Cyclic Stationarity Test

The test formulates a binary hypothesis test where the null hypothesis (H0) states the signal is wide-sense stationary (no cyclostationarity present), while the alternative hypothesis (H1) asserts the signal exhibits cyclostationarity at the candidate cyclic frequency α.

• Test statistic: Derived from the cyclic autocorrelation estimate
• Threshold: Set based on desired false alarm probability
• Decision rule: Compare test statistic magnitude against threshold

The framework provides a rigorous statistical basis for declaring the presence of periodic structure in the signal's second-order moments.

Under the null hypothesis of stationarity, the cyclic autocorrelation estimate is asymptotically complex normal with zero mean. This property enables the construction of test statistics with known distributions.

• Covariance matrix: Depends on the signal's power spectral density
• Chi-squared distribution: Squared magnitude of normalized cyclic autocorrelation follows χ²(2) distribution
• Sample size dependence: Larger observation windows improve the Gaussian approximation

The asymptotic normality holds as the number of time samples approaches infinity, providing a tractable statistical framework for threshold calculation.

Cyclostationarity tests can be categorized by the number of cyclic frequencies evaluated simultaneously.

Single-cycle test:

• Tests for cyclostationarity at one specific cyclic frequency α
• Computationally efficient for known candidate frequencies
• Used when symbol rate or carrier offset is approximately known

Multi-cycle test:

• Jointly tests multiple cyclic frequencies simultaneously
• Higher detection power by aggregating evidence across frequencies
• Essential for blind detection when cyclic frequencies are unknown
• Exploits the fact that cyclostationary signals typically exhibit periodicity at multiple harmonically related frequencies

A key advantage of cyclostationarity tests over energy detection is their inherent robustness to noise uncertainty. Unlike energy detectors that require precise noise power knowledge, cyclostationarity tests exploit the fact that stationary noise exhibits no cyclic correlation at non-zero cyclic frequencies.

• Noise floor independence: Test statistic is theoretically zero-mean for stationary noise regardless of noise power
• Interference rejection: Co-channel interferers with different cyclic frequencies are naturally separated
• Low-SNR operation: Reliable detection possible at negative signal-to-noise ratios

This property makes cyclostationarity tests particularly valuable for spectrum sensing and signal detection in uncertain noise environments.

Practical implementation of cyclostationarity tests requires efficient estimation of the cyclic autocorrelation function from finite data records.

• Time-smoothing method: Averages cyclic periodograms over time using the FFT Accumulation Method (FAM) or Strip Spectral Correlation Analyzer (SSCA)
• Frequency-smoothing method: Averages in the frequency domain to reduce estimator variance
• Covariance estimation: Requires estimation of the asymptotic covariance matrix for proper normalization
• Windowing effects: Choice of data tapering window affects bias-variance tradeoff

Computational complexity scales with the number of cyclic frequencies tested and the desired spectral resolution.

The performance of cyclostationarity tests is characterized by standard detection theory metrics evaluated as functions of observation length and signal-to-noise ratio.

• Probability of detection (Pd): Increases with observation time and SNR; approaches 1 for sufficiently long records
• Probability of false alarm (Pfa): Controlled by threshold selection based on asymptotic distribution
• Receiver operating characteristic (ROC): Quantifies the Pd vs. Pfa tradeoff
• Sample complexity: Minimum number of samples required to achieve target Pd at given Pfa

Empirical evaluation using Monte Carlo simulations is standard practice to validate theoretical performance predictions under realistic channel conditions.