Abstract
In the present paper, we prove that under the assumption of the finite sixth moment for elements of a Wigner matrix, the convergence rate of its empirical spectral distribution to the Wigner semicircular law in probability is $O(n^{-1/2})$ when the dimension $n$ tends to infinity.
Introduction
In random matrix theory, Wigner matrices serve as a foundational model for describing complex systems, ranging from the energy levels of heavy nuclei to correlations in financial markets. A central result in this field is the Wigner semicircular law, which states that as the size of a Wigner matrix ($n$) grows to infinity, the distribution of its eigenvalues—known as the Empirical Spectral Distribution (ESD)—converges to a specific semicircular shape.
While the fact of convergence is well-established, the rate at which this convergence occurs is a subject of intense research. This rate is not merely a theoretical curiosity; it is a vital prerequisite for proving more advanced properties, such as central limit theorems for linear spectral statistics. Historically, achieving the optimal convergence rate of $O(n^{-1/2})$ has required stringent assumptions on the matrix entries, specifically requiring a high number of finite moments (e.g., the 12th moment). This paper addresses the challenge of maintaining this optimal rate while significantly relaxing these moment conditions.
Methodology
The authors utilize the Stieltjes transform method, a standard analytical tool that maps the distribution of eigenvalues into the complex plane, transforming spectral problems into functional analysis problems. The proof is built upon three primary pillars:
- Bai’s Inequality: A critical departure from previous literature is the use of a Berry-Esseen type inequality developed by Zhidong Bai. Earlier proofs often relied on the Götze-Tikhomirov inequality, which involves complex suprema that necessitate higher-order moment conditions to control. Bai’s inequality simplifies the estimation by focusing on the expectation of the difference at specific points rather than over an entire interval.
- Moment Truncation and Centralization: To handle the potential for "heavy tails" in the random variables, the authors employ truncation techniques. This allows them to work with bounded variables while rigorously demonstrating that the error introduced by this truncation is negligible in the limit.
- Recursive Estimation: The difference between the ESD and the semicircular law is decomposed into four distinct components ($S_1$ through $S_4$). These components are bounded systematically using Hölder’s inequality and technical lemmas concerning the traces of matrix inverses.
Key Insights
The primary breakthrough of this work is the realization that the 12th moment condition required in earlier landmark studies (such as Götze et al., 2003) was a technical artifact of the specific inequalities used in those proofs, rather than a fundamental property of the semicircular law itself. By selecting a more efficient inequality (Bai’s Inequality), the authors demonstrate that the "noise" in the eigenvalue distribution dissipates at the $O(n^{-1/2})$ rate even when the matrix entries only possess a finite 6th moment.
Results and Conclusion
The paper successfully establishes that the distance between the empirical distribution and the semicircular law (in probability) is at most a constant times $n^{-1/2}$, provided that $\mathbb{E}|x_{ij}|^6 < \infty$. This result aligns with the best-known convergence rate in the literature but significantly lowers the "entry barrier" for the types of random variables to which the theory can be applied.
Despite this advancement, the authors note that the absolute optimal conditions remain an open question. It is theoretically possible that the $O(n^{-1/2})$ rate could be maintained with as few as four finite moments, or that an even faster rate of convergence exists. These possibilities continue to represent a significant frontier in random matrix theory.
Summary updated