The one-sixth fluctuation scale of the SK spin glass

← All digests

Family 217 · Probability and statistical mechanics
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-11

1. The question

Two pieces of a disordered magnet can have the same size and temperature but different microscopic interactions. Their equilibrium free energies will be close on a per-spin scale, yet not identical. How large are those sample-to-sample differences? After rescaling them, does a stable probability distribution emerge?

The Sherrington–Kirkpatrick, or SK, model is a central mathematical model of this situation. Every spin interacts with every other spin, with random positive or negative strengths. Its leading free energy is described by the Parisi formula, but that formula does not determine the much smaller fluctuations around the mean.

This family’s two manuscripts address the zero-field Gaussian Ising model at every fixed inverse temperature $\beta>1$. They claim that the standard deviation of the log partition function grows like $n^{1/6}$, determine a positive variance prefactor, and establish a limiting law after exact centering and standardization. The scale is much smaller than the $n^{1/2}$ fluctuations familiar from sums of independent terms.

2. Background and essential definitions

There are $n$ spins $\sigma_i\in\{-1,1\}$. For each pair $i<j$, choose an independent standard Gaussian interaction $g_{ij}$, with mean zero and variance one. In the convention used here,

$$H_n(\sigma)=\frac1{\sqrt n}\sum_{i<j}g_{ij}\sigma_i\sigma_j, \qquad Z_n(\beta)=\sum_{\sigma\in\{-1,1\}^n}e^{\beta H_n(\sigma)}.$$

The partition function $Z_n$ sums the statistical weights of all configurations. A Gibbs sample chooses $\sigma$ with probability $e^{\beta H_n(\sigma)}/Z_n$. The log partition function is $F_n=\log Z_n$; physical total free energy is $-F_n/\beta$.

There are two kinds of randomness. Thermal randomness chooses a spin configuration for fixed interactions. Disorder randomness chooses the interactions themselves. All variances and expectations of $F_n$ in the main theorem refer to disorder. They do not measure the energy fluctuations of one Gibbs sample.

Write

$$s_n^2=\operatorname{Var}(F_n), \qquad X_n=\frac{F_n-\mathbb EF_n}{s_n}.$$

The centered variable removes the exact finite-size mean. Standardization divides by the exact finite-size standard deviation. Convergence in distribution means that the probabilities described by $X_n$ approach a common limiting probability law.

For two configurations, their overlap is

$$R(\sigma,\tau)=\frac1n\sum_{i=1}^n\sigma_i\tau_i.$$

It measures similarity: it is one for identical configurations, minus one for opposite configurations, and near zero for weakly related configurations. Two independent Gibbs samples in the same disorder are often called replicas. Their overlap statistics describe the organization of equilibrium states, and are the main bridge between free energy and the Parisi hierarchy.

3. Development history and previous partial results

Sherrington and Kirkpatrick introduced the model in 1975. Parisi replaced the original single-overlap description with a hierarchy of order parameters. Guerra and Toninelli established the thermodynamic limit, Guerra proved a variational bound, and Talagrand proved the matching Parisi formula. These results identify the leading term of order $n$, leaving fluctuations on smaller scales unresolved. The original SK paper, Talagrand’s Parisi formula

The one-sixth prediction developed separately. Kondor’s replica expansion and the work of Crisanti, Paladin, Sommers, and Vulpiani led to the predicted $n^{-5/6}$ fluctuation scale for free energy per spin. Parisi and Rizzo analyzed rare free-energy deviations throughout the low-temperature regime. Inferring the width of typical fluctuations from those rare events requires an additional matching assumption; a large-deviation calculation alone does not supply a variance asymptotic. Parisi–Rizzo

At high temperature, Aizenman, Lebowitz, and Ruelle proved order-one Gaussian fluctuations. Chatterjee’s work on superconcentration showed that fluctuations can be much smaller than a basic concentration bound suggests, and connected this phenomenon to sensitivity under changes of disorder. Critical-temperature results and results with a nonzero external field treat other regimes. Their conclusions cannot simply be continued into fixed low temperature at zero field. Disorder chaos and multiple valleys

The manuscripts identify Aronow and Lopatto’s September 2026 estimates as a direct predecessor. For fixed $\beta>1$, those bounds place the variance between constant multiples of $n^{4/15}$ and $n^{7/15}$. The predicted $n^{1/3}$ lies between them. The present arguments also use Lopatto’s full-support description of the low-temperature Parisi minimizer, together with earlier uniqueness and variational results. Those inputs are substantive parts of the development, rather than consequences of the new fluctuation law. Aronow–Lopatto, Lopatto

Within this family, M2 first obtains the exponent with arbitrarily small power slack. M1 adds a quantitative comparison between different system sizes, which is what allows an actual prefactor and limiting distribution to be extracted.

4. What the new manuscripts claim

The fluctuation exponent and typical magnitude

M2 states, for every fixed finite $\beta>1$,

$$\operatorname{sd}(F_n)=n^{1/6+o(1)}, \qquad \operatorname{Var}(F_n)=n^{1/3+o(1)}.$$

It also gives a typical-magnitude statement: for every $\eta>0$,

$$\mathbb P\!\left( n^{1/6-\eta}\leq|F_n-\mathbb EF_n| \leq n^{1/6+\eta} \right)\longrightarrow1.$$

Thus the variance scale is not explained only by extremely rare samples. The two papers use slightly different deterministic normalizations of the spin sum, but multiplying $Z_n$ by $2^{-n}$ only subtracts $n\log2$, so it does not change centered fluctuations.

A positive constant and a full-sequence limiting law

M1 strengthens the claim to

$$\frac{\operatorname{Var}(F_n)}{n^{1/3}} \longrightarrow a_2(\beta)\in(0,\infty), \qquad X_n\Longrightarrow\nu_\beta.$$

The limit holds as $n$ tends to infinity through all integers. The probability law $\nu_\beta$ has mean zero, variance one, and a finite exponential moment near the origin. It is characterized by convergent finite-system moment formulas. The manuscript does not identify it as a Gaussian, Tracy–Widom, or other named distribution.

A more quantitative theorem introduces an independent Gaussian variable $\zeta_n$ of variance $\beta^2/2$ and sets

$$Y_n=n^{-1/6}\bigl(F_n-\mathbb EF_n+\zeta_n\bigr).$$

For every fixed positive integer $k$, its $k$-th moment converges to a constant $a_k$ with an error bounded by $C_k n^{-c_k}$, where $c_k>0$. Uniform exponential-moment control makes those moments determine a unique law. The added Gaussian disappears on the $n^{1/6}$ scale, so it does not change the final limit.

For physical free-energy density $-F_n/(\beta n)$, the standard-deviation scale is $n^{-5/6}$, and

$$\operatorname{Var}\!\left(-\frac{F_n}{\beta n}\right) \sim\frac{a_2(\beta)}{\beta^2}\,n^{-5/3}.$$

Its standardized limiting law is the reflection of $\nu_\beta$, because physical free energy has the opposite sign.

Two consequences

Using an earlier comparison theorem, M1 extends the same variance constant and limiting law to an added ferromagnetic term

$$\frac{\gamma}{2n}\left(\sum_i\sigma_i\right)^2$$

in the Gibbs exponent, for fixed $0\leq\gamma<1$. The perturbed model uses its own exact mean and variance. The endpoint $\gamma=1$ is not included.

The variance estimate also yields sufficient scales for disorder chaos. If the original interactions are slightly mixed with independent Gaussian interactions, equilibrium replicas from the two systems have vanishing mean squared overlap once the mixing time times $n^{2/3}$ diverges. A corresponding statement holds when a uniformly chosen fraction of edges is replaced and that fraction times $n^{2/3}$ diverges. These are averaged sufficient conditions, not a claimed sharp transition.

5. Why the advance matters

Importance: 5/5. The fluctuation problem is a central refinement of the SK thermodynamic theory. A leading free-energy formula can coexist with many possible fluctuation scales and laws. Resolving the one-sixth scale at every fixed low temperature would explain a longstanding finite-size prediction, and the limiting-law claim goes substantially further.

The positive variance prefactor also has practical mathematical value: it lets comparison results transfer the limiting behavior to a perturbed model and converts general variance inequalities into concrete disorder-sensitivity scales.

