Extended Ghirlanda–Guerra Identities

• Extended Ghirlanda–Guerra identities are asymptotic consistency relations for overlaps in disordered Gibbs measures, enhancing classical identities with a copy-or-refresh mechanism for new replica blocks.
• The derivation employs infinitesimal perturbations, self-averaging, and Gaussian integration by parts to establish moment identities in mean-field models such as mixed p-spin and SK systems.
• These identities underpin key structural consequences including ultrametric organization, Poisson–Dirichlet statistics, and precise identification of the Parisi measure in spin glass theory.

Extended Ghirlanda–Guerra identities are asymptotic consistency relations for overlap arrays generated by replicas sampled from disordered Gibbs measures. They strengthen the classical Ghirlanda–Guerra identities by replacing moment identities for a single added replica with a copy-or-refresh mechanism that applies to arbitrary bounded test functions and, in Talagrand’s pure-state program, to blocks of newly inserted replicas. In the rigorous theory of mean-field spin glasses, these identities function as a structural engine: combined with positivity and concentration, they yield ultrametricity, Poisson–Dirichlet statistics of pure-state weights, and identification of the overlap law with the Parisi measure (Chatterjee, 13 Feb 2026). Closely related forms also appear in mixed $p$-spin models, SK models with Curie–Weiss interaction, randomized models of the Riemann zeta function, and quantum disordered systems (Panchenko, 2010).

1. Formal statement and terminology

A standard asymptotic framework is a random probability measure $G$ on the unit ball of a separable Hilbert space $H$, with replicas $\sigma^1,\sigma^2,\dots$ drawn i.i.d. from $G$ and overlaps

$R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$

In this language, the classical Ghirlanda–Guerra identity asserts that for every $n\ge2$, every bounded measurable $f=f((R_{l,l'})_{1\le l<l'\le n})$, and every bounded measurable $\psi:\mathbb R\to\mathbb R$,

$$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$

Equivalently, conditionally on the overlaps among the first $G$0 replicas, the law of $G$1 is the mixture

$G$2

where $G$3 is the law of $G$4 (Panchenko, 2011).

In Talagrand’s later formulation, the extended identities apply the same copy-or-refresh principle to a whole block of additional replicas. For every $G$5, every bounded measurable $G$6 of the overlaps among replicas $G$7, and every bounded measurable $G$8 of the overlaps between replica $G$9 and a further block of $H$0 new replicas, there is an exact limiting identity expressing

$H$1

as a combination of terms in which each new overlap is either drawn afresh or copies one of $H$2. In particular, all mixed moments

$H$3

decompose into products of lower-order factors with combinatorial weights of order $H$4; the usual one-new-replica identity is the case $H$5 (Chatterjee, 13 Feb 2026).

A stronger formulation, proved for generic mixed $H$6-spin models under positivity, states that conditionally on the overlaps among the first $H$7 replicas, the joint law of

$H$8

is the $H$9-fold product of the empirical mixture

$\sigma^1,\sigma^2,\dots$0

This form is often treated as the operational content of the extended identities in mean-field models (Panchenko, 2010).

The terminology is not perfectly uniform across the literature. In the cited papers, “extended” may refer to arbitrary bounded $\sigma^1,\sigma^2,\dots$1 and bounded $\sigma^1,\sigma^2,\dots$2, to repeated insertion of several new replicas, or to higher-order relations involving $\sigma^1,\sigma^2,\dots$3- and $\sigma^1,\sigma^2,\dots$4-overlaps. These usages are compatible at the level of strengthening the original one-replica moment identity, but they are not identical as formal statements (Chen, 2016).

2. Derivation mechanisms

The principal derivation scheme combines infinitesimal perturbations, self-averaging, and integration by parts. In Talagrand’s pure-state framework, one first adds a suitable infinitesimal, generic Gaussian perturbation to the Hamiltonian; this does not change the limit of $\sigma^1,\sigma^2,\dots$5, but it enforces exact overlap identities in the thermodynamic limit. The only structural inputs emphasized in the chapter are regularity of the covariance function $\sigma^1,\sigma^2,\dots$6 (even, convex on $\sigma^1,\sigma^2,\dots$7), existence of the free-energy limit, and the concentration estimates needed to pass to the limit under small perturbations (Chatterjee, 13 Feb 2026).

