Improving convergence of the Hastings-Metropolis Algorithm with a learning proposal

Didier CHAUVEAU, Pierre VANDEKERKHOVE,

Université de Marne-la-Vallée,
Analyse et Mathématiques Appliquées,
5 Bd Descartes, Champs-sur-Marne, 77454 Marne-la-Vallée Cedex 2, France.

 
 

Abstract: The Hastings-Metropolis algorithm is a general MCMC method for sampling from a density known up to a constant.  Geometric convergence of this algorithm has been proved under conditions relative to the instrumental distribution (or proposal).  We present an inhomogeneous Hastings-Metropolis algorithm for which the proposal density approximates the target density, as the number of iterations increases.  The proposal at the nth step is a nonparametric estimate of the density of the algorithm, and uses an increasing number of iid copies of the Markov chain.  The resulting algorithm converges  (in  n) geometrically faster than a Hastings-Metropolis algorithm with an arbitrary proposal.  The case of a strictly positive density with compact support is presented first, then an extension to more general densities is given.  We conclude by proposing a practical way of implementation for the algorithm, and illustrate it over a simulated example.

Keywords: Hastings-Metropolis algorithm; inhomogeneous Markov chain; MCMC; nonparametric estimation; rate of convergence.

prépublication no 14/99, Université Marne-la-Vallée, 19 pages.