H OMOCLINIC ORBITS AND CHAOTIC BEHAVIOR IN

C LASSICAL M ECHANICS

P EYAM RYAN TABRIZIAN

D ECEMBER 1, 2008

Abstract
This paper is a study of one of the most beautiful phenomena in dynamical
systems: homoclinic orbits. We will first define what a homoclinic orbit is, then
we will study some of the properties of homoclinic points, using methods from
symbolic dynamics. In particular, we will prove that, under certain assumptions,
the existence of one homoclinic point implies the existence of a whole Cantor set of
homoclinic points. We will also implicitly construct the famous Smale horseshoe.
Finally, we will present a topic in physics that is closely related to homoclinic
orbits, namely the restricted 3-body problem.
Introduction
Homoclinic orbits have been introduced by P oincaré more than a century ago,
and since then, they became a fundamental tool in the study of chaos. In this paper, we
are going to study the chaotic dynamics in a neighborhood of a homoclinic orbit. We
will see how the analysis of those orbits, which may seem impossible at first, can be
reduced to the study of the space of sequences on {0, 1, ..., N } (where N ∈ N might
be equal to ∞), which is pretty straightforward . Our study will closely follow the texts
of Ekeland [1] (for the first section) and of Moser [3] (for the remaining sections).

1 Homoclinic orbits
Let φ be a C 1 -diffeomorphism, and O a fixed point of φ. We assume that O is
hyperbolic, i.e. dφ has real eigenvalues λ1 and λ2 such that 0 < |λ1 | < 1 < |λ2 |.
This guarantees the existence of a stable manifold S and an unstable manifold U for φ,
intersecting in O. Also, we assume that φ is an area-preserving transformation, i.e. the
determinant of Dφ is 1.

By definition, of S and U, we have:

n o
S = z | lim φn (z) = O
n→∞

and
 
n
U= z | lim φ (z) = O
n→−∞

Basically, S is the set of points that tend to O (under iteration of φ) and U is the set
of points that emanate from O.
We assume that U and S further intersect in a point H, called homoclinic point (see
Figure 1). More generally, a homoclinic point is defined as follows:

Definition. If r 6= p and φk (r) approaches one periodic orbit p, φ(p), φ2 (p), ..., φm (p) = p ,
as k → ±∞, then r is called homoclinic point
Note. There is also what is called a heteroclinic point in case there are 2 fixed points,
p and q, but we won’t deal with that situation in this paper.
H ∈ S, hence H1 = φ(H), H2 = φ2 (H), etc. all tend to O. We’re about to prove
the following remarkable fact:
Claim. ∀n, a branch of U goes through Hn (see Figure 2)

Proof. Because H ∈ U , we have H1 = φ(H) ∈ φ(U ). Hence, it suffices to prove

that φ(U ) ⊆ U . Let M ∈ φ(U ). By continuity of φ, if M is close to H1 , we have
that M = φ(P ), where P is a point in U close to H1 . And so M −1 = P (where,
in general, M −k = φ−k (M )). Similarly, M −1 and P have the same preimages by φ,

1
 Figure 1: The stable and unstable curves, intersecting in H

which gives M −2 = P −1 (with P −k = φ−k (P )). In the same way, we conclude that
M −3 = P −2 , M −4 = P −3 , and so on. But because P ∈ U , we have, by definition of
U,
lim P−n = O and lim M−n = O
n→∞ n→∞

But this implies, by definition of U, that M ∈ U . Because M was arbitrary, this

proves our claim.
Corollary. There is a doubly infinite sequence of homoclinic points
Proof. Just take the sequence Uk = Hk , k ∈ Z, with Hk = φk (H)!
Moreover, because Dφ has determinant 1, φ also preserves orientations, hence, one
can prove using the intermediate value theorem that, for all n, between Hn and Hn+1
there is another point of intersection between U and S, let’s call those points Kn (with
K0 = K). Obviously, the Kn are homoclinic points as well, so the sequence Vk = Kk
is another sequence of new homoclinic points.

So we started with one homoclinic point, and we arrived at two doubly-infinite se-
quences of homoclinic points! There surely must be something chaotic going on here!

Let’s try connect the Hi . We have to keep the following in mind:

1. U cannot intersect itself. If it did, then by reverse iteration of φ, this would

mean that U would intersect itself in a neighborhood of O, which is not possible
by a theorem on stable/unstable manifolds.
2. U and S cannot intersect themselves tangentially. We would get a similar
contradiction.

