Reduced functions, gradients and Hessians from fixed-point iterations for state equations (Q1610720)

Let \(f:\mathbb{R}^j\times \mathbb{R}^n \to \mathbb{R}^m\) be an objective function as those which appears in design optimization and parameter identification. Let \(w=F(z,x)=0\) be a parameter-dependent system of nonlinear equations, where \(w\in \mathbb{R}^j\), and \(x\) is a vector of parameters. This paper deals with numerical differentiation and the fundamental problem to be considered is the calculation of rapidly converging estimates for the reduced function \(f(z(x),x)\) and its total derivatives with respect to \(x\). Several key assumptions are needed in this work. Firstly, to eliminate the variables \(z\) and \(w\), it is assumed the regularity of the Jacobians of \(F(z,x)\) and \(f(z,x)\) with respect to \(z\in\mathbb{R}^j\) and \(x \in \mathbb{R}^n\). Regularity means that \(F\) and \(f\) are once Lipschitz-continuously differentiable and the square matrix \(F_z\) is nonsingular at the points of a neighborhood. The iterates \(z_k\) to approximate a solution of \(w=0\), satisfy a recurrence of the form \(z_{k+1}=z_k-P_kF(z_k,x)\), where \(P_k\) is the preconditioner which approximates the inverse of the Jacobian \(F_z\). Other assumption states stable convergence in the sense that the preconditioners \(P_k\) satisfy \(\|I-P_k F_z(z,x)\|\leq \rho<1\). Finally, it is also assumed that the derivative tensors \(\partial P_k/(\partial z,\partial x)\) are uniformly bounded over all \(k\), so that derivative recurrences lead to the same convergence. This approach is an application of the adjoint differentiation mode and is verified by some test calculations on an aircraft wing with two responses and two variables.