(February 3) Dr. Kyoung-Sook Moon:
Adaptive Monte Carlo Algorithm for Stochastic Differential Equations
-
I will present an adaptive Monte Carlo algorithm for weak approximations
of stochastic differential equations. The goal is to compute an expected
value of a given function depending on a stochastic process.
Based on an a posteriori error expansion, the adaptive algorithm is proven
to stop with asymptotically optimal number of final time steps and the
approximation error is asymptotically bounded by the specified error
tolerance as the tolerance parameter tends to zero.
Finally, I will show numerical results from computations of barrier
options in financial mathematics. These results show that the adaptive
algorithm achieve the time discretization error of order N^{-1} with N
adaptive time steps, while the error is of order N^{-1/2} for a method
with N uniform time steps.
(February 10) Dr. Francisco Pena:
Some Contributions to the Modeling of Ferro-alloys Production -
Our motivation is to model the production of ferro-alloys in
reduction furnaces. The main part are the electrodes. Heat is
generated in them by the Joule and by the electric arc at the bottom. A sui
table model is the system of equations composed by the Maxwell
equation in low frequency regime (eddy currents) coupled with the
transient heat equation with phase change. This model is related to
the so called ``thermistor'' problems. The main difficulties are the
belonging of the Joule effect to the non-reflexive space $L^1$
and the Stefan problem in the heat equation. We will show a result
about
existence of weak solution for a simpler case, when enthalpy is uni-valued.
(February 17) Dr. R. Bruce Kellogg:
Corner Singularities in the Two Dimensional Compressible
Navier-Stokes System
-
The compressible Navier-Stokes system, and linearized versions
of this system, are not elliptic. When a steady state boundary
value problem for this system is posed on a polygonal domain,
one may ask for the behavior of the solution near a vertex
of the polygon. This talk will discuss some results of
J. R. Kweon and myself, as well as some unresolved questions,
on this problem.
(February 24) Dr. Thomas Russell:
Oh no, not the wiggles again! A revisit of an old problem and a new approach. -
Lumping is often used to control oscillations in schemes that use the method of weighted residuals. Standard lumping procedures add diffusion indiscriminately, resulting in excessive numerical diffusion in solutions. Here it is shown that the mass matrix can be selectively lumped (SLUMPED), with an optimal amount of diffusion added to each row of the mass matrix. The amount of diffusion added is calculated from the right-hand-side vector. The optimal amount of diffusion is found in 4 steps. First, the monotonicity problem is recast in the form of a maximum principle. Second, for a 2x2 element matrix, the amount of diffusion is calculated for an arbitrary right-hand side so that the solution obeys a maximum principle. Third, the result is generalized for larger matrices. Finally, the result is recast to meet the monotonicity requirement. The result is an equation giving the amount of diffusion to be added in terms of a given right-hand-side vector. Selective lumping is shown to be effective for both an Eulerian-Lagrangian Localized Adjoint Method (ELLAM) solution of the transport equation and a finite element solution of the heat equation. In both cases, solutions were monotonic and contained less numerical diffusion than in standard lumping schemes. The SLUMPING method is general and can be applied to any numerical approximation based on the method of weighted residuals. (joint work with Philip J. Binning) (March 2) Dr. Bing Song:
A Fast Algorithm for Variational Level Set Segmentation -
We develop a fast method to solve Variational Level Set Segmentation. This
direct method improves the computational speed dramatically. It differs from
previous methods in that we do not need to solve the Euler-Lagrange equation
of the underlying variational problem. Instead, we calculate the energy
directly and check if the energy is decreased when we change a point inside
the level set to outside or vice versa. We analyze the algorithm and prove
that under most initial conditions, we only need one sweep over the pixels
to converge to the correct solution for 2-phase images. Another advantage of
this method is that the gradient of the functional is not needed. This
enables it to be applied to broader range of optimization problems.
(March 4) Vanessa Lopez:
Periodic Solutions of Chaotic Partial Differential Equations -
We consider the problem of finding relative time-periodic solutions of
chaotic partial differential equations with symmetries. Relative
time-periodic solutions are solutions that are periodic in time, up to a
transformation by an element of the equations' symmetry group. As a model
problem we work with the 1D complex Ginzburg-Landau equation (CGLE), which
is a standard example of an evolution equation that exhibits chaotic
behavior. The problem of finding relative time-periodic solutions
numerically is reduced to one of finding solutions to a system of
nonlinear algebraic equations, obtained after applying a spectral-Galerkin
discretization in space and time to the CGLE. The discretization is
designed to include as an unknown the group element that defines a
relative time-periodic solution. Using this approach, we found a large
collection of distinct relative time-periodic solutions in a chaotic
region of the CGLE. These solutions, all of which have broad temporal and
spatial spectra, were previously unknown. There is a great deal of
variety in their Lyapunov spectra and spatio-temporal profiles.
