Spectral gaps for a Metropolis-Hastings algorithm in infinite dimensions (Q473164): Difference between revisions

It is well known that the most widely used method for general target measures are Markov chain Monte Carlo (MCMC) algorithms. The main aim of this article is to study the complexity of certain sampling algorithms in high dimensions, more precisely, the Metropolis-Hastings algorithm in infinite dimensions. The computational cost of such an algorithm is the number of necessary steps \(\times\) cost of a step. While for most algorithms the cost of one step grows with the dimension, a major result of this paper is an lgorithm which, when applied to measures, defines via a finite-dimensional approximation of a measure defined by a density with respect to a Gaussian random field, requires a number of steps independent of the dimension in order to achieve a given level of accuracy. The authors consider Metropolis-Hastings MCMC methods and focus on cases when the proposal is either a Gaussian random walk (RWM) for which the covariance equals to that of a reference measure or an Orstain-Uhlenbeck proposal (pCN) for which the reference measure is invariant. Among the main results belong: {\parindent=6mm \begin{itemize}\item[{\(\bullet\)}] Description of \(RWM\) and \(pCN\) algorithms and study of their properties, both positive and negative ones. \item[{\(\bullet\)}] Statement of the weak Harris theorem and a discussion of the relationship between exponential convergence in a Wasserstein distance and \(L^{2}_{\mu}\)-spectral gaps. \item[{\(\bullet\)}] Formal confirmation of a conjecture that \(pCN\) has a convergence rate that is independent of the dimension while \(RWM\) has undesirable dimension dependent properties. \end{itemize}} The results are presented for separable Hilbert spaces, but in fact they do also hold on an arbitrary Banach space.