I would like to have a basic models as a ground truth for the numerical solvers. I am looking for systems which have available analytic solution. As an example I know that the closed form solution of IVP: \begin{align} &\dot{y}(t) = -2y(t)\\ &y(0) = 1 \end{align} is: \begin{align} y(t) = e^{-2t} \end{align} When using forward euler I know that numerical solution is given by the recurrence relation: \begin{align} &y_k = (1-2\Delta{t})y_{k-1}\\ &y_0 = y(0) \end{align} This gives me a way of calculating global error for this specific system: \begin{align} e_k = |y_k - y(t_k)| \end{align} The problem is that the above system may not be "relevant" as a benchmark for numerical solvers. I am looking for relevant ones as it is much easier to present the results in that case. There are articles on this topic such as "Hull et al.: Comparing numerical methods for ordinary differential equations" and "Enright et al.: Comparing numerical methods for stiff systems of ODE:s". The problem is that I am looking for a way to present results using a global error. Is there any similar calculating closed form solutions?

Any suggestions and/or critiques are appreciated.

  • $\begingroup$ So you want more complicated systems than the example you presented ? $\endgroup$ – Piyush Grover Feb 20 '17 at 15:53
  • 1
    $\begingroup$ @PiyushGrover They should be more complicated in the sense they reveal strengths and weaknesses of different ODE solvers. Another example may be 2nd order linear system with a large stiffness ratio and so on. I could take the time and try to enumerate them, but I am hoping to find a reference that already did something similar. This was already done in mentioned articles, but with local error as one of criteria instead of global error. $\endgroup$ – Slaven Glumac Feb 20 '17 at 17:28
  • $\begingroup$ Focussing on systems with closed-form solutions pretty much excludes chaotic systems, which are ones where global error can be severe (butterfly effect etc.) $\endgroup$ – Robert Israel Feb 20 '17 at 17:56
  • 1
    $\begingroup$ @Robert Israel I realize that this is the case and this fine for me. I would like to make experiments with such demonstration systems for introduction purposes. $\endgroup$ – Slaven Glumac Feb 20 '17 at 20:28
  • 3
    $\begingroup$ You can easily come up with stiff systems which are analytically solvable. Basically, you need the eigenvalues to be of very different sizes, say $\lambda_1=10^3\lambda_2$. $\endgroup$ – Piyush Grover Feb 20 '17 at 22:13

There is a whole subfield of applied mathematics devoted to developing ODE solvers and understanding their properties. Consequently, there are thousands of relevant papers, and not much more can be said without a more specific question. The canonical reference, which includes lots of methods tested on lots of ODE systems (with global error as a metric) is

Hairer, Norsett, Wanner. Solving ordinary differential equations (2 volumes).


It's easy to come up with a differential equation that has a known general solution, expressed in the form $F(t,y) = constant$ where $F$ is differentiable:

$$ \dfrac{\partial F}{\partial t} + \dfrac{\partial F}{\partial y} y' = 0 $$

Similarly, for a system, expressed in implicit form as

$$ F_i(t, y_1, \ldots, y_n) = c_i,\ i=1 \ldots,n $$ differentiation gives you $$ \dfrac{\partial F_i}{\partial t} + \sum_{j=1}^n \dfrac{\partial F_i}{\partial y_j} y_j' = 0$$ which (when the Jacobian is invertible) you then solve for $y'_1, \ldots, y'_n$.


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.