Improving Markov Chain Monte Carlo Estimators by Coupling to an Approximating Chain

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

We show how large improvements in the accuracy of MCMC estimates for posterior expectations can sometimes be obtained by coupling a Markov chain that samples from the posterior distribution with a chain that samples from a Gaussian approximation to the posterior. Use of this method requires a coupling scheme that produces high correlation between the two chains. An efficient estimator can then be constructed that exploits this correlation, provided an accurate value for the expectation under the Gaussian approximation can be found, which for simple functions can be done analytically. Good coupling schemes are available for many Markov chain samplers, including Gibbs sampling with standard conditional distributions. For many moderate-dimensional problems, the improvement in accuracy using this method will be much greater than the overhead from simulating a second chain.

Technical Report No. 0101, Dept. of Statistics, University of Toronto (February 2001), 13 pages: postscript, pdf.