Abstract
This paper investigates the mathematical properties of Hankel tensors, specifically focusing on the inheritance of positive semi-definiteness (PSD) and sum-of-squares (SOS) properties across different orders. We demonstrate that if a lower-order Hankel tensor is PSD, its higher-order counterpart—generated by the same vector and having an order that is a multiple of the original—is also PSD. Furthermore, we characterize "strong" Hankel tensors through an augmented Vandermonde decomposition and establish bounds for extremal H-eigenvalues based on the associated Hankel matrix. These findings provide a bridge between classical matrix theory and modern multidimensional data analysis.
Introduction
Hankel matrices, characterized by constant entries along their ascending skew-diagonals, are fundamental tools in signal processing, control theory, and numerical analysis. As data becomes increasingly multidimensional, the study of Hankel tensors—the higher-order generalization of these matrices—has become essential.
A critical challenge in tensor analysis is the determination of positive semi-definiteness (PSD), which means the homogeneous polynomial defined by the tensor never takes a negative value. For general tensors, determining PSD-ness is an NP-hard problem. However, the structured nature of Hankel tensors allows for more efficient characterization through "inheritance" properties and structural decompositions.
Inheritance of PSD and SOS Properties
One of the primary contributions of this research is the identification of inheritance properties. The authors prove that properties such as being positive semi-definite, positive definite, or a sum-of-squares (SOS) are preserved as the tensor order increases, provided the generating vector remains the same and the new order is a multiple of the original.
Specifically, if a Hankel tensor of order $m$ is PSD, then the Hankel tensor of order $qm$ (where $q$ is a positive integer) is also PSD. This finding significantly simplifies the verification process for high-order tensors, as one can often check the property on a lower-order version of the same structure. This inheritance also extends to the extremal H-eigenvalues, which are bounded by the eigenvalues of the lower-order counterparts, scaled by specific constants.
Strong Hankel Tensors and Decomposition
The paper distinguishes between general Hankel tensors and "strong" Hankel tensors. A Hankel tensor is defined as strong if its associated Hankel matrix is PSD. While the relationship between a tensor and its associated matrix is complex, strong Hankel tensors possess a highly useful structural property: they can be represented via an augmented Vandermonde decomposition.
Standard Vandermonde decompositions are often used to represent matrices, but they are insufficient to cover the entire cone of strong Hankel tensors because that cone is not closed. To address this, the authors introduce an "augmented" term—specifically $\alpha_r e_n^{\otimes m}$—to account for the closure of the cone. This allows for a concrete SOS decomposition for every strong Hankel tensor, proving that a Hankel tensor is strong if and only if it admits this augmented decomposition with positive coefficients.
Methods and Analytical Techniques
The proofs rely on several advanced mathematical frameworks:
- Convolution Formula: The authors utilize a discrete convolution approach to interpret tensor-vector products, effectively treating tensor operations as polynomial multiplications.
- Anti-circulant Embedding: By embedding Hankel tensors into larger anti-circulant tensors, the researchers leverage the fact that these can be diagonalized using Fourier matrices, simplifying the analysis of their spectral properties.
- Mathematical Induction: This is used to establish that the leading principal submatrices of rank-$r$ PSD Hankel matrices maintain full rank, a necessary condition for the proposed decompositions.
Eigenvalue Relationships
The research further explores the relationship between the H-eigenvalues of a Hankel tensor and the eigenvalues of its associated Hankel matrix. It is proven that a Hankel tensor lacks negative H-eigenvalues if its associated matrix lacks negative eigenvalues. In these cases, the extremal H-eigenvalues of the Hankel tensor are bounded by the extremal eigenvalues of the associated Hankel matrix, multiplied by specific constants. This provides a bridge between classical matrix theory and modern tensor analysis, allowing researchers to infer tensor properties from better-understood matrix characteristics.
Conclusion and Future Directions
While this work provides a comprehensive framework for strong Hankel tensors, it also highlights the "non-strong" gap. There exist Hankel tensors that are PSD but whose associated matrices are not. Characterizing this class remains an open challenge in the field. The authors acknowledge that while examples of such tensors are rare and difficult to construct, a general theory for non-strong PSD Hankel tensors is the next frontier in this line of research. Additionally, the paper poses a third inheritance property regarding general eigenvalue relationships as a conjecture for future study.
Summary updated