Lecture Notes for Math 524

Dr. Yingfei Yi

These notes are based on the lecture notes of Professor James S. Muldowney, the books of Hale,
Copple, Coddington and Levinson, and Perko. They are for the use of students in my graduate
ODE class.

Chapter 4. Global Behaviours of Solutions, Dynamical Systems

4.1. Dynamical systems or flows

Let X be a topological space, a dynamical system or a flow on X is continuous map ϕ : T × X →
X, where T = Z or R, and ϕ(t, x) = ϕt (x), such that

(1) ϕ0 (x) = x, x ∈ X.

(2) ϕs (ϕt (x)) = ϕs+t (x), t, s ∈ R, x ∈ X.

It follows from (1) and (2) that

ϕ−t (ϕt (x)) = ϕ0 (x) = x, t ∈ R, x ∈ X.

Therefore {ϕt } is a one-parameter family of homeomorphisms on X.

As some examples:
1. A linear dynamical system: Let A be an n × n matrix. Define

ϕt (x0 ) = eAt x0 .

It is the f.m. of the linear system x0 = Ax. Verify that ϕt is a flow (exercise.)
2. A nonlinear dynamical system generated by an ODE: Let x = x(t, x0 ) be the solution to the
IVP x0 = f (x), x(0) = x0 . Define
ϕt (x0 ) = x(t, x0 ).
Verify that φt is a flow (assuming that all solutions exist for t ∈ R ).

Question: If we define ϕt (x0 ) = x(t, x0 ) for solutions of the IVP

x0 = f (t, x), x(0) = x0 ,

is ϕt a flow?

1
3. Semi-dynamical systems or semi-flows: if the condition (2) only holds for t, s ∈ R+ , then ϕt is
said to be a semi-dynamical system or a semi-flow. This is the case, when ϕt is generated from an
IVP that only have forward uniqueness of solutions, and in many situations, for discrete dynamical
systems.
4. Discrete dynamical systems: Let F : Rn → Rn be continuous. Define ϕt (x) = F t (x), t ∈ N,
i.e., F 0 (x) = F (x), F t+1 (x) = F (F t (x)), t = 0, 1, 2, · · · . If F is a homeomorphism, then F −1 is
defined, and F t is defined for t ∈ Z. Therefore ϕt is a discrete flow. For instance, let φt (x0 ) is
the continuous flow generated from the IVP x0 = f (x), x(0) = x0 , define F (x0 ) = ϕ1 (x0 ), the
time-one map. Then F is a diffeomorphism, and F n defines a discrete flow. To investigate long
time behaviours of the IVP, one can study the continuous flow φt or the discrete flow F n . On the
other hand, if a map F is not invertible, then F n is only defined for n ≥ 0 and thus F n defines
a discrete semi-flow. For instance, F (x) = x2 + a only generates a discrete semi-flow. A discrete
flow (or semi-flow) F n is also called an iterated map.

More definitions:

The orbit of x is the set {ϕ(x) : t ∈ R}. The positive semi-orbit (or negative semi-orbit) is the
set {ϕt (x) : t ∈ R+ } (or {ϕt (x) : t ∈ R− } ).
A point x is an equilibrium if ϕt (x) = x, t ∈ R. x is a periodic point if ϕω (x) = x for some ω.
In this case, if x is not an equilibrium, the orbit {ϕt (x) : 0 ≤ t < ω} is a simple closed curve in
Rn and called a closed orbit, for continuous flows, or is a finite set of points for discrete flows.

A set K ⊂ X is positively invariant (or negatively invariant) if ϕt (K) ⊂ K, t ≥ 0 (or for t ≤ 0 ),

and K is invariant if ϕt (K) = K, t ∈ R. An equilibrium, periodic orbit, any orbit are examples
of invariant sets.

Our interest is in the nature of the flow near invariant sets, in particular, near equilibria and
periodic orbits.

4.2. Stable manifold theorem

To see what type of results one should expect, we consider some examples.

Example 1. Let h(x1 , x2 ) = (x1 , x2 + x21 /3)T , y = h(x), x = (x1 , x2 ), y = (y1 , y2 ). We find
h−1 (y1 , y2 ) = (y1 , y2 − y12 /3). Consider the system
· ¸0 · ¸
x1 −x1
= x0 = f (x). (1)
x2 x2 + x21

Linearize at equilibrium (0, 0), we have the following system

· ¸0 · ¸· ¸
y1 −1 0 y1 ∂f
= y0 = (0) y. (2)
y2 0 1 y2 ∂x

Make the change of variables y = h(x) in (1), we have

∂h 0 ∂h
y0 = x = f (x)
∂x ∂x

2
or
· ¸0 · ¸ · ¸0 · ¸· ¸
y1 1 0 x1 1 0 −x1
= 2 = 2
y2 3 x1 1 x2 x1 1 x2 + x21
· ¸ 3· ¸ · ¸· ¸
−x1 −y1 −1 0 y1
= = = .
− 32 x21 + x2 + x21 y2 0 1 y2

The homeomorphism y = h(x) transforms system (1) to system (2). If x(t, x0 ) is a solution to
(1) with x(0) = x0 , then y(t) = h(x(t, x0 )) is a solution to (2) with y(0) = h(x0 ). Namely,

h(x(t, x0 )) = y(t, h(x0 )).

