Blind Source Separation

The foundational BSS technique that separates multivariate signals into statistically independent non-Gaussian components. ICA maximizes non-Gaussianity using measures like kurtosis or negentropy, operating on the central limit theorem principle that mixtures are more Gaussian than their sources.

• FastICA: A fixed-point iteration algorithm using Newton's method for rapid convergence
• Infomax ICA: Maximizes mutual information between inputs and outputs using a neural network
• JADE: Joint Approximate Diagonalization of Eigenmatrices, using fourth-order cumulants

Applied extensively to co-channel interference mitigation in wireless systems and artifact removal in EEG/MEG signal processing.

Exploits sparsity in the time-frequency domain rather than statistical independence. SCA assumes that only a small number of sources are active at any given time-frequency point, enabling separation even when sources outnumber sensors (underdetermined BSS).

• DUET (Degenerate Unmixing Estimation Technique): Uses inter-sensor amplitude and delay histograms
• Time-Frequency Masking: Applies binary or soft masks in the spectrogram domain
• Matching Pursuit: Greedy sparse decomposition using overcomplete dictionaries

Critical for separating overlapping radar pulses and multi-speaker audio in the cocktail party problem.

Decomposes a non-negative data matrix into two lower-rank non-negative matrices representing basis vectors and activation coefficients. The non-negativity constraint produces a parts-based, interpretable representation.

• Multiplicative Update Rules: Iterative optimization minimizing KL divergence or Frobenius norm
• Sparse NMF: Adds L1 regularization to enforce sparsity in activations
• Convolutive NMF: Extends to time-shifted basis functions for temporal patterns

Used in spectrogram decomposition for RF emitter identification and audio source separation where power spectra are inherently non-negative.

Leverages second-order statistics (autocorrelation and cross-correlation) rather than higher-order moments. SOBI jointly diagonalizes multiple time-lagged covariance matrices, making it effective for separating temporally correlated sources with distinct power spectra.

• AMUSE: Algorithm for Multiple Unknown Signals Extraction, a computationally efficient precursor
• WASOBI: Weights-adjusted SOBI for improved asymptotic performance
• Robust to Gaussian sources unlike ICA, which requires non-Gaussianity

Ideal for separating narrowband communication signals with different carrier frequencies and biomedical sensor arrays.

An ICA algorithm that operates on fourth-order cumulant tensors, simultaneously diagonalizing a set of eigenmatrices derived from the data. JADE provides excellent separation performance without requiring tuning of learning rate parameters.

• Algebraic solution: No iterative gradient descent, avoiding local minima
• Computationally intensive: O(n⁴) complexity limits use to moderate-dimensional problems
• Robust to outliers due to higher-order statistics

Applied in direction-finding arrays and radar pulse deinterleaving where precise source recovery is critical.

Modern neural network approaches that learn separation mappings directly from data, often outperforming classical methods on complex mixtures. Architectures operate in time-domain or time-frequency domain.

• Conv-TasNet: Fully convolutional time-domain audio separation network
• Deep Clustering: Embeds time-frequency bins into a space where bins belonging to the same source cluster together
• Complex Ratio Masking: Neural estimation of ideal complex masks for phase-aware reconstruction

Increasingly applied to RF signal separation in congested spectrum environments and multi-emitter geolocation tasks.

The core optimization principle behind ICA-based separation. Real-world signals—speech, modulated RF, biological data—exhibit super-Gaussian (spiky, heavy-tailed) or sub-Gaussian (flat, bounded) distributions. By iteratively rotating the mixing matrix to maximize statistical independence, the algorithm isolates sources. Common contrast functions include:

• Kurtosis: Fourth-order cumulant measuring distribution peakedness
• Negentropy: Distance from Gaussianity based on differential entropy
• Log-cosh: Robust approximation of negentropy for outlier resistance

A batch BSS algorithm that operates on fourth-order cumulant tensors rather than iterative optimization. JADE computes the eigenvalue decomposition of the whitened data's cumulant matrix, then jointly diagonalizes a set of eigenmatrices to recover the mixing matrix. Advantages include:

• No convergence issues: Closed-form solution avoids local minima
• Deterministic output: Same result every run unlike stochastic ICA
• Well-suited for small to medium dimensions: Typically fewer than 20 sources Computationally heavier than FastICA for high-dimensional problems.

A BSS technique exploiting temporal structure in source signals using only second-order statistics (autocorrelation). Unlike ICA, SOBI works with Gaussian sources as long as they have distinct power spectra. The algorithm:

• Computes time-lagged covariance matrices at multiple delays
• Performs joint diagonalization of these matrices
• Recovers sources based on their unique temporal signatures Critical for separating correlated RF emitters and biomedical signals like EEG where ICA's independence assumption fails.

A parts-based decomposition method that constrains both the mixing matrix and source signals to be non-negative, making it ideal for magnitude spectra and image data. NMF minimizes the reconstruction error:

• Euclidean distance: Standard least-squares objective
• Kullback-Leibler divergence: Better for Poisson-distributed data
• Itakura-Saito divergence: Scale-invariant, preferred for audio spectrograms Applied to audio source separation where spectrogram magnitudes are inherently non-negative and additive.

A BSS framework that assumes sources are sparse in some transform domain—most coefficients are near zero with occasional large values. SCA operates in two stages:

• Mixing matrix estimation: Clustering scatter plots of mixed observations to identify column directions
• Source recovery: Solving an underdetermined system using ℓ₁-minimization (basis pursuit) or greedy algorithms like Orthogonal Matching Pursuit Enables separation when there are more sources than sensors (underdetermined BSS), impossible with classical ICA.