Nonlinear power-like iteration by polar decomposition and its application to tensor approximation

Abstract

Low rank tensor approximation is an important subject with a wide range of applications. Most prevailing techniques for computing the low rank approximation in the Tucker format often first assemble relevant factors into matrices and then update by turns one factor matrix at a time. In order to improve two factor matrices simultaneously, a special system of nonlinear matrix equations over a certain product Stiefel manifold must be resolved at every update. The solution to the system consists of orbit varieties invariant under the orthogonal group action, which thus imposes challenges on its analysis. This paper proposes a scheme similar to the power method for subspace iterations except that the polar decomposition is used as the normalization process and that the iteration can be applied to both the orbits and the cross-sections. The notion of quotient manifold is employed to factor out the effect of orbital solutions. The dynamics of the iteration is completely characterized. An isometric isomorphism between the tangent spaces of two properly identified Riemannian manifolds is established to lend a hand to the proof of convergence.