Geometric ergodicity of a more efficient conditional Metropolis-Hastings algorithm

Abstract

Despite its extensive application in practice, the Metropolis-Hastings sampler can suffer from slow mixing and, in turn, statistical inefficiency. We introduce a modification to the Metropolis-Hastings algorithm that, under specified conditions, encourages more efficient movement on general state spaces while preserving the overall quality of convergence, geometric ergodicity in particular. We illustrate the modified algorithm and its properties for the Metropolis-Hastings algorithm for a toy univariate Normal model and for the Gibbs sampling algorithm in a toy bivariate Normal model and a Bayesian dynamic spatiotemporal model.