2
 Figure 2: ∀n, a branch of U goes through Hn

3. The areas of each region exterior to S must be equal. Same for the regions
exterior to U. This is because we assume that φ is an area-preserving transfor-
mation.
4. U and S must be smooth curves
5. ∀n, Hn Hn+1 < Hn−1 Hn

See Figure 3 (at the end of this section) for a picture illustrating those ideas.
The most natural way of connecting the Hi is in such a way that the connected path
always remains on one side of the picture, namely on the side of S (see Figure 4). But
this is wrong because, by definition of U and S, any point N very close to S but not
on S will approach O under iteration of φ, but, since N ∈ / S, we have that, eventually,
N will leave O under iteration of φ. Hence, there must be at least part of U that lands
back on S, close to H.
Also note that, under those assumptions, U becomes finer and longer under iteration
of φ. This implies that U will intersect S in infinitely many new points.
And this is not all of it! Namely, the new points of intersection between U and S are
new homoclinic points, and we can repeat this procedure for all points of intersection.
The result is a pretty chaotic picture, which means that the study of homoclinic orbits
is definitely not as straightforward as people in the past century thought it would be.
And this is still not it! Because so far, we only dealt with one side of the picture.
There is still one branch of U and one branch of S we didn’t connect. If we assume that
those two branches intersect in a new (homoclinic) point H’, we get the same picture as
before. So we get, again, a bunch of new homoclinic points (see Figure 6). However,

3
 Figure 3: Things U and S cannot do (errors circled in red)

we will not deal with this kind of situation in this paper, because this problem is easily
reduced to the problem we are considering. Also, it is physically more realistic to
assume that the two branches we didn’t connect blow off to infinity (which illustrates
the impossibility of going backwards in time).

2 Deeper study of the problem

From now on, for more legibility (in particular in order to not confuse points with
sets), H will be denoted by r. U, S, O remain the same as in the previous section.
Consider a quadrilateral R, centered at r, such that one of its sides concides with
part of U and another one with part of S, and the other sides are parallel to the tangents
of U resp. S at r (see Figure 5). We wish to study the dynamics of φ in R. For this, we
must consider the set of points which don’t eventually escape R. Thus, for q ∈ R Let
k = k(q) be the smallest positive integer for which φk (q) ∈ R, if it exists. Let D(φ) e
be the set of q ∈ R for which such a k > 0 exists (this guarantees that we are studying

4
 Figure 4: First attempt to connect the Hi

Figure 5: Chaotic behavior of homoclinic orbits (Part I)

5
Figure 6: Chaotic behavior of homoclinic orbits (Part II)

6
iterates that don’t blow off to infinity) and set:

e = φk (q), q ∈ D(φ)
φ(q) e

φe is called the transversal (or return) map of φ for R. We assume that D(φ)
e is
not empty.

3 Three crucial properties

φe satisfies three crucial properties on R, which will be crucial in our study of
homoclinic orbits.
Property 1. Let A = {1, 2, 3, ..., N } if N ≤ ∞, or N if N = ∞. Then, ∀a ∈ A,
there exist Ua and Va , which are disjoint horizontal resp. vertical strips in D(φ),
e and
φ maps Va homeomorphically to Ua , i.e.
e

φ(V
e a ) = Ua

Furthermore, the vertical boundaries of Va are mapped onto the vertical boundaries of
Ua , and similarly for the horizontal boundaries.
Definition. Given µ such that 0 < µ < 1, a curve y = u(x) is a horizontal curve if
0 ≤ u(x) ≤ 1 for 0 ≤ x ≤ 1, and

|u(x) − u(x0 )| ≤ µ |x − x0 | , 0 ≤ x ≤ x0 ≤ 1
Definition. If u1 (x), u2 (x) are two horizontal curves, and if

0 ≤ u1 (x) < u2 (x) ≤ 1

then the set

U = {(x, y)|0 ≤ x ≤ 1, u1 (x) ≤ y ≤ u2 (x)}

is called a horizontal strip
Definition. If U is a horizontal strip

d(U ) = max (u2 (x) − u1 (x))

