1: \begin{abstract}%

2: %

3: Maximum mean discrepancies (MMDs) like the

4: kernel Stein discrepancy (KSD)

5: have grown central to a wide range of applications,

6: including hypothesis testing, sampler selection, distribution approximation, and variational inference.

7: %

8: In each setting, these kernel-based discrepancy measures are required to $(i)$

9: separate a target $\mathrm{P}$ from other probability measures or even $(ii)$

10: control weak convergence to $\mathrm{P}$.

11: In this article we derive new sufficient and necessary conditions to ensure $(i)$ and $(ii)$.

12: For MMDs on separable metric spaces, we characterize those kernels that separate Bochner embeddable measures and

13: introduce simple conditions for separating all measures with unbounded kernels and for controlling convergence with bounded kernels.

14: We use these results on $\R^d$ to

15: substantially broaden the known conditions

16: for KSD separation and convergence control and to develop the first KSDs known to exactly metrize weak convergence to $\P$.

17: Along the way, we highlight the implications of our results for hypothesis testing, measuring and improving sample quality, and sampling with Stein variational gradient descent.

18: \end{abstract}

19: