In the latest issue of Statistics and Computing (2013, Issue 23, pages 577-587), Iliopoulos and Malefaki published a paper that relates to our vanilla Rao-Blackwellisation paper with Randal Douc. The idea is to derive another approximation to the ideal importance sampling weight using the “accepted” Markov chain. (With Randal, we had a Bernoulli factory representation.) The density g(x) of the accepted chain being unknown; it is represented as the expectation under π of the function
and hence estimated by a self-normalised average based on the whole Markov chain. This means the resulting importance estimate uses twice the output of the algorithm and that it is biased and of order O(n²), thus the same order as our original Rao-Blackwellised estimator (Robert & Casella, 1996)… This also means convergence and CLT are very hard to establish: the main convergence theorem of the paper holds only for finite state spaces, with a surprising smaller asymptotic variance for this self-normalised average than for the ideal importance sampling estimator in the independent Metropolis-Hastings case. (Both are biased by being self-normalised and the paper does not consider the magnitude of those biases.)
Interestingly, the authors also ran a comparison with our parallelised Rao-Blackwellised version (with Pierre Jacob and Murray Smith), but conclude (P.58) at a larger CPU (should be GPU!!) required by the parallelisation, which does not really make sense: when compared with the plain Metropolis-Hastings implementation, run on a single processor, the parallel version only requires an extra random permutation per thread or per processor. I thus suspect a faulty implementation that induces this CPU being linear in the size of the blocks, like maybe only saving one output per block… Also interestingly, the paper re-analyses the Pima Indian probit model Jean-Michel Marin and I (and many others) used as benchmark in several of our papers. As in the most standard examples, the outcome shows a mild reduction in variance when using this estimated importance sampling version. Maybe a comparison with the ideal importance sampler (i.e. the one that does not divide by the sum of the weights since using normalised versions of the target and importance densities) would have helped in the comparison.