Cyclic Bispectrum

The cyclic bispectrum is the definitive tool for detecting quadratic phase coupling (QPC) — a non-linear phenomenon where the phase of a signal at one frequency is a function of phases at two other frequencies. This occurs in signals that have passed through non-linear amplifiers or channels. Unlike the power spectrum or spectral correlation function, the cyclic bispectrum can distinguish between frequency components that are phase-coupled and those that are merely coincidental in frequency. This capability is critical for transmitter fingerprinting and identifying specific non-linear hardware impairments.

The cyclic bispectrum is defined as the 2D Fourier transform of the third-order cyclic cumulant function. Formally, it is expressed as the frequency-frequency representation of the cyclic polyspectrum at order three. It is a function of two spectral frequencies (f1, f2) and one cyclic frequency (α). The key condition is that the signal's third-order moment is periodic with period 1/α. The bispectrum is non-zero only where f1 + f2 = α (or other linear combinations depending on the conjugate choice), creating a deterministic structure in the bifrequency plane that serves as a unique signature for modulated signals.

A primary advantage of the cyclic bispectrum is its theoretical immunity to additive white Gaussian noise (AWGN). Because the bispectrum is a third-order statistic, and the third-order cumulant of a Gaussian process is identically zero, any Gaussian noise component is completely suppressed in the bispectrum domain. This property makes cyclic bispectrum-based classifiers exceptionally robust in low signal-to-noise ratio (SNR) environments where second-order cyclostationary features may be buried. It allows for reliable modulation identification even when the signal power is well below the noise floor.

Different digital modulation schemes produce distinct patterns in the cyclic bispectrum domain. For example, M-PSK signals exhibit peaks at specific cyclic frequencies related to the carrier and symbol rate, while M-QAM signals generate more complex, multi-peak structures due to their amplitude variations. The bispectrum can differentiate between modulation types that appear identical in the power spectrum or even the spectral correlation function. This is particularly useful for classifying higher-order QAM signals (e.g., 64-QAM vs. 256-QAM) where constellation density is high.

Estimating the cyclic bispectrum is computationally intensive, typically requiring a 2D frequency-smoothing approach analogous to the cyclic periodogram. The process involves segmenting the signal, computing the product of three frequency-shifted Fourier transforms, and averaging over time. Efficient algorithms often use the FFT Accumulation Method (FAM) extended to the third order, or direct frequency-domain smoothing. The complexity is O(N³) in the general case, making it suitable for offline or non-real-time analysis, though optimized implementations on FPGAs are an active research area for tactical applications.

The cyclic bispectrum is the Fourier dual of the third-order cyclic cumulant. While the cyclic cumulant operates in the time-lag domain, the bispectrum provides a frequency-frequency representation. This relationship is analogous to how the cyclic spectrum is the Fourier transform of the cyclic autocorrelation function. The cyclic bispectrum can be derived by taking a 2D Fourier transform of the cyclic third-order cumulant with respect to its two lag variables. This connection allows engineers to choose the most computationally advantageous domain for feature extraction.

The statistical framework that underpins the cyclic bispectrum. While second-order statistics (like the Spectral Correlation Function) exploit periodicity in autocorrelation, higher-order methods analyze periodicity in third-order moments and cumulants. This is essential for analyzing signals with quadratic phase coupling or those that appear stationary at second order, providing robustness against Gaussian noise.

A higher-order statistic that generalizes the standard cumulant to cyclostationary signals. The cyclic bispectrum is the Fourier transform of the third-order cyclic cumulant. Key properties include:

• Blind source separation: Additive Gaussian noise is suppressed.
• Phase preservation: Retains information lost in the power spectrum.
• Cyclic frequency (α): Parameterizes the hidden periodicity of the cumulant.

A non-linear phenomenon where two frequency components interact to generate a third component whose phase is the sum of the original phases. The cyclic bispectrum is the definitive tool for detecting QPC in cyclostationary signals. Applications include:

• Fault diagnosis in rotating machinery.
• EEG analysis to identify non-linear brain wave interactions.
• Modulation recognition of continuous-phase modulated signals.

A compact, discriminative representation of a signal used as input to a modulation classifier. Instead of using raw IQ samples, a feature vector is constructed by sampling the cyclic bispectrum at specific cyclic frequencies (α) and bifrequencies. This reduces dimensionality while retaining the unique cyclic signature of the modulation scheme, enabling robust automatic modulation classification.

Cyclostationary features intentionally embedded at the transmitter to aid signal identification. While the cyclic bispectrum can analyze inherent signal properties, it can also detect artificially induced periodicities. Examples include:

• Specific pulse shaping filters.
• Embedded pilot sequences with unique cyclic patterns.
• Transmitter-specific distortion used for RF fingerprinting.

The process of estimating a signal's modulation parameters without prior knowledge. The cyclic bispectrum enables blind extraction of:

• Symbol rate: Detected via cyclic frequency peaks.
• Carrier frequency offset: Estimated from phase shifts in the bispectral plane.
• Modulation type: Classified by matching the extracted cyclic bispectrum signature against a known library of cyclic signatures.