For mixed $\sigma^1,\sigma^2,\dots$8-spin Hamiltonians, Panchenko’s 2010 result identifies a direct route from differentiability of the Parisi functional to strong, non-averaged Ghirlanda–Guerra identities. If the model contains the $\sigma^1,\sigma^2,\dots$9-spin term with $G$0 or even $G$1, then

$G$2

An elementary Gaussian integration by parts then yields the $G$3-th overlap identity for the original, unperturbed Hamiltonian, with no extra averaging over perturbation parameters. By Talagrand’s positivity principle, this is upgraded to the full extended law for blocks of new overlaps (Panchenko, 2010).

A different no-averaging route, due to Chatterjee, works in a broader class of Gaussian disordered systems. If the thermodynamic free energy $G$4 exists and is differentiable at a fixed coupling $G$5, then the classical Ghirlanda–Guerra identity holds at that $G$6 without averaging over a temperature interval and without adding an external perturbation. The proof is based on convexity of the finite-volume free energy, Gaussian concentration, integration by parts, and Chatterjee’s Gaussian–Parseval variance formula (0911.4520).

Panchenko’s 2011 invariance theorem recasts the identities as an exact tilted-measure representation. Given bounded measurable functions $G$7, one defines a perturbation

$G$8

and compensators $G$9. If $R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$0 satisfies the Ghirlanda–Guerra identities, then for every bounded measurable $R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$1,

$R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$2

This representation is obtained by showing that a $R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$3-tilted generating function is constant in $R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$4, and it turns the identities into a flexible invariance principle (Panchenko, 2011).

3. Mean-field spin-glass realizations

In the mixed even-spin SK model with Curie–Weiss interaction, the Hamiltonian is

$R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$5

with centered Gaussian external field $R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$6. On a dense $R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$7-subset $R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$8 of the temperature space, and assuming $R_{l,l'}=\sigma^l\!\cdot\!\sigma^{\,l'}\in[-1,1].$9, the sequence of Gibbs measures satisfies the extended Ghirlanda–Guerra identities: $n\ge2$0 uniformly over $n\ge2$1, for every bounded continuous $n\ge2$2. When $n\ge2$3, the same identities hold with $n\ge2$4 replaced by $n\ge2$5. The proof uses concentration of the Hamiltonian per site, differentiability of the shifted SK free energy, convex analysis for the magnetization maximizer set, and Gaussian integration by parts (Chen, 2011).

That same work links the identities to two further structural facts. First, overlap positivity is established on the same dense parameter set by comparison with the pure SK model in Gaussian field. Second, the squared magnetization gives a lower bound for the overlap in the limit: $n\ge2$6 when $n\ge2$7, and the analogous statement with $n\ge2$8 holds when $n\ge2$9 (Chen, 2011).

For generic mixed $f=f((R_{l,l'})_{1\le l<l'\le n})$0-spin models satisfying the assumptions under which the Parisi formula is differentiable, Panchenko’s strong theorem removes perturbative averaging entirely. The important point is that the identities are obtained for the original Hamiltonian, not only after an auxiliary randomization. This is the setting in which the extended conditional law of a block of new overlaps as a product empirical mixture is derived most explicitly (Panchenko, 2010).

In the bipartite Sherrington–Kirkpatrick model, the free-energy density is proved to satisfy an analogue of the Parisi formula involving both the usual overlap and an additional new type of overlap. The upper bound is obtained by Guerra’s replica symmetry breaking interpolation, while the matching lower bound uses Ghirlanda–Guerra identities together with the Aizenman–Sims–Starr scheme. The same analysis reveals a phase exhibiting partial replica symmetry breaking, realized only in the larger group (Pan et al., 2018). The abstract establishes the role of Ghirlanda–Guerra machinery in this generalized mean-field setting, although the detailed form of the relevant extended identities is not specified in the supplied material.

