1: \begin{abstract}

2: Approximate Markov chain Monte Carlo (MCMC) offers the promise of more rapid sampling at the cost of more biased inference.

3: Since standard MCMC diagnostics fail to detect these biases, researchers have 

4: %

5: developed

6: computable \emph{Stein discrepancy} measures that provably determine %

7: the convergence of a sample to its target distribution.

8: %

9: %

10: %

11: %

12: %

13: %

14: %

15: %

16: %

17: %

18: %

19: %

20: %

21: %

22: %

23: %

24: %

25: %

26: %

27: %

28: %

29: %

30: %

31: %

32: This approach

33: was recently combined with the theory of reproducing kernels

34: %

35: to define a closed-form \emph{kernel Stein discrepancy} (KSD) computable by summing kernel evaluations across pairs of sample points. 

36: We develop a theory of weak convergence for KSDs based on Stein's method, 

37: demonstrate that commonly used KSDs fail to detect non-convergence even for Gaussian targets, 

38: %

39: %

40: and show that kernels with slowly decaying tails provably

41: determine convergence for a large class of target distributions.  The

42: resulting convergence-determining KSDs are suitable for comparing biased,

43: exact, and deterministic sample sequences and simpler to compute and

44: parallelize than alternative Stein discrepancies. 

45: We use our tools to compare biased samplers, select sampler hyperparameters,

46: and improve upon existing KSD approaches to one-sample hypothesis testing and sample quality improvement.

47: \end{abstract}

48: