Abstract:Stein Variational Gradient Descent (SVGD) is a popular variational inference algorithm which simulates an interacting particle system to approximately sample from a target distribution, with impressive empirical performance across various domains. Theoretically, its population (i.e, infinite-particle) limit dynamics is well studied but the behavior of SVGD in the finite-particle regime is much less understood. In this work, we design two computationally efficient variants of SVGD, namely VP-SVGD and GB-SVGD, with provably fast finite-particle convergence rates. We introduce the notion of virtual particles and develop novel stochastic approximations of population-limit SVGD dynamics in the space of probability measures, which are exactly implementable using a finite number of particles. Our algorithms can be viewed as specific random-batch approximations of SVGD, which are computationally more efficient than ordinary SVGD. We show that the $n$ particles output by VP-SVGD and GB-SVGD, run for $T$ steps with batch-size $K$, are at-least as good as i.i.d samples from a distribution whose Kernel Stein Discrepancy to the target is at most $O\left(\tfrac{d^{1/3}}{(KT)^{1/6}}\right)$ under standard assumptions. Our results also hold under a mild growth condition on the potential function, which is much weaker than the isoperimetric (e.g. Poincare Inequality) or information-transport conditions (e.g. Talagrand's Inequality $\mathsf{T}_1$) generally considered in prior works. As a corollary, we consider the convergence of the empirical measure (of the particles output by VP-SVGD and GB-SVGD) to the target distribution and demonstrate a double exponential improvement over the best known finite-particle analysis of SVGD. Beyond this, our results present the first known oracle complexities for this setting with polynomial dimension dependence.

Provably Fast Finite Particle Variants of SVGD via Virtual Particle Stochastic Approximation

DPVI: A Dynamic-Weight Particle-Based Variational Inference Framework

Long-time asymptotics of noisy SVGD outside the population limit

Augmented Message Passing Stein Variational Gradient Descent

Stein Variational Gradient Descent with Local Approximations

Improved Finite-Particle Convergence Rates for Stein Variational Gradient Descent

A Stochastic Version of Stein Variational Gradient Descent for Efficient Sampling

Gradient-free Stein Variational Gradient Descent with Kernel Approximation

Stein Variational Gradient Descent with Learned Direction

Accelerating Convergence of Stein Variational Gradient Descent via Deep Unfolding

On the mean-field limit for Stein variational gradient descent: stability and multilevel approximation

Stein Variational Inference for Discrete Distributions

SVGD: A Virtual Gradients Descent Method for Stochastic Optimization

Stein Variational Evolution Strategies

Stein Variational Gaussian Processes

Particle Optimization in Stochastic Gradient MCMC

Flexible and Efficient Inference with Particles for the Variational Gaussian Approximation

Stein Boltzmann Sampling: A Variational Approach for Global Optimization

Understanding and Accelerating Particle-Based Variational Inference

Particle-based Variational Inference with Generalized Wasserstein Gradient Flow

Stein Variational Gradient Descent on Infinite-Dimensional Space and Applications to Statistical Inverse Problems