One-sided large deviations for the ground-state energy of spin glasses

This paper establishes a large-deviation principle for the maximal energy of a spin glass with $\pm 1$ spins by deriving a Parisi-type formula for fractional moments and leveraging convex duality to show that the rate function is asymptotically quadratic near its minimum if and only if an external magnetic field is present.

The Gist

Imagine you are standing in a vast, foggy mountain range. This isn't a normal mountain range; it's a Spin Glass. In physics, a spin glass is a material where tiny magnetic atoms (spins) are frustrated—they want to align with their neighbors, but the rules of the game are chaotic, so they can't all be happy at once.

In this paper, the authors are trying to answer a very specific question about this mountain: What is the absolute highest peak (the "ground state" energy) you can find, and how likely is it that you'll find a peak even higher than the usual maximum?

Here is a breakdown of their discovery using simple analogies.

1. The Setup: The Chaotic Mountain

Think of the "spins" as hikers on a mountain. Each hiker can face North (+1) or South (-1). The landscape is determined by a complex, random map (the "Gaussian field").

• The Ground State ($L_N$): This is the highest possible altitude the hikers can collectively reach by choosing the best combination of North/South faces. In a typical scenario, there is a "standard" highest peak.
• The Question: What are the odds that, by sheer luck, the hikers find a peak that is significantly higher than this standard maximum? This is called a "Large Deviation."

2. The Main Discovery: The "Un-Inverted" Map

Usually, in physics, we calculate the "average" height of the mountain using a complex formula called the Parisi Formula. This formula is like a recipe that tells you the average height, but it's written in a confusing way (an "infimum" or a "minimum" search). It's like being told, "The highest peak is the lowest point in a valley of possibilities."

The authors did something clever. They took that confusing recipe and flipped it inside out.

• The Analogy: Imagine you have a locked box (the complex formula) that tells you the height. Instead of trying to pick the lock, they found a master key. They turned the "minimum search" into a "maximum search" (a supremum).
• The Result: They found a new, much clearer formula (Theorem 1.1) that describes the highest peak. It looks like a game where you control a "martingale" (a fair game strategy) to maximize your score. This new formula is "un-inverted," meaning it's direct and easier to work with.

3. The Big Surprise: The Role of the "Magnetic Wind"

The most exciting part of the paper is what happens when you add a magnetic field (let's call it a "wind" blowing on the hikers).

• Scenario A: No Wind ($h = 0$)
  If there is no wind, the hikers are perfectly balanced. The landscape is symmetric. The authors found that if you try to find a peak slightly higher than the average, the probability of doing so drops off extremely slowly.

  • The Metaphor: Imagine trying to push a boulder up a hill. Without wind, the hill gets steeper very gradually. The "cost" to go higher is huge, but the shape of the hill is weird—it's not a smooth curve. Mathematically, the "rate function" (the cost of being high) is not quadratic. It's flatter than a parabola near the bottom.

• Scenario B: With Wind ($h \neq 0$)
  If you add a wind (a magnetic field), it pushes all the hikers in one direction. The landscape tilts.

  • The Metaphor: Now, the hill looks like a perfect, smooth bowl (a parabola). If you want to go a little bit higher than the average, the "cost" increases in a very predictable, quadratic way (like $x^2$).
  • The Big Reveal: The authors proved that the rate function is quadratic (a perfect parabola) if and only if there is a magnetic field. If there is no field, the shape is weird and non-quadratic.

4. How They Did It: The "Fractional Moment" Trick

How did they solve this? They used a mathematical magic trick involving fractional moments.

• The Analogy: Imagine you want to know the average height of the tallest person in a crowd. Instead of measuring everyone, you measure the "average of the cubes" or the "average of the square roots" of their heights.
• The authors calculated the "fractional moments" of the partition function (a fancy way of summing up all possible energy states). By playing with these fractional powers, they could derive a formula for the "ground state" (the absolute max) by taking a limit as the power goes to zero.
• They then used Convex Duality (a mathematical tool that swaps "min" for "max") to turn their complex formula into the simple "un-inverted" version involving martingales.

5. Why Does This Matter?

This isn't just about math puzzles.

