Limit theorems for $p$-domain functionals of stationary Gaussian fields

This paper establishes central and non-central limit theorems for functionals of stationary Gaussian fields integrated over growing product domains, analyzing both separable covariance structures with quantitative bounds for Hermite polynomials and extensions to Gneiting-class or additively separable covariances.

The Gist

Imagine you are a weather forecaster trying to predict the future, but instead of looking at a single city, you are looking at a massive, complex map that stretches across space and time. This map isn't just showing temperature; it's showing a "Gaussian field," which is a fancy mathematical way of describing a random, wiggly surface where every point is connected to its neighbors.

Now, imagine you want to measure something specific about this surface, like the total volume of all the "hills" that are higher than a certain height. You do this by integrating (adding up) a function over a specific area.

This paper is about what happens to that total measurement when you expand your map. But here's the twist: you don't just expand the map equally in all directions. Maybe you stretch it 100 miles North, but only 10 miles East. Or maybe you stretch it for 10 years in time, but only 1 year in space.

The authors, Nikolai, Leonardo, Ivan, and Francesca, are asking: "If we stretch our map in these weird, uneven ways, does our measurement eventually settle down into a predictable, bell-curve pattern (like a normal distribution), or does it go crazy and become something unpredictable?"

Here is the breakdown of their findings using some everyday analogies:

1. The "Separable" World (The Lego Block Analogy)

The paper starts with the easiest scenario: Separable Covariance.

• The Analogy: Imagine your map is made of two independent Lego blocks glued together. One block represents "Space" and the other represents "Time." The randomness in Space has nothing to do with the randomness in Time. They are like two separate radio stations playing different songs; the volume of one doesn't affect the other.
• The Discovery: The authors found a beautiful "Reduction Rule." If you have this Lego-block setup, you don't need to analyze the whole giant map to know what happens. You just need to look at the individual blocks.
  • If at least one of the blocks (say, the Time block) is "well-behaved" and settles into a normal bell curve when stretched, then the entire giant map will also settle into a bell curve.
  • It's like saying: "If the music on the Time radio station is smooth, then the combined sound of Time and Space will also be smooth, even if the Space radio station is chaotic."
  • The Catch: If none of the individual blocks are well-behaved (they are all chaotic), then the whole map will be chaotic too, but in a very specific, non-Gaussian way (like a strange, jagged mountain range instead of a smooth hill).

2. The "Non-Separable" World (The Spaghetti Analogy)

Real life is rarely that neat. Often, Space and Time are tangled together. This is the Non-Separable case.

• The Analogy: Imagine the map isn't made of Lego blocks, but of a giant bowl of spaghetti. The noodles (Space and Time) are all mixed up. You can't pull one noodle out without pulling the others.
• The Discovery: In this messy world, the "Reduction Rule" from the Lego world breaks.
  • Just because the "Time" part of the spaghetti looks smooth and the "Space" part looks smooth doesn't mean the whole bowl of spaghetti will be smooth. They might interact in a way that creates a surprise.
  • The authors had to invent new rules for two specific types of "spaghetti":
    1. Gneiting Class: This is a special type of spaghetti where the noodles are tangled, but they are still "sandwiched" between two neat Lego structures. Here, the old rules mostly still work, but you have to be careful.
    2. Additively Separable: This is where the randomness is added together (Space + Time). Here, the rules change completely. The behavior depends on the speed at which you stretch the map. If you stretch Space slowly and Time quickly, the result might be one thing. If you swap the speeds, the result might be totally different. It's like mixing ingredients: adding a cup of flour slowly vs. quickly changes the texture of the dough.

3. Why Does This Matter? (The "Excursion Volume")

Why do we care about these wiggly maps?

• Real World Example: Think of a flood. You want to know the total volume of land that gets flooded above a certain water level.
• The Application: In the past, scientists usually assumed the flood zone grew equally in all directions (a perfect square expanding). But in reality, a flood might spread 100 miles down a river (Time/Length) but only 1 mile wide (Width).
• The Paper's Gift: This paper gives scientists a toolkit to predict the statistics of these "uneven" floods. It tells them: "If the river part of the flood behaves normally, your total volume prediction will be a safe bell curve. But if the river part is chaotic, you need to prepare for a wild, non-normal outcome."

Summary

• The Goal: Predict the behavior of random fields when we look at them in expanding, uneven boxes.
• The Big Win: If the field is "separable" (independent parts), you only need to check the parts. If one part is normal, the whole thing is normal.
• The Warning: If the field is "mixed" (non-separable), you can't just check the parts. The way the parts interact and the speed at which you expand the map matter immensely.
• The Takeaway: Nature is often messy and mixed. This paper provides the mathematical safety net to understand what happens when we zoom in or out on these complex, random systems in real-world, uneven ways.

Technical Summary

Here is a detailed technical summary of the paper "Limit theorems for p-domain functionals of stationary Gaussian fields" by Leonenko, Maini, Nourdin, and Pistolato.

1. Problem Statement

The paper investigates the asymptotic behavior (as $t_1, \dots, t_p \to \infty$) of additive functionals defined over growing domains for stationary, centered Gaussian fields. Specifically, the authors study the normalized functional:
$$\tilde{Y}(t_1, \dots, t_p) := \frac{Y(t_1, \dots, t_p) - \mathbb{E}[Y(t_1, \dots, t_p)]}{\sqrt{\text{Var}(Y(t_1, \dots, t_p))}}$$
where
$$Y(t_1, \dots, t_p) = \int_{t_1 D_1 \times \dots \times t_p D_p} \phi(B_x) \, dx$$
Here, $B = \{B_x\}_{x \in \mathbb{R}^d}$ is a continuous, stationary Gaussian field with unit variance, $\phi$ is a measurable function with finite second moment, and the integration domain is a product of scaled compact sets $t_i D_i \subset \mathbb{R}^{d_i}$.

Key Challenge: While limit theorems for $p=1$ (single growing domain) are well-established (e.g., Breuer-Major and Dobrushin-Major-Taqqu theorems), the behavior of functionals where different dimensions grow at potentially different rates ($p \ge 2$) is less understood. The central question is whether the asymptotic distribution of the $p$-domain functional can be deduced from the behavior of its "marginal" 1-domain functionals, and under what conditions on the covariance function this reduction holds.

2. Methodology

The authors employ a combination of Malliavin Calculus and Stein's Method, specifically leveraging the Fourth Moment Theorem by Nualart and Peccati.

• Wiener Chaos Decomposition: The functional $\phi(B_x)$ is expanded into Hermite polynomials. The functional $Y$ is decomposed into orthogonal Wiener chaoses. The asymptotic behavior is dominated by the term with the lowest Hermite rank $R$ (the first non-zero coefficient in the expansion).
• Separability Assumption: The primary analysis assumes the covariance function $C(x)$ is separable, i.e., $C(x_1, \dots, x_p) = \prod_{i=1}^p C_i(x_i)$. This allows the functional to be expressed as a product of marginal functionals in the context of Wiener-Itô integrals.
• Contraction Norms: The convergence to a Gaussian or non-Gaussian limit is determined by the behavior of the contraction norms of the kernels associated with the Wiener-Itô integrals. The authors analyze how these norms behave for product kernels.
• Quantitative Bounds: For specific cases (Hermite polynomials), the authors derive quantitative bounds on the rate of convergence using total variation ($d_{TV}$) and Wasserstein ($d_W$) distances.

3. Key Contributions and Results

A. The Separable Case

The paper establishes a powerful reduction principle for separable covariance functions.

• Central Limit Theorem (CLT) Reduction (Theorem 1):
  The normalized $p$-domain functional $\tilde{Y}$ converges to a standard Gaussian distribution $N(0,1)$ if and only if at least one of the marginal functionals $\tilde{Y}_i$ (defined on the domain $t_i D_i$) converges to $N(0,1)$.

  • Significance: This implies that if the field is "short-range dependent" in even one dimension (satisfying the Breuer-Major condition), the entire $p$-dimensional functional becomes Gaussian, regardless of long-range dependence in other dimensions.

• Non-Central Limit Theorem (Theorem 2):
  If none of the marginal functionals exhibit Gaussian fluctuations (i.e., all satisfy long-range dependence conditions with regularly varying covariance functions), the limit is non-Gaussian.

  • The limit is identified as a $p$-domain Hermite random variable, defined via a multiple Wiener-Itô integral with respect to a product measure $\nu = \nu_1 \times \dots \times \nu_p$.
  • This generalizes the classical Dobrushin-Major-Taqqu theorem to multi-dimensional, non-uniformly growing domains.

• Quantitative Improvements (Proposition 9 & Corollary 16):
  When $\phi$ is a Hermite polynomial, the authors provide explicit bounds for the distance to the Gaussian limit.

  • Improvement: Previous literature (e.g., [31]) provided bounds where the error terms were additive sums of rates from different dimensions. This paper proves that the error bound is the product of the marginal rates. This is a significant tightening of the bound, clarifying that the convergence speed is multiplicatively determined by the growth rates of the domains.

B. Non-Separable Cases

The authors extend the study to two classes of non-separable covariance functions to test the robustness of the reduction principle.

• Gneiting Class (Theorem 17):
  For Gneiting covariance functions (common in geostatistics), the covariance is "sandwiched" between two separable functions. The authors prove that the reduction principle still holds: if one marginal functional is Gaussian, the joint functional is Gaussian.
• Additively Separable Case (Theorem 18):
  For covariance functions of the form $C(x) = K_1(x_1) + K_2(x_2)$, the situation is more complex. The reduction theorem holds, but the marginal functionals to be considered are different from the standard ones. They depend on the ratio of the variances of the marginal functionals. Specifically, the limit is Gaussian if the marginal corresponding to the "dominant" growth rate (in terms of variance scaling) is Gaussian.

C. Counterexamples

The paper provides a counterexample (Example 5.1) demonstrating that without separability (or specific structural constraints like Gneiting), the Gaussianity of all marginal functionals does not guarantee the Gaussianity of the joint functional. This highlights the necessity of the structural assumptions in the main theorems.

4. Significance and Applications

1. Unification of Theory: The paper unifies the understanding of spatio-temporal and multi-domain functionals, showing that under separability, the complex $p$-dimensional problem reduces to checking simple 1-dimensional conditions.
2. Flexibility in Modeling: By allowing domains $t_i D_i$ to grow at different rates, the results apply to a wider range of physical and statistical scenarios (e.g., observing a process over a long time but a short spatial window, or vice versa).
3. Quantitative Precision: The improvement of error bounds from additive to multiplicative forms is a major technical advancement, offering sharper estimates for convergence rates in applications like fractional Brownian sheets.
4. Non-Gaussian Limits: The explicit construction of the $p$-domain Hermite random variable provides a new class of limiting distributions for fields with strong long-range dependence across multiple dimensions.

5. Conclusion

The paper fundamentally advances the limit theory for functionals of Gaussian fields by establishing that for separable covariances, the asymptotic behavior is entirely determined by the "weakest link" (for CLT) or the collective behavior (for non-CLT) of the marginal dimensions. It bridges the gap between classical 1-domain results and modern multi-dimensional applications, providing both qualitative limit theorems and quantitative convergence rates.