In terms of the flows, let ϕt and φt be the flows generated by (1) and (2), respectively, we have

ψ t ◦ h = h ◦ ϕt

or the following diagram commutes

insert phase diagrams here.

Two flows that satisfy the above property for a homeomorphism h will be called topologically
conjugate. Topologically conjugate flows have the same orbit structure. The example shows that,
near a equilibrium, the nonlinear flow can be topologically conjugate to its linearized flow. The
following theorem gives the general result, a proof can be found in Hartman’s ODE book.
∂f
An equilibrium x = x is said to be hyperbolic if ∂x (x) has no eigenvalues with zero real parts.

Theorem 4.1 (Hartman-Grobman Theorem) Suppose that D ⊂ Rn is open, and let ϕt be

the flow generated by the IVP x0 = f (x), x(0) = x0 . Assume

(1) f ∈ C 1 (D → Rn ).
∂f
(2) x = 0 is a hyperbolic equilibrium. ( A = ∂x (0) )

Then there exist neighbourhoods U, V of 0 and a homeomorphism h : U → V, such that, if x0 ∈ U

then there exists an interval 0 ∈ I0 such that

h ◦ ϕt (x0 ) = eAt h(x0 ), t ∈ I0 .

In another word, the flow ϕt is topologically conjugate to the linear flow eAt near x = 0.

3
Remarks.
1. A generalization of this theorem to the case of non-hyperbolic equilibria is contained in the work
of Palmer and others, see Kirchgraber and Palmer, Pitman Research Notes # 233, 1990.
2. Hartman shows that if f ∈ C 2 , then h can be of C 1 . In general, we don’t expect that h is as
smooth as f, as can be seen in the following example.

Example 2. Consider
x
x01 = −x1 − p 2
log x21 + x22
x (3)
x02 = −x2 + p 1
log x21 + x22
and its linearization at equilibrium (0, 0)

y10 = −y1
(4)
y20 = −y2 .

In polar coordinates, (3) becomes

r0 = −r
1
θ0 =
log r
so that r(t) → 0 and θ(t) → −∞ as t → ∞. Therefore, (0, 0) is a spiral for (3) and a stable
node for (4). There exists no diffeomorphism that preserves the orbit structure (why?).

insert phase diagrams here.

Example 3. Consider
x01 = −x1
x02 = −x2 + x21 (5)
x03 = x3 + x21
and its linearization at equilibrium (0, 0, 0)

y10 = −y1
y20 = −y2 (6)
y30 = y3 .

We can solve both systems to obtain

x1 (t) = x01 e−t

x2 (t) = (x02 + x201 ) e−t − x201 e−2t
x201 t x02 −2t
x3 (t) = (x03 + )e − e
3 3

4
and

y1 (t) = y01 e−t

y2 (t) = y02 e−t
y3 (t) = y03 et .

The stable subspace SA of (6) is the y1 y2 plane, while the unstable subspace UA is the y3 axis.
For the nonlinear system (5), all the initial conditions x(0) = (x01 , x02 , x03 ) that give rise to
exponential decaying solutions limt→∞ x(t) = 0 must satisfy x03 + x201 /3 = 0, which defines a
2-dimensional surface S that is tangent to SA at (0, 0, 0). Moreover, x03 + x201 /3 = 0 implies
x3 (t) + x1 (t)2 /3 = 0 for all t ≥ 0. Thus S is an invariant manifold for (5). Similarly, initial
conditions x(0) that give rise to exponentially growing solutions limt→−∞ x(t) = 0 form a 1-
dimensional negatively invariant manifold U defined by x1 = x2 = 0, which is tangent (the same
as) to UA at (0, 0, 0).

insert phase portraits here.

More generally, we have the following theorem.

Theorem 4.2 (Stable Manifold Theorem) Suppose

(1) f ∈ C 1 (D → Rn ), f (0) = 0.
(2) A = ∂f∂x (0) has k eigenvalues with negative real parts and (n − k0) eigenvalues with positive
real parts.
Then
(a) There is a k -dimensional differentiable manifold S tangent to SA at 0 such that ϕt (S) ⊂
S, t ≥ 0, and x0 ∈ S implies that limt→∞ ϕt (x0 ) = 0.
(b) There is an (n − k) -dimensional differentiable manifold U tangent to UA at 0 such that
ϕt (U) ⊂ U, t ≤ 0, and x0 ∈ U implies that limt→−∞ ϕt (x0 ) = 0.

Recall that SA ( UA ) is the linear span of all generalized eigenvectors of A corresponding to

eigenvalues with negative (positive) real part.

