Arnoldi-Tikhonov regularization methods (Q1008669): Difference between revisions

The problem is to solve a large, ill-conditioned linear system \(Ax=b\) of size \(n\), where \(b=\hat{b}+e\) with \(\hat{b}\) the ``true'' vector and \(e\) some error. Tikhonov regularization minimizes \(\|Ax-b\|^2+\mu^{-1}\|x\|\) with \(\mu\) a regularization parameter. The proposed (range restricted) Arnoldi-Tikhonov regularization looks for the minimizer \(x_{\mu,\ell}\) in the Krylov subspace \(K_\ell(A,Ab)=\text{span}\{Ab,A^2b,\dots,A^{\ell}b\}\) produced by the Arnoldi method. It is shown that, under some conditions, \(\varphi_\ell(\mu)=\|Ax_{\mu,\ell}-b\|^2\) is convex with a unique minimum, and \(\ell\) is taken to be the smallest (or slightly larger) index for which \(\varphi(\mu)<\eta^2\varepsilon^2\) where \(\varepsilon\) is some estimate for \(\|e\|\) and \(\eta>1\) reflects the uncertainty of the estimate \(\varepsilon\). An efficient implementation is described that compares favorably with range restricted generalized minimal residual (GMRES) method (GMRES applied within \(K_\ell(A,Ab)\)), and other regularization methods of the authors which is illustrated by several numerical examples.