Higher-Order Statistics (HOS)

The defining advantage of HOS is its theoretical blindness to Gaussian processes. All cumulants of order greater than two (k>2) are identically zero for any Gaussian-distributed signal. This means that when you compute the bispectrum or trispectrum of a received emitter signal, additive white Gaussian noise (AWGN) is mathematically suppressed in the HOS domain.

• Mechanism: Cumulants of order k≥3 are zero for Gaussian distributions
• Result: The noise floor in the bispectrum is significantly lower than in the power spectrum
• Benefit: Extracts emitter fingerprints from signals buried well below the noise floor where traditional SNR-based methods fail

HOS, particularly the bispectrum, is uniquely capable of detecting and quantifying quadratic phase coupling (QPC)—a phenomenon where two frequency components interact non-linearly to generate a third component whose phase is the sum of the original two phases.

• Why it matters: Power amplifier non-linearities in transmitters generate QPC between harmonic and intermodulation products
• Fingerprint source: The specific pattern of phase coupling is a direct consequence of a device's unique AM-AM and AM-PM distortion characteristics
• Detection: The bicoherence (normalized bispectrum) provides a bounded measure (0 to 1) of coupling strength at each bifrequency pair

HOS quantifies deviations from Gaussianity through two fundamental shape parameters that serve as discriminative features for emitter identification:

• Skewness (3rd-order): Measures the asymmetry of the signal's probability distribution. A non-zero skewness indicates that the I/Q samples are not symmetrically distributed around the mean, often caused by DC offset or local oscillator leakage
• Kurtosis (4th-order): Measures the 'tailedness' of the distribution. Signals with high kurtosis exhibit more extreme outliers, a characteristic of power amplifier saturation and clipping effects
• Application: These scalar moments, computed over short signal segments, form compact feature vectors for real-time device classification

Unlike the power spectrum which discards all phase information by computing only magnitude-squared values, the bispectrum and trispectrum retain Fourier phase relationships. This is critical because hardware impairments manifest as subtle, repeatable phase distortions.

• Power spectrum: Retains only |X(f)|²—phase is lost
• Bispectrum: Retains B(f1, f2) = X(f1)X(f2)X*(f1+f2)—phase relationships are preserved
• Discrimination power: Two emitters with identical power spectra can have radically different bispectra due to unique phase distortion patterns from their analog front-ends
• Result: HOS-based fingerprints achieve higher inter-device separability in the embedding space

A critical property for operational deployment: the cumulant of a signal is invariant to time shifts. If a signal x(t) is delayed by some arbitrary amount, its cumulants remain unchanged.

• Practical implication: The fingerprint is robust to random packet arrival times and asynchronous sampling
• Contrast: Raw I/Q samples and time-domain features shift with every transmission, requiring precise alignment
• Deployment advantage: HOS-based systems eliminate the need for complex time-synchronization pre-processing, enabling blind emitter identification on intercepted bursts with unknown start times

The bispectrum operates in a 2D bifrequency plane, providing exponentially more feature dimensions than the 1D power spectrum. This high-dimensional representation naturally separates emitter classes.

• Bispectrum: A 2D function B(f1, f2) defined over a triangular region of support
• Feature volume: For an N-point FFT, the bispectrum yields O(N²) unique coefficients versus O(N) for the power spectrum
• Integration with deep learning: The bispectrum can be treated as a 2D image and fed directly into a Convolutional Neural Network (CNN) for automatic hierarchical feature learning
• Redundancy exploitation: The symmetry properties of the bispectrum can be leveraged for dimensionality reduction via Principal Component Analysis (PCA) without losing discriminative information

The frequency-domain representation of the third-order cumulant. It transforms a signal to reveal quadratic phase coupling—a phenomenon where two frequency components interact to generate a third, which is invisible to standard power spectrum analysis. The bispectrum is naturally blind to Gaussian noise, making it exceptionally robust for extracting device-specific non-linearities from low-SNR environments. It produces a two-dimensional function B(f1, f2) that serves as a rich, discriminative feature map for emitter classification.

While moments (mean, variance) are the raw statistical averages of a signal, cumulants are the mathematically purified measures that isolate non-Gaussian behavior.

• 2nd-order cumulant: Variance (spread)
• 3rd-order cumulant: Skewness (asymmetry)
• 4th-order cumulant: Kurtosis (tailedness, minus 3 for Gaussian)

For Gaussian processes, all cumulants of order > 2 are identically zero. This property makes cumulants ideal for suppressing Gaussian noise and isolating the deterministic, non-linear hardware impairments that define an emitter's fingerprint.

A non-linear phenomenon where two harmonic components at frequencies f1 and f2 interact within a device's power amplifier to generate a third component at f1 + f2, with a phase that is the sum of the original phases. This phase coherence is a direct product of hardware non-linearity and is undetectable by the power spectrum. The bispectrum explicitly detects this coupling, providing a unique signature of amplifier distortion that varies from device to device due to manufacturing tolerances.

A fundamental advantage of HOS techniques: all cumulants of order ≥ 3 are zero for Gaussian processes. This means that when you compute the bispectrum or trispectrum of a received signal, any additive white Gaussian noise (AWGN) from the channel is theoretically eliminated. The resulting HOS representation isolates only the non-Gaussian signal components—precisely the deterministic hardware impairments introduced by the transmitter's analog front-end. This property is critical for robust fingerprinting in low-SNR tactical environments.

The frequency-domain representation of the fourth-order cumulant, defined over three frequency dimensions T(f1, f2, f3). While computationally more intensive than the bispectrum, the trispectrum can detect cubic phase coupling and characterize more subtle non-linearities. It is particularly useful for analyzing complex modulated signals (e.g., 256-QAM) where higher-order amplifier distortions manifest. The trispectrum also enables the recovery of both magnitude and phase information of a system's transfer function.

The process of converting a bispectrum or trispectrum into a compact, discriminative input for a classifier:

• Integrated Bispectrum: Radially integrate the 2D bispectrum to create a 1D feature vector robust to time shifts
• Axial slices: Extract specific frequency lines from the bispectrum for computational efficiency
• Cumulant values: Use raw 3rd and 4th-order cumulant estimates directly as features
• Bicepstral coefficients: Apply a Fourier transform to the log-bispectrum for compact representation

These vectors feed into CNN or SVM classifiers for device identification.