Cyclic Spectrum

The cyclic spectrum S_x^α(f) is a function of two independent frequency variables. The spectral frequency (f) represents the standard Fourier frequency axis, while the cyclic frequency (α) quantifies the periodicity of the signal's second-order statistics. A non-zero value at a specific (f, α) pair indicates that the spectral components at frequencies f + α/2 and f - α/2 are statistically correlated. This two-dimensional representation is the Fourier transform of the cyclic autocorrelation function, converting a time-domain correlation into a frequency-domain density.

Each digital modulation scheme produces a unique and deterministic cyclic spectrum pattern. For example:

• BPSK exhibits strong cyclic features at cycle frequencies α = k/T_s (symbol rate harmonics) and α = ±2f_c + k/T_s (carrier-related features).
• QPSK suppresses the carrier-related features present in BPSK but retains symbol-rate cyclostationarity.
• 16-QAM displays distinct cyclic features at α = k/T_s with a characteristic amplitude profile across spectral frequency. This uniqueness makes the cyclic spectrum an ideal feature space for automatic modulation classification, as the pattern of spectral correlation peaks serves as a fingerprint for the transmission scheme.

A critical property of the cyclic spectrum is its ability to separate signals based on their cyclic frequencies. Stationary noise and interference exhibit spectral correlation only at α = 0, while modulated signals concentrate their cyclostationary features at non-zero cycle frequencies. This means:

• Gaussian noise contributes only to the α = 0 plane of the cyclic spectrum.
• Co-channel interference with a different symbol rate or modulation type will have cyclic features at distinct α values. By analyzing the cyclic spectrum at α ≠ 0, a classifier can effectively operate in negative signal-to-noise ratio conditions where conventional power spectral density methods fail entirely.

The cyclic spectrum enables blind extraction of physical-layer parameters without prior knowledge of the signal. Key parameters are estimated directly from the locations and magnitudes of cyclic spectrum peaks:

• Symbol Rate (1/T_s): Identified by detecting the cyclic frequency α at which the first harmonic of the symbol-rate feature appears.
• Carrier Frequency (f_c): Estimated from the spectral frequency location of carrier-related cyclostationary features.
• Pulse Shaping Roll-off: The spectral frequency width of cyclic features at α = 1/T_s correlates with the excess bandwidth of the pulse-shaping filter. These estimates provide the foundational parameters required for subsequent demodulation and detailed signal analysis.

The FFT Accumulation Method (FAM) is the dominant algorithm for practical cyclic spectrum estimation. It operates by:

• Channelizing the input signal into narrowband frequency components using a sliding short-time FFT.
• Computing complex demodulates by frequency-shifting the channelized data.
• Accumulating cross-correlations between demodulates separated by the cyclic frequency α. The FAM algorithm reduces computational complexity from O(N²) for direct estimation to O(N log N), making real-time cyclic spectrum analysis feasible on FPGA and GPU hardware. The output is a discretized cyclic spectrum matrix with resolution determined by the channelizer bandwidth and observation time.

The cyclic domain profile is a reduced-dimension representation of the cyclic spectrum, typically computed by integrating the magnitude of S_x^α(f) over spectral frequency f for each cyclic frequency α. This produces a one-dimensional function that:

• Collapses the 2D cyclic spectrum into a plot of cycle frequency versus aggregate spectral correlation strength.
• Reveals all active cyclic frequencies at a glance, including symbol rate harmonics and carrier-related features.
• Serves as a compact feature vector for machine learning classifiers, significantly reducing input dimensionality while preserving the discriminative cyclostationary information. The alpha profile is the most common form of cyclic domain profile, representing a slice at a fixed spectral frequency rather than an integration.

The cyclic spectrum provides a robust method for blind symbol rate estimation without prior knowledge of the transmission. By searching for peaks in the alpha profile (a slice of the cyclic spectrum at a fixed spectral frequency), the symbol rate manifests as a strong cyclic frequency. This technique is resilient to stationary noise and interference, making it superior to simple envelope detection in negative SNR environments. Key steps include:

• Computing the Spectral Correlation Function (SCF) via the FAM Algorithm
• Extracting the alpha profile at the carrier frequency
• Identifying the non-zero cyclic frequency peak corresponding to the symbol rate

Traditional energy detectors fail in the presence of co-channel interference. The Multi-Cycle Detector leverages the cyclic spectrum to distinguish signals based on their unique cyclic signatures. Because different modulation schemes exhibit correlation at distinct cycle frequencies, a detector that integrates energy over multiple alpha-plane peaks can reliably detect a weak signal of interest even when overlapped by a stronger, spectrally coincident interferer. This is a foundational technique for Dynamic Spectrum Awareness in cognitive radio.

