls_dqzUmx-7J2NU62k9NYiZt2FmM0Zuw1reqbhiL4tDyEcZflzA7_xOfG8azeIg45-hZwWbuy7na_ZN2aOo6bu5n7eosjYTiJrEmp6i2pw7sfX9RjLb9fN854ID1EDI-XYYjZVsnYUqvcUQuwU7lOeDwY3FzLHHJ55UKG0xlvgH4o8u4bArkOEJCbCB3UE6gm8k_5Z1AogTL-mc1wKPG20Kthc0URW5sfrkLBabG-YpDuWaIgtZMevn2dfJuFr4mRW5sTXI8_EIEGG852FbYE_yRHNOJcNR59IobGN0__yOYRIkdd0mukGHd-SMO8p40H0YF25MSm-6_3znEP2rd-iJWl67BI7eYGlKbD41DeQWO07Zpt-H-MNenD7AFcGU8RYdDbLdOPxvYjq5SQLqg8Wk8-mt-jLu9-MYx-BkdEzssL4snAbDaT3BVeaqlTGNbg_FPWWxxzeA8nCRC8BQhHHCC7AuoP58aSGIzZkyhljmrS3L5J_opLtlbAbbDhf1LcFMWH_TqTZp7VGaBtkPW8AoIYs-70M9mMRE_mItSS24_1xthq8bpPqhyAP61RFLGVUPB1xlnGGqb3Hj8KHCfoKyr9NZBoc0N6uZ2g https://watermark.silverchair.com/rnx082.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAqMwggKfBgkqhkiG9w0BBwagggKQMIICjAIBADCCAoUGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMpTy2k5dkp22jsiDEAgEQgIICVvOizAs8Wjga_dE8dOauC04x1V7oB1Y7k1XQpn-r8Dofz_2ujgHm2jHaeFwb7m9KZpUS2uWj7pXBgZCzwCp5WXwlDPlS3OGyJcwY3kxWVbPNspZophwj0JD9AyyLr6Y5_oF6lk4UTYJo4znp6UvFG8Hhu98tE_EnY_nGXRK1CI_KCNr5SFTXs14LKrpCKFbFZXnVbLMM-qwrYFRTOEs58I4M66y2nCDZqLFzAuF4rCGHxihWh0tBftZLuUEYa7CpTLfJuU1Nh5JyCBnJBigwhoiAwF7-IkOyVPZgKabDOY9ZcB269cb0bYPACdWTygYCXFod8kTanwHeHW576_GPMvauJhAod3jnS9Kupr_p9adgLI6QPF0622O60YsvWzIx9KjO-bCAGBxqnEhVKZFfHVIV---dNXFNhYx82Bf10xXilKg8Xjaht01DyRrFkqTK9XilujJH6bRmQzaWkhekgbARMt6u-BKwUFHNgnYx8T0YHBPA1Snyc5iYoqkftzZFyX3IfBywKJkigYjDOEyuvRu0DoCnIYLjfFIP8Tis9x3F9szpHOi4BOdH5cBxW0XC9CQLTCbhwyElaT1CKbuu0wxnsi7R1gxR2zNHaQe6AVK4LlMa6ybJ_4Z55jCKW-QWEmVjnPPASJDDmZkPaFOsn4dbmwefF9X1TXPZX3jO5GX1mtR9zB9J_OIEjkhCIdERWG5DxXYkvfzurmtpMybgbIhXvBpztvmKZ78SAzLOSzUEHWeC7-GA-mwgQVgW-OsBOB8-eutF7W7kbCWoVe1D8x5U_i9HZX8
\endverb
\verb{url}
\verb https://doi.org/10.1093/imrn/rnx082%20https://watermark.silverchair.com/rnx082.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAqIwggKeBgkqhkiG9w0BBwagggKPMIICiwIBADCCAoQGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMhQRrmRLL7Ch-IWqgAgEQgIICVVBkabO29YTtlbfXnBznTpzktIa8EOlnOCDVjjfOAAwxPjm9os7k_UiZAa9RiqY7gSMTkNQfyIPwhwFzUMRcV9icWDnBom-jnxOYOqRBa4_dT8DRNHsPX8nh59yslE6RNiFX51in32G6DNSVD6-igyrQqfGhiLYonRuA-lm_dn3mDD4sjP6ZAyC0sls_dqzUmx-7J2NU62k9NYiZt2FmM0Zuw1reqbhiL4tDyEcZflzA7_xOfG8azeIg45-hZwWbuy7na_ZN2aOo6bu5n7eosjYTiJrEmp6i2pw7sfX9RjLb9fN854ID1EDI-XYYjZVsnYUqvcUQuwU7lOeDwY3FzLHHJ55UKG0xlvgH4o8u4bArkOEJCbCB3UE6gm8k_5Z1AogTL-mc1wKPG20Kthc0URW5sfrkLBabG-YpDuWaIgtZMevn2dfJuFr4mRW5sTXI8_EIEGG852FbYE_yRHNOJcNR59IobGN0__yOYRIkdd0mukGHd-SMO8p40H0YF25MSm-6_3znEP2rd-iJWl67BI7eYGlKbD41DeQWO07Zpt-H-MNenD7AFcGU8RYdDbLdOPxvYjq5SQLqg8Wk8-mt-jLu9-MYx-BkdEzssL4snAbDaT3BVeaqlTGNbg_FPWWxxzeA8nCRC8BQhHHCC7AuoP58aSGIzZkyhljmrS3L5J_opLtlbAbbDhf1LcFMWH_TqTZp7VGaBtkPW8AoIYs-70M9mMRE_mItSS24_1xthq8bpPqhyAP61RFLGVUPB1xlnGGqb3Hj8KHCfoKyr9NZBoc0N6uZ2g%20https://watermark.silverchair.com/rnx082.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAqMwggKfBgkqhkiG9w0BBwagggKQMIICjAIBADCCAoUGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMpTy2k5dkp22jsiDEAgEQgIICVvOizAs8Wjga_dE8dOauC04x1V7oB1Y7k1XQpn-r8Dofz_2ujgHm2jHaeFwb7m9KZpUS2uWj7pXBgZCzwCp5WXwlDPlS3OGyJcwY3kxWVbPNspZophwj0JD9AyyLr6Y5_oF6lk4UTYJo4znp6UvFG8Hhu98tE_EnY_nGXRK1CI_KCNr5SFTXs14LKrpCKFbFZXnVbLMM-qwrYFRTOEs58I4M66y2nCDZqLFzAuF4rCGHxihWh0tBftZLuUEYa7CpTLfJuU1Nh5JyCBnJBigwhoiAwF7-IkOyVPZgKabDOY9ZcB269cb0bYPACdWTygYCXFod8kTanwHeHW576_GPMvauJhAod3jnS9Kupr_p9adgLI6QPF0622O60YsvWzIx9KjO-bCAGBxqnEhVKZFfHVIV---dNXFNhYx82Bf10xXilKg8Xjaht01DyRrFkqTK9XilujJH6bRmQzaWkhekgbARMt6u-BKwUFHNgnYx8T0YHBPA1Snyc5iYoqkftzZFyX3IfBywKJkigYjDOEyuvRu0DoCnIYLjfFIP8Tis9x3F9szpHOi4BOdH5cBxW0XC9CQLTCbhwyElaT1CKbuu0wxnsi7R1gxR2zNHaQe6AVK4LlMa6ybJ_4Z55jCKW-QWEmVjnPPASJDDmZkPaFOsn4dbmwefF9X1TXPZX3jO5GX1mtR9zB9J_OIEjkhCIdERWG5DxXYkvfzurmtpMybgbIhXvBpztvmKZ78SAzLOSzUEHWeC7-GA-mwgQVgW-OsBOB8-eutF7W7kbCWoVe1D8x5U_i9HZX8
\endverb
\endentry
\enddatalist
\datalist[entry]{nyt/global//global/global}
\entry{Alon-2000}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=f7a4a7818d41bcfdc8addc9e20cdcbd8}{%
family={Alon},
familyi={A\bibinitperiod},
given={Noga},
giveni={N\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=3f5218bd15d9c14a91245f7ecb0e44cd}{%
family={Sudakov},
familyi={S\bibinitperiod},
given={Benny},
giveni={B\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{1b81cb17765965a6aa6a05b9dc80926c}
\strng{fullhash}{1b81cb17765965a6aa6a05b9dc80926c}
