Unbiased Markov chain Monte Carlo for intractable target distributions (Q2192323)

In the first chapter of the article, starting from a coupling of two Markov chains, \((Z_n) _{n\geq 0}\) and \(({\tilde{Z}}_n)_{n\geq 0}\), for any integer \(k\) and \(m\geq k\), and expectation \(\pi(h)\), the estimator \(H_{k:m}(Z,\tilde{Z})\) is defined. Algorithm 1 summarizes the construction of the unbiased MCMC estimator \(H_{k:m}\) for any choice of \(k\) and \(m\), \(0\geq k\geq \mu\). Algorithm 2 summarizes the sampling from the coupled MH kernel given \((Z_n,{\tilde{Z}}_{n-1})\). Under some special assumptions, one shows that the defined estimator has indeed the expectation \(\pi(h)\), has a finite expected computing time and admits a finite variance. Under additional assumptions one proves that the tail probabilities of the meeting time are polynomially bounded. The efficiency of the new estimator \(H_{k:m}\) is compared to standard MCMC estimators. The second chapter of the paper is devoted to the unbiased pseudo-marginal MCMC. The pseudo-marginal Metropolis-Hastings is studied. Algorithm 3 summarizes the steps to sample from the coupled pseudo-marginal MH kernel. Under new additional assumptions, the authors prove that the coupled pseudo-marginal algorithm returns unbiased estimators with finite variance and finite expected computer time. Results on numerical experiments are largely presented in the third and fourth chapters. Different types of models as the Beta-Bernoulli model, the linear Gaussian state space model as well as a neuroscience experiment are considered. In the fifth chapter, coupled MCMC algorithms to perform inference when the likelihood function is intractable are developed and practical applications as for instance to the inference in a planar lattice Ising model without an external field are shown. Appendix A is dedicated to all proofs. An interesting point: Section A.6 shortly refers to the ``Brexit Referendum Data''.