Numerical methods for second order singular perturbation problems (Q1187805)

The paper introduces some numerical methods for the second order singular perturbation boundary value problem \(\varepsilon y''+s(t)y'+c(t)y=f(t)\), \(t\in [a,b]\), \(y(a)=\alpha\), \(y(b)=\beta\), where \(\varepsilon\) is positive and very small. The discrete problem has the form \(\sigma_ iy_{i-1}+y_ i+\tau_ iy_{i+1}=f_ i\), \(y_ 0=\alpha\), \(y_{n+1}=\beta\), where \(\sigma_ i\), \(\tau_ i\), \(f_ i\) depend on the steps \(h_{i-1}\), \(h_ i\), \(h_{i+1}\) and on the values \(s(t_ i)\), \(c(t_ i)\). In the matrix form, the problem can be written as \(Ty=f\). The authors use sufficient conditions which ensure the well- conditioning of the tridiagonal matrix \(T\). Two theorems which give sufficient conditions are proved for the case of a constant step size, and one for the case of variable step sizes. These considerations are taken into account in the optimal implementation of the methods. Under hypothesis similar to the ones usually used in the continuous theory, methods are derived which use constant step sizes \(h\approx O(\sqrt{\varepsilon})\). In the general case methods with variable step sizes, are defined which solve the problem with a number of steps smaller than \(1/\varepsilon\). All the test equations found in the quoted literature are successfully tested under different choices of the parameter \(\varepsilon\). Examples on standard test problems are shown.