Wavelet-accelerated Tikhonov-Phillips regularization with applications (Q4356950)

Many technical and physical problems can be mathematically modeled by operator equations of the first kind, \[ Ax= y,\tag{1} \] where \(x\) is the searched-for information under observed data \(y\). We consider \(A\) as a compact non-degenerate linear operator acting between the real Hilbert spaces \(X\) and \(Y\). In this setting the problem (1) is ill-posed, that is, its minimum norm solution \(x^+\) does not depend continuously on the right hand side \(y\). Small perturbations in \(y\) may cause dramatic changes in \(x^+\). This instability has to be taken into account by any solution technique for (1). The more as only a perturbation \(y^\delta\) of the exact but unknown data \(y\) is available in general. The perturbation of \(y\) is caused by noise which can not be avoided in real-life applications due to the specific experiment and due to the limitations of the measuring apparatus. The perturbed data \(y^\delta\) are assumed to satisfy \(|y- y^\delta|_Y \leq \delta\) with an \(a\)-priori known noise level \(\delta> 0\). One of the theoretically best understood and most often used stabilization techniques for (1) is the Tikhonov-Phillips regularization where the linear equation (1) is replaced by the minimization problem \[ \text{find } x_\alpha^\delta\in X \text{ which minimizes }T_\alpha(x)= |Ax-y^\delta |_Y^2+\alpha|x|_X^2. \tag{2} \] Here, \(\alpha> 0\) is the regularization parameter. The idea of the Tikhonov-Phillips regularization (2) is to control the influence of the data error in the regularized solution \(x_\alpha^\delta\) by adding a penalty term. The unique minimizer of (2) is given as the unique solution of the regularized normal equation \[ (A^*A+\alpha I)x_\alpha^\delta= A^*y^\delta. \tag{3} \] The high art of regularization is the determination of the regularization parameter \(\alpha= \alpha(\delta,y^\delta)\) such that \(x_\alpha^\delta\) converges to \(x^+\) as \(\delta \to 0\). Examples for such parameter selection strategies are presented. We introduce two methods to speed up the solution process of (3) which can even be combined. Both methods employ wavelet techniques. We present a fast multilevel iteration for the solution of a discrete version of the normal equation (3). The theoretical results we achieve are illustrated by numerical examples where the abstract operator equation (1) will be an integral equation. Finally, we discuss the potential of our multilevel method for solving the 3D-reconstruction problem in computerized tomography. Any iterative scheme for solving (3) requires the multiplication of a vector by the operator \(A^*A+\alpha I\) (resp. a matrix version thereof). Typically, this operator (matrix) will be dense. Therefore, operator compression techniques will speed up any iterative solver. Such methods are considered. First, we study compression schemes from a theoretical point of view and then we discuss two ways of computing such compressions. We report on results obtained by applying this approach to hyperthermia treatment planning. For the entire collection see [Zbl 0871.00041].