Convergence of unadjusted Hamiltonian Monte Carlo for mean-field models (Q6165984)

In this paper, the authors introduce convergence and discretization error bounds that are independent of dimensionality for the unadjusted Hamiltonian Monte Carlo algorithm when applied to high-dimensional probability distributions of mean-field type. Unlike previous approaches, their bounds necessitate a sufficiently small discretization step but do not impose the requirement of strong convexity for either unary or pairwise potential terms inherent in the mean-field model. To address the challenges posed by high dimensionality, the proof employs a particlewise coupling strategy that exhibits contractivity in a complementary particlewise metric. This novel approach enhances the algorithm's ability to handle high-dimensional scenarios effectively. The presented results contribute to a more comprehensive understanding of the Hamiltonian Monte Carlo algorithm's performance in the context of mean-field distributions, especially in scenarios where traditional assumptions of strong convexity may not be applicable.