A cautionary tale on the efficiency of some adaptive Monte Carlo schemes (Q988756)

Adaptive MCMC algorithms are versions of Markov chain Monte Carlo algorithms which employ at each step a different transition kernel \(P_n\). The present paper discusses a class of such algorithms where the kernels \(P_n\) are not invariant with respect to the target distribution \(\pi\) but converge for \(n\to\infty\) to a limiting kernel \(P\) invariant with respect to \(\pi\). Usually, the limiting kernel \(P\) is a very efficient kernel but difficult to implement. The author first defines a general class of adaptive MCMC algorithms which possesses as special cases importance-resampling (IC) MCMC algorithms and equi-energy (EE) samplers. However, the theoretical analysis of the asymptotics is restricted to the sub-class of EE samplers. The author proves a law of large numbers and central limit theorems with random and deterministic centering. It is shown that the asymptotic variance of the adaptive algorithm is at least as large as the asymptotic variance of the limiting kernel. In order to prove the results, the author employs the martingale approximation method. Finally, a numerical example (sampling from a centered bivariate normal distribution) illustrates that the difference of the variances can be substantial. The numerical example is implemented for the adaptive IC and the EE MCMC algorithm, MCMC algorithms based on their respective limiting kernels and the random walk Metropolis (RWM) algorithm. Mean-square errors are considered for the comparison of the algorithms. As expected, it is observed that the MCMC algorithms based on the limiting kernel are much more efficient than the RWM algorithm. Further, the adaptive IC and EE samplers are only slightly more efficient than the RWM algorithm. Taking simulation times into account the author infers that the adaptive methods are not more efficient than the plain RWM.