No label defined (Q4532111)

Iterative methods of gradient type are suggested for approximate solution of nonlinear operator equations of the form \[ F(x)= 0 \] where \(F\:H_1\to H_2\) is a nonlinear operator acting in Hilbert spaces \(H_1\) and \(H_2\). It is assumed that \(F\) is twice differentiable in the sense of Gâteaux. Moreover, \(F'(x)\: H_1 \to H_2\) is a compact operator for every \(x\in \Omega_R\) and \[ \|F'(x)\|\leq N_1,\quad \|F''(x)\|\leq N_2,\quad x\in\Omega_R, \] where \(\Omega_R = \{x\in H_1:\|x - x^{*}\|\leq R\}\), \(R > 0\); \(x^{*}\) denotes a solution to the equation. A sequence of approximate solutions \(\{x_n\}\) is constructed in the form \[ x_0\in\Omega_R,\quad x_{n+1} = \widetilde\Pi_N^{(m)}(x_0)(x_n - \xi - \gamma\widetilde F^{\prime *}(x_n)\widetilde F(x_n)) + \xi, \] where \(\widetilde F\) is an approximation for \(F\), \(\widetilde\Pi_N(x_0)\) is a projection of the operator \(\widetilde F^{\prime *}(x)\widetilde F'(x)\) and \[ \|\widetilde\Pi_N^{(m)}(x_0) - \widetilde\Pi_N(x_0)\|\leq \kappa_m,\quad \lim_{m\to\infty}\kappa_m = 0. \] The authors prove the convergence of the approximate solutions generated by the methods to a neighborhood of the solution with diameter proportional to the magnitude of errors in the input data and a sourcewise representation of the initial residual. This property is guaranteed by a suitable combination of the method of gradient descent for the residual functional with approximate projection onto special finite-dimensional subspaces.