A unifying framework for constructing MCMC algorithms from irreversible diffusion processes

Seminar | April 11 | 3:10-4 p.m. | 1011 Evans Hall

 Yian Ma, U. C. Berkeley

 Department of Statistics

In this talk, I will first present a general recipe for constructing MCMC algorithms from diffusion processes with the desired stationary distributions. The recipe translates the task of finding valid continuous Markov processes into one of choosing two matrices. Importantly, any diffusion process with the target stationary distribution (given an integrability condition) can be represented in our framework. To simulate the irreversible diffusion processes and correct for bias from discretization error, I will turn to MCMC techniques based on jump processes. Generalization of the Metropolis Hastings algorithm will be introduced---with the same ease of implementation---but allowing for the benefits of irreversible dynamics.