Moreover, none bear resemblance to the time-periodic solutions of the CGLE
studied previously. We also briefly discuss the problem of finding
relative time-periodic solutions to the Navier-Stokes equations and
present some preliminary work done towards its solution. (March 9) Dr. Andreas Prohl:
Analysis and Numerics of Total Variation Flow
and Regularized Mumford-Shah Flow -
Automatic denoising and segmentation are principle tasks in image
processing. Mathematically, the goal is to detect/preserve essential
geometric structures (edges, domains,...) during the filtering process to
extract relevant image information. The first part of the talk addresses
analysis of the total variation flow for initial data $u_0 \in
L^2(\Omega)$ by relating it to the prescribed mean curvature flow; then, a
numerical analysis for a fully discrete realization is presented which
leads to proper scaling laws to balance regularization and discretization
effects. The second part of the talk addresses analysis and numerics of
the gradient flow for a $\Gamma$-convergent approximation by Ambrosio and
Tortorelli of the Mumford-Shah functional. I present a numerical scheme
which guarantees convergence of the whole sequence $\{ (\, u_h,
\varphi_h\, )\}_h$ towards weak solutions of the limiting problem; in
particular, this analysis is based on explicit characterization of a
(possible) singularity set. Computational experiments for both scenarios
are presented to illustrate solution's behavior of the discussed problems.
--- This is joint work with X.~Feng (Univ. of Tennessee). (March 16) Dr. Haesun Park:
Dimension Reduction for
Undersampled Problems in Pattern Analysis -
Dimension reduction is imperative for efficient handling of
data represented in a very high dimensional space.
Linear Discriminant Analysis (LDA) has been successfully applied to
dimension reduction problems for many decades.
However, in undersampled problems where the number of data items is smaller
than
the dimension of the data space, it is difficult to apply LDA
due to the potential singularity of the scatter matrices caused by high
dimensionality. In this talk, we examine an optimization criterion that mathematically
models many commonly used dimension reduction methods such as LDA,
Principal Component Analysis (PCA), and Latent Semantic Indexing (LSI).
Then we present efficient new dimension reduction algorithms such as
generalized
LDA based on the generalized singular value decomposition (LDA/GSVD) and
the Orthogonal Centroid Method (OCM), which are applicable regardless of the
relative
dimensions of the data matrices.
By utilizing the GSVD, we establish a relationship between support vector
machines (SVMs) and generalized linear discriminant analysis.
The LDA/GSVD and OCM can easily be extended to nonlinear dimension reduction
methods
by using kernel functions. Also, when LDA/GSVD is reformulated as a
a Mean Squared Error (MSE) problem,
a highly efficient incremental dimension reduction method can be derived.
Our substantial experimental results in text classification,
facial recognition, and fingerprint classification demonstrate
the effectiveness of the proposed methods.
(March 30) Dr. John D. Anderson:
History of the Evolution of the Navier-Stokes Equations And Computational Fluid Dynamics: A Marriage Made in Heaven? -
The Navier-Stokes equations are a result of intellectual thought that begins with early Greek science and mathematics. A brief history of the intellectual evolution of the Navier-Stokes equations is discussed. By the middle of the 19th century, the modern form of these equations was well in hand, but their general solution was not. To the present, no general analytical solution of the Navier-Stokes equations exist. The dust of ages was shaken off these equations, however, when practical computational fluid dynamics (CFD) evolved rather quickly in the middle of the 20th century, and allowed the numerical solution of the Navier-Stokes equations for a host of practical applications. The history of the evolution of CFD is discussed, and its impact on fluid dynamics in general, and the solution of the Navier-Stokes equations in particular, is highlighted. (April 6) Dr. Peter Monk:
A discontinuous Galerkin method for linear symmetric hyperbolic systems
in inhomogeneous media -
The Discontinuous Galerkin (DG) method provides a powerful tool for
approximating hyperbolic problems. We shall present a new space-time DG
method for linear time dependent hyperbolic problems written as a
symmetric system (including the wave equation and
Maxwell's equations). The main features of the scheme are that it can
handle inhomogeneous media, and can be time-stepped by solving a
sequence of small linear systems resulting from applying the method on
small collections of space-time elements. We show that the method is
stable provided the space-time grid is appropriately constructed (this
corresponds to the usual time-step restriction for explicit methods, but
applied locally) and give an error analysis of the scheme. We shall
also provide some simple numerical tests of the algorithm applied to
the wave equation in two space dimensions (plus time).
(April 13) Dr. Georgios Akrivis:
Linearly implicit methods for nonlinear parabolic equations -
We discuss the discretization of nonlinear parabolic equations u'(t) + Au(t) = B(t,u(t)) in a Hilbert space setting, by combinations of rational implicit
and explicit multistep methods.
The resulting schemes are linearly implicit and include as
particular cases implicit--explicit multistep schemes as well as
the combination of implicit Runge--Kutta schemes and extrapolation.
An optimal condition for the stability constant is derived under
which the schemes are locally stable.
Optimal order error estimates are established. The results are extended to equations of the form Lu'(t) + Au(t) = B(t,u(t)) By modifying the backward differentiation formulae, we arrive at
linearly implicit schemes with improved stability properties. (April 20) Dr. Tobias von Petersdorff:
Fast numerical methods for parabolic problems in high dimensions -
Fast numerical methods for parabolic problems in high dimensions We consider a parabolic initial value problem in d space dimensions. One
application is the pricing of options for a basket of d assets where d can
be larger than 20. The initial data are assumed to be rough (H^epsilon),
therefore the solution has singular behavior at t=0. Standard space discretization with mesh size h and time discretization with
M time steps leads to a work of O(M h-d) which is impractical for d>3.
We propose a space discretization with wavelet based sparse grid spaces
requiring only O(h-1) degrees of freedom. For the time discretization we
use an hp discontinuous Galerkin method with exponential convergence.
The resulting method has a total work of O(h-1 log(h)c). We give a complete
error analysis of space discretisation, time discretisation and of the
iterative solution of the resulting linear systems. Numerical results
for dimension up to 25 confirm the theoretical estimates. This is joint work with C. Schwab (ETH Zurich).
(April 27) Dr. Marco Verani:
On the control of output linear functionals -
The purpose of this talk is to explore the question of accurate
approximation of output linear functionals of linear elliptic
PDEs, through the use of adjoint problems and duality arguments.
We present a possible paradigm for an adaptive algorithm designed
to produce specially tuned grids for the computation of the
functional value within a given tolerance. Such an algorithm
stems from an equivalent saddle point formulation of the initial
problem. In particular we make use of an adaptive Uzawa algorithm
formulated in infinite dimensions [Dahmen et al., Nochetto et
al.] (joint work with S. Micheletti and S. Perotto)
(May 6) Dr. Benoît Perthame:
Selection, mutation: adaptive dynamics -
Living systems are subject to constant evolution. The environment, which we consider as a
nutriment available for all the population, allows individuals to develop and thus to select
the `best adapted trait' in the population. On the other hand the population can develop with
small variance on the trait (mutations). We will give a mathematical model of such dynamics and
show that an asymptotic method allows us to describe the `best adapted trait' and eventually to
compute bifurcations which lead to the cohabitation of two different traits.
(May 7) Dr. Benoît Perthame:
Mathematical models for cell motion -
Several transport-diffusion systems arise as simple models in chemotaxis (motion
of bacterias or amebia interacting through a chemical signal) and in
angiogenesis (development of capillary blood vessels from an exhogeneous
chemoattractive signal by solid tumors). These systems describe the evolution
of a density (of cells or blood vessels) coupled with the evolution equation
for a chemical substance, through a nonlinear transport term depending on the
gradient of the chemoattracting substance. Such systems are successful in
recovering various qualitative behavior (chemotactic collapse, ring dynamics).
Endothelial (i.e. cells forming blood vessels) have a tendency to form different
patterns, initiating the vessels shape. Then hyperbolic models seem better
adapted to describe this kind of network formation. We will present these
models, their main mathematical properties (quantitative and qualitative),
numerical simulations and, for bacteria E. Coli, we will give a microscopic
picture based on a kinetic modelling of the interaction (nonlinear scattering
equation). We show that such models can have global solutions that converge in
finite time to the Keller-Segel model, as a scaling parameter vanishes. This
point of view has also the advantage of unifying all the models.
(May 11) Dr. Jian-Guo Liu:
Accurate, Stable and Efficient Navier-Stokes Solvers
Based on Explicit Treatment of the Pressure Term -
In this talk, I will present numerical schemes for the
incompressible Navier-Stokes
equations based on a primitive variable formulation in which the
incompressibility constraint has been replaced by a pressure
Poisson equation. The pressure is treated explicitly in time,
completely decoupling the computation of the momentum and
kinematics equations. The result is a class of extremely
efficient Navier-Stokes solvers. Full time accuracy
achieved for all flow variables. The key to the schemes is a
Neumann boundary condition for the pressure Poisson equation which
enforces the incompressibility condition for the velocity field.
Irrespective of explicit or implicit time discretization of the
viscous term in the momentum equation, the explicit time
discretization of the pressure term does not affect the time step
constraint. Indeed, we show some unconditional stability properties
of the this new formulation for the Stokes equation with explicit
treatment of the pressure term and first or second order implicit
treatment of the viscous term. Systematic numerical experiments
for the full Navier-Stokes equations indicate that a second order
implicit time discretization of the viscous term, with the
pressure and convective terms treated explicitly, is stable under
the standard CFL condition. Additionally, various numerical
examples are presented, including both implicit and explicit time
discretizations, using spectral and finite difference spatial
discretizations, demonstrating the accuracy, flexibility and
efficiency of this class of schemes. In particular, a Galerkin
formulation is presented requiring only C0 elements to implement.
|