Abstract
This paper addresses the development of analytical tools for the computation of the inverse moments of random Gram matrices with one side correlation. Such a question is mainly driven by applications in signal processing and wireless communications wherein such matrices naturally arise. In particular, we derive closed-form expressions for the inverse moments and show that the obtained results can help approximate several performance metrics such as the average estimation error corresponding to the best linear unbiased estimator (BLUE) and the linear minimum mean square error (LMMSE) estimator or also other loss functions used to measure the accuracy of covariance matrix estimates.
Introduction
In wireless communications and signal processing, systems are frequently modeled using random matrices. A particularly significant type is the Gram matrix ($S = H^*\Lambda H$), which appears in multi-antenna (MIMO) channels and sample covariance matrix analysis. Understanding the performance of algorithms—such as the error rate of a receiver—requires calculating the moments of these matrices. While "positive" moments (e.g., mean or variance) are well-understood, inverse moments (the expected value of $S^{-1}, S^{-2}$, etc.) have historically been difficult to calculate for finite-sized, correlated matrices.
Previously, researchers relied on asymptotic analysis, which assumes infinitely large matrices. While useful, these approximations often fail for the finite-sized matrices used in real-world hardware. This paper provides the first exact, closed-form mathematical expressions for the inverse moments of one-sided correlated Gram matrices, bridging the gap between theory and practical application.
Methodology
The authors employ a framework based on the Mellin transform of the marginal density of the matrix's eigenvalues. A major hurdle was a mathematical singularity: standard formulas for these transforms involve Gamma functions that diverge when dealing with negative integers (inverse moments).
The key methodological innovations include:
- Singularity Decomposition: Splitting the complex Mellin transform into two components ($M_1$ and $M_2$).
- Limit Analysis: Rigorously proving that the problematic component ($M_2$) vanishes in the limit, while $M_1$ can be resolved using Taylor expansions and properties of the Gamma function.
- Validation: The theoretical results were validated against $10^4$ Monte Carlo simulations using Bessel correlation models and random positive definite models.
Key Results and Finite-Dimensional Accuracy
The primary contribution is Theorem 1, which provides a closed-form expression for the $r$-th inverse moment $\mu_\Lambda(-r)$ for any finite dimensions $n$ and $m$. Numerical simulations demonstrate that this new formula matches Monte Carlo results perfectly across all tested scenarios.
As shown in Figure 1, while asymptotic approximations (Theorem 2) perform reasonably well for the first inverse moment ($r=-1$), they diverge significantly as the moment order increases (e.g., $r=-3$). The exact formula remains precise regardless of the order, providing a reliable tool for finite-system analysis.
Practical Applications in Signal Processing
The derived moments map directly to the Mean Square Error (MSE) of standard estimators. For instance, the error of the Best Linear Unbiased Estimator (BLUE) is proportional to the first inverse moment.
Beyond BLUE, the authors used these moments to provide highly accurate LMMSE approximations in both high-SNR and low-SNR regimes. They also derived formulas for loss functions in sample covariance matrices, enabling the mathematical determination of optimal "forgetting factors" for tracking time-varying signals.
Limitations and Future Work
The current derivation is valid for $1 \le r \le p$, where $p = \min(m, n-m)$, meaning it cannot calculate arbitrarily high inverse moments if matrix dimensions are too close. Additionally, while closed-form, the expressions involve cofactors and Vandermonde matrix inversions, which may present numerical stability challenges for extremely large matrices. Future research may extend this approach to two-sided correlation models or non-Gaussian entries.
Summary updated