No label defined (Q4532042)

The authors consider numerical methods for solving a discrete analogue of the nonlinear singularly perturbed boundary value problem \[ -\varepsilon^2u''+c(x,u)=0,\quad x\in[0,1],\;\;u(0)=u(1)=0, \] with a small perturbation parameter \(\varepsilon\in(0,\varepsilon_0)\), \(\varepsilon_0\ll 1\). To preserve numerical stability, a combination of the Hermite scheme of the fourth order and the standard central difference scheme of the second order on mesh of Bakhvalov's type is used, giving as a final result the fourth order uniform (in \(\varepsilon\)) accuracy. Choosing a mesh generating function and a scheme for the approximation to \(u''(x)\) at the mesh points, one obtains a discrete analogue in the form of a nonlinear system \(F(y)=0\), where \(F:R^{n+1}\to R^{n+1}\) is a nonlinear mapping of a special form. The obtained nonlinear system is transformed and split into two systems with smaller dimensions and then solved by Newton's method. The new algorithms accelerate the convergence by using an approach which combines solving Newtonian linear systems with symmetric and positive definite matrix employing conjugate gradient and Orthomin methods. Numerical examples show that the number of Newton iterations is reduced and the CPU time in most cases is decreased by the dimension of the system.