Estimating the error of the classic Runge-Kutta formula (Q1805273)

An estimator is introduced for the stepwise error in the popular four- point Runge-Kutta algorithm for approximating the solution to an ordinary differential equation and the estimator is compared theoretically with the estimator obtained by interval doubling and an estimator given by \textit{J. A. Zonneveld} [Automatic numerical integration (1964; Zbl 0139.319)]. In addition a general method is presented to judge effectiveness and cost of an error estimator. The estimator based on doubling requires 11 function evaluations, Zonneveld's requires 8, and the new estimate 7 evaluations. The new estimator may be regarded as no cost in that the 7 function evaluations are exactly those needed to implement the doubling algorithm. If the doubling algorithm is written (i) \(y_{n + 1} = y_n + (h/12)\), \((f_0 + 2f_1 + 2f_2 + f_3 + f_4 + 2f_5 + 2f_6 + f_7)\) then the estimate, \(\text{est}_n \approx y_{n + 1} - y(x_n + h)\) is given by (ii) \(\text{est}_n = (h/72) (-f_1 + 2f_2 - f_3 - 2f_4 + 3f_5 - f_7)\).