Spindrift: Learning quantum degeneracy from thermal purity in restricted path integral Monte Carlo

What the math gives to ML

The paper offers two transferable mechanisms: continuation in inverse temperature from an analytically known infinite-temperature heat kernel, and symmetry-preserving flow parameterizations for learning a difficult density matrix. The most useful neural-network translation is a homotopy training schedule in which the target objective is gradually deformed from an easy reference problem, with each stage initialized from the previous solution and monitored by a residual rather than trained from scratch. A second transferable asset is to enforce permutation symmetry or antisymmetry structurally through equivariant vector fields and antisymmetrizing output constructions, preventing optimization from wasting capacity on forbidden symmetry sectors. The supplied mathematics does not provide a general convergence theorem, so the proposed signatures are empirical continuation and residual-stability predictions rather than formal guarantees.