Non-reversible guided Metropolis-Hastings kernel
We construct a non-reversible Metropolis-Hastings kernel as a multivariate extension of the guided-walk kernel proposed by Gustafson in 1998 by introducing a projection from state space to a locally compact topological group. As a by-product, we construct an efficient reversible Metropolis-Hastings kernel based on the Haar measure which is of interest in its own right. The proposed non-reversible kernel was 10-40 times better than the random-walk Metropolis kernel or the Hamiltonian Monte Carlo kernel for the Gaussian process classification example in terms of effective sample size.
READ FULL TEXT