Posterior-Guided Importance Sampling for Calculating Marginal Likelihoods with Application to Bayesian Exoplanet Searches

James L. Crooks, James O. Berger, Thomas J. Loredo

Duke University, Cornell University

December 2007

This report presents a novel method of high dimensional stochastic integration for calculating marginal likelihoods when a posterior MCMC sample is available. The accuracy and efficiency of importance sampling depends strongly on the importance function having a shape similar to that of the target function. In high dimensions such importance functions can be extremely hard to build. However, in the Bayesian context, the target is proportional to the posterior, so a posterior MCMC sample can be used to guide the choice of importance function. We do this by centering multivariate student distributions around some of the MCMC points and then adapting their parameters to minimize the variance in the importance weights. Our work is done in the context of exoplanet detection, first in determining how many planets are present, and second in scheduling telescope time.

Keywords: Importance Sampling, Bayes Factor, Marginal Likelihood


The manuscript is available in PostScript and PDF formats.