Fine asymptotics of the magnetization of the annealed dilute Curie-Weiss model

This paper establishes sharp cumulant bounds for the magnetization in the annealed dilute Curie-Weiss model under high-temperature conditions with an external magnetic field, thereby proving a central limit theorem with convergence rates, a moderate deviation principle, concentration inequalities, and mod-Gaussian convergence for the regime where $p^3 N^2 \to \infty$.

The Gist

Imagine a giant room filled with N people, each holding a sign that says either +1 (Happy) or -1 (Sad). This is a simplified model of a magnetic material, where the people are "spins" and their signs represent magnetic directions.

In the classic version of this game (the Curie-Weiss model), everyone in the room can see and influence everyone else instantly. If the room is hot (high temperature), everyone is chaotic and random. If it's cold, they all try to agree and pick the same sign.

The New Twist: The "Dilute" Game

In this paper, the authors introduce a twist: The Dilute Curie-Weiss Model.

Imagine the room is now a giant, invisible network (like a social media graph). Two people only influence each other if they are "friends" (connected by an edge). In this specific game, the friendship network is random: any two people become friends with a certain probability, p.

• If p = 1, everyone is friends with everyone (the classic model).
• If p is small, most people are strangers, and the room is "diluted" with empty space between connections.

The authors are studying what happens when the room is hot (high temperature) and there is a magnetic field (an external force, like a loudspeaker shouting "Be Happy!") pushing everyone toward the +1 side.

The Big Question

When you have a huge crowd of people, some acting randomly and some being pushed by a loudspeaker, what does the total mood of the room look like?

• Is it chaotic?
• Does it settle into a predictable pattern?
• How fast does it settle?

The authors prove that even with this messy, random network of friendships, the total mood behaves very nicely. It follows a Bell Curve (the famous "Normal Distribution"). But they didn't just stop at saying "it looks like a bell curve." They wanted to know exactly how close it is to that curve and how fast it gets there as the room gets bigger.

The Secret Weapon: "Cumulants"

To get these precise answers, the authors used a mathematical tool called Cumulants.

• Analogy: Think of the total mood as a song. The "average" mood is the main melody. The "variance" is the volume.
• Cumulants are like the harmonics or the subtle overtones in the music.
  • The first few harmonics tell you if the song is a perfect bell curve.
  • If the higher harmonics (cumulants) are tiny, the song is a perfect bell curve.
  • If they are loud, the song is distorted.

The authors calculated these "harmonics" for their random network and proved they get extremely small very quickly as the room size ($N$) grows.

The Main Discovery: The "Speed Limit"

The paper proves a specific condition (called the Statulevičius condition) that acts like a speed limit for the chaos.

• They found that as long as the network isn't too sparse (specifically, if $p^3 N^2$ gets very large), the "harmonics" die out fast enough.
• What this means: The total mood of the room becomes a perfect Bell Curve very quickly.
• The Result: Because they proved the harmonics are so small, they could derive a whole list of "bonus" results:
  1. How close is the curve? They gave a precise formula for the error (how far off the real mood is from the perfect Bell Curve).
  2. Rare Events: They calculated the odds of the room having a wildly unusual mood (like everyone suddenly flipping to Sad against the loudspeaker).
  3. The "Cramér Correction": They found a way to tweak the Bell Curve formula to make it even more accurate for extreme cases.

Why the "Dilute" Part Matters

In the past, scientists knew this worked for the "everyone is friends" model. But real-world systems (like neurons in a brain or people on a social network) aren't fully connected; they are sparse and random.

This paper bridges the gap. It shows that even if you randomly remove most of the connections between people, as long as you don't remove too many, the system still behaves predictably and follows the same laws as the fully connected one.

The "Magic" Math Trick

To do this, the authors had to use some advanced math (Saddle-Point Method and Complex Analysis).

• The Metaphor: Imagine trying to find the highest point on a foggy mountain range (the "partition function"). Usually, you just walk up the hill. But because the network is random, the landscape is bumpy and foggy.
• The authors had to look at the mountain from a different dimension (using complex numbers) to see the "saddle point" (the perfect balance point) clearly. They proved that even with the fog (randomness), the saddle point is stable and predictable, allowing them to calculate the exact shape of the mood distribution.

Summary

In simple terms: Even in a chaotic, randomly connected crowd, if the temperature is high and there's a gentle push from an external force, the group's collective behavior becomes incredibly predictable and follows a perfect bell curve. The authors didn't just say "it's a bell curve"; they provided the exact recipe for how perfect that curve is and how to calculate the odds of any specific outcome, no matter how rare.

Technical Summary

Here is a detailed technical summary of the paper "Fine asymptotics of the magnetization of the annealed dilute Curie-Weiss model" by Apostel, Döring, and Schubert.

