5
$\begingroup$

Hi!

Let's say that we have a dynamical system described by

$\dot{x} = f(x)$,

where f is some nonlinear function, which has several equilibria. Assume that we have found a continously differentiable Liapunov function V such that

$\dot{V} = 0 \Rightarrow \dot{x} = 0$.

Then, assuming that V is radially unbounded, by the LaSalle invariance principle we should be able to say that the system always converges to an equilibrium point. However, in some works I have seen the additional requirement that in order to show convergence, all equilibrium points must be isolated, otherwise the system could move indefinitely inside a connected set of equilibrium points. Can that really be the case for the situation described above? Doesn't $\dot{x} = 0$ mean that the system has "stopped" (assuming that $x$ completely describes the state of the system)? It seems to me that in my case, the assumption about isolated equilibria is unnecessary.

Kind regards
Olav

$\endgroup$

3 Answers 3

2
$\begingroup$

To illustrate Michael's answet, take the system $$\dot x=(1-x^2-y^2)x-zy,\qquad \dot y=(1-x^2-y^2)y+zx,\qquad \dot z=-z^2.$$ In cylindrical coordinates, it writes $$\dot r=(1-r^2)r,\qquad\dot\theta=z,\qquad \dot z=-z^2.$$ The first equation tells that $r(t)\rightarrow1$ as $t\rightarrow+\infty$. Besides, $$z(t)=\frac{z_0}{1+tz_0}.$$ If $z(0)=z_0$ is positive, the trajectory is defined for all $t>0$, and $\dot\theta$ is not integrable at $+\infty$, so that the solution spins infinitely many times towards the unit circle.

Upon Didier's request. here is a similar example, in the plane. Take two functions $r\mapsto h(r),k(r)$. Consider the system $$\dot r=h(r),\qquad \dot\theta=k(r),$$ which rewrites $$\dot x=\frac{h(r)}{r}x-k(r)y,\qquad\dot y=\frac{h(r)}{r}y+k(r)x.$$ Assume that $(r-1)h(r)<0$ for $r\ne 0,1$. Then you have a Lyapunov function $V(r)$, minimal at $r=1$. The rest points form the circle $\{r=1\}$. If $h$ is flat enough at $r=1$, the convergence $r\rightarrow1$ as $t\rightarrow+\infty$ is algebraic. Then if $k=r^2-1$, a trajectory spins infinitely many times.

$\endgroup$
8
  • $\begingroup$ Is there an obvious Lyapunov function for the system? $\endgroup$ Jan 21, 2011 at 14:13
  • $\begingroup$ Denis: What would be a simple $2$-dimensional version of your example? $\endgroup$
    – Did
    Jan 21, 2011 at 14:30
  • $\begingroup$ You can make a $C^\infty$ vector field in the plane that is 0 on the unit disk, and where all orbits outside the disk spiral counterclockwise inward toward the disk. There's a $C^{infty} Lyapunov function that is identically 0 on the disk, and a monotone function of distance from the origin outside the disk. $\endgroup$ Jan 21, 2011 at 14:44
  • $\begingroup$ @Andrey. Take any smooth function $V(X)=g(|X|^2)$ with the property that $g'<0$ over $(0,1)$ and $g'>0$ over $(1,+\infty)$. It may be flat at $0$. $\endgroup$ Jan 21, 2011 at 15:06
  • $\begingroup$ @Denis Serre: Yes, but $V$ should also satisfy the conditions $V(0)=0$ and $V(X)>0$ for $X\neq 0$. These are not compatible with the property that $g'(x)<0$ for $x\in(0,1)$. $\endgroup$ Jan 21, 2011 at 15:18
2
$\begingroup$

There are also gradient flows with limit cycles. That is $$ \dot x = - \nabla V(x) $$ meaning that $V$ itself is a Lyapunov function by construction. Such an example in polar form $V(r \cos(\theta),r\sin(\theta))=v(r,\theta)$ is given by $$ v(r,\theta) = \begin{cases} \exp\Bigl(-\tfrac{1}{1-r^2}\Bigr) , & r < 1 \\ \exp\Bigl(-\tfrac{1}{r^2-1}\Bigr) \sin\Bigl(\frac{1}{r-1} - \theta\Bigr) , & r>1 \\ 0 & r=0 \end{cases} $$ This is a non-analytic function (at $r=1$). The solution to the gradient flow initially at the origin stays there, for initial points with $r<1$ it converges to $r=1$ with $\theta$ staying constant. For initial $r=1$ solutions are stuck again. For initial $r>1$ there are two cases: For generic initial $\theta$ it converges to an outwards spiraling solution. However, for one specific initial choice $\theta^*=\theta^*(r)$, there is an inwards spiraling solution having the unit circle as limit cycle. See also the picture below.

I'm not sure if one could fix the example to have for generic initial values limit cycles, or if those for gradient systems are somehow special. Polar plot of v

$\endgroup$
1
$\begingroup$

The system stops only in the limit as time goes to infinity. You could, for instance, have a circle consisting of equilibrium point, and as time goes to infinity, solutions spiral towards the circle with continually decreasing speed.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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