Georgia Institute of Technology College of Sciences

The domain decomposition methods considered are preconditioned conjugate gradient methods designed for the very large algebraic systems of equations which often arise in finite element practice. They are designed for massively parallel computer systems and the preconditioners are built from solvers on the substructures into whichthe domain of the given problem is partitioned. In addition, to obtain scalability, there must be a coarse problem, with a small number of degrees of freedom for each substructure. The design of this coarse problem is crucial for obtaining rapidly convergent iterations and poses the most interesting challenge in the analysis.Our work will be illustrated by overlapping Schwarz methods for almost incompressible elasticity approximated by mixed finite element and mixed spectral element methods. These algorithms is now used extensively at the SANDIA, Albuquerque laboratories and were developed in close collaboration with Dr. Clark R. Dohrmann. These results illustrate two roles of the coarse component of the preconditioner.Currently, these algorithms are being actively developed for problems posed in H(curl) and H(div). This work requires the development of new coarse spaces. We will also comment on recent work on extending domain decomposition theory to subdomains with quite irregular boundaries.  This work is relevant because of the use of mesh partitioners in the decomposition of large finite element matrices. 

Many scientific, engineering and financial applications require solving high-dimensional PDEs. However, traditional tensor product based algorithms suffer from the so called "curse of dimensionality".We shall construct a new sparse spectral method for high-dimensional problems, and present, in particular,  rigorous error estimates as well as efficient numerical algorithms for  elliptic equations in both bounded and unbounded domains.As an application, we shall use the proposed sparse spectral method to solve the N-particle electronic  Schrodinger equation.

We consider a variant of Rudin--Osher--Fatemi variational image smoothing that replaces the BV semi-norm in the penalty term with the B^1_\infty(L_1) Besov space semi-norm. The space B^1_\infty(L_1$ differs from BV in a number of ways: It is somewhat larger than BV, so functions inB^1_\infty(L_1) can exhibit more general singularities than exhibited by functions in BV, and, in contrast to BV, affine functions are assigned no penalty in B^1_\infty(L_1). We provide a discrete model that uses a result of Ditzian and Ivanov to compute reliably with moduli of smoothness; we also incorporate some ``geometrical'' considerations into this model. We then present a convergent iterative method for solving the discrete variational problem. The resulting algorithms are multiscale, in that as the amount of smoothing increases, the results are computed using differences over increasingly large pixel distances. Some computational results will be presented. This is joint work with Greg Buzzard, Antonin Chambolle, and Stacey Levine.

 We will discuss the recent developments of fast image reconstrcution with total variation (TV) regularization whose robustness has been justfied by the theory of compressed sensing. However, the solution of TV based reconstruction encounters two main difficulties on the computational aspect of many applications: the inversion matrix can be large, irregular, and severely ill-conditioned, and the objective is nonsmooth. We introduce two algorithms that tackle the problem using variable splitting and optimized step size selection. The algorithms also provide a general framework for solving large and ill-conditioned linear inversion problem with TV regularization. An important and successful application of TV based image reconstruction in magnetic resonance imaging (MRI) known as paratially parallel imaging (PPI) will be discussed. The numerical results demonstrate significantly improved  efficiency and accuracy over the state-of-the-arts.