Dynamical Systems

A Concise Graduate Course

Christian Kuehn

Faculty of Mathematics,
Technical University of Munich,
85748 Garching b. München, Germany

October 15, 2018

1
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Preface
These notes are designed to give a concise introduction to dynamical sys-
tems with a view towards techniques for applications. Many monographs
and extended textbooks exist; see bibliography. The lectures are based
upon quite a number of sources [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 16,
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30]. The goal here is to
extract the key ideas for the course “Dynamical Systems” (4 weekly hours)
held first at TU Munich by the author in the summer semesters 2017 and
2018.
The course considered in these lecture notes is the third part of a five-
course series:

• Dynamical Systems Ia: Ordinary Differential Equations

• Dynamical Systems Ib: Introduction to Nonlinear Dynamics

• Dynamical Systems II: A Concise Graduate Course

• Dynamical Systems III: Numerical Methods

• Dynamical Systems IV: Basics of Structured Systems

Currently not all notes of the four courses are available but may be
updated regularly in the future.
Note carefully: The current version of the notes is a work in progress
and should be used with caution! Please report any errors or inaccuracies
you find to

ckuehn@ma.tum.de

2
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Contents
1 Definitions & Examples 4

2 Stability 9

3 Stable/Unstable Manifolds 15

4 Equivalence 23

5 Poincaré Maps 27

6 Structural Stability 31

7 Hyperbolic Sets 34

8 Local Bifurcations 39

9 Normal Forms 49

10 Center Manifolds 53

11 Higher Codimension 59

12 Global Bifurcations 62

13 The Smale Horseshoe 69

14 Melnikov’s Method 76

15 Chaos 81

16 One-Dimensional Maps 86

17 Renormalization 89

18 Attractors 93

19 Invariant Measures 99

20 Ergodicity 103

21 Mixing 108

22 Entropy 113

3
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 1 Definitions & Examples
Definition 1.1. A dynamical system is a triplet (X , T , φt ), where X is
the phase space (or state space), T is the time set (or time domain),
and φt : X → X for t ∈ T is the family of evolution operators satisfying

(D1) φ0 = Id,

(D2) φt ◦ φs = φt+s ∀t, s ∈ T .

See also Figure 1(a).

Example 1.2. Let X = R2 , T = R, A ∈ R2×2 and consider

∞
tA
X (tA)k
φt (x0 ) = e x0 = x0 for x0 ∈ R2 . (1.1)
k!
k=0

The example (1.1) arises from solving a linear ordinary differential

equation (ODE)

dx
= x′ = Ax, x = x(t), x(0) = x0
dt
by setting φt (x0 ) = x(t) = etA x0 . A very simple nontrivial example in R2
arises if A is a diagonal matrix with Aii 6= 0, which has a simple visualiza-
tion in the phase plane/space; see Figure 1(b). 

g(x)
x2
(a) (b) (c)

φt
x1 x2 x1 x0 x
X

Figure 1: (a) General dynamical system on X . (b) Linear ODE (1.1) with
diagonal matrix A ∈ R2×2 and A11 < 0, A22 > 0. (c) Iterated cubic
map (1.2) including cob-web diagram (dotted lines).

Example 1.3. Take X = R, T = N0 = N ∪ {0}, and consider the mapping

x3 =: g(x), g : R → R, (1.2)

4
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 and define

xk := φk (x0 ) := (g ◦ g ◦ · · · ◦ g)(x0 ) =: g k (x0 ), φ0 = Id, (1.3)

| {z }
k- times

where one has to be careful not to confuse superscripts, k-fold composition,

and powers. Since x 7→ x3 is invertible, we may extend the dynamical
system to T = Z. Simple visualization of the map is obtained by the
cobweb construction as shown in Figure 1(c). 
Definition 1.4. We call (X , T , φt ) a discrete-time dynamical system (or
simply an iterated map) if T has a bijection to N (e.g. T = N0 , Z). If
T is a connected subset of, or equal to, R, we call it a continuous-time
dynamical system.
Example 1.2 is continuous-time and Example 1.3 is discrete-time.
Definition 1.5. A dynamical system is called invertible if T is an additive
group and φt is defined for all t ∈ T , i.e., φ−t is a single-valued map and
φ−t ◦ φt = Id. For the continuous-time case we also speak of flows if T = R
and semiflows if T = R+ = [0, ∞).
Example 1.6. Consider the logistic map

xk+1 = pxk (1 − xk ) = g(xk ; p), p ∈ (0, 4], (1.4)

where we view p as a parameter, which one checks (exercise!) to be well-

defined; see Figure 2.
g(x) (a) g(x) (b)

x 0 1
1
x
2

Figure 2: (a) Sketch of the logistic map (1.4) g(x; p) =: g(x) for p > 2. (b)
Logistic map; draw in I as discussed in the text (exercise!).

However, there exists an open interval I ⊂ [0, 1] such that pre-image

g −1 (I) consists of two disconnected components. Indeed, we have
r
−1 1 1 y
y = g(x; p) = px(1 − x) ⇒ x = g (y) = ± −
2 4 p

5
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 so the inverse is multi-valued. Multi-valued (or set-valued) dynamical
systems will not be discussed here. 
Recall the following classical theorem from a first ODE course:
Theorem 1.7. (Existence, uniqueness and smooth dependence) Let f :
Rd → Rd be smooth and consider the ODEs

x′ = f (x), x = x(t) ∈ Rd , x(0) = x0 . (1.5)

Then there exists a unique map (t, x0 ) 7→ φt (x0 ), smooth in (t, x0 ) such that
(T1) φ0 (x0 ) = x0 ,
(T2) there exists a (maximal) interval T = (−t1 , t2 ), t1,2 = t1,2 (x0 ) > 0
such that for all t ∈ T , φt (x0 ) = x(t), i.e., the ODE (1.5) is satisfied.
Recall that the proof proceeds via re-writing the ODEs in integral form,
defining a suitable norm, and using the Banach fixed-point theorem; see [2,
10].
Theorem 1.8. (Banach Fixed Point Theorem) Let (X , d) be a non-
empty complete metric space and assume g : X → X is a contraction
mapping

d(g(x), g(y)) ≤ λd(x, y), for some λ ∈ [0, 1).

Then g admits a unique fixed point g(x∗ ) = x∗ .

For the ODEs (1.5), if T = R, we have a flow.
Example 1.9. Not all ODEs define flows, just consider x′ = x2 with x ∈ R.
A standard calculation using separation of variables shows that T =
6 R
Z x(t) Z t
1
dy = ds
x0 y2 0

so we get the solution

1
φt (x0 ) = x(t) = 1 ,
x0 −t

which blows up in finite-time. However, we may consider φt as a local

flow on an open time interval. 
The differentiability of φt (x0 ) in Theorem 1.7 directly corresponds to
the differentiability of the input vector field f : Rd → Rd . We always
q with the usual metric induced by the Euclidean norm |x| := |x|2 :=
work
x21 + · · · + x2d for ODEs.

6
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Definition 1.10. Suppose (Rd , R, φt ) is a dynamical system. If φt (x) is
smooth in (t, x), then we say (X , T , φt ), is a smooth dynamical system.
Remark : Similar definitions can be stated for C k (i.e., k-times differentiable,
k ≥ 1), or just continuous C 0 dynamical systems. Usually we shall assume that
the dynamical system is at least C 0 ; however, many of our examples are actually
even smooth.
Our key goal is to analyze and classify possible dynamics.

Definition 1.11. An orbit (or trajectory) of a dynamical system starting

at x0 is the ordered set

Or(x0 ) := {x ∈ X : x = φt (x0 ) for some t ∈ T such that φt (x0 ) is defined}.

Definition 1.12. A point x∗ ∈ X is a called a steady state, equilibrium

or fixed point if φt (x∗ ) = x∗ for all t ∈ T , i.e., the orbit Or(x0 ) is a
singleton.

Fixed point is the more common name for discrete-time systems/maps

while equilibrium point is used for continuous-time systems/flows.

Example 1.13. Let p > 0 be a parameter. Consider the planar van der
Pol equation
3
x′ = y − x3 + x,
(1.6)
y ′ = −px,
An equilibrium of (1.6) must satisfy y ′ = 0 so x = 0, which implies using
y

2
3

1 x
−1

− 32

Figure 3: Phase space R2 = X of (1.6) including both nullclines as thick

curves; cf. (1.7). Using the nullclines, it is relatively straightforward to
sketch the flow (exercise!).

the first equation for x′ that also y = 0. Therefore, (x∗ , y∗ ) = (0, 0) is the
unique equilibrium point; see Figure 3. 

7
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Definition 1.14. If we consider an ODE x′ = f (x), then the curves

{x ∈ Rd : fk (x) = 0}, k ∈ {1, 2, . . . , d} (1.7)

are called nullclines.

At an equilibrium point of an ODEs, all nullclines intersect. A steady

state is a particular case of the following class of sets.

Definition 1.15. A set S ⊂ X is called invariant for φt if x0 ∈ S implies

φt (x0 ) ∈ S for all t ∈ T . S is called positively invariant if x0 ∈ S at
t = 0 implies φt (x0 ) ∈ S for all t > 0; similarly one defines negatively
invariant; see Figure 4(b).

Example 1.16. Let T = Z and X = Σ2 be the space of all bi-infinite

sequences of two symbols, say {0, 1}. A point ω ∈ Σ2 is given by

ω = {. . . , ω−2 , ω−1 , ω0 , ω1 , ω2 , . . .}

or just written as
ω = · · · ω−2 ω−1 ω0 ω1 ω2 · · · .
Define the (one-step, left-)shift map g : Σ2 → Σ2 by

g(ω)j = ωj+1 , (1.8)

i.e., we shift each entry one position to the left. Clearly, we have that

ω = · · · 1111 · · · and ω = · · · 0000 · · ·

are the only two fixed points. However, there are a lot more invariant sets,
which are not fixed points. 

Definition 1.17. A periodic orbit (or cycle) is a nonequilibrium orbit

γ, such that each x0 ∈ γ satisfies

φt+T0 (x0 ) = φt (x0 )

for some fixed T0 > 0 and all t ∈ T ; see also Figure 4(a). We call the
minimal such T0 > 0 the period of γ.

Example 1.18. Consider a circle X = S1 ≃ [0, 1]/(0 ∼ 1), T = N0 and

define the (rational) circle rotation map given by

xk+1 = g(xk ) := (xk + r) mod 1,

8
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 x2
(a) γ (b)

R2 = X
x2 x1
S
x1

Figure 4: (a) Sketch of a periodic orbit for a planar ODE. (b) Sketch of a
positively invariant set S.

where we assume r ∈ Q, r = pq , gcd(p, q) = 1. Then every point x0 ∈ X

lies in some periodic orbit. Indeed, consider
 
p
g q (x0 ) = x0 + q mod 1 = x0 .
q
In fact, each point in X has period q but there are infinitely many distinct
periodic orbits. 
Definition 1.19. The phase portrait of a dynamical systems is the par-
titioning of X into its orbits.
Ideally, we would be given a dynamical system and provide a full un-
derstanding of its phase portrait. Philosophy:

• ≈ 17th-century: Given an ODE, compute its solutions! (“Newton”)

• ≈ 19th-century: Given an ODE, describe its phase portrait! (“Poincaré”)
• during 20th-century: Given no equation, describe its phase portrait!
(→ modern dynamical systems)

2 Stability
Definition 2.1. Let x∗ be a steady state for (X , T , φt ). Then x∗ is called
(Lyapunov) stable if for any given neighbourhood U = U (x∗ ) there exists
another neighbourhood V(x∗ ) = V ⊆ U such that

φt (V) ⊆ U for all t ≥ 0.

If stability does not hold, we say x∗ is unstable. x∗ is called asymp-

totically stable if it is stable and there exists a neighbourhood U such
that
lim φt (x0 ) = x∗ ∀x0 ∈ U .
t→+∞

9
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 From now on, we shall always assume that X has a metric. Since most
examples are related to Rd with Euclidean norm | · |, we just state this case
in our results; generalizations are presented when necessary. For example,
asymptotic stability means

lim |φt (x0 ) − x∗ | = 0 ∀x0 ∈ U .

t→+∞

Furthermore, we use global asymptotic stability to indicate U = X and

employ local asymptotic stability to emphasize that U may have to be
chosen small enough.
Example 2.2. Consider the harmonic oscillator

d2 x
+ x = x′′ + x = 0,
dt2
which can be re-written as a first-order system, setting x′ = y,
 ′   
x 0 1 x
′ = . (2.1)
y −1 0 y

The general solution is given by

(x(t), y(t)) = (c1 cos t + c2 sin t, c2 cos t − c1 sin t)

for constants c1,2 ∈ R. One checks that (x, y) = (0, 0) is an equilibrium

point, which is Lyapunov stable, yet not asymptotically stable; see the
phase portrait in Figure 5(a). 

x2 x2 x2
(a) (b) (c)

x1 x1 x1

Figure 5: (a) Sketch of the phase portrait of (2.1); typical example of a

center equilibrium at the origin. (b) Stable focus. (c) Unstable node.

We prove the following theorems for dynamical systems generated by

an ODE
x′ = f (x), x ∈ Rd . (2.2)
Analogous results for maps exist.

10
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Definition 2.3. Let x∗ be an equilibrium for (2.2) and U be a neighbour-
hood of x∗ . A function L : U → R is called a Lyapunov function if

(D1) L is continuous,

(D2) L(x∗ ) = 0 and L(x) > 0 for x 6= x∗ ,

(D3) L is decreasing along orbits, i.e.,

L(φt0 (x0 )) ≥ L(φt1 (x0 )), (2.3)

for t0 < t1 and φt (x0 ) ∈ U \ {x∗ }.

The Lyapunov function is called strict if equality in (2.3) never occurs; see
Figure 6 for an illustration of the definition.

x1

x2

Figure 6: Sketch of a Lyapunov function L : R2 → R (in grey). The flow has

a stable focus and conditions (D1)-(D3) of Definition 2.3 can be checked.

Theorem 2.4 (Lyapunov Stability Criterion). Suppose x∗ is an equilibrium

and there exists a Lyapunov function L. Then x∗ is stable.

Proof. Let Sδ be the maximal connected component of the sub-level set

{x ∈ U : L(x) ≤ δ}

containing x∗ . Inside any neighbourhood U (x∗ ) = U , take a sufficiently

small ball B(x∗ ; δ), such that ∂U ∩ B(x∗ ; δ) = ∅. We claim that for every
δ > 0, there exists ε > 0 such that

Sε ⊆ B(x∗ ; δ) := {x ∈ Rd : |x − x∗ | < δ} (2.4)

11
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 We argue by contradiction. Therefore, for every n ∈ N, there exists xn ∈
S1/n such that |xn − x∗ | ≥ δ. Since S1/n is connected, we can take xn such
that |xn − x∗ | = δ. Since the sphere is compact, and all xn lie in ∂B(x∗ ; δ),
we extract a convergent subsequence

lim xnm = y ⇒ L(y) = lim L(xnm ) = 0

m→∞ m→∞

since L is continuous by Definition 2.3(D1). Therefore, y = x∗ by Defini-

tion 2.3(D2) and this contradicts |y − x∗ | = δ > 0. Note that Sε is now a
closed set as it does not share boundary points with U . We claim that Sε is
positively invariant. We argue by contradiction and assume there exists t1
such that we leave Sε at φt1 (x0 ) =: x1 . Since Sε is closed it follows x1 ∈ U
and x1 ∈/ ∂U . There exists a ball

B(x1 ; r) ⊆ U

of radius r > 0 such that φt1 +ρ (x0 ) ∈ B(x1 ; r) \ Sε for sufficiently small
ρ > 0. However, this implies

L(φt1 +ρ (x0 )) > L(x1 ), (2.5)

since otherwise Sε would not be the maximal connected component. The

result (2.5) is a contradiction to Definition 2.3(D3), which concludes the
proof.

Certain partial results on the existence of Lyapunov functions exist un-

der the assumption that the equilibrium is asymptotically stable (exercise!).

Definition 2.5. Given any x ∈ X , we define the ω-limit set of x by

 
ω(x) := y ∈ X : ∃tj → +∞ such that lim φtj (x) = y .
j→∞

The α-limit set is given by

 
α(x) := y ∈ X : ∃tj → −∞ such that lim φtj (x) = y .
j→∞

Theorem 2.6 (LaSalle’s Invariance Principle). If a Lyapunov function L

for an equilibrium x∗ is not constant on any orbit lying entirely in U (x∗ ) \
{x∗ }, then x∗ is asymptotically stable.

Proof. For every x with φt (x) ∈ U (x∗ ) for all t ≥ 0, the limit

lim L(φt (x))

t→+∞

12
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 exists by monotonicity. If y ∈ ω(x), it follows that

lim φtj (x) = y for some sequence tj → +∞.

j→∞

Therefore, we have the limit

lim L(φt (x)) = L(y)

t→+∞

Since y ∈ ω(x) and ω(x) is invariant (exercise!), it follows that L(y) is

constant on ω(x) so we must have y = x∗ since L is only constant on the
orbit x∗ . Therefore, ω(x) = {x∗ } so x∗ is asymptotically stable.

If a Lyapunov function is differentiable, we have Definition 2.3(D3) if

and only if L′ (φt (x)) ≤ 0, i.e.,

L′ (φt (x)) = ∇L(φt (x)) · φ′t (x) = ∇L(φt (x)) · f (φt (x)) ≤ 0.

The formula ∇L(x) · f (x) is called the Lie derivative of L along f .

Definition 2.7. A function g : Rd → R, for which the Lie derivative is
zero on every orbit is called a first integral (or constant of motion).
Example 2.8. One easily checks that L(x, y) = x2 + y 2 is a Lyapunov
function for the harmonic oscillator (2.1). In this special case, since
   
2x y
∇L(x) · f (x) = · = 0.
2y −x

it follows that L is even a first integral. 

Example 2.9. Consider a particle at position x = x(t) ∈ Rd of mass 1
moving in a potential V : Rd → R according to Newton’s law x′′ = −∇V (x).
With x′ =: y, we get

x′ = y = ∂y H(x, y),
′ (2.6)
y = −∇V (x) = −∂x H(x, y),

where H : Rd → R, H(x, y) = 12 |y|2 + V (x) is the Hamiltonian of the

system and (2.6) is in the canonical form of Hamilton’s equations. One
checks (exercise!) that H is a first integral but in general not a Lyapunov
function. 
Theorem 2.10 (Linear System Stability). Consider the ODE

x′ = Ax, A ∈ Rd×d (2.7)

with equilibrium x∗ = 0.

13
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (T1) If all eigenvalues λj = aj + ibj of A satisfy Re(λj ) = aj < 0, then 0
is locally asymptotically stable.

(T2) If there exists λj with Re(λj ) > 0 then 0 is unstable.

Proof. Transforming A into Jordan canonical form, say J = Q−1 AQ, yields
that y = Q−1 x satisfies

y ′ = Jy, y ∈ Rd .

Solving the system explicitly using the matrix exponential φt (x) = etJ x
we have to consider all possible Jordan block structures. For (T1), any
diagonal block has solutions

etλj = eaj t (cos(bj t) + i sin(bj t)) → 0 as t → +∞

since aj < 0. For any non-diagonal Jordan block, polynomial pre-factors

are obtained, which are dominated by the exponential. The result (T1)
follows transforming back x = Qy. The proof of (T2) is an exercise.

Definition 2.11. Consider the linear system (2.7) for d = 2 and λ1,2 =
a1,2 + ib1,2 . Then x∗ is called a
• node if b1,2 = 0, a1 a2 > 0;

• saddle if b1,2 = 0, a1 a2 < 0;

• focus if b1,2 6= 0, a1,2 6= 0;

• center if b1,2 6= 0, a1,2 = 0;

we simply refer to the remaining cases as degenerate, they will occur later;
see also Figure 5.
The obvious pre-fixes apply to objects from Definition 2.11, e.g., if
a1 , a2 < 0 and b1,2 = 0, then x∗ = 0 is a stable node; cf. Theorem 2.10.
Consider the linear map

x 7→ Ax x ∈ Rd , x(0) = x0 .

Iterating yields xk = g k (x0 ) = Ak x0 . Using Jordan canonical form J =

Q−1 AQ we have  
lim xk = Q lim J Q−1 x0
k
k→∞ k→∞

converges to the fixed point x∗ = 0 ∈ Rd if the eigenvalues µj of A satisfy

|µj | < 1. In this case, the eigenvalues are called multipliers. Hence, there
is a natural analog for linear systems of ODEs and maps.

14
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 3 Stable/Unstable Manifolds
Before we start, a brief introduction (and/or refresher) on elementary def-
initions from differential geometry/topology.
Definition 3.1. A d-dimensional (topological) manifold M is a second-
countable Hausdorff topological space such that each point x ∈ M has a
neighbourhood, which is locally homeomorphic to an open set in Rd .
One may check that topological manifolds have an open covering {Uα },
for Uα ⊆ M of charts, i.e., homeomorphisms hα : Uα → Vα ⊂ Rd , where
each Vα is an open subset in Rd . The collection {(Uα , hα )} of charts is
called an atlas; see Figure 7(a).
x2
(a) (b1) (b2)
Uα Uβ
x1

hα hβ
(b3) (b4)

Vα Vβ
hα ◦ h−1
β

Figure 7: (a) Sketch of a manifold, charts, and transition maps. (b) Ex-
amples of two-dimensional manifolds: Euclidean space R2 , a square in R2 ,
a torus, and a sphere.

Definition 3.2. A d-dimensional C k -manifold M is a topological mani-

fold, where the atlas satisfies for Uα ∩ Uβ 6= ∅ that the map

hα ◦ h−1 k
β : hβ (Uα ∩ Uβ ) → hα (Uα ∩ Uβ ) is in C , (3.1)

and the atlas is maximal, i.e., if there exists another chart, which is com-
patible in sense of (3.1) with all other charts, then it is included in the
atlas; see also Figure 7(a).
Remark : One may prove that it suffices to exhibit some C k -atlas, where the
smoothness condition for the transition maps is satisfied, to show that the manifold
is C k . Surprising fact: Any topological manifold for d ≤ 3 can be given the
structure of a smooth manifold but this fails for d ≥ 4.

Example 3.3. Clearly, Euclidean space Rd , an interval [0, 1] or the rectan-

gle [0, 1]2 are manifolds. The same holds for the torus Td = Rd /Zd , d ≥ 1

15
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (d = 1 is just the circle), as well as unit sphere Sd = {x ∈ Rd : |x| = 1}. All
these manifolds are C k (they are even smooth k = ∞, and analytic k = ω)
and can be locally written as graphs of smooth functions (exercise!); see
Figure 7(b). 

One naturally obtains the space C k (M1 , M2 ) for maps between smooth
manifolds by postulating that f ∈ C k if it is a smooth map upon application
of coordinate charts.

Definition 3.4. A C k -curve in M is a C k -map γ : (−s, s) → M. The

tangent vector to γ at x = γ(0) is a linear map v : C 1 (M, R) → R defined
by
d
v(f ) := f (γ(t)) .
dt t=0
The linear span of all tangent vectors at x is called the tangent space
Tx M; see Figure 8.

Tx M

S2

Figure 8: Two-dimensional sphere and tangent space at a point x.

In a chart h : U ⊂ M → Rd with local coordinates h(x) = (x1 , x2 , . . . , xd ),

we can view a tangent vector v as
d
X ∂f dxi (γ(t))
v(f ) = vi vi :=
∂xi h(x) dt t=0
i=1

by the chain rule (exercise!) so v is just a vector in Rd with components vi .

Example 3.5. Take M = S2 , then one checks that Tx M ≃ R2 for every x,

i.e., we have just a tangent plane at every point. It holds in more generality
(exercise!) that Tx M ≃ Rd ; see Figure 8. 

16
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 We now start the main dynamical constructions. As before, we mostly
work in this section with dynamical systems generated by the ODE (2.2),
i.e.,
x′ = f (x), x ∈ Rd . (3.2)

Definition 3.6. Consider the ODE (3.2) with f ∈ C 1 . An equilibrium x∗

is called hyperbolic if the Jacobian Df (x∗ ) has no eigenvalues with zero
real parts.

Definition 3.7. Consider an iterated map g ∈ C 1 , x 7→ g(x). A fixed

point x∗ is called hyperbolic if the Jacobian Dg(x∗ ) has no multiplier µ
with |µ| = 1.

Formulated concisely, hyperbolicity for steady states implies that eigen-

values are away from the imaginary axis and multipliers away from the unit
circle. Suppose x∗ is a hyperbolic equilibrium. Wlog x∗ = 0 (using a shift
x̃ = x − x∗ if necessary), so we may use Taylor’s Theorem to write

x′ = f (x) = Df (0)x + R(x), R(x) = o(|x|2 ) as |x| → 0; (3.3)

we would have R(x) = O(|x|2 ) for f ∈ C 2 or even f ∈ C ∞ , which is often

just assumed in dynamical systems; here we shall often write the nonlinear
terms just as O(|x|2 ) with the understanding to be careful in the case of
f ∈ C 1 and f ∈/ C 2 . We expect the linear part A = Df (0), defining the
linearized system

X ′ = Df (0)X = AX, X ∈ Rd

to dominate. Since 0 is hyperbolic, A is a hyperbolic matrix and there

exists a splitting
Rd = E s (0) ⊕ E u (0)
where E s (0) is the stable eigenspace and E u (0) is the unstable eigenspace

E s (0) = {y ∈ Rd : etA y → 0 as t → ∞},

E u (0) = {y ∈ Rd : etA y → 0 as t → −∞}.

The spaces E s = E s (x∗ ), E u = E u (x∗ ) for general equilibria are defined

analogously and come with natural projections

Ps : Rd → E s and Pu : Rd → E u .

However, the nonlinear terms O(|x|2 ) may locally deform the geometry of
the linear spaces near x∗ ; see Figure 9.

17
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 x2 x2
(a) (b)
u
Wloc
Pu

x1 x1
s
Wloc
Eu Eu

Es Es

Figure 9: (a) Stable/unstable eigenspaces for a planar saddle. (b) Sta-

ble/unstable manifolds tangent to the eigenspaces.

