Blind Source Separation

The foundational principle of BSS is the search for components that are statistically independent rather than merely uncorrelated. While decorrelation (second-order statistics) ensures zero covariance, independence requires that all higher-order cross-moments also vanish. Algorithms iteratively rotate the signal space to maximize non-Gaussianity or minimize mutual information, exploiting the Central Limit Theorem: mixtures tend toward Gaussian distributions, so maximizing non-Gaussianity recovers the original sources. This principle is why BSS inherently relies on higher-order cumulants—Gaussian processes have zero cumulants beyond order two, making them invisible to independence-seeking algorithms.

ICA is the most widely deployed BSS framework, formalizing source separation as the recovery of latent independent components from linear mixtures. Core assumptions:

• Source signals are statistically independent
• At most one source may be Gaussian (the Gaussian's cumulant ambiguity)
• The mixing process is linear and instantaneous (or can be made so via time-frequency transforms)
• The number of sensors equals or exceeds the number of sources

Common ICA variants include FastICA (fixed-point iteration maximizing negentropy), Infomax (information maximization in neural networks), and Kernel ICA (non-linear extensions using reproducing kernel Hilbert spaces). In RF applications, ICA separates co-channel emitters by exploiting their distinct higher-order statistical fingerprints.

BSS algorithms require an objective function (contrast) to optimize. Cumulant-based contrasts measure departure from Gaussianity:

• Kurtosis-based contrast: Maximizes the absolute value of fourth-order cumulants. Positive kurtosis indicates super-Gaussian sources (e.g., speech), negative kurtosis indicates sub-Gaussian sources (e.g., QAM signals)
• Negentropy approximation: Uses higher-order cumulants to estimate the Kullback-Leibler divergence from Gaussianity
• Mutual information minimization: Expressed in terms of marginal and joint cumulants
• Cumulant tensor diagonalization: Measures how well the fourth-order cumulant tensor of the separated outputs approximates a diagonal structure

These contrasts are theoretically grounded in the property that statistically independent variables have diagonal higher-order cumulant tensors.

Before applying BSS algorithms, observed mixtures undergo critical preprocessing:

• Whitening (sphering): Transforms the data to have identity covariance, decorrelating the mixtures and normalizing variances. This reduces the mixing matrix to an orthogonal matrix, halving the number of parameters to estimate
• Principal Component Analysis (PCA): Reduces dimensionality by projecting onto the directions of maximum variance, discarding noise-dominated components
• Higher-order whitening: Extends whitening beyond second-order statistics, preparing data for cumulant-based separation
• Centering: Subtracts the mean to produce zero-mean signals, simplifying cumulant estimation

Whitening is mathematically expressed as ( \mathbf{z} = \mathbf{W}\mathbf{x} ) where ( \mathbf{W} = \mathbf{\Sigma}^{-1/2} ) is derived from the covariance matrix ( \mathbf{\Sigma} ). This step is essential because joint diagonalization of cumulant matrices assumes orthogonal mixing.

Standard ICA assumes instantaneous mixing, but real RF environments involve convolutive mixing due to multipath propagation. Convolutive BSS extends separation to the time-domain convolution model:

• Time-domain approaches: Use multi-channel blind deconvolution with FIR filter structures to invert the mixing channels
• Frequency-domain approaches: Transform the problem into multiple instantaneous ICA problems at each frequency bin via Short-Time Fourier Transform (STFT), then solve permutation and scaling ambiguities across bins
• Narrowband assumption: When signal bandwidth is small relative to carrier frequency, instantaneous mixing approximates well
• Higher-order cumulant preservation: Frequency-domain methods must ensure that cumulant structures remain consistent across bins to resolve the permutation problem

This extension is critical for separating co-channel emitters in urban or indoor environments where reflections create multiple delayed copies of each source.

The foundational computational method for blind source separation that decomposes a multivariate signal into additive subcomponents by maximizing their statistical independence. ICA assumes the source signals are non-Gaussian and statistically independent, making it highly effective for separating co-channel emitters in RF environments. The algorithm iteratively finds an unmixing matrix that minimizes the mutual information between estimated sources, exploiting the fact that mixtures tend to be more Gaussian than their individual constituents.

An algebraic technique for blind source separation that simultaneously diagonalizes a set of fourth-order cumulant matrices. Unlike gradient-based ICA methods, this approach provides a direct, closed-form solution by finding a rotation matrix that jointly diagonalizes the cumulant tensor slices. This method is particularly robust in RF applications where signals exhibit strong non-Gaussian characteristics due to hardware impairments, offering computational efficiency and guaranteed convergence without local minima.

A pre-processing transformation that extends beyond standard second-order decorrelation to normalize data using higher-order statistics. While standard whitening ensures uncorrelatedness and unit variance, higher-order whitening forces the data to have diagonalized cumulant tensors up to a specified order. This step is critical for preparing RF mixtures for cumulant-based BSS algorithms, as it removes second-order dependencies and simplifies the subsequent search for independent non-Gaussian components.

A dimensionality reduction technique that projects mixed signal data onto directions maximizing non-Gaussianity, effectively isolating the subspace containing hardware-specific fingerprint information. By identifying and retaining only the most non-Gaussian components, this method separates emitter signatures from Gaussian noise and interference before applying blind source separation. This targeted projection dramatically improves the signal-to-interference ratio for subsequent device identification.

The multi-linear algebraic factorization of higher-order cumulant tensors into interpretable components for blind source separation. Techniques like Higher-Order Singular Value Decomposition (HOSVD) decompose the cumulant tensor into a core tensor and orthogonal factor matrices, enabling efficient statistical feature compression. This tensor-based approach jointly exploits the multi-dimensional structure of higher-order statistics, providing more robust source separation than matrix-based methods in underdetermined RF mixtures.

The exploitation of higher-order statistics' theoretical insensitivity to Gaussian processes to extract non-Gaussian signal features buried below the noise floor. Since cumulants of order three and above are identically zero for Gaussian distributions, BSS algorithms operating on bispectrum or trispectrum representations inherently suppress additive Gaussian noise. This property is invaluable in RF environments where weak emitter signatures must be recovered from thermal noise and Gaussian-like interference.