Convergence properties of pseudo-marginal Markov chain Monte Carlo algorithms (Q2341639)

The convergence properties of the pseudo-marginal Markov chain Monte Carlo algorithms are studied. It is proved that the asymptotic variance of the pseudo-marginal algorithm is always at least as large as that of the marginal algorithm. In the case when the marginal chain admits a right spectral gap and the weights are uniformly bounded it is proved that the pseudo-marginal chain has a spectral gap. Also, the unbounded weight distributions and the recover polynomial convergence rates in more specific cases, when the marginal algorithm is uniformly ergodic, or an independent Metropolis-Hastings algorithm, or a random walk Metropolis targeting, a super-exponential density with regular contours are considered. It is shown that the geometric and the polynomial convergence rates imply the central limit theorems. Under general conditions, it is proved that the asymptotic variance of the pseudo-marginal algorithm converges to the asymptotic variance of the marginal algorithm. In the introduction, the problem to calculate the probability distribution \(\pi\), defined on some measurable space \((X, {\mathcal B}(X))\), is presented. The Markov kernel related to a Metropolis-Hastings algorithm is given. The concepts of the marginal algorithm and the pseudo-marginal algorithm to solve the above problem are reminded. The rates of convergence are discussed. The asymptotic variance of the marginal algorithms is presented. In Section 2, the ordering of the marginal and the pseudo-marginal algorithms is considered. It is shown that the pseudo-marginal and the marginal algorithms are ordered both in the terms of the mean acceptance probability and the asymptotic variance. In Section 3, it is shown that if the marginal chain admits a nonzero right spectral gap and the weights are uniformly bounded, then the pseudo-marginal chain has also a nonzero gap. The proof of this fact is based on an explicit lower bound of the spectral gap. It is shown that the geometric ergodicity of a marginal algorithm is inherited by the pseudo-marginal chain as soon as the weights are uniformly bounded, either through a slight modification or directly in many cases by observing that the pseudo-marginal Markov operator is positive. As a consequence, a result of \textit{C. Andrieu} and \textit{G. O. Roberts} [Ann. Stat. 37, No. 2, 697--725 (2009; Zbl 1185.60083)] is obtained in a more explicit form. In Section 4, the case when the weights are uniformly bounded, is considered. An upper bound of the asymptotic variance of the pseudo-marginal algorithm is presented. This result is generalised to the situation when the weights are unbounded. It is shown how the sub-geometric ergodicity results are essential to establish the main results of this section. In Section 5, the particular case when the pseudo-marginal algorithm is an independent Metropolis-Hastings (IMH) algorithm is considered. It is shown that the pseudo-marginal implementation of this algorithm is also an IMH. Two conditions which ensure the sub-geometric ergodicity of the pseudo-marginal IMH are presented. In Section 6, it is assumed that the marginal algorithm is uniformly ergodic and that the weight distributions are uniformly integrable. Under these conditions, the existence of a Lyapunov function, satisfying a sub-geometric drift condition toward a small step, is shown. In the case when the weight distributions possess finite power moments the polynomial ergodicity is established. In Section 7, the popular random-walk Metropolis (RWM) algorithm is considered. Under the assumption of standard tail conditions on the probability distribution which ensure the geometric ergodicity of the RWM and the existence of uniformly bounded moments, it is shown, that the corresponding pseudo-marginal algorithm is polynomially ergodic. This result is extended to the situation where the moments of the weights are assumed to exist but are not necessarily uniformly bounded. In Section 8, some additional implications of the obtained results such as the existence of central limit theorems, the possibility to compute the quantitative expressions for the asymptotic variance and the analysis of generalisations of pseudo-marginal algorithms are discussed. This paper ends with appendixes, where some preliminary results are proved.