Higher-Order Cumulants

All cumulants of order ν > 2 are identically zero for Gaussian processes. This property allows cumulant-based feature extractors to theoretically suppress additive white Gaussian noise (AWGN) entirely, isolating only the non-Gaussian signal components generated by transmitter hardware impairments.

• Third-order and fourth-order cumulants filter out Gaussian interference
• Enables fingerprint extraction below the noise floor
• Critical for low-SNR environments like distant emitter identification

If two statistically independent random processes are summed, their cumulants add linearly: cum(Y1 + Y2) = cum(Y1) + cum(Y2). This contrasts with moments, which mix cross-terms.

• Simplifies blind source separation of co-channel emitters
• Enables Joint Cumulant Diagonalization for unsupervised signal isolation
• Preserves the individual fingerprint of each transmitter in a mixed signal environment

Higher-order cumulants naturally organize into multi-dimensional arrays (tensors). The fourth-order cumulant, for example, forms a four-dimensional tensor C_ijkl that captures interactions across four signal lags.

• Enables Higher-Order Singular Value Decomposition (HOSVD) for dimensionality reduction
• Preserves multi-modal interactions lost in vectorized representations
• Facilitates extraction of interpretable statistical components through tensor factorization

The bispectrum is the 2D Fourier transform of the third-order cumulant sequence, and the trispectrum is the 3D Fourier transform of the fourth-order cumulant. This duality provides both time-lag and frequency-domain perspectives on non-Gaussianity.

• Cumulants operate in the lag domain for time-series analysis
• Polyspectra reveal frequency-domain phase coupling patterns
• The diagonal slice of the bispectrum reduces computational complexity while retaining key discriminative information

Higher-order cumulants can be normalized to produce scale-invariant features. The bicoherence—a normalized bispectrum—provides a bounded metric between 0 and 1 that measures the proportion of quadratically phase-coupled energy at each bifrequency pair.

• Removes dependence on absolute signal power
• Enables comparison across devices operating at different transmit powers
• Kurtosis and skewness provide compact, interpretable summaries of non-Gaussian distribution shape

For communication signals exhibiting both cyclostationarity (periodic statistics) and non-Gaussianity, cyclic cumulants extend standard cumulants to capture time-varying statistical behavior synchronized with the symbol rate.

• Combines periodicity detection with non-Gaussian characterization
• Provides doubly-robust features for automatic modulation classification
• Enables separation of emitters using different modulation schemes in shared spectrum

A compact statistical fingerprint constructed from estimated higher-order cumulants. Key properties exploited for classification:

• Gaussian noise suppression: Cumulants of order > 2 are theoretically zero for Gaussian processes
• Phase preservation: Unlike the power spectrum, cumulants retain phase information
• Non-linearity sensitivity: Directly quantify deviations from linear amplifier behavior These vectors serve as input to support vector machines and neural network classifiers for emitter identification.

A computational method that decomposes multivariate signals into statistically independent, non-Gaussian components. ICA leverages the fact that higher-order cumulants of independent sources are non-zero, using kurtosis maximization or negentropy as optimization criteria. In RF environments, ICA enables:

• Blind source separation of co-channel emitters
• Recovery of individual device signatures from mixed signals
• Unsupervised isolation of hardware-specific distortion components

Statistical hypothesis tests that determine whether a signal's amplitude distribution deviates from Gaussian. Common tests include:

• Jarque-Bera test: Based on sample skewness and kurtosis
• D'Agostino's K-squared test: Evaluates deviation from expected Gaussian moments
• Shapiro-Wilk test: Assesses normality for smaller sample sizes Validating non-Gaussianity confirms the presence of exploitable hardware fingerprints. If a signal is purely Gaussian, higher-order cumulants provide no additional discriminatory information beyond variance.

A cumulant tensor organizes higher-order cumulants into a multi-dimensional array, enabling joint analysis across time, frequency, and spatial dimensions. Higher-Order Singular Value Decomposition (HOSVD) factorizes this tensor into a core tensor and orthogonal factor matrices, achieving:

• Dimensionality reduction while preserving multi-linear structure
• Extraction of the most discriminative non-Gaussian subspaces
• Efficient compression of statistical fingerprints for real-time classification on edge hardware

A cyclic cumulant captures both the cyclostationary periodicity (e.g., symbol rate, carrier frequency) and the non-Gaussian distribution of communication signals. This dual characterization provides:

• Modulation-specific signatures: Different modulation schemes exhibit distinct cyclic cumulant patterns
• Robustness to stationary noise: Separates cyclostationary signal features from stationary interference
• Joint device and modulation recognition: Simultaneously identifies the transmitter hardware and its operating mode