Real eigenvalues of nonsymmetric tensors (Q1753066): Difference between revisions

The authors first define the space of \(m\)-order \(n\)-dimensional tensors \(T^m(\mathbb R^n)\). For a tensor \(\mathcal A\in T^m(\mathbb R^n)\), they define a real \(Z\)-eigenpair and a real \(H\)-eigenpair of \(\mathcal A\). There are different techniques for computing these eigenpairs depending on the properties of \(T^m(\mathbb R^n)\). The authors focus on the computation of real \(Z\)- and \(H\)-eigenvalues and corresponding eigenvectors. For a symmetric tensor, the computation of real \(Z\)- and \(H\)-eigenpairs is discussed in quite a lot of papers (references are given in the paper). The situation is different for the case of real nonsymmetric tensors. The authors propose numerical methods for computing all real \(Z\)-eigenvalues (if there are finitely many ones) and all real \(H\)-eigenvalues (those are always a finite number). As a motivation, the authors study an \(m\)th order Markov chain that fits observed data by an \(m\)-order nonsymmetric tensor \(\mathcal P\in T^m(\mathbb R^n)\), the so called transition probability tensor. The stationary probability distributions can be obtained by computing the \(Z\)-eigenvalues of the nonsymmetric tensor \(\mathcal P\). In Section 3, the authors reformulate the \(Z\)-eigenvalue computation as a polynomial optimization problem. The set of \(Z\)-eigenvalues of \(\mathcal A\) is denoted \(Z(\mathbb R,\mathcal A)\). If \(Z(\mathbb R,\mathcal A)\) is a finite set the \(Z\)-eigenvalues will be computed monotonically \(\lambda_1<\lambda_2<\dots <\lambda_N\), from the smallest to the largest one. The properties of the set \(Z(\mathbb R,\mathcal A)\) are formulated and proven. The authors propose Lasserre-type semidefinite relaxations for computing \(Z\)-eigenvalues. If there are finitely many real \(Z\)-eigenvalues, each of them can be computed by solving a finite sequence of semidefinite relaxations. Section 4 proposes Lasserre-type semidefinite relaxations for computing all \(H\)-eigenvalues. Each of them can be also computed by solving a finite sequence of semidefinite relaxations. The paper contains algorithms for computing \(Z\) and \(H\)-eigenvalues for nonsymmetric tensors using Lasserre-type semidefinite relaxations. Many examples are provided. The authors also include a comparison with other methods.