Splitting methods for second-order initial value problems (Q1281789)

The authors consider stiff initial value problems for systems of second-order ordinary differential equations of the special form \((*)\) \({d^2y(t)\over dt^2}= f(y(t))\), where \(y, f\in\mathbb{R}^d\). In this context stiff means that the stepsize \(h\) is such that in the neighborhood of the exact solution \(h^2\cdot\rho \left({\partial f\over\partial y}\right)\gg 1\), where \(\rho\) is the spectral radius of the Jacobian matrix. Equations of this type often arise if time-dependent partial differential equations are semidiscretized by the method of lines. Solving such equations by an implicit method implies that one is faced with the problem of solving systems of implicit relations. The paper under review focuses on the construction and analysis of iterative solution methods which are effective in cases where an approximation \(J\) to the Jacobian \({\partial f\over\partial y}\) can be split into a sum of matrices \(J_i\) with an essentially simpler structure than the matrix \(J\). These iterative methods consist of the modified Newton method (outer iteration), in which the linear Newton systems are solved by a second iteration process (the inner iteration) which is based on approximate factorization. The inner-outer iteration process is called by the authors approximate factorization iteration. The authors show to what extent the convergence and stability results valid for first-order ordinary differential equation methods change in the second-order case \((*)\). A number of convergence results is derived for the model situation where the matrices \(J_i\) share the same eigensystem and possess a negative eigenvalue spectrum. Finally, the authors present order of accuracy and stability results in the case of a finite number of inner and outer iterations.