CDS140a

Nonlinear Systems: Local Theory

02/01/2011

Stability and Lyapunov Functions

3.1

Lyapunov Stability

Denition:

x0 of (1) is stable
t (x) N (x0 ).

An equilibrium point
and

t 0,

we have

An equilibrium point

An equilibrium point
that for all

x0 of

if for all

 > 0,

there exists a

>0

such that for all

x N (x0 )

(1) is unstable if it is not stable.

x0 of (1) is asymptotically stable

x N (x0 ) we have limt t (x) = x0 .

if it is stable and if there exists a

>0

such

Remarks:

The about limit being satised does not imply that

From H-G theorem and Stable manifold theorem, it follows that hyperbolic equilibrium points are

x0 is

stable (why?).

either asymptotically stable (sinks) or unstable (sources or saddles).

If

x0

x0 is

Df (x0 )

is stable then no eigenvalue of

has positive real part (why?)

stable but not asymptotically stable, then

x0 is

a non-hyperbolic equilibrium point

Example: Perko 2.9.2 (c) Determine stability of the equilibrium points of :

Equilibrium points are

x 1
x 2


=

4x1 2x2 + 4
x1 x2

(0, 2), (1, 0).

Df (x)

Df (0, 2)

Df (1, 0)

4
x2

2
x1

4
2

2
0

4
0

2
1

What can we say in general about the stability of non-hyperbolic equilibrium points?

3.2

x 1
x 2


=

x2 x1 x2
x1 + x21

Lyapunov Functions

Denition:

Let

derivative of the

f C 1 (E), V C 1 (E) and t the ow of

function V (x) along the solution t (x) is

the dierential equation 1. Then for

xE

the

d
V (t ) d
V (x) = V (t (x)) =
t (x) = DV (x)f (x)
dt
t dt

Theorem (Lyapunov's Direct Method):

f C 1 (E) and that f (x0 ) = 0.

Let

be an open subset of

Rn

x0 . Suppose
V C 1 (E) satisfying

containing

Suppose further that there exists a real valued function

V (x0 ) = 0 and V (x) > 0 if x 6= x0 . Then

V (x) 0 for all x E , x0 is stable;
(b) if V (x) < 0 for all x E\{x0 }, x0 is asymptotically
(c) if V (x) > 0 for all x E\{x0 }, x0 is unstable;
(a) if

stable;

Proof: Without loss of generality, we assume that x0 = 0.

Part (a)
We want to show that

 > 0, there exists a > 0 such that

x N (0) and t 0, we have t (x) N (0)

for all

for all

Br

!"

Outline:
1. Construct a closed set (ball)
that

Br E

Br N ,

such

(i.e., a technicality to make sure

we remain in the domain)

2. Construct

= {x Br |V (x) } (i.e., a
sublevel set of V ) such that
the interior of Br

subset of the

lies in

Figure 1: Sets used in the proof.

Can show that condition (a) implies that

x t (x)
3. Construct
Then since

N (0)

N (0) Br N (0),

we have that

x N t (x) t (x) N (0).

Details:
1. Given any

 > 0,

choose

0 < r ,

such that

Br = {x Rn | |x| r} E.
= min|x|=r V (x) (i.e., the minimum of V in the boundary of Br ). Take 0 < < , =
{x Br |V (x) } . Then it can be easily shown that lies in the interior of Br (if a point a is in
the boundary, then V (a) > ). Notice that for x = 0 (x) , and for all t

2. Let

V (t (x)) V (0 (x))

=
0

d
V (s (x))ds 0
ds

(since

d
V (s (x)) 0)
ds

V (t (x))
and therefore
3. Since

t (x) (t cannot

is continuous,

V (0) = 0

V (0 (x))

exit

Br

since it would mean going through the boundary of

then there exists a

>0

such that

|x| < V (x) < .

Br ).

Therefore for

x N x t (x) t (x) Br t (x) N (0).

So for any

>0

we constructed a

such that for all

x N (0)

and

t 0,

we have

t (x) N (0),

and

therefore the origin is stable.

Part (b) Note:

Intuitively, condition

of 1) implies that as

increases, the

that it eventually goes to

V (x) < 0, x 6= 0 s(i.e., V (x) is strictly decreasing along the trajectories

trajectory moves into lower level sets of V (x). We just need to show

0.

