(September 9) Sren Bartels:
Numerical Analysis of Some Non-Convex Variational Problems -
Non-convex variational problems arise in the mathematical
modeling of phase transitions in crystalline solids, in
particular in a mathematical description of shape memory
alloys. Non-convexity of the minimization problem typically
implies non-existence of classical solutions and enforces
numerical schemes to develop oscillations on mesh dependent
scales. The first part of the talk is based on joint work
with A. Prohl (ETH Zurich) and gives a precise characterization
of those oscillations. A numerical analysis of a well posed
extended (relaxed) formulation of the original problem in the
scalar case is presented in the second part of the talk. This
part states a posteriori error estimates for the adaptive
approximation of measure valued solutions and discusses
efficient solution strategies. The third part of the talk
focuses on aspects of the effective simulation of non-convex
vectorial variational problems and includes approximation
results for upper and lower bounds of quasiconvex envelopes. (September 16) Marcus Calhoun-Lopez:
Numerical Solutions of Hyperbolic Conservation Laws:
Incorporating Multi-Resolution Viscosity Methods into the
Finite Element Framework -
It is well known that the classic Galerkin finite-element method
is unstable when applied to hyperbolic conservation laws, such as
the Euler equations for compressible flow. Adding a diffusion
term to the equations stabilizes the method but sacrifices too
much accuracy to be of any practical use. An elegant solution
devised by Eitan Tadmor for spectral methods is to add diffusion
only to the high frequency modes of the solution, which
stabilizes the method without the sacrifice of accuracy. We
incorporate this idea into the finite-element framework by using
hierarchical functions as a multi-frequency basis. The result is
a new finite element method for solving hyperbolic conservation
laws. For this method, we are able to prove convergence for a
one-dimensional scalar conservation law. Numerical results are
presented for one- and two-dimensional hyperbolic conservation
laws. (September 23) John Zweck:
Euclidean invariant computation of salient closed contours in images
-
Vision is by far the most highly developed sense in humans. Much of our
knowledge of the physical world is based on visual information,
and a large portion of the brain is devoted to visual information
processing. One of the basic tasks of human and computer
vision systems is to complete the boundaries of partially occluded
objects. In this talk I will discuss a biologically plausible
algorithm to complete salient closed contours in images. Our work is different from previous work in three ways.
First, as in primary visual cortex the input image consists of spots not
edges.
Second, our computation is based on a distribution of closed contours
characterized by a random process. Third, although it is implemented
in a discrete neural network our computation is Euclidean invariant
in the continuum. From a numerical point of view, our algorithm was based on
a novel numerical solution of a Fokker-Planck partial
differential equation on the space of positions and directions in
the plane, together with some numerical linear algebra, and a new twist
on the Shannon sampling theorem. This work is joint with Lance Williams.
(September 30) Mark Ainsworth:
Adaptive hp-Finite Element Methods for Maxwell's Equations -
Recently, there has been a dramatic increase in the use of high order
finite element methods for the approximation of Maxwell's equations.
We shall discuss some of our own work in this area. In particular, we
shall present families of hierarchic basis functions for the Galerkin
discretisation of the space $H({\rm curl};\Omega)$ that naturally arises
in the variational formulation of Maxwell equations. The conditioning
and dispersive behaviour of the elements is discussed along with
approximation theory. Numerical examples are shown which demonstrate
the accuracy and efficiency of the methods for computing solutions of
the time-harmonic Maxwell's equations.
(October 7) Richard Falk:
Finite Element Approximation Theory Using Families of Reference Elements -
A standard method for the derivation of error estimates for
the approximation
of smooth functions by families of finite elements indexed by a mesh-size
parameter h is to use a scaling argument based upon the introduction of a
single reference element together with the Bramble-Hilbert lemma. There
are a
number of situations, however, when this method does not apply or depends on
some additional special estimates to make it applicable. In this talk, we
consider a modification of this approach, using a family of reference
elements, that has wider applicability. Several applications of this
modified
approach are discussed.
(October 14) Christoph Schwab:
Finite Elements for Elliptic Problems with Stochastic Data -
We consider the Finite Element Solution of elliptic problems
with random coefficients which are spatially inhomogeneous
with finite second moments. We express these
as Wiener Polynomial Chaos (PC) expansion of Karhunen Loeve (KL) type
in infinitely many random variables. We give a log-linear complexity algorithm to compute
KL expansions for general two-point correlation functions of the data.
New decay estimates for KL Eigenvalues and for pointwise convergence of
KL expansions are presented in terms of the smoothness of the two-point
correlation functions of the data. Truncation of the KL expansion after M terms leads to a deterministic
elliptic problem in M stochastic variables and d physical dimensions.
Analytic dependence of the random solution on the M stochastic variables
is proved. Precise estimates of domains of analyticity show how that
high complexity due to large M can be broken. A polynomial basis is given for which
the O(p^M) deterministic problems can be solved independently and
in parallel, reducing the computational complexity to that of a MC
method, but with exponential convergence in the number of samples,
thereby allowing to handle also random fields with slowly convergent
KL expansions. A-priori error analysis of the fully discrete scheme is provided.
Numerical experiments in physical dimension 2 and 3 with
stochastic dimension M up to 30 taking minutes of CPU on the
ETH BEOWOLF CLUSTER ASGARD are presented. This is joint work with R.A. Todor and P. Frauenfelder at ETHZ.
(November 04) Raul Tempone:
Adaptive Weak approximation of Stochastic Differential Equations with Jumps -
We develop adaptive time stepping methods, based on the Monte Carlo Euler
method, for weak approximation of jump diffusion driven stochastic
differential equations.
