Spectral Correlation Density (SCD)

Unlike the standard Power Spectral Density (PSD) which is a one-dimensional function of frequency, the SCD is a two-dimensional transform $S_x^\alpha(f)$. It measures the correlation between spectral components at frequencies $f + \alpha/2$ and $f - \alpha/2$.

• Frequency axis (f): Represents the standard spectral content.
• Cyclic frequency axis (α): Reveals the hidden periodicities caused by modulation, symbol rates, and coding.
• A signal exhibits cyclostationarity if the SCD is non-zero for $\alpha \neq 0$, creating a unique surface plot that serves as a visual fingerprint for the modulation type.

The SCD's primary advantage in constellation identification is its theoretical immunity to stationary noise and interference. Stationary Gaussian noise has no spectral correlation ($\alpha = 0$ only), meaning its energy collapses to the zero cyclic frequency plane.

• Modulated signals generate distinct correlation patterns at non-zero cyclic frequencies ($\alpha \neq 0$) corresponding to the symbol rate and carrier offset.
• This allows the SCD to cleanly separate a weak signal of interest from a much stronger, overlapping stationary interferer, making it a powerful tool for spectrum awareness in congested environments.

Different modulation families produce SCD peaks at specific cyclic frequencies, acting as a direct identifier for the signal's physical parameters:

• BPSK: Strong peaks at $\alpha = \pm 2f_c$ and $\alpha = \pm 2f_c \pm R_s$ (where $f_c$ is the carrier frequency and $R_s$ is the symbol rate).
• QPSK/QAM: Peaks are concentrated at $\alpha = k/R_s$ (integer multiples of the symbol rate) due to the symbol transitions.
• OFDM: Exhibits a unique cyclic prefix-induced signature at $\alpha = 1/T_u$ (the useful symbol duration).
• These deterministic peak locations allow for blind parameter estimation of the symbol rate and carrier frequency without prior demodulation.

The computationally efficient FFT Accumulation Method (FAM) is the standard algorithm for estimating the SCD from finite IQ sample records. It balances resolution and processing time.

• Channelization: The input signal is first decomposed into narrowband frequency channels using a short-time FFT.
• Decimation: Each channel is downsampled to reduce the computational load.
• Cross-Correlation: The complex envelopes of frequency-separated channels are correlated over time to generate the SCD surface.
• The FAM outputs a matrix where each row is a cyclic spectrum slice at a specific frequency, enabling real-time visualization of cyclostationary features.

While the SCD is a high-dimensional image, it is often reduced to compact feature vectors for input into deep learning classifiers like CNNs or SVMs.

• Alpha Profile: The maximum value of the SCD along the frequency axis for each cyclic frequency, creating a 1D plot that highlights the dominant cycle frequencies.
• Spectral Coherence (SOF): A normalized version of the SCD that is magnitude-independent, making the classifier robust to signal power variations.
• These features are concatenated with cumulants and IQ histograms to form a multi-modal input vector, significantly improving classification accuracy at low signal-to-noise ratios (SNR) compared to using raw IQ alone.

The SCD is inherently resilient to linear time-invariant filtering, including multipath propagation. A frequency-selective fading channel modifies the PSD but preserves the cyclic correlation structure.

• The cyclic frequency locations ($\alpha$) are determined by the transmitter's symbol rate and carrier frequency, which are unaffected by the channel's delay spread.
• The SCD magnitude at these cyclic frequencies is scaled by the channel response, but the presence of the peak is not destroyed.
• This property makes SCD-based classifiers highly effective in non-cooperative environments where channel state information is unknown, such as tactical SIGINT and spectrum enforcement.

The time-domain counterpart to the SCD. It measures the correlation of a signal with a frequency-shifted and conjugated version of itself, averaged over time. The SCD is the Fourier transform of the CAF.

• Quadratic transformation: Like the SCD, the CAF is a second-order statistic.
• Domain duality: The CAF operates in the time-lag domain, while the SCD operates in the spectral-frequency domain.
• Peak locations: Cyclostationary signals exhibit non-zero CAF values at specific cyclic frequencies, corresponding to symbol rates and carrier offsets.

The frequency separation parameter, often denoted by alpha (α), at which spectral components exhibit correlation in a cyclostationary signal. It is the independent variable that distinguishes the SCD from a standard power spectral density.

• Physical origins: Cyclic frequencies arise from periodic phenomena like symbol rates, chip rates, guard intervals, and carrier frequencies.
• Discrete nature: For modulated signals, non-zero SCD values exist only at discrete, predictable cyclic frequencies, forming a unique comb-like signature.
• Stationary noise rejection: Noise is typically stationary and has no cyclic correlation, so its SCD is zero for α ≠ 0, enabling robust signal detection.

A computationally efficient algorithm for estimating the SCD from discrete-time signal samples. It balances time and frequency resolution through channelization.

• Strip spectral correlation: The FAM computes the SCD by first channelizing the signal using a short-time FFT, then correlating the complex outputs of different frequency bins over time.
• Resolution trade-off: The choice of FFT size and block overlap controls the cycle frequency resolution versus spectral frequency resolution.
• Practical implementation: The FAM is the standard method for real-time SCD estimation in software-defined radios due to its FFT-based efficiency.

A higher-order (n > 2) cyclostationary statistic that measures the periodic behavior of a signal's cumulants, which are nonlinear combinations of moments. Cyclic cumulants are immune to Gaussian noise of any color.

• Hierarchical classification: Higher-order cyclic cumulants can distinguish modulation formats that have identical second-order SCD signatures, such as 16-QAM vs. 64-QAM.
• Gaussian immunity: Unlike the SCD, cyclic cumulants of order n > 2 are theoretically zero for any Gaussian process, stationary or not.
• Computational cost: Estimating cyclic cumulants is more complex than the SCD, requiring higher-order products and longer integration times.

The background interference against which the SCD provides a distinct advantage. Stationary processes have a time-invariant autocorrelation function, meaning their spectral components at different frequencies are uncorrelated.

• SCD response: A stationary noise process produces an SCD that is non-zero only on the α = 0 plane, which is the standard power spectral density.
• Feature isolation: By analyzing the SCD at α ≠ 0, the cyclostationary features of a modulated signal are isolated from the stationary noise floor, dramatically improving detection sensitivity.
• Colored noise: This rejection property holds for stationary noise regardless of its spectral shape (white or colored).

A unique, intentionally embedded cyclostationary feature designed for signal identification and network coordination. Unlike inherent modulation features, signatures are artificially created.

• OFDM applications: In cognitive radio, a transmitter can embed a signature by correlating specific subcarriers, creating a unique cyclic frequency peak detectable by other nodes.
• Robust identification: Signatures are designed to be easily distinguishable from naturally occurring cyclostationary features of other signals.
• Low overhead: A signature can be embedded with minimal impact on data throughput by using a small number of dedicated subcarriers or by applying a low-power spreading code.