🔢 mathematics

Gaussian fluctuations of differential observables of stationary lattice fields

This paper establishes a Central Limit Theorem for differential observables of stationary lattice fields on a discrete torus, proving their convergence to a centered Gaussian random distribution in a negative Sobolev space under spectral and cumulant conditions, with applications to the gradient of the Gaussian Free Field and high-temperature Ising and Potts models.

Original authors: Fabio Coppini, Wioletta M. Ruszel

Published 2026-10-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Fabio Coppini, Wioletta M. Ruszel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a vast, flat grid stretching out in every direction, like a checkerboard that never ends. On each square of this grid sits a tiny value, a number that fluctuates randomly, influenced by its neighbors but also by its own chaotic nature. In physics and mathematics, these grids model everything from the magnetic spins in a piece of iron to the height of a rough surface. For decades, scientists have known that if you take a large enough sample of these random values and average them out, the result usually settles into a predictable, bell-shaped pattern known as the normal distribution. This is the central limit theorem, a fundamental rule of probability that explains why so many natural phenomena look the same when viewed from a distance. However, this rule often breaks down when the values on the grid are strongly linked over long distances, or when the patterns are too wild to be tamed by simple averaging.

The challenge becomes even more complex when we stop looking at the raw values and start looking at how they change from one square to the next. In the physical world, this is like measuring the slope of a hill rather than its height, or the speed of a wind gust rather than the air pressure. These changes, or derivatives, are crucial for understanding forces and flows, but they are notoriously difficult to predict in a random system. When the underlying grid has long-range connections, the usual tools for predicting the behavior of these changes often fail, leaving scientists without a clear picture of what happens when the system is zoomed out to a macroscopic scale.

In a recent study, researchers Fabio Coppini and Wioletta Ruszel have provided a new, unified way to understand these fluctuations. They focused on a specific type of random grid where the values are centered around zero and do not change their statistical properties if you shift the grid over. Their goal was to determine what happens when you apply a mathematical operation that measures the rate of change across the grid, and then average those changes over a large area. They discovered that even when the raw data is too chaotic to follow standard rules, the act of measuring the change itself can act as a filter. This filtering process removes the long-range connections that cause the chaos, allowing the system to settle into a clean, predictable pattern that was previously hidden.

The researchers proved that under a wide range of conditions, the fluctuations of these differentiated observables always converge to a specific type of random noise. This noise is not the messy, unpredictable kind; it is a "white noise," a term used to describe a signal that has equal intensity at all frequencies, appearing completely random and uncorrelated from one point to the next. What makes this finding significant is that it works for systems that were previously thought to be too difficult to analyze, including those where the raw data has infinite variance. The study shows that the mathematical operator used to measure the change effectively cancels out the problematic low-frequency noise that usually prevents a clear limit from emerging.

To reach this conclusion, the team developed a method that relies on analyzing the "cumulants" of the system. In simple terms, cumulants are a way of measuring how much a group of random variables deviates from a perfect bell curve. The researchers showed that for their specific setup, these deviations become vanishingly small as the grid gets larger. They demonstrated that if the connections between distant points on the grid decay fast enough, or if the system satisfies certain conditions related to how information spreads through it, the higher-order deviations disappear. This allows the system to behave as if it were made of independent, random pieces, even though the underlying grid is full of complex interactions.

One of the most striking applications of their work is to the study of the Gaussian free field, a mathematical model often used to describe the shape of a fluctuating membrane or the interface between two phases of matter. In this model, the raw values have a property called infinite susceptibility, meaning that a tiny nudge can cause a massive reaction across the entire system. Standard averaging fails completely here. However, the researchers showed that if you look at the gradient, or the slope, of this field instead of the field itself, the infinite susceptibility is neutralized. The slope behaves like a standard white noise, a result that confirms a long-held intuition among physicists but had not been rigorously proven in this general setting before.

