Cyclic Domain Profile

The profile is a 2D map where the x-axis represents spectral frequency (f) and the y-axis represents cycle frequency (α). Correlation magnitude is plotted as intensity. For a stationary signal, all energy collapses onto the α=0 line. For a cyclostationary signal, discrete peaks appear at non-zero α values corresponding to periodicities like the symbol rate or carrier frequency offsets.

Stationary noise and interference have no spectral correlation—their cyclic features exist only at α=0. By analyzing the profile at α ≠ 0, the classifier operates in a domain where thermal noise is theoretically absent. This provides a fundamental signal-to-noise ratio advantage over conventional power spectral density analysis, enabling classification at negative SNR conditions.

Each modulation scheme generates a distinct pattern of cyclic peaks. Key discriminative features include:

• BPSK: Strong peaks at α = ±2f_c and α = symbol rate
• QPSK: Suppressed carrier peaks, dominant symbol-rate features
• 16-QAM: Complex higher-order cyclic cumulant patterns
• OFDM: Peaks at α = subcarrier spacing and cyclic prefix interval
• GMSK: Broadened features due to continuous-phase modulation

An Alpha Profile is a 1D slice of the cyclic domain profile at a fixed spectral frequency f. It plots correlation magnitude versus cycle frequency α. This reduced representation is computationally efficient and captures the essential periodicities of the signal. Peaks in the alpha profile directly correspond to symbol rate, chip rate, or frame rate, enabling blind parameter estimation without demodulation.

Two primary algorithms compute the cyclic domain profile from sampled IQ data:

• FAM (FFT Accumulation Method): Uses a channelizer and short-time FFTs for efficient estimation. Trades spectral resolution for computational speed.
• SSCA (Strip Spectral Correlation Analyzer): Processes data in strips, offering different resolution-complexity trade-offs. Both produce a consistent estimate of the Spectral Correlation Function (SCF) as observation time increases.

For machine learning classifiers, the raw cyclic domain profile is reduced to a Cyclic Feature Vector. This involves:

• Selecting the N most prominent peaks in the alpha profile
• Computing cyclic cumulants at specific (α, f) coordinates
• Extracting the degree of cyclostationarity as a scalar metric
• Applying principal component analysis (PCA) for dimensionality reduction This compact vector feeds directly into SVM, k-NN, or neural network classifiers.

The cyclic domain profile provides a discriminative feature vector for automatic modulation classifiers. By extracting the magnitude of spectral correlation at specific cycle frequencies (the alpha profile), a neural network can distinguish between modulation types like BPSK, QPSK, and 16-QAM without prior knowledge of the signal's carrier frequency or symbol rate. This method is inherently robust to stationary noise and interference, as noise exhibits no cyclostationarity.

Key physical-layer parameters can be extracted directly from the cyclic domain profile without demodulating the signal. This is critical for spectrum surveillance and cognitive radio.

• Symbol Rate Estimation: Identified by the cyclic frequency (α) at which a peak occurs in the cyclic autocorrelation function.
• Carrier Frequency Offset: Estimated by analyzing the symmetry and location of features in the spectral correlation function (SCF).
• OFDM Guard Interval: The length of the cyclic prefix is revealed by a distinct cyclic signature at α = 1/Ts.

Traditional energy detectors fail in low-SNR or high-interference environments. A multi-cycle detector exploits the unique cyclic profile of a signal of interest. By correlating the received signal's cyclic features with a known cyclic signature, the detector can reliably identify the presence of a specific signal even when it is buried beneath stronger, spectrally overlapping interferers. This is a foundational technique for spectrum sharing in dynamic spectrum access networks.

When multiple signals overlap in both time and frequency, conventional linear filters cannot separate them. Frequency-Shift (FRESH) filtering uses the cyclic domain profile to design a Linear Periodically Time-Varying (LPTV) filter. By processing multiple frequency-shifted versions of the input, a FRESH filter can extract a target cyclostationary signal while suppressing spectrally overlapping interferers that have a different cyclic profile. This is a powerful tool for co-channel interference mitigation.

The Cyclic MUSIC algorithm leverages cyclostationarity to overcome the fundamental aperture limit of antenna arrays. Standard MUSIC can resolve at most N-1 sources with N antennas. By exploiting the unique cyclic frequencies of different signals, Cyclic MUSIC can resolve more sources than antennas by separating them in the cyclic domain. This enables compact direction-finding systems for electronic warfare support and spectrum enforcement.

While the cyclic profile primarily identifies the modulation scheme, subtle, transmitter-specific variations in the profile caused by hardware imperfections (e.g., I/Q imbalance, power amplifier non-linearity) can be used for RF fingerprinting. These hardware-induced distortions create unique, measurable perturbations in the higher-order cyclic cumulants, allowing a system to distinguish between two identical radio models transmitting the same modulation. This provides a physical-layer authentication mechanism.