Proof. Step 1. Set-up. Write

f (x) = Ax + F (x)
∂f ∂F
where A = ∂x (0) and F (x) = f (x) − Ax so that F (0) = 0, ∂x (0) = 0. If ² > 0, there exists
δ > 0 such that
|x| ≤ δ, |y| ≤ δ =⇒ |F (x) − F (y)| ≤ ²|x − y|.

5
W.l.o.g., we assume that (after a suitable change of variables y=Px)
· ¸
P 0
A= ,
0 Q

and
P is k × k matrix with eigenvalues λ1 , · · · , λk such that Reλi < 0;
Q is (n − k) × (n − k) matrix with eigenvalues λ1 , · · · , λ(n−k) such that Reλi > 0.
Therefore
· Pt ¸
e 0
eAt = = U (t) + V (t)
0 eQt
· Pt ¸ · ¸
e 0 0 0
U (t) = V (t) =
0 0 0 eQt
U 0 = AU V 0 = AV

We may choose α, σ > 0 sufficiently small that

|U (t)| ≤ Ke−(α+σ)t , t ≥ 0;
σt
|V (t)| ≤ Ke , t ≤ 0.

Step 2: The integral equation. Consider the integral equation

Z t Z ∞
x(t) = U (t)c + U (t − s)F (x(s))ds − V (t − s)F (x(s))ds (7)
0 t

We can verify that any solution to the integral equation satisfies x0 = f (x). Use successive approx-
imation to solve the integral equation:

x0 (t) = 0
Z t Z ∞
xj+1 (t) = U (t)c + U (t − s)F (xj (s))ds − V (t − s)F (xj (s))ds
0 t

We have, if ²K < σ/4, and K|c| ≤ δ,

|x1 (t) − x0 (t)| = |U (t)c| ≤ Ke−αt |c|

K
|xj+1 (t) − xj (t)| ≤ min{ j e−αt |c|, δ}, j ≥ 1, (by induction)
2
This implies
p−1
X p−1
X
−αt 1
|xp (t) − xq (t)| ≤ |xj+1 (t) − xj (t)| ≤ K|c|e , q < p.
2j
j=q j=q

Therefore, for each |c| ≤ δ/K, {xj (t)} uniformly converges to a solution x(t) of the integral
equation (7), and therefore to a solution x(t) of the ODE x0 = f (x). Furthermore, x(t) satisfies

|x(t)| ≤ 2K|c|e−αt , t≥0 (8)

6
if |c| ≤ δ/K.

Step 3: the stable manifold S. From the form of U (t) and the integral equation (7), we observe
that x(t) depends only on c1 , · · · , ck . We can thus set ck+1 = · · · = cn = 0. Furthermore, x(t)
satisfies the following initial condition

xi (0) = ci , i = 1, · · · , k
Z ∞
xi (0) = −( V (−s)F (x(s))ds)i (9)
0
= Ψi (c1 , · · · , ck ), i = k + 1, · · · , n.

Properties of the function Ψ : Rk → Rn−k :

(a) Ψ(0) = 0 (from (8))

(b) Ψ is continuous (since x(t) is continuous)

(c) Ψ is C 1 (see Coddington and Levinson)

∂Ψ
(d) ∂c (0) = 0 (see Coddington and Levinson)

Therefore the set, the stable manifold,

δ
S = {(c1 , · · · , ck , Ψk+1 (c1 , · · · , ck ), · · · , Ψn (c1 , · · · , ck )) : |(c1 , · · · , ck )| < }
2K
is the graph of a C 1 function Ψ and thus is a C 1 manifold. Also, ∂Ψ ∂c (0) = 0 implies that the
tangent space of S at the origin is the c1 · · · ck plane, i.e. SA . It remains to show

(1) solution x(t) of (7) such that |x(t)| ≤ δ is unique.

(2) c ∈ S implies ϕt (c) → 0 as t → ∞.

(3) S is positively invariant.

In fact (2) follows from (8). To show (1), suppose y(t) is another such solution, then
·Z t Z ∞ ¸
−(α+σ)(t−s) σ(t−s)
|x(t) − y(t)| ≤ ²K e |x(s) − y(s)|ds + e |x(s) − y(s)|ds .
0 t

Thus
2²K
sup |x(t) − y(t)| ≤ sup |x(t) − y(t)|,
t≥0 σ t≥0
which implies that x(t) = y(t) since 2²K/σ < 1/2.
To show (3), we prove the following:

a solution x(t) = ϕt (c) of x0 = f (x) satisfies |x(t)| ≤ δ ⇐⇒ c ∈ S.

