Suppose $v$ is a (possibly time-dependent) vector field on a compact manifold $M$.

Its flow is a mapping $g: M \times \mathbb{R} \rightarrow M$, where $g$ satisfies the following conditions (written in the local chart):

$g(x, 0) = x$

$\frac{dg}{dt} (x, t) = v(g(x, t), t)$

I would like to find for a fixed $\epsilon$ such $\delta$ that, for any $\delta$-variation of the vector field, the differential flow for time $1$ would not perturb more than on $\epsilon$.

The norm on vector fields is $\sup_{x\in M}(|v(x)|)$ w.r.t. to some arbitrary riemannian metric on $M$, and the distance between mappings $f$, $g$ is defined as $\sup_{x \in M}(dist(f(x), g(x)))$.

While this fact is almost surely standard, I can not figure out if it follows from the continuous dependence of the solution of ODE on parameters. Does anyone know a good reference?

  • 2
    $\begingroup$ I think that it is a question suitable for MSE rather than for MO. I will give you a hint, however: write the flow as a solution of an integral equation, and apply Grönwall's inequality. But the above holds if $v$ is locally Lipschitz, for example. If $v$ is continuous only, we do not need to have uniqueness. $\endgroup$ – user539887 May 15 '18 at 11:40
  • $\begingroup$ I think the statement should hold for any measurable $v$ which is in $\epsilon$-ball in $l^\infty$, provided solution exists. $\endgroup$ – Asya Rorschach May 15 '18 at 12:27
  • $\begingroup$ @AsyaRorschach Without uniqueness of solutions the question does not make sense. Even continuity is too weak assumption to guarantee uniqueness, so measurability is way too weak. $\endgroup$ – Oleg Eroshkin May 15 '18 at 14:45
  • $\begingroup$ @AsyaRorschach For an example that even under the assumption of continuity one need not select a continuous solution, see Jean Duchon's answer to ODE with Holder drift - Cauchy-Peano theorem. $\endgroup$ – user539887 May 15 '18 at 15:27
  • $\begingroup$ @user539887 will estimate still be in $\sup |v|$, or Lipschitz constant also? $\endgroup$ – Asya Rorschach May 15 '18 at 17:25

Let me consider two autonomous vector fields $X_1, X_2$ on a compact smooth manifold $\mathcal M$ and assume that the Lipschitz condition is true for both of them. The flow $\psi_j$ of $X_j$ is defined on $\mathbb R\times \mathcal M$ by $$ \dot \psi_j(t,m)=X_j(\psi_j(t,m)), \quad \psi_j(0,m)=m. $$ Now we consider for $m$ in a coordinate chart and $t\ge 0$ small enough $$ \psi_2(t,m)-\psi_1(t,m)=\int_0^t\bigl( X_2(\psi_2(s,m))-X_1(\psi_1(s,m))\bigr) ds, $$ so that \begin{multline} \vert\psi_2(t,m)-\psi_1(t,m)\vert\le \int_0^t\bigl\vert X_2(\psi_2(s,m))-X_1(\psi_1(s,m))\bigr\vert ds \\ \le \int_0^t\bigl\vert X_2(\psi_2(s,m))-X_2(\psi_1(s,m))\bigr\vert ds +\int_0^t\bigl\vert X_2(\psi_1(s,m))-X_1(\psi_1(s,m))\bigr\vert ds \\ \le \int_0^t L_2\vert\psi_2(s,m)-\psi_1(s,m)\vert ds +t\Vert X_2-X_1\Vert=R(t). \end{multline} We have thus with $\rho(t)=\vert\psi_2(t,m)-\psi_1(t,m)\vert$, $$ \dot R(t)=L_2\rho(t)+\Vert X_2-X_1\Vert\le L_2 R(t)+\Vert X_2-X_1\Vert, $$ so that $ \frac{d}{dt}\log\bigl(L_2 R(t)+\Vert X_2-X_1\Vert\bigr)\le L_2 $ and $$ \log\bigl(L_2 R(t)+\Vert X_2-X_1\Vert\bigr)\le \log\bigl(\Vert X_2-X_1\Vert\bigr)+ L_2 t, $$ entailing $L_2 \rho(t)+\Vert X_2-X_1\Vert\le L_2 R(t)+\Vert X_2-X_1\Vert\le e^{L_2 t}\Vert X_2-X_1\Vert $ and $$ \rho(t)\le L_2^{-1}(e^{L_2 t}-1)\Vert X_2-X_1\Vert. $$

  • $\begingroup$ Thanks for the proof! Could you please explain how do you use that $t$ is arbitrary small? Also, am I right that $X_1$ Lipschitz constant is not used in the estimate? $\endgroup$ – Asya Rorschach May 16 '18 at 9:16
  • $\begingroup$ @AsyaRorschach I choose $t$ small enough to stay in the coordinate chart. On the other hand, although I did not use a Lipschitz constant for $X_1$, I was glad to have the Lipschitz assumption on $X_1$ to ensure the existence of a flow. You can make various variations on the above inequalities and control $\psi_2(t,m_2)-\psi_1(t,m_1)$, which would make the role of $X_1, X_2$ more symmetrical. Also, you can get flows with an Osgood-type hypothesis "logarithmically close" to Lipschitz, e.g. for log-Lipschitz vector fields and I believe that the reasoning should be similar to the above arguments. $\endgroup$ – Bazin May 16 '18 at 19:37

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.