Maximum of sparsely equicorrelated Gaussian fields and applications

This paper investigates the extreme values of sparse, equicorrelated Gaussian fields on a triangular grid to identify the correlation threshold where the standard Gumbel law fails, utilizing the Chen-Stein method to resolve open questions in high-dimensional statistics and multiple testing.

The Gist

Imagine you are standing in a massive, crowded stadium filled with thousands of people. You want to find the tallest person in the entire crowd.

In a perfectly random crowd where everyone is unrelated, finding the tallest person is a classic problem. Statisticians know exactly how to predict the height of that tallest person; it follows a specific pattern called the Gumbel distribution (think of it as a "standard rulebook" for extreme heights).

However, real life isn't random. People in the same family are related. People in the same office are related. In this paper, the authors are studying a very specific, slightly weird kind of "stadium" where the relationships are structured like a triangular grid.

The Setup: The "Triangle of Friends"

Imagine the people in the stadium are arranged in a triangle.

• If two people sit in the same row or the same column, they are "friends" (correlated). They share a common trait, like wearing the same team jersey.
• If they are in different rows and different columns, they are strangers (independent).

The strength of this "friendship" is controlled by a dial called $r$ (correlation).

• If $r = 0$, everyone is a stranger.
• If $r$ is high, everyone in a row/column is very similar.

For a long time, statisticians believed that if this "friendship" dial ($r$) went above a certain limit (specifically, if $r > 1/3$), the standard rulebook (Gumbel distribution) would break. They thought the "tallest person" would behave unpredictably, or that the whole group would act like a single giant blob rather than a collection of individuals.

The Big Discovery: The "Broken" Rulebook is Actually Fine

The authors of this paper, Heiny, Jiang, Pham, and Qi, discovered something surprising.

They found that the standard rulebook doesn't break just because people are friends. As long as the friendship isn't too intense (specifically, as long as $1 - 2r$ is large enough), the tallest person in the crowd still behaves exactly like the tallest person in a random crowd.

The Analogy:
Imagine you are looking for the tallest person in a room full of twins.

• Old Belief: If there are too many twins, you can't find the "true" tallest person; the group just becomes a blur.
• New Discovery: Even with twins, if the twins aren't identical clones (they have some individuality), you can still predict the height of the tallest person using the old, simple rulebook. The "noise" of the friendship isn't loud enough to drown out the "signal" of the individual heights.

The "Tipping Point": When Things Get Weird

However, the authors also found what happens when you turn the friendship dial up to the absolute maximum (when $r$ gets very close to $1/2$).

At this extreme limit, the old rulebook does break. The tallest person is no longer just one individual standing out. Instead, the "tallest" value becomes a team effort.

• It's like the height of the tallest person is now determined by the sum of the two tallest people in a specific group, rather than just one super-tall outlier.
• The math changes from a simple "Gumbel" shape to a complex mix of random waves (Poisson processes) and normal curves.

Why Does This Matter? (The Real-World Applications)

This isn't just about math puzzles; it fixes problems in three major real-world areas:

1. Measuring Distances in High Dimensions (The "Cosmic Map")

• The Problem: Scientists often measure the distance between data points in massive datasets (like genes or pixels in an image). They want to know the maximum distance between any two points.
• The Fix: Previous studies said, "We can only calculate this if the data is very 'light' (has a low fourth moment)." The authors proved this restriction was unnecessary. You can now calculate the maximum distance even if the data is "heavy" or wild, as long as the underlying structure follows their rules.

2. Finding the "Biggest" Correlation (The "Social Network")

• The Problem: In finance or biology, we look at correlation matrices (who influences whom). We want to find the strongest link.
• The Fix: Previous methods required strict limits on how correlated the groups could be. The authors showed that you can find the strongest link even if the groups are highly correlated, removing a major bottleneck in statistical analysis.

3. Multiple Testing (The "Spam Filter")

• The Problem: Imagine a doctor testing 1,000 different symptoms to see if a patient has a disease. If they just pick the "most extreme" symptom, they might get a false alarm (False Discovery).
• The Fix: To avoid false alarms, you need a precise "threshold" to decide what counts as a real signal. The authors provide a new, highly accurate way to set this threshold, even when the symptoms are related (correlated). This helps doctors and researchers make fewer mistakes.