\strng{bibnamehash}{1b81cb17765965a6aa6a05b9dc80926c}
\strng{authorbibnamehash}{1b81cb17765965a6aa6a05b9dc80926c}
\strng{authornamehash}{1b81cb17765965a6aa6a05b9dc80926c}
\strng{authorfullhash}{1b81cb17765965a6aa6a05b9dc80926c}
\field{sortinit}{A}
\field{sortinithash}{2f401846e2029bad6b3ecc16d50031e2}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{issn}{09635483}
\field{journaltitle}{Combinatorics, Probability and Computing}
\field{number}{1}
\field{title}{Bipartite Subgraphs And The Smallest Eigenvalue}
\field{type}{Journal Article}
\field{volume}{9}
\field{year}{2000}
\field{pages}{1\bibrangedash 12}
\range{pages}{12}
\verb{doi}
\verb 10.1017/s0963548399004071
\endverb
\verb{urlraw}
\verb https://www.cambridge.org/core/journals/combinatorics-probability-and-computing/article/abs/bipartite-subgraphs-and-the-smallest-eigenvalue/B8556CD8A8819800423D0D9A8B6EC6EB
\endverb
\verb{url}
\verb https://www.cambridge.org/core/journals/combinatorics-probability-and-computing/article/abs/bipartite-subgraphs-and-the-smallest-eigenvalue/B8556CD8A8819800423D0D9A8B6EC6EB
\endverb
\endentry
\entry{Brouwer-2011}{book}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=c44c6302e4ceeec9ad7e71fe929ff825}{%
family={Brouwer},
familyi={B\bibinitperiod},
given={Andries\bibnamedelima E.},
giveni={A\bibinitperiod\bibinitdelim E\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=fc2b54fc71a57838bfec1cc92bc42473}{%
family={Haemers},
familyi={H\bibinitperiod},
given={Willem\bibnamedelima H.},
giveni={W\bibinitperiod\bibinitdelim H\bibinitperiod},
givenun=0}}%
}
\list{publisher}{1}{%
{Springer-Verlag New York}%
}
\strng{namehash}{a929539c7c298f5be84afc5a14171d63}
\strng{fullhash}{a929539c7c298f5be84afc5a14171d63}
\strng{bibnamehash}{a929539c7c298f5be84afc5a14171d63}
\strng{authorbibnamehash}{a929539c7c298f5be84afc5a14171d63}
\strng{authornamehash}{a929539c7c298f5be84afc5a14171d63}
\strng{authorfullhash}{a929539c7c298f5be84afc5a14171d63}
\field{sortinit}{B}
\field{sortinithash}{d7095fff47cda75ca2589920aae98399}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{edition}{1}
\field{isbn}{978-1-4614-1938-9 978-1-4614-1939-6}
\field{series}{Universitext}
\field{title}{Spectra of Graphs}
\field{type}{Book}
\field{year}{2012}
\verb{doi}
\verb 10.1007/978-1-4614-1939-6
\endverb
\endentry
\entry{Cvetkovic-2009}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=f6f81f89ff24c3c90560d8900453e1bd}{%
family={Cvetković},
familyi={C\bibinitperiod},
given={Dragoš},
giveni={D\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=662d65e49cb5a1ada6dc0c5e81ae0c13}{%
family={Simić},
familyi={S\bibinitperiod},
given={Slobodan\bibnamedelima K.},
giveni={S\bibinitperiod\bibinitdelim K\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{fullhash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{bibnamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authorbibnamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authornamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authorfullhash}{893ba24e7683cbae6ecba7d19c0a9a75}
\field{extraname}{1}
\field{sortinit}{C}
\field{sortinithash}{4d103a86280481745c9c897c925753c0}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{issn}{0350-1302 1820-7405}
\field{journaltitle}{Publications de l'Institut Mathematique}
\field{number}{99}
\field{title}{Towards A Spectral Theory Of Graphs Based On The Signless {L}aplacian, {I}}
\field{type}{Journal Article}
\field{volume}{85}
\field{year}{2009}
\field{pages}{19\bibrangedash 33}
\range{pages}{15}
\verb{doi}
\verb 10.2298/pim0999019c
\endverb
\verb{urlraw}
\verb http://www.doiserbia.nb.rs/img/doi/0350-1302/2009/0350-13020999019C.pdf
\endverb
\verb{url}
\verb http://www.doiserbia.nb.rs/img/doi/0350-1302/2009/0350-13020999019C.pdf
\endverb
\endentry
\entry{Cvetkovic-2010a}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=f6f81f89ff24c3c90560d8900453e1bd}{%
family={Cvetković},
familyi={C\bibinitperiod},
given={Dragoš},
giveni={D\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=662d65e49cb5a1ada6dc0c5e81ae0c13}{%
family={Simić},
familyi={S\bibinitperiod},
given={Slobodan\bibnamedelima K.},
