E-characteristic polynomials of tensors (Q359704)

The paper is devoted to establish a number of results concerning E-characteristic polynomials of a tensor. Let \(V\) be an \(n\)-dimensional \(\mathbb{R}\)-vector space with an Euclidean inner product. An (\(n\)-dimensional) \(m\)th-order covariant tensor on \(V\) is a multilinear function \(\mathcal{A}:V^m\to\mathbb{R}\). In particular, a first-order covariant tensor is called a covariant vector and is denoted by \(\mathbf{x}\). Given an orthonormal frame \(E:=\{e_1,\dots,e_n\}\) in \(V\), define \(a_{i_1i_2\dots i_m}:=\mathcal{A}(e_{i_1},\dots,e_{i_m})\) for any \((i_1,i_2,\dots, i_m)\in\{1,\dots,n\}^m\) and \(x_i:=\mathbf{x}(e_i)\) for \(1\leq i\leq n\). Then \(A:=(a_{i_1i_2\dots i_m})\) and \(x:=(x_1,\dots,x_n)\) are called the hypermatrix of \(\mathcal{A}\) and \(\mathbf{x}\) with respect to \(E\). To an \(m\)th-order covariant tensor \(\mathcal{A}\) and a covariant vector \(\mathbf{x}\) as above, one may associate a covariant vector \(\mathcal{A}\mathbf{x}^{m-1}\) in the following way: \[ \mathcal{A}\mathbf{x}^{m-1}(e_i):=\sum_{i_2,\dots,i_m=1}^n \mathcal{A}(e_i,e_{i_2},\dots,e_{i_m})\mathbf{x}(e_{i_2}) \dots \mathbf{x}(e_m). \] Further, define \(\mathbf{x}^T\mathbf{x}:=\sum_{i=1}^n(\mathbf{x}(e_i))^2\). If \(\lambda\in\mathbb{R}\) and \(\mathbf{x}\) satisfy the identities \[ \mathcal{A}\mathbf{x}^{m-1}=\lambda \mathbf{x},\quad \mathbf{x}^T\mathbf{x}=1, \] then \(\lambda\) is called an E-eigenvalue and \(\mathbf{x}\) is called an E-eigenvector of \(\mathcal{A}\). All these definitions are orthonormal invariant, namely invariant under changes of orthonormal frames. The second author [J. Symb. Comput. 40, No. 6, 1302--1324 (2005; Zbl 1125.15014); J. Math. Anal. Appl. 325, No. 2, 1363--1377 (2007; Zbl 1113.15020)] introduced a notion of an E-characteristic polynomial \(\psi_{\mathcal{A}}(\lambda)\) of a tensor \(\mathcal{A}\). Such a characteristic polynomial can be expressed as a sum \[ \psi_{\mathcal{A}}(\lambda)=\sum_{j=0}^{h(m,n)}a_j(m,n)\lambda^j, \] where each \(a_j(m,n)\) is a polynomial in the entries \(a_{i_1i_2\dots i_m}:=\mathcal{A}(e_{i_1},\dots,e_{i_m})\) of the hypermatrix \(A\) of \(\mathcal{A}\) in a given frame \(E\). In the paper under review it is shown that all the coefficients of the E-characteristic polynomial of a tensor \(\mathcal{A}\) are orthonormal invariants. Further, formulae for the constant coefficient \(a_0(m,n)\) of \(\psi_{\mathcal{A}}(\lambda)\) are obtained. Determinantal formulas for the E-characteristic polynomial when \(n=2\) are discussed. Finally, the leading coefficient of \(\psi_{\mathcal{A}}(\lambda)\) is discussed. A tensor \(\mathcal{A}\) as above is called irregular if there exists a nonzero covariant vector \(\mathbf{x}\) such that \[ \mathcal{A}\mathbf{x}^{m-1}=0,\quad \mathbf{x}^T\mathbf{x}=0, \] and is called regular if it is not irregular. When a tensor is regular or has an infinite number of eigenpairs then the situation is easy to handle. In order to deal with tensors with a finite number of eigenpairs, a perturbation method is applied which approximates the given tensor by regular ones with a finite number of eigenpairs. For this purpose, the authors show that the set of tensors with an infinite number of eigenpairs and the set of irregular tensors have codimension at least 2 in the projective space of tensors. This allows the authors to conclude that their perturbation can be applied.