Setting up alternating least squares and iterative majorization algorithms for solving various matrix optimization problems_Image_Marked资料.pdf

Computational Statistics Data Analysis 41 (2002) 157–170 /locate/csda Setting up alternating least squares and iterative majorization algorithms for solving various matrix optimization problems Henk A.L. Kiers∗ Heymans Institute (DPMG), University of Groningen, Grote Kruisstraat 2=1, 9712 TS Groningen, The Netherlands Received 1 November 2001; received in revised form 1 March 2002 Abstract A general procedure is described for setting up monotonically convergent algorithms to solve some general matrix optimization problems, if desired, subject to a wide variety of constraints. An overview is given of a number of ready-made building blocks (derived in earlier publica- tions) from which concrete algorithms are set-up with little eort. These algorithms are based on alternating least squares (block relaxation) and iterative majorization. It is demonstrated how the construction of an algorithm for a particular problem that falls in one of the classes of optimization problems under study, reduces to a simple combination of tools. Also, a proce- dure is reviewed for setting up a weighted least squares algorithm for any problem for which an unweighted least squares solution is available. All procedures are illustrated by means of examples. c 2002 Elsevier Science B.V. All rights reserved. 1. Introduction In multivariate statistics, problems of optimization of matrix functions are often en- countered. Sometimes these have straightforward closed-form solutions, but also often such solutions are not a