1. Problem Statement and Model Definition

The paper investigates the dilute Curie-Weiss model on a directed Erdős-Rényi graph $G(N, p)$, specifically focusing on the annealed Gibbs measure.

• Model Setup:

  • Graph: A directed Erdős-Rényi graph with $N$ vertices where each edge $(i, j)$ exists independently with probability $p = p(N)$. Self-loops are allowed.
  • Spins: Variables $\sigma_i \in \{-1, 1\}$ for $i \in \{1, \dots, N\}$.
  • Hamiltonian: The energy of a configuration $\sigma$ is given by:
    $$H_{N,\varepsilon,\beta,h}(\sigma) := -\frac{\beta}{2pN} \sum_{i,j=1}^N \varepsilon_{ij}\sigma_i\sigma_j - h \sum_{i=1}^N \sigma_i$$
    where $\beta > 0$ is the inverse temperature, $h \in \mathbb{R}$ is the external magnetic field, and $\varepsilon_{ij}$ are i.i.d. Bernoulli variables indicating edge presence.
  • Annealed Measure: Unlike the quenched setting (where disorder is fixed), the authors consider the annealed measure, obtained by averaging the Gibbs measure over the disorder $\varepsilon$:
    $$\mu_{N,p,\beta,h}(\sigma) = \frac{\mathbb{E}_\varepsilon [e^{-H_{N,\varepsilon,\beta,h}(\sigma)}]}{\mathbb{E}_\varepsilon [Z_{N,\varepsilon,\beta}(h)]}$$
  • Regime: The study focuses on the high-temperature regime ($0 < \beta < 1$) with an external magnetic field$h \in \mathbb{R}$.
  • Density Condition: The graph is in the "dense" regime where $pN \to \infty$. Crucially, the authors impose a stronger condition $p^3 N^2 \to \infty$ to ensure their analytical techniques hold. This condition is necessary to control the effective temperature parameter derived from the disorder averaging.

• Objective: To derive sharp bounds on the cumulants of the total magnetization $M_N = \sum \sigma_i$ under the annealed measure. These bounds are used to establish quantitative limit theorems beyond the standard qualitative Central Limit Theorem (CLT).

2. Methodology

The authors employ a combination of probabilistic techniques and complex asymptotic analysis.

A. Reduction to a Mean-Field Model

The core technical innovation is transforming the dilute model into an effective classical Curie-Weiss model with $N$-dependent parameters.

1. Hubbard-Stratonovich Transform: By averaging the disorder $\varepsilon_{ij}$ first, the partition function $Z_{N,p,\beta}(h)$ is rewritten. Using the identity $(1-p) + p e^x = a e^{b x}$, the authors derive parameters $a$ and $b$ (dependent on $N$ and $\beta$).
2. Effective Parameters: The annealed partition function becomes:
   $$Z_{N,p,\beta}(h) = a^{N^2} \sum_{\sigma} \exp\left( \frac{b}{2pN} M_N(\sigma)^2 + h M_N(\sigma) \right)$$
   This reveals that the dilute model behaves like a classical Curie-Weiss model with an effective inverse temperature $\beta_{\text{eff}} = b/p$.
3. Asymptotics of $\beta_{\text{eff}}$: The authors prove that under the condition $p^3 N^2 \to \infty$, $\beta_{\text{eff}} \to \beta$. The error term is of order $O(\beta^3 / (p^3 N^2))$.

B. Integral Representation and Saddle-Point Method

To analyze the cumulants, the authors study the finite-volume pressure $\psi_{N,p,\beta}(h) = \frac{1}{N} \log Z_{N,p,\beta}(h)$.

1. Integral Form: Using the Hubbard-Stratonovich transform again, the partition function is expressed as an integral:
   $$Z_{N,p,\beta}(h) \propto \int_{-\infty}^{\infty} \exp\left( N \Phi_N(s, h) \right) ds$$
   where $\Phi_N(s, h) = -\frac{s^2}{2\beta_{\text{eff}}} + \log(2\cosh(h+s))$.
2. Complex Analysis: Since cumulants are derivatives of the pressure, and the authors need bounds for high-order derivatives, they must extend the analysis to the complex plane (specifically, a strip around the real axis).
3. Saddle-Point Analysis: The standard Laplace method is insufficient for complex $h$. The authors use the saddle-point method:
   • They identify a unique holomorphic saddle point $s_N(h)$ solving $\partial_s \Phi_N(s, h) = 0$.
   • They prove that $s_N(h)$ exists and is unique in a complex strip $U_N$ that converges to a limiting strip $U$ as $N \to \infty$.
   • They shift the integration contour to pass through $s_N(h)$ and apply Taylor expansions to derive the asymptotic behavior of the integral.

C. Cumulant Bounds via Cauchy's Estimate

1. Relation to Pressure: The $j$-th cumulant $\kappa_j(M_N)$ is related to the $j$-th derivative of the pressure: $\kappa_j(M_N) = N \cdot \frac{d^j}{dh^j} \psi_{N,p,\beta}(h)$.
2. Cauchy's Estimate: By establishing that the pressure $\psi_{N,p,\beta}(h)$ converges locally uniformly to a holomorphic limit function $\Phi(s(h), h)$ on a complex strip, the authors apply Cauchy's integral formula to bound the derivatives.
3. Statulevičius Condition: They prove that the standardized magnetization $m_N$ satisfies the Statulevičius condition:
   $$|\kappa_j(m_N)| \leq C j! \left( \frac{R}{\sqrt{N}} \right)^{j-2}$$
   for $j \geq 3$, where $R$ is related to the distance to the nearest Lee-Yang zero of the limiting partition function.

3. Key Contributions

1. Sharp Cumulant Bounds for Annealed Dilute Models: The paper provides the first quantitative cumulant bounds for the annealed dilute Curie-Weiss model in the high-temperature regime with an external field. Previous results were mostly qualitative (CLT only) or restricted to the quenched setting.
2. Extension of Classical Results: The results cover the classical Curie-Weiss model ($p=1$) as a special case, establishing quantitative consequences of the Statulevičius condition that were previously missing in the literature for the classical model.
3. Condition $p^3 N^2 \to \infty$: The authors rigorously justify the necessity of this specific density condition. They show that if $p^3 N^2$ is constant, the effective temperature $\beta_{\text{eff}}$ does not converge sufficiently fast to $\beta$, causing a change in behavior that breaks the standard saddle-point analysis.
4. Complex Analytic Techniques: The successful application of the saddle-point method to the annealed partition function in the complex plane, handling the $N$-dependent parameters and ensuring local uniform convergence, is a significant technical achievement.

4. Main Results

Under the assumptions $0 < \beta < 1$,$h \in \mathbb{R}$, and$p^3 N^2 \to \infty$, the standardized magnetization$m_N$ satisfies the following:

1. Quantitative Central Limit Theorem (Berry-Esseen):
   The Kolmogorov distance between the distribution of $m_N$ and the standard normal distribution $Z$ is bounded by $O(1/\sqrt{N})$.
   $$\sup_{x \in \mathbb{R}} |P(m_N \leq x) - P(Z \leq x)| \leq \frac{c}{\sqrt{N}}$$

2. Normal Approximation with Cramér Corrections:
   For $x$ in a moderate range ($0 < x < c\sqrt{N}$), the tail probability admits an expansion involving the Cramér-Petrov series, providing higher-order accuracy than the standard CLT.

3. Concentration Inequality:
   A sharp exponential bound on the tail probabilities of the magnetization is derived, showing sub-Gaussian behavior with specific constants depending on $\beta$ and $h$.

4. Moderate Deviation Principle (MDP):
   For sequences $a_N \to \infty$ with $a_N = o(\sqrt{N})$, the sequence $m_N/a_N$ satisfies an MDP with speed $a_N^2$ and rate function $I(x) = x^2/2$.

5. Mod-Gaussian Convergence:
   The random variable $Y_N$ (a rescaled version of $m_N$) converges mod-Gaussian. This implies the existence of a limiting function $\phi(z)$ such that the characteristic function of $Y_N$ behaves like a Gaussian times $\phi(z)$. This is a strong form of convergence that implies the MDP and local limit theorems.

5. Significance

• Bridging Disordered and Mean-Field Systems: The work demonstrates that despite the disorder introduced by the random graph, the annealed dilute Curie-Weiss model retains the fine asymptotic structure of the classical mean-field model, provided the graph is sufficiently dense ($p^3 N^2 \to \infty$).
• Quantitative Precision: By moving beyond qualitative limit theorems to quantitative bounds (rates of convergence, concentration inequalities, mod-Gaussian convergence), the paper provides tools necessary for understanding finite-size effects in statistical physics models on random graphs.
• Methodological Advancement: The rigorous handling of the annealed measure via complex saddle-point methods sets a precedent for analyzing other disordered systems where the disorder is averaged out, particularly in regimes where the interaction strength depends on system size.
• Lee-Yang Zeros Connection: The radius of convergence for the cumulant bounds is explicitly linked to the distance of the Lee-Yang zeros of the limiting partition function, reinforcing the deep connection between analytic properties of the partition function and probabilistic limit theorems.

In summary, this paper establishes a rigorous framework for understanding the fluctuations of magnetization in annealed dilute spin systems, proving that they obey the same fine-grained asymptotic laws as their classical counterparts, provided the graph density exceeds a specific threshold.