Fast solving of weighted pairing least-squares systems (Q2389536)

The well organized paper deals with so called weighted pairing least-squares (WPLS) as a generalization of the ordinary weighted least-squares problem (WLS). The method is used namely in data alignment problems, where the correspondence between the data points \(X\) and \(Y\) is given by some non-negative measure for each data pair and described by a rectangular matrix of weights. Two fast solutions of the WPLS problem are suggested: Algorithm 1 computes the generalized Cholesky factorization of a given symmetric positive semi-definite matrix. The algorithm provides a suitable factor even if the original matrix is singular. The factor is used to define the minimum norm solution of WPLS. Algorithm 2 is designed to find a \(\{1,2,3\}\)-inverse of the given generalized Cholesky factor. In order to test the performance of the methods, suitable test problems are build. In all cases, the comparison of the results showed that the method based on the \(\{1,2,3\}\)-inverse is faster and more accurate than the fast Moore-Penrose inverse based method. The authors also remark that the rank deficient systems are solved faster then full rank systems. In Appendix, the detailed Matlab codes of both algorithms are given.