Fractional Tikhonov regularization for linear discrete ill-posed problems (Q533711)

For a given discrete ill-posed problem \(Ax \approx b\), the following regularized least-squares problem is proposed: \[ \min_{x \in \mathbb R^n}\{ \|Ax - b\|^2_W + \mu \|x\|^2 \} \] with the weighted norm \(\|y\|_W = (y^T W y)^{1/2}\) for \(W = (AA^T)^{(\alpha-1)/2}\). This formulation requires the specification of two parameters \(\alpha,\mu>0\) and much of the paper is devoted to studying the influence of these parameters on the solution. The corresponding normal equations \[ ( (A^TA)^{(\alpha+1)/2} + \mu I)x = (A^TA)^{(\alpha-1)/2} A^T b \] characterize the solution, implying that \(x\) admits an expansion \[ x = \varphi(\sigma_1) ( u_1^T b)v_1 + \varphi(\sigma_2) ( u_2^T b)v_2 + \cdots, \] where \(\sigma_j,u_j,v_j\) denote the singular values/vectors of \(A\). The so-called filter function \(\varphi\) takes the form \(\varphi(\sigma) = \frac{\sigma^\alpha}{\sigma^{\alpha + 1} + \mu}\). This expansion not only allows for the computation of \(x\) for moderately sized problems (for large-scale problems a Krylov subspace approach is proposed) but also allows for the derivation of various theoretical results. It is shown how the discrepancy principle can be applied to determine a suitable value of \(\mu\). The sensitivity of the resulting choice for \(\mu\) is investigated. Concerning the choice of \(\alpha\), the authors conclude -- based on numerical experiments -- that ``\dots letting \(\alpha\) be smaller than, but close to unity, such as \(\alpha = 0.8\), gives better results than \(\alpha = 1\) for almost all examples. Smaller values of \(\alpha\), such as \(\alpha = 0.5\), may work even better for some examples, at the cost of rendering worse results for others.''