LAUSR

We construct a zig-zag process targeting posterior distributions arising in genetics from the Kingman coalescent and several popular models of mutation. We show that the zig-zag process can lead to efficiency gains of up to several orders of magnitude over classical Metropolis-Hastings, and argue that it is also well suited to parallel computation for coalescent models. Our construction is based on embedding discrete variables into continuous space; a technique previously exploited in the construction of Hamiltonian Monte Carlo algorithms, where it can lead to implementationally and analytically complex boundary crossings. We demonstrate that the continuous-time zig-zag process can largely avoid these complications.