0≤x≤1

is called the diameter of U

Very similar definitions are valid for vertical strips.
Now because φe is C 1 (because φ is), φe enjoys another useful property. To explain
this property, first we represent φe in coordinates by

 x1 = f (x0 , y0 )
y1 = g(x0 , y0 )
(x1 , y1 ) = φ(x
e 0 , y0 )


7
 Figure 7: Two horizontal curves, a horizontal strip, and its diameter

Now in particular, dφe maps the tangent vector (ξ0 , η0 ) at (x0 , y0 ) into the tangent
vector (ξ1 , η1 ) at (x1 , y1 ), with:

ξ1 = fx ξ0 + fy η0
η1 = gx ξ0 + gy η0
Property 2. There exists µ with 0 < µ < 1 such that, if we define the bundle of sectors
S + and S − as:
 +
S := {(ξ, η)| |η| ≤ µ |ξ|}
S − := {(ξ, η)| |ξ| ≤ µ |η|}
Where S + is defined over a∈A Va , and S − over a∈A Ua , then we have:
S S

(
e +) ⊆ S+
dφ(S
d−1 φ(S
e −) ⊆ S−

Moreover, if (ξ1 , η1 ) ∈ S − , and (ξ0 , η0 ) = dφe−1 (ξ1 , η1 ), then we have:

|η0 | ≥ µ−1 |η1 |

Note. Property 2 basically reflects the fact that φe is a very unstable map under itera-
tion, because notice that the horizontal components of a tangent vector get amplified
at least by µ−n under dφen (with n ≥ 1), and the vertical components by µ−n under
dφe−n .
It turns out that Properties 1 and 2 imply the following Property 3.

Property 3. As above, S let A be the set {1, 2, 3, ..., N } if N < ∞, or N if N = ∞. If V

is a vertical strip in a∈A Va , then, ∀a ∈ A, we have that:

8
 φe−1 (V ) ∩ Va = Va0
is a vertical strip (in particular nonempty), and for some ν with 0 < ν < 1:

d(Va0 ) ≤ νd(Va )
S
Similarly, if U is a horizontal strip in a∈A Ua , then, ∀a ∈ A, we have that:

e ) ∩ Ua = U 0
φ(U a

is a horizontal strip, and finally:

d(Ua0 ) ≤ νd(Ua )
Properties 1 and 3 will be crucial later on, when we will prove the main theorem of
this paper.

4 Construction of the horizontal/vertical strips

The horizontal and vertical strips in Property 1 are so important for the study of
homoclinic orbits that, in this section, we will construct them. This section is inde-
pendent of the remaining sections. Also note that, in this section, we will be mainly
concerned with our original map φ, not with the transversal map φ. e
First of all, let’s introduce again local coordinates, O being the origin of the co-
ordinate system, and U, S the coordinate axes. If (x1 , y1 ) = φ(x0 , y0 ), the image of
(x0 , y0 ) under φ, then we have, in the neighborhood of O, that:

x1 = f (x0 , y0 )
y1 = g(x0 , y0 )
with f, g continuously differentiable in this neighborhood, and f (0, y) = g(x, 0) =
0, fx (0, 0) = λ1 , gy (0, 0) = λ2 .
Let’s denote by Q the square

Q : 0 ≤ x ≤ a, 0 ≤ y ≤ a
with a > 0 small enough. Also, it is natural to denote the k th iterate of (x, y) under
φ by (xk , yk ).
Because we can choose the sides of R to be as small as we like, there exist positive
integers l, m, such that
 l
φ (r) ∈ Q
φ−m (r) ∈ Q
Fix l, m, and a, and let δ be such that 0 < δ < a, and let b, c such that 0 < b, c < a2
and φl (r) has coordinates (0, 2b) and φ−m has coordinates (2c, 0). Because we can

9
choose the sides of R to be as small as we like, and because of continuity of φ we can
assume that:
 l
φ (R) ∈ {0 ≤ x ≤ δ, b ≤ y ≤ a}
φ−m (R) ∈ {c ≤ x ≤ a, 0 ≤ y ≤ δ}
We further assume that:

A0 = φl (R)
has one of its sides coinciding with x = 0, and one adjacent side intersecting
x = 0 transversally. And, by choosing the sides of R even smaller if necessary, we
may suppose that both sides adjacent to x = 0 are of the form y = h1 (x), with h1 a
C 1 function. Similarly, we assume that:

