Past

/notices/events/abstracts/topology/lackenby.shtml

/notices/events/abstracts/applied-analysis/ht05/ogden.shtml

Given a countable set of sites S an a transition matrix p(x,y) on that set, we consider a process of particles evolving on S according to the following rule: each particle waits an exponential time and then jumps following a Markov chain governed by p(x,y); the particle keeps jumping until it reaches an empty site where it remains for another exponential time. Unlike most interacting particle systems, this process fails to

have the Feller property. This causes several technical difficulties to study it. We present a method to prove that certain measures are invariant for the process and exploit the Kolmogorov zero or one law to study some of its unusual path properties.

According to the Stokes-Einstein law, microscopic particles subject to intense bombardment by (much smaller) gas molecules perform Brownian motion with a diffusivity inversely proportion to their radius. Smoluchowski, shortly after Einstein's account of Brownian motion, used this model to explain the behaviour of a cloud of such particles when, in addition their diffusive motion, they coagulate on collision. He wrote down a system of evolution equations for the densities of particles of each size, in particular identifying the collision rate as a function of particle size.

We give a rigorous derivation of (a spatially inhomogeneous generalization of) Smoluchowski's equations, as the limit of a sequence of Brownian particle systems with coagulation on collision. The equations are shown to have a unique, mass-preserving solution. A detailed limiting picture emerges describing the ancestral spatial tree of particles making up each particle in the current population. The limit is established at the level of these trees.

Spectral projections enjoy high order convergence for globally smooth functions. However, a single discontinuity introduces O(1) spurious oscillations near the discontinuity and reduces the high order convergence rate to first order, Gibbs' Phenomena. Although a direct expansion of the function in terms of its global moments yields this low order approximation, high resolution information is retained in the global moments. Two techniques for the resolution of the Gibbs' phenomenon are discussed, filtering and reprojection methods. An adaptive filter with optimal joint time-frequency localization is presented, which recovers a function from its N term Fourier projection within the error bound \exp(-Nd(x)), where d(x) is the distance from the point being recovered to the nearest discontinuity. Symmetric filtering, however, must sacrifice accuracy when approaching a discontinuity. To overcome this limitation, Gegenbauer postprocessing was introduced by Gottlieb, Shu, et al, which recovers a function from its N term Fourier projection within the error bound \exp(-N). An extension of Gegenbauer postprocessing with improved convergence and robustness properties is presented, the robust Gibbs complements. Filtering and reprojection methods will be put in a unifying framework, and their properties such as robustness and computational cost contrasted. This research was conducted jointly with Eitan Tadmor and Anne Gelb.

The use of weighted matchings is becoming increasingly standard in the

solution of sparse linear systems. While non-symmetric permutations based on these

matchings have been the state-of-the-art for

several years (especially for direct solvers), approaches for symmetric

matrices have only recently gained attention.

\\

In this talk we discuss results of our work on using weighted matchings in

the preconditioning of symmetric indefinite linear systems, following ideas

introduced by Duff and Gilbert. In order to maintain symmetry,

the weighted matching is symmetrized and the cycle structure of the

resulting matching is used to build reorderings that form small diagonal

blocks from the matched entries.

\\

For the preconditioning we investigated two approaches. One is an

incomplete $LDL^{T}$ preconditioning, that chooses 1x1 or 2x2 diagonal pivots

based on a simple tridiagonal pivoting criterion. The second approach

targets distributed computing, and is based on factorized sparse approximate

inverses, whose existence, in turn, is based on the existence of an $LDL^{T}$

factorization. Results for a number of comprehensive test sets are given,

including comparisons with sparse direct solvers and other preconditioning

approaches.