Reduction to condensed form¶
Computing the eigenvalues or singular values of a matrix, as a general rule, boils down to an iterative procedure. Naive applications of this iterative algorithm would often both require \(O(n^4)\) work for \(n \times n\) matrices and be less numerically stable than first spending \(O(n^3)\) work to condense the matrix to a form where each of roughly \(O(n)\) iterations requires at most \(O(n^2)\) work. In the case of a full eigenvalue decomposition of a Hermitian matrix, a similarity transformation composed of Householder transformations is applied to condense the matrix down to real symmetric tridiagonal form, where an iterative algorithm can be quickly (read: in at most cubic time) applied. Similarly, it is standard practice to condense a general matrix to bidiagonal form in order to compute its Singular Value Decomposition, and to reduce a general square matrix to upper Hessenberg form in order to compute its Schur decomposition. In each of these cases, it is possible cast a significant portion of the reduction to condensed form into high-performance matrix-matrix multiplications [DSH1989].
|[DSH1989]||Jack J. Dongarra, Danny C. Sorensen, and Sven J. Hammarling, Block reduction of matrices to condensed forms for eigenvalue computations, Journal of Computational and Applied Mathematics, Vol. 27, Issues 1–2, pp. 215–227, 1989. DOI: http://dx.doi.org/10.1016/0377-0427(89)90367-1|