Projected Tikhonov regularization of large-scale discrete ill-posed problems (Q368090): Difference between revisions

Least squares problems \(\min_{x \in \mathbb R^m} \| Ax - b \|\), \(A \in \mathbb R^{m \times n}\), \(b \in \mathbb R^m\) are considered where the singular values of \(A\) are clustered at the origin and the vector \(b\) is of the type \(b_{\text{true}} + e\) with a perturbation \(e \in \mathbb R^m\). For the investigation of the influence of the perturbation \(e\) on the solution of the least squares problem confidence intervals for the solution coordinates are determined. The confidence intervals are obtained by solving constrained minimization problems \(\min_{x \in \mathbb R^n} w^Tx\) subject to \(\| Ax - b \| \leq \varepsilon\), \(\| x - d \| \leq \delta\). Here, \(\varepsilon\) is an a priori upper bound for \(\| e \|\), the vector \(d\) is an a priori estimate of the solution, \(w\) is an arbitrary unit vector, and \(\delta\) is a positive constant. The authors consider the situation when confidence intervals are to be determined for many coordinates of the solution of the unperturbed problem. The presented method for solving the constrained minimization problems makes use of a low rank approximation of the matrix \(A\) based on a partial singular value decomposition. The choice of the rank of the partial singular value decomposition is discussed and an algorithm for determining the rank is given. For illustrating the performance of the proposed methods numerical experiments are reported.