In general, my understanding is the global vs. one-step-error of a given method will differ by 1. Assuming this is true and that the one-step-error is order $\alpha+1$ then the error of taking a single step of size $h$ from our known initial condition, say $y(t_0)=y_0$ means that our approximation of $y(t_0+h)$, lets call it $w_1$ is $O(h^{\alpha+1})$.

Now considering cutting $h$ in half and taking 2 steps to approximate $y(t_0+h)$. Let's call this approximation $\bar{w_1}$, then this error should be $O( (h/2)^\alpha)=(1/2)^\alpha O(h^\alpha)$. For example, consider $\alpha=2$ and one step of $h=0.1$ vs two steps of $h=0.05$, then $0.1^3 < 0.05^2$ which seems to imply the accuracy went down by reducing $h$.

I'm use to just thinking that the accuracy goes up by reducing the stepwise, in particular, by a factor of the reduction of the step size to the order of the method. But this assumes that the global error is used in both cases. Considering taking a single step, vs 2, with the respective one-step and global errors, the analysis does not seem as clear.

To try and test it, I used RKmidpoint, an order 2 method, on a problem with an exact answer. Interestingly, comparing one step, vs two steps out to the same final time, generally scaled by $(1/2)^2$ (e.g. $|y(t_f)-\bar{w_1}| \approx (1/4)*|y(t_f) -w_1|)$, as would be expected by just considering global error and not the additional order from the one-step-error perspective.

Should this be expected in general? Was that just due to the problem I considered? Why shouldn't it follow the analysis that considers the one-step-error and the global error?


Consider some interval $[a,b]$ so that $L(b-a)$ is rather small and the variation in the derivatives of the exact solution is not large. Then one can assume that the local error of a single step looks like $$Ch^{α+1}+..$$ and that the errors of the steps simply add to the global error. If the integration over $[a,b]$ is preformed in $N$ steps of step size $h=\frac{b-a)}N$ the dominating term of the accumulated global error is thus $$N(Ch^{α+1}+..)=(b-a)Ch^α+..= (b-a)^{α+1}CN^{-α}+..,$$ in $N=2$ steps thus $$2(Ch^{α+1}+..)=(b-a)Ch^α+..= (b-a)^{α+1}C2^{-α}+..,$$ As this demonstrates, the number of steps combines with one factor $h$ of the local error to the interval length as additional constant factor, thus decreasing the local error order by 1 to give the global error order.

Over larger integration intervals one has to take into account that the accumulated errors of the last step are magnified, as per Grönwall, by a factor of $e^{Lh}$. If one steers the step size in such a way that the local error at step $k$ is always smaller than $\epsilon h_k$, then the accumulated errors satisfy $$e_{k+1}\le e^{Lh_k}e_k+ϵh_k\implies e^{-Lt_{k+1}}e_{k+1}\le e^{-Lt_k}e_k+e^{-Lt_{k+1}}ϵh_k$$ so that in the main terms $$ e_n\leϵ\int_{t_0}^{t_n}e^{L(t_n-t)}dt=\fracϵL\left(e^{L(t_n-t_0)}-1\right) $$

  • $\begingroup$ Well $2^{-\alpha}Ch^{\alpha+1}$ nicely shows considering the one-step-error throughout gives the $2^{-\alpha}$ scaling that would be expected from considering the global. But I'd still like to know why considering the one-step and the global is still wrong? What's the fallacy there? $\endgroup$ – Fractal20 Dec 11 '17 at 18:27
  • $\begingroup$ The fallacy is that you assume that the constant of the big-O terms are all $1$ or at least equal. I hope I have changed the text to be better understandable. $\endgroup$ – LutzL Dec 11 '17 at 22:34
  • $\begingroup$ thanks for the responses. So you would expect that in general I could find specific problems such that playing around with step sizes would be in line with considering the one-step and global error together? Specifically when the big-O terms are approximately equal? $\endgroup$ – Fractal20 Dec 16 '17 at 21:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.