The Secret Weapon: The "Chen-Stein" Magic Trick

How did they solve this? They used a clever mathematical technique called the Chen-Stein method.

The Analogy:
Imagine you are trying to count how many times a rare bird flies over a city. The birds usually fly alone, but sometimes they fly in small flocks.

• The Chen-Stein method is like a sophisticated net that can catch these birds. It allows the mathematicians to pretend the birds are flying completely independently, even though they aren't, as long as the "flocking" behavior is weak enough.
• By using a "truncation" trick (ignoring the extreme outliers that mess up the math), they were able to prove that the "flocking" doesn't actually ruin the prediction until the flock becomes a massive, inseparable cloud.

Summary

In short, this paper tells us:

1. Don't panic about correlation: Even when data points are related, the "extreme" values (the maximums) often still follow the simple, predictable rules we already know.
2. There is a limit: If the correlation gets too strong, the rules change, and the "maximum" becomes a team effort rather than a solo act.
3. Better tools: This new understanding allows scientists to analyze complex data (like brain scans or financial markets) with more confidence and fewer restrictions than before.

They took a complex, "exotic" mathematical shape and showed that, surprisingly, it behaves like a simple, familiar shape for much longer than anyone thought.

Technical Summary

Here is a detailed technical summary of the paper "Maximum of sparsely equicorrelated Gaussian fields and applications" by Heiny, Jiang, Pham, and Qi.

1. Problem Statement

The paper investigates the asymptotic behavior of the maximum of a specific sparse and equicorrelated Gaussian field $G_n = \{G_{ij}\}_{1 \le i < j \le n}$. The correlation structure is defined as:
$$E[G_{ij}G_{kl}] = \begin{cases} 0 & \text{if } |\{i, j\} \cap \{k, l\}| = 0 \\ r & \text{if } |\{i, j\} \cap \{k, l\}| = 1 \\ 1 & \text{if } |\{i, j\} \cap \{k, l\}| = 2 \end{cases}$$
where $r \in [0, 1/2]$ is a correlation parameter.

• Context: This structure arises naturally in high-dimensional statistics, including maximum interpoint distances, sample coefficients of equicorrelated populations, and multiple testing in Gaussian graphical models.
• The Gap: Existing literature (e.g., Fan & Jiang [2019], Heiny & Kleemann [2025]) largely restricts analysis to the case $r \le 1/3$, where the maximum behaves like that of independent and identically distributed (i.i.d.) standard normal variables (converging to a Gumbel distribution). The behavior for $r > 1/3$ (specifically as $r \to 1/2$) was an open problem, particularly regarding the breakdown of the Gumbel law and the nature of the limiting distribution in the critical regime.

2. Methodology

The authors employ a sophisticated combination of probabilistic tools to handle the dependence structure:

• Chen-Stein Method for Poisson Approximation: The core of the proof relies on the Chen-Stein method to approximate the number of exceedances of a high threshold by a Poisson process. This allows the authors to derive the limiting distribution of the maximum.
• Truncation Argument: To manage the strong dependence induced by the parameter $r$, the authors introduce a carefully designed truncation level $t_n$. They decompose the Gaussian field $G_{ij}$ into a common component and an independent component:
  $$G_{ij} = \sqrt{r}(X_i + X_j) + \sqrt{1-2r}Y_{ij}$$
  where $X_i$ and $Y_{ij}$ are independent standard normals. By truncating the common variables $X_i$, they create an asymptotically independent structure that facilitates the application of the Chen-Stein method.
• Point Process Convergence: In the critical regime, the authors utilize the convergence of point processes. They show that the normalized maxima of the common components converge to a Poisson Point Process (PPP) with intensity $e^{-x}dx$. The limiting distribution of the maximum is then expressed as the supremum of points from this PPP combined with independent Gaussian variables.
• Slepian's Lemma: Used to establish bounds and compare the dependent field with fields having slightly different correlation parameters to prove convergence in specific regimes.

3. Key Contributions and Results

The paper establishes three distinct limiting regimes based on the rate at which $(1-2r)$ decays relative to $\log n$:

A. Weakly Dependent Regime (Theorem 2.1)

• Condition: $(1-2r)\sqrt{\log n} / \log \log n \to \infty$.
• Result: The maximum behaves asymptotically like the maximum of i.i.d. standard normals.
• Limiting Distribution: Standard Gumbel law ($\exp(-e^{-x})$).
• Significance: This extends the validity of the Gumbel law beyond the previously known $r \le 1/3$ threshold, showing that the "i.i.d. behavior" persists as long as the correlation is not too close to $1/2$.

B. Critical Regime (Theorem 2.2)

• Condition: $(1-2r)\log n \to \lambda \in (0, \infty)$.
• Result: The i.i.d. behavior breaks down. The limiting distribution is a mixture of a Poisson process component and a Gaussian component.
• Limiting Distribution:
  $$\sup_{i<j} \left( \frac{\eta_i + \eta_j}{\sqrt{2}} + \sqrt{2\lambda} Z_{ij} \right) - \lambda$$
  where $\{\eta_i\}$ are points of a Poisson Point Process (defined via $-\log(\sum_{k=1}^i \zeta_k)$) and $Z_{ij}$ are i.i.d. standard normals.
• Significance: This identifies the precise threshold where the Gumbel law fails and characterizes the new, more complex limiting distribution.

C. Strongly Dependent Regime (Theorem 2.3)

• Condition: $(1-2r)\log n \to 0$ (i.e., $r \to 1/2$ very fast).
• Result: The Gaussian noise term vanishes in the limit.
• Limiting Distribution:
  $$\frac{\eta_1 + \eta_2}{\sqrt{2}}$$
  This corresponds to the sum of the two right-most points of the Poisson process.
• Significance: In the extreme correlation limit, the maximum is determined entirely by the structure of the common factors (the Poisson points), and the independent noise becomes negligible.

4. Applications

The theoretical results are applied to resolve open questions in three major areas:

1. Maximum Interpoint Distance ($D_n$):

   • Problem: The asymptotic distribution of the maximum distance between $p$ points in $\mathbb{R}^n$.
   • Resolution: The authors remove the restrictive assumption $E[\xi^4] \le 5$ (which corresponds to $r \le 1/3$) required in previous works (Heiny & Kleemann [2025], Tang et al. [2022]).
   • New Finding: If the fourth moment diverges at a specific scale (related to $\log p$), the limiting distribution changes from Gumbel to the non-Gumbel forms described in Theorems 2.2 and 2.3.

2. Sample Coefficients of Equicorrelated Populations:

   • Problem: The asymptotic behavior of the largest entries in sample covariance and correlation matrices from equicorrelated populations.
   • Resolution: The authors recover and extend the results of Fan and Jiang [2019].
   • Key Improvement: They remove the technical condition $\limsup \rho < 1/2$ (which was required to keep the effective correlation $r \le 1/3$). They show that for non-Gaussian marginals with diverging fourth moments, phase transitions occur even in weakly dependent regimes, leading to non-Gumbel limits.

3. Family-Wise Error Rate (FWER) Control:

   • Problem: Controlling FWER in multiple testing under a Gaussian graphical model with sparse correlations (e.g., brain imaging data).
   • Resolution: The paper provides asymptotically exact thresholds for rejection.
   • Significance: Unlike standard union bounds which are overly conservative, the derived thresholds based on the Gumbel limit (Theorem 2.1) are precise, provided the correlation decay satisfies the weak dependence condition.

5. Significance and Impact

• Theoretical Breakthrough: The paper resolves the long-standing open problem of the maximum of Gaussian fields with correlation $r \in (1/3, 1/2]$. It precisely maps the phase transition from Gumbel to non-Gumbel limits.
• Methodological Innovation: The use of the Chen-Stein method combined with truncation arguments to handle the "triangle" correlation structure is a novel and powerful technique for high-dimensional extreme value theory.
• Practical Impact: By removing artificial moment constraints (like $E[\xi^4] \le 5$) and correlation bounds (like $\rho < 1/2$), the results make high-dimensional statistical procedures (such as distance-based tests and multiple testing corrections) more robust and applicable to real-world data where heavy tails or strong correlations may exist.
• Unification: The work unifies several disparate problems in high-dimensional statistics under a single theoretical framework, demonstrating that they all reduce to the study of this specific sparse equicorrelated Gaussian field.