Quadratic form of heavy-tailed self-normalized random vector with applications in $\alpha$-heavy Mar\v cenko--Pastur law

This paper establishes that the asymptotic distribution of quadratic forms for self-normalized heavy-tailed random vectors is determined solely by the diagonal entries of the matrix and the stability index $\alpha$, a result applied to derive the atom-free nature of the $\alpha$-heavy Marčenko--Pastur law for heavy-tailed sample correlation matrices.

The Gist

Here is an explanation of the paper "Quadratic form of heavy-tailed self-normalized random vector..." using simple language, analogies, and metaphors.

The Big Picture: Taming the "Wild" Data

Imagine you are trying to understand the behavior of a massive crowd of people (data). In the world of statistics, we usually assume people behave predictably—like a calm crowd at a library. This is called a "light-tailed" distribution. If one person sneezes, it doesn't change the whole room.

But in the real world, data can be "heavy-tailed." This means the crowd is wild. Occasionally, someone might scream, run, or cause a massive disturbance. In math terms, these are outliers with infinite variance. They are rare, but when they happen, they are huge.

This paper tackles a specific problem: What happens when you take this wild, screaming crowd, force them to stand on a circle (normalize them), and then ask a complex question about how they interact with each other?

The Main Characters

1. The Wild Vector ($x$): Imagine a list of $n$ numbers. Most are small, but occasionally, one is astronomically large. These numbers follow a "heavy-tailed" rule (like the Pareto distribution or a t-distribution).
2. The Self-Normalized Vector ($y$): The researchers take that wild list and shrink it so that the total length is exactly 1. It's like taking a chaotic group of people and forcing them to hold hands in a perfect circle. No matter how wild the individuals were, the group as a whole is now a fixed size.
3. The Matrix ($A$): Think of this as a "rulebook" or a "filter." It tells the vector how to interact. Some rules are simple (diagonal entries), and some are complex interactions between different people (off-diagonal entries).
4. The Quadratic Form ($y^T A y$): This is the final score. It's a single number that represents the result of applying the rulebook to the normalized crowd.

The Core Discovery: The "Loud" vs. The "Quiet"

The researchers wanted to know: What is the final score ($Q_n$) when the crowd is wild?

In the past, mathematicians knew what happened when the crowd was calm (light-tailed). The score would settle down to a predictable average. But with wild crowds, the math usually breaks.

The Breakthrough:
The authors discovered a surprising separation between two parts of the rulebook:

• The Off-Diagonal (The Chatter): These are the rules about how person A interacts with person B. The paper proves that in a wild, heavy-tailed setting, this "chatter" actually dies out. It becomes negligible. The noise cancels itself out.
• The Diagonal (The Solo Acts): These are the rules about how person A interacts with themselves. The authors found that this is the only thing that matters.

The Analogy:
Imagine a stadium full of people.

• Light-tailed case: Everyone is whispering. The total noise is just the sum of all whispers.
• Heavy-tailed case: Most people are silent, but one person screams.
  • The researchers found that if you look at how people talk to each other (off-diagonal), the screaming doesn't really change the overall pattern of conversation.
  • However, if you look at how the screamer feels about themselves (diagonal), that single event dominates the result.

The Result:
The final score depends only on the distribution of the diagonal numbers in the rulebook and the "wildness" index ($\alpha$). It doesn't matter how the people interact with each other; the outcome is driven entirely by the individual "loudness" of the diagonal entries.

The "Heavy-Tailed" Marčenko–Pastur Law

The paper applies this discovery to Random Matrix Theory, a field that studies the eigenvalues (the "vibrational frequencies") of huge data matrices.

• The Classic Law (Marčenko–Pastur): When data is calm, the frequencies of a large dataset form a smooth, continuous shape (like a smooth hill). There are no gaps or spikes.
• The New Law ($\alpha$-heavy MP): When data is wild, does the shape change? Does it develop spikes (atoms) where the data gets stuck?

The Mystery:
Previous research suggested that as the data gets extremely wild (approaching a specific limit), the smooth hill might turn into a collection of discrete spikes (like a staircase instead of a ramp).

The Solution:
Using their new formula for the "score," the authors proved that the hill remains smooth.
Even with wild, heavy-tailed data, the distribution of frequencies has no spikes (except possibly at zero). It is a continuous, smooth curve. The "atoms" (spikes) that some feared would appear do not exist.

