Cumulant-Based Feature Vector

The defining advantage of cumulant-based vectors is their theoretical insensitivity to Gaussian noise. While second-order statistics like variance and correlation are corrupted by additive white Gaussian noise (AWGN), third-order and fourth-order cumulants of Gaussian processes are identically zero. This allows the feature vector to isolate the non-Gaussian signal of interest—the transmitter's unique hardware fingerprint—even when it is buried well below the noise floor. This property makes these vectors exceptionally robust for low-SNR tactical and spectrum surveillance applications.

Unlike power spectrum analysis which discards phase, cumulant-based vectors retain critical phase relationships between frequency components. This is essential for capturing non-linear phenomena like quadratic phase coupling, where two frequencies interact to generate a third at their sum or difference. These phase couplings are direct manifestations of amplifier non-linearity and mixer imperfections, providing a rich, discriminative signature that Fourier-based methods completely miss.

Raw higher-order spectra like the bispectrum are high-dimensional and computationally expensive. Feature vectors are made practical through dimensionality reduction techniques:

• Diagonal Slice Spectrum: Extracts only the 1D diagonal of the bispectrum, retaining key non-Gaussian information while collapsing the 2D plane.
• Integrated Polyspectrum: Computes radial or axial integrals of the bispectrum to condense information into manageable feature sets.
• Higher-Order SVD (HOSVD): Applies multi-linear algebra to decompose cumulant tensors into a compact core tensor and orthogonal factor matrices.

Cumulant-based vectors encode specific invariances and sensitivities that define their classification behavior:

• Translation Invariance: Cumulants are naturally invariant to DC offsets and constant signal shifts.
• Scale Sensitivity: Cumulants scale with the k-th power of signal amplitude, making them sensitive to power differences—a useful trait for distinguishing emitters with different amplifier gains.
• Rotation Sensitivity: Higher-order cumulants are sensitive to signal mixing and rotation, which is exploited in Independent Component Analysis (ICA) for blind source separation of co-channel emitters.

Feature vectors are typically constructed in a hierarchical cascade from raw signal samples:

1. Pre-processing: Higher-order whitening transforms data to remove second-order correlations while preserving non-Gaussian structure.
2. Cumulant Estimation: Sample cumulants are estimated from finite data segments, with careful attention to bias-variance tradeoffs.
3. Tensorization: Cumulants are organized into multi-dimensional arrays (tensors) capturing cross-frequency and cross-lag interactions.
4. Decomposition: Tensor decomposition or diagonal slicing reduces dimensionality.
5. Normalization: Features are normalized (e.g., using bicoherence) to remove nuisance variability.

Cumulant-based vectors can be extended to capture higher-order cyclostationarity, combining two powerful signal properties:

• Cyclostationarity: Exploits the periodic statistical behavior imposed by modulation schemes, symbol rates, and framing structures.
• Non-Gaussianity: Captures the distributional deviations caused by hardware impairments. The resulting cyclic cumulant vector provides doubly-robust features that are both modulation-aware and hardware-specific, enabling simultaneous modulation recognition and emitter identification from a single feature set.

Statistical measures beyond second-order variance that quantify deviations from Gaussianity in signal distributions. These form the mathematical foundation for robust RF fingerprint extraction.

• Third-order cumulant (skewness): Captures asymmetry in amplitude distributions caused by amplifier non-linearity
• Fourth-order cumulant (kurtosis): Measures the 'tailedness' of distributions, revealing impulsive hardware impairments
• Gaussian noise suppression: Cumulants of order > 2 are theoretically zero for Gaussian processes, enabling feature extraction below the noise floor

A third-order frequency-domain representation that detects quadratic phase coupling between signal components. Unlike the power spectrum, the bispectrum reveals non-linear interactions invisible to second-order analysis.

• Maps signal energy across bifrequency pairs (f1, f2)
• Preserves phase information critical for non-linear system identification
• Suppresses additive Gaussian noise while highlighting hardware-induced spectral correlations

A normalized bispectrum that measures the proportion of signal energy at a bifrequency pair that is quadratically phase-coupled. Provides a bounded metric (0 to 1) for non-linearity detection.

• Value near 1: Strong quadratic phase coupling, indicating deterministic hardware non-linearity
• Value near 0: No phase coupling, suggesting independent frequency components
• Useful as a compact, normalized feature for classifier input

A computational method that decomposes multivariate signals into statistically independent non-Gaussian components. Widely used for separating co-channel emitters in dense RF environments.

• Leverages higher-order statistics to maximize statistical independence
• Joint Cumulant Diagonalization: An algebraic technique that simultaneously diagonalizes multiple cumulant matrices
• Enables blind source separation without prior knowledge of mixing parameters

Multi-linear algebraic factorization of higher-order data arrays into interpretable components for dimensionality reduction and feature engineering.

• Higher-Order Singular Value Decomposition (HOSVD): Generalizes SVD to decompose cumulant tensors into a core tensor and orthogonal factor matrices
• Compresses high-dimensional cumulant information into compact feature vectors
• Preserves multi-dimensional statistical relationships lost in vectorization

Higher-order statistical functions that capture both the cyclostationary periodicity and non-Gaussian distribution of communication signals. Provides doubly-robust features for modulation and device recognition.

• Exploits the periodic structure inherent in modulated signals (symbol rates, carrier frequencies)
• Combines cyclic frequency analysis with cumulant-based non-Gaussianity detection
• Resilient to stationary noise and interference that may corrupt individual feature domains