Limit theorems for some adaptive MCMC algorithms with subgeometric kernels (Q605038)

The authors investigate the asymptotic behaviour of adaptive Markov chain Monte Carlo algorithms with Markov kernels that are subgeometrically ergodic. Thus, the class of adaptive MCMC algorithms that are accessible for rigorous analysis is enlarged. The contribution of the paper is twofold. On the one hand, it extends previous results to a larger class of kernels, and, on the other hand, this is achieved employing a more sophisticated technical machinery which allows to weaken certain assumptions. Firstly, the authors present conditions implying the ergodicity of the algorithms (convergence of the marginals to the invariant distribution) irrespective of the initial distribution; secondly, introducing stronger assumptions, a strong law of large numbers is proved. The authors believe that their conditions implying ergodicity are close to optimal, assuming only the ``diminishing adaptation condition'' (shown to be necessary) and some ``fairly weak additional assumptions''. On a technical level, there are two contributions originating from the proofs of the two asymptotic results. The coupling technique used to study ergodicity is more careful and the strong law of large numbers is established using a resolvent kernel approach together with martingale theory. A discussion comparing their conditions and results to previous work in relation to the technical machinery is presented. Finally, the authors illustrate the application of their limit theorems first on a toy example and secondly on a adaptive random walk Metropolis algorithm for distributions with subexponential tails.