A1 = φ−m (R)
has two sides that can be written in the form x = h2 (y), with h2 a C 1 function (see
Figure 8).
We can now define the transversal map ψ from A0 to A1 as follows:
If q ∈ A0 and ∃k > 0 such that φk (q) ∈ A1 , and ∀0 < j < k, φj (q) ∈ Q, then we say
that q ∈ D(ψ), and we set:

ψ(q) = φk (q) ∈ A1
with k the smallest such k ≥ 0.
Note. In D(ψ) ∩ D(φ),
e we clearly have

φe = φm ψφl

Now here comes a remarkable fact:

Claim. For ψ, defined in D(ψ) ∈ A0 , ψ(ψ(D)) = ψ 2 (D) intersects A1 in infinitely
many strips Uk (k = 1,2,...). These strips connect the opposite sides of A1
√ √
Proof. For sufficiently small a, we have in Q, that x1 ≥ λ1 x, y1 ≤ λ2 y, so that the
images under φk of the two curves y = h1 (x) intersect the domain {c ≤ x ≤ a, 0 ≤ y ≤ δ}
in a curve connecting x = 0 and x = a. Thus, for large k, the domains φk (A0 ) ∩ A1
will consist of infinitely many quadrilaterals U
fk connecting opposite sides of A1 (see
−1 f
Figure 9). Similarly, Vk = ψ (Uk ) are vertical strips in A0 , and by definition, we
f
have:

ψ(V
fk ) = U
fk

Now set:
(
Uk = φm (U
fk )
−l f
Vk = φ (Vk )

10
Figure 8: Construction of A0 , A1

Figure 9: Construction of Ũi , Ṽi

11
 Figure 10: Construction of Ui , Vi

and we obtain similar strips in R (see Figure 10), which satisfy:

e k ) = φm (ψ(V
φ(V fk )) = Uk

and we get our desired horizontal/vertical strips.

5 Main theorem
Here is the main theorem concerning homoclinic orbits.
Main theorem. There exists an invariant set I in R and a homeomorphism τ of S (the
set of sequences on N symbols) into I such that

φτ
e = τσ

Where σ : S → S is defined by (σ(s))n = sn−1

Proof. We will prove the case where N < ∞. The case N = ∞ will be dealt with in
the next section.
First, define inductively for n ≥ 1,

Vs0 s−1 ...s−n = Vs0 ∩ φe−1 (Vs−1 ...s−n )

Because of Property 3, and by induction, one can show that these are vertical strips.
Also, one can easily see that:

d(Vs0 s−1 ...s−n ) ≤ νd(Vs−1 ...s−n ) ≤ ν n d(Vs−n ) ≤ ν n

12
 Hence

lim d(Vs0 s−1 ...s−n ) = 0

n→∞

By the definition of Vs0 s−1 ...s−n , we have:

n o
Vs0 s−1 ...s−n e | φek (p) ∈ Vs (k = 0, 1, ..., n)
= p ∈ D(φ) −k
n o
e | φe−k (p) ∈ Vs (k = 0, −1, ..., −n)
= p ∈ D(φ) k

Hence:

Vs0 s−1 ...s−n ⊂ Vs0 s−1 ...s−n s−n+1

We need a little lemma (which follows immediately from the compacity of the
family of vertical curves):
Lemma. If V 1 ⊃ V 2 ⊃ V 3 ⊃ ... is a sequence of nested vertical strips, and if
d(V k ) → 0 as k → ∞, then
∞
\
Vk
k=1

defines a vertical curve

From this lemma, it follows that:

∞
\ n o
V (s) = e | φe−k (p) ∈ Vs (k = 0, −1, ...)
Vs0 s−1 ...s−n = p ∈ D(φ) k
n=0

defines a vertical curve, depending on the left half of the sequence s.

Similarly, define inductively, for n ≥ 2,

Us1 s2 ...sn = Us1 ∩ φ(U

e s ...s )
2 n

which, again, are nested horizontal strips whose diameters tend to 0 as k → ∞

Hence, since φ(V
e s ) = Us , we see that:
k k

T∞ n o
U (s) = Us0 s1 ...sn e | φe−k+1 (p) ∈ Us (k = 0, 1, ...)
= p ∈ D(φ)
n=1 k
n o
e | φe−k (p) ∈ Vs (k = 0, 1, ...)
= p ∈ D(φ) k

defines a horizontal curve, depending on the right half of s.

