Canonical correlations and generalized SVD: Applications and new algorithms (Q1122976): Difference between revisions

New algorithms are developed for computing singular value decompositions (SVD) and canonical correlations (CC) [cf. \textit{H. Hotelling}, Biometrika 28, 321-377 (1936; Zbl 0015.40705)]. Three matrices A(n\(\times p)\), H(n\(\times n)\) and K(p\(\times p)\) are given, H and K are symmetric positive definite and two transformations Y and Z are sought such that \(Y^{-1}AZ=D\), where \(Y^ THY=I_ n\) and \(Z^ TKZ=I_ p\) and D(n\(\times p)\) is diagonal. The new algorithms are based on a generalization of the SVD method of the second author [J. Parallel Distrib. Comput. 2, 250-260 (1985)]. Weighted least squares are calculated by the SVD and CC procedure (applied e.g. in aircraft wing flutter analysis). Optimal prediction by minimizing the weighted prediction error by the CC and SVD method is discussed. The product of 3 matrices by SVD is calculated and two algorithms are developed to this aim. The computational complexity is estimated.