giveni={S\bibinitperiod\bibinitdelim K\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{fullhash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{bibnamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authorbibnamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authornamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authorfullhash}{893ba24e7683cbae6ecba7d19c0a9a75}
\field{extraname}{2}
\field{sortinit}{C}
\field{sortinithash}{4d103a86280481745c9c897c925753c0}
\field{extradate}{1}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{abstract}{A spectral graph theory is a theory in which graphs are studied by means of eigenvalues of a matrix M which is in a prescribed way defined for any graph. This theory is called M-theory. We outline a spectral theory of graphs based on the signless Laplacians Q and compare it with other spectral theories, in particular to those based on the adjacency matrix A and the Laplacian L. As demonstrated in the first part, the Q-theory can be constructed in part using various connections to other theories: equivalency with A-theory and L-theory for regular graphs, common features with L-theory for bipartite graphs, general analogies with A-theory and analogies with A-theory via line graphs and subdivision graphs. In this part, we introduce notions of enriched and restricted spectral theories and present results on integral graphs, enumeration of spanning trees, characterizations by eigenvalues, cospectral graphs and graph angles.}
\field{issn}{00243795}
\field{journaltitle}{Linear Algebra and its Applications}
\field{number}{9}
\field{title}{Towards A Spectral Theory Of Graphs Based On The Signless {L}aplacian, {II}}
\field{type}{Journal Article}
\field{volume}{432}
\field{year}{2010}
\field{pages}{2257\bibrangedash 2272}
\range{pages}{16}
\verb{doi}
\verb 10.1016/j.laa.2009.05.020
\endverb
\verb{urlraw}
\verb https://www.sciencedirect.com/science/article/pii/S0024379509002808
\endverb
\verb{url}
\verb https://www.sciencedirect.com/science/article/pii/S0024379509002808
\endverb
\keyw{Graph theory Graph spectra Adjacency matrix Laplacian Signless Laplacian}
\endentry
\entry{Cvetkovic-2010b}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=f6f81f89ff24c3c90560d8900453e1bd}{%
family={Cvetković},
familyi={C\bibinitperiod},
given={Dragoš},
giveni={D\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=662d65e49cb5a1ada6dc0c5e81ae0c13}{%
family={Simić},
familyi={S\bibinitperiod},
given={Slobodan\bibnamedelima K.},
giveni={S\bibinitperiod\bibinitdelim K\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{fullhash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{bibnamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authorbibnamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authornamehash}{893ba24e7683cbae6ecba7d19c0a9a75}
\strng{authorfullhash}{893ba24e7683cbae6ecba7d19c0a9a75}
\field{extraname}{3}
\field{sortinit}{C}
\field{sortinithash}{4d103a86280481745c9c897c925753c0}
\field{extradate}{2}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{abstract}{[This part of our work further extends our project of building a new spectral theory of graphs (based on the signless Laplacian) by some results on graph angles, by several comments and by a short survey of recent results.]}
\field{issn}{14528630, 2406100X}
\field{journaltitle}{Applicable Analysis and Discrete Mathematics}
\field{number}{1}
\field{title}{Towards A Spectral Theory Of Graphs Based On The Signless {L}aplacian, {III}}
\field{type}{Journal Article}
\field{volume}{4}
\field{year}{2010}
\field{pages}{156\bibrangedash 166}
\range{pages}{11}
\verb{doi}
\verb 0.2298/AADM1000001C
\endverb
\verb{urlraw}
\verb http://www.jstor.org/stable/43671298
\endverb
\verb{url}
\verb http://www.jstor.org/stable/43671298
\endverb
\endentry
\entry{Davis-1995}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=f27d50799e4b09903fbd2cac88b3a011}{%
family={Davis},
familyi={D\bibinitperiod},
given={Burgess},
giveni={B\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=c8291e67ecfab6574b93fb053e06fc06}{%
family={McDonald},
familyi={M\bibinitperiod},
given={David},
giveni={D\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{ece47e91e208929b082debbea79d7d4a}
\strng{fullhash}{ece47e91e208929b082debbea79d7d4a}
\strng{bibnamehash}{ece47e91e208929b082debbea79d7d4a}
\strng{authorbibnamehash}{ece47e91e208929b082debbea79d7d4a}
\strng{authornamehash}{ece47e91e208929b082debbea79d7d4a}
