Signature kernel ridge regression gets finite-sample error bounds, and the excess risk still decays on heavy-tailed path data
Synopsis
Path signatures can approximate functions on path space arbitrarily well, but how much data is needed to pin down a suitable signature functional was not settled. Working in the signature kernel ridge regression framework, this work derives finite-sample excess-risk bounds and consistency guarantees under moment conditions on the signature kernel and the response; when the driving path is Brownian motion and the standard moment condition provably fails, a weaker finite-moment assumption with the Fuk-Nagaev inequality yields the same order of decay.
Interpretation
Regression on path space now comes with explicit finite-sample error bounds, not only the expressivity statement that linear functionals of the signature can approximate any function. Earlier universal approximation results showed that linear functionals of the signature are in principle expressive enough, but did not say how much data is needed to identify a suitable signature functional. Under strong moment assumptions on the signature kernel and the response, a Bernstein-inequality argument controls the excess risk with at least a stated probability at a particular scaling of the regularization parameter, and the bound vanishes as the sample size grows.
The error analysis splits the error into an approximation part and a finite-sample statistical part, with the former characterized by the spectral decomposition of an integral operator. The signature kernel ridge regression estimator is a finite-dimensional solution in a reproducing kernel Hilbert space spanned by kernel evaluations, so its error has to be measured relative to the optimal function. On distributions such as time-augmented Brownian motion that satisfy the integrability and density assumptions, the theory is accompanied by three numerical examples showing the asymptotic behavior of the excess risk and of the optimal penalization coefficient.
For signature kernels driven by Brownian motion, the standard moment condition provably fails, so the analysis switches to an assumption requiring only finitely many moments. Bernstein-type concentration requires sub-exponential moment growth, whereas the moments of the time-augmented Brownian signature kernel grow at a rate that violates that condition. Replacing Bernstein with the Fuk-Nagaev inequality under the finite-moment assumption gives excess-risk bounds of the same order as the light-tailed case, at the cost of a stronger lower bound on the regularization parameter.
Perspective
The analysis concerns kernel ridge regression on path space targeting a conditional expectation, with inputs given by signatures or robust signatures of geometric rough paths and error measured as excess risk; the numerical part covers learning the terminal value of a stochastic differential equation, a signature-based model of financial market prices, and several datasets from a time-series extrinsic regression database.
The light-tailed and heavy-tailed excess-risk bounds have the same order, but the heavy-tailed case requires a stronger lower bound on the regularization parameter; the robust signature kernel is bounded, so it needs no distribution-dependent moment assumption, and its bound decays at the same order. The numerical examples show asymptotic behavior and the optimal penalization coefficient rather than a term-by-term quantitative check against the theoretical constants.
