No label defined (Q4389719)

Mainly linear systems of equations \(Ax= b\) are the focus of this thesis which contains slightly more than 180 pages. Here, \(A\) is an (almost) arbitrary real \(n\times n\) matrix and \(b\) is a vector with \(n\) components. The system is solved by means of a multiparametric gradient method which has the form \[ x_{k+ 1}= x_k+ \Biggl(\sum^\ell_{j= 1}\omega^{(k)}_j H_j\Biggr)r_k,\quad r_{k+ 1}= r_k- A\Biggl(\sum^\ell_{j= 1}\omega^{(k)}_j H_j\Biggr)r_k,\quad k= 0,1,\dots. \] The starting vector \(x_0\) is given; \(r_0= b-Ax_0\) is the residual with respect to \(x_0\), the index bound \(\ell\) is a fixed positive integer. The matrices \(H_j\) are to be chosen appropriately and can be interpreted as partial preconditioners which sum up (after having been multiplied by the weights \(\omega^{(k)}_j\)) to the preconditioning matrix \(H^{(k)}= \sum^\ell_{j= 1}\omega^{(k)}_j H_j\). By its additive form the method is suited for parallel processing. The thesis deals with the problem how to choose \(H_j\) and \(\omega^{(k)}_j\) such that the speed of convergence is increased. It uses tools from discrete optimization; in particular, the simplex algorithm is applied. Chapter 1 introduces the subject, Chapter 2 presents auxiliary results from linear algebra, from approximation theory and on roundoff errors. Chapter 3 recalls one-parametric (= classical) gradient methods. It considers their properties and convergence including preconditioning and roundoff error analysis. Chapter 4 deals with multiparametric gradient methods, derives properties and studies the behaviour of convergence by minimizing the spectral norm of the iteration matrix. In addition, it relates the method to known classes of methods by specializing the preconditioning matrices \(H_j\). Chapter 5 touches on one-parametric and multiparametric nonlinear gradient methods and Chapter 6 contains a résumé and prospects.