\strng{authorfullhash}{ece47e91e208929b082debbea79d7d4a}
\field{sortinit}{D}
\field{sortinithash}{6f385f66841fb5e82009dc833c761848}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{abstract}{We give an elementary proof of the local central limit theorem for independent, non-identically distributed, integer valued and vector valued random variables.}
\field{issn}{1572-9230}
\field{journaltitle}{Journal of Theoretical Probability}
\field{number}{3}
\field{title}{An Elementary Proof of The Local Central Limit Theorem}
\field{type}{Journal Article}
\field{volume}{8}
\field{year}{1995}
\field{pages}{693\bibrangedash 701}
\range{pages}{9}
\verb{doi}
\verb 10.1007/BF02218051
\endverb
\verb{urlraw}
\verb https://doi.org/10.1007/BF02218051 https://link.springer.com/article/10.1007/BF02218051
\endverb
\verb{url}
\verb https://doi.org/10.1007/BF02218051%20https://link.springer.com/article/10.1007/BF02218051
\endverb
\endentry
\entry{Desai-1993}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=86463318be96b83b911bb478a33c700b}{%
family={Desai},
familyi={D\bibinitperiod},
given={Madhav\bibnamedelima P.},
giveni={M\bibinitperiod\bibinitdelim P\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=c2e4b3d2f5a66aa2f9c8d5935633fa37}{%
family={Rao},
familyi={R\bibinitperiod},
given={Vasant\bibnamedelima B.},
giveni={V\bibinitperiod\bibinitdelim B\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{ecb369d19f01aba9a5e22c5d783fa4da}
\strng{fullhash}{ecb369d19f01aba9a5e22c5d783fa4da}
\strng{bibnamehash}{ecb369d19f01aba9a5e22c5d783fa4da}
\strng{authorbibnamehash}{ecb369d19f01aba9a5e22c5d783fa4da}
\strng{authornamehash}{ecb369d19f01aba9a5e22c5d783fa4da}
\strng{authorfullhash}{ecb369d19f01aba9a5e22c5d783fa4da}
\field{sortinit}{D}
\field{sortinithash}{6f385f66841fb5e82009dc833c761848}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{issn}{0895-4798 1095-7162}
\field{journaltitle}{SIAM Journal on Matrix Analysis and Applications}
\field{number}{4}
\field{title}{On the Convergence of Reversible {M}arkov Chains}
\field{type}{Journal Article}
\field{volume}{14}
\field{year}{1993}
\field{pages}{950\bibrangedash 966}
\range{pages}{17}
\verb{doi}
\verb 10.1137/0614063
\endverb
\endentry
\entry{Diaconis-1991}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=f5a7dd4f4a0c79d2eaaf29d6dda0f4b3}{%
family={Diaconis},
familyi={D\bibinitperiod},
given={Persi},
giveni={P\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=aa81929f063bb9d99abb8232145aa275}{%
family={Stroock},
familyi={S\bibinitperiod},
given={Daniel},
giveni={D\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{a2efbe206418b980de7f02dada3d2676}
\strng{fullhash}{a2efbe206418b980de7f02dada3d2676}
\strng{bibnamehash}{a2efbe206418b980de7f02dada3d2676}
\strng{authorbibnamehash}{a2efbe206418b980de7f02dada3d2676}
\strng{authornamehash}{a2efbe206418b980de7f02dada3d2676}
\strng{authorfullhash}{a2efbe206418b980de7f02dada3d2676}
\field{sortinit}{D}
\field{sortinithash}{6f385f66841fb5e82009dc833c761848}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{journaltitle}{The Annals of Applied Probability}
\field{number}{1}
\field{title}{Geometric Bounds For Eigenvalues Of {M}arkov Chains}
\field{type}{Journal Article}
\field{volume}{1}
\field{year}{1991}
\field{pages}{36\bibrangedash 61}
\range{pages}{26}
\verb{doi}
\verb 10.1214/aoap/1177005980
\endverb
\verb{urlraw}
\verb https://doi.org/10.1214/aoap/1177005980
\endverb
\verb{url}
\verb https://doi.org/10.1214/aoap/1177005980
\endverb
\endentry
\entry{Landau-1981}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=378ea689326d84243a4c44f1d35790c2}{%
family={Landau},
familyi={L\bibinitperiod},
given={Henry\bibnamedelima J.},
giveni={H\bibinitperiod\bibinitdelim J\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=fe8d3ebce15150423fcc5edb25315ca2}{%
family={Odlyzko},
familyi={O\bibinitperiod},
given={Andrew\bibnamedelima M.},
giveni={A\bibinitperiod\bibinitdelim M\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{e1529692b2fef09eabe5cdc27c233230}
\strng{fullhash}{e1529692b2fef09eabe5cdc27c233230}
\strng{bibnamehash}{e1529692b2fef09eabe5cdc27c233230}
\strng{authorbibnamehash}{e1529692b2fef09eabe5cdc27c233230}
\strng{authornamehash}{e1529692b2fef09eabe5cdc27c233230}
\strng{authorfullhash}{e1529692b2fef09eabe5cdc27c233230}
