Automatic differentiation of large sparse systems (Q2277762)

The authors are concerned with the task of calculating the gradient \(\nabla F(x)\) and the Hessian \(\nabla^ 2F(x)\) of a function F(x), \(x\in {\mathbb{R}}^ n\), by exploiting the features of modern computer languages (for instance, ADA, Pascal-Sc) which allow the definition of operators and data structures. Following previous works by \textit{A. Griewank} [Mathematical programming, Proc. 13th Int. Symp.,Tokyo/Jap. 1988, Math. Appl., Jap. Ser. 6, 83-107 (1989; Zbl 0696.65015)] and \textit{L. B. Rall} [``Optimal implementation of differential arithmetic, Computer arithmetic. Scientific computation and programming language'' (1987; Zbl 0614.65001)] which consider symbolic computation for automatic differentiation, here five new algebras are introduced and discussed. The case of sparsity on the function F(x) is taken into account. The proposed automatic differentiation is then coupled with well-known minimization algorithms, the truncated Newton method as proposed by \textit{R. S. Dembo} [Math. Program. Study 31, 43-71 (1987; Zbl 0635.90072)] and the conjugate gradient algorithm based on the Polak-Ribière formula. Numerical results obtained in solving the extended Dixon problem, the extended Powell problem with \(n=4,8,20,40,80\) and the Dixon-Maany problem with \(n=3000\), are reported. This demonstrates widely the effectiveness of the approach considered by the authors.