The main result is new expansions of the computational error, with
computable leading order term in a posteriori form, based on stochastic
flows and discrete dual backward problems.
The expansions lead to efficient and accurate computation of error estimates.
We describe adaptive algorithms for either stochastic time steps or
quasi-deterministic time steps and present numerical examples illustrating
their behavior.
(November 13) Tom Hou:
Efficient Numerical Computations of Stochastic Partial
Differential Equations -
Stochastic PDE's with solutions depending on multiple scales play
fundamental and important roles in many problems. Numerical simulations
become an important strategy in practical applications. Monte Carlo
methods are used widely, but convergence is usually slow and
accuracy is poor. Here we propose a strategy, which uses Wiener Chaos
expansions, to design efficient numerical methods for stochastic PDE's.
One of the main advantages of this approach is that it separates
randomness from the problems. We can calculate all statistical
properties of the solutions by solving the induced deterministic
coefficient equations. This approach does not use random number
generating in the computations. Thus, well developed numerical
techniques can be used to solve the coefficient equations. We apply
this approach to solve the randomly forced Bousinesq equations and
Burgers' equation. We demonstrate that our approach is more efficient
than the traditional Monte-Carlo method in achieving a comparable
accuracy. Moreover, the method gives better control over the
computational errors. Finally, we have performed careful long time
computations for the randomly forced Bousinesq equations to study the
mixing property of the unstably stratefied fluid flow. (November 14) Tom Hou:
Multiscale Modeling and Computation of Flow in
Heterogeneous Media -
Many problems of fundamental and practical importance contain multiple scale
solutions. Composite materials, flow and transport in porous media, and
turbulent flow are examples of this type. Direct numerical simulations of these
multiscale problems are extremely difficult due to the range of length scales in
the underlying physical problems. In this talk, I will give an overview of the
multiscale finite element method and describe some of its applications,
including composite materials, wave propagation in random media, convection
enhanced diffusion, flow and transport in heterogeneous porous media. It is
important to point out that the multiscale finite element method is designed for
problems with many or continuous spectrum of scales without scale
separation. Further, we introduce a new multiscale analysis for convection
dominated 3-D incompressible flow with multiscale solutions. The main idea is
to construct semi-analytic multiscale solutions locally in space and time, and
use them to construct the coarse grid approximation to the global multiscale
solution. Our multiscale analysis provides an important guideline in designing a
systematic multiscale method for computing incompressible flow with multiscale
solutions. (November 18) Mario Ohlberger:
Error estimates for finite volume approximations of
non-linear conservation laws on bounded domains -
In 1979 Bardos, Leroux and Nedelec gave a characterization of
solutions of non-linear conservation laws on bounded domains
in the class BV using the vanishing viscosity method.
In 1992 Otto gave a uniqueness proof for entropy solutions on bounded
domains in the Linfinity class. Based on the Otto formulation we
derive an a posteriori error bound for finite volume approximations
which is at least of order h1/6 for meshes with uniform mesh size h.
(joint work with Julien Vovelle) (November 25) Madhu Nayakkankuppam:
Solving Large-Scale Semidefinite Programs in Parallel -
Semidefinite programming (SDP), which may be
described as linear programming in the space of symmetric
matrices, has been one of the most intensively researched
topics in optimization in the past decade. In this talk,
we first survey some of the many applications of SDP,
which span numerical linear algebra, combinatorial
optimization, control theory, structural design, and
statistics, just to name a few. Then we discuss
subgradient bundle methods, and how they may be effectively
parallelized to solve large-scale problems. We present
significant computational results on some of the largest
SDP's solved to date.
(December 2) Junping Wang:
Computational Simulation of Surface and Subsurface Fluid Flow -
Simulation of surface and subsurface flow has significant impact on the
economic and urban development of the society. An ideal model for the
simulation is expected to handle fluid bodies and simulating regions of
arbitrary shape. The modeling system usually combines an overland flow model,
a channel flow model, an infiltration model, a sewer flow model
(in urban areas), and a model linking surface flow and sewer flow. Most of the existing computational models are based on approximate or greatly
simplified forms of the Saint Venant equations and the subsurface flow
equation obtained from the Darcy's law. Our study offers a systematic
understanding of the coupling between the surface and subsurface flows.
This talk will introduce the modeling equations and their numerical
approximations associated with the coupled system of surface and subsurface
flows. A modified version of the Saint Venant equation will be presented
and approximated by finite element and finite volume methods combined with the
method of characteristics. A new discretization scheme for the Stokes equation
will be discussed if time permits. (December 9) Eric de Sturler:
Preconditioning Generalized Saddle-Point Problems Using
Cheap Approximations to the Schur Complement -
Recently, for saddle point type problems we have analyzed block
diagonal preconditioners derived from the work by Murphy, Golub, and
Wathen [SISC 21, 2000], and we have derived constraint preconditioners
from these. This gives interesting links between the two types of
preconditioners. For the latter we proved better clustering ofthe
eigenvalues, and we demonstrated for one class of optimization problems that
this translates in significantly faster convergence. However, in these
results we used the exact inverse of a 'Schur complement' type matrix
derived from the splitting that defines our preconditioner. Although this
worked well for the application discussed, it may be expensive for other
problems. Moreover, we did not address systems arising in stabilized finite
element discretizations that have a nonzero (2,2) block. We will address
those issues in this presentation. This is joint work with Joerg Liesen
(TU Berlin) and Chris Siefert (UIUC).
|