Past Computational Mathematics and Applications Seminar

The use of preconditioned Newton-Krylov methods is in many applications mandatory for computing efficiently the solution of large nonlinear systems of equations. However, the available preconditioners are often sub-optimal, due to the changing nature of the linearized operator. This the case, for instance, for quasi-Newton methods where the Jacobian (and its preconditioner) are kept fixed at each non-linear iteration, with the rate of convergence usually degraded from quadratic to linear. Updated Jacobians, on the other hand require updated preconditioners, which may not be readily available. In this work we introduce an adaptive preconditioning technique based on the Krylov subspace information generated at previous steps in the nonlinear iteration. In particular, we use to advantage an adaptive technique suggested for restarted GMRES to enhance existing preconditioners with information about (almost) invariant subspaces constructed by GMRES at previous stages in the nonlinear iteration. We provide guidelines on the choice of invariant-subspace basis used in the construction of our preconditioner and demonstrate the improved performance on various test problems. As a useful general application we consider the case of augmented systems preconditioned by block triangular matrices based on the structure of the system matrix. We show that a sufficiently good solution involving the primal space operator allows for an efficient application of our adaptive technique restricted to the space of dual variables.

In this talk we will review classical multigrid methods and give an overview on algebraic multigrid methods, in particular the "classical" approach to AMG by Ruge and Stueben.

After that we will introduce a new class of multilevel methods. These new AMGs on one hand and exploit information based on filtering vectors and on the other hand, information about the inverse matrix is used to drive the coarsening process.

This new kind of AMG will be discussed and compared with "classical" AMG from a theoretical point of view as well as by showing some numerical examples.

We present a new stabilized mixed finite element method for the linear elasticity problem. The approach is based on the introduction of Galerkin least-squares terms arising from the constitutive and equilibrium equations, and from the relation defining the rotation in terms of the displacement.

We show that the resulting augmented variational formulation and the associated Galerkin scheme are well posed, and that the latter becomes locking-free and asymptotically locking-free for Dirichlet and mixed boundary conditions, respectively. In particular, the discrete scheme allows the utilization of Raviart-Thomas spaces of lowest order for the stress tensor, piecewise linear elements for the displacement, and piecewise constants for the rotation.

In the case of mixed boundary conditions, the essential one (Neumann) is imposed weakly, which yields the introduction of the trace of the displacement as a suitable Lagrange multiplier. This trace is then approximated by piecewise linear elements on an independent partition of the Neumann boundary whose mesh size needs to satisfy a compatibility condition with the mesh size associated to the triangulation of the domain. A reliable and efficient a-posteriori error estimate is also described. Finally, several numerical results illustrating the performance of the augmented scheme are reported.

The Cahn-Hilliard equations provides a model of phase transitions when two or more immiscible fluids interact. When coupled with the Navier-Stokes equations we obtain a model fro the dynamics of multiphase flow. This model takes into account the viscosity and densities of the various fluids present.

A finite element discretisation of the variable density Cahn-Hilliard-Navier-Stokes equations is presented. An analysis of the discretisation and a reliable efficient numerical solution method are presented.

I will describe some application areas for radial basis function, and discuss how the computational problems can be overcome by the use of preconditioning methods and fast evaluation techniques.

This Seminar has been cancelled and will now take place in Trinity Term, Week 3, 11 MAY.

Using the one-dimensional diffusion equation as an example, this seminar looks at ways of constructing approximations to the solution and coefficient functions of differential equations when the coefficients are not fully defined. There may, however, be some information about the solution. The input data, usually given as values of a small number of functionals of the coefficients and the solution, is insufficient for specifying a well-posed problem, and so various extra assumptions are needed. It is argued that looking at these inverse problems as problems in Bayesian statistics is a unifying approach. We show how the standard methods of Tikhonov Regularisation are related to special forms of random field. The numerical approximation of stochastic partial differential Langevin equations to sample generation will be discussed.

Standard finite element or boundary element methods for high frequency scattering problems, with piecewise polynomial approximation spaces, suffer from the limitation that the number of degrees of freedom required to achieve a prescribed level of accuracy grows at least linearly with respect to the frequency. Here we present a new boundary element method for which, by including in the approximation space the products of plane wave basis functions with piecewise polynomials supported on a graded mesh, we can demonstrate a computational cost that grows only logarithmically with respect to the frequency.

The main goal of this talk is to demonstrate connections between the following three big areas of research: the theory of cubature formulas (numerical integration), the discrepancy theory, and nonlinear approximation. First, I will discuss a relation between results on cubature formulas and on discrepancy. In particular, I'll show how standard in the theory of cubature formulas settings can be translated into the discrepancy problem and into a natural generalization of the discrepancy problem. This leads to a concept of the r-discrepancy. Second, I'll present results on a relation between construction of an optimal cubature formula with m knots for a given function class and best nonlinear m-term approximation of a special function determined by the function class. The nonlinear m-term approximation is taken with regard to a redundant dictionary also determined by the function class. Third, I'll give some known results on the lower and the upper estimates of errors of optimal cubature formulas for the class of functions with bounded mixed derivative. One of the important messages (well known in approximation theory) of this talk is that the theory of discrepancy is closely connected with the theory of cubature formulas for the classes of functions with bounded mixed derivative.

