Optimal scaling of random walk Metropolis algorithms with non-Gaussian proposals (Q655925)

The authors study the optimal scaling of heavy-tailed proposal distribution for the Random Walk Metropolis algorithm. So far this problem was considered mainly for Gaussian or symmetric, finite second order proposal distributions. The main focus in the present paper is on Cauchy proposal distribution and the optimisation is with respect to the maximum of the expected squared jumping distance (ESJD), i.e., the second moment of the first jump when the initial distribution is the target distribution. The ESJD is analysed for continuous and discontinuous target densities and compared to Gaussian proposal distributions. The authors find that depending on the target density choosing Cauchy proposal distributions might or might not be an improvement to a Gaussian proposal. In more detail, on the one hand, for continuous target densities it is conjectured that fixed radius spherical, light-tailed distributions, such as a Gaussian, generally outperform heavy-tailed distributions. This claim is support by examples and simulation results. On the other hand, for hypercube (discontinuous) target densities analytic results show that Cauchy proposal distributions always outperform light-tailed proposals.