\field{sortinit}{L}
\field{sortinithash}{7c47d417cecb1f4bd38d1825c427a61a}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{issn}{00243795}
\field{journaltitle}{Linear Algebra and its Applications}
\field{title}{Bounds for Eigenvalues of Certain Stochastic Matrices}
\field{type}{Journal Article}
\field{volume}{38}
\field{year}{1981}
\field{pages}{5\bibrangedash 15}
\range{pages}{11}
\verb{doi}
\verb 10.1016/0024-3795(81)90003-3
\endverb
\endentry
\entry{Levin-2017}{book}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=6134a1a3704cf87d615ec07a8271e311}{%
family={Levin},
familyi={L\bibinitperiod},
given={David\bibnamedelima A},
giveni={D\bibinitperiod\bibinitdelim A\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=be791fcf2bac698f7a6b715cdf5d5d51}{%
family={Peres},
familyi={P\bibinitperiod},
given={Yuval},
giveni={Y\bibinitperiod},
givenun=0}}%
}
\list{publisher}{1}{%
{American Mathematical Society}%
}
\strng{namehash}{7611f59ec0afc64d873af92ca607324e}
\strng{fullhash}{7611f59ec0afc64d873af92ca607324e}
\strng{bibnamehash}{7611f59ec0afc64d873af92ca607324e}
\strng{authorbibnamehash}{7611f59ec0afc64d873af92ca607324e}
\strng{authornamehash}{7611f59ec0afc64d873af92ca607324e}
\strng{authorfullhash}{7611f59ec0afc64d873af92ca607324e}
\field{sortinit}{L}
\field{sortinithash}{7c47d417cecb1f4bd38d1825c427a61a}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{isbn}{1470429624}
\field{title}{{M}arkov Chains and Mixing Times}
\field{type}{Book}
\field{volume}{107}
\field{year}{2017}
\verb{doi}
\verb 10.1090/mbk/107
\endverb
\endentry
\entry{Lyons-2018}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=768832bf8a096d6166cf6869e8f25cfa}{%
family={Lyons},
familyi={L\bibinitperiod},
given={Russell},
giveni={R\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=b2cdc229226b321211a521789842b3f8}{%
family={Oveis\bibnamedelima Gharan},
familyi={O\bibinitperiod\bibinitdelim G\bibinitperiod},
given={Shayan},
giveni={S\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{fullhash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{bibnamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authorbibnamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authornamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authorfullhash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\field{extraname}{1}
\field{sortinit}{L}
\field{sortinithash}{7c47d417cecb1f4bd38d1825c427a61a}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{abstract}{Spectral embedding of graphs uses the top $k$ non-trivial eigenvectors of the random walk matrix to embed the graph into $\mathbb{R}^k$. The primary use of this embedding has been for practical spectral clustering algorithms [39, 43]. Recently, spectral embedding was studied from a theoretical perspective to prove higher order variants of Cheeger’s inequality [27, 30].We use spectral embedding to provide a unifying framework for bounding all the eigenvalues of graphs. For example, we show that for any finite connected graph with $n$ vertices and all $k \ge 2$, the $k$th largest eigenvalue is at most $1-\Omega(k^3/n^3)$, which extends the only other such result known, which is for $k=2$ only and is due to [26]. This upper bound improves to $1-\Omega(k^2/n^2)$ if the graph is regular. We generalize these results, and we provide sharp bounds on the spectral measure of various classes of graphs, including vertex-transitive graphs and infinite graphs, in terms of specific graph parameters like the volume growth.As a consequence, using the entire spectrum, we provide (improved) upper bounds on the return probabilities and mixing time of random walks with considerably shorter and more direct proofs. Our work introduces spectral embedding as a new tool in analysing reversible Markov chains. Furthermore, building on [33], we design a local algorithm to approximate the number of spanning trees of massive graphs.}
\field{issn}{1073-7928 1687-0247}
\field{journaltitle}{International Mathematics Research Notices}
\field{number}{24}
\field{shorthand}{LOG18}
\field{title}{Sharp Bounds On Random Walk Eigenvalues Via Spectral Embedding}
\field{type}{Journal Article}
\field{volume}{2018}
\field{year}{2018}
\field{pages}{7555\bibrangedash 7605}
\range{pages}{51}
\verb{doi}
\verb 10.1093/imrn/rnx082