Why This Matters

1. Real-World Data: Financial markets, internet traffic, and earthquake data are often "heavy-tailed." They have massive outliers. This paper gives statisticians a new, accurate tool to model these systems without assuming the data is calm.
2. Simplicity: It simplifies a very complex problem. Instead of calculating millions of interactions between data points, you only need to look at the individual components.
3. The "Zero" Case: The paper also looks at the extreme edge case where the data is so wild it barely has a variance at all. They showed that in this extreme case, the data behaves like a "Zero-Inflated Poisson" distribution (mostly zeros, with occasional spikes), confirming a long-standing hypothesis.

Summary in One Sentence

By realizing that in a chaotic, heavy-tailed world, the "noise" of interactions cancels out and only the "individual volume" of the data matters, the authors proved that even the wildest data forms a smooth, continuous pattern rather than a jagged, spiky one.

Technical Summary

Here is a detailed technical summary of the paper "Quadratic form of heavy-tailed self-normalized random vector with applications in $\alpha$-heavy Marˇcenko–Pastur law" by Dong, Heiny, and Yao.

1. Problem Statement

The paper addresses the asymptotic behavior of quadratic forms constructed from self-normalized random vectors in the heavy-tailed regime.

• Context: Let $x = (X_1, \dots, X_n)^\top$ be a random vector with i.i.d. components in the domain of attraction of an $\alpha$-stable law with $\alpha \in (0, 2)$. This implies the components have infinite variance ($E[\xi^2] = \infty$).
• Object of Study: The self-normalized vector $y = x / \|x\|_2$ lies on the unit sphere $S^{n-1}$. The authors study the quadratic form $Q_n = y^\top A_n y$, where $A_n$ is a Hermitian matrix (possibly random) independent of $y$.
• Motivation:
  • In light-tailed settings (e.g., sub-Gaussian), $Q_n$ concentrates around its mean, and the limiting spectral distribution (LSD) of sample correlation matrices follows the classical Marčenko–Pastur (MP) law.
  • In heavy-tailed settings ($\alpha < 2$), the denominator $\|x\|_2$ does not concentrate, and the quadratic form behaves differently.
  • A key open problem in Random Matrix Theory (RMT) is characterizing the $\alpha$-heavy Marčenko–Pastur law ($H_{\alpha, \gamma}$), the LSD of sample correlation matrices with heavy-tailed entries. Specifically, it was unknown whether $H_{\alpha, \gamma}$ contains atoms (point masses) for $\alpha \in (0, 2)$. Previous work established the existence of the limit via the method of moments but could not determine its analytic properties (e.g., absolute continuity vs. discrete atoms).

2. Methodology

The authors employ a combination of probabilistic limit theorems, moment analysis, and complex analysis techniques tailored to heavy-tailed distributions.

• Decomposition of Quadratic Forms:
  The quadratic form $Q_n$ is decomposed into diagonal and off-diagonal parts:
  $$Q_n = Q_{n,1} + Q_{n,2} = \sum_{i=1}^n a_{ii} Y_{1i}^2 + \sum_{i \neq j} a_{ij} Y_{1i} Y_{1j}$$
  The authors prove that under mild conditions on the Frobenius norm of the off-diagonal part of $A_n$, the off-diagonal term $Q_{n,2}$ converges to zero in probability. Thus, the asymptotic behavior is governed entirely by the diagonal term $Q_{n,1}$.

• Moment Analysis and Mixed Moments:
  Unlike light-tailed cases where mixed moments decay as $n^{-(k_1+\dots+k_r)}$, the authors utilize the specific asymptotic behavior of mixed moments for self-normalized heavy-tailed vectors (Lemma 2.1):
  $$\lim_{n \to \infty} n^r E[Y_{11}^{2k_1} \dots Y_{1r}^{2k_r}] = C(\alpha, k_1, \dots, k_r)$$
  This "sharp separation" allows the derivation of the limiting distribution of $Q_{n,1}$ based solely on the empirical distribution of the diagonal entries of $A_n$ and the index $\alpha$.

• Stieltjes Transform Representation:
  The limiting law $\mu_{\nu, \alpha}$ (where $\nu$ is the weak limit of the diagonal entries) is characterized via its Stieltjes transform. The authors derive an explicit integral representation for this transform involving the measure $\nu$ and the parameter $\alpha$.

• Application to Sample Correlation Matrices:
  To analyze the sample correlation matrix $R_n = YY^\top$, the authors relate the resolvent diagonal entries $B_{n,ii}(z) = (Y^\top Y - zI)^{-1}_{ii}$ to the quadratic forms studied earlier. They establish that the empirical measure of these diagonal entries converges weakly to a random limit $\psi(z)$.

