Information Newton's flow: second-order optimization method in probability space

Abstract

We introduce a framework for Newton's flows in probability space with information metrics, named information Newton's flows. Here two information metrics are considered, including both the Fisher-Rao metric and the Wasserstein-2 metric. A known fact is that overdamped Langevin dynamics correspond to Wasserstein gradient flows of Kullback-Leibler (KL) divergence. Extending this fact to Wasserstein Newton's flows, we derive Newton's Langevin dynamics. We provide examples of Newton's Langevin dynamics in both one-dimensional space and Gaussian families. For the numerical implementation, we design sampling efficient variational methods in affine models and reproducing kernel Hilbert space (RKHS) to approximate Wasserstein Newton's directions. We also establish convergence results of the proposed information Newton's method with approximated directions. Several numerical examples from Bayesian sampling problems are shown to demonstrate the effectiveness of the proposed method.

arXiv comment62 pages

References in corpus (17)

• Adam: A Method for Stochastic Optimization
• Variational Inference with Normalizing Flows
• Optimizing Neural Networks with Kronecker-factored Approximate Curvature
• Information-Geometric Optimization Algorithms: A Unifying Picture via Invariance Principles
• Stein Variational Gradient Descent as Gradient Flow
• Interacting particle solutions of Fokker-Planck equations through gradient-log-density estimation
• Understanding and Accelerating Particle-Based Variational Inference
• A Stein variational Newton method
• Is There an Analog of Nesterov Acceleration for MCMC?
• Projected Stein Variational Newton: A Fast and Scalable Bayesian Inference Method in High Dimensions
• Transport information geometry I: Riemannian calculus on probability simplex
• Langevin Monte Carlo and JKO splitting
• On Wasserstein geometry of the space of Gaussian measures
• Accelerating Langevin Sampling with Birth-death
• Wasserstein Riemannian Geometry of Positive Definite Matrices
• Proximal Langevin Algorithm: Rapid Convergence Under Isoperimetry
• Accelerated Flow for Probability Distributions

Cited by in corpus (5)

• Practical Quasi-Newton Methods for Training Deep Neural Networks
• Projected Stein Variational Gradient Descent
• Efficient constrained sampling via the mirror-Langevin algorithm
• Momentum Accelerates Evolutionary Dynamics
• Wasserstein Statistics in One-dimensional Location-Scale Model