4
$\begingroup$

Let us assume I have an A-stable numerical scheme. I believe that given any linear equation $y' = Ay$, it means that the numerical scheme applied to this equation is stable (and therefore convergent since it is consistent) if the eigenvalues of A have a negative real part.

My question is then, does this result extend to the non-linear case ? I am interested in particular in a system $y' = f(y,t)$ where the Jacobian of $f$ has negative real eigenvalues $\forall y$.

edit: Also I guess I could give a particular example here (but my question is not limited to that case): for example let us say the ODE describes a 2nd order chemical reaction network, i.e. $f \in R^n \rightarrow R^n$ is a polynomial function of the $y$ of order 2, where $n$ is the number of species. Considering the network I was able to prove that the eigenvalues of the Jacobian are real negative for any $y$, however I don't know if I can make the further argument that any A-stable method will converge (by the way, what about L-stable methods, B-stable methods, ... ?)

$\endgroup$
  • $\begingroup$ Usually, yes. But there are pathological counterexamples. $\endgroup$ – David Ketcheson Mar 8 '13 at 12:23
  • $\begingroup$ Thanks. Could you give me a counter example I could take a look at ? I also edited my question for further details $\endgroup$ – Tibo Mar 8 '13 at 17:20
  • $\begingroup$ regarding your edit: If you are really dealing with a fixed-size system of ODEs (and not a PDE semi-discretization), then any one-step method is guaranteed to converge. You don't need $A$-stability for that. You only need to show that $f$ is Lipschitz continuous in $y$. $\endgroup$ – David Ketcheson Mar 10 '13 at 10:06
6
$\begingroup$

I'm going to answer a more general question than the one you asked: do the eigenvalues of an initial value ODE determine the stability of the solution? Here I'm referring to mathematical stability, not numerical stability. Of course, a "yes" to this question is a necessary condition for a "yes" to your question. And unfortunately, the answer is "no".

In general, for nonlinear and/or non-autonomous problems, studying the eigenvalues of a frozen linearization, i.e. the Jacobian with $t$ fixed, gives very useful insight. However, there are two major caveats:

  1. For non-normal matrices, eigenvalues don't tell the whole story, and one must consider pseudospectra instead. This caveat applies even to linear, constant-coefficient problems. See the book by Trefethen & Embree.

  2. There exist pathological examples of systems where the solution behavior is completely unrelated to the eigenvalues. I will give an example that is attributed to Vinograd:

$$y'(t) = A(t) y(t)$$

with

$$A(t) = \begin{pmatrix} -1-9\cos^2(6t)+6\sin(12t) & 12\cos^2(6t) + \frac{9}{2}\sin(12t) \\ -12\sin^2(6t) + \frac{9}{2}\sin(12t) & -1 - 9 \sin^2(6t)-6\sin(12t) \end{pmatrix}$$

The eigenvalues of $A(t)$ are $\lambda=-1, -10$ -- independent of $t$. But the solution is

$$y(t) = C_1 \exp(2t) \begin{pmatrix} \cos(6t) + 2\sin(6t) \\ 2\cos(6t)-\sin(6t) \end{pmatrix} + C_2 \exp(-13t) \begin{pmatrix} \sin(6t) - 2\cos(6t) \\ 2\sin(6t)+\cos(6t) \end{pmatrix}$$

So in this case, the eigenvalues suggest stability, but the solution exhibits unbounded exponential growth. For a nice discussion of generalizations of this example, see this excellent book. It is out of print, but used copies can sometimes be found -- for a hefty price -- on Amazon.

Some further comments: I don't know of a nonlinear, autonomous example exhibiting this kind of behavior, and it may well be that they don't exist. I also don't know of an example arising in applications. People do generally use the eigenvalues of the frozen jacobian as an indicator of stability in practice.

$\endgroup$
  • $\begingroup$ thanks a lot, that is what I wanted to know. I will try to get a hand on these two books too. In the meantime, I guess I will just make the argument using the eigenvalues of the Jacobian and mention that some counterexamples exist. $\endgroup$ – Tibo Mar 10 '13 at 17:27
  • $\begingroup$ @Tibo did you see my last comment above (on your question)? $\endgroup$ – David Ketcheson Mar 10 '13 at 18:26
  • $\begingroup$ just saw it thanks. You are right I may have not done a correct distinction in my question between convergence (which means that the numerical scheme will give me the right answer when $\Delta t \rightarrow 0$) and stability (which is that the solution of the numerical scheme is bounded). $\endgroup$ – Tibo Mar 10 '13 at 23:06
  • $\begingroup$ For the convergence case of the pure ODE system as in a chemical reaction network, the Lipschitz argument is a bit tricky though: the function f being a polynomial of order 2, it is Lipschitz continous on any interval, but the size of the Lipschitz constant grows with the size of the interval, so I am not sure I can apply the convergence theorem for any proper one step method... $\endgroup$ – Tibo Mar 10 '13 at 23:15
  • $\begingroup$ I mentioned A-stability because I wanted to be sure there was convergence for any $\Delta t$ due to the fact that the eigenvalues of the Jacobian are real negative but unbounded (same reason that the Lipschitz constant $L$ is unbounded) $\endgroup$ – Tibo Mar 10 '13 at 23:20

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.