Higher-Order Statistics (HOS)

The foundational property of HOS: all cumulants of order greater than two are identically zero for Gaussian processes. This means that third-order (skewness) and fourth-order (kurtosis) cumulants of a signal are theoretically uncontaminated by additive white Gaussian noise (AWGN).

• Mechanism: Cumulants extract non-Gaussian information from the signal's probability density function
• Practical impact: Enables signal detection at SNRs as low as -20 dB where energy detection fails
• Key insight: The bispectrum (Fourier transform of the third-order cumulant) suppresses Gaussian noise in the frequency domain, revealing harmonic coupling that a power spectrum cannot show

Unlike the power spectrum (second-order), which discards all phase information, the bispectrum and trispectrum preserve Fourier phase relationships. This enables discrimination between signals with identical power spectra but different phase coupling structures.

• Quadratic Phase Coupling (QPC): Detects when two frequency components are phase-coupled to a sum frequency—a hallmark of non-linear systems
• Application: Distinguishing between linear and non-linear signal sources, identifying specific emitter types by their non-linear front-end characteristics
• Bicepstral analysis: The inverse Fourier transform of the log-bispectrum separates mixed-phase signals that second-order cepstrum cannot resolve

HOS provides a mathematical framework for quantifying deviation from Gaussianity, which is the essential statistical signature of modulated communications signals.

• Skewness (3rd-order): Measures asymmetry in the signal distribution—useful for detecting non-symmetric modulation formats
• Kurtosis (4th-order): Measures the "peakedness" of a distribution relative to Gaussian—excess kurtosis distinguishes sub-Gaussian (BPSK, QPSK) from super-Gaussian (OFDM) signals
• Normalized cumulants: Ratios like |C₄₀|/|C₂₁|² provide modulation-format-specific signatures that are robust to amplitude scaling and phase rotation, forming the basis for automatic modulation classification

HOS enables blind identification of communication channels without requiring training sequences or pilot tones, preserving bandwidth for data transmission.

• Principle: The fourth-order cumulant of the received signal contains sufficient statistics to estimate both channel magnitude and phase response
• Methods: Algorithms like the Cumulant Matrix Pencil and EigenVector-based Algorithm for blind Identification (EVA) exploit the structure of higher-order cumulant tensors
• Advantage over SOS: Second-order statistics alone cannot resolve non-minimum phase channels; HOS can, making it critical for equalization in multipath environments where the direct path is not the strongest

The frequency-domain representation of HOS—polyspectra—reveals non-linear interactions between frequency components that are invisible in the power spectrum.

• Bispectrum B(f₁, f₂): A 2D function mapping every frequency pair to their non-linear coupling strength. Peaks indicate quadratic phase coupling
• Trispectrum T(f₁, f₂, f₃): A 3D function capturing cubic non-linearities, useful for detecting transient events and analyzing OFDM signals with high peak-to-average power ratios
• Bicoherence: A normalized bispectrum (values 0–1) that quantifies the statistical consistency of phase coupling, distinguishing deterministic non-linearities from random fluctuations
• Computational note: Direct bispectrum estimation via the Brillinger-Rosenblatt method averages triple products of FFT coefficients over multiple segments

While HOS is famous for Gaussian noise rejection, it also provides partial immunity to colored Gaussian noise—noise with a non-flat power spectrum that corrupts second-order methods.

• Mechanism: The cumulant of colored Gaussian noise remains zero at orders > 2, regardless of its spectral shape
• Practical scenario: In wideband spectrum sensing, adjacent channel interference and filtered thermal noise create colored Gaussian backgrounds that defeat energy detection but leave HOS-based detectors unaffected
• Limitation: Non-Gaussian interference (e.g., co-channel modulated signals) does contribute to higher-order cumulants and must be addressed through spatial filtering or source separation techniques

A class of statistical techniques, often implemented via Independent Component Analysis (ICA), that separates mixed signals without prior knowledge of the mixing process. HOS is fundamental to many ICA algorithms because maximizing statistical independence requires minimizing cross-cumulants. Second-order statistics (covariance) can only decorrelate signals; higher-order cumulants are required to achieve true independence.

A blind sensing method that tests for the presence of a correlated signal against uncorrelated noise. Standard covariance methods use second-order statistics. HOS-based detection extends this by testing the structure of the fourth-order cumulant matrix, which is theoretically zero for Gaussian noise. This provides a detection statistic with an inherently zero noise floor, unlike eigenvalue-based covariance methods.

Neural architectures that process complex IQ data natively, preserving phase information. When combined with HOS, CVNNs can learn directly from complex cumulants and polyspectra as input features. This avoids the information loss from converting complex values to real-valued magnitude and phase pairs, allowing the network to exploit higher-order phase relationships critical for emitter identification.

Transforms raw IQ streams into time-frequency images via the Short-Time Fourier Transform (STFT). While spectrograms are second-order representations (power spectral density), the bispectrum and trispectrum are their higher-order counterparts. HOS-based spectrogram equivalents map frequency-frequency correlations, revealing quadratic phase coupling that is invisible in standard spectrograms but diagnostic of specific non-linearities and modulation types.

Autonomously identifies the modulation scheme of a received waveform. HOS provides discriminating features that are theoretically invariant to Gaussian noise. For example:

• Cumulant ratio (C40/C42) distinguishes QAM from PSK constellations.
• Bispectrum signatures uniquely identify non-linear modulation types. These features remain stable even when SNR is too low for constellation diagram reconstruction.