A1048
Title: Convergence results of numerically estimated JKO scheme
Authors: Paco Tseng - University of Sydney (Australia) [presenting]
Minh-Ngoc Tran - University of Sydney (Australia)
Abstract: Applications of Wasserstein gradient flow in the field of Bayesian Computation have been a trending topic. The iterative scheme developed under this framework is termed JKOscheme, and the theory states that the solution sequence of the JKO scheme converges to the gradient flow of the KL divergence with respect to the posterior distribution. At each iteration, the solution is determined by a vector field. In particular, the convergence results rely on the tractable vector field. However, in practice, the vector field is often estimated by some numerical methods for various reasons, for instance, large data sets and analytically intractable vector fields. The convergence results of the JKO scheme, when the estimated vector field is in place, are provided.
