A Markov Jump Process for More Efficient Hamiltonian Monte Carlo

Hamiltonian Monte Carlo with discrete time sampling steps is limited by the factthat its transition rates must be less than or equal to 1. By deriving an HMC algorithm that is a continuous time Markov jump process, we are able to overcome this constraint. This allows us to propose new transition rates that lead to improvedmixing, as measured both by both the spectral gap and auto-correlations on several example problems. We release the algorithm as an open source python package.