Cycle Basis Markov chains for the Ising Model

Combinatorics Seminar
Wednesday, May 22, 2013 - 15:05
1 hour (actually 50 minutes)
Skiles 005
National Institute of Standards and Technology
Studying the ferromagnetic Ising model with zero applied field reduces to sampling even subgraphs X of G with probability proportional to \lambda^{|E(X)|}. In this paper we present a class of Markov chains for sampling even subgraphs, which contains the classical single-site dynamics M_G and generalizes it to nonlocal chains. The idea is based on the fact that even subgraphs form a vector space over F_2 generated by a cycle basis of G. Given any cycle basis C of a graph G, we define a Markov chain M(C) whose transitions are defined by symmetric difference with an element of C. We characterize cycle bases into two types: long and short. We show that for any long cycle basis C of any graph G, M(C) requires exponential time to mix when \lambda is small. All fundamental cycle bases of the grid in 2 and 3 dimensions are of this type. Moreover, on the 2-dimensional grid, short bases appear to behave like M_G. In particular, if G has periodic boundary conditions, all short bases yield Markov chains that require exponential time to mix for small enough \lambda. This is joint work with Isabel Beichl, Noah Streib, and Francis Sullivan.