An optimal regularized projection method for solving ill-posed problems via dynamical systems method (Q984793)

The dynamical systems method is considered in a finite-dimensional setting, with the discrepancy principle as parameter choice strategy. The considered equation is \(Ax =y \), where \( A : X \to X \) is a bounded linear operator in a Hilbert space \( X \), and \(y\) belongs to the (non-closed) range of \(A\). It is supposed that the minimum norm solution \( x^+ \) of the equation \(Ax=y \) can be represented in the form \(x^+ = (A^*A)^\nu w, \| w \| \leq \rho \), where \( \nu > 0 \) and \( \rho > 0 \), and \( A^*: X \to X \) denotes the adjoint operator of \( A \). It is moreover assumed that both \(A\) and \( A^* \) are bounded operators from \( X \) into a linear subspace \(X^r \subset X \), where \( X^r \) is equipped with a norm that is stronger than the norm \( \| \cdot \| \) on \(X\). For the discretization, finite-dimensional linear subspaces \( V_0 \subset V_1 \subset \dots \subset X \) are considered. It is assumed that the associated orthogonal projection operators \( P_j: X \to X \) onto \( V_j \) satisfy \( \| I - P_j \|_{X^r \to X} \leq c_r 2^{-rj} \) for \( j = 0, 1, \dots \). The approximations considered in this paper are solutions of the following abstract Cauchy problems: \[ (x_n^\delta)^\prime(t)= g(A_n^*A_n) A_n^* (y^\delta- A_n x_n^\delta(t)), \;t > 0, \;x_n^\delta(0) = 0 \;(n \geq 0). \] Here, \( g(\lambda) = (a+\lambda)^{-1} \), with \( a = \text{const}>0 \), and \[ A_n=\sum_{k=1}^{2n} (P_k-P_{k-1}) A P_{2n-k} + P_0 A P_{2n}: X \to X \] are approximations to \( A \) which require only a small fraction of the scalar products needed to compute the standard discretization \( P_{2n} A P_{2n} \). In addition, \( y^\delta \in X \) is a perturbation of the right-hand side \( y \) satisfying \( \| y^\delta -y \| \leq \delta \), with given noise level \(\delta > 0 \). The parameter choice strategy is as follows: compute approximations \( x_{n(t_l)}^\delta \in V_{n(t_l)} \) for \( t_{l+1} = t_l + q, l = 0, 1, \dots \) until the condition \( \| A_{n(t_l)} x_{n(t_l)}^\delta - P_{2n(t_l)}y^\delta \| \leq b \delta \) is satisfied for the first time, where \( n(t_l) \) is some discretization level depending on \( t_l \), and \( q \) and \( b \) are fixed, and \( t_0 = 0 \). It is shown that the accuracy of this method is of optimal order, i.e., we have \( \| x^+ - x_{n(t_L)}^\delta(t_L) \| \leq c (\rho \delta^{\nu})^{1/(2\nu+1)} \), where \( t_L = t \) denotes the parameter chosen according to the stopping rule. The computational complexity of the numerical scheme is also considered.