Scientific data analysis. An introduction to overdetermined systems

From MaRDI portal





The main title of this interesting book is misleading. Scientific data analysis does not only cover some regression methods but also techniques like cluster analysis, discriminant analysis, canonical correlation, etc. But also the subtitle is not precise. The first three chapters (floating point numbers, norms and condition numbers, sparse matrices) contain a very good introduction into computer-oriented numerical analysis. The problem \(\| Ax-b\|_ p\to \min\) is posed in chapter 4. The author favours \(p=2\) and \(p=1\) and merely states that ``other choices for p are possible, but seldom used. (In practice it turns out that e.g. a value of \(p\approx 1.2\) may combine the advantages of \(p=2\) and \(p=1)\). In chapter 5 (linear least squares) the normal equations approach and orthogonalization methods (Givens, Householder) are presented. The better conditioned second approach needs more operations and thus may produce more rounding errors. I share the author's well-founded opinion that the method of normal equations is better than it was said to be in the past two decades. Iteratively reweighted and constrained least squares are also considered. The method of Barrodale \& Roberts for \(p=1\) is presented in chapter 6, Though I favour this algorithm, too, the faster methods of Abdelmalek (1980), Armstrong/Frome/Kung (1979), and Bloomfield/Steiger (1980,1983) should at least have been mentioned. Chapter 7 considers nonlinear least squares (gradient methods and Nelder/Mead). Finally chapter 8 introduces into the singular value decomposition (does the author know the Golub/Van Loan book on matrix computations?) and into total (orthogonal) least squares. Five out of eight chapters contain listings of FORTRAN programs that are also available on tape (why not on a diskette?). Most of the programs can be found in similar form elsewhere. The references are not always up-to-date. This book contains a lot of remarkable and original hints for a numerical analysist which are based on both a thorough knowledge of theory and application. Thus most of the readers ought to be numerical analysists but not (as intended) scientists or engineers gathering data and (normally) not willing to study such a book. Considering the structure of the book and sometimes when reading in it one got the impression that the author should better have listened to Mozart than to Bruckner, Mahler, and Verdi (as he announces) when making the conception of the book and when writing it. Nevertheless, it is an enrichment for every library.












This page was built for publication: Scientific data analysis. An introduction to overdetermined systems

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3996826)