No Eigenvalues Outside the Limiting Support of Generally Correlated and Noncentral Sample Covariance Matrices

Zeyan Zhuang, Xin Zhang, Dongfang Xu, Shenghui SongarXiv 2025 · arXiv

TL;DR

This paper provides a rigorous theoretical foundation for the spectral behavior of high-dimensional sample covariance matrices that possess both non-zero means (noncentral) and non-identical correlation structures. The authors prove that as matrix dimensions grow, the distribution of eigenvalues converges to a predictable deterministic limit and, crucially, no "stray" eigenvalues appear outside the support of this limit—a property that offers new theoretical guarantees for the performance of MIMO communication systems.

What Problem Does This Paper Address?

In classical statistics, we usually assume we have a massive amount of data ($n$) compared to the number of features ($p$). However, in the modern "big data" era, $p$ and $n$ often grow simultaneously. Random Matrix Theory (RMT) was developed to describe how eigenvalues behave in this high-dimensional regime.

While previous research has covered "centered" matrices (where data has a zero mean) or matrices with identical correlations, real-world data is rarely that simple. For example, in wireless communications, signals often follow a "Rician fading" model, which results in noncentral (non-zero mean) and generally correlated (different correlations for different users) matrices.

The specific "gap" this paper fills is proving the "no-eigenvalue" property for these complex models. It isn't enough to know the general shape of the eigenvalue distribution; engineers need to know if the smallest or largest eigenvalues might "leak" outside the expected range, as these extreme values often dictate system stability and performance.

Methods

The authors employ the Stieltjes transform method, a standard but powerful tool in RMT used to characterize the Empirical Spectral Distribution (ESD).

Key technical steps include:

  • Fixed-Point Equations: They derive a fundamental system of $2n$ fixed-point equations to approximate the "resolvent" (a matrix-valued function related to the covariance matrix).
  • High-Dimensional Asymptotics: They analyze the system as $p, n \to \infty$ while their ratio $p/n$ converges to a constant $c$.
  • Probabilistic Bounds: To prove the "no-eigenvalue" property, the authors use truncation, centralization, and normalization of random variables. They apply advanced tools like the Burkholder inequality and Matrix Bernstein inequality to show that the probability of an eigenvalue appearing outside the predicted support vanishes as the matrix size increases.

Datasets

As a theoretical mathematics and engineering paper, it does not use real-world datasets. Instead, it uses synthetic Monte Carlo simulations to validate the derived theorems:

  • Simulated MIMO Channels: They generate matrices using Rician fading models with heterogeneous spatial correlations.
  • Distribution Types: They test the theory using various underlying distributions for the random noise, including standard complex Gaussian and uniform distributions, to demonstrate that the results are "universal" (independent of the specific distribution type, provided certain moments exist).

Key Insights and Ideas

  • The "No-Outlier" Guarantee: The most significant "aha moment" is the proof that for sufficiently large matrices, no eigenvalues will exist in any closed interval outside the limiting support. This is vital for "spiked" models, where researchers need to distinguish between true signal (eigenvalues outside the support) and noise.
  • Deterministic Approximation: Even when the model is too complex for a simple closed-form distribution (like the Marčenko-Pastur law), the authors show that the resolvent can still be approximated by a deterministic matrix.
  • Physical Interpretability: The math isn't just abstract; it directly maps to the invertibility of matrices. In wireless precoding (Zero-Forcing), if the smallest eigenvalue hits zero, the system fails. This paper provides the conditions under which that won't happen.

Key Results

  • ESD Convergence: Theorem 1 establishes that the trace and bilinear forms of the resolvent converge almost surely to a deterministic limit.
  • Extreme Eigenvalue Bounds: Theorem 2 proves the "no-eigenvalue" property. This ensures that the largest eigenvalue remains bounded and the smallest eigenvalue stays within the predicted support.
  • MIMO Performance: The authors successfully apply their theory to determine the limiting Signal-to-Interference-plus-Noise Ratio (SINR) for multi-user MIMO systems.
  • Zero-Forcing Precoder Validity: They provide a theoretical guarantee (Corollary 4) for the invertibility of the Gram matrix in downlink MIMO, ensuring that Zero-Forcing precoding is mathematically sound even in non-ideal, correlated environments.

Limitations and Future Work

  • Assumption 5: The authors assume the existence of certain limits for the Stieltjes transforms for technical convenience. While they note this is likely provable, it remains a formal assumption in this work.
  • Exact Separation: The paper proves no eigenvalues exist outside the support, but it doesn't explore "exact separation"—the phenomenon where eigenvalues might split into distinct clusters within the support.
  • Fluctuations: While the paper establishes where the eigenvalues are, it doesn't analyze the fluctuations (the "wiggle" or distribution of error) of the extreme eigenvalues around those boundaries. The authors suggest this as a promising direction for future research.

Summary updated

Abstract

Spectral properties of random matrices play an important role in statistics, machine learning, communications, and many other areas. Engaging results regarding the convergence of the empirical spectral distribution (ESD) and the ``no-eigenvalue'' property have been obtained for random matrices with different correlation structures. However, the related spectral analysis for generally correlated and noncentral random matrices is still incomplete, and this paper aims to fill this research gap. Specifically, we consider matrices whose columns are independent but with non-zero means and non-identical correlations. Under high-dimensional asymptotics where both the number of rows and columns grow simultaneously to infinity, we first establish the almost sure convergence of the ESD for the concerned random matrices to a deterministic limit, assuming mild conditions. Furthermore, we prove that with probability 1, no eigenvalues will appear in any closed interval outside the support of the limiting distribution for matrices with sufficiently large dimensions. The above results can be applied to different areas such as statistics, wireless communications, and signal processing. In this paper, we apply the derived results to two communication scenarios: 1) We determine the limiting performance of the signal-to-interference-plus-noise ratio for multi-user multiple-input multiple-output (MIMO) systems with linear minimum mean-square error receivers; and 2) We establish the invertibility of zero-forcing precoding matrices in downlink MIMO systems, providing theoretical guarantees.
BibTeX
@misc{zhuang2025no,
  author = {Zeyan Zhuang and Xin Zhang and Dongfang Xu and Shenghui Song},
  title = {{No Eigenvalues Outside the Limiting Support of Generally Correlated and Noncentral Sample Covariance Matrices}},
  year = {2025},
  eprint = {2507.03356},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.03356},
  howpublished = {arXiv},
}
Figure