Second-Order Cyclostationarity

The defining mathematical property: the autocorrelation function $R_x(t, \tau)$ is periodic in the time variable $t$ with a fundamental period $T_0$.

• Key Equation: $R_x(t + T_0, \tau) = R_x(t, \tau)$
• This periodicity allows the autocorrelation to be expanded into a Fourier series.
• The Fourier coefficients are the Cyclic Autocorrelation Function at cyclic frequencies $\alpha = k/T_0$.
• This contrasts with a wide-sense stationary process, where $R_x(t, \tau)$ is independent of $t$.

In the frequency domain, second-order cyclostationarity manifests as correlation between distinct spectral components of the signal.

• Frequency-shifted versions of the signal, separated by a cyclic frequency $\alpha$, exhibit non-zero correlation.
• This is quantified by the Spectral Correlation Function (SCF).
• For a stationary signal, spectral components at different frequencies are uncorrelated.
• The SCF is a two-dimensional function: $S_x^\alpha(f)$, where $f$ is spectral frequency and $\alpha$ is cyclic frequency.

Second-order cyclostationarity is naturally produced when a stationary signal passes through a Linear Periodically Time-Varying (LPTV) system.

• Many communication operations are inherently LPTV:
  • Modulation: Multiplying a stationary message with a periodic carrier.
  • Multiplexing: Time-division multiplexing imposes a periodic gating structure.
  • Sampling/Scanning: Periodic sampling of a continuous-time signal.
• This causal link makes cyclostationarity an intrinsic fingerprint of man-made modulated signals.

For digitally modulated signals, cyclostationarity is concentrated at specific, discrete cyclic frequencies directly related to the signal's physical parameters.

• Symbol Rate ($R_s$): Cyclic features appear at $\alpha = kR_s$, where $k$ is an integer.
• Carrier Frequency ($f_c$): Features appear at $\alpha = 2f_c$ and $\alpha = 2f_c \pm kR_s$.
• Guard Intervals: OFDM cyclic prefixes induce features at $\alpha = k/T_s$, where $T_s$ is the total symbol duration.
• This discrete structure enables blind parameter estimation by peak detection in the cyclic spectrum.

A primary advantage of second-order cyclostationary processing is its inherent robustness to stationary noise and interference.

• Thermal noise is wide-sense stationary and exhibits no spectral correlation at non-zero cyclic frequencies.
• The Spectral Correlation Function of noise is zero for $\alpha \neq 0$.
• By analyzing the signal at a cyclic frequency $\alpha > 0$, the noise contribution is theoretically eliminated.
• This provides a significant signal-to-noise ratio advantage over conventional energy detection.

The optimal linear estimator for a cyclostationary signal is not time-invariant but periodically time-varying, known as the Cyclic Wiener Filter.

• It exploits the correlation between frequency-shifted signal versions.
• The filter design requires solving Cyclic Wiener-Hopf equations.
• This leads to FRESH (FREquency-SHift) filtering, which jointly filters multiple frequency-shifted copies of the input.
• FRESH filters can separate spectrally overlapping signals if they have distinct cyclic frequencies.

A core application for non-cooperative signal analysis. By computing the cyclic autocorrelation function and searching for peaks at non-zero cyclic frequencies (α), the symbol rate can be extracted without any prior knowledge of the carrier frequency or modulation type.

• Mechanism: A digitally modulated signal exhibits cyclostationarity at cycle frequencies equal to integer multiples of the symbol rate.
• Key Tool: The Dandawate-Giannakis Test is used to statistically confirm the presence of a cyclic feature at a candidate alpha.
• Example: A spectrum monitor intercepts an unknown QPSK signal. The cyclic autocorrelation reveals a strong peak at α = 1.25 MHz, directly indicating the baud rate.

Traditional energy detectors fail in low SNR or when interference overlaps in frequency. A multi-cycle detector exploits the unique cyclic signature of a signal to separate it from stationary noise and co-channel interference.

• Principle: Noise is generally stationary (no cyclostationarity), while modulated signals are not. A detector looking for a specific cyclic frequency can 'see through' the noise.
• Implementation: The Spectral Correlation Function (SCF) is estimated using the FAM Algorithm or SSCA Algorithm to generate a 2D map of spectral correlation.
• Advantage: Can detect and classify signals at negative SNR where an energy detector is completely blind.

Orthogonal Frequency-Division Multiplexing (OFDM) signals are engineered with a cyclic prefix that induces a strong, specific form of second-order cyclostationarity. This is exploited for blind parameter extraction.

• Detection: The repetition of the cyclic prefix creates a correlation peak at a cycle frequency equal to the subcarrier spacing.
• Parameter Estimation: The cyclic autocorrelation can be used to blindly estimate the useful symbol duration (Tu) and the guard interval length (Tg) without demodulating the signal.
• Application: An LTE scanner uses this to identify the presence of a cell and its basic timing parameters before any synchronization.

Before deep learning, and still as a powerful hybrid approach, a cyclic feature vector is constructed from the cyclic spectrum or cyclic cumulants to serve as input to a classifier.

• Feature Extraction: The alpha profile (a slice of the SCF at a fixed f) or the magnitudes of peaks in the cyclic bispectrum provide a compact, discriminative signature.
• Robustness: These features are theoretically immune to any stationary Gaussian noise, making the classifier highly robust.
• Modern Use: These hand-crafted features are often fused with deep learning embeddings to create explainable and data-efficient classifiers.