\documentclass{amsart} \usepackage{amsmath,amsthm,amsfonts,amssymb,graphics} \begin{document} % % The following macro is used to generate the header. % \setlength{\oddsidemargin}{0.25 in} \setlength{\evensidemargin}{-0.25 in} \setlength{\topmargin}{-0.6 in} \setlength{\textwidth}{6.5 in} \setlength{\textheight}{8.5 in} \setlength{\headsep}{0.75 in} \setlength{\parindent}{0 in} \setlength{\parskip}{0.1 in} \newcounter{lecnum} % \newcommand{\lecture}[4]{ \pagestyle{myheadings} \thispagestyle{plain} \newpage \setcounter{lecnum}{#1} \setcounter{page}{1} \noindent \begin{center} \framebox{ \vbox{\vspace{2mm} \hbox to 6.28in { {{\bf 18.317~Combinatorics, Probability, and Computations on Groups} \hfill #2} } \vspace{4mm} \hbox to 6.28in { {\Large \hfill Lecture #1 \hfill} } \vspace{2mm} \hbox to 6.28in { {\it Lecturer: #3 \hfill Scribe: #4} } \vspace{2mm}}} \end{center} \markboth{Lecture #1: #2}{Lecture #1: #2} \vspace*{4mm} } % \theoremstyle {plain} \newtheorem{theorem} {Theorem} [section] \newtheorem {proposition} [theorem] {Proposition} \newtheorem {lemma} [theorem] {Lemma} \newtheorem {corollary} [theorem] {Corollary} \theoremstyle {definition} \newtheorem {definition} [theorem] {Definition} % \def\indent{\hspace*{3em}} \def\newp{\newline\indent} \def\above#1#2{ \begin{array}{c} #1 \\ #2 \end{array}} % \setlength{\parindent}{0em} % % \lecture{11}{3 October 2001}{Igor Pak}{N. Ackerman} % \section{Random Walks on Groups} \label{Random Walks on Groups} % Let $G$ be a finite Group and let $S$ be a set of generators of $G$. We say that $S$ is symmetric if $S = S^{-1}$. In other words $S$ is symmetric if $\forall s\in S$, $s^{-1}\in S$\newline% % \begin{definition} We define $\Gamma = \Gamma(G, S)$ as the \underline{Cayley graph} of $G$ with respect to $S$. This is the graph which has a vertices for each $g\in G$ and edges between each $(g,h)$ such that $g^{-1}h \in S$. \end{definition} % \indent Observe that if $S$ is symmetric then the Cayley graph of $G$ with respect to $S$, $\Gamma(G,S)$ is unoriented. Also observe that if $S$ contains the identity $(id \in S)$ then the Cayley graph of $G$ with respect to $S$, $\Gamma(G,S)$ has loops. \newline% % \begin{definition} On a walk of a Cayley graph we define $x_t$ as the place you reach after $t$ steps. \newline\newline% % We define a \underline{random walk} as just a random walk on the Cayley graph starting at the identity. In other words at each step you choose (randomly and uniformly) which direction to go. ($x_{t+1} = x_t \cdot s$, $x_0 = $ id, $s\in S$ (uniform in $S$))\newline \newline% % We similarly define a \underline{lazy random walk} as a random walk, except before each step you choose first whether to move or stay where you are. Then, if you have decided to move, you decide independently where to move ($x_{t+1} = x_t \cdot s^e$, $s\in S, e\in \{0,1\}$).