Multiscale support vector approach for solving ill-posed problems (Q499265)

Based on compactly supported radial basis functions, the support vector approach (SVA) is a specific version of Tikhonov's regularization method for the stable approximate solution of semi-discretized linear Fredholm integral equations of the first kind on bounded domains in \(\mathbb{R}^d\). In particular, the misfit term in the Tikhonov functional to be minimized uses a sum of Vapnik \(\epsilon\)-intensive distance functions expressing the discrepancy for each data. In this paper, the authors extend this idea to the case of a multiscale support vector approach (MSVA). Using the technique of reproducing kernel Hilbert spaces, convergence for a presented algorithm of (MSVA) is shown when the original problem in infinite-dimensional Hilbert spaces with compact forward operator is moderately ill-posed and satisfies some further technical assumptions. Since in addition to the conventional Tikhonov regularization parameter also Vapnik's \(\epsilon\)-parameter can be controlled, a multi-parameter regularization approach with two regularization parameters occurs. This seems to be an advantage if both parameters can be tuned in a sophisticated way. One section discusses different kinds of simultaneous parameter choices for noise-free and noisy data and derives corresponding error estimates. The authors claim that this paper presents the first result in regularization theory for ill-posed problems, where convergence can be established even if the regularization parameter does not depend on the noise level. However, taking into account Bakushinsky's well-known veto, this can only be the case if a second parameter, here \(\epsilon\), is involved in the parameter choice strategy. Numerical case studies complete the paper.