7
By the variation of constant formula
Z t
At
ϕt (c) = e c + eA(t−s) F (ϕs (c))ds
0
Z t
= [U (t) + V (t)]c + [U (t − s) + V (t − s)]F (ϕs (c))ds (10)
0
Z t Z ∞
= U (t)c + V (t)c + U (t − s)F (ϕs (c))ds − V (t − s)F (ϕs (c))ds
0 t

where Z ∞
c=c+ V (−s)F (ϕs (c))ds.
0
This integral exists if |ϕt (c)| ≤ δ. All the terms in (10) are bounded for t ≥ 0 except possibly
V (t)c, which is bounded if and only if ci = 0, i = k + 1, · · · , n. Therefore

|ϕt (c)| ≤ δ t≥0

⇐⇒ V (t)c = 0
⇐⇒ ϕt (c) is a solution of (7)
⇐⇒ c ∈ S.

This establishes the existence of the stable manifold S. To establish the existence of the unstable
manifold U, use the same procedure for t ≤ 0. ¤

The stable and unstable manifolds obtained in the above theorem are local in nature. One could
extend the stable manifold backwards in time and the unstable manifold forward in time to obtain
global manifolds
S = ∪t≤0 ϕt (S) U = ∪t≥0 ϕt (U).
However, these global manifolds may intersect each other and may also fail to be an embedded
manifold in Rn , see the following example.

Example 4. Consider
x0 = −x − y 2
(11)
y0 = y + x2 .
We have · ¸ · 2¸
−1 0 −y
A= F (x, y) =
0 1 x2
and · −t ¸ · ¸ · ¸
e 0 0 0 c
U (t) = V (t) = , c= 1 .
0 0 0 et 0
To find S, solve the following integral equation
Z t
−t
x(t) = e c1 − e−(t−s) y 2 (s)ds
Z ∞ 0
y(t) = − e(t−s) x2 (s)ds
t

8
by successive approximation. Set

x0 (t) = 0
y0 (t) = 0

we get

x1 (t) = e−t c1
y1 (t) = 0

and
Z t
−t
x2 (t) = e c1 + e−(t−s) 0 ds = e−t c1
0
Z ∞
1
y2 (t) = − e(t−s) e−2s c21 ds = − e−2t c21
t 3
and
Z
c2 t −(t−s) −4s 1
x3 (t) = e−t c1 − 1 e e ds = e−t c1 + (e−4t − e−t )c41
9 0 27
Z ∞
2 1
y3 (t) = −c1 e(t−s) e−2s ds = − e−2t c21
t 3
One can show that
|(x4 (t), y4 (t)) − (x3 (t), y3 (t))| = O(c51 )
and
1
Ψ2 (c1 ) = − c21 + O(c51 ),
3
and thus, near (0, 0),
1
S = {(x, y) : y = Ψ2 (x) = − x2 + O(x5 )}.
3
Similarly,
1
Ψ1 (c2 ) = − c22 + O(c52 ),
3
and thus, near (0, 0),
1
U = {(x, y) : x = Ψ1 (y) = − y 2 + O(y 5 )}.
3
Follow the local stable (unstable) manifold along the flow backward (forward) in time, we get
the global stable and unstable manifolds. Observe that (a) both global manifolds fail to be a
embedded manifold in R2 at (0, 0); (b) the global stable and unstable manifolds coincide and
form a homoclinic orbit of (0, 0).

9
4.3. Center manifold theorem
In the case when the x = 0 is a nonhyperbolic equilibrium, the center space CA of the linearized
system y 0 = Ay, where A = ∂f
∂x (0), is the generalized eigenspace corresponding to those eigenvalues
λ of A with Reλ = 0.

Example 5. The phase portrait of

· 0¸ · ¸· ¸
y1 0 0 y1
0 =
y2 0 −1 y2
shows that CA is the y1 axis, which consists of equilibria.

Example 6. The phase portrait of

 0    
y1 0 −1 0 y1
y20  = 1 0 0 y2 
0
y3 0 0 −1 y3
shows that CA is the y1 y2 plane, which consists of periodic orbits.

Theorem 4.3 (Center Manifold Theorem) Suppose that f ∈ C 1 (D → Rn ) and that

∂f
(1) f (0) = 0, A = ∂x (0).

(2) A has k eigenvalues λ with Reλ < 0, l eigenvalues λ with Reλ > 0, and m = n − k − l
eigenvalues λ with Reλ = 0.
Then there is a k -dimensional stable manifold S, a l -dimensional unstable manifold U as before,
and an m -dimensional invariant manifold C in a neighborhood of 0, which is tangent to CA at
0.

Remark. C contains all equilibria and periodic orbits near 0. C is not necessarily unique, see the
following example.

Example 7. The center manifold of the following system is not unique to the left of the x2 axis.
x01 = x21
x02 = −x2

10
4.4. Flows near a periodic orbit
Consider a nonlinear ODE
x0 = f (x) (12)
where f ∈ C 1 (D → Rn ) and D ⊂ Rn is open. Let x = p(t) be a non constant periodic solution
of period ω, and γ = {p(t) : 0 ≤ t < ω} be its orbit. First of all, we make an observation that
the notion of Lyapunov stability is not appropriate for periodic solutions, since the phase portrait
of a nonlinear pendulum
x00 (t) + sin x = 0
consists of concentric periodic orbits and each of them should be considered as “stable.” However,
each periodic orbit has different period, unlike the linear pendulum (e.g. x00 + x = 0 ). Therefore,
for two periodic orbits that are close to each other, |p(t) − q(t)| will not be small for all t. This
calls for a different notion of stability for periodic solutions.

