🔢 mathematics

Spectral properties of Random Matrices

This paper establishes the theoretical foundations of random matrix theory by utilizing the Stieltjes transform and Dyson equation methods to prove optimal convergence to the Marchenko-Pastur distribution, eigenvalue rigidity, and universal local laws for an interpolating random matrix model.

Original authors: Anastasis Kafetzopoulos

Published 2026-09-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anastasis Kafetzopoulos

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

In the vast landscape of modern mathematics, there is a field dedicated to understanding the hidden order within chaos. This is the study of random matrices, which are simply grids of numbers where every entry is chosen by chance, like rolling dice to fill a spreadsheet. While the individual numbers are unpredictable, mathematicians have long discovered that when these grids become very large, their collective behavior settles into a predictable pattern. Imagine a crowd of people moving randomly; while you cannot predict where any single person will be, you can accurately predict the density of the crowd in any given area. In the world of random matrices, this density is called the spectral distribution, and it reveals the statistical "shape" of the matrix's eigenvalues, which are special numbers that describe how the matrix stretches or compresses space. For decades, scientists have known that for certain types of these random grids, this shape follows a few famous laws, such as a semi-circle for symmetric grids or a perfect circle for grids with complex numbers. However, real-world data is rarely so simple. It often contains correlations, where one number influences another, or it comes from sources that do not follow the standard bell curve of probability. Understanding how these messy, correlated, and non-standard grids behave is crucial for fields ranging from quantum physics to financial modeling, yet the mathematical tools to describe them have often been limited to idealized, perfectly random scenarios.

This thesis by Anastasis Kafetzopoulos tackles the challenge of describing these more complex, realistic random matrices with rigorous precision. The work is divided into two major parts, each introducing a powerful mathematical technique to solve a specific type of problem. The first part focuses on sample covariance matrices, which are grids used to measure how different variables in a dataset relate to one another. In many real-world situations, such as analyzing financial markets or biological data, these matrices have a "hard edge," a boundary where the density of values drops to zero in a sharp, singular way, unlike the smooth curves seen in simpler models. Previous mathematical methods struggled to describe the behavior of eigenvalues right at this sharp edge with high accuracy. Kafetzopoulos developed a refined version of a technique called the Stieltjes transform method. By treating the random matrix as a collection of interacting parts and analyzing the fluctuations of its components, the author proved that the eigenvalues of these complex matrices converge to their expected shape at an optimal speed. Crucially, the work demonstrated that even near the sharp, difficult edge, the eigenvalues are "rigid," meaning they stay extremely close to their predicted positions with very little fluctuation. This result improves upon earlier findings by removing unnecessary mathematical corrections and providing a clearer, more accurate picture of how these matrices behave in their most challenging regions.

The second part of the research moves beyond symmetric or purely random matrices to study a more general and difficult class: non-Hermitian matrices with correlated entries. These are grids where the numbers are not independent, and the matrix does not possess the symmetry that makes many problems easier to solve. Such matrices appear in advanced physics, particularly in the study of quantum chromodynamics, where particles interact in complex ways. To handle this, the author employed a more sophisticated tool known as the Dyson equation method. This approach treats the entire matrix as a single system and derives a deterministic equation that the matrix's properties must satisfy, effectively turning a problem of infinite randomness into a solvable algebraic puzzle. The thesis applies this method to a model that interpolates between several famous laws, creating a spectrum that shifts from a circle to an ellipse and finally to a line, depending on the strength of the correlation between the matrix entries. The research proves that the distribution of eigenvalues for these correlated matrices is "universal." This means that the final shape of the spectrum depends only on the correlation structure and not on the specific details of how the individual numbers were chosen, provided they follow basic statistical rules. The author successfully established that this universal behavior holds even on very small scales, confirming that the local arrangement of eigenvalues is robust and predictable regardless of the initial randomness.

The significance of these findings lies in their ability to bridge the gap between idealized mathematical theory and the messy reality of data. By proving that these complex, correlated systems still obey universal laws, the work provides a reliable foundation for scientists and engineers who rely on random matrix theory to model everything from the vibrations of atomic nuclei to the fluctuations of stock prices. The thesis does not merely suggest these patterns exist; it provides a complete mathematical proof that they hold true under specific, well-defined conditions. The author successfully extended the reach of existing theories to cover "hard edges" and "correlated entries," areas where previous methods were either too approximate or entirely inapplicable. Through careful analysis and the application of advanced probabilistic tools, the research confirms that even in the presence of strong correlations and non-standard distributions, the underlying order of large random systems remains intact, offering a deeper understanding of the statistical laws that govern the complex structures of our world.

Technical Summary: Spectral Properties of Random Matrices

Problem Statement
This thesis addresses the spectral properties of random matrices, specifically focusing on the convergence of empirical spectral distributions (ESD) to deterministic limiting laws and the rigidity of individual eigenvalues. The work is divided into two primary investigations:

Local Marchenko-Pastur Law: Analyzing the sample covariance matrix $X^*X$ (where $X$ is an $N \times N$ matrix with i.i.d. entries) to establish optimal convergence rates and eigenvalue rigidity, particularly near the "hard edge" (zero) where the limiting density exhibits a square-root singularity.
Universality of Correlated Covariance Matrices: Investigating the spectrum of non-Hermitian matrices of the form $X_1 X_2^*$ , where $X_1$ and $X_2$ are correlated random matrices. This model interpolates between the circular law, the elliptic law, and the Marchenko-Pastur law depending on a correlation parameter $\tau$ . The goal is to prove universality results (independence from the specific entry distribution) for these models, extending previous Gaussian-only results.

Methodology
The thesis employs two distinct but rigorous technical frameworks:

The Stieltjes Transform Method (Part I):
- Framework: The empirical spectral measure is analyzed via its Stieltjes transform, $S_N(z) = \frac{1}{N}\text{Tr}(G_z)$ , where $G_z = (X^*X - zI)^{-1}$ is the resolvent.
- Technique: The author derives a self-consistent equation for $S_N(z)$ using resolvent identities (specifically the "stripping lemma" and Woodbury matrix identity). The difference between the random Stieltjes transform and the deterministic limit $S_\rho$ is controlled by analyzing a quadratic error term involving probabilistic fluctuations.
- Key Tools: The proof relies on high-moment bounds for quadratic forms of random vectors. Unlike previous works that often assumed sub-Gaussian decay, this thesis utilizes only a four-moment condition on the matrix entries. The analysis involves a careful induction scheme on the powers of moments and the use of Rosenthal's and Burkholder's inequalities adapted for complex entries.
- Hard Edge Handling: A specific modification is introduced to handle the singularity at zero. The convergence scale is adjusted to $\eta \sim N^{-2}$ near the hard edge, and the bounds are weighted by $|\sqrt{z}|$ to manage the divergence of the Stieltjes transform.
The Dyson Equation Method (Part II):
- Framework: To handle non-Hermitian matrices with correlations, the thesis utilizes the "Hermitization" technique. The spectrum of the non-Hermitian matrix $X_1 X_2^*$ is related to the spectrum of a larger Hermitian "Dirac" matrix $H_\zeta = \begin{pmatrix} 0 & X_1 - \zeta \\ (X_2 - \bar{\zeta})^* & 0 \end{pmatrix}$ .
- Technique: The resolvent of the Hermitized matrix is shown to satisfy a deterministic Dyson equation of the form $I + (z + S[M])M = 0$ , where $S$ is a self-energy operator encoding the correlation structure.
- Key Tools: The proof involves two main steps:
  1. Multivariate Cumulant Expansion: Used to show that the probabilistic equation satisfied by the random resolvent is close to the deterministic Dyson equation (controlling the error matrix).
  2. Stability Analysis: Proving that the linear stability operator associated with the Dyson equation is invertible in the relevant subspace. This requires a detailed spectral analysis of the stability operator for a $4 \times 4$ matrix system, which is more complex than the $2 \times 2$ systems found in standard elliptic law proofs.
- Least Singular Value: A critical component involves controlling the least singular value of the matrix $Y_N - \zeta$ to ensure the resolvent remains bounded, utilizing a bounded density assumption on the matrix entries.

Key Contributions and Results

Optimal Local Marchenko-Pastur Law:
- The thesis establishes optimal bounds for the fluctuation of the Stieltjes transform $S_N(z)$ around the Marchenko-Pastur limit $S_\rho(z)$ .
- Improvement over Literature: It removes logarithmic corrections present in previous works (e.g., Cacciapuoti-Maltsev-Schlein) and improves fluctuation bounds from $\sqrt{\log N}$ to $\log N$ in the counting function.
- Hard Edge Rigidity: It provides new rigidity estimates for eigenvalues near the hard edge ($0$). Specifically, for the $i$ -th eigenvalue $\lambda_i$ near zero, the fluctuation is shown to be of order $\frac{\log N}{N^2}$ (scaled by the local density), improving upon the standard bulk rigidity of $\frac{\sqrt{\log N}}{N}$ .
- Convergence Rate: The rate of convergence of the empirical spectral distribution to the Marchenko-Pastur distribution is quantified with optimal scaling, valid for both the bulk and the edges (soft and hard).
Universality of Correlated Covariance Matrices:
- The thesis proves that the local spectral statistics of the matrix $X_1 X_2^*$ (with correlated entries defined by parameter $\tau$ ) are universal.
- Generalization: It extends the results of Akemann, Byun, and Kang (who assumed Gaussian entries) to general i.i.d. entries with finite moments.
- Local Law: It establishes a local law for the Dirac matrix $H_\zeta$ on the scale of the typical eigenvalue separation ( $O(N^{-1})$ for the Hermitian case, $O(N^{-1/2})$ for the non-Hermitian case).
- Support Characterization: The limiting spectral support is identified as an ellipse (for $\tau \in (0,1)$ ) that transitions to the circular law ( $\tau=0$ ) and the Marchenko-Pastur distribution ( $\tau=1$ ). The proof rigorously derives the shape of this support using the solution to the $4 \times 4$ Dyson equation.

Significance and Claims
The author claims that the primary significance of this work lies in the rigorous extension of local laws to "hard edge" scenarios and correlated non-Hermitian models without relying on Gaussian assumptions.

Methodological Advancement: The thesis demonstrates that the Stieltjes transform method can be refined to handle the singularities at the hard edge of the Marchenko-Pastur distribution, providing a unified treatment for bulk and edge eigenvalues that simplifies previous case-by-case proofs.
Universality: By successfully applying the Dyson equation method to a model with $4 \times 4$ block structures and non-trivial correlations, the work bridges the gap between Gaussian-specific results and general universality classes in non-Hermitian random matrix theory.
Technical Robustness: The results hold under minimal assumptions (finite fourth moments for the covariance case and bounded density for the correlated case), making the findings applicable to a broader class of physical and statistical models, including those arising in Quantum Chromodynamics as noted in the introduction.

The thesis concludes that these new local laws are now considered universal, meaning the spectral properties are independent of the specific initial distribution of the matrix entries, provided the moments and correlation structures are preserved.

More like this