• Physics: It helps us understand how materials behave at extremely low temperatures (near absolute zero).
• Statistics: It tells us how likely extreme events are in complex, random systems.
• The "Quadratic" Connection: In statistics, a "quadratic" rate function usually means the system behaves like a Gaussian (Bell Curve). The authors showed that a magnetic field forces the system to behave "normally" (like a Bell Curve) even at the extreme edges. Without the field, the system behaves "abnormally," with fluctuations that are much harder to predict.

Summary

The authors took a messy, chaotic mountain of data (Spin Glasses), flipped the mathematical map upside down to make it readable, and discovered a simple rule: If you blow a wind (magnetic field) on the mountain, the chance of finding a super-high peak follows a smooth, predictable curve. If there is no wind, the curve is jagged and unpredictable.

They did this by inventing a new way to look at the "average" of the highest peaks, turning a difficult "minimum" problem into an elegant "maximum" problem.

Technical Summary

Here is a detailed technical summary of the paper "One-Sided Large Deviations for the Ground-State Energy of Spin Glasses" by Chen, Guionnet, Ko, Lacroix-A-Chez-Toine, and Mourrat.

1. Problem Statement

The paper investigates the large deviation principle (LDP) for the ground-state energy of spin glass models with $\pm 1$ spins (Ising spins). Specifically, it focuses on the upper tail (deviations above the typical value) of the normalized maximum energy:
$$L_N = \frac{1}{N} \max_{\sigma \in \{-1, 1\}^N} H_N(\sigma)$$
where $H_N(\sigma)$ is a Gaussian random field (the Hamiltonian) with a covariance structure determined by a function $\xi(r) = \sum_{p \ge 2} \beta_p^2 r^p$, plus a deterministic external field $h$.

The authors aim to:

1. Derive an explicit variational formula for the large deviation rate function $\Lambda^*(r)$ for $r \ge g_s$, where $g_s = \lim_{N \to \infty} \mathbb{E}[L_N]$ is the ground-state energy density.
2. Characterize the asymptotic behavior of $\Lambda^*(r)$ near its minimum ($r = g_s$), specifically determining whether it behaves quadratically (Gaussian fluctuations) or non-quadratically.

2. Methodology

The proof strategy relies on a multi-step approach connecting fractional moments, Laplace transforms, and convex duality:

• Fractional Moments and Parisi Formula: The authors first establish a Parisi-type variational formula for the fractional moments of the partition function $Z_N$ (Theorem 2.1). They consider the limit of $\frac{1}{sN} \log \mathbb{E}[Z_N^s]$ for $s \in (0,1)$. This is achieved by coupling the Hamiltonian with a Poisson–Dirichlet cascade and using Guerra's interpolation argument to derive an upper bound, and the Aizenman–Sims–Starr scheme for the lower bound.
• Laplace Transform Limit: Using the relation between fractional moments and the ground state (via $\beta \to \infty$), they derive the limit of the Laplace transform of the ground-state energy, $\Lambda(s) = \lim_{N \to \infty} \frac{1}{N} \log \mathbb{E}[e^{s N L_N}]$ (Theorem 3.1). This is expressed as an infimum over order parameters (Parisi measures) involving a solution to a Hamilton-Jacobi-Bellman type PDE (the Parisi PDE).
• "Un-Inverted" Representation: A crucial step is transforming the infimum representation of $\Lambda(s)$ into a supremum representation involving martingales (Theorem 4.2). This "un-inverted" formula expresses the limit as:
  $$\Lambda(s) = \sup_{\alpha \in \text{Mart}} \left\{ \mathbb{E}\left[h\alpha_0 + \alpha_1 \int_0^1 \sqrt{\xi''(t)} dW_t\right] + \frac{s}{2} \int_0^1 \xi''(t)(\mathbb{E}[\alpha_t^2] - t) dt \right\}$$
  subject to constraints on the martingale $\alpha$. This step utilizes convex duality arguments and the properties of the Parisi PDE.
• Large Deviation Principle: By establishing the $C^1$ regularity of $\Lambda(s)$ and applying the Gärtner-Ellis theorem (specifically a one-sided version), they deduce the LDP. The rate function $\Lambda^*(r)$ is the convex dual of $\Lambda(s)$. The supremum structure of $\Lambda(s)$ allows for an explicit computation of this dual, resulting in a variational formula for $\Lambda^*(r)$ (Theorem 5.1).