Here’s another lemma (whose proof, which uses the intermediate value theorem, is
left to the reader):
Lemma. A horizontal and a vertical curve intersect precisely in one point

13
 Hence, by the above lemma, the intersection
n o
V (s) ∩ U (s) = p ∈ D(φ) e | φe−k (p) ∈ Vs (k = 0, ±1, ±2...)
k

defines precisely one point, say w, in I (the set of points of intersections of those
horizontal and vertical curves, which is clearly invariant under φ)
e
Finally, define τ : S → I by τ (s) = w, where s = (...s1 s0 s1 ...) ∈ S
From construction, it follows that if τ (s) = w, then τ (σ(s)) = φ(w)
e
Hence

τ σ = φτ
e

And it is not very hard to prove that τ is a homeomorphism.

Note. If N < ∞, then it can be shown that τ (S) is a Cantor set. It is therefore
uncountable, but has Lebesgue measure zero.
Note. We can introduce a topology on S by taking as neighborhood basis of

s = (..., s−1 , s0 , s1 , ...)

the sets

Uj = {t ∈ S | tk = sk , |k| < j}

6 An extension
To deal with the case N = ∞, we now discuss an extension of S.
We assume that A = N and that Va is ordered according to increasing x-coordinates
with increasing a (this ordering is possible, because the Va are disjoint). Then Va tend
to a vertical curve V∞ , which agrees with x = 1. Similarly for y = 1.
Now for α, β integers satisfying α ≤ 0, β ≥ 1, let

s = (∞, sα+1 , ..., sβ−1 , ∞), sk ∈ A

Note. The case α = 0, β = 1 corresponds to the symbolic element (∞, ∞)
Now if α = −∞, β = ∞, we get elements in S. Elements with α = −∞, β < ∞
or α > −∞, β = ∞, correspond to half-infinite sequences. These sequences form the
elements of the space S ⊃ S.
Now extend σ to σ, in the domain:

D(σ) = s ∈ S, s0 6= ∞
The range Im(σ) of σ is given by:

R(σ) = s ∈ S, s1 6= ∞

14
Note. We can introduce a topology on S by taking as a neighborhood basis of s =
(..., s−1 , s0 , ..., sβ−1 , ∞), α = ∞, β < ∞ the sets

UK = t ∈ S | tk = sk , −K ≤ k < β, tβ ≥ K
The following theorem is a variation of our main theorem:

Theorem. τ can be extended to τ : S → I such that

φτ
e = τσ

if both sides are restricted to D(σ)

Note. τ S is compact with respect to a certain topology, but not anymore invariant
under φ.e The new sequences correspond to solutions which ’escape’ the homoclinic
orbit for positive or negative time.

7 Consequences
Now witness the power of the previous theorem! There are numerous corollaries,
which are very useful in the understanding of the problem. In this section, we’ll list
and prove some of them.
Corollary. Homoclinic points are dense in R

Proof. Just take sequences that end with ∞ (they correspond to homoclinic points).
Since those sequences are dense in S, we are done.
Corollary. There exist infinitely many points whose orbits are dense in R
Proof. Just take a sequence which contains all blocks of sequences. Since there are
infinitely many such sequences, we get infinitely many points whose orbits are dense
in R.
Corollary. Periodic orbits are dense in R
Proof. Given a sequence s = (..., s−1 , s0 , s1 , ...) ∈ S , consider, for any n, the periodic
sequence s0 = (sn , s−n , s−n+1 , ..., s0 , ..., sn , s−n , s−n+1 , ...) ∈ S, which is clearly
periodic, which gives a periodic sequence that is arbitrarily close to s.
Corollary. Points in R enjoy the property of topological entropy , that is, in any neigh-
borhood of r ∈ R, there exists a point r0 ∈ R such that, under iteration of φ,
e those two
points do not lie in the same neighborhood any more
Proof. Given a sequence s = (..., s−1 , s0 , s1 , ...) ∈ S that corresponds to a point r, just
take r’ to correspond to the sequence s0 = (..., s−1 , s0 , s−1 , ..., sn , s0n+1 , s0n+2 , ...), for
some appropriate n, and such that, for all k > 0, s0n+k 6= sn+k , which geometrically
means that r and r’ first lie in the same horizontal/vertical strips, but that eventually,
under iteration of φ,
e they end up in completely different horizontal/vertical strips.

