Graduate Course on Iterative Methods for Large Scale Problems

This graduate course is organized in the context of the UCL course MAPA3070: SPECIAL TOPICS IN NUMERICAL LINEAR ALGEBRA and the KUL course HP47: NUMERICAL SIMULATION. The seminar series is co-sponsored by the Scientific Research Network of the Fund for Scientific Research Flanders “Advanced Numerical Methods for Mathematical Modelling” and by the Inter University Pole of Attraction V/22 “Dynamical Systems and Control: Computation, Identification and Modelling”.

Seminar Overview
  1.  Krylov Methods - An Introduction
  2.  Krylov Subspace Methods - Practical Variants
  3.  Iterative Methods for Large Linear Eigenvalue Problems
  4.  Preconditioning of Iterative Solvers for PDE’s
  5.  Algebraic Multigrid and Multilevel Methods: A General Introduction

Seminar Details

1. Krylov Methods - An Introduction

April 15, 2005, KUL, Auditorium 200N: 14:00-17:30     Slides
Paul Van Dooren (Université Catholique de Louvain, Belgium)

Abstract:
This introduction derives the general ideas of Krylov based methods starting from two matrix decompositions: one based on orthogonal similarity transformations and one based on arbitrary similarity transformations. From these matrix decompositions we obtain the basic recurrence relations of the Arnoldi and Lanczos algorithms and discuss possible breakdowns. We then show how these lead to different classes of solvers for linear systems of equations and to projection methods for the computation of the spectrum of a matrix. This talk develops the basic background for the more specialized lectures in this series.

 

2. Krylov Subspace Methods - Practical Variants

April 22, 2005, UCL, Auditoire BARB06, Hall Ste Barbe: 14:00-17:30     Slides
Henk van der Vorst (Utrecht University, the Netherlands)

Abstract:
We will present an overview of a number of related iterative methods for the solution of linear systems of equations. These methods are so-called Krylov projection type methods and they include popular methods as Conjugate Gradients, Bi-Conjugate Gradients, QMR, LSQR and GMRES. We will discuss some useful variants of these methods, in particular Bi-CGSTAB and GMRESR. Rounding errors may have serious impact on the results of iterative methods and we will discuss these effects and we will see how these effects can be repaired. Iterative methods are often used in combination with so-called preconditioning operators (approximations for the inverses of the operator of the system to be solved) in order to speed up convergence. We will give a short overview of these techniques, details will be given in other lectures in this series.

 

3. Iterative Methods for Large Linear Eigenvalue Problems

April 29, 2005, KUL, Auditorium 200N: 14:00-17:30     Slides
Henk van der Vorst (Utrecht University, the Netherlands)

Abstract:
Suppose we want to compute one or more eigenvalues and their corresponding eigenvectors of the n x n matrix A. Several iterative methods are available: Jacobi's method, the power method, the method of Lanczos, Arnoldi's method, and Davidson's method. We will discuss how these methods work and how they may be used successfully. Most attention will be given to the methods of Lanczos and Arnoldi. It is quite easy to use these methods, but it is not so easy to select the right information from the produced output. A variant of the Arnoldi method is known as Davidson's method and this method has been reported as being quite successful, most notably in connection with certain symmetric problems in computational chemistry. The success of the method seems to depend quite heavily on (strong) diagonal dominance of A. However, as we will show, Davidson's method has an interesting connection with an old and almost forgotten method of Jacobi. This leads to another view on the method of Davidson, that may help us to explain the behaviour of the method, and that leads to newpowerful algorithms for non-diagonally dominant and unsymmetric matrices as well, the so-called Jacobi-Davidson methods. From these algorithms, we can obtain essentially different approximations for eigenvalues and these can be used to our advantage.

 

4. Preconditioning of Iterative Solvers for PDE’s

May 13, 2005, UCL, Auditoire SUD01, Croix-SUD: 14:00-17:30     Slides
Andy Wathen (Oxford University, UK)

Abstract:
This seminar will describe preconditioning approaches for linear systems deriving from finite difference and finite element approximation of partial differential equations. We will cover self-adjoint problems which give rise to symmetric and positive definite matrices, saddle-point systems which yield symmetric indefinite matrices and non-self-adjoint problems where non-symmetric matrices result. Most of the examples relate to problems of fluid flow, though the coverage in the lectures should be such that familiarity with PDE’s in general and the incompressible Navier-Stokes equations in particular will not be a prerequisite. The approach will be rather algebraic and relevant PDE theory will only be introduced as and when required.

 

5. Algebraic Multigrid and Multilevel Methods: A General Introduction

May 20, 2005, KUL, Auditorium 200N: 14:00-17:30     Slides
Yvan Notay (Université Libre de Bruxelles, Belgium)

Abstract:
Multigrid and multilevel methods, alone or as preconditioner for (Krylov based) iterative methods, are increasingly popular as applications grow in size. They indeed allow to solve discrete PDEs with memory and CPU requirements that remain proportional to the number of unknowns. To face increasingly complex situations, there is in particular a high interest in "algebraic" methods, that can be used robustly in a "black box" fashion, independently of the underlying geometry or discretization.
In this lecture, we first introduce the different type of multigrid & multilevel methods, illustrating their basic principles and ideas with simple "geometric" examples. We next show how each of these methods is easily described algebraically using a few key ingredients. We then browse available theoretical results to identify the necessary and sufficient conditions for fast convergence. In passing, this allows to display some potential weaknesses of "geometric" algorithms. On this basis, we discuss how the different key ingredients may be defined by algebraic manipulations, reviewing several approaches that have been proposed in the literature. We conclude with a few illustrations.