A Class Coupler for Perfect Sampling from Continuous Distributions With and Without Atoms

Abstract

We consider the simulation of distributions that are a mixture of discrete and continuous components. We extend a Metropolis-Hastings-based perfect sampling algorithm of Corcoran and Tweedie to allow for a broader class of transition candidate densities. The resulting algorithm, know as a "class coupler", is fast to implement and is applicable to purely discrete or purely continuous densities as well. Our work is motivated by the study of a composite hypothesis test in a Bayesian setting via posterior simulation and we give simulation results for some problems in this area.