In part (a) we showed that the origin is stable. What we need to show is that

10

 > 0 such that for all x N (0) we have limt t (x) = 0, i.e., there exists a > 0 such
T > 0 such that for all x N (0) and t > T , |t (x)| <  (or t (x) N (0)) .
showed that for all  > 0 we can construct such that N (0), i.e.,

there exists a
that for all

 > 0,

But since we

there exists a

t (x) t (x) N (0).

x N (0)

Therefore, it is sucient to show that for all

lim V (t (x)) = 0

t
(why? because this means that for all

t (x) N (0).)
Since V is a decreasing

>0

there exists a

T > 0,

such that for

t > T , |V (t (x))| < ,

i.e.

function along the trajectories (condition (b)) and bounded below, then

lim V (t (x)) = c 0.

c > 0. Let c = {x Br |V (x) c}. By continuity of V and V (0) = 0, there exists a d > 0, such
Bd = {x Rn | |x| d} c . Since limt V (t (x)) = c, then t (x) lies outside of Bd , i.e., t (x) lies
in the compact set d |x| r , V achieves its maximum in this set. Let = maxd|x|r V (x) > 0. We
have for t > 0
t
d
V (s (x))ds
V (t (x)) = V (0 (x)) +
0 ds
V (0 (x)) t
Assume

that

V (t (x))

eventually

<

c
But we assumed

Part (c)

c > 0,

<

we have a contradiction.

Reverse time (i.e., take

t = t)

then one gets part (b).

Remarks:

satisfying the conditions of the theorem is called a Lyapunov function.

The theorem allows to determine the stability of the equilibrium point without explicitly solving the
dierential equation. In a sense, since

V (x) = DV (x)f (x)

the method converts a dynamics problem (i.e. determining the behavior of the trajectories over time),
into a algebraic one (i.e., verifying inequalities of the form
function

F (x) > 0,

where

is some continuous

F : Rn R)

One can think of the Lyapunov function as a generalization of the idea of the energy of a system.
Then the method studies stability by looking at the rate of change of this measure of energy.

See [1] for a more detailed treatment of Lyapunov functions and nonlinear stability.

The method does not show how to nd a Lyapunov function

Denition: The region of attraction (RoA)

{x Rn | limt t (x) = 0}.
are subsets of the RoA. This way we have a

V.

of an the equilibrium point at the origin for (1) is

procedure for estimating the RoA (by maximizing

More on this later.

11

).

Example 1 (Perko 9.5 (a) [2] )

Lets try

= x + y + xy

= x y x2 y 3

V = x2 + y 2
V

2x(x + y + xy) + 2y(x y x2 y 3 )

2x2 + 4xy 2y 2 2y 4

= 2(x y)2 2y 4
< 0
So the origin is asymptotically stable.

Example 2

Linear Harmonic Oscillator (spring mass)

x
+ kx = 0, k > 0.
x =

Is the origin stable?

= kx

1 2
1
2
Energy E(x, y) = P E + KE =
2 kx + 2 y . This is a good candidate
V (x, y) = E(x, y). Lets check:
2
Let D = R . First V (0, 0) = 0 and V (x, y) > 0 for (x, y) D\(0, 0).

V (x, y)

for a Lyapunov function, i.e.

V x +
V y
x
y
= kxy ykx
=

So it is stable .

What happens if we add a damping term to the equation?


First Jacobian

A=

0
k

1
0

= y

= kx y 3 (1 + x2 )

= i k ,

V (x, y)

so linear analysis not useful. Using same V we get

V x +
V y
x
y
= kxy + y(kx y 3 (1 + x2 ))
=

= y 4 (1 + x2 )
So it is stable for

 > 0.

asymptotically stable for

3.3

Can show, using LaSalle's Invariance Principle (coming up) that it is indeed

>0

and unstable for

 < 0.

Global Stability

Theorem (Barbashin-Krasovskii):
exists a real valued function

V C

1
n
Suppose f C (R ) and that f (0) =
n
(R ) satisfying V (0) = 0 and V (x) > 0 if

|x| V (x)
V (x) < 0, x 6= 0
12

0. Suppose further
x 6= x0 , and

that there

x=0

then

is globally asymptotically stable.

|x| V (x)

Remark: The additional condition

guarantees that the level sets of

V (x)

are

bounded. Why?

p Rn , let c = V (p).