Theorem 3.8 (Stable/Unstable Manifolds (for flows)). Suppose the

ODE (2.2), f ∈ C 1 , has a hyperbolic equilibrium x∗ and Df (x∗ ) has k
real-part negative and d − k real-part positive eigenvalues with correspond-
ing eigenspaces E s (x∗ ) and E u (x∗ ) for the linearized system. Then there
exists a neighbourhood U of x∗ with local stable and unstable manifolds
s (x ) and W u (x )
Wloc ∗ loc ∗
s
Wloc (x∗ ) = {y ∈ U : φt (y) → x∗ as t → ∞ and φt (y) ∈ U ∀t ≥ 0},
u
Wloc (x∗ ) = {y ∈ U : φt (y) → x∗ as t → −∞ and φt (y) ∈ U ∀t ≤ 0}.
s (x ) and
Furthermore, the stable/unstable manifolds are differentiable. Wloc ∗
u s u
Wloc (x∗ ) are tangent to E (x∗ ) and E (x∗ ) at x∗ .
We carry out three steps, the first two are the most involved.
s (x ); the case
Proof. Step 1: We are going to prove the existence of Wloc ∗
u (x ) can be obtained by time-reversal. Wlog x = 0, consider (3.3), let
Wloc ∗ ∗
x(0) = x0 , and define
xs,u := Ps,u x0 , Rs,u (x) := Ps,u R(x), xs,u (t) := Ps,u x(t).
s (0) as a graph over E s (0), i.e., formally
The goal is to construct Wloc
s
Wloc (0) := {x = xs + xu ∈ Rd ∩ U : xu = h(xs )}, h : Rk → Rd−k (3.4)
as shown see Figure 10(a).
There are two main techniques to construct h; here we use functional it-
eration (Lyapunov-Perron method). We re-write (3.3) using the variation-
of-constants formula
Z t
tA
x(t) = e x0 + e(t−r)A R(x(r)) dr. (3.5)
0

18
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (a) (b)

h(x)
Es Es
x
Bεs

Eu Eu
s (0) as a graph with a sufficiently small
Figure 10: (a) Idea to construct Wloc
(ball) domain Bε = B(0; ε) within the stable eigenspace E s = E s (0). (b)
s

Potential problem for the alternative graph transform method mentioned

after the functional iteration method proof presented here.

x(t) remains bounded only for certain x0 = xs + xu . Hence, we project to

the unstable part to obtain
Z t
−tA
xu = e xu (t) − e−rA Ru (x(r)) dr.
0

Suppose |x(t)| is bounded as t → +∞, then taking limits gives

Z ∞
xu = − e−rA Ru (x(r)) dr. (3.6)
0

Plugging (3.6) into (3.5) gives

Z t Z ∞
tA (t−r)A
x(t) = e xs + e Rs (x(r)) dr − e(t−r)A Ru (x(r)) dr. (3.7)
0 t

Define P (t) = Ps for t > 0 and P (t) = −Pu for t ≤ 0, then (3.7) is re-written
as
Z ∞
tA
x(t) = K(x)(t), K(x)(t) = e xs + e(t−r)A P (t − r)R(x(r)) dr. (3.8)
0

In particular, all bounded solutions satisfy the operator fixed-point equa-

tion (3.8). Suppose we could solve (3.8), denote its solution by ψ(t, xs ),
then the definition
h(y) := Pu ψ(0, y)
is going to give Wlocs (0). To solve (3.8), we introduce the Banach space

Cb ([0, ∞), Rd ) with norm

kxk = sup |x(t)|.

t≥0

19
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Since we work in some U , assume |x(t)| ≤ δ for some δ > 0. Observe that
g satisfies
|R(x(t)) − R(y(t))| ≤ ε|x(t) − y(t)| (3.9)
and ε > 0 can be made arbitrarily small if δ > 0 is made small. In
backward time solutions of the linearized system decay exponentially in
E u , and solutions decay in forward time in E s so that the definition of P (t)
implies
ke(t−r)A P (t − r)k ≤ Ce−κ|t−r| (3.10)
for some constant κ > 0, which can be bounded using the spectrum of A
and some constant C > 0. Using (3.9) and (3.10) in (3.8) yields
Z ∞
kK(x) − K(y)k ≤ sup e(t−r)A P (t − r)(R(x(r)) − R(y(r))) dr
t≥0 0
Z ∞
≤ C sup e−κ|t−r| |R(x(r)) − R(y(r))| dr
t≥0 0
Z ∞
2Cε
≤ Cεkx − yk sup e−κ|t−r| dr = kx − yk.
t≥0 0 κ

For ε < κ/(2C), K is a contraction so the Banach Fixed Point Theorem 1.8
s (0) using (3.4).
gives the existence of x(t) and hence also Wloc
Step 2: The next question is differentiability. We know the solution
ψ(t, xs ) is continuous in t. To show continuity ψ ∈ C([0, ∞) × U , Rd ), we
use the triangle inequality

|ψ(t, a) − ψ(s, b)| ≤ |ψ(t, a) − ψ(t, b)| + |ψ(t, b) − ψ(s, b)|, (3.11)

where the second term can be made small by continuity in time. We claim
that the first term is also small. Indeed, note that

K(x)(t) = Kxs (x)(t)

is continuous as an operator in xs since

kKa (x)(t) − Kb (x)(t)k = sup |e−tA a − e−tA b| ≤ Cka − bk.

t≥0

To also obtain ψ ∈ C 1 ([0, ∞) × U , Rd ), we look at partial derivatives. Dif-

ferentiating the fixed-point equation (3.8), we find
Z t
tA
∂t ψ(t, a) = Ae a + ∂t e(t−r)A P (t − r)R(ψ(r, a)) dr
0

20
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 which is continuous since R and P are. The other derivative Da ψ = η, if it
exists, must satisfy the fixed-point equation
Z ∞
tA
η(t, xs ) = e Ps + e(t−r)A P (t − r)η(t, xs ) · ∇R(ψ(r, xs )) dr, (3.12)
0

where the ∇R exists in L1 , as R is Lipschitz on U , hence absolutely contin-

uous, and hence differentiable almost everywhere; furthermore, the domi-
nated convergence theorem has been used to interchange limits. The fixed-
point problem (3.12) can again be solved, now for η using the same Ba-
nach contraction mapping argument as above. Hence, η(t, xs ) exists and
is continuous in xs . Since both partials exist and are continuous, we have
ψ ∈ C 1 ([0, ∞) × U , Rd ).
Step 3: The last question is tangency at x = 0. We evaluate (3.12)
at (t, xs ) = (0, 0), which implies η(0, 0) = Ps since ∇R(ψ(r, 0)) = ∇R(0) is
zero almost everywhere as R = o(|x|) as |x| → 0. Therefore, we have

Dxs h(a)|a=0 = Pu Ps = 0

so E s (0) is indeed tangent to Wloc

s (0).

Differentiability of the stable/unstable manifolds can be improved to

C k if f ∈ C k for k ≥ 1; in this case, even analyticity transfers, i.e., if
s,u
f ∈ C ω then Wloc (x∗ ) are C ω . The manifolds are invariant once they have
been extended under the flow φt of (3.2), i.e.,
[ s,u
W s,u (x∗ ) = φt (Wloc (x∗ )).
t∈T

A natural analogous statement for maps holds.

Theorem 3.9 (Stable/Unstable Manifolds (for maps)). Consider a
C 1 -diffeomorphism
x 7→ g(x) for x ∈ Rd ,
which has a hyperbolic fixed point x∗ and Dg(x∗ ) has k multipliers inside
the unit circle and d − k multipliers outside the unit circle with correspond-
ing eigenspaces E s (x∗ ) and E u (x∗ ) for the linearized system. Then there
exists a neighbourhood U of x∗ with local stable and unstable manifolds
s (x ) and W u (x )
Wloc ∗ loc ∗
s
Wloc (x∗ ) = {y ∈ U : φt (y) → x∗ as t → ∞ and φt (y) ∈ U ∀t ≥ 0},
u
Wloc (x∗ ) = {y ∈ U : φt (y) → x∗ as t → −∞ and φt (y) ∈ U ∀t ≤ 0}.
s (x ) and
Furthermore, the stable/unstable manifolds are differentiable. Wloc ∗
u (x ) are tangent to E s (x ) and E u (x ) at x .
Wloc ∗ ∗ ∗ ∗

21
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 The second major proof technique (Hadamard-Perron method) for
invariant manifold theorems will be briefly illustrated in the context of
Theorem 3.9 for x∗ = 0. Consider the set
Bεs := {y ∈ E s (0) : |y| ≤ ε}.
Let L(L, ε) be the space of Lipschitz-continuous maps h : Bεs → E u (0) with
Lipschitz constant L and h(0) = 0, which is a Banach space under the
supremum norm. The graph transform is constructed as operator
G : L(L, ε) → L(L, ε), G(ϕ) = ϕ̃.
It is defined by considering the maping
ϕ 7→ g −1 (graph(ϕ)). (3.13)
One then proves (nontrivial!) that the resulting set in (3.13) is again the
graph of a Lipschitz map ϕ̃ for sufficiently small ε > 0 over the correct
domain Bεs ; see Figure 10(b). Then one proves (again nontrivial!) that G
is a contraction for suitable choices of (L, ε) and obtains the existence of
s (0) as a fixed point of G.
Wloc
Example 3.10. Consider the (unforced) Duffing equation/oscillator
x′′ + δx′ − x + px3 = 0, x ∈ R,
where p, δ ∈ R are parameters and δ ≥ 0. Re-writing it as a first-order
system gives
x′ = y,
(3.14)
y ′ = −δy + x − px3 .
Obviously, (x, y) = (0, 0) is an equilibrium. If δ = 0, then one checks
dim E s (0) = 1 = dim E u (0). Therefore, there exist one-dimensional stable
and unstable manifolds by Theorem 3.8. 
Example 3.11. Consider the linear map (sometimes called Arnold’s cat
map) defined by
 
21
g(x) = Ax, A= , x ∈ R2 . (3.15)
11
One checks (exercise!) that g preserves the integer lattice Z2 ⊂ R2 and
induces a map on the torus X = T2 = R2 /Z2 ; see Figure 11. The origin
(0, 0) is a hyperbolic saddle fixed point (exercise!) and hence Theorem 3.8
applies to this point. The map g : T2 → T2 is the classical example for a
hyperbolic toral automorphism, i.e., it does not only have a hyperbolic
splitting at (0, 0) but near every point in X it has well-defined contracting
and expanding directions. 

22
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
(a) x2 (b) x2

2 2

1 1

0 x1 0 x1
0 1 2 3 0 1 2 3
Figure 11: (a) Mapping of the fundamental domain [0, 1]2 (grey) via (3.15)
to a polygon (black). (b) Identifying the resulting object back onto the
fundamental domain.

4 Equivalence
Definition 4.1. A dynamical system (X , T , φt ) is called topologically
equivalent to a system (X , T , ψt ) if there exists a homeomorphism h : X →
X mapping orbits from the first system to the second system preserving
the direction of time. If the time parametrization is also preserved, the
systems are called conjugate.

If h, h−1 ∈ C k , k ≥ 1, then the systems are called C k -equivalent

(sometimes this is called smoothly equivalent if the precise k ≥ 1 does
not matter for the current argument). Similarly, for preserved time parametriza-
tions and C k -diffeomorphisms, we have a C k -conjugacy.

Proposition 4.2. Two discrete-time invertible maps

x 7→ f (x), x ∈ Rd , y 7→ g(y), y ∈ Rd ,

are topologically conjugate if and only if there exists a homeomorphism

h : Rd → Rd such that
f = h−1 ◦ g ◦ h. (4.1)

Proof. We only show that topological conjugacy implies (4.1); the converse
is an easy exercise. Topological equivalence implies that there exists a
homeomorphism h : Rd → Rd such that

y = h(x) and h(f (x)) = g(y).

Therefore, h(f (x)) = g(y) = g(h(x)) and since h is invertible, the result
follows.

23
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Proposition 4.3. Consider two continuous-time flows φt and ψt generated
by
x′ = f (x), x ∈ Rd and y ′ = g(y), y ∈ Rd
with f, g ∈ C 1 . Suppose y = h(x) for some C 1 -diffeomorphism h : Rd →
Rd . Then φt , ψt are C 1 -conjugate, and

f (x) = (Dh)−1 (x)g(h(x)). (4.2)

Proof. Regarding (4.2) we make the important calculation

g(h(x)) = g(y) = y ′ = Dh(x)x′ = Dh(x)f (x).

Therefore, h maps solutions to solutions and we have h(φt (x)) = ψt (h(x)).

Definition 4.4. A dynamical system (X , T , φt ) with steady state x∗ is

called locally topologically equivalent to a system (X , T , ψt ) near x∗
if there exists a homeomorphism h : U (x∗ ) → V(y∗ ) mapping orbits from
the first system to the second system preserving the direction of time for
neighbourhoods U (x∗ ), V(y∗ ), where ψt (y∗ ) = y∗ for all t ∈ T .
Theorem 4.5 (Hartman-Grobman Theorem). Consider the ODE x′ =
f (x), x ∈ Rd , f ∈ C 1 , with flow φt and suppose x∗ is a hyperbolic equi-
librium. Then φt is locally topologically conjugate to the flow etA (with
A = Df (x∗ )) generated by the linearized system.

x2 x2

x1 x1

Figure 12: Generic stable node for a nonlinear system conjugated via h
to a node for a linear system. For generic matrices with spectrum in the
left-half complex plane, there exists a weakest contracting eigendirection
(exercise!); see also Definition 6.9.

We are only going to prove the key statement for maps

g(x) = Ax + R(x), x ∈ Rd , g ∈ C 1 , g is invertible, (4.3)

24
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 and then comment on the proof of the Hartman-Grobman Theorem; see
also Figure 12.
Theorem 4.6. Suppose 0 is a hyperbolic fixed point of (4.3). Then there
exists a unique homeomorphism h(x) = x + H(x), H(0) = 0, H bounded,
such that in a neighbourhood U = U (0) ⊂ Rd we have

h ◦ A = g ◦ h, (4.4)

i.e., g is locally conjugate to A.

Proof. Step 1: We have to prove the existence of H. Suppose A has
eigenvalues inside and outside the unit circle; the other cases can be treated
similarly. Split A via As := A|E s (0) and Au := A|E u (0) . Select a matrix norm
k · k such that

α := max(kA−1
u k, kAs k) < 1, |R(x) − R(y)| ≤ ε|x − y|, (4.5)

where we select ε < (1 − α)/2 and the inequality for R holds on U but we
cut off R with a smooth cut-off function outside U so that it holds globally.
The result (4.4) is equivalent to the functional equation

H(Ax) − AH(x) = R(x + H(x)) ∀x ∈ U . (4.6)

Consider the Banach space C(Rd , Rd ) with the supremum norm. Define
two invertible linear operators

(KH)(x) := H(Ax), (AH)(x) := AH(x),

where K preserves the norm, while A has expanding and contracting di-
rections. Setting L := K − A, it follows that (4.6) can be viewed as a
fixed-point problem for L

LH(x) = R(x + H(x)),

where we claim that L is invertible. To prove this, consider the splittings

C(Rd , Rd ) = C(Rd , E s (0)) ⊕ C(Rd , E u (0)), K = Ks ⊕ Ku , L = Ls ⊕ Lu .

We decompose along each direction and find

Lu = Ku − Au = −Au (Id − A−1

u Ku )

so that L−1 −1 −1 −1 −1
u = −(Id − Au Ku ) Au , where the existence of Au holds by
−1
assumption and (Id − Au Ku ) is invertible using the Neumann series
∞
X
(Id − A−1
u Ku )
−1
= (A−1
u Ku )
k

k=0

25
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 as kA−1 −1
u Ku k ≤ α < 1. Furthermore, we have kLu k ≤ 1/(1 − α) using the
geometric series. The same argument applies to Ls so
2
L−1 = (Ku − Au )−1 ⊕ (Ks − As )−1 , kL−1 k ≤ .
1−α

So we may now try to solve H(x) = L−1 R(x + H(x)) for H. Since the goal
is to apply Banach Fixed Point Theorem 1.8, we estimate
2
kL−1 R(Id + H1 ) − L−1 R(Id + H2 )k ≤ kR(Id + H1 ) − R(Id + H2 )k
1−α
2ε
≤ kH1 − H2 k.
1−α
This implies we have a contraction by (4.5) and H exists and is continuous.
Step 2: It remains to prove that h = Id + H is invertible. First, note
that we can construct a map h̃ such that

h̃(x) = x + H̃(x), A ◦ h̃ = h̃ ◦ g. (4.7)

This follows by exchanging the roles of A, g above and repeating the same
line of argument. Therefore, using (4.4) and (4.7) we find

A ◦ h̃ ◦ h = h̃ ◦ g ◦ h = h̃ ◦ h ◦ A. (4.8)

One checks that if R(x) ≡ 0, the homeomorphism is unique in the linear

case so (4.8) implies
h̃ ◦ h = Id = h ◦ h̃
and we obtain h̃ = h−1 .
Step 3: It remains to check that H(0) = 0 and that h is unique. One
evaluates Ah−1 (x) = h−1 (g(x)) at x = 0, which gives the linear equation

Ah−1 (0) = h−1 (0) ⇔ Ah−1 (0) − Idh−1 (0) = 0

so h−1 (0) is an eigenfunction for multiplier 1, but A has no multipliers on

the unit circle by assumption. This means we must have h−1 (0) = 0. The
proof of uniqueness is a (relatively easy) exercise.

The proof of the Hartman-Grobman Theorem 4.5 uses the time-one

map. We let φt denote the flow of x′ = f (x) and define

φ1 (x) = eA x + R(x)

26
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 for some remainder R. Upon localizing the vector field near x = 0 with a
cut-off, one checks that R is indeed small so that Theorem 4.6 applies so
we get a homeomorphism h with

h ◦ eA = φ1 ◦ h. (4.9)

The final step is to define

hs := φs ◦ h ◦ e−sA

and note that hs also satisfies (4.9) for s = 1 (exercise). Using the unique-
ness part of Theorem 4.6 one obtains hs ≡ h completing the construction.

5 Poincaré Maps
Consider the ODE

x′ = f (x), x ∈ Rd , f ∈ C 1 , (5.1)

with flow φt and suppose Γ = Γ(t) is a periodic orbit. The idea of a Poincaré
map is to record returns to a manifold transversal to Γ; see Figure 13.

P (x)
x
y

Figure 13: Sketch of the Poincaré map construction.

Theorem 5.1 (Existence/Smoothness of the Poincaré Map). Consider (5.1),

suppose Γ(t) = Γ has period T > 0 and y ∈ Γ. Let

Σ := {x ∈ Rd : (x − y) · f (y) = 0} (5.2)

be a cross-section (or Poincaré section). Then there exists δ > 0 and

τ : B(y; δ) → R, τ ∈ C 1 , such that

τ (y) = T and φτ (x) (x) ∈ Σ ∀x ∈ B(y; δ).

27
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 In summary, the Poincaré map P (x) := φτ (x) (x) is locally well-defined
and C 1 on B(y; δ) ∩ Σ.

Proof. The idea is to get τ via the implicit function theorem. Define the
function
F (t, x) := (φt (x) − y) · f (y) (5.3)
which is in C 1 since φ ∈ C 1 . Furthermore, F (T, y) = 0 by periodicity of Γ.
Differentiating (5.3) in t, we find

∂t F (T, y) = f (y) · ∂t (φt (x))|(t,x)=(T,y) = |f (y)|2 6= 0

since y is not an equilibrium. Therefore, the implicit function theorem

provides a C 1 map τ : B(y; δ) → R such that

τ (y) = T, 0 = F (τ (x), x) = (φτ (x) (x) − y) · f (y) ⇒ φτ (x) (x) ∈ Σ,

which finishes the proof.

The definition (5.2) imposes that Γ intersects Σ transversally at y.

This construction generalizes naturally if Σ is replaced by a smooth man-
ifold M (exercise!). Up to a coordinate change, we may assume local co-
ordinates x = (x1 , . . . , xd−1 ) on Σ, so that P (0) = 0 and consider the
Jacobian
A = DP (0) with multipliers µ1 , . . . , µd−1 .
In particular, analyzing stability of the periodic orbit is equivalent to sta-
bility of the fixed point via its multipliers.

Proposition 5.2. The multipliers of A are independent of y, of the cross-

section as well as of the local coordinates.

Proof. Suppose there are two points y, ỹ, two sections Σ, Σ̃, and coordinates
x, x̃ with Poincaré maps P, P̃ . Then, using the construction of P and P̃ ,
one finds that there exists a diffeomorphism h(x) = x̃ such that P, P̃ are
locally conjugate so
P = h−1 ◦ P̃ ◦ h.
Differentiating and evaluating at x = 0 gives

A = B −1 ÃB, B := Dh(0), Ã = DP̃ (0).

Then we calculate using the standard characteristic equation

det(A − µId) = det(B −1 ) det(Ã − µId) det(B) = det(Ã − µId)

because det(B −1 ) det(B) = 1 by standard rules for determinants.

28
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Example 5.3. Consider the planar system with parameter p ∈ R given by

x′ = −y + x(p − x2 − y 2 ),
y ′ = x + y(p − x2 − y 2 ),

which turns out to have a limit cycle, where the single multiplier µ can be
calculated exactly (exercise! Hint: polar coordinates). However, in general
cases, this is not possible analytically. 
To analyze the stability of a periodic orbit Γ, one may also link this to
classical ODE theory. Indeed, write

x(t) = Γ(t) + u(t), 0 < |u(t)| ≪ 1,

i.e., u is a small perturbation. Then, we have formally

u′ = x′ − Γ′ = f (Γ(t) + u(t)) − f (Γ(t)) = A(t)u + o(|u|2 )

where A(t) = Df (Γ(t)).

Definition 5.4. The ODE u′ = A(t)u is called the variational equation
about Γ.
Definition 5.5. A matrix M (t) is called a fundamental solution of
a time-dependent linear system x′ = A(t)x if M (0) = Id and M ′ (t) =
A(t)M (t) for all t ∈ T .
The previous definition naturally generalizes to matrices M = M (t, t0 ),
where the starting time is t0 6= 0 but we assume here t0 = 0. If T is
the period of Γ and M solves the variational equation about Γ. Then
u(T ) = M (T )u(0) and M (T ) is called the monodromy matrix.
Theorem 5.6. The monodromy matrix M (T ) has eigenvalues 1, µ1 , . . . , µd−1 ,
where µj are the multipliers of the Poincaré map.
Proof. Fix y ∈ Γ. First, by the rectification theorem from ODEs, we
may assume that

f (x) = (0, . . . , 0, 1)⊤ ∈ Rd for x near y, (5.4)

and take local coordinates (x1 , . . . , xd−1 , z), xd =: z ∈ R, such that the
Poincaré map P (x) = φτ (x) (x) is given solely in the first (d−1) coordinates.
We write φt (x) = φ(t, x) for clarity and differentiate

DP (y) = D φ(τ (x), x)|x=y = ∂t φ(T, y) (∇τ (y))⊤ + Dx φ(T, y)

= f (y) (∇τ (y))⊤ + Dx φ(T, y). (5.5)

29
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 We claim that Dx φ(T, y) = M (T ). Indeed, we have Dx φ(0, y) = Id and we
calculate
(Dx φ(t, y))′ = Dx φ(t, y)f (y)
interchanging derivatives wrt t and x. Therefore, equation (5.5) implies
using (5.4) that  
Dx P (y) v1
M (T ) = (5.6)
v2⊤ m
where v1,2 ∈ Rd−1 are some vectors and m ∈ R. We claim that v1 = 0 and
m = 1, which follows if we can prove M (T )f (y) = f (y), i.e., µ = 1 is an
eigenvalue of the monodromy matrix with eigenvector aligned with tangent
vector to the periodic orbit Γ(t) as easily drawn into Figure 13 (exercise!).
We have
Γ′′ (t) = f (Γ(t))′ = A(t)Γ′ (t)
so Γ′ (t) solves u′ = A(t)u. Therefore, we have

M (T )f (y) = M (T )Γ′ (0) = Γ′ (T ) = Γ′ (0) = f (y)

as required. The structure of the monodromy matrix in (5.6) with v1 = 0

and m = 1 now implies the result.

The construction of the Poincaré map naturally leads to the question,

when we can realize any given map g : Rd → Rd as the time-T map for
some flow φt , i.e.,

g(x) = φT (x), for some T > 0.

Example 5.7. It is easy to check that g(x) = − 21 x has no realization as

the time-T for a flow generated by x′ = f (x) with x ∈ R. 

Theorem 5.8 (Suspension Flow). Given a C 1 -diffeomorphism g : Rd →

Rd , there exists a (d + 1)-dimensional manifold M such that an ODE on
M has g as its time-T map.

Proof. Fix some T > 0 and define the manifold by

M := {(x, t) ∈ Rd × R : t ∈ [0, T ]}/((g(x), 0) ∼ (x, T )).

Then consider the flow φt on M defined by (x′ , t′ ) = (0, 1) (also called the
suspension). The time-T leaves {t = 0} ⊂ M invariant and induces the
map g; see Figure 14.

30
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (x0 , T )
t=T

t=0
(x0 , 0) (g(x0 ), 0)
Figure 14: Illustration of the suspension flow construction.

6 Structural Stability
It would be most elegant to analyze all dynamical systems of a certain class
at once, just restricting to those displaying typical dynamics. Consider a
compact d-dimensional set K ⊂ Rd .

Definition 6.1. The function space C k (K, Rd ) is given by all functions f

such that
k
X
kf kC k = sup sup |Dα fl (x)| < ∞,
l x∈K
|α|=j

where α = (α1 , . . . , αd ) ∈ (N0 )d is a multi-index and Dα = ∂xα11 · · · ∂xαdd .

C k (K, R) is a Banach space, and the definition generalizes to a compact

smooth manifold M replacing K, by taking the supremum over all charts.
For non-compact K we do not obtain Banach spaces but still metric spaces,
i.e., we could set
X supx∈Kj |f (x) − g(x)|
d0 (f, g) := 2−j
supx∈Kj |f (x) − g(x)| + 1
j

where Kj is a countable family of growing compact sets, or we could set

 
d0 (f, g) := min 1, sup |f (x) − g(x)| , (6.1)
x∈K

where both definitions naturally generalize to manifolds, if the manifold has

a metric d(x, y) = |x − y|; here we shall stick to (6.1) until further notice.

31
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Definition 6.2. Let f ∈ C 1 (M) be a vector field. Then f is called struc-
turally stable if there exists ε > 0 such that for all vector fields g ∈ C 1 (M)
with
kf − gkC 1 < ε,
g is topologically equivalent to f . If the f is not structurally stable, it is
called structurally unstable.

Example 6.3. Consider the unit torus T2 = R2 /Z2 with vector field f
defined by
x ′ = ω1 ,
(6.2)
y ′ = ω2 .
for fixed ω1,2 ∈ R. Then f is structurally unstable since for ω1 /ω2 ∈ Q all
orbits are periodic while for ω1 /ω2 ∈ R \ Q, all orbits are aperiodic, i.e.,
are never closed. 

Definition 6.4. Given a dynamical system (X , T , φt ), a point x ∈ X is

called nonwandering if for any neighbourhood U = U (x) of x and for any
t0 ∈ T , there exists a t > t0 such that

φt (U ) ∩ U =
6 ∅.

The set of such points is called the nonwandering set NW(φ); see Fig-
ure 15(a). Other points are called wandering.

(a) φt (b)

φ(U )
M

Figure 15: (a) Sketch of the idea in the definition of a nonwandering point.
(b) Illustration of one aspect of Peixoto’s Theorem: a saddle-to-saddle
connection (grey) is structurally unstable.

Example 6.5. Consider the flow φ generated by (6.2), then we easily check
(exercise!) that NW(φ) = T2 = X . 

32
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Theorem 6.6 (Peixoto’s Theorem). A C 1 -vector field f on a compact two-
dimensional differentiable manifold M is structurally stable if and only if

(T1) the number of periodic orbits and equilibria is finite, and each is hy-
perbolic;

(T2) for saddles x, y (potentially x = y), W s (x) ∩ W u (y) = ∅;

(T3) NW(φ) consists only of equilibria and periodic orbits.

Furthermore, if M is orientable, the set of structurally stable vector fields

is open and dense in C 1 (M).

We shall not prove Theorem 6.6 but see Figure 15(b). Peixoto’s The-
orem for d ≥ 3 is false. We only have a uni-directional statement under
similar assumptions.

Definition 6.7. A Morse-Smale flow φt is one for which

(D1) the number of periodic orbits and equilibria is finite, and each is
hyperbolic;

(D2) stable and unstable manifolds only intersect transversally;

(D3) NW(φ) consists only of equilibria and periodic orbits.

Theorem 6.8 (Morse-Smale Stability). Consider C 1 -vector fields on a

compact differentiable manifold. Then Morse-Smale systems are struc-
turally stable.

Definition 6.9. A property P of elements of a topological space Z is called

generic if it holds on a countable intersection of open dense sets.

A countable intersection of open dense sets is also called a residual.

Peixoto’s Theorem states that on compact two-dimensional orientable man-
ifolds structural stability is a generic property. For d ≥ 3 Morse-Smale
systems are not generic.

Example 6.10. Consider the space of matrices Z = Rd×d with topology

induced by any matrix norm. Then for A ∈ Z the property P of A being
hyperbolic, i.e., A has no zero real part eigenvalues, is generic (exercise!).

33
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 7 Hyperbolic Sets
In this section, we prove, why hyperbolicity is a such a key property to
expect robustness of dynamics and structural stability. Let M be a C 1
Riemannian manifold, i.e., there exists a Riemannian metric differ-
entiable family of symmetric bilinear forms

h·, ·ix : Tx M × Tx M → R
p
where we set |v| := hv, vi. For a curve γ : [a, b] → M, γ ∈ C 1 , we define
its length by
Z b
|γ ′ (s)| ds
a
which induces a metric d(x, y) on M by taking the infimum of the length
over all curves connecting x = γ(a) and y = γ(b). Next, define a discrete-
time dynamical system via a diffeomorphism

g : M → M. (7.1)

Since g ∈ C 1 , we have a well-defined derivative (or tangent map)

Dgx : Tx M → Tg(x) M, Dgx (v)(w) := v(w ◦ g),

which is just the Jacobian in the case M = Rd ; see Figure 16(a).

(a) (b)
v
x M Eu
Tx M Es

Dgx (v)
g(x)
M

Figure 16: (a) Illustration of the derivative/tangent map on a manifold M.

(b) Sketch of a local hyperbolic splitting on M.

Definition 7.1. A compact g-invariant Λ ⊂ M is called hyperbolic if

there exists λ ∈ (0, 1), C > 0, and families of subspaces E s (x), E u (x) ⊂
Tx M such that for every x ∈ Λ
(D1) (“invariant”) Dgx E s (x) = E s (g(x)) and Dgx E u (x) = E u (g(x)).

(D2) (“stable”) |Dgxj v s | ≤ Cλj |v s | for every v s ∈ E s (x) and j ≥ 0;

34
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (D3) (“unstable”) |Dgx−j v u | ≤ Cλj |v u | for every v u ∈ E u (x) and j ≥ 0;

(D4) (“splitting”) Tx M = E s (x) ⊕ E u (x);

One checks that a hyperbolic equilibrium as well as a hyperbolic pe-

riodic orbit are special cases of hyperbolic sets. Furthermore, the sta-
ble distribution E s := (E s (x))x∈M and the unstable distribution
E u := (E u (x))x∈M are checked (exercise!) to be continuous with respect to
x; see Figure 16(b). The stable and unstable distributions are contained in
the collection of all tangent spaces, which is defined in general as follows:

Definition 7.2. Suppose M is a C k -manifold of dimension d. The tangent

bundle [
TM = Tx M
x∈M

is a C k−1 -manifold of dimension 2d. If hα = (x1 , . . . , xd ) : Uα → Vα is a

coordinate chart for M, then (x1 , . . . , xd , dx1 , . . . , dxd ) : TM → R2d is a
coordinate chart for the tangent bundle, where dxi (v) = v(xi ).

