Ritz-regularization versus least-square-regularization. Solution methods for integral equations of the first kind (Q1066622): Difference between revisions

Suppose W is a dense subspace of a real Hilbert space H, \(T: H\to H\) is a linear, compact, positive definite operator, and \(L: W\to H\) is a linear positive operator with a bounded or compact inverse. To approximate the solution of the equation of first kind, \(Tx=y\), the author considers the regularized least squares solution, \(w^{LS}_{\alpha}\), which minimizes the functional \(Q_{\alpha}(w)=\| Tw-y\|^ 2+\alpha \| Lw\|^ 2\), and the regularized Ritz solution, \(w^ R_{\alpha}\), which minimizes the functional \(R_{\alpha}(w)=(Tw,w)- 2(y,w)+\alpha (Lw,w)\), where \(\alpha >0\). In the case where T and L commute, the author shows that the error in approximation \(w^ R_{\alpha}\) is dominated by the error in \(w_{\alpha}^{LS}\). In fact, the same is true for the finite dimensional approximations \(w^{LS}_{\alpha,n}\) and \(w^ R_{\alpha,n}\) obtained by minimization of the respective functionals over the subspace spanned by the first n eigenfunctions. The author also proves a convergence theorem for the Galerkin approximations \(w^ G_{i\alpha}\in W+iW\) satisfying \((Tw+i\alpha Lw-y,v)=0\) for all \(v\in W+iW\) where \(\alpha\) is a nonzero real parameter, and their finite dimensional counterparts, \(w^ G_{i\alpha,n}\). Finally, the author computes the condition numbers for the system of linear equations which determine the approximations \(w^{LS}_{\alpha,n}\), \(w^ R_{\alpha,n}\), and \(w^ G_{i\alpha,n}\) in the case in which T and L commute, the ratios of the eigenvalues of T to those of L form a monotone sequence, and the subspace of coordinate functions consists of the first n eigenfunctions.