For any

c > 0 there is r > 0 such that |x| > r V (x) > c

limt f (t) = ). |x| > r V (x) > c means that if x c (i.e.,
and therefore c Br and therefore bounded.

Condition means that for any

(by denition; similar to denition of

V (x) c)

then

x Br (i.e., |x| r),

Example 3 (Strogatz 7.2.12 [3])

Lets try

x =

x + 2y 3 2y 4

x y + xy

V = x2m + ay 2n
V

2mx2m1 (x + 2y 3 2y 4 ) + 2any 2n1 (x y + xy)

= 2mx2m + 4mx2m1 y 3 4mx2m1 y 4 2any 2n1 x + 2any 2n x 2any 2n

= 2mx2m 2any 2n + 4mx2m1 y 3 4mx2m1 y 4 2any 2n1 x + 2any 2n x


let m=1
= 2x2 2any 2n + 4xy 3 4xy 4 2any 2n1 x + 2any 2n x


let n=2
= 2x2 4ay 4 + 4xy 3 4xy 4 4ay 3 x + 4ay 4 x
let a=1
So

V = x2 + y 4

= 2x2 4y 4
would work, and the origin is globally asymptotically stable.

LaSalle's Theorem (LaSalle's Invariance Principle):

Let

be a compact set that is positively

invariant with respect to the ow of (1). Suppose that there exists a real valued function
that

V (x) 0

D0 .

Then every solution starting in

in

Let

D0

be the set of all points in

approaches

as

V (x) = 0,
t .

where

and

V C 1 (E)

such

the largest invariant set in

Remarks:

The Theorem does not require that

If

One can compute invariant subsets of the RoA of the origin by using the following Lemma:

M = {(0, 0)}

Lemma:
level set

then

is positive denite (

V (x) > 0, x 6= 0).

is in the RoA of the origin.

If there exist a continuously dierentiable real valued function

V, = {x Rn | V (x) }

and

> 0,

such that the

is bounded and

V (0) = 0, V (x) > 0, x 6= 0

V (x) < 0, x 6= 0, x V,
then

V,

is invariant and in the RoA of the origin.

Example 4 (Harmonic Oscillator) Let

x
+ sin x = 0.

a) Can you prove stability of the origin using linearization? Use an appropriate Lyapunov function to prove
that the origin is a stable xed point.
(b) Add a damping term

x
+ x + sin x = 0.

Study the stability of the origin for

Solution: (a)

= y

= sin x
13

 > 0.


Lets look at xed point

(0, 0).

Jacobian

A=

0
1

1
0


,

= i,

so linear analysis not useful. Energy

E(x, y) = P E + KE = 1 cos x + 21 y 2 . This is a good candidate for a Lyapunov

E(x, y). Lets check:
Let E = (, ) R. First V (0, 0) = 0 and V (x, y) > 0 for (x, y) D\(0, 0).
V (x, y)

V x +
V y
x
y
(sin x)y y sin x

function, i.e.,

V (x, y) =

So it is stable .
(b)

= y

= sin x y

Using sameV we get

V (x, y)

V x +
V y
x
y
= (sin x)y + y( sin x (1 x2 )y)
=

= y 2
V 0 and D0 = {(x, y) R2 | V (x, y) = 0} = {(x, y) R2 | y = 0}. But since for (x, y) D0
(y = 0), y = sin x 6= 0 for x 6= k, k Z, the largest invariant subset of D0 is M = {(k, 0), k =
0, 1, 2, . . .}. I.e., the theorem states that the solutions will converge to one of the xed points. If we
choose E = (, ) R, then we get M = {(0, 0)} and therefore the origin is asymptotically stable.
Additionally, any solution starting in = {(x, y) E | V (x, y) } will converge to 0. is an estimate

Take then

of the RoA of the origin.

References
[1] H. K. Khalil. Nonlinear Systems. Prentice Hall,

3rd

edition, 2002.

[2] L. Perko. Dierential Equations and Dynamical Systems. Springer,

[3] S. H. Strogatz.

3rd

edition, 2001.

Nonlinear Dynamics And Chaos: With Applications To Physics, Biology, Chemistry,

And Engineering. Westview Press, 2001.

14