Adaptive discretizations for the choice of a Tikhonov regularization parameter in nonlinear inverse problems

Barbara Kaltenbacher, Alana Kirchner, Boris Vexler

Research output: Contribution to journalArticlepeer-review

31 Scopus citations

Abstract

Parameter identification problems for partial differential equations usually lead to nonlinear inverse problems. A typical property of such problems is their instability, which requires regularization techniques, like, e.g., Tikhonov regularization. The main focus of this paper will be on efficient methods for determining a suitable regularization parameter by using adaptive finite element discretizations based on goal-oriented error estimators. A well-established method for the determination of a regularization parameter is the discrepancy principle where the residual norm, considered as a function i of the regularization parameter, should equal an appropriate multiple of the noise level. We suggest to solve the resulting scalar nonlinear equation by an inexact Newton method, where in each iteration step, a regularized problem is solved at a different discretization level. The proposed algorithm is an extension of the method suggested in Griesbaum A et al (2008 Inverse Problems 24 025025) for linear inverse problems, where goal-oriented error estimators for i and its derivative are used for adaptive refinement strategies in order to keep the discretization level as coarse as possible to save computational effort but fine enough to guarantee global convergence of the inexact Newton method. This concept leads to a highly efficient method for determining the Tikhonov regularization parameter for nonlinear ill-posed problems. Moreover, we prove that with the so-obtained regularization parameter and an also adaptively discretized Tikhonov minimizer, usual convergence and regularization results from the continuous setting can be recovered. As a matter of fact, it is shown that it suffices to use stationary points of the Tikhonov functional. The efficiency of the proposed method is demonstrated by means of numerical experiments.

Original languageEnglish
Article number125008
JournalInverse Problems
Volume27
Issue number12
DOIs
StatePublished - Dec 2011

Fingerprint

Dive into the research topics of 'Adaptive discretizations for the choice of a Tikhonov regularization parameter in nonlinear inverse problems'. Together they form a unique fingerprint.

Cite this