The expected signature kernel measures similarity between probability laws on path space, and it underpins signature maximum mean discrepancies (MMD) used in statistical tests. Friz and Hager build on earlier work by Friz, Hager, and Tapia and by Lemercier, Lyons, and Salvi to extend the reach of these kernels.
Their main result characterizes the expected signature kernel for group-valued inhomogeneous Lévy processes, also called Lévy rough paths, as the scalar component of a solution to an extended Goursat-type system. When the process has no jumps, the system is finite-dimensional and exact. In the presence of jumps, the system becomes infinite-dimensional; the authors prove it is well-posed and derive finite-dimensional truncations with explicit error control.
This characterization has practical consequences for signature-MMD tests. It allows direct evaluation of the reference-dependent terms in a one-sample test without drawing an auxiliary reference sample. For conditionally Gaussian martingales, the authors show that conditioning the usual unbiased empirical squared signature-MMD on sampled quadratic variations yields an unbiased statistic with no larger variance. They derive computable finite-grid versions and demonstrate variance reduction and improved power in illustrations.