Orthogonal Iterations on Companion-Like Pencils

Abstract

We present a class of fast subspace algorithms based on orthogonal iterations for structured matrices/pencils that can be expressed as small rank perturbations of unitary matrices. The representation of the matrix by means of a new data-sparse factorization—named LFR factorization—using orthogonal Hessenberg matrices is at the core of these algorithms. The factorization can be computed at the cost of $$O(n k^2)$$ O ( n k 2 ) arithmetic operations, where n and k are the sizes of the matrix and the small rank perturbation, respectively. At the same cost from the LFR format we can easily obtain suitable QR and RQ factorizations where the orthogonal factor Q is a product of orthogonal Hessenberg matrices and the upper triangular factor R is again given into the LFR format. The orthogonal iteration reduces to a hopping game where Givens plane rotations are moved from one side to the other side of these two factors. The resulting new algorithms approximate an invariant subspace of size s associated with a set of s leading or trailing eigenvalues using only O(nks) operations per iteration. The number of iterations required to reach an invariant subspace depends linearly on the ratio $$|\lambda _{s+1}|/|\lambda _s|$$ | λ s + 1 | / | λ s | . Numerical experiments confirm the effectiveness of our adaptations.