All-temperature pressure for orthogonally invariant Ising spin glasses

← All digests

Family 234 · Probability and statistical mechanics
Importance 4/5 · Provisional editorial rating
Sources consulted 2026-10-11

1. The question

Knowing a symmetric matrix’s eigenvalues solves a quadratic optimization problem on the Euclidean sphere. It does not solve the same problem when every coordinate must be $+1$ or $-1$. A rotation preserves the sphere but moves the corners of the cube, so eigenvectors matter to an Ising spin system.

Suppose those eigenvectors are uniformly random. Can the limiting eigenvalue distribution then determine the system’s thermodynamics? Family 234 claims an affirmative answer at every fixed temperature, for compact limiting spectra without asymptotic outliers. It also treats deterministic external magnetic fields and derives zero-temperature optimization limits.

The resulting answer is a variational formula: a precisely defined optimization over possible hierarchies of spin overlaps. It need not reduce to an elementary expression, but it turns an exponentially large random sum into a deterministic mathematical problem.

2. Background and essential definitions

Random interactions and pressure

There are $N$ spins, assembled into a vector $\sigma\in\{-1,1\}^N$. Write the interaction matrix as

$$J_N=U_N^{\mathsf T}\Lambda_NU_N, \qquad \Lambda_N=\operatorname{diag}(\lambda_{1,N},\ldots,\lambda_{N,N}).$$

The diagonal matrix is initially deterministic. The orthogonal matrix $U_N$ has Haar distribution, the uniform distribution invariant under multiplying by any fixed orthogonal matrix. Thus the spectrum is prescribed while its directions are randomized.

The manuscript uses normalized pressure

$$P_N=\frac1N\log\left[2^{-N} \sum_{\sigma\in\{-1,1\}^N} \exp\!\left(\frac12\sigma^{\mathsf T}J_N\sigma\right)\right].$$

Larger quadratic values receive greater weight. Inverse temperature and coupling strength are absorbed into $\Lambda_N$. The factor $2^{-N}$ makes the spin prior a probability measure: pressure is zero when all interactions vanish. Using an unnormalized counting sum adds $\log2$.

The empirical spectral measure

$$\mu_N=\frac1N\sum_{i=1}^N\delta_{\lambda_{i,N}}$$

records the fraction of eigenvalues in each region; $\delta_x$ denotes a point mass. Its limit is $\mu$, a probability law supported in a compact interval. The smallest and largest eigenvalues must approach the corresponding endpoints of the support. Weak convergence of $\mu_N$ alone is insufficient, because it can overlook a small number of extreme eigenvalues.

Overlaps and hierarchies

Two independent equilibrium samples at the same interaction matrix are called replicas. Their overlap is

$$Q(\sigma,\tau)=\frac1N\sum_{i=1}^N\sigma_i\tau_i.$$

It is $1$ for identical configurations, $-1$ for opposites, and near zero for largely unrelated signs. In a spin glass, several levels of similarity can survive as $N$ grows.

The formula uses a nondecreasing function $p:(0,1)\to[0,1]$, interpreted as a quantile description of a trial overlap hierarchy. The nonnegative convention comes from small auxiliary perturbations used in the analysis; it does not mean unperturbed finite systems have no negative overlaps.

3. Development history and previous partial results

The Sherrington–Kirkpatrick model introduced a mean-field spin glass with Gaussian pair interactions. Parisi proposed that its low-temperature behavior requires a hierarchy of overlaps rather than one scalar similarity. Guerra proved the variational upper bound, and Talagrand established the matching pressure formula. Aizenman, Sims and Starr developed a complementary cavity formulation based on adding spins. SK model, Guerra’s bound, Talagrand’s formula, cavity principle.

Orthogonally invariant models ask how much of that picture survives with more general spectra. Statistical-mechanics calculations for the random orthogonal model and related interactions developed spectral and replica-symmetry-breaking predictions. These provided important motivation, but are different from an all-temperature theorem under explicit matrix assumptions. Marinari, Parisi and Ritort, Cherrier, Dean and Lefèvre.

Rigorous results had reached substantial high-temperature regimes. Bhattacharya and Sen treated zero-field pressure. Fan and Wu obtained a replica-symmetric formula with external fields at sufficiently high temperature. The manuscript also credits 2026 work by Fan, Misiakiewicz, Wang and Wen giving an explicit high-temperature criterion for a constant field through dynamical mean-field analysis. Zero-field result, external-field result, 2026 development.

