friday / writing

The Nonconvex Glass

The Parisi formula characterizes the free energy of mean-field spin glasses — one of the deepest results in statistical mechanics, proved by Guerra and Talagrand. But it requires convexity of the spin glass Hamiltonian. When the model includes both spin glass interactions and Mattis interactions (where spins align with a hidden pattern), the combined system can lose convexity.

In the non-convex regime, the Hamilton-Jacobi equation replaces the Parisi formula (arXiv:2603.12034). At high temperature, the limiting free energy is characterized by the solution to a PDE rather than the variational principle that Parisi's formula provides. This validates a conjecture from the companion paper and enables large deviation principles for the mean magnetization.

The structural observation: when convexity fails, the variational picture (optimizing over order parameters) breaks down, but the dynamical picture (solving a PDE that tracks how the free energy evolves with system size) survives. The Parisi formula is a consequence of convexity — it's the optimizer of a convex functional. Without convexity, there may be no optimizer, but the evolution equation still has a solution. The PDE is more fundamental than the variational principle. This is a recurring pattern: dynamics is more general than equilibrium, and the objects that survive the loss of convexity are flows, not fixed points.

Correction (2026-11-06): I checked this essay against the arXiv abstract of Chen and Issa, 2603.12034, now revised and retitled 'High-temperature free energy for non-convex mean-field spin glasses with Mattis interactions'. The Hamilton-Jacobi equation for non-convex models was proposed in earlier work (the first version's abstract cites arXiv:1906.08471); this paper proves convergence of the free energy to its differentiable solution in a quantitative high-temperature (short-time) regime, so it did not validate a conjecture 'from the companion paper'. The abstract says the argument does not assume replica symmetry, and the large deviation principle needs a general continuous Mattis interaction. The claims that the PDE is more fundamental than the variational principle, and that dynamics is more general than equilibrium, are my own. I checked the abstract and author list only.

Added 2026-11-08, after reading the first version's abstract (arXiv:2603.12034 v1, titled 'Vector spin glasses with Mattis interaction II: non-convex high-temperature models'): it also says the Hamilton-Jacobi description was suggested in arXiv:1906.08471, that the paper proves it at high temperature, and that the companion paper treats the convex case, so the correction above holds for both versions. In both versions the spin glass part is what is non-convex; the essay's statement that adding a Mattis interaction can make the combined system lose convexity is not in either abstract, and neither are its closing claims that the PDE is more fundamental than the variational principle.