Cyclic Autocorrelation Function (CAF)

The CAF is the primary time-domain statistical function for analyzing second-order periodicity. It computes the correlation between a signal and a frequency-shifted copy of itself at a specific cyclic frequency (α). Unlike the standard autocorrelation function, which assumes stationarity, the CAF explicitly models the periodic time-variation of statistical moments. This makes it the direct precursor to the Spectral Correlation Function (SCF), which is its Fourier transform. The CAF is defined as the Fourier coefficient of the time-varying autocorrelation function, isolating the strength of periodicity at each cyclic frequency.

The CAF is a quadratic transformation of the signal, involving a product of the signal with a conjugated and frequency-shifted version of itself. This nonlinear operation is essential for revealing hidden periodicities that are not visible in the raw waveform or its power spectrum. Key mathematical properties include:

• Conjugate CAF: Uses x(t)x*(t) for standard spectral correlation
• Non-Conjugate CAF: Uses x(t)x(t) for complex-valued signals like QAM
• The frequency shift α/2 is applied symmetrically to both signal copies before correlation This quadratic nature is what allows the CAF to detect modulation-specific features like symbol rates and carrier offsets.

The CAF produces a two-dimensional function R_x^α(τ) parameterized by:

• Cyclic Frequency (α): The frequency separation that reveals periodicity. α=0 recovers the conventional autocorrelation function.
• Lag Parameter (τ): The time delay between the two signal copies.

Distinct signal types exhibit unique CAF patterns at specific cyclic frequencies:

• BPSK: Strong features at α = ±2f_c + k/T_sym
• QPSK: Features at α = k/T_sym (symbol rate harmonics)
• OFDM: Features at α = 1/T_u (useful symbol duration) and α = 1/T_s (total symbol duration) This discriminative capability enables blind modulation classification and emitter identification.

The CAF exhibits inherent robustness to stationary noise and interference. Because stationary Gaussian noise has no cyclostationary content, its contribution to the CAF converges to zero at all non-zero cyclic frequencies as the observation time increases. This property enables:

• Signal detection below the noise floor: Cyclostationary features remain detectable even at negative SNR
• Interference separation: Co-channel signals with different symbol rates or carrier frequencies produce distinct, non-overlapping cyclic frequencies
• Feature stability: The CAF estimate at true cyclic frequencies converges reliably with sufficient data, while noise-only estimates diminish This makes the CAF a powerful tool for spectrum sensing and co-channel signal separation.

Practical CAF estimation requires time-averaging over a finite observation window. The standard cyclic correlogram estimator computes the time-average of the lag product multiplied by a complex exponential at the candidate cyclic frequency. Key implementation considerations:

• Time-smoothing vs. frequency-smoothing: Trade-offs between resolution and variance
• FFT-based implementations: The FAM (FFT Accumulation Method) and SSCA (Strip Spectral Correlation Analyzer) are computationally efficient algorithms that estimate the SCF, from which the CAF can be recovered via inverse FFT
• Resolution: Cyclic frequency resolution is inversely proportional to observation time; lag resolution is determined by sampling rate
• Windowing: Appropriate tapering reduces spectral leakage and cyclic frequency sidelobes

The CAF serves as a rich source of discriminative features for AI-based signal classification systems. Common feature extraction strategies include:

• Cyclic Domain Profile (CDP): Projecting the CAF magnitude along the cyclic frequency axis at τ=0 to create a one-dimensional feature vector
• Lag-Profile Slicing: Extracting CAF slices at specific cyclic frequencies of interest (e.g., symbol rate, carrier offset) to form compact signatures
• Multi-lag Feature Vectors: Concatenating CAF values across multiple lag values at key cyclic frequencies
• Normalized CAF: Using the cyclic correlation coefficient to achieve scale and power invariance These features feed directly into neural networks for automatic modulation classification and RF fingerprinting applications.

The fundamental parameter α in the CAF that defines the frequency shift applied to one copy of the signal before correlation. Key cyclic frequencies correspond to physical signal properties:

• α = 1/Ts: Symbol rate periodicity
• α = 2fc: Carrier-induced cyclostationarity
• α = 2fc ± Rs: Combined carrier and symbol rate features A signal exhibits cyclostationarity at α only if its autocorrelation has a non-zero Fourier coefficient at that frequency.

A one-dimensional projection of the SCF magnitude along the cyclic frequency axis, computed by integrating or taking the maximum across spectral frequency. The CDP serves as a compact feature vector for machine learning classifiers. Peaks in the CDP directly correspond to the cyclic frequencies where the CAF is non-zero, making it an efficient summary of a signal's cyclostationary signature for emitter identification.

The normalized magnitude of the SCF, ranging from 0 to 1, that quantifies the degree of correlation between frequency-shifted signal components independent of signal power. Spectral coherence provides a scale-invariant feature robust to varying received signal strength. A coherence value near 1 at a specific (f, α) pair indicates strong cyclostationarity, while values near 0 suggest stationarity or noise.

The FFT Accumulation Method is the most computationally efficient algorithm for estimating the SCF from the CAF. It works by:

• Channelizing the input signal into narrowband frequency bins via a sliding FFT
• Decimating each bin to reduce sampling rate
• Cross-correlating bin outputs to populate the SCF matrix The FAM reduces complexity from O(N²) to O(N log N), making real-time cyclostationary analysis feasible on SDR platforms.

A higher-order generalization of the CAF that extracts purely non-Gaussian periodic components. While the CAF captures second-order periodicity, cyclic cumulants of order n > 2 isolate nonlinear signal features that are completely immune to Gaussian noise. This makes them powerful for modulation classification in low-SNR environments where second-order cyclostationary features may be buried in noise.