Limiting Law of Local Fields for the Mean Field Ghatak-Sherrington Model

Technical Summary: Limiting Law of Local Fields for the Mean Field Ghatak-Sherrington Model

Problem Statement
This paper investigates the limiting distribution of local fields in the mean-field Ghatak-Sherrington (GS) model, a multi-spin extension of the classical Sherrington-Kirkpatrick (SK) model. Unlike the SK model, where spins are binary ($\pm 1$), the GS model allows spin variables to take values in the set $\{0, \pm 1, \dots, \pm S\}$ and includes a crystal field parameter $D$. The study focuses on the high-temperature regime, defined by an inverse temperature $\beta$ sufficiently small to ensure the uniqueness of the solution to the self-consistent equations for the overlap and self-overlap.

The central object of study is the local field $l_N = \frac{1}{\sqrt{N}} \sum_{j<N} g_{jN} \sigma_j$, which represents the effective interaction experienced by a single spin from its neighbors. While the limiting behavior of local fields in the SK model is well-understood (often converging to a mixture of two symmetric Gaussians), the structure of the local field in the GS model is more complex due to the multi-spin nature and the presence of the crystal field. The paper aims to rigorously identify this limiting law and provide quantitative error bounds.

Methodology
The authors employ a combination of the cavity method, Gaussian interpolation, and the moment method. The proof strategy relies heavily on quantitative estimates of overlap and self-overlap quantities established in prior work by Sheng and Wu [19].

1. Cavity Decomposition: The Hamiltonian is decomposed by isolating the last spin $\sigma_N$. This allows the local field $l_N$ to be related to a cavity field $\ell$ defined on the remaining $N-1$ spins. The authors utilize the relation between the Gibbs measure of the $N$-spin system and the $(N-1)$-spin system to derive self-consistency relations.
2. Gaussian Interpolation: To prove the Central Limit Theorem (CLT) for the cavity field, the authors construct an interpolation path between the actual cavity field and a Gaussian random variable. They define an interpolating functional $\psi(t)$ and analyze its derivatives.
3. Overlap Concentration: A critical component of the proof is the control of overlap fluctuations. The authors utilize the quantitative bounds from Sheng-Wu [19], which state that the overlap $R_{1,2}$ and self-overlap $R_{1,1}$ concentrate around their deterministic limits $q$ and $p$ respectively, with moments decaying as $O(N^{-k})$.
4. Recursive Integral Expansion: To handle the singular factors ($t^{-1/2}$) arising in the derivative of the interpolation functional, the authors apply a recursive integral expansion technique (adapted from Chen [5]). This allows them to express the difference between the finite-$N$ expectation and the Gaussian limit as a sum of terms involving overlap fluctuations, which are then bounded using the concentration results.
5. Onsager Correction and Mixture Identification: For the local field, the authors perform a cavity decomposition to derive the Onsager correction term. They show that the limiting law is not a single Gaussian but a random finite mixture of Gaussian distributions, where the mixing weights and means depend on the random variable $\gamma_N$ (a corrected version of the Gibbs average of the local field).

Key Contributions and Results

• Theorem 1.1 (CLT for Cavity Field): The paper establishes a quantitative Central Limit Theorem for the non-centered cavity field $\ell$. Specifically, for any smooth function $U$ with moderate growth, the $2k$-th moment of the difference between the Gibbs average $\langle U(\ell) \rangle$ and the expectation of $U$ under a Gaussian distribution with mean $r$ and variance $p-q$ is bounded by $C/N^k$.
  $$E \left( \langle U(\ell) \rangle - \int_{\mathbb{R}} U(x) \phi_{r, p-q}(x) dx \right)^{2k} \leq \frac{C}{N^k}$$

• Theorem 1.2 (Limiting Law of Local Field): The main result identifies the limiting law of the local field $l_N$ as a random finite mixture of Gaussian distributions. The limiting measure $\nu_N$ is given by:
  $$d\nu_N(x) = \sum_{m=-S}^{S} p_m \phi_{\gamma_N + m\beta(p-q), p-q}(x) dx$$
  where:

  • $p_m$ are random weights depending on the crystal field and the random variable $\gamma_N$.
  • $\gamma_N = \langle l_N \rangle - \beta(p-q)\langle \sigma_N \rangle$ is the corrected cavity field average.
  • The variance of each component in the mixture is $p-q$.
  • The means are shifted by $m\beta(p-q)$, reflecting the discrete spin values.

  The paper provides a quantitative error bound for this convergence, showing that the $2k$-th moment of the difference between the Gibbs average of the local field and the integral against $\nu_N$ is also $O(N^{-k})$.

Significance and Claims
The paper claims to provide a rigorous description of the local field distribution in the GS model, extending previous results known for the SK model.

• Structural Complexity: The authors highlight a key distinction from the SK model: while the SK local field limit is a mixture of two symmetric Gaussians, the GS model yields a more involved structure—a mixture of $2S+1$ Gaussians. This complexity arises naturally from the multi-spin values and the crystal field parameter.
• Quantitative Rigor: The work provides explicit error bounds ($O(N^{-k})$) rather than just asymptotic convergence, offering a precise rate of convergence for the limiting law.
• Methodological Extension: The paper demonstrates that the cavity method, combined with the quantitative overlap estimates of Sheng-Wu [19], can be successfully extended to models with multi-spin variables and crystal fields, validating the approach for deriving TAP (Thouless-Anderson-Palmer) equations in this broader context.

The authors do not propose new experimental applications or future directions beyond the mathematical framework established. The significance is framed strictly within the context of rigorous statistical mechanics and the mathematical understanding of spin glass theory in the high-temperature regime.