A hierarchical formula for the free energy of diluted spin glasses

← All digests

Family 221 · Probability and statistical mechanics
Importance 4/5 · Provisional editorial rating
Sources consulted 2026-10-11

1. The question

In a dense spin glass, each spin interacts with a number of others that grows with the system. In a diluted spin glass, each spin participates in only a finite average number of interactions. The system remains locally irregular no matter how large it becomes.

The cavity method tries to understand such a system by adding one spin and asking how its neighbors influence it. A single effective field may suffice in a simple phase. When many competing states contribute, the proposed description becomes a hierarchy of probability distributions over fields.

The manuscript in this family claims that this hierarchical prescription gives the exact limiting pressure for a specified class of diluted Ising models. Earlier work had established that every finite hierarchy gives an upper bound. The new assertion is that optimizing over all finite depths and all trial laws reaches the actual value.

This is a statement about a variational formula, not a claim that one small hierarchy always suffices or that an optimizer can be computed efficiently.

2. Background and essential definitions

There are $N$ spins $\sigma_i\in\{-1,1\}$. Fix an even interaction size $p\geq2$ and a density $\alpha>0$. The number of interactions is a Poisson random variable $M_N$ with mean $\alpha N$. Each interaction chooses $p$ site indices independently and uniformly with replacement.

Let $\theta_k$ be independent copies of a random function on $p$ spins, and let $h_i$ be independent external fields. The statistical weight is

$$\exp\!\left( \sum_{k=1}^{M_N}\theta_k(\sigma_{i_{1,k}},\ldots,\sigma_{i_{p,k}}) +\sum_{i=1}^N h_i\sigma_i \right).$$

The partition function $Z_N$ sums these weights over all $2^N$ configurations. The pressure is the expected log partition function per spin,

$$F_N=\frac1N\mathbb E\log Z_N.$$

The expectation averages the disorder. If inverse temperature $\beta$ has been included in $\theta$ and $h$, physical free energy per spin is $-F_N/\beta$. This sign explains why an upper pressure bound is often called a lower free-energy bound in the literature.

A message $x\in\mathbb R$ represents a one-spin probability distribution

$$q_x(s)=\frac{e^{xs}}{2\cosh x},\qquad s\in\{-1,1\}.$$

Its mean spin is $\tanh x$. The cavity approximation replaces the influences from different neighboring branches by messages. A random message law captures variation among local environments. A hierarchy adds distributions of message laws, then distributions of those distributions, allowing nested variation among competing states.

The simplest, depth-zero prescription is called replica symmetric. Positive depth incorporates replica symmetry breaking. Here replicas simply mean independent spin samples from the same Gibbs system; the hierarchy is a trial construction rather than an assumed literal decomposition of the original finite system.

3. Development history and previous partial results

The Viana–Bray model supplied a basic diluted two-spin glass. Mézard and Parisi developed a hierarchical cavity description for the Bethe-lattice spin glass, including a one-step symmetry-breaking functional. Their work explained why a distribution of local fields must itself fluctuate across states. The Bethe lattice spin glass revisited

Franz and Leone adapted Guerra’s interpolation method to diluted systems, obtaining replica-symmetric and one-step bounds for even-arity models. Panchenko and Talagrand then established arbitrary finite-level hierarchical bounds for the interaction class used here. Their conjecture asked whether the infimum of those bounds equals the limiting pressure. Panchenko–Talagrand

Other exact variational descriptions already existed. De Sanctis used random multioverlap structures for Viana–Bray models. Panchenko’s spin-distribution approach gave exact formulas over invariant exchangeable spin distributions. These are broader ways of recording the state of the system; an exact formula over such a class does not automatically reduce to the specific finite hierarchy of independent message laws. Spin distributions

Later results clarified pieces of the hierarchy. Panchenko obtained single-site hierarchical representations for finite-symmetry-breaking asymptotic measures of suitably modified models. Other work proved exact replica-symmetric formulas in high-temperature or low-connectivity regimes, and developed variational bounds for random regular graphs.

