Efficiently Maximizing a Homogeneous Polynomial over Unit Sphere without Convex Relaxation

TL;DR

This paper introduces a nonconvex matrix reformulation and a simple iterative method to efficiently find the leading eigenvalue of high-order tensors, improving scalability over traditional semidefinite relaxation techniques.

Contribution

It proposes a novel nonconvex matrix program reformulation for tensor eigenvalue problems and demonstrates an efficient, scalable solution method using alternating direction method.

Findings

Method efficiently computes leading eigenvalues of high-order tensors.

Significant improvement in scalability and efficiency over existing methods.

Able to solve a 3rd-order 500-dimensional tensor in under 2 minutes.

Abstract

This work studies the problem of maximizing a higher degree real homogeneous multivariate polynomial over the unit sphere. This problem is equivalent to finding the leading eigenvalue of the associated symmetric tensor of higher order, which is nonconvex and NP-hard. Recent advances show that semidefinite relaxation is quite effective to find a global solution. However, the solution methods involve full/partial eigenvalue decomposition during the iterates, which heavily limits its efficiency and scalability. On the other hand, for odd degree (odd order) cases, the order has to be increased to even, which potentially reduces the efficiency. To find the global solutions, instead of convexifying the problem, we equivalently reformulate the problem as a nonconvex matrix program based on an equivalence property between symmetric rank-1 tensors and matrices of any order, which is a generalization of the existing results. The program is directly solved by a vanilla alternating direction method, which only involves the computation of leading eigenvalue/singular value of certain matrices, benefiting from the special structure of the program. Although being nonconvex, under certain hypotheses, it is proved that the algorithm converges to a leading eigenvalue of the associated tensor. Numerical experiments on different classes of tensors demonstrate that the proposed approach has a significant improvement in efficiency and scalability, while it can keep the effectiveness of semidefinite relaxation as much as possible. For instance, the proposed method finds the leading eigenpair of a third-order 500 dimensional Hilbert tensor in a personal computer within 100 seconds.