Example 7.3. Consider hyperbolic toral automorphism on M = T2 , con-

sidered already in Example 3.11 defined by
 
21
g(x) = Ax, A= .
11

Calculating eigenvalues and eigenvectors of A, one observes that Defini-

tion 7.1 is satisfied with Λ = M. 

Definition 7.4. If for g defined (7.1), the entire manifold M = Λ is a

hyperbolic set, then we call g an Anosov diffeomorphism.

Theorem 7.5. Let Λ be a hyperbolic set for g : M → M. Then there

exists an open set U containing Λ and ε1 > 0 such that for every ε > 0
there exists δ > 0 such that for any f : M → M with

d0 (Df, Dg) ≤ ε1

and any homeomorphism h : Y → Y of a topological space Y, and any

continuous map ϕ : Y → U satisfying

d0 (ϕ ◦ h, f ◦ ϕ) ≤ δ

there exists a unique continuous map ψ : Y → U with

ψ◦h=f ◦ψ and d0 (ϕ, ψ) < ε.

35
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 U M

f
Λ
g

ϕ
∃ψ s.t. ϕ ≈ ψ
Df ≈ Dg
h
Y Y
Figure 17: Illustration of Theorem 7.5.

The main conclusion of Theorem 7.5 is shown in Figure 17, which

roughly states that for g, and any map f C 1 -close to it, we may turn
an approximate commutative diagram into an exact one due to hyperbolic-
ity. The proof uses some basic differential geometry and the Banach Fixed
Point Theorem to construct ψ, very similar to the proof of Theorem 3.8,
so we shall not carry it out here.

Example 7.6. If Y = {y} is a singleton and h(y) = y, then Theorem 7.5

states that every map f near g satisfies

ψ(y) = f ◦ ψ(y)

so f must have a fixed point. 

Definition 7.7. A δ-pseudo-orbit of a map g : M → M is finite or

infinite sequence (xk ) ⊂ M such that

d(g(xk ), xk+1 ) ≤ δ for all k.

Theorem 7.8 (The Shadowing Theorem). Let Λ be a hyperbolic set of

g : M → M. Then given any ε > 0 there exists δ > 0 such that if (xk ) is
a δ-pseudo-orbit of g and d(xk , Λ) < δ for all k, then there exists

x ∈ Λε := {x ∈ M : d(x, Λ) < ε}

such that Or(x) shadows the pseudo-orbit, i.e., we have

d(g k (x), xk ) < ε for any k.

36
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 .. x3
(a) x0 . (b)
g
g
g x2
g
x1
x2
Λ
g(x0 ) g(x1 )
g
δ
g
d(g(x0 ), x1 ) ≤ δ d(g(x1 ), x2 ) ≤ δ x x1
x0

Figure 18: (a) Illustrating the definition of a pseudo-orbit, which always

stays close to a true orbit. (b) Basic setting for the proof of Theorem 7.8.

Proof. The idea is to apply Theorem 7.5 so we start choosing a neighbour-

hood U satisfying the conclusions of the theorem and such that Λδ ⊂ U . If
(xk ) is not defined for all k ∈ Z, say

(xk ) = (xk− , . . . , xk+ )

augment it pre-images of g for some y− such that d(xk− , y− ) < δ and by

iterates/images of g for some y+ such that d(xk+ , y+ ) < δ. This yields
a bi-infinite sequence, still denoted (xk ), which is a δ-pseudo-orbit lying
inside a δ-neighbourhood of Λ. Define Y = (xk ) with the discrete topology
and note that the left-shift map

h : Y → Y, h(xk ) = xk+1

is a homeomorphism on Y. Now set f = g in Theorem 7.5 and consider

ϕ(xk ) = xk as the usual inclusion. Since we have a δ-pseudo-orbit, we must
have
d(ϕ(h(xk )), g(ϕ(xk ))) = d(xk+1 , g(xk )) ≤ δ
so the assumptions are satisfied and there exists ψ with

ψ◦h=g◦ψ and d0 (ϕ, ψ) < ε.

Now take x = x0 and observe that for any k we have

d(g k (x), xk ) = d(g k (ϕ(x)), ϕ(xk ))

d(g k (ψ(x)), ψ(xk )) = d(ψ(hk (x)), ψ(xk )) = d(ψ(xk ), ψ(x1 )) = 0.

Since ϕ is ε-close to ψ we get d(g k (x), xk ) < ε as required.

37
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 The previous results already hint at the fact that hyperbolic sets should
obey various forms of “persistence” or “robustness” to perturbations.
Theorem 7.9 (Persistence of Hyperbolicity). Let Λ be a hyperbolic set.
Then the following statements hold:
(T1) There is an open set U (Λ) containing Λ and ε1 > 0 such that if
K ⊂ U (Λ) is a compact invariant set of a diffeomorphism f : M → M
with
d1 (f, g) < ε1 , (7.2)
then K is a hyperbolic set.
(T2) For every open set V containing Λ and every ε > 0, there exists δ > 0
such that for every f : M → M with d1 (f, g) < δ, there exists a
hyperbolic set K ⊂ V of f and a homeomorphism ψ : K → Λ such
that
ψ ◦ f |K = g|Λ ◦ ψ
and d0 (ψ, Id) < ε.
Roughly speaking, (T1) say that hyperbolic sets persist under perturba-
tions of the map, while (T2) means that the dynamics on the original and
the perturbed hyperbolic set is topologically conjugate; cf. the Hartman-
Grobman Theorem.

Proof. Step 1: We deal with (T1). By continuity, we may extend the

distributions Egs and Egu to distributions Ẽgs and Ẽgu defined on U (Λ) and
decompose
Tx M = Ẽgs (x) ⊕ Ẽgu (x), v = vs + vu.
Next, consider α > 0 to be chosen below, and define the stable/unstable
cones by

Kαs (x) = {v ∈ Tx M : |v u | ≤ α|v s |},

Kαu (x) = {v ∈ Tx M : |v s | ≤ α|v u |}.

By hyperbolicity we see that there exists α > 0 such that Dg is contracting

on Kαs (x) and expanding on Kαu (x). Since Λ is compact and the unit
tangent bundle
( )
[
T1 M = v ∈ TM = Tx M : |v| = 1
x∈M

is also compact, we have a single λ ∈ (0, 1) such that

|Dgx v| ≤ λ|v| for v ∈ Kαs (x) and |Dgx−1 v| ≤ λ|v| for v ∈ Kαu (x)

38
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 for all x ∈ U (Λ). The same contraction/expansion statement also holds,
upon potentially increasing λ but still λ ∈ (0, 1), for f since it is C 1 -close
to g by (7.2). Now one can define the subspaces
\
E s (x) = D−k K s (f k (x)),
f k (x) α
k≥0
\
u
E (x) = Dkf −k (x) Kαu (f −k (x)),
k≥0

which are precisely satisfying the definitions of a hyperbolic set K for f .

Step 2: Proving (T2) is another application of Theorem 7.5. Let Y =
Λ and h = g|Λ with the usual inclusion ϕ : Λ ֒→ M. Hence, we obtain a
map ψ : Λ → M such that

ψ ◦ g|Λ = f ◦ ψ.

Let K := ψ(Λ). Now apply Theorem 7.5 again but to Y = K, h = f |K , and

ϕ : K ֒→ M to obtain a map ψ̃ : K → M such that

ψ̃ ◦ f |K = g|Λ ◦ ψ

and by uniqueness ψ̃ = ψ −1 . The set K is hyperbolic by (T1).

Corollary 7.10 (Structure of Anosov Diffeomorphisms). The set of Anosov

diffeomorphisms on a given compact manifold M is open in the C 1 -topology.
Furthermore, Anosov diffeomorphisms are structurally stable.

Proof. Openness follows directly from applying (T1) of Theorem 7.9, while
structural stability follows from (T2).

In some sense, we have pushed the concept of (uniform) hyperbolicity

as far as one can go with relatively direct arguments. Next, we are going to
look at cases, where hyperbolicity is lost in parametric families of dynamical
systems.

8 Local Bifurcations
Consider parameters p and the ODE

x′ = f (x, p), x ∈ Rd , p ∈ Rm . (8.1)

Definition 8.1. The appearance of a non-equivalent phase portrait under

parameter variation is called bifurcation.

39
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Example 8.2. Consider p = 1 = d and the ODE

x′ = px − x3 = x(p − x2 ). (8.2)

For p ≤ 0, x∗ = 0 is the unique equilibrium point; in this case, it is globally

asymptotically stable. For p > 0, x∗ = 0 is unstable and two new equi-
√
libria x± = ± p appear, which are locally asymptotically stable; see also
Figure 19(a). The bifurcation at (x, p) = (0, 0) is called pitchfork bifur-
cation. It is commonly found in symmetric systems; note the invariance
of the ODE under x 7→ −x. 
x x
(a) (b)

Sp≤0
Sp>0

p p

Figure 19: (a) (Supercritical) pitchfork bifurcation. (b) Fold bifurcation.

For (a)-(b), unstable invariant sets, here just equilibria, are indicated by
dashed lines, which is a common convention.

It makes sense to slightly tweak/weaken the definition of topological

equivalence for parameter-dependent systems, i.e., comparing (8.1) to

y ′ = g(y, q), y ∈ Rd , q ∈ Rm . (8.3)

Definition 8.3. The parametrized ODEs (8.1) and (8.3) are called (fiber)
topologically equivalent if
(D1) there exists a homeomorphism ξ : Rm → Rm , ξ(p) = q;
(D2) for fixed ξ(p) = q, the ODEs are topologically equivalent.
Fix a parameter p0 and let Sp0 be the maximally connected subset in
parameter space containing p0 with topologically equivalent phase portraits
and call Sp0 a stratum; see Figure 19(a).
Definition 8.4. A bifurcation diagram is a stratification of the param-
eter space (p ∈ Rm ) induced by topological equivalence, together with a
representative phase portrait for each stratum.

40
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Definition 8.5. An unfolding is a parametrized family of dynamical sys-
tems containing a bifurcation in a persistent way. The minimal number of
parameters for an unfolding is called its codimension.
Example 8.6. Consider m = 1 = d and the ODE

x′ = p − x2 = f (x, p). (8.4)

√
For p < 0, there are two equilibrium points x± = ± p. One checks by
looking at the linearization

X ′ = Dx f (x± , p)X = −2x± X = AX, X = X(t),

that x+ is stable and x− is unstable. At p = 0 the two equilibria coincide

and disappear for p > 0. A single eigenvalue λ of A passes through zero
at p = 0 and hyperbolicity of the equilibria is lost at the bifurcation point
(x, p) = (0, 0). The bifurcation diagram is shown in Figure 19(b). Note
carefully that a very similar analysis holds for the ODE

x′ = p + x2 = f (x, p). (8.5)

In both cases, we have a fold bifurcation (or saddle-node bifurcation).

Furthermore, (8.4) is a codimension-one unfolding of the saddle-node bi-
furcation since any sufficiently small perturbation (say in C k , k ≥ 2) of f
still exhibits the same bifurcation. 
The next result provides the key insight that we expect only certain
“lowest-order terms” to matter at a (local) bifurcation point.
Lemma 8.7. In a neighbourhood of (x, p) = (0, 0), the ODE

x′ = p + x2 + O(x3 ) (8.6)

is locally topologically equivalent to

x′ = p + x2 .

Remark : Recall the notation a(x) = O(b(x)) as x → 0 means limx→0 |a(x)|

|b(x)| <
∞. Furthermore, in (8.6) we also allow certain higher-order dependence in the
parameter, e.g., px3 .

Proof. We work in a neighbourhood of the origin. Re-write (8.6) as

y ′ = g(y, p) = p + y 2 + ψ(y, p), ψ = O(y 3 ). (8.7)

The manifold/curve of equilibria can locally be written as

M = {(y, p) : p = −y 2 + G(y)}, G(y) = O(y 3 )

41
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 using the implicit function theorem since ∂p g(0, 0) = 1. This implies there
√
exist equilibria y± (p) of (8.7) close to x± (p) = ± p; see Figure 19(b). To
construct the required homeomorphism hp : R → R, define hp (x) = x for
p ≥ 0 and for p < 0 use a linear transformation
!
hp (x) = a(p) + b(p)x, hp (x± (p)) = y± (p),

where a, b can be solved for implicitly using the last two conditions/equations.
Therefore, we have the required (fiber) topological equivalence of Defini-
tion 8.3.

Theorem 8.8 (Fold Bifurcation of Flows). Any smooth one-dimensional,

one-parameter ODE

x′ = f (x, p), x ∈ R, p ∈ R,

satisfying the conditions

(A1) f (0, 0) = 0, ∂x f (0, 0) = 0,

(A2) ∂xx f (0, 0) 6= 0, ∂p f (0, 0) 6= 0,

is locally near (x, p) = (0, 0) topologically equivalent to one of the two fol-
lowing ODEs
y′ = q ± y2, (8.8)
which are also called the topological norm form of the fold bifurcation.

Proof. First, we expand f in a Taylor series near the origin

f (x, p) = f0 (p) + f1 (p)x + f2 (p)x2 + O(x3 ),

where f0 (0) = f (0, 0) = 0 and f1 (0) = ∂x f (0, 0) = 0 by (A1). For the

geometric idea of the proof see Figure 20.
x y y

p q q

Figure 20: Transformation to normal form for a fold.

42
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Step 1: Consider a linear phase space coordinate change y = x+δ(p) =
x + δ, which yields

y ′ = x′ = f0 (p) + f1 (p)(y − δ) + f2 (p)(y − δ)2 + O((y − δ)3 )

   
= f0 (p) − f1 (p)δ + f2 (p)δ 2 + O(δ 3 ) + f1 (p) − 2f2 (p)δ + O(δ 2 ) y
+ [f2 (p) + O(δ)] y 2 + O(y 3 ).

The linear term in y can be dealt with requiring

  !
F (p, δ) := f1 (p) − 2f2 (p)δ + δ 2 ψ(p, δ) = 0,

where ψ is some smooth function. Note that F (0, 0) = 0, and ∂δ F (0, 0) =

−2f2 (0) = −∂xx f (0, 0) 6= 0 by (A2). Therefore, the Implicit Function
Theorem implies the existence of a unique smooth function δ = δ(p) such
that F (p, δ(p)) = 0 locally and

f1′ (0)
δ(p) = − p + O(p2 ).
2f2 (0)

The remaining ODE is now given by

 
y ′ = f0′ (0)p + O(p2 ) + [f2 (0) + O(p)] y 2 + O(y 3 ). (8.9)

Step 2: Consider as a new parameter q = q(p)

q = f0′ (0)p + p2 µ(p),

for some smooth function µ. Since q(0) = 0, q ′ (0) = f0′ (0) = ∂p f (0, 0) 6= 0
by (A2), the Inverse Function Theorem yields a unique smooth inverse
p = p(q) with p(0) = 0. Therefore, (8.9) can now be written as

y ′ = q + b(q)y 2 + O(y 3 ), (8.10)

where b(q) is smooth and b(0) = f2 (0) = ∂xx f (0, 0) 6= 0.

Step 3: Using a scaling ỹ = |b(q)|y, q̃ = |b(q)|q in (8.10) and dropping
the tildes gives

y ′ = q + sy 2 + O(y 3 ), s = sign(b(0)) = ±1,

so that Lemma 8.7 finishes the proof.

We interpret the conditions (A1)-(A2). In (A1), f (0, 0) = 0 just says we

have an equilibrium at a certain parameter value at a certain point (here

43
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 p = 0 = x but wlog we may always shift coordinates in parameter and
phase space). Hence, the main bifurcation condition is

∂x f (0, 0) = 0

of a single eigenvalue λ of the linearization located at zero so hyperbolicity

is lost. We also say that the fold bifurcation is of codimension one as a
single parameter suffices to describe the bifurcation diagram; see Section 11
for an example of a codimension two bifurcation. For (A2), the condition

∂xx f (0, 0) 6= 0

is a non-degeneracy condition, which ensures that the next non-trivial

Taylor term x2 does not vanish. The condition

∂p f (0, 0) 6= 0,

guarantees (exercise!) that the eigenvalue λ = λ(p) passes through the ori-
gin in C at non-zero speed, which is a so-called transversality condition;
see also Figure 21(a) for a transversality condition in another context. In
summary, one refers to (A2) as genericity conditions.

Example 8.9. Dropping the transversality condition ∂p f (0, 0) = 6 0 but

keeping the other conditions, then one possibility is that we may have the
situation
x′ = x(p − x), x ∈ R, p ∈ R,
which has the trivial equilibrium branch x ≡ 0 and the equilibrium branch
x∗ = p. We find x = 0 is locally asymptotically stable for p < 0 and
unstable for p > 0, while the reverse holds for x∗ = p. At p = 0, a trans-
critical bifurcation occurs, which nicely shows an example for exchange-
of-stability, as can be seen from the bifurcation diagram (exercise: draw
the diagram!). 

From standard linear algebra, it is evident that in a one-parameter

family of matrices, we may not only generically encounter the situation
that a single real eigenvalue λ passes through zero but also the case, when
a complex conjugate pair λ1,2 ∈ C, λ1 = λ2 passes through the imaginary
axis iR ⊂ C; see Figure 21(a).

Theorem 8.10 (Hopf Bifurcation). Any smooth generic two-dimensional,

one-parameter ODE

x′ = f (x, p), x ∈ R2 , p ∈ R,

44
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 y2 y2
(a) Im (b1) (b2)
C

Re p p
p<0 p>0
y1 y1
p=0

Figure 21: Hopf bifurcation. (a) Pair of transversally crossing eigenval-

ues. (b1) Supercritical Hopf bifurcation diagram. (b2) Subcritical Hopf
bifurcation diagram.

satisfying the conditions f (0, 0) = 0, Dx f (0, p) has a complex conjugate

pair of eigenvalues
λ1,2 (p) = µ(p) ± iω(p)
with µ(0) = 0, ω(0) > 0, is locally topologically equivalent to one of the two
following ODEs
 ′     
y1 q −1 y1 2 2 y1
′ = ± (y1 + y2 ) (8.11)
y2 1 q y2 y2

which are also called the topological norm form of the Hopf bifurca-
tion.

The Hopf and fold bifurcations are both of codimension one. Regarding
the requirement that the setting is generic, the correct transversality condi-
tion µ′ (0) 6= 0 for the Hopf bifurcation is easy to guess. The non-degeneracy
condition is a requirement, which guarantees non-vanishing cubic terms; see
Example 8.11. The transformation to normal form is discussed in Section 9.

Example 8.11. We analyze the dynamics of (8.11) denoting the sign of the
cubic term by l1 ∈ R \ {0}; it is called the first Lyapunov coefficient.
The first Lyapunov coefficient turns out to be actually computable from
the original vector field only [7] and the non-degeneracy condition for the
Hopf bifurcation is l1 6= 0. Two coordinate changes are useful to analyze
the Hopf normal form dynamics. First, consider the complexification
z = y1 + iy2 ∈ C and observe

z ′ = y1′ + iy2′ = q(y1 + iy2 ) + i(y1 + iy2 ) + l1 (y1 + iy2 )(y12 + y22 )
= (q + i)z + l1 z|z|2 ,

which is the complex Hopf normal form. Using z = reiϕ and

z ′ = r′ eiϕ + irϕ′ eiϕ = reiϕ (q + i + l1 r2 )

45
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 gives us the polar form
r′ = r(q + l1 r2 ),
(8.12)
ϕ′ = 1.
From (8.12), we obtain the two bifurcation diagrams in Figure 21(b):

• l1 < 0: A supercritical Hopf bifurcation occurs with the transition

from a stable equilibrium for q < 0 to an unstable equilibrium for
q > 0 and stable periodic orbits for q > 0.

• l1 > 0: A subcritical Hopf bifurcation occurs with the transition

from a stable equilibrium for q < 0 to an unstable equilibrium for
q > 0 and unstable periodic orbits for q < 0.

For a Hopf bifurcation not in normal form, the first Lyapunov coefficient is
computable using derivatives of the vector field up to and including order
three. The non-degeneracy condition is then l1 6= 0. 

Next, we briefly consider the codimension one cases for maps

xk+1 = g(xk , p), xk ∈ Rd , d ∈ {1, 2}, p ∈ R, (8.13)

with fixed points x∗ and associated linearized map X 7→ Dg(x∗ , p)X =

A(p)X. Denote the multipliers of A = A(p) by µ = µ(p) ∈ R for d = 1 and
by µ1,2 = µ1,2 (p) ∈ C for d = 2. The condition for loss of hyperbolicity is
now

|µ| = 1 or |µ1,2 | = 1 i.e., multipliers lie on the unit circle.

Theorem 8.12 (Fold Bifurcation of Maps). The topological normal form

for a generic fold bifurcation occurring when g(0, 0) = 0 and µ = 1, is
given by
y 7→ q + y ± y 2 , y ∈ R, q ∈ R, (8.14)
with genericity conditions ∂xx g(0, 0) 6= 0 and ∂p g(0, 0) 6= 0, i.e., under
these conditions, the map (8.13) is locally topologically conjugate to (8.14).

The proof is completely analogous to Theorem 8.8 and the analysis of

fixed points almost identical to Example 8.6. Hence we next focus on the
bifurcation, which has no direct analog for flows.

Theorem 8.13 (Flip Bifurcation). The topological normal form for a

generic flip bifurcation (or period-doubling bifurcation) occurring when
g(0, 0) = 0 and µ = −1, is given by

y 7→ −(1 + q)y ± y 3 , y ∈ R, q ∈ R, (8.15)

46
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 with genericity conditions
1 1
(∂xx g(0, 0))2 + ∂xxx g(0, 0) 6= 0 and ∂xp g(0, 0) 6= 0,
2 3
i.e., under these conditions, the map (8.13) is locally topologically conjugate
to (8.15).

Proof. Since g(0, 0) = 0 and ∂x g(0, 0) = −1 6= 1, we may apply the implicit

function theorem to
g(x, p) − x = 0
to obtain a locally unique fixed point x∗ = x∗ (p). Without loss of generality
we can assume using a preliminary coordinate transformation that x ≡ 0
is a fixed point for |p| sufficiently small. Now Taylor expansion gives

g(x, p) = g1 (p)x + g2 (p)x2 + g3 (x)x3 + O(x4 ),

where g1 (p) = −(1 + κ(p)) for some smooth function κ with κ(0) = 0 since
µ = −1. Using the transversality assumption

κ′ (0) = ∂xp g(0, 0) 6= 0

the inverse function theorem implies locally invertibility and we may con-
sider a new parameter q = κ(p). This yields map

x 7→ µ(q)x + a(q)x2 + b(q)x3 + O(x4 ), (8.16)

where µ(q) = −(1 + q). We consider a smooth coordinate change

x = y + δy 2 , δ = δ(q), (8.17)

which is locally invertible with inverse (exercise!) given by

y = x − δx2 + 2δ 2 x3 + O(x4 ). (8.18)

Inserting (8.17)-(8.18) into (8.16) yields

y 7→ µy + (a + δµ − δµ2 )y 2
+(b + 2δa − 2δµ(δµ + a) + 2δ 2 µ3 )y 3 + O(y 4 ).

The quadratic term is removable for sufficiently small q by setting

a(q)
δ(q) = µ2 (0) − µ(0) = 2 6= 0.
µ2 (q) − µ(q)

47
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 This implies we obtain a map
 
2a2
y 7→ µy + b + 2 y 3 + O(y 4 ) = −(1 + q)y + c(q)y 3 + O(y 4 ).
µ −µ
We can compute the smooth function c at q = 0
 2
2 1 1
c(0) = a (0) + b(0) = ∂xx g(0, 0) + ∂xxx g(0, 0) 6= 0
2 6
by the non-degeneracy assumption. This implies we may use the scaling
p
ỹ = y |c(q)|

and dropping the tildes yields the result up to higher-order terms, which
can be dropped using the same technique as in Lemma 8.7.

Example 8.14. We analyze the normal form (8.15) with the positive sign
for the cubic term. Fixed points satisfy

y = −(1 + q)y + y 3 = y(−1 − q + y 2 ) = f (y, q) (8.19)

so y∗ = 0 is locally for q near 0, the only fixed point with multiplier µ =

−1 − q so it is linearly stable for q < 0 and linearly unstable for q > 0. The
flip bifurcation occurs at q = 0. We consider the second-iterate map to find
periodic orbits
 
f 2 (y, q) = f ◦ f (y, q) = −(1 + q) −(1 + q)y + y 3 + (−(1 + q)y + y 3 )3
 
= (1 + q)2 y − (1 + q)(2 + 2q + q 2 ) y 3 + O(y 5 ).

This map y 7→ f 2 (y, q) has the trivial fixed point y∗ = 0 but also two
non-trivial fixed points for q > 0 sufficiently small
≈ √
y± = ± q + higher-order terms.
√
In fact, one checks explicitly via (8.19) that the points y± = ± 2 are stable
and constitute a periodic orbit of period two as shown in Figure 22(c). 
A similar, yet technically more involved treatment, can be given for the
two-dimensional case

xk+1 = g(xk , p) xk ∈ R2 , p ∈ R,

where a complex conjugate pair of multipliers µ1,2 (q) may pass through
the unit circle |µ1,2 (0)| = 1 (under the assumption µ1 6= µ2 ). In this case,
the bifurcation is called a Neimark-Sacker bifurcation. The Neimark-
Sacker bifurcation leads to a rather complicated invariant curve, which
generically bifurcates at q = 0.

48
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 y g(y) g(y)
(a) (b) (c)

q y y

Figure 22: Flip bifurcation. (a) Bifurcation diagram of the normal

form (8.15) with the positive sign for the cubic term. (b) Cobweb for
q < 0. (c) Cobweb for q > 0 showing a period-two orbit.

Example 8.15. Consider the Hénon map

   
x1 x2
7→ = g(x, p) (8.20)
x2 p1 − p2 x1 − x22

where p = (p1 , p2 ) ∈ R2 are parameters. Fixed points x∗ lying in {x1 = x2 }

can be found solving the quadratic equation x2 = p1 − p2 x2 − x22 so there
are at most two fixed points. Local stability can be computed using the
Jacobian  
0 1
A(p) = Dg(x∗ , p) = .
−p2 −2(x∗ )2
One approach would be to try to calculate the multipliers explicitly and
look for the relevant bifurcation conditions. It is helpful to introduce bifur-
cation functions, which encode the conditions compactly. For example,
for the fold bifurcation we must have

det(A(p) − Id) = 0

while for the flip and Neimark-Sacker bifurcations we may consider

det(A(p) + Id) = 0 and det(A(q)) − 1 = 0.

Then one may work out (exercise!) the algebraic conditions for a multiplier
crossing and try to check, when genericity conditions hold. 

9 Normal Forms
We now discuss a more general complementary approach to previous tech-
niques to compute a normal form for a smooth ODE x′ = f (x), x ∈ Rd .
Definition 9.1. Let Hk denote the space of d-component vector-valued
homogeneous polynomials of degree k.

49
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Assume x = 0 is an equilibrium and Taylor-expand the ODE

x′ = Ax + f (2) (x) + f (3) (x) + · · · , A = Df (0), (9.1)

where f (k) ∈ Hk for k ≥ 2. Consider a polynomial coordinate change

x = y + h(k) (y), y ∈ Rd , h(k) ∈ Hk , k ≥ 2 fixed,

so that y = x − h(k) (x) + O(|x|k+1 ). We compute the transformed ODE

y ′ = x′ − Dx h(k) (x)x′ + O(|x|k+1 )

 
  k
X
= Id − Dx h(k) (x) + O(|x|k ) Ax + f (k) (x) + O(|x|k+1 )
j=2
 
k−1
X  h i
f (k) (y) + f (k) (y) − Dy h(k) (y)Ay − Ah(k) (y)  + O(|y|k+1 ),
 
= Ay +
 }
j=2 | {z
=:Lh(k) (y)

where L : Hk → Hk (exercise!) is a linear operator. The terms of order

less than k have remained invariant. We may aim to eliminate the order-k
terms by solving the homological equation

Lh(k) = f (k) .

Terms in L(Hk ) = range(L) can be eliminated, denote them by g (k) . The

remaining terms r(k) = f (k) − g (k) in a complement Rk to Hk are called
resonant terms.
Theorem 9.2 (Poincaré Normal Form). There exists a local polynomial
near-identity coordinate change

x = y + h(2) (y) + h(3) (y) + · · · + h(k) (y) + O(|x|k+1 ), h(k) ∈ Hk ,

which transforms (9.1) into

y ′ = Ay + r(2) (y) + r(3) (y) + · · · + r(k) (y) + O(|x|k+1 ), r(k) ∈ Rk ,

i.e., only resonant terms remain.

Proof. The discussion before the theorem can be applied recursively, start-
ing with k = 2, then k = 3, and so on.
Remark : The operator L can also be viewed as an adjoint ad(Ax) induced by
Ax via the Lie bracket [·, ·]

Lh = ad(Ax)(h) = [h, Ax] := (Dh)Ax − (DAx)h.

50
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Lemma 9.3. The complement Rk to Hk is orthogonal in a scalar product
h·, ·ik if it is chosen as

ker(L∗ ) = {h ∈ Hk : L∗ u = 0}

where L∗ is the adjoint of L, i.e., hLu, vik = hu, L∗ vik . If there are no
generalized eigenvectors for the zero eigenvalue of L we have that

ker(L) = {h ∈ Hk : Lu = 0}

and range(L) also span Hk , so we can also select Rk = Nk in this case.

Proof. Recall that the Fredholm Alternative Theorem says in our con-
text that either Lh = b has a solution b ∈ range(L) or L∗ h∗ = 0 has a solu-
tion h∗ ∈ Hk with hh∗ , bik 6= 0, which proves the first part. For the second
part, no generalized eigenvectors imply that ker(L) ⊕ range(L) = Hk .

In practice, we identify Hk using coordinates with some RN (k) and L

with a matrix M ∈ RN (k)×N (k) so that we can use the standard Euclidean
inner product.

Theorem 9.4 (Hopf Normal Form). A smooth planar system (9.1) at a

Hopf bifurcation, i.e., the eigenvalues of A are iω with ω > 0, can be locally
transformed into the Poincaré normal form
   
′ 0 −1 2 2 p1 −p2
y = y + (y1 + y2 ) y + O(|y|4 ). (9.2)
1 0 p2 p1

Proof. We write (y1 , y2 ) = (x, y) throughout the proof for notational sim-
plicity. Consider quadratic terms (k = 2), fix an (ordered) polynomial
basis {x2 , xy, y 2 } of H2 . View a vector field h = h(2) as a map h : R2 →
T R2 ≃ R2 with basis coordinates ∂x , ∂y ; for example, write (2xy, x2 )⊤ as
(2xy)∂x + (x2 )∂y . Wlog (upon scaling time if necessary) we have
    
x x
Lh = DA h − (Dh)A
y y
     
0 −1 h1 ∂x h1 ∂y h1 −y
= − .
1 0 h2 ∂x h2 ∂y h2 x

Plugging in the basis elements, we find

 2      2     
x 2xy xy y − x2 y2 −2xy
L = , L = , L = ,
0 x2 0 xy 0 y2

51
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 for the first component, and for the second component
           
0 −x2 0 −xy 0 −y 2
L = , L = , L = .
x2 2xy xy y 2 − x2 y2 −2xy

Of course, it is a lot more convenient to view this in matrix form and

identify L with some M (2) ∈ R6×6 . Using the coordinate ordering in R6 by

(x2 ∂x , xy∂x , y 2 ∂x , x2 ∂y , xy∂y , y 2 ∂y )

we find  
0 −1 0 −1 0 0
 +2
 0 −2 0 −1 0 
 0 +1 0 0 0 −1 
M (2) = .
 +1
 0 0 0 −1 0 
 0 +1 0 +2 0 −2 
0 0 +1 0 +1 0
Since det(M (2) ) = −9 6= 0, all quadratic terms can be eliminated. The
procedure can now be carried out for k = 3. We use the ordering of
coordinates in R8

(x3 ∂x , x2 y∂x , xy 2 ∂x , y 3 ∂x , x3 ∂y , x2 y∂y , xy 2 ∂y , y 3 ∂y ).

The matrix M (3) of L for k = 3 turns out to be (exercise!)

 
0 −1 0 0 −1 0 0 0
 +3
 0 −2 0 0 −1 0 0 

 0
 +2 0 −3 0 0 −1 0 

 0 0 +1 0 0 0 0 −1 
M (3) = .
 +1
 0 0 0 0 −1 0 0 

 0
 +1 0 0 +3 0 −2 0 

 0 0 +1 0 0 +2 0 −3 
0 0 0 +1 0 0 +1 0

The columns 1, 2, 3, 4, 5, 8 are linearly independent. Using the first part of

Lemma 9.3, the adjoint/transposed matrix (M (3) )⊤ has dim(ker((M (3) )⊤ )) =
2 with basis vectors

v ∗ = (3, 0, 1, 0, 0, 1, 0, 3)⊤ and w∗ = (0, −1, 0, −3, 3, 0, 1, 0)⊤

so that R3 = span(v ∗ , w∗ ) gives resonant terms

 3   
(3) 3x + xy 2 −x2 y − 3y 3
r∗ = p1 + p 2 ,
x2 y + 3y 3 3x3 + xy 2

52
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 which does not match (9.2). Using the second part of Lemma 9.3, we find
that ker(M (3) ) is spanned by

v = (1, 0, 1, 0, 0, 1, 0, 1)⊤ and w = (0, −1, 0, −1, 1, 0, 1, 0)⊤ ,

which produces the correct resonant terms in (9.2).

The last proof showed that the Poincaré normal form is not unique
due to a choice of basis. We also verified the algebra necessary for The-
orem 8.10, yet Poincaré normal forms do not yield the first Lyapunov
coefficient. The method of normal forms can be immediately applied to
parameter-dependent systems written in the form

x′ = f (x, p),
p′ = 0.

Theorem 9.5 (Bogdanov-Takens Bifurcation). Any generic planar two-

parameter system with an equilibrium at x = 0 and a linearization having a
double-zero eigenvalue λ1,2 (0) = 0 is locally topologically equivalent to one
of the following normal forms
 ′       
y1 0 0 1 y1 0
= + + . (9.3)
y2′ q1 q2 0 y2 y12 ± y1 y2

The algebraic structure of (9.3) can be derived from the Poincaré normal
form making the ansatz
 
′ 01
x = x + f (2) (x) + f (3) (x) + · · · ,
00

which leads already to quite a substantial calculation. Before we discuss

the dynamics of (9.3), two issues remain. We have to understand whether
we are allowed to drop higher-order terms. This is easy to prove for fold
and Hopf bifurcations explicitly (see Section 8), while for (9.3) we sketch
the proof in Section 11. The second issue is even more fundamental as it
applies to all bifurcations, so we tackle it next.

10 Center Manifolds
So far, we have crucially assumed to work in the minimal dimension for
the smooth ODE

x′ = Ax + f (2) (x) + f (3) (x) + · · · = f (x), x ∈ Rd , f ∈ C k , (10.1)

53
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 where A = Df (0) and x∗ = 0 is an equilibrium. Suppose the spectrum of
A has a center eigenspace

E c (0) = span{vj : Avj = λj vj , Re(λj ) = 0},

which is dc -dimensional. In particular, A has

• dc eigenvalues λc with Re(λc ) = 0;

• ds eigenvalues λs with Re(λs ) < 0;

• du eigenvalues λu with Re(λu ) > 0.

If dc = 0 the Stable/Unstable Manifold Theorem applies; see Figure 23(a)

for the idea, how to deal with the additional center directions.
x2 y
(a) (b)

Es Wc

W c (0, 0)
Ec x1 x

Figure 23: (a) Basic idea of center manifold W c . (b) Illustration of center
manifolds for (10.2) showing lack of uniqueness and smoothness. Note that
the sketch is zoom of the situation near (x, y) = (0, 0).

Theorem 10.1 (Center Manifold Theorem). Consider (10.1) with k ∈ N,

then there exists an invariant dc -dimensional C k -manifold Wloc
c (0) tangent
c
to E (0) locally. Moreover, any trajectory staying inside a neighbourhood
c (0) as t → +∞ (t → −∞).
U = U (0) for t ≥ 0 (t ≤ 0) converges to Wloc
c (0) is called a center manifold.
Definition 10.2. Wloc

The proof of Theorem 10.1 can be accomplished using similar, yet more
subtle, fixed-point arguments as sused in the proof of the Stable/Unstable
Manifold Theorem 3.8. We shall often write Wloc c (0) as W c (0) or just W c .

Theorem 10.3 (Center Manifold Limitations). The following holds:

(T1) Center manifolds need not be unique.

54
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (T2) Center manifolds need not be C ω , even if f ∈ C ω , i.e., center mani-
folds are in general not analytic.
Proof. Consider the counter-example
x′ = x2 ,
(10.2)
y ′ = −y,
with (x(0), y(0)) = (x0 , y0 ). Clearly, (0, 0) is an equilibrium and the eigen-
values of the linearization are λ = 0, −1. The counter-example can be
solved explicitly
1
x(t) = 1 , y(t) = y0 e−t .
x0 − t
See Figure 23(b) for the phase portait. We claim W c (0)p = {(x, y) ∈ R2 :
y = hp (x)} for 

p exp x1 for x < 0,
hp (x) :=
0 for x ≥ 0,
is a one-parameter family (for p ∈ R) of center manifolds. Indeed, the
tangent space is TWpc (0) = {y = 0} = E c (0). Each Wpc (0) is invariant
since for p fixed, {y = 0, x > 0} is a trajectory and for x < 0 we may
simply solve for t in x(t) above to obtain that
y = pe1/x
is also a trajectory for x < 0. Hence, W c (0) is not unique and it is not C ω
for p 6= 0.
Remark : Center manifolds do not even have to be C ∞ , even if f is analytic. We
leave it as an exercise to check that the system
x′1 = −x2 x1 − x31 , x′2 = 0, y ′ = −y + x21
is one possible required counterexample.
Separating the critial and non-critical components, we may write (10.1)
x′ = Bx + F (x, y), F = O(x2 , xy, y 2 )
(10.3)
y ′ = Cy + G(x, y), G = O(x2 , xy, y 2 )

where B ∈ Rdc ×dc with spec(B) ⊂ iR and C ∈ R(ds +du )×(ds +du ) with
spec(C) ∩ iR = ∅.
Theorem 10.4 (Center Manifold Reduction). The ODE (10.3) is locally
topologically equivalent near 0 to
x′ = Bx + F (x, H(x)),
(10.4)
y ′ = Cy,
c (0) = {(x, y) ∈ Rdc × Rds +du : y = H(x)} with H = O(x2 ).
where Wloc

55
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Remark : Essentially analogus center manifold results hold for maps, where the
critical components in E c (0) are associated to multipliers on the unit circle.
In practice, the most important case is du = 0, which we assume for the
rest of this section. We can only apply normal form and bifurcation theory
on W c if we can calculate H, or at least an approximation h : Rdc → Rds .
Making the ansatz y = h(x) in (10.3) yields

y ′ = Dh(x) x′ = Dh(x) [Bx + F (x, h(x))] = Ch(x) + G(x, h(x)).

Definition 10.5. The partial differential equation (PDE)

N (h(x)) = Dh(x) [Bx + F (x, h(x))] − Ch(x) − G(x, h(x)) = 0 (10.5)

for h with boundary conditions h(0) = 0, Dh(0) = 0 is called the (center

manifold) invariance equation.

Theorem 10.6 (Approximation of Center Manifolds). If a map h can be

found such that h(0) = 0, Dh(0) = 0, which satisfies

N (h(x)) = O(|x|p ) for some p > 1 as |x| → 0,

i.e., the invariance equation is solved approximately, then

H(x) = h(x) + O(|x|p ) as |x| → 0, (10.6)

c (0) = {y = H(x)} is the center manifold.
where Wloc

Proof. (Sketch) The proof is an extension of ideas to prove existence of

invariant manifolds, so we only provide the main idea. Consider (10.3),
suppose

H : Rdc → Rds , H(0) = 0, H locally Lipschitz,

is the center manifold. Let x(t, x0 , H) denote the solution of

x′ = Bx + F (x, H(x)), x(0, x0 , H) = x0

as in (10.4). One now checks (non-trivial) that H can be constructed using

Z 0
KH(x0 ) = e−Cs G(x(s, x0 , H), H(x(s, x0 , H))) ds (10.7)
−∞

and looking for a fixed point of the operator K on the space K of Lipschitz
functions with the sup-norm; cf. proof of Theorem 3.8 and note that (10.7)
essentially removes all decaying directions so only the center directions can
remain. To obtain the approximation h, we study a modified fixed point

56
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 problem. Consider θ ∈ Cc1 (Rdc , Rds ) with θ(x) = h(x) for |x| small, and
define
N (x) := Dθ[Bx + F (x, θ(x))] − Cθ(x) − G(x, θ(x))
and observe N (x) = O(|x|p ) as |x| → 0. We know H is a fixed point of K
and define a new operator by

Lw = K(w + θ) − θ.

It can be shown (non-trivial) that L is a contraction mapping on

n o
L := w ∈ K : |w(x)| ≤ c|x|p for all x ∈ Rdc .

We get a fixed point w∗ such that w∗ = K(w∗ + θ) − θ. Since θ(x) = h(x)

locally and K has a fixed point H, the approximation (10.6) follows.

Hence, substituting a formal power series into the invariance equation

leads to a rigorous approximation algorithm for center manifolds.
Example 10.7. Consider the planar ODE

x′ = xy,
(10.8)
y ′ = −y + px2 ,

which has an equilibrium at 0 and is already in the standard form (10.3)

with B ≡ 0 and C ≡ −1. We make a third-order ansatz

y = h(x) = a2 x2 + a3 x3 + O(|x|4 ) (note: h(0) = 0 = h′ (0))

in (10.5) to obtain

0 = h′ (x) [xh(x)] + h(x) − px2

 
= (2a2 x + 3a3 x2 + O(|x|4 ))x a2 x2 + a3 x3 + O(|x|4 )
+(a2 − p)x2 + a3 x3 + O(|x|4 )
= (a2 − p)x2 + a3 x3 + O(|x|4 ).

So we can select a2 = p, then all quadratic terms vanish, and then take
a3 = 0 to eliminate the cubic term to obtain
c
Wloc (0) = {(x, y) ∈ U = U (0) : y = h(x) = px2 + O(|x|4 )}.

The reduced dynamical system on the center manifold is

x′ = px3 + O(x5 )

and the phase portraits for p > 0 and p < 0 are shown in Figure 24. 

57
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 y y
(a) (b)

x x

Figure 24: Sketch of the center manifolds for Example (10.7). (a) p = p1 >
0. (b) p = p2 < 0 with |p2 | < p1 .

Example 10.8. Consider the planar ODE

x′ = y − 1,
(10.9)
y ′ = −(y − 1) + px2 + x(y − 1),

and note that (0, 1) is an equilibrium and the linearized problem

 ′     
X 0 1 X X
′ = =: A
Y 0 −1 Y Y

at this point has eigenvalues 0 and −1. However, the system (10.9) is not
in standard form. Setting (x, y) = (x̃, ỹ + 1), inserting into (10.9), and
dropping the tiles yields

x′ = y,
(10.10)
y ′ = −y + px2 + xy.

Next, we use the eigenvectors of A to diagonalize via the transformation

     
x x̃ 1 1
=T , T = = T −1 .
y ỹ 0 −1

Inserting this coordinate change in (10.10), and dropping the tiles, yields
the standard form
 ′      
x 0 0 x 1 1 0
= + .
y′ 0 −1 y 0 −1 p(x + y)2 − (x + y)y

Now one may proceed analogously to Example 10.7 to find a center manifold
approximation (exercise!). 

58
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 11 Higher Codimension
We return to our goal to unfold the Bogdanov-Takens (double-zero eigen-
value) bifurcation. At the bifurcation point q = 0 the normal form (9.3)
is
y1′ = y2 ,
(11.1)
y2′ = y12 ± y1 y2 ,
where we have dropped higher-order terms. Do third-order Taylor terms
matter? Unfortunately, the situation is not as simple as for the fold in
Lemma 8.7 and it helps to consider a more abstract viewpoint using the
following definition.
Definition 11.1. Let F ∈ C l (Rd , Rd ), F (0) = 0, and consider k ≤ l, k ∈ N.
Then the k-jet Jk F (y) is the Taylor polynomial of F up to and including
order k.
Therefore, our question is, whether the 2-jet (y2 , y12 ±y1 y2 ) determines
the topological type of the vector field, i.e., whether including higher-order
jets yield vector fields locally topologically not equivalent to the 2-jet. The
following example illustrates the technique we need to prove the 2-jet de-
terminacy of (11.1).
Example 11.2. Consider the planar ODE
y1′ = y12 − 2y1 y2 =: F1 (y1 , y2 ),
(11.2)
y2′ = y22 − 2y1 y2 =: F2 (y1 , y2 ).

where 0 = (0, 0) is a non-hyperbolic equilibrium point and F = (F1 , F2 )⊤

is its own 2-jet. Introduce polar coordinates
Φ : S1 × [0, ∞) → R2 , Φ(θ, r) = (r cos θ, r sin θ),
and then one computes the transformed vector field
F̂ (θ, r) = (DΦ−1(θ,r) ◦ F ◦ Φ)(θ, r)
 sin θ cos θ   2 
− r r r (cos2 θ − 2 cos θ sin θ)
=
cos θ sin θ r2 (sin2 θ − 2 cos θ sin θ
 
3r cos θ sin θ(sin θ − cos θ)
= 1 2 .
4 r (cos θ + 3 cos(3θ) + sin θ − 3 sin(3θ))

Note that F̂ has a circle of equilibrium points S1 × {r = 0}. Since Φ is

a diffeomorphism outside of 0, it follows that F̂ on S1 × (0, ∞) and F on
R2 − {0} are topologically conjugate and the same holds for
1
F̄ (θ, r) = F̂ (θ, r)
r

59
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 as division by any power of r just corresponds to a re-parametrization of
time on orbits on the punctured plane (non-trivial exercise!). Yet, one
checks easily that S1 × {r = 0} now contains six equilibria as shown in
Figure 25.
y2

Φ
F F̄

y1

R2 S1 × [0, ∞)

Figure 25: Polar blow-up at (0, 0) of equation (11.2).

All six equilibria are hyperbolic. The Hartman-Grobman Theorem 4.5

implies that near each of these equilibria, the 1-jet determines the vec-
tor field F̄ . Hence, inverting Φ on the punctured plane implies that F is
determined by its 2-jet. 
Definition 11.3. The blow-up (or geometric desingularization) of a
vector field F ∈ Rd is a map
Φ : Sd−1 × [0, ∞) → Rd ,
with associated l-times desingularized blown-up vector field
1
F̄ (θ, r) = (DΦ−1
(θ,r) ◦ F ◦ Φ)(θ, r).
rl
for some l ∈ N. Φ−1 is also known as the blow-down map.
Theorem 11.4. The two-jet in (11.1) is determinate, i.e., k-jets for k ≥ 3
can be dropped in any unfolding of the Bogdanov-Takens bifurcation.
The proof of Theorem 11.4 requires three iterated blow-ups and long
calculations, so we omit it here. However, the idea is precisely the same
as in Example 11.2. The normal form (9.3) is a possible unfolding so it
remains to analyze
y1′ = y2 ,
(11.3)
y2′ = q1 + q2 y1 + y12 − y1 y2 ,
where we picked the negative sign in the last term; the positive sign case
works analogously.

60
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 q2 y2
(a) fold F (b)

q1 y1

homoclinic Hopf
H

Figure 26: (a) Bifurcation diagram of the Bogdanov-Takens bifurcation

normal form (11.3); for the fold and Hopf see Proposition 11.5 and for an
explanation regarding homoclinic bifurcations see Section 12. (b) Phase
portrait of the Hamiltonian system (11.7).

Proposition 11.5. The ODE (11.3) has a curve of fold bifurcations given
by
F = {q ∈ R2 : 4q1 − q22 = 0}, (11.4)
and a curve of supercritical Hopf bifurcations

H = {q ∈ R2 : q1 = 0, q2 < 0}

as shown in Figure 26(a).

Proof. Equilibria y∗± of (11.3) satisfy y2 = 0 and 0 = q1 + q2 y1 + y12 so

p !  
± q2 ± q22 − 4q1 q2 ± ν
y∗ = − , 0 =: − ,0 ,
2 2

which immediately yields the fold curve (11.4) upon inspection. One checks
that crossing F ∩ {q2 > 0} generates an unstable node y∗− and a saddle y∗+ .
Crossing F ∩ {q2 < 0} yields a stable node y∗− and a saddle y∗+ ; note
that this also explains the name saddle-node bifurcation as a synonym
for the fold bifurcation. It is straightforward to check (exercise!) that
y∗− undergoes a Hopf bifurcation on H but is only a neutral saddle for
{q ∈ R2 : q1 = 0, q2 > 0}.

Just moving counter-clockwise in the bifurcation diagram starting at the

neutral saddle line {q1 = 0, q2 > 0}, get the phase portraits in Figure 26(a)

61
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 easily via Proposition 11.5 except that between H and F ∩ {q2 > 0} we
must somehow destroy the periodic orbit generated in the Hopf bifurcation
without any local bifurcations. Let us, formally, motivate what we expect.
For ν > 0, we translate y∗− to the origin via (ỹ1 , ỹ2 ) = (y1 − (y∗− )1 , y2 ).
Upon dropping the tildes, this yields the ODE
y1′ = y2 ,
(11.5)
y2′ = y1 (y1 − ν) − ((y∗− )1 y2 + y1 y2 ).
Near the bifurcation point, we have to deal with small ν so we use a scaling

Y1 = ν −1 y1 , Y2 = ν −3/2 y2 , T = ν 1/2 t, α1 := (y∗− )1 ν −1/2 , α2 := ν 1/2 .

This transformation is essentially another blow-up since it provides a zoom

for ν small near the equilibrium at the origin and gives
Y1′ = Y2 ,
(11.6)
Y2′ = Y1 (Y1 − 1) − (α1 y2 + α2 y1 y2 ).

In (11.5) the distance between the two equilibria is ν while it is precisely 1

in (11.6). Let us assume we are interested in α1,2 = 0, then 11.6 becomes
a Hamiltonian system (cf. Example 2.9)
∂H
Y1′ = Y2 = ∂Y , Y12 Y22 Y13
′
2
∂H H(Y ) = + − . (11.7)
Y2 = Y1 (Y1 − 1) = − ∂Y1 , 2 2 3

The phase portrait is shown in Figure 26(b). In particular, one observes

that the level set {H(Y ) = 1/6} contains an orbit asymptotic in forward
and backward time to the saddle at Y∗+ = (1, 0). Like a the saddle-to-saddle
connections we encountered in Section 6, this situation is not structurally
stable but may occur in parametrized families. This motivates us to look
at certain global bifurcations.

12 Global Bifurcations
Definition 12.1. Let (X , T , φt ) for T = R or T = Z be a dynamical
6 y∗ .
system with steady states x∗ , y∗ with x∗ =
(D1) An orbit γ = γ(t) is called a heteroclinic orbit from x∗ to y∗ if

lim γ(t) = x∗ and lim γ(t) = y∗ .

t→−∞ t→+∞

(D2) An orbit γ = γ(t) is called a homoclinic orbit to x∗ if

lim γ(t) = x∗ .
t→±∞

62
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 W u (0)
(a) (b) (c)

W s (0)
Σu (0)
γ
x∗ γ
x∗ y∗ 0 Σs (0)
Σ

Figure 27: (a) Heteroclinic orbit. (b) Homoclinic orbit. (c) Sketch for
stable/unstable manifolds of a broken homoclinic and for the construction
of the split function.

Examples of heteroclinic/homoclinic orbits and the role of the sta-

ble/unstable manifolds are shown in Figure 27(a)-(b). It is easy to check
(exercise!) that a given homoclinic/heteroclinic orbit to a hyperbolic equi-
librium is structurally unstable. Consider the ODE situation

x′ = f (x, p), x ∈ Rd , p ∈ R, (12.1)

where f is smooth, f (0, p) = 0, and x∗ = 0 is a hyperbolic saddle point.

Definition 12.2. Suppose (12.1) has a homoclinic orbit to x∗ = 0 at p = 0.
Let Σ be a transversal rectangular cross-section to W s (0) in a sufficiently
small neighborhood U = U (0). Let

Σu := W u (0) ∩ Σ and Σs := W s (0) ∩ Σ

Let q be a coordinate on Σ with axis Q ⊂ Σ such that {q = 0} = Σs . Then

a smooth function

ξ(p) : R → R, ξ(p) = q for q ∈ Σu ∩ Q,

with ξ(0) = 0 is called a split function, or split functional.

Remark : There is a choice of direction for the coordinate q but the following
results will be independent of this choice, so we shall not pick a rule to fix a
direction.
In particular, the split function measures a signed distance of the un-
stable manifold on Σ to the stable manifold; see Figure 27(c). The split
function is one example of a bifurcation function; cf. Example (8.15).
Theorem 12.3 (Andronov-Leontovich Theorem). Consider the ODE (12.1)
for d = 2 and let λ1,2 (p) be the eigenvalues of x∗ = 0 with λ1 (0) < 0 <
λ2 (0). Let γ0 (t) be a homoclinic orbit to 0. Assume the following genericity
conditions hold:

63
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (A1) (“non-resonance”) We have σ(0) := λ1 (0) + λ2 (0) 6= 0.

(A2) (“breaking”) The split function ξ(p) = q satisfies ξ ′ (0) 6= 0.

Then a unique limit cycle Γq , lying in a neighbourhood of γ0 ∪{0}, bifurcates

from γ0 upon variation of p. Furthermore, the cycle is stable and exists for
σ(0) < 0, while the cycle is unstable for σ(0) > 0; see Figure 28.

(a) (b) (c)

Γq

γ0

Figure 28: Illustration of Theorem 12.3 for σ(0) < 0. (a)-(c) correspond to
a variation of p through the bifurcation point at p = 0 shown in (b)

Remark : As a particular application, Theorem 12.3 explains, what happens on

the homoclinic bifurcation curve in Figure 26(a) for the Bogdanov-Takens bifur-
cation; calculating the curve precisely in parameter space is possible, yet very
difficult.

Definition 12.4. We call σ(p) := λ1 (p) + λ2 (p) the saddle quantity.

Proof. (of Theorem 12.3) Step 1: We drop the parameter-dependence for

now due to notational simplications and we aim to simplify the system
locally near x∗ = 0 in a neighbourhood V = V(0). Using a linear coordinate
change, we may write (12.1) as

x′1 = λ1 x1 + f1 (x1 , x2 ),
(12.2)
x′2 = λ2 x2 + f2 (x1 , x2 ).

By the Stable/Unstable Manifold Theorem (3.8), we have

W s (0) = {x ∈ V : x2 = S(x1 )}, S(0) = 0 = S ′ (0), (12.3)

u ′
W (0) = {x ∈ V : x1 = U (x2 )}, U (0) = 0 = U (0), (12.4)

so that we may rectify the manifolds using the coordinate change

(y1 , y2 ) = (x1 − U (x2 ), x2 − S(x1 ))

in (12.2); see Figure 29. Upon scaling coordinates, we obtain that in the
standard rectangle Ω = [−1, 1] × [−1, 1], the dynamics is given by

y1′ = λ1 y1 + y1 g1 (y1 , y2 ),
(12.5)
y2′ = λ2 y1 + y2 g2 (y1 , y2 ).

64
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Wu
y2 ζ
τ2 Σh
Φ Ws
τ1
y0 (ξ0 , ζ0 )

y1 ξ

Σv
Ω = [−1, 1] × [−1, 1] Ω = [−1, 1] × [−1, 1]
Figure 29: Local construction near the saddle to prove Theorem 12.3.

where g = (g1 , g2 ) = O(|y|).

Step 2: We claim that there exists a C 1 -linearization near the saddle
so that via a diffeomorphic coordinate change

Φ : Ω → Ω, Φ(y1 , y2 ) = (Φ1 (y), Φ2 (y)) =: (ξ, ζ), (12.6)

one may transform (12.5) into the form

ξ ′ = λ1 ξ,
(12.7)
ζ ′ = λ2 ζ.

To define Φ, fix y0 ∈ Ω in the first quadrant; the other quadrants work

similarly. Let τ1 and τ2 be the absolute values of the first exit times of
Ω of y(t) through the right and top boundaries starting at time t = 0;
see Figure 29. Consider the same construction for (12.7) and obtain Φ by
mapping y0 to (ξ0 , ζ0 ) having the same pair of exit times; see Figure 29. Set
Φ(0) = 0. Φ leaves {y1 = 0} ∪ {y2 = 0} invariant. One checks (exercise!)
that it is a homeomorphism, which is C 1 and invertible away from y = 0.
We claim
DΦ(0) = Id (12.8)
so that Φ is indeed a diffeomorphism. This claim is difficult to check. We
start by considering the linear system (12.7), which can be solved explicitly

ξ(t) = ξ(t0 )eλ1 (t−t0 ) and ζ(t) = ζ(t0 )eλ2 (t−t0 ) .

