Gaussian Mixture Free Energies: Variational Structure, Landau Reduction, and Tricritical Geometry
Year: 2026
Type: preprint
Abstract: Finite Gaussian mixtures provide a concrete setting in which the geometry of component locations, posterior uncertainty, and local free-energy singularities interact. This manuscript establishes an exact variational correspondence between the negative log-density of a common-covariance Gaussian mixture, an entropy-regularized free energy on the probability simplex, and an associated finite-state mean-field model. The correspondence preserves local and global minima, including degenerate minima, and relates Hessian inertia by Schur complements. Posterior cumulants yield a Landau reduction at a stationary point with a one-dimensional Hessian kernel. Relaxing a stable transverse coordinate subtracts a nonnegative contribution from the reduced quartic derivative, so a positive quartic derivative along the kernel line can become negative after relaxation. Under vanishing projected third cumulant, a geometric diameter bound excludes this reversal when the dimensionless squared diameter is below 5/2 + sqrt(6). An explicit planar three-component mixture realizes strict reversal, while a weighted three-component family gives a complementary upper bound of 36/7 on the unresolved optimal threshold. At a special endpoint, the reduced energy has positive leading sixth-order behavior. The manuscript derives a symmetric two-parameter local unfolding, quarter-power amplitude scaling, fifth-root conjugate-field response, and local fold and equal-depth curves. It also identifies the relation of the endpoint to a known three-state Curie–Weiss–Potts butterfly. These are local geometric and mean-field results; they do not establish a new global mode bound or a complete thermodynamic phase diagram. Unrefereed preprint for open scholarly discussion. This record contains the manuscript PDF only; companion computational scripts mentioned in the manuscript are not included in this deposit. No priority is claimed for the mixture–simplex correspondence. The comparison with Okahara and Kobayashi (arXiv:2610.15369) is based on their abstract; a full-text, theorem-by-theorem comparison remains outstanding.
Author: Ma, SeokBin
Institution: Hongik University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7226035606