The remaining issue addressed by this manuscript is the equality for the Poisson model at arbitrary fixed interaction density and temperature, within the stated factorization and positivity class. The distinction between Poisson and fixed-degree graph ensembles is important: the historical Bethe-lattice motivation does not make the theorem an assertion about every regular factor graph.

4. What the new manuscripts claim

The precise interaction class

Write

$$B(\theta)=\max_{s\in\{-1,1\}^p}|\theta(s)|.$$

The integrability assumption is only

$$\mathbb E B(\theta)+\mathbb E|h|<\infty.$$

The interaction must admit the factorization

$$e^{\theta(s_1,\ldots,s_p)} =a\left(1+b\prod_{\ell=1}^p f_\ell(s_\ell)\right),$$

where $a>0$, the $f_\ell$ are independent copies of one random function, and $b$ is independent of them. For every spin tuple, the product term has absolute value strictly below one. The further positivity requirement is

$$\mathbb E[(-b)^k]\geq0\qquad(k=1,2,\ldots).$$

These are substantive restrictions. The theorem is not stated for an arbitrary random interaction on a sparse graph. The multiplier $a$ may depend on the other variables, and no separate moment assumption on $\log a$ is imposed.

The variational equality

For clarity, first describe the cavity terms. If $p$ independent message spins enter one interaction, its average weight is

$$C_\theta(x_1,\ldots,x_p) =\sum_{s_1,\ldots,s_p}e^{\theta(s_1,\ldots,s_p)} \prod_{\ell=1}^p q_{x_\ell}(s_\ell).$$

For a new spin fixed to $s$, let $U_\theta(x_1,\ldots,x_{p-1};s)$ be the logarithm of the same average over the other $p-1$ spins. A new site sees a Poisson number $K$ of interactions with mean $\alpha p$, giving

$$V_{\mathrm{site}}= \frac12\sum_{s=\pm1} \exp\!\left(hs+\sum_{j=1}^K U_{\theta_j}(x_{j1},\ldots,x_{j,p-1};s)\right).$$

The interaction correction uses $V_{\mathrm{edge}}=C_\theta(x_1,\ldots,x_p)$. Its role is to remove the excess counting caused by assigning interactions to their incident sites.

At depth $r$, choose hierarchy parameters $0<m_1<\cdots<m_r<1$ and a law $\zeta$ for the nested message distributions. Starting from the deepest message level, average a positive quantity through successive conditional power means

$$W\longmapsto\bigl(\mathbb E[W^{m_j}\mid\text{earlier message levels}]\bigr)^{1/m_j}.$$

Call the complete operation $\mathcal T_0$. Physical interaction and field disorder is held fixed during these message averages. Different message labels have independently drawn hierarchical roots.

The trial functional is

$$\mathcal B_r(\zeta,m)= \log2+\mathbb E\log\mathcal T_0V_{\mathrm{site}} -\alpha(p-1)\mathbb E\log\mathcal T_0V_{\mathrm{edge}}.$$

At depth zero, omit the power-mean operations and use independent ordinary messages. The manuscript’s main theorem is

$$\lim_{N\to\infty}F_N =\inf_{r\geq0}\ \inf_{\zeta,m}\mathcal B_r(\zeta,m).$$

The limit exists. The infimum ranges over finite depths, but does not require that any finite depth or trial law attain the optimal value. No moment condition is needed on the trial messages themselves.

Examples and zero temperature

The class includes diluted even-spin interactions $\theta=\beta J\prod_\ell s_\ell$ with symmetric coupling $J$, and independent external field $\beta H$, assuming $\mathbb E|J|+\mathbb E|H|<\infty$. The case $p=2$ is Viana–Bray.

It also includes soft random even-K SAT: each violated clause incurs a finite nonnegative energy penalty. Independent nonnegative clause weights with finite mean are allowed. “Soft” means that violations are penalized rather than forbidden outright.

For these named models, the manuscript obtains the limiting expected optimal energy by sending $\beta\to\infty$ after the pressure limit. The error between pressure divided by $\beta$ and the expected optimal score per spin is at most $\log2/\beta$. For unit-weight SAT, this yields the limiting expected minimum violation density.

