Optimal scaling for partially updating MCMC algorithms (Q997939): Difference between revisions

Asymptotic behavior of the random walk Metropolis-within-Gibbs (RWM) and Metropolis adjusted Langevin-within-Gibbs (MALA) algorithms of MCMC sampling are considered in the case when the dimensionality of the sampling space \(d\to\infty\) and the dimensionality of the proposed jump at each step is \(cd\) for some fixed \(0\leq c\leq 1 \). The limit average acceptance rate is evaluated for these algorithms in the case when the target distribution consists of i.i.d. entries and in the case of an exchangeable normal target, e.g., in the i.i.d. case the optimal acceptance rate for RWM is 0.234. The authors conclusion is that while full dimensional Langevein updates are worthwhile, full-dimensional Metropolis ones are asymptotically no better then smaller dimensional updating schemes. Results of simulations are presented.