\endverb
\verb{urlraw}
\verb https://doi.org/10.1093/imrn/rnx082 https://watermark.silverchair.com/rnx082.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAqIwggKeBgkqhkiG9w0BBwagggKPMIICiwIBADCCAoQGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMhQRrmRLL7Ch-IWqgAgEQgIICVVBkabO29YTtlbfXnBznTpzktIa8EOlnOCDVjjfOAAwxPjm9os7k_UiZAa9RiqY7gSMTkNQfyIPwhwFzUMRcV9icWDnBom-jnxOYOqRBa4_dT8DRNHsPX8nh59yslE6RNiFX51in32G6DNSVD6-igyrQqfGhiLYonRuA-lm_dn3mDD4sjP6ZAyC0sls_dqzUmx-7J2NU62k9NYiZt2FmM0Zuw1reqbhiL4tDyEcZflzA7_xOfG8azeIg45-hZwWbuy7na_ZN2aOo6bu5n7eosjYTiJrEmp6i2pw7sfX9RjLb9fN854ID1EDI-XYYjZVsnYUqvcUQuwU7lOeDwY3FzLHHJ55UKG0xlvgH4o8u4bArkOEJCbCB3UE6gm8k_5Z1AogTL-mc1wKPG20Kthc0URW5sfrkLBabG-YpDuWaIgtZMevn2dfJuFr4mRW5sTXI8_EIEGG852FbYE_yRHNOJcNR59IobGN0__yOYRIkdd0mukGHd-SMO8p40H0YF25MSm-6_3znEP2rd-iJWl67BI7eYGlKbD41DeQWO07Zpt-H-MNenD7AFcGU8RYdDbLdOPxvYjq5SQLqg8Wk8-mt-jLu9-MYx-BkdEzssL4snAbDaT3BVeaqlTGNbg_FPWWxxzeA8nCRC8BQhHHCC7AuoP58aSGIzZkyhljmrS3L5J_opLtlbAbbDhf1LcFMWH_TqTZp7VGaBtkPW8AoIYs-70M9mMRE_mItSS24_1xthq8bpPqhyAP61RFLGVUPB1xlnGGqb3Hj8KHCfoKyr9NZBoc0N6uZ2g https://watermark.silverchair.com/rnx082.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAqMwggKfBgkqhkiG9w0BBwagggKQMIICjAIBADCCAoUGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMpTy2k5dkp22jsiDEAgEQgIICVvOizAs8Wjga_dE8dOauC04x1V7oB1Y7k1XQpn-r8Dofz_2ujgHm2jHaeFwb7m9KZpUS2uWj7pXBgZCzwCp5WXwlDPlS3OGyJcwY3kxWVbPNspZophwj0JD9AyyLr6Y5_oF6lk4UTYJo4znp6UvFG8Hhu98tE_EnY_nGXRK1CI_KCNr5SFTXs14LKrpCKFbFZXnVbLMM-qwrYFRTOEs58I4M66y2nCDZqLFzAuF4rCGHxihWh0tBftZLuUEYa7CpTLfJuU1Nh5JyCBnJBigwhoiAwF7-IkOyVPZgKabDOY9ZcB269cb0bYPACdWTygYCXFod8kTanwHeHW576_GPMvauJhAod3jnS9Kupr_p9adgLI6QPF0622O60YsvWzIx9KjO-bCAGBxqnEhVKZFfHVIV---dNXFNhYx82Bf10xXilKg8Xjaht01DyRrFkqTK9XilujJH6bRmQzaWkhekgbARMt6u-BKwUFHNgnYx8T0YHBPA1Snyc5iYoqkftzZFyX3IfBywKJkigYjDOEyuvRu0DoCnIYLjfFIP8Tis9x3F9szpHOi4BOdH5cBxW0XC9CQLTCbhwyElaT1CKbuu0wxnsi7R1gxR2zNHaQe6AVK4LlMa6ybJ_4Z55jCKW-QWEmVjnPPASJDDmZkPaFOsn4dbmwefF9X1TXPZX3jO5GX1mtR9zB9J_OIEjkhCIdERWG5DxXYkvfzurmtpMybgbIhXvBpztvmKZ78SAzLOSzUEHWeC7-GA-mwgQVgW-OsBOB8-eutF7W7kbCWoVe1D8x5U_i9HZX8
\endverb
\verb{url}
\verb https://doi.org/10.1093/imrn/rnx082%20https://watermark.silverchair.com/rnx082.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAqIwggKeBgkqhkiG9w0BBwagggKPMIICiwIBADCCAoQGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMhQRrmRLL7Ch-IWqgAgEQgIICVVBkabO29YTtlbfXnBznTpzktIa8EOlnOCDVjjfOAAwxPjm9os7k_UiZAa9RiqY7gSMTkNQfyIPwhwFzUMRcV9icWDnBom-jnxOYOqRBa4_dT8DRNHsPX8nh59yslE6RNiFX51in32G6DNSVD6-igyrQqfGhiLYonRuA-lm_dn3mDD4sjP6ZAyC0sls_dqzUmx-7J2NU62k9NYiZt2FmM0Zuw1reqbhiL4tDyEcZflzA7_xOfG8azeIg45-hZwWbuy7na_ZN2aOo6bu5n7eosjYTiJrEmp6i2pw7sfX9RjLb9fN854ID1EDI-XYYjZVsnYUqvcUQuwU7lOeDwY3FzLHHJ55UKG0xlvgH4o8u4bArkOEJCbCB3UE6gm8k_5Z1AogTL-mc1wKPG20Kthc0URW5sfrkLBabG-YpDuWaIgtZMevn2dfJuFr4mRW5sTXI8_EIEGG852FbYE_yRHNOJcNR59IobGN0__yOYRIkdd0mukGHd-SMO8p40H0YF25MSm-6_3znEP2rd-iJWl67BI7eYGlKbD41DeQWO07Zpt-H-MNenD7AFcGU8RYdDbLdOPxvYjq5SQLqg8Wk8-mt-jLu9-MYx-BkdEzssL4snAbDaT3BVeaqlTGNbg_FPWWxxzeA8nCRC8BQhHHCC7AuoP58aSGIzZkyhljmrS3L5J_opLtlbAbbDhf1LcFMWH_TqTZp7VGaBtkPW8AoIYs-70M9mMRE_mItSS24_1xthq8bpPqhyAP61RFLGVUPB1xlnGGqb3Hj8KHCfoKyr9NZBoc0N6uZ2g%20https://watermark.silverchair.com/rnx082.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAqMwggKfBgkqhkiG9w0BBwagggKQMIICjAIBADCCAoUGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMpTy2k5dkp22jsiDEAgEQgIICVvOizAs8Wjga_dE8dOauC04x1V7oB1Y7k1XQpn-r8Dofz_2ujgHm2jHaeFwb7m9KZpUS2uWj7pXBgZCzwCp5WXwlDPlS3OGyJcwY3kxWVbPNspZophwj0JD9AyyLr6Y5_oF6lk4UTYJo4znp6UvFG8Hhu98tE_EnY_nGXRK1CI_KCNr5SFTXs14LKrpCKFbFZXnVbLMM-qwrYFRTOEs58I4M66y2nCDZqLFzAuF4rCGHxihWh0tBftZLuUEYa7CpTLfJuU1Nh5JyCBnJBigwhoiAwF7-IkOyVPZgKabDOY9ZcB269cb0bYPACdWTygYCXFod8kTanwHeHW576_GPMvauJhAod3jnS9Kupr_p9adgLI6QPF0622O60YsvWzIx9KjO-bCAGBxqnEhVKZFfHVIV---dNXFNhYx82Bf10xXilKg8Xjaht01DyRrFkqTK9XilujJH6bRmQzaWkhekgbARMt6u-BKwUFHNgnYx8T0YHBPA1Snyc5iYoqkftzZFyX3IfBywKJkigYjDOEyuvRu0DoCnIYLjfFIP8Tis9x3F9szpHOi4BOdH5cBxW0XC9CQLTCbhwyElaT1CKbuu0wxnsi7R1gxR2zNHaQe6AVK4LlMa6ybJ_4Z55jCKW-QWEmVjnPPASJDDmZkPaFOsn4dbmwefF9X1TXPZX3jO5GX1mtR9zB9J_OIEjkhCIdERWG5DxXYkvfzurmtpMybgbIhXvBpztvmKZ78SAzLOSzUEHWeC7-GA-mwgQVgW-OsBOB8-eutF7W7kbCWoVe1D8x5U_i9HZX8
\endverb
\endentry
\entry{Lyons-2021}{misc}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=768832bf8a096d6166cf6869e8f25cfa}{%