The study also clarifies the role of the mathematical operator used to measure the change. The researchers found that the final pattern of the noise is determined by two separate factors multiplied together. One factor comes from the specific way the change was measured, such as whether it was a simple difference or a more complex curvature measurement. The other factor comes from the underlying structure of the random field itself. This separation means that the same underlying random system can produce different types of noise depending on how it is observed, but the mechanism is always the same: the differentiation process strips away the long-range correlations, leaving behind a clean, Gaussian distribution.

The authors also provided practical criteria for determining when this behavior will occur in real-world models. They showed that for many common systems, such as the Ising model of magnetism at high temperatures or the Potts model of colorings, the necessary conditions are automatically satisfied. This means that the fluctuations of magnetization or color in these systems, when viewed through the lens of their spatial changes, will always follow the predictable white noise pattern. The researchers did not rely on special assumptions about the system being Gaussian or having a specific algebraic structure; instead, they worked directly with the statistical properties of the data, making their results applicable to a much broader class of physical systems.

By establishing this central limit theorem for differential observables, the study bridges a gap between the microscopic chaos of random fields and the macroscopic order of statistical laws. It confirms that the act of differentiation is not just a mathematical trick, but a physical mechanism that can restore order to a disordered system. The findings suggest that in many complex systems, the most meaningful signals are not found in the raw values, but in how those values change from place to place. This insight offers a new tool for physicists and mathematicians to analyze systems where traditional methods have struggled, providing a clear path to understanding the large-scale behavior of complex, interacting networks.

Technical Summary: Gaussian Fluctuations of Differential Observables of Stationary Lattice Fields

Problem Statement
This work addresses the Central Limit Theorem (CLT) for fluctuations of stationary random fields on a discrete $d$ -dimensional torus (and subsequently $\mathbb{Z}^d$ ). While classical CLTs for lattice fields typically rely on structural hypotheses such as FKG inequalities, martingale structures, mixing conditions, or explicit Gaussian/chaos representations, this paper seeks to formulate conditions directly in terms of the field's covariance and higher-order cumulants. The specific focus is on differential observables of the form:
$X^P_N(f) := N^{-d/2-m} \sum_{x \in \Lambda_N} \phi^N_x (Pf)(x/N)$
where $P$ is a scalar homogeneous constant-coefficient differential operator of order $m \ge 0$ , and $\phi^N$ is a centered stationary random field. The goal is to establish convergence in law to a Gaussian random distribution in a negative Sobolev space without assuming the field itself is Gaussian or possesses specific correlation structures like FKG.

Methodology
The authors employ a combination of Fourier analysis, the method of moments and cumulants, and a tightness argument. The proof strategy proceeds in three main stages:

Covariance Analysis: The authors derive an exact discrete Fourier representation for the covariance of the observable $X^P_N$ . By analyzing the spectral scaling limit of the covariance density $\hat{\Gamma}_N$ , they identify the limiting covariance profile. A key technical step involves handling the "aliasing" effects of the discrete Fourier transform and establishing uniform decay estimates for the discrete coefficients of smooth test functions.
Cumulant Decay: To prove Gaussianity, the paper utilizes the method of moments and cumulants. It is shown that under specific scaling assumptions, the joint cumulants of order $r \ge 3$ for the observable $X^P_N$ vanish as $N \to \infty$ . This relies on a quantitative smallness condition on the higher-order cumulants of the underlying field $\phi^N$ .
Tightness: To ensure convergence in law within the space of distributions $H^{-s}(\mathbb{T}^d)$ (or $H^{-s}_{loc}(\mathbb{R}^d)$ ), the authors prove tightness by bounding the expected squared norm of the fluctuation field in negative Sobolev spaces. This requires controlling the growth of the spectral density near zero frequency.

The paper also establishes a duality between macroscopic differential operators and microscopic difference operators. This allows the results to be interpreted as the CLT for the microscopic difference field (e.g., the gradient of a field) tested against smooth functions, even when the macroscopic observable itself might degenerate.