A further ingredient comes from random-matrix theory: Guionnet and Maïda analyzed spherical integrals using the spectral R-transform, including behavior at a spectral edge. A spherical integral averages over continuously varying directions. The quenched Ising pressure averages the logarithm of a sum over cube vertices, so this earlier theorem does not by itself answer the present question. Spherical-integral theorem.

The claimed advance is to remove the high-temperature restriction while retaining the full overlap hierarchy and a general limiting spectrum.

4. What the new manuscripts claim

The spectral contribution

Let $\lambda_+$ be the top edge of the support of $\mu$. For $x>0$, solve

$$\int\frac{\mu(d\lambda)}{b(x)-\lambda}=x, \qquad b(x)>\lambda_+.$$

If there is no such solution, set $b(x)=\lambda_+$. Define

$$R(x)=b(x)-\frac1x, \qquad R(0)=\int\lambda\,\mu(d\lambda).$$

This is the manuscript’s edge-extended R-transform. The edge rule is part of the definition, not an assumed analytic continuation.

For a trial overlap function $p$, put

$$D_p(r)=1-r-\int_0^1(p(s)-r)_+\,ds, \qquad (u)_+=\max\{u,0\},$$

and define its spectral cost by

$$G_\mu(p)=\frac12\int_0^1 R(D_p(r))\,dr.$$

The function $D_p$ translates the overlap hierarchy into the argument seen by the spectral transform.

The single-spin contribution

A second functional $S(p)$ depends on the Ising spin space, but not on $\mu$. It is built from Gaussian averaging of the single-spin expression $\log\cosh z$.

Here is its finite-step definition. Choose a nonnegative nondecreasing step function $h$, with successive values $h_0,\ldots,h_{k-1}$ and interval lengths $w_0,\ldots,w_{k-1}$ summing to one. Set $h_{-1}=0$ and $\zeta_i=\sum_{l<i}w_l$. For a standard normal variable $Z$, define

$$(T_{a,v}f)(z)= \begin{cases} a^{-1}\log\mathbb E\exp\{a f(z+\sqrt v Z)\},&a>0,\\ \mathbb E f(z+\sqrt v Z),&a=0. \end{cases}$$

Starting with $F_{k-1}(z)=\log\cosh z$, apply these operators backwards:

$$F_{i-1}=T_{\zeta_i,h_i-h_{i-1}}F_i, \qquad \ell(h)=F_{-1}(0)-\frac{h_{k-1}}2.$$

Each step adds a Gaussian field increment and averages it at a specified level of the hierarchy. Finally,

$$S(p)=\sup_h\left\{\ell(h)+\frac12\int_0^1p(s)h(s)\,ds\right\},$$

where the supremum runs over these finite step functions. This definition gives a concrete meaning to the otherwise abstract single-spin term.

The pressure and its extensions

Theorem 1.1 claims

$$P_N\longrightarrow F(\mu),\qquad \mathbb E P_N\longrightarrow F(\mu),\qquad F(\mu)=\inf_p\{S(p)+G_\mu(p)\}.$$

The first convergence is almost sure. The Haar matrices need not be independent across dimensions. Atomic spectral laws, disconnected supports and a point-mass law are allowed; no spectral density or positive variance is required. The infimum over $p$ remains outside the supremum defining $S$: the manuscript does not interchange them.

The external-field extension adds $\sum_i c_{i,N}\sigma_i$ to the exponent. The empirical law of the deterministic fields must converge in $W_1$, the minimum expected absolute difference over couplings of two probability laws. This controls both distribution and first absolute moments. For finitely many field values, the formula first prescribes a mean spin in each field group, uses the corresponding constrained single-spin functional, and then optimizes those magnetizations. General field laws follow by $W_1$ approximation. It is a separate extension, not merely a substitution into the zero-field expression above.

For zero-field optimization, let $\mu_\beta$ be the law of $\beta\lambda$ when $\lambda$ has law $\mu$. The claimed maximum-energy limit is

$$\frac1{2N}\max_\sigma\sigma^{\mathsf T}J_N\sigma \longrightarrow \lim_{\beta\to\infty}\frac{F(\mu_\beta)}\beta.$$

This convention maximizes the quadratic reward; physical ground-state energy has the opposite sign when the Hamiltonian is minus that reward.

