Free energy of spherical vector spin glasses

Hong-Bin Chen

Oct 31, 2026arXiv:2611.00710v1

math.PRcond-mat.dis-nn

Scorecard· 16/16

6.5/10 impact

Solid, complete extension of Hamilton–Jacobi free-energy formulas to non-convex spherical vector spin glasses via a neat disintegration trick, but incremental within an active programme and of narrow audience.

Abstract

We identify the limit free energy of spherical vector spin glasses by a variational formula. The covariance function need not be convex. Using the global symmetry of the spin configurations, we introduce a family of constraint measures that gives a disintegration of the original spin distribution. We first identify the limit free energy of each constrained model and then apply the Laplace method to recover that of the original model.

AI Impact Assessments

(1 model)

Paper type: Theoretical mathematical physics / probability. There are no numerical experiments, and the results are proofs.

Core Contribution

The paper identifies the limit free energy of spherical vector spin glasses with D-dimensional spins and a general, possibly non-convex, covariance ξ built from PSD tensor series. The covariance may depend on the full D×D overlap matrix. The limit is given by a variational formula: an infimum over self-overlap matrices a (unit diagonal) of a Hopf-type formula f_a(1/2, 0), minus ½ξ(a²) and minus log det a.

The key idea is to split the product-sphere reference measure exactly. The split uses a Wishart-type density on the self-overlap b = a² and a Haar-distributed fibre measure ν_N on {u : uu^T = N·I_D}. Three things follow from this:

Each fixed-a model has an initial condition ψ_{ah} that is convex in the matrix path (Prop. 2.15, via concavity of log det and affine dependence of L_{ℓ,q} on q). This is exactly what the Hamilton–Jacobi/Hopf machinery of Chen–Mourrat–Issa needs.

The −log det a term arises explicitly from the Wishart density.

The earlier ε-neighbourhood constraint of Ko and of Panchenko–Talagrand, whose mass was retained, is replaced by a normalised constraint measure that separates this contribution.

This extends the spherical case from convex covariance to non-convex covariance, and from scalar or multi-species overlaps to full matrix overlaps. This is the main advance over Ko (even p-spin, coupled copies) and Bates–Sohn (multi-species, convex).

Methodological Rigor

The proof is long and appears careful.

Constrained limit. The fibre measure ν_N does not factor over columns, so the paper compares it with block-product models (Lemma 3.1 on rotation invariance, Lemma 3.2 with an M^{-1/4} error, polar normalization) and then invokes the Ho and Chen–Ho–Issa convergence results for compact product measures. The critical-point representation of Chen–Mourrat is passed to M→∞ using a uniform semiconcavity estimate (Lemma 4.3), which improves on the earlier eigenvalue-ratio assumption. The Hopf lower bound comes from Mourrat and Chen–Xia, and convexity of ψ closes the upper bound (Corollary 5.2).

Optimisation over a. The Laplace step uses uniform convergence on {det a² ≥ κ} via net arguments and concentration, and then shows that small-determinant a cannot be near-minimisers.

Verifiable pieces. The explicit Gaussian computations (Lemma 2.6, coercivity and strict convexity of G_v) can be checked directly.

Dependence on several companion papers from 2026, some still preprints (Ho; Chen–Ho–Issa; Chen–Mourrat multi-species spherical), is a vulnerability. The AI-use statement says an AI model identified the key constraint argument. That is unusual but irrelevant to the correctness of the mathematics. The proofs appear complete, with some inherent reliance on cited technical results.

Potential Impact

This is a solid advance for the mathematical spin-glass community and closes a natural gap, namely non-convex spherical vector models with full matrix overlap.

The disintegration-plus-constrained-Hopf strategy is reusable. It is the same logic as Boursier–Mourrat (magnetisation constraint) and Chen–Ho–Issa (empirical measure constraint), now applied to self-overlap, so it could extend to other models with global symmetry constraints.

Applications are indirect: spherical models serve as proxies for high-dimensional landscapes, such as multi-component or tensor PCA-like problems and optimisation, and the zero-temperature limit may give ground-state formulas. No practical deployment is plausible.

The audience is narrow (mathematical physicists and probabilists), though the Hamilton–Jacobi community in mean-field theory will notice it.

Timeliness & Relevance

This fits an active programme to remove convexity assumptions in free-energy variational formulas using Hamilton–Jacobi methods. It completes the picture for the spherical case alongside the Ising and multi-species cases. The timing is good but the contribution is incremental within the programme.

Strengths & Limitations

Strengths:

A clean, explicit variational formula with attained minimisers, covering fully general non-convex ξ.

A conceptually neat exact disintegration yielding the log det term, and a convex initial condition without requiring convexity of ξ.

Complete, detailed proofs (including the Wishart Jacobian and Stirling asymptotics).

Limitations:

The result is a formula that is not simplified further. It is a Hopf formula over matrix-valued paths plus a finite-dimensional optimisation, so structural information (e.g. replica symmetry criteria, a Crisanti–Sommers form) is not extracted.

The paper depends heavily on a chain of recent companion preprints and on earlier machinery, and the new input is largely the constraint device plus its technical verification.

The block comparison gives only an M^{-1/4} rate, which suffices for the limit but is not quantitative.

Little is said about uniqueness or properties of the minimisers a. The paper is dense, with a long length to novelty ratio.

It is restricted to a fixed D and to sphere constraints.

Other observations

There is no empirical component, so reproducibility is in the sense of proof verifiability. The paper's structure is clear, with a good outline section, but the technical sections are heavy.

Rating:6.5/ 10

Significance 6Rigor 7.5Novelty 6.5Clarity 6

Generated Nov 3, 2026

Comparison History (0)

No comparisons yet.

Scorecard· 16/16

6.5/10 impact