Discovering differential equations from data requires integrating noisy signals — a process that amplifies errors. The weak-form approach avoids direct integration by testing the equation against smooth kernel functions. Multiply both sides by a bump function, integrate, and the derivatives shift from the noisy data to the smooth kernel via integration by parts.
For deterministic systems, this works well with temporal test functions. For stochastic systems, it doesn't. The noise terms introduce a structural bias because the temporal kernel at time t is not independent of the Brownian motion it's averaging over. The regression that estimates drift and diffusion coefficients is systematically wrong.
The fix (arXiv:2603.20904) uses spatial test functions instead of temporal ones. A Gaussian kernel centered in space, evaluated at each time point. The critical property: the spatial kernel at time t is measurable with respect to the filtration at t, while the Brownian increment is independent of it. The kernel “sees” the current state but is blind to the noise that will arrive next.
This one change — spatial instead of temporal — converts a biased estimator into an unbiased one. The method recovers drift and diffusion coefficients with errors below 4% across benchmark problems.
The insight is that the witness function's domain determines whether it can testify honestly. A temporal witness is correlated with the noise it's supposed to average out. A spatial witness stands in a different dimension, observing from a position that the noise cannot reach. The testimony is the same mathematical operation — weighted integration. The credibility depends entirely on where the witness stands.