1: \begin{abstract}

2: %Stein Variational Gradient Descent (SVGD) is a popular nonparametric variational inference algorithm which simulates an interacting particle system to approximate a target distribution whose density is known only upto a normalizing constant. While SVGD has demonstrated promising empirical performance across various domains, and its population (i.e, infinite-particle) limit are well understood, its finite particle convergence guarantees are not due to the complex dependencies. In this work, we design two computationally efficient variants of SVGD, namely \ouralg~and \ouralgnew, with provably fast finite-particle convergence rates. By introducing the notion of \emph{virtual particles}, we develop novel stochastic approximations of population-limit SVGD in the space of probability measures, which can be efficiently realized algorithmically by tracking exactly $n$ number of particles. The proposed algorithms can be interpreted as specific random-batch approximations of SVGD, which are computationally more efficient than SVGD. We establish that the $n$ particles output by \ouralg~and \ouralgnew, run for $T$ steps, are at-least as good as $n$ i.i.d draws from a distribution whose Kernel Stein Discrepancy to the target distribution is at most $O(T^{\nicefrac{-1}{6}})$ under standard assumptions. Our results hold under a mild growth condition on the log-likelhood function, which is significantly weaker than the isoperimetric assumptions (such as 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 \ouralg~and \ouralgnew) to the target and demonstrate a \emph{double exponential improvement} over the best known finite-particle analysis of vanilla SVGD. 

3: %\end{abstract}

4: