SMSTC (2024/25)

Dynamical Systems and Conservation Laws

www.smstc.ac.uk

Contents

2 Stable, Unstable and Centre Manifolds 2–1

2.1 Stable and Unstable Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–1
2.1.1 Stable and unstable eigenspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–1
2.1.2 Stable and unstable manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–3
2.2 The Hartman Grobman Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–11
2.3 Steps for drawing phase portraits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–14
2.4 Centre Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–17
2.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–20

(i)
 SMSTC (2024/25)
Dynamical Systems and Conservation Laws
Chapter 2: Stable, Unstable and Centre Manifolds
Lucia Scardia
Original notes: Jack Carr HWU

www.smstc.ac.uk

Contents
2.1 Stable and Unstable Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . 2–1
2.1.1 Stable and unstable eigenspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–1
2.1.2 Stable and unstable manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–3
2.2 The Hartman Grobman Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 2–11
2.3 Steps for drawing phase portraits . . . . . . . . . . . . . . . . . . . . . . . . . 2–14
2.4 Centre Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–17
2.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–20

2.1 Stable and Unstable Manifolds

We now study stability for autonomous systems of the general form x0 (t) = f (x(t)). The first step, as
usual will be to look at the linearised system x0 (t) = Df (xe )x(t).
So, we start by analysing linear systems.
Let xe be an equilibrium of the linear system x0 (t) = Jx(t), where J ∈ Rn×n .

2.1.1 Stable and unstable eigenspaces

We first consider linear systems.
Warm-up case. We start with the simplest case, of a hyperbolic equilibrium, where in addition Df (xe )
has n, distinct, real eigenvalues λ1 , . . . , λn ∈ R \ {0}, and corresponding eigenvectors v1 , . . . , vn ∈ Rn .
Then we define the stable eigenspace E s and the unstable eigenspace E u of the linear system of the linear
system x0 (t) = Df (xe )x(t) as follows:

E s = span{vj : λj < 0} ⊆ Rn ,
E u = span{vj : λj > 0} ⊆ Rn .

General case. Let λ1 , . . . , λn ∈ C for j = 1, . . . , n, and let vj = uj + iwj ∈ Cn , for j = 1, . . . , n be the

corresponding (possibly generalised) eigenvectors. Note that, if λj ∈ R, then vj = uj ∈ Rn . We assume
(with no loss of generality) that there are m conjugate pairs of eigenvalues, and n − 2m real eigenvalues,
so that the eigenvectors are
u1 ± iw1 , . . . , um ± iwm , u2m+1 , . . . , un ,
following the convention of Remark 1.3.14 in Chapter 1. By the same remark we have that

spanR {u1 , w1 , . . . , um , wm , u2m+1 , . . . , un } = Rn . (2.1)

In this case we define

E s = span{uj , wj : <λj < 0} ⊆ Rn ,

E u = span{uj , wj : <λj > 0} ⊆ Rn .

2–1
SMST C: Dynamical Systems and Conservation Laws 2–2

Namely, E s and E u are the subspaces of Rn spanned by the real and imaginary parts of the generalised
eigenvectors vj corresponding to eigenvalues λj with negative or positive real parts, respectively.
If the equilibrium is not hyperbolic, there is the additional subspace
E c = span{uj , wj : <λj = 0} ⊆ Rn ,
spanned by the real and imaginary parts of the generalised eigenvectors vj corresponding to eigenvalues
λj with zero real parts. This is called the centre subspace for the linear system.
By definition and by (2.1) we have that
Rn = E s ⊕ E u ⊕ E c .

Moreover, we have that

lim φt (y) = 0 for every y ∈ E s , and lim φt (y) = 0 for every y ∈ E u .
t→+∞ t→−∞

In other words, trajectories starting from E s reach eventually the equilibrium of the linear system (the
origin), in the forward direction of time, while trajectories starting from E u reach the equilibrium of the
linear system (the origin) in the backward direction of time.

Example 2.1.1 Consider the linear system

 
−2 −1 0
x0 (t) = Jx(t), J = 1 −2 0 .
0 0 3

Find E s , E u and E c .
  
1 0
The eigenvalues of J are λ1,2 = −2 ± i, with eigenvectors  ∓i , and λ3 = 3, with eigenvector  0 .
0 1
Hence
   
 1 0 
Es = span  0  ,  1  = {x = (x1 , x2 , x3 ) ∈ R3 : x3 = 0},
0 0
 
 
0
Eu = span  0  = {x = (x1 , x2 , x3 ) ∈ R3 : x1 = x2 = 0},
1
E c = ∅.
To compute the flow, we compute the solutions of the system:
 −2t
−e−2t sin t
    
e cos t 0
 e−2t sin t  ,  e−2t cos t  ,  0 ,
0 0 e3t
which give the matrix solution
e−2t cos t −e−2t sin t
 
0
(eJt =)  e −2t
sin te −2t
cos t 0 ,
3t
0 0 e
reducing to the identity matrix for t = 0. Hence the flow of the system is
e−2t cos t −e−2t sin t 0
 

φt =  e−2t sin te−2t cos t 0 .

0 0 e3t

Let now y ∈ E s , namely y = (y1 , y2 , 0); then

y1 e−2t cos t − y2 e−2t sin t
     
y1 0
φt  y2  =  y1 e−2t sin t + y2 e−2t cos t  →  0  as t → +∞,
0 0 0
SMST C: Dynamical Systems and Conservation Laws 2–3

while if y ∈ E u , namely y = (0, 0, y3 ), then

     
0 0 0
φt  0  =  0  →  0  as t → −∞.
y3 y3 e3t 0

The stable, unstable and centre subspaces satisfy the following property.

Theorem 2.1 Let J ∈ Rn×n , and let E s , E u and E c be defined as above. Then

etJ E s ⊂ E s , etJ E u ⊂ E u , and etJ E c ⊂ E c for every t ∈ R.

The theorem above states that the stable, unstable and center subspaces of the linear system x0 (t) = Jx(t)
are ‘invariant’ under the flow eJt : Rn → Rn for every t ∈ R.

2.1.2 Stable and unstable manifolds

We can define the ‘nonlinear’ versions of the stable and unstable subspaces E s and E u . In other words,
we will show that near a hyperbolic equilibrium xe , the nonlinear system x0 (t) = f (x(t)) has stable and
unstable manifolds, tangent at xe to the stable and unstable subspaces E s and E u of the linearised system
x0 (t) = Df (xe )x(t).
The precise result is stated in Theorem 2.2. To simplify the exposition, we will from now on assume that
xe = 0, since one can always reduce to this case by means of an affine change of variables, namely by
replacing x with x + xe .

Theorem 2.2 (The Stable Manifold Theorem) Let A ⊂ Rn be an open set containing the origin,
let f ∈ C 1 (A; Rn ), and let φt be the flow of the nonlinear system x0 (t) = f (x(t)). Suppose that the origin
is a hyperbolic equilibrium, and that Df (0) has k eigenvalues with negative real part and n − k eigenvalues
with positive real part.
s
Then there exists a k-dimensional differentiable manifold Wloc (0) tangent to the stable subspace E s of the
linear system x0 (t) = Df (0)x(t) at the origin, and such that for every t ≥ 0 φt (Wloc s s
(0)) ⊂ Wloc (0), and
s
for all y ∈ Wloc (0)
lim φt (y) = 0;
t→+∞
u
and there exists a (n − k)-dimensional differentiable manifold Wloc (0) tangent to the unstable subspace E u
0 u u
of the linear system x (t) = Df (0)x(t) at the origin, and such that for every t ≤ 0 φt (Wloc (0)) ⊂ Wloc (0),
u
and for all y ∈ Wloc (0)
lim φt (y) = 0.
t→−∞

It can be shown that, in a small neighbourhood N of a hyperbolic critical point at the origin, xe = 0, the
local stable and unstable manifolds are given by
s
Wloc (0) = {y ∈ N : φ(t, y) → 0 as t → +∞, φ(t, y) ∈ N for all t ≥ 0} (2.2)
u
Wloc (0) = {y ∈ N : φ(t, y) → 0 as t → −∞, φ(t, y) ∈ N for all t ≤ 0}. (2.3)

The stable manifold theorem is illustrated in Figure 2.1.

Remark 2.1.2 (What is a manifold?) First we need to recall what is a homeomorphism.

Let A, B ⊂ X, where X is a metric space. A homeomorphism of A onto B is a continuous, one-to-one
map h : A → B onto B, such that h−1 : B → A is also continuous. The sets A and B are homeomorphic,
or topologically equivalent, if there is a homeomorphism of A onto B.
An n-dimensional differentiable manifold M is a connected metric space with an open covering {Uα },
namely M = ∪α Uα , such that

(1) For every α there exists hα : Uα → B1 (0) homeomorphism, where B1 (0) denotes the open unit ball
in Rn ;
(2) If Uα ∩ Uβ 6= ∅, then hα (Uα ∩ Uβ ) and hβ (Uα ∩ Uβ ) are subsets of Rn , the map h := hα ◦ h−1
β is
differentiable, and the determinant det Dh(x) 6= 0 for every x ∈ hβ (Uα ∩ Uβ ).
SMST C: Dynamical Systems and Conservation Laws 2–4

u
Wloc

Eu
s
E
s
Wloc

s u
Figure 2.1: Sketch of stable Wloc (xe ) and unstable Wloc (xe ) manifolds and stable E s and unstable E u
eigenspaces local to an equilibrium.

The pair (Uα , hα ) is called a chart.

Proof [of Theorem 2.2] We only give a brief sketch. The main idea is to rewrite the system x0 (t) = f (x(t))
as x0 (t) = Jx(t) + F (x(t)), with J = Df (0) and F (x) = f (x) − Jx. We can assume for simplicity that J
can be written in the form  
P 0
J= ,
0 Q
where P is a k×k-matrix whose eigenvalues λ1 , . . . , λk have negative real part, and Q is an (n−k)×(n−k)-
matrix whose eigenvalues λk+1 , . . . , λn have positive real part. If not, we will have to resort to a change
of variables.
The main aim of the proof is to show the existence of n − k differentiable functions ψj , j = k + 1, . . . , n,
(x1 , . . . , xk ) 7→ ψj (x1 , . . . , xk ) such that the equations
xj = ψj (x1 , . . . , xk ), j = k + 1, . . . , n
define a k-dimensional differentiable manifold, which will be the stable manifold. This step is done by
following the method of successive approximations. More precisely, the strategy is the following:
(1) Define
eP t
   
0 0 0
U (t) = , V (t) = .
0 0 0 eQt
Then U 0 (t) = JU (t), V 0 (t) = JV (t), and eBt = U (t) + V (t).
(2) Rewrite the system x0 (t) = Jx(t) + F (x(t)) in integral form. Namely solve for
Z t Z ∞
u(t, z) = U (t)z + U (t − s)F (u(s, z))ds − V (t − s)F (u(s, z))ds. (2.4)
0 t

Note that if u is a solution of the integral equation, it is also a solution of the system x0 (t) =
Jx(t) + F (x(t)).
(3) Solve the integral equation by successive approximations, by setting
Z t Z ∞
u(0) (t, z) = 0, u(j+1) (t, z) = U (t)z + U (t − s)F (u(j) (s, z))ds − V (t − s)F (u(j) (s, z))ds.
0 t
(2.5)
One can prove that
lim u(j) (t, z) = u(t, z)
j→∞

uniformly for t ≥ 0 and |z| small (depending on F , U and V ), and that u satisfies the integral
equation and hence the original system.
SMST C: Dynamical Systems and Conservation Laws 2–5

(4) By setting
ψj (z1 , . . . , zk ) := uj (0, z1 , . . . , zk , 0, . . . , 0), j = k + 1, . . . , n, (2.6)
one can see that the equations
xj = ψj (x1 , . . . , xk ), j = k + 1, . . . , n.
define a differentiable manifold S, which enjoys the following property: If x is a solution of x0 (t) =
Jx(t) + F (x(t)) with x(0) ∈ S, then x(t) ∈ S for every t ≥ 0. Moreover,
∂ψj
(0) = 0, i = 1, . . . , k, j = k + 1, . . . , n,
∂xi
hence S is tangent at 0 to the stable subspace E s = {y ∈ Rn : y1 , . . . , yk = 0} of the linear system
y 0 (t) = Jy(t).
The existence of an unstable manifold can be done in the same way, by constructing the stable manifold
for the different system x0 (t) = −Jx(t) − F (x(t)). This will be the unstable manifold of the original
system. 

Example 2.1.3 Consider the nonlinear system


0
x1 (t) = −x1 (t),

x02 (t) = −x2 (t) + x21 (t),
 0

x3 (t) = x3 (t) + x21 (t).
The only equilibrium point of the system is at the origin, namely xe = 0. The matrix J = Df (0) is given
by  
−1 0 0
Df (0) =  0 −1 0  ,
0 0 1
   
1 0
which has eigenvalues λ = −1 with multiplicity a = 2 and corresponding eigenvectors  0  and  1 ,
0 0
 
0
and λ = 1, simple, with eigenvector  0 . The linearised system x0 (t) = Jx(t) has stable and unstable
1
subspaces given by
   
 1 0 
E s = span  0  ,  1  = {x = (x1 , x2 , x3 ) ∈ R3 : x3 = 0},
0 0
 
 
0
E u = span  0  = {x = (x1 , x2 , x3 ) ∈ R3 : x1 = x2 = 0}.
1
Now we compute the stable and unstable manifolds of the original system. To this aim we compute the
flow of the nonlinear system.
Note that, once the first equation of the system (linear and homogeneous) is solved, by substituting the
solution into the other two, they become linear inhomogeneous equations, which can also be solved explic-
itly.
More precisely, the ODE x01 (t) = −x1 (t) has general solution satisfying x1 (0) = c1 given by x1 (t) = c1 e−t ,
with c1 ∈ R generic constant.
Substituting in the second ODE gives x02 (t) = −x2 (t) + c21 e−2t . The general solution of this equation,
under the condition that x2 (0) = c2 , is x2 (t) = c2 e−t − c21 e−2t + c21 e−t , for c2 ∈ R generic constant.
Substituting in the third ODE gives x02 (t) = x3 (t) + c21 e−2t . The general solution of this equation, coupled
with x3 (0) = c3 , is x3 (t) = c3 et + 31 c21 (et − e−2t ), for c3 ∈ R generic constant. Hence, the general solution
of the nonlinear system is
c1 e−t
 

x(t) =  c2 e−t − c21 e−2t + c21 e−t  ,

c3 et + 13 c21 (et − e−2t )
SMST C: Dynamical Systems and Conservation Laws 2–6

and the flow is

c1 e−t
 

φt (c) =  c2 e−t − c21 e−2t + c21 e−t  , c = (c1 , c2 , c3 ). (2.7)

c3 et + 31 c21 (et − e−2t )
s
From (2.2) we know that, if c ∈ Wloc (0), then φ(t, c) → 0 as t → +∞, hence we impose that condition in
the expression of the flow (2.7). Note that
 
0
lim φ(t, c) = 0 ⇔ lim  0
 = 0,
t→+∞ t→+∞
c3 + 13 c21 et

since all the other terms are infinitesimal. So, this is only possible if
1
c3 + c21 = 0.
3
This condition defines a surface in R3 (hence a two-dimensional manifold)
 
s 3 1 2
Wloc (0) = c ∈ R : c3 = − c1 .
3

This is a surface of the form c3 = g(c1 , c2 ), and for such surfaces, the normal vector to the tangent plane
at 0 is given by    
∂g ∂g
n = 1, 0, (0) ∧ 0, 1, (0) = (0, 0, 1).
∂c1 ∂c2
The tangent plane at zero is then the plane orthogonal to (0, 0, 1), namely E s .
u
Similarly, from (2.2) we know that, if c ∈ Wloc (0), then φ(t, c) → 0 as t → −∞, hence we impose that
condition in the expression of the flow (2.7). Note that

c1 e−t
 

lim φ(t, c) = 0 ⇔ lim  c2 e−t − c21 e−2t + c21 e−t  ,

t→−∞ t→−∞
− 31 c21 e−2t

since all the other terms are infinitesimal. So, this is only possible if

c1 = 0 = c2 = 0.

This condition defines a line in R3 (hence a one-dimensional manifold) in the x3 -direction,

u
(0) = c ∈ R3 : c1 = c2 = 0 .

Wloc
u
Since Wloc (0) = E u , clearly they have the same dimension, and clearly E u is tangent to Wloc
u
(0) at 0,
and in fact at every point.

We show by means of an example, how we can find approximations to stable and unstable manifolds.

Example 2.1.4 Consider the nonlinear system

(
x01 (t) = −x1 (t) − x22 (t)
(2.8)
x02 (t) = x2 (t) + x21 (t).

Find the approximations u(0) , u(1) , u(2) , u(3) of the solution as defined in (2.5), and use u(3) to approxi-
mate the function ψ2 describing the stable manifold
s
Wloc (0) : x2 = ψ2 (x1 ).

We follow the procedure illustrated in the proof of Theorem 2.2. Note that the system can be written as
x0 (t) = f (x(t)), with f (x1 , x2 ) = (−x1 − x22 , x2 + x21 ), and that
   
−1 −2x2 −1 0
J = Df (0) = (x = 0) = .
2x1 1 0 1
SMST C: Dynamical Systems and Conservation Laws 2–7

Then we identify P = −1 and Q = 1. We do not need a change of variables since P < 0 and Q > 0 (here
k = 1). Moreover, F (x1 , x2 ) = f (x1 , x2 ) − Jx = (−x22 , x21 ). We then set
 −t   
e 0 0 0
U (t) = , V (t) = .
0 0 0 et

We are now ready to build the iteration. We set u(0) (t, z) := 0, and have that
 −t   Z t  −t+s   Z ∞  
(1) e 0 z1 e 0 0 0 0 0
u (t, z) = + ds − t−s ds
0 0 z2 0 0 0 0 t 0 e 0
 −t 
e z1
= ,
0

where we have used that F (u(0) (s, z)) = 0. We iterate further and have
 −t  Z t  −t+s   Z ∞  
(2) e z1 e 0 0 0 0 0
u (t, z) = + −2s 2 ds − ds
0 0 0 e z1 0 et−s e−2s z12
 −t  Z0 ∞    −t
t
e z1 0 e z1
= − t−3s 2 ds = 1 −2t 2 ,
0 t e z 1 − 3e z1

where we have used that F (u(1) (s, z)) = F (e−s z1 , 0) = (0, e−2s z12 ). Finally,
 −t  Z t  −t+s   1 −4s 4  Z ∞
− 91 e−4s z14
 
e z1 e 0 − 9 e z1 0 0
u(3) (t, z) = + −2s 2 ds − ds
0 0 0 0 e z1 t 0 et−s e−2s z12
 −t  Z t  1 −t−3s 4  Z ∞ 
e z1 −9e z1 0
= + ds − t−3s 2 ds
0 0 0 t e z1
 −t 1 4 −4t
− e−t )

e z1 + 27 z1 (e
= ,
− 31 e−2t z12

where we have used that F (u(2) (s, z)) = F (e−s z1 , − 31 e−2s z12 ) = (− 19 e−4s z14 , e−2s z12 ).
By (2.6) (with u replaced by the approximation u(3) , k = 1 and n = 2), we write
(3) (3) 1
ψ2 (x1 ) = u2 (0, x1 , 0) = − x21 .
3
Hence we approximate
1
ψ2 (x1 ) ∼ − x21 ,
3
and
s 1
Wloc (0) ∼ x2 = − x21 .
3
For the unstable manifold, we find it by constructing the stable manifold for the different system x0 (t) =
−Jx(t) − F (x(t)), namely for the system
(
x01 (t) = x1 (t) + x22 (t)
(2.9)
x02 (t) = −x2 (t) − x21 (t).

But (2.9) is exactly the same as (2.8) with x1 and x2 swapped. So we can follow the procedure for (2.8)
and conclude that the stable manifold for (2.9) is

s 1
Wloc (0) ∼ x1 = − x22 .
3
Hence the unstable manifold of (2.8) is

u 1
Wloc (0) ∼ x1 = − x22 .
3
We can use the method above also to provide an approximate solution for the stable and unstable manifold
in Example 2.1.5. The advantage is that the approximation method only requires being able to solve a
linear system, while the exact construction of the manifold required being able to solve the full system
(which was doable in the Example but generally could be hard).
SMST C: Dynamical Systems and Conservation Laws 2–8

Example 2.1.5 Consider the nonlinear system


0
x1 (t) = −x1 (t),

x02 (t) = −x2 (t) + x21 (t),
 0

x3 (t) = x3 (t) + x21 (t).

Find the approximations u(0) , u(1) and u(2) , u(3) of the solution as defined in (2.5), and use u(2) to
approximate the stable manifold.
The system can be written as x0 (t) = f (x(t)), with f (x1 , x2 ) = (−x1 , −x2 + x21 , x3 + x21 ), and that
   
−1 0 0 −1 0 0
J = Df (0) =  2x1 1 0  (x = 0) =  0 −1 0  .
2x1 0 1 0 0 1
 
−1 0
Then we identify P = and Q = 1. We do not need a change of variables since P has
0 −1
negative eigenvalues and Q > 0 (here k = 2). Moreover, F (x1 , x2 , x3 ) = f (x1 , x2 , x3 ) − Jx = (0, x21 , x21 ).
We then set  −t   
e 0 0 0 0 0
U (t) =  0 e−t 0  , V (t) =  0 0 0  .
0 0 0 0 0 et
Recall that eP t , when P is a matrix, is the fundamental matrix solution Y (t) (2 × 2) of the linear system
Y 0 (t) = P Y (t), complemented with the initial condition given by the identity matrix, namely Y (0) = Id.
IN other words, the columns of the matrix eP t = Y (t) are the two linearly independent solutions of the
system y 0 (t) = P y(t). One can see that P has a repeated eigenvalue λ = −1, with corresponding linearly
independent eigenvectors e.g. e1 = (1, 0) and e2 = (0, 1), which leads to
 −t 
Pt e 0
e = Y (t) = , Y (0) = Id,
0 e−t

and hence to the definition of U (t).

We are now ready to build the iteration. We set u(0) (t, z) := 0, and have that
 −t    −t 
e 0 0 z1 e z1
u(1) (t, z) =  0 e−t 0   z2  =  e−t z2  ,
0 0 0 z3 0

where we have used that F (u(0) (s, z)) = 0. We iterate further and have
 −t   −t+s  
e z1 Z t e 0 0 0
u(2) (t, z) =  e−t z2  +  0 e−t+s 0   e−2s z12  ds
0 0 0 0 0 e−2s z12
e−t z1
    
Z ∞ 0 0 0 0
−  0 0 0   e−2s z12  ds =  e−t z2 − e−2t z12 + e−t z12  ,
t 0 0 e t−s
e−2s z12 − 31 e−2t z12

where we have used that F (u(1) (s, z)) = F (e−s z1 , F (e−s z2 , 0) = (0, −e−2s z12 , −e−2s z12 ).
By (2.6) (with u replaced by the approximation u(2) , k = 2 and n = 3), we write

(2) (2) 1
ψ3 (x1 , x2 ) = u3 (0, x1 , x2 , 0) = − x21 .
3
Hence we approximate
1
ψ3 (x1 , x2 ) ∼ − x21 ,
3
and
s 1
Wloc (0) ∼ x3 = − x21 .
3
SMST C: Dynamical Systems and Conservation Laws 2–9

Example 2.1.6 Consider the nonlinear system

(
x01 (t) = −x1 (t) + x22 (t)
x02 (t) = x2 (t) + x22 (t) − x21 (t).

Find the approximations u(0) , u(1) , u(2) , u(3) of the solution as defined in (2.5), and use u(3) to approxi-
mate the function ψ2 describing the stable manifold
s
Wloc (0) : x2 = ψ2 (x1 ).

We follow the procedure illustrated in the proof of Theorem 2.2. Note that the system can be written as
x0 (t) = f (x(t)), with f (x1 , x2 ) = (−x1 + x22 , x2 + x22 − x21 ), and that
   
−1 2x2 −1 0
J = Df (0) = (x = 0) = .
−2x1 1 + 2x2 0 1

Then we identify P = −1 and Q = 1. We do not need a change of variables since P < 0 and Q > 0 (here
k = 1). Moreover, F (x1 , x2 ) = f (x1 , x2 ) − Jx = (x22 , x22 − x21 ). We then set
 −t   
e 0 0 0
U (t) = , V (t) = .
0 0 0 et

We are now ready to build the iteration. We set u(0) (t, z) := 0, and have that
 −t   Z t  −t+s   Z ∞  
e 0 z1 e 0 0 0 0 0
u(1) (t, z) = + ds − ds
0 0 z2 0 0 0 0 t 0 et−s 0
 −t 
e z1
= ,
0

where we have used that F (u(0) (s, z)) = 0. We iterate further and have
 −t  Z t  −t+s   Z ∞  
e z1 e 0 0 0 0 0
u(2) (t, z) = + −2s 2 ds − ds
0 0 0 −e z 1 0 et−s −e−2s z12
 −t  Z0 ∞    −t
t
e z1 0 e z1
= − t−3s 2 ds = 1 −2t 2 ,
0 t −e z 1 3e z1

where we have used that F (u(1) (s, z)) = F (e−s z1 , 0) = (0, −e−2s z12 ). Finally,
 −t  Z t  −t+s  1 −4s 4

(3) e z1 e 0 9 e z1
u (t, z) = + 1 −4s 4 ds
0 0 0 0 9e z1 − e−2s z12
Z ∞  1 −4s 4

0 0 9e z1
− 1 −4s 4 ds
t 0 e t−s
9 e z1 − e−2s z12
 −t  Z t  1 −t−3s 4  Z ∞ 
e z1 e z1 0
= + 9 ds − 1 t−5s 4 ds
0 0 0 t 9e z1 − et−3s z12
!
e−t z1 − 271 4 −4t
z1 (e − e−t )
= 1 −4t 4 1 −2t 2
,
− 45 e z1 + 3 e z1

where we have used that F (u(2) (s, z)) = F (e−s z1 , 13 e−2s z12 ) = ( 19 e−4s z14 , 19 e−4s z14 − e−2s z12 ).
By (2.6) (with u replaced by the approximation u(3) , k = 1 and n = 2), we write

(3) (3) 1 2 1
ψ2 (x1 ) = u2 (0, x1 , 0) = x1 − x41 .
3 45
Hence we approximate
1 2 1
ψ2 (x1 ) ∼ x1 − x41 ,
3 45
and
S 1 2 1
Wloc (0) ∼ x2 = x − x4 .
3 1 45 1
SMST C: Dynamical Systems and Conservation Laws 2–10

For the unstable manifold, we find it by constructing the stable manifold for the different system x0 (t) =
−Jx(t) − F (x(t)), namely for the system
(
x01 (t) = x1 (t) − x22 (t)
x02 (t) = −x2 (t) − x22 (t) + x21 (t).

To put it in the standard form, we set y1 := x2 and y2 := x1 , so that the system for y reads
(
y10 (t) = −y1 (t) − y12 (t) + y22 (t),
y20 (t) = y2 (t) − y12 (t).

The system can be written as y 0 (t) = f (y(t)), with f (y1 , y2 ) = (−y1 − y12 + y22 , y2 − y12 ), and that
   
−1 − 2y1 2y2 −1 0
J = Df (0) = (x = 0) = .
−2y1 1 0 1
Then we identify P = −1 and Q = 1, and F (y1 , y2 ) = f (y1 , y2 ) − Jy = (−y12 + y22 , −y12 ). We then set
 −t   
e 0 0 0
U (t) = , V (t) = .
0 0 0 et

We are now ready to build the iteration. We set u(0) (t, z) := 0, and have that
 −t   Z t  −t+s   Z ∞  
(1) e 0 z1 e 0 0 0 0 0
u (t, z) = + ds − ds
0 0 z2 0 0 0 0 t 0 et−s 0
 −t 
e z1
= ,
0

where we have used that F (u(0) (s, z)) = 0. We iterate further and have
 −t  Z t  −t+s Z ∞
−e−2s z12 −e−2s z12
   
e z1 e 0 0 0
u(2) (t, z) = + −2s 2 ds − ds
0 0 0 0 −e z1 t 0 et−s −e−2s z12
 −t  Z t  ∞
−e−t−s z12
 Z  
e z1 0
= + ds − ds
0 0 0 t −et−3s z12
 −t
e (z1 − z12 ) + z12 e−2t

= 1 −2t 2 ,
3e z1

where we have used that F (u(1) (s, z)) = F (e−s z1 , 0) = (−e−2s z12 , −e−2s z12 ). Finally,
 −t  Z t 
−e−t−s (z1 − z12 )2 − 2e−t−2s z12 (z1 − z12 ) − 98 e−t−3s z14

e z1
u(3) (t, z) = + ds
0 0 0
Z ∞ 
0
− ds
t −et−3s (z1 − z12 )2 − z14 et−5s − 2et−4s z12 (z1 − z12 )
!
e−t z1 + 8 4 −4t
27 z1 (e − e−t ) + (z1 − z12 )2 (e−2t − e−t ) + z12 (z1 − z12 )(e−3t − e−t )
= 1 −2t
,
3e (z1 − z12 )2 + 51 e−4t z14 + 12 e−3t z12 (z1 − z12 )

where we have used that

1
F (u(2) (s, z)) = F (e−s (z1 − z12 ) + z12 e−2s , e−2s z12 )
 3 
−2s 2 2 −3s 2 2 8 −4s 4 −2s 2 2 4 −4s −3s 2 2
= −e (z1 − z1 ) − 2e z1 (z1 − z1 ) − e z1 , −e (z1 − z1 ) − z1 e − 2e z1 (z1 − z1 ) .
9
By (2.6) (with u replaced by the approximation u(3) , k = 1 and n = 2), we write
(3) (3) 1 1 1 1 1 1
ψ2 (y1 ) = u2 (0, y1 , 0) = (y1 − y12 )2 + y14 + y12 (y1 − y12 ) = y12 − y13 + y14 .
3 5 2 3 6 30
Going back to the x-variables via the change of variables x1 := y2 and x2 := y1 , we have that the unstable
manifold is approximated as
u 1 2 1 3 1
Wloc (0) ∼ x1 = x − x + x4 .
3 2 6 2 30 2
SMST C: Dynamical Systems and Conservation Laws 2–11

s u
While Wloc (0) and Wloc (0) are defined locally, in terms of a neighbourhood of the equilibrium, one can
define global stable and unstable manifolds as follows.

Definition 2.1.7 Let φt be the flow of the nonlinear system x0 (t) = f (x(t)). The global stable and
unstable manifolds of the system at the origin are defined by
[ [
W s (0) := s
(0) , W u (0) := u



φt Wloc φt Wloc (0) . (2.10)
t≤0 t≥0

It can be shown that W s (0) and W u (0) are unique, and invariant under the flow (namely φt (W s (0)) ⊂
W s (0) and φt (W u (0)) ⊂ W u (0) for every t ∈ R). Furthermore

lim φt (y) = 0, for every y ∈ W s (0), and lim φt (y) = 0, for every y ∈ W u (0).
t→+∞ t→−∞

2.2 The Hartman Grobman Theorem

We have already informally stated that, near a hyperbolic equilibrium xe , the nonlinear system x0 (t) =
f (f (t)) has the same structure as the linear system x0 (t) = Df (xe )x(t). We will now make this statement
more rigorous.
Again, to simplify the exposition, we will from now on assume that xe = 0, since one can always reduce
to this case by means of the affine change of variables, namely by replacing x with x + xe .

Definition 2.2.1 Two autonomous systems x0 (t) = f (x(t)) and x0 (t) = g(x(t)) are topologically equiv-
alent in a neighbourhood of the origin, if there is a homeomorphism h : U → V between neighbourhoods
U, V of the origin which maps trajectories of the first system in U to trajectories of the second system
in V and preserves their orientation in time (namely, if a trajectory of the first system in U is directed
from P to Q, then its image is directed from h(P ) to h(Q) in V ).

Remark 2.2.2 We comment briefly, with an example in the one-dimensional case, on why homeomor-
phisms are the natural object to connect ODEs with a similar qualitative behaviour (rather than more
regular maps).
Let f, g : R → R be such that f (0) = g(0) = 0, so that 0 is an equilibrium for both x0 (t) = f (x(t)) and
x0 (t) = g(x(t)). If we denote with φ(·, x0 ) the solution of x0 (t) = f (x(t)) and with ψ(·, x0 ) the solution
of x0 (t) = g(x(t)) satisfying x(0) = x0 , then the stated topological equivalence means that there exists an
invertible map h : U → V , continuous with its inverse, which takes solutions of the first equation (in U )
to solutions of the second (in V ), namely satisfies

h(φ(t, x0 )) = ψ(t, h(x0 )). (2.11)

It would be natural to demand that h and the inverse h−1 are C 1 functions, rather than just continuous.
We now show with an example that this is too restrictive. Take f (x) = −x and g(x) = −2x, for which
(2.11) becomes
h(e−t x0 ) = e−2t h(x0 ) .
Differentiating this with respect to x0 and putting x0 = 0 gives e−t h0 (0) = e−2t h0 (0). Since h is invertible
in U and 0 ∈ U , h0 (0) 6= 0 and we are left with the impossible requirement that e−t = e−2t .
We can however construct a homeomorphism, and hence the two ODEs are topologically equivalent. The
homeomorphism is in this case ‘globally’ defined, and is given by the map h : R → R defined as

 −x2 if x < 0,
h(x) = 0 if x = 0,
 2
x if x > 0.

Note that h fails to have a differentiable inverse at x = 0.

Example 2.2.3 As a concrete example, we show that the ODEs

x0 (t) = −3x(t) (2.12)

SMST C: Dynamical Systems and Conservation Laws 2–12

and
x0 (t) = 4 − 2x(t) (2.13)
are equivalent, by constructing a homeomorphism from R to R mapping the solutions of (2.12) to the
solutions of (2.13).
First of all, we observe that the ODEs have the same phase portrait: Indeed, each of them has just one
equilibrium point, xe = 0 for (2.12) and xe = 2 for (2.13). Moreover, both f (x) = −3x and g(x) = 4−2x,
are positive for x < xe and negative for x > xe .
We start by constructing a homeomorphism mapping from (0, +∞) onto (2, +∞). Since we want the
homeomorphism to map solutions onto solutions, it will be constructed in terms of the flows of the ODEs.
For x0 ∈ R, we denote with φ(t, x0 ) = x0 e−3t , for every t ∈ R, the solution of (2.12), and ψ(t, x0 ) =
2 + (x0 − 2)e−2t the solution of (2.13), both with initial condition at t = 0 given by x0 .
We define t̄(x0 ) by, e.g.,
φ(t̄(x0 ), x0 ) = 1, (2.14)
namely, for x0 > 0 we have
t̄(x0 ) = ln(x0 )1/3 .
Note that any A > xe = 0 would work at the right-hand side of (2.14).
Now we define the map h : (0, +∞) → (2, +∞), for x0 > 0, as

h(x0 ) := ψ(−t̄(x0 ), 3) = 2 + (x0 )2/3 .

In this case, instead of 3 we could choose any B > 2.

Similarly, we can construct a map from (−∞, 0) to (−∞, 2) in the same way by taking (for example)
A = −1 and B = 1. The complete map h : R → R is

 2 − x2/3 x<0
h(x) = 2 x=0
2 + x2/3 x>0


The map is clearly continuous. Moreover, its inverse h−1 : R → R is


 (2 − x)3/2 x<2
−1
h (x) = 0 x=2
(x − 2)3/2 x>2


which is also continuous. The map is a homeomorphism.

It remains to show that h maps the solutions of (2.12) to the solutions of (2.13), namely that

h(φ(t, x0 )) = ψ(t, h(x0 )). (2.15)

Note that, for x0 > 0 (and hence φ(t, x0 ) = x0 e−3t > 0)

2/3
h(φ(t, x0 )) = h(x0 e−3t ) = 2 + (x0 e−3t )2/3 = 2 + x0 e−2t

and on the other hand, again for x0 > 0,

2/3 2/3
h(x0 ) = 2 + x0 ⇔ x0 = h(x0 ) − 2,

so
h(φ(t, x0 )) = 2 + (h(x0 ) − 2)e−2t = ψ(t, h(x0 )).
This proves (2.11), and hence the topological equivalence of (2.12) and (2.13).

We can extrapolate from the previous example and produce a method for proving the topological equiv-
alence of ODEs.
SMST C: Dynamical Systems and Conservation Laws 2–13

Remark 2.2.4 Suppose that two scalar, autonomous ODEs x0 (t) = f (x(t)) and x0 (t) = g(x(t)) have the
same phase portraits with a finite number of equilibrium points.
For simplicity, we assume that there is only one equilibrium point for each equation, a, b ∈ R such that
f (a) = g(b) = 0. With no loss of generality, f (x) < 0 for x > a, g(x) < 0 for x > b. We show that
the ODEs are topologically equivalent by constructing a homeomorphism h : (a, ∞) → (b, ∞) that maps
solutions of x0 (t) = f (x(t)) to solutions of x0 (t) = g(x(t)). (Then one can do the same between (−∞, a)
and (−∞, b)).
Let φ(t, x0 ) be the solution of x0 (t) = f (x(t)) and ψ(t, x0 ) the solution of x0 (t) = g(x(t)) for initial datum
x0 . Fix A > a and B > b. For any x0 > a define t̄(x0 ) by φ(t̄(x0 ), x0 ) = A . We define h by

h(x0 ) = ψ(−t̄(x0 ), B) .

Then h is a homeomorphism mapping (a, +∞) to (b, +∞). We have to show that h maps solutions to
solutions, that is (2.11) holds. We have

h(φ(t, x0 )) = ψ(−t̄(φ(t, x0 )), B).

Since t̄(φ(t, x0 )) is defined by φ(t̄(φ(t, x0 )), φ(t, x0 )) = A , it is exactly the time taken by the flow φ to go
from φ(t, x0 ) to A. Also, the flow takes t̄(x0 ) to go from 0 to A (and it takes t to go from x0 to φ(t, x0 )).

t t̄(φ(t, x0 ))
flow
x0 φ(t, x0 ) A

t̄(x0 )

Figure 2.2: Sketching the progress of the flow φ.

In conclusion
t̄(φ(t, x0 )) = t̄(x0 ) − t,
and
h(φ(t, x0 )) = ψ(t − t̄(x0 ), B).
On the other hand,
ψ(t, h(x0 )) = ψ(t, ψ(−t̄(x0 ), B)) = ψ(t − t̄(x0 ), B).
This is because, in general,
ψ(t2 , ψ(t1 , B)) = ψ(t1 + t2 , B).
Hence (2.11) holds.

We are now ready to state the Hartman-Grobman Theorem.

Theorem 2.2.5 (The Hartman-Grobman Theorem) Let A ⊂ Rn be an open set containing the
origin, let f ∈ C 1 (A; Rn ), and let φt be the flow of the nonlinear system x0 (t) = f (x(t)). Suppose
that f (0) = 0, and that J = Df (0) has no eigenvalues with zero real part. Then the nonlinear system
x0 (t) = f (x(t)) is topologically equivalent to its linearisation x0 (t) = Jx(t). More precisely, there exists a
homeomorphism h : U → V between neighbourhoods U, V of the origin such that for each x0 ∈ U there is
an interval I0 ⊂ R containing 0 such that for all t ∈ I0

h(φt (x0 )) = eJt h(x0 ). (2.16)

Remark 2.2.6 Note that (2.16) is nothing but (2.11), where we have used that, for the linear system,
ψt = eJt , and so ψt (h(x0 )) = eJt h(x0 ).
SMST C: Dynamical Systems and Conservation Laws 2–14

2.3 Steps for drawing phase portraits

Here we illustrate how we can draw local phase diagrams and then use this (and some common sense)
to get a global picture, for fixed parameter values. The phase portrait might change as we change
parameters.
Recipe:

• find all the stationary points;

• linearise near each stationary point;
• draw phase portraits local to each stationary point;
• join up “local” phase portraits to get a global picture.

For the last point in the list, saddle points play a special role, namely equilibria for which the origin of
the corresponding linearised system is a saddle point (in two dimensions it means that the eigenvalues of
the linearised system have opposite sign). For saddle points of nonlinear systems the stable and unstable
manifolds are simple, and they are given by curves.

Theorem 2.3 Let xe be a hyperbolic equilibrium of the system x0 (t) = f (x(t)) and assume it is a saddle.
Then there exist two trajectories Γ1 and Γ2 which approach xe as t → +∞ and two trajectories Γ3 and
Γ4 which approach xe as t → −∞. These special trajectories are called separatrices.

The directions at which the trajectories approach the equilibrium xe are the same as the eigenvectors of
the linearised system.
Then for a saddle point xe we have that the stable manifold is Γ1 ∪ Γ2 ∪ {0}, while the unstable manifold
is Γ3 ∪ Γ4 ∪ {0}.

Example 2.3.1 We perform a phase-plane analysis of the Lotka-Volterra model of population dynamics
of two competing species:
(
r0 (t) = f (r(t), s(t)), f (r, s) = r(3 − r − 2s),
0
s (t) = g(r(t), s(t)), g(r, s) = s(2 − r − s).

In the system above, r and s represent the rabbit and the sheep populations, respectively. The model is
formulated under the assumption of logistic growth for each species in the absence of the other, while the
coupling models their competition for food.
We sketch the phase portrait of the system, namely R2 with the trajectories.
As a first step we compute the equilibria of the system. We do so by imposing that
(
r(3 − r − 2s) = 0
s(2 − r − s) = 0.

This gives four equilibrium points: (0, 0), (0, 2), (3, 0), (1, 1).
To study their stability we linearise the system around each equilibrium. We have that
!
∂f ∂f  
∂r ∂s 3 − 2r − 2s −2r
(r, s) = .
∂g ∂g −s 2 − r − 2s
∂r ∂s

• Equilibrium point xe = (0, 0). In this case

   
3 − 2r − 2s −2r 3 0
J= (0, 0) = .
−s 2 − r − 2s 0 2

The eigenvalues are λ1 = 2 and λ2 = 3, both positive, so (0, 0) is unstable.

Calling v1 and v2 the eigenvectors of J corresponding to λ1 and λ2 , respectively, we have that
trajectories leave the origin parallel to the ‘slow’ eigenvector v1 , which corresponds to the smaller
eigenvalue λ1 = 2. Indeed, for the linearised problem the general solution is x(t) = c1 eλ1 t v1 +
SMST C: Dynamical Systems and Conservation Laws 2–15

c2 eλ2 v2 . The trajectories escape from the equilibrium for t → +∞, so they will converge to it as
t → −∞. Since for t << 0 the solution x(t) ∼ c1 eλ1 t v1 , it follows that v1 is the direction of the
trajectory close to the equilibrium (oriented away from it). We can easily compute v1 = (0, 1) and
v2 = (1, 0).
• Equilibrium point xe = (0, 2). In this case
   
3 − 2r − 2s −2r −1 0
J= (0, 2) = .
−s 2 − r − 2s −2 −2

The eigenvalues are λ1 = −2 and λ2 = −1, both negative, so (0, 2) is stable.

Calling v1 and v2 the eigenvectors of J corresponding to λ1 and λ2 , respectively, we have that the
general solution of the linearised problem is x(t) = c1 eλ1 t v1 + c2 eλ2 v2 . The trajectories converge
to the equilibrium as t → +∞. Since for t >> 0 the solution x(t) ∼ c2 eλ2 t v2 , it follows that v2 is
the direction of the trajectory close to the equilibrium (oriented towards it). We then compute the
eigenvector corresponding to λ2 = −1 and obtain that v2 = (1, −2). Note that v1 = (0, 1).

• Equilibrium point xe = (3, 0). In this case

   
3 − 2r − 2s −2r −3 −6
J= (3, 0) = .
−s 2 − r − 2s 0 −1

The eigenvalues are λ1 = −3 and λ2 = −1, both negative, so (3, 0) is stable. As in the previous case,
v2 (the eigenvector to the larger eigenvalue) is the direction of the trajectory close to the equilibrium
(oriented towards it). We then compute the eigenvector corresponding to λ2 = −1 and obtain that
v2 = (3, −1). Note that v1 = (1, 0). In the bottom left, the trajectories of the linearised system are
drawn close to the equilibrium (parabolas in the first 3 cases, hyperbolas in the last case).
• Equilibrium point xe = (1, 1). In this case
   
3 − 2r − 2s −2r −1 −2
J= (1, 1) = .
−s 2 − r − 2s −1 −1
√ √
The eigenvalues are λ1 = −1 − 2 < 0 and λ2 = −1 + 2 > 0, of opposite sign, so (1, 1) is
unstable. Along v1 , which is the eigenvector corresponding to the positive eigenvalue, the directions
will be away from the equilibrium, since eλ1 t → +∞ as t → +∞. Along v1 , which is the eigenvector
corresponding to the negative eigenvalue λ√ 1 , the directions √
will be towards the equilibrium, since
eλ1 t → 0 as t → +∞. We have that v1 = ( 2, 1) and v2 = ( 2, −1).
SMST C: Dynamical Systems and Conservation Laws 2–16

5
2.5

4
2

3
1.5

2
s

s
1

1
0.5

−1
−1 0 1 2 3 4 5 0 0.5 1 1.5 2 2.5 3 3.5
r r
2.5 2.5
basin boundary
s
u

2 2

1.5 1.5 unstable stable manifold

manifold
s

1 1
unstable manifold

0.5 0.5
stable
manifold

s
0 0
u s

0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5

r r

Figure 2.3: In the top left figure the equilibria are plotted. In the top right, the eigenvectors of the
linearised system are added to the equilibria, with the orientation towards and away from the equilibria.
In the bottom left picture we used the knowledge about phase portraits for linear systems to draw the
trajectories of the nonlinear system locally around each equilibrium point. In the bottom right picture
we combine the local pictures into a ‘global’ one. This is beyond the scope of the course. In particular
we sketch separatrices for the saddle point, and the corresponding stable and unstable manifold.
SMST C: Dynamical Systems and Conservation Laws 2–17

2.4 Centre Manifolds

So far we have focused on the case of hyperbolic equilibria. In this section we assume that xe is not
hyperbolic. As usual, we can assume that xe = 0.
The first result of the section is essentially of the type of Theorem 2.2, and establishes the existence of a
centre manifold tangent to the centre subspace E c at the origin.

Theorem 2.4 (The Centre Manifold Theorem) Let A ⊂ Rn be an open set containing the origin,
let f ∈ C 1 (A; Rn ), and let φt be the flow of the nonlinear system x0 (t) = f (x(t)). Suppose that f (0) = 0,
and that Df (0) has s eigenvalues with negative real part, u eigenvalues with positive real part, and
c := n − s − u eigenvalues with zero real part.
c
Then there exists a c-dimensional differentiable manifold Wloc (0) tangent to the centre subspace E c of
0
the linear system x (t) = Df (0)x(t) at the origin, there exists an s-dimensional differentiable manifold
Wlocs
(0) tangent to the stable subspace E s of the linear system x0 (t) = Df (0)x(t) at the origin, and a
u
u-dimensional differentiable manifold Wloc (0) tangent to the unstable subspace E u of the linear system
0
x (t) = Df (0)x(t) at the origin, with c + s + u = n.
c s u
Furthermore, Wloc (0), Wloc (0) and Wloc (0) are invariant under the flow φt of the nonlinear system
0
x (t) = f (x(t)).

While the Hartman-Grobman Theorem completely solves the problem of determining the stability in the
vicinity of a hyperbolic critical point of a nonlinear system, the behaviour near a non-hyperbolic critical
point (namely in an n-dimensional neighbourhood) is determined by its behaviour on the centre manifold
(namely in a c-dimensional neighbourhood, generally with c < n). Since the centre manifold is generally
of smaller dimension than the system, this simplifies the stability analysis near non-hyperbolic critical
points.
To illustrate the idea, let us start by writing the nonlinear system x0 (t) = f (x(t)) in diagonal form, as

0
x (t) = Cx(t) + F (x(t), y(t), z(t)),

0
y (t) = P y(t) + G(x(t), y(t), z(t)), x = (x, y, z) ∈ Rc × Rs × Ru ,
 0

z (t) = Qz(t) + H(x(t), y(t), z(t)),

where C is a square matrix with c eigenvalues with zero real part, P has s eigenvalues with negative
real part, and Q has u eigenvalues with positive real part. Moreover, F (0) = G(0) = H(0) = 0, and
DF (0) = DG(0) = DH(0) = 0.
For simplicity, we assume that u = 0. Then we have the following result, proved by Carr.

Theorem 2.5 (The Local Centre Manifold Theorem) Let A ⊂ Rn be an open set containing the
origin, let f ∈ C r (A; Rn ), with r ≥ 1, and let φt be the flow of the nonlinear system x0 (t) = f (x(t)).
Suppose that f (0) = 0, and that Df (0) has s eigenvalues with negative real part, and c := n−s eigenvalues
with zero real part. Then the system can then be written in diagonal form as
(
x0 (t) = Cx(t) + F (x(t), y(t)),
(x, y) ∈ Rc × Rs , (2.17)
y 0 (t) = P y(t) + G(x(t), y(t)),

where C is a square matrix with c eigenvalues with zero real part, P has s eigenvalues with negative real
part, F (0) = G(0) = 0, and DF (0) = DG(0) = 0. Moreover, there exists δ > 0 and h ∈ C r (Nδ (0)) with
h(0) = 0, Dh(0) = 0, such that h describes the centre manifold as
c
Wloc (0) = {(x, y) ∈ Rc × Rs : y = h(x) for |x| < δ},

and satisfies
Dh(x)(Cx + F (x, h(x))) − P h(x) − G(x, h(x)) = 0, |x| < δ. (2.18)
c
Finally, the flow on the centre manifold W (0) is defined by the system of differential equations

x0 (t) = Cx(t) + F (x, h(x))

for all x ∈ Rc with |x| < δ.

