4:00 pm Monday, April 17, 2017
Colloquium : Analysis based methods for solving linear elliptic PDEs numerically by Gunnar Martinsson (University of Colorado at Boulder and Oxford University) in RLM 5.104
For some of the simplest linear elliptic PDEs, explicit solution formulas are known. Consider for instance, the Poisson equation Lu=f, where L is the Laplace operator. When the domain is free space R^d, and appropriate decay conditions are enforced at infinity, the unique solution is given as a convolution between the load function f and the free space fundamental solution of the Laplace operator. An explicit solution formula such as this can be turned into a powerful numerical tool once it is coupled with an efficient algorithm for rapidly evaluating the integral operator, such as, e.g., the Fast Multipole Method (FMM). The talk describes a set of generalizations to this "FMM based approach", where we numerically build approximations to solution operators for a broad class of elliptic PDEs, such as boundary value problems involving "general" domains, and/or differential operators with variable coefficients. These new so called "fast direct solvers" have many advantages, including improved stability and robustness, and dramatic improvements in speed in certain environments. The talk will also briefly describe methods based on randomized projections for rapidly executing common linear algebraic computations such as, e.g, low-rank approximation, factorization of matrices, solving linear systems, etc. These randomized methods are used to accelerate the direct solvers for elliptic PDEs, but have also proven highly competitive in machine learning, data analysis, etc.
Submitted by