15
Corollary. We have transitivity : Given two neighborhoods U1 , U2 in R, there is a
point r ∈ U1 such that r eventually lies in U2 under iteration of φe
Proof. WLOG, assume that those neighborhoods correspond to the horizontal/vertical
strips, then the sequence s = (..., s−1 , s0 , s1 , s1 , s1 , ...) ∈ S, with s1 6= s0 corresponds
to a point that first lies in the strip V0 , but then lies in the strip U1 , and stays in that strip
for an indefinite amount of time.
Corollary. Taking the definition of Smale [2], φe is chaotic on R.
Proof. Follows directly from the fact that periodic orbits are dense, from topological
entropy, and from transitivity.

Corollary. φ does not possess a real analytic integral in R

Definition. Given a diffeomorphism φ in a domain D ⊂ R2 , a real-valued function f
in D is an integral of φ if f is not constant, and f (φ(p)) = f (p)
Proof. Suppose f is a continuous function on D(φ) e such that f (φ(p)) = f (p). Hence
k k
∀k > 0, f (φ (p)) = f (p). Choosing φ (p) to be a dense orbit (this is possible by one
of the above corollaries), we get, by continuity, that f (q) = f (p), for all q ∈ I.
Now assume that f ∈ C 1 (R), such that f (φ(p)) = f (p) on I. We’ll show that fx =
fy = 0 on I.
In fact, if p∗ = τ (s∗ ) is an arbitrary point in I, then by the first of the claims in the next
section, p∗ lies on a horizontal curve U (s∗ ) and vertical curve V (s∗ ). By modifying the
left tail of the sequence s∗ , we can choose a sequence of points qν ∈ U (s∗ )∩I, q 6= p∗ ,
converging to p∗ . Thus, if we parametrize U (s∗ ) by x, the directional derivative Dx f
of f along the tangent of U (s∗ ) at p∗ can be expressed limit of a difference quotient of
f at p∗ .
Since f (p) = f (qν ), it follows that Dx (f ) = 0 at p∗ , and by the same argument, the
directional derivative Dy (f ) along the tangent of V (s∗ ) vanishes.
Since Dx (f ), Dy (f ) are linearly independent, we conclude that fx = fy = 0 on I.
And if f ∈ C ∞ (Q), then repeat this argument, and conclude that all derivatives of f
vanish on I.
Thus, if f is real analytic, it is the zero function.

8 Further properties
The invariant set I = τ (S) enjoys further geometric properties, which we will
state, but not prove. For this, we need a definition:
Definition. A set P is said to be hyperbolic if one can associate with every point p ∈ P
− −
two linearly independent lines L+ +
p , Lp in the tangent space at p, such that Lp , Lp vary
continuously with p ∈ P , and, furthermore

16
 ±
dφL±
p = Lφ(p)

and that, with some constant λ > 1, and the norm |ζ| = max (|ξ|, |η|), we have:

forζ ∈ L+

|dφ(ζ)| ≥ λ|ζ| p
|dφ−1 (ζ)| ≥ λ|ζ| forζ ∈ L− p

Let ∆ be the Jacobian determinant of dφ at a point p. Also recall the definition of

µ in Property 3.

Claim. If ∆, ∆−1 ≤ 12 µ−2 , then I is a hyperbolic set

Even more is true:
1 1
Claim. If 0 ≤ µ ≤ 12 min(|∆| 2 , |∆|− 2 ), then U(s), V(s) are continuously differntiable
curves whose tangents at points of I agree with the lines of the hyperbolic structure.

9 The restricted 3-body problem

The same study we’ve done in this paper can be modified to analyze to a very
beautiful physical phenomenon: the restricted 3-body problem. In this section as well,
we will state the main results, but not prove them. The setting is as follows:
Suppose we have 2 mass points of equal mass m1 = m2 = 21 , moving under
Newton’s law of attraction in the elliptic orbits while the center of mass is at rest. Now
consider a third mass point, of mass m3 = 0, moving on the line L perpendicular to the
plane of motion s of the first two mass points, and going through the center of mass.
Since m3 = 0, it is clear that the third mass point will remain on L. The problem is to
describe the motion of the third mass point (see Figure 11). To this end, let z be the
coordinate describing the third mass point on L, and t the time parameter. Then we
have the following differential equation:

d2 z z
2
=− 3
dt (z + r2 (t)) 2
2

with
1
r(t) = (1 − cos(t)) + O(2 )
2
and  denotes the eccentricity.
˜
Theorem. There exists a critical escape velocity ż(0) ˜
such that, if ż(0) > ż(0), the
corresponding solutions will tend to infinity
˜
In this section, we will study the solutions near the critical escape velocity ż(0).
Consider a solution z(t) with infinitely many zeroes tk (k = 0, ±1, ±2, ...), such
that tk < tk+1 , then let:

