In linear algebra, matrices possess well-behaved structural properties that guarantee unique solutions to equations, optimization problems, and stability analyses. For instance, if a matrix is positive definite, an associated quadratic form has a unique minimum. If it is an M-matrix or sign-nonsingular, we gain immediate guarantees about system inversion and positivity.
However, when we transition from 2D matrices to higher-order tensors (multi-dimensional arrays), these guarantees shatter. A tensor represents a multi-linear or higher-degree homogeneous polynomial system. Because these systems are inherently non-linear, they can exhibit multiple solutions, chaotic dynamics, or intractable optimization landscapes. To restore predictability, researchers study how classical matrix properties can be extended to tensors to guarantee existence, uniqueness, and stability.
Extending Positivity and Monotonicity
To bridge the matrix-to-tensor gap, we must first redefine what it means for a tensor to be "positive" or "structured." In matrix theory, an M-matrix is a square matrix with non-positive off-diagonal entries and a positive inverse. Qi & Wei extended this concept to higher-order tensors, establishing that semi-positivity and monotonicity can define a class of nonsingular $\mathcal{M}$-tensors. This mapping allows insights from linear matrix theory to apply to multi-dimensional data structures.
Similarly, positive semi-definiteness (PSD) becomes more nuanced in higher dimensions. For instance, Qi & Wei investigated "strong" Hankel tensors—tensors whose underlying matrix representation is PSD. They proved that these tensors possess a sum-of-squares (SOS) decomposition, which makes them computationally tractable. However, characterizing the minimal structural counterexamples—tensors that are PSD but not strong—remains an open boundary problem in understanding tensor-based positivity.
Achieving Uniqueness through Structural Constraints
Without structural constraints, tensor problems are frequently ill-posed. Consider tensor recovery or completion, where we attempt to reconstruct a full tensor from limited observations. Chen et al. demonstrated that to ensure a unique solution when recovering pairwise interaction tensors, specific constraints (such as row or column sums) must be imposed on the constituent matrices to resolve inherent decomposition ambiguities.
Uniqueness can also be guaranteed purely through qualitative structures, such as signs. In matrix theory, a sign-nonsingular (SNS) matrix has a determinant whose sign is invariant regardless of the specific numerical values of its entries, ensuring a unique solution to homogeneous equations. Bu et al. extended this to sign pattern tensors. The authors showed that by verifying specific diagonal configurations or factorization into diagonal and permutation matrices, one can guarantee a sign-stable inverse, ensuring unique solutions based entirely on the qualitative sign distribution.
This quest for structural characterization extends to geometric and factorization contexts as well. For example, in computer vision, Heyden established a unified algebraic framework for multiple view tensors, though determining the exact necessary and sufficient conditions for a set of numbers to represent a valid quadfocal tensor remains an open question. Similarly, in tensor decompositions, Liu et al. defined "s-diagonal tensors" to precisely characterize the internal diagonal structures left underspecified by standard Tensor Singular Value Decompositions (T-SVD).
Even when analyzing eigenvectors, higher-order tensors introduce complexity because a symmetric tensor can have many non-equivalent eigenvectors. Usevich & Robeva explored the conditions under which eigenvectors become "robust" attractors for the tensor power method, conjecturing that specific structural types, like equiangular tensors, can isolate a unique set of robust eigenvectors.
Solving Tensor Complementarity Problems
The ultimate test of these structural properties lies in the Tensor Complementarity Problem (TCP), a non-linear generalization of linear complementarity problems common in game theory and optimization. A TCP asks us to find a vector $x \ge 0$ such that:
$$x \cdot (\mathcal{A}x^{m-1} + q) = 0 \quad \text{and} \quad \mathcal{A}x^{m-1} + q \ge 0$$
Because TCPs are highly non-linear, finding a solution—let alone a unique one—is incredibly difficult without strict tensor structures. Qi et al. demonstrated that restricting the problem to symmetric positive-definite or copositive tensors guarantees that the solution sets are nonempty and compact. Furthermore, the authors highlighted that tensor diagonalizability acts as a critical mathematical lever to guarantee a unique solution, mirroring the role of diagonalizable matrices in linear algebra.
Copositivity—where a tensor yields non-negative values for all non-negative vectors—also arises naturally in physics. Song derived analytic copositivity conditions for fourth-order symmetric tensors used to model scalar dark matter potentials, directly linking tensor copositivity to vacuum stability and the uniqueness of solutions in physical systems.
When it comes to actually computing these solutions, the underlying tensor structure dictates whether an algorithm will succeed. Wang et al. developed a nonlinear gradient dynamical system to solve TCPs efficiently. However, they showed that the guaranteed convergence of such dynamical solvers to a unique solution relies heavily on the tensor being a "strong P-tensor." For unstructured or general tensors, the stability and reliability of these solvers remain completely unresolved.
Key Foundational Work
Exact and Stable Recovery of Pairwise Interaction Tensors (2013) — contributed 2 concept elements to this direction.
A Common Framework for Multiple View Tensors (1998) — contributed 1 concept elements to this direction.
<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:mi mathvariant="script">M</mml:mi></mml:math>-tensors and nonsingular<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:mi mathvariant="script">M</mml:mi></mml:math>-tensors (2013) — contributed 1 concept elements to this direction.