4. Structural consequences

The central structural consequence is ultrametricity. Panchenko’s ultrametricity theorem implies that once the Ghirlanda–Guerra identities hold for arbitrarily many test functions, the limiting overlap array is almost surely ultrametric: $f=f((R_{l,l'})_{1\le l<l'\le n})$1 The chapter on Talagrand’s work emphasizes that, under the extended identities, the Gibbs measure can therefore be organized as a hierarchical tree consistent with Ruelle’s probability cascades, and the law of $f=f((R_{l,l'})_{1\le l<l'\le n})$2 is precisely the Parisi measure (Chatterjee, 13 Feb 2026). Chen’s SKFI analysis states the same implication in the mixed SK $f=f((R_{l,l'})_{1\le l<l'\le n})$3 Curie–Weiss setting, under positivity of the overlap (Chen, 2011).

Talagrand’s pure-state construction isolates a further consequence when the limiting overlap distribution $f=f((R_{l,l'})_{1\le l<l'\le n})$4 has an atom at its maximal support point

$f=f((R_{l,l'})_{1\le l<l'\le n})$5

Under the extended identities, one can construct disjoint random Borel subsets $f=f((R_{l,l'})_{1\le l<l'\le n})$6 such that the Gibbs masses $f=f((R_{l,l'})_{1\le l<l'\le n})$7 converge in distribution to the Poisson–Dirichlet law $f=f((R_{l,l'})_{1\le l<l'\le n})$8. Moreover, conditioned on two replicas lying in the same cluster $f=f((R_{l,l'})_{1\le l<l'\le n})$9, their overlap concentrates near $\psi:\mathbb R\to\mathbb R$0; and after a measurable embedding $\psi:\mathbb R\to\mathbb R$1, the push-forward $\psi:\mathbb R\to\mathbb R$2 is, up to a small remainder, purely atomic: $\psi:\mathbb R\to\mathbb R$3 with $\psi:\mathbb R\to\mathbb R$4 (Chatterjee, 13 Feb 2026).

Panchenko’s invariance representation yields additional consequences in the discrete case. If

$\psi:\mathbb R\to\mathbb R$5

is atomic with non-increasing weights, then the Gram matrix $\psi:\mathbb R\to\mathbb R$6 is weakly exchangeable under finite permutations, and $\psi:\mathbb R\to\mathbb R$7 is independent of the weights $\psi:\mathbb R\to\mathbb R$8. The same formalism yields clustering statements and explicit Poisson–Dirichlet moment identities in the pure orthonormal support case (Panchenko, 2011).

At the level of the Parisi variational theory, the extended identities, together with ultrametricity, identify the overlap distribution with the unique minimizer of the Parisi functional $\psi:\mathbb R\to\mathbb R$9. Equivalently, the cumulative distribution $$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$0 enters the Parisi PDE

$$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$1

so the overlap law is pinned down by the same order parameter that governs the limiting free energy (Chatterjee, 13 Feb 2026).

5. Extensions beyond classical Gaussian mean field

A non-spin-glass but closely parallel realization occurs in the randomized Riemann zeta model. For

$$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$2

with Gibbs measure $$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$3, Arguin and Tai prove that when $$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$4, the two-overlap distribution converges to

$$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$5

Any subsequential weak limit $$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$6 is almost surely a discrete atomic measure on the unit sphere of a Hilbert space with Poisson–Dirichlet weights of parameter $$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$7. For the limiting overlap matrix $$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$8, the exact extended Ghirlanda–Guerra identities take the form

$$E\,\langle\,f\;\psi(R_{1,n+1})\rangle \;=\;\frac1n\,E\,\langle f\rangle\,E\,\langle\psi(R_{1,2})\rangle \;+\;\frac1n\sum_{l=2}^n E\,\langle f\;\psi(R_{1,l})\rangle.$$9

The derivation uses approximate integration by parts over primes, Bovier–Kurkova telescoping, concentration, and passage to the low-temperature limit (Ouimet, 2017).

