Higher-Order Singular Value Decomposition (HOSVD)

Unlike the matrix SVD which guarantees diagonalization, HOSVD produces a core tensor that is all-orthogonal. This means the slices of the core tensor are mutually orthogonal, preserving the statistical independence of extracted features across different modes (time, frequency, phase).

• Mode-n product preserves orthogonality in each dimension
• Factor matrices are unitary, ensuring energy preservation
• Enables robust separation of emitter-specific signatures from channel effects

HOSVD provides the best low-rank approximation of a tensor in the Frobenius norm sense through truncation. By discarding the smallest singular values in each mode, the decomposition achieves maximal variance retention with minimal components.

• Truncated HOSVD yields the Tucker decomposition
• Compression ratios often exceed 100:1 for cumulant tensors
• Preserves the multi-linear structure critical for non-Gaussian signal analysis

HOSVD performs independent dimensionality reduction along each tensor mode. For a 3rd-order cumulant tensor, this means compressing the frequency, time-lag, and observation dimensions separately, each with its own optimal rank.

• Each mode receives a tailored truncation rank
• Prevents over-compression of information-rich dimensions
• Factor matrices serve as learned bases for each modality

When applied to cumulant tensors, HOSVD serves as a pre-processing step for Independent Component Analysis (ICA). The orthogonal factor matrices whiten the data beyond second-order statistics, isolating the non-Gaussian subspace where emitter fingerprints reside.

• Enables separation of co-channel transmitters
• Suppresses Gaussian noise through higher-order processing
• Core tensor entries indicate interaction strength between separated sources

The vectorized core tensor provides a compact, discriminative feature vector for machine learning classifiers. By retaining only the top singular vectors in each mode, HOSVD distills the essential non-Gaussian signature into a manageable representation.

• Typical feature vectors: 50-500 elements from 10^6+ tensor entries
• Preserves quadratic phase coupling information
• Robust to Gaussian measurement noise by construction

HOSVD is computed by performing standard matrix SVD on the mode-n unfoldings of the tensor. This leverages optimized linear algebra libraries and avoids iterative optimization, making it suitable for real-time signal processing pipelines.

• Complexity scales as O(I^3) for an I×I×I tensor
• Parallelizable across modes
• Deterministic result without convergence concerns

A multi-dimensional array organizing higher-order cumulants that serves as the primary input to HOSVD. Unlike a matrix, this tensor preserves the multi-modal correlations between time lags, frequency bins, and spatial channels.

• Stores third-order (skewness) and fourth-order (kurtosis) statistics
• Enables joint blind source separation through multi-linear structure
• Dimensionality: L × L × L for third-order, where L is the number of signal lags

The broader family of multi-linear algebraic factorizations that includes HOSVD, CANDECOMP/PARAFAC (CP), and Tucker decomposition. These methods generalize matrix SVD to N-way arrays, extracting interpretable latent factors.

• CP Decomposition: Expresses tensor as sum of rank-1 components
• Tucker Decomposition: HOSVD is the core computation for this model
• Applications: chemometrics, psychometrics, and blind source separation

An algebraic technique that simultaneously diagonalizes multiple cumulant matrices to achieve blind source separation. HOSVD provides an efficient pre-processing step by compressing the cumulant tensor before diagonalization.

• Avoids iterative gradient-based optimization
• Exploits the multi-linear structure of fourth-order cumulants
• Used in JADE (Joint Approximate Diagonalization of Eigenmatrices) algorithm

A computational method that decomposes multivariate signals into statistically independent non-Gaussian components. HOSVD accelerates ICA by performing initial dimensionality reduction on cumulant tensors.

• Relies on higher-order statistics to measure non-Gaussianity
• Widely used for separating co-channel emitters in RF fingerprinting
• HOSVD pre-whitening improves convergence speed

A dimensionality reduction technique that projects signal data onto directions maximizing non-Gaussianity. HOSVD identifies the dominant singular vectors of the cumulant tensor, isolating the subspace containing hardware-specific fingerprint information.

• Suppresses Gaussian noise components automatically
• Retains only the signal-bearing dimensions of the tensor
• Critical for separating weak emitter signatures from background interference

A pre-processing transformation that decorrelates and normalizes data beyond second-order statistics. HOSVD can be applied to the whitened cumulant tensor to extract the core features used for classification.

• Extends PCA whitening to higher orders
• Prepares signals for cumulant-based feature extraction
• Ensures equal weighting of all statistical modes before decomposition