Covariance Matrix Detection

The core mechanism begins by computing the sample covariance matrix from a finite number of received signal samples. For a sensing receiver with M antennas and N sample snapshots, the received signal matrix X is used to estimate the statistical covariance matrix R̂ = (1/N)XX^H. This matrix captures the correlation structure between antenna elements: under H₀ (noise-only hypothesis), the off-diagonal elements approach zero as N increases, while under H₁ (signal-present hypothesis), the signal's inherent temporal or spatial correlation creates non-zero off-diagonal terms. The computation is performed over a sensing window, and the matrix is typically normalized to ensure numerical stability before test statistic extraction.

Detection decisions rely on extracting a scalar test statistic from the covariance matrix that quantifies the degree of off-diagonal correlation. Common formulations include:

• Covariance Absolute Value (CAV): Ratio of the sum of all matrix element magnitudes to the sum of diagonal elements
• Covariance Frobenius Norm (CFN): Ratio of the matrix Frobenius norm squared to the trace squared
• Maximum-Minimum Eigenvalue (MME): Ratio of the largest to smallest eigenvalue, derived from random matrix theory
• Generalized Likelihood Ratio Test (GLRT): Statistic derived under the assumption of unknown signal parameters Each statistic is compared against a pre-calculated threshold to declare signal presence or absence.

A defining advantage of covariance-based detection is its inherent robustness to noise uncertainty. Traditional energy detection suffers from the SNR wall phenomenon, where unknown noise floor fluctuations create a fundamental detection limit below which reliable sensing becomes impossible. Covariance methods bypass this limitation because they rely on the structural difference between signal and noise rather than absolute energy levels. Since noise samples are statistically independent across antennas or time, their covariance matrix is diagonal, while correlated signals produce off-diagonal energy. This structural distinction persists even when the exact noise power is unknown, enabling reliable detection at significantly lower SNRs than energy detection.

In multi-antenna (MIMO) sensing systems, covariance detection exploits the spatial correlation introduced by the primary user's signal arriving at multiple receive antennas. A transmitted signal impinging on an antenna array creates a rank-1 spatial signature in the covariance matrix, while independent thermal noise at each antenna element contributes only to the diagonal. The technique is particularly effective when:

• Antenna spacing is less than the coherence distance
• The propagation channel exhibits spatial structure
• The number of antennas M is sufficiently large to provide statistical separation This spatial approach eliminates the need for temporal oversampling and works with stationary signal models.

For single-antenna receivers, covariance detection is enabled through temporal oversampling of the received signal. By sampling at a rate higher than the Nyquist rate, adjacent time samples become correlated when a primary user signal is present due to the signal's non-flat power spectral density. The receiver forms a smoothed covariance matrix by stacking consecutive time-shifted sample vectors, effectively converting temporal correlation into a matrix structure. This approach requires:

• An oversampling factor L that determines the matrix dimension
• Sufficient signal bandwidth to create sample-to-sample correlation
• Careful selection of smoothing factor to balance detection performance and complexity Temporal covariance detection enables blind sensing on cost-constrained single-antenna devices.

Setting detection thresholds requires asymptotic random matrix theory (RMT) to analytically characterize the distribution of eigenvalues under the noise-only hypothesis. For large matrix dimensions, the Marchenko-Pastur law describes the limiting spectral distribution of sample covariance matrices of pure noise. Key results include:

• Tracy-Widom distribution: Characterizes fluctuations of the largest eigenvalue, enabling precise false alarm probability control
• Limiting eigenvalue bounds: Provide closed-form expressions for λ_max and λ_min under H₀
• Asymptotic threshold formulas: Allow threshold calculation as a function of matrix dimensions and desired P_fa without requiring noise power knowledge These theoretical foundations ensure that detection thresholds can be set analytically rather than through empirical calibration.