Title: Fast algorithms for data analysis and elliptic partial differential equations
Speaker: Sivaram Ambikasaran (NYU)
Date: 24 June 2015
Large-dense matrices arise in numerous applications: boundary integral formulation for elliptic partial differential equations, covariance matrices in statistics, inverse problems, radial basis function interpolation, multi frontal solvers for sparse linear systems, etc. As the problem size increases, large memory requirements, scaling as O(N^2), and extensive computational time to perform matrix algebra, scaling as O(N^2) or O(N^3), make computations impractical. I will discuss some novel methods for handling these computationally intense problems. In the first half of the talk, I will discuss my contributions to some of the new developments in handling large dense covariance matrices in the context of computational statistics and Bayesian data assimilation. More specifically, I will be discussing how fast dense linear algebra (O(N) algorithms for inversion, determinant computation, symmetric factorisation, etc.) enables us to handle large scale Gaussian processes, thereby providing an attractive approach for big data applications. In the second half of the talk, I will focus on a new algorithm termed Inverse Fast Multipole Method, which permits solving singular integral equations arising out of elliptic PDE’s at a computational cost of O(N).