How to View an MCMC Simulation as a Permutation, with Applications to Parallel Simulation and Improved Importance Sampling

Radford M. Neal, Dept. of Statistics and Dept. of Computer Science, University of Toronto

Consider a Markov chain defined on a finite state space, X, that leaves invariant the uniform distribution on X, and whose transition probabilities are integer multiples of 1/Q, for some integer Q. I show how a simulation of n transitions of this chain starting at x_0 can be viewed as applying a random permutation on the space XxU, where U={0,1,...,Q-1}, to the start state (x_0,u_0), with u_0 drawn uniformly from U. This result can be applied to a non-uniform distribution with probabilities that are integer multiples of 1/P, for some integer P, by representing it as the marginal distribution for X from the uniform distribution on a suitably-defined subset of XxY, where Y={0,1,...,P-1}. By letting Q, P, and the cardinality of X go to infinity, this result can be generalized to non-rational probabilities and to continuous state spaces, with permutations on a finite space replaced by volume-preserving one-to-one maps from a continuous space to itself. These constructions can be efficiently implemented for chains commonly used in Markov chain Monte Carlo (MCMC) simulations. I present two applications in this context - simulation of K realizations of a chain from K initial states, but with transitions defined by a single stream of random numbers, as may be efficient with a vector processor or multiple processors, and use of MCMC to improve an importance sampling distribution that already has substantial overlap with the distribution of interest. I also discuss the implications of this ``permutation MCMC'' method regarding the role of randomness in MCMC simulation, and the potential use of non-random and quasi-random numbers.

Technical Report No. 1201, Dept. of Statistics, University of Toronto (April 2012), 42 pages: postscript, pdf.

Also available at arxiv.org.

You can also get the programs used for the tests in this paper.