3. Key Contributions and Results

A. Explicit Rate Function (Theorem 1.2)

The paper provides an explicit variational formula for the large deviation rate function $\Lambda^*(r)$ for $r \ge g_s$:
$$\Lambda^*(r) = \inf_{\alpha \in \text{Mart}} \left\{ \frac{\left(r - \mathbb{E}[h\alpha_0 + \alpha_1 \int_0^1 \sqrt{\xi''(t)} dW_t]\right)^2}{2 \int_0^1 \xi''(t)(\mathbb{E}[\alpha_t^2] - t) dt} \right\}$$
where the infimum is taken over bounded martingales satisfying specific constraints related to the order parameter. This formula is significantly more explicit than standard convex duals derived from free energy formulas.

B. Ground State Energy Formula (Theorem 1.1)

As a byproduct, the authors derive a "un-inverted" formula for the ground-state energy $g_s$ (the case $s=0$):
$$g_s = \sup_{\alpha \in \text{Mart}} \left\{ \mathbb{E}\left[h\alpha_0 + \alpha_1 \int_0^1 \sqrt{\xi''(t)} dW_t\right] \mid \int_0^1 \xi''(t)(\mathbb{E}[\alpha_t^2] - t) dt \ge 0 \right\}$$
This confirms and extends previous heuristic predictions regarding the structure of the ground state.

C. Quadratic vs. Non-Quadratic Behavior (Theorem 1.3)

The most significant physical insight concerns the behavior of $\Lambda^*(r)$ near the minimum $g_s$:

• Case $h \neq 0$ (External Field Present): The rate function is quadratic near $g_s$. Specifically, $C^{-1}(r-g_s)^2 \le \Lambda^*(r) \le C(r-g_s)^2$. This implies that the ground-state energy fluctuations are of order $N^{-1/2}$ and follow a Gaussian distribution (Central Limit Theorem behavior).
• Case $h = 0$ (No External Field): The rate function is non-quadratic near $g_s$. Specifically, $\lim_{r \to g_s^+} \frac{\Lambda^*(r)}{(r-g_s)^2} = +\infty$. This indicates that the fluctuations are of order smaller than $N^{-1/2}$ (sub-Gaussian), consistent with the physics literature suggesting orders like $N^{-3/4}$ or $N^{-5/6}$ for the SK model.

The paper proves that the quadratic behavior holds if and only if $h \neq 0$. This is linked to the support of the Parisi measure: if $h=0$, the support includes 0, causing the denominator in the rate function formula to vanish for the optimal martingale, leading to the singularity.

4. Significance

• Rigorous Validation of Physics Predictions: The results rigorously confirm long-standing physics predictions (derived via non-rigorous replica methods) regarding the scaling of ground-state fluctuations in spin glasses. It resolves the debate on whether fluctuations are Gaussian ($h \neq 0$) or non-Gaussian ($h=0$).
• New Analytical Tools: The introduction of "un-inverted" formulas (supremum over martingales) for the ground-state energy and its large deviations offers a new perspective on spin glass theory, potentially applicable to multi-type spin glasses with non-convex covariance structures where classical Parisi formulas are difficult to analyze.
• Connection to Random Matrix Theory: The distinction between $h=0$ and $h \neq 0$ mirrors the behavior of the largest eigenvalue of GOE matrices (spherical SK model), where the presence of a field changes the fluctuation scaling from Tracy-Widom ($N^{-2/3}$) to Gaussian ($N^{-1/2}$). This paper extends such insights to Ising spin glasses.
• Methodological Advance: The technique of using fractional moments to access the ground state and then leveraging convex duality to switch between infimum (Parisi) and supremum (martingale) representations provides a robust framework for studying large deviations in disordered systems.

In summary, this paper provides a rigorous mathematical foundation for the large deviation behavior of spin glass ground states, explicitly characterizing the rate function and proving that the presence of an external field is the necessary and sufficient condition for Gaussian fluctuations.