Estimates of the rate of convergence of Runge-Kutta type algorithms with random errors and with consideration of the calculation time (Q2739934)

The paper deals with the Cauchy problem NEWLINE\[NEWLINEy'=f(t,y), \quad y(t_0)=y_0,NEWLINE\]NEWLINE where \(f(t,y)\) is computed in real time with random error \(\varepsilon(t,y)\), \(E\varepsilon(t,y)=0\), \(D\varepsilon(t,y)=\widetilde\sigma^2(t,y)\). Let \(\{\tau_{k}(t,y)\), \(t\geq 0\), \(y\in R^{m}\}\) be a family of non-negative independent for different \(k\) identically distributed random variables. For the solution of the given Cauchy problem the authors consider the recurrent computing procedure NEWLINE\[NEWLINE\widehat y_{k+1,h}=\widehat y_{k,h}+h(\widetilde f(t_{k,h},\widehat y_{k,h})+ \varepsilon_{k}(t_{k,h},\widehat y_{k,h})), \quad \widehat y_{0,h}=y_0,NEWLINE\]NEWLINE where \(\widehat y_{k,h}\) is the approximate value of function \(y(t)\) (explicit solution of the Cauchy problem) at the point \(t_{k,h}\); \(t_{k+1,h}=t_{k,h}+ h\tau_{k}(t_{k,h},\widehat y_{k,h})\), \(t_{0,h}=0\). Let the computing time for \(\widehat y_{k,h}\) depends on \(\tau_{k,h}(t,y)=h\tau_{k}(t,y)\), \(E\tau_{k}(t,y)=m(t,y)\), \(D\tau_{k}(t,y)=\sigma^2(t,y)\). Let \(\widetilde f(t,y)=f(t,y)m(t,y)\) and \(\{\varepsilon_{k}(t,y)\), \(t\geq 0\), \(y\in R^{m}\}\) be the family of independent for different \(k\) random variables with distribution such as the distribution of \(\varepsilon(t,y)\). NEWLINENEWLINENEWLINEThe authors prove that if (i) \(E(\tau^2_{k}(t,y)+\varepsilon_{k}^2(t,y))<C\); (ii) \(D\tau_{k}(t,y)= \sigma^2(t,y)<C\); (iii) for some \(C_{L}\), for all \(L>0\) such, that \(|y_1|\vee|y_2|<L\): \(|m(t,y_1)-m(t,y_2)|+|f(t,y_1)-f(t,y_2)|\leq C_{L} |y_1-y_2)|\), then \(\sup_{t_0\leq t\leq T} |\widehat y_{h}(t)-y(t))|\to 0\) in probability for all \(T\) such that \(y(+\infty)>T\). Here \(\widehat y_{h}(t)=\widehat y_{k,h}\) as \(t\in [t_{k,h}, t_{k+1,h})\).