Eigenvalue ratio detection based on exact moments of smallest and largest eigenvalues

Muhammad Zeeshan Shakir, Wuchen Tang, Anlei Rao, Muhammad Ali Imran, Mohamed-Slim Alouini

Abstract

Spectrum sensing is a critical component of cognitive radio, enabling secondary users to identify and utilize vacant frequency bands. This paper introduces a high-precision Eigenvalue Ratio (ER) detection method that remains reliable even in small-scale deployments with few users and limited signal samples. By deriving the exact analytical moments of extreme eigenvalues and accounting for their statistical dependency via Copula theory, the authors provide a closed-form decision threshold that significantly outperforms standard asymptotic approximations like the Tracy-Widom distribution.

Introduction

In cognitive radio networks, "secondary users" must detect the presence of a "primary user" to avoid harmful interference. Eigenvalue Ratio (ER) detection—the ratio of the largest eigenvalue ($\lambda_1$) to the smallest eigenvalue ($\lambda_K$) of the received signal covariance matrix—is a popular "blind" sensing technique because it requires no prior knowledge of the primary user's signal characteristics.

However, the mathematical distribution of this ratio is notoriously difficult to characterize precisely. Most existing solutions rely on asymptotic assumptions, utilizing mathematical shortcuts like the Tracy-Widom distribution. These approximations only become accurate as the number of users ($K$) and samples ($N$) approach infinity. In practical, small-scale deployments (e.g., a few sensors taking a few dozen samples), these shortcuts lead to inaccurate decision thresholds, causing either missed detections or excessive false alarms.

Methodology

The authors propose a non-asymptotic Gaussian approximation approach to bridge the gap between theory and practical small-scale performance. Their methodology follows four key steps:

  1. Exact Moment Derivation: Instead of relying on limit theorems, the researchers derive the exact analytical $p$-th moments for the largest ($\lambda_1$) and smallest ($\lambda_K$) eigenvalues. This ensures that the underlying statistics are accurate regardless of the system size.
  2. Bivariate Gaussian Modeling: They approximate the joint Probability Density Function (PDF) of these extreme eigenvalues as a bivariate Gaussian distribution.
  3. Dependency Analysis via Copulas: A major contribution of this work is the use of Copula theory to analyze the dependency between $\lambda_1$ and $\lambda_K$. While previous research often assumed these eigenvalues were independent to simplify the mathematics, this paper demonstrates they are significantly correlated in non-asymptotic scenarios.
  4. Closed-Form Threshold Formulation: Using the ratio of two dependent Gaussian variables, the authors derive a specific analytical formula to calculate the decision threshold $\gamma$ for a target False Alarm probability ($P_{fa}$).

Key Insights: The Independence Fallacy

A central finding of the paper is the "Independence Fallacy." The authors prove that the largest and smallest eigenvalues are not independent. For small systems (e.g., $K=2, N=10$), the correlation coefficient $\rho$ can be as high as 0.3. Ignoring this correlation leads to significant errors in threshold calculation. By plotting Copulas (scatter plots of the marginal CDFs), the authors visually demonstrate how the dependency structure of the eigenvalues changes as the system size increases, providing a more intuitive understanding of random matrix behavior.

Results and Performance Evaluation

The proposed model was evaluated using extensive Monte Carlo simulations, generating complex circular Gaussian white noise and signal matrices to create empirical benchmarks.

Accuracy of Analytical Moments

The analytical mean and variance for eigenvalues were found to be nearly identical to simulation results. As shown in Table I, for a configuration of $(K=2, N=4)$, the analytical mean for $\lambda_1$ is 6.1875, matching the simulated value of 6.1857 almost perfectly. In contrast, the Tracy-Widom approximation predicts a mean of 5.2184, representing a significant deviation that would compromise sensing performance.

Table I: Comparison of simulated, Tracy-Widom, and analytical moments for extreme eigenvalues across different (K, N) configurations.

Threshold Reliability

The ultimate test of the method is the accuracy of the decision threshold $\gamma$ relative to the target false alarm rate $P_{fa}$. Figure 3 illustrates that the proposed analytical curve tracks the simulation data perfectly across all tested scenarios. The Tracy-Widom approach, however, fails to match the empirical data for small and moderate values of $K$ and $N$, highlighting its inadequacy for practical cognitive radio design.

Figure 3: Decision threshold gamma vs. false alarm probability P_fa. The proposed analytical method (red dash-dot) aligns closely with simulation results (solid black), while the Tracy-Widom approximation (dashed black) deviates significantly.

Conclusion

This research provides a robust framework for spectrum sensing that is "equivalently good" for any number of users and samples. By moving beyond asymptotic shortcuts and accounting for the inherent dependency between eigenvalues, the authors have developed a versatile tool for real-world cognitive radio design. Future work may focus on calculating the exact correlation coefficient $\rho$ analytically for all configurations and testing the model's performance in non-Gaussian interference environments.

Summary updated

BibTeX
@inproceedings{shakir2011eigenvalue,
  author = {Muhammad Zeeshan Shakir and Wuchen Tang and Anlei Rao and Muhammad Ali Imran and Mohamed-Slim Alouini},
  title = {{Eigenvalue ratio detection based on exact moments of smallest and largest eigenvalues}},
  booktitle = {CrownCom},
  year = {2011},
  doi = {10.4108/icst.crowncom.2011.246151},
}
Figure