The partial total least squares algorithm (Q1098234)

Let (A,B) be an \(m\times (n+d)\)-matrix of rank r. To solve the overdetermined system \(AX=B\) by ``total least-squares (TLS)'' means to look for matrices \(\hat A\), \(\hat B\) with \(R(\hat B)\subseteq R(\hat A)\) such that \(\| (A,B)-(\hat A,\hat B)\|_ F\) is minimal and to solve \(\hat AX=\hat B\). The classical TLS-algorithm performs a singular value decomposition of the matrix (A,B). However, only the right singular subspaces corresponding to the smallest \(n+d-r\) singular values are actually needed. They can be computed effectively with the ``partial singular value decomposition (PSVD)'' algorithm. Replacing the singular value decomposition by PSVD leads to the ``PTLS-algorithm'' suggested by the authors for the TLS-problem. Although the PTLS-algorithm needs more storage space than classical TLS, the operation count is naturally in favor of PTLS. In the examples given in the paper, PTLS is about 2 times faster than classical TLS.