Parallel sparse direct solvers are an interesting alternative to iterative methods for some classes of large sparse systems of linear equations. In the context of a parallel sparse multifrontal solver (MUMPS), we describe a new dynamic scheduling strategy aiming at balancing both the workload and the memory usage. More precisely, this hybrid approach balances the workload under memory constraints. We show that the peak of memory can be significantly reduced, while we have also improved the performance of the solver.

Then, we present preliminary work concerning a parallel out-of-core extension of the solver MUMPS, enabling to solve increasingly large simulation problems.

This is joint work with P.Amestoy, A.Guermouche, S.Pralet and E.Agullo.

Through the advent of enhanced medical imaging computational modelling can now be applied to anatomically correct arterial geometries. However many flow feautures under physiological and pathological flow paraemeters, even in idealised problems, are relatively poorly understood. A commonly studied idealisation of an arterial blockage or stenosis, potentially generated by atherosclerosis, is a sinusoidally varying constricted tube. Although most physiological flow conditions are typically laminar, in this configuration turbulent flow states can arise due to the local increase in sectional Reynolds number. To examine the onset of turbulence in this geometry, under a simple non-reversing pulsatile flows, we have applied Floquet stability analysis and direct
numerical simulation.
As illustrated in the above figure, a period-doubling absolute instability mode associated with alternating tilting of the vortex rings that are ejected out of the stenosis/constriction during each pulse. This primary instability occurs for relatively large reduced velocities associated with long pulse periods (or low Womersley numbers). For lower reduced velocities the primary instability typically manifests itself as azimuthal waves (Widnall instability modes) of low wavenumber that grow on each vortex ring. We have also observed the shear layer of the steady axisymmetric flow is convectively unstable at still shorter temporal periods.
In this presentation we shall outline the challenges of modelling vascular flow problems with a particular focus on idealised stenotic flow. After briefly outlining the numerical analysis methods we shall discuss the flow investigations outlined above and their relation to more classical vortex instabilities.

Image Inpainting turns the art of image restoration, retouching, and disocclusion into a computer-automated trade. Mathematically, it may be viewed as an interpolation problem in BV, SBV, or other fancy function spaces thought suitable for digital images. It has recently drawn the attention of the numerical PDE community, which has led to some impressive results. However, stability restrictions of the suggested explicit schemes so far yield computing times that are next to prohibitive in the realm of interactive digital image processing. We address this issue by constructing an appropriatecontinuous energy functional that combines the possibility of a fast discrete minimization with high perceptible quality of the resulting inpainted images.

The talk will survey the background of the inpainting problem and prominent PDE-based methods before entering the discussion of the suggested new energy functional. Many images will be shown along the way, in parts with online demonstrations.

This is joint work with my student Thomas März.

Alan Turing introduced the sensitivity of the solution of a numerical problem to changes in its data as a way to measure the difficulty of solving the problem accurately. Condition numbers are now considered fundamental to sensitivity analysis. They have been used to measure the mathematical difficulty of many linear algebra problems, including linear systems, linear least-squares, and eigenvalue problems. By definition, unless exact arithmetic is used, it is expected to be difficultto accurately solve an ill-conditioned problem.

In this talk we focus on least-squares problems. After a historical overview of condition number for least-squares, we introduce two related condition numbers. The first is the partial condition number, which measures the sensitivity of a linear combination of the components of the solution. The second is related the truncated SVD solution of the problem, which is often used when the matrix is nearly rank deficient.

Throughout the talk we are interested in three types of results :closed formulas for condition numbers, sharp bounds and statistical estimates.

Direct methods for solving large sparse linear systems of equations are popular because of their generality and robustness. As the requirements of computational scientists for more accurate models increases, so inevitably do the sizes of the systems that must be solved and the amount of memory needed by direct solvers.

For many users, the option of using a computer with a sufficiently large memory is either not available or is too expensive. Using a preconditioned iterative solver may be possible but for the "tough" systems that arise from many practical applications, the difficulties involved in finding and computing a good preconditioner can make iterative methods infeasible. An alternative is to use a direct solver that is able to hold its data structures on disk, that is, an out-of-core solver.

In this talk, we will explain the multifrontal algorithm and discuss the design and development of a new HSL sparse symmetric out-of-core solver that uses it. Both the system matrix A and its factors are stored externally. For the indefinite case, numerical pivoting using 1x1 and 2x2 pivots is incorporated. To minimise storage for the system data, a reverse communication interface is used. Input of A is either by rows or by elements.