\newline \newline% % We define $Q^t(g)=Pr(x_t=g)$ as the probability that after $t$ steps on the walk you will be at vertex $g$. \end{definition} % \begin{proposition} If the Cayley graph is not bipartite then \, $Q^t(g) \to 1/|G|$, as as \, $t \to \infty$. \end{proposition} For example, if $S$ contains the identity $(id\in S)$ then this proposition is true for the lazy random walk. % \underline{Example:} If $G=\mathbb{Z}_m$ and $S=\{\pm 1\}$ then the Cayley graph $\Gamma(G,S)$ is bipartite if and only if $m$ is even. \medskip % \underline{Example:} If $G=S_m$ and $S=\{(i,j)| 1\leq i < j \leq n\}$ \, then the Cayley graph $\Gamma(G,S)$ is bipartite. \newline\newline% % \begin{definition} If $P$ and $Q$ are probability distributions on $G$ then we define the \underline{convolution} of $P$ and $Q$ as % % $$P \ast Q (g) = \sum_{h\in G} P(h)Q(h^{-1}g)$$ \end{definition} % Observe that if $P$ is the probability distribution % $$P(g) =\left\{ \above{1/|S|, g\in S}{0 \text{ otherwise }}\right.$$% then $Q^t = \underbrace{P \ast P \ast \cdots \ast P}$ \newline \indent\indent $t$ times \newline % % \begin{definition} We then define the separation distance after $t$ steps as % $$\text{\rm {sep}}(t) = |G|\cdot \max_{g\in G} (1/|G| - Q^{t}(g))$$ \end{definition} % \begin{proposition} \ \newline a) \text{\rm {sep}}(t+1) $\leq$ \text{\rm {sep}}(t) \newline % b) \text{\rm {sep}}(t+l) $\leq$ \text{\rm {sep}}(t) $\cdot$ \text{\rm {sep}}(l) \newline% c) \text{\rm {sep}}(t)$\sim c\rho^t$ as $t\rightarrow \infty$, $0\leq \rho \leq 1$, \newline \newline % % where \, $f(x) \sim g(x)$ means that $f(x) = g(x \cdot (1+o(1)))$ \end{proposition} % \begin{proof} a) Observe that \text{\rm {sep}}(t) $< \epsilon$ is equivalent to saying that $Q^t = (1-\epsilon)U + \epsilon N$ where $U$ is the uniform distribution, and $N$ is some other distribution. Therefore we know that because $Q^{t+1} = Q^t \ast P$, % $$Q^{t+1} = ((1-\epsilon)U + \epsilon N) \ast P = (1-\epsilon) U \ast P + \epsilon N \ast P$$% But we know $U\ast P$ is still the uniform distribution, and $\min_{g\in G} N \ast P \geq \min_{g\in G} N$ by the construction of $P$. So $\min_{g\in G} Q^{t+1}(g) \geq \min_{g\in G} Q^{t}(g)$. And so finally \text{\rm {sep}}(t+1) $\leq$ \text{\rm {sep}}(t). \newline \newline % b) Let $Q^t = (1-\epsilon) U + \epsilon N_1$ and $Q^l = (1-\delta) U + \delta N_2$. We know that this is equivalent to \text{\rm {sep}}(t) $< \epsilon$ and \text{\rm {sep}}(l) $< \delta$\newline\newline% % We then have % $$Q^{t+l} = ((1-\epsilon) U + \epsilon N_1) \ast ((1-\delta) U + \delta N_2)$$ % and, after we condense terms it is easy to see that % $$Q^{t+l} = (1-\delta\epsilon)U + \delta\epsilon N_1 \ast N_2$$ % % And so we have that \text{\rm {sep}}(t+l) $<$ \text{\rm {sep}}(t) $\cdot$ \text{\rm {sep}}(l) (because \text{\rm {sep}}(t+1)$<$ $\delta\epsilon$ and $\delta$ and $\epsilon$ are arbitrary). \newline\newline% % c) Let $A = (a_{gh})_{g,h\in G}$ be