Abstract:
Generalized hybrid/Hamiltonian Monte Carlo (GHMC) methods differ from hybrid/Hamiltonian Monte Carlo (HMC) techniques in the momentum update step, where a partial refreshment of momentum replaces a complete momentum reset. In order to satisfy detailed balance condition and ensure a stationary distribution, a momentum flip is required upon rejection of a Hamiltonian dynamics proposal step. These momentum swings induce reverse trajectories and might, in principle, slow down mixing and decorrelation of the chain. In this talk we analyze the effect of momentum flips on efficiency and accuracy of several versions of the GHMC applied to molecular and statistical simulations and discuss possible ways for reducing potential negative effects of momentum flips.