Stratified MCMC Sampling for Non-Reversible Dynamics

Stratified MCMC Sampling for Non-Reversible Dynamics

Graduate/faculty Seminar

Gabe Earle (Duke University)

Monday, November 18, 2019 -
12:00pm to 1:00pm
Location: 
119 Physics

Stratified MCMC sampling is an effective tool for analyzing complex probability distributions, particularly when studying tails, distributions with many wells or distributions on manifolds. However, while it is not entirely constrained to reversible systems, integration of stratified sampling with highly irreversible processes is still limited, especially considering the advantages that irreversibility often has regarding sampling speed. Here I will outline a newly proposed stratified algorithm specifically geared towards irreversible Markov chains, and sketch my approach to proving convergence and analyzing performance. I will also outline potential applications, and a possible connection between the analysis of such stratified methods and the theory of Poisson equations.

Last updated: 2019/11/19 - 2:42pm