1: \begin{abstract}

2: Stein discrepancies (SDs) %

3: monitor convergence and non-convergence 

4: in approximate inference when exact integration and sampling are intractable.

5: However, the computation of a Stein discrepancy can be prohibitive

6: if the Stein operator -- often a sum over likelihood terms or potentials -- is expensive to evaluate. 

7: To address this deficiency, we show that \emph{stochastic Stein discrepancies} (SSDs) based on subsampled approximations of the Stein operator inherit the convergence control properties of standard SDs with probability $1$.

8: %

9: %

10: %

11: %

12: %

13: %

14: %

15: %

16: %

17: %

18: %

19: Along the way, we establish the convergence of Stein variational gradient descent (SVGD) on unbounded domains, resolving an open question of Liu (2017).

20: In our experiments with biased Markov chain Monte Carlo (MCMC) hyperparameter tuning, approximate MCMC sampler selection, and stochastic SVGD, 

21: SSDs deliver comparable inferences to standard SDs with orders of magnitude fewer likelihood evaluations.

22: \end{abstract}

23: