Abstract
In many signal processing and wireless communication applications, the stability of matrix inversion is paramount. This paper provides a rigorous mathematical proof that the smallest eigenvalue of generally correlated Gaussian random matrices remains strictly bounded away from zero as matrix dimensions grow. This result solidifies the theoretical foundation for performance analysis in systems employing Zero-Forcing precoding and robust statistical methods, where the invertibility of Gram matrices was previously assumed rather than proven.
Introduction
In the design of multi-user Multiple Input Single Output (MISO) wireless systems and robust signal processing algorithms, researchers frequently encounter Gram matrices formed by Gaussian vectors. A particularly challenging and realistic model is the "generally correlated" case, where matrix columns are independent but each possesses a unique covariance matrix. This heterogeneity reflects real-world scenarios where different mobile users experience distinct scattering environments.
Many foundational algorithms, such as Zero-Forcing (ZF) precoding, require the inversion of these Gram matrices. The numerical stability and performance of these operations depend entirely on the behavior of the smallest eigenvalue. If the smallest eigenvalue approaches zero as the system scales, the inverse becomes unstable, leading to significant performance degradation. While previous work in Random Matrix Theory (RMT) characterized the asymptotic eigenvalue distribution, the specific behavior of the extreme eigenvalues in the generally correlated case remained a critical gap in the literature.
Methodology
The authors bridge this gap by employing advanced tools from RMT and Gaussian calculus. The proof is structured around two primary technical pillars:
1. Bounding the Limiting Support
The first stage involves proving that zero does not belong to the support of the deterministic limiting eigenvalue distribution. By analyzing the Stieltjes transform of the distribution and utilizing standard interference functions, the authors demonstrate that the limiting measure $\mu_N$ leaves a gap $[0, \epsilon]$ for some strictly positive $\epsilon > 0$.
2. Exact Separation
The second stage ensures that no empirical eigenvalues "escape" the support of the limiting distribution as the matrix dimensions $N$ (rows) and $n$ (columns) grow to infinity. Because the entries are Gaussian, the authors leverage the Integration by Parts formula and the Nash-Poincaré inequality. These tools allow for a tight characterization of the variance of the resolvent matrix, effectively controlling the fluctuations of the extreme eigenvalues and proving that they stay within the predicted support almost surely.
Key Insights and Results
Validating Linear Precoding Foundations
This work transforms a common convenience into a mathematical certainty. Previously, researchers analyzing ZF precoding had to assume the smallest eigenvalue was bounded away from zero to ensure the existence of the precoder. This paper provides the rigorous proof necessary to justify those analyses.
Handling Heterogeneity
Unlike the simpler Kronecker model, where all columns share a single correlation structure, this paper addresses the "generally correlated" model. This is a significant step forward for RMT, as it accounts for the statistical diversity inherent in modern communication networks.
Technical Achievements
A core contribution is the quantification of the error when replacing the expected trace of the resolvent matrix with its deterministic equivalent. The authors show this error scales as $\mathcal{O}(1/N^2)$, a result sufficiently tight to prove the absence of eigenvalues outside the limiting support.
Summary of Theorems
- Theorem 1: Establishes that the interval $[0, \epsilon]$ is empty of the support of the deterministic limiting measure.
- Theorem 2: Establishes that with probability one, no eigenvalues of the empirical Gram matrix appear in any closed interval outside that support for sufficiently large $N$.
- Corollary 5: Concludes that for large $N$, the smallest eigenvalue is strictly bounded away from zero, provided the ratio $N/n < 1$ (ensuring the matrix is not rank-deficient) and the individual covariance matrices have bounded eigenvalues.
Limitations and Future Directions
The current proof relies heavily on the properties of Gaussian distributions, specifically the mathematical machinery of Gaussian calculus (e.g., Nash-Poincaré inequality). Consequently, the results are not immediately applicable to non-Gaussian noise or channel models.
Future research may seek to establish "universality" for these results—proving that the smallest eigenvalue remains bounded away from zero for a broader class of independent, non-Gaussian entries. Such an extension would likely require moving away from Gaussian calculus toward more generalized techniques like the Lindeberg principle or moment methods.
Summary updated