Definitions. x = P (t) is said to be orbitally stable if ∀ ² > 0, ∃ δ > 0 such that |x0 − p(0)| < δ
implies d(ϕt (x0 ), γ) < ², where d(x, γ) = miny∈γ |x − y| is the distance from x to γ.
x = p(t) is orbitally asymptotically stable if (a) it is orbitally stable and (b) ∃ δ1 > 0 such that
|x0 − p(0)| < δ1 implies d(ϕt (x0 ), γ) → 0 as t → ∞.
The stability of periodic solutions and the flow near a periodic orbit can be studied by that of
a fixed point of a discrete flow, the Poincaré map, or the first return map. Let x0 = p(0), and
x(t, x0 ) = p(t) = ϕt (x0 ), p(t + ω) = p(t) for all t. Then f (x0 ) 6= 0 since x0 is not an equilibrium.

Proposition 4.4 Let

Σ = {x : (x − x0 )∗ · f (x0 ) = 0}
be the Poincaré section. Then ∃ δ > 0 and a unique function x 7→ τ (x) ∈ R, which is C 1 on
B(x0 , δ) such that τ (x0 ) = ω and

ϕτ (x) (x) ∈ Σ, ∀x ∈ B(x0 , δ).

Proof. Consider
F (t, x) = (ϕt (x) − x0 )∗ · f (x0 ).
Then F ∈ C 1 (R × D), F (ω, x0 ) = 0, and

∂F (ω, x0 )
= f (x0 )∗ · f (x0 ) = kf (x0 )k2 6= 0.
∂t
Implicit Function Theorem implies that ∃ δ > 0 and a τ ∈ C 1 (B(x0 , δ) → R) such that τ (x0 ) = ω
and F (τ (x), x) = 0, x ∈ B(x0 , δ). Therefore ϕτ (x) (x) ∈ Σ. ¤

Definition. The Poincaré map P : Σ → Σ is defined as

P (x) = ϕτ (x) (x) x ∈ Σ.

Remarks.

11
 (1) f ∈ C r =⇒ P ∈ C r .

(2) Fixed points of P m corresponds to periodic solutions of (12) of period close to m

ome . In particular, P (x0 ) = x0 if only if x(t, x0 ) is an ω -periodic solution.

(3) The periodic orbit γ = {p(t) : 0 ≤ t < ω} is o.a.s. if and only if the fixed point x0 = p(0)
is a.s. under the discrete dynamical system P n , and this is the case if all eigenvalues λ of
DP (x0 ) , the derivative of P at x0 , satisfy |λ| < 1.

(4) More generally, the stable, unstable, and the center manifolds of P n at the fixed point x0
determines the stable, unstable, and center manifolds of the periodic orbit γ.

(5) The periodic system

∂f
y 0 (t) =(p(t))y(t), p(t) = ϕt (x0 ) (13)
∂x
has a Floquet multiplier 1 since y = p0 (t) is a nonzero ω -periodic solution. The remaining
n − 1 Floquet multipliers are the eigenvalues of DP (x0 ) and thus determine the dimension
of the stable, unstable, and center manifolds of γ.

To see (5), recall that any f.m. Y (t) of (13) satisfies Y (t + ω) = Y (t)Y (ω). If Y (t) = ∂ϕ
∂x (x0 ),
t

then Y (0) = I and Y (ω) = ∂ϕ 0 00 ∂f 0

∂x (x0 ). Since p (t) = f (p(t)), p (t) = ∂x (p(t)p (t), so that y =
ω

0 0 0
p (t) satisfies (13) with y(0) = p (0) = f (p(0)) = f (x0 ) and p (t) is ω -periodic. Moreover,
p0 (t) = Y (t)p0 (0) = ∂ϕ 0 ∂ϕω
∂x (x0 )p (0) implies f (x0 ) = ∂x (x0 )f (x0 ). Therefore, 1 is an eigenvalue of
t

∂ϕω
Y (ω) = ∂x (x0 ) with eigenvector f (x0 ) = p0 (0), the tangent vector of p(t) at x0 = p(0).

Theorem 4.5 (Stable Manifold Theorem for Periodic Orbits) Suppose that (13) has k Flo-
quet multipliers λ with |λ| < 1, and l = n − k − 1 Floquet multipliers λ with |λ| > 1. Then there
exists a positively (negatively) invariant C 1 manifold S ( U ), such that x ∈ S ( x ∈ U ) implies

d (ϕt (x), γ) → 0, t→∞ (t → −∞).

dim S = k + 1 and dim U = l + 1(= n − k).

The following corollary is immediate.

Corollary 4.6 If (13) has n − 1 Floquet multipliers λ with |λ| < 1, then γ is o.a.s..

