Distributed Sampling Almost-Uniform Graph Coloring with Fewer Colors

Weiming Feng,Thomas P. Hayes,Yitong Yin

2020-01-01

Abstract:Sampling almost-uniform graph colorings is one of the most important problems in the theory of Markov chain Monte Carlo (MCMC) methods. A famous result, discovered independently by Jerrum [27] and Salas and Sokal [31], is that, assuming the Dobrushin’s condition q > (2+ δ)∆, the classic Metropolis chain for proper q-colorings on graphs with n vertices and maximum degree ∆ has mixing time O(n log n). In this paper, we study a distributed variant of the Metropolis chain, called the Lazy Local Metropolis algorithm. The chain achieves an optimal O(log n) mixing time assuming Dobrushin’s condition q > (2 + δ)∆ (discovered independently by Fischer and Ghaffari [17]). We further show that on graphs without small cycles (in particular, girth at least 9), the algorithm achieves the optimal O(log n) mixing time with fewer colors, assuming q ≥ (α∗ + δ)∆ + O(1) where α∗ ≈ 1.763 satisfies α∗ = e1/α . This mixing result with fewer colors is obtained by establishing a so-called local uniformity property for the distributed Markov chain, which asserts that after running for a constant time, the chain already behaves locally like in the stationary. This is the first time the local uniformity property is established on a distributed Markov chain.