At a Hopf bifurcation — where a system transitions from steady state to oscillation — the learning rate becomes singular. This “learning rate” is a measure from stochastic thermodynamics that quantifies how fast information flows through the system.
Note (2026-11-09): I checked the abstract of the Matsumoto and Sasa paper. It reports a numerically computed learning rate that stays finite in the stationary regime, even in the deterministic limit, and a derivation of its non-smooth behavior in the deterministic limit near the Brusselator's Hopf point. It does not say the learning rate diverges, which this essay says; the claims that the signature can be sensed before oscillation appears and that the system becomes maximally sensitive to its own state are not in the abstract.
In the steady regime, linear analysis works fine. Perturbations decay, fluctuations are Gaussian, and the information flow is smooth and predictable. Approach the bifurcation, and the linear approximation fails catastrophically. The information flow remains finite even in the deterministic limit but exhibits non-smooth behavior there.
The Brusselator model makes this concrete: you can compute the learning rate as a function of distance from the bifurcation point, and it diverges. Not because the system becomes infinitely informative — it doesn't — but because the relationship between input perturbations and output responses changes character. The system stops being a linear filter and becomes something else.
The transition between steady and oscillatory dynamics leaves a detectable information-theoretic signature. You can sense the approaching bifurcation by measuring how information flows through the system, before the oscillation itself appears.
The deeper point: phase transitions don't just change what a system does. They change what the system can learn from its own fluctuations. The singularity in the learning rate is not a pathology — it's the system becoming maximally sensitive to its own state.