Proof. First of all, we choose a coordinate system in Rn so that x0 = 0, en = f (x0 ) and

e1 , · · · , en−1 span Σ. we have
∂ϕτ ∂ϕτ ∂τ ∂ϕτ
(x) = + ,
∂x ∂τ ∂x ∂x

12
and thus
∂ϕτ ∂ϕτ ∂τ ∂ϕω ∂τ
(x0 ) = (x0 ) (x0 ) + (x0 ) = f (x0 ) (x0 ) + V
∂x ∂t ∂x  ∂x ∂x
0 ··· 0  
 .. ..  0
 .   .
= .  + v1 · · · vn−1 ..  .
 0 ··· 0 
∂τ ∂τ 1
∂x1 · · · ∂xn

The matrix of ∂P
∂x (x0 ) is the (n − 1) × (n − 1) matrix obtained by deleting the last row and column
∂ϕτ
from ∂x (x0 ). Its eigenvalues are the remaining (n−1) eigenvalues of V = Y (ω), from our special
coordinate system and P = ϕτ |Σ . ¤

Remark. The stability of periodic orbits is characterized by the Floquet multipliers of (13). The
estimation of the (n − 1) Floquet multipliers can be extremely difficult, especially when n > 2.

In the special case when n = 2, the following stability criterion is due to Poincaré.
Theorem 4.7 (Poincaré’s Stability Condition) Assume that n = 2. A periodic solution p(t)
is o.a.s. if Z ω
divf (p(t)) dt < 0. (14)
0

Proof. Let λ1 = 1, λ2 be the two Floquet multipliers, eigenvalues of ∂ϕ ∂x (x0 ). By the Liouville’s
ω

formula,
∂ϕω Rω
λ2 = λ1 λ2 = det (x0 ) = e 0 divf (p(t))dt .
∂x
Therefore, Poincaré’s condition implies 0 < λ2 < 1, and hence the orbital asymptotic stability of
p(t), by Corollary4.6. ¤

Remark. For higher dimensional generalizations of Poincaré’s stability condition, see Muldowney,
Rocky Mountain J. Math. (1990), and Li and Muldowney, Canadian Appl. Math. Quart. (1998).

Example 8. Consider the Liénard Equation

x00 + f (x)x0 + g(x) = 0. (15)
A special case is the van der Pol equation
x00 + µ(1 − x2 )x0 + x = 0. (16)
Here f, g are assumed to be Lipschitz to ensure uniqueness of solutions. We introduce
Z x Z x
F (x) = f (x)dx, G(x) = g(x)dx. (17)
0 0
We convert equation (15) to a first order system by setting y = x0 + F (x). Then
x0 = −F (x) + y (18)
0
y = −g(x). (19)
We prove the following result (see Hale, p.221).

13
Theorem 4.8 Suppose that

(1) ∃x1 < 0 < x2 such that f (x) < 0, x1 < x < x2 ; f (x) > 0, x < x1 or x > x2 .

(2) xg(x) > 0, x 6= 0.

(3) G(x1 ) = G(x2 ) = h.

Then, any periodic solution of (15) is orbitally asymptotically stable.

Remark. Under Theorem 4.8, if a periodic orbit exists for (15), then it must be unique, since any
such orbit must surround the origin and two adjacent periodic orbits can not be both orbitally
asymptotically stable (Why?). (Question: how can one speak of adjacent periodic orbits? how do
we know there are not infinitely many?)

Proof. W first prove the following two claims.

Claim 1: x = 0 is the only solution such that x0 (t) = x00 (t) = 0, for some t. To see this,
suppose x0 (t) = x00 (t) = 0, and x(t) = x. Then g(x) = 0 (from (15)), and thus x = x(t) = 0.
( g(x) = 0 ⇐⇒ x = 0 ) This implies that x(t) satisfies an initial condition x(t) = x0 (t) = 0, hence
x(t) is the zero solution by the uniqueness.
Let E(x, x0 ) = 12 (x0 )2 + G(x) (the total energy). Let x(t) be a nonzero solution. Define
E(t) = E(x(t), x0 (t)). Then

E 0 (t) = x0 x00 + g(x)x0 = −f (x)x02 (by (15)) = 2f (x)(G(x) − E(x)) (by definition of E ).

