Multi-parameter Tikhonov regularization -- an augmented approach (Q741463)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Multi-parameter Tikhonov regularization -- an augmented approach |
scientific article |
Statements
Multi-parameter Tikhonov regularization -- an augmented approach (English)
0 references
12 September 2014
0 references
The authors consider multiparameter regularization techniques for solving inverse problems \( Ku^\dagger = g^\dagger \), where \( K: X \to Y \) is a bounded linear operator between Banach spaces \( X \) and \( Y \), and \( g^\dagger \in Y \) denotes the exact data, and \( u^\dagger \in X \) is the unknown exact solution. The considered techniques are based on the minimization of the functional \(J_{\boldsymbol{\eta}}(u) = \phi(u,g^\delta) + \boldsymbol{\eta} \cdot \boldsymbol{\psi}(u) \) over a closed convex set \( \mathcal{C} \subset X \) of feasible solutions, where \( \phi(u,g^\delta) = \tfrac{1}{2}\| K u-g^\delta \|^2 \), and \( g^\delta \in Y \) denote some noisy data. In addition, \( \boldsymbol{\psi}(u) = (\psi_1(u),\psi_2(u))^{t} \in \mathbb{R}_+^2 \) is a vector-valued penalty term, and \( \boldsymbol{\eta} = (\eta_1,\eta_2)^{t} \in \mathbb{R}_+^2 \) is a regularization parameter vector. The first considered parameter choice strategy is a noise-level free balancing principle. It can be written as a minimization problem for the functional \( \Phi_\gamma(\boldsymbol{\eta}) = F(\boldsymbol{\eta})^{\gamma+2}/(\eta_1\eta_2) \), where \( F(\boldsymbol{\eta}) = \inf_{u \in \mathcal{C}} J_{\boldsymbol{\eta}}(u) \), and \( \gamma > 0 \) denotes some parameter. Error estimates are provided for this strategy, with the Bregman distance used when source conditions are available. The estimates are given in terms of \( \max\{\delta, \delta_*\} \), where \( \delta = \| g^\delta - g^\dagger \| \) denotes the noise level, and \( \delta_* = \| K u_{\boldsymbol{\eta}^*}^\delta-g^\delta \| \) is the residual realized by the parameter choice. This balancing principle is considered also in Hilbert spaces, with the special case \( \psi_i(u) = \tfrac{1}{2}\| L_i u\|^2 \), where \( L_i \) denotes a linear operator. Error estimates with respect to a weighted semi-norm defined by \( \| u\|_t^2 = t \| L_1 u\|^2 + (1-t) \| L_2 u\|^2 \) are provided in this case. The second considered parameter choice strategy is the following hybrid balanced discrepancy principle: \( \phi(u_{\boldsymbol{\eta}}^\delta,g^\delta) = \tfrac{1}{2}c_m^2 \delta^2\), \(\eta_1 \psi_1(u_{\boldsymbol{\eta}}^\delta) = \eta_2 \psi_2(u_{\boldsymbol{\eta}}^\delta) \), where \( c_m \geq 1 \) denotes some constant. Error estimates are provided also for this strategy, with the Bregman distance again used when source conditions are available. Finally, results of some numerical experiments are presented.
0 references
ill-posed problem
0 references
inverse problem
0 references
multiparameter regularization
0 references
augmented Tikhonov regularization
0 references
Hilbert space
0 references
Banach space
0 references
convergence rate
0 references
a posteriori parameter choice
0 references
balanced discrepancy principle
0 references
balancing principle
0 references
Bregman distance
0 references
source condition
0 references
error estimate
0 references
numerical experiment
0 references
0 references
0 references
0 references
0 references
0 references