Hug and hop: a discrete-time, nonreversible Markov chain Monte Carlo algorithm

Summary This article introduces the hug and hop Markov chain Monte Carlo algorithm for estimating expectations with respect to an intractable distribution. The alternates between two kernels, referred as hop. Hug is a nonreversible kernel that repeatedly applies bounce mechanism from recently proposed bouncy particle sampler produce proposal point far current position yet on almost same contour of target density, leading high acceptance probability. complemented by hop, which deliberately proposes jumps contours has efficiency degrades very slowly increasing dimension. There are many parallels Hamiltonian using leapfrog integrator, including order integration scheme, but also able make use local Hessian information without requiring implicit numerical steps, its performance not terminally affected unbounded gradients log-posterior. We test empirically variety toy targets real statistical models, find it can, often does, outperform Carlo.