Cumulant Tensor

A cumulant tensor extends the concept of a covariance matrix into higher orders, arranging fourth-order cumulants into a four-dimensional array. This structure captures joint statistical dependencies across multiple signal lags simultaneously, preserving the multi-linear relationships that matrix-based methods flatten and lose. For an M-channel signal, the fourth-order cumulant tensor has dimensions M×M×M×M, encoding the complete non-Gaussian correlation structure.

Cumulant tensors exhibit supersymmetry, meaning the tensor elements are invariant under any permutation of their indices. For a fourth-order tensor, this implies C(i,j,k,l) = C(j,i,k,l) = C(k,j,i,l) for all permutations. This property reduces the number of unique elements from M⁴ to approximately M⁴/24, significantly lowering computational and storage requirements while preserving all statistical information.

The cumulant tensor can be decomposed using Canonical Polyadic Decomposition (CPD) to perform joint blind source separation. Unlike matrix-based Independent Component Analysis, tensor decomposition exploits the multi-linear structure to achieve unique decomposition under milder conditions. The CPD expresses the cumulant tensor as a sum of rank-1 components, each corresponding to a statistically independent source signal with its associated mixing vector.

A practical approach to tensor-based source separation involves extracting matrix slices from the cumulant tensor and performing joint diagonalization. Each slice represents a linear combination of source cumulants, and the simultaneous diagonalization of multiple slices yields the unmixing matrix. This joint cumulant diagonalization technique is more robust to noise than single-matrix methods and exploits the redundancy inherent in the tensor structure.

HOSVD provides an orthogonal factorization of the cumulant tensor into a core tensor and factor matrices along each mode. This multi-linear generalization of SVD enables dimensionality reduction by truncating the factor matrices, projecting the cumulant information onto a lower-dimensional subspace that retains the dominant non-Gaussian features. The core tensor captures the interactions between the reduced modes.

A defining property of cumulant tensors is their theoretical insensitivity to Gaussian noise. All cumulants of order greater than two are identically zero for Gaussian processes. When a non-Gaussian signal of interest is contaminated by additive Gaussian noise, the cumulant tensor of the mixture equals the cumulant tensor of the signal alone. This property enables feature extraction and source separation at signal-to-noise ratios where second-order methods fail.

The multi-linear algebraic factorization of higher-order data arrays into interpretable components. For cumulant tensors, this enables dimensionality reduction and feature engineering by extracting latent factors that represent independent source signals or hardware impairment signatures.

• CP Decomposition: Factorizes a tensor into a sum of rank-one components
• Tucker Decomposition: Decomposes into a core tensor multiplied by factor matrices along each mode
• Essential for compressing high-dimensional cumulant data while preserving multi-way interactions

An algebraic technique that simultaneously diagonalizes multiple cumulant matrices to achieve blind source separation without requiring gradient-based optimization. This method exploits the fact that cumulant tensors of independent sources are diagonal when represented in the appropriate basis.

• Provides closed-form solutions for separating co-channel emitters
• More computationally efficient than iterative ICA algorithms
• Robust to Gaussian noise due to higher-order statistics' theoretical insensitivity

A multi-linear generalization of standard SVD that decomposes cumulant tensors into a core tensor and orthogonal factor matrices. HOSVD provides optimal low-rank approximation of higher-order data while preserving the multi-dimensional structure critical for statistical feature compression.

• Extends matrix SVD to tensors of arbitrary order
• Enables truncation of least significant components for noise reduction
• Produces orthonormal bases for each mode of the cumulant tensor

A computational method that decomposes multivariate signals into statistically independent non-Gaussian components. ICA leverages cumulant tensors as contrast functions, maximizing higher-order independence to separate co-channel emitters in RF environments.

• Uses kurtosis or negentropy as measures of non-Gaussianity
• Assumes sources are statistically independent and non-Gaussian
• Widely applied in blind source separation for emitter isolation

A dimensionality reduction technique that projects signal data onto directions maximizing non-Gaussianity, isolating the subspace containing hardware-specific fingerprint information. This approach discards Gaussian-dominated dimensions that carry no device-identifying information.

• Identifies the most informative directions for emitter classification
• Suppresses Gaussian noise and interference components
• Serves as a pre-processing step before cumulant-based feature extraction

A pre-processing transformation that decorrelates and normalizes data beyond second-order statistics. While standard whitening removes only pairwise correlations, higher-order whitening prepares signals for cumulant-based feature extraction by ensuring that higher-order dependencies are isolated and accessible.

• Extends PCA whitening to account for non-Gaussian structure
• Simplifies subsequent tensor decomposition by reducing redundancy
• Critical for ensuring robust cumulant estimation in noisy environments