Computational and Applied Math Proseminar

Wednesday, March 28, 2:00 p.m. GWC 604

W. Steven Rosenthal

Dept. Math. & Stats

Discrete Polynomial Least Squares Approximation for Image Reconstruction and Solving Partial Differential Equations

Abstract Fourier spectral methods accurately approximate smooth and periodic functions from equally-spaced data sets. However, the Gibbs phenomenon destroys the Fourier reconstruction of non-periodic and/or piecewise-smooth functions. High-order polynomial interpolations of smooth, non-periodic functions given on equally-spaced points provoke Runge effects near the boundaries. A natural alternative is to use a Chebyshev distribution of nodes; although, in applications to numerical partial differential equations (PDEs), stable time integration is relatively slow. Moreover, the approximation method is incompatible with imaging applications, as data is usually restricted to an equally-spaced grid. Stretching the distribution to be more equally-spaced can accelerate time integration, but with notable loss in accuracy.

We present a suitable alternative using a least-squares approach on equally-spaced points. The method employs a polynomial basis mutually orthogonal with respect to the Freud weight function to obtain a spectrally-accurate approximation of a smooth function. We show how to choose least-squares parameters to minimize the pointwise error in the majority of the domain [-1,1]. Consequently, the Runge effects are partially alleviated. This least-squares method forms the foundation for a multi-domain method to integrate PDEs which yields a less restrictive stability criterion and recovers exponential convergence in [-1,1].

For further information please contact: mittelmann@asu.edu