Hanke-Raus heuristic rule for variational regularization in Banach spaces (Q2820677)

The author analyzes a heuristic parameter choice rule for linear/nonlinear inverse problems in Banach spaces. Consider inverse problems of the form NEWLINE\[NEWLINEF(x) = y,NEWLINE\]NEWLINE where \(F: \mathcal{D}(F) \subset X \to Y\) is an operator between two Banach spaces \(X\) and \(Y\) with its domain \(\mathcal{D}(F)\). Given noisy data \(\tilde y\), one seeks an approximation to the true solution \(x^\dag\) by variational regularization (with \(1<r<\infty\)): NEWLINE\[NEWLINE \tilde x_\alpha \in \arg \min_{x\in\mathcal{D}(F)} \{T_\alpha(x) := \|F(x) -\tilde y\|^r + \alpha \mathcal{R}(x)\}. NEWLINE\]NEWLINE The quality of the approximation \(\tilde x_\alpha\) depends crucially on the regularization parameter \(\alpha>0\). Many rules have been proposed, especially in the Hilbert space setting. The author focuses on the Hanke-Raus rule, which was first proposed by \textit{M. Hanke} and \textit{T. Raus} [SIAM J. Sci. Comput. 17, No. 4, 956--972 (1996; Zbl 0859.65051)] in a Hilbert space setting, and then analyzed by \textit{B. Jin} and \textit{D. A. Lorenz} [SIAM J. Numer. Anal. 48, No. 3, 1208--1229 (2010; Zbl 1215.65100)] for a Banach space setting where \(Y\) remains a Hilbert space. In the present work, the author provides a complete analysis of the following discrete variant in a general setting: Let \(\alpha_0>0\) and \(0<q< 1\) be given numbers and set \(\Delta_q = \{\alpha_0q^j: j = 0,1,\ldots \}\). Then we define \(\alpha_* :=\alpha_* (\tilde y) \in \Delta_q\) such that NEWLINE\[NEWLINE \alpha_* = \arg \min_{\alpha\in \Delta_q} \frac{ \|F(\tilde x_\alpha ) -\tilde y\|^r}{\alpha}. NEWLINE\]NEWLINE The author obtains a posteriori error estimates in terms of Bregman distance, under source conditions formulated as variational inequalities. Further, by imposing certain conditions on the random noise, four convergence results are given: one relies on the source conditions, and the other three do not depend on any source conditions. Numerical results are presented to illustrate the performance.