Setting t0 = 0 and ξ(0) = ξ = Φ1 (y), ζ(0) = ζ = Φ2 (y) we must have

1 = ξ(−τ1 ) = ξe−λ1 τ1 , which yields ξ = eλ1 τ1 . Therefore, we differentiate
ξ = Φ1 (y) to obtain

DΦ1 (y) = D(eλ1 τ1 ) = λ1 eλ1 τ1 Dτ1 . (12.9)

65
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 The solution of the nonlinear system (12.5) can be written as a mild/integral
solution, inserting −τ1 gives
Z −τ1
−λ1 τ1
1 = y1 (−τ1 ) = y1 e + eλ1 (−τ1 −s) y1 (s)g1 (y(s)) ds
0
=: e−λ1 τ1 [y1 + R(τ1 )]
Differentiating y1 = eλ1 τ1 − R(τ1 ) gives
 
λ1 τ1 ′ λ1 τ1 1 −λ1 τ1 ′
(1, 0) = λ1 e Dτ1 − R Dτ1 = λ1 e 1− e R Dτ1 .
λ1
Solving for Dτ1 , and using (12.9) leads to
1
DΦ1 (y) = h i (1, 0). (12.10)
1 −λ1 τ1 ′
1− λ1 e R

By using the Leibniz integral rule we get

Z −τ1
d
e−λ1 τ1 R′ = e−λ1 τ1 e−λ1 s y1 (s)g1 (y(s)) ds = y1 (−τ1 )g1 (y(−τ1 ))
dτ1 0
As y → 0, the integral solution of the nonlinear system shows that we must
have y1 g1 (y) → 0 as the term is higher-order using the bound g = O(|y|)
so e−λ1 τ1 R′ → 0. So we may conclude that DΦ1 (y) → (1, 0) uniformly as
y → 0. Similarly, one may proceed for the derivative DΦ2 to obtain (12.8).
The local coordinate change induced by Φ can be extended to a global
smooth coordinate change.
Step 3: We construct a composition of maps. Define two sections
Σv := {(ξ, ζ) ∈ R2 : ζ ∈ [0, 1], ξ = 1}, Σh := {(ξ, ζ) ∈ R2 : ξ ∈ [0, 1], ζ = 1},
and first consider the local map L : Σv → Σh induced by the flow of (12.7),
which can be explicitly calculated for ζ ∈ (0, 1] as
λ
− λ1
ξ = L(ζ) = ζ 2 , M (0) := 0.
The global map G : Σh → Σv induced by the flow is smooth and has the
general form
ζ = G(ξ) = q + aξ + O(ξ 2 ),
where q is the value of the split function and a > 0 since orbits cannot
intersect. Hence, we have constructed a Poincaré map
λ
− λ1
P (ζ) := G ◦ L(ζ) = q + aζ 2 + ··· .
For |q| small, the fixed points of this map can be analyzed as shown in
Figure 30 and this yields the claimed result.

66
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
(a) P (ζ) (b) P (ζ)

q=0

q=0

ζ ζ

Figure 30: Fixed points of the map P (ζ) upon varying q through q = 0
(solid black curve); the diagonal is shown in grey. (a) An unstable nonzero
fixed point (periodic orbit) occurs. (b) A non-zero stable fixed point (peri-
odic orbit) occurs.

The C 1 -linearization, which we investigated “by-hand” in Step 2, actu-

ally works in R2 without any nonresonance conditions. However, for general
dimensions it is a special case of a very general class of theorems, which
guarantees C k -conjugacy in the absence of resonances. We state one result
for maps but similar results hold for flows.

Theorem 12.5 (Sternberg Linearization). Suppose g ∈ C ∞ (Rd , Rd ) is a

map with fixed point x∗ and Dg(x∗ ) = diag(λ) with eigenvalues {λj }dj=1 and
the nonresonance condition

λj 6= λk for all multiindices k ∈ (N0 )d , |k| ≥ 2,

is satisfied, then g is C ∞ -conjugate to its linear part.

Theorems analogously to the Andronov-Leontovich Theorem also hold

for saddle points with real eigenvalues for d > 2, e.g., by using center-
manifold-type arguments along the entire homoclinic orbit. However, sad-
dles with complex eigenvalues, e.g., a saddle-focus in R3 , can be fundamen-
tally different!

Theorem 12.6 (Shilnikov Homoclinic Orbit). Consider the ODE (12.1)

for d = 3 with a saddle-focus x∗ = 0 with Re(λ2,3 (0)) < 0 < λ1 (0). Let
γ0 (t) be a homoclinic orbit to 0. Assume the following genericity conditions
hold:

(A1) (“Shilnikov condition”) σ(0) := λ1 (0) + Re(λ2,3 (0)) > 0.

(A2) (“nonresonance”) λ2 (0) 6= λ3 (0).

67
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Then for all |p| sufficiently small there exist infinitely many saddle limit
cycles in a neighbourhood of γ0 ∪ {0}.

Proof. We only construct the geometry of the Poincaré map. The final
result is going to follow from results in Section 13. Wlog we may assume
in a neighbourhood U = U (0) that
s
Wloc (0) = {x ∈ U (0) : x1 = 0},
u
Wloc (0) = {x ∈ U (0) : x2 = 0, x3 = 0}.

Introducing two-dimensional cross-sections Σv = {x ∈ U : x1 = cv } and

Σh = {x ∈ U : x2 = ch } for constants cv , ch > 0 sufficiently small, we may
consider the Poincaré map as a composition

P = G ◦ L : Σv ∩ {x1 > 0} → Σv ∩ {x1 > 0},

where G : Σh → Σv and L : Σh → Σv are induced by the flow of (12.1).

Let λ2,3 = α ± ωi with ω > 0 and λ1 = λ, where the parameter-dependence
is implicit. Regarding L, we use a Sternberg-type linearization to obtain

ζ ′ = λζ,
ξ1′ = αξ1 − ωξ2 , (12.11)
ξ2′ = ωξ1 + αξ2 ,

which can be solved explicitly for L (exercise!) mapping the rectangle

Σv ∩ {x1 > 0} to a solid spiral Z ⊂ Σh ; see Figure 31.

x2 G

Σh x2
Σv
L

x1 x3

Σv

x3

Figure 31: Sketch of Shilnikov homoclinic bifurcation and its proof. Phase
space on the left and return maps on Σv on the right.

68
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 The global map G can be given to lowest-order by Taylor expansion as
in the proof of the Andronov-Leontovich Theorem 12.3. By the Shilnikov
condition, G maps the spiral outside of the initial rectangle back to Σv as
shown in Figure 31.

Example 12.7. Homoclinic orbits also appear naturally in many PDEs.

Consider the FitzHugh-Nagumo equation
∂t u = ∂xx u + f (u) − v,
(12.12)
∂t v = ε(u − v),
where f (u) = u(1 − u)(u − p), (x, t) ∈ R × [0, ∞), p ∈ (0, 1/2) and ε are
parameters. Suppose we restrict to (respectively search for) travelling
wave solutions of the form
u(x, t) = u(x − ct) =: u(s), v = v(x − ct) =: v(s),
with wave speed c ∈ R. Using the chain rule we have
du du
= u̇, ∂t u = −c
∂x u = = u̇,
ds ds
and similarly for v. Upon setting u̇ = w, this yields the three-dimensional
ODE
u̇ = w,
ẇ = −cw − f (u) + v, (12.13)
v̇ = ε(u − v).
One checks that there is a unique equilibrium at 0 for (12.13). For the
case ε = 0, the plane P = {v = 0} is invariant and contains the global
equilibrium. The flow on P is
u̇ = w,
(12.14)
ẇ = −cw − f (u),
which is a Hamiltonian system for c = 0 and using the Hamiltonian, one
finds a homoclinic orbit (exercise!). For ε > 0, one can also construct a
substantially more complicated homoclinic orbit to the origin (very difficult
proof). However, it is relatively straightfoward to check that there exist
parameter values (c, a) such that 0 is a saddle-focus and the Shilnikov
condition holds on the eigenvalues. 

13 The Smale Horseshoe

The last map in the the proof of Shilnikov’s Theorem 12.6, leads one nat-
urally to consider maps of the rectangle to itself, which have a horseshoe
geometry as shown in Figure 32.

69
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 k4 k3

k2 H1

k3

k4

k1
H0

k1 k2 g −1

· · · 01 · · · = · · · a0 a1 · · ·
Figure 32: Inverse g −1 of the standard horseshoe map g (exercise: draw
a suitable sketch for g). The unit square Y is stretched horizontally and
compressed vertically, then folded into a horseshoe shape, and finally pasted
back onto the rectangle. The four solid grey horizontal strips form the set
of points g −2 (Y ∩ g(Y) ∩ g 2 (Y)) staying in Y. The corners kj of Y are also
marked under the transformation g −1 . We also indicate the two vertical
strips in Y of the forward image g(Y) ∩ Y and we show the symbolic coding
via the horizontal strips of one point chosen as an example.

Consider the unit square Y = [0, 1] × [0, 1]. Define a map g by (I) lin-
early stretching Y vertically by a factor µ > 2 and linearly compressing
horizontally by a factor λ ∈ (0, 1/2) and (II) folding the resulting rectangle
back so that g(Y) ∩ Y consists of two disjoint vertical strips V0,1 . In par-
ticular, the folded portion lies outside Y so that on Y ∩ g −1 (Y) the map is
linear. The pre-image g −1 (g(Y) ∩ Y) consists of two horizontal strips H0,1 ;
see Figure 32. Now iterate this map
xk+1 = g(xk ), xk ∈ X := {x ∈ X : xk ∈ X for all k ∈ Z}
which is called the Smale horseshoe map. Observe that Y ∩ g(Y) ∩ · · · ∩
g n (Y) is the union of 2n disjoint vertical strips while Y∩g −1 (Y)∩· · ·∩g −n (Y)
is the union of 2n disjoint horizontal strips. The Jacobian of g is constant
and given on H0,1 by  
±λ 0
Dg = (13.1)
0 ±µ
for + on H0 and − on H1 . Most points leave Y under iteration. The remain-
ing points X form a Cantor set (a compact, perfect, totally disconnected
set).

70
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Let Σ2 be the space of bi-infinite sequences a = {aj }∞
j=−∞ with elements
from {0, 1} with the metric
∞
X
d(a, b) = |aj − bj |2−|j|
j=−∞

and also recall the (left-)shift σ : Σ2 → Σ2 defined by

σ(a)j = aj+1 .

Theorem 13.1 (Horseshoe Symbolic Dynamics). g : X → X is topologi-

cally conjugate to σ : Σ2 → Σ2 .
Proof. Let x ∈ X , we construct the conjugacy φ : X → Σ2 by

φ(x) = {aj }∞
j=−∞ with g j (x) ∈ Haj .

See also Figure 32 for an illustration of the labeling of points in X . By

construction, we must have

σ(φ(x))j = σ(a)j = aj+1 = φ(g(x))j .

Let R(b) be the rectangle of all x having a finite central sequence b =

b−m · · · bn , i.e., g j (x) ∈ Hbj . The rectangle R(b) has height µ−(n+1) and
width λm and is obtained by the intersection of a horizontal and a vertical
strip. Since diameter of R(b) converges to 0 as m, n → ∞, injectivity and
continuity of φ follow. Lastly, we claim R(b) is non-empty for any given
b = b−m · · · bn so that φ is surjective; indeed, we have that R(b0 · · · bn )
is a horizontal strip mapped by g vertically across Y so that it intersects
each horizontal strip, which implies R(b0 · · · bn+1 ) is non-empty. A similar
argument holds for backward iteration and R(b−m · · · b−1 ).

Corollary 13.2 (Horseshoe/Shift Dynamics). The following properties of

the horseshoe map g : X → X hold:
(T1) X contains a countable infinite set of saddle-type periodic orbits.

(T2) X contains an uncountable set of non-periodic orbits.

(T3) X contains a dense orbit.

(T4) X is a hyperbolic set.

Proof. The existence of orbits as prescribed in (T1)-(T3) follows from the
conjugacy to the shift map in Theorem 13.1. (T4) and the saddle-type
contraction/expansion follow by looking at (13.1).

71
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 In particular, Theorem 13.2(T4) implies in combination with the per-
sistence of hyperbolic sets Theorem 7.5 that horseshoe dynamics is robust
under perturbations. This observation and Theorem 13.2(T1) almost finish
the proof of Shilnikov’s Theorem 12.6. What we still lack is to generalize
the horseshoe construction to curved structures instead of rectangles for
points (x, y) ∈ Y ⊂ R2 .

Definition 13.3. A vertical curve x = v(y) is a curve for which

0 ≤ v(y) ≤ 1, |v(yl ) − v(yr )| ≤ ρ|yl − yr | for 0 ≤ yl ≤ yr ≤ 1

and some ρ ∈ (0, 1). Similarly, a horizontal curve y = h(x) is a curve for
which

0 ≤ h(x) ≤ 1, |h(xl ) − h(xr )| ≤ ρ|xl − xr | for 0 ≤ xl ≤ xr ≤ 1.

Given two non-intersecting vertical curves v1 (y) < v2 (y), we may define a
vertical strip by

V := {(x, y) ∈ R2 : x ∈ [v1 (y), v2 (y)], y ∈ [0, 1]}

and similarly we may define a horizontal strip

H := {(x, y) ∈ R2 : y ∈ [h1 (x), h2 (x)], x ∈ [0, 1]}

for two non-intersecting horizontal curves h1 (x) < h2 (x).

Theorem 13.4 (Generalized Horsehoe Map). Let I = {0, 1, 2, . . . , N − 1}

and let Hi , Vi for i ∈ I be disjoint horizontal and vertical strips. Consider
g ∈ C 1 (Y, R2 ) with g(Hi ) = Vi . For µ ∈ (0, 1/2) define the (cone-)bundles
( )
[
Ku : = (x, y) ∈ Vi : |x| < µ|y| ,
i∈I
( )
[
Ks : = (x, y) ∈ Hi : |y| < µ|x| .
i∈I

and suppose g satisfies

(A1) (“invariance”) Dg(Ku ) ⊂ Ku , (Dg)−1 (Ks ) ⊂ Ks ;

(A2) (“expansion/contraction”) |(Dgp (x, y))2 | ≥ 1/µ|y|, |(Dgp−1 (x, y))1 | ≥

1/µ|x| uniformly for p ∈ Y;

and | det(Dg)|, | det(Dg −1 )| ≤ 21 µ−2 . Then

72
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (T1) g has a hyperbolic invariant set X .

(T2) On X , g is topologically conjugate to a shift map on the set of bi-

infinite sequences ΣN on N symbols.
The proof of Theorem 13.4 proceeds exactly along the same lines as
the proof of Theorem 13.1 for the standard horseshoe. The next result is
a classical principle to link the geometry of global orbits in a map to the
appearance of a horseshoe.
Theorem 13.5 (Smale-Birkhoff Homoclinic Theorem). Consider a dif-
feomorphism g : Rd → Rd such that 0 is a hyperbolic saddle fixed point.
Suppose there exists q 6= 0 such that W s (0) and W u (0) intersect transver-
sally at q. Then there exists an invariant set X and κ ∈ Z such that g κ is
topologically conjugate to the Smale horseshoe map.
Remark : It is easy to prove (exercise!) that transverse homoclinic points,
such as q above, cannot occur for the stable/unstable manifolds in a flow. Yet,
they may obviously appear in a suitable Poincaré map.
To prove Theorem 13.5, we need a preliminary result, which is of inde-
pendent interest in many other situations:
Lemma 13.6 (λ-Lemma (or Inclination Lemma)). Let 0 be a hyperbolic
fixed point of a diffeomorphism g : Rd → Rd with dim(W u (0)) = k and
dim(W s (0)) = l. Let M be a k-dimensional C 1 -submanifold intersecting
W s (0) transversally at y ∈ W s (0). Then for any r > 0, and any δ > 0
there exists n0 , a k-dimensional disk D ⊂ M, and a point x ∈ W u (0) such
that for all n ≥ n0 we have

d1 (B(x, r) ∩ W u (0), g n (D)) < δ,

i.e., all images of some disk in M are eventually C 1 -close to a ball inside
the unstable manifold; see Figure 33.

Proof. First, note that iterating g and using the invariance of W s (0), im-
plies that if g n (M) intersects W s (0) transversally for any n ≥ 0. Hence, we
may focus on a neighbourhood of 0 upon picking a submanifold D ⊂ M.
Furthermore, it is easy to check that the result holds for C 0 , i.e.,

d0 (B(x, r) ∩ W u (0), g n (D)) < δ,

so the problem is to control the tangent spaces. Without loss of generality,

we write the map locally in a neighbourhood V = V s × V u near 0 as

g(x) = g(xs , xu ) = (As xs + φs (x), Au xu + φu (x))⊤

73
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 W u (0)

g n (D) M
B(x, r)

D
y W s (0)
0

Figure 33: Illustration of the λ-Lemma (or Inclination Lemma).

where (xs , xu ) ∈ V s × V u and

|As | ≤ a < 1, |(Au )−1 | ≤ a < 1, φs (x) = o(|x|) = φu (x).

In the following calculations, it can be shown (non-trivial exercise!) that

carrying around the higher-order terms φu (x), φs (x) does not affect the
result so we shall simply drop them here. Let v0 = (v0s , v0u ) ∈ Tq D, which
has slope |λ0 | := |v0s |/|v0u |, which is non-singular since |v0u | 6= 0 due to
transversality at q. Then define the iterates

q1 = g(q), v1 = Dgq (v0 ), ··· qn = g n (q), vn = Dgqn−1 (vn−1 ).

One then computes

|v1s | |As v s |
λ1 = u = u 0u ≤ a2 λ0 < aλ0 ,
|v1 | |A v0 |

where we stated the last obvious inequality as we are going to lose one
factor of a in the general case with higher-order terms due to cross-terms.
Similarly, one finds
|v1u |
≥ a−2 > a−1 .
|v0u |
Therefore, for any given δ > 0 and r > 0, there exists n0 such that for all
n ≥ n0 , we find
|vnu |
λn ≤ δ and ≥ r,
|v0u |
which implies the result.

74
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Proof. (of Theorem 13.5) We present only the main geometric idea in
R2 . We may again assume that the stable and unstable manifolds W s (0)
and W u (0) are the (x1 , x2 )-coordinate axes in a neigbourhood V = V(0);
cf. (12.3)-(12.4). Since q ∈ W s (0) ∩ W u (0), it follows that there exists n0
such that
g n (q) ∈ V and g −n (q) ∈ V for all n ≥ n0 .
Now pick a thin rectangle R with long horizontal sides parallel to the x1 -
axis but separated from W s (0); see Figure 34.
x2

g k (R)
W u (0)

W s (0) q x1

Figure 34: Geometric sketch for the Smale-Birkhoff Theorem 13.5.

The vertical boundary segments are chosen as straight lines such that
we may find iterates g ns (q) and g ns +1 (q) with ns ≥ n0 close to the hori-
zontal boundary of R. By the λ-Lemma 13.6, the vertical and horizontal
boundary segments of R are transformed to vertical and horizontal curves
for g k (R), which stretches as a thin rectangle parallel to W u (0). Also
for this image, we can pick k to match precisely two iterates g −nu (q) and
g −nu +1 (q) with nu ≥ n0 close to the vertical boundary of R; see Figure 34.
The λ-Lemma 13.6 guarantees that pre-images of vertical and horizontal
boundary segments are transformed to vertical and horizontal curves. This
construction has the required horseshoe geometry viewing R as our basic
domain for the horseshoe analogous to the unit rectangle above. What re-
mains to be checked is that suitable contraction and expansion conditions
hold as prescribed in Theorem 13.4, which is essentially evident (difficult
exercise!) since we may always also ensure by adjusting k that g k is a map
with sufficiently strong contraction and expansion as 0 is a saddle point.

75
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 14 Melnikov’s Method
Although the Smale-Birkhoff Homoclinic Theorem 13.5 provides a crite-
rion, how to prove the existence of a horseshoe, verifying the existence of
transverse homoclinic points is very difficult. In this section, we provide
one special case, where an analytical approach is possible. Consider the
ODE
x′ = f (x) + εg(x, t), x ∈ R2 , (14.1)
where f, g are smooth, g is periodic in t with minimal period T > 0 and
f = (∂x2 H, −∂x1 H)⊤ is Hamiltonian. The key assumption is that the
unperturbed system for ε = 0 contains a homoclinic orbit γ = γ(t) to a
hyperbolic saddle point p, i.e., for ε = 0 we have

lim γ(t) = p, Df (p) has eigenvalues λ1 < 0 < λ2 .

t→±∞

The relevant two-dimensional map is obtained as a Poincaré map

Pεt0 : Σt0 → Σt0 , Σt0 = {(x, t)|t = t0 ∈ [0, T ]} ⊂ R2 × S1 ,

for the three-dimensional autonomous version of (14.1) given by

x′ = f (x) + εg(x, t),

(14.2)
t′ = 1,

and S1 = R/(T Z); see Figure 35.

Lemma 14.1. For ε > 0 sufficiently small, the Poincaré map has a locally
unique hyperbolic saddle point ptε0 = p + O(ε).

Proof. This just follows from the implicit function theorem applied to F :=
Id − Pεt0 and using that p is hyperbolic. Alternatively, one may just appeal
to Theorem 7.9 for a proof.

We may view the family of saddle-points as a hyperbolic periodic orbit

Γε = Γε (t) = p + O(ε).

Lemma 14.2. For ε > 0 sufficiently small, Wloc s (Γ ) and W u (Γ ) are

ε loc ε
k s u
C -close (for some k ≥ 1) to Wloc (Γ0 ) and Wloc (Γ0 ). Furthermore, orbits
γεs (t; t0 ), γεu (t; t0 ) lying in W s (Γε ) and W u (Γε ) starting at time t0 (i.e., in
Σt0 ) can be uniformly approximated as

γεs (t; t0 ) = γ(t − t0 ) + γ1s (t; t0 )ε + O(ε2 ), t ∈ [t0 , ∞),

(14.3)
γεu (t; t0 ) = γ(t − t0 ) + γ1u (t; t0 )ε + O(ε2 ), t ∈ (−∞, t0 ].

76
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 pε
γ(t)

t p

x1

x2 Σt0 L

Figure 35: Phase space for the system (14.2). The solid curves illustrate
the situation in the limit ε = 0, while dashed curves show the breaking of
the family of homoclinic orbits.

Proof. Existence and closeness follows from the stable/unstable manifold

theorem. Pick a sufficiently small neighbourhood U = U (p) so that the
stable and unstable manifolds are O(ε)-close. Consider an orbit γεs (t; t0 ) in
W s (Γε ) starting outside U in a neighbourhood of γ(0); see Figure 35. By
Gronwall’s Lemma, we can follow this orbit until it reaches ∂U , say at time
t1 > t0 , so that

|γεs (t; t0 ) − γ(t − t0 )| = O(ε) for t ∈ [t0 , t1 ].

Since γεs (t; t0 ) remains in U for any t > t1 , the C k -closeness of the stable
manifolds implies

|γεs (t; t0 ) − γ(t − t0 )| = O(ε) for t ∈ (t1 , ∞).

So the first equation in (14.3) follows. For the second one, the same argu-
ment following an orbit backward in time.

In fact, one even has an equation for the first-order approximation using
a variational equation (cf. Definition 5.4) along γ, i.e.,

(γ1s )′ (t; t0 ) = Df (γ(t − t0 ))γ1s (t; t0 ) + g(γ(t − t0 ), t) (14.4)

for t ≥ t0 , and similarly for t ≤ t0 in the case of approximating γ1u (t; t0 ). We

would like to study the distance between W u (ptε0 ) and W s (ptε0 ) on Σt0 as

77
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 shown in Figure 35. As before, consider γ(0) and define a one-dimensional
cross-section/line L via the normal vector

(−f2 (γ(0)), f1 (γ(0)))⊤ ∈ R2

to γ(t) at γ(0). Let γεs (t0 ) := γεs (t0 ; t0 ) ∈ W s (ptε0 ) and γεu (t0 ) := γεu (t0 ; t0 ) ∈
W u (ptε0 ) be the unique points in L closest to γ(0). Define the (signed)
distance by
d(t0 ) := γεu (t0 ) − γεs (t0 ).
Lemma 14.3. The function d(t0 ) can be uniformly approximated by
f (γ(0)) ∧ (γ1u (t0 ) − γ1s (t0 ))
d(t0 ) = ε + O(ε2 ). (14.5)
|f (γ(0))|
Proof. By Lemma 14.2, we may approximate the difference γεu (t0 ) − γεs (t0 )
to first order by γ1u (t0 ) − γ1s (t0 ). Furthermore, recall the definition of the
wedge product a ∧ b = a1 b2 − b1 a2 , which yields that f ∧ (γ1u − γ1s ) is the
projection of γ1u − γ1s onto L and that |f (γ(0))| is a normalization factor;
see Figure 35.

Formula (14.5) suggests to study just the numerator in the O(ε)-term

to study intersections of the stable/unstable manifolds. The following def-
inition turns out to be the correct one to just calculate the intersections
from given data.
Definition 14.4. Define the Melnikov function by
Z ∞
M (t0 ) = f (γ(t − t0 )) ∧ g(γ(t − t0 ), t) dt.
−∞

Theorem 14.5 (Melnikov Zeros for Homoclinic Orbits). If M has simple

zeros and is independent of ε, then W s (ptε0 ) and W u (ptε0 ) intersect transver-
sally. If M has no zeros, then W s (ptε0 ) ∩ W u (ptε0 ) = ∅.
Proof. Consider the time-dependent distance function

δ(t; t0 ) := f (γ(t − t0 )) ∧ (γ1u (t; t0 ) − γ1s (t; t0 ))

= f (γ(t − t0 )) ∧ γ1u (t; t0 ) − f (γ(t − t0 )) ∧ γ1s (t; t0 )
=: δ u (t; t0 ) − δ s (t; t0 ).

By Lemma 14.3, we have d(t0 ) = εδ(t0 ; t0 )/|f (γ(0))| + O(ε2 ). The trick is
to find ODEs for δ u (t; t0 ) and δ s (t; t0 ) and solve them by integration. We
start with the stable part and differentiate
d s
δ (t; t0 ) = Df (γ(t − t0 ))γ ′ (t − t0 ) ∧ γ1s (t; t0 ) + f (γ(t − t0 )) ∧ (γ1s )′ (t; t0 ).
dt

78
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Using the variational equation (14.4), γ ′ = f (γ), and omitting the argu-
ments (t − t0 ) and (t; t0 ) for notational simplicity leads to

(δ s )′ = Df (γ)f (γ) ∧ γ1s + f (γ) ∧ (Df (γ)(γ1s ) + g(γ, t))

= Tr(Df (γ)) δ s + f (γ) ∧ g(γ, t).

However, Tr(Df (γ)) = 0 as the vector field is Hamiltonian and smooth (in
fact: C 2 would suffice here). Integrating the ODE for δ s from t0 to ∞ yields
Z ∞
s s
δ (∞; t0 ) − δ (t0 ; t0 ) = f (γ(t − t0 )) ∧ g(γ(t − t0 ), t) dt. (14.6)
t0

By Lemma 14.2, γ1s (t; t0 ) is bounded for t ≥ t0 so that we may compute

δ s (∞; t0 ) = lim f (γ(t − t0 )) ∧ γ1s (t; t0 ) = f (p) ∧ lim γ1s (t; t0 ) = 0,

t→+∞ t→+∞

so (14.6) gives a formula for δ s (t0 ; t0 ). Similarly, one may prove

Z t0
u
δ (t0 ; t0 ) = f (γ(t − t0 )) ∧ g(γ(t − t0 ), t) dt, (14.7)
−∞

so that adding (14.6) and (14.7) and employing Lemma 14.3 gives us

M (t0 )
d(t0 ) = ε + O(ε2 ), (14.8)
|f (γ(0))|

where we note that |f (γ(0))| = O(1) as ε → 0. If M (t0 ) is independent

of ε, the first term in (14.8) dominates. Suppose M (t0 ) has zeros, then
there exists a time t∗ such that γεs (t∗ ) = γεu (t∗ ), which corresponds to an
intersection in a homoclinic point in the stable/unstable manifolds of the
saddle point of the Poincaré map, i.e.,

W s (ptε∗ ) ∩ W s (ptε∗ ) 6= ∅.

Since Poincaré maps for t0 ∈ [0, T ] are topologically equivalent (exercise!),

W s (ptε0 ) and W s (ptε0 ) must intersect for all t0 ∈ [0, T ]; see Figure 35. If
the zeros of M (t0 ) are simple, then the intersection points are transversal
homoclinic points. In the case, when M has no zeros, the manifolds W s (ptε0 )
and W s (ptε0 ) do not intersect for any t0 .

Example 14.6. Consider the periodically-forced Duffing equation

x′ = y,
(14.9)
y ′ = x − x3 + εα cos(t) − εy,

79
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 where α is a parameter. The system (14.9) is Hamiltonian for ε = 0 with
Hamiltonian function
1 1 1
H(x, y) = y 2 − x2 + x4 .
2 2 4
There is a hyperbolic saddle point at p = (0, 0) and the level set {H = 0}
consists of two homoclinic orbits γ± ; see Figure 36.
y

Figure 36: Phase portrait showing the double-homoclinic loop in the un-
forced (i.e., ε = 0) Duffing equation (14.9).

In this case, one may even explicit calculate the homoclinic orbits and
check that
√ √ 
γ+ (t) = 2 sech(t), − 2 sech(t) tanh(t) , γ− (t) = −γ+ (t).

Therefore, we can also give an explicit formula for the Melnikov function.
We focus on γ+ (t) since the calculation is similar for γ− (t) by symmetry.
y
Let γ+ denote the second component of γ+ , then we have
Z ∞
y y

M (t0 ) = γ+ (t) α cos(t + t0 ) − γ+ (t) dt.
−∞
y
After plugging in the explicit formula for γ+ (t), the resulting integral can
actually be evaluated (non-trivial exercise!) and one obtains
4 √ π 
M (t0 ) = − + 2απ sech sin(t0 ).
3 2
Therefore, after quite a bit of algebraic manipulation (exercise!), one finds
that there exists an amplitude α > 0 of the forcing such that M (t0 ) has
simple zeros. Theorem 14.5 applies and shows the existence of a transverse
homoclinic point. Therefore, the Smale-Birkhoff Theorem can be used to
show that the return map of the Duffing equation for suitable forcing con-
tains horseshoe dynamics. 