An important feature of the package is that all input and output to disk is performed through a set of Fortran subroutines that manage a virtual memory system so that actual i/o occurs only when really necessary. Also important is to design in-core data structures that avoid expensive searches. All these aspects will be discussed.

At the time of writing, we are tuning the code for the positive-definite case and have performance figures for real problems. By the time of the seminar, we hope to have developed the indefinite case, too.

It is well known that the computation of a few extreme eigenvalues, and the corresponding eigenvectors, of a symmetric matrix A can be rewritten as computing extrema of the Rayleigh quotient of A. However, since the Rayleigh quotient is a homogeneous function of degree zero, its extremizers are not isolated. This difficulty can be remedied by restricting the search space to a well-chosen manifold, which brings the extreme eigenvalue problem into the realm of optimization on manifolds. In this presentation, I will show how a recently-proposed generalization of trust-region methods to Riemannian manifolds applies to this problem, and how the resulting algorithms compare with existing ones.

I will also show how the Joint Diagonalization problem (that is, approximately diagonalizing a collection of symmetric matrices via a congruence transformation) can be tackled by a differential geometric approach. This problem has an important application in Independent Component Analysis.

We discuss a method for solving very large structured symmetric indefinite equation systems arising in optimization with interior point methods.

Many real-life economic models involve system dynamics, spatial distribution or uncertainty and lead to large-scale optimization problems. Such problems usually have a hidden structure: they are constructed by replication of some small generic block. The linear algebra subproblems which arise in optimization algorithms for such problems involve matrices which are not only sparse, but they additionally display a block-structure with many smaller blocks sparsely distributed in the large matrix.

We have developed a structure-exploiting parallel interior point solver for optimization problems. Its design uses object-orientated programming techniques. The progress OOPS (Object-Orientated Parallel Solver: http://www.maths.ed.ac.uk/~gondzio/parallel/solver.html) on a number of different computing platforms and achieves scalability on a number of different computing platforms. We illustrate its performance on a collection of problems with sizes reaching 10⁹ variables arising from asset liability management and portfolio optimization.

This is a joint work with Andreas Grothey.

A scale-invariant moving finite element method is proposed for the

adaptive solution of nonlinear partial differential equations. The mesh

movement is based on a finite element discretisation of a scale-invariant

conservation principle incorporating a monitor function, while the time

discretisation of the resulting system of ordinary differential equations

may be carried out using a scale-invariant time-stepping. The accuracy and

reliability of the algorithm is tested against exact self-similar

solutions, where available, and a state-of-the-art $h$-refinement scheme

for a range of second and fourth order problems with moving boundaries.

The monitor functions used are the dependent variable and a monitor

related to the surface area of the solution manifold.

First, I'll give a very brief update on our nonlinear Krylov accelerator for the usual Modified Newton's method. This simple accelerator, which I devised and Neil Carlson implemented as a simple two page Fortran add-on to our implicit stiff ODEs solver, has been robust, simple, cheap, and automatic on all our moving node computations since 1990. I publicize further experience with it here, by us and by others in diverse fields, because it is proving to be of great general usefulness, especially for solving nonlinear evolutionary PDEs or a smooth succession of steady states.

Second, I'll report on some recent work in computerized tomography from X-rays. With colored computer graphics I'll explain how the standard "filtered backprojection" method works for the classical 2D parallel beam problem. Then with that backprojection kernel function H(t) we'll use an integral "change of variables" approach for the 2D fan-beam geometry. Finally, we turn to the tomographic reconstruction of a 3D object f(x,y,z) from a wrapped around cylindical 2D array of detectors opposite a 2D array of sources, such as occurs in PET (positron-emission tomography) or in very-wide-cone-beam tomography with a finely spaced source spiral.

The famous Eckart-Young Theorem states that the (normwise) condition number of a matrix is equal to the reciprocal of its distance to the nearest singular matrix. In a recent paper we proved an extension of this to a number of structures common in matrix analysis, i.e. the structured condition number is equal to the reciprocal of the structured distance to the nearest singular matrix. In this talk we present a number of related results on structured eigenvalue perturbations and structured pseudospectra, for normwise and for componentwise perturbations.

Current methods for globalizing Newton's Method for solving systems of nonlinear equations fall back on steps biased towards the steepest descent direction (e.g. Levenberg/Marquardt, Trust regions, Cauchy point dog-legs etc.), when there is difficulty in making progress. This can occasionally lead to very slow convergence when short steps are repeatedly taken.

This talk looks at alternative strategies based on searching curved arcs related to Davidenko trajectories. Near to manifolds on which the Jacobian matrix is singular, certain conjugate steps are also interleaved, based on identifying a Pareto optimal solution.

Preliminary limited numerical experiments indicate that this approach is very effective, with rapid and ultimately second order convergence in almost all cases. It is hoped to present more detailed numerical evidence when the talk is given. The new ideas can also be incorporated with more recent ideas such as multifilters or nonmonotonic line searches without difficulty, although it may be that there is no longer much to gain by doing this.