Higher-Order Cumulants

Skewness measures the asymmetry of the signal's amplitude distribution. A perfectly linear, ideal transmitter would produce a symmetric distribution (skewness ≈ 0). However, amplifier non-linearity—particularly even-order distortion from push-pull imbalances—introduces a measurable skew. This asymmetry is a direct consequence of the physical semiconductor physics within the power amplifier and is highly stable over time, making it a robust, unclonable identifier for device fingerprinting.

Kurtosis quantifies the 'tailedness' of the signal's distribution—how prone it is to producing outlier amplitude values. A Gaussian process has a kurtosis of 3.0 (excess kurtosis of 0). Transmitter impairments like phase noise and power amplifier memory effects cause the distribution to become leptokurtic (heavy-tailed, kurtosis > 3) or platykurtic (light-tailed, kurtosis < 3). The specific kurtosis value forms a distinctive signature that is resilient to additive white Gaussian noise in the channel.

A critical advantage of cumulants of order n ≥ 3 is their theoretical insensitivity to Gaussian processes. All cumulants of a Gaussian distribution are identically zero for orders greater than two. This means that when extracting features from a received signal, higher-order cumulants automatically suppress the contribution of thermal noise and other Gaussian interference sources, isolating only the non-Gaussian distortion components introduced by the transmitter's hardware impairments. This property makes them exceptionally robust for real-world, low-SNR environments.

The Fourier transforms of the third- and fourth-order cumulants yield the bispectrum and trispectrum, respectively. These higher-order spectra reveal non-linear phase coupling between different frequency components of the signal. Unlike the standard power spectrum, which discards phase information, the bispectrum preserves it, allowing the detection of quadratic phase coupling generated by amplifier non-linearities. This frequency-domain representation provides a rich, two-dimensional feature map for deep learning classifiers.

In practice, cumulants are estimated from finite signal samples using k-statistics (unbiased estimators). For a received signal segment x[n]:

• Second cumulant: variance (σ²)
• Third cumulant: E[(x - μ)³]
• Fourth cumulant: E[(x - μ)⁴] - 3(E[(x - μ)²])² The subtraction of 3σ⁴ in the fourth-order estimate explicitly removes the Gaussian contribution, leaving only the non-Gaussian residual. The variance of these estimators decreases with sample length, requiring sufficient data capture for stable fingerprints.

While marginal cumulants analyze a single signal dimension, joint cumulants capture the statistical dependencies between the in-phase (I) and quadrature (Q) components simultaneously. The cross-cumulant κ₁,₃(I,Q,Q,Q), for example, measures how the I component's distribution is influenced by the cube of the Q component. These joint statistics reveal I/Q imbalance and cross-modulation effects that are invisible to marginal analysis, creating a higher-dimensional fingerprint vector for improved device separability.

The third and fourth standardized moments of a signal's amplitude distribution. Skewness measures distributional asymmetry caused by even-order non-linearities, while kurtosis quantifies the 'tailedness' or peakedness relative to a Gaussian distribution. In RF fingerprinting, deviations from Gaussianity (kurtosis ≠ 3) directly indicate the presence of transmitter-specific impairments like power amplifier compression.

A fundamental property of cumulants of order greater than two: they are identically zero for Gaussian processes. This allows higher-order cumulant-based features to theoretically reject additive white Gaussian noise entirely. In practice, this means fingerprinting algorithms using trispectrum or fourth-order cumulants can extract device signatures at much lower signal-to-noise ratios than second-order methods like correlation.

Higher-order cumulants are essential for characterizing Volterra series models of non-linear transmitter components. The third-order cumulant captures quadratic non-linearities, while the fourth-order cumulant reveals cubic non-linearities. By estimating these cumulants from emitted signals, one can reconstruct the unique AM/AM and AM/PM distortion curves of a specific power amplifier without direct physical access.

The Fourier transform of the fourth-order cumulant, defined in a three-dimensional frequency space. It detects cubic phase coupling between frequency quartets, revealing complex intermodulation products generated by memory effects in power amplifiers. The trispectrum is particularly effective for distinguishing devices with identical spectral shapes but different dynamic non-linear behaviors.

Algorithms that use higher-order cumulants to deconvolve channel effects without requiring a training sequence. By maximizing the kurtosis or minimizing cross-cumulant criteria, these methods can separate the transmitter fingerprint from multipath distortion. This is critical for mobile RF fingerprinting where the channel varies rapidly, enabling channel-robust feature extraction.