80
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 15 Chaos
Although we have already seen the key example of chaotic dynamics, one
may also define the concept abstractly.

Definition 15.1. A dynamical system (X , T , φt ) has sensitive depen-

dence on initial conditions at x0 if there exists r > 0 such that for any
δ > 0, there exists y0 with |x0 − y0 | < δ and a time T > 0 such that

|φT (x0 ) − φT (y0 )| ≥ r.

If S is a set then the system exhibits sensitive dependence on initial

conditions on S if the previous definition applies to any x0 ∈ S.

Example 15.2. Consider the dynamical system on X = [0, 1] with T = N0

given by the doubling map

g(x) = 2x (mod 1), (15.1)

which can be taken as one of the simplest examples, where we can expect
sensitive dependence on initial conditions; see Figure 37(a). 
(a) g(x) (b) g(x)

1 1

1 x 1 x
0 2 1 0 2 1
Figure 37: (a) Sketch of the doubling map 15.1. (b) Sketch of the tent
map (15.3) for p = 2.

Theorem 15.3 (Sensitive Dependence for the Doubling Map). The dou-
bling map (15.1) has sensitive dependence on initial conditions on X =
[0, 1].

Proof. We are going to show that g is expansive with constant 41 , i.e., for
any x0 6= y0 , there exists k such that
1
|g k (x0 ) − g k (y0 )| ≥ ,
4

81
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 which immediately implies the result. Consider two points x, y such that
x < y and y − x < 14 . We claim that

|g(y) − g(x)| ≥ 2|y − x|. (15.2)

If x, y are in the same interval, i.e., x, y ∈ [0, 1/2) or x, y ∈ [1/2, 1) then we

immediately have (15.2). If x, y are in different intervals and y > x, then
the doubling map can be written as

g(x) = 2x, g(y) = 2y − 1.

This observation yields again (15.2) since

|g(y) − g(x)| = |2y − 1 − 2x| = |1 − 2(x − y)| ≥ 1 − 2|x − y| ≥ 2|y − x|

due to |y − x| ≤ 1/4. To finish the proof, we argue by contradiction.

Suppose x0 < y0 and |g k (y0 ) − g k (x0 )| ≤ 1/4 for all k ∈ N0 , then by (15.2)
we have
|g k (y0 ) − g k (x0 )| ≥ 2k |x0 − y0 |,
and selecting k sufficiently large yields the desired result.

The logistic map is another example, where one may prove sensitive
dependence for certain parameters. The calculations are easier in the fol-
lowing simplification.

Example 15.4. Consider the (family of) tent maps


px if x ∈ [0, 1/2)
g(x; p) = (15.3)
p(1 − x) if x ∈ [1/2, 1]

for p ∈ (1, 2]. For example, it is relatively easy to check (exercise!) that the
classical tent map g(x; 2) has sensitive dependence on initial conditions. 

In addition to sensitive dependence, chaos should also contain some

notion of spreading of reaching various parts of phase space.

Definition 15.5. A dynamical system is called topologically transitive

if there exists a point x ∈ X , whose forward orbit is dense in X .

Theorem 15.6 (Tent Map Topological Transitivity). The tent map (15.3)
is topologically transitive for p = 2.
Remark : Similar result and proofs also work for other parameter values and var-
ious other classes of dynamical systems; see Section 16.

82
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Proof. We use symbolic dynamics, similar to the horseshoe analysis.
Let Σ+2 = {s = s0 s1 · · · : sj ∈ {0, 1}} be the space of infinite one-sided
sequences of two symbols. Define the itinerary map h : [0, 1] → Σ+ 2 by

0 if g j (x) ∈ [0, 1/2] =: I0 ,
h(x)j =
1 if g j (x) ∈ [1/2, 1] =: I1 .

We exclude all points x, which are going to eventually hit the point x = 1/2,
which is not going to alter our argument regarding topological transitivity.
Then note that h satisfies h(g(x)) = σ(h(x)), where σ is the left-shift map,
i.e., σ(s)j = sj+1 for j ∈ N0 . We may associate subsets of [0, 1] with subsets
of symbols via

Is0 s1 ···sn := {x ∈ [0, 1] : g j (x) ∈ Isj for j ∈ {0, 1, . . . , n}}.

Observe that each Is0 s1 ···sn is an interval and has length 2−(n+1) . The
sequence of compact intervals

I s0 ⊃ I s0 s1 ⊃ I s0 s1 s2 ⊃ · · ·

has a non-empty intersection by Cantor’s Intersection Theorem and one

easily checks that the intersection is a single point x0 so that we may define
H(s) = x0 . One finds H ◦ h(x0 ) = x0 but observes that H is surjective yet
not injective (see remark below), so it is only a semi-conjugacy H ◦σ(s) =
g ◦ H(s). In any case, having this semi-conjugacy, checking topological
transitivity is very easy: let s∗ be a sequence constructed by concatenating
all possible symbol sequences of length 1, length 2, length 3 and so on.
Define x∗ = H(s∗ ) and observe that eventually g j (x∗ ) ∈ Is0 s1 ···sn for any
given s0 s1 · · · sn . But since the length of Is0 s1 ···sn tends to zero as n → ∞,
density of the forward orbit of x∗ follows.
Remark : The map H : Σ+ 2 can also be viewed as sending a symbol sequence to
the binary form of a real number by
∞
X sj
H(s) = j+1
.
j=0
2

1
It is not injective since, e.g., H(1, 0, 0, 0, · · · ) = 2 = H(0, 1, 1, 1, . . .).

Definition 15.7. An invariant set S ⊆ X is called chaotic if it has sensi-

tive dependence on initial conditions and φt |S is topologically transitive.

Proving the following results is lengthy, yet no new techniques are

needed.

83
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Theorem 15.8 (Chaos in Interval Maps). S = [0, 1] is chaotic for the tent
map (15.3) with p = 2, for the doubling map (15.1), and for the logistic
map x 7→ 4x(1 − x).

Theorem 15.9 (Chaotic Horseshoe). The horsehoe map is chaotic on X ,

where X ⊂ [0, 1] × [0, 1] is the maximal invariant Cantor set as constructed
in Section 13.

Unfortunately, testing for chaos in higher-dimensional maps

xk+1 = g(xk ), g : Rd → Rd , xk ∈ Ω ⊆ Rd

for d ≥ 2 turns out to be difficult; the same problem occurs for flows in
dimension d ≥ 3. The following concept is a good practical indicator.

Definition 15.10. Suppose g ∈ C 2 , fix x0 ∈ Rd , define the matrix

Ak = Dgxk0 = Dgxk−1 · · · Dgx1 Dgx0 ∈ Rd×d (15.4)

(k)
and consider (Ak )⊤ Ak with eigenvalues (λj )2 for j ∈ {1, 2, . . . , d} ordered
decreasingly (upon taking the positive square-root of each eigenvalue) by
(k) (k) (k)
λ1 ≥ λ2 ≥ · · · ≥ λd .

Define the j-th Lyapunov exponent lj at x0 by

1  (k) 
lj (x0 ) = lim ln λj (15.5)
k→∞ k

if this limit exists.

One may interpret, or define and then prove (15.5) afterwards, Lya-
punov exponents as deformation rates of the unit sphere centered at x0
under the linearized dynamics. In particular, looking at Ak (Sd ) and the
resulting axes of the ellipsoid one; see Figure 38.

S2 Ak

Figure 38: Lyapunov exponents can be viewed as a measure for the defor-
mation of the axes of a unit sphere into the axes of an ellipsoid.

84
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Theorem 15.11 (Liouville Formula). Suppose lj (x0 ) exist for all j and
| det Dgp | is the same for every p, then
d
X
lj (x0 ) = ln(| det Dgx0 |).
j=1

Proof. The determinant is the product of the eigenvalues so we have

d
Y (k) (15.4)
(λj )2 = det(Dgxk0 )2 = | det Dgx0 |2k .
j=1

By standard logarithm rules we have

 
d
1 1 Y (k)
ln | det Dgx0 | = ln | det Dgx0 |2k = lim ln  (λj )2 
2k k→∞ 2k
j=1
d d
1 X  (k)  X
= lim ln λj = lj (x0 ),
k→∞ k
j=1 j=1

and the result is proven.

Example 15.12. For the Smale horseshoe, the Jacobian is independent of

x0 and given by (13.1) with λ ∈ (0, 21 ), µ > 2 so that we have

1
l1 = lim ln µk = ln µ > 0, l2 = ln λ < 0, (15.6)
k→∞ k

which leads to the general observation that Lyapunov exponents are directly
related to multipliers for cases, where Dgx0 is independent of x0 . In fact,
one may prove (non-trivial!) that Lyapunov exponents are independent of
the starting point for certain classes of invariant sets. 

Example 15.13. Consider the standard tent map g(x; 2) from (15.3). Ex-
cept at the single point x = 21 , we have l1 (x0 ) = ln(2) > 0. 

The last two examples indicate that a positive Lyapunov exponent

should be a necessary condition for chaos; similarly, one can also charac-
terize the situation for flows and make a definition of Lyapunov exponents
(exercise!) such that again a positive Lyapunov exponent indicates chaos in
the context of ODEs.

85
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 16 One-Dimensional Maps
The tent map, the logistic map, and the doubling map have served as ex-
cellent examples to understand chaotic dynamics. In the case of continuous
maps, we can do even better and understand their orbit structure.

Theorem 16.1 (Li-Yorke Theorem; “period-three implies chaos”). Sup-

pose g : [a, b] → R with a < b, g ∈ C 0 , has a periodic orbit with minimal
period three. Then g has periodic orbits of all periods.

For a proof using transition graphs and symbolic dynamics see [14]. In
fact, one may get a lot finer information on the appearance of different
families of periodic orbits.

Definition 16.2. The Sharkovskii ordering ⊳ on N is defined by

1 ⊳ 2 ⊳ 4 ⊳ 23 ⊳ · · · ⊳ 2n ⊳ 2n+1 · · ·
· · · ⊳ 9 · 2n+1 ⊳ 7 · 2n+1 ⊳ 5 · 2n+1 ⊳ 3 · 2n+1 ⊳ · · ·
· · · ⊳ 9 · 2n ⊳ 7 · 2n ⊳ 5 · 2n ⊳ 3 · 2n ⊳ · · · ⊳ 9 ⊳ 7 ⊳ 5 ⊳ 3.

Theorem 16.3 (Sharkovskii Theorem). Suppose g : [a, b] → R has a peri-

odic orbit with minimal period n. Then g has periodic orbits of all periods
k ⊳ n.

Example 16.4. Consider the tent map (cf. Example 15.4), which can also
be defined as

xk+1 = p min{xk , 1 − xk } = g(xk ; p) = g(xk ), p ∈ (0, 2]. (16.1)

Varying the parameter p, we get the result shown in Figure 39(a).

As an example, consider p = 2, then we get

g(2/7) = 4/7, g(4/7) = 6/7, g(6/7) = 2/7,

so that Sharkovskii’s Theorem and the Li-Yorke Theorem apply. One can
prove that period three orbits do exist for large ranges of parameter values
as indicated already by Figure 39(a). 

The classical tent map (for p = 2) shows strong similarity to the classical
logistic map yk+1 = 4yk (1 − yk ) (exercise: simulate the logistic map!).

Proposition 16.5. The tent map xk+1 = 2 min{xk , 1 − xk } = g(xk ) and

the logistic map yk+1 = 4yk (1 − yk ) = f (yk ) are conjugate.

86
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 1 1
x (a) g(x) (b)
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 2 0 0.2 0.4 0.6 0.8 1
p x

Figure 39: (a) Bifurcation diagram of the tent map (16.1) obtained by
direct simulation for a 500 point equally spaced mesh of the parameter
space p ∈ (0, 2]. Transients have been removed (here: first 500 iterations),
then 40 iterates are plotted for each value of p. (b) Illustration of the
cobweb construction for p = 1.95.

Proof. We claim h(x) = (1 − cos(πx))/2 is a conjugacy. Clearly, h is a

diffeomorphism on [0, 1]. Suppose x ∈ [0, 1/2]. Then we calculate
  
1 − cos(πx) 1 + cos(πx)
f (h(x)) = 4 = 1 − cos2 (πx)
2 2
1 − cos(2πx) 1 − cos(πg(x))
= sin2 (πx) = = = h(g(x)),
2 2
The case x ∈ (1/2, 1] is similar (exercise!) and the result follows.

Although there may be many periodic orbits, two important questions

remain: (a) how are new periodic orbits generated, and (b) what is their
stability? We start with question (b) and return to (a) in Section 17. We
need some preliminary techniques.
Definition 16.6. Let I be an interval in R and suppose g ∈ C 3 (I, R).
Define the Schwarzian derivative of g by
 
g ′′′ (x) 3 g ′′ (x) 2
Sg(x) := ′ − (16.2)
g (x) 2 g ′ (x)
For an isolated critical point x of g we define Sg(x) = limy→x Sg(y).
Lemma 16.7 (Minimum Principle). Suppose g ∈ C 3 (I, I) with g ′ (x) 6= 0
for all x ∈ I. If Sg(x) < 0 for all x ∈ I, then |g ′ (x)| does not attain a local
minimum in the interior of I.

87
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Proof. Suppose z is a critical point of g ′ , which implies g ′′ (z) = 0. Since
Sg < 0, we have by (16.2) that g ′′′ (z)/g ′ (z) < 0. Therefore, g ′′′ (z) and g ′ (z)
have opposite signs. If g ′ (z) < 0, it follows that g ′′′ (z) > 0 so z is a local
minimum of g ′ , so z is also a local maximum of |g ′ | using g ′ (z) < 0. In the
other possible case we have g ′′′ (z) < 0 and g ′ (z) > 0, which yields that z is
a local maximum of g ′ and hence also of |g ′ |. Since g ′ does not vanish in I
by assumption, the result follows.

Theorem 16.8 (Singer’s Theorem). Let I be a closed (potentially un-

bounded) interval and consider g ∈ C 3 (I, I). If Sg < 0 and g has n critical
points then g has at most n + 2 locally asymptotically stable periodic orbits.

Proof. The idea is to check that if there are too many attracting periodic
orbits, then the regions they attract must overlap. Let z be a periodic point
of period m and denote by W (z) the maximal attracting neighbourhood
of z, i.e., g mn (y) → z as n → ∞ for all y ∈ V and W (z) is the maximal
interval inside V (remark: V is also called the basin of attraction of z).
Note we have g m (W (z)) ⊆ W (z) by construction. Suppose

W (z) = (a, b) for a < b and W (z) ∩ ∂I = ∅.

Under this assumption, we claim that g m must have a critical point in

W (z). Since W (z) is maximal, the map g m must preserve the endpoints
a, b of W (z). There are several cases.
Case 1: Suppose g m (a) = g m (b), then g m has a maximum or minimum
in I by the intermediate value theorem and therefore also a critical point.
Case 2: Suppose g m (a) 6= g m (b) then g m permutes a, b. Suppose
g m (a)= a and g m (b) = b. Then (g m )′ ≥ 1 for all points in ∂V since
otherwise a or b would be a locally stable fixed point for g m , whose basin
of attraction overlaps V. By the minimum principle given in Lemma 16.7,
we have that if g m has no critical points in V, then (g m )′ > 1 on V. This
contradicts g m (W (z)) ⊆ W (z).
Combining the two cases, we may now conclude that g m has a critical
point p ∈ W (z). Using the chain rule, this implies that one of the points
p, g(p), . . . g m−1 (p) is a critical point of g. So either W (z) is unbounded, or
it intersects ∂I, or there is a critical point of g, whose orbit meets W (z).
Since there are at most n critical points by assumption and two boundary
points of W (z), the maximum number of locally attracting periodic orbits
is n + 2.

Example 16.9. Consider the logistic map g(x) = px(1−x). Then Sg(x) =
−6/(1 − 2x)2 for x 6= 1/2 and Sg(1/2) = −∞ at the single critical point.

88
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Since Sg < 0 for x ∈ I = [0, 1], it follows that the logistic map has at most
three locally asymptotically stable periodic orbits. 

Singer’s Theorem 16.8 motivates us that we should look closer at differ-

ent notions of attractivity and the concept of an attractor, which is going
to be considered in Section 18.

17 Renormalization
Renormalization is one well-established tool in dynamical systems theory.
Here we illustrate it in two contexts, the second one concerns the question
about generating sequences upon parameter variation for periodic orbits in
unimodal maps as raised in Section 16.

Definition 17.1. Let (X , T , φt ) be a dynamical system. A renormaliza-

tion group (RG) is a group R together with an induced group action by
R on the dynamical system.

Definition 17.1 is not classical but convenient here; in fact, it is difficult

to give a very general definition of renormalization. As we shall see, R is
often required to have connections to scaling and/or self-similarity. The
action of R can change the following: the phase space X , the phase space
X viewed as space of initial conditions, the time domain T , and the flow
φt . The action on φt is often defined by renormalizing an underlying map
g or a vector field f . Of course, a parameter space P can be renormalized
as well since we can always view XP := X × P as the phase space with
φt |P = Id.

Example 17.2. Consider the fold normal form for a map (cf. Theorem 8.12)

y 7→ q + y − y 2 = g(y; q) = g(y), y ∈ R, q ∈ R. (17.1)

Assume q < 0 is small, so that (17.1) has no fixed points, yet is close to
bifurcation. Iterates passing from y > 0 to y < 0 need a long time as
the map is small near y = 0. This effect is known as critical slowing
down (or intermittency). We want to formally determine the number of
iterates N (q) to pass near zero asymptotically as q ր 0, y → 0. Consider
the second iterate map

g 2 (y; q) = (g ◦ g)(y; q) = q + (q + y − y 2 ) − (q + y − y 2 )2
= (2 − q)q + (1 − 2q)y − (2 − 2q)y 2 + 2y 3 − y 4
= 2q + y − 2y 2 + O(q 2 , qy, qy 2 , y 3 ) ≈ 2q + y − 2y 2 .

89
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 If N (q) iterates are required for g, then approximately N (q)/2 iterates are
required for g 2 for the same passage near y = 0. Let Rα (y) := αy =: Y and
Sβ := βq =: Q be scalings. For α = 2 and β = 4 define a transformation
 
R2 g 2 ((R2 )−1 y, (S4 )−1 q) = 2 g 2 (Y /2; Q/4) ≈ Q + Y − Y 2 ,
i.e., the scaling has brought back the mapping, at least approximately, into
the (self-)similar fold normal form; cf. to scaling and blow-up methods in
Section 11. However, this means we must have
N (q)
≈ N (4q) ⇒ N (q) = O(q −1/2 ) as q ր 0
2
giving our desired passage-time asymptotic approximation. (R+ , ·) =: G is
a group so we write the RG action on the space of continuous one-parameter
interval maps Z := C 0 (I × R, I × R) as
R
G × G × Z 7→ Z, R(g) = Rα g(Rα−1 ·, Sβ−1 ·), g ∈ Z.
More abstractly, instead guessing the scales α = 2, β = 4, and using ap-
proximations, one may study exact solutions to the functional equation
(D(2) g)(x; p) := g 2 (y; p) = Rα g(Rα−1 y, Sβ−1 q),

which is an operator equation for g involving the doubling operator D(2) .

This is very similar to functional iteration schemes we used to prove exis-
tence of invariant manifolds in Sections 3 and 10. 
Example 17.2 explains the notions of renormalization, scale, and self-
similarity. The example also crucially used the normal form structure of
the fold to determine the scales α and β.
Example 17.3. Consider the logistic map (1.4) again
xk+1 = pxk (1 − xk ) = g(xk ; p) = g(xk ), p ∈ (0, 4]. (17.2)
One checks the following (exercise!): g has a stable fixed point γ0 = (p−1)/p
for p ∈ (1, 3] = (p0 , p1 ] with multiplier g ′ (γ0 ) = 2 − p. A flip (or period-
doubling) bifurcation occurs√ at p1 and there is a stable two-cycle γ1 =
{γ11 , γ12 } for p ∈ (3, 1 + 6] = (p1 , p2 ] with multiplier
∂g 2 ∂g ∂g
(γ11 ) = (γ11 ) (γ12 ) = 4 + 2p − p2 . (17.3)
∂x ∂x ∂x
Numerically, one easily obtains that the two-cycle bifurcates in a further
flip bifurcation at some p2 > p1 , the resulting four-cycle γ2 bifurcates again
at some p3 > p2 , leading to a period-doubling sequence (or period-
doubling cascade) of bifurcation points {pn }∞ n=0 as shown Figure 40; the
cascade is also known as the period-doubling route-to-chaos. 

90
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 x

d1

P0 P1 P
p0 p1 p2 2 p3
Figure 40: Sketch of the bifurcation diagram of the logistic map (17.2) with
bifurcation points at pj and superstable cycles at Pj ; see also Examples 17.3,
17.5, and 17.7.

To analyze the structure of the period-doubling cascade it helps to con-

sider special cycles.

Definition 17.4. A cycle (or fixed point) of an iterated map is called

superstable if all its multipliers are zero.

Example 17.5. Continuing Example 17.3, we find √ γ0 is superstable for

p = P0 = 2 and γ1 is superstable for p = P1 = 1 + 5 (exercise!). Numerics
indicates the existence of a sequence of superstable cycles at parameters
{Pn }∞
n=0 ; see Figure 40. Generalizing the idea from (17.3) in using the
chain rule, we get for the 2n -cycle, n ∈ N0 , that
n
∂g 2 1 n
(γnj , Pn ) = 0 ⇔ ∃j : γnj = =: xmax ⇒ g 2 (xmax , Pn ) = xmax ,
∂x 2
i.e., the local maximum of g at xmax has to belong to any superstable cycle.
Therefore, a superstable cycle occurs precisely at intersection of the line
{x = xmax , p ∈ (0, 4]} with the bifurcation diagram in Figure 40. Now
define
n−1
dn = g 2 (xmax , Pn ) − xmax , (17.4)
which is actually the closest point on γn to the maximum at xmax (non-
trivial exercise!); see Figure 40. 

The previous construction does not depend upon the precise shape of
the logistic map. It can be carried out for large classes of maps:

91
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Definition 17.6. Let I = [0, 1]. Then g ∈ C 0 (I, I) is called unimodal
if g(0) = 0 = g(1) and g has a unique local maximum xmax ∈ (0, 1).
If in addition, g has negative Schwarzian derivative on I, it is called S-
unimodal.
(a) g(x; P0 ) (b) g 2 (x; P1 ) (c) g 2 (x; P1 )

xmax xl xmaxxr xl xr

Figure 41: Renormalization of a unimodal map g. (a) map g; (b) second-

iterate map g 2 = g ◦ g; (c) zoom into (b) near the fixed point. Note that
(c) is self-similar up to reflection and translation to the original map (a).

Example 17.7. Continuing Example 17.5, Feigenbaum made several nu-

merical (and partially analytical) observations. First, there exist limits
pn − pn−1
p∞ := lim pn , lim = δF ≈ 4.6692 · · · ,
n→∞ n→∞ pn+1 − pn

where δF is actually universal, i.e., the constant did not seem to depend
on the S-unimodal map he considered. Second, the constant is linked to
superstable cycles via the result

Pn = p∞ − κ1 (δF )−n + o((δF )−n ) as n → ∞,

where κ1 > 0 is not universal, i.e., depends upon the particular map.
Thirdly, even for the phase space measure dn defined in (17.4) one finds a
scaling law

dn = κ2 (αF )−n , αF ≈ −2.5029 · · · , as n → ∞,

where κ2 > 0 is not universal but αF is. The constants δF , αF are known
as Feigenbaum constants. The key argument to establish Feigenbaum’s
universality observations is again renormalization. Consider Figure 41, and
observe that the graphs of

g(x, P0 ) and g 2 (x, P1 )

92
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 look similar once we zoom into a certain region around x = xmax and reflect
the graphs shown in Figure 41. In fact, xmax is a superstable fixed point of
both maps. 
Example 17.7 suggests for a general S-unimodal map g = g(x; p) with
superstable cycle sequence Pn to consider the renormalization groups (R− , ·),
(R, +) with actions Rα and Tβ given by multiplication by α and translation
by β. More precisely, we consider the self-similar approximation
x 
g(x, P0 ) ≈ αg 2 ; P1 , P1 = P0 + β0 , (17.5)
α
i.e., we claim the second iterate is, upon conjugacy by scaling induced by
Rα and shifting by Tβ0 for β0 chosen to yield the next superstable parameter
value; cf. Example 17.2. Upon iterating (17.5), we get
n
 x 
g(x, P0 ) ≈ αn g 2 ; Pn , (17.6)
αn
so if we start at P∞ =: Pn , i.e., at the onset of chaos, we get the functional
equation for just the fist step as

g(x) = Rα (D(2) g)(R1/α x; P∞ ) = Rα g 2 (R1/α x; P∞ ). (17.7)

One has to augment (17.7) with boundary conditions to solve it. Upon
translating a-priori to the origin and adjusting the scale of x we see that
g ′ (0) = 0 and g(0) = 1 are reasonable conditions. The fixed point problem
for g turns out to be solvable (very difficult proof!) if and only if α = αF ,
thereby determining αF .
Remark : Linearizing the operator on the right-hand side of (17.7) using a Fréchet
derivative on a suitable function space shows that the fixed point g has one unstable
eigenvalue, which is precisely given by δF . Hence, the renormalization group and
the functional equation (17.7) indeed contain full information about the universal
Feigenbaum constants.

18 Attractors
Having seen that the full orbit structure of one-dimensional maps can be
rather involved (Sections 16-17) we now study the attracting parts of phase
space in general dynamical systems more abstractly. In this section, we
always consider a C 0 dynamical system (X , T , φt ), which is well-defined in
forward time, i.e., for all t ≥ 0, e.g., T = [0, ∞) for flows or T = N0 for
maps.
Definition 18.1. A compact set U is called a trapping region provided
that φt (U ) ⊂ int(U ) for all t > 0.

93
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Example 18.2. Consider the Lorenz system, classically written as

x′1 = σ(x2 − x1 ),
x′2 = ρx1 − x2 − x1 x3 , (18.1)
x′3 = x1 x2 − βx3 ,

with parameters σ, ρ, β > 0. It is easy to prove [14] that the compact

ellipsoid E = U with boundary defined by
 
ρ + σ 2 β(ρ + σ)2
σx21 + x22 + β x3 − = (18.2)
2 4

is a trapping region for (18.1). 

Definition 18.3. An invariant set A is called an attracting set if there

exists a trapping region U such that
\
A= φt (U ), (18.3)
t≥0

see also Figure 42(a). A set A is called an attractor if it is an attracting

set and it is maximal (or indecomposable), i.e., if à =
6 ∅ is an attracting
set and à ⊂ A then à = A.

Theorem 18.4. Consider a dynamical system (X , T , φt ). Let V be a trap-

ping region with attracting set A. Then the following hold:

(T1) A is closed. If T = R or T = Z, then A is invariant;

(T2) if x0 ∈ V then ω(x0 ) ⊆ A;

(T3) if x∗ ∈ A is a hyperbolic steady state, then W u (x∗ ) ⊆ A.

Proof. For (T1), one easily sees that A is closed since it is the intersection
of closed sets. Furthermore, if p ∈ A, then p ∈ φt (V) for all t > 0. One can
re-write the definition (18.3) of A as
\ \
A= φt+s (V) = φt (V)
t≥−s t≥−s

since V is a trapping region so that φ−s (p) ∈ A. (T2) is basically im-

mediate from the definitions (exercise!). Regarding (T3), since x∗ ∈ A it
follows that x∗ ∈ int(V) by the definition of a trapping region, which has
to be mapped into its interior via φt . This implies that the local unstable
u (x ) is contained in V for a sufficiently small neighbourhood
manifold Wloc ∗

94
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 near x∗ . But then we may simply use the definition of the (global) unstable
manifold
[ \[
W u (x∗ ) = u
φt (Wloc (x∗ )) = u
φs ◦ φt (φ−s (Wloc (x∗ )))
t≥0 s≥0 t≥0
\[ \
⊂ φs (φt−s (V)) = φs (V) = A
s≥0 t≥0 s≥0

to finish the proof.

Although Definition 18.3 is very natural, i.e., just taking the forward
intersection of sets under the dynamics, it may not be appropriate for all
situations as the next example shows.

Example 18.5. Consider the circle S1 = [0, 1]/(0 ∼ 1) and let X =

(0, ∞) × S1 be the punctured plane. Consider the vector field

r′ = 1 − r,

(18.4)
θ′ = − θ − 41 θ − 34

in polar coordinates. Clearly, there is a trapping region, say U = S1 ×

[1/2, 3/2] containing the unit circle and there are two equilibrium points
(1, 1/4) and (1, 3/4). One checks that A = {(r, θ) ∈ X : r = 1} but all
initial conditions except for (1, 1/4) are attracted to (1, 1/4). Therefore, it
might be more natural to use a definition to make (1, 1/4) the only attractor
for the given trapping region U ; see also Figure 42(b). 

