Fast Gradient-Based Iterative Regularization Methods for Nonlinear Ill-Posed Problems - Theory and Applications
Year: 2018
Type: dissertation
Abstract: In this thesis, we investigate a class of fast gradient-based iterative regularization methods for nonlinear Inverse and Ill-Posed Problems based on Landweber iteration and Nesterov's acceleration scheme.These so-called Two Point Gradient (TPG) methods have been found to be very useful in practical applications, since they are both easy to implement and lead to a great speedup compared to standard gradient-based methods.While methods utilizing second-order information, which are known for their fast convergence, often become infeasible when dealing with large datasets, gradient-based methods are usually more flexible and able to deal with large datasets, at the disadvantage of requiring a large number of iterations.TPG methods have the potential to bridge the apparent gap between these two classes of methods, being both fast, flexible, and able to deal with large datasets, which are important requirements for any iterative regularization method used for solving inverse problems.This thesis provides a convergence analysis of TPG methods under the common assumption of a tangential cone condition, and covers some well-known choices of commonly used stepsizes.Furthermore, a convergence analysis with the tangential cone condition being replaced by a local convexity assumption more natural to Nesterov's original acceleration idea is performed.These results provide the first successful convergence analysis for TPG methods for the solution of nonlinear ill-posed problems.First of all, I would like to thank my supervisor Prof. Ronny Ramlau, for everything he has done for me so far, be it the opportunity to work under his expert guidance, the interesting topics he suggested, the fruitful discussions we had, as well as the freedom he gave me to pursue my own ideas and his interest in where they eventually led.For all he did for me and allowed me to do, and especially for taking care of many things in the background of which I can only be partly aware, I am most grateful to him.Next, I would like to thank the supervisor of my master thesis, Prof. Andreas Neubauer, for all the discussions we had over the years, for his valuable support especially during the early stages of my PhD, and for continuing to answer all of my many questions.Furthermore, I would like to
Author: Hubmer, Simon Gerhard
Institution: Johannes Kepler University of Linz
Language: English
Subfield: Mathematical Physics
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6981802862
