An Eyring--Kramers Law for the Hypoelliptic Third-Order Langevin Diffusion

What the math gives to ML

The paper provides a principled third-order hypoelliptic Langevin process in which noise injected into an auxiliary variable reaches parameters through a deterministic Hörmander chain, together with a sharp low-temperature metastability law. The transferable asset is not merely Langevin sampling, but the explicit relationship between saddle curvature, the positive unstable eigenvalue of a cubic linearization, and transition-time prefactors. This suggests a momentum-like optimizer with two auxiliary states whose escape and basin-crossing behavior can be tuned using the cubic unstable-rate calculation rather than ad hoc damping. A first experiment should test whether this higher-order dynamics escapes poor nonconvex basins faster than overdamped and underdamped Langevin at matched gradient-evaluation cost while preserving useful training stability.