SMST C: Dynamical Systems and Conservation Laws 2–18

Condition (2.18) comes from differentiating the equation y = h(x) defining the centre manifold with
respect to t, and using (2.17). In principle it gives a formula for h in terms of a PDE, but it is usually
impossible to solve for h. However, it gives a method for approximating the function h to any degree of
accuracy we wish, provided r is large enough. This is achieved by substituting series expansions for the
components of h into (2.18).

Example 2.4.1 Consider the system

(
x0 (t) = x2 (t)y(t) − x5 (t),
y 0 (t) = −y(t) + x2 (t).

Construct the approximate local centre manifold at the origin and determine the flow on W c (0) near the
origin.
Note that the system is of the form (2.17), with c = s = 1. In particular, we have that C = 0, P = −1,
F (x, y) = x2 y − x5 and G(x, y) = x2 .
To find the approximate expression of the centre manifold, we substitute the series approximation h(x) =
ax2 + bx3 + cx4 + O(x5 ) in (2.18), to obtain the (approximate) equality

(2ax + 3bx2 + 4cx3 ) x2 (ax2 + bx3 + cx4 ) − x5 + ax2 + bx3 + cx4 + O(x5 ) − x2 = 0.

By matching the coefficients of the terms of lower order (x2 , x3 and x4 ) we obtain a = 1, b = 0 and
c = 0, and hence
h(x) = x2 + O(x5 ).
The approximate centre manifold is then given by

W c (0) ∼ y = x2 + O(x5 ),

and the (approximate) flow on W c (0) is obtained from x0 (t) = x2 (t)y(t) − x5 (t), by replacing y with h(x),
namely is
x0 (t) = x4 (t) + O(x(t)5 ).
This is a scalar ODE, for which the stability analysis for the origin can be done by studying the sign of
the right-hand side. We conclude that the origin is unstable (a saddle).
Note that if we had brutally approximated W c (0) with E c from the start (note that E c = {y = 0}) we would
have got a different (incorrect) conclusion. Indeed the flow on E c is obtained from x0 (t) = x2 (t)y(t)−x5 (t),
by replacing y with 0, namely is
x0 (t) = −x5 (t),
for which the origin is a stable equilibrium. Hence, in presence of eigenvalues with zero real part, the
linear approximation is insufficient and the stability analysis requires a better approximation of the centre
manifold.

Example 2.4.2 Consider the system

(
x(t)0 = αx3 (t),
(2.19)
y 0 (t) = −y(t) + y(t)2 ,

where α ∈ R is a constant, α 6= 0. Since the equations are uncoupled, we see that the equilibrium point
x = y = 0 of (2.19) is asymptotically stable if and only if α < 0.
Consider now the system obtained by adding to (2.19) some coupling terms, namely
(
x0 (t) = αx3 (t) + x(t)y(t) − x(t)y 2 (t),
. (2.20)
y 0 (t) = −y(t) + βx2 (t) + x2 (t)y(t)

Since the equations are coupled we cannot immediately deduce the stability of the origin. We then proceed
by studying the centre manifold at the origin.
First note that (2.20) is of the form (2.17), with c = s = 1. In particular, we have that C = 0, P = −1,
F (x, y) = αx3 + xy − xy 2 and G(x, y) = βx2 + x2 y.
SMST C: Dynamical Systems and Conservation Laws 2–19

To find the approximate expression of the centre manifold, we substitute the series approximation h(x) =
ax2 + bx3 + cx4 + O(x5 ) in (2.18), to obtain the (approximate) equality

(2ax + 3bx2 ) αx3 + x(ax2 + bx3 ) − x(ax2 + bx3 )2 + ax2 + bx3 + cx4 − βx2 − x2 (ax2 + bx3 ) = 0.

By matching the coefficients of the terms of lower order (x2 , x3 and x4 ) we obtain a = β, b = 0 and
c = β − 2β 2 − 2αβ, and hence

h(x) = βx2 + (β − 2β 2 − 2αβ)x4 + O(x5 ).

The approximate centre manifold is then given by

W c (0) ∼ y = βx2 + (β − 2β 2 − 2αβ)x4 + O(x5 ),

and the (approximate) flow on W c (0) is obtained from x0 (t) = αx3 (t) + x(t)y(t) − x(t)y 2 (t), by replacing
y with h(x), namely is

2
x0 (t) = αx3 (t) + x(t) βx2 (t) + (β − 2β 2 − 2αβ)x4 (t) − x(t) βx2 (t) + (β − 2β 2 − 2αβ)x4 (t) + O(x(t)6 )

= (α + β)x3 (t) + (β − 2αβ − 3β 2 )x5 (t) + O(x(t)6 ).

This is a scalar ODE, for which the stability analysis for the origin can be done by studying the sign of
the right-hand side.
If α + β 6= 0, then the flow on the centre manifold can be approximated by

x0 (t) = (α + β)x3 (t) + O(x4 (t)).

Hence the zero solution of (2.20) is asymptotically stable if α + β < 0 and unstable if α + β > 0. If
α + β = 0 we need to use a better approximation to the invariant manifold in order to decide on the
stability. In particular, if α + β = 0 (and hence α = −β), the flow on the centre manifold can be
approximated by
x0 (t) = β(1 − β)x5 (t) + O(x(t)6 ).
Hence the zero solution of (2.20) is asymptotically stable if β < 0 or β > 1, and unstable if 0 < β < 1.
SMST C: Dynamical Systems and Conservation Laws 2–20

2.5 Exercises
2–1. Show that the scalar ODEs
x0 (t) = 3x(t) (2.21)
and
x0 (t) = x(t) + 1 (2.22)
are topologically equivalent.
2–2. Find the first three successive approximations u(1) , u(2) and u(3) for the system
(
x01 (t) = −x1 (t)
x02 (t) = x2 (t) + x31 (t),

and use u(3) to approximate the local stable manifold near the origin.
2–3. Solve the system below and find the local stable and unstable manifolds of the origin for the system
(
x01 (t) = −x1 (t)
x02 (t) = x2 (t) + x31 (t).

2–4. A model for cancer growth is given by the coupled system of ODEs
(
n0 (t) = n(t)(1 − n(t) − m(t)) − dn(t)
m0 (t) = bm(t)(1 − n(t) − m(t)) − dm(t) − δm(t),

where n denotes normal cells and m denotes mutant cells.

Roughly the parameter 0 < d < 1 is related to a natural death rate, δ is the rate at which the
immune system can kill off the mutant cells and b > 0 is related to how fast mutant cells divide
relative to normal cells.
We do not consider the case where bd = d + δ.

a) Show that if b − d − δ > 0 then the there are three equilibrium points where m ≥ 0 and n ≥ 0.
b) Determine the nature of the equilibrium points and sketch the phase portrait for b = 1, d = 21 ,
δ = 41 .
c) Determine the nature of the equilibrium points and sketch the phase portrait for b = 74 , d = 12 ,
δ = 41 .

Use the phase portraits from parts b) and c) to describe what happens if you start with a small
number of mutant cells m and a large number of normal cells n in each case.
2–5. Study the stability of the equilibrium point xe = (0, 0) for the system
(
x0 (t) = x(t)y(t) + αx(t)3 + βx(t)y(t)2 ,
(2.23)
y 0 (t) = −y(t) + γx2 (t) + δx2 (t)y(t)

by computing the approximate centre manifold and the (reduced) flow on it.
Show in particular that if α + γ = 0, then the stability of the origin depends on the sign of γδ + βγ 2 .