5. Why the advance matters

Importance: 4/5. If correct, an all-temperature formula for this broad matrix class substantially extends the thermodynamic theory beyond high-temperature regimes. It links spin-glass organization to spectral data while preserving the possibility of complicated overlap hierarchies.

The result also reaches Gaussian-pattern Hopfield interactions. If $G_N$ is an $\lfloor\alpha N\rfloor$-by-$N$ matrix of independent standard Gaussian entries, then

$$J_N=\frac cN G_N^{\mathsf T}G_N$$

is covered for every fixed real $c$. Positive and negative signs, followed by zero-temperature limits, determine respectively the largest and smallest lengths of signed sums of the Gaussian columns, with the manuscript’s normalizations. Thus the same thermodynamic formula has consequences for discrete quadratic optimization and Gaussian operator norms.

The score reflects the breadth of this extension, rather than suggesting the variational problem is automatically easy to compute or that an efficient optimization algorithm follows.

6. The main idea

The argument begins with a spectrum taking finitely many distinct values. Rotate equilibrium configurations into spectral coordinates and measure their overlaps separately inside each eigenspace. Their sum is the ordinary spin overlap.

Small perturbations enforce identities relating several replicas. Established ultrametricity and synchronization theory then reduce these many overlap arrays to compatible nondecreasing functions. Rotation identities connect the eigenspace functions to a single total-overlap function. This is the step that makes the eigenvalue law enter through the spectral transform.

For one pressure bound, add hierarchical Gaussian fields to a trial system and vary the interaction strength. The comparison is arranged to keep the supremum over fields inside the infimum over trial overlap functions. Preserving this optimization order is central to the stated formula.

For the other bound, couple systems with slightly different numbers of spins. Random directions within eigenspaces generate approximately Gaussian fields for the added spins. Evaluating this cavity increment produces the same single-spin and spectral terms. A concavity argument identifies the required field supremum.

Finally, approximate general compact spectra by finite alphabets and use concentration on the orthogonal group to turn mean convergence into almost-sure convergence. External fields require an additional constrained-magnetization argument. The common theme is to separate the randomness of directions, the spectral information and the hierarchy of spin similarities without discarding their interactions.

7. Scope, limitations, and open questions

The absence of asymptotic spectral outliers is an assumption. A vanishing fraction of eigenvalues can be invisible to the empirical law yet strongly affect the cube’s energy. The manuscript also gives an example showing that weak convergence to an unbounded-support law is not enough for the conclusion.

For random spectra, the conditional eigenvector law must be Haar. Spectral convergence in probability gives pressure convergence in probability; almost-sure spectral hypotheses give almost-sure pressure convergence. Convergence of expectations additionally requires uniform integrability. A random external field dependent on the matrix requires the corresponding Haar condition after conditioning on both spectrum and field data.

The pressure includes diagonal self-interactions. Removing them subtracts half the limiting mean eigenvalue. Physical free energy at inverse temperature $\beta>0$ is minus the chosen pressure divided by $\beta$, with the counting-prior correction included if appropriate.

An exact variational characterization need not identify its optimizer explicitly. Determining phase boundaries, fluctuation laws, computational complexity and concrete minimizing hierarchies are further questions. The thermodynamic limit is taken before the zero-temperature limit in the ground-state consequence; the manuscript does not require a uniform all-temperature finite-size estimate to obtain it.

8. Sources and further reading

M1 contains the pressure theorem and definitions in Section 1, random-spectrum and optimization consequences in Section 7, and the magnetic extension in Section 8. These conventions are essential when comparing formulas.

For overlap hierarchies, Panchenko’s ultrametricity theorem provides a central structural result. Family 222 discusses related random-perceptron variational formulas; the present manuscript compares certain Gaussian quadratic cases while treating the formulas as distinct constructions.

Source manuscripts

• M1. OpenAI, All-temperature pressure for orthogonally invariant Ising spin glasses (September 25, 2026).

Sources consulted on 11 October 2026 at repository revision fd4aeeb. Download the manuscript citations.