1: \begin{abstract}

2: %Stein Variational Graident Descent (SVGD) is an interacting particle based sampling algorithm, showing rapid convergence with very few particles in practice. While the infinite particle limiting process is well understood, the guarantees for a finite number of particles have been lacking due to the complex nature of interactions. In this work, we design variants of SVGD via virtual particles (\ouralg~and \ouralgnew), where we consider the evolution of an infinite number of particles allowing us to track quantities like KL divergence. Our stochastic approximation ensures that the algorithm only needs to compute a finite number of such particles in order to output. Algorithmically, our variants can be seen as specific types of random batch stochastic approximations of SVGD and are thus computationally more efficient than normal SVGD. We show that the output after $T$ steps is at-least as good as drawing $n$ i.i.d. samples from a distribution which is $O(1/T^{1/6})$ close in Kernel Stein Descrepency to the target distribution. As a corollary, we provide a double exponential improvement for the convergence of the empirical distribution to the actual distribution compared to the best known analysis of vanilla SVGD. 

3: %\end{abstract}

4: