A0206
Title: Integral operator approach for spherical data fitting
Authors: Shao-Bo Lin - Xi'an Jiaotong University (China) [presenting]
Abstract: For kernel interpolation of scattered data on spheres, it is well known that the attainable approximation error and the condition number of the underlying interpolation matrix cannot be made small simultaneously, referred to as the uncertainty phenomenon. An undesirable consequence is that kernel interpolation is susceptible to noisy data. The aim is to develop a novel integral operator approach for deterministic sampling and propose several remedies for the "uncertainty phenomenon". Based on the integral operator approach, it is proven that the popular spectral regularization, distributed learning and random sketching are feasible methods to circumvent the "uncertainty phenomenon". Numerical simulation results are also presented, showing that the mitigation methods are practical and robust in handling noisy data from challenging computing environments.
