A steepest descent algorithm for the global minimization of the Tikhonov functional (Q4531339)

Ill-posed problems of the form \[ F(f,\mu) := Af + B(f,\mu) = y\tag{1} \] are considered, where \( A: X_1 \to Y \) and \( B: X_1 \times X_2 \to Y \) are bounded operators which are linear and bilinear, respectively, and \( X_1, X_2 \) and \( Y \) are Hilbert spaces. The right-hand side \( y \in Y \) is approximately given, and \( f \in X_1 \) and \( \mu \in X_2 \) are unknown. The purpose of this paper is to establish convergence of the steepest descent method with certain variable step sizes to determine a global minimizer of the Tikhonov functional \( \Phi_\alpha(f,\mu) = \|F(f,u) -y^\delta \|^2 + \alpha \|(f,\mu) - (\overline{f},\overline{\mu}) \|^2 \). Here, \( y^\delta \in Y \) is a perturbation of the exact right-hand side \( y \) with \( \|y^\delta - y \|\leq \delta \), and \( (\overline{f},\overline{\mu}) \) is an approximation to a solution of equation (1). It is supposed that there exists a solution \( (f_*,u_*) \) to equation (1) with minimal distance to \( (\overline{f},\overline{\mu}) \) and that the error \( (f_*,u_*) - (\overline{f},\overline{\mu}) \) belongs to the range of the adjoint operator of the Fréchet derivative of \( F \) at \( (f_*,u_*) \). A parameter choice of the form \( \alpha = c \delta \) with some constant \( c \) and an arbitrary starting value \( (f_0,u_0) \) is then considered. It is shown that if \( \alpha \) is sufficiently large, \( \alpha > \alpha_0 \) with some \( \alpha_0 \), then the starting value is sufficiently close to the global minimizer of \( \Phi_{\alpha} \), and the steepest descent method converges to that minimizer. In the other situation (\( \alpha \leq \alpha_0 \)), a modified procedure is proposed where the sequence of regularization parameters \( \alpha_j = q^j \alpha_0 \) for \( j = 0, 1, \ldots, n \) is considered, with \( 0 < q < 1 \) and \( n \) chosen such that \( \alpha_n \approx \alpha \) holds. Here, for \( j = 0, 1, \ldots, n \), the global minimizer of the Tikhonov functional \( \Phi_{\alpha_j} \) is determined by using the steepest descent method, with the global minimizer of \( \Phi_{\alpha_{j-1}} \) taken as starting value (for \( j \geq 1 \); in the case \( j = 0 \), \( (f_0,u_0) \) is taken). Finally, the global minimizer of the Tikhonov functional \( \Phi_{\alpha} \) is then determined by using the steepest descent method, with the global minimizer of \( \Phi_{\alpha_{n}} \) taken as starting value. This procedure is shown to be convergent, and the obtained results are numerically tested for the attenuated Radon transform with applications in medical imaging.