Abstract
The signature kernel is a recent state-of-the-art tool for analyzinghigh-dimensional sequential data, valued for its theoretical guarantees andstrong empirical performance. In this paper, we present a novel method forefficiently computing the signature kernel of long, high-dimensional timeseries via dynamically truncated recursive local power series expansions.Building on the characterization of the signature kernel as the solution of aGoursat PDE, our approach employs tilewise Neumann-series expansions to deriverapidly converging power series approximations of the signature kernel that arelocally defined on subdomains and propagated iteratively across the entiredomain of the Goursat solution by exploiting the geometry of the time series.Algorithmically, this involves solving a system of interdependent local GoursatPDEs by recursively propagating boundary conditions along a directed graph viatopological ordering, with dynamic truncation adaptively terminating each localpower series expansion when coefficients fall below machine precision, strikingan effective balance between computational cost and accuracy. This methodachieves substantial performance improvements over state-of-the-art approachesfor computing the signature kernel, providing (a) adjustable and superioraccuracy, even for time series with very high roughness; (b) drasticallyreduced memory requirements; and (c) scalability to efficiently handle verylong time series (e.g., with up to half a million points or more) on a singleGPU. These advantages make our method particularly well-suited forrough-path-assisted machine learning, financial modeling, and signal processingapplications that involve very long and highly volatile data.