An iterative method for Tikhonov regularization with a general linear regularization operator (Q616671)

To solve a large scale ill-posed linear least squares problem \(Ax=b\) with \(A\in {\mathbb C}^{m\times n}\), a general regularization is proposed in which one has to minimize \(\|Ax-b\|^2+\mu\|Lx\|^2\) with regularization parameter \(\mu>0\) and regularization operator \(L\). It is assumed that \(Ax=\hat{b}\) is consistent and that an upper bound \(\delta\) is known for the perturbation \(b-\hat{b}\). The proposed iterative method is based on a Golub-Kahan bidiagonalization of \(A\). This projects the problem on a \(k\)-dimensional subspace so that one has to solve a \(\mu\)-dependent least squares problem of size \((2k+1)\times k\) with a simple right-hand side that has only one nonzero element. The main idea of the paper is to use a value for \(\mu=\mu(\delta)\) such that with the corresponding solution \(x^{(\mu)}\) one has \(\|Ax^{(\mu)}-b\|=\eta\delta\) with \(\eta>1\) a user-specified constant. This identifies \(\nu=1/\mu\) as a Lagrange multiplier for \(\min \|Lx\|\) with constraint \(\|Ax-b\|=\eta\delta\). This \(\nu\) satisfies in general an equation with a unique solution. The numerical applications given include discretization of integral equations and image processing. In the printed version, the blue graphs for the examples are missing. These are visible in the on-line version.