New estimators of spectral distributions of Wigner matrices

Zhou Wang

TL;DR

This paper introduces nonparametric kernel estimators to study the spectral distributions of Wigner matrices, proving they are consistent even when matrix entries have infinite variances. By applying kernel density estimation to eigenvalues, the author demonstrates that Wigner’s famous semicircle law is more robust than previously thought, holding true as long as the underlying distribution is in the "domain of attraction" of a normal law.

What Problem Does This Paper Address?

In Random Matrix Theory (RMT), Wigner matrices are a fundamental object of study. Traditionally, we know that as the size of these matrices grows, the distribution of their eigenvalues follows the "semicircle law." To estimate this distribution from data, researchers typically use the Empirical Spectral Distribution (ESD)—essentially a step function based on the observed eigenvalues.

However, the ESD has limitations. First, it’s a "rough" estimator (a series of jumps). Second, while we know it converges to the semicircle law, the rate of that convergence is often unclear, and there is no general Central Limit Theorem (CLT) for the difference between the ESD and the limit when matrix elements only have finite moments.

Furthermore, most classical results assume that the entries of the matrix have finite variances. This paper tackles the "heavy-tailed" case: what happens to the spectral distribution when the variance is infinite?

Methods

The author proposes using Kernel Estimators—a popular tool in nonparametric statistics—to smooth out the spectral data.

  1. Density Estimator ($f_n$): Instead of a histogram of eigenvalues ($\mu_i$), the author applies a kernel function $K$ (like a Gaussian or Cauchy kernel) and a bandwidth $h$ to create a smooth density curve: $$f_n(x) = \frac{1}{ph} \sum_{i=1}^{p} K\left(\frac{x - \mu_i}{h}\right)$$
  2. Distribution Estimator ($F_n$): This is simply the integral of the density estimator.
  3. Stieltjes Transform Connection: The author points out a beautiful theoretical link: if you choose a specific kernel (the Cauchy kernel), the density estimator becomes equivalent to the imaginary part of the Stieltjes transform, a powerhouse tool in RMT.

The paper provides rigorous mathematical proofs for the consistency of these estimators—meaning that as the matrix size $n$ goes to infinity, the estimators converge to the true semicircle law.

Datasets

As a theoretical mathematics paper, the "datasets" used are simulated via Monte Carlo methods rather than real-world data. The author evaluates the estimators using:

  • Exponential Distributions: Specifically $P(X_{11} \leq x) = 1 - e^{-(x+1)}$ for $x \geq -1$.
  • Poisson Distributions: $P(X_{11} = k) = e^{-1}/(k+1)!$ for $k = -1, 0, 1, \dots$.
  • Matrix Sizes: Simulations were conducted on $50 \times 50$ and $800 \times 800$ symmetric matrices to observe how the estimators improve with scale.

Key Insights and Ideas

  • Robustness of the Semicircle Law: The most significant "aha" moment is that the semicircle law is a universal attractor. Even if your matrix entries have infinite variance, as long as they are in the "domain of attraction of the normal law" (meaning their sums still eventually look Gaussian), the eigenvalues will still settle into Wigner's semicircle.
  • Smoothing is Better: The simulation study suggests that the kernel distribution estimator ($F_n$) actually performs better (is closer to the limit) than the classical "jumpy" ESD.
  • The Bandwidth Trade-off: Like all kernel methods, the choice of bandwidth $h$ is crucial. The author uses $h = n^{-2/5}$ in simulations, a standard choice that balances bias and variance.

Key Results

  • Theorem 1 & 2: The author proves that the kernel distribution estimator $F_n$ converges "almost surely" to the semicircle law. Crucially, this holds even for infinite variance cases.
  • Theorem 3: The density estimator $f_n(x)$ is also proven to be consistent (converging in probability), though this proof requires stricter conditions (finite 3rd and 6th moments).
  • Visual Confirmation: The simulation plots (Figures 1-4) clearly show that as $n$ increases from 50 to 800, the "wiggly" estimated density curves tighten significantly around the perfect theoretical semicircle.

Limitations and Future Work

  • Moment Requirements for Density: While the distribution estimator works for infinite variance, the proof for the density estimator ($f_n$) requires finite 6th moments for the off-diagonal elements. This suggests a gap in the theory: can we prove density consistency for truly heavy-tailed matrices?
  • Optimal Bandwidth: The paper uses a heuristic for the bandwidth $h$. Future work could investigate "data-driven" bandwidth selection specifically tailored for spectral data.
  • Rate of Convergence: While consistency is proven (it gets there eventually), the paper does not provide a specific "speed limit" (rate of convergence) for these new kernel estimators compared to the classical ESD.

Summary updated

Abstract

We introduce kernel estimators for the semicircle law. In this first part of our general theory on the estimators, we prove the consistency and conduct simulation study to show the performance of the estimators. We also point out that Wigner's semicircle law for our new estimators and the classical empirical spectral distributions is still true when the elements of Wigner matrices don't have finite variances but are in the domain of attraction of normal law.
BibTeX
@article{wang2013new,
  author = {Zhou Wang},
  title = {{New estimators of spectral distributions of Wigner matrices}},
  journal = {Journal of Mathematical Physics},
  year = {2013},
  doi = {10.1063/1.4794075},
}
Figure