Abstract
In the study of multilinear algebra, the spectral properties of Kronecker products are well-understood for matrices: the dominant eigenvector of a product is simply the product of the individual eigenvectors. This paper investigates whether this "decoupling" property extends to tensors. We prove that while decoupling holds for Z-eigenvectors, it fails for H-eigenvectors. Through symbolic algebra and numerical verification, we provide a concrete counterexample where the product's eigenvalue exceeds the product of the individual eigenvalues, while also identifying specific classes of tensors where decoupling is preserved.
Introduction
Tensors serve as higher-dimensional generalizations of matrices. A natural operation in this domain is the Kronecker product ($A \otimes B$), which combines two tensors into a larger structure. For matrices, the spectral behavior of such products is highly predictable. However, for tensors, the outcome depends significantly on the definition of "eigenvector" employed.
Previous research established that Z-eigenvectors (which relate to the standard Euclidean norm) do decouple. This paper addresses the open question regarding H-eigenvectors, where the eigenvalue equation involves a power-sum of vector components. We demonstrate that H-eigenvectors do not necessarily decouple, implying that the dominant behavior of a composite tensor system can be more complex than the sum of its constituent parts.
Methodology
The authors employed a hybrid approach of symbolic algebra and numerical verification to explore these spectral properties:
- Hyperdeterminants: Analytical H-eigenvalues were found using characteristic polynomials derived from tensor hyperdeterminants.
- Numerical Solvers: The researchers utilized
HomotopyContinuation.jlto solve systems of polynomial equations, ensuring all possible H-eigenpairs were identified for their counterexamples. - Rank Analysis: To prove an eigenvector $x$ does not decouple, it was reshaped into a matrix. If the resulting matrix has a rank greater than 1, it cannot be represented as a Kronecker product of two smaller vectors ($u \otimes v$), confirming the failure of decoupling.
- Theoretical Proofs: Formal proofs were provided for specific cases, such as diagonal and nonnegative tensors, where decoupling is maintained.
Key Insights and Results
The study reveals that the geometric property of scale invariance in H-eigenvectors is the primary reason the decoupling property breaks down, unlike the geometry underlying Z-eigenpairs.
The "Larger than the Sum" Effect
In the primary counterexample involving $2 \times 2 \times 2$ symmetric tensors, the largest H-eigenvalue of the Kronecker product ($\approx 1.0352$) was strictly greater than the product of the largest H-eigenvalues of the component tensors ($\approx 1.0345$). This suggests that the Kronecker product can "unlock" spectral growth not present in the individual components.
Decoupling Conditions
Despite the general failure, the authors formally proved that decoupling does hold under specific conditions:
- Diagonal tensors.
- Nonnegative, irreducible tensors: This mirrors the Perron-Frobenius theorem in matrix theory, where nonnegativity guarantees well-behaved dominant eigenvalues.
Conclusion and Future Work
This research settles a fundamental question in tensor spectral theory by showing that dominant H-eigenvectors do not generally decouple. While the paper identifies specific classes where the property holds, the broader landscape of "special cases" remains an open area of study. Future work is needed to develop a general theory explaining exactly when and by how much decoupling fails, particularly for non-symmetric tensors.
Summary updated