Computing extreme eigenvalues of large scale Hankel tensors (Q2399200)

The authors focus on eigenvalues of large scale tensors, which have many important applications in science and engineering, whenever massive data are exploited. Here, in particular, a class of large scale dense tensors with a special Hankel structure is considered. The problem of computing extreme eigenvalues of Hankel tensors is modeled as a nonlinear optimization problem with a unit spherical constraint. Actually, the proposed algorithm is an inexact steepest descent method on the unit sphere. Since there exists a fast algorithm to compute products of a Hankel tensor and vectors, exploiting the fast Fourier transform, the computational cost of each iteration of this new method is quite low and is about \ \(\mathcal{O}(mn\,\log (mn))\), where \( m\) and \(n\) are the order and the dimension, respectively, of the Hankel tensor. The convergence of iterates is analyzed in details. Many numerical experiments are reported to show the effectiveness of this promising method. Remarkably, examples with dimensions up to one million are shown to be computed on a desktop computer in a very short time.