Key Assumptions
The main results rely on two primary hypotheses:

Assumption 2.1 (Spectral Scaling): The spectral density of the covariance, when rescaled by $N^{-2m}$ , converges to a limiting profile $\gamma(k)$ . This assumption accommodates both finite susceptibility (where $\gamma(0) < \infty$ ) and infinite susceptibility (where the spectral density is singular at the origin, but the singularity is compensated by the differential operator of order $m$ ).
Assumption 2.2 (Vanishing Higher Cumulants): The sum of the absolute values of the $r$ -th order cumulants of the field, scaled by $N^{-r(d/2+m)}$ , must vanish as $N \to \infty$ .

Main Results

General CLT (Theorem 2.3 & 2.5): Under the stated assumptions, the random distributions $X^P_N$ converge in law to a centered Gaussian random distribution $X^P$ . The limiting covariance is characterized by the multiplier $|p(k)|^2 \gamma(k)$ , where $p(k)$ is the Fourier symbol of the operator $P$ and $\gamma(k)$ is the limiting spectral profile.
- The limiting covariance separates the deterministic effect of the operator ( $|p|^2$ ) from the probabilistic infrared profile of the field ( $\gamma$ ).
- The result holds for both the torus setting and the whole-space setting ( $\mathbb{Z}^d$ ).
White Noise Limits (Corollary 2.8 & 5.8):
- If the product $|p(k)|^2 \gamma(k)$ is constant (exact cancellation of the spectral singularity by the operator), the limit is a projected white noise.
- Specifically, for the Gaussian Free Field (GFF), the gradient observable converges to a projected white noise, recovering known results for the low-temperature limit of gradient models.
- For Gibbs fields satisfying the Dobrushin uniqueness condition, the paper recovers the white-noise limits for the centered Ising magnetization and Potts colour fluctuation fields at high temperatures.
Sufficient Criteria for Cumulant Assumptions:
- Tree–Graph Bound (Proposition 5.2): The authors provide a checkable sufficient condition where the field satisfies a tree-graph bound on its cumulants. This yields an explicit decay rate for the cumulants of the fluctuation field.
- Dobrushin Uniqueness (Proposition 5.6): It is shown that the Dobrushin uniqueness condition for finite-range Gibbs fields implies the tree-graph bound for bounded single-site observables. This allows the application of the main CLT to a wide class of statistical mechanics models without separate cumulant computations for each model.
Differentiation and Correlation (Section 6): The paper analyzes the "differentiation removes correlation" paradigm. It distinguishes three regimes based on the order $m$ of the operator relative to the spectral singularity exponent $\nu$ :
- Over-differentiation ( $2m > \nu$ ): The variance vanishes; the limit is trivial.
- Exact match ( $2m = \nu$ ): The variance converges to a finite, non-zero limit; the normalization $N^{-d/2-m}$ is correct, and the limit is Gaussian (provided cumulant conditions hold).
- Under-differentiation ( $2m < \nu$ ): The variance diverges; a different normalization ( $N^{-(d+\nu)/2}$ ) is required, potentially leading to non-Gaussian limits (e.g., fractional Gaussian fields) if the field is not Gaussian.

Significance and Claims
The authors claim three novel contributions:

Distribution-Valued Limit: The CLT is established at the level of the random field (as a distribution-valued limit) rather than for a single fixed linear statistic.
Unifying Mechanism: The limiting covariance factorizes as $|p(k)|^2 \gamma(k)$ , unifying the treatment of white noise limits and gradient free fields. This factorization separates the operator's effect from the field's infrared behavior.
Quantitative Differentiation: The work makes quantitative the heuristic that differentiation suppresses long-range correlations. Each derivative multiplies the spectral density by $|k|^2$ , effectively removing the low-frequency mass responsible for singularities. This recovers the white-noise limit for the gradient of the GFF as a special case.

The methodology is described as "deliberately elementary," relying on Fourier analysis and moment/cumulant methods rather than advanced tools like Malliavin calculus or FKG inequalities. The paper emphasizes that for many models (e.g., those in the Dobrushin regime), the necessary cumulant bounds are already available via weighted dependency graphs or cluster expansions, making the distribution-valued and differentiated forms of the CLT accessible without additional structural assumptions.

Technical Summary: Gaussian Fluctuations of Differential Observables of Stationary Lattice Fields

More like this