Metropolis Adjusted Microcanonical Hamiltonian Monte Carlo

For unbiased sampling of distributions with a differentiable density, Hamiltonian Monte Carlo (HMC) and in particular the No-U-Turn Sampler (NUTS) are widely used, especially in the context of Bayesian inference. We propose an alternative sampler to NUTS, the Metropolis-Adjusted Microcanonical sampler (MAMS). The success of MAMS relies on two key innovations. The first is the use of microcanonical dynamics. This has been used in previous Bayesian sampling and molecular dynamics applications without Metropolis adjustment, leading to an asymptotically biased algorithm. Building on this work, we show how to calculate the Metropolis-Hastings ratio and prove that extensions with Langevin noise proposed in the context of HMC straightforwardly transfer to this dynamics. The second is a tuning scheme for step size and trajectory length. We demonstrate that MAMS outperforms NUTS on a variety of benchmark problems.