(a) (b)
U
φt (U )
A

Figure 42: (a) General idea of an attractor A inside a set U , i.e., the flow
φt contracts everything towards/onto A for t > 0. (b) Illustration of a key
example to understand the different attractor definitions.

95
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Definition 18.6. Let A be a closed invariant non-empty set. The basin
of attraction is

BA(A) := {x0 ∈ X : ω(x0 ) ⊂ A}.

Definition 18.7. Let A be a closed invariant set and suppose µ is a given

measure on X . A is called a Milnor attractor if

µ(BA(A)) > 0

and there is no smaller set à ⊂ A such that µ(BA(A) \ BA(Ã)) = 0.

This definition fixes some intuition problems but also creates new ones.

Example 18.8. Regarding the ODE (18.4), we see that (1, 1/4) is its
Milnor attractor. However, if we look at

x′ = −x2 , x = x(t) ∈ R,

then x∗ = 0 is a Milnor attractor, despite the fact that it is (nonlinearly)

unstable as an equilibrium point. 

Hence, we shall simply use the term attractor from Definition 18.3 until
further notice and we shall explicitly write Milnor attractor if we refer to
Definition 18.7. As expected from Sections 13-17, the topology/geometry
of attractors can be extremely intricate.

Example 18.9. Consider the solid torus S1 × D2 , where D2 = {(x, y) ∈

R2 : x2 + y 2 ≤ 1} is the unit disc in R2 . For λ ∈ (0, 1/2), define the map
 
1 1
g(θ, x, y) := 2θ, λx + cos(2πθ), λy + sin(2πθ) , (18.5)
2 2

which stretches the solid torus in the S1 -direction by a factor of 2 and

compresses in the direction of the disk D2 by a factor λ and wraps the
resulting torus twice back inside the original one as shown in Figure 43.
The forward iteration of the map g yields tori stacked inside themselves
as shown on a cross-section in Figure 43. One may prove (exercise!) that
∞
\
A= g k (S1 × D2 )
k=0

is a non-empty attractor and is the product of a Cantor set with an interval.

The set A is called the Smale-Williams solenoid. 

96
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Figure 43: Illustration of the Smale-Williams solenoid showing one forward
iteration from the solid torus to the double-wrapped torus of the map (18.5).
One the right-hand side a cross-section is shown.

Example 18.10. Continuing the Lorenz equations in Example 18.2 with

trapping region U , we immediately obtain an attractor A via Definition 18.3.
For the so-called classical Lorenz parameter values σ = 28, ρ = 10, β = 8/3,
an illustration of the attractor can be found in [14]. 
Definition 18.11. Let A be an attractor. Then A is called
(D1) chaotic attractor if A is also a chaotic set (see Definition 15.7);
(D2) hyperbolic attractor if A is also a hyperbolic set; (see Defini-
tion 7.1);
(D3) strange attractor if A is a chaotic but not a hyperbolic set.
One checks relatively easily that the Smale-Williams solenoid is a hy-
perbolic attractor and with a bit more work that it is chaotic. Although
it may be evident geometrically, it is highly non-trivial to show that the
Lorenz attractor is indeed a strange attractor. In fact, the next result shows
already some of the difficulties.
Theorem 18.12. The Lorenz attractor has volume zero.
Proof. Let f : R3 → R3 be the vector field associated to the ODE (18.1).
Let E be the trapping region/ellipsoid and define
E(t) := {φt (x0 ) : x0 ∈ E}.
Let V (t) be the classical (Riemannian) volume of E(t) as a set inside R3 .
Then we have
Z Z
d d
V (t) = dx = ~n · f (x) dA,
dt dt E(t) ∂E(t)

97
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 where dA is the surface area form for ∂E(t) and ~n is the associated unit
normal vector. Hence, the Divergence Theorem implies
Z
d
V (t) = ∇ · f (x) dx. (18.6)
dt E(t)

The divergence can easily be calculated from (18.1) and we obtain

d
V (t) = −(σ + 1 + β)V (t), (18.7)
dt
which implies the result by solving (18.7).
Remark : The change-of-volume formula (18.6) under a vector field f is valid in
far more generality if f and E(t) are smooth in Rd . It is then proved using Stokes’
Theorem instead of the Divergence Theorem, which itself is evidently a special
case of Stokes’ Theorem.
Also for the Smale-Williams solenoid, one finds that it has three-dimensional
volume equal to zero. These observations motivate an alternative defini-
tion for a strange attractor A, i.e., that A should be chaotic and A should
satisfy a certain “dimension requirement”. To make this precise, we need
a suitable definition of dimension.

Definition 18.13. Let S be a subset in Rd . Let N (r, S) be the minimun

number of non-overlapping unit cubes/boxes of side length r needed to
cover S. Then define the box-counting dimension by

ln N (r, S)
dimB (S) := lim .
r→0 ln(1/r)

For example, we can cover the unit square [0, 1]2 by N = 22j boxes of
side length r = 2−j , so taking the limit j → ∞ yields that dimB ([0, 1]2 ) = 2.

Example 18.14. To show that dimB (S) does capture non-integer consider
the classical middle-third Cantor set. It is obtained by removing the middle
third (1/3, 2/3) from [0, 1], then removing the middle thirds from K1 =
[0, 1/3] ∪ [2/3, 1] to create
       
1 2 1 2 7 8
K2 = 0, ∪ , ∪ , ∪ ,1 ,
9 9 3 3 9 9

and then repeating the process iteratively to create Kj and the resulting
Cantor set \
K= Kj .
j≥0

98
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 At level j for Kj , we need 2j boxes/intervals of length (1/3)j . Hence, we
obtain
ln(2j ) ln 2
dimB (K) = lim j
= ,
j→∞ ln(3 ) ln 3
which is approximately equal to 0.6309, i.e., certainly not an integer. 

Recall that we already encountered Cantor sets as invariant sets for dy-
namics in the case of the Smale-Williams solenoid and the Smale horseshoe.
In addition to box-counting dimension, there are several other notions,
which we shall not discuss here, such as as Lyapunov dimension, Haus-
dorff dimension, and correlation dimension, which can also capture
non-integer dimensions. An alternative definition of a strange attrac-
tor is a chaotic attractor, which is also a fractal set, i.e., of non-integer
dimension.

19 Invariant Measures
Having established the basic theory of hyperbolic, non-hyperbolic and chaotic
dynamical systems, it is difficult to make vast progress using the machinery
we have used so far. Furthermore, the occurrence of chaos naturally leads
to the idea to consider additional structures for a dynamical system, and
we shall look at measure-theoretic tools first.

Definition 19.1. Let (X1 , F1 , µ1 ) and (X2 , F2 , µ2 ) be measure spaces with

Borel probability measures µ1 and µ2 . A measurable map G : X1 → X2 is
called measure-preserving if

µ2 (B2 ) = µ1 (G−1 (B2 )) ∀B2 ∈ F2 .

Definition 19.2. Let (X , F, µ) be a measure space with Borel probability

measure µ. Then µ is called an invariant measure with respect to g :
X → X if g is measure-preserving; we also say µ is g-invariant.

Certain iterated maps as well as flows (via a time-t map) might generate
measure-preserving maps (or measure-preserving transformations). Note
that sets/spaces of finite measure can always be normalized to have measure
one, so just looking at probability measures is not a major restriction.

Example 19.3. Consider the circle S1 = R/Z with Lebesgue measure

Leb = µ induced from R, which is a probability measure in this context
since µ([0, 1]) = 1. As in Example 1.18, consider the rotation

xk+1 = g(xk ) := (xk + r) mod 1,

99
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 for r ∈ R. Since the map just shifts any measurable set and potentially
decomposes it, one easily checks that it is also a measure-preserving trans-
formation. 
An invariant measure can be thought of as describing the distribution
of a typical trajectory of the system.
Theorem 19.4 (Poincaré Recurrence). Let (X , F, µ) be a probability space.
Suppose g : X → X is measure-preserving and fix any B with µ(B) > 0.
Then almost every point in B returns to B infinitely often under forward
iteration of g.
Proof. For N ≥ 0, define
∞
[ ∞
\
BN := g −n (B) and B∗ := BN
n=N N =0

so that B∗ collects all points in X , which enter B infinitely often under

forward iteration of g. So B ∩ B∗ consists of all the points x ∈ B, which re-
enter B infinitely often under forward iteration, i.e., there exists a sequence
0 < n1 < n2 < · · · such that for all i ∈ N

g ni (x) ∈ B and g ni (x) ∈ B ∩ B∗ ,

where the second conclusion follows since g ni −nj (g nj (x)) ∈ B for all j ∈ N;
cf. the proof of Theorem 18.4 (T3). Next, observe that the set B ∩ B∗ could
have small measure but we would like to show that µ(B ∩ B∗ ) = µ(B).
Indeed, we see that g −1 (BN ) = BN +1 implies µ(BN ) = µ(BN +1 ) as g is
measure-preserving. Upon iterating the argument, this yields µ(BN ) =
µ(B0 ) for all N . Furthermore, we observe
∞
!
\
B0 ⊇ B1 ⊇ B2 ⊇ · · · ⇒ µ BN = µ(B0 ),
N =0

which leads to the conclusion µ(B ∩ B∗ ) = µ(B ∩ B0 ). However, upon using

B ⊂ B0 , we obtain µ(B ∩ B0 ) = µ(B), which shows µ(B ∩ B∗ ) = µ(B) and
finishes the proof.

Although it might be nice to have abstract invariant measures, there is

- a priori - no reason to expect that we may be able to find them. The next
result provides a promising partial answer to this question.
Theorem 19.5 (Krylov-Bogulubov). Let X be a compact metric space
and g : X → X be a continuous map. Then there exists a g-invariant Borel
probability measure on X .

100
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 (a) (b) g(x)
1
X

...
B

1 1 x
3 2 1

Figure 44: (a) Illustration of Poincaré recurrence as stated in Theorem 19.4.

(b) Sketch of the Gauss transformation (19.6) from Example 19.6.

Proof. Fix any x ∈ X and consider a continuous function f : X → R (which

we also call an observable). Now define
n−1
1X
Sfn (x) := f (g j (x)). (19.1)
n
j=0

Let F ∗ be a countable dense subset in C(X , R) = C(X ). For any given f ,

the sequence {Sfn (x)}∞
n=1 is bounded since X is compact. We claim that
∗
since F is countable, there exists a sequence nj → ∞ such that the limit
n
lim Sf j (x) =: Sf∞ (x) (19.2)
j→∞

exists for all f ∈ F ∗ . To prove the claim (19.2) one uses a standard di-
agonal argument. Label the elements of F ∗ by f1 , f2 , etc. and first use
that {Sfm1 (x)}∞
m=1 is bounded to extract a subsequence m1 (k) such that
m (k)
Sf1 1 (x) converges as k → ∞. Out of this subsequence extract a further
m (k)
subsequence m2 (k) such that Sf2 2 (x) converges and continue this pro-
cess. Finally, pick the diagonal sequence by setting nj = mj (j) and verify
that (19.2) holds. Next, we use density of F ∗ , which implies that for any
h ∈ C(X ) and any ε > 0, there exists f ∈ F ∗ such that
max |h(y) − f (y)| < ε. (19.3)
y∈X

A direct triangle inequality step now yields that for sufficiently large j, we
may combine (19.2) and (19.3) to obtain
n n n
Sh j (x) − Sf∞ (x) ≤ S|fj−h| (x) + Sf j (x) − Sf∞ (x) ≤ 2ε.
n
Hence, Sh j (x) is a Cauchy sequence and the limit Sh∞ (x) exists for every
h ∈ C(X ) and defines a bounded linear functional
Lx (h) := Sh∞ (x), (19.4)

101
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 which is also positive in the sense that h ≥ 0 implies Lx (h) ≥ 0 by the
construction (19.1), and kLh k = 1. The Riesz Representation Theorem
implies that for a bounded positive linear functional on C(X ), there exists
a measure µx such that
Z
Lx (h) = h(y) dµx (y). (19.5)
X

However, we may simply calculate

n n 1
Sh j (g(x)) − Sh j (x) = |h(g nj (x)) − h(x)| ,
nj
which implies Sh∞ (g(x)) = Sh∞ (x), which can be combined with (19.4)
and (19.5) to conclude that µx is invariant.

Example 19.6. Consider X = [0, 1) and the Gauss transformation


0 if x = 0,
g:X →X g(x) = (19.6)
1/x − [1/x] if x 6= 0,
where [1/x] denotes the greatest integer less or equal to 1/x. The Gauss
transformation maps each interval (1/(n + 1), 1/n] for n ∈ N continuously
and monotonically onto [0, 1) with discontinuities at each 1/n; see Fig-
ure 44. The Gauss transformation is related to the continued fraction
expansion
1
x = [a1 , a2 , a3 , . . .] = 1 , aj ∈ N,
a1 + 1
a2 + a +···
3
 
of x ∈ X . Indeed, one checks if g n−1 (x) 6= 0, then we have aj = 1/g j−1 (x)
for j ≤ n. Furthermore, one observes that
g(x) = [a2 , a3 , . . .]
so that a bit more work shows the Gauss transformation conjugates to the
left-shift map on integer-valued infinite-sequences (exercise!). There exists
an invariant measure, the Gauss measure
Z b
1 1 1+b
µ((a, b)) = dy = (ln 2)−1 ln . (19.7)
ln 2 a 1 + y 1+a
Remark : One should not confuse the Gauss measure with Gaussian measures
given on Rd by
 
1 1 ⊤ −1
p exp − (x − m) Q (x − m) dx
(2π)d det(Q) 2

with positive definite covariance matrix Q ∈ Rd×d and mean vector m ∈ Rd .

102
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 To prove that (19.7) is invariant for (19.6), note that the pre-image
of an interval (a, b) consists of infinitely many intervals. If we restrict to
(1/(n + 1), 1/n), then the pre-image is (1/(n + b), 1/(n + a)). Now one
calculates
∞  !
−1
[ 1 1
µ(g ((a, b))) = µ ,
n+b n+a
n=1
∞  
1 X 1+a+n b+n
= ln · = µ((a, b))
ln 2 a+n 1+b+n
n=1

so µ is indeed a g-invariant measure. 

We have seen in the proof of Theorem 19.5 that it is often helpful to

work with observables f : X → R, which is emphasized again by the next
result.

Lemma 19.7. g : X → X is measure-preserving for the measure µ if and

only if for each f ∈ L1 (X , F, µ) we have
Z Z
f (y) dµ(y) = f (g(y)) dµ(y).
X X

Proving Lemma 19.7 is a direct application of using the definitions and

applying the Monotone Convergence Theorem (exercise!).

20 Ergodicity
Of course, it may happen that a measure-preserving transformation de-
composes into several invariant subsets. However, then it makes sense to
restrict to each subset and study indecomposable objects; see Figure 45(a).

Definition 20.1. Let (X , F, µ) be a probability space. A measure-preserving

transformation/map g : X → X is called ergodic if the only invariant mea-
surable sets have zero or full measure, i.e., g −1 (A) = A implies µ(A) = 0
or µ(A) = 1.

Lemma 20.2. A map g : X → X is ergodic if and only if for every

measurable f : X → R with f ◦ g = f almost everywhere (a.e.) implies f
constant a.e..

We shall not prove Lemma 20.2, which is a (non-trivial!) exercise.

However, one could also simply take the result as an alternative definition
of ergodicity.

103
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22


(a) (b) lim f (x) + f (g(x)) + f (g 2 (x)) + . . .
n→∞
X

=
R
X f (x)dµ

g2 g(x)
x
g1 g
g 2 (x)

X
Figure 45: (a) Illustration of the notion of ergodicity: either we can split
the phase space into two invariant regions so that g restricted to the re-
gions leaves them invariant, or this is not possible and we are ergodic. (b)
Illustration of the Birkhoff Ergodic Theorem 20.7.

Example 20.3. Consider X = S1 ≃ [0, 1]/(0 ∼ 1) and the circle rotation

map given by
xk+1 = g(xk ) := (xk + r) mod 1. (20.1)
We claim that g is ergodic with respect to Lebesgue measure on S1 if and
only if r is irrational. It suffices by a slight extension of Lemma 20.2
(exercise!) to check that if f ∈ L2 (X , R; µ) = L2 is g-invariant, then f is
constant. Consider the Fourier series of f
∞
X
an e2πinx = f (x) in L2 .
n=−∞

Since we also have

∞
X
an e2πin(x+r) = f (g(x)) = f (x) in L2 .
n=−∞

by invariance of f , the uniqueness of Fourier coefficients an implies an =

an e2πinr . Since r is irrational, e2πinr 6= 1 so an = 0 for all n 6= 0,
which means that f is constant. Therefore, irrational r implies ergodic-
ity for (20.1). The converse is equally straightforward to show. 

Ergodicity implies very nice recurrent properties. To uncover these

results, we need some preparation.

Definition 20.4. Let f be a measurable real-valued function on (X , F, µ)

and define the Koopman operator Lg by

(Lg f )(x) = f (g(x)).

104
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 One may study Lg on various function spaces. As an important case,
consider L1 = L1 (X , R; µ) and observe that
Z Z
1 1
Lg L ⊂ L , kLg f kL1 = f (g(x)) dµ(x) = f (y) dµ(y) = kf kL1 .
X X

Also one easily checks that Lg is a positive operator, i.e., f ≥ 0 implies

Lg f ≥ 0.

Theorem 20.5 (Maximal Ergodic Theorem). Consider the Koopman op-

erator Lg =: L on L1 , fix any f ∈ L1 , and define (for n ≥ 1) the following

f0 := 0, fn := f + Lf + L2 f + · · · + Ln−1 f FN := max fn ≥ 0.
0≤n≤N
R
Then it follows that {x:FN (x)>0} f (x) dµ(x) ≥ 0.

Proof. First, note that FN ∈ L1 and FN ≥ fn by construction. Since

the Koopman operator is positive, we find LFN ≥ Lfn so that using the
definition of fn , it follows that LFN +f ≥ fn+1 . Therefore, a direct estimate
yields
if FN (x) > 0
LFN (x) + f (x) ≥ max fn (x) = max fn (x) = FN (x).
1≤n≤N 0≤n≤N

This provides the elegant estimate f ≥ FN − LFN on A := {x ∈ X :

FN (x) > 0}, which we can just integrate over A to obtain
Z Z Z Z Z
f dµ ≥ FN dµ − LFN dµ = FN dµ − LFN dµ (20.2)
A A A X A

where one uses FN = 0 on X \ A for the first term. The second term on the
right-hand side in (20.2) can be estimated from below since FN ≥ 0 implies
LFN ≥ 0, which gives
Z Z Z
f dµ ≥ FN dµ − LFN dµ = 0, (20.3)
A X X

where the last equality is a consequence of g being measure-preserving.

Corollary 20.6. Suppose g : X → X is measure-preserving, h ∈ L1 (X , R; µ),

and define  
 n−1 
1 X
Bα := x ∈ X : sup h(g j (x)) > α . (20.4)
 n≥1 n 
j=0
R
If g −1 A = A then Bα ∩A h dµ ≥ αµ(Bα ∩ A).

105
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Proof. Just restrict g to A so that we can simply apply Theorem 20.5 with
A = X and f = h − α to obtain the result.

Theorem 20.7 (Birkhoff Ergodic Theorem). Let (X , F, µ) be a probability

space with measure-preserving transformation g : X → X and fix any f ∈
L1 = L1 (X , R; µ). The following statements hold for almost every x ∈ X :
(T1) There exists f ∗ ∈ L1 such that
n
1X
lim f (g j (x)) = f ∗ (x). (20.5)
n→∞ n
j=1

R R
(T2) f ∗ ◦ g = f ∗ and f dµ = X f ∗ dµ. X
R
(T3) If g is ergodic, then f ∗ ≡ X f dµ so that
n Z
1X
lim f (g j (x)) = f dµ, (20.6)
n→∞ n X
j=1

i.e., time-average equals space-average.

Proof. We start with (T1), which is actually the most difficult/crucial part.
We define
n−1 n−1
1X 1X
f ∗ (x) := lim sup f (g j (x)) and f∗ (x) := lim inf f (g j (x))
n→∞ n n→∞ n
j=0 j=0

and observe that f ∗ ◦ g = f ∗ and f∗ ◦ g = f∗ . We would like to show that (I)

f ∗ = f∗ so that the limit (20.5) exists and (II) that f ∗ ∈ L1 . For α, β ∈ R,
define Eα,β = {x ∈ X : f∗ (x) < β and α < f∗ (x)}. Since
[
{x ∈ X : f∗ (x) < f ∗ (x)} = Eα,β
α,β∈Q,β<α

it suffices to prove µ(Eα,β ) = 0 for β < α to establish (I). Observe that

g −1 Eα,β = Eα,β , define Bα as in (20.4) and note that Bα ∩ Eα,β = Eα,β .
Hence, Corollary 20.6 applies to yield
Z Z
f dµ = f dµ ≥ αµ(Eα,β ∩ Bα ) = αµ(Eα,β ). (20.7)
Eα,β Eα,β ∩Bα

Replacing f, α, β by −f, −β, −α it is easy to check (exercise!) that we also

obtain the inequality Z
f dµ ≤ βµ(Eα,β ). (20.8)
Eα,β

106
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Combining (20.7)-(20.8) gives
αµ(Eα,β ) ≤ βµ(Eα,β ) ⇒ µ(Eα,β ) = 0 if β < α.
1 Pn−1
This shows (I) and to check (II) let hn (x) := n j=0 f (g j (x)) so Fatou’s
Lemma implies f ∗ ∈ L1 since
Z n−1 Z Z
1X j
lim inf hn dµ ≤ lim inf |f ◦ g | dµ = |f | dµ.
X n→∞ n→∞ n X X j=0

The first statement in (T2), i.e., f ∗ = f ∗ ◦ g is now easy to check. For the
second part, let Dkn := {x ∈ X : k/n ≤ f ∗ (x) ≤ (k + 1)/n}, which are sets
to decompose f ∗ at “scale n” for k ∈ Z, n ∈ N. One may select ε > 0
sufficiently small so that Dkn ∩ B(k/n)−ε = Dkn , which makes Corollary 20.6
applicable again
Z   Z
k n k
f dµ ≥ − ε µ(Dk ) ⇒ f dµ ≥ µ(Dkn ).
n
Dk n n
Dk n

Using this inequality and the definition of Dkn leads to

Z Z
∗ k+1 n 1 n
f dµ ≤ µ(Dk ) ≤ µ(Dk ) + f dµ.
Dkn n n Dkn
R R
Now one sums over k to Rconclude XR f ∗ dµ ≤ n1 + X f dµ and taking
∗
R limit∗ n → R∞ implies X f dµ ≤ X ∗f dµ. Applying this to −f gives
the
RX (−f ) dµ R≤ X (−f ) dµ but since (−f ) = −f∗ by construction we have
X f∗ dµ ≥ X f dµ. Combining this with the previous inequality we have
Z Z Z
f∗ dµ ≥ f dµ ≥ f ∗ dµ
X X X

so that using f ∗ = f∗ finishes the proof of (T2). For (T3), f ∗ is constant

almost everywhere by ergodicity due to Lemma 20.2 so the result follows
by combining (T1) and (T2).

The next result is a nice consequence of the Birkhoff Ergodic Theorem.

Theorem 20.8 (Lp Ergodic Theorem of Von Neumann). Fix p ∈ [1, ∞),
suppose g : X → X is measure-preserving for a probability space (X , F, µ).
Consider f ∈ Lp = Lp (X , R; µ) and the (n-term) average
n−1
1X
(Sn f )(x) := f (g j (x))
n
j=0

then there exists f ∗ ∈ Lp s.t. f ∗ ◦ g = f ∗ and limn→∞ kSn f − f ∗ kLp = 0.

107
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Proof. Fix ε > 0. The proof is a standard ε/3-argument to check that
Sn f is a Cauchy sequence so we provide only the main steps. Consider
h ∈ L∞ such that kh − f kLp ≤ ε/3. Since X is a probability space, it
follows that h ∈ Lp . Therefore, the Birkhoff Ergodic Theorem implies that
Sn h converges in Lp . This yields the existence of a sufficiently large n ∈ N
such that kSn h − Sn+k hkLp ≤ ε/3. Hence, the triangle inequality and
kSn f kLp ≤ kf kLp give

kSn f − Sn+k f kLp ≤ kSn f − Sn hkLp + kSn h − Sn+k hkLp

ε ε ε
+kSn+k h − Sn+k f kLp = + + = ε
3 3 3
and the existence of f ∗ ∈ Lp follows, which is easily checked to be also
g-invariant.

21 Mixing
The next result hints at a crucial consequence of ergodicity, when we are
interested in return probabilities.
Theorem 21.1 (Limit of Cesàro Means). Let (X , F, µ) be a probability
space and suppose g : X → X is a measure-preserving map. Then g is
ergodic if and only if for all A, B ∈ F we have
n−1
1X
lim µ(g −j (A) ∩ B) = µ(A)µ(B). (21.1)
n→∞ n
j=0

Proof. Suppose g is ergodic, then let f = 1A be the indicator function of

A in the Birkhoff Ergodic Theorem (20.7) so that the space average is just
µ(A). Hence, we find
n−1
1X
lim 1A (g j (x)) = µ(A).
n→∞ n
j=0

Multiplying the last equality by 1B and applying the Dominated Conver-

gence Theorem gives
n−1
1X
lim µ(g −j (A) ∩ B) = µ(A)µ(B).
n→∞ n
j=0

The converse of our result assumes (21.1). Suppose S is any invariant set,
g −1 S = S, and let A = S = B. Then (21.1) implies µ(S) = µ(S)2 , so
µ(S) = 0 or µ(S) = 1, which just means ergodicity holds.

108
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 We know by Theorem 21.1 that the sets µ(g −j (A) ∩ B) become asymp-
totically independent “on average” for ergodic maps. Hence, it is very
natural to ask, whether we can do better than just average properties.

Definition 21.2. Let g be a measure-preserving transformation of a prob-

ability space (X , F, µ).

(D1) g is called weak-mixing if for all A, B ∈ F we have

n−1
X
lim µ(g −j (A) ∩ B) − µ(A)µ(B) = 0. (21.2)
1
n→∞ n
j=0

(D2) g is called strong-mixing (or just mixing) if for all A, B ∈ F we

have
lim µ(g −n (A) ∩ B) = µ(A)µ(B). (21.3)
n→∞

Mixing properties entail the - potentially quite slow - decay of corre-

lations. The next result is relatively straightforward to prove (exercise!).

Proposition 21.3. Strong-mixing ⇒ weak-mixing ⇒ ergodic.

The converse implications in Proposition 21.3 are not true; see e.g. Ex-
ample 21.10.

Exercise 21.4. Consider X = S1 ≃ [0, 1]/(0 ∼ 1) and the family of ex-

panding circle endomorphisms given by

xk+1 = g(xk ) := pxk mod 1, p ∈ Z, |p| > 1. (21.4)

One checks that Lebesgue measure µ is g-invariant. Next, consider p = 2

and two intervals A = [p/2i , (p + 1)/2i ] and B = [q/2j , (q + 1)/2j ], where
we may fixed any p, q so that A, B ⊆ [0, 1]. Then g −1 (B) is the union of
2 uniformly spaced intervals of length 1/2j+1 and similarly g −n (B) is the
union of 2n uniformly spaced intervals of length 1/2j+n . Hence, for n > i,
we have that A ∩ g −n (B) consists of 2n−i intervals of length 1/2n+j and we
calculate

µ(A ∩ g −n (B)) = 2n−i (1/2n+j ) = 1/2i+j = µ(A)µ(B)

so that an approximation procedure shows that (21.4) is strong-mixing for

p = 2. ♦

The considerations the last example can easily be turned into a more
general proof of the following result.

109
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Proposition 21.5. Expanding circle endomorphisms are strong-mixing.
For more general maps, proving mixing and/or working with proper-
ties directly becomes almost impossible. However, an operator-theoretic
viewpoint provides a nice characterization.
Definition 21.6. Let Lp (X , R; µ) =: Lp for p ∈ [1, ∞]. View the Koopman
operator as Lg = L : L∞ → L∞ and let Pg = P : L1 → L1 be its adjoint
defined by
Z Z
!
hPf, hiL2 = (Pf )h dµ = f (Lh) dµ = hf, LhiL2 . (21.5)
X X

for f ∈ L1
and h ∈ L∞ . The operator P is called the Perron-Frobenius
operator.
Theorem 21.7 (Operator-Characterization of Mixing). Let (X , F, µ) be a
probability space and g : X → X be measure-preserving. Then g is strong-
mixing if and only if
lim hP n f, hi = hf, 1ih1, hi, f ∈ L1 , h ∈ L∞ , (21.6)
n→∞

where h·, ·i = h·, ·iL2 (X ,R;µ) and P = Pg .

