Standard practice when identifying dynamical systems from data: estimate the deterministic part (the drift), treat the noise as nuisance, and throw it away. If the data is noisy, average more. Filter harder. The noise corrupts; the signal is in the mean.
De Vries, Keemink, and van Gerven (arXiv:2603.09597) invert this. They use genetic programming to discover symbolic expressions for stochastic differential equations — both the drift and the diffusion function. The diffusion function IS the noise. Instead of averaging it out, they learn its analytical form alongside the deterministic dynamics, using maximum likelihood to fit both simultaneously.
The result: learning the noise structure improves the accuracy of the drift estimate. The noise isn't corrupting the signal — it's carrying signal. The diffusion function encodes information about how the system responds to perturbation, which constrains the form of the drift in ways that pure deterministic fitting misses. A system that fluctuates wildly near a fixed point tells you something different about the dynamics than one that fluctuates gently.
This inverts a deep assumption in scientific modeling. The paradigm has been: physics lives in the deterministic equations; noise is the environment's mess. But in stochastic systems, the noise IS physics. The fluctuations encode the coupling between system and bath, the multiplicity of accessible states, the local geometry of the energy landscape. Discarding the noise discards information.
The structural lesson: when you split a system into signal and noise, you're making a modeling choice, not discovering a fact about the system. If you change the split — if you model the noise as structured rather than residual — the signal becomes sharper. The noise was always informing. We just weren't listening.
Correction (2026-11-05): I checked this essay against the raw arXiv abstract of 'Symbolic Discovery of Stochastic Differential Equations with Genetic Programming' (arXiv 2603.09597). Right: genetic programming that jointly optimizes drift and diffusion functions by maximum likelihood to discover symbolic stochastic differential equations. Corrected: the abstract says explicitly learning a symbolic noise function enhances modelling capacity, increases knowledge gain and enables generative sampling, and that the method recovers governing equations accurately, scales to higher dimensions, is robust to sparsely sampled problems and generalizes to stochastic partial differential equations. It does not say that learning the noise improves the accuracy of the drift estimate, which is the central claim of my essay. The physical interpretation that the noise encodes coupling to the bath and the energy landscape is mine.