When signals overlap in both time and frequency, linear time-invariant filters cannot separate them. FRESH (FREquency-SHift) filtering exploits the cyclic spectrum to perform signal separation. By processing multiple frequency-shifted versions of the received waveform and weighting them adaptively, a FRESH filter can extract a target signal while suppressing spectrally overlapping interference. This is modeled as a Linear Periodically Time-Varying (LPTV) System, where the periodicity of the filter matches the cyclostationarity of the desired signal.

The cyclic prefix in OFDM signals induces a strong second-order cyclostationarity at the OFDM symbol rate. The cyclic spectrum reveals a distinct pattern of spectral correlation at cycle frequencies equal to integer multiples of the subcarrier spacing. This allows for:

• Blind OFDM signal identification without demodulation
• Estimation of the useful symbol duration and guard interval length
• Distinguishing between different OFDM-based standards (e.g., LTE vs. Wi-Fi) based on their unique timing parameters

The Cyclic MUSIC algorithm extends the standard MUSIC direction-finding technique by exploiting signal cyclostationarity. By focusing on the signal subspace at a specific cyclic frequency, Cyclic MUSIC can resolve more source signals than the number of antenna elements in the array. This is because it can separate signals with overlapping spectra but distinct cyclic signatures, effectively increasing the degrees of freedom of the array and enabling high-resolution spatial processing in dense electromagnetic environments.

The Spectral Correlation Function is the fundamental two-dimensional transform that defines the cyclic spectrum. It measures the density of correlation between two frequency-shifted versions of a signal, specifically at frequencies f + α/2 and f - α/2. For a cyclostationary signal, the SCF is non-zero only for specific discrete values of the cyclic frequency (α), creating a unique, machine-readable signature. This function is the primary tool for separating overlapping signals in the frequency domain because stationary noise and interference exhibit no spectral correlation (SCF is zero for α ≠ 0).

The time-domain counterpart to the cyclic spectrum. The Cyclic Autocorrelation Function is computed as the Fourier coefficient of the time-varying autocorrelation function. It quantifies the correlation between a signal and a frequency-shifted version of itself, parameterized by the cyclic frequency (α). The Fourier transform of this function yields the cyclic spectrum. Key properties include:

• For α = 0, it reduces to the conventional autocorrelation.
• Discrete non-zero values for α indicate the presence of second-order cyclostationarity.
• It is the optimal feature basis for blind symbol rate estimation.

The FFT Accumulation Method (FAM) is the most widely used computationally efficient estimator for the cyclic spectrum. It works by:

• Channelizer: Decomposing the input signal into narrowband frequency channels via a short-time FFT.
• Decimation: Downsampling each channel output to the data rate.
• Cross-Correlation: Computing the complex demodulates between channel pairs separated by the cyclic frequency α. This striping technique trades off temporal resolution for spectral resolution, making real-time cyclic spectrum estimation feasible on FPGAs and DSPs.

A normalized version of the Spectral Correlation Function that provides a dimensionless measure of correlation strength. The Spectral Coherence magnitude is bounded between 0 and 1, where:

• 1 indicates perfect correlation between frequency-shifted components.
• 0 indicates no correlation (typical of stationary noise). This normalization removes the dependency on the signal's power spectral density, making it an ideal feature for modulation classification in environments with variable signal-to-noise ratios. It is often used as the input feature vector for deep learning classifiers.

A higher-order statistic that extends the concept of cumulants to cyclostationary signals. While the cyclic spectrum is a second-order measure, cyclic cumulants capture the periodic behavior of third-order and fourth-order moments. They are critical for:

• Classifying modulation schemes that have identical second-order cyclostationary features (e.g., QPSK vs. 16QAM).
• Providing absolute robustness against Gaussian noise, as all cumulants of order > 2 are theoretically zero for Gaussian processes.
• Forming the basis for higher-order cyclostationarity analysis.

A one-dimensional slice of the cyclic spectrum at a fixed spectral frequency f. The Alpha Profile plots the magnitude of spectral correlation as a function of the cyclic frequency α. This representation is extremely powerful for blind parameter extraction:

• Peaks at α = k/Tₛ reveal the symbol rate.
• Peaks at α = 2f_c reveal the carrier frequency offset.
• The pattern of peaks serves as a compressed cyclic signature for rapid modulation identification without requiring the full two-dimensional SCF computation.