On principal angles between subspaces of Euclidean space (Q2706249)

Given two matrices \(X\) and \(Y\), with full column rank and common row dimension, the calculation of the cosines of the principal angles between the two subspaces generated by them can be efficiently carried out, as shown by \textit{A. Björck} and \textit{G. H. Golub} [Math. Comput. 27, 579-594 (1973; Zbl 0282.65031)], by means of the singular value decomposition of \(Q*T_x Q_y\), where \(Q_x\) and \(Q_y\) are the respective orthogonal factors in the \(QR\) decomposition of \(X\) and \(Y\). The present paper examines at first the numerical stability of the Björck-Golub algorithm and it is proved that it is stable in the framework of backward error analysis. After this, a new algorithm is proposed. Instead of directly computing \(Q_x\) and \(Q_y\), \(LU\) Gaussian factorizations of \(X\) and \(Y\) with complete (or partial) pivoting are calculated and the columns of unit lower factors are the new matrices whose columns are to be orthogonalized. NEWLINENEWLINENEWLINEThis idea, which goes back to \textit{M. J. D. Powell} and \textit{J. K. Reid} [Inform. Process., Proc. IFIP Congr., Edinburgh 1968, No. 1, 122-126 (1969; Zbl 0154.47002)], improves the original Björck-Golub algorithm in the following way. First of all, it is backward stable too. Then, although both algorithms are sensitive to certain forms of ill-conditioning in \(X\) and \(Y\), the improved algorithm usually retains its accuracy properties, while the original one does not. This is shown with a minute analysis of a nice set of interesting numerical examples.