[Paper Review] A Subspace Estimator for Fixed Rank Perturbations of Large Random Matrices

Walid Hachem, Philippe Loubaton+3 more|arXiv (Cornell University)|Jun 8, 2011

Random Matrices and Applications22 references4 citations

TL;DR

This paper proposes a novel subspace estimator for parameter estimation in large-dimensional signal processing, leveraging large random matrix theory to handle fixed-rank perturbations of empirical covariance matrices when both signal dimension $N$ and sample size $n$ grow asymptotically. It extends the MUSIC algorithm to the double-asymptotic regime ($N,n \to \infty$ with $N/n \to c \in (0,\infty)$), proving almost sure convergence and central limit theorems for eigenvalue and eigenspace estimators, with explicit first- and second-order asymptotic behavior derived via recent random matrix results.

ABSTRACT

This paper deals with the problem of parameter estimation based on certain eigenspaces of the empirical covariance matrix of an observed multidimensional time series, in the case where the time series dimension and the observation window grow to infinity at the same pace. In the area of large random matrix theory, recent contributions studied the behavior of the extreme eigenvalues of a random matrix and their associated eigenspaces when this matrix is subject to a fixed-rank perturbation. The present work is concerned with the situation where the parameters to be estimated determine the eigenspace structure of a certain fixed-rank perturbation of the empirical covariance matrix. An estimation algorithm in the spirit of the well-known MUSIC algorithm for parameter estimation is developed. It relies on an approach recently developed by Benaych-Georges and Nadakuditi, relating the eigenspaces of extreme eigenvalues of the empirical covariance matrix with eigenspaces of the perturbation matrix. First and second order analyses of the new algorithm are performed.

Motivation & Objective

• To develop a robust parameter estimation algorithm for multidimensional time series when both signal dimension $N$ and observation window $n$ grow to infinity at the same rate.
• To address the failure of classical MUSIC in the double-asymptotic regime, where the spectral norm of the projection error $\|\widehat{\Pi} - \Pi\|$ does not vanish.
• To extend the MUSIC framework using recent large random matrix theory results on extreme eigenvalues and eigenspaces of fixed-rank perturbations.
• To provide first- and second-order asymptotic analysis of the proposed estimator, including almost sure convergence and central limit theorems.

Proposed method

• The method relies on the eigenspace structure of extreme eigenvalues of the empirical covariance matrix $\Sigma_n\Sigma_n^*$, which are linked to the perturbation matrix $P_n$ via recent results by Benaych-Georges and Nadakuditi.
• It constructs a subspace estimator by projecting onto the eigenspace associated with the $r$ largest eigenvalues of $\Sigma_n\Sigma_n^*$, replacing the classical projection $\Pi$ with $\widehat{\Pi}$.
• The asymptotic behavior of the estimator is analyzed using the Stieltjes transform and spectral measure convergence, particularly focusing on the edge eigenvalues that emerge outside the bulk spectrum due to fixed-rank perturbations.
• The paper derives the almost sure convergence of the eigenvalues $\widehat{\lambda}_{k,n}$ to their deterministic equivalents $\rho_{k}$, and establishes a central limit theorem for the normalized eigenvector fluctuations.
• It uses a perturbation expansion of the empirical covariance matrix $\Sigma_n\Sigma_n^* = X_nX_n^* + P_nP_n^*$, where $X_n$ is a standard Gaussian matrix and $P_n$ is a fixed-rank signal matrix.
• Theoretical results are validated through asymptotic analysis under assumptions A1–A6, including convergence of $S_n^*S_n \to O > 0$, $B^*B \to I_r$, and $\sqrt{n}(B^*B S_n^*S_n - O) = \mathcal{O}(1)$.

Experimental results

Research questions

• RQ1 How do the extreme eigenvalues and eigenspaces of the empirical covariance matrix behave when both $N$ and $n$ grow to infinity at the same rate, under a fixed-rank perturbation?
• RQ2 Can the MUSIC algorithm be extended to the double-asymptotic regime where $N,n \to \infty$ with $N/n \to c \in (0,\infty)$, and what are the limiting behaviors of the estimator?
• RQ3 What is the first-order asymptotic behavior of the eigenspace estimator in this regime, and does it converge almost surely to the true signal subspace?
• RQ4 What is the second-order asymptotic behavior, and does the normalized eigenvector fluctuation converge to a Gaussian distribution?
• RQ5 How do the asymptotic distributions of the localization function $\chi_{\text{new}}(\varphi)$ compare to the classical MUSIC function in finite-sample and large-sample regimes?

Key findings

• The eigenvalues $\widehat{\lambda}_{k,n}$ of the empirical covariance matrix $\Sigma_n\Sigma_n^*$ converge almost surely to deterministic limits $\rho_k$, which are solutions to the equation $g(\rho) = 1/\omega_{k,n}^2$, where $g$ is the Stieltjes transform of the limiting spectral measure.
• The normalized eigenvector fluctuations $\sqrt{n}(\widehat{\Pi} - \Pi)$ converge in distribution to a complex Gaussian vector with covariance matrix $R$, as shown via a central limit theorem for linear spectral statistics.
• The proposed estimator maintains consistency in the double-asymptotic regime, with $\|\widehat{\Pi} - \Pi\| \to 0$ almost surely, despite the non-vanishing spectral norm in finite $N,n$.
• The asymptotic distribution of the localization function $\chi_{\text{new}}(\varphi)$ is derived, showing that its maximum values (at the true angles $\varphi_k$) are asymptotically normal with variance depending on the geometry of the steering vectors.
• The first-order asymptotic bias of the estimator is shown to be $\mathcal{O}(1/\sqrt{n})$, and the second-order fluctuations are characterized by a CLT with explicit covariance structure.
• The convergence of $B^*B \to I_r$, $n^{-1}B^*B' \to -\imath cD/2 \, I_r$, and $n^{-2}(B')^*B' \to c^2D^2/3 \, I_r$ implies that the perturbation matrix $B$ behaves asymptotically like an isometry, enabling the derivation of the limiting eigenspace behavior.