Spectral Correlation Function (SCF)

The SCF maps a signal's energy as a function of two independent variables: spectral frequency (f) and cyclic frequency (α). This dual-frequency plane exposes correlations between spectral components separated by α/2, which are invisible in a standard power spectral density plot. The result is a unique surface plot where non-zero values at α ≠ 0 directly indicate the presence of cyclostationarity.

A critical property of the SCF is its ability to separate signals based on their unique cyclic signatures. Stationary noise and interference have no spectral correlation, meaning their SCF is zero for all α ≠ 0. By evaluating the SCF at a target signal's known cyclic frequency (e.g., the symbol rate), the function naturally filters out stationary background noise, enabling robust feature extraction in highly negative Signal-to-Noise Ratio (SNR) environments.

Each modulation scheme generates a distinct pattern of spectral correlation. The SCF acts as a unique fingerprint:

• BPSK: Strong features at α = 2f_c + kR_s
• QPSK/OQPSK: Features at cycle frequencies related to the symbol rate, but with different harmonic structures.
• MSK/GMSK: Specific patterns tied to the frequency deviation and Gaussian filter parameters. These deterministic patterns make the SCF a foundational tool for Automatic Modulation Classification (AMC).

The SCF enables the extraction of critical signal parameters without prior knowledge. By detecting the location of peaks in the cyclic frequency (α) axis, one can directly estimate:

• Symbol Rate (R_s): Identified by the first strong peak at α = R_s.
• Carrier Frequency (f_c): Estimated by analyzing the symmetry of features around the carrier in the spectral frequency (f) axis. This blind estimation capability is essential for spectrum monitoring and cognitive radio.

Direct computation of the SCF is intensive. The FFT Accumulation Method (FAM) is the standard estimator, using a channelizer and short-time FFTs to trade off resolution and complexity. Key properties of the estimate include:

• Resolution Product: Controlled by Δf (spectral resolution) and Δα (cyclic resolution), where ΔfΔα ≈ M/N for an N-point FFT and M channels.
• Reliability: The estimate's variance decreases as the time-frequency smoothing product increases, requiring careful parameter selection for reliable cyclic feature vectors.

The SCF is the Fourier transform of the cyclic autocorrelation function (CAF). This Wiener-Khinchin-like relation for cyclostationary processes means:

• The SCF, denoted S_x^α(f), represents the density of correlation in the frequency domain.
• The CAF, denoted R_x^α(τ), represents the correlation in the time domain.
• A peak in the SCF at (f, α) corresponds directly to a periodic component in the time-varying autocorrelation, providing a dual-domain view of the signal's hidden periodicity.

A normalized version of the SCF that quantifies the degree of correlation between frequency-shifted signal components on a scale from 0 to 1. Unlike the raw SCF, coherence removes the influence of signal power, making it a scale-invariant feature ideal for modulation classifiers operating under varying signal-to-noise ratios. Values near 1 indicate strong cyclostationarity at a given cyclic frequency.

The time-domain counterpart to the SCF, measuring the correlation of a signal with a frequency-shifted and conjugated version of itself. Parameterized by cyclic frequency (α) and time lag (τ), it reveals hidden periodicities in the signal's second-order statistics. The SCF is derived directly from this function via Fourier transform over the time variable.

The FFT Accumulation Method is the workhorse for practical SCF estimation. It uses a channelizer to decompose the signal into narrowband components, then computes correlations between frequency-shifted channels via short-time FFTs. Key advantages:

• Computationally efficient compared to direct estimation
• Enables real-time cyclic feature extraction
• Trades off cycle frequency resolution against spectral frequency resolution

A compact set of discriminative features derived from the SCF at specific cycle frequencies characteristic of a modulation scheme. For example, a BPSK signal exhibits strong cyclostationarity at cycle frequency α = 2fc ± symbol rate. These vectors serve as input to statistical classifiers or neural networks, providing robust features that are immune to stationary noise.

A one-dimensional slice of the SCF taken at a fixed spectral frequency (f), showing the magnitude of spectral correlation across all cyclic frequencies (α). Alpha profiles serve as compact signatures for modulation recognition:

• BPSK: Strong peaks at α = 2fc and α = 2fc ± Rs
• QPSK: Peak at α = 2fc, but suppressed at symbol rate
• 16-QAM: Distinct pattern of higher-order cyclic features

A statistical hypothesis test formulated in the frequency domain to detect the presence of cyclostationarity at a specific cyclic frequency. It constructs a chi-squared test statistic from the estimated cyclic spectrum, providing a principled method for determining whether an observed spectral correlation is statistically significant or merely a noise artifact. Essential for blind signal detection.