Improved Adaptive Rejection Metropolis Sampling Algorithms

Abstract

Markov Chain Monte Carlo (MCMC) methods, such as the Metropolis-Hastings (MH) algorithm, are widely used for Bayesian inference. One of the most important issues for any MCMC method is the convergence of the Markov chain, which depends crucially on a suitable choice of the proposal density. Adaptive Rejection Metropolis Sampling (ARMS) is a well-known MH scheme that generates samples from one-dimensional target densities making use of adaptive piecewise proposals constructed using support points taken from rejected samples. In this work we pinpoint a crucial drawback in the adaptive procedure in ARMS: support points might never be added inside regions where the proposal is below the target. When this happens in many regions it leads to a poor performance of ARMS, with the proposal never converging to the target. In order to overcome this limitation we propose two improved adaptive schemes for constructing the proposal. The first one is a direct modification of the ARMS procedure that incorporates support points inside regions where the proposal is below the target, while satisfying the diminishing adaptation property, one of the required conditions to assure the convergence of the Markov chain. The second one is an adaptive independent MH algorithm with the ability to learn from all previous samples except for the current state of the chain, thus also guaranteeing the convergence to the invariant density. These two new schemes improve the adaptive strategy of ARMS, thus simplifying the complexity in the construction of the proposals. Numerical results show that the new techniques provide better performance w.r.t. the standard ARMS.