Spectral Independence and Local-to-Global Techniques for Optimal Mixing of Markov Chains

Abstract

This monograph is an exposition on an exciting new technique known as spectral independence, which has been instrumental in analyzing the convergence rate of Markov Chain Monte Carlo (MCMC) algorithms. For a high-dimensional distribution defined on labelings of the vertices of an $n$-vertex graph, the spectral independence condition, introduced by Anari, Liu, and Oveis Gharan (2020), is a bound on the maximum eigenvalue of the influence matrix capturing the influence between pairs of vertices (closely related to the covariance between the variables). In the first part of the monograph, we present results showing that spectral independence (and related techniques) imply fast mixing of simple Markov chains such as the Glauber dynamics (aka Gibbs sampler). These proofs rely on local-to-global theorems relating local walks encoding pairwise correlations to variance decay and mixing properties of Markov chains. We focus on two applications: the hard-core model on independent sets of a graph (which is a combinatorial example of a binary graphical model) and random bases of a matroid. We apply the techniques presented in this monograph to show recent results of fast mixing of the Glauber dynamics on general graphs in the so-called tree-uniqueness region, polynomial-time mixing on general graphs at the critical point for the uniqueness threshold, polynomial-time mixing on random regular graphs beyond the uniqueness threshold, and fast mixing of the bases-exchange walk for generating a random basis of an arbitrary matroid. Our focus in this monograph is on the analysis of the spectral gap of the associated Markov chains from a functional analysis perspective; we present proofs of the associated local-to-global theorems and the Trickle-Down Theorem from this same Markov chain perspective. The monograph is self-contained and aims to present the proofs in a unified fashion.