Lagrangian Dynamical Monte Carlo

Shiwei Lan,Vassilios Stathopoulos,Babak Shahbaba,Mark Girolami

DOI: https://doi.org/10.1080/10618600.2014.902764

2012-11-21

Abstract:Hamiltonian Monte Carlo (HMC) improves the computational efficiency of the Metropolis algorithm by reducing its random walk behavior. Riemannian Manifold HMC (RMHMC) further improves HMC's performance by exploiting the geometric properties of the parameter space. However, the geometric integrator used for RMHMC involves implicit equations that require costly numerical analysis (e.g., fixed-point iteration). In some cases, the computational overhead for solving implicit equations undermines RMHMC's benefits. To avoid this problem, we propose an explicit geometric integrator that replaces the momentum variable in RMHMC by velocity. We show that the resulting transformation is equivalent to transforming Riemannian Hamilton dynamics to Lagrangian dynamics. Experimental results show that our method improves RMHMC's overall computational efficiency. All computer programs and data sets are available online (<a class="link-external link-http" href="http://www.ics.uci.edu/~babaks/Site/Codes.html" rel="external noopener nofollow">this http URL</a>) in order to allow replications of the results reported in this paper.

Computation,Dynamical Systems,Computational Physics

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the low computational efficiency in the Riemannian Manifold Hamiltonian Monte Carlo (RMHMC) method due to the need to solve implicit equations. Specifically, RMHMC improves the sampling efficiency of the standard HMC by exploiting the geometric properties of the parameter space, but this improvement comes at the cost of increased computational complexity. Because the geometric integrator of RMHMC involves two implicit equations, these equations require additional numerical analysis (such as fixed - point iteration), thus increasing the computational cost and simulation error.

To solve this problem, the author proposes a new integrator that is completely explicit: by replacing momentum with velocity in the Riemannian Manifold Hamilton dynamics, one of the implicit steps is eliminated. Then, a time - symmetric integrator is constructed to remove the remaining implicit step, thus proposing a sampling scheme that contains only explicit equations, called e - RMHMC (explicit Riemannian Manifold Hamiltonian Monte Carlo). This improvement aims to reduce the computational overhead while maintaining or improving the efficiency and accuracy of sampling. Experimental results show that this method indeed improves the overall computational efficiency of RMHMC.