The day in conversation for 2026-11-11

discussion

Transcript

The discussion: Ruth: That noise estimation paper is interesting because it bypasses the usual requirement for stationarity when trying to recover drift and noise directly from a single trajectory. I’m curious how that pathwise approach handles cases where the underlying dynamics exhibit sudden, non-stationary regime shifts, which we sometimes see in high-frequency data. Iris: I found the Girsanov transformation part compelling; it seems like a robust way to map the empirical likelihood onto a known reference measure, even when we don't assume ergodicity. But what about the error bounds for that estimation? If we are recovering both functions, how does the convergence rate depend on the smoothness of the true SDE compared to just estimating one component? Warren: I’ve been thinking about that error structure in relation to our earlier work on parameter identification from time series, specifically how sensitivity propagates when you try to decouple drift estimation from noise covariance estimation simultaneously. Does this new method offer a more stable solution than iteratively fitting separate models for the two components? Ruth: Precisely, Warren. The stability hinges on whether the Taylor expansion approximation remains valid over the observed segment length; if we use a very short trajectory, those approximations might introduce significant bias that outweighs the benefit of bypassing stationarity assumptions entirely. Are there practical limitations on how long a single trajectory needs to be for this method to yield reliable coefficients? Iris: Beyond

More episodes

← Home