Automatic differentiation of the vector that solves a parametric linear system (Q1176796): Difference between revisions

The broader use of formulae manipulation languages nowadays requires also more studies of techniques of formal (automatic) differentiation. The paper under review treats the automatic differentiation of the solution of a linear system of equations, if the matrix and thus also the solution depends on parameters. If \(A\) is an \((n\times n)\)-matrix whose entries depend on parameters \(s\in S\subset\mathbb{R}^ p\) such that \(A(s)\) is regular for these parameters, then the solution \(x\) of \(Ax = b\) may be regarded as a function \(x: S\rightarrow \mathbb{R}^ n\), even when the right hand side \(b\) also depends on the same set of parameters. The question investigated in the paper is how to compute automatically the first and second derivatives of \(x\) with respect to the various components of \(s\). The paper starts with some general remarks on how to build up a programming system which can handle automatic differentiation taking elementary and library functions into account and starting with the obvious rules for differentiation of composite functions. A first section on automatic differentiation of \(x\) is devoted to the solution of \(Ax = b\) by elimination methods using certain decompositions of \(A\) or by other finite methods like Cramer's rule. Another section is used to explain how to differentiate \(x\) when \(x\) is found by an infinite iterative process like Jacobi's method. In the end there is a sketch for differentiating \(x\) which solves the more general problem \(K(s,x(s))=0\).