5. Why the advance matters

Importance: 4/5. The claim turns an established family of hierarchical bounds into an exact thermodynamic formula for a broad class of sparse disordered models. It would justify a central cavity prescription beyond the regimes where a single message distribution is already known to suffice.

The first-moment scope is useful: the result permits unbounded interaction and external-field laws as long as their mean absolute sizes are finite. The zero-temperature consequence connects equilibrium pressure with optimization on random sparse instances.

The rating also reflects the theorem’s boundaries. It identifies a variational value for a specific even-arity Poisson class, without solving the optimization explicitly or covering all sparse constraint models.

6. The main idea

The previously known interpolation inequality supplies one side: every admissible trial functional lies above the pressure. The new task is to extract trial messages from a large physical system whose value comes arbitrarily close to that pressure.

Retain more than pairwise overlaps

For replicas indexed by a finite set $S$, define the multioverlap

$$R_S=\frac1N\sum_{i=1}^N\prod_{a\in S}\sigma_i^a.$$

Pair overlaps use sets of size two. A diluted interaction can depend on a whole pattern of replica spins at a site, and the probabilities of those patterns involve multioverlaps of every fixed finite order. Pairwise information alone is therefore insufficient.

The manuscript adds a sublinear number of bounded auxiliary perturbations, so their contribution per spin vanishes. These perturbations yield identities for replicas with prescribed branching patterns in a finite auxiliary hierarchy. The hierarchy is introduced as a tool; no ultrametric organization of the physical Gibbs states is assumed.

Turn a reservoir into independent messages

A large Gibbs system serves as a reservoir for comparing the partition functions before and after one spin is inserted. Concentration identities control multioverlaps averaged over branching depths. Moving branching points one level later gives an induction that controls the necessary conditional fluctuations.

The control need not be uniform over every possible tree. The cavity expansion averages over finite trees, and the exceptional branching patterns receive vanishing total weight. This matches the estimate to the quantity actually being computed.

At a uniformly sampled reservoir site, successive conditional laws encode a hierarchical message. Multioverlap concentration then permits the message roots attached to different sites to be redrawn independently without losing the cavity increment. The resulting object belongs to the prescribed Panchenko–Talagrand trial class and nearly achieves the pressure.

Remove boundedness without changing the target

The construction first handles bounded interactions. A scalar shift removes the potentially troublesome multiplier $a$ from the factorization while preserving the relevant pressure difference. Truncation and estimates uniform over all trial laws then pass to first-moment interactions and fields.

The order of limits matters: system size tends to infinity at each fixed hierarchy depth, and only afterward is the depth increased. This establishes equality of variational values without constructing one infinite-depth Gibbs hierarchy or asserting attainment.

7. Scope, limitations, and open questions

Even arity is part of the theorem. In particular, the soft-SAT application does not cover random 3-SAT. The manuscript explicitly distinguishes the separate companion formula for that model, whose trial messages have a different form.

Interactions are counted by a Poisson variable, and their site positions are sampled with replacement. The result should not be silently transferred to every fixed-degree graph ensemble or every conditioned hypergraph model.

The formula is exact but not automatically an effective numerical algorithm. It does not bound the depth needed for a prescribed accuracy, prove existence of an optimal finite trial law, or identify which parameters exhibit a particular number of symmetry-breaking levels.

The zero-temperature conclusion concerns an expected optimal energy density. A vanishing minimum violation density does not itself determine whether a random SAT instance is completely satisfiable with high probability. Exact thresholds require additional information about rare violations and finite-instance probabilities.

8. Sources and further reading

The source manuscript states the interaction assumptions, full power-mean functional, equality theorem, and named-model applications. Its main conceptual distinction is between an exact formula over general spin distributions and reduction to the specific finite hierarchical message class.

Read Mézard–Parisi for the cavity proposal, Panchenko–Talagrand for the hierarchical bounds and equality question, and Panchenko’s spin-distribution formulation for the broader variational viewpoint that precedes the new reduction.

Source manuscripts

• M1. OpenAI, The Mézard–Parisi formula for diluted spin glasses (September 23, 2026).

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