Abstract
We propose a procedure for optimising the friction matrix of underdamped Langevin dynamics when used for continuous time Markov Chain Monte Carlo. Starting from a central limit theorem for the ergodic average, we present a new expression of the gradient of the asymptotic variance with respect to friction matrix. In addition, we present an approximation method that uses simulations of the associated first variation/tangent process. Our algorithm is applied to a variety of numerical examples such as toy problems with tractable asymptotic variance, diffusion bridge sampling and Bayesian inference problems for high dimensional logistic regression.
Funder
Engineering and Physical Sciences Research Council
JPMorgan Chase and Company
Leverhulme Trust
Reference70 articles.
1. Aliprantis C.D. and Border K.C., Infinite Dimensional Analysis: A Hitchhiker’s Guide. 3rd edition. Springer, Berlin (2006).
2. Some Remarks on Preconditioning Molecular Dynamics
3. A tutorial on adaptive MCMC
4. Angeli L., Crisan D. and Ottobre M., Uniform in time convergence of numerical schemes for stochastic differential equations via Strong Exponential stability: Euler methods. Split-Step and Tamed Schemes. Preprint arXiv:2303.15463 (2023).
5. Control variates for stochastic gradient MCMC