17
 Figure 11: The restricted 3-body problem

 
tk+1 − tk
sk =
2π
Basically, sk measures the number of complete revolutions of the primaries be-
tween two zeroes of z(t). Hence, we can associate to every such solution a doubly
infinite sequence of integers.
The surprising fact is that the converse holds too:
Theorem. For all but finitely many eccentricites 0 <  < 1, there exists an integer
m = m() such that any sequence s with sk ≥ m corresponds to a solution of the
above differential equation.

Any solution z(t) 6= 0 is determined by giving its velocity z˙0 and time t0 when
z(t0 ) = 0. Such a zero always exists, otherwise, if z(t) > 0 for all t, hence z¨0 < 0, so
z(t) is concave for all t, and positive, which is impossible.
Hence, we can describe an arbitrary orbit by giving t0 [2π] and z˙0 . Since the differential
equation is invariant under the reflection z 7→ −z, we may set v0 = |ż(t0 )|. Now we
can define a map φ, which maps (v0 , t0 ) to (v1 , t1 ), where t1 is the next zero (if it
exists), and v1 = |ż(t1 )|.
Note. The point (0, ∞) can be viewed in this case as a hyperbolic fixed point of φ, thus
the importance of studying solutions that blow off to infinity.
We have the following very important lemmas:
Lemma. There exists a real analytic simple closed curve in R2 in whose interior D0
the mapping φ is defined. For (v0 , t0 ) outside D0 , the corresponding solutions escape.

18
Lemma. φ maps D0 onto a domain D1 = φ(D0 ), which agrees with the reflected
domain D1 = ρ(D0 ), where ρ(v, t) = (v, −t) (see Figure 12). Moreover, φ preserves
the area element vdvdt, and:

φ−1 = ρ−1 φρ

Lemma. If  > 0 is small enough, then D0 6= D1 , and the boundary curves ∂D0 , ∂D1
intersect on the symmetry line nontangentially at a point P (see Figure 12).
Note. P plays the same role as a homoclinic point

Figure 12: Construction of D0 , D1 , P

Here we come with the main theorem concerning the restricted 3-body-problem:
Theorem. There exists a homeomorphism τ of S into D0 such that

φτ = τ σ̄

Moreover, there exists a hyperbolic invariant set I homeomorphic to S on which φ is

topologically equivalent to σ
The idea of the proof is the similar as for the main theorem: Just consider a rectan-
gle R around P, and verify Properties 1 and 3. In this case, the horizontal strips are just
the elements of φ(R) ∩ R, and the vertical strips are the reflections of those horizontal
strips.
We conclude that the movement of the third mass point is very chaotic.

Conclusion
In order to study the dynamics near a homoclinic point H, we introduced a transver-
sal map φ.
e This map satisfies 3 properties, which led us to the main theorem. This
theorem reduced our study to a simple analysis of sequences, which helped us to prove

19
corollaries that seem too good to be true. For example, among those corollaries, we
have that homoclinic orbits form a chaotic system (according to Smale), and that they
do not possess an analytic integral, which completely rules out any regular/nonchaotic
behavior. Thus, Moser’s paper opened some windows in the understanding of chaos.
But nonetheless, this field of study has not yet been completely understood, and will
probably one of the hot topics of mathematics in the 21st century.

References
[1] Ivar Ekeland. Le Calcul, l’Imprévu : les figures du temps de Kepler á Thom. Paris:
Editions du Seuil, 1984.
[2] Stephen Smale et.al. Differential Equations, Dynamical Systems, and an Introduc-
tion to Chaos. Elsevier Academic Press, 2004.

[3] Jurgen Moser. Stable and Random Motions in Dynamical Systems. Princeton
University Press, 2001.

20