a matrix such that $a_{gh} = P(hg^{-1})$. We then let $A^t = A * \cdots * A$. Then observe that % $$Q^{t} = \left[ A^t \cdot \left( \begin{array}{c} 1 \\ 0 \\ 0 \\ \vdots \\ 0 \end{array} \right)\right]_g $$% We then have% $$A^t \cdot % \left(\begin{array}{c} 1 \\ 0 \\ 0 \\ \vdots \\ 0 \end{array} \right)% = \left(\begin{array}{c} 1/|G| \\ 1/|G| \\ \vdots \\ 1/|G| \end{array} \right)% + \lambda_1^t(v_1) + \cdots + \lambda_n^t(v_n)$$ % % Where $v_i$ and $\lambda_i$ are eigenvectors and eigenvalues. And so % $$Q^{t}(g) = \lambda_1^t(w_1) + \cdots + \lambda_n^t(w_n)$$ % Now, if we let $q_t = \min_{g\in G} Q^{t}(g)$ then $q_t = 1/|G| + w_1 \lambda^t_1 + \cdots$ and % $$1/|G| - q_t = w_1\lambda^t_1 + \sigma C$$% % Also observe that if we do this same thing for the lazy random walk we have ${\lambda'}_i = 1/2(1+\lambda_i)$. \newline\newline % \end{proof} % \begin{definition} We define the \underline{relaxation time} as: \ $\tau_1 = 1/(1-\lambda_1)$ \newline% We then define the \underline{mixing time} as minimum time for the separation time to be less than one half: \ $\tau_2 = \min \{t:$ \text{\rm {sep}}(t)$ \leq 1/2\}$. \newline% Finally, we define the \underline{optimal time} as: \ $\tau_3 = \sum^{\infty}_{t = 0}$ \text{\rm sep}(t)% \end{definition} % \begin{proposition} $1/2 \tau_3 < \tau_2 < 2\tau_3$ \end{proposition} \begin{proof} Now, we can see from the definitions that \newline% $\tau_3 \geq \underbrace{1/2 + 1/2 + \cdots + 1/2} \geq 1/2 \tau_2$ \newline% \indent \indent $\tau_2$ \newline% Now we also know from the definitions that \newline% $\tau_3 \leq \underbrace{1+1 + \cdots + 1} + \underbrace{1/2+1/2 + \cdots + 1/2} + \underbrace{1/4+1/4 + \cdots + 1/4} + \cdots$ \newline% \indent\indent $\tau_2$ \indent \indent $\tau_2$ \indent\indent\indent $\tau_2$ \newline% $= \tau_2 (1 + 1/2 + 1/4 + \cdots) = 2\tau_2$ \newline % And this completes the proof. \end{proof} % \begin{proposition} $\tau_2< \tau_1 \log(|G|)$ \end{proposition} \begin{proof} First, note that % $$|1/|G| - Q^t(g)| < \sum w_i \lambda_i^t < |G| \lambda_1^t$$% (Note that $\lambda = \rho$ if $s=s^{-1}$). \newline \newline% Likewise, from the definitions we see that: \newline% $|G|\cdot |1/|G| - Q^t(g)| < |G|^2 (1-1/\tau_1)^t \leq 1/e < 1/2 \cdot q_t, t=2\log(|G|)$ \newline% \end{proof} % \underline{Example} If $G = \mathbb{Z}_m$ and $S = \{\pm 1\}$ then $$A =\left[% \begin{array}{ccccc} 0 & 1/2 & 0 & 0 & 1/2 \\ % 1/2 & 0 & \ddots & 0 & 0 \\ % 0 & \ddots & \ddots & \ddots & 0 \\ % 0 & 0 & \ddots & 0 & 1/2 \\ % 1/2 & 0 & 0 & 1/2 & 0 \\ % \end{array}\right] $$ One can show that \ $\lambda_j = \cos(2\pi j/m)$, so $$\lambda_1 = 1 - \frac{(2 \pi)^2}{m^2} + o\left(\frac{1}{m^4}\right) = % 1- \frac{c}{m^2} + o\left(\frac{1}{m^4}\right).$$ Now Proposition 1.9 implies that \ $\text{\bf {mix}} = o(m^2 \log m)$. In the future we will show that \ $\text{\bf {mix}} = \theta(m^2)$. % % \end{document}