On the efficiency of pseudo-marginal random walk Metropolis algorithms (Q2338926)

The Markov chain Monte Carlo (MCMC) algorithms have been proved to be particullary successful in statistics for an investigation of the posterior distributions in Bayesian analysis of complex models. Almost all MCMC methods are based on the Metropolis-Hastings (MH) algorithm which owes much of its success to its tremendous flexibility. The pseudo-marginal Metropolis-Hastings algorithm provides a general recipe for circumventing the need for target density evaluation. In the present paper, the behaviour of the pseudo-marginal random walk Metropolis (PsMRWM) algorithm is examined. Under relatively general conditions on the target distribution, limiting formulae for the acceptance rate and for the expected squared jump distance, as the dimension of the target approaches infinitely, are obtained. The overall efficiency of the algorithm, in the terms of both speed of mixing and computational time is considered. Under the assumption that the additive noise is Gaussian and inversely proportional to the number of unbiased estimates, it is proved that the algorithm is optimally efficient when the variance of the noise is approximately \(3.283\) and the acceptance rate approximately \(7.001\). In the introduction, the concepts of the pseudo-marginal Metropolis-Hastings algorithm and the efficiency of the random walk Metropolis algorithm are briefly reminded. Here \(\chi \subset {\mathbb R}^d\) is a state space, \(\pi(\cdot)\) is the distribution on \(\chi\). The concept that the MH provides algorithms for obtaining an approximation of the target distribution \(\pi(\cdot)\) by constructing a Markov chain is developed. In Section 2, the pseudo-marginal walk Metropolis in high distributions is considered. In Subsection 2.1, the isotropic Gausssian proposal \({\mathbf X}^* = {\mathbf x} + \lambda {\mathbf Z}\), where \({\mathbf Z} \sim N(0, I)\), \( \lambda > 0\), is the scaling parameter and \(I\) is the \(d \times d\) identity matrix is considered. In Subsection 2.2, the concept of the noise in the estimate of the log-target is developed. In Subsection 2.3, the concept of the high-dimensional target distribution is given. In Subsection 2.4, the concept of the expected squared jump distance is considered. In Theorem 1, under the assumption of additive independent noise in the log-target, expressions for limiting expected squared jump distance are obtained and the asymptotic acceptance rate is shown. The obtained results are graphically illustrated. In Subsection 2.5, it is proved that the PsMRWM can be well approximated by an appropriate diffusion limit. In Theorem 2 in the case of a target with independent and identically distributed components, the diffusion limit is proved. The efficiency of the algorithm is given by the speed of this limiting diffusion. In Section 3, the question of the optimizing the PsMRWM is considered. In Subsection 3.1, the case in which the additive noise follows a Gaussian distribution, i.e., the standard asymptotic regime (SAR), is considered. In Subsection 3.2, an optimization under the SAR is realized. In Section 4, the SAR of Subsection 3.1 is investigated. Here the partical marginal RWM algorithm is used to perform an exact inference for the Lodka-Volterra predator-prey model. In Section 5, the main result -- Theorems 1 and 2 -- are proved. In Section 6, some conclusions about the behaviour of the PsMRWM algorithm, the behaviour of the Markov chain on the target, the noise, obtaining a limiting diffusion for the first component of a target are realized. The paper ends with two appendixes, where some preliminary results are proved.