Levenberg-Marquardt method in Banach spaces with general convex regularization terms (Q303637): Difference between revisions
From MaRDI portal
Revision as of 10:44, 12 July 2024
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Levenberg-Marquardt method in Banach spaces with general convex regularization terms |
scientific article |
Statements
Levenberg-Marquardt method in Banach spaces with general convex regularization terms (English)
0 references
22 August 2016
0 references
The article deals with the equation \(F(x) = y\) with the operator \(F\) from a reflexive Banach space \({\mathcal X}\) into a uniformly smooth Banach space \({\mathcal Y}\). It is assumed that a convex function \(\Theta:\;{\mathcal X} \to (-\infty,+\infty]\) and \[ \partial \Theta(x)\;:= \{\xi \in {\mathcal X}^*:\;\Theta(\bar{x}) - \Theta(x) - \langle \xi,\bar{x} - x \rangle \geq 0 \;\text{for all} \;\bar{x} \in {\mathcal X}\}. \] About the operator \(F\) it is asumed that there exists a function \(L:\;{\mathcal X} \to {\mathcal L}({\mathcal X},{\mathcal Y})\) such that for some \(\eta \in [0,1)\) \[ \|F(\bar{x}) - F(x) - L(x)(\bar{x}- x)\| \leq \eta\|\bar{x} - x\|. \] The authors study the following Levenberg-Marquardt algorithm of approximate solving the equation \(F(x) = Y\): {\parindent=0.6cm\begin{itemize}\item[{\(\bullet\)}] Take \(x_0 \in {\mathcal X}\) and \(\xi_0 \in {\mathcal X}^*\) such that \(\xi_0 \in \partial \Theta(x_0)\). Pick \(0< \mu \leq \mu_1<1\) and \(\tau>1\). \item[{\(\bullet\)}] Let \(x_0^\delta\;:= x_0\) and \(\xi_0^\delta\;:= \xi_0\). Assume that \(x_n^\delta\) and \(\xi_n^\delta\) are well-defined, we then define \(x_{n+1}^\delta\) and \(\xi_{n+1}^\delta\) as follows: (a) For each \(\alpha > 0\) we define \(x_n(\alpha,y^\delta)\) and \(xi_n(\alpha,y^\delta)\) as \[ x_n(\alpha,y^\delta) = \arg \min_{x \in {\mathcal X}} \bigg\{\frac1r \, \|y^\delta - F(x_n^\delta) - L(x_n^\delta)(x - x_n^\delta)\|^r + \alpha D_{\xi_n^\delta} \Theta(x,x_n^\delta)\bigg\}, \] \[ \xi_n(\alpha,y^\delta) = \xi_n^\delta + \frac1\alpha \, L(x_n^\delta)^* J_r^{\mathcal Y} (y^\delta - F(x_n^\delta) - L(x_n^\delta)(x_n(\alpha,y^\delta) - x_n^\delta)); \] (b) Take \(\alpha_n(y^\delta) > 0\) to be a number \(\alpha\) such that \[ \mu_0\|y^\delta - F(x_n^\delta)\| \leq \|y^\delta - F(x_n^\delta) - L(x_n^\delta)(x_n(\alpha,y^\delta) - x_n^\delta)\| \leq \mu_1\|y^\delta - F(x_n^\delta)\|; \] (c) Define \(x_{n+1}^\delta\;:= x_n(\alpha_n(y^\delta),y^\delta)\) and \(\xi_{n+1}^\delta\;:= \xi_n(\alpha_n(\delta),y^\delta)\), \item[{\(\bullet\)}] Let \(n_\delta\) be the first integer such that \(\|y^\delta - F(x_{n_\delta}^\delta)\| \leq \tau\delta\) and \(x_{n_\delta}^\delta\) an approximate solution. \end{itemize}} The main results are a theorem about the convergence of these approximations in the usual for ill-posed problems sense under natural assumptions about the operator \(F\). At the end of the article numerical examples are presented.
0 references
Levenberg-Marquardt method with exact and noisy data
0 references
nonlinear inverse probl
0 references
numerical example
0 references
0 references
0 references
0 references
0 references
0 references
0 references
0 references
0 references
0 references
0 references