Overview of Constrained PARAFAC Models

In this paper, we present an overview of constrained PARAFAC models where the constraints model linear dependencies among columns of the factor matrices of the tensor decomposition, or alternatively, the pattern of interactions between different modes of the tensor which are captured by the equivalent core tensor. Some tensor prerequisites with a particular emphasis on mode combination using Kronecker products of canonical vectors that makes easier matricization operations, are first introduced. This Kronecker product based approach is also formulated in terms of the index notation, which provides an original and concise formalism for both matricizing tensors and writing tensor models. Then, after a brief reminder of PARAFAC and Tucker models, two families of constrained tensor models, the co-called PARALIND/CONFAC and PARATUCK models, are described in a unified framework, for $N^{th}$ order tensors. New tensor models, called nested Tucker models and block PARALIND/CONFAC models, are also introduced. A link between PARATUCK models and constrained PARAFAC models is then established. Finally, new uniqueness properties of PARATUCK models are deduced from sufficient conditions for essential uniqueness of their associated constrained PARAFAC models.

💡 Research Summary

This paper provides a comprehensive overview of constrained PARAFAC (parallel factor) models, focusing on how linear dependencies among factor‑matrix columns—or, equivalently, structured interactions among tensor modes—can be incorporated into tensor decompositions. The authors begin by reviewing essential tensor algebra, emphasizing a Kronecker‑product based approach to mode combination. By expressing mode concatenation through Kronecker products of canonical vectors, they obtain compact matrix‑unfolding formulas that simplify the derivation of matricized tensor models. An index‑notation formalism is introduced, offering a concise way to write both tensor matricizations and model equations.

After recalling the classic PARAFAC (also known as CANDECOMP/CP) and Tucker decompositions for N‑th‑order tensors, the paper concentrates on two families of constrained models: PARALIND/CONFAC and PARATUCK. In PARALIND/CONFAC models, each factor matrix A⁽ⁿ⁾ is premultiplied by a constraint matrix Φ⁽ⁿ⁾, thereby imposing linear relationships among its columns. These constraint matrices may have block‑diagonal, Vandermonde, Toeplitz, or other structured forms, reflecting physical, design, or allocation constraints. PARATUCK models, on the other hand, are Tucker‑type decompositions where the core tensor G is endowed with a specific structure (e.g., block‑diagonal, diagonal, or patterned). The authors show that a PARATUCK model can be recast as a constrained PARAFAC model, establishing a direct link between the two families.

Beyond these established families, the paper introduces two novel constructions. Nested Tucker models consist of multiple hierarchical core tensors, enabling the representation of complex multi‑stage interactions such as those found in nonlinear system identification. Block PARALIND/CONFAC models concatenate several independent PARALIND blocks into a larger block‑diagonal tensor, which is useful for data exhibiting block structure (e.g., multi‑user MIMO, block‑coded communications).

A major contribution lies in the uniqueness analysis. By transforming constrained PARAFAC and PARATUCK models into equivalent unconstrained forms, the authors extend Kruskal’s classic uniqueness condition to accommodate the presence of constraint matrices. They derive sufficient conditions under which the factor matrices remain essentially unique despite the imposed linear dependencies. These results are then applied to PARATUCK models, yielding new essential‑uniqueness theorems that guarantee identifiability of the underlying parameters.

The paper concludes with a discussion of practical implications. Constrained tensor models have been successfully applied in wireless communications (e.g., block‑Tucker CDMA schemes, Volterra‑based nonlinear modeling), blind source separation using cumulant tensors, chemometrics, hyperspectral imaging, and other domains where physical interpretability, reduced parameter count, or structured sparsity are desirable. By providing a unified mathematical framework, novel model extensions, and rigorous uniqueness conditions, this work advances both the theory and the applicability of constrained multi‑way tensor decompositions.