family={Lyons},
familyi={L\bibinitperiod},
given={Russell},
giveni={R\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=b2cdc229226b321211a521789842b3f8}{%
family={Oveis\bibnamedelima Gharan},
familyi={O\bibinitperiod\bibinitdelim G\bibinitperiod},
given={Shayan},
giveni={S\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{fullhash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{bibnamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authorbibnamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authornamehash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\strng{authorfullhash}{88f518e8b5ca19e2e6d5fe798d1d7ff3}
\field{extraname}{2}
\field{sortinit}{L}
\field{sortinithash}{7c47d417cecb1f4bd38d1825c427a61a}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{note}{Available at \url{https://rdlyons.pages.iu.edu/errata/peigs.pdf}}
\field{title}{Errata to ``Sharp Bounds on Random Walk Eigenvalues via Spectral Embedding''}
\field{type}{Manuscript}
\field{year}{2021}
\endentry
\entry{Lyons-2017}{book}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=768832bf8a096d6166cf6869e8f25cfa}{%
family={Lyons},
familyi={L\bibinitperiod},
given={Russell},
giveni={R\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=be791fcf2bac698f7a6b715cdf5d5d51}{%
family={Peres},
familyi={P\bibinitperiod},
given={Yuval},
giveni={Y\bibinitperiod},
givenun=0}}%
}
\list{publisher}{1}{%
{Cambridge University Press, New York}%
}
\strng{namehash}{a12173e5bdd8a5200ced01216607495b}
\strng{fullhash}{a12173e5bdd8a5200ced01216607495b}
\strng{bibnamehash}{a12173e5bdd8a5200ced01216607495b}
\strng{authorbibnamehash}{a12173e5bdd8a5200ced01216607495b}
\strng{authornamehash}{a12173e5bdd8a5200ced01216607495b}
\strng{authorfullhash}{a12173e5bdd8a5200ced01216607495b}
\field{sortinit}{L}
\field{sortinithash}{7c47d417cecb1f4bd38d1825c427a61a}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{isbn}{9781316672815}
\field{note}{Available at \url{https://rdlyons.pages.iu.edu/}}
\field{series}{Cambridge Series in Statistical and Probabilistic Mathematics}
\field{title}{Probability on Trees and Networks}
\field{type}{Book}
\field{volume}{42}
\field{year}{2017}
\verb{doi}
\verb 10.1017/9781316672815
\endverb
\verb{urlraw}
\verb http://dx.doi.org/10.1017/9781316672815
\endverb
\verb{url}
\verb http://dx.doi.org/10.1017/9781316672815
\endverb
\endentry
\entry{Mohar-1989}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=984a9c4e45e3489881e4a5220c9d1254}{%
family={Mohar},
familyi={M\bibinitperiod},
given={Bojan},
giveni={B\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=08dc1887cfc4e137166d98664c0ebfec}{%
family={Woess},
familyi={W\bibinitperiod},
given={Wolfgang},
giveni={W\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{60f3525e1bf20d054cd68b4d6c812039}
\strng{fullhash}{60f3525e1bf20d054cd68b4d6c812039}
\strng{bibnamehash}{60f3525e1bf20d054cd68b4d6c812039}
\strng{authorbibnamehash}{60f3525e1bf20d054cd68b4d6c812039}
\strng{authornamehash}{60f3525e1bf20d054cd68b4d6c812039}
\strng{authorfullhash}{60f3525e1bf20d054cd68b4d6c812039}
\field{sortinit}{M}
\field{sortinithash}{4625c616857f13d17ce56f7d4f97d451}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{issn}{00246093}
\field{journaltitle}{Bulletin of the London Mathematical Society}
\field{number}{3}
\field{title}{A Survey on Spectra of Infinite Graphs}
\field{type}{Journal Article}
\field{volume}{21}
\field{year}{1989}
\field{pages}{209\bibrangedash 234}
\range{pages}{26}
\verb{doi}
\verb 10.1112/blms/21.3.209
\endverb
\verb{urlraw}
\verb https://londmathsoc.onlinelibrary.wiley.com/doi/abs/10.1112/blms/21.3.209
\endverb
\verb{url}
\verb https://londmathsoc.onlinelibrary.wiley.com/doi/abs/10.1112/blms/21.3.209
\endverb
\endentry
\entry{Oliveira-2019}{inproceedings}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=5fa97e7a2689b91efc2fddbc775c53d1}{%
family={Oliveira},
familyi={O\bibinitperiod},
given={Roberto},
giveni={R\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=be791fcf2bac698f7a6b715cdf5d5d51}{%
family={Peres},
familyi={P\bibinitperiod},
given={Yuval},
giveni={Y\bibinitperiod},
givenun=0}}%
}
\list{publisher}{1}{%
{SIAM}%
}
\strng{namehash}{fafcd98aea957793b90d4a9bdd9331e7}
\strng{fullhash}{fafcd98aea957793b90d4a9bdd9331e7}
