A family of rules for parameter choice in Tikhonov regularization of ill-posed problems with inexact noise level (Q2428129): Difference between revisions

The subject of this paper is a class of a posteriori parameter choice strategies for Tikhonov's method to regularize linear ill-posed problems in Hilbert spaces. The considered class includes well-known rules like the discrepancy principle, a modified discrepancy principle and the monotone error rule. Under- and overestimated noise levels are also covered. We give some details, and for this purpose consider an operator equation \( A u = f \), where \( A: H \to F \) is a bounded linear operator between Hilbert spaces \( H \) and \( F \). In addition, \( f \) is supposed to belong to the range of \( A \) which may be non-closed, in general. For the stable solution of \( Au=f \), Tikhonov regularization \( u_\alpha = (\alpha I + A^* A)^{-1} A^* \tilde{f}, \, \alpha > 0 \), is considered, where \( A^*: F \to H \) denotes the adjoint of the operator \( A \), and \( \tilde{f} \in F \) is an approximation to the right-hand side \( f \). The considered parameter choice rule is of the form \[ d(\alpha,q,k,s) := \frac{\kappa_\alpha^{2k+s_0-s} \| D_\alpha^k B_\alpha (Au_\alpha - \tilde{f}) \|^2} {\| D_\alpha^{sq/2} B_\alpha^{2q-2} (Au_\alpha - \tilde{f}) \|^{2/q}} = b \delta^{2-2/q}, \tag{\(*\)} \] where \( q, k, s \) and \( b \) are some parameters, with \( s \geq 0, \, k \geq s/2 \), and \( b \) is sufficiently large. In addition, \( \delta > 0 \) represents some noise level for the right-hand side of \( Au= f\), and this noise level may be under- or overestimated, in general. Moreover we have \( \kappa_\alpha = 1 + \alpha \| A \|^{-2}, \, B_\alpha = \sqrt{\alpha}(\alpha I + A A^*)^{-1/2} \) and \( D_\alpha = \alpha^{-1} AA^* B_\alpha^2 \;(\alpha > 0) \). Finally, there holds \( s_0 = 0 \) if \( 2k = s \), and \( s_0 = 1 \) otherwise. In the course of this paper, the special case \( q = 3/2 \) is considered in more detail. For this setting, the following three results are presented: a) if \( \alpha=\alpha(\delta) \) is the smallest solution of equation \( (*) \) and \( \| \tilde{f} - f \| \leq \delta \) holds, then \( \| u_{\alpha(\delta)} - u_* \| \to 0 \) as \( \delta \to 0 \), where \( u_* \in H \) denotes the minimum norm solution of the equation \( Au=f \); b) if \( \alpha=\alpha(\delta) \) is the largest solution of equation \( (*) \) and \( \| \tilde{f} - f \| \leq c_* \delta \) holds for some constant \( c_* \), then \( \| u_{\alpha(\delta)} - u_* \| \to 0 \) as \( \delta \to 0 \); c) if \( \alpha=\alpha(\delta) \) is the smallest solution of equation \( (*) \), and if in addition \( \alpha(\delta)\) can be bounded from above by some constant specified in the paper, then in the case \( \| \tilde{f} - f \| \leq \delta \) the order optimal estimate \( \| u_{\alpha(\delta)} - u_* \| \leq C\inf_{\alpha>0} \{ \| u_{\alpha}^+ - u_* \| + \tfrac{\delta}{2\sqrt{\alpha}} \} \) holds. Here the notation \( u_\alpha^+ = (\alpha I + A^* A)^{-1} A^* f, \, \alpha > 0 \), is used, and \( C \) is some positive constant. In b), \( k > s/2 \) is additionally required, and in c), \( k \leq 3s/4 \) is assumed. The paper concludes with the presentation of some numerical experiments.