Towards Perfect Sampling for Bayesian Mixture Priors


by Duncan Murdoch (University of Western Ontario, Canada) and Xiao-Li Meng (University of Chicago, USA)


Perfect sampling using the coupling from the past (CFTP) algorithm was introduced by Propp and Wilson in 1996. In much the way rejection sampling allows one to convert samplers from one distribution into samplers from another, CFTP allows one to convert Markov chain Monte Carlo algorithms from approximate samplers of the steady-state distribution into perfect ones. Since 1996 CFTP has been applied to many different Markov chains. However, its use in routine Bayesian computation is still in the early stages of development. This paper provides a couple of building blocks for its potentially routine application in Bayesian mixture priors, including a t mixture coupler, and demonstrates the types of difficulties that currently prevent CFTP from being applied routinely in Bayesian computation.

Keywords: Coupling from the past, Exact sampling, MCMC, Mixtures


Back to Duncan Murdoch's home page.