\strng{bibnamehash}{fafcd98aea957793b90d4a9bdd9331e7}
\strng{authorbibnamehash}{fafcd98aea957793b90d4a9bdd9331e7}
\strng{authornamehash}{fafcd98aea957793b90d4a9bdd9331e7}
\strng{authorfullhash}{fafcd98aea957793b90d4a9bdd9331e7}
\field{sortinit}{O}
\field{sortinithash}{2cd7140a07aea5341f9e2771efe90aae}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{booktitle}{2019 Proceedings of the Sixteenth Workshop on Analytic Algorithmics and Combinatorics (ANALCO)}
\field{title}{Random walks on graphs: new bounds on hitting, meeting, coalescing and returning}
\field{type}{Conference Proceedings}
\field{year}{2019}
\field{pages}{119\bibrangedash 126}
\range{pages}{8}
\verb{doi}
\verb 10.1137/1.9781611975505
\endverb
\endentry
\entry{Tetali-2005}{article}{}
\name{author}{2}{}{%
{{un=0,uniquepart=base,hash=171e5dea80d8e25fa980154d64d58fa8}{%
family={Tetali},
familyi={T\bibinitperiod},
given={P.},
giveni={P\bibinitperiod},
givenun=0}}%
{{un=0,uniquepart=base,hash=4e772db9e103adc2864b791b5c4ddaa0}{%
family={Montenegro},
familyi={M\bibinitperiod},
given={R.},
giveni={R\bibinitperiod},
givenun=0}}%
}
\strng{namehash}{404a98928c58b6287568799bb7cc1f82}
\strng{fullhash}{404a98928c58b6287568799bb7cc1f82}
\strng{bibnamehash}{404a98928c58b6287568799bb7cc1f82}
\strng{authorbibnamehash}{404a98928c58b6287568799bb7cc1f82}
\strng{authornamehash}{404a98928c58b6287568799bb7cc1f82}
\strng{authorfullhash}{404a98928c58b6287568799bb7cc1f82}
\field{sortinit}{T}
\field{sortinithash}{9af77f0292593c26bde9a56e688eaee9}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{issn}{1551-305X 1551-3068}
\field{journaltitle}{Foundations and Trends in Theoretical Computer Science}
\field{number}{3}
\field{title}{Mathematical Aspects of Mixing Times in {M}arkov Chains}
\field{type}{Journal Article}
\field{volume}{1}
\field{year}{2005}
\field{pages}{237\bibrangedash 354}
\range{pages}{118}
\verb{doi}
\verb 10.1561/0400000003
\endverb
\endentry
\entry{Luxburg-2007}{article}{}
\name{author}{1}{}{%
{{un=0,uniquepart=base,hash=86714cf89fed80215d15171e9ed5f412}{%
family={Luxburg},
familyi={L\bibinitperiod},
given={Ulrike},
giveni={U\bibinitperiod},
givenun=0,
prefix={von},
prefixi={v\bibinitperiod}}}%
}
\strng{namehash}{86714cf89fed80215d15171e9ed5f412}
\strng{fullhash}{86714cf89fed80215d15171e9ed5f412}
\strng{bibnamehash}{86714cf89fed80215d15171e9ed5f412}
\strng{authorbibnamehash}{86714cf89fed80215d15171e9ed5f412}
\strng{authornamehash}{86714cf89fed80215d15171e9ed5f412}
\strng{authorfullhash}{86714cf89fed80215d15171e9ed5f412}
\field{sortinit}{v}
\field{sortinithash}{afb52128e5b4dc4b843768c0113d673b}
\field{extradatescope}{labelyear}
\field{labeldatesource}{year}
\field{labelnamesource}{author}
\field{labeltitlesource}{title}
\field{abstract}{In recent years, spectral clustering has become one of the most popular modern clustering algorithms. It is simple to implement, can be solved efficiently by standard linear algebra software, and very often outperforms traditional clustering algorithms such as the k-means algorithm. On the first glance spectral clustering appears slightly mysterious, and it is not obvious to see why it works at all and what it really does. The goal of this tutorial is to give some intuition on those questions. We describe different graph Laplacians and their basic properties, present the most common spectral clustering algorithms, and derive those algorithms from scratch by several different approaches. Advantages and disadvantages of the different spectral clustering algorithms are discussed.}
\field{issn}{1573-1375}
\field{journaltitle}{Statistics and Computing}
\field{number}{4}
\field{title}{A Tutorial On Spectral Clustering}
\field{type}{Journal Article}
\field{volume}{17}
\field{year}{2007}
\field{pages}{395\bibrangedash 416}
\range{pages}{22}
\verb{doi}
\verb 10.1007/s11222-007-9033-z
\endverb
\verb{urlraw}
\verb https://doi.org/10.1007/s11222-007-9033-z https://link.springer.com/article/10.1007/s11222-007-9033-z https://link.springer.com/article/10.1007%2Fs11222-007-9033-z
\endverb
\verb{url}
\verb https://doi.org/10.1007/s11222-007-9033-z%20https://link.springer.com/article/10.1007/s11222-007-9033-z%20https://link.springer.com/article/10.1007%2Fs11222-007-9033-z
\endverb
\endentry
\enddatalist
\endrefsection
\endinput