A GCV based Arnoldi-Tikhonov regularization method (Q398632): Difference between revisions

This paper considers a linear discrete ill-posed problem \(Ax=b\) with the square system matrix and the right-hand side affected by noise; the (discrete) Picard condition is assumed. The problem is solved using Tikhonov regularization \(\min_x\{\|Ax-b\|^2+\lambda^2\|Lx\|^2\}\), where the regularization matrix \(L\) can be, in general, different from the identity. The authors use the standard Arnoldi algorithm for an iterative projection of \(A\) (and also of the whole minimization problem) to a lower-dimensional subspace; the so-called Arnoldi-Tikhonov method. The regularization parameter \(\lambda\) is in each step obtained by the generalized cross validation (GCV) approach. The authors analyze the convergence properties of the whole procees, using the (generalized) singular value decomposition of the matrix \(A\), or the matrix pair \((A,L)\). The results are illustrated on numerical examples from the regularization toolbox and on image restoration problem.