Proof. We start by showing that strong-mixing yields (21.6). The mixing
condition (21.3) can be written in integral form as
Z Z Z
lim 1A (x)1A (g n (x)) dµ(x) = 1A (x) dµ(x) 1B (x) dµ(x)
n→∞ X X X
so that using the definition of the Koopman and Perron-Frobenius operators
gives
h1A , 1ih1, 1B i = lim h1A , Ln 1B i = lim hP n 1A , 1B i (21.7)
n→∞ n→∞

so the desired equality holds for indicator functions, and by linearity also
for linear combinations of indicator functions, which are just the standard
simple functions. We approximate h ∈ L∞ by simple functions hk such
that hk → h, and we approximate f ∈ L1 by simple functions such that
fk → f . Then we have
|hP n f, hi − hf, 1ih1, hi| ≤ |hP n f, hi − hPfk , hk i|
+|hP n fk , hk i − hfk , 1ih1, hk i| (21.8)
+|hfk , 1ih1, hk i − hf, 1ih1, hi|
so that a standard ε/3-argument yields the result as all three terms on the
right-hand side of (21.8) can be made small. The converse of the result is
easier since if (21.6) holds then setting f = 1A and h = 1B gives strong-
mixing.

110
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 The last results naturally point into the direction of using invariants of
the Koopman operator L (respectively the Perron Frobenius operator P)
to characterize/classify measure-preserving transformations. Observe that
Lg = L : L2 → L2 is a unitary operator since it preserves the L2 inner
product. Hence, if Lf = λf for some λ ∈ C, i.e., λ is an eigenvalue and f
is an eigenfunction, then λ ∈ {z ∈ C : |z| = 1}.
Definition 21.8. Consider a probability space (X , F, µ) and let L2 =
L2 (X , F, µ). The Koopman operator L = Lg : L2 → L2 , defined by
Lf (x) = f (g(x)), has:
(D1) pure-point spectrum if the eigenfunctions of L span L2 ;
(D2) continuous spectrum if 1 is the only eigenvalue.
Remark : Of course, many linear operators have both types of spectra, i.e., some
eigenvalues and then some remaining part consisting of continuous spectrum. This
is highly relevant for the dynamical systems analysis of PDEs [15].
The next result illustrates that one can characterize a lot of dynamical
effects of g : X → X purely by looking at the spectrum of Lg = L.
Theorem 21.9 (Spectral Characterization). The following results hold:
(T1) g is ergodic if and only if 1 is a simple eigenvalue of L.
(T2) If g is ergodic, then every eigenvalue of L is simple.
(T3) Suppose g is invertible then g is weak-mixing if and only if g has
continuous spectrum.
Proof. (T1) is easy to check (exercise!). Regarding (T2), if g is ergodic,
then every eigenfunction is constant a.e., so that if there are two eigenfunc-
tions f1 and f2 with the same eigenvalue then f1 /f2 is also an eigenfunction
with eigenvalue 1 so f1 /f2 is also constant a.e., which proves the result. For
(T3), g weak-mixing implying continuous spectrum is relatively easy (exer-
cise!). The converse is difficult, i.e., we now assume that g has continuous
spectrum. An equivalent characterization of weak-mixing is
n−1
1X
lim |hLj f, f i − hf, 1ih1, f i|2 = 0 for f ∈ L2 .
n→∞ n
j=0

For a constant f , this is obvious, so we only need to check it in the orthog-

onal complement to the constants in L2 , i.e., when hf, 1i = 0. Therefore,
we have to prove
n−1
1X
lim |hLj f, f i|2 = 0 for f ∈ L2 . (21.9)
n→∞ n
j=0

111
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Since L is a unitary operator on the Hilbert space L2 , we may apply the
Spectral Theorem, which guarantees for every f ∈ L2 the existence of a
unique finite Borel measure µf on the spectrum K such that
Z
n
hL f, f i = λn dµf (λ). (21.10)
K

Furthermore, since L has only continuous spectrum, the Spectral Theorem

guarantees that µf has no atoms. Combining (21.9) and (21.10) means we
just have to prove
n−1 Z 2
1X
lim λn dµf (λ) =0 for f ∈ L2 . (21.11)
n→∞ n K
j=0

Since for points λ in the spectrum we have λ̄ = λ−1 , we may re-write the
sum in (21.11) as
n−1 Z n−1 Z Z 
1X n 1X 2 n −n
| λ dµf (λ)| = λ dµf (λ) λ dµf (λ)
n K n K K
j=0 j=0
n−1 Z
1X
= (λτ̄ )n d(µf × µf )(λ, τ )
n K×K
j=0
Z n−1
1X
= (λτ̄ )n d(µf × µf )(λ, τ ) (21.12)
K×K n
j=0

where the second equality step follows from a re-labelling of the second
variable and using Fubini’s Theorem; essentially, the last argument is a
classical case of doubling variables. For points off the diagonal in K × K,
we get
n−1  
1X n 1 1 − (λτ̄ )n
lim (λτ̄ ) = lim = 0. (21.13)
n→∞ n n→∞ n 1 − (λτ̄ )
j=0

Since µf is non-atomic, the diagonal has measure zero, the limit in (21.13)
holds a.e., and the Dominated Convergence Theorem can be applied to (21.12)
to get the result.

Example 21.10. Consider the circle rotation g(x) = x + r (mod 1) for

x ∈ S1 = R/Z. For any n ∈ Z, the function fn = exp(2πinx) is an
eigenfunction of L since

Lfn (x) = fn (g(x)) = e2πinr e2πinx = e2πinr fn (x).

112
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 If r ∈ R\Q, then the eigenfunctions span L2 and Theorem 21.9(T3) implies
that the irrational circle rotation is not weak-mixing. However, it is shown
in Example 20.3 that irrational circle rotations are ergodic; cf. Proposi-
tion 21.3. 

22 Entropy
Spectral properties provide one possible way to distinguish two dynamical
systems. Hence, a natural question to ask is, which other useful invariants
one may define? Borrowing the concept of entropy from statistical physics
has turned to yield one very helpful invariant. There are two related ap-
proaches to entropy, a measure-theoretic and a topological. We start with
the measure-theoretic and consider a measure-preserving transformation

g:X →X

on a probability space (X , F, µ).

Definition 22.1. A (finite) partition A = {A1 , . . . , Am } of (X , F, µ) is

a disjoint collection of elements of F, whose union is X . The join (or
common refinement) A ∨ B of two partitions A and B is the partition
into intersections Am ∩ Bn .

Usually we shall assume that a partition A is directly related to a sub-

(σ-)algebra, i.e., by considering all elements in F which can be written as
unions of elements of A. Similarly we may relate a given sub-algebra of F
to a partition (exercise!). Furthermore, we write A ⊆ B to mean that B is
a refinement (or finer partition), i.e., elements of A can be constructed
via unions of elements of B.

Definition 22.2. For a partition A = {A1 , . . . , Am } define the entropy

m
X
H(A) := − µ(Aj ) ln µ(Aj )
j=1

Entropy has a natural interpretation as average information of the el-

ements of a partition. Indeed, let I(C) := − ln µ(C) be the information
of the event (or measurable set) C, then the information is a non-negative
decreasing function, i.e., observing a lower probability event gives more
information. Furthermore, the trivial event X has I(X ) = 0 and for inde-
pendent events, say µ(C1 ∩ C2 ) = µ(C1 )µ(C2 ), we have

I(C1 ∩ C2 ) = I(C1 ) + I(C2 ), (22.1)

113
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 so information is additive; in fact, using a logarithm (up to a constant)
is the only choice if one requires the last observations from a definition of
information.

Definition 22.3. Consider a measure-preserving transformation g : X →

X and a partition A. Define the entropy of g with respect to A as
 
n−1
1  _ −j 
h(g, A) := lim H g A (22.2)
n→+∞ n
j=0

where g −j A = {g −j A1 , . . . , g −j Am }. Finally, define the entropy of g by

taking the supremum over all finite partitions

h(g) := sup h(g, A).

A finite

It can be shown (non-trivial exercise!) that the limit in (22.2) actually

exists. The next result is expected (exercise!) since we can just insert a
conjugating map in the last definition without too much trouble.

Theorem 22.4 (Entropy as Conjugacy Invariant). Two measure-preserving

transformations, which are topologically conjugate, have the same entropy.

The main problem is that entropy is still almost impossible to calculate

from the definition of h(g) for concrete examples. A technical tool is needed
to “sort” the information into smaller pieces. It is a natural concept to
condition upon previous knowledge, i.e., for measurable sets C1 , C2 with
µ(C2 ) > 0 define the conditional probability as

µ(C1 ∩ C2 )
µ(C1 |C2 ) := .
µ(C2 )
Hence, we can now average the usual entropy conditionally over another
partition.

Definition 22.5. For two partitions A, B define the conditional entropy

X
H(A|B) := − µ(Bj )µ(Ai |Bj ) ln µ(Ai |Bj ).
i,j

Theorem 22.6 (Properties of Conditional Entropy). If A, B, C are finite

sub-algebras of F, then we have

(T1c) H(A ∨ B) = H(A) + H(B|A);

(T2c) A ⊆ B ⇒ H(A) = H(B);

114
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
(T3c) C ⊆ B ⇒ H(A|C) ≥ H(A|B);

(T4c) H(A ∨ B|C) ≤ H(A|C) + H(B|C);

(T5c) H(g −1 A|g −1 B) = H(A|B) and H(g −1 A) = H(A).

Remark : The properties are all quite intuitive, e.g., (T1) claims that the entropy
(“information”) contained in the join of A, B is the entropy of A plus the additional
information in B.

Proof. Most of the properties can be proved by (quite lengthy) definition-

chasing. However, let us prove (T3) to see, how the structure of the entropy
is used. Consider the function ψ : [0, ∞) → R defined by ψ(x) := x ln x
with the obvious limit point definition ψ(0) = 0. Observe that ψ is strictly
convex so that, for fixed i, j, we have
!
X µ(Bk ∩ Cj ) µ(Ai ∩ Bk ) X µ(Bk ∩ Cj )  µ(Ai ∩ Bk ) 
ψ ≤ ψ .
µ(Cj ) µ(Bk ) µ(Cj ) µ(Bk )
k k

Since C ⊂ B, the left-hand side of the last inequality simplifies to

 
µ(Ai ∩ Cj ) µ(Ai ∩ Cj ) µ(Ai ∩ Cj )
ψ = ln .
µ(Cj ) µ(Cj ) µ(Cj )

Using these last two results, multiplying both sides by µ(Cj ) and summing
over the indices i, j we get
X µ(Ai ∩ Cj ) X µ(Ai ∩ Bk ) µ(Ai ∩ Bk )
µ(Ai ∩ Cj ) ln ≤ µ(Bk ∩ Cj ) ln
µ(Cj ) µ(Bk ) µ(Bk )
i,j i,j,k
X µ(Ai ∩ Bk ) µ(Ai ∩ Bk )
= µ(Bk ) ln ,
µ(Bk ) µ(Bk )
i,k

which just means −H(A|C) ≤ −H(A|B) and (T3) follows.

Several properties of Theorem 22.6 have natural consequences for the

entropy of a transformation g.
Theorem 22.7 (Properties of Entropy). The following hold:
(T1) h(g, A) ≤ H(A);

(T2) h(g, A ∨ B) ≤ h(g, A) + h(g, B);

(T3) h(g, A) ≤ h(g, B) + H(A|B);

W 
k
(T4) if g is invertible, then h(g, A) = h j=−k gj A .

115
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Proof. As above, proofs are lengthy but not conceptually difficult. Let us
just show that, how the results of Theorem 22.6 lead to (T3). We want to
use the definition in (22.2) and start with the estimate
     
n−1
_ (T2c) n−1
_ n−1
_
H g −j A ≤ H  g −j A ∨  g −j B 
j=0 j=0 j=0
   
n−1
_ n−1
_ n−1
_
(T1c)
= H g −j B  + H  g −j A g −j B  .
j=0 j=0 j=0

The second term on the right can be be estimated from above as follows
   
n−1
_ n−1
_ (T4c) n−1
X n−1
_
H g −j A g −j B  ≤ H g −j A g −j B 
j=0 j=0 j=0 j=0

(T3c) n−1
X
≤ H(g −j A|g −j B)
j=0
(T5c)
= nH(A|B).
Hence, combining this estimate with the previous one, dividing by n, and
taking the limit n → ∞ yields the result.
Finally we can introduce a crucial theorem, how we may calculate en-
tropy in certain cases as it allows us to select certain sub-algebras respec-
tively certain types of partitions.
Theorem 22.8 (Kolmogorov-Sinai). Suppose g is a measure-preserving
invertible transformation on (X , F, µ). Let A be a finite sub-algebra of F
such that
n
_
lim g j A =: lim An = F,
n→∞ n→∞
j=−n
i.e., A is generating, then we may conclude h(g) = h(g, A).
Proof. The idea of the proof is essentially an approximation procedure in-
volving multiple steps with error terms in the form of conditional entropies,
which have to be controlled. So let us first find the error term.
Step 1: (error term identification) To conclude our result, it is going to
sufficefrom the definition of h(g) to consider any finite partition B ⊆ F and
prove that h(g, B) ≤ h(g, A). Using Theorem 22.7, we the main estimate
(T 3)
h(g, B) ≤ h(g, An ) + H(B|An )
(T 4)
= h(g, A) + H(B|An ).

116
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 So if we could show that limn→∞ H(B|An ) = 0 for a generating partition
A, then the result would follow.
Step 2: (partition approximation) Recall the symmetric difference
between two sets C1 , C2

C1 △C2 := (C1 \ C2 ) ∪ (C2 \ C1 ) .

We claim that given ε > 0 and two finite equal-length partitions D, B we

have the existence of a δ > 0 such that
X
d(D, B) := µ(Dj △Bj ) < δ ⇒ ρ(D, B) := H(D|B)+H(B|D) < ε,
j
(22.3)
where ρ is also known as the Rokhlin metric. SConsider a new partition
C constructed by sets Di ∩ Bj (i 6= j) and sets j (Dj ∩ Bj ). One notices
D∨B = D∨C and D∨B = B∨C; we focus on the last equality, theSother case
follows by a similar argument. Since for i 6= j, we have Di ∩Bj ⊂ j Dj △Bj ,
our assumption in (22.3) implies
 
[
µ(Di ∩ Cj ) < δ for i 6= j, and µ  (Dj ∩ Cj ) > 1 − δ.
j

This implies we can calculate an entropy estimate

ε
H(C) ≤ κδ ln δ − (1 − δ) ln(1 − δ) < for some κ > 1,
2
where the last estimate follows by making δ sufficiently small. Therefore,
we find
ε
H(B) + H(D|B) ≤ H(D ∨ B) = H(B ∨ C) ≤ H(B) + H(C) < H(B) + ,
2
which implies H(D|B) < 2ε . By the same line of argument, one also gets
H(B|D) < 2ε and (22.3) follows. Otherwise said, if the partitions do not
differ too much in the sense of symmetric differences, the Rokhlin metric is
small, i.e., there is a uniformly continuous inclusion map from the metric
space of partitions under the metric d to the metric ρ.
Step 3: (sub-algebra approximation) Recall that An is a sequence of
finite subalgebras generating F eventually. Let us take a finite subalgebra
B of F. We claim that we can find a finite algebra D contained in A∞ such
that
H(D|B) + H(B|D) ≤ ε, (22.4)

117
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 which is going to follow once we have verified the symmetric difference
condition in (22.3). Note that the statement (22.4) is already very close to,
what we want, i.e., it gives us control over arbitrary subalgebras just using
our generating sequence An . To prove (22.4), first let us select (A∞ )j such
that µ(Bj ∩ (A∞ )j ) < δ̃ for some δ̃ > 0, which can be chosen sufficiently
small in the end of the argument since A∞ does generate F by assumption.
Then we set [
Dj := (A∞ )j \ ((A∞ )i ∩ (A∞ )k ),
i6=k

S partition of the same length as B. Then we add one

which gives us a finite
more set D∗ := X \ j Dj and let D = {D1 , . . . , Dj , . . . , D∗ } be our partition.
Then it is possible to check (exercise!) that D indeed satisfies (22.4) as we
can select δ̃ sufficiently small.
Step 4: (limit of the approximation) Finally, we can tackle the error
term H(B|An ). By (22.4), we can find D such that
H(B|D) < ε
but D is completely contained in An0 for some n0 as D is finite. For n ≥ n0 ,
we finally get by Theorem 22.6
(T 3c) (T 3c)
H(B|An ) ≤ H(B|An0 ) ≤ H(B|D) < ε
and since ε > 0 was arbitrary we may take the limit n → ∞ to conclude
that the error term derived in the first step vanishes.

With the Kolmogorov-Sinai Theorem in hand, we are finally ready to

use the entropy as aa conjugacy invariant to prove that certain dynamical
systems cannot be topologically conjugate.
P
Example 22.9. Consider a vector p = (p0 , . . . , pk−1 ) with j pj = 1. Let
X = Σk be the space of bi-infinite sequences of symbols from {0, 1, . . . , k −
1}. A suitable sigma-algebra F is generated by the cylinder sets
Cα := {x ∈ X : xi = αi , −m ≤ i ≤ n},
where α is a given finite sequence of symbols and the associated measure µ
is just
n
Y
µ(Cα ) := p αi ,
i=−m
which associates a probability of occurrence to a sequence using the given
vector p. The dynamics will be given just by the usual left-shift map
g(x; p)n := xn+1 , g : (X , F, µ) → (X , F, µ),

118
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 which is a measure-preserving transformation. If we want to compute the
entropy via the Kolmogorov-Sinai Theorem 22.8 requires the choice of a
helpful generating subalgebra. One natural choice is

Aj := {x : x0 = j}, A = (A0 , . . . , Ak−1 )

since the sets are simple and one verifies quite quickly that we can obtain
all cylinder sets upon using shift dynamics so that
∞
_
g j A = F.
j=−∞

Observe that the typical elements of An have measure that are easy to
calculate
 
µ Aj0 ∩ g −1 Aj1 ∩ g −(n−1) Ajn−1 = µ ({x : x0 = j0 , . . . xn−1 = jn−1 })
= pj0 · · · pjn−1 .

Hence, we can also determine entropy of the generating subalgebra by sum-

mation
X
H(An ) = − (pj0 · · · pjn−1 ) ln(pj0 · · · pjn−1 )
k−1
X

=− (pj0 · · · pjn−1 ) ln pj0 + · · · + ln pjn−1
j0 ,...,jn−1 =0
k−1
X
= −n pj ln pj
j=0

and the Kolmogorov-Sinai Theorem 22.8 yields

k−1
1 X
h(g) = h(g, A) = lim H(An ) = pj ln pj .
n→∞ n
j=0

Therefore, the Bernoulli shift on two symbols with p = ( 12 , 12 ) and the

Bernoulli shift on three symbols with p = ( 13 , 13 , 13 ) are not topologically
conjugate by Theorem 22.4 since their entropies are ln 2 and ln 3 respec-
tively. 

The last results on entropy show that it is actually quite difficult to

coarse-grain dynamical systems, i.e., to extract useful low-dimensional in-
variants, which can be used to partition systems into different conjugacy
classes.

119
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 References
[1] K.T. Alligood, T.D. Sauer, and J.A. Yorke. Chaos: An Introduction to Dynamical Systems.
Springer, 1996.
[2] V.I. Arnold. Ordinary Differential Equations. MIT Press, 1973.
[3] V.I. Arnold. Geometrical Methods in the Theory of Ordinary Differential Equations.
Springer, New York, NY, 1983.
[4] M. Brin and G. Stuck. Introduction to Dynamical Systems. CUP, 2002.
[5] C. Chicone. Ordinary Differential Equations with Applications. Texts in Applied Mathe-
matics. Springer, 2nd edition, 2010.
[6] P. Glendinning. Stability, Instability and Chaos. CUP, 1994.
[7] J. Guckenheimer and P. Holmes. Nonlinear Oscillations, Dynamical Systems, and Bifur-
cations of Vector Fields. Springer, New York, NY, 1983.
[8] J.K. Hale. Ordinary Differential Equations. Dover, New York, NY, 2009.
[9] J.K. Hale and H. Kocak. Dynamics and Bifurcations. Springer, New York, NY, 1991.
[10] M.W. Hirsch, S. Smale, and R. Devaney. Differential Equations, Dynamical Systems, and
an Introduction to Chaos. Academic Press, 2nd edition, 2003.
[11] G. Iooss and D.D. Joseph. Elementary Stability and Bifurcation Theory. Springer, 1997.
[12] A. Katok and B. Hasselblatt. Introduction to the Modern Theory of Dynamical Systems.
CUP, 1995.
[13] C. Kuehn. Multiple Time Scale Dynamics. Springer, 2015. 814 pp.
[14] C. Kuehn. Dynamical Systems I: ODEs & Introduction to Nonlinear Systems. AMS Open-
Math Notes, 2018.
[15] C. Kuehn. PDE Dynamics: An Introduction. SIAM, 2019.
[16] Yu.A. Kuznetsov. Elements of Applied Bifurcation Theory. Springer, New York, NY, 3rd
edition, 2004.
[17] A. Lasota and M.C. Mackey. Chaos, Fractals and Noise. Springer, 1994.
[18] G.C. Layek. An Introduction to Dynamical Systems and Chaos. Springer, 2015.
[19] A.J. Lichtenberg and M.A. Lieberman. Regular and Chaotic Dynamics. Springer, 1992.
[20] J.D. Meiss. Differential Dynamical Systems. SIAM, 2007.
[21] L. Perko. Differential Equations and Dynamical Systems. Springer, 2001.
[22] R.C. Robinson. An Introduction to Dynamical Systems: Continuous and Discrete. AMS,
2013.
[23] R. Seydel. Practical Bifurcation and Stability Analysis. Springer, 1994.
[24] S. Sternberg. Dynamical Systems. Dover, 2010.
[25] S.H. Strogatz. Nonlinear Dynamics and Chaos. Westview Press, 2000.
[26] G. Teschl. Ordinary Differential Equations and Dynamical Systems. AMS, 2012.
[27] F. Verhulst. Nonlinear Differential Equations and Dynamical Systems. Springer, 2006.
[28] P. Walters. An Introduction to Ergodic Theory. Springer, 2000.
[29] S. Wiggins. Global Bifurcations and Chaos. Springer, 1998.
[30] S. Wiggins. Introduction to Applied Nonlinear Dynamical Systems and Chaos. Springer,
New York, NY, 2nd edition, 2003.

120
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Index
ε/3-argument, 108 Cesàro mean, 108
chaotic
adjoint, 50 attractor, 97
Anosov diffeomorphism, 35 chaotic set, 83
aperiodic, 32 chart, 15
Arnold’s cat map, 22 circle endomorphism, 109
atlas, 15 circle rotation, 8, 104
attracting set, 94 cobweb, 5
attractor, 94 codimension, 44
chaotic, 97 common refinement, 113
hyperbolic, 97 complexification, 45
Milnor, 96 conditional
strange, 97, 99 entropy, 114
Banach Fixed Point Theorem, 6 probability, 114
basin of attraction, 88, 96 cones
bifurcation, 39 stable/unstable, 38
Bogdanov-Takens, 53 conjugate, 23
condition, 44 constant of motion, 13
diagram, 40 continued fraction expansion, 102
flip, 46 continuous spectrum, 111
fold, 41, 42, 46 continuous-time, 5
function, 49, 63 contraction mapping, 6
Hopf, 45 critical slowing down, 89
Neimark-Sacker, 48 curve, 16
period-doubling, 46 cycle, 8
pitchfork, 40 cylinder set, 118
saddle-node, 41 decay of correlations, 109
transcritical, 44 determinacy, 59
unfolding, 41 diagonal argument, 101
Birkhoff Ergodic Theorem, 106 differentiable manifold, 15
blow-down, 60 dimension
blow-up, 6, 60 box-counting, 98
Bogdanov-Takens discrete-time, 5
bifurcation, 53 doubling map, 81
box-counting dimension, 98 doubling operator, 90
Duffing equation, 22, 79
cat map, 22
dynamical system, 4
center, 14
manifold, 54 eigenfunction, 111

121
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 eigenvalue harmonic oscillator, 10
of an operator, 111 Hartman-Grobman Theorem, 24
entropy, 113, 114 heteroclinic orbit, 62
equilibrium, 7 homological equation, 50
equivalence, 23 Hopf
ergodic, 103 bifurcation, 45
Ergodic Theorem, 106, 107 horizontal curve, 72
expanding circle endomorphism, 109 horizontal strip, 72
hyperbolic, 17
Feigenbaum constants, 92 attractor, 97
fiber topological equivalence, 40 set, 34
first integral, 13 toral automorphism, 22
first Lyapunov coefficient, 45
fixed point, 6, 7 inclination lemma, 73
flip indecomposable, 94
bifurcation, 46 information, 113
flow, 5 intermittency, 89
focus, 14 invariance equation, 56
fold invariant, 8, 113
bifurcation, 42, 46 negatively, 8
fold bifurcation, 41 positively, 8
Fourier series, 104 invariant measure, 99
fractal, 99 iterated map, 5
Fredholm alternative, 51
functional equation, 25, 90 jet, 59
fundamental solution, 29 join, 113

Gauss Koopman operator, 104

measure, 102 Krylov-Bogulubov Theorem, 100
transformation, 102 lambda-lemma, 73
Gaussian measure, 102 LaSalle’s invariance principle, 12
generating subalgebra, 116 length, 34
generic, 33 Li-Yorke Theorem, 86
genericity Lie bracket, 50
conditions, 44 Lie derivative, 13
geometric desingularization, 60 linearized
graph transform, 22 system, 17
Hénon map, 49 local
Hadamard-Perron method, 22 flow, 6
Hamilton’s equations, 13 topological equivalence, 24
Hamiltonian, 13 logistic map, 5
Lorenz system, 94

122
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Lyapunov period, 8
coefficient, 45 period-doubling
exponent, 84 bifurcation, 46
function, 11 period-doubling sequence, 90
stability, 9 periodic
stability criterion, 11 orbit, 8
Lyapunov-Perron method, 18 Perron-Frobenius operator, 110
phase
manifold, 15 portrait, 9
stable/unstable, 18, 21 phase space, 4
maximal, 94 pitchfork bifurcation, 40
Maximal Ergodic Theorem, 105 Poincaré
measure-preserving, 99 map, 28
Melnikov function, 78 normal form, 50
Milnor attractor, 96 section, 27
minimum principle, 87 Poincaré
mixing, 109 recurrence, 100
strong, 109 point
weak, 109 spectrum, 111
Morse-Smale flow, 33 positive operator, 105
multi-valued, 6 positively invariant, 8
multipliers, 14 pseudo-orbit, 36
near identity, 50 rectification theorem, 29
negatively invariant, 8 recurrence, 100
Neimark-Sacker refinement, 113
bifurcation, 48 renormalization, 89
Neumann series, 25 residual set, 33
node, 14 resonant terms, 50
non-degeneracy condition, 44 Riemannian
nonresonance condition, 67 manifold, 34
nonwandering, 32 metric, 34
normal form, 42, 45 Riesz Representation Theorem, 102
Poincaré, 50 Rokhlin metric, 117
nullcline, 8 rotation map, 8, 104
observable, 101 route-to-chaos, 90
ODE S-unimodal, 92
linear, 4 saddle, 14
orbit, 7 saddle quantity, 64
partition, 113 saddle-node bifurcation, 41
Peixoto’s Theorem, 33 scaling, 62, 89

123
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Schwarzian derivative, 87 vector, 16
self-similar, 89 tangent bundle, 35
semiflow, 5 tangent map, 34
sensitive dependence, 81 tangent space, 16
shadowing, 36 tent map, 82
theorem, 36 time, 4
Sharkovskii time average, 106
ordering, 86 time-T map, 30
Sharkovskii Theorem, 86 time-one map, 26
shift map, 8 topological
Singer’s Theorem, 88 normal form, 42, 45
Smale horseshoe, 70 topological equivalence, 23
Smale-Williams solenoid, 96 fiber, 40
solenoid, 96 topologically transitive, 82
space average, 106 trajectory, 7
Spectral Theorem, 112 transcritical
spectrum, 111 bifurcation, 44
split function, 63 transversality
stability condition, 44
Lyapunov, 9 transverse
stable homoclinic point, 73
cone, 38 transverse intersection, 28
distribution, 35 trapping region, 93
eigenspace, 17 travelling wave, 69
manifold, 18, 21
structurally, 32 unfolding, 41
state space, 4 unimodal, 92
steady state, 7 unit tangent bundle, 38
strange universal, 92
attractor, 97 unstable, 9
strange attractor, 99 cone, 38
stratum, 40 distribution, 35
strict Lyapunov function, 11 eigenspace, 17
strong-mixing, 109 manifold, 18, 21
structurally stable, 32 structurally, 32
subcritical Hopf, 46
van der Pol, 7
supercritical Hopf, 46
variation-of-constants, 18
superstable, 91
variational equation, 29
suspension, 30
vector field, 6
symmetric difference, 117
vertical curve, 72
tangent vertical strip, 72

124
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22
 Von Neumann Ergodic Theorem, 107

wave
speed, 69
weak-mixing, 109

125
AMS Open Math Notes: Works in Progress; Reference # OMN:201709.110713; Last Revised: 2019-01-05 10:35:22