The rating concerns these consequences if the claims hold. It does not imply that the limiting distribution has an explicit density or that the same law has been established for every spin glass.

6. The main idea

Compare overlapping hierarchies of Gaussian disorder

Gaussian interpolation gradually changes the covariance of the interactions while differentiating free-energy quantities. The derivatives can be expressed through overlaps. A hierarchy lets several replicas share an initial portion of revealed disorder and then branch into independent continuations.

The Parisi minimizer supplies a scalar reference description along this hierarchy. At the thermodynamic scale, replacing the full system by that scalar description can be sufficiently accurate. At the fluctuation scale it loses too much. The manuscripts instead optimize short comparisons and track the difference between actual overlap statistics and the scalar reference.

A cavity step removes or adds one spin. Several apparently large correction terms cancel only when their signed contributions are combined before taking absolute values. Extra test replicas organize these cancellations. This is the mechanism by which the comparison retains information smaller than the leading free energy.

First find the scale, then compare sizes

M2 controls positive Laplace transforms, quantities of the form

$$\mathbb E\exp\!\left(s(F_n-\mathbb EF_n)\right).$$

They measure fluctuations through exponential tests. Transform estimates alone can miss how probability is distributed near the center, so the argument adds an auxiliary exponential variable and uses log-concavity to obtain moment and small-interval bounds. The auxiliary randomness is removed at the end. This yields both standard deviation and typical magnitude with the one-sixth exponent.

M1 then compares a system of size $N$ with every size between $N$ and $2N$, after scaling by the appropriate one-sixth power. It adjusts covariance and field parameters, changes size, and reverses the adjustments. Carefully combined pressure terms cancel the leading contributions, leaving an error with a positive power saving.

Such an error can be summed over dyadic scales $N,2N,4N,\ldots$. The resulting moment sequence is Cauchy, giving a limit rather than only bounded subsequences. Uniform exponential tails identify the probability law from its moments, while the lower fluctuation bounds ensure that the variance limit is positive.

The distinction is useful beyond this example: knowing the exponent of a sequence does not force its normalized values to converge. The cross-size comparison is the extra ingredient that rules out persistent oscillation between different subsequential behaviors.

7. Scope, limitations, and open questions

The temperature is fixed with $1<\beta<\infty$. Constants may depend on it, and neither manuscript claims uniform control as $\beta$ approaches one or infinity. The critical window and zero-temperature ground-state fluctuations require separate arguments.

The disorder is Gaussian, the spins are Ising, and the external field is zero. The results do not automatically extend to non-Gaussian interactions, mixed higher-order models, or nonzero field. Those changes can affect the fluctuation problem even when the leading thermodynamic theory has a broader formulation.

Exact finite-size centering matters. Replacing $\mathbb EF_n$ by only its leading Parisi term could leave a deterministic correction larger than the fluctuation scale. The theorem does not authorize that substitution without additional control.

The limiting law is specified through moments of finite systems, rather than a closed formula. Identifying a simpler description, determining detailed tails, and understanding universality across disorders remain natural questions beyond the stated result. The disorder-chaos corollary gives sufficient averaged perturbation scales, not matching lower bounds or a statement for each typical disorder separately.

Finally, this is an equilibrium sample-fluctuation result. The relaxation and mixing-time questions discussed in family 227 concern a different aspect of the same model.

8. Sources and further reading

M2 develops the fluctuation-scale estimates; M1 uses and strengthens that framework to obtain a positive variance constant and a limiting law. Their introductions distinguish fixed low temperature from critical-window and external-field results.

For the thermodynamic foundation, see Talagrand’s Parisi formula. For the physical fluctuation prediction, see Parisi–Rizzo. The immediate quantitative predecessor is Aronow–Lopatto, while Chatterjee explains the connection between small free-energy variance and disorder chaos.

Source manuscripts

• M1. OpenAI, The low-temperature Sherrington–Kirkpatrick free-energy limiting law (September 24, 2026).
• M2. OpenAI, The low-temperature Sherrington–Kirkpatrick fluctuation scale (September 24, 2026).

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