Dandawate-Giannakis Test

The test frames cyclostationarity detection as a choice between two competing hypotheses:

• H₀ (Null Hypothesis): The signal exhibits no cyclostationarity at the candidate cyclic frequency α. The cyclic spectral density is zero.
• H₁ (Alternative Hypothesis): The signal is cyclostationary at cyclic frequency α. The cyclic spectral density is non-zero.

This rigorous statistical framing allows the test to output a confidence level (p-value) rather than a simple heuristic metric, making it suitable for autonomous spectrum analysis where false alarms must be strictly controlled.

The Dandawate-Giannakis test operates entirely in the frequency domain, constructing a test statistic from the estimated cyclic spectral density. The process involves:

• Computing the cyclic periodogram from finite-time Fourier transforms of the signal.
• Forming a vector of cyclic spectral estimates across multiple frequency bins.
• Deriving an asymptotic covariance matrix of these estimates under the null hypothesis.
• Constructing a chi-squared distributed test statistic that measures the normalized distance of the observed cyclic spectrum from zero.

This frequency-domain approach naturally handles the inherent correlation between spectral estimates at different frequencies.

A critical property of the test is that its statistic converges to a central chi-squared distribution under the null hypothesis as the observation length increases. Key implications:

• The degrees of freedom equal twice the number of frequency bins used in the cyclic spectrum estimate.
• This known distribution enables precise p-value calculation and threshold setting for any desired false alarm rate.
• The asymptotic guarantee holds for a wide class of mixing processes, not just ideal Gaussian signals.

This property transforms cyclostationarity detection from an ad-hoc peak-finding exercise into a principled statistical decision.

The test's power derives from its careful handling of the covariance structure of cyclic spectral estimates. The method:

• Derives a closed-form expression for the asymptotic covariance matrix of the cyclic periodogram.
• Accounts for the correlation between different frequency bins that arises from finite observation windows.
• Uses a consistent estimator of this covariance matrix computed directly from the data.
• Avoids the assumption of independent spectral estimates, which would lead to incorrect threshold settings.

This rigorous covariance modeling is what distinguishes the Dandawate-Giannakis test from simpler, less reliable ad-hoc detection methods.

The test is designed to function without requiring prior knowledge of the noise power spectral density. This blind operation is achieved because:

• Under H₀, the cyclic spectral estimate has an expected value of zero regardless of the stationary noise floor.
• The covariance matrix estimation procedure automatically captures the local noise characteristics from the data.
• The test remains valid for colored and non-Gaussian noise, provided the process satisfies mild mixing conditions.

This robustness makes it ideal for practical spectrum monitoring where the noise environment is dynamic and unknown.

The framework supports both focused and comprehensive detection strategies:

• Single-cycle test: Evaluates evidence for cyclostationarity at one specific candidate cyclic frequency α. Used when searching for a known signal type with a characteristic cycle frequency.
• Multi-cycle test: Jointly tests for the presence of cyclostationarity across a set of candidate cyclic frequencies. This combines evidence from multiple cycle frequencies to improve detection sensitivity for signals with rich cyclic signatures.

The multi-cycle variant is particularly powerful for modulation classification, where different schemes exhibit distinct patterns of cyclic frequencies.