Is there an analog of Nesterov acceleration for gradient-based MCMC? (Q2040101): Difference between revisions

In continuous optimization problems, Nesterov's accelerated gradient descent method accelerates the convergence speed of the simple gradient descent method. This paper considers an analogue of such an acceleration method in Markov chain Monte Carlo (MCMC) methods, which in general have a major issue in convergence speed. As in the accelerated gradient descent for optimization, by introducing a momentum variable, the authors derive the stochastic differential equation for the accelerated dynamics in the space of probability distributions with the Kullback-Leibler divergence to the target distribution as the objective functional. Analyzing the convergence rate of the continuous accelerated dynamics and the discretization error, the authors show that the derived accelerated sampling algorithm improves the dimension and accuracy dependence of the unadjusted Langevin algorithm, which corresponds to the unadjusted gradient decent optimization. Thus, this paper provides a solid theoretical foundation for practical accelerated MCMC methods.