Quantum analogues exist for short-range disordered spin systems. Itoi studies a Gaussian short-range quantum model on a cubic lattice and defines overlap operators

$G$00

Theorem 1.1 gives a quantum stochastic-stability identity, and Theorem 1.2 gives a quantum Ghirlanda–Guerra identity in Duhamel expectation for all $G$01 outside a Lebesgue measure-zero exceptional set $G$02. These relations imply nontrivial sum-rules and inequalities for overlap operators and show that the same linear structure familiar from the classical case survives in the thermodynamic limit (Itoi, 2016).

A related quantum application is the transverse and longitudinal random-field Ising model. After a Lie–Trotter–Suzuki path-integral representation and an artificial Gaussian perturbation, one obtains classical Ghirlanda–Guerra identities for the path-integral overlaps $G$03 for almost every perturbation parameter $G$04. Continuity in $G$05 at $G$06, together with Harris’s inequality comparing Duhamel and ordinary expectations, yields vanishing overlap variance for the quantum overlaps $G$07 and hence absence of replica symmetry breaking in any dimension (Itoi, 2017).

Universality results show that the identities are not confined to Gaussian disorder. For mixed $G$08-spin models, the standard $G$09-overlap Ghirlanda–Guerra identities hold under i.i.d. disorder with zero mean and finite variance. Under the genericity condition that $G$10 spans a dense subspace of $G$11, and assuming the first two moments match a standard Gaussian, the finite-dimensional spin distributions converge weakly to those of the Gaussian model. As a corollary, the extended $G$12- and $G$13-overlap identities proved in the Gaussian case also hold universally under two matching moments. With an additional mild Viana–Bray dilution and three matching moments, the same universality is upgraded to total-variation convergence of finite spin marginals (Chen, 2016).

6. Scope, limitations, and failures

The literature distinguishes carefully between what is proved exactly and what is derived only after auxiliary input. Chatterjee’s no-averaging theorem establishes the classical one-new-replica identity at every differentiability point of the limiting free energy for a large class of Gaussian disordered models, including the Edwards–Anderson model, the random-field Ising model, and the SK model with Gaussian random field. However, that paper also states explicitly that this result does not by itself prove the full extended conditional-law statement

$G$14

in an almost sure sense (0911.4520). In this respect, “no averaging” and “extended” are related but not interchangeable assertions.

The cited papers also show that the classical Ghirlanda–Guerra framework is not universal across all ultrametric or overlap-based models. In flat-landscape models of natural evolution, the one-parent model and the species-formation model display non-self-averaging overlap statistics, yet the standard Ghirlanda–Guerra and Aizenman–Contucci constraints fail in the first and third models. Instead, an infinite family of new exact polynomial identities is derived from the explicit overlap distribution in the one-parent model, and numerical evidence indicates that these new identities remain accurate in the species-formation model and in genomic datasets. The homogeneous population model is replica symmetric, and in that case the classical constraints hold trivially (Agliari et al., 2023).

The mechanism of failure in those evolutionary models is identified in the paper itself: the overlap measure arises from a genealogical process rather than from a Boltzmann weight, full replica-exchangeability is lost, and the usual Ghirlanda–Guerra and Aizenman–Contucci cancellations no longer occur (Agliari et al., 2023). This suggests that ultrametric organization or non-self-averaging alone does not force the classical extended identities outside the Gibbsian spin-glass setting.

Within the rigorous spin-glass program, by contrast, the extended identities remain one of the principal bridges from thermodynamic input—free-energy differentiability, perturbative stability, and self-averaging—to the full asymptotic geometry of states: ultrametric overlap arrays, Poisson–Dirichlet cluster masses, and Parisi-order-parameter control of the Gibbs measure (Chatterjee, 13 Feb 2026).