Minimising MCMC variance via diffusion limits, with an application to simulated tempering (Q2443188)

If a Markov chain \(\{X_n\}\) has a stationary distribution \(\pi\), then \(\int h(x) \pi(dx)\) can often be estimated by \(n^{-1}\sum_{i=1}^n h(X_i)\) for suitably large \(n\). The efficiency of the estimate can be measured in terms of the asymptotic variance \(\mathrm{Var}(h,P)=\lim_{n \to \infty}n^{-1}\mathrm{Var}(\sum_{i=1}^n h(X_i))\), where \(P\) is transition kernel of the Markov chain. Given two Markov chain kernels \(P_1,P_2\), both having the same invariant measure \(\pi\), we say that \(P_1\) dominates \(P_2\) if \(\mathrm{Var}(h,P_1)\leq \mathrm{Var}(h,P_2)\) for all admissible \(h\). In the article, this approach is developed for the comparison of the asymptotic variance of Langevin diffusions. Sufficient conditions are given. The results are then applied to simulated tempering algorithms in high-dimensional spaces. Their limiting descriptions are proved to be Langevin diffusions, and the most efficient Markov chain Monte Carlo algorithm is described.