APPROXIMATING POSTERIOR DISTRIBUTIONS BY MIXTURES

Kernel density estimation techniques are used to smooth simulated samples from importance sampling function approximations to posterior distributions, resulting in revised approximations that are mixtures of standard parametric forms, usually multivariate normal or T-distributions. Adaptive refinement of such mixture approximations involves repeating this process to home-in successively on the posterior. In fairly low dimensional problems, this provides a general and automatic method of approximating posteriors by mixtures, so that marginal densities and other summaries may be easily computed. This is discussed and illustrated, with comment on variations and extensions suited to sequential Bayesian updating of Monte Carlo approximations, an area in which existing and alternative numerical methods are difficult to apply.