Claim 2: x0 (t1 ) 6= 0 if E(t) has a local extrema at t1 . Otherwise, x0 (t1 ) = 0 and x00 (t1 ) 6= 0 (from
Claim 1). Hence, x(t) has a strict local extrema at t1 , and thus x(t) < x(t1 ) (or x(t) > x(t1 ) )
for t near t1 , t 6= t1 . Therefore f (x(t)) is of one sign near t1 , and thus E 0 (t) = −f (x)x02 is of
one sign near t1 , E(t1 ) can not be a local extrema. The contradiction establishes the claim.
Now, along a periodic solution x(t), E(t) assumes its local minimum at t = t1 . Therefore,
0 = E 0 (t) = −f (x(t1 ))x0 (t1 )2 , and Claim 2 implies that f (x(t1 )) = 0, and thus x(t1 ) = x1 or x2 .
Hence, G(x(t1 )) = G(x1 ) = h. This implies E(t) ≥ E(t1 ) = 21 x0 (t1 )2 + G(x(t1 )) ≥ 21 x0 (t1 )2 + h > h
for all t. Integrating the following identity

E0 2f (G − E) G−h
+ 2f = + 2f = 2f ,
E−h E−h E−h
we get Z Z
ω ω
G−h
f (x(t))dt = f dt > 0,
0 0 E−h

14
since E > h and f (G − h) > 0 ( G = h ⇐⇒ f = 0 ). Therefore, the second Floquet multiplier
Rω
λ = exp− 0 f (x(t))dt

satisfies 0 < λ < 1. By the Poincaré’s condition, x(t) is o.a.s. ¤

Theorem 4.9 (Bendixson’s criterion) Assume n = 2. Let D ⊂ R2 be a simply connected region.

If
div f (x) < 0, (or > 0 ) x ∈ D, (20)
then no periodic orbits can lie entirely in D.

Proof. Let x = (u, v) and f (x) = (P (u, v), Q(u, v)). Suppose a periodic orbit γ = {(u(t), v(t)) :
0 ≤ t < ω} ⊂ D. Let G be the region enclosed in the interior of γ. Under the right orientation of
γ, Green’s Theorem implies
ZZ I Z ω Z ω
∂P ∂Q
0> ( + )dudv = (P dv − Qdu) = (P v 0 − Qu0 )dt = (P Q − P Q)dt = 0,
G ∂u ∂v γ 0 0

and the contradiction establishes the theorem. ¤

Remarks. (a) Bendixson’s condition requires that div(f ) does not change sign in D.
(b) From the proof, we can see that Bendixson’s condition rules out periodic orbits, homoclinic
orbits, and heteroclinic cycles.
Exercise. Construct an example to show that the Bendixson’s criterion as stated in Theorem4.9
is no longer true in Rn for n > 2, namely, negative divergence condition may not be able to rule
out periodic orbits in dimensions higher than 2.
However, suitable conditions (other than the divergence condition) can be derived in higher
dimensions that rule out periodic orbits. For recent research on higher dimensional generalization
of Theorem4.9, see Li and Muldowney, J. Differential Equations 106 (1993), 27-39.
Exercise. Show that if the Bendixson’s condition holds in an annular region D ∈ R2 , then D
can contain at most one periodic orbit. Can you generalize this result?

Corollary 4.10 (Dulac’s Criteria) Assume that n = 2 and D ⊂ R2 is simply connected. If there
is a scalar-valued function α(x) such that

div (αf )(x) < 0 (or > 0 ) x ∈ D,

then D contains no periodic orbits.

Proof. Follow the same proof of Theorem4.9, apply the Green’s Theorem to (αP, αQ). ¤

Definitions. Let ϕt be the flow generated by x0 = f (x).

(i) The ω -limit set ω (x0 ) of x0 is defined as ω (x0 ) = ∩T ≥0 ∪t≥T ϕt (x0 ). Equivalently, y ∈
ω (x0 ) ⇐⇒ ∃ tn → ∞, such that x(tn ) → y, as n → ∞.

15
 (ii) The α -limit set α (x0 ) of x0 is defined as α (x0 ) = ∩T ≥0 ∪t≥T ϕ−t (x0 ). Equivalently, y ∈
α (x0 ) ⇐⇒ ∃ tn → −∞, such that x(tn ) → y, as n → ∞.

Properties of Limit Sets

(a) If ϕt (x0 ) is bounded for t ≥ 0 ( t ≤ 0 ), then ω (x0 ) ( α (x0 ) ) is nonempty and compact.

(b) Both ω(x0 ) and α(x0 ) are invariant and connected.

(c) If ω(x0 ) = {x}, a singleton, then limt→∞ x(t) = x.

Proof. Exercise.
Limit sets can be very complicated, as in the case of Lorenz attractor. Characterization of all
limit sets is a primary goal of the qualitative theory of differential equations. For two-dimensional
systems, we have the following result, which also holds on compact two-dimensional manifolds.

Theorem 4.11 (Poincaré-Bendixson Theorem) Assume that n = 2. Let L be a nonempty

compact limit set (of x0 = f (x) ). If L contains no equilibria, then L is a periodic orbit.

Proof. See Hale, Hartman.

Example 9. (An SIR Model in Mathematical Epidemiology) The following system of differential
equations describes the spread of an infectious disease in a host population, by investigating the
dynamical transfer of populations among the three compartments: susceptible ( S ), infectious ( I )
and recovered ( R ):

S 0 = b − λIS − bS + δR
I 0 = λIS − (b + γ)I (21)
R0 = γI − (b + δ)R.

