Non-Linear System Identification

The core goal is to mathematically characterize the non-linear relationship between a transmitter's ideal baseband input and its distorted RF output. Unlike linear models, this approach captures harmonic generation and intermodulation distortion caused by analog components like power amplifiers and mixers. The system is often modeled using a Volterra series, which represents the output as a sum of multi-dimensional convolution integrals. Higher-order statistics are essential here because they can identify the quadratic and cubic phase couplings that define the specific non-linearity, acting as a unique hardware fingerprint.

A foundational advantage of this approach is the theoretical insensitivity of higher-order cumulants (third-order and above) to additive white Gaussian noise (AWGN). In a linear system driven by a Gaussian input, all cumulants above the second order are zero. Therefore, by analyzing the bispectrum or trispectrum of a received signal, the Gaussian thermal noise component is mathematically suppressed. This allows the extraction of the non-Gaussian signal of interest, which contains the transmitter's non-linear distortion signature, even when it is buried well below the noise floor.

A key non-linear phenomenon is quadratic phase coupling (QPC), where two frequency components, f1 and f2, interact to generate a third component at their sum or difference frequency (f1 ± f2) with a phase that is the sum or difference of the original phases. This is a hallmark of non-linear systems and is invisible to the power spectrum. The bispectrum is the definitive tool for detecting QPC, as it measures the statistical correlation between a frequency pair and their sum frequency. This phase coherence is a direct product of the transmitter's specific hardware impairments.

The Volterra series is the dominant mathematical model for a causal, time-invariant non-linear system with memory. It extends the linear convolution integral by adding higher-order terms:

• First-order kernel: Linear impulse response.
• Second-order kernel: Captures quadratic non-linear interactions between two time instances.
• Third-order kernel: Captures cubic interactions, essential for modeling power amplifier saturation. The polyspectra (bispectrum, trispectrum) are the multi-dimensional Fourier transforms of these higher-order Volterra kernels, directly linking the frequency-domain statistical analysis to the physical system model.

Standard second-order statistics like the power spectrum and autocorrelation are phase-blind; they discard all phase information. Non-linear system identification requires phase to distinguish between different non-linear mechanisms. Higher-order spectra inherently preserve phase relationships. For example, the bispectrum is a complex-valued function whose phase is the biphase. This biphase captures the combined phase of the interacting frequency components, providing a rich, discriminative feature space for identifying individual emitters that a simple power spectrum could never reveal.

A powerful application is the ability to characterize the non-linear system without knowing the input signal. Blind system identification techniques use only the observed output signal. By assuming the input is a statistically independent, non-Gaussian process (a valid assumption for many communication signals), algorithms can estimate the system's linear and non-linear components. This is achieved through methods like blind source separation and cumulant-based deconvolution, which exploit the statistical independence of the input to isolate the deterministic, hardware-specific distortion channel.

A third-order frequency-domain representation that detects quadratic phase coupling between signal components. Unlike the power spectrum, the bispectrum reveals non-Gaussian signatures generated by transmitter hardware non-linearities. It is computed as the 2D Fourier transform of the third-order cumulant sequence, providing a frequency-frequency map where peaks indicate specific non-linear interactions.

Statistical measures extending beyond second-order variance that quantify deviations from Gaussianity in signal amplitude distributions. Key cumulants include:

• Skewness (3rd order): Measures distribution asymmetry, revealing directional amplifier biases
• Kurtosis (4th order): Quantifies distribution tailedness, where excess kurtosis indicates non-Gaussian hardware fingerprints
• These cumulants are theoretically insensitive to additive Gaussian noise, making them robust features for low-SNR emitter identification.

A non-linear phenomenon where two frequency components (f1, f2) interact within a transmitter's analog components to generate a third component at their sum or difference frequency (f1±f2). This coupling preserves a consistent phase relationship that is detectable via the bispectrum. The specific coupling patterns serve as a distinctive, hardware-induced fingerprint that cannot be easily cloned or spoofed by an adversary.

A normalized bispectrum that measures the proportion of signal energy at a bifrequency pair that is quadratically phase-coupled. Bicoherence values range from 0 to 1, providing a bounded, interpretable metric for non-linearity detection. It is particularly useful for distinguishing between coupled spectral components generated by hardware non-linearities and those arising from random noise or linear processes.

A compact statistical fingerprint constructed from estimated higher-order cumulants that serves as input to machine learning classifiers. The process involves:

• Estimating 2nd, 3rd, and 4th order cumulants from received signal samples
• Organizing them into a fixed-dimensional vector
• Using this vector as a robust, noise-resistant feature set for emitter identification
• These vectors capture the unique non-linear distortion profile of each transmitter's analog chain.

A computational method that decomposes multivariate signals into statistically independent non-Gaussian components. In RF fingerprinting, ICA is widely used for separating co-channel emitters by exploiting the non-Gaussianity of individual hardware signatures. It relies on higher-order statistics to maximize statistical independence, enabling the isolation of individual transmitter fingerprints from overlapping signal mixtures.