Spectral Coherence

Spectral coherence is defined as the spectral correlation function normalized by the geometric mean of the power spectral densities at the two frequency-shifted components. This normalization constrains the output to the range [0, 1], where 1 indicates perfect correlation and 0 indicates no correlation. The formula is:

C_X^α(f) = S_X^α(f) / sqrt(S_X(f + α/2) * S_X(f - α/2))

• S_X^α(f): Spectral correlation density at cyclic frequency α and spectral frequency f
• S_X(f ± α/2): Power spectral density at the shifted frequencies
• Scale-invariant: Removes dependency on absolute signal power

Unlike the raw spectral correlation function, spectral coherence is independent of signal amplitude. This makes it a robust feature for emitter identification where received power varies dramatically due to:

• Distance changes: Path loss attenuation over varying ranges
• Fading: Multipath-induced amplitude fluctuations
• Amplifier gain: Different transmit power settings

The coherence value remains stable as long as the underlying cyclostationary structure is preserved, enabling reliable fingerprinting without power normalization preprocessing.

Spectral coherence maps serve as powerful input features for deep learning classifiers. The two-dimensional coherence plane (cyclic frequency α vs. spectral frequency f) reveals:

• Modulation-specific patterns: Unique coherence peaks at symbol rate and carrier offset multiples
• Hardware impairment signatures: Subtle coherence structures from I/Q imbalance and phase noise
• Dimensionality reduction: Key cyclic frequencies can be sampled to form compact cyclic feature vectors

Convolutional neural networks trained on coherence images achieve high accuracy in automatic modulation classification and specific emitter identification tasks.

Spectral coherence provides inherent noise immunity because stationary Gaussian noise has zero cyclostationarity. Key benefits include:

• Noise floor suppression: Additive white Gaussian noise contributes only at α=0, leaving non-zero cyclic frequencies clean
• Interference separation: Co-channel signals with different symbol rates appear at distinct cyclic frequencies
• Low-SNR operation: Reliable feature extraction even at negative signal-to-noise ratios where conventional methods fail

This makes coherence-based detection superior to energy detection for spectrum sensing and signal identification in contested electromagnetic environments.

The Cyclic Domain Profile (CDP) is a one-dimensional projection of spectral coherence obtained by integrating or maximizing along the spectral frequency axis:

CDP(α) = max_f |C_X^α(f)|

This compression:

• Reduces dimensionality: Collapses the 2D coherence map into a 1D vector
• Preserves cyclic frequency information: Peaks remain at symbol rate, carrier offset, and frame rate multiples
• Enables rapid classification: Compact feature vectors suitable for lightweight classifiers on edge hardware

The CDP serves as a computationally efficient alternative to full coherence analysis for real-time emitter identification.

Practical spectral coherence estimation requires time and frequency smoothing to reduce variance from finite observation windows. Two primary algorithms exist:

• FAM (FFT Accumulation Method): Channelizes the signal into narrowband bins, then correlates frequency-shifted outputs. Efficient for wideband analysis
• SSCA (Strip Spectral Correlation Analyzer): Computes complex demodulates and correlates with the original signal. Lower complexity for sparse cyclic frequency sets

Both methods trade temporal resolution against cyclic frequency resolution, governed by the uncertainty principle. The coherence estimate reliability improves with observation length, making it suitable for steady-state signal analysis.

The time-domain counterpart to the SCF that computes the correlation between a signal and a frequency-shifted version of itself at a specific cyclic frequency (alpha). The CAF is non-zero only when the shift matches an inherent periodicity of the signal, such as the symbol rate or chip rate. Fourier transformation of the CAF yields the SCF, establishing the Wiener-Khinchin theorem for cyclostationary processes.

A one-dimensional projection of the spectral coherence magnitude along the cyclic frequency axis. The CDP compresses the two-dimensional SCF into a compact feature vector by integrating or maximizing over the spectral frequency dimension. This representation is highly effective for:

• Modulation recognition using peak pattern matching
• Signal detection with reduced computational complexity
• Emitter identification via distinctive cyclic signature profiles

The FFT Accumulation Method is a computationally efficient channelized algorithm for estimating the SCF. It works by:

• Decomposing the input signal into narrowband frequency channels via a sliding FFT
• Computing cross-correlations between channelized outputs at offsets matching the cyclic frequency
• Accumulating results over time to reduce variance The FAM algorithm trades resolution for speed, making real-time spectral coherence estimation feasible on FPGA and SDR platforms.

A higher-order statistical function that extracts the purely non-Gaussian periodic components of a signal. Unlike second-order methods like spectral coherence, cyclic cumulants of order three and above are theoretically immune to additive Gaussian noise. This makes them exceptionally robust features for:

• Automatic modulation classification in low-SNR environments
• Emitter fingerprinting where noise floors are high
• Nonlinear system identification

FREquency-SHift filtering exploits cyclostationarity by linearly combining multiple frequency-shifted versions of a received signal to separate spectrally overlapping interferers. Unlike conventional notch filters, FRESH filters can extract a signal of interest even when it completely overlaps an interferer in both time and frequency, provided they have distinct cyclic frequencies. This technique is critical for co-channel interference suppression in congested spectrum environments.