Here, S(t), I(t), and R(t) denote the size of subpopulations in each compartment at time t, and
b, α, δ, γ are positive constants. Our interest is in the limiting behaviours (as t → ∞ ) of solutions
(S(t), I(t), R(t)) for a given initial condition (S0 , I0 , R0 ).

The reduction of dimension: Adding all equations in (21), we obtain

(S + I + R)0 = b(1 − S − I − R),

and thus S + I + R = 1 + (S0 + I0 + R0 − 1) e−bt → 1 as t → ∞. Thus all ω -limit sets of (21)

must lie on the hyperplane S + I + R = 1. To investigate the limiting behaviours, one can simply
study the restriction of (21) on the hyperplane, namely assume that S + I + R = 1. Substitute
R = 1 − S − I into (21), we arrive at a two-dimensional system

S 0 = b − λIS − bS + δ(1 − S − I) (22)

0
I = λIS − (b + γ)I

16
We study this system in the following region
D = {(S, I) : 0 < S, I, S + I ≤ 1}
due to biological interpretation of S and I. One can verify that D is positively invariant for
system (22).

Equilibria: The system (22) has two equilibria: P0 = (1, 0), which corresponds to the case of no
disease and is called the disease-free equilibrium, and P ∗ = (S ∗ , I ∗ ), where
b+γ b+γ
S∗ = > 0, I∗ = (1 − S ∗ ) > 0.
λ b+γ+δ
P ∗ is called the endemic equilibrium since it corresponds to the case when the disease persists in
the population. Notice that P0 always lies on the boundary of D, while P ∗ ∈ D when λ > b + γ,
and P ∗ ∈/ D when λ < b + γ. Set
λ
R0 =
b+γ
∗
so that S = 1/R0 . Then the number of equilibria in D depends on the value of R0 : if R0 ≤ 1,
there exists only one equilibrium P0 in D, while there is an additional equilibrium P ∗ ∈ D if
R0 > 1. ( R0 is thus a bifurcation parameter).

Stability of P0 : One can linearize at P0 and show that both eigenvalues are real and negative if
R0 < 1, and one eigenvalue becomes positive if R0 > 1. This shows that P0 is locally asymptot-
ically stable if R0 < 1 and unstable if R0 > 1. It also implies that one eigenvalue will be 0 if
R0 = 1. In this case, P0 is not hyperbolic so the method of linearization does not apply. In the
following, we use a Lyapunov function to show the following result.

Theorem 4.12 If R0 ≤ 1, then P0 is asymptotically stable and all solutions in D converges to

P0 (in this case, we say P0 is globally asymptotically stable in D. If R0 > 1, then P0 is unstable.

Proof. Set V (S, I) = I. Then

V ∗ (t) = I 0 (t) = λIS − (b + γ)I = I[λS − (b + γ)] ≤ 0,
if R0 ≤ 1. This implies that P0 is locally stable. Since V (t) will strictly decrease as long as
V ∗ (t) < 0, we can conclude that all ω -limit sets must be contained in the set where V ∗ = 0,
namely in the set G = {(S, I) : I = 0 or S = 1(R0 = 1)}. The only possible candidate in G for
limit sets (compact invariant sets) is the singleton {P0 }. Therefore all solutions in D converge
to P0 . If R0 > 1, then V ∗ > 0 for I > 0 and S < 1 and (S, I) ∈ D sufficiently close to P0 ,
therefore, P0 is unstable. ¤

Stability of P ∗ when R0 > 1 : Now assume R0 > 1, so that P ∗ ∈ D. Linearize (22) at P ∗ , we

have the Jacobian matrix
· ¸
−b − δ − λI ∗ −λS ∗ − δ
J= .
λI ∗ λS ∗ − b − γ (= 0)
Since trJ = −b − δ − λI ∗ < 0 and detJ = λI ∗ (λS ∗ + δ) > 0, we know P ∗ is asymptotically stable,
if R0 > 1. In fact we have the following global stability result.

17
Theorem 4.13 If R0 > 1, then P ∗ is globally asymptotically stable in D.

Proof. It remains to show that all solutions in D converge to P ∗ , or equivalently, all ω -limit sets
in D is the singleton {P ∗ }. (In this case, we say that P ∗ attracts all points in D. ) We want to
show that a Dulac condition holds in D. Let (P (S, I), Q(S, I)) denote the right hand side of (22).
Choose the Dulac multiplier α(S, I) = 1/I. Then

∂(αP ) ∂(αQ) ∂(P/I) ∂(Q/I) ∂P 1 ∂(λS − b − γ) ∂P 1 −b − δ − λI

+ = + = + = = <0
∂S ∂I ∂S ∂I ∂S I ∂I ∂S I I
for all (S, I) ∈ D. Therefore, no periodic solutions can exist. Poincaré-Bendixson Theorem then
implies that all ω -limit sets in D must contain the only equilibrium P ∗ , and therefore be equal to
the singleton {P ∗ } since P ∗ is locally asymptotically stable. (Question: why wouldn’t an ω -limit
set contain the boundary equilibrium P0 ? ) ¤

18