Abstract
This paper introduces and analyses interacting underdamped Langevinalgorithms, termed Kinetic Interacting Particle Langevin Monte Carlo (KIPLMC)methods, for statistical inference in latent variable models. We propose adiffusion process that evolves jointly in the space of parameters and latentvariables and exploit the fact that the stationary distribution of thisdiffusion concentrates around the maximum marginal likelihood estimate of theparameters. We then provide two explicit discretisations of this diffusion aspractical algorithms to estimate parameters of statistical models. For eachalgorithm, we obtain nonasymptotic rates of convergence for the case where thejoint log-likelihood is strongly concave with respect to latent variables andparameters. In particular, we provide convergence analysis for the diffusiontogether with the discretisation error, providing convergence rate estimatesfor the algorithms in Wasserstein-2 distance. We achieve acceleratedconvergence rates clearly demonstrating improvement in dimension dependence,similar to the underdamped samplers. To demonstrate the utility of theintroduced methodology, we provide numerical experiments that demonstrate theeffectiveness of the proposed diffusion for statistical inference and thestability of the numerical integrators utilised for discretisation. Our settingcovers a broad number of applications, including unsupervised learning,statistical inference, and inverse problems.