• Holomorphic Extension and Contradiction:
  To prove the absence of atoms in $H_{\alpha, \gamma}$, they derive an implicit equation for the Stieltjes transform of $H_{\alpha, \gamma}$ involving the expectation of a function of the random limit $\psi(z)$. They then use a contradiction argument: assuming an atom exists at $u > 0$ leads to incompatible asymptotic behaviors of the imaginary part of the Stieltjes transform as $\text{Im}(z) \to 0$.

3. Key Contributions and Results

A. Limiting Law of Quadratic Forms

• Theorem 2.4 & 2.12: The authors establish that if the empirical distribution of the diagonal entries of $A_n$ converges weakly to a deterministic measure $\nu$, then $Q_n$ converges in distribution to a non-degenerate law $\mu_{\nu, \alpha}$.
• Stieltjes Transform: The transform is given by:
  $$s_{\mu_{\nu, \alpha}}(z) = - \frac{\int (z-x)^{\frac{\alpha}{2}-1} \nu(dx)}{\int (z-x)^{\frac{\alpha}{2}} \nu(dx)}$$
• Density and Atoms:
  • Theorem 2.10: If $\nu$ is non-degenerate, $\mu_{\nu, \alpha}$ is atom-free (continuous) and possesses an explicit density function $f_{\nu, \alpha}(x)$.
  • Tail Behavior: The paper analyzes the tail behavior of $\mu_{\nu, \alpha}$, showing it inherits the tail index of $\nu$ if $\nu$ is polynomially heavy-tailed, but exhibits Gamma-like tails if $\nu$ is exponentially decaying.
• Boundary Cases:
  • As $\alpha \uparrow 2$ (finite variance), the law degenerates to a point mass at the mean of $\nu$.
  • As $\alpha \downarrow 0$ (extremely heavy tails), the law converges to $\nu$ itself.

B. The $\alpha$-Heavy Marčenko–Pastur Law ($H_{\alpha, \gamma}$)

• Implicit Representation (Theorem 3.3): The Stieltjes transform of the LSD of the sample correlation matrix is derived as:
  $$s_{H_{\alpha, \gamma}}(z) = - \frac{E[(1 + \psi(z))^{\frac{\alpha}{2}-1}]}{z E[(1 + \psi(z))^{\frac{\alpha}{2}}]}$$
  where $\psi(z)$ is the weak limit of the resolvent's diagonal entries.
• Absence of Atoms (Proposition 3.5): The paper proves that for $\alpha \in (0, 2)$, the law $H_{\alpha, \gamma}$ has no atoms on the positive real line $(0, \infty)$. This resolves a long-standing open question regarding whether the "countably many atoms" observed in the $\alpha \to 0$ limit persist for $\alpha > 0$.
• Slowly Varying Case ($\alpha = 0$): For the boundary case where the tail is slowly varying (e.g., $P(|\xi|>x) \sim L(x)$), the authors prove the LSD converges to a zero-inflated Poisson distribution, confirming the discrete nature of the limit at $\alpha=0$.

C. Light-Tailed Comparison

• The Appendix provides a Hanson–Wright type concentration inequality for $Q_n$ when $\xi$ is sub-Gaussian, contrasting the sharp concentration in the light-tailed case with the lack thereof in the heavy-tailed setting.

4. Significance

• Theoretical Breakthrough: This work bridges the gap between the moment-based existence proofs of heavy-tailed limits and their analytic characterization. It moves beyond knowing that a limit exists to understanding what the limit looks like (density, support, atoms).
• Resolution of the Atom Problem: By proving $H_{\alpha, \gamma}$ is atom-free for $\alpha \in (0, 2)$, the authors clarify the phase transition in random matrix theory. The discrete atoms observed at $\alpha=0$ are a singular phenomenon that disappears immediately as soon as the tail index becomes positive, even if the variance remains infinite.
• Methodological Innovation: The technique of separating diagonal and off-diagonal contributions in self-normalized vectors and linking the resolvent diagonal to quadratic forms provides a robust framework for analyzing other heavy-tailed random matrix models where standard local laws fail.
• Applications: The results are crucial for high-dimensional statistics involving heavy-tailed data (e.g., financial returns, network traffic), where sample correlation matrices are standard tools but their spectral properties were previously poorly understood in the infinite variance regime.

In summary, the paper provides a complete characterization of the limiting spectral distribution for heavy-tailed sample correlation matrices, proving their absolute continuity and providing explicit formulas for their Stieltjes transforms and densities.